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  49  * are no longer valid.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         put_cred(req->r_cred);
 837         if (req->r_pagelist)
 838                 ceph_pagelist_release(req->r_pagelist);
 839         put_request_session(req);
 840         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
 841         WARN_ON_ONCE(!list_empty(&req->r_wait));
 842         kmem_cache_free(ceph_mds_request_cachep, req);
 843 }
 844
 845 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
 846
 847 /*
 848  * lookup session, bump ref if found.
 849  *
 850  * called under mdsc->mutex.
 851  */
 852 static struct ceph_mds_request *
 853 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
 854 {
 855         struct ceph_mds_request *req;
 856
 857         req = lookup_request(&mdsc->request_tree, tid);
 858         if (req)
 859                 ceph_mdsc_get_request(req);
 860
 861         return req;
 862 }
 863
 864 /*
 865  * Register an in-flight request, and assign a tid.  Link to directory
 866  * are modifying (if any).
 867  *
 868  * Called under mdsc->mutex.
 869  */
 870 static void __register_request(struct ceph_mds_client *mdsc,
 871                                struct ceph_mds_request *req,
 872                                struct inode *dir)
 873 {
 874         int ret = 0;
 875
 876         req->r_tid = ++mdsc->last_tid;
 877         if (req->r_num_caps) {
 878                 ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
 879                                         req->r_num_caps);
 880                 if (ret < 0) {
 881                         pr_err("__register_request %p "
 882                                "failed to reserve caps: %d\n", req, ret);
 883                         /* set req->r_err to fail early from __do_request */
 884                         req->r_err = ret;
 885                         return;
 886                 }
 887         }
 888         dout("__register_request %p tid %lld\n", req, req->r_tid);
 889         ceph_mdsc_get_request(req);
 890         insert_request(&mdsc->request_tree, req);
 891
 892         req->r_cred = get_current_cred();
 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(3);
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,
2546                                                  req->r_cred->fsuid));
2547         head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns,
2548                                                  req->r_cred->fsgid));
2549         head->ino = cpu_to_le64(req->r_deleg_ino);
2550         head->args = req->r_args;
2551
2552         ceph_encode_filepath(&p, end, ino1, path1);
2553         ceph_encode_filepath(&p, end, ino2, path2);
2554
2555         /* make note of release offset, in case we need to replay */
2556         req->r_request_release_offset = p - msg->front.iov_base;
2557
2558         /* cap releases */
2559         releases = 0;
2560         if (req->r_inode_drop)
2561                 releases += ceph_encode_inode_release(&p,
2562                       req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2563                       mds, req->r_inode_drop, req->r_inode_unless,
2564                       req->r_op == CEPH_MDS_OP_READDIR);
2565         if (req->r_dentry_drop)
2566                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
2567                                 req->r_parent, mds, req->r_dentry_drop,
2568                                 req->r_dentry_unless);
2569         if (req->r_old_dentry_drop)
2570                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2571                                 req->r_old_dentry_dir, mds,
2572                                 req->r_old_dentry_drop,
2573                                 req->r_old_dentry_unless);
2574         if (req->r_old_inode_drop)
2575                 releases += ceph_encode_inode_release(&p,
2576                       d_inode(req->r_old_dentry),
2577                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2578
2579         if (drop_cap_releases) {
2580                 releases = 0;
2581                 p = msg->front.iov_base + req->r_request_release_offset;
2582         }
2583
2584         head->num_releases = cpu_to_le16(releases);
2585
2586         /* time stamp */
2587         {
2588                 struct ceph_timespec ts;
2589                 ceph_encode_timespec64(&ts, &req->r_stamp);
2590                 ceph_encode_copy(&p, &ts, sizeof(ts));
2591         }
2592
2593         if (WARN_ON_ONCE(p > end)) {
2594                 ceph_msg_put(msg);
2595                 msg = ERR_PTR(-ERANGE);
2596                 goto out_free2;
2597         }
2598
2599         msg->front.iov_len = p - msg->front.iov_base;
2600         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2601
2602         if (req->r_pagelist) {
2603                 struct ceph_pagelist *pagelist = req->r_pagelist;
2604                 ceph_msg_data_add_pagelist(msg, pagelist);
2605                 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2606         } else {
2607                 msg->hdr.data_len = 0;
2608         }
2609
2610         msg->hdr.data_off = cpu_to_le16(0);
2611
2612 out_free2:
2613         if (freepath2)
2614                 ceph_mdsc_free_path((char *)path2, pathlen2);
2615 out_free1:
2616         if (freepath1)
2617                 ceph_mdsc_free_path((char *)path1, pathlen1);
2618 out:
2619         return msg;
2620 }
2621
2622 /*
2623  * called under mdsc->mutex if error, under no mutex if
2624  * success.
2625  */
2626 static void complete_request(struct ceph_mds_client *mdsc,
2627                              struct ceph_mds_request *req)
2628 {
2629         req->r_end_latency = ktime_get();
2630
2631         if (req->r_callback)
2632                 req->r_callback(mdsc, req);
2633         complete_all(&req->r_completion);
2634 }
2635
2636 /*
2637  * called under mdsc->mutex
2638  */
2639 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2640                                   struct ceph_mds_request *req,
2641                                   int mds, bool drop_cap_releases)
2642 {
2643         struct ceph_mds_request_head *rhead;
2644         struct ceph_msg *msg;
2645         int flags = 0;
2646
2647         req->r_attempts++;
2648         if (req->r_inode) {
2649                 struct ceph_cap *cap =
2650                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2651
2652                 if (cap)
2653                         req->r_sent_on_mseq = cap->mseq;
2654                 else
2655                         req->r_sent_on_mseq = -1;
2656         }
2657         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2658              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2659
2660         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2661                 void *p;
2662                 /*
2663                  * Replay.  Do not regenerate message (and rebuild
2664                  * paths, etc.); just use the original message.
2665                  * Rebuilding paths will break for renames because
2666                  * d_move mangles the src name.
2667                  */
2668                 msg = req->r_request;
2669                 rhead = msg->front.iov_base;
2670
2671                 flags = le32_to_cpu(rhead->flags);
2672                 flags |= CEPH_MDS_FLAG_REPLAY;
2673                 rhead->flags = cpu_to_le32(flags);
2674
2675                 if (req->r_target_inode)
2676                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2677
2678                 rhead->num_retry = req->r_attempts - 1;
2679
2680                 /* remove cap/dentry releases from message */
2681                 rhead->num_releases = 0;
2682
2683                 /* time stamp */
2684                 p = msg->front.iov_base + req->r_request_release_offset;
2685                 {
2686                         struct ceph_timespec ts;
2687                         ceph_encode_timespec64(&ts, &req->r_stamp);
2688                         ceph_encode_copy(&p, &ts, sizeof(ts));
2689                 }
2690
2691                 msg->front.iov_len = p - msg->front.iov_base;
2692                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2693                 return 0;
2694         }
2695
2696         if (req->r_request) {
2697                 ceph_msg_put(req->r_request);
2698                 req->r_request = NULL;
2699         }
2700         msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2701         if (IS_ERR(msg)) {
2702                 req->r_err = PTR_ERR(msg);
2703                 return PTR_ERR(msg);
2704         }
2705         req->r_request = msg;
2706
2707         rhead = msg->front.iov_base;
2708         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2709         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2710                 flags |= CEPH_MDS_FLAG_REPLAY;
2711         if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags))
2712                 flags |= CEPH_MDS_FLAG_ASYNC;
2713         if (req->r_parent)
2714                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2715         rhead->flags = cpu_to_le32(flags);
2716         rhead->num_fwd = req->r_num_fwd;
2717         rhead->num_retry = req->r_attempts - 1;
2718
2719         dout(" r_parent = %p\n", req->r_parent);
2720         return 0;
2721 }
2722
2723 /*
2724  * called under mdsc->mutex
2725  */
2726 static int __send_request(struct ceph_mds_client *mdsc,
2727                           struct ceph_mds_session *session,
2728                           struct ceph_mds_request *req,
2729                           bool drop_cap_releases)
2730 {
2731         int err;
2732
2733         err = __prepare_send_request(mdsc, req, session->s_mds,
2734                                      drop_cap_releases);
2735         if (!err) {
2736                 ceph_msg_get(req->r_request);
2737                 ceph_con_send(&session->s_con, req->r_request);
2738         }
2739
2740         return err;
2741 }
2742
2743 /*
2744  * send request, or put it on the appropriate wait list.
2745  */
2746 static void __do_request(struct ceph_mds_client *mdsc,
2747                         struct ceph_mds_request *req)
2748 {
2749         struct ceph_mds_session *session = NULL;
2750         int mds = -1;
2751         int err = 0;
2752         bool random;
2753
2754         if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2755                 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2756                         __unregister_request(mdsc, req);
2757                 return;
2758         }
2759
2760         if (req->r_timeout &&
2761             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2762                 dout("do_request timed out\n");
2763                 err = -ETIMEDOUT;
2764                 goto finish;
2765         }
2766         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2767                 dout("do_request forced umount\n");
2768                 err = -EIO;
2769                 goto finish;
2770         }
2771         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2772                 if (mdsc->mdsmap_err) {
2773                         err = mdsc->mdsmap_err;
2774                         dout("do_request mdsmap err %d\n", err);
2775                         goto finish;
2776                 }
2777                 if (mdsc->mdsmap->m_epoch == 0) {
2778                         dout("do_request no mdsmap, waiting for map\n");
2779                         list_add(&req->r_wait, &mdsc->waiting_for_map);
2780                         return;
2781                 }
2782                 if (!(mdsc->fsc->mount_options->flags &
2783                       CEPH_MOUNT_OPT_MOUNTWAIT) &&
2784                     !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2785                         err = -EHOSTUNREACH;
2786                         goto finish;
2787                 }
2788         }
2789
2790         put_request_session(req);
2791
2792         mds = __choose_mds(mdsc, req, &random);
2793         if (mds < 0 ||
2794             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2795                 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2796                         err = -EJUKEBOX;
2797                         goto finish;
2798                 }
2799                 dout("do_request no mds or not active, waiting for map\n");
2800                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2801                 return;
2802         }
2803
2804         /* get, open session */
2805         session = __ceph_lookup_mds_session(mdsc, mds);
2806         if (!session) {
2807                 session = register_session(mdsc, mds);
2808                 if (IS_ERR(session)) {
2809                         err = PTR_ERR(session);
2810                         goto finish;
2811                 }
2812         }
2813         req->r_session = ceph_get_mds_session(session);
2814
2815         dout("do_request mds%d session %p state %s\n", mds, session,
2816              ceph_session_state_name(session->s_state));
2817         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2818             session->s_state != CEPH_MDS_SESSION_HUNG) {
2819                 /*
2820                  * We cannot queue async requests since the caps and delegated
2821                  * inodes are bound to the session. Just return -EJUKEBOX and
2822                  * let the caller retry a sync request in that case.
2823                  */
2824                 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2825                         err = -EJUKEBOX;
2826                         goto out_session;
2827                 }
2828
2829                 /*
2830                  * If the session has been REJECTED, then return a hard error,
2831                  * unless it's a CLEANRECOVER mount, in which case we'll queue
2832                  * it to the mdsc queue.
