linux/fs/nfs/write.c
<<
>>
Prefs
   1/*
   2 * linux/fs/nfs/write.c
   3 *
   4 * Write file data over NFS.
   5 *
   6 * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
   7 */
   8
   9#include <linux/types.h>
  10#include <linux/slab.h>
  11#include <linux/mm.h>
  12#include <linux/pagemap.h>
  13#include <linux/file.h>
  14#include <linux/writeback.h>
  15#include <linux/swap.h>
  16#include <linux/migrate.h>
  17
  18#include <linux/sunrpc/clnt.h>
  19#include <linux/nfs_fs.h>
  20#include <linux/nfs_mount.h>
  21#include <linux/nfs_page.h>
  22#include <linux/backing-dev.h>
  23#include <linux/export.h>
  24#include <linux/freezer.h>
  25#include <linux/wait.h>
  26#include <linux/iversion.h>
  27
  28#include <linux/uaccess.h>
  29#include <linux/sched/mm.h>
  30
  31#include "delegation.h"
  32#include "internal.h"
  33#include "iostat.h"
  34#include "nfs4_fs.h"
  35#include "fscache.h"
  36#include "pnfs.h"
  37
  38#include "nfstrace.h"
  39
  40#define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  41
  42#define MIN_POOL_WRITE          (32)
  43#define MIN_POOL_COMMIT         (4)
  44
  45struct nfs_io_completion {
  46        void (*complete)(void *data);
  47        void *data;
  48        struct kref refcount;
  49};
  50
  51/*
  52 * Local function declarations
  53 */
  54static void nfs_redirty_request(struct nfs_page *req);
  55static const struct rpc_call_ops nfs_commit_ops;
  56static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
  57static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
  58static const struct nfs_rw_ops nfs_rw_write_ops;
  59static void nfs_clear_request_commit(struct nfs_page *req);
  60static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
  61                                      struct inode *inode);
  62static struct nfs_page *
  63nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
  64                                                struct page *page);
  65
  66static struct kmem_cache *nfs_wdata_cachep;
  67static mempool_t *nfs_wdata_mempool;
  68static struct kmem_cache *nfs_cdata_cachep;
  69static mempool_t *nfs_commit_mempool;
  70
  71struct nfs_commit_data *nfs_commitdata_alloc(bool never_fail)
  72{
  73        struct nfs_commit_data *p;
  74
  75        if (never_fail)
  76                p = mempool_alloc(nfs_commit_mempool, GFP_NOIO);
  77        else {
  78                /* It is OK to do some reclaim, not no safe to wait
  79                 * for anything to be returned to the pool.
  80                 * mempool_alloc() cannot handle that particular combination,
  81                 * so we need two separate attempts.
  82                 */
  83                p = mempool_alloc(nfs_commit_mempool, GFP_NOWAIT);
  84                if (!p)
  85                        p = kmem_cache_alloc(nfs_cdata_cachep, GFP_NOIO |
  86                                             __GFP_NOWARN | __GFP_NORETRY);
  87                if (!p)
  88                        return NULL;
  89        }
  90
  91        memset(p, 0, sizeof(*p));
  92        INIT_LIST_HEAD(&p->pages);
  93        return p;
  94}
  95EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
  96
  97void nfs_commit_free(struct nfs_commit_data *p)
  98{
  99        mempool_free(p, nfs_commit_mempool);
 100}
 101EXPORT_SYMBOL_GPL(nfs_commit_free);
 102
 103static struct nfs_pgio_header *nfs_writehdr_alloc(void)
 104{
 105        struct nfs_pgio_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOIO);
 106
 107        memset(p, 0, sizeof(*p));
 108        p->rw_mode = FMODE_WRITE;
 109        return p;
 110}
 111
 112static void nfs_writehdr_free(struct nfs_pgio_header *hdr)
 113{
 114        mempool_free(hdr, nfs_wdata_mempool);
 115}
 116
 117static struct nfs_io_completion *nfs_io_completion_alloc(gfp_t gfp_flags)
 118{
 119        return kmalloc(sizeof(struct nfs_io_completion), gfp_flags);
 120}
 121
 122static void nfs_io_completion_init(struct nfs_io_completion *ioc,
 123                void (*complete)(void *), void *data)
 124{
 125        ioc->complete = complete;
 126        ioc->data = data;
 127        kref_init(&ioc->refcount);
 128}
 129
 130static void nfs_io_completion_release(struct kref *kref)
 131{
 132        struct nfs_io_completion *ioc = container_of(kref,
 133                        struct nfs_io_completion, refcount);
 134        ioc->complete(ioc->data);
 135        kfree(ioc);
 136}
 137
 138static void nfs_io_completion_get(struct nfs_io_completion *ioc)
 139{
 140        if (ioc != NULL)
 141                kref_get(&ioc->refcount);
 142}
 143
 144static void nfs_io_completion_put(struct nfs_io_completion *ioc)
 145{
 146        if (ioc != NULL)
 147                kref_put(&ioc->refcount, nfs_io_completion_release);
 148}
 149
 150static struct nfs_page *
 151nfs_page_private_request(struct page *page)
 152{
 153        if (!PagePrivate(page))
 154                return NULL;
 155        return (struct nfs_page *)page_private(page);
 156}
 157
 158/*
 159 * nfs_page_find_head_request_locked - find head request associated with @page
 160 *
 161 * must be called while holding the inode lock.
 162 *
 163 * returns matching head request with reference held, or NULL if not found.
 164 */
 165static struct nfs_page *
 166nfs_page_find_private_request(struct page *page)
 167{
 168        struct address_space *mapping = page_file_mapping(page);
 169        struct nfs_page *req;
 170
 171        if (!PagePrivate(page))
 172                return NULL;
 173        spin_lock(&mapping->private_lock);
 174        req = nfs_page_private_request(page);
 175        if (req) {
 176                WARN_ON_ONCE(req->wb_head != req);
 177                kref_get(&req->wb_kref);
 178        }
 179        spin_unlock(&mapping->private_lock);
 180        return req;
 181}
 182
 183static struct nfs_page *
 184nfs_page_find_swap_request(struct page *page)
 185{
 186        struct inode *inode = page_file_mapping(page)->host;
 187        struct nfs_inode *nfsi = NFS_I(inode);
 188        struct nfs_page *req = NULL;
 189        if (!PageSwapCache(page))
 190                return NULL;
 191        mutex_lock(&nfsi->commit_mutex);
 192        if (PageSwapCache(page)) {
 193                req = nfs_page_search_commits_for_head_request_locked(nfsi,
 194                        page);
 195                if (req) {
 196                        WARN_ON_ONCE(req->wb_head != req);
 197                        kref_get(&req->wb_kref);
 198                }
 199        }
 200        mutex_unlock(&nfsi->commit_mutex);
 201        return req;
 202}
 203
 204/*
 205 * nfs_page_find_head_request - find head request associated with @page
 206 *
 207 * returns matching head request with reference held, or NULL if not found.
 208 */
 209static struct nfs_page *nfs_page_find_head_request(struct page *page)
 210{
 211        struct nfs_page *req;
 212
 213        req = nfs_page_find_private_request(page);
 214        if (!req)
 215                req = nfs_page_find_swap_request(page);
 216        return req;
 217}
 218
 219/* Adjust the file length if we're writing beyond the end */
 220static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 221{
 222        struct inode *inode = page_file_mapping(page)->host;
 223        loff_t end, i_size;
 224        pgoff_t end_index;
 225
 226        spin_lock(&inode->i_lock);
 227        i_size = i_size_read(inode);
 228        end_index = (i_size - 1) >> PAGE_SHIFT;
 229        if (i_size > 0 && page_index(page) < end_index)
 230                goto out;
 231        end = page_file_offset(page) + ((loff_t)offset+count);
 232        if (i_size >= end)
 233                goto out;
 234        i_size_write(inode, end);
 235        NFS_I(inode)->cache_validity &= ~NFS_INO_INVALID_SIZE;
 236        nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 237out:
 238        spin_unlock(&inode->i_lock);
 239}
 240
 241/* A writeback failed: mark the page as bad, and invalidate the page cache */
 242static void nfs_set_pageerror(struct address_space *mapping)
 243{
 244        nfs_zap_mapping(mapping->host, mapping);
 245}
 246
 247/*
 248 * nfs_page_group_search_locked
 249 * @head - head request of page group
 250 * @page_offset - offset into page
 251 *
 252 * Search page group with head @head to find a request that contains the
 253 * page offset @page_offset.
 254 *
 255 * Returns a pointer to the first matching nfs request, or NULL if no
 256 * match is found.
