root/fs/ceph/caps.c

/* [<][>][^][v][top][bottom][index][help] */

DEFINITIONS

This source file includes following definitions.
  1. gcap_string
  2. ceph_cap_string
  3. ceph_caps_init
  4. ceph_caps_finalize
  5. ceph_adjust_caps_max_min
  6. __ceph_unreserve_caps
  7. ceph_reserve_caps
  8. ceph_unreserve_caps
  9. ceph_get_cap
  10. ceph_put_cap
  11. ceph_reservation_status
  12. __get_cap_for_mds
  13. ceph_get_cap_for_mds
  14. __insert_cap_node
  15. __cap_set_timeouts
  16. __cap_delay_requeue
  17. __cap_delay_requeue_front
  18. __cap_delay_cancel
  19. __check_cap_issue
  20. ceph_add_cap
  21. __cap_is_valid
  22. __ceph_caps_issued
  23. __ceph_caps_issued_other
  24. __touch_cap
  25. __ceph_caps_issued_mask
  26. __ceph_caps_revoking_other
  27. ceph_caps_revoking
  28. __ceph_caps_used
  29. __ceph_caps_file_wanted
  30. __ceph_caps_mds_wanted
  31. __ceph_is_single_caps
  32. __ceph_is_any_caps
  33. ceph_is_any_caps
  34. drop_inode_snap_realm
  35. __ceph_remove_cap
  36. send_cap_msg
  37. __ceph_remove_caps
  38. __send_cap
  39. __send_flush_snap
  40. __ceph_flush_snaps
  41. ceph_flush_snaps
  42. __ceph_mark_dirty_caps
  43. ceph_alloc_cap_flush
  44. ceph_free_cap_flush
  45. __get_oldest_flush_tid
  46. __finish_cap_flush
  47. __mark_caps_flushing
  48. try_nonblocking_invalidate
  49. __ceph_should_report_size
  50. ceph_check_caps
  51. try_flush_caps
  52. caps_are_flushed
  53. unsafe_request_wait
  54. ceph_fsync
  55. ceph_write_inode
  56. __kick_flushing_caps
  57. ceph_early_kick_flushing_caps
  58. ceph_kick_flushing_caps
  59. kick_flushing_inode_caps
  60. __take_cap_refs
  61. try_get_cap_refs
  62. check_max_size
  63. ceph_try_get_caps
  64. ceph_get_caps
  65. ceph_get_cap_refs
  66. ceph_try_drop_cap_snap
  67. ceph_put_cap_refs
  68. ceph_put_wrbuffer_cap_refs
  69. invalidate_aliases
  70. handle_cap_grant
  71. handle_cap_flush_ack
  72. handle_cap_flushsnap_ack
  73. handle_cap_trunc
  74. handle_cap_export
  75. handle_cap_import
  76. ceph_handle_caps
  77. ceph_check_delayed_caps
  78. ceph_flush_dirty_caps
  79. __ceph_get_fmode
  80. ceph_put_fmode
  81. ceph_drop_caps_for_unlink
  82. ceph_encode_inode_release
  83. ceph_encode_dentry_release

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/ceph/ceph_debug.h>
   3 
   4 #include <linux/fs.h>
   5 #include <linux/kernel.h>
   6 #include <linux/sched/signal.h>
   7 #include <linux/slab.h>
   8 #include <linux/vmalloc.h>
   9 #include <linux/wait.h>
  10 #include <linux/writeback.h>
  11 #include <linux/iversion.h>
  12 
  13 #include "super.h"
  14 #include "mds_client.h"
  15 #include "cache.h"
  16 #include <linux/ceph/decode.h>
  17 #include <linux/ceph/messenger.h>
  18 
  19 /*
  20  * Capability management
  21  *
  22  * The Ceph metadata servers control client access to inode metadata
  23  * and file data by issuing capabilities, granting clients permission
  24  * to read and/or write both inode field and file data to OSDs
  25  * (storage nodes).  Each capability consists of a set of bits
  26  * indicating which operations are allowed.
  27  *
  28  * If the client holds a *_SHARED cap, the client has a coherent value
  29  * that can be safely read from the cached inode.
  30  *
  31  * In the case of a *_EXCL (exclusive) or FILE_WR capabilities, the
  32  * client is allowed to change inode attributes (e.g., file size,
  33  * mtime), note its dirty state in the ceph_cap, and asynchronously
  34  * flush that metadata change to the MDS.
  35  *
  36  * In the event of a conflicting operation (perhaps by another
  37  * client), the MDS will revoke the conflicting client capabilities.
  38  *
  39  * In order for a client to cache an inode, it must hold a capability
  40  * with at least one MDS server.  When inodes are released, release
  41  * notifications are batched and periodically sent en masse to the MDS
  42  * cluster to release server state.
  43  */
  44 
  45 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc);
  46 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
  47                                  struct ceph_mds_session *session,
  48                                  struct ceph_inode_info *ci,
  49                                  u64 oldest_flush_tid);
  50 
  51 /*
  52  * Generate readable cap strings for debugging output.
  53  */
  54 #define MAX_CAP_STR 20
  55 static char cap_str[MAX_CAP_STR][40];
  56 static DEFINE_SPINLOCK(cap_str_lock);
  57 static int last_cap_str;
  58 
  59 static char *gcap_string(char *s, int c)
  60 {
  61         if (c & CEPH_CAP_GSHARED)
  62                 *s++ = 's';
  63         if (c & CEPH_CAP_GEXCL)
  64                 *s++ = 'x';
  65         if (c & CEPH_CAP_GCACHE)
  66                 *s++ = 'c';
  67         if (c & CEPH_CAP_GRD)
  68                 *s++ = 'r';
  69         if (c & CEPH_CAP_GWR)
  70                 *s++ = 'w';
  71         if (c & CEPH_CAP_GBUFFER)
  72                 *s++ = 'b';
  73         if (c & CEPH_CAP_GWREXTEND)
  74                 *s++ = 'a';
  75         if (c & CEPH_CAP_GLAZYIO)
  76                 *s++ = 'l';
  77         return s;
  78 }
  79 
  80 const char *ceph_cap_string(int caps)
  81 {
  82         int i;
  83         char *s;
  84         int c;
  85 
  86         spin_lock(&cap_str_lock);
  87         i = last_cap_str++;
  88         if (last_cap_str == MAX_CAP_STR)
  89                 last_cap_str = 0;
  90         spin_unlock(&cap_str_lock);
  91 
  92         s = cap_str[i];
  93 
  94         if (caps & CEPH_CAP_PIN)
  95                 *s++ = 'p';
  96 
  97         c = (caps >> CEPH_CAP_SAUTH) & 3;
  98         if (c) {
  99                 *s++ = 'A';
 100                 s = gcap_string(s, c);
 101         }
 102 
 103         c = (caps >> CEPH_CAP_SLINK) & 3;
 104         if (c) {
 105                 *s++ = 'L';
 106                 s = gcap_string(s, c);
 107         }
 108 
 109         c = (caps >> CEPH_CAP_SXATTR) & 3;
 110         if (c) {
 111                 *s++ = 'X';
 112                 s = gcap_string(s, c);
 113         }
 114 
 115         c = caps >> CEPH_CAP_SFILE;
 116         if (c) {
 117                 *s++ = 'F';
 118                 s = gcap_string(s, c);
 119         }
 120 
 121         if (s == cap_str[i])
 122                 *s++ = '-';
 123         *s = 0;
 124         return cap_str[i];
 125 }
 126 
 127 void ceph_caps_init(struct ceph_mds_client *mdsc)
 128 {
 129         INIT_LIST_HEAD(&mdsc->caps_list);
 130         spin_lock_init(&mdsc->caps_list_lock);
 131 }
 132 
 133 void ceph_caps_finalize(struct ceph_mds_client *mdsc)
 134 {
 135         struct ceph_cap *cap;
 136 
 137         spin_lock(&mdsc->caps_list_lock);
 138         while (!list_empty(&mdsc->caps_list)) {
 139                 cap = list_first_entry(&mdsc->caps_list,
 140                                        struct ceph_cap, caps_item);
 141                 list_del(&cap->caps_item);
 142                 kmem_cache_free(ceph_cap_cachep, cap);
 143         }
 144         mdsc->caps_total_count = 0;
 145         mdsc->caps_avail_count = 0;
 146         mdsc->caps_use_count = 0;
 147         mdsc->caps_reserve_count = 0;
 148         mdsc->caps_min_count = 0;
 149         spin_unlock(&mdsc->caps_list_lock);
 150 }
 151 
 152 void ceph_adjust_caps_max_min(struct ceph_mds_client *mdsc,
 153                               struct ceph_mount_options *fsopt)
 154 {
 155         spin_lock(&mdsc->caps_list_lock);
 156         mdsc->caps_min_count = fsopt->max_readdir;
 157         if (mdsc->caps_min_count < 1024)
 158                 mdsc->caps_min_count = 1024;
 159         mdsc->caps_use_max = fsopt->caps_max;
 160         if (mdsc->caps_use_max > 0 &&
 161             mdsc->caps_use_max < mdsc->caps_min_count)
 162                 mdsc->caps_use_max = mdsc->caps_min_count;
 163         spin_unlock(&mdsc->caps_list_lock);
 164 }
 165 
 166 static void __ceph_unreserve_caps(struct ceph_mds_client *mdsc, int nr_caps)
 167 {
 168         struct ceph_cap *cap;
 169         int i;
 170 
 171         if (nr_caps) {
 172                 BUG_ON(mdsc->caps_reserve_count < nr_caps);
 173                 mdsc->caps_reserve_count -= nr_caps;
 174                 if (mdsc->caps_avail_count >=
 175                     mdsc->caps_reserve_count + mdsc->caps_min_count) {
 176                         mdsc->caps_total_count -= nr_caps;
 177                         for (i = 0; i < nr_caps; i++) {
 178                                 cap = list_first_entry(&mdsc->caps_list,
 179                                         struct ceph_cap, caps_item);
 180                                 list_del(&cap->caps_item);
 181                                 kmem_cache_free(ceph_cap_cachep, cap);
 182                         }
 183                 } else {
 184                         mdsc->caps_avail_count += nr_caps;
 185                 }
 186 
 187                 dout("%s: caps %d = %d used + %d resv + %d avail\n",
 188                      __func__,
 189                      mdsc->caps_total_count, mdsc->caps_use_count,
 190                      mdsc->caps_reserve_count, mdsc->caps_avail_count);
 191                 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
 192                                                  mdsc->caps_reserve_count +
 193                                                  mdsc->caps_avail_count);
 194         }
 195 }
 196 
 197 /*
 198  * Called under mdsc->mutex.
 199  */
 200 int ceph_reserve_caps(struct ceph_mds_client *mdsc,
 201                       struct ceph_cap_reservation *ctx, int need)
 202 {
 203         int i, j;
 204         struct ceph_cap *cap;
 205         int have;
 206         int alloc = 0;
 207         int max_caps;
 208         int err = 0;
 209         bool trimmed = false;
 210         struct ceph_mds_session *s;
 211         LIST_HEAD(newcaps);
 212 
 213         dout("reserve caps ctx=%p need=%d\n", ctx, need);
 214 
 215         /* first reserve any caps that are already allocated */
 216         spin_lock(&mdsc->caps_list_lock);
 217         if (mdsc->caps_avail_count >= need)
 218                 have = need;
 219         else
 220                 have = mdsc->caps_avail_count;
 221         mdsc->caps_avail_count -= have;
 222         mdsc->caps_reserve_count += have;
 223         BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
 224                                          mdsc->caps_reserve_count +
 225                                          mdsc->caps_avail_count);
 226         spin_unlock(&mdsc->caps_list_lock);
 227 
 228         for (i = have; i < need; ) {
 229                 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
 230                 if (cap) {
 231                         list_add(&cap->caps_item, &newcaps);
 232                         alloc++;
 233                         i++;
 234                         continue;
 235                 }
 236 
 237                 if (!trimmed) {
 238                         for (j = 0; j < mdsc->max_sessions; j++) {
 239                                 s = __ceph_lookup_mds_session(mdsc, j);
 240                                 if (!s)
 241                                         continue;
 242                                 mutex_unlock(&mdsc->mutex);
 243 
 244                                 mutex_lock(&s->s_mutex);
 245                                 max_caps = s->s_nr_caps - (need - i);
 246                                 ceph_trim_caps(mdsc, s, max_caps);
 247                                 mutex_unlock(&s->s_mutex);
 248 
 249                                 ceph_put_mds_session(s);
 250                                 mutex_lock(&mdsc->mutex);
 251                         }
 252                         trimmed = true;
 253 
 254                         spin_lock(&mdsc->caps_list_lock);
 255                         if (mdsc->caps_avail_count) {
 256                                 int more_have;
 257                                 if (mdsc->caps_avail_count >= need - i)
 258                                         more_have = need - i;
 259                                 else
 260                                         more_have = mdsc->caps_avail_count;
 261 
 262                                 i += more_have;
 263                                 have += more_have;
 264                                 mdsc->caps_avail_count -= more_have;
 265                                 mdsc->caps_reserve_count += more_have;
 266 
 267                         }
 268                         spin_unlock(&mdsc->caps_list_lock);
 269 
 270                         continue;
 271                 }
 272 
 273                 pr_warn("reserve caps ctx=%p ENOMEM need=%d got=%d\n",
 274                         ctx, need, have + alloc);
 275                 err = -ENOMEM;
 276                 break;
 277         }
 278 
 279         if (!err) {
 280                 BUG_ON(have + alloc != need);
 281                 ctx->count = need;
 282                 ctx->used = 0;
 283         }
 284 
 285         spin_lock(&mdsc->caps_list_lock);
 286         mdsc->caps_total_count += alloc;
 287         mdsc->caps_reserve_count += alloc;
 288         list_splice(&newcaps, &mdsc->caps_list);
 289 
 290         BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
 291                                          mdsc->caps_reserve_count +
 292                                          mdsc->caps_avail_count);
 293 
 294         if (err)
 295                 __ceph_unreserve_caps(mdsc, have + alloc);
 296 
 297         spin_unlock(&mdsc->caps_list_lock);
 298 
 299         dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n",
 300              ctx, mdsc->caps_total_count, mdsc->caps_use_count,
 301              mdsc->caps_reserve_count, mdsc->caps_avail_count);
 302         return err;
 303 }
 304 
 305 void ceph_unreserve_caps(struct ceph_mds_client *mdsc,
 306                          struct ceph_cap_reservation *ctx)
 307 {
 308         bool reclaim = false;
 309         if (!ctx->count)
 310                 return;
 311 
 312         dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count);
 313         spin_lock(&mdsc->caps_list_lock);
 314         __ceph_unreserve_caps(mdsc, ctx->count);
 315         ctx->count = 0;
 316 
 317         if (mdsc->caps_use_max > 0 &&
 318             mdsc->caps_use_count > mdsc->caps_use_max)
 319                 reclaim = true;
 320         spin_unlock(&mdsc->caps_list_lock);
 321 
 322         if (reclaim)
 323                 ceph_reclaim_caps_nr(mdsc, ctx->used);
 324 }
 325 
 326 struct ceph_cap *ceph_get_cap(struct ceph_mds_client *mdsc,
 327                               struct ceph_cap_reservation *ctx)
 328 {
 329         struct ceph_cap *cap = NULL;
 330 
 331         /* temporary, until we do something about cap import/export */
 332         if (!ctx) {
 333                 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
 334                 if (cap) {
 335                         spin_lock(&mdsc->caps_list_lock);
 336                         mdsc->caps_use_count++;
 337                         mdsc->caps_total_count++;
 338                         spin_unlock(&mdsc->caps_list_lock);
 339                 } else {
 340                         spin_lock(&mdsc->caps_list_lock);
 341                         if (mdsc->caps_avail_count) {
 342                                 BUG_ON(list_empty(&mdsc->caps_list));
 343 
 344                                 mdsc->caps_avail_count--;
 345                                 mdsc->caps_use_count++;
 346                                 cap = list_first_entry(&mdsc->caps_list,
 347                                                 struct ceph_cap, caps_item);
 348                                 list_del(&cap->caps_item);
 349 
 350                                 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
 351                                        mdsc->caps_reserve_count + mdsc->caps_avail_count);
 352                         }
 353                         spin_unlock(&mdsc->caps_list_lock);
 354                 }
 355 
 356                 return cap;
 357         }
 358 
 359         spin_lock(&mdsc->caps_list_lock);
 360         dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n",
 361              ctx, ctx->count, mdsc->caps_total_count, mdsc->caps_use_count,
 362              mdsc->caps_reserve_count, mdsc->caps_avail_count);
 363         BUG_ON(!ctx->count);
 364         BUG_ON(ctx->count > mdsc->caps_reserve_count);
 365         BUG_ON(list_empty(&mdsc->caps_list));
 366 
 367         ctx->count--;
 368         ctx->used++;
 369         mdsc->caps_reserve_count--;
 370         mdsc->caps_use_count++;
 371 
 372         cap = list_first_entry(&mdsc->caps_list, struct ceph_cap, caps_item);
 373         list_del(&cap->caps_item);
 374 
 375         BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
 376                mdsc->caps_reserve_count + mdsc->caps_avail_count);
 377         spin_unlock(&mdsc->caps_list_lock);
 378         return cap;
 379 }
 380 
 381 void ceph_put_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap)
 382 {
 383         spin_lock(&mdsc->caps_list_lock);
 384         dout("put_cap %p %d = %d used + %d resv + %d avail\n",
 385              cap, mdsc->caps_total_count, mdsc->caps_use_count,
 386              mdsc->caps_reserve_count, mdsc->caps_avail_count);
 387         mdsc->caps_use_count--;
 388         /*
 389          * Keep some preallocated caps around (ceph_min_count), to
 390          * avoid lots of free/alloc churn.
 391          */
 392         if (mdsc->caps_avail_count >= mdsc->caps_reserve_count +
 393                                       mdsc->caps_min_count) {
 394                 mdsc->caps_total_count--;
 395                 kmem_cache_free(ceph_cap_cachep, cap);
 396         } else {
 397                 mdsc->caps_avail_count++;
 398                 list_add(&cap->caps_item, &mdsc->caps_list);
 399         }
 400 
 401         BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
 402                mdsc->caps_reserve_count + mdsc->caps_avail_count);
 403         spin_unlock(&mdsc->caps_list_lock);
 404 }
 405 
 406 void ceph_reservation_status(struct ceph_fs_client *fsc,
 407                              int *total, int *avail, int *used, int *reserved,
 408                              int *min)
 409 {
 410         struct ceph_mds_client *mdsc = fsc->mdsc;
 411 
 412         spin_lock(&mdsc->caps_list_lock);
 413 
 414         if (total)
 415                 *total = mdsc->caps_total_count;
 416         if (avail)
 417                 *avail = mdsc->caps_avail_count;
 418         if (used)
 419                 *used = mdsc->caps_use_count;
 420         if (reserved)
 421                 *reserved = mdsc->caps_reserve_count;
 422         if (min)
 423                 *min = mdsc->caps_min_count;
 424 
 425         spin_unlock(&mdsc->caps_list_lock);
 426 }
 427 
 428 /*
 429  * Find ceph_cap for given mds, if any.
 430  *
 431  * Called with i_ceph_lock held.
 432  */
 433 static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds)
 434 {
 435         struct ceph_cap *cap;
 436         struct rb_node *n = ci->i_caps.rb_node;
 437 
 438         while (n) {
 439                 cap = rb_entry(n, struct ceph_cap, ci_node);
 440                 if (mds < cap->mds)
 441                         n = n->rb_left;
 442                 else if (mds > cap->mds)
 443                         n = n->rb_right;
 444                 else
 445                         return cap;
 446         }
 447         return NULL;
 448 }
 449 
 450 struct ceph_cap *ceph_get_cap_for_mds(struct ceph_inode_info *ci, int mds)
 451 {
 452         struct ceph_cap *cap;
 453 
 454         spin_lock(&ci->i_ceph_lock);
 455         cap = __get_cap_for_mds(ci, mds);
 456         spin_unlock(&ci->i_ceph_lock);
 457         return cap;
 458 }
 459 
 460 /*
 461  * Called under i_ceph_lock.
 462  */
 463 static void __insert_cap_node(struct ceph_inode_info *ci,
 464                               struct ceph_cap *new)
 465 {
 466         struct rb_node **p = &ci->i_caps.rb_node;
 467         struct rb_node *parent = NULL;
 468         struct ceph_cap *cap = NULL;
 469 
 470         while (*p) {
 471                 parent = *p;
 472                 cap = rb_entry(parent, struct ceph_cap, ci_node);
 473                 if (new->mds < cap->mds)
 474                         p = &(*p)->rb_left;
 475                 else if (new->mds > cap->mds)
 476                         p = &(*p)->rb_right;
 477                 else
 478                         BUG();
 479         }
 480 
 481         rb_link_node(&new->ci_node, parent, p);
 482         rb_insert_color(&new->ci_node, &ci->i_caps);
 483 }
 484 
 485 /*
 486  * (re)set cap hold timeouts, which control the delayed release
 487  * of unused caps back to the MDS.  Should be called on cap use.