2833                  */
2834                 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2835                         if (ceph_test_mount_opt(mdsc->fsc, CLEANRECOVER))
2836                                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2837                         else
2838                                 err = -EACCES;
2839                         goto out_session;
2840                 }
2841
2842                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2843                     session->s_state == CEPH_MDS_SESSION_CLOSING) {
2844                         err = __open_session(mdsc, session);
2845                         if (err)
2846                                 goto out_session;
2847                         /* retry the same mds later */
2848                         if (random)
2849                                 req->r_resend_mds = mds;
2850                 }
2851                 list_add(&req->r_wait, &session->s_waiting);
2852                 goto out_session;
2853         }
2854
2855         /* send request */
2856         req->r_resend_mds = -1;   /* forget any previous mds hint */
2857
2858         if (req->r_request_started == 0)   /* note request start time */
2859                 req->r_request_started = jiffies;
2860
2861         err = __send_request(mdsc, session, req, false);
2862
2863 out_session:
2864         ceph_put_mds_session(session);
2865 finish:
2866         if (err) {
2867                 dout("__do_request early error %d\n", err);
2868                 req->r_err = err;
2869                 complete_request(mdsc, req);
2870                 __unregister_request(mdsc, req);
2871         }
2872         return;
2873 }
2874
2875 /*
2876  * called under mdsc->mutex
2877  */
2878 static void __wake_requests(struct ceph_mds_client *mdsc,
2879                             struct list_head *head)
2880 {
2881         struct ceph_mds_request *req;
2882         LIST_HEAD(tmp_list);
2883
2884         list_splice_init(head, &tmp_list);
2885
2886         while (!list_empty(&tmp_list)) {
2887                 req = list_entry(tmp_list.next,
2888                                  struct ceph_mds_request, r_wait);
2889                 list_del_init(&req->r_wait);
2890                 dout(" wake request %p tid %llu\n", req, req->r_tid);
2891                 __do_request(mdsc, req);
2892         }
2893 }
2894
2895 /*
2896  * Wake up threads with requests pending for @mds, so that they can
2897  * resubmit their requests to a possibly different mds.
2898  */
2899 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2900 {
2901         struct ceph_mds_request *req;
2902         struct rb_node *p = rb_first(&mdsc->request_tree);
2903
2904         dout("kick_requests mds%d\n", mds);
2905         while (p) {
2906                 req = rb_entry(p, struct ceph_mds_request, r_node);
2907                 p = rb_next(p);
2908                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2909                         continue;
2910                 if (req->r_attempts > 0)
2911                         continue; /* only new requests */
2912                 if (req->r_session &&
2913                     req->r_session->s_mds == mds) {
2914                         dout(" kicking tid %llu\n", req->r_tid);
2915                         list_del_init(&req->r_wait);
2916                         __do_request(mdsc, req);
2917                 }
2918         }
2919 }
2920
2921 int ceph_mdsc_submit_request(struct ceph_mds_client *mdsc, struct inode *dir,
2922                               struct ceph_mds_request *req)
2923 {
2924         int err = 0;
2925
2926         /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2927         if (req->r_inode)
2928                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2929         if (req->r_parent) {
2930                 struct ceph_inode_info *ci = ceph_inode(req->r_parent);
2931                 int fmode = (req->r_op & CEPH_MDS_OP_WRITE) ?
2932                             CEPH_FILE_MODE_WR : CEPH_FILE_MODE_RD;
2933                 spin_lock(&ci->i_ceph_lock);
2934                 ceph_take_cap_refs(ci, CEPH_CAP_PIN, false);
2935                 __ceph_touch_fmode(ci, mdsc, fmode);
2936                 spin_unlock(&ci->i_ceph_lock);
2937                 ihold(req->r_parent);
2938         }
2939         if (req->r_old_dentry_dir)
2940                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2941                                   CEPH_CAP_PIN);
2942
2943         if (req->r_inode) {
2944                 err = ceph_wait_on_async_create(req->r_inode);
2945                 if (err) {
2946                         dout("%s: wait for async create returned: %d\n",
2947                              __func__, err);
2948                         return err;
2949                 }
2950         }
2951
2952         if (!err && req->r_old_inode) {
2953                 err = ceph_wait_on_async_create(req->r_old_inode);
2954                 if (err) {
2955                         dout("%s: wait for async create returned: %d\n",
2956                              __func__, err);
2957                         return err;
2958                 }
2959         }
2960
2961         dout("submit_request on %p for inode %p\n", req, dir);
2962         mutex_lock(&mdsc->mutex);
2963         __register_request(mdsc, req, dir);
2964         __do_request(mdsc, req);
2965         err = req->r_err;
2966         mutex_unlock(&mdsc->mutex);
2967         return err;
2968 }
2969
2970 static int ceph_mdsc_wait_request(struct ceph_mds_client *mdsc,
2971                                   struct ceph_mds_request *req)
2972 {
2973         int err;
2974
2975         /* wait */
2976         dout("do_request waiting\n");
2977         if (!req->r_timeout && req->r_wait_for_completion) {
2978                 err = req->r_wait_for_completion(mdsc, req);
2979         } else {
2980                 long timeleft = wait_for_completion_killable_timeout(
2981                                         &req->r_completion,
2982                                         ceph_timeout_jiffies(req->r_timeout));
2983                 if (timeleft > 0)
2984                         err = 0;
2985                 else if (!timeleft)
2986                         err = -ETIMEDOUT;  /* timed out */
2987                 else
2988                         err = timeleft;  /* killed */
2989         }
2990         dout("do_request waited, got %d\n", err);
2991         mutex_lock(&mdsc->mutex);
2992
2993         /* only abort if we didn't race with a real reply */
2994         if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2995                 err = le32_to_cpu(req->r_reply_info.head->result);
2996         } else if (err < 0) {
2997                 dout("aborted request %lld with %d\n", req->r_tid, err);
2998
2999                 /*
3000                  * ensure we aren't running concurrently with
3001                  * ceph_fill_trace or ceph_readdir_prepopulate, which
3002                  * rely on locks (dir mutex) held by our caller.
3003                  */
3004                 mutex_lock(&req->r_fill_mutex);
3005                 req->r_err = err;
3006                 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
3007                 mutex_unlock(&req->r_fill_mutex);
3008
3009                 if (req->r_parent &&
3010                     (req->r_op & CEPH_MDS_OP_WRITE))
3011                         ceph_invalidate_dir_request(req);
3012         } else {
3013                 err = req->r_err;
3014         }
3015
3016         mutex_unlock(&mdsc->mutex);
3017         return err;
3018 }
3019
3020 /*
3021  * Synchrously perform an mds request.  Take care of all of the
3022  * session setup, forwarding, retry details.
3023  */
3024 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
3025                          struct inode *dir,
3026                          struct ceph_mds_request *req)
3027 {
3028         int err;
3029
3030         dout("do_request on %p\n", req);
3031
3032         /* issue */
3033         err = ceph_mdsc_submit_request(mdsc, dir, req);
3034         if (!err)
3035                 err = ceph_mdsc_wait_request(mdsc, req);
3036         dout("do_request %p done, result %d\n", req, err);
3037         return err;
3038 }
3039
3040 /*
3041  * Invalidate dir's completeness, dentry lease state on an aborted MDS
3042  * namespace request.
3043  */
3044 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
3045 {
3046         struct inode *dir = req->r_parent;
3047         struct inode *old_dir = req->r_old_dentry_dir;
3048
3049         dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
3050
3051         ceph_dir_clear_complete(dir);
3052         if (old_dir)
3053                 ceph_dir_clear_complete(old_dir);
3054         if (req->r_dentry)
3055                 ceph_invalidate_dentry_lease(req->r_dentry);
3056         if (req->r_old_dentry)
3057                 ceph_invalidate_dentry_lease(req->r_old_dentry);
3058 }
3059
3060 /*
3061  * Handle mds reply.
3062  *
3063  * We take the session mutex and parse and process the reply immediately.
3064  * This preserves the logical ordering of replies, capabilities, etc., sent
3065  * by the MDS as they are applied to our local cache.
3066  */
3067 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
3068 {
3069         struct ceph_mds_client *mdsc = session->s_mdsc;
3070         struct ceph_mds_request *req;
3071         struct ceph_mds_reply_head *head = msg->front.iov_base;
3072         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
3073         struct ceph_snap_realm *realm;
3074         u64 tid;
3075         int err, result;
3076         int mds = session->s_mds;
3077
3078         if (msg->front.iov_len < sizeof(*head)) {
3079                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
3080                 ceph_msg_dump(msg);
3081                 return;
3082         }
3083
3084         /* get request, session */
3085         tid = le64_to_cpu(msg->hdr.tid);
3086         mutex_lock(&mdsc->mutex);
3087         req = lookup_get_request(mdsc, tid);
3088         if (!req) {
3089                 dout("handle_reply on unknown tid %llu\n", tid);
3090                 mutex_unlock(&mdsc->mutex);
3091                 return;
3092         }
3093         dout("handle_reply %p\n", req);
3094
3095         /* correct session? */
3096         if (req->r_session != session) {
3097                 pr_err("mdsc_handle_reply got %llu on session mds%d"
3098                        " not mds%d\n", tid, session->s_mds,
3099                        req->r_session ? req->r_session->s_mds : -1);
3100                 mutex_unlock(&mdsc->mutex);
3101                 goto out;
3102         }
3103
3104         /* dup? */
3105         if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
3106             (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
3107                 pr_warn("got a dup %s reply on %llu from mds%d\n",
3108                            head->safe ? "safe" : "unsafe", tid, mds);
3109                 mutex_unlock(&mdsc->mutex);
3110                 goto out;
3111         }
3112         if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
3113                 pr_warn("got unsafe after safe on %llu from mds%d\n",
3114                            tid, mds);
3115                 mutex_unlock(&mdsc->mutex);
3116                 goto out;
3117         }
3118
3119         result = le32_to_cpu(head->result);
3120
3121         /*
3122          * Handle an ESTALE
3123          * if we're not talking to the authority, send to them
3124          * if the authority has changed while we weren't looking,
3125          * send to new authority
3126          * Otherwise we just have to return an ESTALE
3127          */
3128         if (result == -ESTALE) {
3129                 dout("got ESTALE on request %llu\n", req->r_tid);
3130                 req->r_resend_mds = -1;
3131                 if (req->r_direct_mode != USE_AUTH_MDS) {
3132                         dout("not using auth, setting for that now\n");
3133                         req->r_direct_mode = USE_AUTH_MDS;
3134                         __do_request(mdsc, req);
3135                         mutex_unlock(&mdsc->mutex);
3136                         goto out;
3137                 } else  {
3138                         int mds = __choose_mds(mdsc, req, NULL);
3139                         if (mds >= 0 && mds != req->r_session->s_mds) {
3140                                 dout("but auth changed, so resending\n");
3141                                 __do_request(mdsc, req);
3142                                 mutex_unlock(&mdsc->mutex);
3143                                 goto out;
3144                         }
3145                 }
3146                 dout("have to return ESTALE on request %llu\n", req->r_tid);
3147         }
3148
3149
3150         if (head->safe) {
3151                 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
3152                 __unregister_request(mdsc, req);
3153
3154                 /* last request during umount? */
3155                 if (mdsc->stopping && !__get_oldest_req(mdsc))
3156                         complete_all(&mdsc->safe_umount_waiters);
3157
3158                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3159                         /*
3160                          * We already handled the unsafe response, now do the
3161                          * cleanup.  No need to examine the response; the MDS
3162                          * doesn't include any result info in the safe
3163                          * response.  And even if it did, there is nothing
3164                          * useful we could do with a revised return value.