 257 *
 258 * Must be called with the page group lock held
 259 */
 260static struct nfs_page *
 261nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
 262{
 263        struct nfs_page *req;
 264
 265        req = head;
 266        do {
 267                if (page_offset >= req->wb_pgbase &&
 268                    page_offset < (req->wb_pgbase + req->wb_bytes))
 269                        return req;
 270
 271                req = req->wb_this_page;
 272        } while (req != head);
 273
 274        return NULL;
 275}
 276
 277/*
 278 * nfs_page_group_covers_page
 279 * @head - head request of page group
 280 *
 281 * Return true if the page group with head @head covers the whole page,
 282 * returns false otherwise
 283 */
 284static bool nfs_page_group_covers_page(struct nfs_page *req)
 285{
 286        struct nfs_page *tmp;
 287        unsigned int pos = 0;
 288        unsigned int len = nfs_page_length(req->wb_page);
 289
 290        nfs_page_group_lock(req);
 291
 292        for (;;) {
 293                tmp = nfs_page_group_search_locked(req->wb_head, pos);
 294                if (!tmp)
 295                        break;
 296                pos = tmp->wb_pgbase + tmp->wb_bytes;
 297        }
 298
 299        nfs_page_group_unlock(req);
 300        return pos >= len;
 301}
 302
 303/* We can set the PG_uptodate flag if we see that a write request
 304 * covers the full page.
 305 */
 306static void nfs_mark_uptodate(struct nfs_page *req)
 307{
 308        if (PageUptodate(req->wb_page))
 309                return;
 310        if (!nfs_page_group_covers_page(req))
 311                return;
 312        SetPageUptodate(req->wb_page);
 313}
 314
 315static int wb_priority(struct writeback_control *wbc)
 316{
 317        int ret = 0;
 318
 319        if (wbc->sync_mode == WB_SYNC_ALL)
 320                ret = FLUSH_COND_STABLE;
 321        return ret;
 322}
 323
 324/*
 325 * NFS congestion control
 326 */
 327
 328int nfs_congestion_kb;
 329
 330#define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 331#define NFS_CONGESTION_OFF_THRESH       \
 332        (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 333
 334static void nfs_set_page_writeback(struct page *page)
 335{
 336        struct inode *inode = page_file_mapping(page)->host;
 337        struct nfs_server *nfss = NFS_SERVER(inode);
 338        int ret = test_set_page_writeback(page);
 339
 340        WARN_ON_ONCE(ret != 0);
 341
 342        if (atomic_long_inc_return(&nfss->writeback) >
 343                        NFS_CONGESTION_ON_THRESH)
 344                set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 345}
 346
 347static void nfs_end_page_writeback(struct nfs_page *req)
 348{
 349        struct inode *inode = page_file_mapping(req->wb_page)->host;
 350        struct nfs_server *nfss = NFS_SERVER(inode);
 351        bool is_done;
 352
 353        is_done = nfs_page_group_sync_on_bit(req, PG_WB_END);
 354        nfs_unlock_request(req);
 355        if (!is_done)
 356                return;
 357
 358        end_page_writeback(req->wb_page);
 359        if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
 360                clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 361}
 362
 363/*
 364 * nfs_unroll_locks_and_wait -  unlock all newly locked reqs and wait on @req
 365 *
 366 * this is a helper function for nfs_lock_and_join_requests
 367 *
 368 * @inode - inode associated with request page group, must be holding inode lock
 369 * @head  - head request of page group, must be holding head lock
 370 * @req   - request that couldn't lock and needs to wait on the req bit lock
 371 *
 372 * NOTE: this must be called holding page_group bit lock
 373 *       which will be released before returning.
 374 *
 375 * returns 0 on success, < 0 on error.
 376 */
 377static void
 378nfs_unroll_locks(struct inode *inode, struct nfs_page *head,
 379                          struct nfs_page *req)
 380{
 381        struct nfs_page *tmp;
 382
 383        /* relinquish all the locks successfully grabbed this run */
 384        for (tmp = head->wb_this_page ; tmp != req; tmp = tmp->wb_this_page) {
 385                if (!kref_read(&tmp->wb_kref))
 386                        continue;
 387                nfs_unlock_and_release_request(tmp);
 388        }
 389}
 390
 391/*
 392 * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
 393 *
 394 * @destroy_list - request list (using wb_this_page) terminated by @old_head
 395 * @old_head - the old head of the list
 396 *
 397 * All subrequests must be locked and removed from all lists, so at this point
 398 * they are only "active" in this function, and possibly in nfs_wait_on_request
 399 * with a reference held by some other context.
 400 */
 401static void
 402nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
 403                                 struct nfs_page *old_head,
 404                                 struct inode *inode)
 405{
 406        while (destroy_list) {
 407                struct nfs_page *subreq = destroy_list;
 408
 409                destroy_list = (subreq->wb_this_page == old_head) ?
 410                                   NULL : subreq->wb_this_page;
 411
 412                WARN_ON_ONCE(old_head != subreq->wb_head);
 413
 414                /* make sure old group is not used */
 415                subreq->wb_this_page = subreq;
 416
 417                clear_bit(PG_REMOVE, &subreq->wb_flags);
 418
 419                /* Note: races with nfs_page_group_destroy() */
 420                if (!kref_read(&subreq->wb_kref)) {
 421                        /* Check if we raced with nfs_page_group_destroy() */
 422                        if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags))
 423                                nfs_free_request(subreq);
 424                        continue;
 425                }
 426
 427                subreq->wb_head = subreq;
 428
 429                if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
 430                        nfs_release_request(subreq);
 431                        atomic_long_dec(&NFS_I(inode)->nrequests);
 432                }
 433
 434                /* subreq is now totally disconnected from page group or any
 435                 * write / commit lists. last chance to wake any waiters */
 436                nfs_unlock_and_release_request(subreq);
 437        }
 438}
 439
 440/*
 441 * nfs_lock_and_join_requests - join all subreqs to the head req and return
 442 *                              a locked reference, cancelling any pending
 443 *                              operations for this page.
 444 *
 445 * @page - the page used to lookup the "page group" of nfs_page structures
 446 *
 447 * This function joins all sub requests to the head request by first
 448 * locking all requests in the group, cancelling any pending operations
 449 * and finally updating the head request to cover the whole range covered by
 450 * the (former) group.  All subrequests are removed from any write or commit
 451 * lists, unlinked from the group and destroyed.
 452 *
 453 * Returns a locked, referenced pointer to the head request - which after
 454 * this call is guaranteed to be the only request associated with the page.
 455 * Returns NULL if no requests are found for @page, or a ERR_PTR if an
 456 * error was encountered.
 457 */
 458static struct nfs_page *
 459nfs_lock_and_join_requests(struct page *page)
 460{
 461        struct inode *inode = page_file_mapping(page)->host;
 462        struct nfs_page *head, *subreq;
 463        struct nfs_page *destroy_list = NULL;
 464        unsigned int total_bytes;
 465        int ret;
 466
 467try_again:
 468        /*
 469         * A reference is taken only on the head request which acts as a
 470         * reference to the whole page group - the group will not be destroyed
 471         * until the head reference is released.
 472         */
 473        head = nfs_page_find_head_request(page);
 474        if (!head)
 475                return NULL;
 476
 477        /* lock the page head first in order to avoid an ABBA inefficiency */
 478        if (!nfs_lock_request(head)) {
 479                ret = nfs_wait_on_request(head);
 480                nfs_release_request(head);
 481                if (ret < 0)
 482                        return ERR_PTR(ret);
 483                goto try_again;
 484        }
 485
 486        /* Ensure that nobody removed the request before we locked it */
 487        if (head != nfs_page_private_request(page) && !PageSwapCache(page)) {
 488                nfs_unlock_and_release_request(head);
 489                goto try_again;
 490        }
 491
 492        ret = nfs_page_group_lock(head);
 493        if (ret < 0)
 494                goto release_request;
 495
 496        /* lock each request in the page group */
 497        total_bytes = head->wb_bytes;
 498        for (subreq = head->wb_this_page; subreq != head;
 499                        subreq = subreq->wb_this_page) {
 500
 501                if (!kref_get_unless_zero(&subreq->wb_kref)) {
 502                        if (subreq->wb_offset == head->wb_offset + total_bytes)
 503                                total_bytes += subreq->wb_bytes;
 504                        continue;
 505                }
 506
 507                while (!nfs_lock_request(subreq)) {
 508                        /*
 509                         * Unlock page to allow nfs_page_group_sync_on_bit()
 510                         * to succeed
 511                         */
 512                        nfs_page_group_unlock(head);
 513                        ret = nfs_wait_on_request(subreq);
 514                        if (!ret)
 515                                ret = nfs_page_group_lock(head);
 516                        if (ret < 0) {
 517                                nfs_unroll_locks(inode, head, subreq);
 518                                nfs_release_request(subreq);
 519                                goto release_request;
 520                        }
 521                }
 522                /*
 523                 * Subrequests are always contiguous, non overlapping
 524                 * and in order - but may be repeated (mirrored writes).
 525                 */
 526                if (subreq->wb_offset == (head->wb_offset + total_bytes)) {
 527                        /* keep track of how many bytes this group covers */
 528                        total_bytes += subreq->wb_bytes;
 529                } else if (WARN_ON_ONCE(subreq->wb_offset < head->wb_offset ||
 530                            ((subreq->wb_offset + subreq->wb_bytes) >
 531                             (head->wb_offset + total_bytes)))) {
 532                        nfs_page_group_unlock(head);
 533                        nfs_unroll_locks(inode, head, subreq);
 534                        nfs_unlock_and_release_request(subreq);
 535                        ret = -EIO;
 536                        goto release_request;
 537                }
 538        }
 539
 540        /* Now that all requests are locked, make sure they aren't on any list.
