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  49  * are no longer valid.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                         u8 struct_v, struct_compat;
 520                         u32 len;
 521
 522                         info->has_create_ino = true;
 523                         ceph_decode_8_safe(p, end, struct_v, bad);
 524                         ceph_decode_8_safe(p, end, struct_compat, bad);
 525                         ceph_decode_32_safe(p, end, len, bad);
 526                         ceph_decode_64_safe(p, end, info->ino, bad);
 527                         ret = ceph_parse_deleg_inos(p, end, s);
 528                         if (ret)
 529                                 return ret;
 530                 } else {
 531                         /* legacy */
 532                         ceph_decode_64_safe(p, end, info->ino, bad);
 533                         info->has_create_ino = true;
 534                 }
 535         } else {
 536                 if (*p != end)
 537                         goto bad;
 538         }
 539
 540         /* Skip over any unrecognized fields */
 541         *p = end;
 542         return 0;
 543 bad:
 544         return -EIO;
 545 }
 546
 547 /*
 548  * parse extra results
 549  */
 550 static int parse_reply_info_extra(void **p, void *end,
 551                                   struct ceph_mds_reply_info_parsed *info,
 552                                   u64 features, struct ceph_mds_session *s)
 553 {
 554         u32 op = le32_to_cpu(info->head->op);
 555
 556         if (op == CEPH_MDS_OP_GETFILELOCK)
 557                 return parse_reply_info_filelock(p, end, info, features);
 558         else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
 559                 return parse_reply_info_readdir(p, end, info, features);
 560         else if (op == CEPH_MDS_OP_CREATE)
 561                 return parse_reply_info_create(p, end, info, features, s);
 562         else
 563                 return -EIO;
 564 }
 565
 566 /*
 567  * parse entire mds reply
 568  */
 569 static int parse_reply_info(struct ceph_mds_session *s, struct ceph_msg *msg,
 570                             struct ceph_mds_reply_info_parsed *info,
 571                             u64 features)
 572 {
 573         void *p, *end;
 574         u32 len;
 575         int err;
 576
 577         info->head = msg->front.iov_base;
 578         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 579         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 580
 581         /* trace */
 582         ceph_decode_32_safe(&p, end, len, bad);
 583         if (len > 0) {
 584                 ceph_decode_need(&p, end, len, bad);
 585                 err = parse_reply_info_trace(&p, p+len, info, features);
 586                 if (err < 0)
 587                         goto out_bad;
 588         }
 589
 590         /* extra */
 591         ceph_decode_32_safe(&p, end, len, bad);
 592         if (len > 0) {
 593                 ceph_decode_need(&p, end, len, bad);
 594                 err = parse_reply_info_extra(&p, p+len, info, features, s);
 595                 if (err < 0)
 596                         goto out_bad;
 597         }
 598
 599         /* snap blob */
 600         ceph_decode_32_safe(&p, end, len, bad);
 601         info->snapblob_len = len;
 602         info->snapblob = p;
 603         p += len;
 604
 605         if (p != end)
 606                 goto bad;
 607         return 0;
 608
 609 bad:
 610         err = -EIO;
 611 out_bad:
 612         pr_err("mds parse_reply err %d\n", err);
 613         return err;
 614 }
 615
 616 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
 617 {
 618         if (!info->dir_entries)
 619                 return;
 620         free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
 621 }
 622
 623
 624 /*
 625  * sessions
 626  */
 627 const char *ceph_session_state_name(int s)
 628 {
 629         switch (s) {
 630         case CEPH_MDS_SESSION_NEW: return "new";
 631         case CEPH_MDS_SESSION_OPENING: return "opening";
 632         case CEPH_MDS_SESSION_OPEN: return "open";
 633         case CEPH_MDS_SESSION_HUNG: return "hung";
 634         case CEPH_MDS_SESSION_CLOSING: return "closing";
 635         case CEPH_MDS_SESSION_CLOSED: return "closed";
 636         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
 637         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
 638         case CEPH_MDS_SESSION_REJECTED: return "rejected";
 639         default: return "???";
 640         }
 641 }
 642
 643 struct ceph_mds_session *ceph_get_mds_session(struct ceph_mds_session *s)
 644 {
 645         if (refcount_inc_not_zero(&s->s_ref)) {
 646                 dout("mdsc get_session %p %d -> %d\n", s,
 647                      refcount_read(&s->s_ref)-1, refcount_read(&s->s_ref));
 648                 return s;
 649         } else {
 650                 dout("mdsc get_session %p 0 -- FAIL\n", s);
 651                 return NULL;
 652         }
 653 }
 654
 655 void ceph_put_mds_session(struct ceph_mds_session *s)
 656 {
 657         dout("mdsc put_session %p %d -> %d\n", s,
 658              refcount_read(&s->s_ref), refcount_read(&s->s_ref)-1);
 659         if (refcount_dec_and_test(&s->s_ref)) {
 660                 if (s->s_auth.authorizer)
 661                         ceph_auth_destroy_authorizer(s->s_auth.authorizer);
 662                 WARN_ON(mutex_is_locked(&s->s_mutex));
 663                 xa_destroy(&s->s_delegated_inos);
 664                 kfree(s);
 665         }
 666 }
 667
 668 /*
 669  * called under mdsc->mutex
 670  */
 671 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
 672                                                    int mds)
 673 {
 674         if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
 675                 return NULL;
 676         return ceph_get_mds_session(mdsc->sessions[mds]);
 677 }
 678
 679 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
 680 {
 681         if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
 682                 return false;
 683         else
 684                 return true;
 685 }
 686
 687 static int __verify_registered_session(struct ceph_mds_client *mdsc,
 688                                        struct ceph_mds_session *s)
 689 {
 690         if (s->s_mds >= mdsc->max_sessions ||
 691             mdsc->sessions[s->s_mds] != s)
 692                 return -ENOENT;
 693         return 0;
 694 }
 695
 696 /*
 697  * create+register a new session for given mds.
 698  * called under mdsc->mutex.
 699  */
 700 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 701                                                  int mds)
 702 {
 703         struct ceph_mds_session *s;
 704
 705         if (mds >= mdsc->mdsmap->possible_max_rank)
 706                 return ERR_PTR(-EINVAL);
 707
 708         s = kzalloc(sizeof(*s), GFP_NOFS);
 709         if (!s)
 710                 return ERR_PTR(-ENOMEM);
 711
 712         if (mds >= mdsc->max_sessions) {
 713                 int newmax = 1 << get_count_order(mds + 1);
 714                 struct ceph_mds_session **sa;
 715
 716                 dout("%s: realloc to %d\n", __func__, newmax);
 717                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 718                 if (!sa)
 719                         goto fail_realloc;
 720                 if (mdsc->sessions) {
 721                         memcpy(sa, mdsc->sessions,
 722                                mdsc->max_sessions * sizeof(void *));
 723                         kfree(mdsc->sessions);
 724                 }
 725                 mdsc->sessions = sa;
 726                 mdsc->max_sessions = newmax;
 727         }
 728
 729         dout("%s: mds%d\n", __func__, mds);
 730         s->s_mdsc = mdsc;
 731         s->s_mds = mds;
 732         s->s_state = CEPH_MDS_SESSION_NEW;
 733         s->s_ttl = 0;
 734         s->s_seq = 0;
 735         mutex_init(&s->s_mutex);
 736
 737         ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
 738
 739         spin_lock_init(&s->s_gen_ttl_lock);
 740         s->s_cap_gen = 1;
 741         s->s_cap_ttl = jiffies - 1;
 742
 743         spin_lock_init(&s->s_cap_lock);
 744         s->s_renew_requested = 0;
 745         s->s_renew_seq = 0;
 746         INIT_LIST_HEAD(&s->s_caps);
 747         s->s_nr_caps = 0;
 748         refcount_set(&s->s_ref, 1);
 749         INIT_LIST_HEAD(&s->s_waiting);
 750         INIT_LIST_HEAD(&s->s_unsafe);
 751         xa_init(&s->s_delegated_inos);
 752         s->s_num_cap_releases = 0;
 753         s->s_cap_reconnect = 0;
 754         s->s_cap_iterator = NULL;
 755         INIT_LIST_HEAD(&s->s_cap_releases);
 756         INIT_WORK(&s->s_cap_release_work, ceph_cap_release_work);
 757
 758         INIT_LIST_HEAD(&s->s_cap_dirty);
 759         INIT_LIST_HEAD(&s->s_cap_flushing);
 760
 761         mdsc->sessions[mds] = s;
 762         atomic_inc(&mdsc->num_sessions);
 763         refcount_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
 764
 765         ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
 766                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
 767
 768         return s;
 769
 770 fail_realloc:
 771         kfree(s);
 772         return ERR_PTR(-ENOMEM);
 773 }
 774
 775 /*
 776  * called under mdsc->mutex
 777  */
 778 static void __unregister_session(struct ceph_mds_client *mdsc,
 779                                struct ceph_mds_session *s)
 780 {
 781         dout("__unregister_session mds%d %p\n", s->s_mds, s);
 782         BUG_ON(mdsc->sessions[s->s_mds] != s);
 783         mdsc->sessions[s->s_mds] = NULL;
 784         ceph_con_close(&s->s_con);
 785         ceph_put_mds_session(s);
 786         atomic_dec(&mdsc->num_sessions);
 787 }
 788
 789 /*
 790  * drop session refs in request.
 791  *
 792  * should be last request ref, or hold mdsc->mutex
 793  */
 794 static void put_request_session(struct ceph_mds_request *req)
 795 {
 796         if (req->r_session) {
 797                 ceph_put_mds_session(req->r_session);
 798                 req->r_session = NULL;
 799         }
 800 }
 801
 802 void ceph_mdsc_release_request(struct kref *kref)
 803 {
 804         struct ceph_mds_request *req = container_of(kref,
 805                                                     struct ceph_mds_request,
 806                                                     r_kref);
 807         ceph_mdsc_release_dir_caps_no_check(req);
 808         destroy_reply_info(&req->r_reply_info);
 809         if (req->r_request)
 810                 ceph_msg_put(req->r_request);
 811         if (req->r_reply)
 812                 ceph_msg_put(req->r_reply);
 813         if (req->r_inode) {
 814                 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
 815                 /* avoid calling iput_final() in mds dispatch threads */
 816                 ceph_async_iput(req->r_inode);
 817         }
 818         if (req->r_parent) {
 819                 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
 820                 ceph_async_iput(req->r_parent);
 821         }
 822         ceph_async_iput(req->r_target_inode);
 823         if (req->r_dentry)
 824                 dput(req->r_dentry);
 825         if (req->r_old_dentry)
 826                 dput(req->r_old_dentry);
 827         if (req->r_old_dentry_dir) {
 828                 /*
 829                  * track (and drop pins for) r_old_dentry_dir
 830                  * separately, since r_old_dentry's d_parent may have
 831                  * changed between the dir mutex being dropped and
 832                  * this request being freed.
 833                  */
 834                 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
 835                                   CEPH_CAP_PIN);
 836                 ceph_async_iput(req->r_old_dentry_dir);
 837         }
 838         kfree(req->r_path1);
 839         kfree(req->r_path2);
 840         if (req->r_pagelist)
 841                 ceph_pagelist_release(req->r_pagelist);
 842         put_request_session(req);
 843         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
 844         WARN_ON_ONCE(!list_empty(&req->r_wait));
 845         kmem_cache_free(ceph_mds_request_cachep, req);
 846 }
 847
 848 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
 849
 850 /*
 851  * lookup session, bump ref if found.
 852  *
 853  * called under mdsc->mutex.
 854  */
 855 static struct ceph_mds_request *
 856 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
 857 {
 858         struct ceph_mds_request *req;
 859
 860         req = lookup_request(&mdsc->request_tree, tid);
 861         if (req)
 862                 ceph_mdsc_get_request(req);
 863
 864         return req;
 865 }
 866
 867 /*
 868  * Register an in-flight request, and assign a tid.  Link to directory
 869  * are modifying (if any).
 870  *
 871  * Called under mdsc->mutex.
 872  */
 873 static void __register_request(struct ceph_mds_client *mdsc,
 874                                struct ceph_mds_request *req,
 875                                struct inode *dir)
 876 {
 877         int ret = 0;
 878
 879         req->r_tid = ++mdsc->last_tid;
 880         if (req->r_num_caps) {
 881                 ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
 882                                         req->r_num_caps);
 883                 if (ret < 0) {
 884                         pr_err("__register_request %p "
 885                                "failed to reserve caps: %d\n", req, ret);
 886                         /* set req->r_err to fail early from __do_request */
 887                         req->r_err = ret;
 888                         return;
 889                 }
 890         }
 891         dout("__register_request %p tid %lld\n", req, req->r_tid);
 892         ceph_mdsc_get_request(req);
 893         insert_request(&mdsc->request_tree, req);
 894
 895         req->r_uid = current_fsuid();
 896         req->r_gid = current_fsgid();
 897
 898         if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
 899                 mdsc->oldest_tid = req->r_tid;
 900
 901         if (dir) {
 902                 struct ceph_inode_info *ci = ceph_inode(dir);
 903
 904                 ihold(dir);
 905                 req->r_unsafe_dir = dir;
 906                 spin_lock(&ci->i_unsafe_lock);
 907                 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
 908                 spin_unlock(&ci->i_unsafe_lock);
 909         }
 910 }
 911
 912 static void __unregister_request(struct ceph_mds_client *mdsc,
 913                                  struct ceph_mds_request *req)
 914 {
 915         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 916
 917         /* Never leave an unregistered request on an unsafe list! */
 918         list_del_init(&req->r_unsafe_item);
 919
 920         if (req->r_tid == mdsc->oldest_tid) {
 921                 struct rb_node *p = rb_next(&req->r_node);
 922                 mdsc->oldest_tid = 0;
 923                 while (p) {
 924                         struct ceph_mds_request *next_req =
 925                                 rb_entry(p, struct ceph_mds_request, r_node);
 926                         if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
 927                                 mdsc->oldest_tid = next_req->r_tid;
 928                                 break;
 929                         }
 930                         p = rb_next(p);
 931                 }
 932         }
 933
 934         erase_request(&mdsc->request_tree, req);
 935
 936         if (req->r_unsafe_dir) {
 937                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 938                 spin_lock(&ci->i_unsafe_lock);
 939                 list_del_init(&req->r_unsafe_dir_item);
 940                 spin_unlock(&ci->i_unsafe_lock);
 941         }
 942         if (req->r_target_inode &&
 943             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
 944                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
 945                 spin_lock(&ci->i_unsafe_lock);
 946                 list_del_init(&req->r_unsafe_target_item);
 947                 spin_unlock(&ci->i_unsafe_lock);
 948         }
 949
 950         if (req->r_unsafe_dir) {
 951                 /* avoid calling iput_final() in mds dispatch threads */
 952                 ceph_async_iput(req->r_unsafe_dir);
 953                 req->r_unsafe_dir = NULL;
 954         }
 955
 956         complete_all(&req->r_safe_completion);
 957
 958         ceph_mdsc_put_request(req);
 959 }
 960
 961 /*
 962  * Walk back up the dentry tree until we hit a dentry representing a
 963  * non-snapshot inode. We do this using the rcu_read_lock (which must be held
 964  * when calling this) to ensure that the objects won't disappear while we're
 965  * working with them. Once we hit a candidate dentry, we attempt to take a
 966  * reference to it, and return that as the result.
 967  */
 968 static struct inode *get_nonsnap_parent(struct dentry *dentry)
 969 {
 970         struct inode *inode = NULL;
 971
 972         while (dentry && !IS_ROOT(dentry)) {
 973                 inode = d_inode_rcu(dentry);
 974                 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
 975                         break;
 976                 dentry = dentry->d_parent;
 977         }
 978         if (inode)
 979                 inode = igrab(inode);
 980         return inode;
 981 }
 982
 983 /*
 984  * Choose mds to send request to next.  If there is a hint set in the
 985  * request (e.g., due to a prior forward hint from the mds), use that.
