linux/fs/ceph/snap.c
<<
>>
Prefs
   1// SPDX-License-Identifier: GPL-2.0
   2#include <linux/ceph/ceph_debug.h>
   3
   4#include <linux/sort.h>
   5#include <linux/slab.h>
   6#include <linux/iversion.h>
   7#include "super.h"
   8#include "mds_client.h"
   9#include <linux/ceph/decode.h>
  10
  11/* unused map expires after 5 minutes */
  12#define CEPH_SNAPID_MAP_TIMEOUT (5 * 60 * HZ)
  13
  14/*
  15 * Snapshots in ceph are driven in large part by cooperation from the
  16 * client.  In contrast to local file systems or file servers that
  17 * implement snapshots at a single point in the system, ceph's
  18 * distributed access to storage requires clients to help decide
  19 * whether a write logically occurs before or after a recently created
  20 * snapshot.
  21 *
  22 * This provides a perfect instantanous client-wide snapshot.  Between
  23 * clients, however, snapshots may appear to be applied at slightly
  24 * different points in time, depending on delays in delivering the
  25 * snapshot notification.
  26 *
  27 * Snapshots are _not_ file system-wide.  Instead, each snapshot
  28 * applies to the subdirectory nested beneath some directory.  This
  29 * effectively divides the hierarchy into multiple "realms," where all
  30 * of the files contained by each realm share the same set of
  31 * snapshots.  An individual realm's snap set contains snapshots
  32 * explicitly created on that realm, as well as any snaps in its
  33 * parent's snap set _after_ the point at which the parent became it's
  34 * parent (due to, say, a rename).  Similarly, snaps from prior parents
  35 * during the time intervals during which they were the parent are included.
  36 *
  37 * The client is spared most of this detail, fortunately... it must only
  38 * maintains a hierarchy of realms reflecting the current parent/child
  39 * realm relationship, and for each realm has an explicit list of snaps
  40 * inherited from prior parents.
  41 *
  42 * A snap_realm struct is maintained for realms containing every inode
  43 * with an open cap in the system.  (The needed snap realm information is
  44 * provided by the MDS whenever a cap is issued, i.e., on open.)  A 'seq'
  45 * version number is used to ensure that as realm parameters change (new
  46 * snapshot, new parent, etc.) the client's realm hierarchy is updated.
  47 *
  48 * The realm hierarchy drives the generation of a 'snap context' for each
  49 * realm, which simply lists the resulting set of snaps for the realm.  This
  50 * is attached to any writes sent to OSDs.
  51 */
  52/*
  53 * Unfortunately error handling is a bit mixed here.  If we get a snap
  54 * update, but don't have enough memory to update our realm hierarchy,
  55 * it's not clear what we can do about it (besides complaining to the
  56 * console).
  57 */
  58
  59
  60/*
  61 * increase ref count for the realm
  62 *
  63 * caller must hold snap_rwsem for write.
  64 */
  65void ceph_get_snap_realm(struct ceph_mds_client *mdsc,
  66                         struct ceph_snap_realm *realm)
  67{
  68        dout("get_realm %p %d -> %d\n", realm,
  69             atomic_read(&realm->nref), atomic_read(&realm->nref)+1);
  70        /*
  71         * since we _only_ increment realm refs or empty the empty
  72         * list with snap_rwsem held, adjusting the empty list here is
  73         * safe.  we do need to protect against concurrent empty list
  74         * additions, however.
  75         */
  76        if (atomic_inc_return(&realm->nref) == 1) {
  77                spin_lock(&mdsc->snap_empty_lock);
  78                list_del_init(&realm->empty_item);
  79                spin_unlock(&mdsc->snap_empty_lock);
  80        }
  81}
  82
  83static void __insert_snap_realm(struct rb_root *root,
  84                                struct ceph_snap_realm *new)
  85{
  86        struct rb_node **p = &root->rb_node;
  87        struct rb_node *parent = NULL;
  88        struct ceph_snap_realm *r = NULL;
  89
  90        while (*p) {
  91                parent = *p;
  92                r = rb_entry(parent, struct ceph_snap_realm, node);
  93                if (new->ino < r->ino)
  94                        p = &(*p)->rb_left;
  95                else if (new->ino > r->ino)
  96                        p = &(*p)->rb_right;
  97                else
  98                        BUG();
  99        }
 100
 101        rb_link_node(&new->node, parent, p);
 102        rb_insert_color(&new->node, root);
 103}
 104
 105/*
 106 * create and get the realm rooted at @ino and bump its ref count.
 107 *
 108 * caller must hold snap_rwsem for write.
 109 */
 110static struct ceph_snap_realm *ceph_create_snap_realm(
 111        struct ceph_mds_client *mdsc,
 112        u64 ino)
 113{
 114        struct ceph_snap_realm *realm;
 115
 116        realm = kzalloc(sizeof(*realm), GFP_NOFS);
 117        if (!realm)
 118                return ERR_PTR(-ENOMEM);
 119
 120        atomic_set(&realm->nref, 1);    /* for caller */
 121        realm->ino = ino;
 122        INIT_LIST_HEAD(&realm->children);
 123        INIT_LIST_HEAD(&realm->child_item);
 124        INIT_LIST_HEAD(&realm->empty_item);
 125        INIT_LIST_HEAD(&realm->dirty_item);
 126        INIT_LIST_HEAD(&realm->inodes_with_caps);
 127        spin_lock_init(&realm->inodes_with_caps_lock);
 128        __insert_snap_realm(&mdsc->snap_realms, realm);
 129        mdsc->num_snap_realms++;
 130
 131        dout("create_snap_realm %llx %p\n", realm->ino, realm);
 132        return realm;
 133}
 134
 135/*
 136 * lookup the realm rooted at @ino.
 137 *
 138 * caller must hold snap_rwsem for write.