 488  */
 489 static void __cap_set_timeouts(struct ceph_mds_client *mdsc,
 490                                struct ceph_inode_info *ci)
 491 {
 492         struct ceph_mount_options *opt = mdsc->fsc->mount_options;
 493 
 494         ci->i_hold_caps_min = round_jiffies(jiffies +
 495                                             opt->caps_wanted_delay_min * HZ);
 496         ci->i_hold_caps_max = round_jiffies(jiffies +
 497                                             opt->caps_wanted_delay_max * HZ);
 498         dout("__cap_set_timeouts %p min %lu max %lu\n", &ci->vfs_inode,
 499              ci->i_hold_caps_min - jiffies, ci->i_hold_caps_max - jiffies);
 500 }
 501 
 502 /*
 503  * (Re)queue cap at the end of the delayed cap release list.
 504  *
 505  * If I_FLUSH is set, leave the inode at the front of the list.
 506  *
 507  * Caller holds i_ceph_lock
 508  *    -> we take mdsc->cap_delay_lock
 509  */
 510 static void __cap_delay_requeue(struct ceph_mds_client *mdsc,
 511                                 struct ceph_inode_info *ci,
 512                                 bool set_timeout)
 513 {
 514         dout("__cap_delay_requeue %p flags %d at %lu\n", &ci->vfs_inode,
 515              ci->i_ceph_flags, ci->i_hold_caps_max);
 516         if (!mdsc->stopping) {
 517                 spin_lock(&mdsc->cap_delay_lock);
 518                 if (!list_empty(&ci->i_cap_delay_list)) {
 519                         if (ci->i_ceph_flags & CEPH_I_FLUSH)
 520                                 goto no_change;
 521                         list_del_init(&ci->i_cap_delay_list);
 522                 }
 523                 if (set_timeout)
 524                         __cap_set_timeouts(mdsc, ci);
 525                 list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
 526 no_change:
 527                 spin_unlock(&mdsc->cap_delay_lock);
 528         }
 529 }
 530 
 531 /*
 532  * Queue an inode for immediate writeback.  Mark inode with I_FLUSH,
 533  * indicating we should send a cap message to flush dirty metadata
 534  * asap, and move to the front of the delayed cap list.
 535  */
 536 static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc,
 537                                       struct ceph_inode_info *ci)
 538 {
 539         dout("__cap_delay_requeue_front %p\n", &ci->vfs_inode);
 540         spin_lock(&mdsc->cap_delay_lock);
 541         ci->i_ceph_flags |= CEPH_I_FLUSH;
 542         if (!list_empty(&ci->i_cap_delay_list))
 543                 list_del_init(&ci->i_cap_delay_list);
 544         list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
 545         spin_unlock(&mdsc->cap_delay_lock);
 546 }
 547 
 548 /*
 549  * Cancel delayed work on cap.
 550  *
 551  * Caller must hold i_ceph_lock.
 552  */
 553 static void __cap_delay_cancel(struct ceph_mds_client *mdsc,
 554                                struct ceph_inode_info *ci)
 555 {
 556         dout("__cap_delay_cancel %p\n", &ci->vfs_inode);
 557         if (list_empty(&ci->i_cap_delay_list))
 558                 return;
 559         spin_lock(&mdsc->cap_delay_lock);
 560         list_del_init(&ci->i_cap_delay_list);
 561         spin_unlock(&mdsc->cap_delay_lock);
 562 }
 563 
 564 /*
 565  * Common issue checks for add_cap, handle_cap_grant.
 566  */
 567 static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap,
 568                               unsigned issued)
 569 {
 570         unsigned had = __ceph_caps_issued(ci, NULL);
 571 
 572         /*
 573          * Each time we receive FILE_CACHE anew, we increment
 574          * i_rdcache_gen.
 575          */
 576         if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
 577             (had & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) {
 578                 ci->i_rdcache_gen++;
 579         }
 580 
 581         /*
 582          * If FILE_SHARED is newly issued, mark dir not complete. We don't
 583          * know what happened to this directory while we didn't have the cap.
 584          * If FILE_SHARED is being revoked, also mark dir not complete. It
 585          * stops on-going cached readdir.
 586          */
 587         if ((issued & CEPH_CAP_FILE_SHARED) != (had & CEPH_CAP_FILE_SHARED)) {
 588                 if (issued & CEPH_CAP_FILE_SHARED)
 589                         atomic_inc(&ci->i_shared_gen);
 590                 if (S_ISDIR(ci->vfs_inode.i_mode)) {
 591                         dout(" marking %p NOT complete\n", &ci->vfs_inode);
 592                         __ceph_dir_clear_complete(ci);
 593                 }
 594         }
 595 }
 596 
 597 /*
 598  * Add a capability under the given MDS session.
 599  *
 600  * Caller should hold session snap_rwsem (read) and ci->i_ceph_lock
 601  *
 602  * @fmode is the open file mode, if we are opening a file, otherwise
 603  * it is < 0.  (This is so we can atomically add the cap and add an
 604  * open file reference to it.)
 605  */
 606 void ceph_add_cap(struct inode *inode,
 607                   struct ceph_mds_session *session, u64 cap_id,
 608                   int fmode, unsigned issued, unsigned wanted,
 609                   unsigned seq, unsigned mseq, u64 realmino, int flags,
 610                   struct ceph_cap **new_cap)
 611 {
 612         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
 613         struct ceph_inode_info *ci = ceph_inode(inode);
 614         struct ceph_cap *cap;
 615         int mds = session->s_mds;
 616         int actual_wanted;
 617         u32 gen;
 618 
 619         lockdep_assert_held(&ci->i_ceph_lock);
 620 
 621         dout("add_cap %p mds%d cap %llx %s seq %d\n", inode,
 622              session->s_mds, cap_id, ceph_cap_string(issued), seq);
 623 
 624         /*
 625          * If we are opening the file, include file mode wanted bits
 626          * in wanted.
 627          */
 628         if (fmode >= 0)
 629                 wanted |= ceph_caps_for_mode(fmode);
 630 
 631         spin_lock(&session->s_gen_ttl_lock);
 632         gen = session->s_cap_gen;
 633         spin_unlock(&session->s_gen_ttl_lock);
 634 
 635         cap = __get_cap_for_mds(ci, mds);
 636         if (!cap) {
 637                 cap = *new_cap;
 638                 *new_cap = NULL;
 639 
 640                 cap->issued = 0;
 641                 cap->implemented = 0;
 642                 cap->mds = mds;
 643                 cap->mds_wanted = 0;
 644                 cap->mseq = 0;
 645 
 646                 cap->ci = ci;
 647                 __insert_cap_node(ci, cap);
 648 
 649                 /* add to session cap list */
 650                 cap->session = session;
 651                 spin_lock(&session->s_cap_lock);
 652                 list_add_tail(&cap->session_caps, &session->s_caps);
 653                 session->s_nr_caps++;
 654                 spin_unlock(&session->s_cap_lock);
 655         } else {
 656                 spin_lock(&session->s_cap_lock);
 657                 list_move_tail(&cap->session_caps, &session->s_caps);
 658                 spin_unlock(&session->s_cap_lock);
 659 
 660                 if (cap->cap_gen < gen)
 661                         cap->issued = cap->implemented = CEPH_CAP_PIN;
 662 
 663                 /*
 664                  * auth mds of the inode changed. we received the cap export
 665                  * message, but still haven't received the cap import message.
 666                  * handle_cap_export() updated the new auth MDS' cap.
 667                  *
 668                  * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing
 669                  * a message that was send before the cap import message. So
 670                  * don't remove caps.
 671                  */
 672                 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
 673                         WARN_ON(cap != ci->i_auth_cap);
 674                         WARN_ON(cap->cap_id != cap_id);
 675                         seq = cap->seq;
 676                         mseq = cap->mseq;
 677                         issued |= cap->issued;
 678                         flags |= CEPH_CAP_FLAG_AUTH;
 679                 }
 680         }
 681 
 682         if (!ci->i_snap_realm ||
 683             ((flags & CEPH_CAP_FLAG_AUTH) &&
 684              realmino != (u64)-1 && ci->i_snap_realm->ino != realmino)) {
 685                 /*
 686                  * add this inode to the appropriate snap realm
 687                  */
 688                 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc,
 689                                                                realmino);
 690                 if (realm) {
 691                         struct ceph_snap_realm *oldrealm = ci->i_snap_realm;
 692                         if (oldrealm) {
 693                                 spin_lock(&oldrealm->inodes_with_caps_lock);
 694                                 list_del_init(&ci->i_snap_realm_item);
 695                                 spin_unlock(&oldrealm->inodes_with_caps_lock);
 696                         }
 697 
 698                         spin_lock(&realm->inodes_with_caps_lock);
 699                         list_add(&ci->i_snap_realm_item,
 700                                  &realm->inodes_with_caps);
 701                         ci->i_snap_realm = realm;
 702                         if (realm->ino == ci->i_vino.ino)
 703                                 realm->inode = inode;
 704                         spin_unlock(&realm->inodes_with_caps_lock);
 705 
 706                         if (oldrealm)
 707                                 ceph_put_snap_realm(mdsc, oldrealm);
 708                 } else {
 709                         pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
 710                                realmino);
 711                         WARN_ON(!realm);
 712                 }
 713         }
 714 
 715         __check_cap_issue(ci, cap, issued);
 716 
 717         /*
 718          * If we are issued caps we don't want, or the mds' wanted
 719          * value appears to be off, queue a check so we'll release
 720          * later and/or update the mds wanted value.
 721          */
 722         actual_wanted = __ceph_caps_wanted(ci);
 723         if ((wanted & ~actual_wanted) ||
 724             (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
 725                 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
 726                      ceph_cap_string(issued), ceph_cap_string(wanted),
 727                      ceph_cap_string(actual_wanted));
 728                 __cap_delay_requeue(mdsc, ci, true);
 729         }
 730 
 731         if (flags & CEPH_CAP_FLAG_AUTH) {
 732                 if (!ci->i_auth_cap ||
 733                     ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) {
 734                         ci->i_auth_cap = cap;
 735                         cap->mds_wanted = wanted;
 736                 }
 737         } else {
 738                 WARN_ON(ci->i_auth_cap == cap);
 739         }
 740 
 741         dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
 742              inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
 743              ceph_cap_string(issued|cap->issued), seq, mds);
 744         cap->cap_id = cap_id;
 745         cap->issued = issued;
 746         cap->implemented |= issued;
 747         if (ceph_seq_cmp(mseq, cap->mseq) > 0)
 748                 cap->mds_wanted = wanted;
 749         else
 750                 cap->mds_wanted |= wanted;
 751         cap->seq = seq;
 752         cap->issue_seq = seq;
 753         cap->mseq = mseq;
 754         cap->cap_gen = gen;
 755 
 756         if (fmode >= 0)
 757                 __ceph_get_fmode(ci, fmode);
 758 }
 759 
 760 /*
 761  * Return true if cap has not timed out and belongs to the current
 762  * generation of the MDS session (i.e. has not gone 'stale' due to
 763  * us losing touch with the mds).
 764  */
 765 static int __cap_is_valid(struct ceph_cap *cap)
 766 {
 767         unsigned long ttl;
 768         u32 gen;
 769 
 770         spin_lock(&cap->session->s_gen_ttl_lock);
 771         gen = cap->session->s_cap_gen;
 772         ttl = cap->session->s_cap_ttl;
 773         spin_unlock(&cap->session->s_gen_ttl_lock);
 774 
 775         if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
 776                 dout("__cap_is_valid %p cap %p issued %s "
 777                      "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode,
 778                      cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
 779                 return 0;
 780         }
 781 
 782         return 1;
 783 }
 784 
 785 /*
 786  * Return set of valid cap bits issued to us.  Note that caps time
 787  * out, and may be invalidated in bulk if the client session times out
 788  * and session->s_cap_gen is bumped.
 789  */
 790 int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
 791 {
 792         int have = ci->i_snap_caps;
 793         struct ceph_cap *cap;
 794         struct rb_node *p;
 795 
 796         if (implemented)
 797                 *implemented = 0;
 798         for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
 799                 cap = rb_entry(p, struct ceph_cap, ci_node);
 800                 if (!__cap_is_valid(cap))
 801                         continue;
 802                 dout("__ceph_caps_issued %p cap %p issued %s\n",
 803                      &ci->vfs_inode, cap, ceph_cap_string(cap->issued));
 804                 have |= cap->issued;
 805                 if (implemented)
 806                         *implemented |= cap->implemented;
 807         }
 808         /*
 809          * exclude caps issued by non-auth MDS, but are been revoking
 810          * by the auth MDS. The non-auth MDS should be revoking/exporting
 811          * these caps, but the message is delayed.
 812          */
 813         if (ci->i_auth_cap) {
 814                 cap = ci->i_auth_cap;
 815                 have &= ~cap->implemented | cap->issued;
 816         }
 817         return have;
 818 }
 819 
 820 /*
 821  * Get cap bits issued by caps other than @ocap
 822  */
 823 int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
 824 {
 825         int have = ci->i_snap_caps;
 826         struct ceph_cap *cap;
 827         struct rb_node *p;
 828 
 829         for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
 830                 cap = rb_entry(p, struct ceph_cap, ci_node);
 831                 if (cap == ocap)
 832                         continue;
 833                 if (!__cap_is_valid(cap))
 834                         continue;
 835                 have |= cap->issued;
 836         }
 837         return have;
 838 }
 839 
 840 /*
 841  * Move a cap to the end of the LRU (oldest caps at list head, newest
 842  * at list tail).
 843  */
 844 static void __touch_cap(struct ceph_cap *cap)
 845 {
 846         struct ceph_mds_session *s = cap->session;
 847 
 848         spin_lock(&s->s_cap_lock);
 849         if (!s->s_cap_iterator) {
 850                 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap,
 851                      s->s_mds);
 852                 list_move_tail(&cap->session_caps, &s->s_caps);
 853         } else {
 854                 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
 855                      &cap->ci->vfs_inode, cap, s->s_mds);
 856         }
 857         spin_unlock(&s->s_cap_lock);
 858 }
 859 
 860 /*
 861  * Check if we hold the given mask.  If so, move the cap(s) to the
 862  * front of their respective LRUs.  (This is the preferred way for
 863  * callers to check for caps they want.)
 864  */
 865 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
 866 {
 867         struct ceph_cap *cap;
 868         struct rb_node *p;
 869         int have = ci->i_snap_caps;
 870 
 871         if ((have & mask) == mask) {
 872                 dout("__ceph_caps_issued_mask ino 0x%lx snap issued %s"
 873                      " (mask %s)\n", ci->vfs_inode.i_ino,
 874                      ceph_cap_string(have),
 875                      ceph_cap_string(mask));
 876                 return 1;
 877         }
 878 
 879         for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
 880                 cap = rb_entry(p, struct ceph_cap, ci_node);
 881                 if (!__cap_is_valid(cap))
 882                         continue;
 883                 if ((cap->issued & mask) == mask) {
 884                         dout("__ceph_caps_issued_mask ino 0x%lx cap %p issued %s"
 885                              " (mask %s)\n", ci->vfs_inode.i_ino, cap,
 886                              ceph_cap_string(cap->issued),
 887                              ceph_cap_string(mask));
 888                         if (touch)
 889                                 __touch_cap(cap);
 890                         return 1;
 891                 }
 892 
 893                 /* does a combination of caps satisfy mask? */
 894                 have |= cap->issued;
 895                 if ((have & mask) == mask) {
 896                         dout("__ceph_caps_issued_mask ino 0x%lx combo issued %s"
 897                              " (mask %s)\n", ci->vfs_inode.i_ino,
 898                              ceph_cap_string(cap->issued),
 899                              ceph_cap_string(mask));
 900                         if (touch) {
 901                                 struct rb_node *q;
 902 
 903                                 /* touch this + preceding caps */
 904                                 __touch_cap(cap);
 905                                 for (q = rb_first(&ci->i_caps); q != p;
 906                                      q = rb_next(q)) {
 907                                         cap = rb_entry(q, struct ceph_cap,
 908                                                        ci_node);
 909                                         if (!__cap_is_valid(cap))
 910                                                 continue;
 911                                         __touch_cap(cap);
 912                                 }
 913                         }
 914                         return 1;
 915                 }
 916         }
 917 
 918         return 0;
 919 }
 920 
 921 /*
 922  * Return true if mask caps are currently being revoked by an MDS.
 923  */
 924 int __ceph_caps_revoking_other(struct ceph_inode_info *ci,
 925                                struct ceph_cap *ocap, int mask)
 926 {
 927         struct ceph_cap *cap;
 928         struct rb_node *p;
 929 
 930         for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
 931                 cap = rb_entry(p, struct ceph_cap, ci_node);
 932                 if (cap != ocap &&
 933                     (cap->implemented & ~cap->issued & mask))
 934                         return 1;
 935         }
 936         return 0;
 937 }
 938 
 939 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
 940 {
 941         struct inode *inode = &ci->vfs_inode;
 942         int ret;
 943 
 944         spin_lock(&ci->i_ceph_lock);
 945         ret = __ceph_caps_revoking_other(ci, NULL, mask);
 946         spin_unlock(&ci->i_ceph_lock);
 947         dout("ceph_caps_revoking %p %s = %d\n", inode,
 948              ceph_cap_string(mask), ret);
 949         return ret;
 950 }
 951 
 952 int __ceph_caps_used(struct ceph_inode_info *ci)
 953 {
 954         int used = 0;
 955         if (ci->i_pin_ref)
 956                 used |= CEPH_CAP_PIN;
 957         if (ci->i_rd_ref)
 958                 used |= CEPH_CAP_FILE_RD;
 959         if (ci->i_rdcache_ref ||
 960             (!S_ISDIR(ci->vfs_inode.i_mode) && /* ignore readdir cache */
 961              ci->vfs_inode.i_data.nrpages))
 962                 used |= CEPH_CAP_FILE_CACHE;
 963         if (ci->i_wr_ref)
 964                 used |= CEPH_CAP_FILE_WR;
 965         if (ci->i_wb_ref || ci->i_wrbuffer_ref)
 966                 used |= CEPH_CAP_FILE_BUFFER;
 967         return used;
 968 }
 969 
 970 /*
 971  * wanted, by virtue of open file modes
 972  */
 973 int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
 974 {
 975         int i, bits = 0;
 976         for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
 977                 if (ci->i_nr_by_mode[i])
 978                         bits |= 1 << i;
 979         }
 980         if (bits == 0)
 981                 return 0;
 982         return ceph_caps_for_mode(bits >> 1);
 983 }
 984 
 985 /*
 986  * Return caps we have registered with the MDS(s) as 'wanted'.
 987  */
 988 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci, bool check)
 989 {
 990         struct ceph_cap *cap;
 991         struct rb_node *p;
 992         int mds_wanted = 0;
 993 
 994         for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
 995                 cap = rb_entry(p, struct ceph_cap, ci_node);
 996                 if (check && !__cap_is_valid(cap))
 997                         continue;
 998                 if (cap == ci->i_auth_cap)
 999                         mds_wanted |= cap->mds_wanted;
1000                 else
1001                         mds_wanted |= (cap->mds_wanted & ~CEPH_CAP_ANY_FILE_WR);
1002         }
1003         return mds_wanted;
1004 }
1005 
1006 /*
1007  * called under i_ceph_lock
1008  */
1009 static int __ceph_is_single_caps(struct ceph_inode_info *ci)
1010 {
1011         return rb_first(&ci->i_caps) == rb_last(&ci->i_caps);
1012 }
1013 
1014 static int __ceph_is_any_caps(struct ceph_inode_info *ci)
1015 {
1016         return !RB_EMPTY_ROOT(&ci->i_caps);
1017 }
1018 
1019 int ceph_is_any_caps(struct inode *inode)
1020 {
1021         struct ceph_inode_info *ci = ceph_inode(inode);
1022         int ret;
1023 
1024         spin_lock(&ci->i_ceph_lock);
1025         ret = __ceph_is_any_caps(ci);
1026         spin_unlock(&ci->i_ceph_lock);
1027 
1028         return ret;
1029 }
1030 
1031 static void drop_inode_snap_realm(struct ceph_inode_info *ci)
1032 {
1033         struct ceph_snap_realm *realm = ci->i_snap_realm;
1034         spin_lock(&realm->inodes_with_caps_lock);
1035         list_del_init(&ci->i_snap_realm_item);
1036         ci->i_snap_realm_counter++;
1037         ci->i_snap_realm = NULL;
1038         if (realm->ino == ci->i_vino.ino)
1039                 realm->inode = NULL;
1040         spin_unlock(&realm->inodes_with_caps_lock);
1041         ceph_put_snap_realm(ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc,
1042                             realm);
1043 }
1044 
1045 /*
1046  * Remove a cap.  Take steps to deal with a racing iterate_session_caps.