 986  * Otherwise, consult frag tree and/or caps to identify the
 987  * appropriate mds.  If all else fails, choose randomly.
 988  *
 989  * Called under mdsc->mutex.
 990  */
 991 static int __choose_mds(struct ceph_mds_client *mdsc,
 992                         struct ceph_mds_request *req,
 993                         bool *random)
 994 {
 995         struct inode *inode;
 996         struct ceph_inode_info *ci;
 997         struct ceph_cap *cap;
 998         int mode = req->r_direct_mode;
 999         int mds = -1;
1000         u32 hash = req->r_direct_hash;
1001         bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
1002
1003         if (random)
1004                 *random = false;
1005
1006         /*
1007          * is there a specific mds we should try?  ignore hint if we have
1008          * no session and the mds is not up (active or recovering).
1009          */
1010         if (req->r_resend_mds >= 0 &&
1011             (__have_session(mdsc, req->r_resend_mds) ||
1012              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
1013                 dout("%s using resend_mds mds%d\n", __func__,
1014                      req->r_resend_mds);
1015                 return req->r_resend_mds;
1016         }
1017
1018         if (mode == USE_RANDOM_MDS)
1019                 goto random;
1020
1021         inode = NULL;
1022         if (req->r_inode) {
1023                 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
1024                         inode = req->r_inode;
1025                         ihold(inode);
1026                 } else {
1027                         /* req->r_dentry is non-null for LSSNAP request */
1028                         rcu_read_lock();
1029                         inode = get_nonsnap_parent(req->r_dentry);
1030                         rcu_read_unlock();
1031                         dout("%s using snapdir's parent %p\n", __func__, inode);
1032                 }
1033         } else if (req->r_dentry) {
1034                 /* ignore race with rename; old or new d_parent is okay */
1035                 struct dentry *parent;
1036                 struct inode *dir;
1037
1038                 rcu_read_lock();
1039                 parent = READ_ONCE(req->r_dentry->d_parent);
1040                 dir = req->r_parent ? : d_inode_rcu(parent);
1041
1042                 if (!dir || dir->i_sb != mdsc->fsc->sb) {
1043                         /*  not this fs or parent went negative */
1044                         inode = d_inode(req->r_dentry);
1045                         if (inode)
1046                                 ihold(inode);
1047                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
1048                         /* direct snapped/virtual snapdir requests
1049                          * based on parent dir inode */
1050                         inode = get_nonsnap_parent(parent);
1051                         dout("%s using nonsnap parent %p\n", __func__, inode);
1052                 } else {
1053                         /* dentry target */
1054                         inode = d_inode(req->r_dentry);
1055                         if (!inode || mode == USE_AUTH_MDS) {
1056                                 /* dir + name */
1057                                 inode = igrab(dir);
1058                                 hash = ceph_dentry_hash(dir, req->r_dentry);
1059                                 is_hash = true;
1060                         } else {
1061                                 ihold(inode);
1062                         }
1063                 }
1064                 rcu_read_unlock();
1065         }
1066
1067         dout("%s %p is_hash=%d (0x%x) mode %d\n", __func__, inode, (int)is_hash,
1068              hash, mode);
1069         if (!inode)
1070                 goto random;
1071         ci = ceph_inode(inode);
1072
1073         if (is_hash && S_ISDIR(inode->i_mode)) {
1074                 struct ceph_inode_frag frag;
1075                 int found;
1076
1077                 ceph_choose_frag(ci, hash, &frag, &found);
1078                 if (found) {
1079                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
1080                                 u8 r;
1081
1082                                 /* choose a random replica */
1083                                 get_random_bytes(&r, 1);
1084                                 r %= frag.ndist;
1085                                 mds = frag.dist[r];
1086                                 dout("%s %p %llx.%llx frag %u mds%d (%d/%d)\n",
1087                                      __func__, inode, ceph_vinop(inode),
1088                                      frag.frag, mds, (int)r, frag.ndist);
1089                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1090                                     CEPH_MDS_STATE_ACTIVE &&
1091                                     !ceph_mdsmap_is_laggy(mdsc->mdsmap, mds))
1092                                         goto out;
1093                         }
1094
1095                         /* since this file/dir wasn't known to be
1096                          * replicated, then we want to look for the
1097                          * authoritative mds. */
1098                         if (frag.mds >= 0) {
1099                                 /* choose auth mds */
1100                                 mds = frag.mds;
1101                                 dout("%s %p %llx.%llx frag %u mds%d (auth)\n",
1102                                      __func__, inode, ceph_vinop(inode),
1103                                      frag.frag, mds);
1104                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1105                                     CEPH_MDS_STATE_ACTIVE) {
1106                                         if (mode == USE_ANY_MDS &&
1107                                             !ceph_mdsmap_is_laggy(mdsc->mdsmap,
1108                                                                   mds))
1109                                                 goto out;
1110                                 }
1111                         }
1112                         mode = USE_AUTH_MDS;
1113                 }
1114         }
1115
1116         spin_lock(&ci->i_ceph_lock);
1117         cap = NULL;
1118         if (mode == USE_AUTH_MDS)
1119                 cap = ci->i_auth_cap;
1120         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
1121                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
1122         if (!cap) {
1123                 spin_unlock(&ci->i_ceph_lock);
1124                 ceph_async_iput(inode);
1125                 goto random;
1126         }
1127         mds = cap->session->s_mds;
1128         dout("%s %p %llx.%llx mds%d (%scap %p)\n", __func__,
1129              inode, ceph_vinop(inode), mds,
1130              cap == ci->i_auth_cap ? "auth " : "", cap);
1131         spin_unlock(&ci->i_ceph_lock);
1132 out:
1133         /* avoid calling iput_final() while holding mdsc->mutex or
1134          * in mds dispatch threads */
1135         ceph_async_iput(inode);
1136         return mds;
1137
1138 random:
1139         if (random)
1140                 *random = true;
1141
1142         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
1143         dout("%s chose random mds%d\n", __func__, mds);
1144         return mds;
1145 }
1146
1147
1148 /*
1149  * session messages
1150  */
1151 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
1152 {
1153         struct ceph_msg *msg;
1154         struct ceph_mds_session_head *h;
1155
1156         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
1157                            false);
1158         if (!msg) {
1159                 pr_err("create_session_msg ENOMEM creating msg\n");
1160                 return NULL;
1161         }
1162         h = msg->front.iov_base;
1163         h->op = cpu_to_le32(op);
1164         h->seq = cpu_to_le64(seq);
1165
1166         return msg;
1167 }
1168
1169 static const unsigned char feature_bits[] = CEPHFS_FEATURES_CLIENT_SUPPORTED;
1170 #define FEATURE_BYTES(c) (DIV_ROUND_UP((size_t)feature_bits[c - 1] + 1, 64) * 8)
1171 static int encode_supported_features(void **p, void *end)
1172 {
1173         static const size_t count = ARRAY_SIZE(feature_bits);
1174
1175         if (count > 0) {
1176                 size_t i;
1177                 size_t size = FEATURE_BYTES(count);
1178
1179                 if (WARN_ON_ONCE(*p + 4 + size > end))
1180                         return -ERANGE;
1181
1182                 ceph_encode_32(p, size);
1183                 memset(*p, 0, size);
1184                 for (i = 0; i < count; i++)
1185                         ((unsigned char*)(*p))[i / 8] |= BIT(feature_bits[i] % 8);
1186                 *p += size;
1187         } else {
1188                 if (WARN_ON_ONCE(*p + 4 > end))
1189                         return -ERANGE;
1190
1191                 ceph_encode_32(p, 0);
1192         }
1193
1194         return 0;
1195 }
1196
1197 /*
1198  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
1199  * to include additional client metadata fields.
1200  */
1201 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
1202 {
1203         struct ceph_msg *msg;
1204         struct ceph_mds_session_head *h;
1205         int i = -1;
1206         int extra_bytes = 0;
1207         int metadata_key_count = 0;
1208         struct ceph_options *opt = mdsc->fsc->client->options;
1209         struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
1210         size_t size, count;
1211         void *p, *end;
1212         int ret;
1213
1214         const char* metadata[][2] = {
1215                 {"hostname", mdsc->nodename},
1216                 {"kernel_version", init_utsname()->release},
1217                 {"entity_id", opt->name ? : ""},
1218                 {"root", fsopt->server_path ? : "/"},
1219                 {NULL, NULL}
1220         };
1221
1222         /* Calculate serialized length of metadata */
1223         extra_bytes = 4;  /* map length */
1224         for (i = 0; metadata[i][0]; ++i) {
1225                 extra_bytes += 8 + strlen(metadata[i][0]) +
1226                         strlen(metadata[i][1]);
1227                 metadata_key_count++;
1228         }
1229
1230         /* supported feature */
1231         size = 0;
1232         count = ARRAY_SIZE(feature_bits);
1233         if (count > 0)
1234                 size = FEATURE_BYTES(count);
1235         extra_bytes += 4 + size;
1236
1237         /* Allocate the message */
1238         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + extra_bytes,
1239                            GFP_NOFS, false);
1240         if (!msg) {
1241                 pr_err("create_session_msg ENOMEM creating msg\n");
1242                 return ERR_PTR(-ENOMEM);
1243         }
1244         p = msg->front.iov_base;
1245         end = p + msg->front.iov_len;
1246
1247         h = p;
1248         h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
1249         h->seq = cpu_to_le64(seq);
1250
1251         /*
1252          * Serialize client metadata into waiting buffer space, using
1253          * the format that userspace expects for map<string, string>
1254          *
1255          * ClientSession messages with metadata are v3
1256          */
1257         msg->hdr.version = cpu_to_le16(3);
1258         msg->hdr.compat_version = cpu_to_le16(1);
1259
1260         /* The write pointer, following the session_head structure */
1261         p += sizeof(*h);
1262
1263         /* Number of entries in the map */
1264         ceph_encode_32(&p, metadata_key_count);
1265
1266         /* Two length-prefixed strings for each entry in the map */
1267         for (i = 0; metadata[i][0]; ++i) {
1268                 size_t const key_len = strlen(metadata[i][0]);
1269                 size_t const val_len = strlen(metadata[i][1]);
1270
1271                 ceph_encode_32(&p, key_len);
1272                 memcpy(p, metadata[i][0], key_len);
1273                 p += key_len;
1274                 ceph_encode_32(&p, val_len);
1275                 memcpy(p, metadata[i][1], val_len);
1276                 p += val_len;
1277         }
1278
1279         ret = encode_supported_features(&p, end);
1280         if (ret) {
1281                 pr_err("encode_supported_features failed!\n");
1282                 ceph_msg_put(msg);
1283                 return ERR_PTR(ret);
1284         }
1285
1286         msg->front.iov_len = p - msg->front.iov_base;
1287         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1288
1289         return msg;
1290 }
1291
1292 /*
1293  * send session open request.
1294  *
1295  * called under mdsc->mutex
1296  */
1297 static int __open_session(struct ceph_mds_client *mdsc,
1298                           struct ceph_mds_session *session)
1299 {
1300         struct ceph_msg *msg;
1301         int mstate;
1302         int mds = session->s_mds;
1303
1304         /* wait for mds to go active? */
1305         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
1306         dout("open_session to mds%d (%s)\n", mds,
1307              ceph_mds_state_name(mstate));
1308         session->s_state = CEPH_MDS_SESSION_OPENING;
1309         session->s_renew_requested = jiffies;
1310
1311         /* send connect message */
1312         msg = create_session_open_msg(mdsc, session->s_seq);
1313         if (IS_ERR(msg))
1314                 return PTR_ERR(msg);
1315         ceph_con_send(&session->s_con, msg);
1316         return 0;
1317 }
1318
1319 /*
1320  * open sessions for any export targets for the given mds
1321  *
1322  * called under mdsc->mutex
1323  */
1324 static struct ceph_mds_session *
1325 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
1326 {
1327         struct ceph_mds_session *session;
1328         int ret;
1329
1330         session = __ceph_lookup_mds_session(mdsc, target);
1331         if (!session) {
1332                 session = register_session(mdsc, target);
1333                 if (IS_ERR(session))
1334                         return session;
1335         }
1336         if (session->s_state == CEPH_MDS_SESSION_NEW ||
1337             session->s_state == CEPH_MDS_SESSION_CLOSING) {
1338                 ret = __open_session(mdsc, session);
1339                 if (ret)
1340                         return ERR_PTR(ret);
1341         }
1342
1343         return session;
1344 }
1345
1346 struct ceph_mds_session *
1347 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1348 {
1349         struct ceph_mds_session *session;
1350
1351         dout("open_export_target_session to mds%d\n", target);
1352
1353         mutex_lock(&mdsc->mutex);
1354         session = __open_export_target_session(mdsc, target);
1355         mutex_unlock(&mdsc->mutex);
1356
1357         return session;
1358 }
1359
1360 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1361                                           struct ceph_mds_session *session)
1362 {
1363         struct ceph_mds_info *mi;
1364         struct ceph_mds_session *ts;
1365         int i, mds = session->s_mds;
1366
1367         if (mds >= mdsc->mdsmap->possible_max_rank)
1368                 return;
1369
1370         mi = &mdsc->mdsmap->m_info[mds];
1371         dout("open_export_target_sessions for mds%d (%d targets)\n",
1372              session->s_mds, mi->num_export_targets);
1373
1374         for (i = 0; i < mi->num_export_targets; i++) {
1375                 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1376                 if (!IS_ERR(ts))
1377                         ceph_put_mds_session(ts);
1378         }
1379 }
1380
1381 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1382                                            struct ceph_mds_session *session)
1383 {
1384         mutex_lock(&mdsc->mutex);
1385         __open_export_target_sessions(mdsc, session);
1386         mutex_unlock(&mdsc->mutex);
1387 }
1388
1389 /*
1390  * session caps
1391  */
1392
1393 static void detach_cap_releases(struct ceph_mds_session *session,
1394                                 struct list_head *target)
1395 {
1396         lockdep_assert_held(&session->s_cap_lock);
1397
1398         list_splice_init(&session->s_cap_releases, target);
1399         session->s_num_cap_releases = 0;
1400         dout("dispose_cap_releases mds%d\n", session->s_mds);
1401 }
1402
1403 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1404                                  struct list_head *dispose)
1405 {
1406         while (!list_empty(dispose)) {
1407                 struct ceph_cap *cap;
1408                 /* zero out the in-progress message */
1409                 cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1410                 list_del(&cap->session_caps);
1411                 ceph_put_cap(mdsc, cap);
1412         }
1413 }
1414
1415 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1416                                      struct ceph_mds_session *session)
1417 {
1418         struct ceph_mds_request *req;
1419         struct rb_node *p;
1420         struct ceph_inode_info *ci;
1421
1422         dout("cleanup_session_requests mds%d\n", session->s_mds);
1423         mutex_lock(&mdsc->mutex);
1424         while (!list_empty(&session->s_unsafe)) {
1425                 req = list_first_entry(&session->s_unsafe,
1426                                        struct ceph_mds_request, r_unsafe_item);
1427                 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1428                                     req->r_tid);
1429                 if (req->r_target_inode) {
1430                         /* dropping unsafe change of inode's attributes */
1431                         ci = ceph_inode(req->r_target_inode);
1432                         errseq_set(&ci->i_meta_err, -EIO);
1433                 }
1434                 if (req->r_unsafe_dir) {
1435                         /* dropping unsafe directory operation */
1436                         ci = ceph_inode(req->r_unsafe_dir);
1437                         errseq_set(&ci->i_meta_err, -EIO);
1438                 }
1439                 __unregister_request(mdsc, req);
1440         }
1441         /* zero r_attempts, so kick_requests() will re-send requests */
1442         p = rb_first(&mdsc->request_tree);
1443         while (p) {
1444                 req = rb_entry(p, struct ceph_mds_request, r_node);
1445                 p = rb_next(p);
1446                 if (req->r_session &&
1447                     req->r_session->s_mds == session->s_mds)
1448                         req->r_attempts = 0;
1449         }
1450         mutex_unlock(&mdsc->mutex);
1451 }
1452
1453 /*
1454  * Helper to safely iterate over all caps associated with a session, with
1455  * special care taken to handle a racing __ceph_remove_cap().