3165                          */
3166                         dout("got safe reply %llu, mds%d\n", tid, mds);
3167
3168                         mutex_unlock(&mdsc->mutex);
3169                         goto out;
3170                 }
3171         } else {
3172                 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
3173                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
3174         }
3175
3176         dout("handle_reply tid %lld result %d\n", tid, result);
3177         rinfo = &req->r_reply_info;
3178         if (test_bit(CEPHFS_FEATURE_REPLY_ENCODING, &session->s_features))
3179                 err = parse_reply_info(session, msg, rinfo, (u64)-1);
3180         else
3181                 err = parse_reply_info(session, msg, rinfo, session->s_con.peer_features);
3182         mutex_unlock(&mdsc->mutex);
3183
3184         /* Must find target inode outside of mutexes to avoid deadlocks */
3185         if ((err >= 0) && rinfo->head->is_target) {
3186                 struct inode *in;
3187                 struct ceph_vino tvino = {
3188                         .ino  = le64_to_cpu(rinfo->targeti.in->ino),
3189                         .snap = le64_to_cpu(rinfo->targeti.in->snapid)
3190                 };
3191
3192                 in = ceph_get_inode(mdsc->fsc->sb, tvino);
3193                 if (IS_ERR(in)) {
3194                         err = PTR_ERR(in);
3195                         mutex_lock(&session->s_mutex);
3196                         goto out_err;
3197                 }
3198                 req->r_target_inode = in;
3199         }
3200
3201         mutex_lock(&session->s_mutex);
3202         if (err < 0) {
3203                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
3204                 ceph_msg_dump(msg);
3205                 goto out_err;
3206         }
3207
3208         /* snap trace */
3209         realm = NULL;
3210         if (rinfo->snapblob_len) {
3211                 down_write(&mdsc->snap_rwsem);
3212                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
3213                                 rinfo->snapblob + rinfo->snapblob_len,
3214                                 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
3215                                 &realm);
3216                 downgrade_write(&mdsc->snap_rwsem);
3217         } else {
3218                 down_read(&mdsc->snap_rwsem);
3219         }
3220
3221         /* insert trace into our cache */
3222         mutex_lock(&req->r_fill_mutex);
3223         current->journal_info = req;
3224         err = ceph_fill_trace(mdsc->fsc->sb, req);
3225         if (err == 0) {
3226                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
3227                                     req->r_op == CEPH_MDS_OP_LSSNAP))
3228                         ceph_readdir_prepopulate(req, req->r_session);
3229         }
3230         current->journal_info = NULL;
3231         mutex_unlock(&req->r_fill_mutex);
3232
3233         up_read(&mdsc->snap_rwsem);
3234         if (realm)
3235                 ceph_put_snap_realm(mdsc, realm);
3236
3237         if (err == 0) {
3238                 if (req->r_target_inode &&
3239                     test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3240                         struct ceph_inode_info *ci =
3241                                 ceph_inode(req->r_target_inode);
3242                         spin_lock(&ci->i_unsafe_lock);
3243                         list_add_tail(&req->r_unsafe_target_item,
3244                                       &ci->i_unsafe_iops);
3245                         spin_unlock(&ci->i_unsafe_lock);
3246                 }
3247
3248                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
3249         }
3250 out_err:
3251         mutex_lock(&mdsc->mutex);
3252         if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3253                 if (err) {
3254                         req->r_err = err;
3255                 } else {
3256                         req->r_reply =  ceph_msg_get(msg);
3257                         set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
3258                 }
3259         } else {
3260                 dout("reply arrived after request %lld was aborted\n", tid);
3261         }
3262         mutex_unlock(&mdsc->mutex);
3263
3264         mutex_unlock(&session->s_mutex);
3265
3266         /* kick calling process */
3267         complete_request(mdsc, req);
3268
3269         ceph_update_metadata_latency(&mdsc->metric, req->r_start_latency,
3270                                      req->r_end_latency, err);
3271 out:
3272         ceph_mdsc_put_request(req);
3273         return;
3274 }
3275
3276
3277
3278 /*
3279  * handle mds notification that our request has been forwarded.
3280  */
3281 static void handle_forward(struct ceph_mds_client *mdsc,
3282                            struct ceph_mds_session *session,
3283                            struct ceph_msg *msg)
3284 {
3285         struct ceph_mds_request *req;
3286         u64 tid = le64_to_cpu(msg->hdr.tid);
3287         u32 next_mds;
3288         u32 fwd_seq;
3289         int err = -EINVAL;
3290         void *p = msg->front.iov_base;
3291         void *end = p + msg->front.iov_len;
3292
3293         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
3294         next_mds = ceph_decode_32(&p);
3295         fwd_seq = ceph_decode_32(&p);
3296
3297         mutex_lock(&mdsc->mutex);
3298         req = lookup_get_request(mdsc, tid);
3299         if (!req) {
3300                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
3301                 goto out;  /* dup reply? */
3302         }
3303
3304         if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3305                 dout("forward tid %llu aborted, unregistering\n", tid);
3306                 __unregister_request(mdsc, req);
3307         } else if (fwd_seq <= req->r_num_fwd) {
3308                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
3309                      tid, next_mds, req->r_num_fwd, fwd_seq);
3310         } else {
3311                 /* resend. forward race not possible; mds would drop */
3312                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
3313                 BUG_ON(req->r_err);
3314                 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
3315                 req->r_attempts = 0;
3316                 req->r_num_fwd = fwd_seq;
3317                 req->r_resend_mds = next_mds;
3318                 put_request_session(req);
3319                 __do_request(mdsc, req);
3320         }
3321         ceph_mdsc_put_request(req);
3322 out:
3323         mutex_unlock(&mdsc->mutex);
3324         return;
3325
3326 bad:
3327         pr_err("mdsc_handle_forward decode error err=%d\n", err);
3328 }
3329
3330 static int __decode_session_metadata(void **p, void *end,
3331                                      bool *blocklisted)
3332 {
3333         /* map<string,string> */
3334         u32 n;
3335         bool err_str;
3336         ceph_decode_32_safe(p, end, n, bad);
3337         while (n-- > 0) {
3338                 u32 len;
3339                 ceph_decode_32_safe(p, end, len, bad);
3340                 ceph_decode_need(p, end, len, bad);
3341                 err_str = !strncmp(*p, "error_string", len);
3342                 *p += len;
3343                 ceph_decode_32_safe(p, end, len, bad);
3344                 ceph_decode_need(p, end, len, bad);
3345                 /*
3346                  * Match "blocklisted (blacklisted)" from newer MDSes,
3347                  * or "blacklisted" from older MDSes.
3348                  */
3349                 if (err_str && strnstr(*p, "blacklisted", len))
3350                         *blocklisted = true;
3351                 *p += len;
3352         }
3353         return 0;
3354 bad:
3355         return -1;
3356 }
3357
3358 /*
3359  * handle a mds session control message
3360  */
3361 static void handle_session(struct ceph_mds_session *session,
3362                            struct ceph_msg *msg)
3363 {
3364         struct ceph_mds_client *mdsc = session->s_mdsc;
3365         int mds = session->s_mds;
3366         int msg_version = le16_to_cpu(msg->hdr.version);
3367         void *p = msg->front.iov_base;
3368         void *end = p + msg->front.iov_len;
3369         struct ceph_mds_session_head *h;
3370         u32 op;
3371         u64 seq, features = 0;
3372         int wake = 0;
3373         bool blocklisted = false;
3374
3375         /* decode */
3376         ceph_decode_need(&p, end, sizeof(*h), bad);
3377         h = p;
3378         p += sizeof(*h);
3379
3380         op = le32_to_cpu(h->op);
3381         seq = le64_to_cpu(h->seq);
3382
3383         if (msg_version >= 3) {
3384                 u32 len;
3385                 /* version >= 2, metadata */
3386                 if (__decode_session_metadata(&p, end, &blocklisted) < 0)
3387                         goto bad;
3388                 /* version >= 3, feature bits */
3389                 ceph_decode_32_safe(&p, end, len, bad);
3390                 if (len) {
3391                         ceph_decode_64_safe(&p, end, features, bad);
3392                         p += len - sizeof(features);
3393                 }
3394         }
3395
3396         mutex_lock(&mdsc->mutex);
3397         if (op == CEPH_SESSION_CLOSE) {
3398                 ceph_get_mds_session(session);
3399                 __unregister_session(mdsc, session);
3400         }
3401         /* FIXME: this ttl calculation is generous */
3402         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
3403         mutex_unlock(&mdsc->mutex);
3404
3405         mutex_lock(&session->s_mutex);
3406
3407         dout("handle_session mds%d %s %p state %s seq %llu\n",
3408              mds, ceph_session_op_name(op), session,
3409              ceph_session_state_name(session->s_state), seq);
3410
3411         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
3412                 session->s_state = CEPH_MDS_SESSION_OPEN;
3413                 pr_info("mds%d came back\n", session->s_mds);
3414         }
3415
3416         switch (op) {
3417         case CEPH_SESSION_OPEN:
3418                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3419                         pr_info("mds%d reconnect success\n", session->s_mds);
3420                 session->s_state = CEPH_MDS_SESSION_OPEN;
3421                 session->s_features = features;
3422                 renewed_caps(mdsc, session, 0);
3423                 if (test_bit(CEPHFS_FEATURE_METRIC_COLLECT, &session->s_features))
3424                         metric_schedule_delayed(&mdsc->metric);
3425                 wake = 1;
3426                 if (mdsc->stopping)
3427                         __close_session(mdsc, session);
3428                 break;
3429
3430         case CEPH_SESSION_RENEWCAPS:
3431                 if (session->s_renew_seq == seq)
3432                         renewed_caps(mdsc, session, 1);
3433                 break;
3434
3435         case CEPH_SESSION_CLOSE:
3436                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3437                         pr_info("mds%d reconnect denied\n", session->s_mds);
3438                 session->s_state = CEPH_MDS_SESSION_CLOSED;
3439                 cleanup_session_requests(mdsc, session);
3440                 remove_session_caps(session);
3441                 wake = 2; /* for good measure */
3442                 wake_up_all(&mdsc->session_close_wq);
3443                 break;
3444
3445         case CEPH_SESSION_STALE:
3446                 pr_info("mds%d caps went stale, renewing\n",
3447                         session->s_mds);
3448                 spin_lock(&session->s_gen_ttl_lock);
3449                 session->s_cap_gen++;
3450                 session->s_cap_ttl = jiffies - 1;
3451                 spin_unlock(&session->s_gen_ttl_lock);
3452                 send_renew_caps(mdsc, session);
3453                 break;
3454
3455         case CEPH_SESSION_RECALL_STATE:
3456                 ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
3457                 break;
3458
3459         case CEPH_SESSION_FLUSHMSG:
3460                 send_flushmsg_ack(mdsc, session, seq);
3461                 break;
3462
3463         case CEPH_SESSION_FORCE_RO:
3464                 dout("force_session_readonly %p\n", session);
3465                 spin_lock(&session->s_cap_lock);
3466                 session->s_readonly = true;
3467                 spin_unlock(&session->s_cap_lock);
3468                 wake_up_session_caps(session, FORCE_RO);
3469                 break;
3470
3471         case CEPH_SESSION_REJECT:
3472                 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
3473                 pr_info("mds%d rejected session\n", session->s_mds);
3474                 session->s_state = CEPH_MDS_SESSION_REJECTED;
3475                 cleanup_session_requests(mdsc, session);
3476                 remove_session_caps(session);
3477                 if (blocklisted)
3478                         mdsc->fsc->blocklisted = true;
3479                 wake = 2; /* for good measure */
3480                 break;
3481
3482         default:
3483                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
3484                 WARN_ON(1);
3485         }
3486
3487         mutex_unlock(&session->s_mutex);
3488         if (wake) {
3489                 mutex_lock(&mdsc->mutex);
3490                 __wake_requests(mdsc, &session->s_waiting);
3491                 if (wake == 2)
3492                         kick_requests(mdsc, mds);
3493                 mutex_unlock(&mdsc->mutex);
3494         }
3495         if (op == CEPH_SESSION_CLOSE)
3496                 ceph_put_mds_session(session);
3497         return;
3498
3499 bad:
3500         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
3501                (int)msg->front.iov_len);
3502         ceph_msg_dump(msg);
3503         return;
3504 }
3505
3506 void ceph_mdsc_release_dir_caps(struct ceph_mds_request *req)
3507 {
3508         int dcaps;
3509
3510         dcaps = xchg(&req->r_dir_caps, 0);
3511         if (dcaps) {
3512                 dout("releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
3513                 ceph_put_cap_refs(ceph_inode(req->r_parent), dcaps);
3514         }
3515 }
3516
3517 void ceph_mdsc_release_dir_caps_no_check(struct ceph_mds_request *req)
3518 {
3519         int dcaps;
3520
3521         dcaps = xchg(&req->r_dir_caps, 0);
3522         if (dcaps) {
3523                 dout("releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
3524                 ceph_put_cap_refs_no_check_caps(ceph_inode(req->r_parent),
3525                                                 dcaps);
3526         }
3527 }
3528
3529 /*
3530  * called under session->mutex.