1047  *
1048  * caller should hold i_ceph_lock.
1049  * caller will not hold session s_mutex if called from destroy_inode.
1050  */
1051 void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
1052 {
1053         struct ceph_mds_session *session = cap->session;
1054         struct ceph_inode_info *ci = cap->ci;
1055         struct ceph_mds_client *mdsc =
1056                 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1057         int removed = 0;
1058 
1059         dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
1060 
1061         /* remove from inode's cap rbtree, and clear auth cap */
1062         rb_erase(&cap->ci_node, &ci->i_caps);
1063         if (ci->i_auth_cap == cap)
1064                 ci->i_auth_cap = NULL;
1065 
1066         /* remove from session list */
1067         spin_lock(&session->s_cap_lock);
1068         if (session->s_cap_iterator == cap) {
1069                 /* not yet, we are iterating over this very cap */
1070                 dout("__ceph_remove_cap  delaying %p removal from session %p\n",
1071                      cap, cap->session);
1072         } else {
1073                 list_del_init(&cap->session_caps);
1074                 session->s_nr_caps--;
1075                 cap->session = NULL;
1076                 removed = 1;
1077         }
1078         /* protect backpointer with s_cap_lock: see iterate_session_caps */
1079         cap->ci = NULL;
1080 
1081         /*
1082          * s_cap_reconnect is protected by s_cap_lock. no one changes
1083          * s_cap_gen while session is in the reconnect state.
1084          */
1085         if (queue_release &&
1086             (!session->s_cap_reconnect || cap->cap_gen == session->s_cap_gen)) {
1087                 cap->queue_release = 1;
1088                 if (removed) {
1089                         __ceph_queue_cap_release(session, cap);
1090                         removed = 0;
1091                 }
1092         } else {
1093                 cap->queue_release = 0;
1094         }
1095         cap->cap_ino = ci->i_vino.ino;
1096 
1097         spin_unlock(&session->s_cap_lock);
1098 
1099         if (removed)
1100                 ceph_put_cap(mdsc, cap);
1101 
1102         /* when reconnect denied, we remove session caps forcibly,
1103          * i_wr_ref can be non-zero. If there are ongoing write,
1104          * keep i_snap_realm.
1105          */
1106         if (!__ceph_is_any_caps(ci) && ci->i_wr_ref == 0 && ci->i_snap_realm)
1107                 drop_inode_snap_realm(ci);
1108 
1109         if (!__ceph_is_any_real_caps(ci))
1110                 __cap_delay_cancel(mdsc, ci);
1111 }
1112 
1113 struct cap_msg_args {
1114         struct ceph_mds_session *session;
1115         u64                     ino, cid, follows;
1116         u64                     flush_tid, oldest_flush_tid, size, max_size;
1117         u64                     xattr_version;
1118         u64                     change_attr;
1119         struct ceph_buffer      *xattr_buf;
1120         struct timespec64       atime, mtime, ctime, btime;
1121         int                     op, caps, wanted, dirty;
1122         u32                     seq, issue_seq, mseq, time_warp_seq;
1123         u32                     flags;
1124         kuid_t                  uid;
1125         kgid_t                  gid;
1126         umode_t                 mode;
1127         bool                    inline_data;
1128 };
1129 
1130 /*
1131  * Build and send a cap message to the given MDS.
1132  *
1133  * Caller should be holding s_mutex.
1134  */
1135 static int send_cap_msg(struct cap_msg_args *arg)
1136 {
1137         struct ceph_mds_caps *fc;
1138         struct ceph_msg *msg;
1139         void *p;
1140         size_t extra_len;
1141         struct ceph_osd_client *osdc = &arg->session->s_mdsc->fsc->client->osdc;
1142 
1143         dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
1144              " seq %u/%u tid %llu/%llu mseq %u follows %lld size %llu/%llu"
1145              " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(arg->op),
1146              arg->cid, arg->ino, ceph_cap_string(arg->caps),
1147              ceph_cap_string(arg->wanted), ceph_cap_string(arg->dirty),
1148              arg->seq, arg->issue_seq, arg->flush_tid, arg->oldest_flush_tid,
1149              arg->mseq, arg->follows, arg->size, arg->max_size,
1150              arg->xattr_version,
1151              arg->xattr_buf ? (int)arg->xattr_buf->vec.iov_len : 0);
1152 
1153         /* flock buffer size + inline version + inline data size +
1154          * osd_epoch_barrier + oldest_flush_tid */
1155         extra_len = 4 + 8 + 4 + 4 + 8 + 4 + 4 + 4 + 8 + 8 + 4;
1156         msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc) + extra_len,
1157                            GFP_NOFS, false);
1158         if (!msg)
1159                 return -ENOMEM;
1160 
1161         msg->hdr.version = cpu_to_le16(10);
1162         msg->hdr.tid = cpu_to_le64(arg->flush_tid);
1163 
1164         fc = msg->front.iov_base;
1165         memset(fc, 0, sizeof(*fc));
1166 
1167         fc->cap_id = cpu_to_le64(arg->cid);
1168         fc->op = cpu_to_le32(arg->op);
1169         fc->seq = cpu_to_le32(arg->seq);
1170         fc->issue_seq = cpu_to_le32(arg->issue_seq);
1171         fc->migrate_seq = cpu_to_le32(arg->mseq);
1172         fc->caps = cpu_to_le32(arg->caps);
1173         fc->wanted = cpu_to_le32(arg->wanted);
1174         fc->dirty = cpu_to_le32(arg->dirty);
1175         fc->ino = cpu_to_le64(arg->ino);
1176         fc->snap_follows = cpu_to_le64(arg->follows);
1177 
1178         fc->size = cpu_to_le64(arg->size);
1179         fc->max_size = cpu_to_le64(arg->max_size);
1180         ceph_encode_timespec64(&fc->mtime, &arg->mtime);
1181         ceph_encode_timespec64(&fc->atime, &arg->atime);
1182         ceph_encode_timespec64(&fc->ctime, &arg->ctime);
1183         fc->time_warp_seq = cpu_to_le32(arg->time_warp_seq);
1184 
1185         fc->uid = cpu_to_le32(from_kuid(&init_user_ns, arg->uid));
1186         fc->gid = cpu_to_le32(from_kgid(&init_user_ns, arg->gid));
1187         fc->mode = cpu_to_le32(arg->mode);
1188 
1189         fc->xattr_version = cpu_to_le64(arg->xattr_version);
1190         if (arg->xattr_buf) {
1191                 msg->middle = ceph_buffer_get(arg->xattr_buf);
1192                 fc->xattr_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1193                 msg->hdr.middle_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1194         }
1195 
1196         p = fc + 1;
1197         /* flock buffer size (version 2) */
1198         ceph_encode_32(&p, 0);
1199         /* inline version (version 4) */
1200         ceph_encode_64(&p, arg->inline_data ? 0 : CEPH_INLINE_NONE);
1201         /* inline data size */
1202         ceph_encode_32(&p, 0);
1203         /*
1204          * osd_epoch_barrier (version 5)
1205          * The epoch_barrier is protected osdc->lock, so READ_ONCE here in
1206          * case it was recently changed
1207          */
1208         ceph_encode_32(&p, READ_ONCE(osdc->epoch_barrier));
1209         /* oldest_flush_tid (version 6) */
1210         ceph_encode_64(&p, arg->oldest_flush_tid);
1211 
1212         /*
1213          * caller_uid/caller_gid (version 7)
1214          *
1215          * Currently, we don't properly track which caller dirtied the caps
1216          * last, and force a flush of them when there is a conflict. For now,
1217          * just set this to 0:0, to emulate how the MDS has worked up to now.
1218          */
1219         ceph_encode_32(&p, 0);
1220         ceph_encode_32(&p, 0);
1221 
1222         /* pool namespace (version 8) (mds always ignores this) */
1223         ceph_encode_32(&p, 0);
1224 
1225         /* btime and change_attr (version 9) */
1226         ceph_encode_timespec64(p, &arg->btime);
1227         p += sizeof(struct ceph_timespec);
1228         ceph_encode_64(&p, arg->change_attr);
1229 
1230         /* Advisory flags (version 10) */
1231         ceph_encode_32(&p, arg->flags);
1232 
1233         ceph_con_send(&arg->session->s_con, msg);
1234         return 0;
1235 }
1236 
1237 /*
1238  * Queue cap releases when an inode is dropped from our cache.
1239  */
1240 void __ceph_remove_caps(struct ceph_inode_info *ci)
1241 {
1242         struct rb_node *p;
1243 
1244         /* lock i_ceph_lock, because ceph_d_revalidate(..., LOOKUP_RCU)
1245          * may call __ceph_caps_issued_mask() on a freeing inode. */
1246         spin_lock(&ci->i_ceph_lock);
1247         p = rb_first(&ci->i_caps);
1248         while (p) {
1249                 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1250                 p = rb_next(p);
1251                 __ceph_remove_cap(cap, true);
1252         }
1253         spin_unlock(&ci->i_ceph_lock);
1254 }
1255 
1256 /*
1257  * Send a cap msg on the given inode.  Update our caps state, then
1258  * drop i_ceph_lock and send the message.
1259  *
1260  * Make note of max_size reported/requested from mds, revoked caps
1261  * that have now been implemented.
1262  *
1263  * Return non-zero if delayed release, or we experienced an error
1264  * such that the caller should requeue + retry later.
1265  *
1266  * called with i_ceph_lock, then drops it.
1267  * caller should hold snap_rwsem (read), s_mutex.
1268  */
1269 static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1270                       int op, int flags, int used, int want, int retain,
1271                       int flushing, u64 flush_tid, u64 oldest_flush_tid)
1272         __releases(cap->ci->i_ceph_lock)
1273 {
1274         struct ceph_inode_info *ci = cap->ci;
1275         struct inode *inode = &ci->vfs_inode;
1276         struct ceph_buffer *old_blob = NULL;
1277         struct cap_msg_args arg;
1278         int held, revoking;
1279         int wake = 0;
1280         int delayed = 0;
1281         int ret;
1282 
1283         held = cap->issued | cap->implemented;
1284         revoking = cap->implemented & ~cap->issued;
1285         retain &= ~revoking;
1286 
1287         dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n",
1288              inode, cap, cap->session,
1289              ceph_cap_string(held), ceph_cap_string(held & retain),
1290              ceph_cap_string(revoking));
1291         BUG_ON((retain & CEPH_CAP_PIN) == 0);
1292 
1293         arg.session = cap->session;
1294 
1295         /* don't release wanted unless we've waited a bit. */
1296         if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1297             time_before(jiffies, ci->i_hold_caps_min)) {
1298                 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n",
1299                      ceph_cap_string(cap->issued),
1300                      ceph_cap_string(cap->issued & retain),
1301                      ceph_cap_string(cap->mds_wanted),
1302                      ceph_cap_string(want));
1303                 want |= cap->mds_wanted;
1304                 retain |= cap->issued;
1305                 delayed = 1;
1306         }
1307         ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH);
1308         if (want & ~cap->mds_wanted) {
1309                 /* user space may open/close single file frequently.
1310                  * This avoids droping mds_wanted immediately after
1311                  * requesting new mds_wanted.
1312                  */
1313                 __cap_set_timeouts(mdsc, ci);
1314         }
1315 
1316         cap->issued &= retain;  /* drop bits we don't want */
1317         if (cap->implemented & ~cap->issued) {
1318                 /*
1319                  * Wake up any waiters on wanted -> needed transition.
1320                  * This is due to the weird transition from buffered
1321                  * to sync IO... we need to flush dirty pages _before_
1322                  * allowing sync writes to avoid reordering.
1323                  */
1324                 wake = 1;
1325         }
1326         cap->implemented &= cap->issued | used;
1327         cap->mds_wanted = want;
1328 
1329         arg.ino = ceph_vino(inode).ino;
1330         arg.cid = cap->cap_id;
1331         arg.follows = flushing ? ci->i_head_snapc->seq : 0;
1332         arg.flush_tid = flush_tid;
1333         arg.oldest_flush_tid = oldest_flush_tid;
1334 
1335         arg.size = inode->i_size;
1336         ci->i_reported_size = arg.size;
1337         arg.max_size = ci->i_wanted_max_size;
1338         ci->i_requested_max_size = arg.max_size;
1339 
1340         if (flushing & CEPH_CAP_XATTR_EXCL) {
1341                 old_blob = __ceph_build_xattrs_blob(ci);
1342                 arg.xattr_version = ci->i_xattrs.version;
1343                 arg.xattr_buf = ci->i_xattrs.blob;
1344         } else {
1345                 arg.xattr_buf = NULL;
1346         }
1347 
1348         arg.mtime = inode->i_mtime;
1349         arg.atime = inode->i_atime;
1350         arg.ctime = inode->i_ctime;
1351         arg.btime = ci->i_btime;
1352         arg.change_attr = inode_peek_iversion_raw(inode);
1353 
1354         arg.op = op;
1355         arg.caps = cap->implemented;
1356         arg.wanted = want;
1357         arg.dirty = flushing;
1358 
1359         arg.seq = cap->seq;
1360         arg.issue_seq = cap->issue_seq;
1361         arg.mseq = cap->mseq;
1362         arg.time_warp_seq = ci->i_time_warp_seq;
1363 
1364         arg.uid = inode->i_uid;
1365         arg.gid = inode->i_gid;
1366         arg.mode = inode->i_mode;
1367 
1368         arg.inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
1369         if (!(flags & CEPH_CLIENT_CAPS_PENDING_CAPSNAP) &&
1370             !list_empty(&ci->i_cap_snaps)) {
1371                 struct ceph_cap_snap *capsnap;
1372                 list_for_each_entry_reverse(capsnap, &ci->i_cap_snaps, ci_item) {
1373                         if (capsnap->cap_flush.tid)
1374                                 break;
1375                         if (capsnap->need_flush) {
1376                                 flags |= CEPH_CLIENT_CAPS_PENDING_CAPSNAP;
1377                                 break;
1378                         }
1379                 }
1380         }
1381         arg.flags = flags;
1382 
1383         spin_unlock(&ci->i_ceph_lock);
1384 
1385         ceph_buffer_put(old_blob);
1386 
1387         ret = send_cap_msg(&arg);
1388         if (ret < 0) {
1389                 dout("error sending cap msg, must requeue %p\n", inode);
1390                 delayed = 1;
1391         }
1392 
1393         if (wake)
1394                 wake_up_all(&ci->i_cap_wq);
1395 
1396         return delayed;
1397 }
1398 
1399 static inline int __send_flush_snap(struct inode *inode,
1400                                     struct ceph_mds_session *session,
1401                                     struct ceph_cap_snap *capsnap,
1402                                     u32 mseq, u64 oldest_flush_tid)
1403 {
1404         struct cap_msg_args     arg;
1405 
1406         arg.session = session;
1407         arg.ino = ceph_vino(inode).ino;
1408         arg.cid = 0;
1409         arg.follows = capsnap->follows;
1410         arg.flush_tid = capsnap->cap_flush.tid;
1411         arg.oldest_flush_tid = oldest_flush_tid;
1412 
1413         arg.size = capsnap->size;
1414         arg.max_size = 0;
1415         arg.xattr_version = capsnap->xattr_version;
1416         arg.xattr_buf = capsnap->xattr_blob;
1417 
1418         arg.atime = capsnap->atime;
1419         arg.mtime = capsnap->mtime;
1420         arg.ctime = capsnap->ctime;
1421         arg.btime = capsnap->btime;
1422         arg.change_attr = capsnap->change_attr;
1423 
1424         arg.op = CEPH_CAP_OP_FLUSHSNAP;
1425         arg.caps = capsnap->issued;
1426         arg.wanted = 0;
1427         arg.dirty = capsnap->dirty;
1428 
1429         arg.seq = 0;
1430         arg.issue_seq = 0;
1431         arg.mseq = mseq;
1432         arg.time_warp_seq = capsnap->time_warp_seq;
1433 
1434         arg.uid = capsnap->uid;
1435         arg.gid = capsnap->gid;
1436         arg.mode = capsnap->mode;
1437 
1438         arg.inline_data = capsnap->inline_data;
1439         arg.flags = 0;
1440 
1441         return send_cap_msg(&arg);
1442 }
1443 
1444 /*
1445  * When a snapshot is taken, clients accumulate dirty metadata on
1446  * inodes with capabilities in ceph_cap_snaps to describe the file
1447  * state at the time the snapshot was taken.  This must be flushed
1448  * asynchronously back to the MDS once sync writes complete and dirty
1449  * data is written out.
1450  *
1451  * Called under i_ceph_lock.  Takes s_mutex as needed.
1452  */
1453 static void __ceph_flush_snaps(struct ceph_inode_info *ci,
1454                                struct ceph_mds_session *session)
1455                 __releases(ci->i_ceph_lock)
1456                 __acquires(ci->i_ceph_lock)
1457 {
1458         struct inode *inode = &ci->vfs_inode;
1459         struct ceph_mds_client *mdsc = session->s_mdsc;
1460         struct ceph_cap_snap *capsnap;
1461         u64 oldest_flush_tid = 0;
1462         u64 first_tid = 1, last_tid = 0;
1463 
1464         dout("__flush_snaps %p session %p\n", inode, session);
1465 
1466         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1467                 /*
1468                  * we need to wait for sync writes to complete and for dirty
1469                  * pages to be written out.
1470                  */
1471                 if (capsnap->dirty_pages || capsnap->writing)
1472                         break;
1473 
1474                 /* should be removed by ceph_try_drop_cap_snap() */
1475                 BUG_ON(!capsnap->need_flush);
1476 
1477                 /* only flush each capsnap once */
1478                 if (capsnap->cap_flush.tid > 0) {
1479                         dout(" already flushed %p, skipping\n", capsnap);
1480                         continue;
1481                 }
1482 
1483                 spin_lock(&mdsc->cap_dirty_lock);
1484                 capsnap->cap_flush.tid = ++mdsc->last_cap_flush_tid;
1485                 list_add_tail(&capsnap->cap_flush.g_list,
1486                               &mdsc->cap_flush_list);
1487                 if (oldest_flush_tid == 0)
1488                         oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1489                 if (list_empty(&ci->i_flushing_item)) {
1490                         list_add_tail(&ci->i_flushing_item,
1491                                       &session->s_cap_flushing);
1492                 }
1493                 spin_unlock(&mdsc->cap_dirty_lock);
1494 
1495                 list_add_tail(&capsnap->cap_flush.i_list,
1496                               &ci->i_cap_flush_list);
1497 
1498                 if (first_tid == 1)
1499                         first_tid = capsnap->cap_flush.tid;
1500                 last_tid = capsnap->cap_flush.tid;
1501         }
1502 
1503         ci->i_ceph_flags &= ~CEPH_I_FLUSH_SNAPS;
1504 
1505         while (first_tid <= last_tid) {
1506                 struct ceph_cap *cap = ci->i_auth_cap;
1507                 struct ceph_cap_flush *cf;
1508                 int ret;
1509 
1510                 if (!(cap && cap->session == session)) {
1511                         dout("__flush_snaps %p auth cap %p not mds%d, "
1512                              "stop\n", inode, cap, session->s_mds);
1513                         break;
1514                 }
1515 
1516                 ret = -ENOENT;
1517                 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
1518                         if (cf->tid >= first_tid) {
1519                                 ret = 0;
1520                                 break;
1521                         }
1522                 }
1523                 if (ret < 0)
1524                         break;
1525 
1526                 first_tid = cf->tid + 1;
1527 
1528                 capsnap = container_of(cf, struct ceph_cap_snap, cap_flush);
1529                 refcount_inc(&capsnap->nref);
1530                 spin_unlock(&ci->i_ceph_lock);
1531 
1532                 dout("__flush_snaps %p capsnap %p tid %llu %s\n",
1533                      inode, capsnap, cf->tid, ceph_cap_string(capsnap->dirty));
1534 
1535                 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
1536                                         oldest_flush_tid);
1537                 if (ret < 0) {
1538                         pr_err("__flush_snaps: error sending cap flushsnap, "
1539                                "ino (%llx.%llx) tid %llu follows %llu\n",
1540                                 ceph_vinop(inode), cf->tid, capsnap->follows);
1541                 }
1542 
1543                 ceph_put_cap_snap(capsnap);
1544                 spin_lock(&ci->i_ceph_lock);
1545         }
1546 }
1547 
1548 void ceph_flush_snaps(struct ceph_inode_info *ci,
1549                       struct ceph_mds_session **psession)
1550 {
1551         struct inode *inode = &ci->vfs_inode;
1552         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1553         struct ceph_mds_session *session = NULL;
1554         int mds;
1555 
1556         dout("ceph_flush_snaps %p\n", inode);
1557         if (psession)
1558                 session = *psession;
1559 retry:
1560         spin_lock(&ci->i_ceph_lock);
1561         if (!(ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)) {
1562                 dout(" no capsnap needs flush, doing nothing\n");
1563                 goto out;
1564         }
1565         if (!ci->i_auth_cap) {
1566                 dout(" no auth cap (migrating?), doing nothing\n");
1567                 goto out;
1568         }
1569 
1570         mds = ci->i_auth_cap->session->s_mds;
1571         if (session && session->s_mds != mds) {
1572                 dout(" oops, wrong session %p mutex\n", session);
1573                 mutex_unlock(&session->s_mutex);
1574                 ceph_put_mds_session(session);
1575                 session = NULL;
1576         }
1577         if (!session) {
1578                 spin_unlock(&ci->i_ceph_lock);
1579                 mutex_lock(&mdsc->mutex);
1580                 session = __ceph_lookup_mds_session(mdsc, mds);
1581                 mutex_unlock(&mdsc->mutex);
1582                 if (session) {
1583                         dout(" inverting session/ino locks on %p\n", session);
1584                         mutex_lock(&session->s_mutex);
1585                 }
1586                 goto retry;
1587         }
1588 
1589         // make sure flushsnap messages are sent in proper order.