 139 */
 140static struct ceph_snap_realm *__lookup_snap_realm(struct ceph_mds_client *mdsc,
 141                                                   u64 ino)
 142{
 143        struct rb_node *n = mdsc->snap_realms.rb_node;
 144        struct ceph_snap_realm *r;
 145
 146        while (n) {
 147                r = rb_entry(n, struct ceph_snap_realm, node);
 148                if (ino < r->ino)
 149                        n = n->rb_left;
 150                else if (ino > r->ino)
 151                        n = n->rb_right;
 152                else {
 153                        dout("lookup_snap_realm %llx %p\n", r->ino, r);
 154                        return r;
 155                }
 156        }
 157        return NULL;
 158}
 159
 160struct ceph_snap_realm *ceph_lookup_snap_realm(struct ceph_mds_client *mdsc,
 161                                               u64 ino)
 162{
 163        struct ceph_snap_realm *r;
 164        r = __lookup_snap_realm(mdsc, ino);
 165        if (r)
 166                ceph_get_snap_realm(mdsc, r);
 167        return r;
 168}
 169
 170static void __put_snap_realm(struct ceph_mds_client *mdsc,
 171                             struct ceph_snap_realm *realm);
 172
 173/*
 174 * called with snap_rwsem (write)
 175 */
 176static void __destroy_snap_realm(struct ceph_mds_client *mdsc,
 177                                 struct ceph_snap_realm *realm)
 178{
 179        dout("__destroy_snap_realm %p %llx\n", realm, realm->ino);
 180
 181        rb_erase(&realm->node, &mdsc->snap_realms);
 182        mdsc->num_snap_realms--;
 183
 184        if (realm->parent) {
 185                list_del_init(&realm->child_item);
 186                __put_snap_realm(mdsc, realm->parent);
 187        }
 188
 189        kfree(realm->prior_parent_snaps);
 190        kfree(realm->snaps);
 191        ceph_put_snap_context(realm->cached_context);
 192        kfree(realm);
 193}
 194
 195/*
 196 * caller holds snap_rwsem (write)
 197 */
 198static void __put_snap_realm(struct ceph_mds_client *mdsc,
 199                             struct ceph_snap_realm *realm)
 200{
 201        dout("__put_snap_realm %llx %p %d -> %d\n", realm->ino, realm,
 202             atomic_read(&realm->nref), atomic_read(&realm->nref)-1);
 203        if (atomic_dec_and_test(&realm->nref))
 204                __destroy_snap_realm(mdsc, realm);
 205}
 206
 207/*
 208 * caller needn't hold any locks
 209 */
 210void ceph_put_snap_realm(struct ceph_mds_client *mdsc,
 211                         struct ceph_snap_realm *realm)
 212{
 213        dout("put_snap_realm %llx %p %d -> %d\n", realm->ino, realm,
 214             atomic_read(&realm->nref), atomic_read(&realm->nref)-1);
 215        if (!atomic_dec_and_test(&realm->nref))
 216                return;
 217
 218        if (down_write_trylock(&mdsc->snap_rwsem)) {
 219                __destroy_snap_realm(mdsc, realm);
 220                up_write(&mdsc->snap_rwsem);
 221        } else {
 222                spin_lock(&mdsc->snap_empty_lock);
 223                list_add(&realm->empty_item, &mdsc->snap_empty);
 224                spin_unlock(&mdsc->snap_empty_lock);
 225        }
 226}
 227
 228/*
 229 * Clean up any realms whose ref counts have dropped to zero.  Note
 230 * that this does not include realms who were created but not yet
 231 * used.
 232 *
 233 * Called under snap_rwsem (write)
 234 */
 235static void __cleanup_empty_realms(struct ceph_mds_client *mdsc)
 236{
 237        struct ceph_snap_realm *realm;
 238
 239        spin_lock(&mdsc->snap_empty_lock);
 240        while (!list_empty(&mdsc->snap_empty)) {
 241                realm = list_first_entry(&mdsc->snap_empty,
 242                                   struct ceph_snap_realm, empty_item);
 243                list_del(&realm->empty_item);
 244                spin_unlock(&mdsc->snap_empty_lock);
 245                __destroy_snap_realm(mdsc, realm);
 246                spin_lock(&mdsc->snap_empty_lock);
 247        }
 248        spin_unlock(&mdsc->snap_empty_lock);
 249}
 250
 251void ceph_cleanup_empty_realms(struct ceph_mds_client *mdsc)
 252{
 253        down_write(&mdsc->snap_rwsem);
 254        __cleanup_empty_realms(mdsc);
 255        up_write(&mdsc->snap_rwsem);
 256}
 257
 258/*
 259 * adjust the parent realm of a given @realm.  adjust child list, and parent
 260 * pointers, and ref counts appropriately.
 261 *
 262 * return true if parent was changed, 0 if unchanged, <0 on error.
 263 *
 264 * caller must hold snap_rwsem for write.
 265 */
 266static int adjust_snap_realm_parent(struct ceph_mds_client *mdsc,
 267                                    struct ceph_snap_realm *realm,
 268                                    u64 parentino)
 269{
 270        struct ceph_snap_realm *parent;
 271
 272        if (realm->parent_ino == parentino)
 273                return 0;
 274
 275        parent = ceph_lookup_snap_realm(mdsc, parentino);
 276        if (!parent) {
 277                parent = ceph_create_snap_realm(mdsc, parentino);
 278                if (IS_ERR(parent))
 279                        return PTR_ERR(parent);
 280        }
 281        dout("adjust_snap_realm_parent %llx %p: %llx %p -> %llx %p\n",
 282             realm->ino, realm, realm->parent_ino, realm->parent,
 283             parentino, parent);
 284        if (realm->parent) {
 285                list_del_init(&realm->child_item);
 286                ceph_put_snap_realm(mdsc, realm->parent);
 287        }
 288        realm->parent_ino = parentino;
 289        realm->parent = parent;
 290        list_add(&realm->child_item, &parent->children);
 291        return 1;
 292}
 293
 294
 295static int cmpu64_rev(const void *a, const void *b)
 296{
 297        if (*(u64 *)a < *(u64 *)b)
 298                return 1;
 299        if (*(u64 *)a > *(u64 *)b)
 300                return -1;
 301        return 0;
 302}
 303
 304
 305/*
 306 * build the snap context for a given realm.