3531  */
3532 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
3533                                    struct ceph_mds_session *session)
3534 {
3535         struct ceph_mds_request *req, *nreq;
3536         struct rb_node *p;
3537
3538         dout("replay_unsafe_requests mds%d\n", session->s_mds);
3539
3540         mutex_lock(&mdsc->mutex);
3541         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item)
3542                 __send_request(mdsc, session, req, true);
3543
3544         /*
3545          * also re-send old requests when MDS enters reconnect stage. So that MDS
3546          * can process completed request in clientreplay stage.
3547          */
3548         p = rb_first(&mdsc->request_tree);
3549         while (p) {
3550                 req = rb_entry(p, struct ceph_mds_request, r_node);
3551                 p = rb_next(p);
3552                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3553                         continue;
3554                 if (req->r_attempts == 0)
3555                         continue; /* only old requests */
3556                 if (!req->r_session)
3557                         continue;
3558                 if (req->r_session->s_mds != session->s_mds)
3559                         continue;
3560
3561                 ceph_mdsc_release_dir_caps_no_check(req);
3562
3563                 __send_request(mdsc, session, req, true);
3564         }
3565         mutex_unlock(&mdsc->mutex);
3566 }
3567
3568 static int send_reconnect_partial(struct ceph_reconnect_state *recon_state)
3569 {
3570         struct ceph_msg *reply;
3571         struct ceph_pagelist *_pagelist;
3572         struct page *page;
3573         __le32 *addr;
3574         int err = -ENOMEM;
3575
3576         if (!recon_state->allow_multi)
3577                 return -ENOSPC;
3578
3579         /* can't handle message that contains both caps and realm */
3580         BUG_ON(!recon_state->nr_caps == !recon_state->nr_realms);
3581
3582         /* pre-allocate new pagelist */
3583         _pagelist = ceph_pagelist_alloc(GFP_NOFS);
3584         if (!_pagelist)
3585                 return -ENOMEM;
3586
3587         reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3588         if (!reply)
3589                 goto fail_msg;
3590
3591         /* placeholder for nr_caps */
3592         err = ceph_pagelist_encode_32(_pagelist, 0);
3593         if (err < 0)
3594                 goto fail;
3595
3596         if (recon_state->nr_caps) {
3597                 /* currently encoding caps */
3598                 err = ceph_pagelist_encode_32(recon_state->pagelist, 0);
3599                 if (err)
3600                         goto fail;
3601         } else {
3602                 /* placeholder for nr_realms (currently encoding relams) */
3603                 err = ceph_pagelist_encode_32(_pagelist, 0);
3604                 if (err < 0)
3605                         goto fail;
3606         }
3607
3608         err = ceph_pagelist_encode_8(recon_state->pagelist, 1);
3609         if (err)
3610                 goto fail;
3611
3612         page = list_first_entry(&recon_state->pagelist->head, struct page, lru);
3613         addr = kmap_atomic(page);
3614         if (recon_state->nr_caps) {
3615                 /* currently encoding caps */
3616                 *addr = cpu_to_le32(recon_state->nr_caps);
3617         } else {
3618                 /* currently encoding relams */
3619                 *(addr + 1) = cpu_to_le32(recon_state->nr_realms);
3620         }
3621         kunmap_atomic(addr);
3622
3623         reply->hdr.version = cpu_to_le16(5);
3624         reply->hdr.compat_version = cpu_to_le16(4);
3625
3626         reply->hdr.data_len = cpu_to_le32(recon_state->pagelist->length);
3627         ceph_msg_data_add_pagelist(reply, recon_state->pagelist);
3628
3629         ceph_con_send(&recon_state->session->s_con, reply);
3630         ceph_pagelist_release(recon_state->pagelist);
3631
3632         recon_state->pagelist = _pagelist;
3633         recon_state->nr_caps = 0;
3634         recon_state->nr_realms = 0;
3635         recon_state->msg_version = 5;
3636         return 0;
3637 fail:
3638         ceph_msg_put(reply);
3639 fail_msg:
3640         ceph_pagelist_release(_pagelist);
3641         return err;
3642 }
3643
3644 static struct dentry* d_find_primary(struct inode *inode)
3645 {
3646         struct dentry *alias, *dn = NULL;
3647
3648         if (hlist_empty(&inode->i_dentry))
3649                 return NULL;
3650
3651         spin_lock(&inode->i_lock);
3652         if (hlist_empty(&inode->i_dentry))
3653                 goto out_unlock;
3654
3655         if (S_ISDIR(inode->i_mode)) {
3656                 alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias);
3657                 if (!IS_ROOT(alias))
3658                         dn = dget(alias);
3659                 goto out_unlock;
3660         }
3661
3662         hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
3663                 spin_lock(&alias->d_lock);
3664                 if (!d_unhashed(alias) &&
3665                     (ceph_dentry(alias)->flags & CEPH_DENTRY_PRIMARY_LINK)) {
3666                         dn = dget_dlock(alias);
3667                 }
3668                 spin_unlock(&alias->d_lock);
3669                 if (dn)
3670                         break;
3671         }
3672 out_unlock:
3673         spin_unlock(&inode->i_lock);
3674         return dn;
3675 }
3676
3677 /*
3678  * Encode information about a cap for a reconnect with the MDS.
3679  */
3680 static int reconnect_caps_cb(struct inode *inode, struct ceph_cap *cap,
3681                           void *arg)
3682 {
3683         union {
3684                 struct ceph_mds_cap_reconnect v2;
3685                 struct ceph_mds_cap_reconnect_v1 v1;
3686         } rec;
3687         struct ceph_inode_info *ci = cap->ci;
3688         struct ceph_reconnect_state *recon_state = arg;
3689         struct ceph_pagelist *pagelist = recon_state->pagelist;
3690         struct dentry *dentry;
3691         char *path;
3692         int pathlen, err;
3693         u64 pathbase;
3694         u64 snap_follows;
3695
3696         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
3697              inode, ceph_vinop(inode), cap, cap->cap_id,
3698              ceph_cap_string(cap->issued));
3699
3700         dentry = d_find_primary(inode);
3701         if (dentry) {
3702                 /* set pathbase to parent dir when msg_version >= 2 */
3703                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase,
3704                                             recon_state->msg_version >= 2);
3705                 dput(dentry);
3706                 if (IS_ERR(path)) {
3707                         err = PTR_ERR(path);
3708                         goto out_err;
3709                 }
3710         } else {
3711                 path = NULL;
3712                 pathlen = 0;
3713                 pathbase = 0;
3714         }
3715
3716         spin_lock(&ci->i_ceph_lock);
3717         cap->seq = 0;        /* reset cap seq */
3718         cap->issue_seq = 0;  /* and issue_seq */
3719         cap->mseq = 0;       /* and migrate_seq */
3720         cap->cap_gen = cap->session->s_cap_gen;
3721
3722         /* These are lost when the session goes away */
3723         if (S_ISDIR(inode->i_mode)) {
3724                 if (cap->issued & CEPH_CAP_DIR_CREATE) {
3725                         ceph_put_string(rcu_dereference_raw(ci->i_cached_layout.pool_ns));
3726                         memset(&ci->i_cached_layout, 0, sizeof(ci->i_cached_layout));
3727                 }
3728                 cap->issued &= ~CEPH_CAP_ANY_DIR_OPS;
3729         }
3730
3731         if (recon_state->msg_version >= 2) {
3732                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
3733                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3734                 rec.v2.issued = cpu_to_le32(cap->issued);
3735                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3736                 rec.v2.pathbase = cpu_to_le64(pathbase);
3737                 rec.v2.flock_len = (__force __le32)
3738                         ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
3739         } else {
3740                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
3741                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3742                 rec.v1.issued = cpu_to_le32(cap->issued);
3743                 rec.v1.size = cpu_to_le64(inode->i_size);
3744                 ceph_encode_timespec64(&rec.v1.mtime, &inode->i_mtime);
3745                 ceph_encode_timespec64(&rec.v1.atime, &inode->i_atime);
3746                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3747                 rec.v1.pathbase = cpu_to_le64(pathbase);
3748         }
3749
3750         if (list_empty(&ci->i_cap_snaps)) {
3751                 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
3752         } else {
3753                 struct ceph_cap_snap *capsnap =
3754                         list_first_entry(&ci->i_cap_snaps,
3755                                          struct ceph_cap_snap, ci_item);
3756                 snap_follows = capsnap->follows;
3757         }
3758         spin_unlock(&ci->i_ceph_lock);
3759
3760         if (recon_state->msg_version >= 2) {
3761                 int num_fcntl_locks, num_flock_locks;
3762                 struct ceph_filelock *flocks = NULL;
3763                 size_t struct_len, total_len = sizeof(u64);
3764                 u8 struct_v = 0;
3765
3766 encode_again:
3767                 if (rec.v2.flock_len) {
3768                         ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
3769                 } else {
3770                         num_fcntl_locks = 0;
3771                         num_flock_locks = 0;
3772                 }
3773                 if (num_fcntl_locks + num_flock_locks > 0) {
3774                         flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
3775                                                sizeof(struct ceph_filelock),
3776                                                GFP_NOFS);
3777                         if (!flocks) {
3778                                 err = -ENOMEM;
3779                                 goto out_err;
3780                         }
3781                         err = ceph_encode_locks_to_buffer(inode, flocks,
3782                                                           num_fcntl_locks,
3783                                                           num_flock_locks);
3784                         if (err) {
3785                                 kfree(flocks);
3786                                 flocks = NULL;
3787                                 if (err == -ENOSPC)
3788                                         goto encode_again;
3789                                 goto out_err;
3790                         }
3791                 } else {
3792                         kfree(flocks);
3793                         flocks = NULL;
3794                 }
3795
3796                 if (recon_state->msg_version >= 3) {
3797                         /* version, compat_version and struct_len */
3798                         total_len += 2 * sizeof(u8) + sizeof(u32);
3799                         struct_v = 2;
3800                 }
3801                 /*
3802                  * number of encoded locks is stable, so copy to pagelist
3803                  */
3804                 struct_len = 2 * sizeof(u32) +
3805                             (num_fcntl_locks + num_flock_locks) *
3806                             sizeof(struct ceph_filelock);
3807                 rec.v2.flock_len = cpu_to_le32(struct_len);
3808
3809                 struct_len += sizeof(u32) + pathlen + sizeof(rec.v2);
3810
3811                 if (struct_v >= 2)
3812                         struct_len += sizeof(u64); /* snap_follows */
3813
3814                 total_len += struct_len;
3815
3816                 if (pagelist->length + total_len > RECONNECT_MAX_SIZE) {
3817                         err = send_reconnect_partial(recon_state);
3818                         if (err)
3819                                 goto out_freeflocks;
3820                         pagelist = recon_state->pagelist;
3821                 }
3822
3823                 err = ceph_pagelist_reserve(pagelist, total_len);
3824                 if (err)
3825                         goto out_freeflocks;
3826
3827                 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3828                 if (recon_state->msg_version >= 3) {
3829                         ceph_pagelist_encode_8(pagelist, struct_v);
3830                         ceph_pagelist_encode_8(pagelist, 1);
3831                         ceph_pagelist_encode_32(pagelist, struct_len);
3832                 }
3833                 ceph_pagelist_encode_string(pagelist, path, pathlen);
3834                 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
3835                 ceph_locks_to_pagelist(flocks, pagelist,
3836                                        num_fcntl_locks, num_flock_locks);
3837                 if (struct_v >= 2)
3838                         ceph_pagelist_encode_64(pagelist, snap_follows);
3839 out_freeflocks:
3840                 kfree(flocks);
3841         } else {
3842                 err = ceph_pagelist_reserve(pagelist,
3843                                             sizeof(u64) + sizeof(u32) +
3844                                             pathlen + sizeof(rec.v1));
3845                 if (err)
3846                         goto out_err;
3847
3848                 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3849                 ceph_pagelist_encode_string(pagelist, path, pathlen);
3850                 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
3851         }
3852
3853 out_err:
3854         ceph_mdsc_free_path(path, pathlen);
3855         if (!err)
3856                 recon_state->nr_caps++;
3857         return err;
3858 }
3859
3860 static int encode_snap_realms(struct ceph_mds_client *mdsc,
3861                               struct ceph_reconnect_state *recon_state)
3862 {
3863         struct rb_node *p;
3864         struct ceph_pagelist *pagelist = recon_state->pagelist;
3865         int err = 0;
3866
3867         if (recon_state->msg_version >= 4) {
3868                 err = ceph_pagelist_encode_32(pagelist, mdsc->num_snap_realms);
3869                 if (err < 0)
3870                         goto fail;
3871         }
3872
3873         /*
3874          * snaprealms.  we provide mds with the ino, seq (version), and
3875          * parent for all of our realms.  If the mds has any newer info,
3876          * it will tell us.