1590         if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
1591                 __kick_flushing_caps(mdsc, session, ci, 0);
1592 
1593         __ceph_flush_snaps(ci, session);
1594 out:
1595         spin_unlock(&ci->i_ceph_lock);
1596 
1597         if (psession) {
1598                 *psession = session;
1599         } else if (session) {
1600                 mutex_unlock(&session->s_mutex);
1601                 ceph_put_mds_session(session);
1602         }
1603         /* we flushed them all; remove this inode from the queue */
1604         spin_lock(&mdsc->snap_flush_lock);
1605         list_del_init(&ci->i_snap_flush_item);
1606         spin_unlock(&mdsc->snap_flush_lock);
1607 }
1608 
1609 /*
1610  * Mark caps dirty.  If inode is newly dirty, return the dirty flags.
1611  * Caller is then responsible for calling __mark_inode_dirty with the
1612  * returned flags value.
1613  */
1614 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
1615                            struct ceph_cap_flush **pcf)
1616 {
1617         struct ceph_mds_client *mdsc =
1618                 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1619         struct inode *inode = &ci->vfs_inode;
1620         int was = ci->i_dirty_caps;
1621         int dirty = 0;
1622 
1623         if (!ci->i_auth_cap) {
1624                 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1625                         "but no auth cap (session was closed?)\n",
1626                         inode, ceph_ino(inode), ceph_cap_string(mask));
1627                 return 0;
1628         }
1629 
1630         dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1631              ceph_cap_string(mask), ceph_cap_string(was),
1632              ceph_cap_string(was | mask));
1633         ci->i_dirty_caps |= mask;
1634         if (was == 0) {
1635                 WARN_ON_ONCE(ci->i_prealloc_cap_flush);
1636                 swap(ci->i_prealloc_cap_flush, *pcf);
1637 
1638                 if (!ci->i_head_snapc) {
1639                         WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1640                         ci->i_head_snapc = ceph_get_snap_context(
1641                                 ci->i_snap_realm->cached_context);
1642                 }
1643                 dout(" inode %p now dirty snapc %p auth cap %p\n",
1644                      &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1645                 BUG_ON(!list_empty(&ci->i_dirty_item));
1646                 spin_lock(&mdsc->cap_dirty_lock);
1647                 list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
1648                 spin_unlock(&mdsc->cap_dirty_lock);
1649                 if (ci->i_flushing_caps == 0) {
1650                         ihold(inode);
1651                         dirty |= I_DIRTY_SYNC;
1652                 }
1653         } else {
1654                 WARN_ON_ONCE(!ci->i_prealloc_cap_flush);
1655         }
1656         BUG_ON(list_empty(&ci->i_dirty_item));
1657         if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1658             (mask & CEPH_CAP_FILE_BUFFER))
1659                 dirty |= I_DIRTY_DATASYNC;
1660         __cap_delay_requeue(mdsc, ci, true);
1661         return dirty;
1662 }
1663 
1664 struct ceph_cap_flush *ceph_alloc_cap_flush(void)
1665 {
1666         return kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1667 }
1668 
1669 void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1670 {
1671         if (cf)
1672                 kmem_cache_free(ceph_cap_flush_cachep, cf);
1673 }
1674 
1675 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1676 {
1677         if (!list_empty(&mdsc->cap_flush_list)) {
1678                 struct ceph_cap_flush *cf =
1679                         list_first_entry(&mdsc->cap_flush_list,
1680                                          struct ceph_cap_flush, g_list);
1681                 return cf->tid;
1682         }
1683         return 0;
1684 }
1685 
1686 /*
1687  * Remove cap_flush from the mdsc's or inode's flushing cap list.
1688  * Return true if caller needs to wake up flush waiters.
1689  */
1690 static bool __finish_cap_flush(struct ceph_mds_client *mdsc,
1691                                struct ceph_inode_info *ci,
1692                                struct ceph_cap_flush *cf)
1693 {
1694         struct ceph_cap_flush *prev;
1695         bool wake = cf->wake;
1696         if (mdsc) {
1697                 /* are there older pending cap flushes? */
1698                 if (wake && cf->g_list.prev != &mdsc->cap_flush_list) {
1699                         prev = list_prev_entry(cf, g_list);
1700                         prev->wake = true;
1701                         wake = false;
1702                 }
1703                 list_del(&cf->g_list);
1704         } else if (ci) {
1705                 if (wake && cf->i_list.prev != &ci->i_cap_flush_list) {
1706                         prev = list_prev_entry(cf, i_list);
1707                         prev->wake = true;
1708                         wake = false;
1709                 }
1710                 list_del(&cf->i_list);
1711         } else {
1712                 BUG_ON(1);
1713         }
1714         return wake;
1715 }
1716 
1717 /*
1718  * Add dirty inode to the flushing list.  Assigned a seq number so we
1719  * can wait for caps to flush without starving.
1720  *
1721  * Called under i_ceph_lock. Returns the flush tid.
1722  */
1723 static u64 __mark_caps_flushing(struct inode *inode,
1724                                 struct ceph_mds_session *session, bool wake,
1725                                 u64 *oldest_flush_tid)
1726 {
1727         struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1728         struct ceph_inode_info *ci = ceph_inode(inode);
1729         struct ceph_cap_flush *cf = NULL;
1730         int flushing;
1731 
1732         BUG_ON(ci->i_dirty_caps == 0);
1733         BUG_ON(list_empty(&ci->i_dirty_item));
1734         BUG_ON(!ci->i_prealloc_cap_flush);
1735 
1736         flushing = ci->i_dirty_caps;
1737         dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1738              ceph_cap_string(flushing),
1739              ceph_cap_string(ci->i_flushing_caps),
1740              ceph_cap_string(ci->i_flushing_caps | flushing));
1741         ci->i_flushing_caps |= flushing;
1742         ci->i_dirty_caps = 0;
1743         dout(" inode %p now !dirty\n", inode);
1744 
1745         swap(cf, ci->i_prealloc_cap_flush);
1746         cf->caps = flushing;
1747         cf->wake = wake;
1748 
1749         spin_lock(&mdsc->cap_dirty_lock);
1750         list_del_init(&ci->i_dirty_item);
1751 
1752         cf->tid = ++mdsc->last_cap_flush_tid;
1753         list_add_tail(&cf->g_list, &mdsc->cap_flush_list);
1754         *oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1755 
1756         if (list_empty(&ci->i_flushing_item)) {
1757                 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1758                 mdsc->num_cap_flushing++;
1759         }
1760         spin_unlock(&mdsc->cap_dirty_lock);
1761 
1762         list_add_tail(&cf->i_list, &ci->i_cap_flush_list);
1763 
1764         return cf->tid;
1765 }
1766 
1767 /*
1768  * try to invalidate mapping pages without blocking.
1769  */
1770 static int try_nonblocking_invalidate(struct inode *inode)
1771 {
1772         struct ceph_inode_info *ci = ceph_inode(inode);
1773         u32 invalidating_gen = ci->i_rdcache_gen;
1774 
1775         spin_unlock(&ci->i_ceph_lock);
1776         invalidate_mapping_pages(&inode->i_data, 0, -1);
1777         spin_lock(&ci->i_ceph_lock);
1778 
1779         if (inode->i_data.nrpages == 0 &&
1780             invalidating_gen == ci->i_rdcache_gen) {
1781                 /* success. */
1782                 dout("try_nonblocking_invalidate %p success\n", inode);
1783                 /* save any racing async invalidate some trouble */
1784                 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1785                 return 0;
1786         }
1787         dout("try_nonblocking_invalidate %p failed\n", inode);
1788         return -1;
1789 }
1790 
1791 bool __ceph_should_report_size(struct ceph_inode_info *ci)
1792 {
1793         loff_t size = ci->vfs_inode.i_size;
1794         /* mds will adjust max size according to the reported size */
1795         if (ci->i_flushing_caps & CEPH_CAP_FILE_WR)
1796                 return false;
1797         if (size >= ci->i_max_size)
1798                 return true;
1799         /* half of previous max_size increment has been used */
1800         if (ci->i_max_size > ci->i_reported_size &&
1801             (size << 1) >= ci->i_max_size + ci->i_reported_size)
1802                 return true;
1803         return false;
1804 }
1805 
1806 /*
1807  * Swiss army knife function to examine currently used and wanted
1808  * versus held caps.  Release, flush, ack revoked caps to mds as
1809  * appropriate.
1810  *
1811  *  CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1812  *    cap release further.
1813  *  CHECK_CAPS_AUTHONLY - we should only check the auth cap
1814  *  CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1815  *    further delay.
1816  */
1817 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1818                      struct ceph_mds_session *session)
1819 {
1820         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1821         struct ceph_mds_client *mdsc = fsc->mdsc;
1822         struct inode *inode = &ci->vfs_inode;
1823         struct ceph_cap *cap;
1824         u64 flush_tid, oldest_flush_tid;
1825         int file_wanted, used, cap_used;
1826         int took_snap_rwsem = 0;             /* true if mdsc->snap_rwsem held */
1827         int issued, implemented, want, retain, revoking, flushing = 0;
1828         int mds = -1;   /* keep track of how far we've gone through i_caps list
1829                            to avoid an infinite loop on retry */
1830         struct rb_node *p;
1831         int delayed = 0, sent = 0;
1832         bool no_delay = flags & CHECK_CAPS_NODELAY;
1833         bool queue_invalidate = false;
1834         bool tried_invalidate = false;
1835 
1836         /* if we are unmounting, flush any unused caps immediately. */
1837         if (mdsc->stopping)
1838                 no_delay = true;
1839 
1840         spin_lock(&ci->i_ceph_lock);
1841 
1842         if (ci->i_ceph_flags & CEPH_I_FLUSH)
1843                 flags |= CHECK_CAPS_FLUSH;
1844 
1845         if (!(flags & CHECK_CAPS_AUTHONLY) ||
1846             (ci->i_auth_cap && __ceph_is_single_caps(ci)))
1847                 __cap_delay_cancel(mdsc, ci);
1848 
1849         goto retry_locked;
1850 retry:
1851         spin_lock(&ci->i_ceph_lock);
1852 retry_locked:
1853         file_wanted = __ceph_caps_file_wanted(ci);
1854         used = __ceph_caps_used(ci);
1855         issued = __ceph_caps_issued(ci, &implemented);
1856         revoking = implemented & ~issued;
1857 
1858         want = file_wanted;
1859         retain = file_wanted | used | CEPH_CAP_PIN;
1860         if (!mdsc->stopping && inode->i_nlink > 0) {
1861                 if (file_wanted) {
1862                         retain |= CEPH_CAP_ANY;       /* be greedy */
1863                 } else if (S_ISDIR(inode->i_mode) &&
1864                            (issued & CEPH_CAP_FILE_SHARED) &&
1865                            __ceph_dir_is_complete(ci)) {
1866                         /*
1867                          * If a directory is complete, we want to keep
1868                          * the exclusive cap. So that MDS does not end up
1869                          * revoking the shared cap on every create/unlink
1870                          * operation.
1871                          */
1872                         if (IS_RDONLY(inode))
1873                                 want = CEPH_CAP_ANY_SHARED;
1874                         else
1875                                 want = CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1876                         retain |= want;
1877                 } else {
1878 
1879                         retain |= CEPH_CAP_ANY_SHARED;
1880                         /*
1881                          * keep RD only if we didn't have the file open RW,
1882                          * because then the mds would revoke it anyway to
1883                          * journal max_size=0.
1884                          */
1885                         if (ci->i_max_size == 0)
1886                                 retain |= CEPH_CAP_ANY_RD;
1887                 }
1888         }
1889 
1890         dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1891              " issued %s revoking %s retain %s %s%s%s\n", inode,
1892              ceph_cap_string(file_wanted),
1893              ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1894              ceph_cap_string(ci->i_flushing_caps),
1895              ceph_cap_string(issued), ceph_cap_string(revoking),
1896              ceph_cap_string(retain),
1897              (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1898              (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "",
1899              (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1900 
1901         /*
1902          * If we no longer need to hold onto old our caps, and we may
1903          * have cached pages, but don't want them, then try to invalidate.
1904          * If we fail, it's because pages are locked.... try again later.
1905          */
1906         if ((!no_delay || mdsc->stopping) &&
1907             !S_ISDIR(inode->i_mode) &&          /* ignore readdir cache */
1908             !(ci->i_wb_ref || ci->i_wrbuffer_ref) &&   /* no dirty pages... */
1909             inode->i_data.nrpages &&            /* have cached pages */
1910             (revoking & (CEPH_CAP_FILE_CACHE|
1911                          CEPH_CAP_FILE_LAZYIO)) && /*  or revoking cache */
1912             !tried_invalidate) {
1913                 dout("check_caps trying to invalidate on %p\n", inode);
1914                 if (try_nonblocking_invalidate(inode) < 0) {
1915                         dout("check_caps queuing invalidate\n");
1916                         queue_invalidate = true;
1917                         ci->i_rdcache_revoking = ci->i_rdcache_gen;
1918                 }
1919                 tried_invalidate = true;
1920                 goto retry_locked;
1921         }
1922 
1923         for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1924                 cap = rb_entry(p, struct ceph_cap, ci_node);
1925 
1926                 /* avoid looping forever */
1927                 if (mds >= cap->mds ||
1928                     ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
1929                         continue;
1930 
1931                 /* NOTE: no side-effects allowed, until we take s_mutex */
1932 
1933                 cap_used = used;
1934                 if (ci->i_auth_cap && cap != ci->i_auth_cap)
1935                         cap_used &= ~ci->i_auth_cap->issued;
1936 
1937                 revoking = cap->implemented & ~cap->issued;
1938                 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
1939                      cap->mds, cap, ceph_cap_string(cap_used),
1940                      ceph_cap_string(cap->issued),
1941                      ceph_cap_string(cap->implemented),
1942                      ceph_cap_string(revoking));
1943 
1944                 if (cap == ci->i_auth_cap &&
1945                     (cap->issued & CEPH_CAP_FILE_WR)) {
1946                         /* request larger max_size from MDS? */
1947                         if (ci->i_wanted_max_size > ci->i_max_size &&
1948                             ci->i_wanted_max_size > ci->i_requested_max_size) {
1949                                 dout("requesting new max_size\n");
1950                                 goto ack;
1951                         }
1952 
1953                         /* approaching file_max? */
1954                         if (__ceph_should_report_size(ci)) {
1955                                 dout("i_size approaching max_size\n");
1956                                 goto ack;
1957                         }
1958                 }
1959                 /* flush anything dirty? */
1960                 if (cap == ci->i_auth_cap) {
1961                         if ((flags & CHECK_CAPS_FLUSH) && ci->i_dirty_caps) {
1962                                 dout("flushing dirty caps\n");
1963                                 goto ack;
1964                         }
1965                         if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) {
1966                                 dout("flushing snap caps\n");
1967                                 goto ack;
1968                         }
1969                 }
1970 
1971                 /* completed revocation? going down and there are no caps? */
1972                 if (revoking && (revoking & cap_used) == 0) {
1973                         dout("completed revocation of %s\n",
1974                              ceph_cap_string(cap->implemented & ~cap->issued));
1975                         goto ack;
1976                 }
1977 
1978                 /* want more caps from mds? */
1979                 if (want & ~cap->mds_wanted) {
1980                         if (want & ~(cap->mds_wanted | cap->issued))
1981                                 goto ack;
1982                         if (!__cap_is_valid(cap))
1983                                 goto ack;
1984                 }
1985 
1986                 /* things we might delay */
1987                 if ((cap->issued & ~retain) == 0)
1988                         continue;     /* nope, all good */
1989 
1990                 if (no_delay)
1991                         goto ack;
1992 
1993                 /* delay? */
1994                 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1995                     time_before(jiffies, ci->i_hold_caps_max)) {
1996                         dout(" delaying issued %s -> %s, wanted %s -> %s\n",
1997                              ceph_cap_string(cap->issued),
1998                              ceph_cap_string(cap->issued & retain),
1999                              ceph_cap_string(cap->mds_wanted),
2000                              ceph_cap_string(want));
2001                         delayed++;
2002                         continue;
2003                 }
2004 
2005 ack:
2006                 if (session && session != cap->session) {
2007                         dout("oops, wrong session %p mutex\n", session);
2008                         mutex_unlock(&session->s_mutex);
2009                         session = NULL;
2010                 }
2011                 if (!session) {
2012                         session = cap->session;
2013                         if (mutex_trylock(&session->s_mutex) == 0) {
2014                                 dout("inverting session/ino locks on %p\n",
2015                                      session);
2016                                 spin_unlock(&ci->i_ceph_lock);
2017                                 if (took_snap_rwsem) {
2018                                         up_read(&mdsc->snap_rwsem);
2019                                         took_snap_rwsem = 0;
2020                                 }
2021                                 mutex_lock(&session->s_mutex);
2022                                 goto retry;
2023                         }
2024                 }
2025 
2026                 /* kick flushing and flush snaps before sending normal
2027                  * cap message */
2028                 if (cap == ci->i_auth_cap &&
2029                     (ci->i_ceph_flags &
2030                      (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS))) {
2031                         if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2032                                 __kick_flushing_caps(mdsc, session, ci, 0);
2033                         if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2034                                 __ceph_flush_snaps(ci, session);
2035 
2036                         goto retry_locked;
2037                 }
2038 
2039                 /* take snap_rwsem after session mutex */
2040                 if (!took_snap_rwsem) {
2041                         if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
2042                                 dout("inverting snap/in locks on %p\n",
2043                                      inode);
2044                                 spin_unlock(&ci->i_ceph_lock);
2045                                 down_read(&mdsc->snap_rwsem);
2046                                 took_snap_rwsem = 1;
2047                                 goto retry;
2048                         }
2049                         took_snap_rwsem = 1;
2050                 }
2051 
2052                 if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
2053                         flushing = ci->i_dirty_caps;
2054                         flush_tid = __mark_caps_flushing(inode, session, false,
2055                                                          &oldest_flush_tid);
2056                 } else {
2057                         flushing = 0;
2058                         flush_tid = 0;
2059                         spin_lock(&mdsc->cap_dirty_lock);
2060                         oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2061                         spin_unlock(&mdsc->cap_dirty_lock);
2062                 }
2063 
2064                 mds = cap->mds;  /* remember mds, so we don't repeat */
2065                 sent++;
2066 
2067                 /* __send_cap drops i_ceph_lock */
2068                 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, 0,
2069                                 cap_used, want, retain, flushing,
2070                                 flush_tid, oldest_flush_tid);
2071                 goto retry; /* retake i_ceph_lock and restart our cap scan. */
2072         }
2073 
2074         /* Reschedule delayed caps release if we delayed anything */
2075         if (delayed)
2076                 __cap_delay_requeue(mdsc, ci, false);
2077 
2078         spin_unlock(&ci->i_ceph_lock);
2079 
2080         if (queue_invalidate)
2081                 ceph_queue_invalidate(inode);
2082 
2083         if (session)
2084                 mutex_unlock(&session->s_mutex);
2085         if (took_snap_rwsem)
2086                 up_read(&mdsc->snap_rwsem);
2087 }
2088 
2089 /*
2090  * Try to flush dirty caps back to the auth mds.
