linux/drivers/md/raid5.c
<<
>>
Prefs
   1// SPDX-License-Identifier: GPL-2.0-or-later
   2/*
   3 * raid5.c : Multiple Devices driver for Linux
   4 *         Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   5 *         Copyright (C) 1999, 2000 Ingo Molnar
   6 *         Copyright (C) 2002, 2003 H. Peter Anvin
   7 *
   8 * RAID-4/5/6 management functions.
   9 * Thanks to Penguin Computing for making the RAID-6 development possible
  10 * by donating a test server!
  11 */
  12
  13/*
  14 * BITMAP UNPLUGGING:
  15 *
  16 * The sequencing for updating the bitmap reliably is a little
  17 * subtle (and I got it wrong the first time) so it deserves some
  18 * explanation.
  19 *
  20 * We group bitmap updates into batches.  Each batch has a number.
  21 * We may write out several batches at once, but that isn't very important.
  22 * conf->seq_write is the number of the last batch successfully written.
  23 * conf->seq_flush is the number of the last batch that was closed to
  24 *    new additions.
  25 * When we discover that we will need to write to any block in a stripe
  26 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
  27 * the number of the batch it will be in. This is seq_flush+1.
  28 * When we are ready to do a write, if that batch hasn't been written yet,
  29 *   we plug the array and queue the stripe for later.
  30 * When an unplug happens, we increment bm_flush, thus closing the current
  31 *   batch.
  32 * When we notice that bm_flush > bm_write, we write out all pending updates
  33 * to the bitmap, and advance bm_write to where bm_flush was.
  34 * This may occasionally write a bit out twice, but is sure never to
  35 * miss any bits.
  36 */
  37
  38#include <linux/blkdev.h>
  39#include <linux/kthread.h>
  40#include <linux/raid/pq.h>
  41#include <linux/async_tx.h>
  42#include <linux/module.h>
  43#include <linux/async.h>
  44#include <linux/seq_file.h>
  45#include <linux/cpu.h>
  46#include <linux/slab.h>
  47#include <linux/ratelimit.h>
  48#include <linux/nodemask.h>
  49
  50#include <trace/events/block.h>
  51#include <linux/list_sort.h>
  52
  53#include "md.h"
  54#include "raid5.h"
  55#include "raid0.h"
  56#include "md-bitmap.h"
  57#include "raid5-log.h"
  58
  59#define UNSUPPORTED_MDDEV_FLAGS (1L << MD_FAILFAST_SUPPORTED)
  60
  61#define cpu_to_group(cpu) cpu_to_node(cpu)
  62#define ANY_GROUP NUMA_NO_NODE
  63
  64static bool devices_handle_discard_safely = false;
  65module_param(devices_handle_discard_safely, bool, 0644);
  66MODULE_PARM_DESC(devices_handle_discard_safely,
  67                 "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
  68static struct workqueue_struct *raid5_wq;
  69
  70static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
  71{
  72        int hash = (sect >> RAID5_STRIPE_SHIFT(conf)) & HASH_MASK;
  73        return &conf->stripe_hashtbl[hash];
  74}
  75
  76static inline int stripe_hash_locks_hash(struct r5conf *conf, sector_t sect)
  77{
  78        return (sect >> RAID5_STRIPE_SHIFT(conf)) & STRIPE_HASH_LOCKS_MASK;
  79}
  80
  81static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
  82{
  83        spin_lock_irq(conf->hash_locks + hash);
  84        spin_lock(&conf->device_lock);
  85}
  86
  87static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
  88{
  89        spin_unlock(&conf->device_lock);
  90        spin_unlock_irq(conf->hash_locks + hash);
  91}
  92
  93static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
  94{
  95        int i;
  96        spin_lock_irq(conf->hash_locks);
  97        for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
  98                spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
  99        spin_lock(&conf->device_lock);
 100}
 101
 102static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
 103{
 104        int i;
 105        spin_unlock(&conf->device_lock);
 106        for (i = NR_STRIPE_HASH_LOCKS - 1; i; i--)
 107                spin_unlock(conf->hash_locks + i);
 108        spin_unlock_irq(conf->hash_locks);
 109}
 110
 111/* Find first data disk in a raid6 stripe */
 112static inline int raid6_d0(struct stripe_head *sh)
 113{
 114        if (sh->ddf_layout)
 115                /* ddf always start from first device */
 116                return 0;
 117        /* md starts just after Q block */
 118        if (sh->qd_idx == sh->disks - 1)
 119                return 0;
 120        else
 121                return sh->qd_idx + 1;
 122}
 123static inline int raid6_next_disk(int disk, int raid_disks)
 124{
 125        disk++;
 126        return (disk < raid_disks) ? disk : 0;
 127}
 128
 129/* When walking through the disks in a raid5, starting at raid6_d0,
 130 * We need to map each disk to a 'slot', where the data disks are slot
 131 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
 132 * is raid_disks-1.  This help does that mapping.
 133 */
 134static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
 135                             int *count, int syndrome_disks)
 136{
 137        int slot = *count;
 138
 139        if (sh->ddf_layout)
 140                (*count)++;
 141        if (idx == sh->pd_idx)
 142                return syndrome_disks;
 143        if (idx == sh->qd_idx)
 144                return syndrome_disks + 1;
 145        if (!sh->ddf_layout)
 146                (*count)++;
 147        return slot;
 148}
 149
 150static void print_raid5_conf (struct r5conf *conf);
 151
 152static int stripe_operations_active(struct stripe_head *sh)
 153{
 154        return sh->check_state || sh->reconstruct_state ||
 155               test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
 156               test_bit(STRIPE_COMPUTE_RUN, &sh->state);
 157}
 158
 159static bool stripe_is_lowprio(struct stripe_head *sh)
 160{
 161        return (test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) ||
 162                test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) &&
 163               !test_bit(STRIPE_R5C_CACHING, &sh->state);
 164}
 165
 166static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
 167{
 168        struct r5conf *conf = sh->raid_conf;
 169        struct r5worker_group *group;
 170        int thread_cnt;
 171        int i, cpu = sh->cpu;
 172
 173        if (!cpu_online(cpu)) {
 174                cpu = cpumask_any(cpu_online_mask);
 175                sh->cpu = cpu;
 176        }
 177
 178        if (list_empty(&sh->lru)) {
 179                struct r5worker_group *group;
 180                group = conf->worker_groups + cpu_to_group(cpu);
 181                if (stripe_is_lowprio(sh))
 182                        list_add_tail(&sh->lru, &group->loprio_list);
 183                else
 184                        list_add_tail(&sh->lru, &group->handle_list);
 185                group->stripes_cnt++;
 186                sh->group = group;
 187        }
 188
 189        if (conf->worker_cnt_per_group == 0) {
 190                md_wakeup_thread(conf->mddev->thread);
 191                return;
 192        }
 193
 194        group = conf->worker_groups + cpu_to_group(sh->cpu);
 195
 196        group->workers[0].working = true;
 197        /* at least one worker should run to avoid race */
 198        queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
 199
 200        thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
 201        /* wakeup more workers */
 202        for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
 203                if (group->workers[i].working == false) {
 204                        group->workers[i].working = true;
 205                        queue_work_on(sh->cpu, raid5_wq,
 206                                      &group->workers[i].work);
 207                        thread_cnt--;
 208                }
 209        }
 210}
 211
 212static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
 213                              struct list_head *temp_inactive_list)
 214{
 215        int i;
 216        int injournal = 0;      /* number of date pages with R5_InJournal */
 217
 218        BUG_ON(!list_empty(&sh->lru));
 219        BUG_ON(atomic_read(&conf->active_stripes)==0);
 220
 221        if (r5c_is_writeback(conf->log))
 222                for (i = sh->disks; i--; )
 223                        if (test_bit(R5_InJournal, &sh->dev[i].flags))
 224                                injournal++;
 225        /*
 226         * In the following cases, the stripe cannot be released to cached
 227         * lists. Therefore, we make the stripe write out and set
 228         * STRIPE_HANDLE:
 229         *   1. when quiesce in r5c write back;
 230         *   2. when resync is requested fot the stripe.
 231         */
 232        if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) ||
 233            (conf->quiesce && r5c_is_writeback(conf->log) &&
 234             !test_bit(STRIPE_HANDLE, &sh->state) && injournal != 0)) {
 235                if (test_bit(STRIPE_R5C_CACHING, &sh->state))
 236                        r5c_make_stripe_write_out(sh);
 237                set_bit(STRIPE_HANDLE, &sh->state);
 238        }
 239
 240        if (test_bit(STRIPE_HANDLE, &sh->state)) {
 241                if (test_bit(STRIPE_DELAYED, &sh->state) &&
 242                    !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 243                        list_add_tail(&sh->lru, &conf->delayed_list);
 244                else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
 245                           sh->bm_seq - conf->seq_write > 0)
 246                        list_add_tail(&sh->lru, &conf->bitmap_list);
 247                else {
 248                        clear_bit(STRIPE_DELAYED, &sh->state);
 249                        clear_bit(STRIPE_BIT_DELAY, &sh->state);
 250                        if (conf->worker_cnt_per_group == 0) {
 251                                if (stripe_is_lowprio(sh))
 252                                        list_add_tail(&sh->lru,
 253                                                        &conf->loprio_list);
 254                                else
 255                                        list_add_tail(&sh->lru,
 256                                                        &conf->handle_list);
 257                        } else {
 258                                raid5_wakeup_stripe_thread(sh);
 259                                return;
 260                        }
 261                }
 262                md_wakeup_thread(conf->mddev->thread);
 263        } else {
 264                BUG_ON(stripe_operations_active(sh));
 265                if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 266                        if (atomic_dec_return(&conf->preread_active_stripes)
 267                            < IO_THRESHOLD)
 268                                md_wakeup_thread(conf->mddev->thread);
 269                atomic_dec(&conf->active_stripes);
 270                if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
 271                        if (!r5c_is_writeback(conf->log))
 272                                list_add_tail(&sh->lru, temp_inactive_list);
 273                        else {
 274                                WARN_ON(test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags));
 275                                if (injournal == 0)
 276                                        list_add_tail(&sh->lru, temp_inactive_list);
 277                                else if (injournal == conf->raid_disks - conf->max_degraded) {
 278                                        /* full stripe */
 279                                        if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE, &sh->state))
 280                                                atomic_inc(&conf->r5c_cached_full_stripes);
 281                                        if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
 282                                                atomic_dec(&conf->r5c_cached_partial_stripes);
 283                                        list_add_tail(&sh->lru, &conf->r5c_full_stripe_list);
 284                                        r5c_check_cached_full_stripe(conf);
 285                                } else
 286                                        /*
 287                                         * STRIPE_R5C_PARTIAL_STRIPE is set in
 288                                         * r5c_try_caching_write(). No need to
 289                                         * set it again.
 290                                         */
 291                                        list_add_tail(&sh->lru, &conf->r5c_partial_stripe_list);
 292                        }
 293                }
 294        }
 295}
 296
 297static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
 298                             struct list_head *temp_inactive_list)
 299{
 300        if (atomic_dec_and_test(&sh->count))
 301                do_release_stripe(conf, sh, temp_inactive_list);
 302}
 303
 304/*
 305 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
 306 *
 307 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
 308 * given time. Adding stripes only takes device lock, while deleting stripes
 309 * only takes hash lock.
 310 */
 311static void release_inactive_stripe_list(struct r5conf *conf,
 312                                         struct list_head *temp_inactive_list,
 313                                         int hash)
 314{
 315        int size;
 316        bool do_wakeup = false;
 317        unsigned long flags;
 318
 319        if (hash == NR_STRIPE_HASH_LOCKS) {
 320                size = NR_STRIPE_HASH_LOCKS;
 321                hash = NR_STRIPE_HASH_LOCKS - 1;
 322        } else
 323                size = 1;
 324        while (size) {
 325                struct list_head *list = &temp_inactive_list[size - 1];
 326
 327                /*
 328                 * We don't hold any lock here yet, raid5_get_active_stripe() might
 329                 * remove stripes from the list
 330                 */
 331                if (!list_empty_careful(list)) {
 332                        spin_lock_irqsave(conf->hash_locks + hash, flags);
 333                        if (list_empty(conf->inactive_list + hash) &&
 334                            !list_empty(list))
 335                                atomic_dec(&conf->empty_inactive_list_nr);
 336                        list_splice_tail_init(list, conf->inactive_list + hash);
 337                        do_wakeup = true;
 338                        spin_unlock_irqrestore(conf->hash_locks + hash, flags);
 339                }
 340                size--;
 341                hash--;
 342        }
 343
 344        if (do_wakeup) {
 345                wake_up(&conf->wait_for_stripe);
 346                if (atomic_read(&conf->active_stripes) == 0)
 347                        wake_up(&conf->wait_for_quiescent);
 348                if (conf->retry_read_aligned)
 349                        md_wakeup_thread(conf->mddev->thread);
 350        }
 351}
 352
 353/* should hold conf->device_lock already */
 354static int release_stripe_list(struct r5conf *conf,
 355                               struct list_head *temp_inactive_list)
 356{
 357        struct stripe_head *sh, *t;
 358        int count = 0;
 359        struct llist_node *head;
 360
 361        head = llist_del_all(&conf->released_stripes);
 362        head = llist_reverse_order(head);
 363        llist_for_each_entry_safe(sh, t, head, release_list) {
 364                int hash;
 365
 366                /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
 367                smp_mb();
 368                clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
 369                /*
 370                 * Don't worry the bit is set here, because if the bit is set
 371                 * again, the count is always > 1. This is true for
 372                 * STRIPE_ON_UNPLUG_LIST bit too.
 373                 */
 374                hash = sh->hash_lock_index;
 375                __release_stripe(conf, sh, &temp_inactive_list[hash]);
 376                count++;
 377        }
 378
 379        return count;
 380}
 381
 382void raid5_release_stripe(struct stripe_head *sh)
 383{
 384        struct r5conf *conf = sh->raid_conf;
 385        unsigned long flags;
 386        struct list_head list;
 387        int hash;
 388        bool wakeup;
 389
 390        /* Avoid release_list until the last reference.
 391         */
 392        if (atomic_add_unless(&sh->count, -1, 1))
 393                return;
 394
 395        if (unlikely(!conf->mddev->thread) ||
 396                test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
 397                goto slow_path;
 398        wakeup = llist_add(&sh->release_list, &conf->released_stripes);
 399        if (wakeup)
 400                md_wakeup_thread(conf->mddev->thread);
 401        return;
 402slow_path:
 403        /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
 404        if (atomic_dec_and_lock_irqsave(&sh->count, &conf->device_lock, flags)) {
 405                INIT_LIST_HEAD(&list);
 406                hash = sh->hash_lock_index;
 407                do_release_stripe(conf, sh, &list);
 408                spin_unlock_irqrestore(&conf->device_lock, flags);
 409                release_inactive_stripe_list(conf, &list, hash);
 410        }
 411}
 412
 413static inline void remove_hash(struct stripe_head *sh)
 414{
 415        pr_debug("remove_hash(), stripe %llu\n",
 416                (unsigned long long)sh->sector);
 417
 418        hlist_del_init(&sh->hash);
 419}
 420
 421static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
 422{
 423        struct hlist_head *hp = stripe_hash(conf, sh->sector);
 424
 425        pr_debug("insert_hash(), stripe %llu\n",
 426                (unsigned long long)sh->sector);
 427
 428        hlist_add_head(&sh->hash, hp);
 429}
 430
 431/* find an idle stripe, make sure it is unhashed, and return it. */
 432static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
 433{
 434        struct stripe_head *sh = NULL;
 435        struct list_head *first;
 436
 437        if (list_empty(conf->inactive_list + hash))
 438                goto out;
 439        first = (conf->inactive_list + hash)->next;
 440        sh = list_entry(first, struct stripe_head, lru);
 441        list_del_init(first);
 442        remove_hash(sh);
 443        atomic_inc(&conf->active_stripes);
 444        BUG_ON(hash != sh->hash_lock_index);
 445        if (list_empty(conf->inactive_list + hash))
 446                atomic_inc(&conf->empty_inactive_list_nr);
 447out:
 448        return sh;
 449}
 450
 451#if PAGE_SIZE != DEFAULT_STRIPE_SIZE
 452static void free_stripe_pages(struct stripe_head *sh)
 453{
 454        int i;
 455        struct page *p;
 456
 457        /* Have not allocate page pool */
 458        if (!sh->pages)
 459                return;
 460
 461        for (i = 0; i < sh->nr_pages; i++) {
 462                p = sh->pages[i];
 463                if (p)
 464                        put_page(p);
 465                sh->pages[i] = NULL;
 466        }
 467}
 468
 469static int alloc_stripe_pages(struct stripe_head *sh, gfp_t gfp)
 470{
 471        int i;
 472        struct page *p;
 473
 474        for (i = 0; i < sh->nr_pages; i++) {
 475                /* The page have allocated. */
 476                if (sh->pages[i])
 477                        continue;
 478
 479                p = alloc_page(gfp);
 480                if (!p) {
 481                        free_stripe_pages(sh);
 482                        return -ENOMEM;
 483                }
 484                sh->pages[i] = p;
 485        }
 486        return 0;
 487}
 488
 489static int
 490init_stripe_shared_pages(struct stripe_head *sh, struct r5conf *conf, int disks)
 491{
 492        int nr_pages, cnt;
 493
 494        if (sh->pages)
 495                return 0;
 496
 497        /* Each of the sh->dev[i] need one conf->stripe_size */
 498        cnt = PAGE_SIZE / conf->stripe_size;
 499        nr_pages = (disks + cnt - 1) / cnt;
 500
 501        sh->pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
 502        if (!sh->pages)
 503                return -ENOMEM;
 504        sh->nr_pages = nr_pages;
 505        sh->stripes_per_page = cnt;
 506        return 0;
 507}
 508#endif
 509
 510static void shrink_buffers(struct stripe_head *sh)
 511{
 512        int i;
 513        int num = sh->raid_conf->pool_size;
 514
 515#if PAGE_SIZE == DEFAULT_STRIPE_SIZE
 516        for (i = 0; i < num ; i++) {
 517                struct page *p;
 518
 519                WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
 520                p = sh->dev[i].page;
 521                if (!p)
 522                        continue;
 523                sh->dev[i].page = NULL;
 524                put_page(p);
 525        }
 526#else
 527        for (i = 0; i < num; i++)
 528                sh->dev[i].page = NULL;
 529        free_stripe_pages(sh); /* Free pages */
 530#endif
 531}
 532
 533static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
 534{
 535        int i;
 536        int num = sh->raid_conf->pool_size;
 537
 538#if PAGE_SIZE == DEFAULT_STRIPE_SIZE
 539        for (i = 0; i < num; i++) {
 540                struct page *page;
 541
 542                if (!(page = alloc_page(gfp))) {
 543                        return 1;
 544                }
 545                sh->dev[i].page = page;
 546                sh->dev[i].orig_page = page;
 547                sh->dev[i].offset = 0;
 548        }
 549#else
 550        if (alloc_stripe_pages(sh, gfp))
 551                return -ENOMEM;
 552
 553        for (i = 0; i < num; i++) {
 554                sh->dev[i].page = raid5_get_dev_page(sh, i);
 555                sh->dev[i].orig_page = sh->dev[i].page;
 556                sh->dev[i].offset = raid5_get_page_offset(sh, i);
 557        }
 558#endif
 559        return 0;
 560}
 561
 562static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
 563                            struct stripe_head *sh);
 564
 565static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
 566{
 567        struct r5conf *conf = sh->raid_conf;
 568        int i, seq;
 569
 570        BUG_ON(atomic_read(&sh->count) != 0);
 571        BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
 572        BUG_ON(stripe_operations_active(sh));
 573        BUG_ON(sh->batch_head);
 574
 575        pr_debug("init_stripe called, stripe %llu\n",
 576                (unsigned long long)sector);
 577retry:
 578        seq = read_seqcount_begin(&conf->gen_lock);
 579        sh->generation = conf->generation - previous;
 580        sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
 581        sh->sector = sector;
 582        stripe_set_idx(sector, conf, previous, sh);
 583        sh->state = 0;
 584
 585        for (i = sh->disks; i--; ) {
 586                struct r5dev *dev = &sh->dev[i];
 587
 588                if (dev->toread || dev->read || dev->towrite || dev->written ||
 589                    test_bit(R5_LOCKED, &dev->flags)) {
 590                        pr_err("sector=%llx i=%d %p %p %p %p %d\n",
 591                               (unsigned long long)sh->sector, i, dev->toread,
 592                               dev->read, dev->towrite, dev->written,
 593                               test_bit(R5_LOCKED, &dev->flags));
 594                        WARN_ON(1);
 595                }
 596                dev->flags = 0;
 597                dev->sector = raid5_compute_blocknr(sh, i, previous);
 598        }
 599        if (read_seqcount_retry(&conf->gen_lock, seq))
 600                goto retry;
 601        sh->overwrite_disks = 0;
 602        insert_hash(conf, sh);
 603        sh->cpu = smp_processor_id();
 604        set_bit(STRIPE_BATCH_READY, &sh->state);
 605}
 606
 607static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
 608                                         short generation)
 609{
 610        struct stripe_head *sh;
 611
 612        pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
 613        hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
 614                if (sh->sector == sector && sh->generation == generation)
 615                        return sh;
 616        pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
 617        return NULL;
 618}
 619
 620/*
 621 * Need to check if array has failed when deciding whether to:
 622 *  - start an array
 623 *  - remove non-faulty devices
 624 *  - add a spare
 625 *  - allow a reshape
 626 * This determination is simple when no reshape is happening.
 627 * However if there is a reshape, we need to carefully check
 628 * both the before and after sections.
 629 * This is because some failed devices may only affect one
 630 * of the two sections, and some non-in_sync devices may
 631 * be insync in the section most affected by failed devices.
 632 */
 633int raid5_calc_degraded(struct r5conf *conf)
 634{
 635        int degraded, degraded2;
 636        int i;
 637
 638        rcu_read_lock();
 639        degraded = 0;
 640        for (i = 0; i < conf->previous_raid_disks; i++) {
 641                struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 642                if (rdev && test_bit(Faulty, &rdev->flags))
 643                        rdev = rcu_dereference(conf->disks[i].replacement);
 644                if (!rdev || test_bit(Faulty, &rdev->flags))
 645                        degraded++;
 646                else if (test_bit(In_sync, &rdev->flags))
 647                        ;
 648                else
 649                        /* not in-sync or faulty.
 650                         * If the reshape increases the number of devices,
 651                         * this is being recovered by the reshape, so
 652                         * this 'previous' section is not in_sync.
 653                         * If the number of devices is being reduced however,
 654                         * the device can only be part of the array if
 655                         * we are reverting a reshape, so this section will
 656                         * be in-sync.
