fs/ceph/mds_client.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/ceph/ceph_debug.h>
   3
   4 #include <linux/fs.h>
   5 #include <linux/wait.h>
   6 #include <linux/slab.h>
   7 #include <linux/gfp.h>
   8 #include <linux/sched.h>
   9 #include <linux/debugfs.h>
  10 #include <linux/seq_file.h>
  11 #include <linux/ratelimit.h>
  12 #include <linux/bits.h>
  13 #include <linux/ktime.h>
  14
  15 #include "super.h"
  16 #include "mds_client.h"
  17
  18 #include <linux/ceph/ceph_features.h>
  19 #include <linux/ceph/messenger.h>
  20 #include <linux/ceph/decode.h>
  21 #include <linux/ceph/pagelist.h>
  22 #include <linux/ceph/auth.h>
  23 #include <linux/ceph/debugfs.h>
  24
  25 #define RECONNECT_MAX_SIZE (INT_MAX - PAGE_SIZE)
  26
  27 /*
  28  * A cluster of MDS (metadata server) daemons is responsible for
  29  * managing the file system namespace (the directory hierarchy and
  30  * inodes) and for coordinating shared access to storage.  Metadata is
  31  * partitioning hierarchically across a number of servers, and that
  32  * partition varies over time as the cluster adjusts the distribution
  33  * in order to balance load.
  34  *
  35  * The MDS client is primarily responsible to managing synchronous
  36  * metadata requests for operations like open, unlink, and so forth.
  37  * If there is a MDS failure, we find out about it when we (possibly
  38  * request and) receive a new MDS map, and can resubmit affected
  39  * requests.
  40  *
  41  * For the most part, though, we take advantage of a lossless
  42  * communications channel to the MDS, and do not need to worry about
  43  * timing out or resubmitting requests.
  44  *
  45  * We maintain a stateful "session" with each MDS we interact with.
  46  * Within each session, we sent periodic heartbeat messages to ensure
  47  * any capabilities or leases we have been issues remain valid.  If
  48  * the session times out and goes stale, our leases and capabilities
  49  * are no longer valid.
  50  */
  51
  52 struct ceph_reconnect_state {
  53         struct ceph_mds_session *session;
  54         int nr_caps, nr_realms;
  55         struct ceph_pagelist *pagelist;
  56         unsigned msg_version;
  57         bool allow_multi;
  58 };
  59
  60 static void __wake_requests(struct ceph_mds_client *mdsc,
  61                             struct list_head *head);
  62 static void ceph_cap_release_work(struct work_struct *work);
  63 static void ceph_cap_reclaim_work(struct work_struct *work);
  64
  65 static const struct ceph_connection_operations mds_con_ops;
  66
  67
  68 /*
  69  * mds reply parsing
  70  */
  71
  72 static int parse_reply_info_quota(void **p, void *end,
  73                                   struct ceph_mds_reply_info_in *info)
  74 {
  75         u8 struct_v, struct_compat;
  76         u32 struct_len;
  77
  78         ceph_decode_8_safe(p, end, struct_v, bad);
  79         ceph_decode_8_safe(p, end, struct_compat, bad);
  80         /* struct_v is expected to be >= 1. we only
  81          * understand encoding with struct_compat == 1. */
  82         if (!struct_v || struct_compat != 1)
  83                 goto bad;
  84         ceph_decode_32_safe(p, end, struct_len, bad);
  85         ceph_decode_need(p, end, struct_len, bad);
  86         end = *p + struct_len;
  87         ceph_decode_64_safe(p, end, info->max_bytes, bad);
  88         ceph_decode_64_safe(p, end, info->max_files, bad);
  89         *p = end;
  90         return 0;
  91 bad:
  92         return -EIO;
  93 }
  94
  95 /*
  96  * parse individual inode info
  97  */
  98 static int parse_reply_info_in(void **p, void *end,
  99                                struct ceph_mds_reply_info_in *info,
 100                                u64 features)
 101 {
 102         int err = 0;
 103         u8 struct_v = 0;
 104
 105         if (features == (u64)-1) {
 106                 u32 struct_len;
 107                 u8 struct_compat;
 108                 ceph_decode_8_safe(p, end, struct_v, bad);
 109                 ceph_decode_8_safe(p, end, struct_compat, bad);
 110                 /* struct_v is expected to be >= 1. we only understand
 111                  * encoding with struct_compat == 1. */
 112                 if (!struct_v || struct_compat != 1)
 113                         goto bad;
 114                 ceph_decode_32_safe(p, end, struct_len, bad);
 115                 ceph_decode_need(p, end, struct_len, bad);
 116                 end = *p + struct_len;
 117         }
 118
 119         ceph_decode_need(p, end, sizeof(struct ceph_mds_reply_inode), bad);
 120         info->in = *p;
 121         *p += sizeof(struct ceph_mds_reply_inode) +
 122                 sizeof(*info->in->fragtree.splits) *
 123                 le32_to_cpu(info->in->fragtree.nsplits);
 124
 125         ceph_decode_32_safe(p, end, info->symlink_len, bad);
 126         ceph_decode_need(p, end, info->symlink_len, bad);
 127         info->symlink = *p;
 128         *p += info->symlink_len;
 129
 130         ceph_decode_copy_safe(p, end, &info->dir_layout,
 131                               sizeof(info->dir_layout), bad);
 132         ceph_decode_32_safe(p, end, info->xattr_len, bad);
 133         ceph_decode_need(p, end, info->xattr_len, bad);
 134         info->xattr_data = *p;
 135         *p += info->xattr_len;
 136
 137         if (features == (u64)-1) {
 138                 /* inline data */
 139                 ceph_decode_64_safe(p, end, info->inline_version, bad);
 140                 ceph_decode_32_safe(p, end, info->inline_len, bad);
 141                 ceph_decode_need(p, end, info->inline_len, bad);
 142                 info->inline_data = *p;
 143                 *p += info->inline_len;
 144                 /* quota */
 145                 err = parse_reply_info_quota(p, end, info);
 146                 if (err < 0)
 147                         goto out_bad;
 148                 /* pool namespace */
 149                 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
 150                 if (info->pool_ns_len > 0) {
 151                         ceph_decode_need(p, end, info->pool_ns_len, bad);
 152                         info->pool_ns_data = *p;
 153                         *p += info->pool_ns_len;
 154                 }
 155
 156                 /* btime */
 157                 ceph_decode_need(p, end, sizeof(info->btime), bad);
 158                 ceph_decode_copy(p, &info->btime, sizeof(info->btime));
 159
 160                 /* change attribute */
 161                 ceph_decode_64_safe(p, end, info->change_attr, bad);
 162
 163                 /* dir pin */
 164                 if (struct_v >= 2) {
 165                         ceph_decode_32_safe(p, end, info->dir_pin, bad);
 166                 } else {
 167                         info->dir_pin = -ENODATA;
 168                 }
 169
 170                 /* snapshot birth time, remains zero for v<=2 */
 171                 if (struct_v >= 3) {
 172                         ceph_decode_need(p, end, sizeof(info->snap_btime), bad);
 173                         ceph_decode_copy(p, &info->snap_btime,
 174                                          sizeof(info->snap_btime));
 175                 } else {
 176                         memset(&info->snap_btime, 0, sizeof(info->snap_btime));
 177                 }
 178
 179                 *p = end;
 180         } else {
 181                 if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
 182                         ceph_decode_64_safe(p, end, info->inline_version, bad);
 183                         ceph_decode_32_safe(p, end, info->inline_len, bad);
 184                         ceph_decode_need(p, end, info->inline_len, bad);
 185                         info->inline_data = *p;
 186                         *p += info->inline_len;
 187                 } else
 188                         info->inline_version = CEPH_INLINE_NONE;
 189
 190                 if (features & CEPH_FEATURE_MDS_QUOTA) {
 191                         err = parse_reply_info_quota(p, end, info);
 192                         if (err < 0)
 193                                 goto out_bad;
 194                 } else {
 195                         info->max_bytes = 0;
 196                         info->max_files = 0;
 197                 }
 198
 199                 info->pool_ns_len = 0;
 200                 info->pool_ns_data = NULL;
 201                 if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
 202                         ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
 203                         if (info->pool_ns_len > 0) {
 204                                 ceph_decode_need(p, end, info->pool_ns_len, bad);
 205                                 info->pool_ns_data = *p;
 206                                 *p += info->pool_ns_len;
 207                         }
 208                 }
 209
 210                 if (features & CEPH_FEATURE_FS_BTIME) {
 211                         ceph_decode_need(p, end, sizeof(info->btime), bad);
 212                         ceph_decode_copy(p, &info->btime, sizeof(info->btime));
 213                         ceph_decode_64_safe(p, end, info->change_attr, bad);
 214                 }
 215
 216                 info->dir_pin = -ENODATA;
 217                 /* info->snap_btime remains zero */
 218         }
 219         return 0;
 220 bad:
 221         err = -EIO;
 222 out_bad:
 223         return err;
 224 }
 225
 226 static int parse_reply_info_dir(void **p, void *end,
 227                                 struct ceph_mds_reply_dirfrag **dirfrag,
 228                                 u64 features)
 229 {
 230         if (features == (u64)-1) {
 231                 u8 struct_v, struct_compat;
 232                 u32 struct_len;
 233                 ceph_decode_8_safe(p, end, struct_v, bad);
 234                 ceph_decode_8_safe(p, end, struct_compat, bad);
 235                 /* struct_v is expected to be >= 1. we only understand
 236                  * encoding whose struct_compat == 1. */
 237                 if (!struct_v || struct_compat != 1)
 238                         goto bad;
 239                 ceph_decode_32_safe(p, end, struct_len, bad);
 240                 ceph_decode_need(p, end, struct_len, bad);
 241                 end = *p + struct_len;
 242         }
 243
 244         ceph_decode_need(p, end, sizeof(**dirfrag), bad);
 245         *dirfrag = *p;
 246         *p += sizeof(**dirfrag) + sizeof(u32) * le32_to_cpu((*dirfrag)->ndist);
 247         if (unlikely(*p > end))
 248                 goto bad;
 249         if (features == (u64)-1)
 250                 *p = end;
 251         return 0;
 252 bad:
 253         return -EIO;
 254 }
 255
 256 static int parse_reply_info_lease(void **p, void *end,
 257                                   struct ceph_mds_reply_lease **lease,
 258                                   u64 features)
 259 {
 260         if (features == (u64)-1) {
 261                 u8 struct_v, struct_compat;
 262                 u32 struct_len;
 263                 ceph_decode_8_safe(p, end, struct_v, bad);
 264                 ceph_decode_8_safe(p, end, struct_compat, bad);
 265                 /* struct_v is expected to be >= 1. we only understand
 266                  * encoding whose struct_compat == 1. */
 267                 if (!struct_v || struct_compat != 1)
 268                         goto bad;
 269                 ceph_decode_32_safe(p, end, struct_len, bad);
 270                 ceph_decode_need(p, end, struct_len, bad);
 271                 end = *p + struct_len;
 272         }
 273
 274         ceph_decode_need(p, end, sizeof(**lease), bad);
 275         *lease = *p;
 276         *p += sizeof(**lease);
 277         if (features == (u64)-1)
 278                 *p = end;
 279         return 0;
 280 bad:
 281         return -EIO;
 282 }
 283
 284 /*
 285  * parse a normal reply, which may contain a (dir+)dentry and/or a
 286  * target inode.
 287  */
 288 static int parse_reply_info_trace(void **p, void *end,
 289                                   struct ceph_mds_reply_info_parsed *info,
 290                                   u64 features)
 291 {
 292         int err;
 293
 294         if (info->head->is_dentry) {
 295                 err = parse_reply_info_in(p, end, &info->diri, features);
 296                 if (err < 0)
 297                         goto out_bad;
 298
 299                 err = parse_reply_info_dir(p, end, &info->dirfrag, features);
 300                 if (err < 0)
 301                         goto out_bad;
 302
 303                 ceph_decode_32_safe(p, end, info->dname_len, bad);
 304                 ceph_decode_need(p, end, info->dname_len, bad);
 305                 info->dname = *p;
 306                 *p += info->dname_len;
 307
 308                 err = parse_reply_info_lease(p, end, &info->dlease, features);
 309                 if (err < 0)
 310                         goto out_bad;
 311         }
 312
 313         if (info->head->is_target) {
 314                 err = parse_reply_info_in(p, end, &info->targeti, features);
 315                 if (err < 0)
 316                         goto out_bad;
 317         }
 318
 319         if (unlikely(*p != end))
 320                 goto bad;
 321         return 0;
 322
 323 bad:
 324         err = -EIO;
 325 out_bad:
 326         pr_err("problem parsing mds trace %d\n", err);
 327         return err;
 328 }
 329
 330 /*
 331  * parse readdir results
 332  */
 333 static int parse_reply_info_readdir(void **p, void *end,
 334                                 struct ceph_mds_reply_info_parsed *info,
 335                                 u64 features)
 336 {
 337         u32 num, i = 0;
 338         int err;
 339
 340         err = parse_reply_info_dir(p, end, &info->dir_dir, features);
 341         if (err < 0)
 342                 goto out_bad;
 343
 344         ceph_decode_need(p, end, sizeof(num) + 2, bad);
 345         num = ceph_decode_32(p);
 346         {
 347                 u16 flags = ceph_decode_16(p);
 348                 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
 349                 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
 350                 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
 351                 info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
 352         }
 353         if (num == 0)
 354                 goto done;
 355
 356         BUG_ON(!info->dir_entries);
 357         if ((unsigned long)(info->dir_entries + num) >
 358             (unsigned long)info->dir_entries + info->dir_buf_size) {
 359                 pr_err("dir contents are larger than expected\n");
 360                 WARN_ON(1);
 361                 goto bad;
 362         }
 363
 364         info->dir_nr = num;
 365         while (num) {
 366                 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
 367                 /* dentry */
 368                 ceph_decode_32_safe(p, end, rde->name_len, bad);
 369                 ceph_decode_need(p, end, rde->name_len, bad);
 370                 rde->name = *p;
 371                 *p += rde->name_len;
 372                 dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
 373
 374                 /* dentry lease */
 375                 err = parse_reply_info_lease(p, end, &rde->lease, features);
 376                 if (err)
 377                         goto out_bad;
 378                 /* inode */
 379                 err = parse_reply_info_in(p, end, &rde->inode, features);
 380                 if (err < 0)
 381                         goto out_bad;
 382                 /* ceph_readdir_prepopulate() will update it */
 383                 rde->offset = 0;
 384                 i++;
 385                 num--;
 386         }
 387
 388 done:
 389         /* Skip over any unrecognized fields */
 390         *p = end;
 391         return 0;
 392
 393 bad:
 394         err = -EIO;
 395 out_bad:
 396         pr_err("problem parsing dir contents %d\n", err);
 397         return err;
 398 }
 399
 400 /*
 401  * parse fcntl F_GETLK results
 402  */
 403 static int parse_reply_info_filelock(void **p, void *end,
 404                                      struct ceph_mds_reply_info_parsed *info,
 405                                      u64 features)
 406 {
 407         if (*p + sizeof(*info->filelock_reply) > end)
 408                 goto bad;
 409
 410         info->filelock_reply = *p;
 411
 412         /* Skip over any unrecognized fields */
 413         *p = end;
 414         return 0;
 415 bad:
 416         return -EIO;
 417 }
 418
 419
 420 #if BITS_PER_LONG == 64
 421
 422 #define DELEGATED_INO_AVAILABLE         xa_mk_value(1)
 423
 424 static int ceph_parse_deleg_inos(void **p, void *end,
 425                                  struct ceph_mds_session *s)
 426 {
 427         u32 sets;
 428
 429         ceph_decode_32_safe(p, end, sets, bad);
 430         dout("got %u sets of delegated inodes\n", sets);
 431         while (sets--) {
 432                 u64 start, len, ino;
 433
 434                 ceph_decode_64_safe(p, end, start, bad);
 435                 ceph_decode_64_safe(p, end, len, bad);
 436                 while (len--) {
 437                         int err = xa_insert(&s->s_delegated_inos, ino = start++,
 438                                             DELEGATED_INO_AVAILABLE,
 439                                             GFP_KERNEL);
 440                         if (!err) {
 441                                 dout("added delegated inode 0x%llx\n",
 442                                      start - 1);
 443                         } else if (err == -EBUSY) {
 444                                 pr_warn("ceph: MDS delegated inode 0x%llx more than once.\n",
 445                                         start - 1);
 446                         } else {
 447                                 return err;
 448                         }
 449                 }
 450         }
 451         return 0;
 452 bad:
 453         return -EIO;
 454 }
 455
 456 u64 ceph_get_deleg_ino(struct ceph_mds_session *s)
 457 {
 458         unsigned long ino;
 459         void *val;
 460
 461         xa_for_each(&s->s_delegated_inos, ino, val) {
 462                 val = xa_erase(&s->s_delegated_inos, ino);
 463                 if (val == DELEGATED_INO_AVAILABLE)
 464                         return ino;
 465         }
 466         return 0;
 467 }
 468
 469 int ceph_restore_deleg_ino(struct ceph_mds_session *s, u64 ino)
 470 {
 471         return xa_insert(&s->s_delegated_inos, ino, DELEGATED_INO_AVAILABLE,
 472                          GFP_KERNEL);
 473 }
 474 #else /* BITS_PER_LONG == 64 */
 475 /*
 476  * FIXME: xarrays can't handle 64-bit indexes on a 32-bit arch. For now, just
 477  * ignore delegated_inos on 32 bit arch. Maybe eventually add xarrays for top
 478  * and bottom words?
 479  */
 480 static int ceph_parse_deleg_inos(void **p, void *end,
 481                                  struct ceph_mds_session *s)
 482 {
 483         u32 sets;
 484
 485         ceph_decode_32_safe(p, end, sets, bad);
 486         if (sets)
 487                 ceph_decode_skip_n(p, end, sets * 2 * sizeof(__le64), bad);
 488         return 0;
 489 bad:
 490         return -EIO;
 491 }
 492
 493 u64 ceph_get_deleg_ino(struct ceph_mds_session *s)
 494 {
 495         return 0;
 496 }
 497
 498 int ceph_restore_deleg_ino(struct ceph_mds_session *s, u64 ino)
 499 {
 500         return 0;
 501 }
 502 #endif /* BITS_PER_LONG == 64 */
 503
 504 /*
 505  * parse create results
 506  */
 507 static int parse_reply_info_create(void **p, void *end,
 508                                   struct ceph_mds_reply_info_parsed *info,
 509                                   u64 features, struct ceph_mds_session *s)
 510 {
 511         int ret;
 512
 513         if (features == (u64)-1 ||
 514             (features & CEPH_FEATURE_REPLY_CREATE_INODE)) {
 515                 if (*p == end) {
 516                         /* Malformed reply? */
 517                         info->has_create_ino = false;
 518                 } else if (test_bit(CEPHFS_FEATURE_DELEG_INO, &s->s_features)) {
 519                         info->has_create_ino = true;
 520                         /* struct_v, struct_compat, and len */
 521                         ceph_decode_skip_n(p, end, 2 + sizeof(u32), bad);
 522                         ceph_decode_64_safe(p, end, info->ino, bad);
 523                         ret = ceph_parse_deleg_inos(p, end, s);
 524                         if (ret)
 525                                 return ret;
 526                 } else {
 527                         /* legacy */
 528                         ceph_decode_64_safe(p, end, info->ino, bad);
 529                         info->has_create_ino = true;
 530                 }
 531         } else {
 532                 if (*p != end)
 533                         goto bad;
 534         }
 535
 536         /* Skip over any unrecognized fields */
 537         *p = end;
 538         return 0;
 539 bad:
 540         return -EIO;
 541 }
 542
 543 /*
 544  * parse extra results
 545  */
 546 static int parse_reply_info_extra(void **p, void *end,
 547                                   struct ceph_mds_reply_info_parsed *info,
 548                                   u64 features, struct ceph_mds_session *s)
 549 {
 550         u32 op = le32_to_cpu(info->head->op);
 551
 552         if (op == CEPH_MDS_OP_GETFILELOCK)
 553                 return parse_reply_info_filelock(p, end, info, features);
 554         else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
 555                 return parse_reply_info_readdir(p, end, info, features);
 556         else if (op == CEPH_MDS_OP_CREATE)
 557                 return parse_reply_info_create(p, end, info, features, s);
 558         else
 559                 return -EIO;
 560 }
 561
 562 /*
 563  * parse entire mds reply
 564  */
 565 static int parse_reply_info(struct ceph_mds_session *s, struct ceph_msg *msg,
 566                             struct ceph_mds_reply_info_parsed *info,
 567                             u64 features)
 568 {
 569         void *p, *end;
 570         u32 len;
 571         int err;
 572
 573         info->head = msg->front.iov_base;
 574         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 575         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 576
 577         /* trace */
 578         ceph_decode_32_safe(&p, end, len, bad);
 579         if (len > 0) {
 580                 ceph_decode_need(&p, end, len, bad);
 581                 err = parse_reply_info_trace(&p, p+len, info, features);
 582                 if (err < 0)
 583                         goto out_bad;
 584         }
 585
 586         /* extra */
 587         ceph_decode_32_safe(&p, end, len, bad);
 588         if (len > 0) {
 589                 ceph_decode_need(&p, end, len, bad);
 590                 err = parse_reply_info_extra(&p, p+len, info, features, s);
 591                 if (err < 0)
 592                         goto out_bad;
 593         }
 594
 595         /* snap blob */
 596         ceph_decode_32_safe(&p, end, len, bad);
 597         info->snapblob_len = len;
 598         info->snapblob = p;
 599         p += len;
 600
 601         if (p != end)
 602                 goto bad;
 603         return 0;
 604
 605 bad:
 606         err = -EIO;
 607 out_bad:
 608         pr_err("mds parse_reply err %d\n", err);
 609         return err;
 610 }
 611
 612 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
 613 {
 614         if (!info->dir_entries)
 615                 return;
 616         free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
 617 }
 618
 619
 620 /*
 621  * sessions
 622  */
 623 const char *ceph_session_state_name(int s)
 624 {
 625         switch (s) {
 626         case CEPH_MDS_SESSION_NEW: return "new";
 627         case CEPH_MDS_SESSION_OPENING: return "opening";
 628         case CEPH_MDS_SESSION_OPEN: return "open";
 629         case CEPH_MDS_SESSION_HUNG: return "hung";
 630         case CEPH_MDS_SESSION_CLOSING: return "closing";
 631         case CEPH_MDS_SESSION_CLOSED: return "closed";
 632         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
 633         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
 634         case CEPH_MDS_SESSION_REJECTED: return "rejected";
 635         default: return "???";
 636         }
 637 }
 638
 639 struct ceph_mds_session *ceph_get_mds_session(struct ceph_mds_session *s)
 640 {
 641         if (refcount_inc_not_zero(&s->s_ref)) {
 642                 dout("mdsc get_session %p %d -> %d\n", s,
 643                      refcount_read(&s->s_ref)-1, refcount_read(&s->s_ref));
 644                 return s;
 645         } else {
 646                 dout("mdsc get_session %p 0 -- FAIL\n", s);
 647                 return NULL;
 648         }
 649 }
 650
 651 void ceph_put_mds_session(struct ceph_mds_session *s)
 652 {
 653         dout("mdsc put_session %p %d -> %d\n", s,
 654              refcount_read(&s->s_ref), refcount_read(&s->s_ref)-1);
 655         if (refcount_dec_and_test(&s->s_ref)) {
 656                 if (s->s_auth.authorizer)
 657                         ceph_auth_destroy_authorizer(s->s_auth.authorizer);
 658                 WARN_ON(mutex_is_locked(&s->s_mutex));
 659                 xa_destroy(&s->s_delegated_inos);
 660                 kfree(s);
 661         }
 662 }
 663
 664 /*
 665  * called under mdsc->mutex
 666  */
 667 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
 668                                                    int mds)
 669 {
 670         if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
 671                 return NULL;
 672         return ceph_get_mds_session(mdsc->sessions[mds]);
 673 }
 674
 675 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
 676 {
 677         if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
 678                 return false;
 679         else
 680                 return true;
 681 }
 682
 683 static int __verify_registered_session(struct ceph_mds_client *mdsc,
 684                                        struct ceph_mds_session *s)
 685 {
 686         if (s->s_mds >= mdsc->max_sessions ||
 687             mdsc->sessions[s->s_mds] != s)
 688                 return -ENOENT;
 689         return 0;
 690 }
 691
 692 /*
 693  * create+register a new session for given mds.