1456  *
1457  * Caller must hold session s_mutex.
1458  */
1459 int ceph_iterate_session_caps(struct ceph_mds_session *session,
1460                               int (*cb)(struct inode *, struct ceph_cap *,
1461                                         void *), void *arg)
1462 {
1463         struct list_head *p;
1464         struct ceph_cap *cap;
1465         struct inode *inode, *last_inode = NULL;
1466         struct ceph_cap *old_cap = NULL;
1467         int ret;
1468
1469         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1470         spin_lock(&session->s_cap_lock);
1471         p = session->s_caps.next;
1472         while (p != &session->s_caps) {
1473                 cap = list_entry(p, struct ceph_cap, session_caps);
1474                 inode = igrab(&cap->ci->vfs_inode);
1475                 if (!inode) {
1476                         p = p->next;
1477                         continue;
1478                 }
1479                 session->s_cap_iterator = cap;
1480                 spin_unlock(&session->s_cap_lock);
1481
1482                 if (last_inode) {
1483                         /* avoid calling iput_final() while holding
1484                          * s_mutex or in mds dispatch threads */
1485                         ceph_async_iput(last_inode);
1486                         last_inode = NULL;
1487                 }
1488                 if (old_cap) {
1489                         ceph_put_cap(session->s_mdsc, old_cap);
1490                         old_cap = NULL;
1491                 }
1492
1493                 ret = cb(inode, cap, arg);
1494                 last_inode = inode;
1495
1496                 spin_lock(&session->s_cap_lock);
1497                 p = p->next;
1498                 if (!cap->ci) {
1499                         dout("iterate_session_caps  finishing cap %p removal\n",
1500                              cap);
1501                         BUG_ON(cap->session != session);
1502                         cap->session = NULL;
1503                         list_del_init(&cap->session_caps);
1504                         session->s_nr_caps--;
1505                         atomic64_dec(&session->s_mdsc->metric.total_caps);
1506                         if (cap->queue_release)
1507                                 __ceph_queue_cap_release(session, cap);
1508                         else
1509                                 old_cap = cap;  /* put_cap it w/o locks held */
1510                 }
1511                 if (ret < 0)
1512                         goto out;
1513         }
1514         ret = 0;
1515 out:
1516         session->s_cap_iterator = NULL;
1517         spin_unlock(&session->s_cap_lock);
1518
1519         ceph_async_iput(last_inode);
1520         if (old_cap)
1521                 ceph_put_cap(session->s_mdsc, old_cap);
1522
1523         return ret;
1524 }
1525
1526 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1527                                   void *arg)
1528 {
1529         struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1530         struct ceph_inode_info *ci = ceph_inode(inode);
1531         LIST_HEAD(to_remove);
1532         bool dirty_dropped = false;
1533         bool invalidate = false;
1534
1535         dout("removing cap %p, ci is %p, inode is %p\n",
1536              cap, ci, &ci->vfs_inode);
1537         spin_lock(&ci->i_ceph_lock);
1538         __ceph_remove_cap(cap, false);
1539         if (!ci->i_auth_cap) {
1540                 struct ceph_cap_flush *cf;
1541                 struct ceph_mds_client *mdsc = fsc->mdsc;
1542
1543                 if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1544                         if (inode->i_data.nrpages > 0)
1545                                 invalidate = true;
1546                         if (ci->i_wrbuffer_ref > 0)
1547                                 mapping_set_error(&inode->i_data, -EIO);
1548                 }
1549
1550                 while (!list_empty(&ci->i_cap_flush_list)) {
1551                         cf = list_first_entry(&ci->i_cap_flush_list,
1552                                               struct ceph_cap_flush, i_list);
1553                         list_move(&cf->i_list, &to_remove);
1554                 }
1555
1556                 spin_lock(&mdsc->cap_dirty_lock);
1557
1558                 list_for_each_entry(cf, &to_remove, i_list)
1559                         list_del(&cf->g_list);
1560
1561                 if (!list_empty(&ci->i_dirty_item)) {
1562                         pr_warn_ratelimited(
1563                                 " dropping dirty %s state for %p %lld\n",
1564                                 ceph_cap_string(ci->i_dirty_caps),
1565                                 inode, ceph_ino(inode));
1566                         ci->i_dirty_caps = 0;
1567                         list_del_init(&ci->i_dirty_item);
1568                         dirty_dropped = true;
1569                 }
1570                 if (!list_empty(&ci->i_flushing_item)) {
1571                         pr_warn_ratelimited(
1572                                 " dropping dirty+flushing %s state for %p %lld\n",
1573                                 ceph_cap_string(ci->i_flushing_caps),
1574                                 inode, ceph_ino(inode));
1575                         ci->i_flushing_caps = 0;
1576                         list_del_init(&ci->i_flushing_item);
1577                         mdsc->num_cap_flushing--;
1578                         dirty_dropped = true;
1579                 }
1580                 spin_unlock(&mdsc->cap_dirty_lock);
1581
1582                 if (dirty_dropped) {
1583                         errseq_set(&ci->i_meta_err, -EIO);
1584
1585                         if (ci->i_wrbuffer_ref_head == 0 &&
1586                             ci->i_wr_ref == 0 &&
1587                             ci->i_dirty_caps == 0 &&
1588                             ci->i_flushing_caps == 0) {
1589                                 ceph_put_snap_context(ci->i_head_snapc);
1590                                 ci->i_head_snapc = NULL;
1591                         }
1592                 }
1593
1594                 if (atomic_read(&ci->i_filelock_ref) > 0) {
1595                         /* make further file lock syscall return -EIO */
1596                         ci->i_ceph_flags |= CEPH_I_ERROR_FILELOCK;
1597                         pr_warn_ratelimited(" dropping file locks for %p %lld\n",
1598                                             inode, ceph_ino(inode));
1599                 }
1600
1601                 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1602                         list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1603                         ci->i_prealloc_cap_flush = NULL;
1604                 }
1605         }
1606         spin_unlock(&ci->i_ceph_lock);
1607         while (!list_empty(&to_remove)) {
1608                 struct ceph_cap_flush *cf;
1609                 cf = list_first_entry(&to_remove,
1610                                       struct ceph_cap_flush, i_list);
1611                 list_del(&cf->i_list);
1612                 ceph_free_cap_flush(cf);
1613         }
1614
1615         wake_up_all(&ci->i_cap_wq);
1616         if (invalidate)
1617                 ceph_queue_invalidate(inode);
1618         if (dirty_dropped)
1619                 iput(inode);
1620         return 0;
1621 }
1622
1623 /*
1624  * caller must hold session s_mutex
1625  */
1626 static void remove_session_caps(struct ceph_mds_session *session)
1627 {
1628         struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1629         struct super_block *sb = fsc->sb;
1630         LIST_HEAD(dispose);
1631
1632         dout("remove_session_caps on %p\n", session);
1633         ceph_iterate_session_caps(session, remove_session_caps_cb, fsc);
1634
1635         wake_up_all(&fsc->mdsc->cap_flushing_wq);
1636
1637         spin_lock(&session->s_cap_lock);
1638         if (session->s_nr_caps > 0) {
1639                 struct inode *inode;
1640                 struct ceph_cap *cap, *prev = NULL;
1641                 struct ceph_vino vino;
1642                 /*
1643                  * iterate_session_caps() skips inodes that are being
1644                  * deleted, we need to wait until deletions are complete.
1645                  * __wait_on_freeing_inode() is designed for the job,
1646                  * but it is not exported, so use lookup inode function
1647                  * to access it.
1648                  */
1649                 while (!list_empty(&session->s_caps)) {
1650                         cap = list_entry(session->s_caps.next,
1651                                          struct ceph_cap, session_caps);
1652                         if (cap == prev)
1653                                 break;
1654                         prev = cap;
1655                         vino = cap->ci->i_vino;
1656                         spin_unlock(&session->s_cap_lock);
1657
1658                         inode = ceph_find_inode(sb, vino);
1659                          /* avoid calling iput_final() while holding s_mutex */
1660                         ceph_async_iput(inode);
1661
1662                         spin_lock(&session->s_cap_lock);
1663                 }
1664         }
1665
1666         // drop cap expires and unlock s_cap_lock
1667         detach_cap_releases(session, &dispose);
1668
1669         BUG_ON(session->s_nr_caps > 0);
1670         BUG_ON(!list_empty(&session->s_cap_flushing));
1671         spin_unlock(&session->s_cap_lock);
1672         dispose_cap_releases(session->s_mdsc, &dispose);
1673 }
1674
1675 enum {
1676         RECONNECT,
1677         RENEWCAPS,
1678         FORCE_RO,
1679 };
1680
1681 /*
1682  * wake up any threads waiting on this session's caps.  if the cap is
1683  * old (didn't get renewed on the client reconnect), remove it now.
1684  *
1685  * caller must hold s_mutex.
1686  */
1687 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1688                               void *arg)
1689 {
1690         struct ceph_inode_info *ci = ceph_inode(inode);
1691         unsigned long ev = (unsigned long)arg;
1692
1693         if (ev == RECONNECT) {
1694                 spin_lock(&ci->i_ceph_lock);
1695                 ci->i_wanted_max_size = 0;
1696                 ci->i_requested_max_size = 0;
1697                 spin_unlock(&ci->i_ceph_lock);
1698         } else if (ev == RENEWCAPS) {
1699                 if (cap->cap_gen < cap->session->s_cap_gen) {
1700                         /* mds did not re-issue stale cap */
1701                         spin_lock(&ci->i_ceph_lock);
1702                         cap->issued = cap->implemented = CEPH_CAP_PIN;
1703                         spin_unlock(&ci->i_ceph_lock);
1704                 }
1705         } else if (ev == FORCE_RO) {
1706         }
1707         wake_up_all(&ci->i_cap_wq);
1708         return 0;
1709 }
1710
1711 static void wake_up_session_caps(struct ceph_mds_session *session, int ev)
1712 {
1713         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1714         ceph_iterate_session_caps(session, wake_up_session_cb,
1715                                   (void *)(unsigned long)ev);
1716 }
1717
1718 /*
1719  * Send periodic message to MDS renewing all currently held caps.  The
1720  * ack will reset the expiration for all caps from this session.
1721  *
1722  * caller holds s_mutex
1723  */
1724 static int send_renew_caps(struct ceph_mds_client *mdsc,
1725                            struct ceph_mds_session *session)
1726 {
1727         struct ceph_msg *msg;
1728         int state;
1729
1730         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1731             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1732                 pr_info("mds%d caps stale\n", session->s_mds);
1733         session->s_renew_requested = jiffies;
1734
1735         /* do not try to renew caps until a recovering mds has reconnected
1736          * with its clients. */
1737         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1738         if (state < CEPH_MDS_STATE_RECONNECT) {
1739                 dout("send_renew_caps ignoring mds%d (%s)\n",
1740                      session->s_mds, ceph_mds_state_name(state));
1741                 return 0;
1742         }
1743
1744         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1745                 ceph_mds_state_name(state));
1746         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1747                                  ++session->s_renew_seq);
1748         if (!msg)
1749                 return -ENOMEM;
1750         ceph_con_send(&session->s_con, msg);
1751         return 0;
1752 }
1753
1754 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1755                              struct ceph_mds_session *session, u64 seq)
1756 {
1757         struct ceph_msg *msg;
1758
1759         dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1760              session->s_mds, ceph_session_state_name(session->s_state), seq);
1761         msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1762         if (!msg)
1763                 return -ENOMEM;
1764         ceph_con_send(&session->s_con, msg);
1765         return 0;
1766 }
1767
1768
1769 /*
1770  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1771  *
1772  * Called under session->s_mutex
1773  */
1774 static void renewed_caps(struct ceph_mds_client *mdsc,
1775                          struct ceph_mds_session *session, int is_renew)
1776 {
1777         int was_stale;
1778         int wake = 0;
1779
1780         spin_lock(&session->s_cap_lock);
1781         was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1782
1783         session->s_cap_ttl = session->s_renew_requested +
1784                 mdsc->mdsmap->m_session_timeout*HZ;
1785
1786         if (was_stale) {
1787                 if (time_before(jiffies, session->s_cap_ttl)) {
1788                         pr_info("mds%d caps renewed\n", session->s_mds);
1789                         wake = 1;
1790                 } else {
1791                         pr_info("mds%d caps still stale\n", session->s_mds);
1792                 }
1793         }
1794         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1795              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1796              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1797         spin_unlock(&session->s_cap_lock);
1798
1799         if (wake)
1800                 wake_up_session_caps(session, RENEWCAPS);
1801 }
1802
1803 /*
1804  * send a session close request
1805  */
1806 static int request_close_session(struct ceph_mds_session *session)
1807 {
1808         struct ceph_msg *msg;
1809
1810         dout("request_close_session mds%d state %s seq %lld\n",
1811              session->s_mds, ceph_session_state_name(session->s_state),
1812              session->s_seq);
1813         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1814         if (!msg)
1815                 return -ENOMEM;
1816         ceph_con_send(&session->s_con, msg);
1817         return 1;
1818 }
1819
1820 /*
1821  * Called with s_mutex held.
1822  */
1823 static int __close_session(struct ceph_mds_client *mdsc,
1824                          struct ceph_mds_session *session)
1825 {
1826         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1827                 return 0;
1828         session->s_state = CEPH_MDS_SESSION_CLOSING;
1829         return request_close_session(session);
1830 }
1831
1832 static bool drop_negative_children(struct dentry *dentry)
1833 {
1834         struct dentry *child;
1835         bool all_negative = true;
1836
1837         if (!d_is_dir(dentry))
1838                 goto out;
1839
1840         spin_lock(&dentry->d_lock);
1841         list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1842                 if (d_really_is_positive(child)) {
1843                         all_negative = false;
1844                         break;
1845                 }
1846         }
1847         spin_unlock(&dentry->d_lock);
1848
1849         if (all_negative)
1850                 shrink_dcache_parent(dentry);
1851 out:
1852         return all_negative;
1853 }
1854
1855 /*
1856  * Trim old(er) caps.
1857  *
1858  * Because we can't cache an inode without one or more caps, we do
1859  * this indirectly: if a cap is unused, we prune its aliases, at which
1860  * point the inode will hopefully get dropped to.
1861  *
1862  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1863  * memory pressure from the MDS, though, so it needn't be perfect.
1864  */
1865 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1866 {
1867         int *remaining = arg;
1868         struct ceph_inode_info *ci = ceph_inode(inode);
1869         int used, wanted, oissued, mine;
1870
1871         if (*remaining <= 0)
1872                 return -1;
1873
1874         spin_lock(&ci->i_ceph_lock);
1875         mine = cap->issued | cap->implemented;
1876         used = __ceph_caps_used(ci);
1877         wanted = __ceph_caps_file_wanted(ci);
1878         oissued = __ceph_caps_issued_other(ci, cap);
1879
1880         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1881              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1882              ceph_cap_string(used), ceph_cap_string(wanted));
1883         if (cap == ci->i_auth_cap) {
1884                 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1885                     !list_empty(&ci->i_cap_snaps))
1886                         goto out;
1887                 if ((used | wanted) & CEPH_CAP_ANY_WR)
1888                         goto out;
1889                 /* Note: it's possible that i_filelock_ref becomes non-zero
1890                  * after dropping auth caps. It doesn't hurt because reply
1891                  * of lock mds request will re-add auth caps. */
1892                 if (atomic_read(&ci->i_filelock_ref) > 0)
1893                         goto out;
1894         }
1895         /* The inode has cached pages, but it's no longer used.