2091  */
2092 static int try_flush_caps(struct inode *inode, u64 *ptid)
2093 {
2094         struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2095         struct ceph_inode_info *ci = ceph_inode(inode);
2096         struct ceph_mds_session *session = NULL;
2097         int flushing = 0;
2098         u64 flush_tid = 0, oldest_flush_tid = 0;
2099 
2100 retry:
2101         spin_lock(&ci->i_ceph_lock);
2102 retry_locked:
2103         if (ci->i_dirty_caps && ci->i_auth_cap) {
2104                 struct ceph_cap *cap = ci->i_auth_cap;
2105                 int delayed;
2106 
2107                 if (session != cap->session) {
2108                         spin_unlock(&ci->i_ceph_lock);
2109                         if (session)
2110                                 mutex_unlock(&session->s_mutex);
2111                         session = cap->session;
2112                         mutex_lock(&session->s_mutex);
2113                         goto retry;
2114                 }
2115                 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN) {
2116                         spin_unlock(&ci->i_ceph_lock);
2117                         goto out;
2118                 }
2119 
2120                 if (ci->i_ceph_flags &
2121                     (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS)) {
2122                         if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2123                                 __kick_flushing_caps(mdsc, session, ci, 0);
2124                         if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2125                                 __ceph_flush_snaps(ci, session);
2126                         goto retry_locked;
2127                 }
2128 
2129                 flushing = ci->i_dirty_caps;
2130                 flush_tid = __mark_caps_flushing(inode, session, true,
2131                                                  &oldest_flush_tid);
2132 
2133                 /* __send_cap drops i_ceph_lock */
2134                 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
2135                                      CEPH_CLIENT_CAPS_SYNC,
2136                                      __ceph_caps_used(ci),
2137                                      __ceph_caps_wanted(ci),
2138                                      (cap->issued | cap->implemented),
2139                                      flushing, flush_tid, oldest_flush_tid);
2140 
2141                 if (delayed) {
2142                         spin_lock(&ci->i_ceph_lock);
2143                         __cap_delay_requeue(mdsc, ci, true);
2144                         spin_unlock(&ci->i_ceph_lock);
2145                 }
2146         } else {
2147                 if (!list_empty(&ci->i_cap_flush_list)) {
2148                         struct ceph_cap_flush *cf =
2149                                 list_last_entry(&ci->i_cap_flush_list,
2150                                                 struct ceph_cap_flush, i_list);
2151                         cf->wake = true;
2152                         flush_tid = cf->tid;
2153                 }
2154                 flushing = ci->i_flushing_caps;
2155                 spin_unlock(&ci->i_ceph_lock);
2156         }
2157 out:
2158         if (session)
2159                 mutex_unlock(&session->s_mutex);
2160 
2161         *ptid = flush_tid;
2162         return flushing;
2163 }
2164 
2165 /*
2166  * Return true if we've flushed caps through the given flush_tid.
2167  */
2168 static int caps_are_flushed(struct inode *inode, u64 flush_tid)
2169 {
2170         struct ceph_inode_info *ci = ceph_inode(inode);
2171         int ret = 1;
2172 
2173         spin_lock(&ci->i_ceph_lock);
2174         if (!list_empty(&ci->i_cap_flush_list)) {
2175                 struct ceph_cap_flush * cf =
2176                         list_first_entry(&ci->i_cap_flush_list,
2177                                          struct ceph_cap_flush, i_list);
2178                 if (cf->tid <= flush_tid)
2179                         ret = 0;
2180         }
2181         spin_unlock(&ci->i_ceph_lock);
2182         return ret;
2183 }
2184 
2185 /*
2186  * wait for any unsafe requests to complete.
2187  */
2188 static int unsafe_request_wait(struct inode *inode)
2189 {
2190         struct ceph_inode_info *ci = ceph_inode(inode);
2191         struct ceph_mds_request *req1 = NULL, *req2 = NULL;
2192         int ret, err = 0;
2193 
2194         spin_lock(&ci->i_unsafe_lock);
2195         if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) {
2196                 req1 = list_last_entry(&ci->i_unsafe_dirops,
2197                                         struct ceph_mds_request,
2198                                         r_unsafe_dir_item);
2199                 ceph_mdsc_get_request(req1);
2200         }
2201         if (!list_empty(&ci->i_unsafe_iops)) {
2202                 req2 = list_last_entry(&ci->i_unsafe_iops,
2203                                         struct ceph_mds_request,
2204                                         r_unsafe_target_item);
2205                 ceph_mdsc_get_request(req2);
2206         }
2207         spin_unlock(&ci->i_unsafe_lock);
2208 
2209         dout("unsafe_request_wait %p wait on tid %llu %llu\n",
2210              inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL);
2211         if (req1) {
2212                 ret = !wait_for_completion_timeout(&req1->r_safe_completion,
2213                                         ceph_timeout_jiffies(req1->r_timeout));
2214                 if (ret)
2215                         err = -EIO;
2216                 ceph_mdsc_put_request(req1);
2217         }
2218         if (req2) {
2219                 ret = !wait_for_completion_timeout(&req2->r_safe_completion,
2220                                         ceph_timeout_jiffies(req2->r_timeout));
2221                 if (ret)
2222                         err = -EIO;
2223                 ceph_mdsc_put_request(req2);
2224         }
2225         return err;
2226 }
2227 
2228 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2229 {
2230         struct ceph_file_info *fi = file->private_data;
2231         struct inode *inode = file->f_mapping->host;
2232         struct ceph_inode_info *ci = ceph_inode(inode);
2233         u64 flush_tid;
2234         int ret, err;
2235         int dirty;
2236 
2237         dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
2238 
2239         ret = file_write_and_wait_range(file, start, end);
2240         if (datasync)
2241                 goto out;
2242 
2243         dirty = try_flush_caps(inode, &flush_tid);
2244         dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
2245 
2246         err = unsafe_request_wait(inode);
2247 
2248         /*
2249          * only wait on non-file metadata writeback (the mds
2250          * can recover size and mtime, so we don't need to
2251          * wait for that)
2252          */
2253         if (!err && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
2254                 err = wait_event_interruptible(ci->i_cap_wq,
2255                                         caps_are_flushed(inode, flush_tid));
2256         }
2257 
2258         if (err < 0)
2259                 ret = err;
2260 
2261         if (errseq_check(&ci->i_meta_err, READ_ONCE(fi->meta_err))) {
2262                 spin_lock(&file->f_lock);
2263                 err = errseq_check_and_advance(&ci->i_meta_err,
2264                                                &fi->meta_err);
2265                 spin_unlock(&file->f_lock);
2266                 if (err < 0)
2267                         ret = err;
2268         }
2269 out:
2270         dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
2271         return ret;
2272 }
2273 
2274 /*
2275  * Flush any dirty caps back to the mds.  If we aren't asked to wait,
2276  * queue inode for flush but don't do so immediately, because we can
2277  * get by with fewer MDS messages if we wait for data writeback to
2278  * complete first.
2279  */
2280 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
2281 {
2282         struct ceph_inode_info *ci = ceph_inode(inode);
2283         u64 flush_tid;
2284         int err = 0;
2285         int dirty;
2286         int wait = (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync);
2287 
2288         dout("write_inode %p wait=%d\n", inode, wait);
2289         if (wait) {
2290                 dirty = try_flush_caps(inode, &flush_tid);
2291                 if (dirty)
2292                         err = wait_event_interruptible(ci->i_cap_wq,
2293                                        caps_are_flushed(inode, flush_tid));
2294         } else {
2295                 struct ceph_mds_client *mdsc =
2296                         ceph_sb_to_client(inode->i_sb)->mdsc;
2297 
2298                 spin_lock(&ci->i_ceph_lock);
2299                 if (__ceph_caps_dirty(ci))
2300                         __cap_delay_requeue_front(mdsc, ci);
2301                 spin_unlock(&ci->i_ceph_lock);
2302         }
2303         return err;
2304 }
2305 
2306 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
2307                                  struct ceph_mds_session *session,
2308                                  struct ceph_inode_info *ci,
2309                                  u64 oldest_flush_tid)
2310         __releases(ci->i_ceph_lock)
2311         __acquires(ci->i_ceph_lock)
2312 {
2313         struct inode *inode = &ci->vfs_inode;
2314         struct ceph_cap *cap;
2315         struct ceph_cap_flush *cf;
2316         int ret;
2317         u64 first_tid = 0;
2318         u64 last_snap_flush = 0;
2319 
2320         ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2321 
2322         list_for_each_entry_reverse(cf, &ci->i_cap_flush_list, i_list) {
2323                 if (!cf->caps) {
2324                         last_snap_flush = cf->tid;
2325                         break;
2326                 }
2327         }
2328 
2329         list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
2330                 if (cf->tid < first_tid)
2331                         continue;
2332 
2333                 cap = ci->i_auth_cap;
2334                 if (!(cap && cap->session == session)) {
2335                         pr_err("%p auth cap %p not mds%d ???\n",
2336                                inode, cap, session->s_mds);
2337                         break;
2338                 }
2339 
2340                 first_tid = cf->tid + 1;
2341 
2342                 if (cf->caps) {
2343                         dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2344                              inode, cap, cf->tid, ceph_cap_string(cf->caps));
2345                         ci->i_ceph_flags |= CEPH_I_NODELAY;
2346 
2347                         ret = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
2348                                          (cf->tid < last_snap_flush ?
2349                                           CEPH_CLIENT_CAPS_PENDING_CAPSNAP : 0),
2350                                           __ceph_caps_used(ci),
2351                                           __ceph_caps_wanted(ci),
2352                                           (cap->issued | cap->implemented),
2353                                           cf->caps, cf->tid, oldest_flush_tid);
2354                         if (ret) {
2355                                 pr_err("kick_flushing_caps: error sending "
2356                                         "cap flush, ino (%llx.%llx) "
2357                                         "tid %llu flushing %s\n",
2358                                         ceph_vinop(inode), cf->tid,
2359                                         ceph_cap_string(cf->caps));
2360                         }
2361                 } else {
2362                         struct ceph_cap_snap *capsnap =
2363                                         container_of(cf, struct ceph_cap_snap,
2364                                                     cap_flush);
2365                         dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2366                              inode, capsnap, cf->tid,
2367                              ceph_cap_string(capsnap->dirty));
2368 
2369                         refcount_inc(&capsnap->nref);
2370                         spin_unlock(&ci->i_ceph_lock);
2371 
2372                         ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
2373                                                 oldest_flush_tid);
2374                         if (ret < 0) {
2375                                 pr_err("kick_flushing_caps: error sending "
2376                                         "cap flushsnap, ino (%llx.%llx) "
2377                                         "tid %llu follows %llu\n",
2378                                         ceph_vinop(inode), cf->tid,
2379                                         capsnap->follows);
2380                         }
2381 
2382                         ceph_put_cap_snap(capsnap);
2383                 }
2384 
2385                 spin_lock(&ci->i_ceph_lock);
2386         }
2387 }
2388 
2389 void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
2390                                    struct ceph_mds_session *session)
2391 {
2392         struct ceph_inode_info *ci;
2393         struct ceph_cap *cap;
2394         u64 oldest_flush_tid;
2395 
2396         dout("early_kick_flushing_caps mds%d\n", session->s_mds);
2397 
2398         spin_lock(&mdsc->cap_dirty_lock);
2399         oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2400         spin_unlock(&mdsc->cap_dirty_lock);
2401 
2402         list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2403                 spin_lock(&ci->i_ceph_lock);
2404                 cap = ci->i_auth_cap;
2405                 if (!(cap && cap->session == session)) {
2406                         pr_err("%p auth cap %p not mds%d ???\n",
2407                                 &ci->vfs_inode, cap, session->s_mds);
2408                         spin_unlock(&ci->i_ceph_lock);
2409                         continue;
2410                 }
2411 
2412 
2413                 /*
2414                  * if flushing caps were revoked, we re-send the cap flush
2415                  * in client reconnect stage. This guarantees MDS * processes
2416                  * the cap flush message before issuing the flushing caps to
2417                  * other client.
2418                  */
2419                 if ((cap->issued & ci->i_flushing_caps) !=
2420                     ci->i_flushing_caps) {
2421                         /* encode_caps_cb() also will reset these sequence
2422                          * numbers. make sure sequence numbers in cap flush
2423                          * message match later reconnect message */
2424                         cap->seq = 0;
2425                         cap->issue_seq = 0;
2426                         cap->mseq = 0;
2427                         __kick_flushing_caps(mdsc, session, ci,
2428                                              oldest_flush_tid);
2429                 } else {
2430                         ci->i_ceph_flags |= CEPH_I_KICK_FLUSH;
2431                 }
2432 
2433                 spin_unlock(&ci->i_ceph_lock);
2434         }
2435 }
2436 
2437 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2438                              struct ceph_mds_session *session)
2439 {
2440         struct ceph_inode_info *ci;
2441         struct ceph_cap *cap;
2442         u64 oldest_flush_tid;
2443 
2444         dout("kick_flushing_caps mds%d\n", session->s_mds);
2445 
2446         spin_lock(&mdsc->cap_dirty_lock);
2447         oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2448         spin_unlock(&mdsc->cap_dirty_lock);
2449 
2450         list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2451                 spin_lock(&ci->i_ceph_lock);
2452                 cap = ci->i_auth_cap;
2453                 if (!(cap && cap->session == session)) {
2454                         pr_err("%p auth cap %p not mds%d ???\n",
2455                                 &ci->vfs_inode, cap, session->s_mds);
2456                         spin_unlock(&ci->i_ceph_lock);
2457                         continue;
2458                 }
2459                 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
2460                         __kick_flushing_caps(mdsc, session, ci,
2461                                              oldest_flush_tid);
2462                 }
2463                 spin_unlock(&ci->i_ceph_lock);
2464         }
2465 }
2466 
2467 static void kick_flushing_inode_caps(struct ceph_mds_client *mdsc,
2468                                      struct ceph_mds_session *session,
2469                                      struct inode *inode)
2470         __releases(ci->i_ceph_lock)
2471 {
2472         struct ceph_inode_info *ci = ceph_inode(inode);
2473         struct ceph_cap *cap;
2474 
2475         cap = ci->i_auth_cap;
2476         dout("kick_flushing_inode_caps %p flushing %s\n", inode,
2477              ceph_cap_string(ci->i_flushing_caps));
2478 
2479         if (!list_empty(&ci->i_cap_flush_list)) {
2480                 u64 oldest_flush_tid;
2481                 spin_lock(&mdsc->cap_dirty_lock);
2482                 list_move_tail(&ci->i_flushing_item,
2483                                &cap->session->s_cap_flushing);
2484                 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2485                 spin_unlock(&mdsc->cap_dirty_lock);
2486 
2487                 __kick_flushing_caps(mdsc, session, ci, oldest_flush_tid);
2488                 spin_unlock(&ci->i_ceph_lock);
2489         } else {
2490                 spin_unlock(&ci->i_ceph_lock);
2491         }
2492 }
2493 
2494 
2495 /*
2496  * Take references to capabilities we hold, so that we don't release
2497  * them to the MDS prematurely.
2498  *
2499  * Protected by i_ceph_lock.
2500  */
2501 static void __take_cap_refs(struct ceph_inode_info *ci, int got,
2502                             bool snap_rwsem_locked)
2503 {
2504         if (got & CEPH_CAP_PIN)
2505                 ci->i_pin_ref++;
2506         if (got & CEPH_CAP_FILE_RD)
2507                 ci->i_rd_ref++;
2508         if (got & CEPH_CAP_FILE_CACHE)
2509                 ci->i_rdcache_ref++;
2510         if (got & CEPH_CAP_FILE_WR) {
2511                 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2512                         BUG_ON(!snap_rwsem_locked);
2513                         ci->i_head_snapc = ceph_get_snap_context(
2514                                         ci->i_snap_realm->cached_context);
2515                 }
2516                 ci->i_wr_ref++;
2517         }
2518         if (got & CEPH_CAP_FILE_BUFFER) {
2519                 if (ci->i_wb_ref == 0)
2520                         ihold(&ci->vfs_inode);
2521                 ci->i_wb_ref++;
2522                 dout("__take_cap_refs %p wb %d -> %d (?)\n",
2523                      &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2524         }
2525 }
2526 
2527 /*
2528  * Try to grab cap references.  Specify those refs we @want, and the
2529  * minimal set we @need.  Also include the larger offset we are writing
2530  * to (when applicable), and check against max_size here as well.
2531  * Note that caller is responsible for ensuring max_size increases are
2532  * requested from the MDS.
2533  *
2534  * Returns 0 if caps were not able to be acquired (yet), a 1 if they were,
2535  * or a negative error code.
2536  *
2537  * FIXME: how does a 0 return differ from -EAGAIN?
2538  */
2539 enum {
2540         NON_BLOCKING    = 1,
2541         CHECK_FILELOCK  = 2,
2542 };
2543 
2544 static int try_get_cap_refs(struct inode *inode, int need, int want,
2545                             loff_t endoff, int flags, int *got)
2546 {
2547         struct ceph_inode_info *ci = ceph_inode(inode);
2548         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2549         int ret = 0;
2550         int have, implemented;
2551         int file_wanted;
2552         bool snap_rwsem_locked = false;
2553 
2554         dout("get_cap_refs %p need %s want %s\n", inode,
2555              ceph_cap_string(need), ceph_cap_string(want));
2556 
2557 again:
2558         spin_lock(&ci->i_ceph_lock);
2559 
2560         if ((flags & CHECK_FILELOCK) &&
2561             (ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK)) {
2562                 dout("try_get_cap_refs %p error filelock\n", inode);
2563                 ret = -EIO;
2564                 goto out_unlock;
2565         }
2566 
2567         /* make sure file is actually open */
2568         file_wanted = __ceph_caps_file_wanted(ci);
2569         if ((file_wanted & need) != need) {
2570                 dout("try_get_cap_refs need %s file_wanted %s, EBADF\n",
2571                      ceph_cap_string(need), ceph_cap_string(file_wanted));
2572                 ret = -EBADF;
2573                 goto out_unlock;
2574         }
2575 
2576         /* finish pending truncate */
2577         while (ci->i_truncate_pending) {
2578                 spin_unlock(&ci->i_ceph_lock);
2579                 if (snap_rwsem_locked) {
2580                         up_read(&mdsc->snap_rwsem);
2581                         snap_rwsem_locked = false;
2582                 }
2583                 __ceph_do_pending_vmtruncate(inode);
2584                 spin_lock(&ci->i_ceph_lock);
2585         }
2586 
2587         have = __ceph_caps_issued(ci, &implemented);
2588 
2589         if (have & need & CEPH_CAP_FILE_WR) {
2590                 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2591                         dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2592                              inode, endoff, ci->i_max_size);
2593                         if (endoff > ci->i_requested_max_size)
2594                                 ret = -EAGAIN;
2595                         goto out_unlock;
2596                 }
2597                 /*
2598                  * If a sync write is in progress, we must wait, so that we
2599                  * can get a final snapshot value for size+mtime.
2600                  */
2601                 if (__ceph_have_pending_cap_snap(ci)) {
2602                         dout("get_cap_refs %p cap_snap_pending\n", inode);
2603                         goto out_unlock;
2604                 }
2605         }
2606 
2607         if ((have & need) == need) {
2608                 /*
2609                  * Look at (implemented & ~have & not) so that we keep waiting
2610                  * on transition from wanted -> needed caps.  This is needed
2611                  * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2612                  * going before a prior buffered writeback happens.
2613                  */
2614                 int not = want & ~(have & need);
2615                 int revoking = implemented & ~have;
2616                 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2617                      inode, ceph_cap_string(have), ceph_cap_string(not),
2618                      ceph_cap_string(revoking));
2619                 if ((revoking & not) == 0) {
2620                         if (!snap_rwsem_locked &&
2621                             !ci->i_head_snapc &&
2622                             (need & CEPH_CAP_FILE_WR)) {
2623                                 if (!down_read_trylock(&mdsc->snap_rwsem)) {
2624                                         /*
2625                                          * we can not call down_read() when
2626                                          * task isn't in TASK_RUNNING state
2627                                          */
2628                                         if (flags & NON_BLOCKING) {
2629                                                 ret = -EAGAIN;
2630                                                 goto out_unlock;
2631                                         }
2632 
2633                                         spin_unlock(&ci->i_ceph_lock);
2634                                         down_read(&mdsc->snap_rwsem);
2635                                         snap_rwsem_locked = true;
2636                                         goto again;
2637                                 }
2638                                 snap_rwsem_locked = true;
2639                         }
2640                         *got = need | (have & want);
2641                         if ((need & CEPH_CAP_FILE_RD) &&
2642                             !(*got & CEPH_CAP_FILE_CACHE))
2643                                 ceph_disable_fscache_readpage(ci);
2644                         __take_cap_refs(ci, *got, true);
2645                         ret = 1;
2646                 }
2647         } else {
2648                 int session_readonly = false;
2649                 if ((need & CEPH_CAP_FILE_WR) && ci->i_auth_cap) {
2650                         struct ceph_mds_session *s = ci->i_auth_cap->session;
2651                         spin_lock(&s->s_cap_lock);
2652                         session_readonly = s->s_readonly;
2653                         spin_unlock(&s->s_cap_lock);
2654                 }
2655                 if (session_readonly) {
2656                         dout("get_cap_refs %p needed %s but mds%d readonly\n",
2657                              inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2658                         ret = -EROFS;
2659                         goto out_unlock;
2660                 }
2661 
2662                 if (ci->i_ceph_flags & CEPH_I_CAP_DROPPED) {
2663                         int mds_wanted;
2664                         if (READ_ONCE(mdsc->fsc->mount_state) ==
2665                             CEPH_MOUNT_SHUTDOWN) {
2666                                 dout("get_cap_refs %p forced umount\n", inode);
2667                                 ret = -EIO;
2668                                 goto out_unlock;
2669                         }
2670                         mds_wanted = __ceph_caps_mds_wanted(ci, false);
2671                         if (need & ~(mds_wanted & need)) {
2672                                 dout("get_cap_refs %p caps were dropped"
2673                                      " (session killed?)\n", inode);
2674                                 ret = -ESTALE;
2675                                 goto out_unlock;
2676                         }
2677                         if (!(file_wanted & ~mds_wanted))
2678                                 ci->i_ceph_flags &= ~CEPH_I_CAP_DROPPED;
2679                 }
2680 
2681                 dout("get_cap_refs %p have %s needed %s\n", inode,
2682                      ceph_cap_string(have), ceph_cap_string(need));
2683         }
2684 out_unlock:
2685         spin_unlock(&ci->i_ceph_lock);
2686         if (snap_rwsem_locked)
2687                 up_read(&mdsc->snap_rwsem);
2688 
2689         dout("get_cap_refs %p ret %d got %s\n", inode,
2690              ret, ceph_cap_string(*got));
2691         return ret;
2692 }
2693 
2694 /*
2695  * Check the offset we are writing up to against our current
2696  * max_size.  If necessary, tell the MDS we want to write to
2697  * a larger offset.