 694  * called under mdsc->mutex.
 695  */
 696 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 697                                                  int mds)
 698 {
 699         struct ceph_mds_session *s;
 700
 701         if (mds >= mdsc->mdsmap->possible_max_rank)
 702                 return ERR_PTR(-EINVAL);
 703
 704         s = kzalloc(sizeof(*s), GFP_NOFS);
 705         if (!s)
 706                 return ERR_PTR(-ENOMEM);
 707
 708         if (mds >= mdsc->max_sessions) {
 709                 int newmax = 1 << get_count_order(mds + 1);
 710                 struct ceph_mds_session **sa;
 711
 712                 dout("%s: realloc to %d\n", __func__, newmax);
 713                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 714                 if (!sa)
 715                         goto fail_realloc;
 716                 if (mdsc->sessions) {
 717                         memcpy(sa, mdsc->sessions,
 718                                mdsc->max_sessions * sizeof(void *));
 719                         kfree(mdsc->sessions);
 720                 }
 721                 mdsc->sessions = sa;
 722                 mdsc->max_sessions = newmax;
 723         }
 724
 725         dout("%s: mds%d\n", __func__, mds);
 726         s->s_mdsc = mdsc;
 727         s->s_mds = mds;
 728         s->s_state = CEPH_MDS_SESSION_NEW;
 729         s->s_ttl = 0;
 730         s->s_seq = 0;
 731         mutex_init(&s->s_mutex);
 732
 733         ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
 734
 735         spin_lock_init(&s->s_gen_ttl_lock);
 736         s->s_cap_gen = 1;
 737         s->s_cap_ttl = jiffies - 1;
 738
 739         spin_lock_init(&s->s_cap_lock);
 740         s->s_renew_requested = 0;
 741         s->s_renew_seq = 0;
 742         INIT_LIST_HEAD(&s->s_caps);
 743         s->s_nr_caps = 0;
 744         refcount_set(&s->s_ref, 1);
 745         INIT_LIST_HEAD(&s->s_waiting);
 746         INIT_LIST_HEAD(&s->s_unsafe);
 747         xa_init(&s->s_delegated_inos);
 748         s->s_num_cap_releases = 0;
 749         s->s_cap_reconnect = 0;
 750         s->s_cap_iterator = NULL;
 751         INIT_LIST_HEAD(&s->s_cap_releases);
 752         INIT_WORK(&s->s_cap_release_work, ceph_cap_release_work);
 753
 754         INIT_LIST_HEAD(&s->s_cap_dirty);
 755         INIT_LIST_HEAD(&s->s_cap_flushing);
 756
 757         mdsc->sessions[mds] = s;
 758         atomic_inc(&mdsc->num_sessions);
 759         refcount_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
 760
 761         ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
 762                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
 763
 764         return s;
 765
 766 fail_realloc:
 767         kfree(s);
 768         return ERR_PTR(-ENOMEM);
 769 }
 770
 771 /*
 772  * called under mdsc->mutex
 773  */
 774 static void __unregister_session(struct ceph_mds_client *mdsc,
 775                                struct ceph_mds_session *s)
 776 {
 777         dout("__unregister_session mds%d %p\n", s->s_mds, s);
 778         BUG_ON(mdsc->sessions[s->s_mds] != s);
 779         mdsc->sessions[s->s_mds] = NULL;
 780         ceph_con_close(&s->s_con);
 781         ceph_put_mds_session(s);
 782         atomic_dec(&mdsc->num_sessions);
 783 }
 784
 785 /*
 786  * drop session refs in request.
 787  *
 788  * should be last request ref, or hold mdsc->mutex
 789  */
 790 static void put_request_session(struct ceph_mds_request *req)
 791 {
 792         if (req->r_session) {
 793                 ceph_put_mds_session(req->r_session);
 794                 req->r_session = NULL;
 795         }
 796 }
 797
 798 void ceph_mdsc_release_request(struct kref *kref)
 799 {
 800         struct ceph_mds_request *req = container_of(kref,
 801                                                     struct ceph_mds_request,
 802                                                     r_kref);
 803         ceph_mdsc_release_dir_caps_no_check(req);
 804         destroy_reply_info(&req->r_reply_info);
 805         if (req->r_request)
 806                 ceph_msg_put(req->r_request);
 807         if (req->r_reply)
 808                 ceph_msg_put(req->r_reply);
 809         if (req->r_inode) {
 810                 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
 811                 /* avoid calling iput_final() in mds dispatch threads */
 812                 ceph_async_iput(req->r_inode);
 813         }
 814         if (req->r_parent) {
 815                 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
 816                 ceph_async_iput(req->r_parent);
 817         }
 818         ceph_async_iput(req->r_target_inode);
 819         if (req->r_dentry)
 820                 dput(req->r_dentry);
 821         if (req->r_old_dentry)
 822                 dput(req->r_old_dentry);
 823         if (req->r_old_dentry_dir) {
 824                 /*
 825                  * track (and drop pins for) r_old_dentry_dir
 826                  * separately, since r_old_dentry's d_parent may have
 827                  * changed between the dir mutex being dropped and
 828                  * this request being freed.
 829                  */
 830                 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
 831                                   CEPH_CAP_PIN);
 832                 ceph_async_iput(req->r_old_dentry_dir);
 833         }
 834         kfree(req->r_path1);
 835         kfree(req->r_path2);
 836         if (req->r_pagelist)
 837                 ceph_pagelist_release(req->r_pagelist);
 838         put_request_session(req);
 839         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
 840         WARN_ON_ONCE(!list_empty(&req->r_wait));
 841         kmem_cache_free(ceph_mds_request_cachep, req);
 842 }
 843
 844 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
 845
 846 /*
 847  * lookup session, bump ref if found.
 848  *
 849  * called under mdsc->mutex.
 850  */
 851 static struct ceph_mds_request *
 852 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
 853 {
 854         struct ceph_mds_request *req;
 855
 856         req = lookup_request(&mdsc->request_tree, tid);
 857         if (req)
 858                 ceph_mdsc_get_request(req);
 859
 860         return req;
 861 }
 862
 863 /*
 864  * Register an in-flight request, and assign a tid.  Link to directory
 865  * are modifying (if any).
 866  *
 867  * Called under mdsc->mutex.
 868  */
 869 static void __register_request(struct ceph_mds_client *mdsc,
 870                                struct ceph_mds_request *req,
 871                                struct inode *dir)
 872 {
 873         int ret = 0;
 874
 875         req->r_tid = ++mdsc->last_tid;
 876         if (req->r_num_caps) {
 877                 ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
 878                                         req->r_num_caps);
 879                 if (ret < 0) {
 880                         pr_err("__register_request %p "
 881                                "failed to reserve caps: %d\n", req, ret);
 882                         /* set req->r_err to fail early from __do_request */
 883                         req->r_err = ret;
 884                         return;
 885                 }
 886         }
 887         dout("__register_request %p tid %lld\n", req, req->r_tid);
 888         ceph_mdsc_get_request(req);
 889         insert_request(&mdsc->request_tree, req);
 890
 891         req->r_uid = current_fsuid();
 892         req->r_gid = current_fsgid();
 893
 894         if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
 895                 mdsc->oldest_tid = req->r_tid;
 896
 897         if (dir) {
 898                 struct ceph_inode_info *ci = ceph_inode(dir);
 899
 900                 ihold(dir);
 901                 req->r_unsafe_dir = dir;
 902                 spin_lock(&ci->i_unsafe_lock);
 903                 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
 904                 spin_unlock(&ci->i_unsafe_lock);
 905         }
 906 }
 907
 908 static void __unregister_request(struct ceph_mds_client *mdsc,
 909                                  struct ceph_mds_request *req)
 910 {
 911         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 912
 913         /* Never leave an unregistered request on an unsafe list! */
 914         list_del_init(&req->r_unsafe_item);
 915
 916         if (req->r_tid == mdsc->oldest_tid) {
 917                 struct rb_node *p = rb_next(&req->r_node);
 918                 mdsc->oldest_tid = 0;
 919                 while (p) {
 920                         struct ceph_mds_request *next_req =
 921                                 rb_entry(p, struct ceph_mds_request, r_node);
 922                         if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
 923                                 mdsc->oldest_tid = next_req->r_tid;
 924                                 break;
 925                         }
 926                         p = rb_next(p);
 927                 }
 928         }
 929
 930         erase_request(&mdsc->request_tree, req);
 931
 932         if (req->r_unsafe_dir) {
 933                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 934                 spin_lock(&ci->i_unsafe_lock);
 935                 list_del_init(&req->r_unsafe_dir_item);
 936                 spin_unlock(&ci->i_unsafe_lock);
 937         }
 938         if (req->r_target_inode &&
 939             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
 940                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
 941                 spin_lock(&ci->i_unsafe_lock);
 942                 list_del_init(&req->r_unsafe_target_item);
 943                 spin_unlock(&ci->i_unsafe_lock);
 944         }
 945
 946         if (req->r_unsafe_dir) {
 947                 /* avoid calling iput_final() in mds dispatch threads */
 948                 ceph_async_iput(req->r_unsafe_dir);
 949                 req->r_unsafe_dir = NULL;
 950         }
 951
 952         complete_all(&req->r_safe_completion);
 953
 954         ceph_mdsc_put_request(req);
 955 }
 956
 957 /*
 958  * Walk back up the dentry tree until we hit a dentry representing a
 959  * non-snapshot inode. We do this using the rcu_read_lock (which must be held
 960  * when calling this) to ensure that the objects won't disappear while we're
 961  * working with them. Once we hit a candidate dentry, we attempt to take a
 962  * reference to it, and return that as the result.
 963  */
 964 static struct inode *get_nonsnap_parent(struct dentry *dentry)
 965 {
 966         struct inode *inode = NULL;
 967
 968         while (dentry && !IS_ROOT(dentry)) {
 969                 inode = d_inode_rcu(dentry);
 970                 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
 971                         break;
 972                 dentry = dentry->d_parent;
 973         }
 974         if (inode)
 975                 inode = igrab(inode);
 976         return inode;
 977 }
 978
 979 /*
 980  * Choose mds to send request to next.  If there is a hint set in the
 981  * request (e.g., due to a prior forward hint from the mds), use that.
 982  * Otherwise, consult frag tree and/or caps to identify the
 983  * appropriate mds.  If all else fails, choose randomly.
 984  *
 985  * Called under mdsc->mutex.
 986  */
 987 static int __choose_mds(struct ceph_mds_client *mdsc,
 988                         struct ceph_mds_request *req,
 989                         bool *random)
 990 {
 991         struct inode *inode;
 992         struct ceph_inode_info *ci;
 993         struct ceph_cap *cap;
 994         int mode = req->r_direct_mode;
 995         int mds = -1;
 996         u32 hash = req->r_direct_hash;
 997         bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
 998
 999         if (random)
1000                 *random = false;
1001
1002         /*
1003          * is there a specific mds we should try?  ignore hint if we have
1004          * no session and the mds is not up (active or recovering).
1005          */
1006         if (req->r_resend_mds >= 0 &&
1007             (__have_session(mdsc, req->r_resend_mds) ||
1008              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
1009                 dout("%s using resend_mds mds%d\n", __func__,
1010                      req->r_resend_mds);
1011                 return req->r_resend_mds;
1012         }
1013
1014         if (mode == USE_RANDOM_MDS)
1015                 goto random;
1016
1017         inode = NULL;
1018         if (req->r_inode) {
1019                 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
1020                         inode = req->r_inode;
1021                         ihold(inode);
1022                 } else {
1023                         /* req->r_dentry is non-null for LSSNAP request */
1024                         rcu_read_lock();
1025                         inode = get_nonsnap_parent(req->r_dentry);
1026                         rcu_read_unlock();
1027                         dout("%s using snapdir's parent %p\n", __func__, inode);
1028                 }
1029         } else if (req->r_dentry) {
1030                 /* ignore race with rename; old or new d_parent is okay */
1031                 struct dentry *parent;
1032                 struct inode *dir;
1033
1034                 rcu_read_lock();
1035                 parent = READ_ONCE(req->r_dentry->d_parent);
1036                 dir = req->r_parent ? : d_inode_rcu(parent);
1037
1038                 if (!dir || dir->i_sb != mdsc->fsc->sb) {
1039                         /*  not this fs or parent went negative */
1040                         inode = d_inode(req->r_dentry);
1041                         if (inode)
1042                                 ihold(inode);
1043                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
1044                         /* direct snapped/virtual snapdir requests
1045                          * based on parent dir inode */
1046                         inode = get_nonsnap_parent(parent);
1047                         dout("%s using nonsnap parent %p\n", __func__, inode);
1048                 } else {
1049                         /* dentry target */
1050                         inode = d_inode(req->r_dentry);
1051                         if (!inode || mode == USE_AUTH_MDS) {
1052                                 /* dir + name */
1053                                 inode = igrab(dir);
1054                                 hash = ceph_dentry_hash(dir, req->r_dentry);
1055                                 is_hash = true;
1056                         } else {
1057                                 ihold(inode);
1058                         }
1059                 }
1060                 rcu_read_unlock();
1061         }
1062
1063         dout("%s %p is_hash=%d (0x%x) mode %d\n", __func__, inode, (int)is_hash,
1064              hash, mode);
1065         if (!inode)
1066                 goto random;
1067         ci = ceph_inode(inode);
1068
1069         if (is_hash && S_ISDIR(inode->i_mode)) {
1070                 struct ceph_inode_frag frag;
1071                 int found;
1072
1073                 ceph_choose_frag(ci, hash, &frag, &found);
1074                 if (found) {
1075                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
1076                                 u8 r;
1077
1078                                 /* choose a random replica */
1079                                 get_random_bytes(&r, 1);
1080                                 r %= frag.ndist;
1081                                 mds = frag.dist[r];
1082                                 dout("%s %p %llx.%llx frag %u mds%d (%d/%d)\n",
1083                                      __func__, inode, ceph_vinop(inode),
1084                                      frag.frag, mds, (int)r, frag.ndist);
1085                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1086                                     CEPH_MDS_STATE_ACTIVE &&
1087                                     !ceph_mdsmap_is_laggy(mdsc->mdsmap, mds))
1088                                         goto out;
1089                         }
1090
1091                         /* since this file/dir wasn't known to be
1092                          * replicated, then we want to look for the
1093                          * authoritative mds. */
1094                         if (frag.mds >= 0) {
1095                                 /* choose auth mds */
1096                                 mds = frag.mds;
1097                                 dout("%s %p %llx.%llx frag %u mds%d (auth)\n",
1098                                      __func__, inode, ceph_vinop(inode),
1099                                      frag.frag, mds);
1100                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1101                                     CEPH_MDS_STATE_ACTIVE) {
1102                                         if (!ceph_mdsmap_is_laggy(mdsc->mdsmap,
1103                                                                   mds))
1104                                                 goto out;
1105                                 }
1106                         }
1107                         mode = USE_AUTH_MDS;
1108                 }
1109         }
1110
1111         spin_lock(&ci->i_ceph_lock);
1112         cap = NULL;
1113         if (mode == USE_AUTH_MDS)
1114                 cap = ci->i_auth_cap;
1115         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
1116                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
1117         if (!cap) {
1118                 spin_unlock(&ci->i_ceph_lock);
1119                 ceph_async_iput(inode);
1120                 goto random;
1121         }
1122         mds = cap->session->s_mds;
1123         dout("%s %p %llx.%llx mds%d (%scap %p)\n", __func__,
1124              inode, ceph_vinop(inode), mds,
1125              cap == ci->i_auth_cap ? "auth " : "", cap);
1126         spin_unlock(&ci->i_ceph_lock);
1127 out:
1128         /* avoid calling iput_final() while holding mdsc->mutex or
1129          * in mds dispatch threads */
1130         ceph_async_iput(inode);
1131         return mds;
1132
1133 random:
1134         if (random)
1135                 *random = true;
1136
1137         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
1138         dout("%s chose random mds%d\n", __func__, mds);
1139         return mds;
1140 }
1141
1142
1143 /*
1144  * session messages
1145  */
1146 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
1147 {
1148         struct ceph_msg *msg;
1149         struct ceph_mds_session_head *h;
1150
1151         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
1152                            false);
1153         if (!msg) {
1154                 pr_err("create_session_msg ENOMEM creating msg\n");
1155                 return NULL;
1156         }
1157         h = msg->front.iov_base;
1158         h->op = cpu_to_le32(op);
1159         h->seq = cpu_to_le64(seq);
1160
1161         return msg;
1162 }
1163
1164 static const unsigned char feature_bits[] = CEPHFS_FEATURES_CLIENT_SUPPORTED;
1165 #define FEATURE_BYTES(c) (DIV_ROUND_UP((size_t)feature_bits[c - 1] + 1, 64) * 8)
1166 static int encode_supported_features(void **p, void *end)
1167 {
1168         static const size_t count = ARRAY_SIZE(feature_bits);
1169
1170         if (count > 0) {
1171                 size_t i;
1172                 size_t size = FEATURE_BYTES(count);
1173
1174                 if (WARN_ON_ONCE(*p + 4 + size > end))
1175                         return -ERANGE;
1176
1177                 ceph_encode_32(p, size);
1178                 memset(*p, 0, size);
1179                 for (i = 0; i < count; i++)
1180                         ((unsigned char*)(*p))[i / 8] |= BIT(feature_bits[i] % 8);
1181                 *p += size;
1182         } else {
1183                 if (WARN_ON_ONCE(*p + 4 > end))
1184                         return -ERANGE;
1185
1186                 ceph_encode_32(p, 0);
1187         }
1188
1189         return 0;
1190 }
1191
1192 static const unsigned char metric_bits[] = CEPHFS_METRIC_SPEC_CLIENT_SUPPORTED;
1193 #define METRIC_BYTES(cnt) (DIV_ROUND_UP((size_t)metric_bits[cnt - 1] + 1, 64) * 8)
1194 static int encode_metric_spec(void **p, void *end)
1195 {
1196         static const size_t count = ARRAY_SIZE(metric_bits);
1197
1198         /* header */
1199         if (WARN_ON_ONCE(*p + 2 > end))
1200                 return -ERANGE;
1201
1202         ceph_encode_8(p, 1); /* version */
1203         ceph_encode_8(p, 1); /* compat */
1204
1205         if (count > 0) {
1206                 size_t i;
1207                 size_t size = METRIC_BYTES(count);
1208
1209                 if (WARN_ON_ONCE(*p + 4 + 4 + size > end))
1210                         return -ERANGE;
1211
1212                 /* metric spec info length */
1213                 ceph_encode_32(p, 4 + size);
1214
1215                 /* metric spec */
1216                 ceph_encode_32(p, size);
1217                 memset(*p, 0, size);
1218                 for (i = 0; i < count; i++)
1219                         ((unsigned char *)(*p))[i / 8] |= BIT(metric_bits[i] % 8);
1220                 *p += size;
1221         } else {
1222                 if (WARN_ON_ONCE(*p + 4 + 4 > end))
1223                         return -ERANGE;
1224
1225                 /* metric spec info length */
1226                 ceph_encode_32(p, 4);
1227                 /* metric spec */
1228                 ceph_encode_32(p, 0);
1229         }
1230
1231         return 0;
1232 }
1233
1234 /*
1235  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
1236  * to include additional client metadata fields.