1896          * we can safely drop it */
1897         if (S_ISREG(inode->i_mode) &&
1898             wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1899             !(oissued & CEPH_CAP_FILE_CACHE)) {
1900           used = 0;
1901           oissued = 0;
1902         }
1903         if ((used | wanted) & ~oissued & mine)
1904                 goto out;   /* we need these caps */
1905
1906         if (oissued) {
1907                 /* we aren't the only cap.. just remove us */
1908                 __ceph_remove_cap(cap, true);
1909                 (*remaining)--;
1910         } else {
1911                 struct dentry *dentry;
1912                 /* try dropping referring dentries */
1913                 spin_unlock(&ci->i_ceph_lock);
1914                 dentry = d_find_any_alias(inode);
1915                 if (dentry && drop_negative_children(dentry)) {
1916                         int count;
1917                         dput(dentry);
1918                         d_prune_aliases(inode);
1919                         count = atomic_read(&inode->i_count);
1920                         if (count == 1)
1921                                 (*remaining)--;
1922                         dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1923                              inode, cap, count);
1924                 } else {
1925                         dput(dentry);
1926                 }
1927                 return 0;
1928         }
1929
1930 out:
1931         spin_unlock(&ci->i_ceph_lock);
1932         return 0;
1933 }
1934
1935 /*
1936  * Trim session cap count down to some max number.
1937  */
1938 int ceph_trim_caps(struct ceph_mds_client *mdsc,
1939                    struct ceph_mds_session *session,
1940                    int max_caps)
1941 {
1942         int trim_caps = session->s_nr_caps - max_caps;
1943
1944         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1945              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1946         if (trim_caps > 0) {
1947                 int remaining = trim_caps;
1948
1949                 ceph_iterate_session_caps(session, trim_caps_cb, &remaining);
1950                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1951                      session->s_mds, session->s_nr_caps, max_caps,
1952                         trim_caps - remaining);
1953         }
1954
1955         ceph_flush_cap_releases(mdsc, session);
1956         return 0;
1957 }
1958
1959 static int check_caps_flush(struct ceph_mds_client *mdsc,
1960                             u64 want_flush_tid)
1961 {
1962         int ret = 1;
1963
1964         spin_lock(&mdsc->cap_dirty_lock);
1965         if (!list_empty(&mdsc->cap_flush_list)) {
1966                 struct ceph_cap_flush *cf =
1967                         list_first_entry(&mdsc->cap_flush_list,
1968                                          struct ceph_cap_flush, g_list);
1969                 if (cf->tid <= want_flush_tid) {
1970                         dout("check_caps_flush still flushing tid "
1971                              "%llu <= %llu\n", cf->tid, want_flush_tid);
1972                         ret = 0;
1973                 }
1974         }
1975         spin_unlock(&mdsc->cap_dirty_lock);
1976         return ret;
1977 }
1978
1979 /*
1980  * flush all dirty inode data to disk.
1981  *
1982  * returns true if we've flushed through want_flush_tid
1983  */
1984 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1985                             u64 want_flush_tid)
1986 {
1987         dout("check_caps_flush want %llu\n", want_flush_tid);
1988
1989         wait_event(mdsc->cap_flushing_wq,
1990                    check_caps_flush(mdsc, want_flush_tid));
1991
1992         dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1993 }
1994
1995 /*
1996  * called under s_mutex
1997  */
1998 static void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1999                                    struct ceph_mds_session *session)
2000 {
2001         struct ceph_msg *msg = NULL;
2002         struct ceph_mds_cap_release *head;
2003         struct ceph_mds_cap_item *item;
2004         struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
2005         struct ceph_cap *cap;
2006         LIST_HEAD(tmp_list);
2007         int num_cap_releases;
2008         __le32  barrier, *cap_barrier;
2009
2010         down_read(&osdc->lock);
2011         barrier = cpu_to_le32(osdc->epoch_barrier);
2012         up_read(&osdc->lock);
2013
2014         spin_lock(&session->s_cap_lock);
2015 again:
2016         list_splice_init(&session->s_cap_releases, &tmp_list);
2017         num_cap_releases = session->s_num_cap_releases;
2018         session->s_num_cap_releases = 0;
2019         spin_unlock(&session->s_cap_lock);
2020
2021         while (!list_empty(&tmp_list)) {
2022                 if (!msg) {
2023                         msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
2024                                         PAGE_SIZE, GFP_NOFS, false);
2025                         if (!msg)
2026                                 goto out_err;
2027                         head = msg->front.iov_base;
2028                         head->num = cpu_to_le32(0);
2029                         msg->front.iov_len = sizeof(*head);
2030
2031                         msg->hdr.version = cpu_to_le16(2);
2032                         msg->hdr.compat_version = cpu_to_le16(1);
2033                 }
2034
2035                 cap = list_first_entry(&tmp_list, struct ceph_cap,
2036                                         session_caps);
2037                 list_del(&cap->session_caps);
2038                 num_cap_releases--;
2039
2040                 head = msg->front.iov_base;
2041                 put_unaligned_le32(get_unaligned_le32(&head->num) + 1,
2042                                    &head->num);
2043                 item = msg->front.iov_base + msg->front.iov_len;
2044                 item->ino = cpu_to_le64(cap->cap_ino);
2045                 item->cap_id = cpu_to_le64(cap->cap_id);
2046                 item->migrate_seq = cpu_to_le32(cap->mseq);
2047                 item->seq = cpu_to_le32(cap->issue_seq);
2048                 msg->front.iov_len += sizeof(*item);
2049
2050                 ceph_put_cap(mdsc, cap);
2051
2052                 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
2053                         // Append cap_barrier field
2054                         cap_barrier = msg->front.iov_base + msg->front.iov_len;
2055                         *cap_barrier = barrier;
2056                         msg->front.iov_len += sizeof(*cap_barrier);
2057
2058                         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2059                         dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
2060                         ceph_con_send(&session->s_con, msg);
2061                         msg = NULL;
2062                 }
2063         }
2064
2065         BUG_ON(num_cap_releases != 0);
2066
2067         spin_lock(&session->s_cap_lock);
2068         if (!list_empty(&session->s_cap_releases))
2069                 goto again;
2070         spin_unlock(&session->s_cap_lock);
2071
2072         if (msg) {
2073                 // Append cap_barrier field
2074                 cap_barrier = msg->front.iov_base + msg->front.iov_len;
2075                 *cap_barrier = barrier;
2076                 msg->front.iov_len += sizeof(*cap_barrier);
2077
2078                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2079                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
2080                 ceph_con_send(&session->s_con, msg);
2081         }
2082         return;
2083 out_err:
2084         pr_err("send_cap_releases mds%d, failed to allocate message\n",
2085                 session->s_mds);
2086         spin_lock(&session->s_cap_lock);
2087         list_splice(&tmp_list, &session->s_cap_releases);
2088         session->s_num_cap_releases += num_cap_releases;
2089         spin_unlock(&session->s_cap_lock);
2090 }
2091
2092 static void ceph_cap_release_work(struct work_struct *work)
2093 {
2094         struct ceph_mds_session *session =
2095                 container_of(work, struct ceph_mds_session, s_cap_release_work);
2096
2097         mutex_lock(&session->s_mutex);
2098         if (session->s_state == CEPH_MDS_SESSION_OPEN ||
2099             session->s_state == CEPH_MDS_SESSION_HUNG)
2100                 ceph_send_cap_releases(session->s_mdsc, session);
2101         mutex_unlock(&session->s_mutex);
2102         ceph_put_mds_session(session);
2103 }
2104
2105 void ceph_flush_cap_releases(struct ceph_mds_client *mdsc,
2106                              struct ceph_mds_session *session)
2107 {
2108         if (mdsc->stopping)
2109                 return;
2110
2111         ceph_get_mds_session(session);
2112         if (queue_work(mdsc->fsc->cap_wq,
2113                        &session->s_cap_release_work)) {
2114                 dout("cap release work queued\n");
2115         } else {
2116                 ceph_put_mds_session(session);
2117                 dout("failed to queue cap release work\n");
2118         }
2119 }
2120
2121 /*
2122  * caller holds session->s_cap_lock
2123  */
2124 void __ceph_queue_cap_release(struct ceph_mds_session *session,
2125                               struct ceph_cap *cap)
2126 {
2127         list_add_tail(&cap->session_caps, &session->s_cap_releases);
2128         session->s_num_cap_releases++;
2129
2130         if (!(session->s_num_cap_releases % CEPH_CAPS_PER_RELEASE))
2131                 ceph_flush_cap_releases(session->s_mdsc, session);
2132 }
2133
2134 static void ceph_cap_reclaim_work(struct work_struct *work)
2135 {
2136         struct ceph_mds_client *mdsc =
2137                 container_of(work, struct ceph_mds_client, cap_reclaim_work);
2138         int ret = ceph_trim_dentries(mdsc);
2139         if (ret == -EAGAIN)
2140                 ceph_queue_cap_reclaim_work(mdsc);
2141 }
2142
2143 void ceph_queue_cap_reclaim_work(struct ceph_mds_client *mdsc)
2144 {
2145         if (mdsc->stopping)
2146                 return;
2147
2148         if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_reclaim_work)) {
2149                 dout("caps reclaim work queued\n");
2150         } else {
2151                 dout("failed to queue caps release work\n");
2152         }
2153 }
2154
2155 void ceph_reclaim_caps_nr(struct ceph_mds_client *mdsc, int nr)
2156 {
2157         int val;
2158         if (!nr)
2159                 return;
2160         val = atomic_add_return(nr, &mdsc->cap_reclaim_pending);
2161         if ((val % CEPH_CAPS_PER_RELEASE) < nr) {
2162                 atomic_set(&mdsc->cap_reclaim_pending, 0);
2163                 ceph_queue_cap_reclaim_work(mdsc);
2164         }
2165 }
2166
2167 /*
2168  * requests
2169  */
2170
2171 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
2172                                     struct inode *dir)
2173 {
2174         struct ceph_inode_info *ci = ceph_inode(dir);
2175         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
2176         struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
2177         size_t size = sizeof(struct ceph_mds_reply_dir_entry);
2178         unsigned int num_entries;
2179         int order;
2180
2181         spin_lock(&ci->i_ceph_lock);
2182         num_entries = ci->i_files + ci->i_subdirs;
2183         spin_unlock(&ci->i_ceph_lock);
2184         num_entries = max(num_entries, 1U);
2185         num_entries = min(num_entries, opt->max_readdir);
2186
2187         order = get_order(size * num_entries);
2188         while (order >= 0) {
2189                 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
2190                                                              __GFP_NOWARN,
2191                                                              order);
2192                 if (rinfo->dir_entries)
2193                         break;
2194                 order--;
2195         }
2196         if (!rinfo->dir_entries)
2197                 return -ENOMEM;
2198
2199         num_entries = (PAGE_SIZE << order) / size;
2200         num_entries = min(num_entries, opt->max_readdir);
2201
2202         rinfo->dir_buf_size = PAGE_SIZE << order;
2203         req->r_num_caps = num_entries + 1;
2204         req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
2205         req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
2206         return 0;
2207 }
2208
2209 /*
2210  * Create an mds request.
2211  */
2212 struct ceph_mds_request *
2213 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
2214 {
2215         struct ceph_mds_request *req;
2216
2217         req = kmem_cache_zalloc(ceph_mds_request_cachep, GFP_NOFS);
2218         if (!req)
2219                 return ERR_PTR(-ENOMEM);
2220
2221         mutex_init(&req->r_fill_mutex);
2222         req->r_mdsc = mdsc;
2223         req->r_started = jiffies;
2224         req->r_start_latency = ktime_get();
2225         req->r_resend_mds = -1;
2226         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
2227         INIT_LIST_HEAD(&req->r_unsafe_target_item);
2228         req->r_fmode = -1;
2229         kref_init(&req->r_kref);
2230         RB_CLEAR_NODE(&req->r_node);
2231         INIT_LIST_HEAD(&req->r_wait);
2232         init_completion(&req->r_completion);
2233         init_completion(&req->r_safe_completion);
2234         INIT_LIST_HEAD(&req->r_unsafe_item);
2235
2236         ktime_get_coarse_real_ts64(&req->r_stamp);
2237
2238         req->r_op = op;
2239         req->r_direct_mode = mode;
2240         return req;
2241 }
2242
2243 /*
2244  * return oldest (lowest) request, tid in request tree, 0 if none.
2245  *
2246  * called under mdsc->mutex.
2247  */
2248 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
2249 {
2250         if (RB_EMPTY_ROOT(&mdsc->request_tree))
2251                 return NULL;
2252         return rb_entry(rb_first(&mdsc->request_tree),
2253                         struct ceph_mds_request, r_node);
2254 }
2255
2256 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
2257 {
2258         return mdsc->oldest_tid;
2259 }
2260
2261 /*
2262  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
2263  * on build_path_from_dentry in fs/cifs/dir.c.
2264  *
2265  * If @stop_on_nosnap, generate path relative to the first non-snapped
2266  * inode.
2267  *
2268  * Encode hidden .snap dirs as a double /, i.e.
2269  *   foo/.snap/bar -> foo//bar
2270  */
2271 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *pbase,
2272                            int stop_on_nosnap)
2273 {
2274         struct dentry *temp;
2275         char *path;
2276         int pos;
2277         unsigned seq;
2278         u64 base;
2279
2280         if (!dentry)
2281                 return ERR_PTR(-EINVAL);
2282
2283         path = __getname();
2284         if (!path)
2285                 return ERR_PTR(-ENOMEM);
2286 retry:
2287         pos = PATH_MAX - 1;
2288         path[pos] = '\0';
2289
2290         seq = read_seqbegin(&rename_lock);
2291         rcu_read_lock();
2292         temp = dentry;
2293         for (;;) {
2294                 struct inode *inode;
2295
2296                 spin_lock(&temp->d_lock);
2297                 inode = d_inode(temp);
2298                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
2299                         dout("build_path path+%d: %p SNAPDIR\n",
2300                              pos, temp);
2301                 } else if (stop_on_nosnap && inode && dentry != temp &&
2302                            ceph_snap(inode) == CEPH_NOSNAP) {
2303                         spin_unlock(&temp->d_lock);
2304                         pos++; /* get rid of any prepended '/' */
2305                         break;
2306                 } else {
2307                         pos -= temp->d_name.len;
2308                         if (pos < 0) {
2309                                 spin_unlock(&temp->d_lock);
2310                                 break;
2311                         }
2312                         memcpy(path + pos, temp->d_name.name, temp->d_name.len);
2313                 }
2314                 spin_unlock(&temp->d_lock);
2315                 temp = READ_ONCE(temp->d_parent);
2316
2317                 /* Are we at the root? */
2318                 if (IS_ROOT(temp))
2319                         break;
2320
2321                 /* Are we out of buffer? */
2322                 if (--pos < 0)
2323                         break;
2324
2325                 path[pos] = '/';
2326         }
2327         base = ceph_ino(d_inode(temp));
2328         rcu_read_unlock();
2329
2330         if (read_seqretry(&rename_lock, seq))
2331                 goto retry;
2332
2333         if (pos < 0) {
2334                 /*
2335                  * A rename didn't occur, but somehow we didn't end up where
2336                  * we thought we would. Throw a warning and try again.