 657                         */
 658                        if (conf->raid_disks >= conf->previous_raid_disks)
 659                                degraded++;
 660        }
 661        rcu_read_unlock();
 662        if (conf->raid_disks == conf->previous_raid_disks)
 663                return degraded;
 664        rcu_read_lock();
 665        degraded2 = 0;
 666        for (i = 0; i < conf->raid_disks; i++) {
 667                struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 668                if (rdev && test_bit(Faulty, &rdev->flags))
 669                        rdev = rcu_dereference(conf->disks[i].replacement);
 670                if (!rdev || test_bit(Faulty, &rdev->flags))
 671                        degraded2++;
 672                else if (test_bit(In_sync, &rdev->flags))
 673                        ;
 674                else
 675                        /* not in-sync or faulty.
 676                         * If reshape increases the number of devices, this
 677                         * section has already been recovered, else it
 678                         * almost certainly hasn't.
 679                         */
 680                        if (conf->raid_disks <= conf->previous_raid_disks)
 681                                degraded2++;
 682        }
 683        rcu_read_unlock();
 684        if (degraded2 > degraded)
 685                return degraded2;
 686        return degraded;
 687}
 688
 689static int has_failed(struct r5conf *conf)
 690{
 691        int degraded;
 692
 693        if (conf->mddev->reshape_position == MaxSector)
 694                return conf->mddev->degraded > conf->max_degraded;
 695
 696        degraded = raid5_calc_degraded(conf);
 697        if (degraded > conf->max_degraded)
 698                return 1;
 699        return 0;
 700}
 701
 702struct stripe_head *
 703raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
 704                        int previous, int noblock, int noquiesce)
 705{
 706        struct stripe_head *sh;
 707        int hash = stripe_hash_locks_hash(conf, sector);
 708        int inc_empty_inactive_list_flag;
 709
 710        pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
 711
 712        spin_lock_irq(conf->hash_locks + hash);
 713
 714        do {
 715                wait_event_lock_irq(conf->wait_for_quiescent,
 716                                    conf->quiesce == 0 || noquiesce,
 717                                    *(conf->hash_locks + hash));
 718                sh = __find_stripe(conf, sector, conf->generation - previous);
 719                if (!sh) {
 720                        if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
 721                                sh = get_free_stripe(conf, hash);
 722                                if (!sh && !test_bit(R5_DID_ALLOC,
 723                                                     &conf->cache_state))
 724                                        set_bit(R5_ALLOC_MORE,
 725                                                &conf->cache_state);
 726                        }
 727                        if (noblock && sh == NULL)
 728                                break;
 729
 730                        r5c_check_stripe_cache_usage(conf);
 731                        if (!sh) {
 732                                set_bit(R5_INACTIVE_BLOCKED,
 733                                        &conf->cache_state);
 734                                r5l_wake_reclaim(conf->log, 0);
 735                                wait_event_lock_irq(
 736                                        conf->wait_for_stripe,
 737                                        !list_empty(conf->inactive_list + hash) &&
 738                                        (atomic_read(&conf->active_stripes)
 739                                         < (conf->max_nr_stripes * 3 / 4)
 740                                         || !test_bit(R5_INACTIVE_BLOCKED,
 741                                                      &conf->cache_state)),
 742                                        *(conf->hash_locks + hash));
 743                                clear_bit(R5_INACTIVE_BLOCKED,
 744                                          &conf->cache_state);
 745                        } else {
 746                                init_stripe(sh, sector, previous);
 747                                atomic_inc(&sh->count);
 748                        }
 749                } else if (!atomic_inc_not_zero(&sh->count)) {
 750                        spin_lock(&conf->device_lock);
 751                        if (!atomic_read(&sh->count)) {
 752                                if (!test_bit(STRIPE_HANDLE, &sh->state))
 753                                        atomic_inc(&conf->active_stripes);
 754                                BUG_ON(list_empty(&sh->lru) &&
 755                                       !test_bit(STRIPE_EXPANDING, &sh->state));
 756                                inc_empty_inactive_list_flag = 0;
 757                                if (!list_empty(conf->inactive_list + hash))
 758                                        inc_empty_inactive_list_flag = 1;
 759                                list_del_init(&sh->lru);
 760                                if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
 761                                        atomic_inc(&conf->empty_inactive_list_nr);
 762                                if (sh->group) {
 763                                        sh->group->stripes_cnt--;
 764                                        sh->group = NULL;
 765                                }
 766                        }
 767                        atomic_inc(&sh->count);
 768                        spin_unlock(&conf->device_lock);
 769                }
 770        } while (sh == NULL);
 771
 772        spin_unlock_irq(conf->hash_locks + hash);
 773        return sh;
 774}
 775
 776static bool is_full_stripe_write(struct stripe_head *sh)
 777{
 778        BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
 779        return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
 780}
 781
 782static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 783                __acquires(&sh1->stripe_lock)
 784                __acquires(&sh2->stripe_lock)
 785{
 786        if (sh1 > sh2) {
 787                spin_lock_irq(&sh2->stripe_lock);
 788                spin_lock_nested(&sh1->stripe_lock, 1);
 789        } else {
 790                spin_lock_irq(&sh1->stripe_lock);
 791                spin_lock_nested(&sh2->stripe_lock, 1);
 792        }
 793}
 794
 795static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 796                __releases(&sh1->stripe_lock)
 797                __releases(&sh2->stripe_lock)
 798{
 799        spin_unlock(&sh1->stripe_lock);
 800        spin_unlock_irq(&sh2->stripe_lock);
 801}
 802
 803/* Only freshly new full stripe normal write stripe can be added to a batch list */
 804static bool stripe_can_batch(struct stripe_head *sh)
 805{
 806        struct r5conf *conf = sh->raid_conf;
 807
 808        if (raid5_has_log(conf) || raid5_has_ppl(conf))
 809                return false;
 810        return test_bit(STRIPE_BATCH_READY, &sh->state) &&
 811                !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
 812                is_full_stripe_write(sh);
 813}
 814
 815/* we only do back search */
 816static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
 817{
 818        struct stripe_head *head;
 819        sector_t head_sector, tmp_sec;
 820        int hash;
 821        int dd_idx;
 822        int inc_empty_inactive_list_flag;
 823
 824        /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
 825        tmp_sec = sh->sector;
 826        if (!sector_div(tmp_sec, conf->chunk_sectors))
 827                return;
 828        head_sector = sh->sector - RAID5_STRIPE_SECTORS(conf);
 829
 830        hash = stripe_hash_locks_hash(conf, head_sector);
 831        spin_lock_irq(conf->hash_locks + hash);
 832        head = __find_stripe(conf, head_sector, conf->generation);
 833        if (head && !atomic_inc_not_zero(&head->count)) {
 834                spin_lock(&conf->device_lock);
 835                if (!atomic_read(&head->count)) {
 836                        if (!test_bit(STRIPE_HANDLE, &head->state))
 837                                atomic_inc(&conf->active_stripes);
 838                        BUG_ON(list_empty(&head->lru) &&
 839                               !test_bit(STRIPE_EXPANDING, &head->state));
 840                        inc_empty_inactive_list_flag = 0;
 841                        if (!list_empty(conf->inactive_list + hash))
 842                                inc_empty_inactive_list_flag = 1;
 843                        list_del_init(&head->lru);
 844                        if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
 845                                atomic_inc(&conf->empty_inactive_list_nr);
 846                        if (head->group) {
 847                                head->group->stripes_cnt--;
 848                                head->group = NULL;
 849                        }
 850                }
 851                atomic_inc(&head->count);
 852                spin_unlock(&conf->device_lock);
 853        }
 854        spin_unlock_irq(conf->hash_locks + hash);
 855
 856        if (!head)
 857                return;
 858        if (!stripe_can_batch(head))
 859                goto out;
 860
 861        lock_two_stripes(head, sh);
 862        /* clear_batch_ready clear the flag */
 863        if (!stripe_can_batch(head) || !stripe_can_batch(sh))
 864                goto unlock_out;
 865
 866        if (sh->batch_head)
 867                goto unlock_out;
 868
 869        dd_idx = 0;
 870        while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
 871                dd_idx++;
 872        if (head->dev[dd_idx].towrite->bi_opf != sh->dev[dd_idx].towrite->bi_opf ||
 873            bio_op(head->dev[dd_idx].towrite) != bio_op(sh->dev[dd_idx].towrite))
 874                goto unlock_out;
 875
 876        if (head->batch_head) {
 877                spin_lock(&head->batch_head->batch_lock);
 878                /* This batch list is already running */
 879                if (!stripe_can_batch(head)) {
 880                        spin_unlock(&head->batch_head->batch_lock);
 881                        goto unlock_out;
 882                }
 883                /*
 884                 * We must assign batch_head of this stripe within the
 885                 * batch_lock, otherwise clear_batch_ready of batch head
 886                 * stripe could clear BATCH_READY bit of this stripe and
 887                 * this stripe->batch_head doesn't get assigned, which
 888                 * could confuse clear_batch_ready for this stripe
 889                 */
 890                sh->batch_head = head->batch_head;
 891
 892                /*
 893                 * at this point, head's BATCH_READY could be cleared, but we
 894                 * can still add the stripe to batch list
 895                 */
 896                list_add(&sh->batch_list, &head->batch_list);
 897                spin_unlock(&head->batch_head->batch_lock);
 898        } else {
 899                head->batch_head = head;
 900                sh->batch_head = head->batch_head;
 901                spin_lock(&head->batch_lock);
 902                list_add_tail(&sh->batch_list, &head->batch_list);
 903                spin_unlock(&head->batch_lock);
 904        }
 905
 906        if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 907                if (atomic_dec_return(&conf->preread_active_stripes)
 908                    < IO_THRESHOLD)
 909                        md_wakeup_thread(conf->mddev->thread);
 910
 911        if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
 912                int seq = sh->bm_seq;
 913                if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
 914                    sh->batch_head->bm_seq > seq)
 915                        seq = sh->batch_head->bm_seq;
 916                set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
 917                sh->batch_head->bm_seq = seq;
 918        }
 919
 920        atomic_inc(&sh->count);
 921unlock_out:
 922        unlock_two_stripes(head, sh);
 923out:
 924        raid5_release_stripe(head);
 925}
 926
 927/* Determine if 'data_offset' or 'new_data_offset' should be used
 928 * in this stripe_head.
 929 */
 930static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
 931{
 932        sector_t progress = conf->reshape_progress;
 933        /* Need a memory barrier to make sure we see the value
 934         * of conf->generation, or ->data_offset that was set before
 935         * reshape_progress was updated.
 936         */
 937        smp_rmb();
 938        if (progress == MaxSector)
 939                return 0;
 940        if (sh->generation == conf->generation - 1)
 941                return 0;
 942        /* We are in a reshape, and this is a new-generation stripe,
 943         * so use new_data_offset.
 944         */
 945        return 1;
 946}
 947
 948static void dispatch_bio_list(struct bio_list *tmp)
 949{
 950        struct bio *bio;
 951
 952        while ((bio = bio_list_pop(tmp)))
 953                submit_bio_noacct(bio);
 954}
 955
 956static int cmp_stripe(void *priv, const struct list_head *a,
 957                      const struct list_head *b)
 958{
 959        const struct r5pending_data *da = list_entry(a,
 960                                struct r5pending_data, sibling);
 961        const struct r5pending_data *db = list_entry(b,
 962                                struct r5pending_data, sibling);
 963        if (da->sector > db->sector)
 964                return 1;
 965        if (da->sector < db->sector)
 966                return -1;
 967        return 0;
 968}
 969
 970static void dispatch_defer_bios(struct r5conf *conf, int target,
 971                                struct bio_list *list)
 972{
 973        struct r5pending_data *data;
 974        struct list_head *first, *next = NULL;
 975        int cnt = 0;
 976
 977        if (conf->pending_data_cnt == 0)
 978                return;
 979
 980        list_sort(NULL, &conf->pending_list, cmp_stripe);
 981
 982        first = conf->pending_list.next;
 983
 984        /* temporarily move the head */
 985        if (conf->next_pending_data)
 986                list_move_tail(&conf->pending_list,
 987                                &conf->next_pending_data->sibling);
 988
 989        while (!list_empty(&conf->pending_list)) {
 990                data = list_first_entry(&conf->pending_list,
 991                        struct r5pending_data, sibling);
 992                if (&data->sibling == first)
 993                        first = data->sibling.next;
 994                next = data->sibling.next;
 995
 996                bio_list_merge(list, &data->bios);
 997                list_move(&data->sibling, &conf->free_list);
 998                cnt++;
 999                if (cnt >= target)
1000                        break;
1001        }
1002        conf->pending_data_cnt -= cnt;
1003        BUG_ON(conf->pending_data_cnt < 0 || cnt < target);
1004
1005        if (next != &conf->pending_list)
1006                conf->next_pending_data = list_entry(next,
1007                                struct r5pending_data, sibling);
1008        else
1009                conf->next_pending_data = NULL;
1010        /* list isn't empty */
1011        if (first != &conf->pending_list)
1012                list_move_tail(&conf->pending_list, first);
1013}
1014
1015static void flush_deferred_bios(struct r5conf *conf)
1016{
1017        struct bio_list tmp = BIO_EMPTY_LIST;
1018
1019        if (conf->pending_data_cnt == 0)
1020                return;
1021
1022        spin_lock(&conf->pending_bios_lock);
1023        dispatch_defer_bios(conf, conf->pending_data_cnt, &tmp);
1024        BUG_ON(conf->pending_data_cnt != 0);
1025        spin_unlock(&conf->pending_bios_lock);
1026
1027        dispatch_bio_list(&tmp);
1028}
1029
1030static void defer_issue_bios(struct r5conf *conf, sector_t sector,
1031                                struct bio_list *bios)
1032{
1033        struct bio_list tmp = BIO_EMPTY_LIST;
1034        struct r5pending_data *ent;
1035
1036        spin_lock(&conf->pending_bios_lock);
1037        ent = list_first_entry(&conf->free_list, struct r5pending_data,
1038                                                        sibling);
1039        list_move_tail(&ent->sibling, &conf->pending_list);
1040        ent->sector = sector;
1041        bio_list_init(&ent->bios);
1042        bio_list_merge(&ent->bios, bios);
1043        conf->pending_data_cnt++;
1044        if (conf->pending_data_cnt >= PENDING_IO_MAX)
1045                dispatch_defer_bios(conf, PENDING_IO_ONE_FLUSH, &tmp);
1046
1047        spin_unlock(&conf->pending_bios_lock);
1048
1049        dispatch_bio_list(&tmp);
1050}
1051
1052static void
1053raid5_end_read_request(struct bio *bi);
1054static void
1055raid5_end_write_request(struct bio *bi);
1056
1057static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
1058{
1059        struct r5conf *conf = sh->raid_conf;
1060        int i, disks = sh->disks;
1061        struct stripe_head *head_sh = sh;
1062        struct bio_list pending_bios = BIO_EMPTY_LIST;
1063        bool should_defer;
1064
1065        might_sleep();
1066
1067        if (log_stripe(sh, s) == 0)
1068                return;
1069
1070        should_defer = conf->batch_bio_dispatch && conf->group_cnt;
1071
1072        for (i = disks; i--; ) {
1073                int op, op_flags = 0;
1074                int replace_only = 0;
1075                struct bio *bi, *rbi;
1076                struct md_rdev *rdev, *rrdev = NULL;
1077
1078                sh = head_sh;
1079                if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
1080                        op = REQ_OP_WRITE;
1081                        if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
1082                                op_flags = REQ_FUA;
1083                        if (test_bit(R5_Discard, &sh->dev[i].flags))
1084                                op = REQ_OP_DISCARD;
1085                } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1086                        op = REQ_OP_READ;
1087                else if (test_and_clear_bit(R5_WantReplace,
1088                                            &sh->dev[i].flags)) {
1089                        op = REQ_OP_WRITE;
1090                        replace_only = 1;
1091                } else
1092                        continue;
1093                if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
1094                        op_flags |= REQ_SYNC;
1095
1096again:
1097                bi = &sh->dev[i].req;
1098                rbi = &sh->dev[i].rreq; /* For writing to replacement */
1099
1100                rcu_read_lock();
1101                rrdev = rcu_dereference(conf->disks[i].replacement);
1102                smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
1103                rdev = rcu_dereference(conf->disks[i].rdev);
1104                if (!rdev) {
1105                        rdev = rrdev;
1106                        rrdev = NULL;
1107                }
1108                if (op_is_write(op)) {
1109                        if (replace_only)
1110                                rdev = NULL;
1111                        if (rdev == rrdev)
1112                                /* We raced and saw duplicates */
1113                                rrdev = NULL;
1114                } else {
1115                        if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
1116                                rdev = rrdev;
1117                        rrdev = NULL;
1118                }
1119
1120                if (rdev && test_bit(Faulty, &rdev->flags))
1121                        rdev = NULL;
1122                if (rdev)
1123                        atomic_inc(&rdev->nr_pending);
1124                if (rrdev && test_bit(Faulty, &rrdev->flags))
1125                        rrdev = NULL;
1126                if (rrdev)
1127                        atomic_inc(&rrdev->nr_pending);
1128                rcu_read_unlock();
1129
1130                /* We have already checked bad blocks for reads.  Now
1131                 * need to check for writes.  We never accept write errors
1132                 * on the replacement, so we don't to check rrdev.
1133                 */
1134                while (op_is_write(op) && rdev &&
1135                       test_bit(WriteErrorSeen, &rdev->flags)) {
1136                        sector_t first_bad;
1137                        int bad_sectors;
1138                        int bad = is_badblock(rdev, sh->sector, RAID5_STRIPE_SECTORS(conf),
1139                                              &first_bad, &bad_sectors);
1140                        if (!bad)
1141                                break;
1142
1143                        if (bad < 0) {
1144                                set_bit(BlockedBadBlocks, &rdev->flags);
1145                                if (!conf->mddev->external &&
1146                                    conf->mddev->sb_flags) {
1147                                        /* It is very unlikely, but we might
1148                                         * still need to write out the
1149                                         * bad block log - better give it
1150                                         * a chance*/
1151                                        md_check_recovery(conf->mddev);
1152                                }
1153                                /*
1154                                 * Because md_wait_for_blocked_rdev
1155                                 * will dec nr_pending, we must
1156                                 * increment it first.