1237  */
1238 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
1239 {
1240         struct ceph_msg *msg;
1241         struct ceph_mds_session_head *h;
1242         int i;
1243         int extra_bytes = 0;
1244         int metadata_key_count = 0;
1245         struct ceph_options *opt = mdsc->fsc->client->options;
1246         struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
1247         size_t size, count;
1248         void *p, *end;
1249         int ret;
1250
1251         const char* metadata[][2] = {
1252                 {"hostname", mdsc->nodename},
1253                 {"kernel_version", init_utsname()->release},
1254                 {"entity_id", opt->name ? : ""},
1255                 {"root", fsopt->server_path ? : "/"},
1256                 {NULL, NULL}
1257         };
1258
1259         /* Calculate serialized length of metadata */
1260         extra_bytes = 4;  /* map length */
1261         for (i = 0; metadata[i][0]; ++i) {
1262                 extra_bytes += 8 + strlen(metadata[i][0]) +
1263                         strlen(metadata[i][1]);
1264                 metadata_key_count++;
1265         }
1266
1267         /* supported feature */
1268         size = 0;
1269         count = ARRAY_SIZE(feature_bits);
1270         if (count > 0)
1271                 size = FEATURE_BYTES(count);
1272         extra_bytes += 4 + size;
1273
1274         /* metric spec */
1275         size = 0;
1276         count = ARRAY_SIZE(metric_bits);
1277         if (count > 0)
1278                 size = METRIC_BYTES(count);
1279         extra_bytes += 2 + 4 + 4 + size;
1280
1281         /* Allocate the message */
1282         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + extra_bytes,
1283                            GFP_NOFS, false);
1284         if (!msg) {
1285                 pr_err("create_session_msg ENOMEM creating msg\n");
1286                 return ERR_PTR(-ENOMEM);
1287         }
1288         p = msg->front.iov_base;
1289         end = p + msg->front.iov_len;
1290
1291         h = p;
1292         h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
1293         h->seq = cpu_to_le64(seq);
1294
1295         /*
1296          * Serialize client metadata into waiting buffer space, using
1297          * the format that userspace expects for map<string, string>
1298          *
1299          * ClientSession messages with metadata are v4
1300          */
1301         msg->hdr.version = cpu_to_le16(4);
1302         msg->hdr.compat_version = cpu_to_le16(1);
1303
1304         /* The write pointer, following the session_head structure */
1305         p += sizeof(*h);
1306
1307         /* Number of entries in the map */
1308         ceph_encode_32(&p, metadata_key_count);
1309
1310         /* Two length-prefixed strings for each entry in the map */
1311         for (i = 0; metadata[i][0]; ++i) {
1312                 size_t const key_len = strlen(metadata[i][0]);
1313                 size_t const val_len = strlen(metadata[i][1]);
1314
1315                 ceph_encode_32(&p, key_len);
1316                 memcpy(p, metadata[i][0], key_len);
1317                 p += key_len;
1318                 ceph_encode_32(&p, val_len);
1319                 memcpy(p, metadata[i][1], val_len);
1320                 p += val_len;
1321         }
1322
1323         ret = encode_supported_features(&p, end);
1324         if (ret) {
1325                 pr_err("encode_supported_features failed!\n");
1326                 ceph_msg_put(msg);
1327                 return ERR_PTR(ret);
1328         }
1329
1330         ret = encode_metric_spec(&p, end);
1331         if (ret) {
1332                 pr_err("encode_metric_spec failed!\n");
1333                 ceph_msg_put(msg);
1334                 return ERR_PTR(ret);
1335         }
1336
1337         msg->front.iov_len = p - msg->front.iov_base;
1338         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1339
1340         return msg;
1341 }
1342
1343 /*
1344  * send session open request.
1345  *
1346  * called under mdsc->mutex
1347  */
1348 static int __open_session(struct ceph_mds_client *mdsc,
1349                           struct ceph_mds_session *session)
1350 {
1351         struct ceph_msg *msg;
1352         int mstate;
1353         int mds = session->s_mds;
1354
1355         /* wait for mds to go active? */
1356         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
1357         dout("open_session to mds%d (%s)\n", mds,
1358              ceph_mds_state_name(mstate));
1359         session->s_state = CEPH_MDS_SESSION_OPENING;
1360         session->s_renew_requested = jiffies;
1361
1362         /* send connect message */
1363         msg = create_session_open_msg(mdsc, session->s_seq);
1364         if (IS_ERR(msg))
1365                 return PTR_ERR(msg);
1366         ceph_con_send(&session->s_con, msg);
1367         return 0;
1368 }
1369
1370 /*
1371  * open sessions for any export targets for the given mds
1372  *
1373  * called under mdsc->mutex
1374  */
1375 static struct ceph_mds_session *
1376 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
1377 {
1378         struct ceph_mds_session *session;
1379         int ret;
1380
1381         session = __ceph_lookup_mds_session(mdsc, target);
1382         if (!session) {
1383                 session = register_session(mdsc, target);
1384                 if (IS_ERR(session))
1385                         return session;
1386         }
1387         if (session->s_state == CEPH_MDS_SESSION_NEW ||
1388             session->s_state == CEPH_MDS_SESSION_CLOSING) {
1389                 ret = __open_session(mdsc, session);
1390                 if (ret)
1391                         return ERR_PTR(ret);
1392         }
1393
1394         return session;
1395 }
1396
1397 struct ceph_mds_session *
1398 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1399 {
1400         struct ceph_mds_session *session;
1401
1402         dout("open_export_target_session to mds%d\n", target);
1403
1404         mutex_lock(&mdsc->mutex);
1405         session = __open_export_target_session(mdsc, target);
1406         mutex_unlock(&mdsc->mutex);
1407
1408         return session;
1409 }
1410
1411 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1412                                           struct ceph_mds_session *session)
1413 {
1414         struct ceph_mds_info *mi;
1415         struct ceph_mds_session *ts;
1416         int i, mds = session->s_mds;
1417
1418         if (mds >= mdsc->mdsmap->possible_max_rank)
1419                 return;
1420
1421         mi = &mdsc->mdsmap->m_info[mds];
1422         dout("open_export_target_sessions for mds%d (%d targets)\n",
1423              session->s_mds, mi->num_export_targets);
1424
1425         for (i = 0; i < mi->num_export_targets; i++) {
1426                 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1427                 if (!IS_ERR(ts))
1428                         ceph_put_mds_session(ts);
1429         }
1430 }
1431
1432 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1433                                            struct ceph_mds_session *session)
1434 {
1435         mutex_lock(&mdsc->mutex);
1436         __open_export_target_sessions(mdsc, session);
1437         mutex_unlock(&mdsc->mutex);
1438 }
1439
1440 /*
1441  * session caps
1442  */
1443
1444 static void detach_cap_releases(struct ceph_mds_session *session,
1445                                 struct list_head *target)
1446 {
1447         lockdep_assert_held(&session->s_cap_lock);
1448
1449         list_splice_init(&session->s_cap_releases, target);
1450         session->s_num_cap_releases = 0;
1451         dout("dispose_cap_releases mds%d\n", session->s_mds);
1452 }
1453
1454 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1455                                  struct list_head *dispose)
1456 {
1457         while (!list_empty(dispose)) {
1458                 struct ceph_cap *cap;
1459                 /* zero out the in-progress message */
1460                 cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1461                 list_del(&cap->session_caps);
1462                 ceph_put_cap(mdsc, cap);
1463         }
1464 }
1465
1466 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1467                                      struct ceph_mds_session *session)
1468 {
1469         struct ceph_mds_request *req;
1470         struct rb_node *p;
1471         struct ceph_inode_info *ci;
1472
1473         dout("cleanup_session_requests mds%d\n", session->s_mds);
1474         mutex_lock(&mdsc->mutex);
1475         while (!list_empty(&session->s_unsafe)) {
1476                 req = list_first_entry(&session->s_unsafe,
1477                                        struct ceph_mds_request, r_unsafe_item);
1478                 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1479                                     req->r_tid);
1480                 if (req->r_target_inode) {
1481                         /* dropping unsafe change of inode's attributes */
1482                         ci = ceph_inode(req->r_target_inode);
1483                         errseq_set(&ci->i_meta_err, -EIO);
1484                 }
1485                 if (req->r_unsafe_dir) {
1486                         /* dropping unsafe directory operation */
1487                         ci = ceph_inode(req->r_unsafe_dir);
1488                         errseq_set(&ci->i_meta_err, -EIO);
1489                 }
1490                 __unregister_request(mdsc, req);
1491         }
1492         /* zero r_attempts, so kick_requests() will re-send requests */
1493         p = rb_first(&mdsc->request_tree);
1494         while (p) {
1495                 req = rb_entry(p, struct ceph_mds_request, r_node);
1496                 p = rb_next(p);
1497                 if (req->r_session &&
1498                     req->r_session->s_mds == session->s_mds)
1499                         req->r_attempts = 0;
1500         }
1501         mutex_unlock(&mdsc->mutex);
1502 }
1503
1504 /*
1505  * Helper to safely iterate over all caps associated with a session, with
1506  * special care taken to handle a racing __ceph_remove_cap().
1507  *
1508  * Caller must hold session s_mutex.
1509  */
1510 int ceph_iterate_session_caps(struct ceph_mds_session *session,
1511                               int (*cb)(struct inode *, struct ceph_cap *,
1512                                         void *), void *arg)
1513 {
1514         struct list_head *p;
1515         struct ceph_cap *cap;
1516         struct inode *inode, *last_inode = NULL;
1517         struct ceph_cap *old_cap = NULL;
1518         int ret;
1519
1520         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1521         spin_lock(&session->s_cap_lock);
1522         p = session->s_caps.next;
1523         while (p != &session->s_caps) {
1524                 cap = list_entry(p, struct ceph_cap, session_caps);
1525                 inode = igrab(&cap->ci->vfs_inode);
1526                 if (!inode) {
1527                         p = p->next;
1528                         continue;
1529                 }
1530                 session->s_cap_iterator = cap;
1531                 spin_unlock(&session->s_cap_lock);
1532
1533                 if (last_inode) {
1534                         /* avoid calling iput_final() while holding
1535                          * s_mutex or in mds dispatch threads */
1536                         ceph_async_iput(last_inode);
1537                         last_inode = NULL;
1538                 }
1539                 if (old_cap) {
1540                         ceph_put_cap(session->s_mdsc, old_cap);
1541                         old_cap = NULL;
1542                 }
1543
1544                 ret = cb(inode, cap, arg);
1545                 last_inode = inode;
1546
1547                 spin_lock(&session->s_cap_lock);
1548                 p = p->next;
1549                 if (!cap->ci) {
1550                         dout("iterate_session_caps  finishing cap %p removal\n",
1551                              cap);
1552                         BUG_ON(cap->session != session);
1553                         cap->session = NULL;
1554                         list_del_init(&cap->session_caps);
1555                         session->s_nr_caps--;
1556                         atomic64_dec(&session->s_mdsc->metric.total_caps);
1557                         if (cap->queue_release)
1558                                 __ceph_queue_cap_release(session, cap);
1559                         else
1560                                 old_cap = cap;  /* put_cap it w/o locks held */
1561                 }
1562                 if (ret < 0)
1563                         goto out;
1564         }
1565         ret = 0;
1566 out:
1567         session->s_cap_iterator = NULL;
1568         spin_unlock(&session->s_cap_lock);
1569
1570         ceph_async_iput(last_inode);
1571         if (old_cap)
1572                 ceph_put_cap(session->s_mdsc, old_cap);
1573
1574         return ret;
1575 }
1576
1577 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1578                                   void *arg)
1579 {
1580         struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1581         struct ceph_inode_info *ci = ceph_inode(inode);
1582         LIST_HEAD(to_remove);
1583         bool dirty_dropped = false;
1584         bool invalidate = false;
1585
1586         dout("removing cap %p, ci is %p, inode is %p\n",
1587              cap, ci, &ci->vfs_inode);
1588         spin_lock(&ci->i_ceph_lock);
1589         __ceph_remove_cap(cap, false);
1590         if (!ci->i_auth_cap) {
1591                 struct ceph_cap_flush *cf;
1592                 struct ceph_mds_client *mdsc = fsc->mdsc;
1593
1594                 if (READ_ONCE(fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN) {
1595                         if (inode->i_data.nrpages > 0)
1596                                 invalidate = true;
1597                         if (ci->i_wrbuffer_ref > 0)
1598                                 mapping_set_error(&inode->i_data, -EIO);
1599                 }
1600
1601                 while (!list_empty(&ci->i_cap_flush_list)) {
1602                         cf = list_first_entry(&ci->i_cap_flush_list,
1603                                               struct ceph_cap_flush, i_list);
1604                         list_move(&cf->i_list, &to_remove);
1605                 }
1606
1607                 spin_lock(&mdsc->cap_dirty_lock);
1608
1609                 list_for_each_entry(cf, &to_remove, i_list)
1610                         list_del(&cf->g_list);
1611
1612                 if (!list_empty(&ci->i_dirty_item)) {
1613                         pr_warn_ratelimited(
1614                                 " dropping dirty %s state for %p %lld\n",
1615                                 ceph_cap_string(ci->i_dirty_caps),
1616                                 inode, ceph_ino(inode));
1617                         ci->i_dirty_caps = 0;
1618                         list_del_init(&ci->i_dirty_item);
1619                         dirty_dropped = true;
1620                 }
1621                 if (!list_empty(&ci->i_flushing_item)) {
1622                         pr_warn_ratelimited(
1623                                 " dropping dirty+flushing %s state for %p %lld\n",
1624                                 ceph_cap_string(ci->i_flushing_caps),
1625                                 inode, ceph_ino(inode));
1626                         ci->i_flushing_caps = 0;
1627                         list_del_init(&ci->i_flushing_item);
1628                         mdsc->num_cap_flushing--;
1629                         dirty_dropped = true;
1630                 }
1631                 spin_unlock(&mdsc->cap_dirty_lock);
1632
1633                 if (dirty_dropped) {
1634                         errseq_set(&ci->i_meta_err, -EIO);
1635
1636                         if (ci->i_wrbuffer_ref_head == 0 &&
1637                             ci->i_wr_ref == 0 &&
1638                             ci->i_dirty_caps == 0 &&
1639                             ci->i_flushing_caps == 0) {
1640                                 ceph_put_snap_context(ci->i_head_snapc);
1641                                 ci->i_head_snapc = NULL;
1642                         }
1643                 }
1644
1645                 if (atomic_read(&ci->i_filelock_ref) > 0) {
1646                         /* make further file lock syscall return -EIO */
1647                         ci->i_ceph_flags |= CEPH_I_ERROR_FILELOCK;
1648                         pr_warn_ratelimited(" dropping file locks for %p %lld\n",
1649                                             inode, ceph_ino(inode));
1650                 }
1651
1652                 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1653                         list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1654                         ci->i_prealloc_cap_flush = NULL;
1655                 }
1656         }
1657         spin_unlock(&ci->i_ceph_lock);
1658         while (!list_empty(&to_remove)) {
1659                 struct ceph_cap_flush *cf;
1660                 cf = list_first_entry(&to_remove,
1661                                       struct ceph_cap_flush, i_list);
1662                 list_del(&cf->i_list);
1663                 ceph_free_cap_flush(cf);
1664         }
1665
1666         wake_up_all(&ci->i_cap_wq);
1667         if (invalidate)
1668                 ceph_queue_invalidate(inode);
1669         if (dirty_dropped)
1670                 iput(inode);
1671         return 0;
1672 }
1673
1674 /*
1675  * caller must hold session s_mutex
1676  */
1677 static void remove_session_caps(struct ceph_mds_session *session)
1678 {
1679         struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1680         struct super_block *sb = fsc->sb;
1681         LIST_HEAD(dispose);
1682
1683         dout("remove_session_caps on %p\n", session);
1684         ceph_iterate_session_caps(session, remove_session_caps_cb, fsc);
1685
1686         wake_up_all(&fsc->mdsc->cap_flushing_wq);
1687
1688         spin_lock(&session->s_cap_lock);
1689         if (session->s_nr_caps > 0) {
1690                 struct inode *inode;
1691                 struct ceph_cap *cap, *prev = NULL;
1692                 struct ceph_vino vino;
1693                 /*
1694                  * iterate_session_caps() skips inodes that are being
1695                  * deleted, we need to wait until deletions are complete.
1696                  * __wait_on_freeing_inode() is designed for the job,
1697                  * but it is not exported, so use lookup inode function
1698                  * to access it.
1699                  */
1700                 while (!list_empty(&session->s_caps)) {
1701                         cap = list_entry(session->s_caps.next,
1702                                          struct ceph_cap, session_caps);
1703                         if (cap == prev)
1704                                 break;
1705                         prev = cap;
1706                         vino = cap->ci->i_vino;
1707                         spin_unlock(&session->s_cap_lock);
1708
1709                         inode = ceph_find_inode(sb, vino);
1710                          /* avoid calling iput_final() while holding s_mutex */
1711                         ceph_async_iput(inode);
1712
1713                         spin_lock(&session->s_cap_lock);
1714                 }
1715         }
1716
1717         // drop cap expires and unlock s_cap_lock
1718         detach_cap_releases(session, &dispose);
1719
1720         BUG_ON(session->s_nr_caps > 0);
1721         BUG_ON(!list_empty(&session->s_cap_flushing));
1722         spin_unlock(&session->s_cap_lock);
1723         dispose_cap_releases(session->s_mdsc, &dispose);
1724 }
1725
1726 enum {
1727         RECONNECT,
1728         RENEWCAPS,
1729         FORCE_RO,
1730 };
1731
1732 /*
1733  * wake up any threads waiting on this session's caps.  if the cap is
1734  * old (didn't get renewed on the client reconnect), remove it now.
1735  *
1736  * caller must hold s_mutex.
1737  */
1738 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1739                               void *arg)
1740 {
1741         struct ceph_inode_info *ci = ceph_inode(inode);
1742         unsigned long ev = (unsigned long)arg;
1743
1744         if (ev == RECONNECT) {
1745                 spin_lock(&ci->i_ceph_lock);
1746                 ci->i_wanted_max_size = 0;
1747                 ci->i_requested_max_size = 0;
1748                 spin_unlock(&ci->i_ceph_lock);
1749         } else if (ev == RENEWCAPS) {
1750                 if (cap->cap_gen < cap->session->s_cap_gen) {
1751                         /* mds did not re-issue stale cap */
1752                         spin_lock(&ci->i_ceph_lock);
1753                         cap->issued = cap->implemented = CEPH_CAP_PIN;
1754                         spin_unlock(&ci->i_ceph_lock);
1755                 }
1756         } else if (ev == FORCE_RO) {
1757         }
1758         wake_up_all(&ci->i_cap_wq);
1759         return 0;
1760 }
1761
1762 static void wake_up_session_caps(struct ceph_mds_session *session, int ev)
1763 {
1764         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1765         ceph_iterate_session_caps(session, wake_up_session_cb,
1766                                   (void *)(unsigned long)ev);
1767 }
1768
1769 /*
1770  * Send periodic message to MDS renewing all currently held caps.  The
1771  * ack will reset the expiration for all caps from this session.
1772  *
1773  * caller holds s_mutex
1774  */
1775 static int send_renew_caps(struct ceph_mds_client *mdsc,
1776                            struct ceph_mds_session *session)
1777 {
1778         struct ceph_msg *msg;
1779         int state;
1780
1781         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1782             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1783                 pr_info("mds%d caps stale\n", session->s_mds);
1784         session->s_renew_requested = jiffies;
1785
1786         /* do not try to renew caps until a recovering mds has reconnected
1787          * with its clients. */
1788         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1789         if (state < CEPH_MDS_STATE_RECONNECT) {
1790                 dout("send_renew_caps ignoring mds%d (%s)\n",
1791                      session->s_mds, ceph_mds_state_name(state));
1792                 return 0;
1793         }
1794
1795         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1796                 ceph_mds_state_name(state));
1797         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1798                                  ++session->s_renew_seq);
1799         if (!msg)
1800                 return -ENOMEM;
1801         ceph_con_send(&session->s_con, msg);
1802         return 0;
1803 }
1804
1805 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1806                              struct ceph_mds_session *session, u64 seq)
1807 {
1808         struct ceph_msg *msg;
1809
1810         dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1811              session->s_mds, ceph_session_state_name(session->s_state), seq);
1812         msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1813         if (!msg)
1814                 return -ENOMEM;
1815         ceph_con_send(&session->s_con, msg);
1816         return 0;
1817 }
1818
1819
1820 /*
1821  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1822  *
1823  * Called under session->s_mutex
1824  */
1825 static void renewed_caps(struct ceph_mds_client *mdsc,
1826                          struct ceph_mds_session *session, int is_renew)
1827 {
1828         int was_stale;
1829         int wake = 0;
1830
1831         spin_lock(&session->s_cap_lock);
1832         was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1833
1834         session->s_cap_ttl = session->s_renew_requested +
1835                 mdsc->mdsmap->m_session_timeout*HZ;
1836
1837         if (was_stale) {
1838                 if (time_before(jiffies, session->s_cap_ttl)) {
1839                         pr_info("mds%d caps renewed\n", session->s_mds);
1840                         wake = 1;
1841                 } else {
1842                         pr_info("mds%d caps still stale\n", session->s_mds);
1843                 }
1844         }
1845         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1846              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1847              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1848         spin_unlock(&session->s_cap_lock);
1849
1850         if (wake)
1851                 wake_up_session_caps(session, RENEWCAPS);
1852 }
1853
1854 /*
1855  * send a session close request
1856  */
1857 static int request_close_session(struct ceph_mds_session *session)
1858 {
1859         struct ceph_msg *msg;
1860
1861         dout("request_close_session mds%d state %s seq %lld\n",
1862              session->s_mds, ceph_session_state_name(session->s_state),
1863              session->s_seq);
1864         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1865         if (!msg)
1866                 return -ENOMEM;
1867         ceph_con_send(&session->s_con, msg);
1868         return 1;
1869 }
1870
1871 /*
1872  * Called with s_mutex held.
1873  */
1874 static int __close_session(struct ceph_mds_client *mdsc,
1875                          struct ceph_mds_session *session)
1876 {
1877         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1878                 return 0;
1879         session->s_state = CEPH_MDS_SESSION_CLOSING;
1880         return request_close_session(session);
1881 }
1882
1883 static bool drop_negative_children(struct dentry *dentry)
1884 {
1885         struct dentry *child;
1886         bool all_negative = true;
1887
1888         if (!d_is_dir(dentry))
1889                 goto out;
1890
1891         spin_lock(&dentry->d_lock);
1892         list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1893                 if (d_really_is_positive(child)) {
1894                         all_negative = false;
1895                         break;
1896                 }
1897         }
1898         spin_unlock(&dentry->d_lock);
1899
1900         if (all_negative)
1901                 shrink_dcache_parent(dentry);
1902 out:
1903         return all_negative;
1904 }
1905
1906 /*
1907  * Trim old(er) caps.
1908  *
1909  * Because we can't cache an inode without one or more caps, we do
1910  * this indirectly: if a cap is unused, we prune its aliases, at which
1911  * point the inode will hopefully get dropped to.
1912  *
1913  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1914  * memory pressure from the MDS, though, so it needn't be perfect.
1915  */
1916 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1917 {
1918         int *remaining = arg;
1919         struct ceph_inode_info *ci = ceph_inode(inode);
1920         int used, wanted, oissued, mine;
1921
1922         if (*remaining <= 0)
1923                 return -1;
1924
1925         spin_lock(&ci->i_ceph_lock);
1926         mine = cap->issued | cap->implemented;
1927         used = __ceph_caps_used(ci);
1928         wanted = __ceph_caps_file_wanted(ci);
1929         oissued = __ceph_caps_issued_other(ci, cap);
1930
1931         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1932              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1933              ceph_cap_string(used), ceph_cap_string(wanted));
1934         if (cap == ci->i_auth_cap) {
1935                 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1936                     !list_empty(&ci->i_cap_snaps))
1937                         goto out;
1938                 if ((used | wanted) & CEPH_CAP_ANY_WR)
1939                         goto out;
1940                 /* Note: it's possible that i_filelock_ref becomes non-zero
1941                  * after dropping auth caps. It doesn't hurt because reply
1942                  * of lock mds request will re-add auth caps. */
1943                 if (atomic_read(&ci->i_filelock_ref) > 0)
1944                         goto out;
1945         }
1946         /* The inode has cached pages, but it's no longer used.