 307 */
 308static int build_snap_context(struct ceph_snap_realm *realm,
 309                              struct list_head* dirty_realms)
 310{
 311        struct ceph_snap_realm *parent = realm->parent;
 312        struct ceph_snap_context *snapc;
 313        int err = 0;
 314        u32 num = realm->num_prior_parent_snaps + realm->num_snaps;
 315
 316        /*
 317         * build parent context, if it hasn't been built.
 318         * conservatively estimate that all parent snaps might be
 319         * included by us.
 320         */
 321        if (parent) {
 322                if (!parent->cached_context) {
 323                        err = build_snap_context(parent, dirty_realms);
 324                        if (err)
 325                                goto fail;
 326                }
 327                num += parent->cached_context->num_snaps;
 328        }
 329
 330        /* do i actually need to update?  not if my context seq
 331           matches realm seq, and my parents' does to.  (this works
 332           because we rebuild_snap_realms() works _downward_ in
 333           hierarchy after each update.) */
 334        if (realm->cached_context &&
 335            realm->cached_context->seq == realm->seq &&
 336            (!parent ||
 337             realm->cached_context->seq >= parent->cached_context->seq)) {
 338                dout("build_snap_context %llx %p: %p seq %lld (%u snaps)"
 339                     " (unchanged)\n",
 340                     realm->ino, realm, realm->cached_context,
 341                     realm->cached_context->seq,
 342                     (unsigned int)realm->cached_context->num_snaps);
 343                return 0;
 344        }
 345
 346        /* alloc new snap context */
 347        err = -ENOMEM;
 348        if (num > (SIZE_MAX - sizeof(*snapc)) / sizeof(u64))
 349                goto fail;
 350        snapc = ceph_create_snap_context(num, GFP_NOFS);
 351        if (!snapc)
 352                goto fail;
 353
 354        /* build (reverse sorted) snap vector */
 355        num = 0;
 356        snapc->seq = realm->seq;
 357        if (parent) {
 358                u32 i;
 359
 360                /* include any of parent's snaps occurring _after_ my
 361                   parent became my parent */
 362                for (i = 0; i < parent->cached_context->num_snaps; i++)
 363                        if (parent->cached_context->snaps[i] >=
 364                            realm->parent_since)
 365                                snapc->snaps[num++] =
 366                                        parent->cached_context->snaps[i];
 367                if (parent->cached_context->seq > snapc->seq)
 368                        snapc->seq = parent->cached_context->seq;
 369        }
 370        memcpy(snapc->snaps + num, realm->snaps,
 371               sizeof(u64)*realm->num_snaps);
 372        num += realm->num_snaps;
 373        memcpy(snapc->snaps + num, realm->prior_parent_snaps,
 374               sizeof(u64)*realm->num_prior_parent_snaps);
 375        num += realm->num_prior_parent_snaps;
 376
 377        sort(snapc->snaps, num, sizeof(u64), cmpu64_rev, NULL);
 378        snapc->num_snaps = num;
 379        dout("build_snap_context %llx %p: %p seq %lld (%u snaps)\n",
 380             realm->ino, realm, snapc, snapc->seq,
 381             (unsigned int) snapc->num_snaps);
 382
 383        ceph_put_snap_context(realm->cached_context);
 384        realm->cached_context = snapc;
 385        /* queue realm for cap_snap creation */
 386        list_add_tail(&realm->dirty_item, dirty_realms);
 387        return 0;
 388
 389fail:
 390        /*
 391         * if we fail, clear old (incorrect) cached_context... hopefully
 392         * we'll have better luck building it later
 393         */
 394        if (realm->cached_context) {
 395                ceph_put_snap_context(realm->cached_context);
 396                realm->cached_context = NULL;
 397        }
 398        pr_err("build_snap_context %llx %p fail %d\n", realm->ino,
 399               realm, err);
 400        return err;
 401}
 402
 403/*
 404 * rebuild snap context for the given realm and all of its children.
 405 */
 406static void rebuild_snap_realms(struct ceph_snap_realm *realm,
 407                                struct list_head *dirty_realms)
 408{
 409        struct ceph_snap_realm *child;
 410
 411        dout("rebuild_snap_realms %llx %p\n", realm->ino, realm);
 412        build_snap_context(realm, dirty_realms);
 413
 414        list_for_each_entry(child, &realm->children, child_item)
 415                rebuild_snap_realms(child, dirty_realms);
 416}
 417
 418
 419/*
 420 * helper to allocate and decode an array of snapids.  free prior
 421 * instance, if any.
 422 */
 423static int dup_array(u64 **dst, __le64 *src, u32 num)
 424{
 425        u32 i;
 426
 427        kfree(*dst);
 428        if (num) {
 429                *dst = kcalloc(num, sizeof(u64), GFP_NOFS);
 430                if (!*dst)
 431                        return -ENOMEM;
 432                for (i = 0; i < num; i++)
 433                        (*dst)[i] = get_unaligned_le64(src + i);
 434        } else {
 435                *dst = NULL;
 436        }
 437        return 0;
 438}
 439
 440static bool has_new_snaps(struct ceph_snap_context *o,
 441                          struct ceph_snap_context *n)
 442{
 443        if (n->num_snaps == 0)
 444                return false;
 445        /* snaps are in descending order */
 446        return n->snaps[0] > o->seq;
 447}
 448
 449/*
 450 * When a snapshot is applied, the size/mtime inode metadata is queued
 451 * in a ceph_cap_snap (one for each snapshot) until writeback
 452 * completes and the metadata can be flushed back to the MDS.