1157                                 */
1158                                atomic_inc(&rdev->nr_pending);
1159                                md_wait_for_blocked_rdev(rdev, conf->mddev);
1160                        } else {
1161                                /* Acknowledged bad block - skip the write */
1162                                rdev_dec_pending(rdev, conf->mddev);
1163                                rdev = NULL;
1164                        }
1165                }
1166
1167                if (rdev) {
1168                        if (s->syncing || s->expanding || s->expanded
1169                            || s->replacing)
1170                                md_sync_acct(rdev->bdev, RAID5_STRIPE_SECTORS(conf));
1171
1172                        set_bit(STRIPE_IO_STARTED, &sh->state);
1173
1174                        bio_set_dev(bi, rdev->bdev);
1175                        bio_set_op_attrs(bi, op, op_flags);
1176                        bi->bi_end_io = op_is_write(op)
1177                                ? raid5_end_write_request
1178                                : raid5_end_read_request;
1179                        bi->bi_private = sh;
1180
1181                        pr_debug("%s: for %llu schedule op %d on disc %d\n",
1182                                __func__, (unsigned long long)sh->sector,
1183                                bi->bi_opf, i);
1184                        atomic_inc(&sh->count);
1185                        if (sh != head_sh)
1186                                atomic_inc(&head_sh->count);
1187                        if (use_new_offset(conf, sh))
1188                                bi->bi_iter.bi_sector = (sh->sector
1189                                                 + rdev->new_data_offset);
1190                        else
1191                                bi->bi_iter.bi_sector = (sh->sector
1192                                                 + rdev->data_offset);
1193                        if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1194                                bi->bi_opf |= REQ_NOMERGE;
1195
1196                        if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1197                                WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1198
1199                        if (!op_is_write(op) &&
1200                            test_bit(R5_InJournal, &sh->dev[i].flags))
1201                                /*
1202                                 * issuing read for a page in journal, this
1203                                 * must be preparing for prexor in rmw; read
1204                                 * the data into orig_page
1205                                 */
1206                                sh->dev[i].vec.bv_page = sh->dev[i].orig_page;
1207                        else
1208                                sh->dev[i].vec.bv_page = sh->dev[i].page;
1209                        bi->bi_vcnt = 1;
1210                        bi->bi_io_vec[0].bv_len = RAID5_STRIPE_SIZE(conf);
1211                        bi->bi_io_vec[0].bv_offset = sh->dev[i].offset;
1212                        bi->bi_iter.bi_size = RAID5_STRIPE_SIZE(conf);
1213                        bi->bi_write_hint = sh->dev[i].write_hint;
1214                        if (!rrdev)
1215                                sh->dev[i].write_hint = RWH_WRITE_LIFE_NOT_SET;
1216                        /*
1217                         * If this is discard request, set bi_vcnt 0. We don't
1218                         * want to confuse SCSI because SCSI will replace payload
1219                         */
1220                        if (op == REQ_OP_DISCARD)
1221                                bi->bi_vcnt = 0;
1222                        if (rrdev)
1223                                set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1224
1225                        if (conf->mddev->gendisk)
1226                                trace_block_bio_remap(bi,
1227                                                disk_devt(conf->mddev->gendisk),
1228                                                sh->dev[i].sector);
1229                        if (should_defer && op_is_write(op))
1230                                bio_list_add(&pending_bios, bi);
1231                        else
1232                                submit_bio_noacct(bi);
1233                }
1234                if (rrdev) {
1235                        if (s->syncing || s->expanding || s->expanded
1236                            || s->replacing)
1237                                md_sync_acct(rrdev->bdev, RAID5_STRIPE_SECTORS(conf));
1238
1239                        set_bit(STRIPE_IO_STARTED, &sh->state);
1240
1241                        bio_set_dev(rbi, rrdev->bdev);
1242                        bio_set_op_attrs(rbi, op, op_flags);
1243                        BUG_ON(!op_is_write(op));
1244                        rbi->bi_end_io = raid5_end_write_request;
1245                        rbi->bi_private = sh;
1246
1247                        pr_debug("%s: for %llu schedule op %d on "
1248                                 "replacement disc %d\n",
1249                                __func__, (unsigned long long)sh->sector,
1250                                rbi->bi_opf, i);
1251                        atomic_inc(&sh->count);
1252                        if (sh != head_sh)
1253                                atomic_inc(&head_sh->count);
1254                        if (use_new_offset(conf, sh))
1255                                rbi->bi_iter.bi_sector = (sh->sector
1256                                                  + rrdev->new_data_offset);
1257                        else
1258                                rbi->bi_iter.bi_sector = (sh->sector
1259                                                  + rrdev->data_offset);
1260                        if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1261                                WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1262                        sh->dev[i].rvec.bv_page = sh->dev[i].page;
1263                        rbi->bi_vcnt = 1;
1264                        rbi->bi_io_vec[0].bv_len = RAID5_STRIPE_SIZE(conf);
1265                        rbi->bi_io_vec[0].bv_offset = sh->dev[i].offset;
1266                        rbi->bi_iter.bi_size = RAID5_STRIPE_SIZE(conf);
1267                        rbi->bi_write_hint = sh->dev[i].write_hint;
1268                        sh->dev[i].write_hint = RWH_WRITE_LIFE_NOT_SET;
1269                        /*
1270                         * If this is discard request, set bi_vcnt 0. We don't
1271                         * want to confuse SCSI because SCSI will replace payload
1272                         */
1273                        if (op == REQ_OP_DISCARD)
1274                                rbi->bi_vcnt = 0;
1275                        if (conf->mddev->gendisk)
1276                                trace_block_bio_remap(rbi,
1277                                                disk_devt(conf->mddev->gendisk),
1278                                                sh->dev[i].sector);
1279                        if (should_defer && op_is_write(op))
1280                                bio_list_add(&pending_bios, rbi);
1281                        else
1282                                submit_bio_noacct(rbi);
1283                }
1284                if (!rdev && !rrdev) {
1285                        if (op_is_write(op))
1286                                set_bit(STRIPE_DEGRADED, &sh->state);
1287                        pr_debug("skip op %d on disc %d for sector %llu\n",
1288                                bi->bi_opf, i, (unsigned long long)sh->sector);
1289                        clear_bit(R5_LOCKED, &sh->dev[i].flags);
1290                        set_bit(STRIPE_HANDLE, &sh->state);
1291                }
1292
1293                if (!head_sh->batch_head)
1294                        continue;
1295                sh = list_first_entry(&sh->batch_list, struct stripe_head,
1296                                      batch_list);
1297                if (sh != head_sh)
1298                        goto again;
1299        }
1300
1301        if (should_defer && !bio_list_empty(&pending_bios))
1302                defer_issue_bios(conf, head_sh->sector, &pending_bios);
1303}
1304
1305static struct dma_async_tx_descriptor *
1306async_copy_data(int frombio, struct bio *bio, struct page **page,
1307        unsigned int poff, sector_t sector, struct dma_async_tx_descriptor *tx,
1308        struct stripe_head *sh, int no_skipcopy)
1309{
1310        struct bio_vec bvl;
1311        struct bvec_iter iter;
1312        struct page *bio_page;
1313        int page_offset;
1314        struct async_submit_ctl submit;
1315        enum async_tx_flags flags = 0;
1316        struct r5conf *conf = sh->raid_conf;
1317
1318        if (bio->bi_iter.bi_sector >= sector)
1319                page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1320        else
1321                page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1322
1323        if (frombio)
1324                flags |= ASYNC_TX_FENCE;
1325        init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1326
1327        bio_for_each_segment(bvl, bio, iter) {
1328                int len = bvl.bv_len;
1329                int clen;
1330                int b_offset = 0;
1331
1332                if (page_offset < 0) {
1333                        b_offset = -page_offset;
1334                        page_offset += b_offset;
1335                        len -= b_offset;
1336                }
1337
1338                if (len > 0 && page_offset + len > RAID5_STRIPE_SIZE(conf))
1339                        clen = RAID5_STRIPE_SIZE(conf) - page_offset;
1340                else
1341                        clen = len;
1342
1343                if (clen > 0) {
1344                        b_offset += bvl.bv_offset;
1345                        bio_page = bvl.bv_page;
1346                        if (frombio) {
1347                                if (conf->skip_copy &&
1348                                    b_offset == 0 && page_offset == 0 &&
1349                                    clen == RAID5_STRIPE_SIZE(conf) &&
1350                                    !no_skipcopy)
1351                                        *page = bio_page;
1352                                else
1353                                        tx = async_memcpy(*page, bio_page, page_offset + poff,
1354                                                  b_offset, clen, &submit);
1355                        } else
1356                                tx = async_memcpy(bio_page, *page, b_offset,
1357                                                  page_offset + poff, clen, &submit);
1358                }
1359                /* chain the operations */
1360                submit.depend_tx = tx;
1361
1362                if (clen < len) /* hit end of page */
1363                        break;
1364                page_offset +=  len;
1365        }
1366
1367        return tx;
1368}
1369
1370static void ops_complete_biofill(void *stripe_head_ref)
1371{
1372        struct stripe_head *sh = stripe_head_ref;
1373        int i;
1374        struct r5conf *conf = sh->raid_conf;
1375
1376        pr_debug("%s: stripe %llu\n", __func__,
1377                (unsigned long long)sh->sector);
1378
1379        /* clear completed biofills */
1380        for (i = sh->disks; i--; ) {
1381                struct r5dev *dev = &sh->dev[i];
1382
1383                /* acknowledge completion of a biofill operation */
1384                /* and check if we need to reply to a read request,
1385                 * new R5_Wantfill requests are held off until
1386                 * !STRIPE_BIOFILL_RUN
1387                 */
1388                if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1389                        struct bio *rbi, *rbi2;
1390
1391                        BUG_ON(!dev->read);
1392                        rbi = dev->read;
1393                        dev->read = NULL;
1394                        while (rbi && rbi->bi_iter.bi_sector <
1395                                dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1396                                rbi2 = r5_next_bio(conf, rbi, dev->sector);
1397                                bio_endio(rbi);
1398                                rbi = rbi2;
1399                        }
1400                }
1401        }
1402        clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1403
1404        set_bit(STRIPE_HANDLE, &sh->state);
1405        raid5_release_stripe(sh);
1406}
1407
1408static void ops_run_biofill(struct stripe_head *sh)
1409{
1410        struct dma_async_tx_descriptor *tx = NULL;
1411        struct async_submit_ctl submit;
1412        int i;
1413        struct r5conf *conf = sh->raid_conf;
1414
1415        BUG_ON(sh->batch_head);
1416        pr_debug("%s: stripe %llu\n", __func__,
1417                (unsigned long long)sh->sector);
1418
1419        for (i = sh->disks; i--; ) {
1420                struct r5dev *dev = &sh->dev[i];
1421                if (test_bit(R5_Wantfill, &dev->flags)) {
1422                        struct bio *rbi;
1423                        spin_lock_irq(&sh->stripe_lock);
1424                        dev->read = rbi = dev->toread;
1425                        dev->toread = NULL;
1426                        spin_unlock_irq(&sh->stripe_lock);
1427                        while (rbi && rbi->bi_iter.bi_sector <
1428                                dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1429                                tx = async_copy_data(0, rbi, &dev->page,
1430                                                     dev->offset,
1431                                                     dev->sector, tx, sh, 0);
1432                                rbi = r5_next_bio(conf, rbi, dev->sector);
1433                        }
1434                }
1435        }
1436
1437        atomic_inc(&sh->count);
1438        init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1439        async_trigger_callback(&submit);
1440}
1441
1442static void mark_target_uptodate(struct stripe_head *sh, int target)
1443{
1444        struct r5dev *tgt;
1445
1446        if (target < 0)
1447                return;
1448
1449        tgt = &sh->dev[target];
1450        set_bit(R5_UPTODATE, &tgt->flags);
1451        BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1452        clear_bit(R5_Wantcompute, &tgt->flags);
1453}
1454
1455static void ops_complete_compute(void *stripe_head_ref)
1456{
1457        struct stripe_head *sh = stripe_head_ref;
1458
1459        pr_debug("%s: stripe %llu\n", __func__,
1460                (unsigned long long)sh->sector);
1461
1462        /* mark the computed target(s) as uptodate */
1463        mark_target_uptodate(sh, sh->ops.target);
1464        mark_target_uptodate(sh, sh->ops.target2);
1465
1466        clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1467        if (sh->check_state == check_state_compute_run)
1468                sh->check_state = check_state_compute_result;
1469        set_bit(STRIPE_HANDLE, &sh->state);
1470        raid5_release_stripe(sh);
1471}
1472
1473/* return a pointer to the address conversion region of the scribble buffer */
1474static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1475{
1476        return percpu->scribble + i * percpu->scribble_obj_size;
1477}
1478
1479/* return a pointer to the address conversion region of the scribble buffer */
1480static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1481                                 struct raid5_percpu *percpu, int i)
1482{
1483        return (void *) (to_addr_page(percpu, i) + sh->disks + 2);
1484}
1485
1486/*
1487 * Return a pointer to record offset address.
1488 */
1489static unsigned int *
1490to_addr_offs(struct stripe_head *sh, struct raid5_percpu *percpu)
1491{
1492        return (unsigned int *) (to_addr_conv(sh, percpu, 0) + sh->disks + 2);
1493}
1494
1495static struct dma_async_tx_descriptor *
1496ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1497{
1498        int disks = sh->disks;
1499        struct page **xor_srcs = to_addr_page(percpu, 0);
1500        unsigned int *off_srcs = to_addr_offs(sh, percpu);
1501        int target = sh->ops.target;
1502        struct r5dev *tgt = &sh->dev[target];
1503        struct page *xor_dest = tgt->page;
1504        unsigned int off_dest = tgt->offset;
1505        int count = 0;
1506        struct dma_async_tx_descriptor *tx;
1507        struct async_submit_ctl submit;
1508        int i;
1509
1510        BUG_ON(sh->batch_head);
1511
1512        pr_debug("%s: stripe %llu block: %d\n",
1513                __func__, (unsigned long long)sh->sector, target);
1514        BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1515
1516        for (i = disks; i--; ) {
1517                if (i != target) {
1518                        off_srcs[count] = sh->dev[i].offset;
1519                        xor_srcs[count++] = sh->dev[i].page;
1520                }
1521        }
1522
1523        atomic_inc(&sh->count);
1524
1525        init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1526                          ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1527        if (unlikely(count == 1))
1528                tx = async_memcpy(xor_dest, xor_srcs[0], off_dest, off_srcs[0],
1529                                RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1530        else
1531                tx = async_xor_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
1532                                RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1533
1534        return tx;
1535}
1536
1537/* set_syndrome_sources - populate source buffers for gen_syndrome
1538 * @srcs - (struct page *) array of size sh->disks
1539 * @offs - (unsigned int) array of offset for each page
1540 * @sh - stripe_head to parse
1541 *
1542 * Populates srcs in proper layout order for the stripe and returns the
1543 * 'count' of sources to be used in a call to async_gen_syndrome.  The P
1544 * destination buffer is recorded in srcs[count] and the Q destination
1545 * is recorded in srcs[count+1]].
1546 */
1547static int set_syndrome_sources(struct page **srcs,
1548                                unsigned int *offs,
1549                                struct stripe_head *sh,
1550                                int srctype)
1551{
1552        int disks = sh->disks;
1553        int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1554        int d0_idx = raid6_d0(sh);
1555        int count;
1556        int i;
1557
1558        for (i = 0; i < disks; i++)
1559                srcs[i] = NULL;
1560
1561        count = 0;
1562        i = d0_idx;
1563        do {
1564                int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1565                struct r5dev *dev = &sh->dev[i];
1566
1567                if (i == sh->qd_idx || i == sh->pd_idx ||
1568                    (srctype == SYNDROME_SRC_ALL) ||
1569                    (srctype == SYNDROME_SRC_WANT_DRAIN &&
1570                     (test_bit(R5_Wantdrain, &dev->flags) ||
1571                      test_bit(R5_InJournal, &dev->flags))) ||
1572                    (srctype == SYNDROME_SRC_WRITTEN &&
1573                     (dev->written ||
1574                      test_bit(R5_InJournal, &dev->flags)))) {
1575                        if (test_bit(R5_InJournal, &dev->flags))
1576                                srcs[slot] = sh->dev[i].orig_page;
1577                        else
1578                                srcs[slot] = sh->dev[i].page;
1579                        /*
1580                         * For R5_InJournal, PAGE_SIZE must be 4KB and will
1581                         * not shared page. In that case, dev[i].offset
1582                         * is 0.
1583                         */
1584                        offs[slot] = sh->dev[i].offset;
1585                }
1586                i = raid6_next_disk(i, disks);
1587        } while (i != d0_idx);
1588
1589        return syndrome_disks;
1590}
1591
1592static struct dma_async_tx_descriptor *
1593ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1594{
1595        int disks = sh->disks;
1596        struct page **blocks = to_addr_page(percpu, 0);
1597        unsigned int *offs = to_addr_offs(sh, percpu);
1598        int target;
1599        int qd_idx = sh->qd_idx;
1600        struct dma_async_tx_descriptor *tx;
1601        struct async_submit_ctl submit;
1602        struct r5dev *tgt;
1603        struct page *dest;
1604        unsigned int dest_off;
1605        int i;
1606        int count;
1607
1608        BUG_ON(sh->batch_head);
1609        if (sh->ops.target < 0)
1610                target = sh->ops.target2;
1611        else if (sh->ops.target2 < 0)
1612                target = sh->ops.target;
1613        else
1614                /* we should only have one valid target */
1615                BUG();
1616        BUG_ON(target < 0);
1617        pr_debug("%s: stripe %llu block: %d\n",
1618                __func__, (unsigned long long)sh->sector, target);
1619
1620        tgt = &sh->dev[target];
1621        BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1622        dest = tgt->page;
1623        dest_off = tgt->offset;
1624
1625        atomic_inc(&sh->count);
1626
1627        if (target == qd_idx) {
1628                count = set_syndrome_sources(blocks, offs, sh, SYNDROME_SRC_ALL);
1629                blocks[count] = NULL; /* regenerating p is not necessary */
1630                BUG_ON(blocks[count+1] != dest); /* q should already be set */
1631                init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1632                                  ops_complete_compute, sh,
1633                                  to_addr_conv(sh, percpu, 0));
1634                tx = async_gen_syndrome(blocks, offs, count+2,
1635                                RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1636        } else {
1637                /* Compute any data- or p-drive using XOR */
1638                count = 0;
1639                for (i = disks; i-- ; ) {
1640                        if (i == target || i == qd_idx)
1641                                continue;
1642                        offs[count] = sh->dev[i].offset;
1643                        blocks[count++] = sh->dev[i].page;
1644                }
1645
1646                init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1647                                  NULL, ops_complete_compute, sh,
1648                                  to_addr_conv(sh, percpu, 0));
1649                tx = async_xor_offs(dest, dest_off, blocks, offs, count,
1650                                RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1651        }
1652
1653        return tx;
1654}
1655
1656static struct dma_async_tx_descriptor *
1657ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1658{
1659        int i, count, disks = sh->disks;
1660        int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1661        int d0_idx = raid6_d0(sh);
1662        int faila = -1, failb = -1;
1663        int target = sh->ops.target;
1664        int target2 = sh->ops.target2;
1665        struct r5dev *tgt = &sh->dev[target];
1666        struct r5dev *tgt2 = &sh->dev[target2];
1667        struct dma_async_tx_descriptor *tx;
1668        struct page **blocks = to_addr_page(percpu, 0);
1669        unsigned int *offs = to_addr_offs(sh, percpu);
1670        struct async_submit_ctl submit;
1671
1672        BUG_ON(sh->batch_head);
1673        pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1674                 __func__, (unsigned long long)sh->sector, target, target2);
1675        BUG_ON(target < 0 || target2 < 0);
1676        BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1677        BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1678
1679        /* we need to open-code set_syndrome_sources to handle the
1680         * slot number conversion for 'faila' and 'failb'
1681         */
1682        for (i = 0; i < disks ; i++) {
1683                offs[i] = 0;
1684                blocks[i] = NULL;
1685        }
1686        count = 0;
1687        i = d0_idx;
1688        do {
1689                int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1690
1691                offs[slot] = sh->dev[i].offset;
1692                blocks[slot] = sh->dev[i].page;
1693
1694                if (i == target)
1695                        faila = slot;
1696                if (i == target2)
1697                        failb = slot;
1698                i = raid6_next_disk(i, disks);
1699        } while (i != d0_idx);
1700
1701        BUG_ON(faila == failb);
1702        if (failb < faila)
1703                swap(faila, failb);
1704        pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1705                 __func__, (unsigned long long)sh->sector, faila, failb);
1706
1707        atomic_inc(&sh->count);
1708
1709        if (failb == syndrome_disks+1) {
1710                /* Q disk is one of the missing disks */
1711                if (faila == syndrome_disks) {
1712                        /* Missing P+Q, just recompute */
1713                        init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1714                                          ops_complete_compute, sh,
1715                                          to_addr_conv(sh, percpu, 0));
1716                        return async_gen_syndrome(blocks, offs, syndrome_disks+2,
1717                                                  RAID5_STRIPE_SIZE(sh->raid_conf),
1718                                                  &submit);
1719                } else {
1720                        struct page *dest;
1721                        unsigned int dest_off;
1722                        int data_target;
1723                        int qd_idx = sh->qd_idx;
1724
1725                        /* Missing D+Q: recompute D from P, then recompute Q */
1726                        if (target == qd_idx)
1727                                data_target = target2;
1728                        else
1729                                data_target = target;
1730
1731                        count = 0;
1732                        for (i = disks; i-- ; ) {
1733                                if (i == data_target || i == qd_idx)
1734                                        continue;
1735                                offs[count] = sh->dev[i].offset;
1736                                blocks[count++] = sh->dev[i].page;
1737                        }
1738                        dest = sh->dev[data_target].page;
1739                        dest_off = sh->dev[data_target].offset;
1740                        init_async_submit(&submit,
1741                                          ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1742                                          NULL, NULL, NULL,
1743                                          to_addr_conv(sh, percpu, 0));
1744                        tx = async_xor_offs(dest, dest_off, blocks, offs, count,
1745                                       RAID5_STRIPE_SIZE(sh->raid_conf),
1746                                       &submit);
1747
1748                        count = set_syndrome_sources(blocks, offs, sh, SYNDROME_SRC_ALL);
1749                        init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1750                                          ops_complete_compute, sh,
1751                                          to_addr_conv(sh, percpu, 0));
1752                        return async_gen_syndrome(blocks, offs, count+2,
1753                                                  RAID5_STRIPE_SIZE(sh->raid_conf),
1754                                                  &submit);
1755                }
1756        } else {
1757                init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1758                                  ops_complete_compute, sh,
1759                                  to_addr_conv(sh, percpu, 0));
1760                if (failb == syndrome_disks) {
1761                        /* We're missing D+P. */
1762                        return async_raid6_datap_recov(syndrome_disks+2,
1763                                                RAID5_STRIPE_SIZE(sh->raid_conf),
1764                                                faila,
1765                                                blocks, offs, &submit);
1766                } else {
1767                        /* We're missing D+D. */
1768                        return async_raid6_2data_recov(syndrome_disks+2,
1769                                                RAID5_STRIPE_SIZE(sh->raid_conf),
1770                                                faila, failb,
1771                                                blocks, offs, &submit);
1772                }
1773        }
1774}
1775
1776static void ops_complete_prexor(void *stripe_head_ref)
1777{
1778        struct stripe_head *sh = stripe_head_ref;
1779
1780        pr_debug("%s: stripe %llu\n", __func__,
1781                (unsigned long long)sh->sector);
1782
1783        if (r5c_is_writeback(sh->raid_conf->log))
1784                /*
1785                 * raid5-cache write back uses orig_page during prexor.
1786                 * After prexor, it is time to free orig_page
1787                 */
1788                r5c_release_extra_page(sh);
1789}
1790
1791static struct dma_async_tx_descriptor *
1792ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1793                struct dma_async_tx_descriptor *tx)
1794{
1795        int disks = sh->disks;
1796        struct page **xor_srcs = to_addr_page(percpu, 0);
1797        unsigned int *off_srcs = to_addr_offs(sh, percpu);
1798        int count = 0, pd_idx = sh->pd_idx, i;
1799        struct async_submit_ctl submit;
1800
1801        /* existing parity data subtracted */
1802        unsigned int off_dest = off_srcs[count] = sh->dev[pd_idx].offset;
1803        struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1804
1805        BUG_ON(sh->batch_head);
1806        pr_debug("%s: stripe %llu\n", __func__,
1807                (unsigned long long)sh->sector);
1808
1809        for (i = disks; i--; ) {
1810                struct r5dev *dev = &sh->dev[i];
1811                /* Only process blocks that are known to be uptodate */
1812                if (test_bit(R5_InJournal, &dev->flags)) {
1813                        /*
1814                         * For this case, PAGE_SIZE must be equal to 4KB and
1815                         * page offset is zero.