1947          * we can safely drop it */
1948         if (S_ISREG(inode->i_mode) &&
1949             wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1950             !(oissued & CEPH_CAP_FILE_CACHE)) {
1951           used = 0;
1952           oissued = 0;
1953         }
1954         if ((used | wanted) & ~oissued & mine)
1955                 goto out;   /* we need these caps */
1956
1957         if (oissued) {
1958                 /* we aren't the only cap.. just remove us */
1959                 __ceph_remove_cap(cap, true);
1960                 (*remaining)--;
1961         } else {
1962                 struct dentry *dentry;
1963                 /* try dropping referring dentries */
1964                 spin_unlock(&ci->i_ceph_lock);
1965                 dentry = d_find_any_alias(inode);
1966                 if (dentry && drop_negative_children(dentry)) {
1967                         int count;
1968                         dput(dentry);
1969                         d_prune_aliases(inode);
1970                         count = atomic_read(&inode->i_count);
1971                         if (count == 1)
1972                                 (*remaining)--;
1973                         dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1974                              inode, cap, count);
1975                 } else {
1976                         dput(dentry);
1977                 }
1978                 return 0;
1979         }
1980
1981 out:
1982         spin_unlock(&ci->i_ceph_lock);
1983         return 0;
1984 }
1985
1986 /*
1987  * Trim session cap count down to some max number.
1988  */
1989 int ceph_trim_caps(struct ceph_mds_client *mdsc,
1990                    struct ceph_mds_session *session,
1991                    int max_caps)
1992 {
1993         int trim_caps = session->s_nr_caps - max_caps;
1994
1995         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1996              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1997         if (trim_caps > 0) {
1998                 int remaining = trim_caps;
1999
2000                 ceph_iterate_session_caps(session, trim_caps_cb, &remaining);
2001                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
2002                      session->s_mds, session->s_nr_caps, max_caps,
2003                         trim_caps - remaining);
2004         }
2005
2006         ceph_flush_cap_releases(mdsc, session);
2007         return 0;
2008 }
2009
2010 static int check_caps_flush(struct ceph_mds_client *mdsc,
2011                             u64 want_flush_tid)
2012 {
2013         int ret = 1;
2014
2015         spin_lock(&mdsc->cap_dirty_lock);
2016         if (!list_empty(&mdsc->cap_flush_list)) {
2017                 struct ceph_cap_flush *cf =
2018                         list_first_entry(&mdsc->cap_flush_list,
2019                                          struct ceph_cap_flush, g_list);
2020                 if (cf->tid <= want_flush_tid) {
2021                         dout("check_caps_flush still flushing tid "
2022                              "%llu <= %llu\n", cf->tid, want_flush_tid);
2023                         ret = 0;
2024                 }
2025         }
2026         spin_unlock(&mdsc->cap_dirty_lock);
2027         return ret;
2028 }
2029
2030 /*
2031  * flush all dirty inode data to disk.
2032  *
2033  * returns true if we've flushed through want_flush_tid
2034  */
2035 static void wait_caps_flush(struct ceph_mds_client *mdsc,
2036                             u64 want_flush_tid)
2037 {
2038         dout("check_caps_flush want %llu\n", want_flush_tid);
2039
2040         wait_event(mdsc->cap_flushing_wq,
2041                    check_caps_flush(mdsc, want_flush_tid));
2042
2043         dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
2044 }
2045
2046 /*
2047  * called under s_mutex
2048  */
2049 static void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
2050                                    struct ceph_mds_session *session)
2051 {
2052         struct ceph_msg *msg = NULL;
2053         struct ceph_mds_cap_release *head;
2054         struct ceph_mds_cap_item *item;
2055         struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
2056         struct ceph_cap *cap;
2057         LIST_HEAD(tmp_list);
2058         int num_cap_releases;
2059         __le32  barrier, *cap_barrier;
2060
2061         down_read(&osdc->lock);
2062         barrier = cpu_to_le32(osdc->epoch_barrier);
2063         up_read(&osdc->lock);
2064
2065         spin_lock(&session->s_cap_lock);
2066 again:
2067         list_splice_init(&session->s_cap_releases, &tmp_list);
2068         num_cap_releases = session->s_num_cap_releases;
2069         session->s_num_cap_releases = 0;
2070         spin_unlock(&session->s_cap_lock);
2071
2072         while (!list_empty(&tmp_list)) {
2073                 if (!msg) {
2074                         msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
2075                                         PAGE_SIZE, GFP_NOFS, false);
2076                         if (!msg)
2077                                 goto out_err;
2078                         head = msg->front.iov_base;
2079                         head->num = cpu_to_le32(0);
2080                         msg->front.iov_len = sizeof(*head);
2081
2082                         msg->hdr.version = cpu_to_le16(2);
2083                         msg->hdr.compat_version = cpu_to_le16(1);
2084                 }
2085
2086                 cap = list_first_entry(&tmp_list, struct ceph_cap,
2087                                         session_caps);
2088                 list_del(&cap->session_caps);
2089                 num_cap_releases--;
2090
2091                 head = msg->front.iov_base;
2092                 put_unaligned_le32(get_unaligned_le32(&head->num) + 1,
2093                                    &head->num);
2094                 item = msg->front.iov_base + msg->front.iov_len;
2095                 item->ino = cpu_to_le64(cap->cap_ino);
2096                 item->cap_id = cpu_to_le64(cap->cap_id);
2097                 item->migrate_seq = cpu_to_le32(cap->mseq);
2098                 item->seq = cpu_to_le32(cap->issue_seq);
2099                 msg->front.iov_len += sizeof(*item);
2100
2101                 ceph_put_cap(mdsc, cap);
2102
2103                 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
2104                         // Append cap_barrier field
2105                         cap_barrier = msg->front.iov_base + msg->front.iov_len;
2106                         *cap_barrier = barrier;
2107                         msg->front.iov_len += sizeof(*cap_barrier);
2108
2109                         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2110                         dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
2111                         ceph_con_send(&session->s_con, msg);
2112                         msg = NULL;
2113                 }
2114         }
2115
2116         BUG_ON(num_cap_releases != 0);
2117
2118         spin_lock(&session->s_cap_lock);
2119         if (!list_empty(&session->s_cap_releases))
2120                 goto again;
2121         spin_unlock(&session->s_cap_lock);
2122
2123         if (msg) {
2124                 // Append cap_barrier field
2125                 cap_barrier = msg->front.iov_base + msg->front.iov_len;
2126                 *cap_barrier = barrier;
2127                 msg->front.iov_len += sizeof(*cap_barrier);
2128
2129                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2130                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
2131                 ceph_con_send(&session->s_con, msg);
2132         }
2133         return;
2134 out_err:
2135         pr_err("send_cap_releases mds%d, failed to allocate message\n",
2136                 session->s_mds);
2137         spin_lock(&session->s_cap_lock);
2138         list_splice(&tmp_list, &session->s_cap_releases);
2139         session->s_num_cap_releases += num_cap_releases;
2140         spin_unlock(&session->s_cap_lock);
2141 }
2142
2143 static void ceph_cap_release_work(struct work_struct *work)
2144 {
2145         struct ceph_mds_session *session =
2146                 container_of(work, struct ceph_mds_session, s_cap_release_work);
2147
2148         mutex_lock(&session->s_mutex);
2149         if (session->s_state == CEPH_MDS_SESSION_OPEN ||
2150             session->s_state == CEPH_MDS_SESSION_HUNG)
2151                 ceph_send_cap_releases(session->s_mdsc, session);
2152         mutex_unlock(&session->s_mutex);
2153         ceph_put_mds_session(session);
2154 }
2155
2156 void ceph_flush_cap_releases(struct ceph_mds_client *mdsc,
2157                              struct ceph_mds_session *session)
2158 {
2159         if (mdsc->stopping)
2160                 return;
2161
2162         ceph_get_mds_session(session);
2163         if (queue_work(mdsc->fsc->cap_wq,
2164                        &session->s_cap_release_work)) {
2165                 dout("cap release work queued\n");
2166         } else {
2167                 ceph_put_mds_session(session);
2168                 dout("failed to queue cap release work\n");
2169         }
2170 }
2171
2172 /*
2173  * caller holds session->s_cap_lock
2174  */
2175 void __ceph_queue_cap_release(struct ceph_mds_session *session,
2176                               struct ceph_cap *cap)
2177 {
2178         list_add_tail(&cap->session_caps, &session->s_cap_releases);
2179         session->s_num_cap_releases++;
2180
2181         if (!(session->s_num_cap_releases % CEPH_CAPS_PER_RELEASE))
2182                 ceph_flush_cap_releases(session->s_mdsc, session);
2183 }
2184
2185 static void ceph_cap_reclaim_work(struct work_struct *work)
2186 {
2187         struct ceph_mds_client *mdsc =
2188                 container_of(work, struct ceph_mds_client, cap_reclaim_work);
2189         int ret = ceph_trim_dentries(mdsc);
2190         if (ret == -EAGAIN)
2191                 ceph_queue_cap_reclaim_work(mdsc);
2192 }
2193
2194 void ceph_queue_cap_reclaim_work(struct ceph_mds_client *mdsc)
2195 {
2196         if (mdsc->stopping)
2197                 return;
2198
2199         if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_reclaim_work)) {
2200                 dout("caps reclaim work queued\n");
2201         } else {
2202                 dout("failed to queue caps release work\n");
2203         }
2204 }
2205
2206 void ceph_reclaim_caps_nr(struct ceph_mds_client *mdsc, int nr)
2207 {
2208         int val;
2209         if (!nr)
2210                 return;
2211         val = atomic_add_return(nr, &mdsc->cap_reclaim_pending);
2212         if ((val % CEPH_CAPS_PER_RELEASE) < nr) {
2213                 atomic_set(&mdsc->cap_reclaim_pending, 0);
2214                 ceph_queue_cap_reclaim_work(mdsc);
2215         }
2216 }
2217
2218 /*
2219  * requests
2220  */
2221
2222 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
2223                                     struct inode *dir)
2224 {
2225         struct ceph_inode_info *ci = ceph_inode(dir);
2226         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
2227         struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
2228         size_t size = sizeof(struct ceph_mds_reply_dir_entry);
2229         unsigned int num_entries;
2230         int order;
2231
2232         spin_lock(&ci->i_ceph_lock);
2233         num_entries = ci->i_files + ci->i_subdirs;
2234         spin_unlock(&ci->i_ceph_lock);
2235         num_entries = max(num_entries, 1U);
2236         num_entries = min(num_entries, opt->max_readdir);
2237
2238         order = get_order(size * num_entries);
2239         while (order >= 0) {
2240                 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
2241                                                              __GFP_NOWARN,
2242                                                              order);
2243                 if (rinfo->dir_entries)
2244                         break;
2245                 order--;
2246         }
2247         if (!rinfo->dir_entries)
2248                 return -ENOMEM;
2249
2250         num_entries = (PAGE_SIZE << order) / size;
2251         num_entries = min(num_entries, opt->max_readdir);
2252
2253         rinfo->dir_buf_size = PAGE_SIZE << order;
2254         req->r_num_caps = num_entries + 1;
2255         req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
2256         req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
2257         return 0;
2258 }
2259
2260 /*
2261  * Create an mds request.
2262  */
2263 struct ceph_mds_request *
2264 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
2265 {
2266         struct ceph_mds_request *req;
2267
2268         req = kmem_cache_zalloc(ceph_mds_request_cachep, GFP_NOFS);
2269         if (!req)
2270                 return ERR_PTR(-ENOMEM);
2271
2272         mutex_init(&req->r_fill_mutex);
2273         req->r_mdsc = mdsc;
2274         req->r_started = jiffies;
2275         req->r_start_latency = ktime_get();
2276         req->r_resend_mds = -1;
2277         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
2278         INIT_LIST_HEAD(&req->r_unsafe_target_item);
2279         req->r_fmode = -1;
2280         kref_init(&req->r_kref);
2281         RB_CLEAR_NODE(&req->r_node);
2282         INIT_LIST_HEAD(&req->r_wait);
2283         init_completion(&req->r_completion);
2284         init_completion(&req->r_safe_completion);
2285         INIT_LIST_HEAD(&req->r_unsafe_item);
2286
2287         ktime_get_coarse_real_ts64(&req->r_stamp);
2288
2289         req->r_op = op;
2290         req->r_direct_mode = mode;
2291         return req;
2292 }
2293
2294 /*
2295  * return oldest (lowest) request, tid in request tree, 0 if none.
2296  *
2297  * called under mdsc->mutex.
2298  */
2299 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
2300 {
2301         if (RB_EMPTY_ROOT(&mdsc->request_tree))
2302                 return NULL;
2303         return rb_entry(rb_first(&mdsc->request_tree),
2304                         struct ceph_mds_request, r_node);
2305 }
2306
2307 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
2308 {
2309         return mdsc->oldest_tid;
2310 }
2311
2312 /*
2313  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
2314  * on build_path_from_dentry in fs/cifs/dir.c.
2315  *
2316  * If @stop_on_nosnap, generate path relative to the first non-snapped
2317  * inode.
2318  *
2319  * Encode hidden .snap dirs as a double /, i.e.
2320  *   foo/.snap/bar -> foo//bar
2321  */
2322 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *pbase,
2323                            int stop_on_nosnap)
2324 {
2325         struct dentry *temp;
2326         char *path;
2327         int pos;
2328         unsigned seq;
2329         u64 base;
2330
2331         if (!dentry)
2332                 return ERR_PTR(-EINVAL);
2333
2334         path = __getname();
2335         if (!path)
2336                 return ERR_PTR(-ENOMEM);
2337 retry:
2338         pos = PATH_MAX - 1;
2339         path[pos] = '\0';
2340
2341         seq = read_seqbegin(&rename_lock);
2342         rcu_read_lock();
2343         temp = dentry;
2344         for (;;) {
2345                 struct inode *inode;
2346
2347                 spin_lock(&temp->d_lock);
2348                 inode = d_inode(temp);
2349                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
2350                         dout("build_path path+%d: %p SNAPDIR\n",
2351                              pos, temp);
2352                 } else if (stop_on_nosnap && inode && dentry != temp &&
2353                            ceph_snap(inode) == CEPH_NOSNAP) {
2354                         spin_unlock(&temp->d_lock);
2355                         pos++; /* get rid of any prepended '/' */
2356                         break;
2357                 } else {
2358                         pos -= temp->d_name.len;
2359                         if (pos < 0) {
2360                                 spin_unlock(&temp->d_lock);
2361                                 break;
2362                         }
2363                         memcpy(path + pos, temp->d_name.name, temp->d_name.len);
2364                 }
2365                 spin_unlock(&temp->d_lock);
2366                 temp = READ_ONCE(temp->d_parent);
2367
2368                 /* Are we at the root? */
2369                 if (IS_ROOT(temp))
2370                         break;
2371
2372                 /* Are we out of buffer? */
2373                 if (--pos < 0)
2374                         break;
2375
2376                 path[pos] = '/';
2377         }
2378         base = ceph_ino(d_inode(temp));
2379         rcu_read_unlock();
2380
2381         if (read_seqretry(&rename_lock, seq))
2382                 goto retry;
2383
2384         if (pos < 0) {
2385                 /*
2386                  * A rename didn't occur, but somehow we didn't end up where
2387                  * we thought we would. Throw a warning and try again.
2388                  */
2389                 pr_warn("build_path did not end path lookup where "
2390                         "expected, pos is %d\n", pos);
2391                 goto retry;
2392         }
2393
2394         *pbase = base;
2395         *plen = PATH_MAX - 1 - pos;
2396         dout("build_path on %p %d built %llx '%.*s'\n",
2397              dentry, d_count(dentry), base, *plen, path + pos);
2398         return path + pos;
2399 }
2400
2401 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
2402                              const char **ppath, int *ppathlen, u64 *pino,
2403                              bool *pfreepath, bool parent_locked)
2404 {
2405         char *path;
2406
2407         rcu_read_lock();
2408         if (!dir)
2409                 dir = d_inode_rcu(dentry->d_parent);
2410         if (dir && parent_locked && ceph_snap(dir) == CEPH_NOSNAP) {
2411                 *pino = ceph_ino(dir);
2412                 rcu_read_unlock();
2413                 *ppath = dentry->d_name.name;
2414                 *ppathlen = dentry->d_name.len;
2415                 return 0;
2416         }
2417         rcu_read_unlock();
2418         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2419         if (IS_ERR(path))
2420                 return PTR_ERR(path);
2421         *ppath = path;
2422         *pfreepath = true;
2423         return 0;
2424 }
2425
2426 static int build_inode_path(struct inode *inode,
2427                             const char **ppath, int *ppathlen, u64 *pino,
2428                             bool *pfreepath)
2429 {
2430         struct dentry *dentry;
2431         char *path;
2432
2433         if (ceph_snap(inode) == CEPH_NOSNAP) {
2434                 *pino = ceph_ino(inode);
2435                 *ppathlen = 0;
2436                 return 0;
2437         }
2438         dentry = d_find_alias(inode);
2439         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2440         dput(dentry);
2441         if (IS_ERR(path))
2442                 return PTR_ERR(path);
2443         *ppath = path;
2444         *pfreepath = true;
2445         return 0;
2446 }
2447
2448 /*
2449  * request arguments may be specified via an inode *, a dentry *, or
2450  * an explicit ino+path.
2451  */
2452 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
2453                                   struct inode *rdiri, const char *rpath,
2454                                   u64 rino, const char **ppath, int *pathlen,
2455                                   u64 *ino, bool *freepath, bool parent_locked)
2456 {
2457         int r = 0;
2458
2459         if (rinode) {
2460                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2461                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2462                      ceph_snap(rinode));
2463         } else if (rdentry) {
2464                 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
2465                                         freepath, parent_locked);
2466                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
2467                      *ppath);
2468         } else if (rpath || rino) {
2469                 *ino = rino;
2470                 *ppath = rpath;
2471                 *pathlen = rpath ? strlen(rpath) : 0;
2472                 dout(" path %.*s\n", *pathlen, rpath);
2473         }
2474
2475         return r;
2476 }
2477
2478 /*
2479  * called under mdsc->mutex
2480  */
2481 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
2482                                                struct ceph_mds_request *req,
2483                                                int mds, bool drop_cap_releases)
2484 {
2485         struct ceph_msg *msg;
2486         struct ceph_mds_request_head *head;
2487         const char *path1 = NULL;
2488         const char *path2 = NULL;
2489         u64 ino1 = 0, ino2 = 0;
2490         int pathlen1 = 0, pathlen2 = 0;
2491         bool freepath1 = false, freepath2 = false;
2492         int len;
2493         u16 releases;
2494         void *p, *end;
2495         int ret;
2496
2497         ret = set_request_path_attr(req->r_inode, req->r_dentry,
2498                               req->r_parent, req->r_path1, req->r_ino1.ino,
2499                               &path1, &pathlen1, &ino1, &freepath1,
2500                               test_bit(CEPH_MDS_R_PARENT_LOCKED,
2501                                         &req->r_req_flags));
2502         if (ret < 0) {
2503                 msg = ERR_PTR(ret);
2504                 goto out;
2505         }
2506
2507         /* If r_old_dentry is set, then assume that its parent is locked */
2508         ret = set_request_path_attr(NULL, req->r_old_dentry,
2509                               req->r_old_dentry_dir,
2510                               req->r_path2, req->r_ino2.ino,
2511                               &path2, &pathlen2, &ino2, &freepath2, true);
2512         if (ret < 0) {
2513                 msg = ERR_PTR(ret);
2514                 goto out_free1;
2515         }
2516
2517         len = sizeof(*head) +
2518                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2519                 sizeof(struct ceph_timespec);
2520
2521         /* calculate (max) length for cap releases */
2522         len += sizeof(struct ceph_mds_request_release) *
2523                 (!!req->r_inode_drop + !!req->r_dentry_drop +
2524                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2525         if (req->r_dentry_drop)
2526                 len += pathlen1;
2527         if (req->r_old_dentry_drop)
2528                 len += pathlen2;
2529
2530         msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
2531         if (!msg) {
2532                 msg = ERR_PTR(-ENOMEM);
2533                 goto out_free2;
2534         }
2535
2536         msg->hdr.version = cpu_to_le16(2);
2537         msg->hdr.tid = cpu_to_le64(req->r_tid);
2538
2539         head = msg->front.iov_base;
2540         p = msg->front.iov_base + sizeof(*head);
2541         end = msg->front.iov_base + msg->front.iov_len;
2542
2543         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2544         head->op = cpu_to_le32(req->r_op);
2545         head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
2546         head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
2547         head->ino = cpu_to_le64(req->r_deleg_ino);
2548         head->args = req->r_args;
2549
2550         ceph_encode_filepath(&p, end, ino1, path1);
2551         ceph_encode_filepath(&p, end, ino2, path2);
2552
2553         /* make note of release offset, in case we need to replay */
2554         req->r_request_release_offset = p - msg->front.iov_base;
2555
2556         /* cap releases */
2557         releases = 0;
2558         if (req->r_inode_drop)
2559                 releases += ceph_encode_inode_release(&p,
2560                       req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2561                       mds, req->r_inode_drop, req->r_inode_unless,
2562                       req->r_op == CEPH_MDS_OP_READDIR);
2563         if (req->r_dentry_drop)
2564                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
2565                                 req->r_parent, mds, req->r_dentry_drop,
2566                                 req->r_dentry_unless);
2567         if (req->r_old_dentry_drop)
2568                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2569                                 req->r_old_dentry_dir, mds,
2570                                 req->r_old_dentry_drop,
2571                                 req->r_old_dentry_unless);
2572         if (req->r_old_inode_drop)
2573                 releases += ceph_encode_inode_release(&p,
2574                       d_inode(req->r_old_dentry),
2575                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2576
2577         if (drop_cap_releases) {
2578                 releases = 0;
2579                 p = msg->front.iov_base + req->r_request_release_offset;
2580         }
2581
2582         head->num_releases = cpu_to_le16(releases);
2583
2584         /* time stamp */
2585         {
2586                 struct ceph_timespec ts;
2587                 ceph_encode_timespec64(&ts, &req->r_stamp);
2588                 ceph_encode_copy(&p, &ts, sizeof(ts));
2589         }
2590
2591         if (WARN_ON_ONCE(p > end)) {
2592                 ceph_msg_put(msg);
2593                 msg = ERR_PTR(-ERANGE);
2594                 goto out_free2;
2595         }
2596
2597         msg->front.iov_len = p - msg->front.iov_base;
2598         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2599
2600         if (req->r_pagelist) {
2601                 struct ceph_pagelist *pagelist = req->r_pagelist;
2602                 ceph_msg_data_add_pagelist(msg, pagelist);
2603                 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2604         } else {
2605                 msg->hdr.data_len = 0;
2606         }
2607
2608         msg->hdr.data_off = cpu_to_le16(0);
2609
2610 out_free2:
2611         if (freepath2)
2612                 ceph_mdsc_free_path((char *)path2, pathlen2);
2613 out_free1:
2614         if (freepath1)
2615                 ceph_mdsc_free_path((char *)path1, pathlen1);
2616 out:
2617         return msg;
2618 }
2619
2620 /*
2621  * called under mdsc->mutex if error, under no mutex if
2622  * success.