 541         * Commit list removal accounting is done after locks are dropped */
 542        subreq = head;
 543        do {
 544                nfs_clear_request_commit(subreq);
 545                subreq = subreq->wb_this_page;
 546        } while (subreq != head);
 547
 548        /* unlink subrequests from head, destroy them later */
 549        if (head->wb_this_page != head) {
 550                /* destroy list will be terminated by head */
 551                destroy_list = head->wb_this_page;
 552                head->wb_this_page = head;
 553
 554                /* change head request to cover whole range that
 555                 * the former page group covered */
 556                head->wb_bytes = total_bytes;
 557        }
 558
 559        /* Postpone destruction of this request */
 560        if (test_and_clear_bit(PG_REMOVE, &head->wb_flags)) {
 561                set_bit(PG_INODE_REF, &head->wb_flags);
 562                kref_get(&head->wb_kref);
 563                atomic_long_inc(&NFS_I(inode)->nrequests);
 564        }
 565
 566        nfs_page_group_unlock(head);
 567
 568        nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
 569
 570        /* Did we lose a race with nfs_inode_remove_request()? */
 571        if (!(PagePrivate(page) || PageSwapCache(page))) {
 572                nfs_unlock_and_release_request(head);
 573                return NULL;
 574        }
 575
 576        /* still holds ref on head from nfs_page_find_head_request
 577         * and still has lock on head from lock loop */
 578        return head;
 579
 580release_request:
 581        nfs_unlock_and_release_request(head);
 582        return ERR_PTR(ret);
 583}
 584
 585static void nfs_write_error_remove_page(struct nfs_page *req)
 586{
 587        nfs_end_page_writeback(req);
 588        generic_error_remove_page(page_file_mapping(req->wb_page),
 589                                  req->wb_page);
 590        nfs_release_request(req);
 591}
 592
 593static bool
 594nfs_error_is_fatal_on_server(int err)
 595{
 596        switch (err) {
 597        case 0:
 598        case -ERESTARTSYS:
 599        case -EINTR:
 600                return false;
 601        }
 602        return nfs_error_is_fatal(err);
 603}
 604
 605/*
 606 * Find an associated nfs write request, and prepare to flush it out
 607 * May return an error if the user signalled nfs_wait_on_request().
 608 */
 609static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 610                                struct page *page)
 611{
 612        struct nfs_page *req;
 613        int ret = 0;
 614
 615        req = nfs_lock_and_join_requests(page);
 616        if (!req)
 617                goto out;
 618        ret = PTR_ERR(req);
 619        if (IS_ERR(req))
 620                goto out;
 621
 622        nfs_set_page_writeback(page);
 623        WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
 624
 625        ret = req->wb_context->error;
 626        /* If there is a fatal error that covers this write, just exit */
 627        if (nfs_error_is_fatal_on_server(ret))
 628                goto out_launder;
 629
 630        ret = 0;
 631        if (!nfs_pageio_add_request(pgio, req)) {
 632                ret = pgio->pg_error;
 633                /*
 634                 * Remove the problematic req upon fatal errors on the server
 635                 */
 636                if (nfs_error_is_fatal(ret)) {
 637                        nfs_context_set_write_error(req->wb_context, ret);
 638                        if (nfs_error_is_fatal_on_server(ret))
 639                                goto out_launder;
 640                } else
 641                        ret = -EAGAIN;
 642                nfs_redirty_request(req);
 643        } else
 644                nfs_add_stats(page_file_mapping(page)->host,
 645                                NFSIOS_WRITEPAGES, 1);
 646out:
 647        return ret;
 648out_launder:
 649        nfs_write_error_remove_page(req);
 650        return ret;
 651}
 652
 653static int nfs_do_writepage(struct page *page, struct writeback_control *wbc,
 654                            struct nfs_pageio_descriptor *pgio)
 655{
 656        int ret;
 657
 658        nfs_pageio_cond_complete(pgio, page_index(page));
 659        ret = nfs_page_async_flush(pgio, page);
 660        if (ret == -EAGAIN) {
 661                redirty_page_for_writepage(wbc, page);
 662                ret = 0;
 663        }
 664        return ret;
 665}
 666
 667/*
 668 * Write an mmapped page to the server.
 669 */
 670static int nfs_writepage_locked(struct page *page,
 671                                struct writeback_control *wbc)
 672{
 673        struct nfs_pageio_descriptor pgio;
 674        struct inode *inode = page_file_mapping(page)->host;
 675        int err;
 676
 677        nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 678        nfs_pageio_init_write(&pgio, inode, 0,
 679                                false, &nfs_async_write_completion_ops);
 680        err = nfs_do_writepage(page, wbc, &pgio);
 681        nfs_pageio_complete(&pgio);
 682        if (err < 0)
 683                return err;
 684        if (pgio.pg_error < 0)
 685                return pgio.pg_error;
 686        return 0;
 687}
 688
 689int nfs_writepage(struct page *page, struct writeback_control *wbc)
 690{
 691        int ret;
 692
 693        ret = nfs_writepage_locked(page, wbc);
 694        unlock_page(page);
 695        return ret;
 696}
 697
 698static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 699{
 700        int ret;
 701
 702        ret = nfs_do_writepage(page, wbc, data);
 703        unlock_page(page);
 704        return ret;
 705}
 706
 707static void nfs_io_completion_commit(void *inode)
 708{
 709        nfs_commit_inode(inode, 0);
 710}
 711
 712int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 713{
 714        struct inode *inode = mapping->host;
 715        struct nfs_pageio_descriptor pgio;
 716        struct nfs_io_completion *ioc;
 717        unsigned int pflags = memalloc_nofs_save();
 718        int err;
 719
 720        nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 721
 722        ioc = nfs_io_completion_alloc(GFP_NOFS);
 723        if (ioc)
 724                nfs_io_completion_init(ioc, nfs_io_completion_commit, inode);
 725
 726        nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
 727                                &nfs_async_write_completion_ops);
 728        pgio.pg_io_completion = ioc;
 729        err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
 730        nfs_pageio_complete(&pgio);
 731        nfs_io_completion_put(ioc);
 732
 733        memalloc_nofs_restore(pflags);
 734
 735        if (err < 0)
 736                goto out_err;
 737        err = pgio.pg_error;
 738        if (err < 0)
 739                goto out_err;
 740        return 0;
 741out_err:
 742        return err;
 743}
 744
 745/*
 746 * Insert a write request into an inode
 747 */
 748static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 749{
 750        struct address_space *mapping = page_file_mapping(req->wb_page);
 751        struct nfs_inode *nfsi = NFS_I(inode);
 752
 753        WARN_ON_ONCE(req->wb_this_page != req);
 754
 755        /* Lock the request! */
 756        nfs_lock_request(req);
 757
 758        /*
 759         * Swap-space should not get truncated. Hence no need to plug the race
 760         * with invalidate/truncate.
 761         */
 762        spin_lock(&mapping->private_lock);
 763        if (!nfs_have_writebacks(inode) &&
 764            NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
 765                inode_inc_iversion_raw(inode);
 766        if (likely(!PageSwapCache(req->wb_page))) {
 767                set_bit(PG_MAPPED, &req->wb_flags);
 768                SetPagePrivate(req->wb_page);
 769                set_page_private(req->wb_page, (unsigned long)req);
 770        }
 771        spin_unlock(&mapping->private_lock);
 772        atomic_long_inc(&nfsi->nrequests);
 773        /* this a head request for a page group - mark it as having an
 774         * extra reference so sub groups can follow suit.
 775         * This flag also informs pgio layer when to bump nrequests when
 776         * adding subrequests. */
 777        WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
 778        kref_get(&req->wb_kref);
 779}
 780
 781/*
 782 * Remove a write request from an inode
 783 */
 784static void nfs_inode_remove_request(struct nfs_page *req)
 785{
 786        struct address_space *mapping = page_file_mapping(req->wb_page);
 787        struct inode *inode = mapping->host;
 788        struct nfs_inode *nfsi = NFS_I(inode);
 789        struct nfs_page *head;
 790
 791        atomic_long_dec(&nfsi->nrequests);
 792        if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
 793                head = req->wb_head;
 794
 795                spin_lock(&mapping->private_lock);
 796                if (likely(head->wb_page && !PageSwapCache(head->wb_page))) {
 797                        set_page_private(head->wb_page, 0);
 798                        ClearPagePrivate(head->wb_page);
 799                        clear_bit(PG_MAPPED, &head->wb_flags);
 800                }
 801                spin_unlock(&mapping->private_lock);
 802        }
 803
 804        if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags))
 805                nfs_release_request(req);
 806}
 807
 808static void
 809nfs_mark_request_dirty(struct nfs_page *req)
 810{
 811        if (req->wb_page)
 812                __set_page_dirty_nobuffers(req->wb_page);
 813}
 814
 815/*
 816 * nfs_page_search_commits_for_head_request_locked
 817 *
 818 * Search through commit lists on @inode for the head request for @page.