 453 *
 454 * However, if a (sync) write is currently in-progress when we apply
 455 * the snapshot, we have to wait until the write succeeds or fails
 456 * (and a final size/mtime is known).  In this case the
 457 * cap_snap->writing = 1, and is said to be "pending."  When the write
 458 * finishes, we __ceph_finish_cap_snap().
 459 *
 460 * Caller must hold snap_rwsem for read (i.e., the realm topology won't
 461 * change).
 462 */
 463void ceph_queue_cap_snap(struct ceph_inode_info *ci)
 464{
 465        struct inode *inode = &ci->vfs_inode;
 466        struct ceph_cap_snap *capsnap;
 467        struct ceph_snap_context *old_snapc, *new_snapc;
 468        struct ceph_buffer *old_blob = NULL;
 469        int used, dirty;
 470
 471        capsnap = kzalloc(sizeof(*capsnap), GFP_NOFS);
 472        if (!capsnap) {
 473                pr_err("ENOMEM allocating ceph_cap_snap on %p\n", inode);
 474                return;
 475        }
 476
 477        spin_lock(&ci->i_ceph_lock);
 478        used = __ceph_caps_used(ci);
 479        dirty = __ceph_caps_dirty(ci);
 480
 481        old_snapc = ci->i_head_snapc;
 482        new_snapc = ci->i_snap_realm->cached_context;
 483
 484        /*
 485         * If there is a write in progress, treat that as a dirty Fw,
 486         * even though it hasn't completed yet; by the time we finish
 487         * up this capsnap it will be.
 488         */
 489        if (used & CEPH_CAP_FILE_WR)
 490                dirty |= CEPH_CAP_FILE_WR;
 491
 492        if (__ceph_have_pending_cap_snap(ci)) {
 493                /* there is no point in queuing multiple "pending" cap_snaps,
 494                   as no new writes are allowed to start when pending, so any
 495                   writes in progress now were started before the previous
 496                   cap_snap.  lucky us. */
 497                dout("queue_cap_snap %p already pending\n", inode);
 498                goto update_snapc;
 499        }
 500        if (ci->i_wrbuffer_ref_head == 0 &&
 501            !(dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))) {
 502                dout("queue_cap_snap %p nothing dirty|writing\n", inode);
 503                goto update_snapc;
 504        }
 505
 506        BUG_ON(!old_snapc);
 507
 508        /*
 509         * There is no need to send FLUSHSNAP message to MDS if there is
 510         * no new snapshot. But when there is dirty pages or on-going
 511         * writes, we still need to create cap_snap. cap_snap is needed
 512         * by the write path and page writeback path.
 513         *
 514         * also see ceph_try_drop_cap_snap()
 515         */
 516        if (has_new_snaps(old_snapc, new_snapc)) {
 517                if (dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))
 518                        capsnap->need_flush = true;
 519        } else {
 520                if (!(used & CEPH_CAP_FILE_WR) &&
 521                    ci->i_wrbuffer_ref_head == 0) {
 522                        dout("queue_cap_snap %p "
 523                             "no new_snap|dirty_page|writing\n", inode);
 524                        goto update_snapc;
 525                }
 526        }
 527
 528        dout("queue_cap_snap %p cap_snap %p queuing under %p %s %s\n",
 529             inode, capsnap, old_snapc, ceph_cap_string(dirty),
 530             capsnap->need_flush ? "" : "no_flush");
 531        ihold(inode);
 532
 533        refcount_set(&capsnap->nref, 1);
 534        INIT_LIST_HEAD(&capsnap->ci_item);
 535
 536        capsnap->follows = old_snapc->seq;
 537        capsnap->issued = __ceph_caps_issued(ci, NULL);
 538        capsnap->dirty = dirty;
 539
 540        capsnap->mode = inode->i_mode;
 541        capsnap->uid = inode->i_uid;
 542        capsnap->gid = inode->i_gid;
 543
 544        if (dirty & CEPH_CAP_XATTR_EXCL) {
 545                old_blob = __ceph_build_xattrs_blob(ci);
 546                capsnap->xattr_blob =
 547                        ceph_buffer_get(ci->i_xattrs.blob);
 548                capsnap->xattr_version = ci->i_xattrs.version;
 549        } else {
 550                capsnap->xattr_blob = NULL;
 551                capsnap->xattr_version = 0;
 552        }
 553
 554        capsnap->inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
 555
 556        /* dirty page count moved from _head to this cap_snap;
 557           all subsequent writes page dirties occur _after_ this
 558           snapshot. */
 559        capsnap->dirty_pages = ci->i_wrbuffer_ref_head;
 560        ci->i_wrbuffer_ref_head = 0;
 561        capsnap->context = old_snapc;
 562        list_add_tail(&capsnap->ci_item, &ci->i_cap_snaps);
 563
 564        if (used & CEPH_CAP_FILE_WR) {
 565                dout("queue_cap_snap %p cap_snap %p snapc %p"
 566                     " seq %llu used WR, now pending\n", inode,
 567                     capsnap, old_snapc, old_snapc->seq);
 568                capsnap->writing = 1;
 569        } else {
 570                /* note mtime, size NOW. */
 571                __ceph_finish_cap_snap(ci, capsnap);
 572        }
 573        capsnap = NULL;
 574        old_snapc = NULL;
 575
 576update_snapc:
 577       if (ci->i_wrbuffer_ref_head == 0 &&
 578           ci->i_wr_ref == 0 &&
 579           ci->i_dirty_caps == 0 &&
 580           ci->i_flushing_caps == 0) {
 581               ci->i_head_snapc = NULL;
 582       } else {
 583                ci->i_head_snapc = ceph_get_snap_context(new_snapc);
 584                dout(" new snapc is %p\n", new_snapc);
 585        }
 586        spin_unlock(&ci->i_ceph_lock);
 587
 588        ceph_buffer_put(old_blob);
 589        kfree(capsnap);
 590        ceph_put_snap_context(old_snapc);
 591}
 592
 593/*
 594 * Finalize the size, mtime for a cap_snap.. that is, settle on final values
 595 * to be used for the snapshot, to be flushed back to the mds.