1816                         */
1817                        off_srcs[count] = dev->offset;
1818                        xor_srcs[count++] = dev->orig_page;
1819                } else if (test_bit(R5_Wantdrain, &dev->flags)) {
1820                        off_srcs[count] = dev->offset;
1821                        xor_srcs[count++] = dev->page;
1822                }
1823        }
1824
1825        init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1826                          ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1827        tx = async_xor_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
1828                        RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1829
1830        return tx;
1831}
1832
1833static struct dma_async_tx_descriptor *
1834ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1835                struct dma_async_tx_descriptor *tx)
1836{
1837        struct page **blocks = to_addr_page(percpu, 0);
1838        unsigned int *offs = to_addr_offs(sh, percpu);
1839        int count;
1840        struct async_submit_ctl submit;
1841
1842        pr_debug("%s: stripe %llu\n", __func__,
1843                (unsigned long long)sh->sector);
1844
1845        count = set_syndrome_sources(blocks, offs, sh, SYNDROME_SRC_WANT_DRAIN);
1846
1847        init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1848                          ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1849        tx = async_gen_syndrome(blocks, offs, count+2,
1850                        RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1851
1852        return tx;
1853}
1854
1855static struct dma_async_tx_descriptor *
1856ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1857{
1858        struct r5conf *conf = sh->raid_conf;
1859        int disks = sh->disks;
1860        int i;
1861        struct stripe_head *head_sh = sh;
1862
1863        pr_debug("%s: stripe %llu\n", __func__,
1864                (unsigned long long)sh->sector);
1865
1866        for (i = disks; i--; ) {
1867                struct r5dev *dev;
1868                struct bio *chosen;
1869
1870                sh = head_sh;
1871                if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1872                        struct bio *wbi;
1873
1874again:
1875                        dev = &sh->dev[i];
1876                        /*
1877                         * clear R5_InJournal, so when rewriting a page in
1878                         * journal, it is not skipped by r5l_log_stripe()
1879                         */
1880                        clear_bit(R5_InJournal, &dev->flags);
1881                        spin_lock_irq(&sh->stripe_lock);
1882                        chosen = dev->towrite;
1883                        dev->towrite = NULL;
1884                        sh->overwrite_disks = 0;
1885                        BUG_ON(dev->written);
1886                        wbi = dev->written = chosen;
1887                        spin_unlock_irq(&sh->stripe_lock);
1888                        WARN_ON(dev->page != dev->orig_page);
1889
1890                        while (wbi && wbi->bi_iter.bi_sector <
1891                                dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1892                                if (wbi->bi_opf & REQ_FUA)
1893                                        set_bit(R5_WantFUA, &dev->flags);
1894                                if (wbi->bi_opf & REQ_SYNC)
1895                                        set_bit(R5_SyncIO, &dev->flags);
1896                                if (bio_op(wbi) == REQ_OP_DISCARD)
1897                                        set_bit(R5_Discard, &dev->flags);
1898                                else {
1899                                        tx = async_copy_data(1, wbi, &dev->page,
1900                                                             dev->offset,
1901                                                             dev->sector, tx, sh,
1902                                                             r5c_is_writeback(conf->log));
1903                                        if (dev->page != dev->orig_page &&
1904                                            !r5c_is_writeback(conf->log)) {
1905                                                set_bit(R5_SkipCopy, &dev->flags);
1906                                                clear_bit(R5_UPTODATE, &dev->flags);
1907                                                clear_bit(R5_OVERWRITE, &dev->flags);
1908                                        }
1909                                }
1910                                wbi = r5_next_bio(conf, wbi, dev->sector);
1911                        }
1912
1913                        if (head_sh->batch_head) {
1914                                sh = list_first_entry(&sh->batch_list,
1915                                                      struct stripe_head,
1916                                                      batch_list);
1917                                if (sh == head_sh)
1918                                        continue;
1919                                goto again;
1920                        }
1921                }
1922        }
1923
1924        return tx;
1925}
1926
1927static void ops_complete_reconstruct(void *stripe_head_ref)
1928{
1929        struct stripe_head *sh = stripe_head_ref;
1930        int disks = sh->disks;
1931        int pd_idx = sh->pd_idx;
1932        int qd_idx = sh->qd_idx;
1933        int i;
1934        bool fua = false, sync = false, discard = false;
1935
1936        pr_debug("%s: stripe %llu\n", __func__,
1937                (unsigned long long)sh->sector);
1938
1939        for (i = disks; i--; ) {
1940                fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1941                sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1942                discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1943        }
1944
1945        for (i = disks; i--; ) {
1946                struct r5dev *dev = &sh->dev[i];
1947
1948                if (dev->written || i == pd_idx || i == qd_idx) {
1949                        if (!discard && !test_bit(R5_SkipCopy, &dev->flags)) {
1950                                set_bit(R5_UPTODATE, &dev->flags);
1951                                if (test_bit(STRIPE_EXPAND_READY, &sh->state))
1952                                        set_bit(R5_Expanded, &dev->flags);
1953                        }
1954                        if (fua)
1955                                set_bit(R5_WantFUA, &dev->flags);
1956                        if (sync)
1957                                set_bit(R5_SyncIO, &dev->flags);
1958                }
1959        }
1960
1961        if (sh->reconstruct_state == reconstruct_state_drain_run)
1962                sh->reconstruct_state = reconstruct_state_drain_result;
1963        else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1964                sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1965        else {
1966                BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1967                sh->reconstruct_state = reconstruct_state_result;
1968        }
1969
1970        set_bit(STRIPE_HANDLE, &sh->state);
1971        raid5_release_stripe(sh);
1972}
1973
1974static void
1975ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1976                     struct dma_async_tx_descriptor *tx)
1977{
1978        int disks = sh->disks;
1979        struct page **xor_srcs;
1980        unsigned int *off_srcs;
1981        struct async_submit_ctl submit;
1982        int count, pd_idx = sh->pd_idx, i;
1983        struct page *xor_dest;
1984        unsigned int off_dest;
1985        int prexor = 0;
1986        unsigned long flags;
1987        int j = 0;
1988        struct stripe_head *head_sh = sh;
1989        int last_stripe;
1990
1991        pr_debug("%s: stripe %llu\n", __func__,
1992                (unsigned long long)sh->sector);
1993
1994        for (i = 0; i < sh->disks; i++) {
1995                if (pd_idx == i)
1996                        continue;
1997                if (!test_bit(R5_Discard, &sh->dev[i].flags))
1998                        break;
1999        }
2000        if (i >= sh->disks) {
2001                atomic_inc(&sh->count);
2002                set_bit(R5_Discard, &sh->dev[pd_idx].flags);
2003                ops_complete_reconstruct(sh);
2004                return;
2005        }
2006again:
2007        count = 0;
2008        xor_srcs = to_addr_page(percpu, j);
2009        off_srcs = to_addr_offs(sh, percpu);
2010        /* check if prexor is active which means only process blocks
2011         * that are part of a read-modify-write (written)
2012         */
2013        if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
2014                prexor = 1;
2015                off_dest = off_srcs[count] = sh->dev[pd_idx].offset;
2016                xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
2017                for (i = disks; i--; ) {
2018                        struct r5dev *dev = &sh->dev[i];
2019                        if (head_sh->dev[i].written ||
2020                            test_bit(R5_InJournal, &head_sh->dev[i].flags)) {
2021                                off_srcs[count] = dev->offset;
2022                                xor_srcs[count++] = dev->page;
2023                        }
2024                }
2025        } else {
2026                xor_dest = sh->dev[pd_idx].page;
2027                off_dest = sh->dev[pd_idx].offset;
2028                for (i = disks; i--; ) {
2029                        struct r5dev *dev = &sh->dev[i];
2030                        if (i != pd_idx) {
2031                                off_srcs[count] = dev->offset;
2032                                xor_srcs[count++] = dev->page;
2033                        }
2034                }
2035        }
2036
2037        /* 1/ if we prexor'd then the dest is reused as a source
2038         * 2/ if we did not prexor then we are redoing the parity
2039         * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
2040         * for the synchronous xor case
2041         */
2042        last_stripe = !head_sh->batch_head ||
2043                list_first_entry(&sh->batch_list,
2044                                 struct stripe_head, batch_list) == head_sh;
2045        if (last_stripe) {
2046                flags = ASYNC_TX_ACK |
2047                        (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
2048
2049                atomic_inc(&head_sh->count);
2050                init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
2051                                  to_addr_conv(sh, percpu, j));
2052        } else {
2053                flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
2054                init_async_submit(&submit, flags, tx, NULL, NULL,
2055                                  to_addr_conv(sh, percpu, j));
2056        }
2057
2058        if (unlikely(count == 1))
2059                tx = async_memcpy(xor_dest, xor_srcs[0], off_dest, off_srcs[0],
2060                                RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
2061        else
2062                tx = async_xor_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
2063                                RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
2064        if (!last_stripe) {
2065                j++;
2066                sh = list_first_entry(&sh->batch_list, struct stripe_head,
2067                                      batch_list);
2068                goto again;
2069        }
2070}
2071
2072static void
2073ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
2074                     struct dma_async_tx_descriptor *tx)
2075{
2076        struct async_submit_ctl submit;
2077        struct page **blocks;
2078        unsigned int *offs;
2079        int count, i, j = 0;
2080        struct stripe_head *head_sh = sh;
2081        int last_stripe;
2082        int synflags;
2083        unsigned long txflags;
2084
2085        pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
2086
2087        for (i = 0; i < sh->disks; i++) {
2088                if (sh->pd_idx == i || sh->qd_idx == i)
2089                        continue;
2090                if (!test_bit(R5_Discard, &sh->dev[i].flags))
2091                        break;
2092        }
2093        if (i >= sh->disks) {
2094                atomic_inc(&sh->count);
2095                set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
2096                set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
2097                ops_complete_reconstruct(sh);
2098                return;
2099        }
2100
2101again:
2102        blocks = to_addr_page(percpu, j);
2103        offs = to_addr_offs(sh, percpu);
2104
2105        if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
2106                synflags = SYNDROME_SRC_WRITTEN;
2107                txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
2108        } else {
2109                synflags = SYNDROME_SRC_ALL;
2110                txflags = ASYNC_TX_ACK;
2111        }
2112
2113        count = set_syndrome_sources(blocks, offs, sh, synflags);
2114        last_stripe = !head_sh->batch_head ||
2115                list_first_entry(&sh->batch_list,
2116                                 struct stripe_head, batch_list) == head_sh;
2117
2118        if (last_stripe) {
2119                atomic_inc(&head_sh->count);
2120                init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
2121                                  head_sh, to_addr_conv(sh, percpu, j));
2122        } else
2123                init_async_submit(&submit, 0, tx, NULL, NULL,
2124                                  to_addr_conv(sh, percpu, j));
2125        tx = async_gen_syndrome(blocks, offs, count+2,
2126                        RAID5_STRIPE_SIZE(sh->raid_conf),  &submit);
2127        if (!last_stripe) {
2128                j++;
2129                sh = list_first_entry(&sh->batch_list, struct stripe_head,
2130                                      batch_list);
2131                goto again;
2132        }
2133}
2134
2135static void ops_complete_check(void *stripe_head_ref)
2136{
2137        struct stripe_head *sh = stripe_head_ref;
2138
2139        pr_debug("%s: stripe %llu\n", __func__,
2140                (unsigned long long)sh->sector);
2141
2142        sh->check_state = check_state_check_result;
2143        set_bit(STRIPE_HANDLE, &sh->state);
2144        raid5_release_stripe(sh);
2145}
2146
2147static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
2148{
2149        int disks = sh->disks;
2150        int pd_idx = sh->pd_idx;
2151        int qd_idx = sh->qd_idx;
2152        struct page *xor_dest;
2153        unsigned int off_dest;
2154        struct page **xor_srcs = to_addr_page(percpu, 0);
2155        unsigned int *off_srcs = to_addr_offs(sh, percpu);
2156        struct dma_async_tx_descriptor *tx;
2157        struct async_submit_ctl submit;
2158        int count;
2159        int i;
2160
2161        pr_debug("%s: stripe %llu\n", __func__,
2162                (unsigned long long)sh->sector);
2163
2164        BUG_ON(sh->batch_head);
2165        count = 0;
2166        xor_dest = sh->dev[pd_idx].page;
2167        off_dest = sh->dev[pd_idx].offset;
2168        off_srcs[count] = off_dest;
2169        xor_srcs[count++] = xor_dest;
2170        for (i = disks; i--; ) {
2171                if (i == pd_idx || i == qd_idx)
2172                        continue;
2173                off_srcs[count] = sh->dev[i].offset;
2174                xor_srcs[count++] = sh->dev[i].page;
2175        }
2176
2177        init_async_submit(&submit, 0, NULL, NULL, NULL,
2178                          to_addr_conv(sh, percpu, 0));
2179        tx = async_xor_val_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
2180                           RAID5_STRIPE_SIZE(sh->raid_conf),
2181                           &sh->ops.zero_sum_result, &submit);
2182
2183        atomic_inc(&sh->count);
2184        init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
2185        tx = async_trigger_callback(&submit);
2186}
2187
2188static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
2189{
2190        struct page **srcs = to_addr_page(percpu, 0);
2191        unsigned int *offs = to_addr_offs(sh, percpu);
2192        struct async_submit_ctl submit;
2193        int count;
2194
2195        pr_debug("%s: stripe %llu checkp: %d\n", __func__,
2196                (unsigned long long)sh->sector, checkp);
2197
2198        BUG_ON(sh->batch_head);
2199        count = set_syndrome_sources(srcs, offs, sh, SYNDROME_SRC_ALL);
2200        if (!checkp)
2201                srcs[count] = NULL;
2202
2203        atomic_inc(&sh->count);
2204        init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
2205                          sh, to_addr_conv(sh, percpu, 0));
2206        async_syndrome_val(srcs, offs, count+2,
2207                           RAID5_STRIPE_SIZE(sh->raid_conf),
2208                           &sh->ops.zero_sum_result, percpu->spare_page, 0, &submit);
2209}
2210
2211static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
2212{
2213        int overlap_clear = 0, i, disks = sh->disks;
2214        struct dma_async_tx_descriptor *tx = NULL;
2215        struct r5conf *conf = sh->raid_conf;
2216        int level = conf->level;
2217        struct raid5_percpu *percpu;
2218        unsigned long cpu;
2219
2220        cpu = get_cpu();
2221        percpu = per_cpu_ptr(conf->percpu, cpu);
2222        if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
2223                ops_run_biofill(sh);
2224                overlap_clear++;
2225        }
2226
2227        if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
2228                if (level < 6)
2229                        tx = ops_run_compute5(sh, percpu);
2230                else {
2231                        if (sh->ops.target2 < 0 || sh->ops.target < 0)
2232                                tx = ops_run_compute6_1(sh, percpu);
2233                        else
2234                                tx = ops_run_compute6_2(sh, percpu);
2235                }
2236                /* terminate the chain if reconstruct is not set to be run */
2237                if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
2238                        async_tx_ack(tx);
2239        }
2240
2241        if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
2242                if (level < 6)
2243                        tx = ops_run_prexor5(sh, percpu, tx);
2244                else
2245                        tx = ops_run_prexor6(sh, percpu, tx);
2246        }
2247
2248        if (test_bit(STRIPE_OP_PARTIAL_PARITY, &ops_request))
2249                tx = ops_run_partial_parity(sh, percpu, tx);
2250
2251        if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
2252                tx = ops_run_biodrain(sh, tx);
2253                overlap_clear++;
2254        }
2255
2256        if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
2257                if (level < 6)
2258                        ops_run_reconstruct5(sh, percpu, tx);
2259                else
2260                        ops_run_reconstruct6(sh, percpu, tx);
2261        }
2262
2263        if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
2264                if (sh->check_state == check_state_run)
2265                        ops_run_check_p(sh, percpu);
2266                else if (sh->check_state == check_state_run_q)
2267                        ops_run_check_pq(sh, percpu, 0);
2268                else if (sh->check_state == check_state_run_pq)
2269                        ops_run_check_pq(sh, percpu, 1);
2270                else
2271                        BUG();
2272        }
2273
2274        if (overlap_clear && !sh->batch_head)
2275                for (i = disks; i--; ) {
2276                        struct r5dev *dev = &sh->dev[i];
2277                        if (test_and_clear_bit(R5_Overlap, &dev->flags))
2278                                wake_up(&sh->raid_conf->wait_for_overlap);
2279                }
2280        put_cpu();
2281}
2282
2283static void free_stripe(struct kmem_cache *sc, struct stripe_head *sh)
2284{
2285#if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2286        kfree(sh->pages);
2287#endif
2288        if (sh->ppl_page)
2289                __free_page(sh->ppl_page);
2290        kmem_cache_free(sc, sh);
2291}
2292
2293static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp,
2294        int disks, struct r5conf *conf)
2295{
2296        struct stripe_head *sh;
2297        int i;
2298
2299        sh = kmem_cache_zalloc(sc, gfp);
2300        if (sh) {
2301                spin_lock_init(&sh->stripe_lock);
2302                spin_lock_init(&sh->batch_lock);
2303                INIT_LIST_HEAD(&sh->batch_list);
2304                INIT_LIST_HEAD(&sh->lru);
2305                INIT_LIST_HEAD(&sh->r5c);
2306                INIT_LIST_HEAD(&sh->log_list);
2307                atomic_set(&sh->count, 1);
2308                sh->raid_conf = conf;
2309                sh->log_start = MaxSector;
2310                for (i = 0; i < disks; i++) {
2311                        struct r5dev *dev = &sh->dev[i];
2312
2313                        bio_init(&dev->req, &dev->vec, 1);
2314                        bio_init(&dev->rreq, &dev->rvec, 1);
2315                }
2316
2317                if (raid5_has_ppl(conf)) {
2318                        sh->ppl_page = alloc_page(gfp);
2319                        if (!sh->ppl_page) {
2320                                free_stripe(sc, sh);
2321                                return NULL;
2322                        }
2323                }
2324#if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2325                if (init_stripe_shared_pages(sh, conf, disks)) {
2326                        free_stripe(sc, sh);
2327                        return NULL;
2328                }
2329#endif
2330        }
2331        return sh;
2332}
2333static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2334{
2335        struct stripe_head *sh;
2336
2337        sh = alloc_stripe(conf->slab_cache, gfp, conf->pool_size, conf);
2338        if (!sh)
2339                return 0;
2340
2341        if (grow_buffers(sh, gfp)) {
2342                shrink_buffers(sh);
2343                free_stripe(conf->slab_cache, sh);
2344                return 0;
2345        }
2346        sh->hash_lock_index =
2347                conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2348        /* we just created an active stripe so... */
2349        atomic_inc(&conf->active_stripes);
2350
2351        raid5_release_stripe(sh);
2352        conf->max_nr_stripes++;
2353        return 1;
2354}
2355
2356static int grow_stripes(struct r5conf *conf, int num)
2357{
2358        struct kmem_cache *sc;
2359        size_t namelen = sizeof(conf->cache_name[0]);
2360        int devs = max(conf->raid_disks, conf->previous_raid_disks);
2361
2362        if (conf->mddev->gendisk)
2363                snprintf(conf->cache_name[0], namelen,
2364                        "raid%d-%s", conf->level, mdname(conf->mddev));
2365        else
2366                snprintf(conf->cache_name[0], namelen,
2367                        "raid%d-%p", conf->level, conf->mddev);
2368        snprintf(conf->cache_name[1], namelen, "%.27s-alt", conf->cache_name[0]);
2369
2370        conf->active_name = 0;
2371        sc = kmem_cache_create(conf->cache_name[conf->active_name],
2372                               sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2373                               0, 0, NULL);
2374        if (!sc)
2375                return 1;
2376        conf->slab_cache = sc;
2377        conf->pool_size = devs;
2378        while (num--)
2379                if (!grow_one_stripe(conf, GFP_KERNEL))
2380                        return 1;
2381
2382        return 0;
2383}
2384
2385/**
2386 * scribble_alloc - allocate percpu scribble buffer for required size
2387 *                  of the scribble region
2388 * @percpu: from for_each_present_cpu() of the caller
2389 * @num: total number of disks in the array
2390 * @cnt: scribble objs count for required size of the scribble region
2391 *
2392 * The scribble buffer size must be enough to contain:
2393 * 1/ a struct page pointer for each device in the array +2
2394 * 2/ room to convert each entry in (1) to its corresponding dma
2395 *    (dma_map_page()) or page (page_address()) address.
2396 *
2397 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2398 * calculate over all devices (not just the data blocks), using zeros in place
2399 * of the P and Q blocks.
2400 */
2401static int scribble_alloc(struct raid5_percpu *percpu,
2402                          int num, int cnt)
2403{
2404        size_t obj_size =
2405                sizeof(struct page *) * (num + 2) +
2406                sizeof(addr_conv_t) * (num + 2) +
2407                sizeof(unsigned int) * (num + 2);
2408        void *scribble;
2409
2410        /*
2411         * If here is in raid array suspend context, it is in memalloc noio
2412         * context as well, there is no potential recursive memory reclaim
2413         * I/Os with the GFP_KERNEL flag.
2414         */
2415        scribble = kvmalloc_array(cnt, obj_size, GFP_KERNEL);
2416        if (!scribble)
2417                return -ENOMEM;
2418
2419        kvfree(percpu->scribble);
2420
2421        percpu->scribble = scribble;
2422        percpu->scribble_obj_size = obj_size;
2423        return 0;
2424}
2425
2426static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2427{
2428        unsigned long cpu;
2429        int err = 0;
2430
2431        /*
2432         * Never shrink. And mddev_suspend() could deadlock if this is called
2433         * from raid5d. In that case, scribble_disks and scribble_sectors
2434         * should equal to new_disks and new_sectors
2435         */
2436        if (conf->scribble_disks >= new_disks &&
2437            conf->scribble_sectors >= new_sectors)
2438                return 0;
2439        mddev_suspend(conf->mddev);
2440        cpus_read_lock();
2441
2442        for_each_present_cpu(cpu) {
2443                struct raid5_percpu *percpu;
2444
2445                percpu = per_cpu_ptr(conf->percpu, cpu);
2446                err = scribble_alloc(percpu, new_disks,
2447                                     new_sectors / RAID5_STRIPE_SECTORS(conf));
2448                if (err)
2449                        break;
2450        }
2451
2452        cpus_read_unlock();
2453        mddev_resume(conf->mddev);
2454        if (!err) {
2455                conf->scribble_disks = new_disks;
2456                conf->scribble_sectors = new_sectors;
2457        }
2458        return err;
2459}
2460
2461static int resize_stripes(struct r5conf *conf, int newsize)
2462{
2463        /* Make all the stripes able to hold 'newsize' devices.
2464         * New slots in each stripe get 'page' set to a new page.
2465         *
2466         * This happens in stages:
2467         * 1/ create a new kmem_cache and allocate the required number of
2468         *    stripe_heads.
2469         * 2/ gather all the old stripe_heads and transfer the pages across
2470         *    to the new stripe_heads.  This will have the side effect of
2471         *    freezing the array as once all stripe_heads have been collected,
2472         *    no IO will be possible.  Old stripe heads are freed once their
2473         *    pages have been transferred over, and the old kmem_cache is
2474         *    freed when all stripes are done.
2475         * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
2476         *    we simple return a failure status - no need to clean anything up.
2477         * 4/ allocate new pages for the new slots in the new stripe_heads.
2478         *    If this fails, we don't bother trying the shrink the
2479         *    stripe_heads down again, we just leave them as they are.
2480         *    As each stripe_head is processed the new one is released into
2481         *    active service.
2482         *
2483         * Once step2 is started, we cannot afford to wait for a write,
2484         * so we use GFP_NOIO allocations.
2485         */
2486        struct stripe_head *osh, *nsh;
2487        LIST_HEAD(newstripes);
2488        struct disk_info *ndisks;
2489        int err = 0;
2490        struct kmem_cache *sc;
2491        int i;
2492        int hash, cnt;
2493
2494        md_allow_write(conf->mddev);
2495
2496        /* Step 1 */
2497        sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2498                               sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2499                               0, 0, NULL);
2500        if (!sc)
2501                return -ENOMEM;
2502
2503        /* Need to ensure auto-resizing doesn't interfere */
2504        mutex_lock(&conf->cache_size_mutex);
2505
2506        for (i = conf->max_nr_stripes; i; i--) {
2507                nsh = alloc_stripe(sc, GFP_KERNEL, newsize, conf);
2508                if (!nsh)
2509                        break;
2510
2511                list_add(&nsh->lru, &newstripes);
2512        }
2513        if (i) {
2514                /* didn't get enough, give up */
2515                while (!list_empty(&newstripes)) {
2516                        nsh = list_entry(newstripes.next, struct stripe_head, lru);
2517                        list_del(&nsh->lru);
2518                        free_stripe(sc, nsh);
2519                }
2520                kmem_cache_destroy(sc);
2521                mutex_unlock(&conf->cache_size_mutex);
2522                return -ENOMEM;
2523        }
2524        /* Step 2 - Must use GFP_NOIO now.
2525         * OK, we have enough stripes, start collecting inactive
2526         * stripes and copying them over
2527         */
2528        hash = 0;
2529        cnt = 0;
2530        list_for_each_entry(nsh, &newstripes, lru) {
2531                lock_device_hash_lock(conf, hash);
2532                wait_event_cmd(conf->wait_for_stripe,
2533                                    !list_empty(conf->inactive_list + hash),
2534                                    unlock_device_hash_lock(conf, hash),
2535                                    lock_device_hash_lock(conf, hash));
2536                osh = get_free_stripe(conf, hash);
2537                unlock_device_hash_lock(conf, hash);
2538
2539#if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2540        for (i = 0; i < osh->nr_pages; i++) {
2541                nsh->pages[i] = osh->pages[i];
2542                osh->pages[i] = NULL;
2543        }
2544#endif
2545                for(i=0; i<conf->pool_size; i++) {
2546                        nsh->dev[i].page = osh->dev[i].page;
2547                        nsh->dev[i].orig_page = osh->dev[i].page;
2548                        nsh->dev[i].offset = osh->dev[i].offset;
2549                }
2550                nsh->hash_lock_index = hash;
2551                free_stripe(conf->slab_cache, osh);
2552                cnt++;
2553                if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2554                    !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2555                        hash++;
2556                        cnt = 0;
2557                }
2558        }
2559        kmem_cache_destroy(conf->slab_cache);
2560
2561        /* Step 3.