2698  */
2699 static void check_max_size(struct inode *inode, loff_t endoff)
2700 {
2701         struct ceph_inode_info *ci = ceph_inode(inode);
2702         int check = 0;
2703 
2704         /* do we need to explicitly request a larger max_size? */
2705         spin_lock(&ci->i_ceph_lock);
2706         if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2707                 dout("write %p at large endoff %llu, req max_size\n",
2708                      inode, endoff);
2709                 ci->i_wanted_max_size = endoff;
2710         }
2711         /* duplicate ceph_check_caps()'s logic */
2712         if (ci->i_auth_cap &&
2713             (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2714             ci->i_wanted_max_size > ci->i_max_size &&
2715             ci->i_wanted_max_size > ci->i_requested_max_size)
2716                 check = 1;
2717         spin_unlock(&ci->i_ceph_lock);
2718         if (check)
2719                 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2720 }
2721 
2722 int ceph_try_get_caps(struct inode *inode, int need, int want,
2723                       bool nonblock, int *got)
2724 {
2725         int ret;
2726 
2727         BUG_ON(need & ~CEPH_CAP_FILE_RD);
2728         BUG_ON(want & ~(CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO|CEPH_CAP_FILE_SHARED));
2729         ret = ceph_pool_perm_check(inode, need);
2730         if (ret < 0)
2731                 return ret;
2732 
2733         ret = try_get_cap_refs(inode, need, want, 0,
2734                                (nonblock ? NON_BLOCKING : 0), got);
2735         return ret == -EAGAIN ? 0 : ret;
2736 }
2737 
2738 /*
2739  * Wait for caps, and take cap references.  If we can't get a WR cap
2740  * due to a small max_size, make sure we check_max_size (and possibly
2741  * ask the mds) so we don't get hung up indefinitely.
2742  */
2743 int ceph_get_caps(struct file *filp, int need, int want,
2744                   loff_t endoff, int *got, struct page **pinned_page)
2745 {
2746         struct ceph_file_info *fi = filp->private_data;
2747         struct inode *inode = file_inode(filp);
2748         struct ceph_inode_info *ci = ceph_inode(inode);
2749         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
2750         int ret, _got, flags;
2751 
2752         ret = ceph_pool_perm_check(inode, need);
2753         if (ret < 0)
2754                 return ret;
2755 
2756         if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2757             fi->filp_gen != READ_ONCE(fsc->filp_gen))
2758                 return -EBADF;
2759 
2760         while (true) {
2761                 if (endoff > 0)
2762                         check_max_size(inode, endoff);
2763 
2764                 flags = atomic_read(&fi->num_locks) ? CHECK_FILELOCK : 0;
2765                 _got = 0;
2766                 ret = try_get_cap_refs(inode, need, want, endoff,
2767                                        flags, &_got);
2768                 if (ret == -EAGAIN)
2769                         continue;
2770                 if (!ret) {
2771                         DEFINE_WAIT_FUNC(wait, woken_wake_function);
2772                         add_wait_queue(&ci->i_cap_wq, &wait);
2773 
2774                         flags |= NON_BLOCKING;
2775                         while (!(ret = try_get_cap_refs(inode, need, want,
2776                                                         endoff, flags, &_got))) {
2777                                 if (signal_pending(current)) {
2778                                         ret = -ERESTARTSYS;
2779                                         break;
2780                                 }
2781                                 wait_woken(&wait, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2782                         }
2783 
2784                         remove_wait_queue(&ci->i_cap_wq, &wait);
2785                         if (ret == -EAGAIN)
2786                                 continue;
2787                 }
2788 
2789                 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2790                     fi->filp_gen != READ_ONCE(fsc->filp_gen)) {
2791                         if (ret >= 0 && _got)
2792                                 ceph_put_cap_refs(ci, _got);
2793                         return -EBADF;
2794                 }
2795 
2796                 if (ret < 0) {
2797                         if (ret == -ESTALE) {
2798                                 /* session was killed, try renew caps */
2799                                 ret = ceph_renew_caps(inode);
2800                                 if (ret == 0)
2801                                         continue;
2802                         }
2803                         return ret;
2804                 }
2805 
2806                 if (ci->i_inline_version != CEPH_INLINE_NONE &&
2807                     (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2808                     i_size_read(inode) > 0) {
2809                         struct page *page =
2810                                 find_get_page(inode->i_mapping, 0);
2811                         if (page) {
2812                                 if (PageUptodate(page)) {
2813                                         *pinned_page = page;
2814                                         break;
2815                                 }
2816                                 put_page(page);
2817                         }
2818                         /*
2819                          * drop cap refs first because getattr while
2820                          * holding * caps refs can cause deadlock.
2821                          */
2822                         ceph_put_cap_refs(ci, _got);
2823                         _got = 0;
2824 
2825                         /*
2826                          * getattr request will bring inline data into
2827                          * page cache
2828                          */
2829                         ret = __ceph_do_getattr(inode, NULL,
2830                                                 CEPH_STAT_CAP_INLINE_DATA,
2831                                                 true);
2832                         if (ret < 0)
2833                                 return ret;
2834                         continue;
2835                 }
2836                 break;
2837         }
2838 
2839         if ((_got & CEPH_CAP_FILE_RD) && (_got & CEPH_CAP_FILE_CACHE))
2840                 ceph_fscache_revalidate_cookie(ci);
2841 
2842         *got = _got;
2843         return 0;
2844 }
2845 
2846 /*
2847  * Take cap refs.  Caller must already know we hold at least one ref
2848  * on the caps in question or we don't know this is safe.
2849  */
2850 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2851 {
2852         spin_lock(&ci->i_ceph_lock);
2853         __take_cap_refs(ci, caps, false);
2854         spin_unlock(&ci->i_ceph_lock);
2855 }
2856 
2857 
2858 /*
2859  * drop cap_snap that is not associated with any snapshot.
2860  * we don't need to send FLUSHSNAP message for it.
2861  */
2862 static int ceph_try_drop_cap_snap(struct ceph_inode_info *ci,
2863                                   struct ceph_cap_snap *capsnap)
2864 {
2865         if (!capsnap->need_flush &&
2866             !capsnap->writing && !capsnap->dirty_pages) {
2867                 dout("dropping cap_snap %p follows %llu\n",
2868                      capsnap, capsnap->follows);
2869                 BUG_ON(capsnap->cap_flush.tid > 0);
2870                 ceph_put_snap_context(capsnap->context);
2871                 if (!list_is_last(&capsnap->ci_item, &ci->i_cap_snaps))
2872                         ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
2873 
2874                 list_del(&capsnap->ci_item);
2875                 ceph_put_cap_snap(capsnap);
2876                 return 1;
2877         }
2878         return 0;
2879 }
2880 
2881 /*
2882  * Release cap refs.
2883  *
2884  * If we released the last ref on any given cap, call ceph_check_caps
2885  * to release (or schedule a release).
2886  *
2887  * If we are releasing a WR cap (from a sync write), finalize any affected
2888  * cap_snap, and wake up any waiters.
2889  */
2890 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2891 {
2892         struct inode *inode = &ci->vfs_inode;
2893         int last = 0, put = 0, flushsnaps = 0, wake = 0;
2894 
2895         spin_lock(&ci->i_ceph_lock);
2896         if (had & CEPH_CAP_PIN)
2897                 --ci->i_pin_ref;
2898         if (had & CEPH_CAP_FILE_RD)
2899                 if (--ci->i_rd_ref == 0)
2900                         last++;
2901         if (had & CEPH_CAP_FILE_CACHE)
2902                 if (--ci->i_rdcache_ref == 0)
2903                         last++;
2904         if (had & CEPH_CAP_FILE_BUFFER) {
2905                 if (--ci->i_wb_ref == 0) {
2906                         last++;
2907                         put++;
2908                 }
2909                 dout("put_cap_refs %p wb %d -> %d (?)\n",
2910                      inode, ci->i_wb_ref+1, ci->i_wb_ref);
2911         }
2912         if (had & CEPH_CAP_FILE_WR)
2913                 if (--ci->i_wr_ref == 0) {
2914                         last++;
2915                         if (__ceph_have_pending_cap_snap(ci)) {
2916                                 struct ceph_cap_snap *capsnap =
2917                                         list_last_entry(&ci->i_cap_snaps,
2918                                                         struct ceph_cap_snap,
2919                                                         ci_item);
2920                                 capsnap->writing = 0;
2921                                 if (ceph_try_drop_cap_snap(ci, capsnap))
2922                                         put++;
2923                                 else if (__ceph_finish_cap_snap(ci, capsnap))
2924                                         flushsnaps = 1;
2925                                 wake = 1;
2926                         }
2927                         if (ci->i_wrbuffer_ref_head == 0 &&
2928                             ci->i_dirty_caps == 0 &&
2929                             ci->i_flushing_caps == 0) {
2930                                 BUG_ON(!ci->i_head_snapc);
2931                                 ceph_put_snap_context(ci->i_head_snapc);
2932                                 ci->i_head_snapc = NULL;
2933                         }
2934                         /* see comment in __ceph_remove_cap() */
2935                         if (!__ceph_is_any_caps(ci) && ci->i_snap_realm)
2936                                 drop_inode_snap_realm(ci);
2937                 }
2938         spin_unlock(&ci->i_ceph_lock);
2939 
2940         dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
2941              last ? " last" : "", put ? " put" : "");
2942 
2943         if (last && !flushsnaps)
2944                 ceph_check_caps(ci, 0, NULL);
2945         else if (flushsnaps)
2946                 ceph_flush_snaps(ci, NULL);
2947         if (wake)
2948                 wake_up_all(&ci->i_cap_wq);
2949         while (put-- > 0)
2950                 iput(inode);
2951 }
2952 
2953 /*
2954  * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2955  * context.  Adjust per-snap dirty page accounting as appropriate.
2956  * Once all dirty data for a cap_snap is flushed, flush snapped file
2957  * metadata back to the MDS.  If we dropped the last ref, call
2958  * ceph_check_caps.
2959  */
2960 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
2961                                 struct ceph_snap_context *snapc)
2962 {
2963         struct inode *inode = &ci->vfs_inode;
2964         struct ceph_cap_snap *capsnap = NULL;
2965         int put = 0;
2966         bool last = false;
2967         bool found = false;
2968         bool flush_snaps = false;
2969         bool complete_capsnap = false;
2970 
2971         spin_lock(&ci->i_ceph_lock);
2972         ci->i_wrbuffer_ref -= nr;
2973         if (ci->i_wrbuffer_ref == 0) {
2974                 last = true;
2975                 put++;
2976         }
2977 
2978         if (ci->i_head_snapc == snapc) {
2979                 ci->i_wrbuffer_ref_head -= nr;
2980                 if (ci->i_wrbuffer_ref_head == 0 &&
2981                     ci->i_wr_ref == 0 &&
2982                     ci->i_dirty_caps == 0 &&
2983                     ci->i_flushing_caps == 0) {
2984                         BUG_ON(!ci->i_head_snapc);
2985                         ceph_put_snap_context(ci->i_head_snapc);
2986                         ci->i_head_snapc = NULL;
2987                 }
2988                 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
2989                      inode,
2990                      ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
2991                      ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
2992                      last ? " LAST" : "");
2993         } else {
2994                 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2995                         if (capsnap->context == snapc) {
2996                                 found = true;
2997                                 break;
2998                         }
2999                 }
3000                 BUG_ON(!found);
3001                 capsnap->dirty_pages -= nr;
3002                 if (capsnap->dirty_pages == 0) {
3003                         complete_capsnap = true;
3004                         if (!capsnap->writing) {
3005                                 if (ceph_try_drop_cap_snap(ci, capsnap)) {
3006                                         put++;
3007                                 } else {
3008                                         ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
3009                                         flush_snaps = true;
3010                                 }
3011                         }
3012                 }
3013                 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
3014                      " snap %lld %d/%d -> %d/%d %s%s\n",
3015                      inode, capsnap, capsnap->context->seq,
3016                      ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
3017                      ci->i_wrbuffer_ref, capsnap->dirty_pages,
3018                      last ? " (wrbuffer last)" : "",
3019                      complete_capsnap ? " (complete capsnap)" : "");
3020         }
3021 
3022         spin_unlock(&ci->i_ceph_lock);
3023 
3024         if (last) {
3025                 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
3026         } else if (flush_snaps) {
3027                 ceph_flush_snaps(ci, NULL);
3028         }
3029         if (complete_capsnap)
3030                 wake_up_all(&ci->i_cap_wq);
3031         while (put-- > 0) {
3032                 /* avoid calling iput_final() in osd dispatch threads */
3033                 ceph_async_iput(inode);
3034         }
3035 }
3036 
3037 /*
3038  * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
3039  */
3040 static void invalidate_aliases(struct inode *inode)
3041 {
3042         struct dentry *dn, *prev = NULL;
3043 
3044         dout("invalidate_aliases inode %p\n", inode);
3045         d_prune_aliases(inode);
3046         /*
3047          * For non-directory inode, d_find_alias() only returns
3048          * hashed dentry. After calling d_invalidate(), the
3049          * dentry becomes unhashed.
3050          *
3051          * For directory inode, d_find_alias() can return
3052          * unhashed dentry. But directory inode should have
3053          * one alias at most.
3054          */
3055         while ((dn = d_find_alias(inode))) {
3056                 if (dn == prev) {
3057                         dput(dn);
3058                         break;
3059                 }
3060                 d_invalidate(dn);
3061                 if (prev)
3062                         dput(prev);
3063                 prev = dn;
3064         }
3065         if (prev)
3066                 dput(prev);
3067 }
3068 
3069 struct cap_extra_info {
3070         struct ceph_string *pool_ns;
3071         /* inline data */
3072         u64 inline_version;
3073         void *inline_data;
3074         u32 inline_len;
3075         /* dirstat */
3076         bool dirstat_valid;
3077         u64 nfiles;
3078         u64 nsubdirs;
3079         u64 change_attr;
3080         /* currently issued */
3081         int issued;
3082         struct timespec64 btime;
3083 };
3084 
3085 /*
3086  * Handle a cap GRANT message from the MDS.  (Note that a GRANT may
3087  * actually be a revocation if it specifies a smaller cap set.)
3088  *
3089  * caller holds s_mutex and i_ceph_lock, we drop both.
3090  */
3091 static void handle_cap_grant(struct inode *inode,
3092                              struct ceph_mds_session *session,
3093                              struct ceph_cap *cap,
3094                              struct ceph_mds_caps *grant,
3095                              struct ceph_buffer *xattr_buf,
3096                              struct cap_extra_info *extra_info)
3097         __releases(ci->i_ceph_lock)
3098         __releases(session->s_mdsc->snap_rwsem)
3099 {
3100         struct ceph_inode_info *ci = ceph_inode(inode);
3101         int seq = le32_to_cpu(grant->seq);
3102         int newcaps = le32_to_cpu(grant->caps);
3103         int used, wanted, dirty;
3104         u64 size = le64_to_cpu(grant->size);
3105         u64 max_size = le64_to_cpu(grant->max_size);
3106         unsigned char check_caps = 0;
3107         bool was_stale = cap->cap_gen < session->s_cap_gen;
3108         bool wake = false;
3109         bool writeback = false;
3110         bool queue_trunc = false;
3111         bool queue_invalidate = false;
3112         bool deleted_inode = false;
3113         bool fill_inline = false;
3114 
3115         dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
3116              inode, cap, session->s_mds, seq, ceph_cap_string(newcaps));
3117         dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
3118                 inode->i_size);
3119 
3120 
3121         /*
3122          * If CACHE is being revoked, and we have no dirty buffers,
3123          * try to invalidate (once).  (If there are dirty buffers, we
3124          * will invalidate _after_ writeback.)
3125          */
3126         if (!S_ISDIR(inode->i_mode) && /* don't invalidate readdir cache */
3127             ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
3128             (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3129             !(ci->i_wrbuffer_ref || ci->i_wb_ref)) {
3130                 if (try_nonblocking_invalidate(inode)) {
3131                         /* there were locked pages.. invalidate later
3132                            in a separate thread. */
3133                         if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
3134                                 queue_invalidate = true;
3135                                 ci->i_rdcache_revoking = ci->i_rdcache_gen;
3136                         }
3137                 }
3138         }
3139 
3140         if (was_stale)
3141                 cap->issued = cap->implemented = CEPH_CAP_PIN;
3142 
3143         /*
3144          * auth mds of the inode changed. we received the cap export message,
3145          * but still haven't received the cap import message. handle_cap_export
3146          * updated the new auth MDS' cap.
3147          *
3148          * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
3149          * that was sent before the cap import message. So don't remove caps.
3150          */
3151         if (ceph_seq_cmp(seq, cap->seq) <= 0) {
3152                 WARN_ON(cap != ci->i_auth_cap);
3153                 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
3154                 seq = cap->seq;
3155                 newcaps |= cap->issued;
3156         }
3157 
3158         /* side effects now are allowed */
3159         cap->cap_gen = session->s_cap_gen;
3160         cap->seq = seq;
3161 
3162         __check_cap_issue(ci, cap, newcaps);
3163 
3164         inode_set_max_iversion_raw(inode, extra_info->change_attr);
3165 
3166         if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
3167             (extra_info->issued & CEPH_CAP_AUTH_EXCL) == 0) {
3168                 inode->i_mode = le32_to_cpu(grant->mode);
3169                 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
3170                 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
3171                 ci->i_btime = extra_info->btime;
3172                 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
3173                      from_kuid(&init_user_ns, inode->i_uid),
3174                      from_kgid(&init_user_ns, inode->i_gid));
3175         }
3176 
3177         if ((newcaps & CEPH_CAP_LINK_SHARED) &&
3178             (extra_info->issued & CEPH_CAP_LINK_EXCL) == 0) {
3179                 set_nlink(inode, le32_to_cpu(grant->nlink));
3180                 if (inode->i_nlink == 0 &&
3181                     (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
3182                         deleted_inode = true;
3183         }
3184 
3185         if ((extra_info->issued & CEPH_CAP_XATTR_EXCL) == 0 &&
3186             grant->xattr_len) {
3187                 int len = le32_to_cpu(grant->xattr_len);
3188                 u64 version = le64_to_cpu(grant->xattr_version);
3189 
3190                 if (version > ci->i_xattrs.version) {
3191                         dout(" got new xattrs v%llu on %p len %d\n",
3192                              version, inode, len);
3193                         if (ci->i_xattrs.blob)
3194                                 ceph_buffer_put(ci->i_xattrs.blob);
3195                         ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
3196                         ci->i_xattrs.version = version;
3197                         ceph_forget_all_cached_acls(inode);
3198                         ceph_security_invalidate_secctx(inode);
3199                 }
3200         }
3201 
3202         if (newcaps & CEPH_CAP_ANY_RD) {
3203                 struct timespec64 mtime, atime, ctime;
3204                 /* ctime/mtime/atime? */
3205                 ceph_decode_timespec64(&mtime, &grant->mtime);
3206                 ceph_decode_timespec64(&atime, &grant->atime);
3207                 ceph_decode_timespec64(&ctime, &grant->ctime);
3208                 ceph_fill_file_time(inode, extra_info->issued,
3209                                     le32_to_cpu(grant->time_warp_seq),
3210                                     &ctime, &mtime, &atime);
3211         }
3212 
3213         if ((newcaps & CEPH_CAP_FILE_SHARED) && extra_info->dirstat_valid) {
3214                 ci->i_files = extra_info->nfiles;
3215                 ci->i_subdirs = extra_info->nsubdirs;
3216         }
3217 
3218         if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
3219                 /* file layout may have changed */
3220                 s64 old_pool = ci->i_layout.pool_id;
3221                 struct ceph_string *old_ns;
3222 
3223                 ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout);
3224                 old_ns = rcu_dereference_protected(ci->i_layout.pool_ns,
3225                                         lockdep_is_held(&ci->i_ceph_lock));
3226                 rcu_assign_pointer(ci->i_layout.pool_ns, extra_info->pool_ns);
3227 
3228                 if (ci->i_layout.pool_id != old_pool ||
3229                     extra_info->pool_ns != old_ns)
3230                         ci->i_ceph_flags &= ~CEPH_I_POOL_PERM;
3231 
3232                 extra_info->pool_ns = old_ns;
3233 
3234                 /* size/truncate_seq? */
3235                 queue_trunc = ceph_fill_file_size(inode, extra_info->issued,
3236                                         le32_to_cpu(grant->truncate_seq),
3237                                         le64_to_cpu(grant->truncate_size),
3238                                         size);
3239         }
3240 
3241         if (ci->i_auth_cap == cap && (newcaps & CEPH_CAP_ANY_FILE_WR)) {
3242                 if (max_size != ci->i_max_size) {
3243                         dout("max_size %lld -> %llu\n",
3244                              ci->i_max_size, max_size);
3245                         ci->i_max_size = max_size;
3246                         if (max_size >= ci->i_wanted_max_size) {
3247                                 ci->i_wanted_max_size = 0;  /* reset */
3248                                 ci->i_requested_max_size = 0;
3249                         }
3250                         wake = true;
3251                 } else if (ci->i_wanted_max_size > ci->i_max_size &&
3252                            ci->i_wanted_max_size > ci->i_requested_max_size) {
3253                         /* CEPH_CAP_OP_IMPORT */
3254                         wake = true;
3255                 }
3256         }
3257 
3258         /* check cap bits */
3259         wanted = __ceph_caps_wanted(ci);
3260         used = __ceph_caps_used(ci);
3261         dirty = __ceph_caps_dirty(ci);
3262         dout(" my wanted = %s, used = %s, dirty %s\n",
3263              ceph_cap_string(wanted),
3264              ceph_cap_string(used),
3265              ceph_cap_string(dirty));
3266 
3267         if ((was_stale || le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) &&
3268             (wanted & ~(cap->mds_wanted | newcaps))) {
3269                 /*
3270                  * If mds is importing cap, prior cap messages that update
3271                  * 'wanted' may get dropped by mds (migrate seq mismatch).