 596 *
 597 * If capsnap can now be flushed, add to snap_flush list, and return 1.
 598 *
 599 * Caller must hold i_ceph_lock.
 600 */
 601int __ceph_finish_cap_snap(struct ceph_inode_info *ci,
 602                            struct ceph_cap_snap *capsnap)
 603{
 604        struct inode *inode = &ci->vfs_inode;
 605        struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
 606
 607        BUG_ON(capsnap->writing);
 608        capsnap->size = inode->i_size;
 609        capsnap->mtime = inode->i_mtime;
 610        capsnap->atime = inode->i_atime;
 611        capsnap->ctime = inode->i_ctime;
 612        capsnap->btime = ci->i_btime;
 613        capsnap->change_attr = inode_peek_iversion_raw(inode);
 614        capsnap->time_warp_seq = ci->i_time_warp_seq;
 615        capsnap->truncate_size = ci->i_truncate_size;
 616        capsnap->truncate_seq = ci->i_truncate_seq;
 617        if (capsnap->dirty_pages) {
 618                dout("finish_cap_snap %p cap_snap %p snapc %p %llu %s s=%llu "
 619                     "still has %d dirty pages\n", inode, capsnap,
 620                     capsnap->context, capsnap->context->seq,
 621                     ceph_cap_string(capsnap->dirty), capsnap->size,
 622                     capsnap->dirty_pages);
 623                return 0;
 624        }
 625
 626        ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
 627        dout("finish_cap_snap %p cap_snap %p snapc %p %llu %s s=%llu\n",
 628             inode, capsnap, capsnap->context,
 629             capsnap->context->seq, ceph_cap_string(capsnap->dirty),
 630             capsnap->size);
 631
 632        spin_lock(&mdsc->snap_flush_lock);
 633        if (list_empty(&ci->i_snap_flush_item))
 634                list_add_tail(&ci->i_snap_flush_item, &mdsc->snap_flush_list);
 635        spin_unlock(&mdsc->snap_flush_lock);
 636        return 1;  /* caller may want to ceph_flush_snaps */
 637}
 638
 639/*
 640 * Queue cap_snaps for snap writeback for this realm and its children.
 641 * Called under snap_rwsem, so realm topology won't change.
 642 */
 643static void queue_realm_cap_snaps(struct ceph_snap_realm *realm)
 644{
 645        struct ceph_inode_info *ci;
 646        struct inode *lastinode = NULL;
 647
 648        dout("queue_realm_cap_snaps %p %llx inodes\n", realm, realm->ino);
 649
 650        spin_lock(&realm->inodes_with_caps_lock);
 651        list_for_each_entry(ci, &realm->inodes_with_caps, i_snap_realm_item) {
 652                struct inode *inode = igrab(&ci->vfs_inode);
 653                if (!inode)
 654                        continue;
 655                spin_unlock(&realm->inodes_with_caps_lock);
 656                /* avoid calling iput_final() while holding
 657                 * mdsc->snap_rwsem or in mds dispatch threads */
 658                ceph_async_iput(lastinode);
 659                lastinode = inode;
 660                ceph_queue_cap_snap(ci);
 661                spin_lock(&realm->inodes_with_caps_lock);
 662        }
 663        spin_unlock(&realm->inodes_with_caps_lock);
 664        ceph_async_iput(lastinode);
 665
 666        dout("queue_realm_cap_snaps %p %llx done\n", realm, realm->ino);
 667}
 668
 669/*
 670 * Parse and apply a snapblob "snap trace" from the MDS.  This specifies
 671 * the snap realm parameters from a given realm and all of its ancestors,
 672 * up to the root.
 673 *
 674 * Caller must hold snap_rwsem for write.
 675 */
 676int ceph_update_snap_trace(struct ceph_mds_client *mdsc,
 677                           void *p, void *e, bool deletion,
 678                           struct ceph_snap_realm **realm_ret)
 679{
 680        struct ceph_mds_snap_realm *ri;    /* encoded */
 681        __le64 *snaps;                     /* encoded */
 682        __le64 *prior_parent_snaps;        /* encoded */
 683        struct ceph_snap_realm *realm = NULL;
 684        struct ceph_snap_realm *first_realm = NULL;
 685        int invalidate = 0;
 686        int err = -ENOMEM;
 687        LIST_HEAD(dirty_realms);
 688
 689        dout("update_snap_trace deletion=%d\n", deletion);
 690more:
 691        ceph_decode_need(&p, e, sizeof(*ri), bad);
 692        ri = p;
 693        p += sizeof(*ri);
 694        ceph_decode_need(&p, e, sizeof(u64)*(le32_to_cpu(ri->num_snaps) +
 695                            le32_to_cpu(ri->num_prior_parent_snaps)), bad);
 696        snaps = p;
 697        p += sizeof(u64) * le32_to_cpu(ri->num_snaps);
 698        prior_parent_snaps = p;
 699        p += sizeof(u64) * le32_to_cpu(ri->num_prior_parent_snaps);
 700
 701        realm = ceph_lookup_snap_realm(mdsc, le64_to_cpu(ri->ino));
 702        if (!realm) {
 703                realm = ceph_create_snap_realm(mdsc, le64_to_cpu(ri->ino));
 704                if (IS_ERR(realm)) {
 705                        err = PTR_ERR(realm);
 706                        goto fail;
 707                }
 708        }
 709
 710        /* ensure the parent is correct */
 711        err = adjust_snap_realm_parent(mdsc, realm, le64_to_cpu(ri->parent));