3877          */
3878         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3879                 struct ceph_snap_realm *realm =
3880                        rb_entry(p, struct ceph_snap_realm, node);
3881                 struct ceph_mds_snaprealm_reconnect sr_rec;
3882
3883                 if (recon_state->msg_version >= 4) {
3884                         size_t need = sizeof(u8) * 2 + sizeof(u32) +
3885                                       sizeof(sr_rec);
3886
3887                         if (pagelist->length + need > RECONNECT_MAX_SIZE) {
3888                                 err = send_reconnect_partial(recon_state);
3889                                 if (err)
3890                                         goto fail;
3891                                 pagelist = recon_state->pagelist;
3892                         }
3893
3894                         err = ceph_pagelist_reserve(pagelist, need);
3895                         if (err)
3896                                 goto fail;
3897
3898                         ceph_pagelist_encode_8(pagelist, 1);
3899                         ceph_pagelist_encode_8(pagelist, 1);
3900                         ceph_pagelist_encode_32(pagelist, sizeof(sr_rec));
3901                 }
3902
3903                 dout(" adding snap realm %llx seq %lld parent %llx\n",
3904                      realm->ino, realm->seq, realm->parent_ino);
3905                 sr_rec.ino = cpu_to_le64(realm->ino);
3906                 sr_rec.seq = cpu_to_le64(realm->seq);
3907                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3908
3909                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3910                 if (err)
3911                         goto fail;
3912
3913                 recon_state->nr_realms++;
3914         }
3915 fail:
3916         return err;
3917 }
3918
3919
3920 /*
3921  * If an MDS fails and recovers, clients need to reconnect in order to
3922  * reestablish shared state.  This includes all caps issued through
3923  * this session _and_ the snap_realm hierarchy.  Because it's not
3924  * clear which snap realms the mds cares about, we send everything we
3925  * know about.. that ensures we'll then get any new info the
3926  * recovering MDS might have.
3927  *
3928  * This is a relatively heavyweight operation, but it's rare.
3929  */
3930 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3931                                struct ceph_mds_session *session)
3932 {
3933         struct ceph_msg *reply;
3934         int mds = session->s_mds;
3935         int err = -ENOMEM;
3936         struct ceph_reconnect_state recon_state = {
3937                 .session = session,
3938         };
3939         LIST_HEAD(dispose);
3940
3941         pr_info("mds%d reconnect start\n", mds);
3942
3943         recon_state.pagelist = ceph_pagelist_alloc(GFP_NOFS);
3944         if (!recon_state.pagelist)
3945                 goto fail_nopagelist;
3946
3947         reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3948         if (!reply)
3949                 goto fail_nomsg;
3950
3951         xa_destroy(&session->s_delegated_inos);
3952
3953         mutex_lock(&session->s_mutex);
3954         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3955         session->s_seq = 0;
3956
3957         dout("session %p state %s\n", session,
3958              ceph_session_state_name(session->s_state));
3959
3960         spin_lock(&session->s_gen_ttl_lock);
3961         session->s_cap_gen++;
3962         spin_unlock(&session->s_gen_ttl_lock);
3963
3964         spin_lock(&session->s_cap_lock);
3965         /* don't know if session is readonly */
3966         session->s_readonly = 0;
3967         /*
3968          * notify __ceph_remove_cap() that we are composing cap reconnect.
3969          * If a cap get released before being added to the cap reconnect,
3970          * __ceph_remove_cap() should skip queuing cap release.
3971          */
3972         session->s_cap_reconnect = 1;
3973         /* drop old cap expires; we're about to reestablish that state */
3974         detach_cap_releases(session, &dispose);
3975         spin_unlock(&session->s_cap_lock);
3976         dispose_cap_releases(mdsc, &dispose);
3977
3978         /* trim unused caps to reduce MDS's cache rejoin time */
3979         if (mdsc->fsc->sb->s_root)
3980                 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3981
3982         ceph_con_close(&session->s_con);
3983         ceph_con_open(&session->s_con,
3984                       CEPH_ENTITY_TYPE_MDS, mds,
3985                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3986
3987         /* replay unsafe requests */
3988         replay_unsafe_requests(mdsc, session);
3989
3990         ceph_early_kick_flushing_caps(mdsc, session);
3991
3992         down_read(&mdsc->snap_rwsem);
3993
3994         /* placeholder for nr_caps */
3995         err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
3996         if (err)
3997                 goto fail;
3998
3999         if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
4000                 recon_state.msg_version = 3;
4001                 recon_state.allow_multi = true;
4002         } else if (session->s_con.peer_features & CEPH_FEATURE_MDSENC) {
4003                 recon_state.msg_version = 3;
4004         } else {
4005                 recon_state.msg_version = 2;
4006         }
4007         /* trsaverse this session's caps */
4008         err = ceph_iterate_session_caps(session, reconnect_caps_cb, &recon_state);
4009
4010         spin_lock(&session->s_cap_lock);
4011         session->s_cap_reconnect = 0;
4012         spin_unlock(&session->s_cap_lock);
4013
4014         if (err < 0)
4015                 goto fail;
4016
4017         /* check if all realms can be encoded into current message */
4018         if (mdsc->num_snap_realms) {
4019                 size_t total_len =
4020                         recon_state.pagelist->length +
4021                         mdsc->num_snap_realms *
4022                         sizeof(struct ceph_mds_snaprealm_reconnect);
4023                 if (recon_state.msg_version >= 4) {
4024                         /* number of realms */
4025                         total_len += sizeof(u32);
4026                         /* version, compat_version and struct_len */
4027                         total_len += mdsc->num_snap_realms *
4028                                      (2 * sizeof(u8) + sizeof(u32));
4029                 }
4030                 if (total_len > RECONNECT_MAX_SIZE) {
4031                         if (!recon_state.allow_multi) {
4032                                 err = -ENOSPC;
4033                                 goto fail;
4034                         }
4035                         if (recon_state.nr_caps) {
4036                                 err = send_reconnect_partial(&recon_state);
4037                                 if (err)
4038                                         goto fail;
4039                         }
4040                         recon_state.msg_version = 5;
4041                 }
4042         }
4043
4044         err = encode_snap_realms(mdsc, &recon_state);
4045         if (err < 0)
4046                 goto fail;
4047
4048         if (recon_state.msg_version >= 5) {
4049                 err = ceph_pagelist_encode_8(recon_state.pagelist, 0);
4050                 if (err < 0)
4051                         goto fail;
4052         }
4053
4054         if (recon_state.nr_caps || recon_state.nr_realms) {
4055                 struct page *page =
4056                         list_first_entry(&recon_state.pagelist->head,
4057                                         struct page, lru);
4058                 __le32 *addr = kmap_atomic(page);
4059                 if (recon_state.nr_caps) {
4060                         WARN_ON(recon_state.nr_realms != mdsc->num_snap_realms);
4061                         *addr = cpu_to_le32(recon_state.nr_caps);
4062                 } else if (recon_state.msg_version >= 4) {
4063                         *(addr + 1) = cpu_to_le32(recon_state.nr_realms);
4064                 }
4065                 kunmap_atomic(addr);
4066         }
4067
4068         reply->hdr.version = cpu_to_le16(recon_state.msg_version);
4069         if (recon_state.msg_version >= 4)
4070                 reply->hdr.compat_version = cpu_to_le16(4);
4071
4072         reply->hdr.data_len = cpu_to_le32(recon_state.pagelist->length);
4073         ceph_msg_data_add_pagelist(reply, recon_state.pagelist);
4074
4075         ceph_con_send(&session->s_con, reply);
4076
4077         mutex_unlock(&session->s_mutex);
4078
4079         mutex_lock(&mdsc->mutex);
4080         __wake_requests(mdsc, &session->s_waiting);
4081         mutex_unlock(&mdsc->mutex);
4082
4083         up_read(&mdsc->snap_rwsem);
4084         ceph_pagelist_release(recon_state.pagelist);
4085         return;
4086
4087 fail:
4088         ceph_msg_put(reply);
4089         up_read(&mdsc->snap_rwsem);
4090         mutex_unlock(&session->s_mutex);
4091 fail_nomsg:
4092         ceph_pagelist_release(recon_state.pagelist);
4093 fail_nopagelist:
4094         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
4095         return;
4096 }
4097
4098
4099 /*
4100  * compare old and new mdsmaps, kicking requests
4101  * and closing out old connections as necessary
4102  *
4103  * called under mdsc->mutex.
4104  */
4105 static void check_new_map(struct ceph_mds_client *mdsc,
4106                           struct ceph_mdsmap *newmap,
4107                           struct ceph_mdsmap *oldmap)
4108 {
4109         int i;
4110         int oldstate, newstate;
4111         struct ceph_mds_session *s;
4112
4113         dout("check_new_map new %u old %u\n",
4114              newmap->m_epoch, oldmap->m_epoch);
4115
4116         for (i = 0; i < oldmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4117                 if (!mdsc->sessions[i])
4118                         continue;
4119                 s = mdsc->sessions[i];
4120                 oldstate = ceph_mdsmap_get_state(oldmap, i);
4121                 newstate = ceph_mdsmap_get_state(newmap, i);
4122
4123                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
4124                      i, ceph_mds_state_name(oldstate),
4125                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
4126                      ceph_mds_state_name(newstate),
4127                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
4128                      ceph_session_state_name(s->s_state));
4129
4130                 if (i >= newmap->possible_max_rank) {
4131                         /* force close session for stopped mds */
4132                         ceph_get_mds_session(s);
4133                         __unregister_session(mdsc, s);
4134                         __wake_requests(mdsc, &s->s_waiting);
4135                         mutex_unlock(&mdsc->mutex);
4136
4137                         mutex_lock(&s->s_mutex);
4138                         cleanup_session_requests(mdsc, s);
4139                         remove_session_caps(s);
4140                         mutex_unlock(&s->s_mutex);
4141
4142                         ceph_put_mds_session(s);
4143
4144                         mutex_lock(&mdsc->mutex);
4145                         kick_requests(mdsc, i);
4146                         continue;
4147                 }
4148
4149                 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
4150                            ceph_mdsmap_get_addr(newmap, i),
4151                            sizeof(struct ceph_entity_addr))) {
4152                         /* just close it */
4153                         mutex_unlock(&mdsc->mutex);
4154                         mutex_lock(&s->s_mutex);
4155                         mutex_lock(&mdsc->mutex);
4156                         ceph_con_close(&s->s_con);
4157                         mutex_unlock(&s->s_mutex);
4158                         s->s_state = CEPH_MDS_SESSION_RESTARTING;
4159                 } else if (oldstate == newstate) {
4160                         continue;  /* nothing new with this mds */
4161                 }
4162
4163                 /*
4164                  * send reconnect?
4165                  */
4166                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
4167                     newstate >= CEPH_MDS_STATE_RECONNECT) {
4168                         mutex_unlock(&mdsc->mutex);
4169                         send_mds_reconnect(mdsc, s);
4170                         mutex_lock(&mdsc->mutex);
4171                 }
4172
4173                 /*
4174                  * kick request on any mds that has gone active.