2562         * At this point, we are holding all the stripes so the array
2563         * is completely stalled, so now is a good time to resize
2564         * conf->disks and the scribble region
2565         */
2566        ndisks = kcalloc(newsize, sizeof(struct disk_info), GFP_NOIO);
2567        if (ndisks) {
2568                for (i = 0; i < conf->pool_size; i++)
2569                        ndisks[i] = conf->disks[i];
2570
2571                for (i = conf->pool_size; i < newsize; i++) {
2572                        ndisks[i].extra_page = alloc_page(GFP_NOIO);
2573                        if (!ndisks[i].extra_page)
2574                                err = -ENOMEM;
2575                }
2576
2577                if (err) {
2578                        for (i = conf->pool_size; i < newsize; i++)
2579                                if (ndisks[i].extra_page)
2580                                        put_page(ndisks[i].extra_page);
2581                        kfree(ndisks);
2582                } else {
2583                        kfree(conf->disks);
2584                        conf->disks = ndisks;
2585                }
2586        } else
2587                err = -ENOMEM;
2588
2589        conf->slab_cache = sc;
2590        conf->active_name = 1-conf->active_name;
2591
2592        /* Step 4, return new stripes to service */
2593        while(!list_empty(&newstripes)) {
2594                nsh = list_entry(newstripes.next, struct stripe_head, lru);
2595                list_del_init(&nsh->lru);
2596
2597#if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2598                for (i = 0; i < nsh->nr_pages; i++) {
2599                        if (nsh->pages[i])
2600                                continue;
2601                        nsh->pages[i] = alloc_page(GFP_NOIO);
2602                        if (!nsh->pages[i])
2603                                err = -ENOMEM;
2604                }
2605
2606                for (i = conf->raid_disks; i < newsize; i++) {
2607                        if (nsh->dev[i].page)
2608                                continue;
2609                        nsh->dev[i].page = raid5_get_dev_page(nsh, i);
2610                        nsh->dev[i].orig_page = nsh->dev[i].page;
2611                        nsh->dev[i].offset = raid5_get_page_offset(nsh, i);
2612                }
2613#else
2614                for (i=conf->raid_disks; i < newsize; i++)
2615                        if (nsh->dev[i].page == NULL) {
2616                                struct page *p = alloc_page(GFP_NOIO);
2617                                nsh->dev[i].page = p;
2618                                nsh->dev[i].orig_page = p;
2619                                nsh->dev[i].offset = 0;
2620                                if (!p)
2621                                        err = -ENOMEM;
2622                        }
2623#endif
2624                raid5_release_stripe(nsh);
2625        }
2626        /* critical section pass, GFP_NOIO no longer needed */
2627
2628        if (!err)
2629                conf->pool_size = newsize;
2630        mutex_unlock(&conf->cache_size_mutex);
2631
2632        return err;
2633}
2634
2635static int drop_one_stripe(struct r5conf *conf)
2636{
2637        struct stripe_head *sh;
2638        int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2639
2640        spin_lock_irq(conf->hash_locks + hash);
2641        sh = get_free_stripe(conf, hash);
2642        spin_unlock_irq(conf->hash_locks + hash);
2643        if (!sh)
2644                return 0;
2645        BUG_ON(atomic_read(&sh->count));
2646        shrink_buffers(sh);
2647        free_stripe(conf->slab_cache, sh);
2648        atomic_dec(&conf->active_stripes);
2649        conf->max_nr_stripes--;
2650        return 1;
2651}
2652
2653static void shrink_stripes(struct r5conf *conf)
2654{
2655        while (conf->max_nr_stripes &&
2656               drop_one_stripe(conf))
2657                ;
2658
2659        kmem_cache_destroy(conf->slab_cache);
2660        conf->slab_cache = NULL;
2661}
2662
2663static void raid5_end_read_request(struct bio * bi)
2664{
2665        struct stripe_head *sh = bi->bi_private;
2666        struct r5conf *conf = sh->raid_conf;
2667        int disks = sh->disks, i;
2668        char b[BDEVNAME_SIZE];
2669        struct md_rdev *rdev = NULL;
2670        sector_t s;
2671
2672        for (i=0 ; i<disks; i++)
2673                if (bi == &sh->dev[i].req)
2674                        break;
2675
2676        pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2677                (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2678                bi->bi_status);
2679        if (i == disks) {
2680                bio_reset(bi);
2681                BUG();
2682                return;
2683        }
2684        if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2685                /* If replacement finished while this request was outstanding,
2686                 * 'replacement' might be NULL already.
2687                 * In that case it moved down to 'rdev'.
2688                 * rdev is not removed until all requests are finished.
2689                 */
2690                rdev = conf->disks[i].replacement;
2691        if (!rdev)
2692                rdev = conf->disks[i].rdev;
2693
2694        if (use_new_offset(conf, sh))
2695                s = sh->sector + rdev->new_data_offset;
2696        else
2697                s = sh->sector + rdev->data_offset;
2698        if (!bi->bi_status) {
2699                set_bit(R5_UPTODATE, &sh->dev[i].flags);
2700                if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2701                        /* Note that this cannot happen on a
2702                         * replacement device.  We just fail those on
2703                         * any error
2704                         */
2705                        pr_info_ratelimited(
2706                                "md/raid:%s: read error corrected (%lu sectors at %llu on %s)\n",
2707                                mdname(conf->mddev), RAID5_STRIPE_SECTORS(conf),
2708                                (unsigned long long)s,
2709                                bdevname(rdev->bdev, b));
2710                        atomic_add(RAID5_STRIPE_SECTORS(conf), &rdev->corrected_errors);
2711                        clear_bit(R5_ReadError, &sh->dev[i].flags);
2712                        clear_bit(R5_ReWrite, &sh->dev[i].flags);
2713                } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2714                        clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2715
2716                if (test_bit(R5_InJournal, &sh->dev[i].flags))
2717                        /*
2718                         * end read for a page in journal, this
2719                         * must be preparing for prexor in rmw
2720                         */
2721                        set_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2722
2723                if (atomic_read(&rdev->read_errors))
2724                        atomic_set(&rdev->read_errors, 0);
2725        } else {
2726                const char *bdn = bdevname(rdev->bdev, b);
2727                int retry = 0;
2728                int set_bad = 0;
2729
2730                clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2731                if (!(bi->bi_status == BLK_STS_PROTECTION))
2732                        atomic_inc(&rdev->read_errors);
2733                if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2734                        pr_warn_ratelimited(
2735                                "md/raid:%s: read error on replacement device (sector %llu on %s).\n",
2736                                mdname(conf->mddev),
2737                                (unsigned long long)s,
2738                                bdn);
2739                else if (conf->mddev->degraded >= conf->max_degraded) {
2740                        set_bad = 1;
2741                        pr_warn_ratelimited(
2742                                "md/raid:%s: read error not correctable (sector %llu on %s).\n",
2743                                mdname(conf->mddev),
2744                                (unsigned long long)s,
2745                                bdn);
2746                } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2747                        /* Oh, no!!! */
2748                        set_bad = 1;
2749                        pr_warn_ratelimited(
2750                                "md/raid:%s: read error NOT corrected!! (sector %llu on %s).\n",
2751                                mdname(conf->mddev),
2752                                (unsigned long long)s,
2753                                bdn);
2754                } else if (atomic_read(&rdev->read_errors)
2755                         > conf->max_nr_stripes) {
2756                        if (!test_bit(Faulty, &rdev->flags)) {
2757                                pr_warn("md/raid:%s: %d read_errors > %d stripes\n",
2758                                    mdname(conf->mddev),
2759                                    atomic_read(&rdev->read_errors),
2760                                    conf->max_nr_stripes);
2761                                pr_warn("md/raid:%s: Too many read errors, failing device %s.\n",
2762                                    mdname(conf->mddev), bdn);
2763                        }
2764                } else
2765                        retry = 1;
2766                if (set_bad && test_bit(In_sync, &rdev->flags)
2767                    && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2768                        retry = 1;
2769                if (retry)
2770                        if (sh->qd_idx >= 0 && sh->pd_idx == i)
2771                                set_bit(R5_ReadError, &sh->dev[i].flags);
2772                        else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2773                                set_bit(R5_ReadError, &sh->dev[i].flags);
2774                                clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2775                        } else
2776                                set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2777                else {
2778                        clear_bit(R5_ReadError, &sh->dev[i].flags);
2779                        clear_bit(R5_ReWrite, &sh->dev[i].flags);
2780                        if (!(set_bad
2781                              && test_bit(In_sync, &rdev->flags)
2782                              && rdev_set_badblocks(
2783                                      rdev, sh->sector, RAID5_STRIPE_SECTORS(conf), 0)))
2784                                md_error(conf->mddev, rdev);
2785                }
2786        }
2787        rdev_dec_pending(rdev, conf->mddev);
2788        bio_reset(bi);
2789        clear_bit(R5_LOCKED, &sh->dev[i].flags);
2790        set_bit(STRIPE_HANDLE, &sh->state);
2791        raid5_release_stripe(sh);
2792}
2793
2794static void raid5_end_write_request(struct bio *bi)
2795{
2796        struct stripe_head *sh = bi->bi_private;
2797        struct r5conf *conf = sh->raid_conf;
2798        int disks = sh->disks, i;
2799        struct md_rdev *rdev;
2800        sector_t first_bad;
2801        int bad_sectors;
2802        int replacement = 0;
2803
2804        for (i = 0 ; i < disks; i++) {
2805                if (bi == &sh->dev[i].req) {
2806                        rdev = conf->disks[i].rdev;
2807                        break;
2808                }
2809                if (bi == &sh->dev[i].rreq) {
2810                        rdev = conf->disks[i].replacement;
2811                        if (rdev)
2812                                replacement = 1;
2813                        else
2814                                /* rdev was removed and 'replacement'
2815                                 * replaced it.  rdev is not removed
2816                                 * until all requests are finished.
2817                                 */
2818                                rdev = conf->disks[i].rdev;
2819                        break;
2820                }
2821        }
2822        pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2823                (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2824                bi->bi_status);
2825        if (i == disks) {
2826                bio_reset(bi);
2827                BUG();
2828                return;
2829        }
2830
2831        if (replacement) {
2832                if (bi->bi_status)
2833                        md_error(conf->mddev, rdev);
2834                else if (is_badblock(rdev, sh->sector,
2835                                     RAID5_STRIPE_SECTORS(conf),
2836                                     &first_bad, &bad_sectors))
2837                        set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2838        } else {
2839                if (bi->bi_status) {
2840                        set_bit(STRIPE_DEGRADED, &sh->state);
2841                        set_bit(WriteErrorSeen, &rdev->flags);
2842                        set_bit(R5_WriteError, &sh->dev[i].flags);
2843                        if (!test_and_set_bit(WantReplacement, &rdev->flags))
2844                                set_bit(MD_RECOVERY_NEEDED,
2845                                        &rdev->mddev->recovery);
2846                } else if (is_badblock(rdev, sh->sector,
2847                                       RAID5_STRIPE_SECTORS(conf),
2848                                       &first_bad, &bad_sectors)) {
2849                        set_bit(R5_MadeGood, &sh->dev[i].flags);
2850                        if (test_bit(R5_ReadError, &sh->dev[i].flags))
2851                                /* That was a successful write so make
2852                                 * sure it looks like we already did
2853                                 * a re-write.
2854                                 */
2855                                set_bit(R5_ReWrite, &sh->dev[i].flags);
2856                }
2857        }
2858        rdev_dec_pending(rdev, conf->mddev);
2859
2860        if (sh->batch_head && bi->bi_status && !replacement)
2861                set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2862
2863        bio_reset(bi);
2864        if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2865                clear_bit(R5_LOCKED, &sh->dev[i].flags);
2866        set_bit(STRIPE_HANDLE, &sh->state);
2867        raid5_release_stripe(sh);
2868
2869        if (sh->batch_head && sh != sh->batch_head)
2870                raid5_release_stripe(sh->batch_head);
2871}
2872
2873static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2874{
2875        char b[BDEVNAME_SIZE];
2876        struct r5conf *conf = mddev->private;
2877        unsigned long flags;
2878        pr_debug("raid456: error called\n");
2879
2880        spin_lock_irqsave(&conf->device_lock, flags);
2881
2882        if (test_bit(In_sync, &rdev->flags) &&
2883            mddev->degraded == conf->max_degraded) {
2884                /*
2885                 * Don't allow to achieve failed state
2886                 * Don't try to recover this device
2887                 */
2888                conf->recovery_disabled = mddev->recovery_disabled;
2889                spin_unlock_irqrestore(&conf->device_lock, flags);
2890                return;
2891        }
2892
2893        set_bit(Faulty, &rdev->flags);
2894        clear_bit(In_sync, &rdev->flags);
2895        mddev->degraded = raid5_calc_degraded(conf);
2896        spin_unlock_irqrestore(&conf->device_lock, flags);
2897        set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2898
2899        set_bit(Blocked, &rdev->flags);
2900        set_mask_bits(&mddev->sb_flags, 0,
2901                      BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2902        pr_crit("md/raid:%s: Disk failure on %s, disabling device.\n"
2903                "md/raid:%s: Operation continuing on %d devices.\n",
2904                mdname(mddev),
2905                bdevname(rdev->bdev, b),
2906                mdname(mddev),
2907                conf->raid_disks - mddev->degraded);
2908        r5c_update_on_rdev_error(mddev, rdev);
2909}
2910
2911/*
2912 * Input: a 'big' sector number,
2913 * Output: index of the data and parity disk, and the sector # in them.
2914 */
2915sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2916                              int previous, int *dd_idx,
2917                              struct stripe_head *sh)
2918{
2919        sector_t stripe, stripe2;
2920        sector_t chunk_number;
2921        unsigned int chunk_offset;
2922        int pd_idx, qd_idx;
2923        int ddf_layout = 0;
2924        sector_t new_sector;
2925        int algorithm = previous ? conf->prev_algo
2926                                 : conf->algorithm;
2927        int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2928                                         : conf->chunk_sectors;
2929        int raid_disks = previous ? conf->previous_raid_disks
2930                                  : conf->raid_disks;
2931        int data_disks = raid_disks - conf->max_degraded;
2932
2933        /* First compute the information on this sector */
2934
2935        /*
2936         * Compute the chunk number and the sector offset inside the chunk
2937         */
2938        chunk_offset = sector_div(r_sector, sectors_per_chunk);
2939        chunk_number = r_sector;
2940
2941        /*
2942         * Compute the stripe number
2943         */
2944        stripe = chunk_number;
2945        *dd_idx = sector_div(stripe, data_disks);
2946        stripe2 = stripe;
2947        /*
2948         * Select the parity disk based on the user selected algorithm.
2949         */
2950        pd_idx = qd_idx = -1;
2951        switch(conf->level) {
2952        case 4:
2953                pd_idx = data_disks;
2954                break;
2955        case 5:
2956                switch (algorithm) {
2957                case ALGORITHM_LEFT_ASYMMETRIC:
2958                        pd_idx = data_disks - sector_div(stripe2, raid_disks);
2959                        if (*dd_idx >= pd_idx)
2960                                (*dd_idx)++;
2961                        break;
2962                case ALGORITHM_RIGHT_ASYMMETRIC:
2963                        pd_idx = sector_div(stripe2, raid_disks);
2964                        if (*dd_idx >= pd_idx)
2965                                (*dd_idx)++;
2966                        break;
2967                case ALGORITHM_LEFT_SYMMETRIC:
2968                        pd_idx = data_disks - sector_div(stripe2, raid_disks);
2969                        *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2970                        break;
2971                case ALGORITHM_RIGHT_SYMMETRIC:
2972                        pd_idx = sector_div(stripe2, raid_disks);
2973                        *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2974                        break;
2975                case ALGORITHM_PARITY_0:
2976                        pd_idx = 0;
2977                        (*dd_idx)++;
2978                        break;
2979                case ALGORITHM_PARITY_N:
2980                        pd_idx = data_disks;
2981                        break;
2982                default:
2983                        BUG();
2984                }
2985                break;
2986        case 6:
2987
2988                switch (algorithm) {
2989                case ALGORITHM_LEFT_ASYMMETRIC:
2990                        pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2991                        qd_idx = pd_idx + 1;
2992                        if (pd_idx == raid_disks-1) {
2993                                (*dd_idx)++;    /* Q D D D P */
2994                                qd_idx = 0;
2995                        } else if (*dd_idx >= pd_idx)
2996                                (*dd_idx) += 2; /* D D P Q D */
2997                        break;
2998                case ALGORITHM_RIGHT_ASYMMETRIC:
2999                        pd_idx = sector_div(stripe2, raid_disks);
3000                        qd_idx = pd_idx + 1;
3001                        if (pd_idx == raid_disks-1) {
3002                                (*dd_idx)++;    /* Q D D D P */
3003                                qd_idx = 0;
3004                        } else if (*dd_idx >= pd_idx)
3005                                (*dd_idx) += 2; /* D D P Q D */
3006                        break;
3007                case ALGORITHM_LEFT_SYMMETRIC:
3008                        pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3009                        qd_idx = (pd_idx + 1) % raid_disks;
3010                        *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
3011                        break;
3012                case ALGORITHM_RIGHT_SYMMETRIC:
3013                        pd_idx = sector_div(stripe2, raid_disks);
3014                        qd_idx = (pd_idx + 1) % raid_disks;
3015                        *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
3016                        break;
3017
3018                case ALGORITHM_PARITY_0:
3019                        pd_idx = 0;
3020                        qd_idx = 1;
3021                        (*dd_idx) += 2;
3022                        break;
3023                case ALGORITHM_PARITY_N:
3024                        pd_idx = data_disks;
3025                        qd_idx = data_disks + 1;
3026                        break;
3027
3028                case ALGORITHM_ROTATING_ZERO_RESTART:
3029                        /* Exactly the same as RIGHT_ASYMMETRIC, but or
3030                         * of blocks for computing Q is different.
3031                         */
3032                        pd_idx = sector_div(stripe2, raid_disks);
3033                        qd_idx = pd_idx + 1;
3034                        if (pd_idx == raid_disks-1) {
3035                                (*dd_idx)++;    /* Q D D D P */
3036                                qd_idx = 0;
3037                        } else if (*dd_idx >= pd_idx)
3038                                (*dd_idx) += 2; /* D D P Q D */
3039                        ddf_layout = 1;
3040                        break;
3041
3042                case ALGORITHM_ROTATING_N_RESTART:
3043                        /* Same a left_asymmetric, by first stripe is
3044                         * D D D P Q  rather than
3045                         * Q D D D P
3046                         */
3047                        stripe2 += 1;
3048                        pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3049                        qd_idx = pd_idx + 1;
3050                        if (pd_idx == raid_disks-1) {
3051                                (*dd_idx)++;    /* Q D D D P */
3052                                qd_idx = 0;
3053                        } else if (*dd_idx >= pd_idx)
3054                                (*dd_idx) += 2; /* D D P Q D */
3055                        ddf_layout = 1;
3056                        break;
3057
3058                case ALGORITHM_ROTATING_N_CONTINUE:
3059                        /* Same as left_symmetric but Q is before P */
3060                        pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3061                        qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
3062                        *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
3063                        ddf_layout = 1;
3064                        break;
3065
3066                case ALGORITHM_LEFT_ASYMMETRIC_6:
3067                        /* RAID5 left_asymmetric, with Q on last device */
3068                        pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
3069                        if (*dd_idx >= pd_idx)
3070                                (*dd_idx)++;
3071                        qd_idx = raid_disks - 1;
3072                        break;
3073
3074                case ALGORITHM_RIGHT_ASYMMETRIC_6:
3075                        pd_idx = sector_div(stripe2, raid_disks-1);
3076                        if (*dd_idx >= pd_idx)
3077                                (*dd_idx)++;
3078                        qd_idx = raid_disks - 1;
3079                        break;
3080
3081                case ALGORITHM_LEFT_SYMMETRIC_6:
3082                        pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
3083                        *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
3084                        qd_idx = raid_disks - 1;
3085                        break;
3086
3087                case ALGORITHM_RIGHT_SYMMETRIC_6:
3088                        pd_idx = sector_div(stripe2, raid_disks-1);
3089                        *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
3090                        qd_idx = raid_disks - 1;
3091                        break;
3092
3093                case ALGORITHM_PARITY_0_6:
3094                        pd_idx = 0;
3095                        (*dd_idx)++;
3096                        qd_idx = raid_disks - 1;
3097                        break;
3098
3099                default:
3100                        BUG();
3101                }
3102                break;
3103        }
3104
3105        if (sh) {
3106                sh->pd_idx = pd_idx;
3107                sh->qd_idx = qd_idx;
3108                sh->ddf_layout = ddf_layout;
3109        }
3110        /*
3111         * Finally, compute the new sector number
3112         */
3113        new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
3114        return new_sector;
3115}
3116
3117sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
3118{
3119        struct r5conf *conf = sh->raid_conf;
3120        int raid_disks = sh->disks;
3121        int data_disks = raid_disks - conf->max_degraded;
3122        sector_t new_sector = sh->sector, check;
3123        int sectors_per_chunk = previous ? conf->prev_chunk_sectors
3124                                         : conf->chunk_sectors;
3125        int algorithm = previous ? conf->prev_algo
3126                                 : conf->algorithm;
3127        sector_t stripe;
3128        int chunk_offset;
3129        sector_t chunk_number;
3130        int dummy1, dd_idx = i;
3131        sector_t r_sector;
3132        struct stripe_head sh2;
3133
3134        chunk_offset = sector_div(new_sector, sectors_per_chunk);
3135        stripe = new_sector;
3136
3137        if (i == sh->pd_idx)
3138                return 0;
3139        switch(conf->level) {
3140        case 4: break;
3141        case 5:
3142                switch (algorithm) {
3143                case ALGORITHM_LEFT_ASYMMETRIC:
3144                case ALGORITHM_RIGHT_ASYMMETRIC:
3145                        if (i > sh->pd_idx)
3146                                i--;
3147                        break;
3148                case ALGORITHM_LEFT_SYMMETRIC:
3149                case ALGORITHM_RIGHT_SYMMETRIC:
3150                        if (i < sh->pd_idx)
3151                                i += raid_disks;
3152                        i -= (sh->pd_idx + 1);
3153                        break;
3154                case ALGORITHM_PARITY_0:
3155                        i -= 1;
3156                        break;
3157                case ALGORITHM_PARITY_N:
3158                        break;
3159                default:
3160                        BUG();
3161                }
3162                break;
3163        case 6:
3164                if (i == sh->qd_idx)
3165                        return 0; /* It is the Q disk */
3166                switch (algorithm) {
3167                case ALGORITHM_LEFT_ASYMMETRIC:
3168                case ALGORITHM_RIGHT_ASYMMETRIC:
3169                case ALGORITHM_ROTATING_ZERO_RESTART:
3170                case ALGORITHM_ROTATING_N_RESTART:
3171                        if (sh->pd_idx == raid_disks-1)
3172                                i--;    /* Q D D D P */
3173                        else if (i > sh->pd_idx)
3174                                i -= 2; /* D D P Q D */
3175                        break;
3176                case ALGORITHM_LEFT_SYMMETRIC:
3177                case ALGORITHM_RIGHT_SYMMETRIC:
3178                        if (sh->pd_idx == raid_disks-1)
3179                                i--; /* Q D D D P */
3180                        else {
3181                                /* D D P Q D */
3182                                if (i < sh->pd_idx)
3183                                        i += raid_disks;
3184                                i -= (sh->pd_idx + 2);
3185                        }
3186                        break;
3187                case ALGORITHM_PARITY_0:
3188                        i -= 2;
3189                        break;
3190                case ALGORITHM_PARITY_N:
3191                        break;
3192                case ALGORITHM_ROTATING_N_CONTINUE:
3193                        /* Like left_symmetric, but P is before Q */
3194                        if (sh->pd_idx == 0)
3195                                i--;    /* P D D D Q */
3196                        else {
3197                                /* D D Q P D */
3198                                if (i < sh->pd_idx)
3199                                        i += raid_disks;
3200                                i -= (sh->pd_idx + 1);
3201                        }
3202                        break;
3203                case ALGORITHM_LEFT_ASYMMETRIC_6:
3204                case ALGORITHM_RIGHT_ASYMMETRIC_6:
3205                        if (i > sh->pd_idx)
3206                                i--;
3207                        break;
3208                case ALGORITHM_LEFT_SYMMETRIC_6:
3209                case ALGORITHM_RIGHT_SYMMETRIC_6:
3210                        if (i < sh->pd_idx)
3211                                i += data_disks + 1;
3212                        i -= (sh->pd_idx + 1);
3213                        break;
3214                case ALGORITHM_PARITY_0_6:
3215                        i -= 1;
3216                        break;
3217                default:
3218                        BUG();
3219                }
3220                break;
3221        }
3222
3223        chunk_number = stripe * data_disks + i;
3224        r_sector = chunk_number * sectors_per_chunk + chunk_offset;
3225
3226        check = raid5_compute_sector(conf, r_sector,
3227                                     previous, &dummy1, &sh2);
3228        if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
3229                || sh2.qd_idx != sh->qd_idx) {
3230                pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
3231                        mdname(conf->mddev));
3232                return 0;
3233        }
3234        return r_sector;
3235}
3236
3237/*
3238 * There are cases where we want handle_stripe_dirtying() and
3239 * schedule_reconstruction() to delay towrite to some dev of a stripe.