2623  */
2624 static void complete_request(struct ceph_mds_client *mdsc,
2625                              struct ceph_mds_request *req)
2626 {
2627         req->r_end_latency = ktime_get();
2628
2629         if (req->r_callback)
2630                 req->r_callback(mdsc, req);
2631         complete_all(&req->r_completion);
2632 }
2633
2634 /*
2635  * called under mdsc->mutex
2636  */
2637 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2638                                   struct ceph_mds_request *req,
2639                                   int mds, bool drop_cap_releases)
2640 {
2641         struct ceph_mds_request_head *rhead;
2642         struct ceph_msg *msg;
2643         int flags = 0;
2644
2645         req->r_attempts++;
2646         if (req->r_inode) {
2647                 struct ceph_cap *cap =
2648                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2649
2650                 if (cap)
2651                         req->r_sent_on_mseq = cap->mseq;
2652                 else
2653                         req->r_sent_on_mseq = -1;
2654         }
2655         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2656              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2657
2658         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2659                 void *p;
2660                 /*
2661                  * Replay.  Do not regenerate message (and rebuild
2662                  * paths, etc.); just use the original message.
2663                  * Rebuilding paths will break for renames because
2664                  * d_move mangles the src name.
2665                  */
2666                 msg = req->r_request;
2667                 rhead = msg->front.iov_base;
2668
2669                 flags = le32_to_cpu(rhead->flags);
2670                 flags |= CEPH_MDS_FLAG_REPLAY;
2671                 rhead->flags = cpu_to_le32(flags);
2672
2673                 if (req->r_target_inode)
2674                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2675
2676                 rhead->num_retry = req->r_attempts - 1;
2677
2678                 /* remove cap/dentry releases from message */
2679                 rhead->num_releases = 0;
2680
2681                 /* time stamp */
2682                 p = msg->front.iov_base + req->r_request_release_offset;
2683                 {
2684                         struct ceph_timespec ts;
2685                         ceph_encode_timespec64(&ts, &req->r_stamp);
2686                         ceph_encode_copy(&p, &ts, sizeof(ts));
2687                 }
2688
2689                 msg->front.iov_len = p - msg->front.iov_base;
2690                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2691                 return 0;
2692         }
2693
2694         if (req->r_request) {
2695                 ceph_msg_put(req->r_request);
2696                 req->r_request = NULL;
2697         }
2698         msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2699         if (IS_ERR(msg)) {
2700                 req->r_err = PTR_ERR(msg);
2701                 return PTR_ERR(msg);
2702         }
2703         req->r_request = msg;
2704
2705         rhead = msg->front.iov_base;
2706         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2707         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2708                 flags |= CEPH_MDS_FLAG_REPLAY;
2709         if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags))
2710                 flags |= CEPH_MDS_FLAG_ASYNC;
2711         if (req->r_parent)
2712                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2713         rhead->flags = cpu_to_le32(flags);
2714         rhead->num_fwd = req->r_num_fwd;
2715         rhead->num_retry = req->r_attempts - 1;
2716
2717         dout(" r_parent = %p\n", req->r_parent);
2718         return 0;
2719 }
2720
2721 /*
2722  * called under mdsc->mutex
2723  */
2724 static int __send_request(struct ceph_mds_client *mdsc,
2725                           struct ceph_mds_session *session,
2726                           struct ceph_mds_request *req,
2727                           bool drop_cap_releases)
2728 {
2729         int err;
2730
2731         err = __prepare_send_request(mdsc, req, session->s_mds,
2732                                      drop_cap_releases);
2733         if (!err) {
2734                 ceph_msg_get(req->r_request);
2735                 ceph_con_send(&session->s_con, req->r_request);
2736         }
2737
2738         return err;
2739 }
2740
2741 /*
2742  * send request, or put it on the appropriate wait list.
2743  */
2744 static void __do_request(struct ceph_mds_client *mdsc,
2745                         struct ceph_mds_request *req)
2746 {
2747         struct ceph_mds_session *session = NULL;
2748         int mds = -1;
2749         int err = 0;
2750         bool random;
2751
2752         if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2753                 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2754                         __unregister_request(mdsc, req);
2755                 return;
2756         }
2757
2758         if (req->r_timeout &&
2759             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2760                 dout("do_request timed out\n");
2761                 err = -ETIMEDOUT;
2762                 goto finish;
2763         }
2764         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2765                 dout("do_request forced umount\n");
2766                 err = -EIO;
2767                 goto finish;
2768         }
2769         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2770                 if (mdsc->mdsmap_err) {
2771                         err = mdsc->mdsmap_err;
2772                         dout("do_request mdsmap err %d\n", err);
2773                         goto finish;
2774                 }
2775                 if (mdsc->mdsmap->m_epoch == 0) {
2776                         dout("do_request no mdsmap, waiting for map\n");
2777                         list_add(&req->r_wait, &mdsc->waiting_for_map);
2778                         return;
2779                 }
2780                 if (!(mdsc->fsc->mount_options->flags &
2781                       CEPH_MOUNT_OPT_MOUNTWAIT) &&
2782                     !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2783                         err = -EHOSTUNREACH;
2784                         goto finish;
2785                 }
2786         }
2787
2788         put_request_session(req);
2789
2790         mds = __choose_mds(mdsc, req, &random);
2791         if (mds < 0 ||
2792             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2793                 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2794                         err = -EJUKEBOX;
2795                         goto finish;
2796                 }
2797                 dout("do_request no mds or not active, waiting for map\n");
2798                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2799                 return;
2800         }
2801
2802         /* get, open session */
2803         session = __ceph_lookup_mds_session(mdsc, mds);
2804         if (!session) {
2805                 session = register_session(mdsc, mds);
2806                 if (IS_ERR(session)) {
2807                         err = PTR_ERR(session);
2808                         goto finish;
2809                 }
2810         }
2811         req->r_session = ceph_get_mds_session(session);
2812
2813         dout("do_request mds%d session %p state %s\n", mds, session,
2814              ceph_session_state_name(session->s_state));
2815         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2816             session->s_state != CEPH_MDS_SESSION_HUNG) {
2817                 /*
2818                  * We cannot queue async requests since the caps and delegated
2819                  * inodes are bound to the session. Just return -EJUKEBOX and
2820                  * let the caller retry a sync request in that case.
2821                  */
2822                 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2823                         err = -EJUKEBOX;
2824                         goto out_session;
2825                 }
2826
2827                 /*
2828                  * If the session has been REJECTED, then return a hard error,
2829                  * unless it's a CLEANRECOVER mount, in which case we'll queue
2830                  * it to the mdsc queue.
2831                  */
2832                 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2833                         if (ceph_test_mount_opt(mdsc->fsc, CLEANRECOVER))
2834                                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2835                         else
2836                                 err = -EACCES;
2837                         goto out_session;
2838                 }
2839
2840                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2841                     session->s_state == CEPH_MDS_SESSION_CLOSING) {
2842                         err = __open_session(mdsc, session);
2843                         if (err)
2844                                 goto out_session;
2845                         /* retry the same mds later */
2846                         if (random)
2847                                 req->r_resend_mds = mds;
2848                 }
2849                 list_add(&req->r_wait, &session->s_waiting);
2850                 goto out_session;
2851         }
2852
2853         /* send request */
2854         req->r_resend_mds = -1;   /* forget any previous mds hint */
2855
2856         if (req->r_request_started == 0)   /* note request start time */
2857                 req->r_request_started = jiffies;
2858
2859         err = __send_request(mdsc, session, req, false);
2860
2861 out_session:
2862         ceph_put_mds_session(session);
2863 finish:
2864         if (err) {
2865                 dout("__do_request early error %d\n", err);
2866                 req->r_err = err;
2867                 complete_request(mdsc, req);
2868                 __unregister_request(mdsc, req);
2869         }
2870         return;
2871 }
2872
2873 /*
2874  * called under mdsc->mutex
2875  */
2876 static void __wake_requests(struct ceph_mds_client *mdsc,
2877                             struct list_head *head)
2878 {
2879         struct ceph_mds_request *req;
2880         LIST_HEAD(tmp_list);
2881
2882         list_splice_init(head, &tmp_list);
2883
2884         while (!list_empty(&tmp_list)) {
2885                 req = list_entry(tmp_list.next,
2886                                  struct ceph_mds_request, r_wait);
2887                 list_del_init(&req->r_wait);
2888                 dout(" wake request %p tid %llu\n", req, req->r_tid);
2889                 __do_request(mdsc, req);
2890         }
2891 }
2892
2893 /*
2894  * Wake up threads with requests pending for @mds, so that they can
2895  * resubmit their requests to a possibly different mds.
2896  */
2897 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2898 {
2899         struct ceph_mds_request *req;
2900         struct rb_node *p = rb_first(&mdsc->request_tree);
2901
2902         dout("kick_requests mds%d\n", mds);
2903         while (p) {
2904                 req = rb_entry(p, struct ceph_mds_request, r_node);
2905                 p = rb_next(p);
2906                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2907                         continue;
2908                 if (req->r_attempts > 0)
2909                         continue; /* only new requests */
2910                 if (req->r_session &&
2911                     req->r_session->s_mds == mds) {
2912                         dout(" kicking tid %llu\n", req->r_tid);
2913                         list_del_init(&req->r_wait);
2914                         __do_request(mdsc, req);
2915                 }
2916         }
2917 }
2918
2919 int ceph_mdsc_submit_request(struct ceph_mds_client *mdsc, struct inode *dir,
2920                               struct ceph_mds_request *req)
2921 {
2922         int err = 0;
2923
2924         /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2925         if (req->r_inode)
2926                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2927         if (req->r_parent) {
2928                 struct ceph_inode_info *ci = ceph_inode(req->r_parent);
2929                 int fmode = (req->r_op & CEPH_MDS_OP_WRITE) ?
2930                             CEPH_FILE_MODE_WR : CEPH_FILE_MODE_RD;
2931                 spin_lock(&ci->i_ceph_lock);
2932                 ceph_take_cap_refs(ci, CEPH_CAP_PIN, false);
2933                 __ceph_touch_fmode(ci, mdsc, fmode);
2934                 spin_unlock(&ci->i_ceph_lock);
2935                 ihold(req->r_parent);
2936         }
2937         if (req->r_old_dentry_dir)
2938                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2939                                   CEPH_CAP_PIN);
2940
2941         if (req->r_inode) {
2942                 err = ceph_wait_on_async_create(req->r_inode);
2943                 if (err) {
2944                         dout("%s: wait for async create returned: %d\n",
2945                              __func__, err);
2946                         return err;
2947                 }
2948         }
2949
2950         if (!err && req->r_old_inode) {
2951                 err = ceph_wait_on_async_create(req->r_old_inode);
2952                 if (err) {
2953                         dout("%s: wait for async create returned: %d\n",
2954                              __func__, err);
2955                         return err;
2956                 }
2957         }
2958
2959         dout("submit_request on %p for inode %p\n", req, dir);
2960         mutex_lock(&mdsc->mutex);
2961         __register_request(mdsc, req, dir);
2962         __do_request(mdsc, req);
2963         err = req->r_err;
2964         mutex_unlock(&mdsc->mutex);
2965         return err;
2966 }
2967
2968 static int ceph_mdsc_wait_request(struct ceph_mds_client *mdsc,
2969                                   struct ceph_mds_request *req)
2970 {
2971         int err;
2972
2973         /* wait */
2974         dout("do_request waiting\n");
2975         if (!req->r_timeout && req->r_wait_for_completion) {
2976                 err = req->r_wait_for_completion(mdsc, req);
2977         } else {
2978                 long timeleft = wait_for_completion_killable_timeout(
2979                                         &req->r_completion,
2980                                         ceph_timeout_jiffies(req->r_timeout));
2981                 if (timeleft > 0)
2982                         err = 0;
2983                 else if (!timeleft)
2984                         err = -ETIMEDOUT;  /* timed out */
2985                 else
2986                         err = timeleft;  /* killed */
2987         }
2988         dout("do_request waited, got %d\n", err);
2989         mutex_lock(&mdsc->mutex);
2990
2991         /* only abort if we didn't race with a real reply */
2992         if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2993                 err = le32_to_cpu(req->r_reply_info.head->result);
2994         } else if (err < 0) {
2995                 dout("aborted request %lld with %d\n", req->r_tid, err);
2996
2997                 /*
2998                  * ensure we aren't running concurrently with
2999                  * ceph_fill_trace or ceph_readdir_prepopulate, which
3000                  * rely on locks (dir mutex) held by our caller.
3001                  */
3002                 mutex_lock(&req->r_fill_mutex);
3003                 req->r_err = err;
3004                 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
3005                 mutex_unlock(&req->r_fill_mutex);
3006
3007                 if (req->r_parent &&
3008                     (req->r_op & CEPH_MDS_OP_WRITE))
3009                         ceph_invalidate_dir_request(req);
3010         } else {
3011                 err = req->r_err;
3012         }
3013
3014         mutex_unlock(&mdsc->mutex);
3015         return err;
3016 }
3017
3018 /*
3019  * Synchrously perform an mds request.  Take care of all of the
3020  * session setup, forwarding, retry details.
3021  */
3022 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
3023                          struct inode *dir,
3024                          struct ceph_mds_request *req)
3025 {
3026         int err;
3027
3028         dout("do_request on %p\n", req);
3029
3030         /* issue */
3031         err = ceph_mdsc_submit_request(mdsc, dir, req);
3032         if (!err)
3033                 err = ceph_mdsc_wait_request(mdsc, req);
3034         dout("do_request %p done, result %d\n", req, err);
3035         return err;
3036 }
3037
3038 /*
3039  * Invalidate dir's completeness, dentry lease state on an aborted MDS
3040  * namespace request.
3041  */
3042 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
3043 {
3044         struct inode *dir = req->r_parent;
3045         struct inode *old_dir = req->r_old_dentry_dir;
3046
3047         dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
3048
3049         ceph_dir_clear_complete(dir);
3050         if (old_dir)
3051                 ceph_dir_clear_complete(old_dir);
3052         if (req->r_dentry)
3053                 ceph_invalidate_dentry_lease(req->r_dentry);
3054         if (req->r_old_dentry)
3055                 ceph_invalidate_dentry_lease(req->r_old_dentry);
3056 }
3057
3058 /*
3059  * Handle mds reply.
3060  *
3061  * We take the session mutex and parse and process the reply immediately.
3062  * This preserves the logical ordering of replies, capabilities, etc., sent
3063  * by the MDS as they are applied to our local cache.
3064  */
3065 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
3066 {
3067         struct ceph_mds_client *mdsc = session->s_mdsc;
3068         struct ceph_mds_request *req;
3069         struct ceph_mds_reply_head *head = msg->front.iov_base;
3070         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
3071         struct ceph_snap_realm *realm;
3072         u64 tid;
3073         int err, result;
3074         int mds = session->s_mds;
3075
3076         if (msg->front.iov_len < sizeof(*head)) {
3077                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
3078                 ceph_msg_dump(msg);
3079                 return;
3080         }
3081
3082         /* get request, session */
3083         tid = le64_to_cpu(msg->hdr.tid);
3084         mutex_lock(&mdsc->mutex);
3085         req = lookup_get_request(mdsc, tid);
3086         if (!req) {
3087                 dout("handle_reply on unknown tid %llu\n", tid);
3088                 mutex_unlock(&mdsc->mutex);
3089                 return;
3090         }
3091         dout("handle_reply %p\n", req);
3092
3093         /* correct session? */
3094         if (req->r_session != session) {
3095                 pr_err("mdsc_handle_reply got %llu on session mds%d"
3096                        " not mds%d\n", tid, session->s_mds,
3097                        req->r_session ? req->r_session->s_mds : -1);
3098                 mutex_unlock(&mdsc->mutex);
3099                 goto out;
3100         }
3101
3102         /* dup? */
3103         if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
3104             (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
3105                 pr_warn("got a dup %s reply on %llu from mds%d\n",
3106                            head->safe ? "safe" : "unsafe", tid, mds);
3107                 mutex_unlock(&mdsc->mutex);
3108                 goto out;
3109         }
3110         if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
3111                 pr_warn("got unsafe after safe on %llu from mds%d\n",
3112                            tid, mds);
3113                 mutex_unlock(&mdsc->mutex);
3114                 goto out;
3115         }
3116
3117         result = le32_to_cpu(head->result);
3118
3119         /*
3120          * Handle an ESTALE
3121          * if we're not talking to the authority, send to them
3122          * if the authority has changed while we weren't looking,
3123          * send to new authority
3124          * Otherwise we just have to return an ESTALE
3125          */
3126         if (result == -ESTALE) {
3127                 dout("got ESTALE on request %llu\n", req->r_tid);
3128                 req->r_resend_mds = -1;
3129                 if (req->r_direct_mode != USE_AUTH_MDS) {
3130                         dout("not using auth, setting for that now\n");
3131                         req->r_direct_mode = USE_AUTH_MDS;
3132                         __do_request(mdsc, req);
3133                         mutex_unlock(&mdsc->mutex);
3134                         goto out;
3135                 } else  {
3136                         int mds = __choose_mds(mdsc, req, NULL);
3137                         if (mds >= 0 && mds != req->r_session->s_mds) {
3138                                 dout("but auth changed, so resending\n");
3139                                 __do_request(mdsc, req);
3140                                 mutex_unlock(&mdsc->mutex);
3141                                 goto out;
3142                         }
3143                 }
3144                 dout("have to return ESTALE on request %llu\n", req->r_tid);
3145         }
3146
3147
3148         if (head->safe) {
3149                 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
3150                 __unregister_request(mdsc, req);
3151
3152                 /* last request during umount? */
3153                 if (mdsc->stopping && !__get_oldest_req(mdsc))
3154                         complete_all(&mdsc->safe_umount_waiters);
3155
3156                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3157                         /*
3158                          * We already handled the unsafe response, now do the
3159                          * cleanup.  No need to examine the response; the MDS
3160                          * doesn't include any result info in the safe
3161                          * response.  And even if it did, there is nothing
3162                          * useful we could do with a revised return value.
3163                          */
3164                         dout("got safe reply %llu, mds%d\n", tid, mds);
3165
3166                         mutex_unlock(&mdsc->mutex);
3167                         goto out;
3168                 }
3169         } else {
3170                 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
3171                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
3172         }
3173
3174         dout("handle_reply tid %lld result %d\n", tid, result);
3175         rinfo = &req->r_reply_info;
3176         if (test_bit(CEPHFS_FEATURE_REPLY_ENCODING, &session->s_features))
3177                 err = parse_reply_info(session, msg, rinfo, (u64)-1);
3178         else
3179                 err = parse_reply_info(session, msg, rinfo, session->s_con.peer_features);
3180         mutex_unlock(&mdsc->mutex);
3181
3182         /* Must find target inode outside of mutexes to avoid deadlocks */
3183         if ((err >= 0) && rinfo->head->is_target) {
3184                 struct inode *in;
3185                 struct ceph_vino tvino = {
3186                         .ino  = le64_to_cpu(rinfo->targeti.in->ino),
3187                         .snap = le64_to_cpu(rinfo->targeti.in->snapid)
3188                 };
3189
3190                 in = ceph_get_inode(mdsc->fsc->sb, tvino);
3191                 if (IS_ERR(in)) {
3192                         err = PTR_ERR(in);
3193                         mutex_lock(&session->s_mutex);
3194                         goto out_err;
3195                 }
3196                 req->r_target_inode = in;
3197         }
3198
3199         mutex_lock(&session->s_mutex);
3200         if (err < 0) {
3201                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
3202                 ceph_msg_dump(msg);
3203                 goto out_err;
3204         }
3205
3206         /* snap trace */
3207         realm = NULL;
3208         if (rinfo->snapblob_len) {
3209                 down_write(&mdsc->snap_rwsem);
3210                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
3211                                 rinfo->snapblob + rinfo->snapblob_len,
3212                                 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
3213                                 &realm);
3214                 downgrade_write(&mdsc->snap_rwsem);
3215         } else {
3216                 down_read(&mdsc->snap_rwsem);
3217         }
3218
3219         /* insert trace into our cache */
3220         mutex_lock(&req->r_fill_mutex);
3221         current->journal_info = req;
3222         err = ceph_fill_trace(mdsc->fsc->sb, req);
3223         if (err == 0) {
3224                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
3225                                     req->r_op == CEPH_MDS_OP_LSSNAP))
3226                         ceph_readdir_prepopulate(req, req->r_session);
3227         }
3228         current->journal_info = NULL;
3229         mutex_unlock(&req->r_fill_mutex);
3230
3231         up_read(&mdsc->snap_rwsem);
3232         if (realm)
3233                 ceph_put_snap_realm(mdsc, realm);
3234
3235         if (err == 0) {
3236                 if (req->r_target_inode &&
3237                     test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3238                         struct ceph_inode_info *ci =
3239                                 ceph_inode(req->r_target_inode);
3240                         spin_lock(&ci->i_unsafe_lock);
3241                         list_add_tail(&req->r_unsafe_target_item,
3242                                       &ci->i_unsafe_iops);
3243                         spin_unlock(&ci->i_unsafe_lock);
3244                 }
3245
3246                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
3247         }
3248 out_err:
3249         mutex_lock(&mdsc->mutex);
3250         if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3251                 if (err) {
3252                         req->r_err = err;
3253                 } else {
3254                         req->r_reply =  ceph_msg_get(msg);
3255                         set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
3256                 }
3257         } else {
3258                 dout("reply arrived after request %lld was aborted\n", tid);
3259         }
3260         mutex_unlock(&mdsc->mutex);
3261
3262         mutex_unlock(&session->s_mutex);
3263
3264         /* kick calling process */
3265         complete_request(mdsc, req);
3266
3267         ceph_update_metadata_latency(&mdsc->metric, req->r_start_latency,
3268                                      req->r_end_latency, err);
3269 out:
3270         ceph_mdsc_put_request(req);
3271         return;
3272 }
3273
3274
3275
3276 /*
3277  * handle mds notification that our request has been forwarded.