 712        if (err < 0)
 713                goto fail;
 714        invalidate += err;
 715
 716        if (le64_to_cpu(ri->seq) > realm->seq) {
 717                dout("update_snap_trace updating %llx %p %lld -> %lld\n",
 718                     realm->ino, realm, realm->seq, le64_to_cpu(ri->seq));
 719                /* update realm parameters, snap lists */
 720                realm->seq = le64_to_cpu(ri->seq);
 721                realm->created = le64_to_cpu(ri->created);
 722                realm->parent_since = le64_to_cpu(ri->parent_since);
 723
 724                realm->num_snaps = le32_to_cpu(ri->num_snaps);
 725                err = dup_array(&realm->snaps, snaps, realm->num_snaps);
 726                if (err < 0)
 727                        goto fail;
 728
 729                realm->num_prior_parent_snaps =
 730                        le32_to_cpu(ri->num_prior_parent_snaps);
 731                err = dup_array(&realm->prior_parent_snaps, prior_parent_snaps,
 732                                realm->num_prior_parent_snaps);
 733                if (err < 0)
 734                        goto fail;
 735
 736                if (realm->seq > mdsc->last_snap_seq)
 737                        mdsc->last_snap_seq = realm->seq;
 738
 739                invalidate = 1;
 740        } else if (!realm->cached_context) {
 741                dout("update_snap_trace %llx %p seq %lld new\n",
 742                     realm->ino, realm, realm->seq);
 743                invalidate = 1;
 744        } else {
 745                dout("update_snap_trace %llx %p seq %lld unchanged\n",
 746                     realm->ino, realm, realm->seq);
 747        }
 748
 749        dout("done with %llx %p, invalidated=%d, %p %p\n", realm->ino,
 750             realm, invalidate, p, e);
 751
 752        /* invalidate when we reach the _end_ (root) of the trace */
 753        if (invalidate && p >= e)
 754                rebuild_snap_realms(realm, &dirty_realms);
 755
 756        if (!first_realm)
 757                first_realm = realm;
 758        else
 759                ceph_put_snap_realm(mdsc, realm);
 760
 761        if (p < e)
 762                goto more;
 763
 764        /*
 765         * queue cap snaps _after_ we've built the new snap contexts,
 766         * so that i_head_snapc can be set appropriately.
 767         */
 768        while (!list_empty(&dirty_realms)) {
 769                realm = list_first_entry(&dirty_realms, struct ceph_snap_realm,
 770                                         dirty_item);
 771                list_del_init(&realm->dirty_item);
 772                queue_realm_cap_snaps(realm);
 773        }
 774
 775        if (realm_ret)
 776                *realm_ret = first_realm;
 777        else
 778                ceph_put_snap_realm(mdsc, first_realm);
 779
 780        __cleanup_empty_realms(mdsc);
 781        return 0;
 782
 783bad:
 784        err = -EINVAL;
 785fail:
 786        if (realm && !IS_ERR(realm))
 787                ceph_put_snap_realm(mdsc, realm);
 788        if (first_realm)
 789                ceph_put_snap_realm(mdsc, first_realm);
 790        pr_err("update_snap_trace error %d\n", err);
 791        return err;
 792}
 793
 794
 795/*
 796 * Send any cap_snaps that are queued for flush.  Try to carry
 797 * s_mutex across multiple snap flushes to avoid locking overhead.
 798 *
 799 * Caller holds no locks.
 800 */
 801static void flush_snaps(struct ceph_mds_client *mdsc)
 802{
 803        struct ceph_inode_info *ci;
 804        struct inode *inode;
 805        struct ceph_mds_session *session = NULL;
 806
 807        dout("flush_snaps\n");
 808        spin_lock(&mdsc->snap_flush_lock);
 809        while (!list_empty(&mdsc->snap_flush_list)) {
 810                ci = list_first_entry(&mdsc->snap_flush_list,
 811                                struct ceph_inode_info, i_snap_flush_item);
 812                inode = &ci->vfs_inode;
 813                ihold(inode);
 814                spin_unlock(&mdsc->snap_flush_lock);
 815                ceph_flush_snaps(ci, &session);
 816                /* avoid calling iput_final() while holding
 817                 * session->s_mutex or in mds dispatch threads */
 818                ceph_async_iput(inode);
 819                spin_lock(&mdsc->snap_flush_lock);
 820        }
 821        spin_unlock(&mdsc->snap_flush_lock);
 822
 823        if (session) {
 824                mutex_unlock(&session->s_mutex);
 825                ceph_put_mds_session(session);
 826        }
 827        dout("flush_snaps done\n");
 828}
 829
 830
 831/*
 832 * Handle a snap notification from the MDS.
 833 *
 834 * This can take two basic forms: the simplest is just a snap creation
 835 * or deletion notification on an existing realm.  This should update the
 836 * realm and its children.
 837 *
 838 * The more difficult case is realm creation, due to snap creation at a
 839 * new point in the file hierarchy, or due to a rename that moves a file or
 840 * directory into another realm.