2337                  */
2338                 pr_warn("build_path did not end path lookup where "
2339                         "expected, pos is %d\n", pos);
2340                 goto retry;
2341         }
2342
2343         *pbase = base;
2344         *plen = PATH_MAX - 1 - pos;
2345         dout("build_path on %p %d built %llx '%.*s'\n",
2346              dentry, d_count(dentry), base, *plen, path + pos);
2347         return path + pos;
2348 }
2349
2350 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
2351                              const char **ppath, int *ppathlen, u64 *pino,
2352                              bool *pfreepath, bool parent_locked)
2353 {
2354         char *path;
2355
2356         rcu_read_lock();
2357         if (!dir)
2358                 dir = d_inode_rcu(dentry->d_parent);
2359         if (dir && parent_locked && ceph_snap(dir) == CEPH_NOSNAP) {
2360                 *pino = ceph_ino(dir);
2361                 rcu_read_unlock();
2362                 *ppath = dentry->d_name.name;
2363                 *ppathlen = dentry->d_name.len;
2364                 return 0;
2365         }
2366         rcu_read_unlock();
2367         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2368         if (IS_ERR(path))
2369                 return PTR_ERR(path);
2370         *ppath = path;
2371         *pfreepath = true;
2372         return 0;
2373 }
2374
2375 static int build_inode_path(struct inode *inode,
2376                             const char **ppath, int *ppathlen, u64 *pino,
2377                             bool *pfreepath)
2378 {
2379         struct dentry *dentry;
2380         char *path;
2381
2382         if (ceph_snap(inode) == CEPH_NOSNAP) {
2383                 *pino = ceph_ino(inode);
2384                 *ppathlen = 0;
2385                 return 0;
2386         }
2387         dentry = d_find_alias(inode);
2388         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2389         dput(dentry);
2390         if (IS_ERR(path))
2391                 return PTR_ERR(path);
2392         *ppath = path;
2393         *pfreepath = true;
2394         return 0;
2395 }
2396
2397 /*
2398  * request arguments may be specified via an inode *, a dentry *, or
2399  * an explicit ino+path.
2400  */
2401 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
2402                                   struct inode *rdiri, const char *rpath,
2403                                   u64 rino, const char **ppath, int *pathlen,
2404                                   u64 *ino, bool *freepath, bool parent_locked)
2405 {
2406         int r = 0;
2407
2408         if (rinode) {
2409                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2410                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2411                      ceph_snap(rinode));
2412         } else if (rdentry) {
2413                 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
2414                                         freepath, parent_locked);
2415                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
2416                      *ppath);
2417         } else if (rpath || rino) {
2418                 *ino = rino;
2419                 *ppath = rpath;
2420                 *pathlen = rpath ? strlen(rpath) : 0;
2421                 dout(" path %.*s\n", *pathlen, rpath);
2422         }
2423
2424         return r;
2425 }
2426
2427 /*
2428  * called under mdsc->mutex
2429  */
2430 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
2431                                                struct ceph_mds_request *req,
2432                                                int mds, bool drop_cap_releases)
2433 {
2434         struct ceph_msg *msg;
2435         struct ceph_mds_request_head *head;
2436         const char *path1 = NULL;
2437         const char *path2 = NULL;
2438         u64 ino1 = 0, ino2 = 0;
2439         int pathlen1 = 0, pathlen2 = 0;
2440         bool freepath1 = false, freepath2 = false;
2441         int len;
2442         u16 releases;
2443         void *p, *end;
2444         int ret;
2445
2446         ret = set_request_path_attr(req->r_inode, req->r_dentry,
2447                               req->r_parent, req->r_path1, req->r_ino1.ino,
2448                               &path1, &pathlen1, &ino1, &freepath1,
2449                               test_bit(CEPH_MDS_R_PARENT_LOCKED,
2450                                         &req->r_req_flags));
2451         if (ret < 0) {
2452                 msg = ERR_PTR(ret);
2453                 goto out;
2454         }
2455
2456         /* If r_old_dentry is set, then assume that its parent is locked */
2457         ret = set_request_path_attr(NULL, req->r_old_dentry,
2458                               req->r_old_dentry_dir,
2459                               req->r_path2, req->r_ino2.ino,
2460                               &path2, &pathlen2, &ino2, &freepath2, true);
2461         if (ret < 0) {
2462                 msg = ERR_PTR(ret);
2463                 goto out_free1;
2464         }
2465
2466         len = sizeof(*head) +
2467                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2468                 sizeof(struct ceph_timespec);
2469
2470         /* calculate (max) length for cap releases */
2471         len += sizeof(struct ceph_mds_request_release) *
2472                 (!!req->r_inode_drop + !!req->r_dentry_drop +
2473                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2474         if (req->r_dentry_drop)
2475                 len += pathlen1;
2476         if (req->r_old_dentry_drop)
2477                 len += pathlen2;
2478
2479         msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
2480         if (!msg) {
2481                 msg = ERR_PTR(-ENOMEM);
2482                 goto out_free2;
2483         }
2484
2485         msg->hdr.version = cpu_to_le16(2);
2486         msg->hdr.tid = cpu_to_le64(req->r_tid);
2487
2488         head = msg->front.iov_base;
2489         p = msg->front.iov_base + sizeof(*head);
2490         end = msg->front.iov_base + msg->front.iov_len;
2491
2492         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2493         head->op = cpu_to_le32(req->r_op);
2494         head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
2495         head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
2496         head->ino = cpu_to_le64(req->r_deleg_ino);
2497         head->args = req->r_args;
2498
2499         ceph_encode_filepath(&p, end, ino1, path1);
2500         ceph_encode_filepath(&p, end, ino2, path2);
2501
2502         /* make note of release offset, in case we need to replay */
2503         req->r_request_release_offset = p - msg->front.iov_base;
2504
2505         /* cap releases */
2506         releases = 0;
2507         if (req->r_inode_drop)
2508                 releases += ceph_encode_inode_release(&p,
2509                       req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2510                       mds, req->r_inode_drop, req->r_inode_unless,
2511                       req->r_op == CEPH_MDS_OP_READDIR);
2512         if (req->r_dentry_drop)
2513                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
2514                                 req->r_parent, mds, req->r_dentry_drop,
2515                                 req->r_dentry_unless);
2516         if (req->r_old_dentry_drop)
2517                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2518                                 req->r_old_dentry_dir, mds,
2519                                 req->r_old_dentry_drop,
2520                                 req->r_old_dentry_unless);
2521         if (req->r_old_inode_drop)
2522                 releases += ceph_encode_inode_release(&p,
2523                       d_inode(req->r_old_dentry),
2524                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2525
2526         if (drop_cap_releases) {
2527                 releases = 0;
2528                 p = msg->front.iov_base + req->r_request_release_offset;
2529         }
2530
2531         head->num_releases = cpu_to_le16(releases);
2532
2533         /* time stamp */
2534         {
2535                 struct ceph_timespec ts;
2536                 ceph_encode_timespec64(&ts, &req->r_stamp);
2537                 ceph_encode_copy(&p, &ts, sizeof(ts));
2538         }
2539
2540         if (WARN_ON_ONCE(p > end)) {
2541                 ceph_msg_put(msg);
2542                 msg = ERR_PTR(-ERANGE);
2543                 goto out_free2;
2544         }
2545
2546         msg->front.iov_len = p - msg->front.iov_base;
2547         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2548
2549         if (req->r_pagelist) {
2550                 struct ceph_pagelist *pagelist = req->r_pagelist;
2551                 ceph_msg_data_add_pagelist(msg, pagelist);
2552                 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2553         } else {
2554                 msg->hdr.data_len = 0;
2555         }
2556
2557         msg->hdr.data_off = cpu_to_le16(0);
2558
2559 out_free2:
2560         if (freepath2)
2561                 ceph_mdsc_free_path((char *)path2, pathlen2);
2562 out_free1:
2563         if (freepath1)
2564                 ceph_mdsc_free_path((char *)path1, pathlen1);
2565 out:
2566         return msg;
2567 }
2568
2569 /*
2570  * called under mdsc->mutex if error, under no mutex if
2571  * success.
2572  */
2573 static void complete_request(struct ceph_mds_client *mdsc,
2574                              struct ceph_mds_request *req)
2575 {
2576         req->r_end_latency = ktime_get();
2577
2578         if (req->r_callback)
2579                 req->r_callback(mdsc, req);
2580         complete_all(&req->r_completion);
2581 }
2582
2583 /*
2584  * called under mdsc->mutex
2585  */
2586 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2587                                   struct ceph_mds_request *req,
2588                                   int mds, bool drop_cap_releases)
2589 {
2590         struct ceph_mds_request_head *rhead;
2591         struct ceph_msg *msg;
2592         int flags = 0;
2593
2594         req->r_attempts++;
2595         if (req->r_inode) {
2596                 struct ceph_cap *cap =
2597                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2598
2599                 if (cap)
2600                         req->r_sent_on_mseq = cap->mseq;
2601                 else
2602                         req->r_sent_on_mseq = -1;
2603         }
2604         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2605              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2606
2607         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2608                 void *p;
2609                 /*
2610                  * Replay.  Do not regenerate message (and rebuild
2611                  * paths, etc.); just use the original message.
2612                  * Rebuilding paths will break for renames because
2613                  * d_move mangles the src name.
2614                  */
2615                 msg = req->r_request;
2616                 rhead = msg->front.iov_base;
2617
2618                 flags = le32_to_cpu(rhead->flags);
2619                 flags |= CEPH_MDS_FLAG_REPLAY;
2620                 rhead->flags = cpu_to_le32(flags);
2621
2622                 if (req->r_target_inode)
2623                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2624
2625                 rhead->num_retry = req->r_attempts - 1;
2626
2627                 /* remove cap/dentry releases from message */
2628                 rhead->num_releases = 0;
2629
2630                 /* time stamp */
2631                 p = msg->front.iov_base + req->r_request_release_offset;
2632                 {
2633                         struct ceph_timespec ts;
2634                         ceph_encode_timespec64(&ts, &req->r_stamp);
2635                         ceph_encode_copy(&p, &ts, sizeof(ts));
2636                 }
2637
2638                 msg->front.iov_len = p - msg->front.iov_base;
2639                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2640                 return 0;
2641         }
2642
2643         if (req->r_request) {
2644                 ceph_msg_put(req->r_request);
2645                 req->r_request = NULL;
2646         }
2647         msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2648         if (IS_ERR(msg)) {
2649                 req->r_err = PTR_ERR(msg);
2650                 return PTR_ERR(msg);
2651         }
2652         req->r_request = msg;
2653
2654         rhead = msg->front.iov_base;
2655         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2656         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2657                 flags |= CEPH_MDS_FLAG_REPLAY;
2658         if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags))
2659                 flags |= CEPH_MDS_FLAG_ASYNC;
2660         if (req->r_parent)
2661                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2662         rhead->flags = cpu_to_le32(flags);
2663         rhead->num_fwd = req->r_num_fwd;
2664         rhead->num_retry = req->r_attempts - 1;
2665
2666         dout(" r_parent = %p\n", req->r_parent);
2667         return 0;
2668 }
2669
2670 /*
2671  * called under mdsc->mutex
2672  */
2673 static int __send_request(struct ceph_mds_client *mdsc,
2674                           struct ceph_mds_session *session,
2675                           struct ceph_mds_request *req,
2676                           bool drop_cap_releases)
2677 {
2678         int err;
2679
2680         err = __prepare_send_request(mdsc, req, session->s_mds,
2681                                      drop_cap_releases);
2682         if (!err) {
2683                 ceph_msg_get(req->r_request);
2684                 ceph_con_send(&session->s_con, req->r_request);
2685         }
2686
2687         return err;
2688 }
2689
2690 /*
2691  * send request, or put it on the appropriate wait list.
2692  */
2693 static void __do_request(struct ceph_mds_client *mdsc,
2694                         struct ceph_mds_request *req)
2695 {
2696         struct ceph_mds_session *session = NULL;
2697         int mds = -1;
2698         int err = 0;
2699         bool random;
2700
2701         if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2702                 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2703                         __unregister_request(mdsc, req);
2704                 return;
2705         }
2706
2707         if (req->r_timeout &&
2708             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2709                 dout("do_request timed out\n");
2710                 err = -ETIMEDOUT;
2711                 goto finish;
2712         }
2713         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2714                 dout("do_request forced umount\n");
2715                 err = -EIO;
2716                 goto finish;
2717         }
2718         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2719                 if (mdsc->mdsmap_err) {
2720                         err = mdsc->mdsmap_err;
2721                         dout("do_request mdsmap err %d\n", err);
2722                         goto finish;
2723                 }
2724                 if (mdsc->mdsmap->m_epoch == 0) {
2725                         dout("do_request no mdsmap, waiting for map\n");
2726                         list_add(&req->r_wait, &mdsc->waiting_for_map);
2727                         return;
2728                 }
2729                 if (!(mdsc->fsc->mount_options->flags &
2730                       CEPH_MOUNT_OPT_MOUNTWAIT) &&
2731                     !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2732                         err = -EHOSTUNREACH;
2733                         goto finish;
2734                 }
2735         }
2736
2737         put_request_session(req);
2738
2739         mds = __choose_mds(mdsc, req, &random);
2740         if (mds < 0 ||
2741             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2742                 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2743                         err = -EJUKEBOX;
2744                         goto finish;
2745                 }
2746                 dout("do_request no mds or not active, waiting for map\n");
2747                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2748                 return;
2749         }
2750
2751         /* get, open session */
2752         session = __ceph_lookup_mds_session(mdsc, mds);
2753         if (!session) {
2754                 session = register_session(mdsc, mds);
2755                 if (IS_ERR(session)) {
2756                         err = PTR_ERR(session);
2757                         goto finish;
2758                 }
2759         }
2760         req->r_session = ceph_get_mds_session(session);
2761
2762         dout("do_request mds%d session %p state %s\n", mds, session,
2763              ceph_session_state_name(session->s_state));
2764         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2765             session->s_state != CEPH_MDS_SESSION_HUNG) {
2766                 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2767                         err = -EACCES;
2768                         goto out_session;
2769                 }
2770                 /*
2771                  * We cannot queue async requests since the caps and delegated
2772                  * inodes are bound to the session. Just return -EJUKEBOX and
2773                  * let the caller retry a sync request in that case.
2774                  */
2775                 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2776                         err = -EJUKEBOX;
2777                         goto out_session;
2778                 }
2779                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2780                     session->s_state == CEPH_MDS_SESSION_CLOSING) {
2781                         err = __open_session(mdsc, session);
2782                         if (err)
2783                                 goto out_session;
2784                         /* retry the same mds later */
2785                         if (random)
2786                                 req->r_resend_mds = mds;
2787                 }
2788                 list_add(&req->r_wait, &session->s_waiting);
2789                 goto out_session;
2790         }
2791
2792         /* send request */
2793         req->r_resend_mds = -1;   /* forget any previous mds hint */
2794
2795         if (req->r_request_started == 0)   /* note request start time */
2796                 req->r_request_started = jiffies;
2797
2798         err = __send_request(mdsc, session, req, false);
2799
2800 out_session:
2801         ceph_put_mds_session(session);
2802 finish:
2803         if (err) {
2804                 dout("__do_request early error %d\n", err);
2805                 req->r_err = err;
2806                 complete_request(mdsc, req);
2807                 __unregister_request(mdsc, req);
2808         }
2809         return;
2810 }
2811
2812 /*
2813  * called under mdsc->mutex
2814  */
2815 static void __wake_requests(struct ceph_mds_client *mdsc,
2816                             struct list_head *head)
2817 {
2818         struct ceph_mds_request *req;
2819         LIST_HEAD(tmp_list);
2820
2821         list_splice_init(head, &tmp_list);
2822
2823         while (!list_empty(&tmp_list)) {
2824                 req = list_entry(tmp_list.next,
2825                                  struct ceph_mds_request, r_wait);
2826                 list_del_init(&req->r_wait);
2827                 dout(" wake request %p tid %llu\n", req, req->r_tid);
2828                 __do_request(mdsc, req);
2829         }
2830 }
2831
2832 /*
2833  * Wake up threads with requests pending for @mds, so that they can
2834  * resubmit their requests to a possibly different mds.
2835  */
2836 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2837 {
2838         struct ceph_mds_request *req;
2839         struct rb_node *p = rb_first(&mdsc->request_tree);
2840
2841         dout("kick_requests mds%d\n", mds);
2842         while (p) {
2843                 req = rb_entry(p, struct ceph_mds_request, r_node);
2844                 p = rb_next(p);
2845                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2846                         continue;
2847                 if (req->r_attempts > 0)
2848                         continue; /* only new requests */
2849                 if (req->r_session &&
2850                     req->r_session->s_mds == mds) {
2851                         dout(" kicking tid %llu\n", req->r_tid);
2852                         list_del_init(&req->r_wait);
2853                         __do_request(mdsc, req);
2854                 }
2855         }
2856 }
2857
2858 int ceph_mdsc_submit_request(struct ceph_mds_client *mdsc, struct inode *dir,
2859                               struct ceph_mds_request *req)
2860 {
2861         int err = 0;
2862
2863         /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2864         if (req->r_inode)
2865                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2866         if (req->r_parent) {
2867                 struct ceph_inode_info *ci = ceph_inode(req->r_parent);
2868                 int fmode = (req->r_op & CEPH_MDS_OP_WRITE) ?