4175                  */
4176                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
4177                     newstate >= CEPH_MDS_STATE_ACTIVE) {
4178                         if (oldstate != CEPH_MDS_STATE_CREATING &&
4179                             oldstate != CEPH_MDS_STATE_STARTING)
4180                                 pr_info("mds%d recovery completed\n", s->s_mds);
4181                         kick_requests(mdsc, i);
4182                         mutex_unlock(&mdsc->mutex);
4183                         mutex_lock(&s->s_mutex);
4184                         mutex_lock(&mdsc->mutex);
4185                         ceph_kick_flushing_caps(mdsc, s);
4186                         mutex_unlock(&s->s_mutex);
4187                         wake_up_session_caps(s, RECONNECT);
4188                 }
4189         }
4190
4191         for (i = 0; i < newmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4192                 s = mdsc->sessions[i];
4193                 if (!s)
4194                         continue;
4195                 if (!ceph_mdsmap_is_laggy(newmap, i))
4196                         continue;
4197                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4198                     s->s_state == CEPH_MDS_SESSION_HUNG ||
4199                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
4200                         dout(" connecting to export targets of laggy mds%d\n",
4201                              i);
4202                         __open_export_target_sessions(mdsc, s);
4203                 }
4204         }
4205 }
4206
4207
4208
4209 /*
4210  * leases
4211  */
4212
4213 /*
4214  * caller must hold session s_mutex, dentry->d_lock
4215  */
4216 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
4217 {
4218         struct ceph_dentry_info *di = ceph_dentry(dentry);
4219
4220         ceph_put_mds_session(di->lease_session);
4221         di->lease_session = NULL;
4222 }
4223
4224 static void handle_lease(struct ceph_mds_client *mdsc,
4225                          struct ceph_mds_session *session,
4226                          struct ceph_msg *msg)
4227 {
4228         struct super_block *sb = mdsc->fsc->sb;
4229         struct inode *inode;
4230         struct dentry *parent, *dentry;
4231         struct ceph_dentry_info *di;
4232         int mds = session->s_mds;
4233         struct ceph_mds_lease *h = msg->front.iov_base;
4234         u32 seq;
4235         struct ceph_vino vino;
4236         struct qstr dname;
4237         int release = 0;
4238
4239         dout("handle_lease from mds%d\n", mds);
4240
4241         /* decode */
4242         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
4243                 goto bad;
4244         vino.ino = le64_to_cpu(h->ino);
4245         vino.snap = CEPH_NOSNAP;
4246         seq = le32_to_cpu(h->seq);
4247         dname.len = get_unaligned_le32(h + 1);
4248         if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
4249                 goto bad;
4250         dname.name = (void *)(h + 1) + sizeof(u32);
4251
4252         /* lookup inode */
4253         inode = ceph_find_inode(sb, vino);
4254         dout("handle_lease %s, ino %llx %p %.*s\n",
4255              ceph_lease_op_name(h->action), vino.ino, inode,
4256              dname.len, dname.name);
4257
4258         mutex_lock(&session->s_mutex);
4259         inc_session_sequence(session);
4260
4261         if (!inode) {
4262                 dout("handle_lease no inode %llx\n", vino.ino);
4263                 goto release;
4264         }
4265
4266         /* dentry */
4267         parent = d_find_alias(inode);
4268         if (!parent) {
4269                 dout("no parent dentry on inode %p\n", inode);
4270                 WARN_ON(1);
4271                 goto release;  /* hrm... */
4272         }
4273         dname.hash = full_name_hash(parent, dname.name, dname.len);
4274         dentry = d_lookup(parent, &dname);
4275         dput(parent);
4276         if (!dentry)
4277                 goto release;
4278
4279         spin_lock(&dentry->d_lock);
4280         di = ceph_dentry(dentry);
4281         switch (h->action) {
4282         case CEPH_MDS_LEASE_REVOKE:
4283                 if (di->lease_session == session) {
4284                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
4285                                 h->seq = cpu_to_le32(di->lease_seq);
4286                         __ceph_mdsc_drop_dentry_lease(dentry);
4287                 }
4288                 release = 1;
4289                 break;
4290
4291         case CEPH_MDS_LEASE_RENEW:
4292                 if (di->lease_session == session &&
4293                     di->lease_gen == session->s_cap_gen &&
4294                     di->lease_renew_from &&
4295                     di->lease_renew_after == 0) {
4296                         unsigned long duration =
4297                                 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
4298
4299                         di->lease_seq = seq;
4300                         di->time = di->lease_renew_from + duration;
4301                         di->lease_renew_after = di->lease_renew_from +
4302                                 (duration >> 1);
4303                         di->lease_renew_from = 0;
4304                 }
4305                 break;
4306         }
4307         spin_unlock(&dentry->d_lock);
4308         dput(dentry);
4309
4310         if (!release)
4311                 goto out;
4312
4313 release:
4314         /* let's just reuse the same message */
4315         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
4316         ceph_msg_get(msg);
4317         ceph_con_send(&session->s_con, msg);
4318
4319 out:
4320         mutex_unlock(&session->s_mutex);
4321         /* avoid calling iput_final() in mds dispatch threads */
4322         ceph_async_iput(inode);
4323         return;
4324
4325 bad:
4326         pr_err("corrupt lease message\n");
4327         ceph_msg_dump(msg);
4328 }
4329
4330 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
4331                               struct dentry *dentry, char action,
4332                               u32 seq)
4333 {
4334         struct ceph_msg *msg;
4335         struct ceph_mds_lease *lease;
4336         struct inode *dir;
4337         int len = sizeof(*lease) + sizeof(u32) + NAME_MAX;
4338
4339         dout("lease_send_msg identry %p %s to mds%d\n",
4340              dentry, ceph_lease_op_name(action), session->s_mds);
4341
4342         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
4343         if (!msg)
4344                 return;
4345         lease = msg->front.iov_base;
4346         lease->action = action;
4347         lease->seq = cpu_to_le32(seq);
4348
4349         spin_lock(&dentry->d_lock);
4350         dir = d_inode(dentry->d_parent);
4351         lease->ino = cpu_to_le64(ceph_ino(dir));
4352         lease->first = lease->last = cpu_to_le64(ceph_snap(dir));
4353
4354         put_unaligned_le32(dentry->d_name.len, lease + 1);
4355         memcpy((void *)(lease + 1) + 4,
4356                dentry->d_name.name, dentry->d_name.len);
4357         spin_unlock(&dentry->d_lock);
4358         /*
4359          * if this is a preemptive lease RELEASE, no need to
4360          * flush request stream, since the actual request will
4361          * soon follow.
4362          */
4363         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
4364
4365         ceph_con_send(&session->s_con, msg);
4366 }
4367
4368 /*
4369  * lock unlock sessions, to wait ongoing session activities
4370  */
4371 static void lock_unlock_sessions(struct ceph_mds_client *mdsc)
4372 {
4373         int i;
4374
4375         mutex_lock(&mdsc->mutex);
4376         for (i = 0; i < mdsc->max_sessions; i++) {
4377                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4378                 if (!s)
4379                         continue;
4380                 mutex_unlock(&mdsc->mutex);
4381                 mutex_lock(&s->s_mutex);
4382                 mutex_unlock(&s->s_mutex);
4383                 ceph_put_mds_session(s);
4384                 mutex_lock(&mdsc->mutex);
4385         }
4386         mutex_unlock(&mdsc->mutex);
4387 }
4388
4389 static void maybe_recover_session(struct ceph_mds_client *mdsc)
4390 {
4391         struct ceph_fs_client *fsc = mdsc->fsc;
4392
4393         if (!ceph_test_mount_opt(fsc, CLEANRECOVER))
4394                 return;
4395
4396         if (READ_ONCE(fsc->mount_state) != CEPH_MOUNT_MOUNTED)
4397                 return;
4398
4399         if (!READ_ONCE(fsc->blocklisted))
4400                 return;
4401
4402         pr_info("auto reconnect after blocklisted\n");
4403         ceph_force_reconnect(fsc->sb);
4404 }
4405
4406 bool check_session_state(struct ceph_mds_session *s)
4407 {
4408         switch (s->s_state) {
4409         case CEPH_MDS_SESSION_OPEN:
4410                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
4411                         s->s_state = CEPH_MDS_SESSION_HUNG;
4412                         pr_info("mds%d hung\n", s->s_mds);
4413                 }
4414                 break;
4415         case CEPH_MDS_SESSION_CLOSING:
4416                 /* Should never reach this when we're unmounting */
4417                 WARN_ON_ONCE(true);
4418                 fallthrough;
4419         case CEPH_MDS_SESSION_NEW:
4420         case CEPH_MDS_SESSION_RESTARTING:
4421         case CEPH_MDS_SESSION_CLOSED:
4422         case CEPH_MDS_SESSION_REJECTED:
4423                 return false;
4424         }
4425
4426         return true;
4427 }
4428
4429 /*
4430  * If the sequence is incremented while we're waiting on a REQUEST_CLOSE reply,
4431  * then we need to retransmit that request.