3278  */
3279 static void handle_forward(struct ceph_mds_client *mdsc,
3280                            struct ceph_mds_session *session,
3281                            struct ceph_msg *msg)
3282 {
3283         struct ceph_mds_request *req;
3284         u64 tid = le64_to_cpu(msg->hdr.tid);
3285         u32 next_mds;
3286         u32 fwd_seq;
3287         int err = -EINVAL;
3288         void *p = msg->front.iov_base;
3289         void *end = p + msg->front.iov_len;
3290
3291         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
3292         next_mds = ceph_decode_32(&p);
3293         fwd_seq = ceph_decode_32(&p);
3294
3295         mutex_lock(&mdsc->mutex);
3296         req = lookup_get_request(mdsc, tid);
3297         if (!req) {
3298                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
3299                 goto out;  /* dup reply? */
3300         }
3301
3302         if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3303                 dout("forward tid %llu aborted, unregistering\n", tid);
3304                 __unregister_request(mdsc, req);
3305         } else if (fwd_seq <= req->r_num_fwd) {
3306                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
3307                      tid, next_mds, req->r_num_fwd, fwd_seq);
3308         } else {
3309                 /* resend. forward race not possible; mds would drop */
3310                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
3311                 BUG_ON(req->r_err);
3312                 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
3313                 req->r_attempts = 0;
3314                 req->r_num_fwd = fwd_seq;
3315                 req->r_resend_mds = next_mds;
3316                 put_request_session(req);
3317                 __do_request(mdsc, req);
3318         }
3319         ceph_mdsc_put_request(req);
3320 out:
3321         mutex_unlock(&mdsc->mutex);
3322         return;
3323
3324 bad:
3325         pr_err("mdsc_handle_forward decode error err=%d\n", err);
3326 }
3327
3328 static int __decode_session_metadata(void **p, void *end,
3329                                      bool *blocklisted)
3330 {
3331         /* map<string,string> */
3332         u32 n;
3333         bool err_str;
3334         ceph_decode_32_safe(p, end, n, bad);
3335         while (n-- > 0) {
3336                 u32 len;
3337                 ceph_decode_32_safe(p, end, len, bad);
3338                 ceph_decode_need(p, end, len, bad);
3339                 err_str = !strncmp(*p, "error_string", len);
3340                 *p += len;
3341                 ceph_decode_32_safe(p, end, len, bad);
3342                 ceph_decode_need(p, end, len, bad);
3343                 /*
3344                  * Match "blocklisted (blacklisted)" from newer MDSes,
3345                  * or "blacklisted" from older MDSes.
3346                  */
3347                 if (err_str && strnstr(*p, "blacklisted", len))
3348                         *blocklisted = true;
3349                 *p += len;
3350         }
3351         return 0;
3352 bad:
3353         return -1;
3354 }
3355
3356 /*
3357  * handle a mds session control message
3358  */
3359 static void handle_session(struct ceph_mds_session *session,
3360                            struct ceph_msg *msg)
3361 {
3362         struct ceph_mds_client *mdsc = session->s_mdsc;
3363         int mds = session->s_mds;
3364         int msg_version = le16_to_cpu(msg->hdr.version);
3365         void *p = msg->front.iov_base;
3366         void *end = p + msg->front.iov_len;
3367         struct ceph_mds_session_head *h;
3368         u32 op;
3369         u64 seq, features = 0;
3370         int wake = 0;
3371         bool blocklisted = false;
3372
3373         /* decode */
3374         ceph_decode_need(&p, end, sizeof(*h), bad);
3375         h = p;
3376         p += sizeof(*h);
3377
3378         op = le32_to_cpu(h->op);
3379         seq = le64_to_cpu(h->seq);
3380
3381         if (msg_version >= 3) {
3382                 u32 len;
3383                 /* version >= 2, metadata */
3384                 if (__decode_session_metadata(&p, end, &blocklisted) < 0)
3385                         goto bad;
3386                 /* version >= 3, feature bits */
3387                 ceph_decode_32_safe(&p, end, len, bad);
3388                 if (len) {
3389                         ceph_decode_64_safe(&p, end, features, bad);
3390                         p += len - sizeof(features);
3391                 }
3392         }
3393
3394         mutex_lock(&mdsc->mutex);
3395         if (op == CEPH_SESSION_CLOSE) {
3396                 ceph_get_mds_session(session);
3397                 __unregister_session(mdsc, session);
3398         }
3399         /* FIXME: this ttl calculation is generous */
3400         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
3401         mutex_unlock(&mdsc->mutex);
3402
3403         mutex_lock(&session->s_mutex);
3404
3405         dout("handle_session mds%d %s %p state %s seq %llu\n",
3406              mds, ceph_session_op_name(op), session,
3407              ceph_session_state_name(session->s_state), seq);
3408
3409         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
3410                 session->s_state = CEPH_MDS_SESSION_OPEN;
3411                 pr_info("mds%d came back\n", session->s_mds);
3412         }
3413
3414         switch (op) {
3415         case CEPH_SESSION_OPEN:
3416                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3417                         pr_info("mds%d reconnect success\n", session->s_mds);
3418                 session->s_state = CEPH_MDS_SESSION_OPEN;
3419                 session->s_features = features;
3420                 renewed_caps(mdsc, session, 0);
3421                 if (test_bit(CEPHFS_FEATURE_METRIC_COLLECT, &session->s_features))
3422                         metric_schedule_delayed(&mdsc->metric);
3423                 wake = 1;
3424                 if (mdsc->stopping)
3425                         __close_session(mdsc, session);
3426                 break;
3427
3428         case CEPH_SESSION_RENEWCAPS:
3429                 if (session->s_renew_seq == seq)
3430                         renewed_caps(mdsc, session, 1);
3431                 break;
3432
3433         case CEPH_SESSION_CLOSE:
3434                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3435                         pr_info("mds%d reconnect denied\n", session->s_mds);
3436                 session->s_state = CEPH_MDS_SESSION_CLOSED;
3437                 cleanup_session_requests(mdsc, session);
3438                 remove_session_caps(session);
3439                 wake = 2; /* for good measure */
3440                 wake_up_all(&mdsc->session_close_wq);
3441                 break;
3442
3443         case CEPH_SESSION_STALE:
3444                 pr_info("mds%d caps went stale, renewing\n",
3445                         session->s_mds);
3446                 spin_lock(&session->s_gen_ttl_lock);
3447                 session->s_cap_gen++;
3448                 session->s_cap_ttl = jiffies - 1;
3449                 spin_unlock(&session->s_gen_ttl_lock);
3450                 send_renew_caps(mdsc, session);
3451                 break;
3452
3453         case CEPH_SESSION_RECALL_STATE:
3454                 ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
3455                 break;
3456
3457         case CEPH_SESSION_FLUSHMSG:
3458                 send_flushmsg_ack(mdsc, session, seq);
3459                 break;
3460
3461         case CEPH_SESSION_FORCE_RO:
3462                 dout("force_session_readonly %p\n", session);
3463                 spin_lock(&session->s_cap_lock);
3464                 session->s_readonly = true;
3465                 spin_unlock(&session->s_cap_lock);
3466                 wake_up_session_caps(session, FORCE_RO);
3467                 break;
3468
3469         case CEPH_SESSION_REJECT:
3470                 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
3471                 pr_info("mds%d rejected session\n", session->s_mds);
3472                 session->s_state = CEPH_MDS_SESSION_REJECTED;
3473                 cleanup_session_requests(mdsc, session);
3474                 remove_session_caps(session);
3475                 if (blocklisted)
3476                         mdsc->fsc->blocklisted = true;
3477                 wake = 2; /* for good measure */
3478                 break;
3479
3480         default:
3481                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
3482                 WARN_ON(1);
3483         }
3484
3485         mutex_unlock(&session->s_mutex);
3486         if (wake) {
3487                 mutex_lock(&mdsc->mutex);
3488                 __wake_requests(mdsc, &session->s_waiting);
3489                 if (wake == 2)
3490                         kick_requests(mdsc, mds);
3491                 mutex_unlock(&mdsc->mutex);
3492         }
3493         if (op == CEPH_SESSION_CLOSE)
3494                 ceph_put_mds_session(session);
3495         return;
3496
3497 bad:
3498         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
3499                (int)msg->front.iov_len);
3500         ceph_msg_dump(msg);
3501         return;
3502 }
3503
3504 void ceph_mdsc_release_dir_caps(struct ceph_mds_request *req)
3505 {
3506         int dcaps;
3507
3508         dcaps = xchg(&req->r_dir_caps, 0);
3509         if (dcaps) {
3510                 dout("releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
3511                 ceph_put_cap_refs(ceph_inode(req->r_parent), dcaps);
3512         }
3513 }
3514
3515 void ceph_mdsc_release_dir_caps_no_check(struct ceph_mds_request *req)
3516 {
3517         int dcaps;
3518
3519         dcaps = xchg(&req->r_dir_caps, 0);
3520         if (dcaps) {
3521                 dout("releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
3522                 ceph_put_cap_refs_no_check_caps(ceph_inode(req->r_parent),
3523                                                 dcaps);
3524         }
3525 }
3526
3527 /*
3528  * called under session->mutex.
3529  */
3530 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
3531                                    struct ceph_mds_session *session)
3532 {
3533         struct ceph_mds_request *req, *nreq;
3534         struct rb_node *p;
3535
3536         dout("replay_unsafe_requests mds%d\n", session->s_mds);
3537
3538         mutex_lock(&mdsc->mutex);
3539         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item)
3540                 __send_request(mdsc, session, req, true);
3541
3542         /*
3543          * also re-send old requests when MDS enters reconnect stage. So that MDS
3544          * can process completed request in clientreplay stage.
3545          */
3546         p = rb_first(&mdsc->request_tree);
3547         while (p) {
3548                 req = rb_entry(p, struct ceph_mds_request, r_node);
3549                 p = rb_next(p);
3550                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3551                         continue;
3552                 if (req->r_attempts == 0)
3553                         continue; /* only old requests */
3554                 if (!req->r_session)
3555                         continue;
3556                 if (req->r_session->s_mds != session->s_mds)
3557                         continue;
3558
3559                 ceph_mdsc_release_dir_caps_no_check(req);
3560
3561                 __send_request(mdsc, session, req, true);
3562         }
3563         mutex_unlock(&mdsc->mutex);
3564 }
3565
3566 static int send_reconnect_partial(struct ceph_reconnect_state *recon_state)
3567 {
3568         struct ceph_msg *reply;
3569         struct ceph_pagelist *_pagelist;
3570         struct page *page;
3571         __le32 *addr;
3572         int err = -ENOMEM;
3573
3574         if (!recon_state->allow_multi)
3575                 return -ENOSPC;
3576
3577         /* can't handle message that contains both caps and realm */
3578         BUG_ON(!recon_state->nr_caps == !recon_state->nr_realms);
3579
3580         /* pre-allocate new pagelist */
3581         _pagelist = ceph_pagelist_alloc(GFP_NOFS);
3582         if (!_pagelist)
3583                 return -ENOMEM;
3584
3585         reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3586         if (!reply)
3587                 goto fail_msg;
3588
3589         /* placeholder for nr_caps */
3590         err = ceph_pagelist_encode_32(_pagelist, 0);
3591         if (err < 0)
3592                 goto fail;
3593
3594         if (recon_state->nr_caps) {
3595                 /* currently encoding caps */
3596                 err = ceph_pagelist_encode_32(recon_state->pagelist, 0);
3597                 if (err)
3598                         goto fail;
3599         } else {
3600                 /* placeholder for nr_realms (currently encoding relams) */
3601                 err = ceph_pagelist_encode_32(_pagelist, 0);
3602                 if (err < 0)
3603                         goto fail;
3604         }
3605
3606         err = ceph_pagelist_encode_8(recon_state->pagelist, 1);
3607         if (err)
3608                 goto fail;
3609
3610         page = list_first_entry(&recon_state->pagelist->head, struct page, lru);
3611         addr = kmap_atomic(page);
3612         if (recon_state->nr_caps) {
3613                 /* currently encoding caps */
3614                 *addr = cpu_to_le32(recon_state->nr_caps);
3615         } else {
3616                 /* currently encoding relams */
3617                 *(addr + 1) = cpu_to_le32(recon_state->nr_realms);
3618         }
3619         kunmap_atomic(addr);
3620
3621         reply->hdr.version = cpu_to_le16(5);
3622         reply->hdr.compat_version = cpu_to_le16(4);
3623
3624         reply->hdr.data_len = cpu_to_le32(recon_state->pagelist->length);
3625         ceph_msg_data_add_pagelist(reply, recon_state->pagelist);
3626
3627         ceph_con_send(&recon_state->session->s_con, reply);
3628         ceph_pagelist_release(recon_state->pagelist);
3629
3630         recon_state->pagelist = _pagelist;
3631         recon_state->nr_caps = 0;
3632         recon_state->nr_realms = 0;
3633         recon_state->msg_version = 5;
3634         return 0;
3635 fail:
3636         ceph_msg_put(reply);
3637 fail_msg:
3638         ceph_pagelist_release(_pagelist);
3639         return err;
3640 }
3641
3642 static struct dentry* d_find_primary(struct inode *inode)
3643 {
3644         struct dentry *alias, *dn = NULL;
3645
3646         if (hlist_empty(&inode->i_dentry))
3647                 return NULL;
3648
3649         spin_lock(&inode->i_lock);
3650         if (hlist_empty(&inode->i_dentry))
3651                 goto out_unlock;
3652
3653         if (S_ISDIR(inode->i_mode)) {
3654                 alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias);
3655                 if (!IS_ROOT(alias))
3656                         dn = dget(alias);
3657                 goto out_unlock;
3658         }
3659
3660         hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
3661                 spin_lock(&alias->d_lock);
3662                 if (!d_unhashed(alias) &&
3663                     (ceph_dentry(alias)->flags & CEPH_DENTRY_PRIMARY_LINK)) {
3664                         dn = dget_dlock(alias);
3665                 }
3666                 spin_unlock(&alias->d_lock);
3667                 if (dn)
3668                         break;
3669         }
3670 out_unlock:
3671         spin_unlock(&inode->i_lock);
3672         return dn;
3673 }
3674
3675 /*
3676  * Encode information about a cap for a reconnect with the MDS.
3677  */
3678 static int reconnect_caps_cb(struct inode *inode, struct ceph_cap *cap,
3679                           void *arg)
3680 {
3681         union {
3682                 struct ceph_mds_cap_reconnect v2;
3683                 struct ceph_mds_cap_reconnect_v1 v1;
3684         } rec;
3685         struct ceph_inode_info *ci = cap->ci;
3686         struct ceph_reconnect_state *recon_state = arg;
3687         struct ceph_pagelist *pagelist = recon_state->pagelist;
3688         struct dentry *dentry;
3689         char *path;
3690         int pathlen, err;
3691         u64 pathbase;
3692         u64 snap_follows;
3693
3694         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
3695              inode, ceph_vinop(inode), cap, cap->cap_id,
3696              ceph_cap_string(cap->issued));
3697
3698         dentry = d_find_primary(inode);
3699         if (dentry) {
3700                 /* set pathbase to parent dir when msg_version >= 2 */
3701                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase,
3702                                             recon_state->msg_version >= 2);
3703                 dput(dentry);
3704                 if (IS_ERR(path)) {
3705                         err = PTR_ERR(path);
3706                         goto out_err;
3707                 }
3708         } else {
3709                 path = NULL;
3710                 pathlen = 0;
3711                 pathbase = 0;
3712         }
3713
3714         spin_lock(&ci->i_ceph_lock);
3715         cap->seq = 0;        /* reset cap seq */
3716         cap->issue_seq = 0;  /* and issue_seq */
3717         cap->mseq = 0;       /* and migrate_seq */
3718         cap->cap_gen = cap->session->s_cap_gen;
3719
3720         /* These are lost when the session goes away */
3721         if (S_ISDIR(inode->i_mode)) {
3722                 if (cap->issued & CEPH_CAP_DIR_CREATE) {
3723                         ceph_put_string(rcu_dereference_raw(ci->i_cached_layout.pool_ns));
3724                         memset(&ci->i_cached_layout, 0, sizeof(ci->i_cached_layout));
3725                 }
3726                 cap->issued &= ~CEPH_CAP_ANY_DIR_OPS;
3727         }
3728
3729         if (recon_state->msg_version >= 2) {
3730                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
3731                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3732                 rec.v2.issued = cpu_to_le32(cap->issued);
3733                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3734                 rec.v2.pathbase = cpu_to_le64(pathbase);
3735                 rec.v2.flock_len = (__force __le32)
3736                         ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
3737         } else {
3738                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
3739                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3740                 rec.v1.issued = cpu_to_le32(cap->issued);
3741                 rec.v1.size = cpu_to_le64(inode->i_size);
3742                 ceph_encode_timespec64(&rec.v1.mtime, &inode->i_mtime);
3743                 ceph_encode_timespec64(&rec.v1.atime, &inode->i_atime);
3744                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3745                 rec.v1.pathbase = cpu_to_le64(pathbase);
3746         }
3747
3748         if (list_empty(&ci->i_cap_snaps)) {
3749                 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
3750         } else {
3751                 struct ceph_cap_snap *capsnap =
3752                         list_first_entry(&ci->i_cap_snaps,
3753                                          struct ceph_cap_snap, ci_item);
3754                 snap_follows = capsnap->follows;
3755         }
3756         spin_unlock(&ci->i_ceph_lock);
3757
3758         if (recon_state->msg_version >= 2) {
3759                 int num_fcntl_locks, num_flock_locks;
3760                 struct ceph_filelock *flocks = NULL;
3761                 size_t struct_len, total_len = sizeof(u64);
3762                 u8 struct_v = 0;
3763
3764 encode_again:
3765                 if (rec.v2.flock_len) {
3766                         ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
3767                 } else {
3768                         num_fcntl_locks = 0;
3769                         num_flock_locks = 0;
3770                 }
3771                 if (num_fcntl_locks + num_flock_locks > 0) {
3772                         flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
3773                                                sizeof(struct ceph_filelock),
3774                                                GFP_NOFS);
3775                         if (!flocks) {
3776                                 err = -ENOMEM;
3777                                 goto out_err;
3778                         }
3779                         err = ceph_encode_locks_to_buffer(inode, flocks,
3780                                                           num_fcntl_locks,
3781                                                           num_flock_locks);
3782                         if (err) {
3783                                 kfree(flocks);
3784                                 flocks = NULL;
3785                                 if (err == -ENOSPC)
3786                                         goto encode_again;
3787                                 goto out_err;
3788                         }
3789                 } else {
3790                         kfree(flocks);
3791                         flocks = NULL;
3792                 }
3793
3794                 if (recon_state->msg_version >= 3) {
3795                         /* version, compat_version and struct_len */
3796                         total_len += 2 * sizeof(u8) + sizeof(u32);
3797                         struct_v = 2;
3798                 }
3799                 /*
3800                  * number of encoded locks is stable, so copy to pagelist
3801                  */
3802                 struct_len = 2 * sizeof(u32) +
3803                             (num_fcntl_locks + num_flock_locks) *
3804                             sizeof(struct ceph_filelock);
3805                 rec.v2.flock_len = cpu_to_le32(struct_len);
3806
3807                 struct_len += sizeof(u32) + pathlen + sizeof(rec.v2);
3808
3809                 if (struct_v >= 2)
3810                         struct_len += sizeof(u64); /* snap_follows */
3811
3812                 total_len += struct_len;
3813
3814                 if (pagelist->length + total_len > RECONNECT_MAX_SIZE) {
3815                         err = send_reconnect_partial(recon_state);
3816                         if (err)
3817                                 goto out_freeflocks;
3818                         pagelist = recon_state->pagelist;
3819                 }
3820
3821                 err = ceph_pagelist_reserve(pagelist, total_len);
3822                 if (err)
3823                         goto out_freeflocks;
3824
3825                 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3826                 if (recon_state->msg_version >= 3) {
3827                         ceph_pagelist_encode_8(pagelist, struct_v);
3828                         ceph_pagelist_encode_8(pagelist, 1);
3829                         ceph_pagelist_encode_32(pagelist, struct_len);
3830                 }
3831                 ceph_pagelist_encode_string(pagelist, path, pathlen);
3832                 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
3833                 ceph_locks_to_pagelist(flocks, pagelist,
3834                                        num_fcntl_locks, num_flock_locks);
3835                 if (struct_v >= 2)
3836                         ceph_pagelist_encode_64(pagelist, snap_follows);
3837 out_freeflocks:
3838                 kfree(flocks);
3839         } else {
3840                 err = ceph_pagelist_reserve(pagelist,
3841                                             sizeof(u64) + sizeof(u32) +
3842                                             pathlen + sizeof(rec.v1));
3843                 if (err)
3844                         goto out_err;
3845
3846                 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3847                 ceph_pagelist_encode_string(pagelist, path, pathlen);
3848                 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
3849         }
3850
3851 out_err:
3852         ceph_mdsc_free_path(path, pathlen);
3853         if (!err)
3854                 recon_state->nr_caps++;
3855         return err;
3856 }
3857
3858 static int encode_snap_realms(struct ceph_mds_client *mdsc,
3859                               struct ceph_reconnect_state *recon_state)
3860 {
3861         struct rb_node *p;
3862         struct ceph_pagelist *pagelist = recon_state->pagelist;
3863         int err = 0;
3864
3865         if (recon_state->msg_version >= 4) {
3866                 err = ceph_pagelist_encode_32(pagelist, mdsc->num_snap_realms);
3867                 if (err < 0)
3868                         goto fail;
3869         }
3870
3871         /*
3872          * snaprealms.  we provide mds with the ino, seq (version), and
3873          * parent for all of our realms.  If the mds has any newer info,
3874          * it will tell us.
3875          */
3876         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3877                 struct ceph_snap_realm *realm =
3878                        rb_entry(p, struct ceph_snap_realm, node);
3879                 struct ceph_mds_snaprealm_reconnect sr_rec;
3880
3881                 if (recon_state->msg_version >= 4) {
3882                         size_t need = sizeof(u8) * 2 + sizeof(u32) +
3883                                       sizeof(sr_rec);
3884
3885                         if (pagelist->length + need > RECONNECT_MAX_SIZE) {
3886                                 err = send_reconnect_partial(recon_state);
3887                                 if (err)
3888                                         goto fail;
3889                                 pagelist = recon_state->pagelist;
3890                         }
3891
3892                         err = ceph_pagelist_reserve(pagelist, need);
3893                         if (err)
3894                                 goto fail;
3895
3896                         ceph_pagelist_encode_8(pagelist, 1);
3897                         ceph_pagelist_encode_8(pagelist, 1);
3898                         ceph_pagelist_encode_32(pagelist, sizeof(sr_rec));
3899                 }
3900
3901                 dout(" adding snap realm %llx seq %lld parent %llx\n",
3902                      realm->ino, realm->seq, realm->parent_ino);
3903                 sr_rec.ino = cpu_to_le64(realm->ino);
3904                 sr_rec.seq = cpu_to_le64(realm->seq);
3905                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3906
3907                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3908                 if (err)
3909                         goto fail;
3910
3911                 recon_state->nr_realms++;
3912         }
3913 fail:
3914         return err;
3915 }
3916
3917
3918 /*
3919  * If an MDS fails and recovers, clients need to reconnect in order to
3920  * reestablish shared state.  This includes all caps issued through
3921  * this session _and_ the snap_realm hierarchy.  Because it's not
3922  * clear which snap realms the mds cares about, we send everything we
3923  * know about.. that ensures we'll then get any new info the
3924  * recovering MDS might have.