3272                  *
3273                  * We don't send cap message to update 'wanted' if what we
3274                  * want are already issued. If mds revokes caps, cap message
3275                  * that releases caps also tells mds what we want. But if
3276                  * caps got revoked by mds forcedly (session stale). We may
3277                  * haven't told mds what we want.
3278                  */
3279                 check_caps = 1;
3280         }
3281 
3282         /* revocation, grant, or no-op? */
3283         if (cap->issued & ~newcaps) {
3284                 int revoking = cap->issued & ~newcaps;
3285 
3286                 dout("revocation: %s -> %s (revoking %s)\n",
3287                      ceph_cap_string(cap->issued),
3288                      ceph_cap_string(newcaps),
3289                      ceph_cap_string(revoking));
3290                 if (revoking & used & CEPH_CAP_FILE_BUFFER)
3291                         writeback = true;  /* initiate writeback; will delay ack */
3292                 else if (revoking == CEPH_CAP_FILE_CACHE &&
3293                          (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3294                          queue_invalidate)
3295                         ; /* do nothing yet, invalidation will be queued */
3296                 else if (cap == ci->i_auth_cap)
3297                         check_caps = 1; /* check auth cap only */
3298                 else
3299                         check_caps = 2; /* check all caps */
3300                 cap->issued = newcaps;
3301                 cap->implemented |= newcaps;
3302         } else if (cap->issued == newcaps) {
3303                 dout("caps unchanged: %s -> %s\n",
3304                      ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
3305         } else {
3306                 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
3307                      ceph_cap_string(newcaps));
3308                 /* non-auth MDS is revoking the newly grant caps ? */
3309                 if (cap == ci->i_auth_cap &&
3310                     __ceph_caps_revoking_other(ci, cap, newcaps))
3311                     check_caps = 2;
3312 
3313                 cap->issued = newcaps;
3314                 cap->implemented |= newcaps; /* add bits only, to
3315                                               * avoid stepping on a
3316                                               * pending revocation */
3317                 wake = true;
3318         }
3319         BUG_ON(cap->issued & ~cap->implemented);
3320 
3321         if (extra_info->inline_version > 0 &&
3322             extra_info->inline_version >= ci->i_inline_version) {
3323                 ci->i_inline_version = extra_info->inline_version;
3324                 if (ci->i_inline_version != CEPH_INLINE_NONE &&
3325                     (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
3326                         fill_inline = true;
3327         }
3328 
3329         if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
3330                 if (newcaps & ~extra_info->issued)
3331                         wake = true;
3332                 kick_flushing_inode_caps(session->s_mdsc, session, inode);
3333                 up_read(&session->s_mdsc->snap_rwsem);
3334         } else {
3335                 spin_unlock(&ci->i_ceph_lock);
3336         }
3337 
3338         if (fill_inline)
3339                 ceph_fill_inline_data(inode, NULL, extra_info->inline_data,
3340                                       extra_info->inline_len);
3341 
3342         if (queue_trunc)
3343                 ceph_queue_vmtruncate(inode);
3344 
3345         if (writeback)
3346                 /*
3347                  * queue inode for writeback: we can't actually call
3348                  * filemap_write_and_wait, etc. from message handler
3349                  * context.
3350                  */
3351                 ceph_queue_writeback(inode);
3352         if (queue_invalidate)
3353                 ceph_queue_invalidate(inode);
3354         if (deleted_inode)
3355                 invalidate_aliases(inode);
3356         if (wake)
3357                 wake_up_all(&ci->i_cap_wq);
3358 
3359         if (check_caps == 1)
3360                 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY,
3361                                 session);
3362         else if (check_caps == 2)
3363                 ceph_check_caps(ci, CHECK_CAPS_NODELAY, session);
3364         else
3365                 mutex_unlock(&session->s_mutex);
3366 }
3367 
3368 /*
3369  * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3370  * MDS has been safely committed.
3371  */
3372 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
3373                                  struct ceph_mds_caps *m,
3374                                  struct ceph_mds_session *session,
3375                                  struct ceph_cap *cap)
3376         __releases(ci->i_ceph_lock)
3377 {
3378         struct ceph_inode_info *ci = ceph_inode(inode);
3379         struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3380         struct ceph_cap_flush *cf, *tmp_cf;
3381         LIST_HEAD(to_remove);
3382         unsigned seq = le32_to_cpu(m->seq);
3383         int dirty = le32_to_cpu(m->dirty);
3384         int cleaned = 0;
3385         bool drop = false;
3386         bool wake_ci = false;
3387         bool wake_mdsc = false;
3388 
3389         list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) {
3390                 if (cf->tid == flush_tid)
3391                         cleaned = cf->caps;
3392                 if (cf->caps == 0) /* capsnap */
3393                         continue;
3394                 if (cf->tid <= flush_tid) {
3395                         if (__finish_cap_flush(NULL, ci, cf))
3396                                 wake_ci = true;
3397                         list_add_tail(&cf->i_list, &to_remove);
3398                 } else {
3399                         cleaned &= ~cf->caps;
3400                         if (!cleaned)
3401                                 break;
3402                 }
3403         }
3404 
3405         dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3406              " flushing %s -> %s\n",
3407              inode, session->s_mds, seq, ceph_cap_string(dirty),
3408              ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
3409              ceph_cap_string(ci->i_flushing_caps & ~cleaned));
3410 
3411         if (list_empty(&to_remove) && !cleaned)
3412                 goto out;
3413 
3414         ci->i_flushing_caps &= ~cleaned;
3415 
3416         spin_lock(&mdsc->cap_dirty_lock);
3417 
3418         list_for_each_entry(cf, &to_remove, i_list) {
3419                 if (__finish_cap_flush(mdsc, NULL, cf))
3420                         wake_mdsc = true;
3421         }
3422 
3423         if (ci->i_flushing_caps == 0) {
3424                 if (list_empty(&ci->i_cap_flush_list)) {
3425                         list_del_init(&ci->i_flushing_item);
3426                         if (!list_empty(&session->s_cap_flushing)) {
3427                                 dout(" mds%d still flushing cap on %p\n",
3428                                      session->s_mds,
3429                                      &list_first_entry(&session->s_cap_flushing,
3430                                                 struct ceph_inode_info,
3431                                                 i_flushing_item)->vfs_inode);
3432                         }
3433                 }
3434                 mdsc->num_cap_flushing--;
3435                 dout(" inode %p now !flushing\n", inode);
3436 
3437                 if (ci->i_dirty_caps == 0) {
3438                         dout(" inode %p now clean\n", inode);
3439                         BUG_ON(!list_empty(&ci->i_dirty_item));
3440                         drop = true;
3441                         if (ci->i_wr_ref == 0 &&
3442                             ci->i_wrbuffer_ref_head == 0) {
3443                                 BUG_ON(!ci->i_head_snapc);
3444                                 ceph_put_snap_context(ci->i_head_snapc);
3445                                 ci->i_head_snapc = NULL;
3446                         }
3447                 } else {
3448                         BUG_ON(list_empty(&ci->i_dirty_item));
3449                 }
3450         }
3451         spin_unlock(&mdsc->cap_dirty_lock);
3452 
3453 out:
3454         spin_unlock(&ci->i_ceph_lock);
3455 
3456         while (!list_empty(&to_remove)) {
3457                 cf = list_first_entry(&to_remove,
3458                                       struct ceph_cap_flush, i_list);
3459                 list_del(&cf->i_list);
3460                 ceph_free_cap_flush(cf);
3461         }
3462 
3463         if (wake_ci)
3464                 wake_up_all(&ci->i_cap_wq);
3465         if (wake_mdsc)
3466                 wake_up_all(&mdsc->cap_flushing_wq);
3467         if (drop)
3468                 iput(inode);
3469 }
3470 
3471 /*
3472  * Handle FLUSHSNAP_ACK.  MDS has flushed snap data to disk and we can
3473  * throw away our cap_snap.
3474  *
3475  * Caller hold s_mutex.
3476  */
3477 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
3478                                      struct ceph_mds_caps *m,
3479                                      struct ceph_mds_session *session)
3480 {
3481         struct ceph_inode_info *ci = ceph_inode(inode);
3482         struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3483         u64 follows = le64_to_cpu(m->snap_follows);
3484         struct ceph_cap_snap *capsnap;
3485         bool flushed = false;
3486         bool wake_ci = false;
3487         bool wake_mdsc = false;
3488 
3489         dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3490              inode, ci, session->s_mds, follows);
3491 
3492         spin_lock(&ci->i_ceph_lock);
3493         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3494                 if (capsnap->follows == follows) {
3495                         if (capsnap->cap_flush.tid != flush_tid) {
3496                                 dout(" cap_snap %p follows %lld tid %lld !="
3497                                      " %lld\n", capsnap, follows,
3498                                      flush_tid, capsnap->cap_flush.tid);
3499                                 break;
3500                         }
3501                         flushed = true;
3502                         break;
3503                 } else {
3504                         dout(" skipping cap_snap %p follows %lld\n",
3505                              capsnap, capsnap->follows);
3506                 }
3507         }
3508         if (flushed) {
3509                 WARN_ON(capsnap->dirty_pages || capsnap->writing);
3510                 dout(" removing %p cap_snap %p follows %lld\n",
3511                      inode, capsnap, follows);
3512                 list_del(&capsnap->ci_item);
3513                 if (__finish_cap_flush(NULL, ci, &capsnap->cap_flush))
3514                         wake_ci = true;
3515 
3516                 spin_lock(&mdsc->cap_dirty_lock);
3517 
3518                 if (list_empty(&ci->i_cap_flush_list))
3519                         list_del_init(&ci->i_flushing_item);
3520 
3521                 if (__finish_cap_flush(mdsc, NULL, &capsnap->cap_flush))
3522                         wake_mdsc = true;
3523 
3524                 spin_unlock(&mdsc->cap_dirty_lock);
3525         }
3526         spin_unlock(&ci->i_ceph_lock);
3527         if (flushed) {
3528                 ceph_put_snap_context(capsnap->context);
3529                 ceph_put_cap_snap(capsnap);
3530                 if (wake_ci)
3531                         wake_up_all(&ci->i_cap_wq);
3532                 if (wake_mdsc)
3533                         wake_up_all(&mdsc->cap_flushing_wq);
3534                 iput(inode);
3535         }
3536 }
3537 
3538 /*
3539  * Handle TRUNC from MDS, indicating file truncation.
3540  *
3541  * caller hold s_mutex.
3542  */
3543 static void handle_cap_trunc(struct inode *inode,
3544                              struct ceph_mds_caps *trunc,
3545                              struct ceph_mds_session *session)
3546         __releases(ci->i_ceph_lock)
3547 {
3548         struct ceph_inode_info *ci = ceph_inode(inode);
3549         int mds = session->s_mds;
3550         int seq = le32_to_cpu(trunc->seq);
3551         u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
3552         u64 truncate_size = le64_to_cpu(trunc->truncate_size);
3553         u64 size = le64_to_cpu(trunc->size);
3554         int implemented = 0;
3555         int dirty = __ceph_caps_dirty(ci);
3556         int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
3557         int queue_trunc = 0;
3558 
3559         issued |= implemented | dirty;
3560 
3561         dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3562              inode, mds, seq, truncate_size, truncate_seq);
3563         queue_trunc = ceph_fill_file_size(inode, issued,
3564                                           truncate_seq, truncate_size, size);
3565         spin_unlock(&ci->i_ceph_lock);
3566 
3567         if (queue_trunc)
3568                 ceph_queue_vmtruncate(inode);
3569 }
3570 
3571 /*
3572  * Handle EXPORT from MDS.  Cap is being migrated _from_ this mds to a
3573  * different one.  If we are the most recent migration we've seen (as
3574  * indicated by mseq), make note of the migrating cap bits for the
3575  * duration (until we see the corresponding IMPORT).
3576  *
3577  * caller holds s_mutex
3578  */
3579 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
3580                               struct ceph_mds_cap_peer *ph,
3581                               struct ceph_mds_session *session)
3582 {
3583         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
3584         struct ceph_mds_session *tsession = NULL;
3585         struct ceph_cap *cap, *tcap, *new_cap = NULL;
3586         struct ceph_inode_info *ci = ceph_inode(inode);
3587         u64 t_cap_id;
3588         unsigned mseq = le32_to_cpu(ex->migrate_seq);
3589         unsigned t_seq, t_mseq;
3590         int target, issued;
3591         int mds = session->s_mds;
3592 
3593         if (ph) {
3594                 t_cap_id = le64_to_cpu(ph->cap_id);
3595                 t_seq = le32_to_cpu(ph->seq);
3596                 t_mseq = le32_to_cpu(ph->mseq);
3597                 target = le32_to_cpu(ph->mds);
3598         } else {
3599                 t_cap_id = t_seq = t_mseq = 0;
3600                 target = -1;
3601         }
3602 
3603         dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3604              inode, ci, mds, mseq, target);
3605 retry:
3606         spin_lock(&ci->i_ceph_lock);
3607         cap = __get_cap_for_mds(ci, mds);
3608         if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
3609                 goto out_unlock;
3610 
3611         if (target < 0) {
3612                 if (cap->mds_wanted | cap->issued)
3613                         ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
3614                 __ceph_remove_cap(cap, false);
3615                 goto out_unlock;
3616         }
3617 
3618         /*
3619          * now we know we haven't received the cap import message yet
3620          * because the exported cap still exist.
3621          */
3622 
3623         issued = cap->issued;
3624         if (issued != cap->implemented)
3625                 pr_err_ratelimited("handle_cap_export: issued != implemented: "
3626                                 "ino (%llx.%llx) mds%d seq %d mseq %d "
3627                                 "issued %s implemented %s\n",
3628                                 ceph_vinop(inode), mds, cap->seq, cap->mseq,
3629                                 ceph_cap_string(issued),
3630                                 ceph_cap_string(cap->implemented));
3631 
3632 
3633         tcap = __get_cap_for_mds(ci, target);
3634         if (tcap) {
3635                 /* already have caps from the target */
3636                 if (tcap->cap_id == t_cap_id &&
3637                     ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3638                         dout(" updating import cap %p mds%d\n", tcap, target);
3639                         tcap->cap_id = t_cap_id;
3640                         tcap->seq = t_seq - 1;
3641                         tcap->issue_seq = t_seq - 1;
3642                         tcap->issued |= issued;
3643                         tcap->implemented |= issued;
3644                         if (cap == ci->i_auth_cap)
3645                                 ci->i_auth_cap = tcap;
3646 
3647                         if (!list_empty(&ci->i_cap_flush_list) &&
3648                             ci->i_auth_cap == tcap) {
3649                                 spin_lock(&mdsc->cap_dirty_lock);
3650                                 list_move_tail(&ci->i_flushing_item,
3651                                                &tcap->session->s_cap_flushing);
3652                                 spin_unlock(&mdsc->cap_dirty_lock);
3653                         }
3654                 }
3655                 __ceph_remove_cap(cap, false);
3656                 goto out_unlock;
3657         } else if (tsession) {
3658                 /* add placeholder for the export tagert */
3659                 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3660                 tcap = new_cap;
3661                 ceph_add_cap(inode, tsession, t_cap_id, -1, issued, 0,
3662                              t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3663 
3664                 if (!list_empty(&ci->i_cap_flush_list) &&
3665                     ci->i_auth_cap == tcap) {
3666                         spin_lock(&mdsc->cap_dirty_lock);
3667                         list_move_tail(&ci->i_flushing_item,
3668                                        &tcap->session->s_cap_flushing);
3669                         spin_unlock(&mdsc->cap_dirty_lock);
3670                 }
3671 
3672                 __ceph_remove_cap(cap, false);
3673                 goto out_unlock;
3674         }
3675 
3676         spin_unlock(&ci->i_ceph_lock);
3677         mutex_unlock(&session->s_mutex);
3678 
3679         /* open target session */
3680         tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3681         if (!IS_ERR(tsession)) {
3682                 if (mds > target) {
3683                         mutex_lock(&session->s_mutex);
3684                         mutex_lock_nested(&tsession->s_mutex,
3685                                           SINGLE_DEPTH_NESTING);
3686                 } else {
3687                         mutex_lock(&tsession->s_mutex);
3688                         mutex_lock_nested(&session->s_mutex,
3689                                           SINGLE_DEPTH_NESTING);
3690                 }
3691                 new_cap = ceph_get_cap(mdsc, NULL);
3692         } else {
3693                 WARN_ON(1);
3694                 tsession = NULL;
3695                 target = -1;
3696                 mutex_lock(&session->s_mutex);
3697         }
3698         goto retry;
3699 
3700 out_unlock:
3701         spin_unlock(&ci->i_ceph_lock);
3702         mutex_unlock(&session->s_mutex);
3703         if (tsession) {
3704                 mutex_unlock(&tsession->s_mutex);
3705                 ceph_put_mds_session(tsession);
3706         }
3707         if (new_cap)
3708                 ceph_put_cap(mdsc, new_cap);
3709 }
3710 
3711 /*
3712  * Handle cap IMPORT.
3713  *
3714  * caller holds s_mutex. acquires i_ceph_lock
3715  */
3716 static void handle_cap_import(struct ceph_mds_client *mdsc,
3717                               struct inode *inode, struct ceph_mds_caps *im,
3718                               struct ceph_mds_cap_peer *ph,
3719                               struct ceph_mds_session *session,
3720                               struct ceph_cap **target_cap, int *old_issued)
3721         __acquires(ci->i_ceph_lock)
3722 {
3723         struct ceph_inode_info *ci = ceph_inode(inode);
3724         struct ceph_cap *cap, *ocap, *new_cap = NULL;
3725         int mds = session->s_mds;
3726         int issued;
3727         unsigned caps = le32_to_cpu(im->caps);
3728         unsigned wanted = le32_to_cpu(im->wanted);
3729         unsigned seq = le32_to_cpu(im->seq);
3730         unsigned mseq = le32_to_cpu(im->migrate_seq);
3731         u64 realmino = le64_to_cpu(im->realm);
3732         u64 cap_id = le64_to_cpu(im->cap_id);
3733         u64 p_cap_id;
3734         int peer;
3735 
3736         if (ph) {
3737                 p_cap_id = le64_to_cpu(ph->cap_id);
3738                 peer = le32_to_cpu(ph->mds);
3739         } else {
3740                 p_cap_id = 0;
3741                 peer = -1;
3742         }
3743 
3744         dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
3745              inode, ci, mds, mseq, peer);
3746 
3747 retry:
3748         spin_lock(&ci->i_ceph_lock);
3749         cap = __get_cap_for_mds(ci, mds);
3750         if (!cap) {
3751                 if (!new_cap) {
3752                         spin_unlock(&ci->i_ceph_lock);
3753                         new_cap = ceph_get_cap(mdsc, NULL);
3754                         goto retry;
3755                 }
3756                 cap = new_cap;
3757         } else {
3758                 if (new_cap) {
3759                         ceph_put_cap(mdsc, new_cap);
3760                         new_cap = NULL;
3761                 }
3762         }
3763 
3764         __ceph_caps_issued(ci, &issued);
3765         issued |= __ceph_caps_dirty(ci);
3766 
3767         ceph_add_cap(inode, session, cap_id, -1, caps, wanted, seq, mseq,
3768                      realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
3769 
3770         ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
3771         if (ocap && ocap->cap_id == p_cap_id) {
3772                 dout(" remove export cap %p mds%d flags %d\n",
3773                      ocap, peer, ph->flags);
3774                 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
3775                     (ocap->seq != le32_to_cpu(ph->seq) ||
3776                      ocap->mseq != le32_to_cpu(ph->mseq))) {
3777                         pr_err_ratelimited("handle_cap_import: "
3778                                         "mismatched seq/mseq: ino (%llx.%llx) "
3779                                         "mds%d seq %d mseq %d importer mds%d "
3780                                         "has peer seq %d mseq %d\n",
3781                                         ceph_vinop(inode), peer, ocap->seq,
3782                                         ocap->mseq, mds, le32_to_cpu(ph->seq),
3783                                         le32_to_cpu(ph->mseq));
3784                 }
3785                 __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
3786         }
3787 
3788         /* make sure we re-request max_size, if necessary */
3789         ci->i_requested_max_size = 0;
3790 
3791         *old_issued = issued;
3792         *target_cap = cap;
3793 }
3794 
3795 /*
3796  * Handle a caps message from the MDS.