 819 * Must be called while holding the inode (which is cinfo) lock.
 820 *
 821 * Returns the head request if found, or NULL if not found.
 822 */
 823static struct nfs_page *
 824nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
 825                                                struct page *page)
 826{
 827        struct nfs_page *freq, *t;
 828        struct nfs_commit_info cinfo;
 829        struct inode *inode = &nfsi->vfs_inode;
 830
 831        nfs_init_cinfo_from_inode(&cinfo, inode);
 832
 833        /* search through pnfs commit lists */
 834        freq = pnfs_search_commit_reqs(inode, &cinfo, page);
 835        if (freq)
 836                return freq->wb_head;
 837
 838        /* Linearly search the commit list for the correct request */
 839        list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
 840                if (freq->wb_page == page)
 841                        return freq->wb_head;
 842        }
 843
 844        return NULL;
 845}
 846
 847/**
 848 * nfs_request_add_commit_list_locked - add request to a commit list
 849 * @req: pointer to a struct nfs_page
 850 * @dst: commit list head
 851 * @cinfo: holds list lock and accounting info
 852 *
 853 * This sets the PG_CLEAN bit, updates the cinfo count of
 854 * number of outstanding requests requiring a commit as well as
 855 * the MM page stats.
 856 *
 857 * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
 858 * nfs_page lock.
 859 */
 860void
 861nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
 862                            struct nfs_commit_info *cinfo)
 863{
 864        set_bit(PG_CLEAN, &req->wb_flags);
 865        nfs_list_add_request(req, dst);
 866        atomic_long_inc(&cinfo->mds->ncommit);
 867}
 868EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
 869
 870/**
 871 * nfs_request_add_commit_list - add request to a commit list
 872 * @req: pointer to a struct nfs_page
 873 * @cinfo: holds list lock and accounting info
 874 *
 875 * This sets the PG_CLEAN bit, updates the cinfo count of
 876 * number of outstanding requests requiring a commit as well as
 877 * the MM page stats.
 878 *
 879 * The caller must _not_ hold the cinfo->lock, but must be
 880 * holding the nfs_page lock.
 881 */
 882void
 883nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
 884{
 885        mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 886        nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
 887        mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
 888        if (req->wb_page)
 889                nfs_mark_page_unstable(req->wb_page, cinfo);
 890}
 891EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
 892
 893/**
 894 * nfs_request_remove_commit_list - Remove request from a commit list
 895 * @req: pointer to a nfs_page
 896 * @cinfo: holds list lock and accounting info
 897 *
 898 * This clears the PG_CLEAN bit, and updates the cinfo's count of
 899 * number of outstanding requests requiring a commit
 900 * It does not update the MM page stats.
 901 *
 902 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
 903 */
 904void
 905nfs_request_remove_commit_list(struct nfs_page *req,
 906                               struct nfs_commit_info *cinfo)
 907{
 908        if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
 909                return;
 910        nfs_list_remove_request(req);
 911        atomic_long_dec(&cinfo->mds->ncommit);
 912}
 913EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
 914
 915static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 916                                      struct inode *inode)
 917{
 918        cinfo->inode = inode;
 919        cinfo->mds = &NFS_I(inode)->commit_info;
 920        cinfo->ds = pnfs_get_ds_info(inode);
 921        cinfo->dreq = NULL;
 922        cinfo->completion_ops = &nfs_commit_completion_ops;
 923}
 924
 925void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 926                    struct inode *inode,
 927                    struct nfs_direct_req *dreq)
 928{
 929        if (dreq)
 930                nfs_init_cinfo_from_dreq(cinfo, dreq);
 931        else
 932                nfs_init_cinfo_from_inode(cinfo, inode);
 933}
 934EXPORT_SYMBOL_GPL(nfs_init_cinfo);
 935
 936/*
 937 * Add a request to the inode's commit list.
 938 */
 939void
 940nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 941                        struct nfs_commit_info *cinfo, u32 ds_commit_idx)
 942{
 943        if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
 944                return;
 945        nfs_request_add_commit_list(req, cinfo);
 946}
 947
 948static void
 949nfs_clear_page_commit(struct page *page)
 950{
 951        dec_node_page_state(page, NR_UNSTABLE_NFS);
 952        dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
 953                    WB_RECLAIMABLE);
 954}
 955
 956/* Called holding the request lock on @req */
 957static void
 958nfs_clear_request_commit(struct nfs_page *req)
 959{
 960        if (test_bit(PG_CLEAN, &req->wb_flags)) {
 961                struct inode *inode = d_inode(req->wb_context->dentry);
 962                struct nfs_commit_info cinfo;
 963
 964                nfs_init_cinfo_from_inode(&cinfo, inode);
 965                mutex_lock(&NFS_I(inode)->commit_mutex);
 966                if (!pnfs_clear_request_commit(req, &cinfo)) {
 967                        nfs_request_remove_commit_list(req, &cinfo);
 968                }
 969                mutex_unlock(&NFS_I(inode)->commit_mutex);
 970                nfs_clear_page_commit(req->wb_page);
 971        }
 972}
 973
 974int nfs_write_need_commit(struct nfs_pgio_header *hdr)
 975{
 976        if (hdr->verf.committed == NFS_DATA_SYNC)
 977                return hdr->lseg == NULL;
 978        return hdr->verf.committed != NFS_FILE_SYNC;
 979}
 980
 981static void nfs_async_write_init(struct nfs_pgio_header *hdr)
 982{
 983        nfs_io_completion_get(hdr->io_completion);
 984}
 985
 986static void nfs_write_completion(struct nfs_pgio_header *hdr)
 987{
 988        struct nfs_commit_info cinfo;
 989        unsigned long bytes = 0;
 990
 991        if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 992                goto out;
 993        nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
 994        while (!list_empty(&hdr->pages)) {
 995                struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 996
 997                bytes += req->wb_bytes;
 998                nfs_list_remove_request(req);
 999                if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
1000                    (hdr->good_bytes < bytes)) {
1001                        nfs_set_pageerror(page_file_mapping(req->wb_page));
1002                        nfs_context_set_write_error(req->wb_context, hdr->error);
1003                        goto remove_req;
1004                }
1005                if (nfs_write_need_commit(hdr)) {
1006                        memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1007                        nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1008                                hdr->pgio_mirror_idx);
1009                        goto next;
1010                }
1011remove_req:
1012                nfs_inode_remove_request(req);
1013next:
1014                nfs_end_page_writeback(req);
1015                nfs_release_request(req);
1016        }
1017out:
1018        nfs_io_completion_put(hdr->io_completion);
1019        hdr->release(hdr);
1020}
1021
1022unsigned long
1023nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1024{
1025        return atomic_long_read(&cinfo->mds->ncommit);
1026}
1027
1028/* NFS_I(cinfo->inode)->commit_mutex held by caller */
1029int
1030nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1031                     struct nfs_commit_info *cinfo, int max)
1032{
1033        struct nfs_page *req, *tmp;
1034        int ret = 0;
1035
1036restart:
1037        list_for_each_entry_safe(req, tmp, src, wb_list) {
1038                kref_get(&req->wb_kref);
1039                if (!nfs_lock_request(req)) {
1040                        int status;
1041
1042                        /* Prevent deadlock with nfs_lock_and_join_requests */
1043                        if (!list_empty(dst)) {
1044                                nfs_release_request(req);
1045                                continue;
1046                        }
1047                        /* Ensure we make progress to prevent livelock */
1048                        mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1049                        status = nfs_wait_on_request(req);
1050                        nfs_release_request(req);
1051                        mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1052                        if (status < 0)
1053                                break;
1054                        goto restart;
1055                }
1056                nfs_request_remove_commit_list(req, cinfo);
1057                clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1058                nfs_list_add_request(req, dst);
1059                ret++;
1060                if ((ret == max) && !cinfo->dreq)
1061                        break;
1062                cond_resched();
1063        }
1064        return ret;
1065}
1066EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1067
1068/*
1069 * nfs_scan_commit - Scan an inode for commit requests
1070 * @inode: NFS inode to scan
1071 * @dst: mds destination list
1072 * @cinfo: mds and ds lists of reqs ready to commit
1073 *
1074 * Moves requests from the inode's 'commit' request list.
1075 * The requests are *not* checked to ensure that they form a contiguous set.
1076 */
1077int
1078nfs_scan_commit(struct inode *inode, struct list_head *dst,
1079                struct nfs_commit_info *cinfo)
1080{
1081        int ret = 0;
1082
1083        if (!atomic_long_read(&cinfo->mds->ncommit))
1084                return 0;
1085        mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1086        if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1087                const int max = INT_MAX;
1088
1089                ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1090                                           cinfo, max);
1091                ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1092        }
1093        mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1094        return ret;
1095}
1096
1097/*
1098 * Search for an existing write request, and attempt to update
1099 * it to reflect a new dirty region on a given page.