3240 *
3241 * This function checks whether we want to delay the towrite. Specifically,
3242 * we delay the towrite when:
3243 *
3244 *   1. degraded stripe has a non-overwrite to the missing dev, AND this
3245 *      stripe has data in journal (for other devices).
3246 *
3247 *      In this case, when reading data for the non-overwrite dev, it is
3248 *      necessary to handle complex rmw of write back cache (prexor with
3249 *      orig_page, and xor with page). To keep read path simple, we would
3250 *      like to flush data in journal to RAID disks first, so complex rmw
3251 *      is handled in the write patch (handle_stripe_dirtying).
3252 *
3253 *   2. when journal space is critical (R5C_LOG_CRITICAL=1)
3254 *
3255 *      It is important to be able to flush all stripes in raid5-cache.
3256 *      Therefore, we need reserve some space on the journal device for
3257 *      these flushes. If flush operation includes pending writes to the
3258 *      stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
3259 *      for the flush out. If we exclude these pending writes from flush
3260 *      operation, we only need (conf->max_degraded + 1) pages per stripe.
3261 *      Therefore, excluding pending writes in these cases enables more
3262 *      efficient use of the journal device.
3263 *
3264 *      Note: To make sure the stripe makes progress, we only delay
3265 *      towrite for stripes with data already in journal (injournal > 0).
3266 *      When LOG_CRITICAL, stripes with injournal == 0 will be sent to
3267 *      no_space_stripes list.
3268 *
3269 *   3. during journal failure
3270 *      In journal failure, we try to flush all cached data to raid disks
3271 *      based on data in stripe cache. The array is read-only to upper
3272 *      layers, so we would skip all pending writes.
3273 *
3274 */
3275static inline bool delay_towrite(struct r5conf *conf,
3276                                 struct r5dev *dev,
3277                                 struct stripe_head_state *s)
3278{
3279        /* case 1 above */
3280        if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3281            !test_bit(R5_Insync, &dev->flags) && s->injournal)
3282                return true;
3283        /* case 2 above */
3284        if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
3285            s->injournal > 0)
3286                return true;
3287        /* case 3 above */
3288        if (s->log_failed && s->injournal)
3289                return true;
3290        return false;
3291}
3292
3293static void
3294schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
3295                         int rcw, int expand)
3296{
3297        int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
3298        struct r5conf *conf = sh->raid_conf;
3299        int level = conf->level;
3300
3301        if (rcw) {
3302                /*
3303                 * In some cases, handle_stripe_dirtying initially decided to
3304                 * run rmw and allocates extra page for prexor. However, rcw is
3305                 * cheaper later on. We need to free the extra page now,
3306                 * because we won't be able to do that in ops_complete_prexor().
3307                 */
3308                r5c_release_extra_page(sh);
3309
3310                for (i = disks; i--; ) {
3311                        struct r5dev *dev = &sh->dev[i];
3312
3313                        if (dev->towrite && !delay_towrite(conf, dev, s)) {
3314                                set_bit(R5_LOCKED, &dev->flags);
3315                                set_bit(R5_Wantdrain, &dev->flags);
3316                                if (!expand)
3317                                        clear_bit(R5_UPTODATE, &dev->flags);
3318                                s->locked++;
3319                        } else if (test_bit(R5_InJournal, &dev->flags)) {
3320                                set_bit(R5_LOCKED, &dev->flags);
3321                                s->locked++;
3322                        }
3323                }
3324                /* if we are not expanding this is a proper write request, and
3325                 * there will be bios with new data to be drained into the
3326                 * stripe cache
3327                 */
3328                if (!expand) {
3329                        if (!s->locked)
3330                                /* False alarm, nothing to do */
3331                                return;
3332                        sh->reconstruct_state = reconstruct_state_drain_run;
3333                        set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3334                } else
3335                        sh->reconstruct_state = reconstruct_state_run;
3336
3337                set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3338
3339                if (s->locked + conf->max_degraded == disks)
3340                        if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
3341                                atomic_inc(&conf->pending_full_writes);
3342        } else {
3343                BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
3344                        test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
3345                BUG_ON(level == 6 &&
3346                        (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
3347                           test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
3348
3349                for (i = disks; i--; ) {
3350                        struct r5dev *dev = &sh->dev[i];
3351                        if (i == pd_idx || i == qd_idx)
3352                                continue;
3353
3354                        if (dev->towrite &&
3355                            (test_bit(R5_UPTODATE, &dev->flags) ||
3356                             test_bit(R5_Wantcompute, &dev->flags))) {
3357                                set_bit(R5_Wantdrain, &dev->flags);
3358                                set_bit(R5_LOCKED, &dev->flags);
3359                                clear_bit(R5_UPTODATE, &dev->flags);
3360                                s->locked++;
3361                        } else if (test_bit(R5_InJournal, &dev->flags)) {
3362                                set_bit(R5_LOCKED, &dev->flags);
3363                                s->locked++;
3364                        }
3365                }
3366                if (!s->locked)
3367                        /* False alarm - nothing to do */
3368                        return;
3369                sh->reconstruct_state = reconstruct_state_prexor_drain_run;
3370                set_bit(STRIPE_OP_PREXOR, &s->ops_request);
3371                set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3372                set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3373        }
3374
3375        /* keep the parity disk(s) locked while asynchronous operations
3376         * are in flight
3377         */
3378        set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
3379        clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3380        s->locked++;
3381
3382        if (level == 6) {
3383                int qd_idx = sh->qd_idx;
3384                struct r5dev *dev = &sh->dev[qd_idx];
3385
3386                set_bit(R5_LOCKED, &dev->flags);
3387                clear_bit(R5_UPTODATE, &dev->flags);
3388                s->locked++;
3389        }
3390
3391        if (raid5_has_ppl(sh->raid_conf) && sh->ppl_page &&
3392            test_bit(STRIPE_OP_BIODRAIN, &s->ops_request) &&
3393            !test_bit(STRIPE_FULL_WRITE, &sh->state) &&
3394            test_bit(R5_Insync, &sh->dev[pd_idx].flags))
3395                set_bit(STRIPE_OP_PARTIAL_PARITY, &s->ops_request);
3396
3397        pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3398                __func__, (unsigned long long)sh->sector,
3399                s->locked, s->ops_request);
3400}
3401
3402/*
3403 * Each stripe/dev can have one or more bion attached.
3404 * toread/towrite point to the first in a chain.
3405 * The bi_next chain must be in order.
3406 */
3407static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
3408                          int forwrite, int previous)
3409{
3410        struct bio **bip;
3411        struct r5conf *conf = sh->raid_conf;
3412        int firstwrite=0;
3413
3414        pr_debug("adding bi b#%llu to stripe s#%llu\n",
3415                (unsigned long long)bi->bi_iter.bi_sector,
3416                (unsigned long long)sh->sector);
3417
3418        spin_lock_irq(&sh->stripe_lock);
3419        sh->dev[dd_idx].write_hint = bi->bi_write_hint;
3420        /* Don't allow new IO added to stripes in batch list */
3421        if (sh->batch_head)
3422                goto overlap;
3423        if (forwrite) {
3424                bip = &sh->dev[dd_idx].towrite;
3425                if (*bip == NULL)
3426                        firstwrite = 1;
3427        } else
3428                bip = &sh->dev[dd_idx].toread;
3429        while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
3430                if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3431                        goto overlap;
3432                bip = & (*bip)->bi_next;
3433        }
3434        if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3435                goto overlap;
3436
3437        if (forwrite && raid5_has_ppl(conf)) {
3438                /*
3439                 * With PPL only writes to consecutive data chunks within a
3440                 * stripe are allowed because for a single stripe_head we can
3441                 * only have one PPL entry at a time, which describes one data
3442                 * range. Not really an overlap, but wait_for_overlap can be
3443                 * used to handle this.
3444                 */
3445                sector_t sector;
3446                sector_t first = 0;
3447                sector_t last = 0;
3448                int count = 0;
3449                int i;
3450
3451                for (i = 0; i < sh->disks; i++) {
3452                        if (i != sh->pd_idx &&
3453                            (i == dd_idx || sh->dev[i].towrite)) {
3454                                sector = sh->dev[i].sector;
3455                                if (count == 0 || sector < first)
3456                                        first = sector;
3457                                if (sector > last)
3458                                        last = sector;
3459                                count++;
3460                        }
3461                }
3462
3463                if (first + conf->chunk_sectors * (count - 1) != last)
3464                        goto overlap;
3465        }
3466
3467        if (!forwrite || previous)
3468                clear_bit(STRIPE_BATCH_READY, &sh->state);
3469
3470        BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3471        if (*bip)
3472                bi->bi_next = *bip;
3473        *bip = bi;
3474        bio_inc_remaining(bi);
3475        md_write_inc(conf->mddev, bi);
3476
3477        if (forwrite) {
3478                /* check if page is covered */
3479                sector_t sector = sh->dev[dd_idx].sector;
3480                for (bi=sh->dev[dd_idx].towrite;
3481                     sector < sh->dev[dd_idx].sector + RAID5_STRIPE_SECTORS(conf) &&
3482                             bi && bi->bi_iter.bi_sector <= sector;
3483                     bi = r5_next_bio(conf, bi, sh->dev[dd_idx].sector)) {
3484                        if (bio_end_sector(bi) >= sector)
3485                                sector = bio_end_sector(bi);
3486                }
3487                if (sector >= sh->dev[dd_idx].sector + RAID5_STRIPE_SECTORS(conf))
3488                        if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3489                                sh->overwrite_disks++;
3490        }
3491
3492        pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3493                (unsigned long long)(*bip)->bi_iter.bi_sector,
3494                (unsigned long long)sh->sector, dd_idx);
3495
3496        if (conf->mddev->bitmap && firstwrite) {
3497                /* Cannot hold spinlock over bitmap_startwrite,
3498                 * but must ensure this isn't added to a batch until
3499                 * we have added to the bitmap and set bm_seq.
3500                 * So set STRIPE_BITMAP_PENDING to prevent
3501                 * batching.
3502                 * If multiple add_stripe_bio() calls race here they
3503                 * much all set STRIPE_BITMAP_PENDING.  So only the first one
3504                 * to complete "bitmap_startwrite" gets to set
3505                 * STRIPE_BIT_DELAY.  This is important as once a stripe
3506                 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3507                 * any more.
3508                 */
3509                set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3510                spin_unlock_irq(&sh->stripe_lock);
3511                md_bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3512                                     RAID5_STRIPE_SECTORS(conf), 0);
3513                spin_lock_irq(&sh->stripe_lock);
3514                clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3515                if (!sh->batch_head) {
3516                        sh->bm_seq = conf->seq_flush+1;
3517                        set_bit(STRIPE_BIT_DELAY, &sh->state);
3518                }
3519        }
3520        spin_unlock_irq(&sh->stripe_lock);
3521
3522        if (stripe_can_batch(sh))
3523                stripe_add_to_batch_list(conf, sh);
3524        return 1;
3525
3526 overlap:
3527        set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3528        spin_unlock_irq(&sh->stripe_lock);
3529        return 0;
3530}
3531
3532static void end_reshape(struct r5conf *conf);
3533
3534static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3535                            struct stripe_head *sh)
3536{
3537        int sectors_per_chunk =
3538                previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3539        int dd_idx;
3540        int chunk_offset = sector_div(stripe, sectors_per_chunk);
3541        int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3542
3543        raid5_compute_sector(conf,
3544                             stripe * (disks - conf->max_degraded)
3545                             *sectors_per_chunk + chunk_offset,
3546                             previous,
3547                             &dd_idx, sh);
3548}
3549
3550static void
3551handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3552                     struct stripe_head_state *s, int disks)
3553{
3554        int i;
3555        BUG_ON(sh->batch_head);
3556        for (i = disks; i--; ) {
3557                struct bio *bi;
3558                int bitmap_end = 0;
3559
3560                if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3561                        struct md_rdev *rdev;
3562                        rcu_read_lock();
3563                        rdev = rcu_dereference(conf->disks[i].rdev);
3564                        if (rdev && test_bit(In_sync, &rdev->flags) &&
3565                            !test_bit(Faulty, &rdev->flags))
3566                                atomic_inc(&rdev->nr_pending);
3567                        else
3568                                rdev = NULL;
3569                        rcu_read_unlock();
3570                        if (rdev) {
3571                                if (!rdev_set_badblocks(
3572                                            rdev,
3573                                            sh->sector,
3574                                            RAID5_STRIPE_SECTORS(conf), 0))
3575                                        md_error(conf->mddev, rdev);
3576                                rdev_dec_pending(rdev, conf->mddev);
3577                        }
3578                }
3579                spin_lock_irq(&sh->stripe_lock);
3580                /* fail all writes first */
3581                bi = sh->dev[i].towrite;
3582                sh->dev[i].towrite = NULL;
3583                sh->overwrite_disks = 0;
3584                spin_unlock_irq(&sh->stripe_lock);
3585                if (bi)
3586                        bitmap_end = 1;
3587
3588                log_stripe_write_finished(sh);
3589
3590                if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3591                        wake_up(&conf->wait_for_overlap);
3592
3593                while (bi && bi->bi_iter.bi_sector <
3594                        sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3595                        struct bio *nextbi = r5_next_bio(conf, bi, sh->dev[i].sector);
3596
3597                        md_write_end(conf->mddev);
3598                        bio_io_error(bi);
3599                        bi = nextbi;
3600                }
3601                if (bitmap_end)
3602                        md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3603                                           RAID5_STRIPE_SECTORS(conf), 0, 0);
3604                bitmap_end = 0;
3605                /* and fail all 'written' */
3606                bi = sh->dev[i].written;
3607                sh->dev[i].written = NULL;
3608                if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3609                        WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3610                        sh->dev[i].page = sh->dev[i].orig_page;
3611                }
3612
3613                if (bi) bitmap_end = 1;
3614                while (bi && bi->bi_iter.bi_sector <
3615                       sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3616                        struct bio *bi2 = r5_next_bio(conf, bi, sh->dev[i].sector);
3617
3618                        md_write_end(conf->mddev);
3619                        bio_io_error(bi);
3620                        bi = bi2;
3621                }
3622
3623                /* fail any reads if this device is non-operational and
3624                 * the data has not reached the cache yet.
3625                 */
3626                if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3627                    s->failed > conf->max_degraded &&
3628                    (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3629                      test_bit(R5_ReadError, &sh->dev[i].flags))) {
3630                        spin_lock_irq(&sh->stripe_lock);
3631                        bi = sh->dev[i].toread;
3632                        sh->dev[i].toread = NULL;
3633                        spin_unlock_irq(&sh->stripe_lock);
3634                        if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3635                                wake_up(&conf->wait_for_overlap);
3636                        if (bi)
3637                                s->to_read--;
3638                        while (bi && bi->bi_iter.bi_sector <
3639                               sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3640                                struct bio *nextbi =
3641                                        r5_next_bio(conf, bi, sh->dev[i].sector);
3642
3643                                bio_io_error(bi);
3644                                bi = nextbi;
3645                        }
3646                }
3647                if (bitmap_end)
3648                        md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3649                                           RAID5_STRIPE_SECTORS(conf), 0, 0);
3650                /* If we were in the middle of a write the parity block might
3651                 * still be locked - so just clear all R5_LOCKED flags
3652                 */
3653                clear_bit(R5_LOCKED, &sh->dev[i].flags);
3654        }
3655        s->to_write = 0;
3656        s->written = 0;
3657
3658        if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3659                if (atomic_dec_and_test(&conf->pending_full_writes))
3660                        md_wakeup_thread(conf->mddev->thread);
3661}
3662
3663static void
3664handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3665                   struct stripe_head_state *s)
3666{
3667        int abort = 0;
3668        int i;
3669
3670        BUG_ON(sh->batch_head);
3671        clear_bit(STRIPE_SYNCING, &sh->state);
3672        if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3673                wake_up(&conf->wait_for_overlap);
3674        s->syncing = 0;
3675        s->replacing = 0;
3676        /* There is nothing more to do for sync/check/repair.
3677         * Don't even need to abort as that is handled elsewhere
3678         * if needed, and not always wanted e.g. if there is a known
3679         * bad block here.
3680         * For recover/replace we need to record a bad block on all
3681         * non-sync devices, or abort the recovery
3682         */
3683        if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3684                /* During recovery devices cannot be removed, so
3685                 * locking and refcounting of rdevs is not needed
3686                 */
3687                rcu_read_lock();
3688                for (i = 0; i < conf->raid_disks; i++) {
3689                        struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
3690                        if (rdev
3691                            && !test_bit(Faulty, &rdev->flags)
3692                            && !test_bit(In_sync, &rdev->flags)
3693                            && !rdev_set_badblocks(rdev, sh->sector,
3694                                                   RAID5_STRIPE_SECTORS(conf), 0))
3695                                abort = 1;
3696                        rdev = rcu_dereference(conf->disks[i].replacement);
3697                        if (rdev
3698                            && !test_bit(Faulty, &rdev->flags)
3699                            && !test_bit(In_sync, &rdev->flags)
3700                            && !rdev_set_badblocks(rdev, sh->sector,
3701                                                   RAID5_STRIPE_SECTORS(conf), 0))
3702                                abort = 1;
3703                }
3704                rcu_read_unlock();
3705                if (abort)
3706                        conf->recovery_disabled =
3707                                conf->mddev->recovery_disabled;
3708        }
3709        md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), !abort);
3710}
3711
3712static int want_replace(struct stripe_head *sh, int disk_idx)
3713{
3714        struct md_rdev *rdev;
3715        int rv = 0;
3716
3717        rcu_read_lock();
3718        rdev = rcu_dereference(sh->raid_conf->disks[disk_idx].replacement);
3719        if (rdev
3720            && !test_bit(Faulty, &rdev->flags)
3721            && !test_bit(In_sync, &rdev->flags)
3722            && (rdev->recovery_offset <= sh->sector
3723                || rdev->mddev->recovery_cp <= sh->sector))
3724                rv = 1;
3725        rcu_read_unlock();
3726        return rv;
3727}
3728
3729static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3730                           int disk_idx, int disks)
3731{
3732        struct r5dev *dev = &sh->dev[disk_idx];
3733        struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3734                                  &sh->dev[s->failed_num[1]] };
3735        int i;
3736        bool force_rcw = (sh->raid_conf->rmw_level == PARITY_DISABLE_RMW);
3737
3738
3739        if (test_bit(R5_LOCKED, &dev->flags) ||
3740            test_bit(R5_UPTODATE, &dev->flags))
3741                /* No point reading this as we already have it or have
3742                 * decided to get it.
3743                 */
3744                return 0;
3745
3746        if (dev->toread ||
3747            (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3748                /* We need this block to directly satisfy a request */
3749                return 1;
3750
3751        if (s->syncing || s->expanding ||
3752            (s->replacing && want_replace(sh, disk_idx)))
3753                /* When syncing, or expanding we read everything.
3754                 * When replacing, we need the replaced block.
3755                 */
3756                return 1;
3757
3758        if ((s->failed >= 1 && fdev[0]->toread) ||
3759            (s->failed >= 2 && fdev[1]->toread))
3760                /* If we want to read from a failed device, then
3761                 * we need to actually read every other device.
3762                 */
3763                return 1;
3764
3765        /* Sometimes neither read-modify-write nor reconstruct-write
3766         * cycles can work.  In those cases we read every block we
3767         * can.  Then the parity-update is certain to have enough to
3768         * work with.
3769         * This can only be a problem when we need to write something,
3770         * and some device has failed.  If either of those tests
3771         * fail we need look no further.
3772         */
3773        if (!s->failed || !s->to_write)
3774                return 0;
3775
3776        if (test_bit(R5_Insync, &dev->flags) &&
3777            !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3778                /* Pre-reads at not permitted until after short delay
3779                 * to gather multiple requests.  However if this
3780                 * device is no Insync, the block could only be computed
3781                 * and there is no need to delay that.