3797  *
3798  * Identify the appropriate session, inode, and call the right handler
3799  * based on the cap op.
3800  */
3801 void ceph_handle_caps(struct ceph_mds_session *session,
3802                       struct ceph_msg *msg)
3803 {
3804         struct ceph_mds_client *mdsc = session->s_mdsc;
3805         struct inode *inode;
3806         struct ceph_inode_info *ci;
3807         struct ceph_cap *cap;
3808         struct ceph_mds_caps *h;
3809         struct ceph_mds_cap_peer *peer = NULL;
3810         struct ceph_snap_realm *realm = NULL;
3811         int op;
3812         int msg_version = le16_to_cpu(msg->hdr.version);
3813         u32 seq, mseq;
3814         struct ceph_vino vino;
3815         void *snaptrace;
3816         size_t snaptrace_len;
3817         void *p, *end;
3818         struct cap_extra_info extra_info = {};
3819 
3820         dout("handle_caps from mds%d\n", session->s_mds);
3821 
3822         /* decode */
3823         end = msg->front.iov_base + msg->front.iov_len;
3824         if (msg->front.iov_len < sizeof(*h))
3825                 goto bad;
3826         h = msg->front.iov_base;
3827         op = le32_to_cpu(h->op);
3828         vino.ino = le64_to_cpu(h->ino);
3829         vino.snap = CEPH_NOSNAP;
3830         seq = le32_to_cpu(h->seq);
3831         mseq = le32_to_cpu(h->migrate_seq);
3832 
3833         snaptrace = h + 1;
3834         snaptrace_len = le32_to_cpu(h->snap_trace_len);
3835         p = snaptrace + snaptrace_len;
3836 
3837         if (msg_version >= 2) {
3838                 u32 flock_len;
3839                 ceph_decode_32_safe(&p, end, flock_len, bad);
3840                 if (p + flock_len > end)
3841                         goto bad;
3842                 p += flock_len;
3843         }
3844 
3845         if (msg_version >= 3) {
3846                 if (op == CEPH_CAP_OP_IMPORT) {
3847                         if (p + sizeof(*peer) > end)
3848                                 goto bad;
3849                         peer = p;
3850                         p += sizeof(*peer);
3851                 } else if (op == CEPH_CAP_OP_EXPORT) {
3852                         /* recorded in unused fields */
3853                         peer = (void *)&h->size;
3854                 }
3855         }
3856 
3857         if (msg_version >= 4) {
3858                 ceph_decode_64_safe(&p, end, extra_info.inline_version, bad);
3859                 ceph_decode_32_safe(&p, end, extra_info.inline_len, bad);
3860                 if (p + extra_info.inline_len > end)
3861                         goto bad;
3862                 extra_info.inline_data = p;
3863                 p += extra_info.inline_len;
3864         }
3865 
3866         if (msg_version >= 5) {
3867                 struct ceph_osd_client  *osdc = &mdsc->fsc->client->osdc;
3868                 u32                     epoch_barrier;
3869 
3870                 ceph_decode_32_safe(&p, end, epoch_barrier, bad);
3871                 ceph_osdc_update_epoch_barrier(osdc, epoch_barrier);
3872         }
3873 
3874         if (msg_version >= 8) {
3875                 u64 flush_tid;
3876                 u32 caller_uid, caller_gid;
3877                 u32 pool_ns_len;
3878 
3879                 /* version >= 6 */
3880                 ceph_decode_64_safe(&p, end, flush_tid, bad);
3881                 /* version >= 7 */
3882                 ceph_decode_32_safe(&p, end, caller_uid, bad);
3883                 ceph_decode_32_safe(&p, end, caller_gid, bad);
3884                 /* version >= 8 */
3885                 ceph_decode_32_safe(&p, end, pool_ns_len, bad);
3886                 if (pool_ns_len > 0) {
3887                         ceph_decode_need(&p, end, pool_ns_len, bad);
3888                         extra_info.pool_ns =
3889                                 ceph_find_or_create_string(p, pool_ns_len);
3890                         p += pool_ns_len;
3891                 }
3892         }
3893 
3894         if (msg_version >= 9) {
3895                 struct ceph_timespec *btime;
3896 
3897                 if (p + sizeof(*btime) > end)
3898                         goto bad;
3899                 btime = p;
3900                 ceph_decode_timespec64(&extra_info.btime, btime);
3901                 p += sizeof(*btime);
3902                 ceph_decode_64_safe(&p, end, extra_info.change_attr, bad);
3903         }
3904 
3905         if (msg_version >= 11) {
3906                 u32 flags;
3907                 /* version >= 10 */
3908                 ceph_decode_32_safe(&p, end, flags, bad);
3909                 /* version >= 11 */
3910                 extra_info.dirstat_valid = true;
3911                 ceph_decode_64_safe(&p, end, extra_info.nfiles, bad);
3912                 ceph_decode_64_safe(&p, end, extra_info.nsubdirs, bad);
3913         }
3914 
3915         /* lookup ino */
3916         inode = ceph_find_inode(mdsc->fsc->sb, vino);
3917         ci = ceph_inode(inode);
3918         dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
3919              vino.snap, inode);
3920 
3921         mutex_lock(&session->s_mutex);
3922         session->s_seq++;
3923         dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
3924              (unsigned)seq);
3925 
3926         if (!inode) {
3927                 dout(" i don't have ino %llx\n", vino.ino);
3928 
3929                 if (op == CEPH_CAP_OP_IMPORT) {
3930                         cap = ceph_get_cap(mdsc, NULL);
3931                         cap->cap_ino = vino.ino;
3932                         cap->queue_release = 1;
3933                         cap->cap_id = le64_to_cpu(h->cap_id);
3934                         cap->mseq = mseq;
3935                         cap->seq = seq;
3936                         cap->issue_seq = seq;
3937                         spin_lock(&session->s_cap_lock);
3938                         __ceph_queue_cap_release(session, cap);
3939                         spin_unlock(&session->s_cap_lock);
3940                 }
3941                 goto flush_cap_releases;
3942         }
3943 
3944         /* these will work even if we don't have a cap yet */
3945         switch (op) {
3946         case CEPH_CAP_OP_FLUSHSNAP_ACK:
3947                 handle_cap_flushsnap_ack(inode, le64_to_cpu(msg->hdr.tid),
3948                                          h, session);
3949                 goto done;
3950 
3951         case CEPH_CAP_OP_EXPORT:
3952                 handle_cap_export(inode, h, peer, session);
3953                 goto done_unlocked;
3954 
3955         case CEPH_CAP_OP_IMPORT:
3956                 realm = NULL;
3957                 if (snaptrace_len) {
3958                         down_write(&mdsc->snap_rwsem);
3959                         ceph_update_snap_trace(mdsc, snaptrace,
3960                                                snaptrace + snaptrace_len,
3961                                                false, &realm);
3962                         downgrade_write(&mdsc->snap_rwsem);
3963                 } else {
3964                         down_read(&mdsc->snap_rwsem);
3965                 }
3966                 handle_cap_import(mdsc, inode, h, peer, session,
3967                                   &cap, &extra_info.issued);
3968                 handle_cap_grant(inode, session, cap,
3969                                  h, msg->middle, &extra_info);
3970                 if (realm)
3971                         ceph_put_snap_realm(mdsc, realm);
3972                 goto done_unlocked;
3973         }
3974 
3975         /* the rest require a cap */
3976         spin_lock(&ci->i_ceph_lock);
3977         cap = __get_cap_for_mds(ceph_inode(inode), session->s_mds);
3978         if (!cap) {
3979                 dout(" no cap on %p ino %llx.%llx from mds%d\n",
3980                      inode, ceph_ino(inode), ceph_snap(inode),
3981                      session->s_mds);
3982                 spin_unlock(&ci->i_ceph_lock);
3983                 goto flush_cap_releases;
3984         }
3985 
3986         /* note that each of these drops i_ceph_lock for us */
3987         switch (op) {
3988         case CEPH_CAP_OP_REVOKE:
3989         case CEPH_CAP_OP_GRANT:
3990                 __ceph_caps_issued(ci, &extra_info.issued);
3991                 extra_info.issued |= __ceph_caps_dirty(ci);
3992                 handle_cap_grant(inode, session, cap,
3993                                  h, msg->middle, &extra_info);
3994                 goto done_unlocked;
3995 
3996         case CEPH_CAP_OP_FLUSH_ACK:
3997                 handle_cap_flush_ack(inode, le64_to_cpu(msg->hdr.tid),
3998                                      h, session, cap);
3999                 break;
4000 
4001         case CEPH_CAP_OP_TRUNC:
4002                 handle_cap_trunc(inode, h, session);
4003                 break;
4004 
4005         default:
4006                 spin_unlock(&ci->i_ceph_lock);
4007                 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
4008                        ceph_cap_op_name(op));
4009         }
4010 
4011 done:
4012         mutex_unlock(&session->s_mutex);
4013 done_unlocked:
4014         ceph_put_string(extra_info.pool_ns);
4015         /* avoid calling iput_final() in mds dispatch threads */
4016         ceph_async_iput(inode);
4017         return;
4018 
4019 flush_cap_releases:
4020         /*
4021          * send any cap release message to try to move things
4022          * along for the mds (who clearly thinks we still have this
4023          * cap).
4024          */
4025         ceph_flush_cap_releases(mdsc, session);
4026         goto done;
4027 
4028 bad:
4029         pr_err("ceph_handle_caps: corrupt message\n");
4030         ceph_msg_dump(msg);
4031         return;
4032 }
4033 
4034 /*
4035  * Delayed work handler to process end of delayed cap release LRU list.
4036  */
4037 void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
4038 {
4039         struct inode *inode;
4040         struct ceph_inode_info *ci;
4041         int flags = CHECK_CAPS_NODELAY;
4042 
4043         dout("check_delayed_caps\n");
4044         while (1) {
4045                 spin_lock(&mdsc->cap_delay_lock);
4046                 if (list_empty(&mdsc->cap_delay_list))
4047                         break;
4048                 ci = list_first_entry(&mdsc->cap_delay_list,
4049                                       struct ceph_inode_info,
4050                                       i_cap_delay_list);
4051                 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
4052                     time_before(jiffies, ci->i_hold_caps_max))
4053                         break;
4054                 list_del_init(&ci->i_cap_delay_list);
4055 
4056                 inode = igrab(&ci->vfs_inode);
4057                 spin_unlock(&mdsc->cap_delay_lock);
4058 
4059                 if (inode) {
4060                         dout("check_delayed_caps on %p\n", inode);
4061                         ceph_check_caps(ci, flags, NULL);
4062                         /* avoid calling iput_final() in tick thread */
4063                         ceph_async_iput(inode);
4064                 }
4065         }
4066         spin_unlock(&mdsc->cap_delay_lock);
4067 }
4068 
4069 /*
4070  * Flush all dirty caps to the mds
4071  */
4072 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
4073 {
4074         struct ceph_inode_info *ci;
4075         struct inode *inode;
4076 
4077         dout("flush_dirty_caps\n");
4078         spin_lock(&mdsc->cap_dirty_lock);
4079         while (!list_empty(&mdsc->cap_dirty)) {
4080                 ci = list_first_entry(&mdsc->cap_dirty, struct ceph_inode_info,
4081                                       i_dirty_item);
4082                 inode = &ci->vfs_inode;
4083                 ihold(inode);
4084                 dout("flush_dirty_caps %p\n", inode);
4085                 spin_unlock(&mdsc->cap_dirty_lock);
4086                 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH, NULL);
4087                 iput(inode);
4088                 spin_lock(&mdsc->cap_dirty_lock);
4089         }
4090         spin_unlock(&mdsc->cap_dirty_lock);
4091         dout("flush_dirty_caps done\n");
4092 }
4093 
4094 void __ceph_get_fmode(struct ceph_inode_info *ci, int fmode)
4095 {
4096         int i;
4097         int bits = (fmode << 1) | 1;
4098         for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4099                 if (bits & (1 << i))
4100                         ci->i_nr_by_mode[i]++;
4101         }
4102 }
4103 
4104 /*
4105  * Drop open file reference.  If we were the last open file,
4106  * we may need to release capabilities to the MDS (or schedule
4107  * their delayed release).
4108  */
4109 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode)
4110 {
4111         int i, last = 0;
4112         int bits = (fmode << 1) | 1;
4113         spin_lock(&ci->i_ceph_lock);
4114         for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4115                 if (bits & (1 << i)) {
4116                         BUG_ON(ci->i_nr_by_mode[i] == 0);
4117                         if (--ci->i_nr_by_mode[i] == 0)
4118                                 last++;
4119                 }
4120         }
4121         dout("put_fmode %p fmode %d {%d,%d,%d,%d}\n",
4122              &ci->vfs_inode, fmode,
4123              ci->i_nr_by_mode[0], ci->i_nr_by_mode[1],
4124              ci->i_nr_by_mode[2], ci->i_nr_by_mode[3]);
4125         spin_unlock(&ci->i_ceph_lock);
4126 
4127         if (last && ci->i_vino.snap == CEPH_NOSNAP)
4128                 ceph_check_caps(ci, 0, NULL);
4129 }
4130 
4131 /*
4132  * For a soon-to-be unlinked file, drop the LINK caps. If it
4133  * looks like the link count will hit 0, drop any other caps (other
4134  * than PIN) we don't specifically want (due to the file still being
4135  * open).
4136  */
4137 int ceph_drop_caps_for_unlink(struct inode *inode)
4138 {
4139         struct ceph_inode_info *ci = ceph_inode(inode);
4140         int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
4141 
4142         spin_lock(&ci->i_ceph_lock);
4143         if (inode->i_nlink == 1) {
4144                 drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN);
4145 
4146                 ci->i_ceph_flags |= CEPH_I_NODELAY;
4147                 if (__ceph_caps_dirty(ci)) {
4148                         struct ceph_mds_client *mdsc =
4149                                 ceph_inode_to_client(inode)->mdsc;
4150                         __cap_delay_requeue_front(mdsc, ci);
4151                 }
4152         }
4153         spin_unlock(&ci->i_ceph_lock);
4154         return drop;
4155 }
4156 
4157 /*
4158  * Helpers for embedding cap and dentry lease releases into mds
4159  * requests.
4160  *
4161  * @force is used by dentry_release (below) to force inclusion of a
4162  * record for the directory inode, even when there aren't any caps to
4163  * drop.
4164  */
4165 int ceph_encode_inode_release(void **p, struct inode *inode,
4166                               int mds, int drop, int unless, int force)
4167 {
4168         struct ceph_inode_info *ci = ceph_inode(inode);
4169         struct ceph_cap *cap;
4170         struct ceph_mds_request_release *rel = *p;
4171         int used, dirty;
4172         int ret = 0;
4173 
4174         spin_lock(&ci->i_ceph_lock);
4175         used = __ceph_caps_used(ci);
4176         dirty = __ceph_caps_dirty(ci);
4177 
4178         dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
4179              inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
4180              ceph_cap_string(unless));
4181 
4182         /* only drop unused, clean caps */
4183         drop &= ~(used | dirty);
4184 
4185         cap = __get_cap_for_mds(ci, mds);
4186         if (cap && __cap_is_valid(cap)) {
4187                 unless &= cap->issued;
4188                 if (unless) {
4189                         if (unless & CEPH_CAP_AUTH_EXCL)
4190                                 drop &= ~CEPH_CAP_AUTH_SHARED;
4191                         if (unless & CEPH_CAP_LINK_EXCL)
4192                                 drop &= ~CEPH_CAP_LINK_SHARED;
4193                         if (unless & CEPH_CAP_XATTR_EXCL)
4194                                 drop &= ~CEPH_CAP_XATTR_SHARED;
4195                         if (unless & CEPH_CAP_FILE_EXCL)
4196                                 drop &= ~CEPH_CAP_FILE_SHARED;
4197                 }
4198 
4199                 if (force || (cap->issued & drop)) {
4200                         if (cap->issued & drop) {
4201                                 int wanted = __ceph_caps_wanted(ci);
4202                                 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0)
4203                                         wanted |= cap->mds_wanted;
4204                                 dout("encode_inode_release %p cap %p "
4205                                      "%s -> %s, wanted %s -> %s\n", inode, cap,
4206                                      ceph_cap_string(cap->issued),
4207                                      ceph_cap_string(cap->issued & ~drop),
4208                                      ceph_cap_string(cap->mds_wanted),
4209                                      ceph_cap_string(wanted));
4210 
4211                                 cap->issued &= ~drop;
4212                                 cap->implemented &= ~drop;
4213                                 cap->mds_wanted = wanted;
4214                         } else {
4215                                 dout("encode_inode_release %p cap %p %s"
4216                                      " (force)\n", inode, cap,
4217                                      ceph_cap_string(cap->issued));
4218                         }
4219 
4220                         rel->ino = cpu_to_le64(ceph_ino(inode));
4221                         rel->cap_id = cpu_to_le64(cap->cap_id);
4222                         rel->seq = cpu_to_le32(cap->seq);
4223                         rel->issue_seq = cpu_to_le32(cap->issue_seq);
4224                         rel->mseq = cpu_to_le32(cap->mseq);
4225                         rel->caps = cpu_to_le32(cap->implemented);
4226                         rel->wanted = cpu_to_le32(cap->mds_wanted);
4227                         rel->dname_len = 0;
4228                         rel->dname_seq = 0;
4229                         *p += sizeof(*rel);
4230                         ret = 1;
4231                 } else {
4232                         dout("encode_inode_release %p cap %p %s (noop)\n",
4233                              inode, cap, ceph_cap_string(cap->issued));
4234                 }
4235         }
4236         spin_unlock(&ci->i_ceph_lock);
4237         return ret;
4238 }
4239 
4240 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
4241                                struct inode *dir,
4242                                int mds, int drop, int unless)
4243 {
4244         struct dentry *parent = NULL;
4245         struct ceph_mds_request_release *rel = *p;
4246         struct ceph_dentry_info *di = ceph_dentry(dentry);
4247         int force = 0;
4248         int ret;
4249 
4250         /*
4251          * force an record for the directory caps if we have a dentry lease.
4252          * this is racy (can't take i_ceph_lock and d_lock together), but it
4253          * doesn't have to be perfect; the mds will revoke anything we don't
4254          * release.
4255          */
4256         spin_lock(&dentry->d_lock);
4257         if (di->lease_session && di->lease_session->s_mds == mds)
4258                 force = 1;
4259         if (!dir) {
4260                 parent = dget(dentry->d_parent);
4261                 dir = d_inode(parent);
4262         }
4263         spin_unlock(&dentry->d_lock);
4264 
4265         ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
4266         dput(parent);
4267 
4268         spin_lock(&dentry->d_lock);
4269         if (ret && di->lease_session && di->lease_session->s_mds == mds) {
4270                 dout("encode_dentry_release %p mds%d seq %d\n",
4271                      dentry, mds, (int)di->lease_seq);
4272                 rel->dname_len = cpu_to_le32(dentry->d_name.len);
4273                 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
4274                 *p += dentry->d_name.len;
4275                 rel->dname_seq = cpu_to_le32(di->lease_seq);
4276                 __ceph_mdsc_drop_dentry_lease(dentry);
4277         }
4278         spin_unlock(&dentry->d_lock);
4279         return ret;
4280 }

/* [<][>][^][v][top][bottom][index][help] */