1100 *
1101 * If the attempt fails, then the existing request is flushed out
1102 * to disk.
1103 */
1104static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1105                struct page *page,
1106                unsigned int offset,
1107                unsigned int bytes)
1108{
1109        struct nfs_page *req;
1110        unsigned int rqend;
1111        unsigned int end;
1112        int error;
1113
1114        end = offset + bytes;
1115
1116        req = nfs_lock_and_join_requests(page);
1117        if (IS_ERR_OR_NULL(req))
1118                return req;
1119
1120        rqend = req->wb_offset + req->wb_bytes;
1121        /*
1122         * Tell the caller to flush out the request if
1123         * the offsets are non-contiguous.
1124         * Note: nfs_flush_incompatible() will already
1125         * have flushed out requests having wrong owners.
1126         */
1127        if (offset > rqend || end < req->wb_offset)
1128                goto out_flushme;
1129
1130        /* Okay, the request matches. Update the region */
1131        if (offset < req->wb_offset) {
1132                req->wb_offset = offset;
1133                req->wb_pgbase = offset;
1134        }
1135        if (end > rqend)
1136                req->wb_bytes = end - req->wb_offset;
1137        else
1138                req->wb_bytes = rqend - req->wb_offset;
1139        return req;
1140out_flushme:
1141        /*
1142         * Note: we mark the request dirty here because
1143         * nfs_lock_and_join_requests() cannot preserve
1144         * commit flags, so we have to replay the write.
1145         */
1146        nfs_mark_request_dirty(req);
1147        nfs_unlock_and_release_request(req);
1148        error = nfs_wb_page(inode, page);
1149        return (error < 0) ? ERR_PTR(error) : NULL;
1150}
1151
1152/*
1153 * Try to update an existing write request, or create one if there is none.
1154 *
1155 * Note: Should always be called with the Page Lock held to prevent races
1156 * if we have to add a new request. Also assumes that the caller has
1157 * already called nfs_flush_incompatible() if necessary.
1158 */
1159static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1160                struct page *page, unsigned int offset, unsigned int bytes)
1161{
1162        struct inode *inode = page_file_mapping(page)->host;
1163        struct nfs_page *req;
1164
1165        req = nfs_try_to_update_request(inode, page, offset, bytes);
1166        if (req != NULL)
1167                goto out;
1168        req = nfs_create_request(ctx, page, NULL, offset, bytes);
1169        if (IS_ERR(req))
1170                goto out;
1171        nfs_inode_add_request(inode, req);
1172out:
1173        return req;
1174}
1175
1176static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1177                unsigned int offset, unsigned int count)
1178{
1179        struct nfs_page *req;
1180
1181        req = nfs_setup_write_request(ctx, page, offset, count);
1182        if (IS_ERR(req))
1183                return PTR_ERR(req);
1184        /* Update file length */
1185        nfs_grow_file(page, offset, count);
1186        nfs_mark_uptodate(req);
1187        nfs_mark_request_dirty(req);
1188        nfs_unlock_and_release_request(req);
1189        return 0;
1190}
1191
1192int nfs_flush_incompatible(struct file *file, struct page *page)
1193{
1194        struct nfs_open_context *ctx = nfs_file_open_context(file);
1195        struct nfs_lock_context *l_ctx;
1196        struct file_lock_context *flctx = file_inode(file)->i_flctx;
1197        struct nfs_page *req;
1198        int do_flush, status;
1199        /*
1200         * Look for a request corresponding to this page. If there
1201         * is one, and it belongs to another file, we flush it out
1202         * before we try to copy anything into the page. Do this
1203         * due to the lack of an ACCESS-type call in NFSv2.
1204         * Also do the same if we find a request from an existing
1205         * dropped page.
1206         */
1207        do {
1208                req = nfs_page_find_head_request(page);
1209                if (req == NULL)
1210                        return 0;
1211                l_ctx = req->wb_lock_context;
1212                do_flush = req->wb_page != page ||
1213                        !nfs_match_open_context(req->wb_context, ctx);
1214                if (l_ctx && flctx &&
1215                    !(list_empty_careful(&flctx->flc_posix) &&
1216                      list_empty_careful(&flctx->flc_flock))) {
1217                        do_flush |= l_ctx->lockowner != current->files;
1218                }
1219                nfs_release_request(req);
1220                if (!do_flush)
1221                        return 0;
1222                status = nfs_wb_page(page_file_mapping(page)->host, page);
1223        } while (status == 0);
1224        return status;
1225}
1226
1227/*
1228 * Avoid buffered writes when a open context credential's key would
1229 * expire soon.
1230 *
1231 * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1232 *
1233 * Return 0 and set a credential flag which triggers the inode to flush
1234 * and performs  NFS_FILE_SYNC writes if the key will expired within
1235 * RPC_KEY_EXPIRE_TIMEO.
1236 */
1237int
1238nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1239{
1240        struct nfs_open_context *ctx = nfs_file_open_context(filp);
1241
1242        if (nfs_ctx_key_to_expire(ctx, inode) &&
1243            !ctx->ll_cred)
1244                /* Already expired! */
1245                return -EACCES;
1246        return 0;
1247}
1248
1249/*
1250 * Test if the open context credential key is marked to expire soon.
1251 */
1252bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1253{
1254        struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1255        struct rpc_cred *cred = ctx->ll_cred;
1256        struct auth_cred acred = {
1257                .cred = ctx->cred,
1258        };
1259
1260        if (cred && !cred->cr_ops->crmatch(&acred, cred, 0)) {
1261                put_rpccred(cred);
1262                ctx->ll_cred = NULL;
1263                cred = NULL;
1264        }
1265        if (!cred)
1266                cred = auth->au_ops->lookup_cred(auth, &acred, 0);
1267        if (!cred || IS_ERR(cred))
1268                return true;
1269        ctx->ll_cred = cred;
1270        return !!(cred->cr_ops->crkey_timeout &&
1271                  cred->cr_ops->crkey_timeout(cred));
1272}
1273
1274/*
1275 * If the page cache is marked as unsafe or invalid, then we can't rely on
1276 * the PageUptodate() flag. In this case, we will need to turn off
1277 * write optimisations that depend on the page contents being correct.
1278 */
1279static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
1280{
1281        struct nfs_inode *nfsi = NFS_I(inode);
1282
1283        if (nfs_have_delegated_attributes(inode))
1284                goto out;
1285        if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE)
1286                return false;
1287        smp_rmb();
1288        if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
1289                return false;
1290out:
1291        if (nfsi->cache_validity & NFS_INO_INVALID_DATA)
1292                return false;
1293        return PageUptodate(page) != 0;
1294}
1295
1296static bool
1297is_whole_file_wrlock(struct file_lock *fl)
1298{
1299        return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1300                        fl->fl_type == F_WRLCK;
1301}
1302
1303/* If we know the page is up to date, and we're not using byte range locks (or
1304 * if we have the whole file locked for writing), it may be more efficient to
1305 * extend the write to cover the entire page in order to avoid fragmentation
1306 * inefficiencies.
1307 *
1308 * If the file is opened for synchronous writes then we can just skip the rest
1309 * of the checks.
1310 */
1311static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
1312{
1313        int ret;
1314        struct file_lock_context *flctx = inode->i_flctx;
1315        struct file_lock *fl;
1316
1317        if (file->f_flags & O_DSYNC)
1318                return 0;
1319        if (!nfs_write_pageuptodate(page, inode))
1320                return 0;
1321        if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1322                return 1;
1323        if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1324                       list_empty_careful(&flctx->flc_posix)))
1325                return 1;
1326
1327        /* Check to see if there are whole file write locks */
1328        ret = 0;
1329        spin_lock(&flctx->flc_lock);
1330        if (!list_empty(&flctx->flc_posix)) {
1331                fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1332                                        fl_list);
1333                if (is_whole_file_wrlock(fl))
1334                        ret = 1;
1335        } else if (!list_empty(&flctx->flc_flock)) {
1336                fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1337                                        fl_list);
1338                if (fl->fl_type == F_WRLCK)
1339                        ret = 1;
1340        }
1341        spin_unlock(&flctx->flc_lock);
1342        return ret;
1343}
1344
1345/*
1346 * Update and possibly write a cached page of an NFS file.
1347 *
1348 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1349 * things with a page scheduled for an RPC call (e.g. invalidate it).