 841 */
 842void ceph_handle_snap(struct ceph_mds_client *mdsc,
 843                      struct ceph_mds_session *session,
 844                      struct ceph_msg *msg)
 845{
 846        struct super_block *sb = mdsc->fsc->sb;
 847        int mds = session->s_mds;
 848        u64 split;
 849        int op;
 850        int trace_len;
 851        struct ceph_snap_realm *realm = NULL;
 852        void *p = msg->front.iov_base;
 853        void *e = p + msg->front.iov_len;
 854        struct ceph_mds_snap_head *h;
 855        int num_split_inos, num_split_realms;
 856        __le64 *split_inos = NULL, *split_realms = NULL;
 857        int i;
 858        int locked_rwsem = 0;
 859
 860        /* decode */
 861        if (msg->front.iov_len < sizeof(*h))
 862                goto bad;
 863        h = p;
 864        op = le32_to_cpu(h->op);
 865        split = le64_to_cpu(h->split);   /* non-zero if we are splitting an
 866                                          * existing realm */
 867        num_split_inos = le32_to_cpu(h->num_split_inos);
 868        num_split_realms = le32_to_cpu(h->num_split_realms);
 869        trace_len = le32_to_cpu(h->trace_len);
 870        p += sizeof(*h);
 871
 872        dout("handle_snap from mds%d op %s split %llx tracelen %d\n", mds,
 873             ceph_snap_op_name(op), split, trace_len);
 874
 875        mutex_lock(&session->s_mutex);
 876        session->s_seq++;
 877        mutex_unlock(&session->s_mutex);
 878
 879        down_write(&mdsc->snap_rwsem);
 880        locked_rwsem = 1;
 881
 882        if (op == CEPH_SNAP_OP_SPLIT) {
 883                struct ceph_mds_snap_realm *ri;
 884
 885                /*
 886                 * A "split" breaks part of an existing realm off into
 887                 * a new realm.  The MDS provides a list of inodes
 888                 * (with caps) and child realms that belong to the new
 889                 * child.
 890                 */
 891                split_inos = p;
 892                p += sizeof(u64) * num_split_inos;
 893                split_realms = p;
 894                p += sizeof(u64) * num_split_realms;
 895                ceph_decode_need(&p, e, sizeof(*ri), bad);
 896                /* we will peek at realm info here, but will _not_
 897                 * advance p, as the realm update will occur below in
 898                 * ceph_update_snap_trace. */
 899                ri = p;
 900
 901                realm = ceph_lookup_snap_realm(mdsc, split);
 902                if (!realm) {
 903                        realm = ceph_create_snap_realm(mdsc, split);
 904                        if (IS_ERR(realm))
 905                                goto out;
 906                }
 907
 908                dout("splitting snap_realm %llx %p\n", realm->ino, realm);
 909                for (i = 0; i < num_split_inos; i++) {
 910                        struct ceph_vino vino = {
 911                                .ino = le64_to_cpu(split_inos[i]),
 912                                .snap = CEPH_NOSNAP,
 913                        };
 914                        struct inode *inode = ceph_find_inode(sb, vino);
 915                        struct ceph_inode_info *ci;
 916                        struct ceph_snap_realm *oldrealm;
 917
 918                        if (!inode)
 919                                continue;
 920                        ci = ceph_inode(inode);
 921
 922                        spin_lock(&ci->i_ceph_lock);
 923                        if (!ci->i_snap_realm)
 924                                goto skip_inode;
 925                        /*
 926                         * If this inode belongs to a realm that was
 927                         * created after our new realm, we experienced
 928                         * a race (due to another split notifications
 929                         * arriving from a different MDS).  So skip
 930                         * this inode.
 931                         */
 932                        if (ci->i_snap_realm->created >
 933                            le64_to_cpu(ri->created)) {
 934                                dout(" leaving %p in newer realm %llx %p\n",
 935                                     inode, ci->i_snap_realm->ino,
 936                                     ci->i_snap_realm);
 937                                goto skip_inode;
 938                        }
 939                        dout(" will move %p to split realm %llx %p\n",
 940                             inode, realm->ino, realm);
 941                        /*
 942                         * Move the inode to the new realm
 943                         */
 944                        oldrealm = ci->i_snap_realm;
 945                        spin_lock(&oldrealm->inodes_with_caps_lock);
 946                        list_del_init(&ci->i_snap_realm_item);
 947                        spin_unlock(&oldrealm->inodes_with_caps_lock);
 948
 949                        spin_lock(&realm->inodes_with_caps_lock);
 950                        list_add(&ci->i_snap_realm_item,
 951                                 &realm->inodes_with_caps);
 952                        ci->i_snap_realm = realm;
 953                        if (realm->ino == ci->i_vino.ino)
 954                                realm->inode = inode;
 955                        spin_unlock(&realm->inodes_with_caps_lock);
 956
 957                        spin_unlock(&ci->i_ceph_lock);
 958
 959                        ceph_get_snap_realm(mdsc, realm);
 960                        ceph_put_snap_realm(mdsc, oldrealm);
 961
 962                        /* avoid calling iput_final() while holding
 963                         * mdsc->snap_rwsem or mds in dispatch threads */
 964                        ceph_async_iput(inode);
 965                        continue;
 966
 967skip_inode:
 968                        spin_unlock(&ci->i_ceph_lock);
 969                        ceph_async_iput(inode);
 970                }
 971
 972                /* we may have taken some of the old realm's children. */
 973                for (i = 0; i < num_split_realms; i++) {
 974                        struct ceph_snap_realm *child =
 975                                __lookup_snap_realm(mdsc,
 976                                           le64_to_cpu(split_realms[i]));
 977                        if (!child)
 978                                continue;
 979                        adjust_snap_realm_parent(mdsc, child, realm->ino);
 980                }
 981        }
 982
 983        /*
 984         * update using the provided snap trace. if we are deleting a
 985         * snap, we can avoid queueing cap_snaps.