3782                 */
3783                return 0;
3784
3785        for (i = 0; i < s->failed && i < 2; i++) {
3786                if (fdev[i]->towrite &&
3787                    !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3788                    !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3789                        /* If we have a partial write to a failed
3790                         * device, then we will need to reconstruct
3791                         * the content of that device, so all other
3792                         * devices must be read.
3793                         */
3794                        return 1;
3795
3796                if (s->failed >= 2 &&
3797                    (fdev[i]->towrite ||
3798                     s->failed_num[i] == sh->pd_idx ||
3799                     s->failed_num[i] == sh->qd_idx) &&
3800                    !test_bit(R5_UPTODATE, &fdev[i]->flags))
3801                        /* In max degraded raid6, If the failed disk is P, Q,
3802                         * or we want to read the failed disk, we need to do
3803                         * reconstruct-write.
3804                         */
3805                        force_rcw = true;
3806        }
3807
3808        /* If we are forced to do a reconstruct-write, because parity
3809         * cannot be trusted and we are currently recovering it, there
3810         * is extra need to be careful.
3811         * If one of the devices that we would need to read, because
3812         * it is not being overwritten (and maybe not written at all)
3813         * is missing/faulty, then we need to read everything we can.
3814         */
3815        if (!force_rcw &&
3816            sh->sector < sh->raid_conf->mddev->recovery_cp)
3817                /* reconstruct-write isn't being forced */
3818                return 0;
3819        for (i = 0; i < s->failed && i < 2; i++) {
3820                if (s->failed_num[i] != sh->pd_idx &&
3821                    s->failed_num[i] != sh->qd_idx &&
3822                    !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3823                    !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3824                        return 1;
3825        }
3826
3827        return 0;
3828}
3829
3830/* fetch_block - checks the given member device to see if its data needs
3831 * to be read or computed to satisfy a request.
3832 *
3833 * Returns 1 when no more member devices need to be checked, otherwise returns
3834 * 0 to tell the loop in handle_stripe_fill to continue
3835 */
3836static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3837                       int disk_idx, int disks)
3838{
3839        struct r5dev *dev = &sh->dev[disk_idx];
3840
3841        /* is the data in this block needed, and can we get it? */
3842        if (need_this_block(sh, s, disk_idx, disks)) {
3843                /* we would like to get this block, possibly by computing it,
3844                 * otherwise read it if the backing disk is insync
3845                 */
3846                BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3847                BUG_ON(test_bit(R5_Wantread, &dev->flags));
3848                BUG_ON(sh->batch_head);
3849
3850                /*
3851                 * In the raid6 case if the only non-uptodate disk is P
3852                 * then we already trusted P to compute the other failed
3853                 * drives. It is safe to compute rather than re-read P.
3854                 * In other cases we only compute blocks from failed
3855                 * devices, otherwise check/repair might fail to detect
3856                 * a real inconsistency.
3857                 */
3858
3859                if ((s->uptodate == disks - 1) &&
3860                    ((sh->qd_idx >= 0 && sh->pd_idx == disk_idx) ||
3861                    (s->failed && (disk_idx == s->failed_num[0] ||
3862                                   disk_idx == s->failed_num[1])))) {
3863                        /* have disk failed, and we're requested to fetch it;
3864                         * do compute it
3865                         */
3866                        pr_debug("Computing stripe %llu block %d\n",
3867                               (unsigned long long)sh->sector, disk_idx);
3868                        set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3869                        set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3870                        set_bit(R5_Wantcompute, &dev->flags);
3871                        sh->ops.target = disk_idx;
3872                        sh->ops.target2 = -1; /* no 2nd target */
3873                        s->req_compute = 1;
3874                        /* Careful: from this point on 'uptodate' is in the eye
3875                         * of raid_run_ops which services 'compute' operations
3876                         * before writes. R5_Wantcompute flags a block that will
3877                         * be R5_UPTODATE by the time it is needed for a
3878                         * subsequent operation.
3879                         */
3880                        s->uptodate++;
3881                        return 1;
3882                } else if (s->uptodate == disks-2 && s->failed >= 2) {
3883                        /* Computing 2-failure is *very* expensive; only
3884                         * do it if failed >= 2
3885                         */
3886                        int other;
3887                        for (other = disks; other--; ) {
3888                                if (other == disk_idx)
3889                                        continue;
3890                                if (!test_bit(R5_UPTODATE,
3891                                      &sh->dev[other].flags))
3892                                        break;
3893                        }
3894                        BUG_ON(other < 0);
3895                        pr_debug("Computing stripe %llu blocks %d,%d\n",
3896                               (unsigned long long)sh->sector,
3897                               disk_idx, other);
3898                        set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3899                        set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3900                        set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3901                        set_bit(R5_Wantcompute, &sh->dev[other].flags);
3902                        sh->ops.target = disk_idx;
3903                        sh->ops.target2 = other;
3904                        s->uptodate += 2;
3905                        s->req_compute = 1;
3906                        return 1;
3907                } else if (test_bit(R5_Insync, &dev->flags)) {
3908                        set_bit(R5_LOCKED, &dev->flags);
3909                        set_bit(R5_Wantread, &dev->flags);
3910                        s->locked++;
3911                        pr_debug("Reading block %d (sync=%d)\n",
3912                                disk_idx, s->syncing);
3913                }
3914        }
3915
3916        return 0;
3917}
3918
3919/*
3920 * handle_stripe_fill - read or compute data to satisfy pending requests.
3921 */
3922static void handle_stripe_fill(struct stripe_head *sh,
3923                               struct stripe_head_state *s,
3924                               int disks)
3925{
3926        int i;
3927
3928        /* look for blocks to read/compute, skip this if a compute
3929         * is already in flight, or if the stripe contents are in the
3930         * midst of changing due to a write
3931         */
3932        if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3933            !sh->reconstruct_state) {
3934
3935                /*
3936                 * For degraded stripe with data in journal, do not handle
3937                 * read requests yet, instead, flush the stripe to raid
3938                 * disks first, this avoids handling complex rmw of write
3939                 * back cache (prexor with orig_page, and then xor with
3940                 * page) in the read path
3941                 */
3942                if (s->injournal && s->failed) {
3943                        if (test_bit(STRIPE_R5C_CACHING, &sh->state))
3944                                r5c_make_stripe_write_out(sh);
3945                        goto out;
3946                }
3947
3948                for (i = disks; i--; )
3949                        if (fetch_block(sh, s, i, disks))
3950                                break;
3951        }
3952out:
3953        set_bit(STRIPE_HANDLE, &sh->state);
3954}
3955
3956static void break_stripe_batch_list(struct stripe_head *head_sh,
3957                                    unsigned long handle_flags);
3958/* handle_stripe_clean_event
3959 * any written block on an uptodate or failed drive can be returned.
3960 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3961 * never LOCKED, so we don't need to test 'failed' directly.
3962 */
3963static void handle_stripe_clean_event(struct r5conf *conf,
3964        struct stripe_head *sh, int disks)
3965{
3966        int i;
3967        struct r5dev *dev;
3968        int discard_pending = 0;
3969        struct stripe_head *head_sh = sh;
3970        bool do_endio = false;
3971
3972        for (i = disks; i--; )
3973                if (sh->dev[i].written) {
3974                        dev = &sh->dev[i];
3975                        if (!test_bit(R5_LOCKED, &dev->flags) &&
3976                            (test_bit(R5_UPTODATE, &dev->flags) ||
3977                             test_bit(R5_Discard, &dev->flags) ||
3978                             test_bit(R5_SkipCopy, &dev->flags))) {
3979                                /* We can return any write requests */
3980                                struct bio *wbi, *wbi2;
3981                                pr_debug("Return write for disc %d\n", i);
3982                                if (test_and_clear_bit(R5_Discard, &dev->flags))
3983                                        clear_bit(R5_UPTODATE, &dev->flags);
3984                                if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3985                                        WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3986                                }
3987                                do_endio = true;
3988
3989returnbi:
3990                                dev->page = dev->orig_page;
3991                                wbi = dev->written;
3992                                dev->written = NULL;
3993                                while (wbi && wbi->bi_iter.bi_sector <
3994                                        dev->sector + RAID5_STRIPE_SECTORS(conf)) {
3995                                        wbi2 = r5_next_bio(conf, wbi, dev->sector);
3996                                        md_write_end(conf->mddev);
3997                                        bio_endio(wbi);
3998                                        wbi = wbi2;
3999                                }
4000                                md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
4001                                                   RAID5_STRIPE_SECTORS(conf),
4002                                                   !test_bit(STRIPE_DEGRADED, &sh->state),
4003                                                   0);
4004                                if (head_sh->batch_head) {
4005                                        sh = list_first_entry(&sh->batch_list,
4006                                                              struct stripe_head,
4007                                                              batch_list);
4008                                        if (sh != head_sh) {
4009                                                dev = &sh->dev[i];
4010                                                goto returnbi;
4011                                        }
4012                                }
4013                                sh = head_sh;
4014                                dev = &sh->dev[i];
4015                        } else if (test_bit(R5_Discard, &dev->flags))
4016                                discard_pending = 1;
4017                }
4018
4019        log_stripe_write_finished(sh);
4020
4021        if (!discard_pending &&
4022            test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
4023                int hash;
4024                clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
4025                clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4026                if (sh->qd_idx >= 0) {
4027                        clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
4028                        clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
4029                }
4030                /* now that discard is done we can proceed with any sync */
4031                clear_bit(STRIPE_DISCARD, &sh->state);
4032                /*
4033                 * SCSI discard will change some bio fields and the stripe has
4034                 * no updated data, so remove it from hash list and the stripe
4035                 * will be reinitialized
4036                 */
4037unhash:
4038                hash = sh->hash_lock_index;
4039                spin_lock_irq(conf->hash_locks + hash);
4040                remove_hash(sh);
4041                spin_unlock_irq(conf->hash_locks + hash);
4042                if (head_sh->batch_head) {
4043                        sh = list_first_entry(&sh->batch_list,
4044                                              struct stripe_head, batch_list);
4045                        if (sh != head_sh)
4046                                        goto unhash;
4047                }
4048                sh = head_sh;
4049
4050                if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
4051                        set_bit(STRIPE_HANDLE, &sh->state);
4052
4053        }
4054
4055        if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
4056                if (atomic_dec_and_test(&conf->pending_full_writes))
4057                        md_wakeup_thread(conf->mddev->thread);
4058
4059        if (head_sh->batch_head && do_endio)
4060                break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
4061}
4062
4063/*
4064 * For RMW in write back cache, we need extra page in prexor to store the
4065 * old data. This page is stored in dev->orig_page.
4066 *
4067 * This function checks whether we have data for prexor. The exact logic
4068 * is:
4069 *       R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
4070 */
4071static inline bool uptodate_for_rmw(struct r5dev *dev)
4072{
4073        return (test_bit(R5_UPTODATE, &dev->flags)) &&
4074                (!test_bit(R5_InJournal, &dev->flags) ||
4075                 test_bit(R5_OrigPageUPTDODATE, &dev->flags));
4076}
4077
4078static int handle_stripe_dirtying(struct r5conf *conf,
4079                                  struct stripe_head *sh,
4080                                  struct stripe_head_state *s,
4081                                  int disks)
4082{
4083        int rmw = 0, rcw = 0, i;
4084        sector_t recovery_cp = conf->mddev->recovery_cp;
4085
4086        /* Check whether resync is now happening or should start.
4087         * If yes, then the array is dirty (after unclean shutdown or
4088         * initial creation), so parity in some stripes might be inconsistent.
4089         * In this case, we need to always do reconstruct-write, to ensure
4090         * that in case of drive failure or read-error correction, we
4091         * generate correct data from the parity.
4092         */
4093        if (conf->rmw_level == PARITY_DISABLE_RMW ||
4094            (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
4095             s->failed == 0)) {
4096                /* Calculate the real rcw later - for now make it
4097                 * look like rcw is cheaper
4098                 */
4099                rcw = 1; rmw = 2;
4100                pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
4101                         conf->rmw_level, (unsigned long long)recovery_cp,
4102                         (unsigned long long)sh->sector);
4103        } else for (i = disks; i--; ) {
4104                /* would I have to read this buffer for read_modify_write */
4105                struct r5dev *dev = &sh->dev[i];
4106                if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
4107                     i == sh->pd_idx || i == sh->qd_idx ||
4108                     test_bit(R5_InJournal, &dev->flags)) &&
4109                    !test_bit(R5_LOCKED, &dev->flags) &&
4110                    !(uptodate_for_rmw(dev) ||
4111                      test_bit(R5_Wantcompute, &dev->flags))) {
4112                        if (test_bit(R5_Insync, &dev->flags))
4113                                rmw++;
4114                        else
4115                                rmw += 2*disks;  /* cannot read it */
4116                }
4117                /* Would I have to read this buffer for reconstruct_write */
4118                if (!test_bit(R5_OVERWRITE, &dev->flags) &&
4119                    i != sh->pd_idx && i != sh->qd_idx &&
4120                    !test_bit(R5_LOCKED, &dev->flags) &&
4121                    !(test_bit(R5_UPTODATE, &dev->flags) ||
4122                      test_bit(R5_Wantcompute, &dev->flags))) {
4123                        if (test_bit(R5_Insync, &dev->flags))
4124                                rcw++;
4125                        else
4126                                rcw += 2*disks;
4127                }
4128        }
4129
4130        pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
4131                 (unsigned long long)sh->sector, sh->state, rmw, rcw);
4132        set_bit(STRIPE_HANDLE, &sh->state);
4133        if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
4134                /* prefer read-modify-write, but need to get some data */
4135                if (conf->mddev->queue)
4136                        blk_add_trace_msg(conf->mddev->queue,
4137                                          "raid5 rmw %llu %d",
4138                                          (unsigned long long)sh->sector, rmw);
4139                for (i = disks; i--; ) {
4140                        struct r5dev *dev = &sh->dev[i];
4141                        if (test_bit(R5_InJournal, &dev->flags) &&
4142                            dev->page == dev->orig_page &&
4143                            !test_bit(R5_LOCKED, &sh->dev[sh->pd_idx].flags)) {
4144                                /* alloc page for prexor */
4145                                struct page *p = alloc_page(GFP_NOIO);
4146
4147                                if (p) {
4148                                        dev->orig_page = p;
4149                                        continue;
4150                                }
4151
4152                                /*
4153                                 * alloc_page() failed, try use
4154                                 * disk_info->extra_page
4155                                 */
4156                                if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE,
4157                                                      &conf->cache_state)) {
4158                                        r5c_use_extra_page(sh);
4159                                        break;
4160                                }
4161
4162                                /* extra_page in use, add to delayed_list */
4163                                set_bit(STRIPE_DELAYED, &sh->state);
4164                                s->waiting_extra_page = 1;
4165                                return -EAGAIN;
4166                        }
4167                }
4168
4169                for (i = disks; i--; ) {
4170                        struct r5dev *dev = &sh->dev[i];
4171                        if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
4172                             i == sh->pd_idx || i == sh->qd_idx ||
4173                             test_bit(R5_InJournal, &dev->flags)) &&
4174                            !test_bit(R5_LOCKED, &dev->flags) &&
4175                            !(uptodate_for_rmw(dev) ||
4176                              test_bit(R5_Wantcompute, &dev->flags)) &&
4177                            test_bit(R5_Insync, &dev->flags)) {
4178                                if (test_bit(STRIPE_PREREAD_ACTIVE,
4179                                             &sh->state)) {
4180                                        pr_debug("Read_old block %d for r-m-w\n",
4181                                                 i);
4182                                        set_bit(R5_LOCKED, &dev->flags);
4183                                        set_bit(R5_Wantread, &dev->flags);
4184                                        s->locked++;
4185                                } else
4186                                        set_bit(STRIPE_DELAYED, &sh->state);
4187                        }
4188                }
4189        }
4190        if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
4191                /* want reconstruct write, but need to get some data */
4192                int qread =0;
4193                rcw = 0;
4194                for (i = disks; i--; ) {
4195                        struct r5dev *dev = &sh->dev[i];
4196                        if (!test_bit(R5_OVERWRITE, &dev->flags) &&
4197                            i != sh->pd_idx && i != sh->qd_idx &&
4198                            !test_bit(R5_LOCKED, &dev->flags) &&
4199                            !(test_bit(R5_UPTODATE, &dev->flags) ||
4200                              test_bit(R5_Wantcompute, &dev->flags))) {
4201                                rcw++;
4202                                if (test_bit(R5_Insync, &dev->flags) &&
4203                                    test_bit(STRIPE_PREREAD_ACTIVE,
4204                                             &sh->state)) {
4205                                        pr_debug("Read_old block "
4206                                                "%d for Reconstruct\n", i);
4207                                        set_bit(R5_LOCKED, &dev->flags);
4208                                        set_bit(R5_Wantread, &dev->flags);
4209                                        s->locked++;
4210                                        qread++;
4211                                } else
4212                                        set_bit(STRIPE_DELAYED, &sh->state);
4213                        }
4214                }
4215                if (rcw && conf->mddev->queue)
4216                        blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
4217                                          (unsigned long long)sh->sector,
4218                                          rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
4219        }
4220
4221        if (rcw > disks && rmw > disks &&
4222            !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4223                set_bit(STRIPE_DELAYED, &sh->state);
4224
4225        /* now if nothing is locked, and if we have enough data,
4226         * we can start a write request
4227         */
4228        /* since handle_stripe can be called at any time we need to handle the
4229         * case where a compute block operation has been submitted and then a
4230         * subsequent call wants to start a write request.  raid_run_ops only
4231         * handles the case where compute block and reconstruct are requested
4232         * simultaneously.  If this is not the case then new writes need to be
4233         * held off until the compute completes.
4234         */
4235        if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
4236            (s->locked == 0 && (rcw == 0 || rmw == 0) &&
4237             !test_bit(STRIPE_BIT_DELAY, &sh->state)))
4238                schedule_reconstruction(sh, s, rcw == 0, 0);
4239        return 0;
4240}
4241
4242static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
4243                                struct stripe_head_state *s, int disks)
4244{
4245        struct r5dev *dev = NULL;
4246
4247        BUG_ON(sh->batch_head);
4248        set_bit(STRIPE_HANDLE, &sh->state);
4249
4250        switch (sh->check_state) {
4251        case check_state_idle:
4252                /* start a new check operation if there are no failures */
4253                if (s->failed == 0) {
4254                        BUG_ON(s->uptodate != disks);
4255                        sh->check_state = check_state_run;
4256                        set_bit(STRIPE_OP_CHECK, &s->ops_request);
4257                        clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4258                        s->uptodate--;
4259                        break;
4260                }
4261                dev = &sh->dev[s->failed_num[0]];
4262                fallthrough;
4263        case check_state_compute_result:
4264                sh->check_state = check_state_idle;
4265                if (!dev)
4266                        dev = &sh->dev[sh->pd_idx];
4267
4268                /* check that a write has not made the stripe insync */
4269                if (test_bit(STRIPE_INSYNC, &sh->state))
4270                        break;
4271
4272                /* either failed parity check, or recovery is happening */
4273                BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
4274                BUG_ON(s->uptodate != disks);
4275
4276                set_bit(R5_LOCKED, &dev->flags);
4277                s->locked++;
4278                set_bit(R5_Wantwrite, &dev->flags);
4279
4280                clear_bit(STRIPE_DEGRADED, &sh->state);
4281                set_bit(STRIPE_INSYNC, &sh->state);
4282                break;
4283        case check_state_run:
4284                break; /* we will be called again upon completion */
4285        case check_state_check_result:
4286                sh->check_state = check_state_idle;
4287
4288                /* if a failure occurred during the check operation, leave
4289                 * STRIPE_INSYNC not set and let the stripe be handled again
4290                 */
4291                if (s->failed)
4292                        break;
4293
4294                /* handle a successful check operation, if parity is correct
4295                 * we are done.  Otherwise update the mismatch count and repair
4296                 * parity if !MD_RECOVERY_CHECK
4297                 */
4298                if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
4299                        /* parity is correct (on disc,
4300                         * not in buffer any more)
4301                         */
4302                        set_bit(STRIPE_INSYNC, &sh->state);
4303                else {
4304                        atomic64_add(RAID5_STRIPE_SECTORS(conf), &conf->mddev->resync_mismatches);
4305                        if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4306                                /* don't try to repair!! */
4307                                set_bit(STRIPE_INSYNC, &sh->state);
4308                                pr_warn_ratelimited("%s: mismatch sector in range "
4309                                                    "%llu-%llu\n", mdname(conf->mddev),
4310                                                    (unsigned long long) sh->sector,
4311                                                    (unsigned long long) sh->sector +
4312                                                    RAID5_STRIPE_SECTORS(conf));
4313                        } else {
4314                                sh->check_state = check_state_compute_run;
4315                                set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4316                                set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4317                                set_bit(R5_Wantcompute,
4318                                        &sh->dev[sh->pd_idx].flags);
4319                                sh->ops.target = sh->pd_idx;
4320                                sh->ops.target2 = -1;
4321                                s->uptodate++;
4322                        }
4323                }
4324                break;
4325        case check_state_compute_run:
4326                break;
4327        default:
4328                pr_err("%s: unknown check_state: %d sector: %llu\n",
4329                       __func__, sh->check_state,
4330                       (unsigned long long) sh->sector);
4331                BUG();
4332        }
4333}
4334
4335static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
4336                                  struct stripe_head_state *s,
4337                                  int disks)
4338{
4339        int pd_idx = sh->pd_idx;
4340        int qd_idx = sh->qd_idx;
4341        struct r5dev *dev;
4342
4343        BUG_ON(sh->batch_head);
4344        set_bit(STRIPE_HANDLE, &sh->state);
4345
4346        BUG_ON(s->failed > 2);
4347
4348        /* Want to check and possibly repair P and Q.
4349         * However there could be one 'failed' device, in which
4350         * case we can only check one of them, possibly using the
4351         * other to generate missing data
4352         */
4353
4354        switch (sh->check_state) {
4355        case check_state_idle:
4356                /* start a new check operation if there are < 2 failures */
4357                if (s->failed == s->q_failed) {
4358                        /* The only possible failed device holds Q, so it
4359                         * makes sense to check P (If anything else were failed,
4360                         * we would have used P to recreate it).