1350 */
1351int nfs_updatepage(struct file *file, struct page *page,
1352                unsigned int offset, unsigned int count)
1353{
1354        struct nfs_open_context *ctx = nfs_file_open_context(file);
1355        struct address_space *mapping = page_file_mapping(page);
1356        struct inode    *inode = mapping->host;
1357        int             status = 0;
1358
1359        nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1360
1361        dprintk("NFS:       nfs_updatepage(%pD2 %d@%lld)\n",
1362                file, count, (long long)(page_file_offset(page) + offset));
1363
1364        if (!count)
1365                goto out;
1366
1367        if (nfs_can_extend_write(file, page, inode)) {
1368                count = max(count + offset, nfs_page_length(page));
1369                offset = 0;
1370        }
1371
1372        status = nfs_writepage_setup(ctx, page, offset, count);
1373        if (status < 0)
1374                nfs_set_pageerror(mapping);
1375        else
1376                __set_page_dirty_nobuffers(page);
1377out:
1378        dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
1379                        status, (long long)i_size_read(inode));
1380        return status;
1381}
1382
1383static int flush_task_priority(int how)
1384{
1385        switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1386                case FLUSH_HIGHPRI:
1387                        return RPC_PRIORITY_HIGH;
1388                case FLUSH_LOWPRI:
1389                        return RPC_PRIORITY_LOW;
1390        }
1391        return RPC_PRIORITY_NORMAL;
1392}
1393
1394static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1395                               struct rpc_message *msg,
1396                               const struct nfs_rpc_ops *rpc_ops,
1397                               struct rpc_task_setup *task_setup_data, int how)
1398{
1399        int priority = flush_task_priority(how);
1400
1401        task_setup_data->priority = priority;
1402        rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1403        trace_nfs_initiate_write(hdr->inode, hdr->io_start, hdr->good_bytes,
1404                                 hdr->args.stable);
1405}
1406
1407/* If a nfs_flush_* function fails, it should remove reqs from @head and
1408 * call this on each, which will prepare them to be retried on next
1409 * writeback using standard nfs.
1410 */
1411static void nfs_redirty_request(struct nfs_page *req)
1412{
1413        nfs_mark_request_dirty(req);
1414        set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1415        nfs_end_page_writeback(req);
1416        nfs_release_request(req);
1417}
1418
1419static void nfs_async_write_error(struct list_head *head, int error)
1420{
1421        struct nfs_page *req;
1422
1423        while (!list_empty(head)) {
1424                req = nfs_list_entry(head->next);
1425                nfs_list_remove_request(req);
1426                if (nfs_error_is_fatal(error)) {
1427                        nfs_context_set_write_error(req->wb_context, error);
1428                        if (nfs_error_is_fatal_on_server(error)) {
1429                                nfs_write_error_remove_page(req);
1430                                continue;
1431                        }
1432                }
1433                nfs_redirty_request(req);
1434        }
1435}
1436
1437static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1438{
1439        nfs_async_write_error(&hdr->pages, 0);
1440        filemap_fdatawrite_range(hdr->inode->i_mapping, hdr->args.offset,
1441                        hdr->args.offset + hdr->args.count - 1);
1442}
1443
1444static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1445        .init_hdr = nfs_async_write_init,
1446        .error_cleanup = nfs_async_write_error,
1447        .completion = nfs_write_completion,
1448        .reschedule_io = nfs_async_write_reschedule_io,
1449};
1450
1451void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1452                               struct inode *inode, int ioflags, bool force_mds,
1453                               const struct nfs_pgio_completion_ops *compl_ops)
1454{
1455        struct nfs_server *server = NFS_SERVER(inode);
1456        const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1457
1458#ifdef CONFIG_NFS_V4_1
1459        if (server->pnfs_curr_ld && !force_mds)
1460                pg_ops = server->pnfs_curr_ld->pg_write_ops;
1461#endif
1462        nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1463                        server->wsize, ioflags);
1464}
1465EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1466
1467void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1468{
1469        struct nfs_pgio_mirror *mirror;
1470
1471        if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1472                pgio->pg_ops->pg_cleanup(pgio);
1473
1474        pgio->pg_ops = &nfs_pgio_rw_ops;
1475
1476        nfs_pageio_stop_mirroring(pgio);
1477
1478        mirror = &pgio->pg_mirrors[0];
1479        mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1480}
1481EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1482
1483
1484void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1485{
1486        struct nfs_commit_data *data = calldata;
1487
1488        NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1489}
1490
1491/*
1492 * Special version of should_remove_suid() that ignores capabilities.
1493 */
1494static int nfs_should_remove_suid(const struct inode *inode)
1495{
1496        umode_t mode = inode->i_mode;
1497        int kill = 0;
1498
1499        /* suid always must be killed */
1500        if (unlikely(mode & S_ISUID))
1501                kill = ATTR_KILL_SUID;
1502
1503        /*
1504         * sgid without any exec bits is just a mandatory locking mark; leave
1505         * it alone.  If some exec bits are set, it's a real sgid; kill it.
1506         */
1507        if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1508                kill |= ATTR_KILL_SGID;
1509
1510        if (unlikely(kill && S_ISREG(mode)))
1511                return kill;
1512
1513        return 0;
1514}
1515
1516static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1517                struct nfs_fattr *fattr)
1518{
1519        struct nfs_pgio_args *argp = &hdr->args;
1520        struct nfs_pgio_res *resp = &hdr->res;
1521        u64 size = argp->offset + resp->count;
1522
1523        if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1524                fattr->size = size;
1525        if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1526                fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1527                return;
1528        }
1529        if (size != fattr->size)
1530                return;
1531        /* Set attribute barrier */
1532        nfs_fattr_set_barrier(fattr);
1533        /* ...and update size */
1534        fattr->valid |= NFS_ATTR_FATTR_SIZE;
1535}
1536
1537void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1538{
1539        struct nfs_fattr *fattr = &hdr->fattr;
1540        struct inode *inode = hdr->inode;
1541
1542        spin_lock(&inode->i_lock);
1543        nfs_writeback_check_extend(hdr, fattr);
1544        nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1545        spin_unlock(&inode->i_lock);
1546}
1547EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1548
1549/*
1550 * This function is called when the WRITE call is complete.
1551 */
1552static int nfs_writeback_done(struct rpc_task *task,
1553                              struct nfs_pgio_header *hdr,
1554                              struct inode *inode)
1555{
1556        int status;
1557
1558        /*
1559         * ->write_done will attempt to use post-op attributes to detect
1560         * conflicting writes by other clients.  A strict interpretation
1561         * of close-to-open would allow us to continue caching even if
1562         * another writer had changed the file, but some applications
1563         * depend on tighter cache coherency when writing.
1564         */
1565        status = NFS_PROTO(inode)->write_done(task, hdr);
1566        if (status != 0)
1567                return status;
1568
1569        nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1570        trace_nfs_writeback_done(inode, task->tk_status,
1571                                 hdr->args.offset, hdr->res.verf);
1572
1573        if (hdr->res.verf->committed < hdr->args.stable &&
1574            task->tk_status >= 0) {
1575                /* We tried a write call, but the server did not
1576                 * commit data to stable storage even though we
1577                 * requested it.
1578                 * Note: There is a known bug in Tru64 < 5.0 in which
1579                 *       the server reports NFS_DATA_SYNC, but performs
1580                 *       NFS_FILE_SYNC. We therefore implement this checking
1581                 *       as a dprintk() in order to avoid filling syslog.
1582                 */
1583                static unsigned long    complain;
1584
1585                /* Note this will print the MDS for a DS write */
1586                if (time_before(complain, jiffies)) {
1587                        dprintk("NFS:       faulty NFS server %s:"
1588                                " (committed = %d) != (stable = %d)\n",
1589                                NFS_SERVER(inode)->nfs_client->cl_hostname,
1590                                hdr->res.verf->committed, hdr->args.stable);
1591                        complain = jiffies + 300 * HZ;
1592                }
1593        }
1594
1595        /* Deal with the suid/sgid bit corner case */
1596        if (nfs_should_remove_suid(inode)) {
1597                spin_lock(&inode->i_lock);
1598                NFS_I(inode)->cache_validity |= NFS_INO_INVALID_OTHER;
1599                spin_unlock(&inode->i_lock);
1600        }
1601        return 0;
1602}
1603
1604/*
1605 * This function is called when the WRITE call is complete.
1606 */
1607static void nfs_writeback_result(struct rpc_task *task,
1608                                 struct nfs_pgio_header *hdr)
1609{
1610        struct nfs_pgio_args    *argp = &hdr->args;
1611        struct nfs_pgio_res     *resp = &hdr->res;
1612
1613        if (resp->count < argp->count) {
1614                static unsigned long    complain;
1615
1616                /* This a short write! */
1617                nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1618
1619                /* Has the server at least made some progress? */
1620                if (resp->count == 0) {
1621                        if (time_before(complain, jiffies)) {
1622                                printk(KERN_WARNING
1623                                       "NFS: Server wrote zero bytes, expected %u.\n",
1624                                       argp->count);
1625                                complain = jiffies + 300 * HZ;
1626                        }
1627                        nfs_set_pgio_error(hdr, -EIO, argp->offset);
1628                        task->tk_status = -EIO;
1629                        return;
1630                }
1631
1632                /* For non rpc-based layout drivers, retry-through-MDS */
1633                if (!task->tk_ops) {
1634                        hdr->pnfs_error = -EAGAIN;
1635                        return;
1636                }
1637
1638                /* Was this an NFSv2 write or an NFSv3 stable write? */
1639                if (resp->verf->committed != NFS_UNSTABLE) {
1640                        /* Resend from where the server left off */
1641                        hdr->mds_offset += resp->count;
1642                        argp->offset += resp->count;
1643                        argp->pgbase += resp->count;
1644                        argp->count -= resp->count;
1645                } else {
1646                        /* Resend as a stable write in order to avoid
1647                         * headaches in the case of a server crash.