4432  */
4433 void inc_session_sequence(struct ceph_mds_session *s)
4434 {
4435         lockdep_assert_held(&s->s_mutex);
4436
4437         s->s_seq++;
4438
4439         if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
4440                 int ret;
4441
4442                 dout("resending session close request for mds%d\n", s->s_mds);
4443                 ret = request_close_session(s);
4444                 if (ret < 0)
4445                         pr_err("unable to close session to mds%d: %d\n",
4446                                s->s_mds, ret);
4447         }
4448 }
4449
4450 /*
4451  * delayed work -- periodically trim expired leases, renew caps with mds
4452  */
4453 static void schedule_delayed(struct ceph_mds_client *mdsc)
4454 {
4455         int delay = 5;
4456         unsigned hz = round_jiffies_relative(HZ * delay);
4457         schedule_delayed_work(&mdsc->delayed_work, hz);
4458 }
4459
4460 static void delayed_work(struct work_struct *work)
4461 {
4462         int i;
4463         struct ceph_mds_client *mdsc =
4464                 container_of(work, struct ceph_mds_client, delayed_work.work);
4465         int renew_interval;
4466         int renew_caps;
4467
4468         dout("mdsc delayed_work\n");
4469
4470         if (mdsc->stopping)
4471                 return;
4472
4473         mutex_lock(&mdsc->mutex);
4474         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
4475         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
4476                                    mdsc->last_renew_caps);
4477         if (renew_caps)
4478                 mdsc->last_renew_caps = jiffies;
4479
4480         for (i = 0; i < mdsc->max_sessions; i++) {
4481                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4482                 if (!s)
4483                         continue;
4484
4485                 if (!check_session_state(s)) {
4486                         ceph_put_mds_session(s);
4487                         continue;
4488                 }
4489                 mutex_unlock(&mdsc->mutex);
4490
4491                 mutex_lock(&s->s_mutex);
4492                 if (renew_caps)
4493                         send_renew_caps(mdsc, s);
4494                 else
4495                         ceph_con_keepalive(&s->s_con);
4496                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4497                     s->s_state == CEPH_MDS_SESSION_HUNG)
4498                         ceph_send_cap_releases(mdsc, s);
4499                 mutex_unlock(&s->s_mutex);
4500                 ceph_put_mds_session(s);
4501
4502                 mutex_lock(&mdsc->mutex);
4503         }
4504         mutex_unlock(&mdsc->mutex);
4505
4506         ceph_check_delayed_caps(mdsc);
4507
4508         ceph_queue_cap_reclaim_work(mdsc);
4509
4510         ceph_trim_snapid_map(mdsc);
4511
4512         maybe_recover_session(mdsc);
4513
4514         schedule_delayed(mdsc);
4515 }
4516
4517 int ceph_mdsc_init(struct ceph_fs_client *fsc)
4518
4519 {
4520         struct ceph_mds_client *mdsc;
4521         int err;
4522
4523         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
4524         if (!mdsc)
4525                 return -ENOMEM;
4526         mdsc->fsc = fsc;
4527         mutex_init(&mdsc->mutex);
4528         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
4529         if (!mdsc->mdsmap) {
4530                 err = -ENOMEM;
4531                 goto err_mdsc;
4532         }
4533
4534         init_completion(&mdsc->safe_umount_waiters);
4535         init_waitqueue_head(&mdsc->session_close_wq);
4536         INIT_LIST_HEAD(&mdsc->waiting_for_map);
4537         mdsc->sessions = NULL;
4538         atomic_set(&mdsc->num_sessions, 0);
4539         mdsc->max_sessions = 0;
4540         mdsc->stopping = 0;
4541         atomic64_set(&mdsc->quotarealms_count, 0);
4542         mdsc->quotarealms_inodes = RB_ROOT;
4543         mutex_init(&mdsc->quotarealms_inodes_mutex);
4544         mdsc->last_snap_seq = 0;
4545         init_rwsem(&mdsc->snap_rwsem);
4546         mdsc->snap_realms = RB_ROOT;
4547         INIT_LIST_HEAD(&mdsc->snap_empty);
4548         mdsc->num_snap_realms = 0;
4549         spin_lock_init(&mdsc->snap_empty_lock);
4550         mdsc->last_tid = 0;
4551         mdsc->oldest_tid = 0;
4552         mdsc->request_tree = RB_ROOT;
4553         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
4554         mdsc->last_renew_caps = jiffies;
4555         INIT_LIST_HEAD(&mdsc->cap_delay_list);
4556         INIT_LIST_HEAD(&mdsc->cap_wait_list);
4557         spin_lock_init(&mdsc->cap_delay_lock);
4558         INIT_LIST_HEAD(&mdsc->snap_flush_list);
4559         spin_lock_init(&mdsc->snap_flush_lock);
4560         mdsc->last_cap_flush_tid = 1;
4561         INIT_LIST_HEAD(&mdsc->cap_flush_list);
4562         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
4563         mdsc->num_cap_flushing = 0;
4564         spin_lock_init(&mdsc->cap_dirty_lock);
4565         init_waitqueue_head(&mdsc->cap_flushing_wq);
4566         INIT_WORK(&mdsc->cap_reclaim_work, ceph_cap_reclaim_work);
4567         atomic_set(&mdsc->cap_reclaim_pending, 0);
4568         err = ceph_metric_init(&mdsc->metric);
4569         if (err)
4570                 goto err_mdsmap;
4571
4572         spin_lock_init(&mdsc->dentry_list_lock);
4573         INIT_LIST_HEAD(&mdsc->dentry_leases);
4574         INIT_LIST_HEAD(&mdsc->dentry_dir_leases);
4575
4576         ceph_caps_init(mdsc);
4577         ceph_adjust_caps_max_min(mdsc, fsc->mount_options);
4578
4579         spin_lock_init(&mdsc->snapid_map_lock);
4580         mdsc->snapid_map_tree = RB_ROOT;
4581         INIT_LIST_HEAD(&mdsc->snapid_map_lru);
4582
4583         init_rwsem(&mdsc->pool_perm_rwsem);
4584         mdsc->pool_perm_tree = RB_ROOT;
4585
4586         strscpy(mdsc->nodename, utsname()->nodename,
4587                 sizeof(mdsc->nodename));
4588
4589         fsc->mdsc = mdsc;
4590         return 0;
4591
4592 err_mdsmap:
4593         kfree(mdsc->mdsmap);
4594 err_mdsc:
4595         kfree(mdsc);
4596         return err;
4597 }
4598
4599 /*
4600  * Wait for safe replies on open mds requests.  If we time out, drop
4601  * all requests from the tree to avoid dangling dentry refs.
4602  */
4603 static void wait_requests(struct ceph_mds_client *mdsc)
4604 {
4605         struct ceph_options *opts = mdsc->fsc->client->options;
4606         struct ceph_mds_request *req;
4607
4608         mutex_lock(&mdsc->mutex);
4609         if (__get_oldest_req(mdsc)) {
4610                 mutex_unlock(&mdsc->mutex);
4611
4612                 dout("wait_requests waiting for requests\n");
4613                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
4614                                     ceph_timeout_jiffies(opts->mount_timeout));
4615
4616                 /* tear down remaining requests */
4617                 mutex_lock(&mdsc->mutex);
4618                 while ((req = __get_oldest_req(mdsc))) {
4619                         dout("wait_requests timed out on tid %llu\n",
4620                              req->r_tid);
4621                         list_del_init(&req->r_wait);
4622                         __unregister_request(mdsc, req);
4623                 }
4624         }
4625         mutex_unlock(&mdsc->mutex);
4626         dout("wait_requests done\n");
4627 }
4628
4629 /*
4630  * called before mount is ro, and before dentries are torn down.
4631  * (hmm, does this still race with new lookups?)
4632  */
4633 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
4634 {
4635         dout("pre_umount\n");
4636         mdsc->stopping = 1;
4637
4638         lock_unlock_sessions(mdsc);
4639         ceph_flush_dirty_caps(mdsc);
4640         wait_requests(mdsc);
4641
4642         /*
4643          * wait for reply handlers to drop their request refs and
4644          * their inode/dcache refs
4645          */
4646         ceph_msgr_flush();
4647
4648         ceph_cleanup_quotarealms_inodes(mdsc);
4649 }
4650
4651 /*
4652  * wait for all write mds requests to flush.
4653  */
4654 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
4655 {
4656         struct ceph_mds_request *req = NULL, *nextreq;
4657         struct rb_node *n;
4658
4659         mutex_lock(&mdsc->mutex);
4660         dout("wait_unsafe_requests want %lld\n", want_tid);
4661 restart:
4662         req = __get_oldest_req(mdsc);
4663         while (req && req->r_tid <= want_tid) {
4664                 /* find next request */
4665                 n = rb_next(&req->r_node);
4666                 if (n)
4667                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
4668                 else
4669                         nextreq = NULL;
4670                 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
4671                     (req->r_op & CEPH_MDS_OP_WRITE)) {
4672                         /* write op */
4673                         ceph_mdsc_get_request(req);
4674                         if (nextreq)
4675                                 ceph_mdsc_get_request(nextreq);
4676                         mutex_unlock(&mdsc->mutex);
4677                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
4678                              req->r_tid, want_tid);
4679                         wait_for_completion(&req->r_safe_completion);
4680                         mutex_lock(&mdsc->mutex);
4681                         ceph_mdsc_put_request(req);
4682                         if (!nextreq)
4683                                 break;  /* next dne before, so we're done! */
4684                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
4685                                 /* next request was removed from tree */
4686                                 ceph_mdsc_put_request(nextreq);
4687                                 goto restart;
4688                         }
4689                         ceph_mdsc_put_request(nextreq);  /* won't go away */
4690                 }
4691                 req = nextreq;
4692         }
4693         mutex_unlock(&mdsc->mutex);
4694         dout("wait_unsafe_requests done\n");
4695 }
4696
4697 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
4698 {
4699         u64 want_tid, want_flush;
4700
4701         if (READ_ONCE(mdsc->fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN)
4702                 return;
4703
4704         dout("sync\n");
4705         mutex_lock(&mdsc->mutex);
4706         want_tid = mdsc->last_tid;
4707         mutex_unlock(&mdsc->mutex);
4708
4709         ceph_flush_dirty_caps(mdsc);
4710         spin_lock(&mdsc->cap_dirty_lock);
4711         want_flush = mdsc->last_cap_flush_tid;
4712         if (!list_empty(&mdsc->cap_flush_list)) {
4713                 struct ceph_cap_flush *cf =
4714                         list_last_entry(&mdsc->cap_flush_list,
4715                                         struct ceph_cap_flush, g_list);
4716                 cf->wake = true;
4717         }
4718         spin_unlock(&mdsc->cap_dirty_lock);
4719
4720         dout("sync want tid %lld flush_seq %lld\n",
4721              want_tid, want_flush);
4722
4723         wait_unsafe_requests(mdsc, want_tid);
4724         wait_caps_flush(mdsc, want_flush);
4725 }
4726
4727 /*
4728  * true if all sessions are closed, or we force unmount
4729  */
4730 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
4731 {
4732         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4733                 return true;
4734         return atomic_read(&mdsc->num_sessions) <= skipped;
4735 }
4736
4737 /*
4738  * called after sb is ro.
4739  */
4740 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
4741 {
4742         struct ceph_options *opts = mdsc->fsc->client->options;
4743         struct ceph_mds_session *session;
4744         int i;
4745         int skipped = 0;
4746
4747         dout("close_sessions\n");
4748
4749         /* close sessions */
4750         mutex_lock(&mdsc->mutex);
4751         for (i = 0; i < mdsc->max_sessions; i++) {
4752                 session = __ceph_lookup_mds_session(mdsc, i);
4753                 if (!session)
4754                         continue;
4755                 mutex_unlock(&mdsc->mutex);
4756                 mutex_lock(&session->s_mutex);
4757                 if (__close_session(mdsc, session) <= 0)
4758                         skipped++;
4759                 mutex_unlock(&session->s_mutex);
4760                 ceph_put_mds_session(session);
4761                 mutex_lock(&mdsc->mutex);
4762         }
4763         mutex_unlock(&mdsc->mutex);
4764
4765         dout("waiting for sessions to close\n");
4766         wait_event_timeout(mdsc->session_close_wq,
4767                            done_closing_sessions(mdsc, skipped),
4768                            ceph_timeout_jiffies(opts->mount_timeout));
4769
4770         /* tear down remaining sessions */
4771         mutex_lock(&mdsc->mutex);
4772         for (i = 0; i < mdsc->max_sessions; i++) {
4773                 if (mdsc->sessions[i]) {
4774                         session = ceph_get_mds_session(mdsc->sessions[i]);
4775                         __unregister_session(mdsc, session);
4776                         mutex_unlock(&mdsc->mutex);
4777                         mutex_lock(&session->s_mutex);
4778                         remove_session_caps(session);
4779                         mutex_unlock(&session->s_mutex);
4780                         ceph_put_mds_session(session);
4781                         mutex_lock(&mdsc->mutex);
4782                 }
4783         }
4784         WARN_ON(!list_empty(&mdsc->cap_delay_list));
4785         mutex_unlock(&mdsc->mutex);
4786
4787         ceph_cleanup_snapid_map(mdsc);
4788         ceph_cleanup_empty_realms(mdsc);
4789
4790         cancel_work_sync(&mdsc->cap_reclaim_work);
4791         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4792
4793         dout("stopped\n");
4794 }
4795
4796 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
4797 {
4798         struct ceph_mds_session *session;
4799         int mds;
4800
4801         dout("force umount\n");
4802
4803         mutex_lock(&mdsc->mutex);
4804         for (mds = 0; mds < mdsc->max_sessions; mds++) {
4805                 session = __ceph_lookup_mds_session(mdsc, mds);
4806                 if (!session)
4807                         continue;
4808
4809                 if (session->s_state == CEPH_MDS_SESSION_REJECTED)
4810                         __unregister_session(mdsc, session);
4811                 __wake_requests(mdsc, &session->s_waiting);
4812                 mutex_unlock(&mdsc->mutex);
4813
4814                 mutex_lock(&session->s_mutex);
4815                 __close_session(mdsc, session);
4816                 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
4817                         cleanup_session_requests(mdsc, session);
4818                         remove_session_caps(session);
4819                 }
4820                 mutex_unlock(&session->s_mutex);
4821                 ceph_put_mds_session(session);
4822
4823                 mutex_lock(&mdsc->mutex);
4824                 kick_requests(mdsc, mds);
4825         }
4826         __wake_requests(mdsc, &mdsc->waiting_for_map);
4827         mutex_unlock(&mdsc->mutex);
4828 }
4829
4830 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
4831 {
4832         dout("stop\n");
4833         /*
4834          * Make sure the delayed work stopped before releasing
4835          * the resources.
4836          *
4837          * Because the cancel_delayed_work_sync() will only
4838          * guarantee that the work finishes executing. But the
4839          * delayed work will re-arm itself again after that.