3925  *
3926  * This is a relatively heavyweight operation, but it's rare.
3927  */
3928 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3929                                struct ceph_mds_session *session)
3930 {
3931         struct ceph_msg *reply;
3932         int mds = session->s_mds;
3933         int err = -ENOMEM;
3934         struct ceph_reconnect_state recon_state = {
3935                 .session = session,
3936         };
3937         LIST_HEAD(dispose);
3938
3939         pr_info("mds%d reconnect start\n", mds);
3940
3941         recon_state.pagelist = ceph_pagelist_alloc(GFP_NOFS);
3942         if (!recon_state.pagelist)
3943                 goto fail_nopagelist;
3944
3945         reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3946         if (!reply)
3947                 goto fail_nomsg;
3948
3949         xa_destroy(&session->s_delegated_inos);
3950
3951         mutex_lock(&session->s_mutex);
3952         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3953         session->s_seq = 0;
3954
3955         dout("session %p state %s\n", session,
3956              ceph_session_state_name(session->s_state));
3957
3958         spin_lock(&session->s_gen_ttl_lock);
3959         session->s_cap_gen++;
3960         spin_unlock(&session->s_gen_ttl_lock);
3961
3962         spin_lock(&session->s_cap_lock);
3963         /* don't know if session is readonly */
3964         session->s_readonly = 0;
3965         /*
3966          * notify __ceph_remove_cap() that we are composing cap reconnect.
3967          * If a cap get released before being added to the cap reconnect,
3968          * __ceph_remove_cap() should skip queuing cap release.
3969          */
3970         session->s_cap_reconnect = 1;
3971         /* drop old cap expires; we're about to reestablish that state */
3972         detach_cap_releases(session, &dispose);
3973         spin_unlock(&session->s_cap_lock);
3974         dispose_cap_releases(mdsc, &dispose);
3975
3976         /* trim unused caps to reduce MDS's cache rejoin time */
3977         if (mdsc->fsc->sb->s_root)
3978                 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3979
3980         ceph_con_close(&session->s_con);
3981         ceph_con_open(&session->s_con,
3982                       CEPH_ENTITY_TYPE_MDS, mds,
3983                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3984
3985         /* replay unsafe requests */
3986         replay_unsafe_requests(mdsc, session);
3987
3988         ceph_early_kick_flushing_caps(mdsc, session);
3989
3990         down_read(&mdsc->snap_rwsem);
3991
3992         /* placeholder for nr_caps */
3993         err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
3994         if (err)
3995                 goto fail;
3996
3997         if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
3998                 recon_state.msg_version = 3;
3999                 recon_state.allow_multi = true;
4000         } else if (session->s_con.peer_features & CEPH_FEATURE_MDSENC) {
4001                 recon_state.msg_version = 3;
4002         } else {
4003                 recon_state.msg_version = 2;
4004         }
4005         /* trsaverse this session's caps */
4006         err = ceph_iterate_session_caps(session, reconnect_caps_cb, &recon_state);
4007
4008         spin_lock(&session->s_cap_lock);
4009         session->s_cap_reconnect = 0;
4010         spin_unlock(&session->s_cap_lock);
4011
4012         if (err < 0)
4013                 goto fail;
4014
4015         /* check if all realms can be encoded into current message */
4016         if (mdsc->num_snap_realms) {
4017                 size_t total_len =
4018                         recon_state.pagelist->length +
4019                         mdsc->num_snap_realms *
4020                         sizeof(struct ceph_mds_snaprealm_reconnect);
4021                 if (recon_state.msg_version >= 4) {
4022                         /* number of realms */
4023                         total_len += sizeof(u32);
4024                         /* version, compat_version and struct_len */
4025                         total_len += mdsc->num_snap_realms *
4026                                      (2 * sizeof(u8) + sizeof(u32));
4027                 }
4028                 if (total_len > RECONNECT_MAX_SIZE) {
4029                         if (!recon_state.allow_multi) {
4030                                 err = -ENOSPC;
4031                                 goto fail;
4032                         }
4033                         if (recon_state.nr_caps) {
4034                                 err = send_reconnect_partial(&recon_state);
4035                                 if (err)
4036                                         goto fail;
4037                         }
4038                         recon_state.msg_version = 5;
4039                 }
4040         }
4041
4042         err = encode_snap_realms(mdsc, &recon_state);
4043         if (err < 0)
4044                 goto fail;
4045
4046         if (recon_state.msg_version >= 5) {
4047                 err = ceph_pagelist_encode_8(recon_state.pagelist, 0);
4048                 if (err < 0)
4049                         goto fail;
4050         }
4051
4052         if (recon_state.nr_caps || recon_state.nr_realms) {
4053                 struct page *page =
4054                         list_first_entry(&recon_state.pagelist->head,
4055                                         struct page, lru);
4056                 __le32 *addr = kmap_atomic(page);
4057                 if (recon_state.nr_caps) {
4058                         WARN_ON(recon_state.nr_realms != mdsc->num_snap_realms);
4059                         *addr = cpu_to_le32(recon_state.nr_caps);
4060                 } else if (recon_state.msg_version >= 4) {
4061                         *(addr + 1) = cpu_to_le32(recon_state.nr_realms);
4062                 }
4063                 kunmap_atomic(addr);
4064         }
4065
4066         reply->hdr.version = cpu_to_le16(recon_state.msg_version);
4067         if (recon_state.msg_version >= 4)
4068                 reply->hdr.compat_version = cpu_to_le16(4);
4069
4070         reply->hdr.data_len = cpu_to_le32(recon_state.pagelist->length);
4071         ceph_msg_data_add_pagelist(reply, recon_state.pagelist);
4072
4073         ceph_con_send(&session->s_con, reply);
4074
4075         mutex_unlock(&session->s_mutex);
4076
4077         mutex_lock(&mdsc->mutex);
4078         __wake_requests(mdsc, &session->s_waiting);
4079         mutex_unlock(&mdsc->mutex);
4080
4081         up_read(&mdsc->snap_rwsem);
4082         ceph_pagelist_release(recon_state.pagelist);
4083         return;
4084
4085 fail:
4086         ceph_msg_put(reply);
4087         up_read(&mdsc->snap_rwsem);
4088         mutex_unlock(&session->s_mutex);
4089 fail_nomsg:
4090         ceph_pagelist_release(recon_state.pagelist);
4091 fail_nopagelist:
4092         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
4093         return;
4094 }
4095
4096
4097 /*
4098  * compare old and new mdsmaps, kicking requests
4099  * and closing out old connections as necessary
4100  *
4101  * called under mdsc->mutex.
4102  */
4103 static void check_new_map(struct ceph_mds_client *mdsc,
4104                           struct ceph_mdsmap *newmap,
4105                           struct ceph_mdsmap *oldmap)
4106 {
4107         int i;
4108         int oldstate, newstate;
4109         struct ceph_mds_session *s;
4110
4111         dout("check_new_map new %u old %u\n",
4112              newmap->m_epoch, oldmap->m_epoch);
4113
4114         for (i = 0; i < oldmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4115                 if (!mdsc->sessions[i])
4116                         continue;
4117                 s = mdsc->sessions[i];
4118                 oldstate = ceph_mdsmap_get_state(oldmap, i);
4119                 newstate = ceph_mdsmap_get_state(newmap, i);
4120
4121                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
4122                      i, ceph_mds_state_name(oldstate),
4123                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
4124                      ceph_mds_state_name(newstate),
4125                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
4126                      ceph_session_state_name(s->s_state));
4127
4128                 if (i >= newmap->possible_max_rank) {
4129                         /* force close session for stopped mds */
4130                         ceph_get_mds_session(s);
4131                         __unregister_session(mdsc, s);
4132                         __wake_requests(mdsc, &s->s_waiting);
4133                         mutex_unlock(&mdsc->mutex);
4134
4135                         mutex_lock(&s->s_mutex);
4136                         cleanup_session_requests(mdsc, s);
4137                         remove_session_caps(s);
4138                         mutex_unlock(&s->s_mutex);
4139
4140                         ceph_put_mds_session(s);
4141
4142                         mutex_lock(&mdsc->mutex);
4143                         kick_requests(mdsc, i);
4144                         continue;
4145                 }
4146
4147                 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
4148                            ceph_mdsmap_get_addr(newmap, i),
4149                            sizeof(struct ceph_entity_addr))) {
4150                         /* just close it */
4151                         mutex_unlock(&mdsc->mutex);
4152                         mutex_lock(&s->s_mutex);
4153                         mutex_lock(&mdsc->mutex);
4154                         ceph_con_close(&s->s_con);
4155                         mutex_unlock(&s->s_mutex);
4156                         s->s_state = CEPH_MDS_SESSION_RESTARTING;
4157                 } else if (oldstate == newstate) {
4158                         continue;  /* nothing new with this mds */
4159                 }
4160
4161                 /*
4162                  * send reconnect?
4163                  */
4164                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
4165                     newstate >= CEPH_MDS_STATE_RECONNECT) {
4166                         mutex_unlock(&mdsc->mutex);
4167                         send_mds_reconnect(mdsc, s);
4168                         mutex_lock(&mdsc->mutex);
4169                 }
4170
4171                 /*
4172                  * kick request on any mds that has gone active.
4173                  */
4174                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
4175                     newstate >= CEPH_MDS_STATE_ACTIVE) {
4176                         if (oldstate != CEPH_MDS_STATE_CREATING &&
4177                             oldstate != CEPH_MDS_STATE_STARTING)
4178                                 pr_info("mds%d recovery completed\n", s->s_mds);
4179                         kick_requests(mdsc, i);
4180                         mutex_unlock(&mdsc->mutex);
4181                         mutex_lock(&s->s_mutex);
4182                         mutex_lock(&mdsc->mutex);
4183                         ceph_kick_flushing_caps(mdsc, s);
4184                         mutex_unlock(&s->s_mutex);
4185                         wake_up_session_caps(s, RECONNECT);
4186                 }
4187         }
4188
4189         for (i = 0; i < newmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4190                 s = mdsc->sessions[i];
4191                 if (!s)
4192                         continue;
4193                 if (!ceph_mdsmap_is_laggy(newmap, i))
4194                         continue;
4195                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4196                     s->s_state == CEPH_MDS_SESSION_HUNG ||
4197                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
4198                         dout(" connecting to export targets of laggy mds%d\n",
4199                              i);
4200                         __open_export_target_sessions(mdsc, s);
4201                 }
4202         }
4203 }
4204
4205
4206
4207 /*
4208  * leases
4209  */
4210
4211 /*
4212  * caller must hold session s_mutex, dentry->d_lock
4213  */
4214 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
4215 {
4216         struct ceph_dentry_info *di = ceph_dentry(dentry);
4217
4218         ceph_put_mds_session(di->lease_session);
4219         di->lease_session = NULL;
4220 }
4221
4222 static void handle_lease(struct ceph_mds_client *mdsc,
4223                          struct ceph_mds_session *session,
4224                          struct ceph_msg *msg)
4225 {
4226         struct super_block *sb = mdsc->fsc->sb;
4227         struct inode *inode;
4228         struct dentry *parent, *dentry;
4229         struct ceph_dentry_info *di;
4230         int mds = session->s_mds;
4231         struct ceph_mds_lease *h = msg->front.iov_base;
4232         u32 seq;
4233         struct ceph_vino vino;
4234         struct qstr dname;
4235         int release = 0;
4236
4237         dout("handle_lease from mds%d\n", mds);
4238
4239         /* decode */
4240         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
4241                 goto bad;
4242         vino.ino = le64_to_cpu(h->ino);
4243         vino.snap = CEPH_NOSNAP;
4244         seq = le32_to_cpu(h->seq);
4245         dname.len = get_unaligned_le32(h + 1);
4246         if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
4247                 goto bad;
4248         dname.name = (void *)(h + 1) + sizeof(u32);
4249
4250         /* lookup inode */
4251         inode = ceph_find_inode(sb, vino);
4252         dout("handle_lease %s, ino %llx %p %.*s\n",
4253              ceph_lease_op_name(h->action), vino.ino, inode,
4254              dname.len, dname.name);
4255
4256         mutex_lock(&session->s_mutex);
4257         inc_session_sequence(session);
4258
4259         if (!inode) {
4260                 dout("handle_lease no inode %llx\n", vino.ino);
4261                 goto release;
4262         }
4263
4264         /* dentry */
4265         parent = d_find_alias(inode);
4266         if (!parent) {
4267                 dout("no parent dentry on inode %p\n", inode);
4268                 WARN_ON(1);
4269                 goto release;  /* hrm... */
4270         }
4271         dname.hash = full_name_hash(parent, dname.name, dname.len);
4272         dentry = d_lookup(parent, &dname);
4273         dput(parent);
4274         if (!dentry)
4275                 goto release;
4276
4277         spin_lock(&dentry->d_lock);
4278         di = ceph_dentry(dentry);
4279         switch (h->action) {
4280         case CEPH_MDS_LEASE_REVOKE:
4281                 if (di->lease_session == session) {
4282                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
4283                                 h->seq = cpu_to_le32(di->lease_seq);
4284                         __ceph_mdsc_drop_dentry_lease(dentry);
4285                 }
4286                 release = 1;
4287                 break;
4288
4289         case CEPH_MDS_LEASE_RENEW:
4290                 if (di->lease_session == session &&
4291                     di->lease_gen == session->s_cap_gen &&
4292                     di->lease_renew_from &&
4293                     di->lease_renew_after == 0) {
4294                         unsigned long duration =
4295                                 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
4296
4297                         di->lease_seq = seq;
4298                         di->time = di->lease_renew_from + duration;
4299                         di->lease_renew_after = di->lease_renew_from +
4300                                 (duration >> 1);
4301                         di->lease_renew_from = 0;
4302                 }
4303                 break;
4304         }
4305         spin_unlock(&dentry->d_lock);
4306         dput(dentry);
4307
4308         if (!release)
4309                 goto out;
4310
4311 release:
4312         /* let's just reuse the same message */
4313         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
4314         ceph_msg_get(msg);
4315         ceph_con_send(&session->s_con, msg);
4316
4317 out:
4318         mutex_unlock(&session->s_mutex);
4319         /* avoid calling iput_final() in mds dispatch threads */
4320         ceph_async_iput(inode);
4321         return;
4322
4323 bad:
4324         pr_err("corrupt lease message\n");
4325         ceph_msg_dump(msg);
4326 }
4327
4328 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
4329                               struct dentry *dentry, char action,
4330                               u32 seq)
4331 {
4332         struct ceph_msg *msg;
4333         struct ceph_mds_lease *lease;
4334         struct inode *dir;
4335         int len = sizeof(*lease) + sizeof(u32) + NAME_MAX;
4336
4337         dout("lease_send_msg identry %p %s to mds%d\n",
4338              dentry, ceph_lease_op_name(action), session->s_mds);
4339
4340         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
4341         if (!msg)
4342                 return;
4343         lease = msg->front.iov_base;
4344         lease->action = action;
4345         lease->seq = cpu_to_le32(seq);
4346
4347         spin_lock(&dentry->d_lock);
4348         dir = d_inode(dentry->d_parent);
4349         lease->ino = cpu_to_le64(ceph_ino(dir));
4350         lease->first = lease->last = cpu_to_le64(ceph_snap(dir));
4351
4352         put_unaligned_le32(dentry->d_name.len, lease + 1);
4353         memcpy((void *)(lease + 1) + 4,
4354                dentry->d_name.name, dentry->d_name.len);
4355         spin_unlock(&dentry->d_lock);
4356         /*
4357          * if this is a preemptive lease RELEASE, no need to
4358          * flush request stream, since the actual request will
4359          * soon follow.
4360          */
4361         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
4362
4363         ceph_con_send(&session->s_con, msg);
4364 }
4365
4366 /*
4367  * lock unlock sessions, to wait ongoing session activities
4368  */
4369 static void lock_unlock_sessions(struct ceph_mds_client *mdsc)
4370 {
4371         int i;
4372
4373         mutex_lock(&mdsc->mutex);
4374         for (i = 0; i < mdsc->max_sessions; i++) {
4375                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4376                 if (!s)
4377                         continue;
4378                 mutex_unlock(&mdsc->mutex);
4379                 mutex_lock(&s->s_mutex);
4380                 mutex_unlock(&s->s_mutex);
4381                 ceph_put_mds_session(s);
4382                 mutex_lock(&mdsc->mutex);
4383         }
4384         mutex_unlock(&mdsc->mutex);
4385 }
4386
4387 static void maybe_recover_session(struct ceph_mds_client *mdsc)
4388 {
4389         struct ceph_fs_client *fsc = mdsc->fsc;
4390
4391         if (!ceph_test_mount_opt(fsc, CLEANRECOVER))
4392                 return;
4393
4394         if (READ_ONCE(fsc->mount_state) != CEPH_MOUNT_MOUNTED)
4395                 return;
4396
4397         if (!READ_ONCE(fsc->blocklisted))
4398                 return;
4399
4400         pr_info("auto reconnect after blocklisted\n");
4401         ceph_force_reconnect(fsc->sb);
4402 }
4403
4404 bool check_session_state(struct ceph_mds_session *s)
4405 {
4406         switch (s->s_state) {
4407         case CEPH_MDS_SESSION_OPEN:
4408                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
4409                         s->s_state = CEPH_MDS_SESSION_HUNG;
4410                         pr_info("mds%d hung\n", s->s_mds);
4411                 }
4412                 break;
4413         case CEPH_MDS_SESSION_CLOSING:
4414                 /* Should never reach this when we're unmounting */
4415                 WARN_ON_ONCE(true);
4416                 fallthrough;
4417         case CEPH_MDS_SESSION_NEW:
4418         case CEPH_MDS_SESSION_RESTARTING:
4419         case CEPH_MDS_SESSION_CLOSED:
4420         case CEPH_MDS_SESSION_REJECTED:
4421                 return false;
4422         }
4423
4424         return true;
4425 }
4426
4427 /*
4428  * If the sequence is incremented while we're waiting on a REQUEST_CLOSE reply,
4429  * then we need to retransmit that request.
4430  */
4431 void inc_session_sequence(struct ceph_mds_session *s)
4432 {
4433         lockdep_assert_held(&s->s_mutex);
4434
4435         s->s_seq++;
4436
4437         if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
4438                 int ret;
4439
4440                 dout("resending session close request for mds%d\n", s->s_mds);
4441                 ret = request_close_session(s);
4442                 if (ret < 0)
4443                         pr_err("unable to close session to mds%d: %d\n",
4444                                s->s_mds, ret);
4445         }
4446 }
4447
4448 /*
4449  * delayed work -- periodically trim expired leases, renew caps with mds
4450  */
4451 static void schedule_delayed(struct ceph_mds_client *mdsc)
4452 {
4453         int delay = 5;
4454         unsigned hz = round_jiffies_relative(HZ * delay);
4455         schedule_delayed_work(&mdsc->delayed_work, hz);
4456 }
4457
4458 static void delayed_work(struct work_struct *work)
4459 {
4460         int i;
4461         struct ceph_mds_client *mdsc =
4462                 container_of(work, struct ceph_mds_client, delayed_work.work);
4463         int renew_interval;
4464         int renew_caps;
4465
4466         dout("mdsc delayed_work\n");
4467
4468         if (mdsc->stopping)
4469                 return;
4470
4471         mutex_lock(&mdsc->mutex);
4472         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
4473         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
4474                                    mdsc->last_renew_caps);
4475         if (renew_caps)
4476                 mdsc->last_renew_caps = jiffies;
4477
4478         for (i = 0; i < mdsc->max_sessions; i++) {
4479                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4480                 if (!s)
4481                         continue;
4482
4483                 if (!check_session_state(s)) {
4484                         ceph_put_mds_session(s);
4485                         continue;
4486                 }
4487                 mutex_unlock(&mdsc->mutex);
4488
4489                 mutex_lock(&s->s_mutex);
4490                 if (renew_caps)
4491                         send_renew_caps(mdsc, s);
4492                 else
4493                         ceph_con_keepalive(&s->s_con);
4494                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4495                     s->s_state == CEPH_MDS_SESSION_HUNG)
4496                         ceph_send_cap_releases(mdsc, s);
4497                 mutex_unlock(&s->s_mutex);
4498                 ceph_put_mds_session(s);
4499
4500                 mutex_lock(&mdsc->mutex);
4501         }
4502         mutex_unlock(&mdsc->mutex);
4503
4504         ceph_check_delayed_caps(mdsc);
4505
4506         ceph_queue_cap_reclaim_work(mdsc);
4507
4508         ceph_trim_snapid_map(mdsc);
4509
4510         maybe_recover_session(mdsc);
4511
4512         schedule_delayed(mdsc);
4513 }
4514
4515 int ceph_mdsc_init(struct ceph_fs_client *fsc)
4516
4517 {
4518         struct ceph_mds_client *mdsc;
4519         int err;
4520
4521         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
4522         if (!mdsc)
4523                 return -ENOMEM;
4524         mdsc->fsc = fsc;
4525         mutex_init(&mdsc->mutex);
4526         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
4527         if (!mdsc->mdsmap) {
4528                 err = -ENOMEM;
4529                 goto err_mdsc;
4530         }
4531
4532         init_completion(&mdsc->safe_umount_waiters);
4533         init_waitqueue_head(&mdsc->session_close_wq);
4534         INIT_LIST_HEAD(&mdsc->waiting_for_map);
4535         mdsc->sessions = NULL;
4536         atomic_set(&mdsc->num_sessions, 0);
4537         mdsc->max_sessions = 0;
4538         mdsc->stopping = 0;
4539         atomic64_set(&mdsc->quotarealms_count, 0);
4540         mdsc->quotarealms_inodes = RB_ROOT;
4541         mutex_init(&mdsc->quotarealms_inodes_mutex);
4542         mdsc->last_snap_seq = 0;
4543         init_rwsem(&mdsc->snap_rwsem);
4544         mdsc->snap_realms = RB_ROOT;
4545         INIT_LIST_HEAD(&mdsc->snap_empty);
4546         mdsc->num_snap_realms = 0;
4547         spin_lock_init(&mdsc->snap_empty_lock);
4548         mdsc->last_tid = 0;
4549         mdsc->oldest_tid = 0;
4550         mdsc->request_tree = RB_ROOT;
4551         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
4552         mdsc->last_renew_caps = jiffies;
4553         INIT_LIST_HEAD(&mdsc->cap_delay_list);
4554         INIT_LIST_HEAD(&mdsc->cap_wait_list);
4555         spin_lock_init(&mdsc->cap_delay_lock);
4556         INIT_LIST_HEAD(&mdsc->snap_flush_list);
4557         spin_lock_init(&mdsc->snap_flush_lock);
4558         mdsc->last_cap_flush_tid = 1;
4559         INIT_LIST_HEAD(&mdsc->cap_flush_list);
4560         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
4561         mdsc->num_cap_flushing = 0;
4562         spin_lock_init(&mdsc->cap_dirty_lock);
4563         init_waitqueue_head(&mdsc->cap_flushing_wq);
4564         INIT_WORK(&mdsc->cap_reclaim_work, ceph_cap_reclaim_work);
4565         atomic_set(&mdsc->cap_reclaim_pending, 0);
4566         err = ceph_metric_init(&mdsc->metric);
4567         if (err)
4568                 goto err_mdsmap;
4569
4570         spin_lock_init(&mdsc->dentry_list_lock);
4571         INIT_LIST_HEAD(&mdsc->dentry_leases);
4572         INIT_LIST_HEAD(&mdsc->dentry_dir_leases);
4573
4574         ceph_caps_init(mdsc);
4575         ceph_adjust_caps_max_min(mdsc, fsc->mount_options);
4576
4577         spin_lock_init(&mdsc->snapid_map_lock);
4578         mdsc->snapid_map_tree = RB_ROOT;
4579         INIT_LIST_HEAD(&mdsc->snapid_map_lru);
4580
4581         init_rwsem(&mdsc->pool_perm_rwsem);
4582         mdsc->pool_perm_tree = RB_ROOT;
4583
4584         strscpy(mdsc->nodename, utsname()->nodename,
4585                 sizeof(mdsc->nodename));
4586
4587         fsc->mdsc = mdsc;
4588         return 0;
4589
4590 err_mdsmap:
4591         kfree(mdsc->mdsmap);
4592 err_mdsc:
4593         kfree(mdsc);
4594         return err;
4595 }
4596
4597 /*
4598  * Wait for safe replies on open mds requests.  If we time out, drop
4599  * all requests from the tree to avoid dangling dentry refs.