1648                         */
1649                        argp->stable = NFS_FILE_SYNC;
1650                }
1651                rpc_restart_call_prepare(task);
1652        }
1653}
1654
1655static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1656{
1657        return wait_var_event_killable(&cinfo->rpcs_out,
1658                                       !atomic_read(&cinfo->rpcs_out));
1659}
1660
1661static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1662{
1663        atomic_inc(&cinfo->rpcs_out);
1664}
1665
1666static void nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1667{
1668        if (atomic_dec_and_test(&cinfo->rpcs_out))
1669                wake_up_var(&cinfo->rpcs_out);
1670}
1671
1672void nfs_commitdata_release(struct nfs_commit_data *data)
1673{
1674        put_nfs_open_context(data->context);
1675        nfs_commit_free(data);
1676}
1677EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1678
1679int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1680                        const struct nfs_rpc_ops *nfs_ops,
1681                        const struct rpc_call_ops *call_ops,
1682                        int how, int flags)
1683{
1684        struct rpc_task *task;
1685        int priority = flush_task_priority(how);
1686        struct rpc_message msg = {
1687                .rpc_argp = &data->args,
1688                .rpc_resp = &data->res,
1689                .rpc_cred = data->cred,
1690        };
1691        struct rpc_task_setup task_setup_data = {
1692                .task = &data->task,
1693                .rpc_client = clnt,
1694                .rpc_message = &msg,
1695                .callback_ops = call_ops,
1696                .callback_data = data,
1697                .workqueue = nfsiod_workqueue,
1698                .flags = RPC_TASK_ASYNC | flags,
1699                .priority = priority,
1700        };
1701        /* Set up the initial task struct.  */
1702        nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1703        trace_nfs_initiate_commit(data);
1704
1705        dprintk("NFS: initiated commit call\n");
1706
1707        task = rpc_run_task(&task_setup_data);
1708        if (IS_ERR(task))
1709                return PTR_ERR(task);
1710        if (how & FLUSH_SYNC)
1711                rpc_wait_for_completion_task(task);
1712        rpc_put_task(task);
1713        return 0;
1714}
1715EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1716
1717static loff_t nfs_get_lwb(struct list_head *head)
1718{
1719        loff_t lwb = 0;
1720        struct nfs_page *req;
1721
1722        list_for_each_entry(req, head, wb_list)
1723                if (lwb < (req_offset(req) + req->wb_bytes))
1724                        lwb = req_offset(req) + req->wb_bytes;
1725
1726        return lwb;
1727}
1728
1729/*
1730 * Set up the argument/result storage required for the RPC call.
1731 */
1732void nfs_init_commit(struct nfs_commit_data *data,
1733                     struct list_head *head,
1734                     struct pnfs_layout_segment *lseg,
1735                     struct nfs_commit_info *cinfo)
1736{
1737        struct nfs_page *first = nfs_list_entry(head->next);
1738        struct inode *inode = d_inode(first->wb_context->dentry);
1739
1740        /* Set up the RPC argument and reply structs
1741         * NB: take care not to mess about with data->commit et al. */
1742
1743        list_splice_init(head, &data->pages);
1744
1745        data->inode       = inode;
1746        data->cred        = first->wb_context->cred;
1747        data->lseg        = lseg; /* reference transferred */
1748        /* only set lwb for pnfs commit */
1749        if (lseg)
1750                data->lwb = nfs_get_lwb(&data->pages);
1751        data->mds_ops     = &nfs_commit_ops;
1752        data->completion_ops = cinfo->completion_ops;
1753        data->dreq        = cinfo->dreq;
1754
1755        data->args.fh     = NFS_FH(data->inode);
1756        /* Note: we always request a commit of the entire inode */
1757        data->args.offset = 0;
1758        data->args.count  = 0;
1759        data->context     = get_nfs_open_context(first->wb_context);
1760        data->res.fattr   = &data->fattr;
1761        data->res.verf    = &data->verf;
1762        nfs_fattr_init(&data->fattr);
1763}
1764EXPORT_SYMBOL_GPL(nfs_init_commit);
1765
1766void nfs_retry_commit(struct list_head *page_list,
1767                      struct pnfs_layout_segment *lseg,
1768                      struct nfs_commit_info *cinfo,
1769                      u32 ds_commit_idx)
1770{
1771        struct nfs_page *req;
1772
1773        while (!list_empty(page_list)) {
1774                req = nfs_list_entry(page_list->next);
1775                nfs_list_remove_request(req);
1776                nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1777                if (!cinfo->dreq)
1778                        nfs_clear_page_commit(req->wb_page);
1779                nfs_unlock_and_release_request(req);
1780        }
1781}
1782EXPORT_SYMBOL_GPL(nfs_retry_commit);
1783
1784static void
1785nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1786                struct nfs_page *req)
1787{
1788        __set_page_dirty_nobuffers(req->wb_page);
1789}
1790
1791/*
1792 * Commit dirty pages
1793 */
1794static int
1795nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1796                struct nfs_commit_info *cinfo)
1797{
1798        struct nfs_commit_data  *data;
1799
1800        /* another commit raced with us */
1801        if (list_empty(head))
1802                return 0;
1803
1804        data = nfs_commitdata_alloc(true);
1805
1806        /* Set up the argument struct */
1807        nfs_init_commit(data, head, NULL, cinfo);
1808        atomic_inc(&cinfo->mds->rpcs_out);
1809        return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1810                                   data->mds_ops, how, 0);
1811}
1812
1813/*
1814 * COMMIT call returned
1815 */
1816static void nfs_commit_done(struct rpc_task *task, void *calldata)
1817{
1818        struct nfs_commit_data  *data = calldata;
1819
1820        dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1821                                task->tk_pid, task->tk_status);
1822
1823        /* Call the NFS version-specific code */
1824        NFS_PROTO(data->inode)->commit_done(task, data);
1825        trace_nfs_commit_done(data);
1826}
1827
1828static void nfs_commit_release_pages(struct nfs_commit_data *data)
1829{
1830        struct nfs_page *req;
1831        int status = data->task.tk_status;
1832        struct nfs_commit_info cinfo;
1833        struct nfs_server *nfss;
1834
1835        while (!list_empty(&data->pages)) {
1836                req = nfs_list_entry(data->pages.next);
1837                nfs_list_remove_request(req);
1838                if (req->wb_page)
1839                        nfs_clear_page_commit(req->wb_page);
1840
1841                dprintk("NFS:       commit (%s/%llu %d@%lld)",
1842                        req->wb_context->dentry->d_sb->s_id,
1843                        (unsigned long long)NFS_FILEID(d_inode(req->wb_context->dentry)),
1844                        req->wb_bytes,
1845                        (long long)req_offset(req));
1846                if (status < 0) {
1847                        nfs_context_set_write_error(req->wb_context, status);
1848                        if (req->wb_page)
1849                                nfs_inode_remove_request(req);
1850                        dprintk_cont(", error = %d\n", status);
1851                        goto next;
1852                }
1853
1854                /* Okay, COMMIT succeeded, apparently. Check the verifier
1855                 * returned by the server against all stored verfs. */
1856                if (!nfs_write_verifier_cmp(&req->wb_verf, &data->verf.verifier)) {
1857                        /* We have a match */
1858                        if (req->wb_page)
1859                                nfs_inode_remove_request(req);
1860                        dprintk_cont(" OK\n");
1861                        goto next;
1862                }
1863                /* We have a mismatch. Write the page again */
1864                dprintk_cont(" mismatch\n");
1865                nfs_mark_request_dirty(req);
1866                set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1867        next:
1868                nfs_unlock_and_release_request(req);
1869                /* Latency breaker */
1870                cond_resched();
1871        }
1872        nfss = NFS_SERVER(data->inode);
1873        if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1874                clear_bdi_congested(inode_to_bdi(data->inode), BLK_RW_ASYNC);
1875
1876        nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1877        nfs_commit_end(cinfo.mds);
1878}
1879
1880static void nfs_commit_release(void *calldata)
1881{
1882        struct nfs_commit_data *data = calldata;
1883
1884        data->completion_ops->completion(data);
1885        nfs_commitdata_release(calldata);
1886}
1887
1888static const struct rpc_call_ops nfs_commit_ops = {
1889        .rpc_call_prepare = nfs_commit_prepare,
1890        .rpc_call_done = nfs_commit_done,
1891        .rpc_release = nfs_commit_release,
1892};
1893
1894static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1895        .completion = nfs_commit_release_pages,
1896        .resched_write = nfs_commit_resched_write,
1897};
1898
1899int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1900                            int how, struct nfs_commit_info *cinfo)
1901{
1902        int status;
1903
1904        status = pnfs_commit_list(inode, head, how, cinfo);
1905        if (status == PNFS_NOT_ATTEMPTED)
1906                status = nfs_commit_list(inode, head, how, cinfo);
1907        return status;
1908}
1909
1910static int __nfs_commit_inode(struct inode *inode, int how,
1911                struct writeback_control *wbc)
1912{
1913        LIST_HEAD(head);
1914        struct nfs_commit_info cinfo;
1915        int may_wait = how & FLUSH_SYNC;
1916        int ret, nscan;
1917
1918        nfs_init_cinfo_from_inode(&cinfo, inode);
1919        nfs_commit_begin(cinfo.mds);
1920        for (;;) {
1921                ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1922                if (ret <= 0)
1923                        break;
1924                ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1925                if (ret < 0)
1926                        break;
1927                ret = 0;
1928                if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1929                        if (nscan < wbc->nr_to_write)
1930                                wbc->nr_to_write -= nscan;
1931                        else
1932                                wbc->nr_to_write = 0;
1933                }
1934                if (nscan < INT_MAX)
1935                        break;
1936                cond_resched();
1937        }
1938        nfs_commit_end(cinfo.mds);
1939        if (ret || !may_wait)
1940                return ret;
1941        return wait_on_commit(cinfo.mds);
1942}
1943
1944int nfs_commit_inode(struct inode *inode, int how)
1945{
1946        return __nfs_commit_inode(inode, how, NULL);
1947}
1948EXPORT_SYMBOL_GPL(nfs_commit_inode);
1949
1950int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1951{
1952        struct nfs_inode *nfsi = NFS_I(inode);
1953        int flags = FLUSH_SYNC;
1954        int ret = 0;
1955
1956        if (wbc->sync_mode == WB_SYNC_NONE) {
1957                /* no commits means nothing needs to be done */
1958                if (!atomic_long_read(&nfsi->commit_info.ncommit))
1959                        goto check_requests_outstanding;
1960
1961                /* Don't commit yet if this is a non-blocking flush and there
1962                 * are a lot of outstanding writes for this mapping.