2869                             CEPH_FILE_MODE_WR : CEPH_FILE_MODE_RD;
2870                 spin_lock(&ci->i_ceph_lock);
2871                 ceph_take_cap_refs(ci, CEPH_CAP_PIN, false);
2872                 __ceph_touch_fmode(ci, mdsc, fmode);
2873                 spin_unlock(&ci->i_ceph_lock);
2874                 ihold(req->r_parent);
2875         }
2876         if (req->r_old_dentry_dir)
2877                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2878                                   CEPH_CAP_PIN);
2879
2880         if (req->r_inode) {
2881                 err = ceph_wait_on_async_create(req->r_inode);
2882                 if (err) {
2883                         dout("%s: wait for async create returned: %d\n",
2884                              __func__, err);
2885                         return err;
2886                 }
2887         }
2888
2889         if (!err && req->r_old_inode) {
2890                 err = ceph_wait_on_async_create(req->r_old_inode);
2891                 if (err) {
2892                         dout("%s: wait for async create returned: %d\n",
2893                              __func__, err);
2894                         return err;
2895                 }
2896         }
2897
2898         dout("submit_request on %p for inode %p\n", req, dir);
2899         mutex_lock(&mdsc->mutex);
2900         __register_request(mdsc, req, dir);
2901         __do_request(mdsc, req);
2902         err = req->r_err;
2903         mutex_unlock(&mdsc->mutex);
2904         return err;
2905 }
2906
2907 static int ceph_mdsc_wait_request(struct ceph_mds_client *mdsc,
2908                                   struct ceph_mds_request *req)
2909 {
2910         int err;
2911
2912         /* wait */
2913         dout("do_request waiting\n");
2914         if (!req->r_timeout && req->r_wait_for_completion) {
2915                 err = req->r_wait_for_completion(mdsc, req);
2916         } else {
2917                 long timeleft = wait_for_completion_killable_timeout(
2918                                         &req->r_completion,
2919                                         ceph_timeout_jiffies(req->r_timeout));
2920                 if (timeleft > 0)
2921                         err = 0;
2922                 else if (!timeleft)
2923                         err = -ETIMEDOUT;  /* timed out */
2924                 else
2925                         err = timeleft;  /* killed */
2926         }
2927         dout("do_request waited, got %d\n", err);
2928         mutex_lock(&mdsc->mutex);
2929
2930         /* only abort if we didn't race with a real reply */
2931         if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2932                 err = le32_to_cpu(req->r_reply_info.head->result);
2933         } else if (err < 0) {
2934                 dout("aborted request %lld with %d\n", req->r_tid, err);
2935
2936                 /*
2937                  * ensure we aren't running concurrently with
2938                  * ceph_fill_trace or ceph_readdir_prepopulate, which
2939                  * rely on locks (dir mutex) held by our caller.
2940                  */
2941                 mutex_lock(&req->r_fill_mutex);
2942                 req->r_err = err;
2943                 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2944                 mutex_unlock(&req->r_fill_mutex);
2945
2946                 if (req->r_parent &&
2947                     (req->r_op & CEPH_MDS_OP_WRITE))
2948                         ceph_invalidate_dir_request(req);
2949         } else {
2950                 err = req->r_err;
2951         }
2952
2953         mutex_unlock(&mdsc->mutex);
2954         return err;
2955 }
2956
2957 /*
2958  * Synchrously perform an mds request.  Take care of all of the
2959  * session setup, forwarding, retry details.
2960  */
2961 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2962                          struct inode *dir,
2963                          struct ceph_mds_request *req)
2964 {
2965         int err;
2966
2967         dout("do_request on %p\n", req);
2968
2969         /* issue */
2970         err = ceph_mdsc_submit_request(mdsc, dir, req);
2971         if (!err)
2972                 err = ceph_mdsc_wait_request(mdsc, req);
2973         dout("do_request %p done, result %d\n", req, err);
2974         return err;
2975 }
2976
2977 /*
2978  * Invalidate dir's completeness, dentry lease state on an aborted MDS
2979  * namespace request.
2980  */
2981 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2982 {
2983         struct inode *dir = req->r_parent;
2984         struct inode *old_dir = req->r_old_dentry_dir;
2985
2986         dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
2987
2988         ceph_dir_clear_complete(dir);
2989         if (old_dir)
2990                 ceph_dir_clear_complete(old_dir);
2991         if (req->r_dentry)
2992                 ceph_invalidate_dentry_lease(req->r_dentry);
2993         if (req->r_old_dentry)
2994                 ceph_invalidate_dentry_lease(req->r_old_dentry);
2995 }
2996
2997 /*
2998  * Handle mds reply.
2999  *
3000  * We take the session mutex and parse and process the reply immediately.
3001  * This preserves the logical ordering of replies, capabilities, etc., sent
3002  * by the MDS as they are applied to our local cache.
3003  */
3004 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
3005 {
3006         struct ceph_mds_client *mdsc = session->s_mdsc;
3007         struct ceph_mds_request *req;
3008         struct ceph_mds_reply_head *head = msg->front.iov_base;
3009         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
3010         struct ceph_snap_realm *realm;
3011         u64 tid;
3012         int err, result;
3013         int mds = session->s_mds;
3014
3015         if (msg->front.iov_len < sizeof(*head)) {
3016                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
3017                 ceph_msg_dump(msg);
3018                 return;
3019         }
3020
3021         /* get request, session */
3022         tid = le64_to_cpu(msg->hdr.tid);
3023         mutex_lock(&mdsc->mutex);
3024         req = lookup_get_request(mdsc, tid);
3025         if (!req) {
3026                 dout("handle_reply on unknown tid %llu\n", tid);
3027                 mutex_unlock(&mdsc->mutex);
3028                 return;
3029         }
3030         dout("handle_reply %p\n", req);
3031
3032         /* correct session? */
3033         if (req->r_session != session) {
3034                 pr_err("mdsc_handle_reply got %llu on session mds%d"
3035                        " not mds%d\n", tid, session->s_mds,
3036                        req->r_session ? req->r_session->s_mds : -1);
3037                 mutex_unlock(&mdsc->mutex);
3038                 goto out;
3039         }
3040
3041         /* dup? */
3042         if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
3043             (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
3044                 pr_warn("got a dup %s reply on %llu from mds%d\n",
3045                            head->safe ? "safe" : "unsafe", tid, mds);
3046                 mutex_unlock(&mdsc->mutex);
3047                 goto out;
3048         }
3049         if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
3050                 pr_warn("got unsafe after safe on %llu from mds%d\n",
3051                            tid, mds);
3052                 mutex_unlock(&mdsc->mutex);
3053                 goto out;
3054         }
3055
3056         result = le32_to_cpu(head->result);
3057
3058         /*
3059          * Handle an ESTALE
3060          * if we're not talking to the authority, send to them
3061          * if the authority has changed while we weren't looking,
3062          * send to new authority
3063          * Otherwise we just have to return an ESTALE
3064          */
3065         if (result == -ESTALE) {
3066                 dout("got ESTALE on request %llu\n", req->r_tid);
3067                 req->r_resend_mds = -1;
3068                 if (req->r_direct_mode != USE_AUTH_MDS) {
3069                         dout("not using auth, setting for that now\n");
3070                         req->r_direct_mode = USE_AUTH_MDS;
3071                         __do_request(mdsc, req);
3072                         mutex_unlock(&mdsc->mutex);
3073                         goto out;
3074                 } else  {
3075                         int mds = __choose_mds(mdsc, req, NULL);
3076                         if (mds >= 0 && mds != req->r_session->s_mds) {
3077                                 dout("but auth changed, so resending\n");
3078                                 __do_request(mdsc, req);
3079                                 mutex_unlock(&mdsc->mutex);
3080                                 goto out;
3081                         }
3082                 }
3083                 dout("have to return ESTALE on request %llu\n", req->r_tid);
3084         }
3085
3086
3087         if (head->safe) {
3088                 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
3089                 __unregister_request(mdsc, req);
3090
3091                 /* last request during umount? */
3092                 if (mdsc->stopping && !__get_oldest_req(mdsc))
3093                         complete_all(&mdsc->safe_umount_waiters);
3094
3095                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3096                         /*
3097                          * We already handled the unsafe response, now do the
3098                          * cleanup.  No need to examine the response; the MDS
3099                          * doesn't include any result info in the safe
3100                          * response.  And even if it did, there is nothing
3101                          * useful we could do with a revised return value.
3102                          */
3103                         dout("got safe reply %llu, mds%d\n", tid, mds);
3104
3105                         mutex_unlock(&mdsc->mutex);
3106                         goto out;
3107                 }
3108         } else {
3109                 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
3110                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
3111         }
3112
3113         dout("handle_reply tid %lld result %d\n", tid, result);
3114         rinfo = &req->r_reply_info;
3115         if (test_bit(CEPHFS_FEATURE_REPLY_ENCODING, &session->s_features))
3116                 err = parse_reply_info(session, msg, rinfo, (u64)-1);
3117         else
3118                 err = parse_reply_info(session, msg, rinfo, session->s_con.peer_features);
3119         mutex_unlock(&mdsc->mutex);
3120
3121         mutex_lock(&session->s_mutex);
3122         if (err < 0) {
3123                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
3124                 ceph_msg_dump(msg);
3125                 goto out_err;
3126         }
3127
3128         /* snap trace */
3129         realm = NULL;
3130         if (rinfo->snapblob_len) {
3131                 down_write(&mdsc->snap_rwsem);
3132                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
3133                                 rinfo->snapblob + rinfo->snapblob_len,
3134                                 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
3135                                 &realm);
3136                 downgrade_write(&mdsc->snap_rwsem);
3137         } else {
3138                 down_read(&mdsc->snap_rwsem);
3139         }
3140
3141         /* insert trace into our cache */
3142         mutex_lock(&req->r_fill_mutex);
3143         current->journal_info = req;
3144         err = ceph_fill_trace(mdsc->fsc->sb, req);
3145         if (err == 0) {
3146                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
3147                                     req->r_op == CEPH_MDS_OP_LSSNAP))
3148                         ceph_readdir_prepopulate(req, req->r_session);
3149         }
3150         current->journal_info = NULL;
3151         mutex_unlock(&req->r_fill_mutex);
3152
3153         up_read(&mdsc->snap_rwsem);
3154         if (realm)
3155                 ceph_put_snap_realm(mdsc, realm);
3156
3157         if (err == 0) {
3158                 if (req->r_target_inode &&
3159                     test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3160                         struct ceph_inode_info *ci =
3161                                 ceph_inode(req->r_target_inode);
3162                         spin_lock(&ci->i_unsafe_lock);
3163                         list_add_tail(&req->r_unsafe_target_item,
3164                                       &ci->i_unsafe_iops);
3165                         spin_unlock(&ci->i_unsafe_lock);
3166                 }
3167
3168                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
3169         }
3170 out_err:
3171         mutex_lock(&mdsc->mutex);
3172         if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3173                 if (err) {
3174                         req->r_err = err;
3175                 } else {
3176                         req->r_reply =  ceph_msg_get(msg);
3177                         set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
3178                 }
3179         } else {
3180                 dout("reply arrived after request %lld was aborted\n", tid);
3181         }
3182         mutex_unlock(&mdsc->mutex);
3183
3184         mutex_unlock(&session->s_mutex);
3185
3186         /* kick calling process */
3187         complete_request(mdsc, req);
3188
3189         ceph_update_metadata_latency(&mdsc->metric, req->r_start_latency,
3190                                      req->r_end_latency, err);
3191 out:
3192         ceph_mdsc_put_request(req);
3193         return;
3194 }
3195
3196
3197
3198 /*
3199  * handle mds notification that our request has been forwarded.