4361                         */
4362                        sh->check_state = check_state_run;
4363                }
4364                if (!s->q_failed && s->failed < 2) {
4365                        /* Q is not failed, and we didn't use it to generate
4366                         * anything, so it makes sense to check it
4367                         */
4368                        if (sh->check_state == check_state_run)
4369                                sh->check_state = check_state_run_pq;
4370                        else
4371                                sh->check_state = check_state_run_q;
4372                }
4373
4374                /* discard potentially stale zero_sum_result */
4375                sh->ops.zero_sum_result = 0;
4376
4377                if (sh->check_state == check_state_run) {
4378                        /* async_xor_zero_sum destroys the contents of P */
4379                        clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
4380                        s->uptodate--;
4381                }
4382                if (sh->check_state >= check_state_run &&
4383                    sh->check_state <= check_state_run_pq) {
4384                        /* async_syndrome_zero_sum preserves P and Q, so
4385                         * no need to mark them !uptodate here
4386                         */
4387                        set_bit(STRIPE_OP_CHECK, &s->ops_request);
4388                        break;
4389                }
4390
4391                /* we have 2-disk failure */
4392                BUG_ON(s->failed != 2);
4393                fallthrough;
4394        case check_state_compute_result:
4395                sh->check_state = check_state_idle;
4396
4397                /* check that a write has not made the stripe insync */
4398                if (test_bit(STRIPE_INSYNC, &sh->state))
4399                        break;
4400
4401                /* now write out any block on a failed drive,
4402                 * or P or Q if they were recomputed
4403                 */
4404                dev = NULL;
4405                if (s->failed == 2) {
4406                        dev = &sh->dev[s->failed_num[1]];
4407                        s->locked++;
4408                        set_bit(R5_LOCKED, &dev->flags);
4409                        set_bit(R5_Wantwrite, &dev->flags);
4410                }
4411                if (s->failed >= 1) {
4412                        dev = &sh->dev[s->failed_num[0]];
4413                        s->locked++;
4414                        set_bit(R5_LOCKED, &dev->flags);
4415                        set_bit(R5_Wantwrite, &dev->flags);
4416                }
4417                if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4418                        dev = &sh->dev[pd_idx];
4419                        s->locked++;
4420                        set_bit(R5_LOCKED, &dev->flags);
4421                        set_bit(R5_Wantwrite, &dev->flags);
4422                }
4423                if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4424                        dev = &sh->dev[qd_idx];
4425                        s->locked++;
4426                        set_bit(R5_LOCKED, &dev->flags);
4427                        set_bit(R5_Wantwrite, &dev->flags);
4428                }
4429                if (WARN_ONCE(dev && !test_bit(R5_UPTODATE, &dev->flags),
4430                              "%s: disk%td not up to date\n",
4431                              mdname(conf->mddev),
4432                              dev - (struct r5dev *) &sh->dev)) {
4433                        clear_bit(R5_LOCKED, &dev->flags);
4434                        clear_bit(R5_Wantwrite, &dev->flags);
4435                        s->locked--;
4436                }
4437                clear_bit(STRIPE_DEGRADED, &sh->state);
4438
4439                set_bit(STRIPE_INSYNC, &sh->state);
4440                break;
4441        case check_state_run:
4442        case check_state_run_q:
4443        case check_state_run_pq:
4444                break; /* we will be called again upon completion */
4445        case check_state_check_result:
4446                sh->check_state = check_state_idle;
4447
4448                /* handle a successful check operation, if parity is correct
4449                 * we are done.  Otherwise update the mismatch count and repair
4450                 * parity if !MD_RECOVERY_CHECK
4451                 */
4452                if (sh->ops.zero_sum_result == 0) {
4453                        /* both parities are correct */
4454                        if (!s->failed)
4455                                set_bit(STRIPE_INSYNC, &sh->state);
4456                        else {
4457                                /* in contrast to the raid5 case we can validate
4458                                 * parity, but still have a failure to write
4459                                 * back
4460                                 */
4461                                sh->check_state = check_state_compute_result;
4462                                /* Returning at this point means that we may go
4463                                 * off and bring p and/or q uptodate again so
4464                                 * we make sure to check zero_sum_result again
4465                                 * to verify if p or q need writeback
4466                                 */
4467                        }
4468                } else {
4469                        atomic64_add(RAID5_STRIPE_SECTORS(conf), &conf->mddev->resync_mismatches);
4470                        if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4471                                /* don't try to repair!! */
4472                                set_bit(STRIPE_INSYNC, &sh->state);
4473                                pr_warn_ratelimited("%s: mismatch sector in range "
4474                                                    "%llu-%llu\n", mdname(conf->mddev),
4475                                                    (unsigned long long) sh->sector,
4476                                                    (unsigned long long) sh->sector +
4477                                                    RAID5_STRIPE_SECTORS(conf));
4478                        } else {
4479                                int *target = &sh->ops.target;
4480
4481                                sh->ops.target = -1;
4482                                sh->ops.target2 = -1;
4483                                sh->check_state = check_state_compute_run;
4484                                set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4485                                set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4486                                if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4487                                        set_bit(R5_Wantcompute,
4488                                                &sh->dev[pd_idx].flags);
4489                                        *target = pd_idx;
4490                                        target = &sh->ops.target2;
4491                                        s->uptodate++;
4492                                }
4493                                if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4494                                        set_bit(R5_Wantcompute,
4495                                                &sh->dev[qd_idx].flags);
4496                                        *target = qd_idx;
4497                                        s->uptodate++;
4498                                }
4499                        }
4500                }
4501                break;
4502        case check_state_compute_run:
4503                break;
4504        default:
4505                pr_warn("%s: unknown check_state: %d sector: %llu\n",
4506                        __func__, sh->check_state,
4507                        (unsigned long long) sh->sector);
4508                BUG();
4509        }
4510}
4511
4512static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
4513{
4514        int i;
4515
4516        /* We have read all the blocks in this stripe and now we need to
4517         * copy some of them into a target stripe for expand.
4518         */
4519        struct dma_async_tx_descriptor *tx = NULL;
4520        BUG_ON(sh->batch_head);
4521        clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4522        for (i = 0; i < sh->disks; i++)
4523                if (i != sh->pd_idx && i != sh->qd_idx) {
4524                        int dd_idx, j;
4525                        struct stripe_head *sh2;
4526                        struct async_submit_ctl submit;
4527
4528                        sector_t bn = raid5_compute_blocknr(sh, i, 1);
4529                        sector_t s = raid5_compute_sector(conf, bn, 0,
4530                                                          &dd_idx, NULL);
4531                        sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
4532                        if (sh2 == NULL)
4533                                /* so far only the early blocks of this stripe
4534                                 * have been requested.  When later blocks
4535                                 * get requested, we will try again
4536                                 */
4537                                continue;
4538                        if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
4539                           test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
4540                                /* must have already done this block */
4541                                raid5_release_stripe(sh2);
4542                                continue;
4543                        }
4544
4545                        /* place all the copies on one channel */
4546                        init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
4547                        tx = async_memcpy(sh2->dev[dd_idx].page,
4548                                          sh->dev[i].page, sh2->dev[dd_idx].offset,
4549                                          sh->dev[i].offset, RAID5_STRIPE_SIZE(conf),
4550                                          &submit);
4551
4552                        set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
4553                        set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
4554                        for (j = 0; j < conf->raid_disks; j++)
4555                                if (j != sh2->pd_idx &&
4556                                    j != sh2->qd_idx &&
4557                                    !test_bit(R5_Expanded, &sh2->dev[j].flags))
4558                                        break;
4559                        if (j == conf->raid_disks) {
4560                                set_bit(STRIPE_EXPAND_READY, &sh2->state);
4561                                set_bit(STRIPE_HANDLE, &sh2->state);
4562                        }
4563                        raid5_release_stripe(sh2);
4564
4565                }
4566        /* done submitting copies, wait for them to complete */
4567        async_tx_quiesce(&tx);
4568}
4569
4570/*
4571 * handle_stripe - do things to a stripe.
4572 *
4573 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4574 * state of various bits to see what needs to be done.
4575 * Possible results:
4576 *    return some read requests which now have data
4577 *    return some write requests which are safely on storage
4578 *    schedule a read on some buffers
4579 *    schedule a write of some buffers
4580 *    return confirmation of parity correctness
4581 *
4582 */
4583
4584static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4585{
4586        struct r5conf *conf = sh->raid_conf;
4587        int disks = sh->disks;
4588        struct r5dev *dev;
4589        int i;
4590        int do_recovery = 0;
4591
4592        memset(s, 0, sizeof(*s));
4593
4594        s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4595        s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4596        s->failed_num[0] = -1;
4597        s->failed_num[1] = -1;
4598        s->log_failed = r5l_log_disk_error(conf);
4599
4600        /* Now to look around and see what can be done */
4601        rcu_read_lock();
4602        for (i=disks; i--; ) {
4603                struct md_rdev *rdev;
4604                sector_t first_bad;
4605                int bad_sectors;
4606                int is_bad = 0;
4607
4608                dev = &sh->dev[i];
4609
4610                pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4611                         i, dev->flags,
4612                         dev->toread, dev->towrite, dev->written);
4613                /* maybe we can reply to a read
4614                 *
4615                 * new wantfill requests are only permitted while
4616                 * ops_complete_biofill is guaranteed to be inactive
4617                 */
4618                if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4619                    !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4620                        set_bit(R5_Wantfill, &dev->flags);
4621
4622                /* now count some things */
4623                if (test_bit(R5_LOCKED, &dev->flags))
4624                        s->locked++;
4625                if (test_bit(R5_UPTODATE, &dev->flags))
4626                        s->uptodate++;
4627                if (test_bit(R5_Wantcompute, &dev->flags)) {
4628                        s->compute++;
4629                        BUG_ON(s->compute > 2);
4630                }
4631
4632                if (test_bit(R5_Wantfill, &dev->flags))
4633                        s->to_fill++;
4634                else if (dev->toread)
4635                        s->to_read++;
4636                if (dev->towrite) {
4637                        s->to_write++;
4638                        if (!test_bit(R5_OVERWRITE, &dev->flags))
4639                                s->non_overwrite++;
4640                }
4641                if (dev->written)
4642                        s->written++;
4643                /* Prefer to use the replacement for reads, but only
4644                 * if it is recovered enough and has no bad blocks.
4645                 */
4646                rdev = rcu_dereference(conf->disks[i].replacement);
4647                if (rdev && !test_bit(Faulty, &rdev->flags) &&
4648                    rdev->recovery_offset >= sh->sector + RAID5_STRIPE_SECTORS(conf) &&
4649                    !is_badblock(rdev, sh->sector, RAID5_STRIPE_SECTORS(conf),
4650                                 &first_bad, &bad_sectors))
4651                        set_bit(R5_ReadRepl, &dev->flags);
4652                else {
4653                        if (rdev && !test_bit(Faulty, &rdev->flags))
4654                                set_bit(R5_NeedReplace, &dev->flags);
4655                        else
4656                                clear_bit(R5_NeedReplace, &dev->flags);
4657                        rdev = rcu_dereference(conf->disks[i].rdev);
4658                        clear_bit(R5_ReadRepl, &dev->flags);
4659                }
4660                if (rdev && test_bit(Faulty, &rdev->flags))
4661                        rdev = NULL;
4662                if (rdev) {
4663                        is_bad = is_badblock(rdev, sh->sector, RAID5_STRIPE_SECTORS(conf),
4664                                             &first_bad, &bad_sectors);
4665                        if (s->blocked_rdev == NULL
4666                            && (test_bit(Blocked, &rdev->flags)
4667                                || is_bad < 0)) {
4668                                if (is_bad < 0)
4669                                        set_bit(BlockedBadBlocks,
4670                                                &rdev->flags);
4671                                s->blocked_rdev = rdev;
4672                                atomic_inc(&rdev->nr_pending);
4673                        }
4674                }
4675                clear_bit(R5_Insync, &dev->flags);
4676                if (!rdev)
4677                        /* Not in-sync */;
4678                else if (is_bad) {
4679                        /* also not in-sync */
4680                        if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4681                            test_bit(R5_UPTODATE, &dev->flags)) {
4682                                /* treat as in-sync, but with a read error
4683                                 * which we can now try to correct
4684                                 */
4685                                set_bit(R5_Insync, &dev->flags);
4686                                set_bit(R5_ReadError, &dev->flags);
4687                        }
4688                } else if (test_bit(In_sync, &rdev->flags))
4689                        set_bit(R5_Insync, &dev->flags);
4690                else if (sh->sector + RAID5_STRIPE_SECTORS(conf) <= rdev->recovery_offset)
4691                        /* in sync if before recovery_offset */
4692                        set_bit(R5_Insync, &dev->flags);
4693                else if (test_bit(R5_UPTODATE, &dev->flags) &&
4694                         test_bit(R5_Expanded, &dev->flags))
4695                        /* If we've reshaped into here, we assume it is Insync.
4696                         * We will shortly update recovery_offset to make
4697                         * it official.
4698                         */
4699                        set_bit(R5_Insync, &dev->flags);
4700
4701                if (test_bit(R5_WriteError, &dev->flags)) {
4702                        /* This flag does not apply to '.replacement'
4703                         * only to .rdev, so make sure to check that*/
4704                        struct md_rdev *rdev2 = rcu_dereference(
4705                                conf->disks[i].rdev);
4706                        if (rdev2 == rdev)
4707                                clear_bit(R5_Insync, &dev->flags);
4708                        if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4709                                s->handle_bad_blocks = 1;
4710                                atomic_inc(&rdev2->nr_pending);
4711                        } else
4712                                clear_bit(R5_WriteError, &dev->flags);
4713                }
4714                if (test_bit(R5_MadeGood, &dev->flags)) {
4715                        /* This flag does not apply to '.replacement'
4716                         * only to .rdev, so make sure to check that*/
4717                        struct md_rdev *rdev2 = rcu_dereference(
4718                                conf->disks[i].rdev);
4719                        if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4720                                s->handle_bad_blocks = 1;
4721                                atomic_inc(&rdev2->nr_pending);
4722                        } else
4723                                clear_bit(R5_MadeGood, &dev->flags);
4724                }
4725                if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4726                        struct md_rdev *rdev2 = rcu_dereference(
4727                                conf->disks[i].replacement);
4728                        if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4729                                s->handle_bad_blocks = 1;
4730                                atomic_inc(&rdev2->nr_pending);
4731                        } else
4732                                clear_bit(R5_MadeGoodRepl, &dev->flags);
4733                }
4734                if (!test_bit(R5_Insync, &dev->flags)) {
4735                        /* The ReadError flag will just be confusing now */
4736                        clear_bit(R5_ReadError, &dev->flags);
4737                        clear_bit(R5_ReWrite, &dev->flags);
4738                }
4739                if (test_bit(R5_ReadError, &dev->flags))
4740                        clear_bit(R5_Insync, &dev->flags);
4741                if (!test_bit(R5_Insync, &dev->flags)) {
4742                        if (s->failed < 2)
4743                                s->failed_num[s->failed] = i;
4744                        s->failed++;
4745                        if (rdev && !test_bit(Faulty, &rdev->flags))
4746                                do_recovery = 1;
4747                        else if (!rdev) {
4748                                rdev = rcu_dereference(
4749                                    conf->disks[i].replacement);
4750                                if (rdev && !test_bit(Faulty, &rdev->flags))
4751                                        do_recovery = 1;
4752                        }
4753                }
4754
4755                if (test_bit(R5_InJournal, &dev->flags))
4756                        s->injournal++;
4757                if (test_bit(R5_InJournal, &dev->flags) && dev->written)
4758                        s->just_cached++;
4759        }
4760        if (test_bit(STRIPE_SYNCING, &sh->state)) {
4761                /* If there is a failed device being replaced,
4762                 *     we must be recovering.
4763                 * else if we are after recovery_cp, we must be syncing
4764                 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4765                 * else we can only be replacing
4766                 * sync and recovery both need to read all devices, and so
4767                 * use the same flag.
4768                 */
4769                if (do_recovery ||
4770                    sh->sector >= conf->mddev->recovery_cp ||
4771                    test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4772                        s->syncing = 1;
4773                else
4774                        s->replacing = 1;
4775        }
4776        rcu_read_unlock();
4777}
4778
4779/*
4780 * Return '1' if this is a member of batch, or '0' if it is a lone stripe or
4781 * a head which can now be handled.
4782 */
4783static int clear_batch_ready(struct stripe_head *sh)
4784{
4785        struct stripe_head *tmp;
4786        if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4787                return (sh->batch_head && sh->batch_head != sh);
4788        spin_lock(&sh->stripe_lock);
4789        if (!sh->batch_head) {
4790                spin_unlock(&sh->stripe_lock);
4791                return 0;
4792        }
4793
4794        /*
4795         * this stripe could be added to a batch list before we check
4796         * BATCH_READY, skips it
4797         */
4798        if (sh->batch_head != sh) {
4799                spin_unlock(&sh->stripe_lock);
4800                return 1;
4801        }
4802        spin_lock(&sh->batch_lock);
4803        list_for_each_entry(tmp, &sh->batch_list, batch_list)
4804                clear_bit(STRIPE_BATCH_READY, &tmp->state);
4805        spin_unlock(&sh->batch_lock);
4806        spin_unlock(&sh->stripe_lock);
4807
4808        /*
4809         * BATCH_READY is cleared, no new stripes can be added.
4810         * batch_list can be accessed without lock
4811         */
4812        return 0;
4813}
4814
4815static void break_stripe_batch_list(struct stripe_head *head_sh,
4816                                    unsigned long handle_flags)
4817{
4818        struct stripe_head *sh, *next;
4819        int i;
4820        int do_wakeup = 0;
4821
4822        list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4823
4824                list_del_init(&sh->batch_list);
4825
4826                WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4827                                          (1 << STRIPE_SYNCING) |
4828                                          (1 << STRIPE_REPLACED) |
4829                                          (1 << STRIPE_DELAYED) |
4830                                          (1 << STRIPE_BIT_DELAY) |
4831                                          (1 << STRIPE_FULL_WRITE) |
4832                                          (1 << STRIPE_BIOFILL_RUN) |
4833                                          (1 << STRIPE_COMPUTE_RUN)  |
4834                                          (1 << STRIPE_DISCARD) |
4835                                          (1 << STRIPE_BATCH_READY) |
4836                                          (1 << STRIPE_BATCH_ERR) |
4837                                          (1 << STRIPE_BITMAP_PENDING)),
4838                        "stripe state: %lx\n", sh->state);
4839                WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4840                                              (1 << STRIPE_REPLACED)),
4841                        "head stripe state: %lx\n", head_sh->state);
4842
4843                set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4844                                            (1 << STRIPE_PREREAD_ACTIVE) |
4845                                            (1 << STRIPE_DEGRADED) |
4846                                            (1 << STRIPE_ON_UNPLUG_LIST)),
4847                              head_sh->state & (1 << STRIPE_INSYNC));
4848
4849                sh->check_state = head_sh->check_state;
4850                sh->reconstruct_state = head_sh->reconstruct_state;
4851                spin_lock_irq(&sh->stripe_lock);
4852                sh->batch_head = NULL;
4853                spin_unlock_irq(&sh->stripe_lock);
4854                for (i = 0; i < sh->disks; i++) {
4855                        if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4856                                do_wakeup = 1;
4857                        sh->dev[i].flags = head_sh->dev[i].flags &
4858                                (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4859                }
4860                if (handle_flags == 0 ||
4861                    sh->state & handle_flags)
4862                        set_bit(STRIPE_HANDLE, &sh->state);
4863                raid5_release_stripe(sh);
4864        }
4865        spin_lock_irq(&head_sh->stripe_lock);
4866        head_sh->batch_head = NULL;
4867        spin_unlock_irq(&head_sh->stripe_lock);
4868        for (i = 0; i < head_sh->disks; i++)
4869                if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4870                        do_wakeup = 1;
4871        if (head_sh->state & handle_flags)
4872                set_bit(STRIPE_HANDLE, &head_sh->state);
4873
4874        if (do_wakeup)
4875                wake_up(&head_sh->raid_conf->wait_for_overlap);
4876}
4877
4878static void handle_stripe(struct stripe_head *sh)
4879{
4880        struct stripe_head_state s;
4881        struct r5conf *conf = sh->raid_conf;
4882        int i;
4883        int prexor;
4884        int disks = sh->disks;
4885        struct r5dev *pdev, *qdev;
4886
4887        clear_bit(STRIPE_HANDLE, &sh->state);
4888
4889        /*
4890         * handle_stripe should not continue handle the batched stripe, only
4891         * the head of batch list or lone stripe can continue. Otherwise we
4892         * could see break_stripe_batch_list warns about the STRIPE_ACTIVE
4893         * is set for the batched stripe.
4894         */
4895        if (clear_batch_ready(sh))
4896                return;
4897
4898        if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4899                /* already being handled, ensure it gets handled
4900                 * again when current action finishes */
4901                set_bit(STRIPE_HANDLE, &sh->state);
4902                return;
4903        }
4904
4905        if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4906                break_stripe_batch_list(sh, 0);
4907
4908        if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4909                spin_lock(&sh->stripe_lock);
4910                /*
4911                 * Cannot process 'sync' concurrently with 'discard'.
4912                 * Flush data in r5cache before 'sync'.
4913                 */
4914                if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
4915                    !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) &&
4916                    !test_bit(STRIPE_DISCARD, &sh->state) &&
4917                    test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4918                        set_bit(STRIPE_SYNCING, &sh->state);
4919                        clear_bit(STRIPE_INSYNC, &sh->state);
4920                        clear_bit(STRIPE_REPLACED, &sh->state);
4921                }
4922                spin_unlock(&sh->stripe_lock);
4923        }
4924        clear_bit(STRIPE_DELAYED, &sh->state);
4925
4926        pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4927                "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4928               (unsigned long long)sh->sector, sh->state,
4929               atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4930               sh->check_state, sh->reconstruct_state);
4931
4932        analyse_stripe(sh, &s);
4933
4934        if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4935                goto finish;
4936
4937        if (s.handle_bad_blocks ||
4938            test_bit(MD_SB_CHANGE_PENDING, &conf->mddev->sb_flags)) {
4939                set_bit(STRIPE_HANDLE, &sh->state);
4940                goto finish;
4941        }
4942
4943        if (unlikely(s.blocked_rdev)) {
4944                if (s.syncing || s.expanding || s.expanded ||
4945                    s.replacing || s.to_write || s.written) {
4946                        set_bit(STRIPE_HANDLE, &sh->state);
4947                        goto finish;
4948                }
4949                /* There is nothing for the blocked_rdev to block */
4950                rdev_dec_pending(s.blocked_rdev, conf->mddev);
4951                s.blocked_rdev = NULL;
4952        }
4953
4954        if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4955                set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4956                set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4957        }
4958
4959        pr_debug("locked=%d uptodate=%d to_read=%d"
4960               " to_write=%d failed=%d failed_num=%d,%d\n",
4961               s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4962               s.failed_num[0], s.failed_num[1]);
4963        /*
4964         * check if the array has lost more than max_degraded devices and,
4965         * if so, some requests might need to be failed.
4966         *
4967         * When journal device failed (log_failed), we will only process
4968         * the stripe if there is data need write to raid disks
4969         */
4970        if (s.failed > conf->max_degraded ||
4971            (s.log_failed && s.injournal == 0)) {
4972                sh->check_state = 0;
4973                sh->reconstruct_state = 0;
4974                break_stripe_batch_list(sh, 0);
4975                if (s.to_read+s.to_write+s.written)
4976                        handle_failed_stripe(conf, sh, &s,