1963                 */
1964                if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1965                        goto out_mark_dirty;
1966
1967                /* don't wait for the COMMIT response */
1968                flags = 0;
1969        }
1970
1971        ret = __nfs_commit_inode(inode, flags, wbc);
1972        if (!ret) {
1973                if (flags & FLUSH_SYNC)
1974                        return 0;
1975        } else if (atomic_long_read(&nfsi->commit_info.ncommit))
1976                goto out_mark_dirty;
1977
1978check_requests_outstanding:
1979        if (!atomic_read(&nfsi->commit_info.rpcs_out))
1980                return ret;
1981out_mark_dirty:
1982        __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1983        return ret;
1984}
1985EXPORT_SYMBOL_GPL(nfs_write_inode);
1986
1987/*
1988 * Wrapper for filemap_write_and_wait_range()
1989 *
1990 * Needed for pNFS in order to ensure data becomes visible to the
1991 * client.
1992 */
1993int nfs_filemap_write_and_wait_range(struct address_space *mapping,
1994                loff_t lstart, loff_t lend)
1995{
1996        int ret;
1997
1998        ret = filemap_write_and_wait_range(mapping, lstart, lend);
1999        if (ret == 0)
2000                ret = pnfs_sync_inode(mapping->host, true);
2001        return ret;
2002}
2003EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
2004
2005/*
2006 * flush the inode to disk.
2007 */
2008int nfs_wb_all(struct inode *inode)
2009{
2010        int ret;
2011
2012        trace_nfs_writeback_inode_enter(inode);
2013
2014        ret = filemap_write_and_wait(inode->i_mapping);
2015        if (ret)
2016                goto out;
2017        ret = nfs_commit_inode(inode, FLUSH_SYNC);
2018        if (ret < 0)
2019                goto out;
2020        pnfs_sync_inode(inode, true);
2021        ret = 0;
2022
2023out:
2024        trace_nfs_writeback_inode_exit(inode, ret);
2025        return ret;
2026}
2027EXPORT_SYMBOL_GPL(nfs_wb_all);
2028
2029int nfs_wb_page_cancel(struct inode *inode, struct page *page)
2030{
2031        struct nfs_page *req;
2032        int ret = 0;
2033
2034        wait_on_page_writeback(page);
2035
2036        /* blocking call to cancel all requests and join to a single (head)
2037         * request */
2038        req = nfs_lock_and_join_requests(page);
2039
2040        if (IS_ERR(req)) {
2041                ret = PTR_ERR(req);
2042        } else if (req) {
2043                /* all requests from this page have been cancelled by
2044                 * nfs_lock_and_join_requests, so just remove the head
2045                 * request from the inode / page_private pointer and
2046                 * release it */
2047                nfs_inode_remove_request(req);
2048                nfs_unlock_and_release_request(req);
2049        }
2050
2051        return ret;
2052}
2053
2054/*
2055 * Write back all requests on one page - we do this before reading it.
2056 */
2057int nfs_wb_page(struct inode *inode, struct page *page)
2058{
2059        loff_t range_start = page_file_offset(page);
2060        loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
2061        struct writeback_control wbc = {
2062                .sync_mode = WB_SYNC_ALL,
2063                .nr_to_write = 0,
2064                .range_start = range_start,
2065                .range_end = range_end,
2066        };
2067        int ret;
2068
2069        trace_nfs_writeback_page_enter(inode);
2070
2071        for (;;) {
2072                wait_on_page_writeback(page);
2073                if (clear_page_dirty_for_io(page)) {
2074                        ret = nfs_writepage_locked(page, &wbc);
2075                        if (ret < 0)
2076                                goto out_error;
2077                        continue;
2078                }
2079                ret = 0;
2080                if (!PagePrivate(page))
2081                        break;
2082                ret = nfs_commit_inode(inode, FLUSH_SYNC);
2083                if (ret < 0)
2084                        goto out_error;
2085        }
2086out_error:
2087        trace_nfs_writeback_page_exit(inode, ret);
2088        return ret;
2089}
2090
2091#ifdef CONFIG_MIGRATION
2092int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
2093                struct page *page, enum migrate_mode mode)
2094{
2095        /*
2096         * If PagePrivate is set, then the page is currently associated with
2097         * an in-progress read or write request. Don't try to migrate it.
2098         *
2099         * FIXME: we could do this in principle, but we'll need a way to ensure
2100         *        that we can safely release the inode reference while holding
2101         *        the page lock.
2102         */
2103        if (PagePrivate(page))
2104                return -EBUSY;
2105
2106        if (!nfs_fscache_release_page(page, GFP_KERNEL))
2107                return -EBUSY;
2108
2109        return migrate_page(mapping, newpage, page, mode);
2110}
2111#endif
2112
2113int __init nfs_init_writepagecache(void)
2114{
2115        nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2116                                             sizeof(struct nfs_pgio_header),
2117                                             0, SLAB_HWCACHE_ALIGN,
2118                                             NULL);
2119        if (nfs_wdata_cachep == NULL)
2120                return -ENOMEM;
2121
2122        nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2123                                                     nfs_wdata_cachep);
2124        if (nfs_wdata_mempool == NULL)
2125                goto out_destroy_write_cache;
2126
2127        nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2128                                             sizeof(struct nfs_commit_data),
2129                                             0, SLAB_HWCACHE_ALIGN,
2130                                             NULL);
2131        if (nfs_cdata_cachep == NULL)
2132                goto out_destroy_write_mempool;
2133
2134        nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2135                                                      nfs_cdata_cachep);
2136        if (nfs_commit_mempool == NULL)
2137                goto out_destroy_commit_cache;
2138
2139        /*
2140         * NFS congestion size, scale with available memory.
2141         *
2142         *  64MB:    8192k
2143         * 128MB:   11585k
2144         * 256MB:   16384k
2145         * 512MB:   23170k
2146         *   1GB:   32768k
2147         *   2GB:   46340k
2148         *   4GB:   65536k
2149         *   8GB:   92681k
2150         *  16GB:  131072k
2151         *
2152         * This allows larger machines to have larger/more transfers.
2153         * Limit the default to 256M
2154         */
2155        nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
2156        if (nfs_congestion_kb > 256*1024)
2157                nfs_congestion_kb = 256*1024;
2158
2159        return 0;
2160
2161out_destroy_commit_cache:
2162        kmem_cache_destroy(nfs_cdata_cachep);
2163out_destroy_write_mempool:
2164        mempool_destroy(nfs_wdata_mempool);
2165out_destroy_write_cache:
2166        kmem_cache_destroy(nfs_wdata_cachep);
2167        return -ENOMEM;
2168}
2169
2170void nfs_destroy_writepagecache(void)
2171{
2172        mempool_destroy(nfs_commit_mempool);
2173        kmem_cache_destroy(nfs_cdata_cachep);
2174        mempool_destroy(nfs_wdata_mempool);
2175        kmem_cache_destroy(nfs_wdata_cachep);
2176}
2177
2178static const struct nfs_rw_ops nfs_rw_write_ops = {
2179        .rw_alloc_header        = nfs_writehdr_alloc,
2180        .rw_free_header         = nfs_writehdr_free,
2181        .rw_done                = nfs_writeback_done,
2182        .rw_result              = nfs_writeback_result,
2183        .rw_initiate            = nfs_initiate_write,
2184};
2185