3200  */
3201 static void handle_forward(struct ceph_mds_client *mdsc,
3202                            struct ceph_mds_session *session,
3203                            struct ceph_msg *msg)
3204 {
3205         struct ceph_mds_request *req;
3206         u64 tid = le64_to_cpu(msg->hdr.tid);
3207         u32 next_mds;
3208         u32 fwd_seq;
3209         int err = -EINVAL;
3210         void *p = msg->front.iov_base;
3211         void *end = p + msg->front.iov_len;
3212
3213         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
3214         next_mds = ceph_decode_32(&p);
3215         fwd_seq = ceph_decode_32(&p);
3216
3217         mutex_lock(&mdsc->mutex);
3218         req = lookup_get_request(mdsc, tid);
3219         if (!req) {
3220                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
3221                 goto out;  /* dup reply? */
3222         }
3223
3224         if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3225                 dout("forward tid %llu aborted, unregistering\n", tid);
3226                 __unregister_request(mdsc, req);
3227         } else if (fwd_seq <= req->r_num_fwd) {
3228                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
3229                      tid, next_mds, req->r_num_fwd, fwd_seq);
3230         } else {
3231                 /* resend. forward race not possible; mds would drop */
3232                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
3233                 BUG_ON(req->r_err);
3234                 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
3235                 req->r_attempts = 0;
3236                 req->r_num_fwd = fwd_seq;
3237                 req->r_resend_mds = next_mds;
3238                 put_request_session(req);
3239                 __do_request(mdsc, req);
3240         }
3241         ceph_mdsc_put_request(req);
3242 out:
3243         mutex_unlock(&mdsc->mutex);
3244         return;
3245
3246 bad:
3247         pr_err("mdsc_handle_forward decode error err=%d\n", err);
3248 }
3249
3250 static int __decode_session_metadata(void **p, void *end,
3251                                      bool *blacklisted)
3252 {
3253         /* map<string,string> */
3254         u32 n;
3255         bool err_str;
3256         ceph_decode_32_safe(p, end, n, bad);
3257         while (n-- > 0) {
3258                 u32 len;
3259                 ceph_decode_32_safe(p, end, len, bad);
3260                 ceph_decode_need(p, end, len, bad);
3261                 err_str = !strncmp(*p, "error_string", len);
3262                 *p += len;
3263                 ceph_decode_32_safe(p, end, len, bad);
3264                 ceph_decode_need(p, end, len, bad);
3265                 if (err_str && strnstr(*p, "blacklisted", len))
3266                         *blacklisted = true;
3267                 *p += len;
3268         }
3269         return 0;
3270 bad:
3271         return -1;
3272 }
3273
3274 /*
3275  * handle a mds session control message
3276  */
3277 static void handle_session(struct ceph_mds_session *session,
3278                            struct ceph_msg *msg)
3279 {
3280         struct ceph_mds_client *mdsc = session->s_mdsc;
3281         int mds = session->s_mds;
3282         int msg_version = le16_to_cpu(msg->hdr.version);
3283         void *p = msg->front.iov_base;
3284         void *end = p + msg->front.iov_len;
3285         struct ceph_mds_session_head *h;
3286         u32 op;
3287         u64 seq, features = 0;
3288         int wake = 0;
3289         bool blacklisted = false;
3290
3291         /* decode */
3292         ceph_decode_need(&p, end, sizeof(*h), bad);
3293         h = p;
3294         p += sizeof(*h);
3295
3296         op = le32_to_cpu(h->op);
3297         seq = le64_to_cpu(h->seq);
3298
3299         if (msg_version >= 3) {
3300                 u32 len;
3301                 /* version >= 2, metadata */
3302                 if (__decode_session_metadata(&p, end, &blacklisted) < 0)
3303                         goto bad;
3304                 /* version >= 3, feature bits */
3305                 ceph_decode_32_safe(&p, end, len, bad);
3306                 ceph_decode_64_safe(&p, end, features, bad);
3307                 p += len - sizeof(features);
3308         }
3309
3310         mutex_lock(&mdsc->mutex);
3311         if (op == CEPH_SESSION_CLOSE) {
3312                 ceph_get_mds_session(session);
3313                 __unregister_session(mdsc, session);
3314         }
3315         /* FIXME: this ttl calculation is generous */
3316         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
3317         mutex_unlock(&mdsc->mutex);
3318
3319         mutex_lock(&session->s_mutex);
3320
3321         dout("handle_session mds%d %s %p state %s seq %llu\n",
3322              mds, ceph_session_op_name(op), session,
3323              ceph_session_state_name(session->s_state), seq);
3324
3325         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
3326                 session->s_state = CEPH_MDS_SESSION_OPEN;
3327                 pr_info("mds%d came back\n", session->s_mds);
3328         }
3329
3330         switch (op) {
3331         case CEPH_SESSION_OPEN:
3332                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3333                         pr_info("mds%d reconnect success\n", session->s_mds);
3334                 session->s_state = CEPH_MDS_SESSION_OPEN;
3335                 session->s_features = features;
3336                 renewed_caps(mdsc, session, 0);
3337                 if (test_bit(CEPHFS_FEATURE_METRIC_COLLECT, &session->s_features))
3338                         metric_schedule_delayed(&mdsc->metric);
3339                 wake = 1;
3340                 if (mdsc->stopping)
3341                         __close_session(mdsc, session);
3342                 break;
3343
3344         case CEPH_SESSION_RENEWCAPS:
3345                 if (session->s_renew_seq == seq)
3346                         renewed_caps(mdsc, session, 1);
3347                 break;
3348
3349         case CEPH_SESSION_CLOSE:
3350                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3351                         pr_info("mds%d reconnect denied\n", session->s_mds);
3352                 session->s_state = CEPH_MDS_SESSION_CLOSED;
3353                 cleanup_session_requests(mdsc, session);
3354                 remove_session_caps(session);
3355                 wake = 2; /* for good measure */
3356                 wake_up_all(&mdsc->session_close_wq);
3357                 break;
3358
3359         case CEPH_SESSION_STALE:
3360                 pr_info("mds%d caps went stale, renewing\n",
3361                         session->s_mds);
3362                 spin_lock(&session->s_gen_ttl_lock);
3363                 session->s_cap_gen++;
3364                 session->s_cap_ttl = jiffies - 1;
3365                 spin_unlock(&session->s_gen_ttl_lock);
3366                 send_renew_caps(mdsc, session);
3367                 break;
3368
3369         case CEPH_SESSION_RECALL_STATE:
3370                 ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
3371                 break;
3372
3373         case CEPH_SESSION_FLUSHMSG:
3374                 send_flushmsg_ack(mdsc, session, seq);
3375                 break;
3376
3377         case CEPH_SESSION_FORCE_RO:
3378                 dout("force_session_readonly %p\n", session);
3379                 spin_lock(&session->s_cap_lock);
3380                 session->s_readonly = true;
3381                 spin_unlock(&session->s_cap_lock);
3382                 wake_up_session_caps(session, FORCE_RO);
3383                 break;
3384
3385         case CEPH_SESSION_REJECT:
3386                 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
3387                 pr_info("mds%d rejected session\n", session->s_mds);
3388                 session->s_state = CEPH_MDS_SESSION_REJECTED;
3389                 cleanup_session_requests(mdsc, session);
3390                 remove_session_caps(session);
3391                 if (blacklisted)
3392                         mdsc->fsc->blacklisted = true;
3393                 wake = 2; /* for good measure */
3394                 break;
3395
3396         default:
3397                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
3398                 WARN_ON(1);
3399         }
3400
3401         mutex_unlock(&session->s_mutex);
3402         if (wake) {
3403                 mutex_lock(&mdsc->mutex);
3404                 __wake_requests(mdsc, &session->s_waiting);
3405                 if (wake == 2)
3406                         kick_requests(mdsc, mds);
3407                 mutex_unlock(&mdsc->mutex);
3408         }
3409         if (op == CEPH_SESSION_CLOSE)
3410                 ceph_put_mds_session(session);
3411         return;
3412
3413 bad:
3414         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
3415                (int)msg->front.iov_len);
3416         ceph_msg_dump(msg);
3417         return;
3418 }
3419
3420 void ceph_mdsc_release_dir_caps(struct ceph_mds_request *req)
3421 {
3422         int dcaps;
3423
3424         dcaps = xchg(&req->r_dir_caps, 0);
3425         if (dcaps) {
3426                 dout("releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
3427                 ceph_put_cap_refs(ceph_inode(req->r_parent), dcaps);
3428         }
3429 }
3430
3431 void ceph_mdsc_release_dir_caps_no_check(struct ceph_mds_request *req)
3432 {
3433         int dcaps;
3434
3435         dcaps = xchg(&req->r_dir_caps, 0);
3436         if (dcaps) {
3437                 dout("releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
3438                 ceph_put_cap_refs_no_check_caps(ceph_inode(req->r_parent),
3439                                                 dcaps);
3440         }
3441 }
3442
3443 /*
3444  * called under session->mutex.
3445  */
3446 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
3447                                    struct ceph_mds_session *session)
3448 {
3449         struct ceph_mds_request *req, *nreq;
3450         struct rb_node *p;
3451
3452         dout("replay_unsafe_requests mds%d\n", session->s_mds);
3453
3454         mutex_lock(&mdsc->mutex);
3455         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item)
3456                 __send_request(mdsc, session, req, true);
3457
3458         /*
3459          * also re-send old requests when MDS enters reconnect stage. So that MDS
3460          * can process completed request in clientreplay stage.
3461          */
3462         p = rb_first(&mdsc->request_tree);
3463         while (p) {
3464                 req = rb_entry(p, struct ceph_mds_request, r_node);
3465                 p = rb_next(p);
3466                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3467                         continue;
3468                 if (req->r_attempts == 0)
3469                         continue; /* only old requests */
3470                 if (!req->r_session)
3471                         continue;
3472                 if (req->r_session->s_mds != session->s_mds)
3473                         continue;
3474
3475                 ceph_mdsc_release_dir_caps_no_check(req);
3476
3477                 __send_request(mdsc, session, req, true);
3478         }
3479         mutex_unlock(&mdsc->mutex);
3480 }
3481
3482 static int send_reconnect_partial(struct ceph_reconnect_state *recon_state)
3483 {
3484         struct ceph_msg *reply;
3485         struct ceph_pagelist *_pagelist;
3486         struct page *page;
3487         __le32 *addr;
3488         int err = -ENOMEM;
3489
3490         if (!recon_state->allow_multi)
3491                 return -ENOSPC;
3492
3493         /* can't handle message that contains both caps and realm */
3494         BUG_ON(!recon_state->nr_caps == !recon_state->nr_realms);
3495
3496         /* pre-allocate new pagelist */
3497         _pagelist = ceph_pagelist_alloc(GFP_NOFS);
3498         if (!_pagelist)
3499                 return -ENOMEM;
3500
3501         reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3502         if (!reply)
3503                 goto fail_msg;
3504
3505         /* placeholder for nr_caps */
3506         err = ceph_pagelist_encode_32(_pagelist, 0);
3507         if (err < 0)
3508                 goto fail;
3509
3510         if (recon_state->nr_caps) {
3511                 /* currently encoding caps */
3512                 err = ceph_pagelist_encode_32(recon_state->pagelist, 0);
3513                 if (err)
3514                         goto fail;
3515         } else {
3516                 /* placeholder for nr_realms (currently encoding relams) */
3517                 err = ceph_pagelist_encode_32(_pagelist, 0);
3518                 if (err < 0)
3519                         goto fail;
3520         }
3521
3522         err = ceph_pagelist_encode_8(recon_state->pagelist, 1);
3523         if (err)
3524                 goto fail;
3525
3526         page = list_first_entry(&recon_state->pagelist->head, struct page, lru);
3527         addr = kmap_atomic(page);
3528         if (recon_state->nr_caps) {
3529                 /* currently encoding caps */
3530                 *addr = cpu_to_le32(recon_state->nr_caps);
3531         } else {
3532                 /* currently encoding relams */
3533                 *(addr + 1) = cpu_to_le32(recon_state->nr_realms);
3534         }
3535         kunmap_atomic(addr);
3536
3537         reply->hdr.version = cpu_to_le16(5);
3538         reply->hdr.compat_version = cpu_to_le16(4);
3539
3540         reply->hdr.data_len = cpu_to_le32(recon_state->pagelist->length);
3541         ceph_msg_data_add_pagelist(reply, recon_state->pagelist);
3542
3543         ceph_con_send(&recon_state->session->s_con, reply);
3544         ceph_pagelist_release(recon_state->pagelist);
3545
3546         recon_state->pagelist = _pagelist;
3547         recon_state->nr_caps = 0;
3548         recon_state->nr_realms = 0;
3549         recon_state->msg_version = 5;
3550         return 0;
3551 fail:
3552         ceph_msg_put(reply);
3553 fail_msg:
3554         ceph_pagelist_release(_pagelist);
3555         return err;
3556 }
3557
3558 /*
3559  * Encode information about a cap for a reconnect with the MDS.
3560  */
3561 static int reconnect_caps_cb(struct inode *inode, struct ceph_cap *cap,
3562                           void *arg)
3563 {
3564         union {
3565                 struct ceph_mds_cap_reconnect v2;
3566                 struct ceph_mds_cap_reconnect_v1 v1;
3567         } rec;
3568         struct ceph_inode_info *ci = cap->ci;
3569         struct ceph_reconnect_state *recon_state = arg;
3570         struct ceph_pagelist *pagelist = recon_state->pagelist;
3571         int err;
3572         u64 snap_follows;
3573
3574         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
3575              inode, ceph_vinop(inode), cap, cap->cap_id,
3576              ceph_cap_string(cap->issued));
3577
3578         spin_lock(&ci->i_ceph_lock);
3579         cap->seq = 0;        /* reset cap seq */
3580         cap->issue_seq = 0;  /* and issue_seq */
3581         cap->mseq = 0;       /* and migrate_seq */
3582         cap->cap_gen = cap->session->s_cap_gen;
3583
3584         /* These are lost when the session goes away */
3585         if (S_ISDIR(inode->i_mode)) {
3586                 if (cap->issued & CEPH_CAP_DIR_CREATE) {
3587                         ceph_put_string(rcu_dereference_raw(ci->i_cached_layout.pool_ns));
3588                         memset(&ci->i_cached_layout, 0, sizeof(ci->i_cached_layout));
3589                 }
3590                 cap->issued &= ~CEPH_CAP_ANY_DIR_OPS;
3591         }
3592
3593         if (recon_state->msg_version >= 2) {
3594                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
3595                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3596                 rec.v2.issued = cpu_to_le32(cap->issued);
3597                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3598                 rec.v2.pathbase = 0;
3599                 rec.v2.flock_len = (__force __le32)
3600                         ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
3601         } else {
3602                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
3603                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3604                 rec.v1.issued = cpu_to_le32(cap->issued);
3605                 rec.v1.size = cpu_to_le64(inode->i_size);
3606                 ceph_encode_timespec64(&rec.v1.mtime, &inode->i_mtime);
3607                 ceph_encode_timespec64(&rec.v1.atime, &inode->i_atime);
3608                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3609                 rec.v1.pathbase = 0;
3610         }
3611
3612         if (list_empty(&ci->i_cap_snaps)) {
3613                 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
3614         } else {
3615                 struct ceph_cap_snap *capsnap =
3616                         list_first_entry(&ci->i_cap_snaps,
3617                                          struct ceph_cap_snap, ci_item);
3618                 snap_follows = capsnap->follows;
3619         }
3620         spin_unlock(&ci->i_ceph_lock);
3621
3622         if (recon_state->msg_version >= 2) {
3623                 int num_fcntl_locks, num_flock_locks;
3624                 struct ceph_filelock *flocks = NULL;
3625                 size_t struct_len, total_len = sizeof(u64);
3626                 u8 struct_v = 0;
3627
3628 encode_again:
3629                 if (rec.v2.flock_len) {
3630                         ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
3631                 } else {
3632                         num_fcntl_locks = 0;
3633                         num_flock_locks = 0;
3634                 }
3635                 if (num_fcntl_locks + num_flock_locks > 0) {
3636                         flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
3637                                                sizeof(struct ceph_filelock),
3638                                                GFP_NOFS);
3639                         if (!flocks) {
3640                                 err = -ENOMEM;
3641                                 goto out_err;
3642                         }
3643                         err = ceph_encode_locks_to_buffer(inode, flocks,
3644                                                           num_fcntl_locks,
3645                                                           num_flock_locks);
3646                         if (err) {
3647                                 kfree(flocks);
3648                                 flocks = NULL;
3649                                 if (err == -ENOSPC)
3650                                         goto encode_again;
3651                                 goto out_err;
3652                         }
3653                 } else {
3654                         kfree(flocks);
3655                         flocks = NULL;
3656                 }
3657
3658                 if (recon_state->msg_version >= 3) {
3659                         /* version, compat_version and struct_len */
3660                         total_len += 2 * sizeof(u8) + sizeof(u32);
3661                         struct_v = 2;
3662                 }
3663                 /*
3664                  * number of encoded locks is stable, so copy to pagelist
3665                  */
3666                 struct_len = 2 * sizeof(u32) +
3667                             (num_fcntl_locks + num_flock_locks) *
3668                             sizeof(struct ceph_filelock);
3669                 rec.v2.flock_len = cpu_to_le32(struct_len);
3670
3671                 struct_len += sizeof(u32) + sizeof(rec.v2);
3672
3673                 if (struct_v >= 2)
3674                         struct_len += sizeof(u64); /* snap_follows */
3675
3676                 total_len += struct_len;
3677
3678                 if (pagelist->length + total_len > RECONNECT_MAX_SIZE) {
3679                         err = send_reconnect_partial(recon_state);
3680                         if (err)
3681                                 goto out_freeflocks;
3682                         pagelist = recon_state->pagelist;
3683                 }
3684
3685                 err = ceph_pagelist_reserve(pagelist, total_len);
3686                 if (err)
3687                         goto out_freeflocks;
3688
3689                 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3690                 if (recon_state->msg_version >= 3) {
3691                         ceph_pagelist_encode_8(pagelist, struct_v);
3692                         ceph_pagelist_encode_8(pagelist, 1);
3693                         ceph_pagelist_encode_32(pagelist, struct_len);
3694                 }
3695                 ceph_pagelist_encode_string(pagelist, NULL, 0);
3696                 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
3697                 ceph_locks_to_pagelist(flocks, pagelist,
3698                                        num_fcntl_locks, num_flock_locks);
3699                 if (struct_v >= 2)
3700                         ceph_pagelist_encode_64(pagelist, snap_follows);
3701 out_freeflocks:
3702                 kfree(flocks);
3703         } else {
3704                 u64 pathbase = 0;
3705                 int pathlen = 0;
3706                 char *path = NULL;
3707                 struct dentry *dentry;
3708
3709                 dentry = d_find_alias(inode);
3710                 if (dentry) {
3711                         path = ceph_mdsc_build_path(dentry,
3712                                                 &pathlen, &pathbase, 0);
3713                         dput(dentry);
3714                         if (IS_ERR(path)) {
3715                                 err = PTR_ERR(path);
3716                                 goto out_err;
3717                         }
3718                         rec.v1.pathbase = cpu_to_le64(pathbase);
3719                 }
3720
3721                 err = ceph_pagelist_reserve(pagelist,
3722                                             sizeof(u64) + sizeof(u32) +
3723                                             pathlen + sizeof(rec.v1));
3724                 if (err) {
3725                         goto out_freepath;
3726                 }
3727
3728                 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3729                 ceph_pagelist_encode_string(pagelist, path, pathlen);
3730                 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
3731 out_freepath:
3732                 ceph_mdsc_free_path(path, pathlen);
3733         }
3734
3735 out_err:
3736         if (err >= 0)
3737                 recon_state->nr_caps++;
3738         return err;
3739 }
3740
3741 static int encode_snap_realms(struct ceph_mds_client *mdsc,
3742                               struct ceph_reconnect_state *recon_state)
3743 {
3744         struct rb_node *p;
3745         struct ceph_pagelist *pagelist = recon_state->pagelist;
3746         int err = 0;
3747
3748         if (recon_state->msg_version >= 4) {
3749                 err = ceph_pagelist_encode_32(pagelist, mdsc->num_snap_realms);
3750                 if (err < 0)
3751                         goto fail;
3752         }
3753
3754         /*
3755          * snaprealms.  we provide mds with the ino, seq (version), and
3756          * parent for all of our realms.  If the mds has any newer info,
3757          * it will tell us.