 986         */
 987        ceph_update_snap_trace(mdsc, p, e,
 988                               op == CEPH_SNAP_OP_DESTROY, NULL);
 989
 990        if (op == CEPH_SNAP_OP_SPLIT)
 991                /* we took a reference when we created the realm, above */
 992                ceph_put_snap_realm(mdsc, realm);
 993
 994        __cleanup_empty_realms(mdsc);
 995
 996        up_write(&mdsc->snap_rwsem);
 997
 998        flush_snaps(mdsc);
 999        return;
1000
1001bad:
1002        pr_err("corrupt snap message from mds%d\n", mds);
1003        ceph_msg_dump(msg);
1004out:
1005        if (locked_rwsem)
1006                up_write(&mdsc->snap_rwsem);
1007        return;
1008}
1009
1010struct ceph_snapid_map* ceph_get_snapid_map(struct ceph_mds_client *mdsc,
1011                                            u64 snap)
1012{
1013        struct ceph_snapid_map *sm, *exist;
1014        struct rb_node **p, *parent;
1015        int ret;
1016
1017        exist = NULL;
1018        spin_lock(&mdsc->snapid_map_lock);
1019        p = &mdsc->snapid_map_tree.rb_node;
1020        while (*p) {
1021                exist = rb_entry(*p, struct ceph_snapid_map, node);
1022                if (snap > exist->snap) {
1023                        p = &(*p)->rb_left;
1024                } else if (snap < exist->snap) {
1025                        p = &(*p)->rb_right;
1026                } else {
1027                        if (atomic_inc_return(&exist->ref) == 1)
1028                                list_del_init(&exist->lru);
1029                        break;
1030                }
1031                exist = NULL;
1032        }
1033        spin_unlock(&mdsc->snapid_map_lock);
1034        if (exist) {
1035                dout("found snapid map %llx -> %x\n", exist->snap, exist->dev);
1036                return exist;
1037        }
1038
1039        sm = kmalloc(sizeof(*sm), GFP_NOFS);
1040        if (!sm)
1041                return NULL;
1042
1043        ret = get_anon_bdev(&sm->dev);
1044        if (ret < 0) {
1045                kfree(sm);
1046                return NULL;
1047        }
1048
1049        INIT_LIST_HEAD(&sm->lru);
1050        atomic_set(&sm->ref, 1);
1051        sm->snap = snap;
1052
1053        exist = NULL;
1054        parent = NULL;
1055        p = &mdsc->snapid_map_tree.rb_node;
1056        spin_lock(&mdsc->snapid_map_lock);
1057        while (*p) {
1058                parent = *p;
1059                exist = rb_entry(*p, struct ceph_snapid_map, node);
1060                if (snap > exist->snap)
1061                        p = &(*p)->rb_left;
1062                else if (snap < exist->snap)
1063                        p = &(*p)->rb_right;
1064                else
1065                        break;
1066                exist = NULL;
1067        }
1068        if (exist) {
1069                if (atomic_inc_return(&exist->ref) == 1)
1070                        list_del_init(&exist->lru);
1071        } else {
1072                rb_link_node(&sm->node, parent, p);
1073                rb_insert_color(&sm->node, &mdsc->snapid_map_tree);
1074        }
1075        spin_unlock(&mdsc->snapid_map_lock);
1076        if (exist) {
1077                free_anon_bdev(sm->dev);
1078                kfree(sm);
1079                dout("found snapid map %llx -> %x\n", exist->snap, exist->dev);
1080                return exist;
1081        }
1082
1083        dout("create snapid map %llx -> %x\n", sm->snap, sm->dev);
1084        return sm;
1085}
1086
1087void ceph_put_snapid_map(struct ceph_mds_client* mdsc,
1088                         struct ceph_snapid_map *sm)
1089{
1090        if (!sm)
1091                return;
1092        if (atomic_dec_and_lock(&sm->ref, &mdsc->snapid_map_lock)) {
1093                if (!RB_EMPTY_NODE(&sm->node)) {
1094                        sm->last_used = jiffies;
1095                        list_add_tail(&sm->lru, &mdsc->snapid_map_lru);
1096                        spin_unlock(&mdsc->snapid_map_lock);
1097                } else {
1098                        /* already cleaned up by
1099                         * ceph_cleanup_snapid_map() */
1100                        spin_unlock(&mdsc->snapid_map_lock);
1101                        kfree(sm);
1102                }
1103        }
1104}
1105
1106void ceph_trim_snapid_map(struct ceph_mds_client *mdsc)
1107{
1108        struct ceph_snapid_map *sm;
1109        unsigned long now;
1110        LIST_HEAD(to_free);
1111
1112        spin_lock(&mdsc->snapid_map_lock);
1113        now = jiffies;
1114
1115        while (!list_empty(&mdsc->snapid_map_lru)) {
1116                sm = list_first_entry(&mdsc->snapid_map_lru,
1117                                      struct ceph_snapid_map, lru);
1118                if (time_after(sm->last_used + CEPH_SNAPID_MAP_TIMEOUT, now))
1119                        break;
1120
1121                rb_erase(&sm->node, &mdsc->snapid_map_tree);
1122                list_move(&sm->lru, &to_free);
1123        }
1124        spin_unlock(&mdsc->snapid_map_lock);
1125
1126        while (!list_empty(&to_free)) {
1127                sm = list_first_entry(&to_free, struct ceph_snapid_map, lru);
1128                list_del(&sm->lru);
1129                dout("trim snapid map %llx -> %x\n", sm->snap, sm->dev);
1130                free_anon_bdev(sm->dev);
1131                kfree(sm);
1132        }
1133}
1134
1135void ceph_cleanup_snapid_map(struct ceph_mds_client *mdsc)
1136{
1137        struct ceph_snapid_map *sm;
1138        struct rb_node *p;
1139        LIST_HEAD(to_free);
1140
1141        spin_lock(&mdsc->snapid_map_lock);
1142        while ((p = rb_first(&mdsc->snapid_map_tree))) {
1143                sm = rb_entry(p, struct ceph_snapid_map, node);
1144                rb_erase(p, &mdsc->snapid_map_tree);
1145                RB_CLEAR_NODE(p);
1146                list_move(&sm->lru, &to_free);
1147        }
1148        spin_unlock(&mdsc->snapid_map_lock);
1149
1150        while (!list_empty(&to_free)) {
1151                sm = list_first_entry(&to_free, struct ceph_snapid_map, lru);
1152                list_del(&sm->lru);
1153                free_anon_bdev(sm->dev);
1154                if (WARN_ON_ONCE(atomic_read(&sm->ref))) {
1155                        pr_err("snapid map %llx -> %x still in use\n",
1156                               sm->snap, sm->dev);
1157                }
1158                kfree(sm);
1159        }
1160}
1161