4840          */
4841         flush_delayed_work(&mdsc->delayed_work);
4842
4843         if (mdsc->mdsmap)
4844                 ceph_mdsmap_destroy(mdsc->mdsmap);
4845         kfree(mdsc->sessions);
4846         ceph_caps_finalize(mdsc);
4847         ceph_pool_perm_destroy(mdsc);
4848 }
4849
4850 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
4851 {
4852         struct ceph_mds_client *mdsc = fsc->mdsc;
4853         dout("mdsc_destroy %p\n", mdsc);
4854
4855         if (!mdsc)
4856                 return;
4857
4858         /* flush out any connection work with references to us */
4859         ceph_msgr_flush();
4860
4861         ceph_mdsc_stop(mdsc);
4862
4863         ceph_metric_destroy(&mdsc->metric);
4864
4865         flush_delayed_work(&mdsc->metric.delayed_work);
4866         fsc->mdsc = NULL;
4867         kfree(mdsc);
4868         dout("mdsc_destroy %p done\n", mdsc);
4869 }
4870
4871 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4872 {
4873         struct ceph_fs_client *fsc = mdsc->fsc;
4874         const char *mds_namespace = fsc->mount_options->mds_namespace;
4875         void *p = msg->front.iov_base;
4876         void *end = p + msg->front.iov_len;
4877         u32 epoch;
4878         u32 num_fs;
4879         u32 mount_fscid = (u32)-1;
4880         int err = -EINVAL;
4881
4882         ceph_decode_need(&p, end, sizeof(u32), bad);
4883         epoch = ceph_decode_32(&p);
4884
4885         dout("handle_fsmap epoch %u\n", epoch);
4886
4887         /* struct_v, struct_cv, map_len, epoch, legacy_client_fscid */
4888         ceph_decode_skip_n(&p, end, 2 + sizeof(u32) * 3, bad);
4889
4890         ceph_decode_32_safe(&p, end, num_fs, bad);
4891         while (num_fs-- > 0) {
4892                 void *info_p, *info_end;
4893                 u32 info_len;
4894                 u32 fscid, namelen;
4895
4896                 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4897                 p += 2;         // info_v, info_cv
4898                 info_len = ceph_decode_32(&p);
4899                 ceph_decode_need(&p, end, info_len, bad);
4900                 info_p = p;
4901                 info_end = p + info_len;
4902                 p = info_end;
4903
4904                 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
4905                 fscid = ceph_decode_32(&info_p);
4906                 namelen = ceph_decode_32(&info_p);
4907                 ceph_decode_need(&info_p, info_end, namelen, bad);
4908
4909                 if (mds_namespace &&
4910                     strlen(mds_namespace) == namelen &&
4911                     !strncmp(mds_namespace, (char *)info_p, namelen)) {
4912                         mount_fscid = fscid;
4913                         break;
4914                 }
4915         }
4916
4917         ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
4918         if (mount_fscid != (u32)-1) {
4919                 fsc->client->monc.fs_cluster_id = mount_fscid;
4920                 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
4921                                    0, true);
4922                 ceph_monc_renew_subs(&fsc->client->monc);
4923         } else {
4924                 err = -ENOENT;
4925                 goto err_out;
4926         }
4927         return;
4928
4929 bad:
4930         pr_err("error decoding fsmap\n");
4931 err_out:
4932         mutex_lock(&mdsc->mutex);
4933         mdsc->mdsmap_err = err;
4934         __wake_requests(mdsc, &mdsc->waiting_for_map);
4935         mutex_unlock(&mdsc->mutex);
4936 }
4937
4938 /*
4939  * handle mds map update.
4940  */
4941 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4942 {
4943         u32 epoch;
4944         u32 maplen;
4945         void *p = msg->front.iov_base;
4946         void *end = p + msg->front.iov_len;
4947         struct ceph_mdsmap *newmap, *oldmap;
4948         struct ceph_fsid fsid;
4949         int err = -EINVAL;
4950
4951         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
4952         ceph_decode_copy(&p, &fsid, sizeof(fsid));
4953         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
4954                 return;
4955         epoch = ceph_decode_32(&p);
4956         maplen = ceph_decode_32(&p);
4957         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
4958
4959         /* do we need it? */
4960         mutex_lock(&mdsc->mutex);
4961         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
4962                 dout("handle_map epoch %u <= our %u\n",
4963                      epoch, mdsc->mdsmap->m_epoch);
4964                 mutex_unlock(&mdsc->mutex);
4965                 return;
4966         }
4967
4968         newmap = ceph_mdsmap_decode(&p, end);
4969         if (IS_ERR(newmap)) {
4970                 err = PTR_ERR(newmap);
4971                 goto bad_unlock;
4972         }
4973
4974         /* swap into place */
4975         if (mdsc->mdsmap) {
4976                 oldmap = mdsc->mdsmap;
4977                 mdsc->mdsmap = newmap;
4978                 check_new_map(mdsc, newmap, oldmap);
4979                 ceph_mdsmap_destroy(oldmap);
4980         } else {
4981                 mdsc->mdsmap = newmap;  /* first mds map */
4982         }
4983         mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
4984                                         MAX_LFS_FILESIZE);
4985
4986         __wake_requests(mdsc, &mdsc->waiting_for_map);
4987         ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
4988                           mdsc->mdsmap->m_epoch);
4989
4990         mutex_unlock(&mdsc->mutex);
4991         schedule_delayed(mdsc);
4992         return;
4993
4994 bad_unlock:
4995         mutex_unlock(&mdsc->mutex);
4996 bad:
4997         pr_err("error decoding mdsmap %d\n", err);
4998         return;
4999 }
5000
5001 static struct ceph_connection *con_get(struct ceph_connection *con)
5002 {
5003         struct ceph_mds_session *s = con->private;
5004
5005         if (ceph_get_mds_session(s))
5006                 return con;
5007         return NULL;
5008 }
5009
5010 static void con_put(struct ceph_connection *con)
5011 {
5012         struct ceph_mds_session *s = con->private;
5013
5014         ceph_put_mds_session(s);
5015 }
5016
5017 /*
5018  * if the client is unresponsive for long enough, the mds will kill
5019  * the session entirely.
5020  */
5021 static void peer_reset(struct ceph_connection *con)
5022 {
5023         struct ceph_mds_session *s = con->private;
5024         struct ceph_mds_client *mdsc = s->s_mdsc;
5025
5026         pr_warn("mds%d closed our session\n", s->s_mds);
5027         send_mds_reconnect(mdsc, s);
5028 }
5029
5030 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
5031 {
5032         struct ceph_mds_session *s = con->private;
5033         struct ceph_mds_client *mdsc = s->s_mdsc;
5034         int type = le16_to_cpu(msg->hdr.type);
5035
5036         mutex_lock(&mdsc->mutex);
5037         if (__verify_registered_session(mdsc, s) < 0) {
5038                 mutex_unlock(&mdsc->mutex);
5039                 goto out;
5040         }
5041         mutex_unlock(&mdsc->mutex);
5042
5043         switch (type) {
5044         case CEPH_MSG_MDS_MAP:
5045                 ceph_mdsc_handle_mdsmap(mdsc, msg);
5046                 break;
5047         case CEPH_MSG_FS_MAP_USER:
5048                 ceph_mdsc_handle_fsmap(mdsc, msg);
5049                 break;
5050         case CEPH_MSG_CLIENT_SESSION:
5051                 handle_session(s, msg);
5052                 break;
5053         case CEPH_MSG_CLIENT_REPLY:
5054                 handle_reply(s, msg);
5055                 break;
5056         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
5057                 handle_forward(mdsc, s, msg);
5058                 break;
5059         case CEPH_MSG_CLIENT_CAPS:
5060                 ceph_handle_caps(s, msg);
5061                 break;
5062         case CEPH_MSG_CLIENT_SNAP:
5063                 ceph_handle_snap(mdsc, s, msg);
5064                 break;
5065         case CEPH_MSG_CLIENT_LEASE:
5066                 handle_lease(mdsc, s, msg);
5067                 break;
5068         case CEPH_MSG_CLIENT_QUOTA:
5069                 ceph_handle_quota(mdsc, s, msg);
5070                 break;
5071
5072         default:
5073                 pr_err("received unknown message type %d %s\n", type,
5074                        ceph_msg_type_name(type));
5075         }
5076 out:
5077         ceph_msg_put(msg);
5078 }
5079
5080 /*
5081  * authentication
5082  */
5083
5084 /*
5085  * Note: returned pointer is the address of a structure that's
5086  * managed separately.  Caller must *not* attempt to free it.
5087  */
5088 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
5089                                         int *proto, int force_new)
5090 {
5091         struct ceph_mds_session *s = con->private;
5092         struct ceph_mds_client *mdsc = s->s_mdsc;
5093         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5094         struct ceph_auth_handshake *auth = &s->s_auth;
5095
5096         if (force_new && auth->authorizer) {
5097                 ceph_auth_destroy_authorizer(auth->authorizer);
5098                 auth->authorizer = NULL;
5099         }
5100         if (!auth->authorizer) {
5101                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
5102                                                       auth);
5103                 if (ret)
5104                         return ERR_PTR(ret);
5105         } else {
5106                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
5107                                                       auth);
5108                 if (ret)
5109                         return ERR_PTR(ret);
5110         }
5111         *proto = ac->protocol;
5112
5113         return auth;
5114 }
5115
5116 static int add_authorizer_challenge(struct ceph_connection *con,
5117                                     void *challenge_buf, int challenge_buf_len)
5118 {
5119         struct ceph_mds_session *s = con->private;
5120         struct ceph_mds_client *mdsc = s->s_mdsc;
5121         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5122
5123         return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
5124                                             challenge_buf, challenge_buf_len);
5125 }
5126
5127 static int verify_authorizer_reply(struct ceph_connection *con)
5128 {
5129         struct ceph_mds_session *s = con->private;
5130         struct ceph_mds_client *mdsc = s->s_mdsc;
5131         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5132
5133         return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
5134 }
5135
5136 static int invalidate_authorizer(struct ceph_connection *con)
5137 {
5138         struct ceph_mds_session *s = con->private;
5139         struct ceph_mds_client *mdsc = s->s_mdsc;
5140         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5141
5142         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
5143
5144         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
5145 }
5146
5147 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
5148                                 struct ceph_msg_header *hdr, int *skip)
5149 {
5150         struct ceph_msg *msg;
5151         int type = (int) le16_to_cpu(hdr->type);
5152         int front_len = (int) le32_to_cpu(hdr->front_len);
5153
5154         if (con->in_msg)
5155                 return con->in_msg;
5156
5157         *skip = 0;
5158         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
5159         if (!msg) {
5160                 pr_err("unable to allocate msg type %d len %d\n",
5161                        type, front_len);
5162                 return NULL;
5163         }
5164
5165         return msg;
5166 }
5167
5168 static int mds_sign_message(struct ceph_msg *msg)
5169 {
5170        struct ceph_mds_session *s = msg->con->private;
5171        struct ceph_auth_handshake *auth = &s->s_auth;
5172
5173        return ceph_auth_sign_message(auth, msg);
5174 }
5175
5176 static int mds_check_message_signature(struct ceph_msg *msg)
5177 {
5178        struct ceph_mds_session *s = msg->con->private;
5179        struct ceph_auth_handshake *auth = &s->s_auth;
5180
5181        return ceph_auth_check_message_signature(auth, msg);
5182 }
5183
5184 static const struct ceph_connection_operations mds_con_ops = {
5185         .get = con_get,
5186         .put = con_put,
5187         .dispatch = dispatch,
5188         .get_authorizer = get_authorizer,
5189         .add_authorizer_challenge = add_authorizer_challenge,
5190         .verify_authorizer_reply = verify_authorizer_reply,
5191         .invalidate_authorizer = invalidate_authorizer,
5192         .peer_reset = peer_reset,
5193         .alloc_msg = mds_alloc_msg,
5194         .sign_message = mds_sign_message,
5195         .check_message_signature = mds_check_message_signature,
5196 };
5197
5198 /* eof */