4600  */
4601 static void wait_requests(struct ceph_mds_client *mdsc)
4602 {
4603         struct ceph_options *opts = mdsc->fsc->client->options;
4604         struct ceph_mds_request *req;
4605
4606         mutex_lock(&mdsc->mutex);
4607         if (__get_oldest_req(mdsc)) {
4608                 mutex_unlock(&mdsc->mutex);
4609
4610                 dout("wait_requests waiting for requests\n");
4611                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
4612                                     ceph_timeout_jiffies(opts->mount_timeout));
4613
4614                 /* tear down remaining requests */
4615                 mutex_lock(&mdsc->mutex);
4616                 while ((req = __get_oldest_req(mdsc))) {
4617                         dout("wait_requests timed out on tid %llu\n",
4618                              req->r_tid);
4619                         list_del_init(&req->r_wait);
4620                         __unregister_request(mdsc, req);
4621                 }
4622         }
4623         mutex_unlock(&mdsc->mutex);
4624         dout("wait_requests done\n");
4625 }
4626
4627 /*
4628  * called before mount is ro, and before dentries are torn down.
4629  * (hmm, does this still race with new lookups?)
4630  */
4631 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
4632 {
4633         dout("pre_umount\n");
4634         mdsc->stopping = 1;
4635
4636         lock_unlock_sessions(mdsc);
4637         ceph_flush_dirty_caps(mdsc);
4638         wait_requests(mdsc);
4639
4640         /*
4641          * wait for reply handlers to drop their request refs and
4642          * their inode/dcache refs
4643          */
4644         ceph_msgr_flush();
4645
4646         ceph_cleanup_quotarealms_inodes(mdsc);
4647 }
4648
4649 /*
4650  * wait for all write mds requests to flush.
4651  */
4652 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
4653 {
4654         struct ceph_mds_request *req = NULL, *nextreq;
4655         struct rb_node *n;
4656
4657         mutex_lock(&mdsc->mutex);
4658         dout("wait_unsafe_requests want %lld\n", want_tid);
4659 restart:
4660         req = __get_oldest_req(mdsc);
4661         while (req && req->r_tid <= want_tid) {
4662                 /* find next request */
4663                 n = rb_next(&req->r_node);
4664                 if (n)
4665                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
4666                 else
4667                         nextreq = NULL;
4668                 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
4669                     (req->r_op & CEPH_MDS_OP_WRITE)) {
4670                         /* write op */
4671                         ceph_mdsc_get_request(req);
4672                         if (nextreq)
4673                                 ceph_mdsc_get_request(nextreq);
4674                         mutex_unlock(&mdsc->mutex);
4675                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
4676                              req->r_tid, want_tid);
4677                         wait_for_completion(&req->r_safe_completion);
4678                         mutex_lock(&mdsc->mutex);
4679                         ceph_mdsc_put_request(req);
4680                         if (!nextreq)
4681                                 break;  /* next dne before, so we're done! */
4682                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
4683                                 /* next request was removed from tree */
4684                                 ceph_mdsc_put_request(nextreq);
4685                                 goto restart;
4686                         }
4687                         ceph_mdsc_put_request(nextreq);  /* won't go away */
4688                 }
4689                 req = nextreq;
4690         }
4691         mutex_unlock(&mdsc->mutex);
4692         dout("wait_unsafe_requests done\n");
4693 }
4694
4695 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
4696 {
4697         u64 want_tid, want_flush;
4698
4699         if (READ_ONCE(mdsc->fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN)
4700                 return;
4701
4702         dout("sync\n");
4703         mutex_lock(&mdsc->mutex);
4704         want_tid = mdsc->last_tid;
4705         mutex_unlock(&mdsc->mutex);
4706
4707         ceph_flush_dirty_caps(mdsc);
4708         spin_lock(&mdsc->cap_dirty_lock);
4709         want_flush = mdsc->last_cap_flush_tid;
4710         if (!list_empty(&mdsc->cap_flush_list)) {
4711                 struct ceph_cap_flush *cf =
4712                         list_last_entry(&mdsc->cap_flush_list,
4713                                         struct ceph_cap_flush, g_list);
4714                 cf->wake = true;
4715         }
4716         spin_unlock(&mdsc->cap_dirty_lock);
4717
4718         dout("sync want tid %lld flush_seq %lld\n",
4719              want_tid, want_flush);
4720
4721         wait_unsafe_requests(mdsc, want_tid);
4722         wait_caps_flush(mdsc, want_flush);
4723 }
4724
4725 /*
4726  * true if all sessions are closed, or we force unmount
4727  */
4728 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
4729 {
4730         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4731                 return true;
4732         return atomic_read(&mdsc->num_sessions) <= skipped;
4733 }
4734
4735 /*
4736  * called after sb is ro.
4737  */
4738 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
4739 {
4740         struct ceph_options *opts = mdsc->fsc->client->options;
4741         struct ceph_mds_session *session;
4742         int i;
4743         int skipped = 0;
4744
4745         dout("close_sessions\n");
4746
4747         /* close sessions */
4748         mutex_lock(&mdsc->mutex);
4749         for (i = 0; i < mdsc->max_sessions; i++) {
4750                 session = __ceph_lookup_mds_session(mdsc, i);
4751                 if (!session)
4752                         continue;
4753                 mutex_unlock(&mdsc->mutex);
4754                 mutex_lock(&session->s_mutex);
4755                 if (__close_session(mdsc, session) <= 0)
4756                         skipped++;
4757                 mutex_unlock(&session->s_mutex);
4758                 ceph_put_mds_session(session);
4759                 mutex_lock(&mdsc->mutex);
4760         }
4761         mutex_unlock(&mdsc->mutex);
4762
4763         dout("waiting for sessions to close\n");
4764         wait_event_timeout(mdsc->session_close_wq,
4765                            done_closing_sessions(mdsc, skipped),
4766                            ceph_timeout_jiffies(opts->mount_timeout));
4767
4768         /* tear down remaining sessions */
4769         mutex_lock(&mdsc->mutex);
4770         for (i = 0; i < mdsc->max_sessions; i++) {
4771                 if (mdsc->sessions[i]) {
4772                         session = ceph_get_mds_session(mdsc->sessions[i]);
4773                         __unregister_session(mdsc, session);
4774                         mutex_unlock(&mdsc->mutex);
4775                         mutex_lock(&session->s_mutex);
4776                         remove_session_caps(session);
4777                         mutex_unlock(&session->s_mutex);
4778                         ceph_put_mds_session(session);
4779                         mutex_lock(&mdsc->mutex);
4780                 }
4781         }
4782         WARN_ON(!list_empty(&mdsc->cap_delay_list));
4783         mutex_unlock(&mdsc->mutex);
4784
4785         ceph_cleanup_snapid_map(mdsc);
4786         ceph_cleanup_empty_realms(mdsc);
4787
4788         cancel_work_sync(&mdsc->cap_reclaim_work);
4789         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4790
4791         dout("stopped\n");
4792 }
4793
4794 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
4795 {
4796         struct ceph_mds_session *session;
4797         int mds;
4798
4799         dout("force umount\n");
4800
4801         mutex_lock(&mdsc->mutex);
4802         for (mds = 0; mds < mdsc->max_sessions; mds++) {
4803                 session = __ceph_lookup_mds_session(mdsc, mds);
4804                 if (!session)
4805                         continue;
4806
4807                 if (session->s_state == CEPH_MDS_SESSION_REJECTED)
4808                         __unregister_session(mdsc, session);
4809                 __wake_requests(mdsc, &session->s_waiting);
4810                 mutex_unlock(&mdsc->mutex);
4811
4812                 mutex_lock(&session->s_mutex);
4813                 __close_session(mdsc, session);
4814                 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
4815                         cleanup_session_requests(mdsc, session);
4816                         remove_session_caps(session);
4817                 }
4818                 mutex_unlock(&session->s_mutex);
4819                 ceph_put_mds_session(session);
4820
4821                 mutex_lock(&mdsc->mutex);
4822                 kick_requests(mdsc, mds);
4823         }
4824         __wake_requests(mdsc, &mdsc->waiting_for_map);
4825         mutex_unlock(&mdsc->mutex);
4826 }
4827
4828 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
4829 {
4830         dout("stop\n");
4831         /*
4832          * Make sure the delayed work stopped before releasing
4833          * the resources.
4834          *
4835          * Because the cancel_delayed_work_sync() will only
4836          * guarantee that the work finishes executing. But the
4837          * delayed work will re-arm itself again after that.
4838          */
4839         flush_delayed_work(&mdsc->delayed_work);
4840
4841         if (mdsc->mdsmap)
4842                 ceph_mdsmap_destroy(mdsc->mdsmap);
4843         kfree(mdsc->sessions);
4844         ceph_caps_finalize(mdsc);
4845         ceph_pool_perm_destroy(mdsc);
4846 }
4847
4848 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
4849 {
4850         struct ceph_mds_client *mdsc = fsc->mdsc;
4851         dout("mdsc_destroy %p\n", mdsc);
4852
4853         if (!mdsc)
4854                 return;
4855
4856         /* flush out any connection work with references to us */
4857         ceph_msgr_flush();
4858
4859         ceph_mdsc_stop(mdsc);
4860
4861         ceph_metric_destroy(&mdsc->metric);
4862
4863         flush_delayed_work(&mdsc->metric.delayed_work);
4864         fsc->mdsc = NULL;
4865         kfree(mdsc);
4866         dout("mdsc_destroy %p done\n", mdsc);
4867 }
4868
4869 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4870 {
4871         struct ceph_fs_client *fsc = mdsc->fsc;
4872         const char *mds_namespace = fsc->mount_options->mds_namespace;
4873         void *p = msg->front.iov_base;
4874         void *end = p + msg->front.iov_len;
4875         u32 epoch;
4876         u32 num_fs;
4877         u32 mount_fscid = (u32)-1;
4878         int err = -EINVAL;
4879
4880         ceph_decode_need(&p, end, sizeof(u32), bad);
4881         epoch = ceph_decode_32(&p);
4882
4883         dout("handle_fsmap epoch %u\n", epoch);
4884
4885         /* struct_v, struct_cv, map_len, epoch, legacy_client_fscid */
4886         ceph_decode_skip_n(&p, end, 2 + sizeof(u32) * 3, bad);
4887
4888         ceph_decode_32_safe(&p, end, num_fs, bad);
4889         while (num_fs-- > 0) {
4890                 void *info_p, *info_end;
4891                 u32 info_len;
4892                 u32 fscid, namelen;
4893
4894                 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4895                 p += 2;         // info_v, info_cv
4896                 info_len = ceph_decode_32(&p);
4897                 ceph_decode_need(&p, end, info_len, bad);
4898                 info_p = p;
4899                 info_end = p + info_len;
4900                 p = info_end;
4901
4902                 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
4903                 fscid = ceph_decode_32(&info_p);
4904                 namelen = ceph_decode_32(&info_p);
4905                 ceph_decode_need(&info_p, info_end, namelen, bad);
4906
4907                 if (mds_namespace &&
4908                     strlen(mds_namespace) == namelen &&
4909                     !strncmp(mds_namespace, (char *)info_p, namelen)) {
4910                         mount_fscid = fscid;
4911                         break;
4912                 }
4913         }
4914
4915         ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
4916         if (mount_fscid != (u32)-1) {
4917                 fsc->client->monc.fs_cluster_id = mount_fscid;
4918                 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
4919                                    0, true);
4920                 ceph_monc_renew_subs(&fsc->client->monc);
4921         } else {
4922                 err = -ENOENT;
4923                 goto err_out;
4924         }
4925         return;
4926
4927 bad:
4928         pr_err("error decoding fsmap\n");
4929 err_out:
4930         mutex_lock(&mdsc->mutex);
4931         mdsc->mdsmap_err = err;
4932         __wake_requests(mdsc, &mdsc->waiting_for_map);
4933         mutex_unlock(&mdsc->mutex);
4934 }
4935
4936 /*
4937  * handle mds map update.
4938  */
4939 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4940 {
4941         u32 epoch;
4942         u32 maplen;
4943         void *p = msg->front.iov_base;
4944         void *end = p + msg->front.iov_len;
4945         struct ceph_mdsmap *newmap, *oldmap;
4946         struct ceph_fsid fsid;
4947         int err = -EINVAL;
4948
4949         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
4950         ceph_decode_copy(&p, &fsid, sizeof(fsid));
4951         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
4952                 return;
4953         epoch = ceph_decode_32(&p);
4954         maplen = ceph_decode_32(&p);
4955         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
4956
4957         /* do we need it? */
4958         mutex_lock(&mdsc->mutex);
4959         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
4960                 dout("handle_map epoch %u <= our %u\n",
4961                      epoch, mdsc->mdsmap->m_epoch);
4962                 mutex_unlock(&mdsc->mutex);
4963                 return;
4964         }
4965
4966         newmap = ceph_mdsmap_decode(&p, end);
4967         if (IS_ERR(newmap)) {
4968                 err = PTR_ERR(newmap);
4969                 goto bad_unlock;
4970         }
4971
4972         /* swap into place */
4973         if (mdsc->mdsmap) {
4974                 oldmap = mdsc->mdsmap;
4975                 mdsc->mdsmap = newmap;
4976                 check_new_map(mdsc, newmap, oldmap);
4977                 ceph_mdsmap_destroy(oldmap);
4978         } else {
4979                 mdsc->mdsmap = newmap;  /* first mds map */
4980         }
4981         mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
4982                                         MAX_LFS_FILESIZE);
4983
4984         __wake_requests(mdsc, &mdsc->waiting_for_map);
4985         ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
4986                           mdsc->mdsmap->m_epoch);
4987
4988         mutex_unlock(&mdsc->mutex);
4989         schedule_delayed(mdsc);
4990         return;
4991
4992 bad_unlock:
4993         mutex_unlock(&mdsc->mutex);
4994 bad:
4995         pr_err("error decoding mdsmap %d\n", err);
4996         return;
4997 }
4998
4999 static struct ceph_connection *con_get(struct ceph_connection *con)
5000 {
5001         struct ceph_mds_session *s = con->private;
5002
5003         if (ceph_get_mds_session(s))
5004                 return con;
5005         return NULL;
5006 }
5007
5008 static void con_put(struct ceph_connection *con)
5009 {
5010         struct ceph_mds_session *s = con->private;
5011
5012         ceph_put_mds_session(s);
5013 }
5014
5015 /*
5016  * if the client is unresponsive for long enough, the mds will kill
5017  * the session entirely.
5018  */
5019 static void peer_reset(struct ceph_connection *con)
5020 {
5021         struct ceph_mds_session *s = con->private;
5022         struct ceph_mds_client *mdsc = s->s_mdsc;
5023
5024         pr_warn("mds%d closed our session\n", s->s_mds);
5025         send_mds_reconnect(mdsc, s);
5026 }
5027
5028 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
5029 {
5030         struct ceph_mds_session *s = con->private;
5031         struct ceph_mds_client *mdsc = s->s_mdsc;
5032         int type = le16_to_cpu(msg->hdr.type);
5033
5034         mutex_lock(&mdsc->mutex);
5035         if (__verify_registered_session(mdsc, s) < 0) {
5036                 mutex_unlock(&mdsc->mutex);
5037                 goto out;
5038         }
5039         mutex_unlock(&mdsc->mutex);
5040
5041         switch (type) {
5042         case CEPH_MSG_MDS_MAP:
5043                 ceph_mdsc_handle_mdsmap(mdsc, msg);
5044                 break;
5045         case CEPH_MSG_FS_MAP_USER:
5046                 ceph_mdsc_handle_fsmap(mdsc, msg);
5047                 break;
5048         case CEPH_MSG_CLIENT_SESSION:
5049                 handle_session(s, msg);
5050                 break;
5051         case CEPH_MSG_CLIENT_REPLY:
5052                 handle_reply(s, msg);
5053                 break;
5054         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
5055                 handle_forward(mdsc, s, msg);
5056                 break;
5057         case CEPH_MSG_CLIENT_CAPS:
5058                 ceph_handle_caps(s, msg);
5059                 break;
5060         case CEPH_MSG_CLIENT_SNAP:
5061                 ceph_handle_snap(mdsc, s, msg);
5062                 break;
5063         case CEPH_MSG_CLIENT_LEASE:
5064                 handle_lease(mdsc, s, msg);
5065                 break;
5066         case CEPH_MSG_CLIENT_QUOTA:
5067                 ceph_handle_quota(mdsc, s, msg);
5068                 break;
5069
5070         default:
5071                 pr_err("received unknown message type %d %s\n", type,
5072                        ceph_msg_type_name(type));
5073         }
5074 out:
5075         ceph_msg_put(msg);
5076 }
5077
5078 /*
5079  * authentication
5080  */
5081
5082 /*
5083  * Note: returned pointer is the address of a structure that's
5084  * managed separately.  Caller must *not* attempt to free it.
5085  */
5086 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
5087                                         int *proto, int force_new)
5088 {
5089         struct ceph_mds_session *s = con->private;
5090         struct ceph_mds_client *mdsc = s->s_mdsc;
5091         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5092         struct ceph_auth_handshake *auth = &s->s_auth;
5093
5094         if (force_new && auth->authorizer) {
5095                 ceph_auth_destroy_authorizer(auth->authorizer);
5096                 auth->authorizer = NULL;
5097         }
5098         if (!auth->authorizer) {
5099                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
5100                                                       auth);
5101                 if (ret)
5102                         return ERR_PTR(ret);
5103         } else {
5104                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
5105                                                       auth);
5106                 if (ret)
5107                         return ERR_PTR(ret);
5108         }
5109         *proto = ac->protocol;
5110
5111         return auth;
5112 }
5113
5114 static int add_authorizer_challenge(struct ceph_connection *con,
5115                                     void *challenge_buf, int challenge_buf_len)
5116 {
5117         struct ceph_mds_session *s = con->private;
5118         struct ceph_mds_client *mdsc = s->s_mdsc;
5119         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5120
5121         return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
5122                                             challenge_buf, challenge_buf_len);
5123 }
5124
5125 static int verify_authorizer_reply(struct ceph_connection *con)
5126 {
5127         struct ceph_mds_session *s = con->private;
5128         struct ceph_mds_client *mdsc = s->s_mdsc;
5129         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5130
5131         return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
5132 }
5133
5134 static int invalidate_authorizer(struct ceph_connection *con)
5135 {
5136         struct ceph_mds_session *s = con->private;
5137         struct ceph_mds_client *mdsc = s->s_mdsc;
5138         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5139
5140         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
5141
5142         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
5143 }
5144
5145 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
5146                                 struct ceph_msg_header *hdr, int *skip)
5147 {
5148         struct ceph_msg *msg;
5149         int type = (int) le16_to_cpu(hdr->type);
5150         int front_len = (int) le32_to_cpu(hdr->front_len);
5151
5152         if (con->in_msg)
5153                 return con->in_msg;
5154
5155         *skip = 0;
5156         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
5157         if (!msg) {
5158                 pr_err("unable to allocate msg type %d len %d\n",
5159                        type, front_len);
5160                 return NULL;
5161         }
5162
5163         return msg;
5164 }
5165
5166 static int mds_sign_message(struct ceph_msg *msg)
5167 {
5168        struct ceph_mds_session *s = msg->con->private;
5169        struct ceph_auth_handshake *auth = &s->s_auth;
5170
5171        return ceph_auth_sign_message(auth, msg);
5172 }
5173
5174 static int mds_check_message_signature(struct ceph_msg *msg)
5175 {
5176        struct ceph_mds_session *s = msg->con->private;
5177        struct ceph_auth_handshake *auth = &s->s_auth;
5178
5179        return ceph_auth_check_message_signature(auth, msg);
5180 }
5181
5182 static const struct ceph_connection_operations mds_con_ops = {
5183         .get = con_get,
5184         .put = con_put,
5185         .dispatch = dispatch,
5186         .get_authorizer = get_authorizer,
5187         .add_authorizer_challenge = add_authorizer_challenge,
5188         .verify_authorizer_reply = verify_authorizer_reply,
5189         .invalidate_authorizer = invalidate_authorizer,
5190         .peer_reset = peer_reset,
5191         .alloc_msg = mds_alloc_msg,
5192         .sign_message = mds_sign_message,
5193         .check_message_signature = mds_check_message_signature,
5194 };
5195
5196 /* eof */