3758          */
3759         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3760                 struct ceph_snap_realm *realm =
3761                        rb_entry(p, struct ceph_snap_realm, node);
3762                 struct ceph_mds_snaprealm_reconnect sr_rec;
3763
3764                 if (recon_state->msg_version >= 4) {
3765                         size_t need = sizeof(u8) * 2 + sizeof(u32) +
3766                                       sizeof(sr_rec);
3767
3768                         if (pagelist->length + need > RECONNECT_MAX_SIZE) {
3769                                 err = send_reconnect_partial(recon_state);
3770                                 if (err)
3771                                         goto fail;
3772                                 pagelist = recon_state->pagelist;
3773                         }
3774
3775                         err = ceph_pagelist_reserve(pagelist, need);
3776                         if (err)
3777                                 goto fail;
3778
3779                         ceph_pagelist_encode_8(pagelist, 1);
3780                         ceph_pagelist_encode_8(pagelist, 1);
3781                         ceph_pagelist_encode_32(pagelist, sizeof(sr_rec));
3782                 }
3783
3784                 dout(" adding snap realm %llx seq %lld parent %llx\n",
3785                      realm->ino, realm->seq, realm->parent_ino);
3786                 sr_rec.ino = cpu_to_le64(realm->ino);
3787                 sr_rec.seq = cpu_to_le64(realm->seq);
3788                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3789
3790                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3791                 if (err)
3792                         goto fail;
3793
3794                 recon_state->nr_realms++;
3795         }
3796 fail:
3797         return err;
3798 }
3799
3800
3801 /*
3802  * If an MDS fails and recovers, clients need to reconnect in order to
3803  * reestablish shared state.  This includes all caps issued through
3804  * this session _and_ the snap_realm hierarchy.  Because it's not
3805  * clear which snap realms the mds cares about, we send everything we
3806  * know about.. that ensures we'll then get any new info the
3807  * recovering MDS might have.
3808  *
3809  * This is a relatively heavyweight operation, but it's rare.
3810  */
3811 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3812                                struct ceph_mds_session *session)
3813 {
3814         struct ceph_msg *reply;
3815         int mds = session->s_mds;
3816         int err = -ENOMEM;
3817         struct ceph_reconnect_state recon_state = {
3818                 .session = session,
3819         };
3820         LIST_HEAD(dispose);
3821
3822         pr_info("mds%d reconnect start\n", mds);
3823
3824         recon_state.pagelist = ceph_pagelist_alloc(GFP_NOFS);
3825         if (!recon_state.pagelist)
3826                 goto fail_nopagelist;
3827
3828         reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3829         if (!reply)
3830                 goto fail_nomsg;
3831
3832         xa_destroy(&session->s_delegated_inos);
3833
3834         mutex_lock(&session->s_mutex);
3835         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3836         session->s_seq = 0;
3837
3838         dout("session %p state %s\n", session,
3839              ceph_session_state_name(session->s_state));
3840
3841         spin_lock(&session->s_gen_ttl_lock);
3842         session->s_cap_gen++;
3843         spin_unlock(&session->s_gen_ttl_lock);
3844
3845         spin_lock(&session->s_cap_lock);
3846         /* don't know if session is readonly */
3847         session->s_readonly = 0;
3848         /*
3849          * notify __ceph_remove_cap() that we are composing cap reconnect.
3850          * If a cap get released before being added to the cap reconnect,
3851          * __ceph_remove_cap() should skip queuing cap release.
3852          */
3853         session->s_cap_reconnect = 1;
3854         /* drop old cap expires; we're about to reestablish that state */
3855         detach_cap_releases(session, &dispose);
3856         spin_unlock(&session->s_cap_lock);
3857         dispose_cap_releases(mdsc, &dispose);
3858
3859         /* trim unused caps to reduce MDS's cache rejoin time */
3860         if (mdsc->fsc->sb->s_root)
3861                 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3862
3863         ceph_con_close(&session->s_con);
3864         ceph_con_open(&session->s_con,
3865                       CEPH_ENTITY_TYPE_MDS, mds,
3866                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3867
3868         /* replay unsafe requests */
3869         replay_unsafe_requests(mdsc, session);
3870
3871         ceph_early_kick_flushing_caps(mdsc, session);
3872
3873         down_read(&mdsc->snap_rwsem);
3874
3875         /* placeholder for nr_caps */
3876         err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
3877         if (err)
3878                 goto fail;
3879
3880         if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
3881                 recon_state.msg_version = 3;
3882                 recon_state.allow_multi = true;
3883         } else if (session->s_con.peer_features & CEPH_FEATURE_MDSENC) {
3884                 recon_state.msg_version = 3;
3885         } else {
3886                 recon_state.msg_version = 2;
3887         }
3888         /* trsaverse this session's caps */
3889         err = ceph_iterate_session_caps(session, reconnect_caps_cb, &recon_state);
3890
3891         spin_lock(&session->s_cap_lock);
3892         session->s_cap_reconnect = 0;
3893         spin_unlock(&session->s_cap_lock);
3894
3895         if (err < 0)
3896                 goto fail;
3897
3898         /* check if all realms can be encoded into current message */
3899         if (mdsc->num_snap_realms) {
3900                 size_t total_len =
3901                         recon_state.pagelist->length +
3902                         mdsc->num_snap_realms *
3903                         sizeof(struct ceph_mds_snaprealm_reconnect);
3904                 if (recon_state.msg_version >= 4) {
3905                         /* number of realms */
3906                         total_len += sizeof(u32);
3907                         /* version, compat_version and struct_len */
3908                         total_len += mdsc->num_snap_realms *
3909                                      (2 * sizeof(u8) + sizeof(u32));
3910                 }
3911                 if (total_len > RECONNECT_MAX_SIZE) {
3912                         if (!recon_state.allow_multi) {
3913                                 err = -ENOSPC;
3914                                 goto fail;
3915                         }
3916                         if (recon_state.nr_caps) {
3917                                 err = send_reconnect_partial(&recon_state);
3918                                 if (err)
3919                                         goto fail;
3920                         }
3921                         recon_state.msg_version = 5;
3922                 }
3923         }
3924
3925         err = encode_snap_realms(mdsc, &recon_state);
3926         if (err < 0)
3927                 goto fail;
3928
3929         if (recon_state.msg_version >= 5) {
3930                 err = ceph_pagelist_encode_8(recon_state.pagelist, 0);
3931                 if (err < 0)
3932                         goto fail;
3933         }
3934
3935         if (recon_state.nr_caps || recon_state.nr_realms) {
3936                 struct page *page =
3937                         list_first_entry(&recon_state.pagelist->head,
3938                                         struct page, lru);
3939                 __le32 *addr = kmap_atomic(page);
3940                 if (recon_state.nr_caps) {
3941                         WARN_ON(recon_state.nr_realms != mdsc->num_snap_realms);
3942                         *addr = cpu_to_le32(recon_state.nr_caps);
3943                 } else if (recon_state.msg_version >= 4) {
3944                         *(addr + 1) = cpu_to_le32(recon_state.nr_realms);
3945                 }
3946                 kunmap_atomic(addr);
3947         }
3948
3949         reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3950         if (recon_state.msg_version >= 4)
3951                 reply->hdr.compat_version = cpu_to_le16(4);
3952
3953         reply->hdr.data_len = cpu_to_le32(recon_state.pagelist->length);
3954         ceph_msg_data_add_pagelist(reply, recon_state.pagelist);
3955
3956         ceph_con_send(&session->s_con, reply);
3957
3958         mutex_unlock(&session->s_mutex);
3959
3960         mutex_lock(&mdsc->mutex);
3961         __wake_requests(mdsc, &session->s_waiting);
3962         mutex_unlock(&mdsc->mutex);
3963
3964         up_read(&mdsc->snap_rwsem);
3965         ceph_pagelist_release(recon_state.pagelist);
3966         return;
3967
3968 fail:
3969         ceph_msg_put(reply);
3970         up_read(&mdsc->snap_rwsem);
3971         mutex_unlock(&session->s_mutex);
3972 fail_nomsg:
3973         ceph_pagelist_release(recon_state.pagelist);
3974 fail_nopagelist:
3975         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3976         return;
3977 }
3978
3979
3980 /*
3981  * compare old and new mdsmaps, kicking requests
3982  * and closing out old connections as necessary
3983  *
3984  * called under mdsc->mutex.
3985  */
3986 static void check_new_map(struct ceph_mds_client *mdsc,
3987                           struct ceph_mdsmap *newmap,
3988                           struct ceph_mdsmap *oldmap)
3989 {
3990         int i;
3991         int oldstate, newstate;
3992         struct ceph_mds_session *s;
3993
3994         dout("check_new_map new %u old %u\n",
3995              newmap->m_epoch, oldmap->m_epoch);
3996
3997         for (i = 0; i < oldmap->possible_max_rank && i < mdsc->max_sessions; i++) {
3998                 if (!mdsc->sessions[i])
3999                         continue;
4000                 s = mdsc->sessions[i];
4001                 oldstate = ceph_mdsmap_get_state(oldmap, i);
4002                 newstate = ceph_mdsmap_get_state(newmap, i);
4003
4004                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
4005                      i, ceph_mds_state_name(oldstate),
4006                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
4007                      ceph_mds_state_name(newstate),
4008                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
4009                      ceph_session_state_name(s->s_state));
4010
4011                 if (i >= newmap->possible_max_rank) {
4012                         /* force close session for stopped mds */
4013                         ceph_get_mds_session(s);
4014                         __unregister_session(mdsc, s);
4015                         __wake_requests(mdsc, &s->s_waiting);
4016                         mutex_unlock(&mdsc->mutex);
4017
4018                         mutex_lock(&s->s_mutex);
4019                         cleanup_session_requests(mdsc, s);
4020                         remove_session_caps(s);
4021                         mutex_unlock(&s->s_mutex);
4022
4023                         ceph_put_mds_session(s);
4024
4025                         mutex_lock(&mdsc->mutex);
4026                         kick_requests(mdsc, i);
4027                         continue;
4028                 }
4029
4030                 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
4031                            ceph_mdsmap_get_addr(newmap, i),
4032                            sizeof(struct ceph_entity_addr))) {
4033                         /* just close it */
4034                         mutex_unlock(&mdsc->mutex);
4035                         mutex_lock(&s->s_mutex);
4036                         mutex_lock(&mdsc->mutex);
4037                         ceph_con_close(&s->s_con);
4038                         mutex_unlock(&s->s_mutex);
4039                         s->s_state = CEPH_MDS_SESSION_RESTARTING;
4040                 } else if (oldstate == newstate) {
4041                         continue;  /* nothing new with this mds */
4042                 }
4043
4044                 /*
4045                  * send reconnect?
4046                  */
4047                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
4048                     newstate >= CEPH_MDS_STATE_RECONNECT) {
4049                         mutex_unlock(&mdsc->mutex);
4050                         send_mds_reconnect(mdsc, s);
4051                         mutex_lock(&mdsc->mutex);
4052                 }
4053
4054                 /*
4055                  * kick request on any mds that has gone active.
4056                  */
4057                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
4058                     newstate >= CEPH_MDS_STATE_ACTIVE) {
4059                         if (oldstate != CEPH_MDS_STATE_CREATING &&
4060                             oldstate != CEPH_MDS_STATE_STARTING)
4061                                 pr_info("mds%d recovery completed\n", s->s_mds);
4062                         kick_requests(mdsc, i);
4063                         mutex_unlock(&mdsc->mutex);
4064                         mutex_lock(&s->s_mutex);
4065                         mutex_lock(&mdsc->mutex);
4066                         ceph_kick_flushing_caps(mdsc, s);
4067                         mutex_unlock(&s->s_mutex);
4068                         wake_up_session_caps(s, RECONNECT);
4069                 }
4070         }
4071
4072         for (i = 0; i < newmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4073                 s = mdsc->sessions[i];
4074                 if (!s)
4075                         continue;
4076                 if (!ceph_mdsmap_is_laggy(newmap, i))
4077                         continue;
4078                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4079                     s->s_state == CEPH_MDS_SESSION_HUNG ||
4080                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
4081                         dout(" connecting to export targets of laggy mds%d\n",
4082                              i);
4083                         __open_export_target_sessions(mdsc, s);
4084                 }
4085         }
4086 }
4087
4088
4089
4090 /*
4091  * leases
4092  */
4093
4094 /*
4095  * caller must hold session s_mutex, dentry->d_lock
4096  */
4097 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
4098 {
4099         struct ceph_dentry_info *di = ceph_dentry(dentry);
4100
4101         ceph_put_mds_session(di->lease_session);
4102         di->lease_session = NULL;
4103 }
4104
4105 static void handle_lease(struct ceph_mds_client *mdsc,
4106                          struct ceph_mds_session *session,
4107                          struct ceph_msg *msg)
4108 {
4109         struct super_block *sb = mdsc->fsc->sb;
4110         struct inode *inode;
4111         struct dentry *parent, *dentry;
4112         struct ceph_dentry_info *di;
4113         int mds = session->s_mds;
4114         struct ceph_mds_lease *h = msg->front.iov_base;
4115         u32 seq;
4116         struct ceph_vino vino;
4117         struct qstr dname;
4118         int release = 0;
4119
4120         dout("handle_lease from mds%d\n", mds);
4121
4122         /* decode */
4123         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
4124                 goto bad;
4125         vino.ino = le64_to_cpu(h->ino);
4126         vino.snap = CEPH_NOSNAP;
4127         seq = le32_to_cpu(h->seq);
4128         dname.len = get_unaligned_le32(h + 1);
4129         if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
4130                 goto bad;
4131         dname.name = (void *)(h + 1) + sizeof(u32);
4132
4133         /* lookup inode */
4134         inode = ceph_find_inode(sb, vino);
4135         dout("handle_lease %s, ino %llx %p %.*s\n",
4136              ceph_lease_op_name(h->action), vino.ino, inode,
4137              dname.len, dname.name);
4138
4139         mutex_lock(&session->s_mutex);
4140         session->s_seq++;
4141
4142         if (!inode) {
4143                 dout("handle_lease no inode %llx\n", vino.ino);
4144                 goto release;
4145         }
4146
4147         /* dentry */
4148         parent = d_find_alias(inode);
4149         if (!parent) {
4150                 dout("no parent dentry on inode %p\n", inode);
4151                 WARN_ON(1);
4152                 goto release;  /* hrm... */
4153         }
4154         dname.hash = full_name_hash(parent, dname.name, dname.len);
4155         dentry = d_lookup(parent, &dname);
4156         dput(parent);
4157         if (!dentry)
4158                 goto release;
4159
4160         spin_lock(&dentry->d_lock);
4161         di = ceph_dentry(dentry);
4162         switch (h->action) {
4163         case CEPH_MDS_LEASE_REVOKE:
4164                 if (di->lease_session == session) {
4165                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
4166                                 h->seq = cpu_to_le32(di->lease_seq);
4167                         __ceph_mdsc_drop_dentry_lease(dentry);
4168                 }
4169                 release = 1;
4170                 break;
4171
4172         case CEPH_MDS_LEASE_RENEW:
4173                 if (di->lease_session == session &&
4174                     di->lease_gen == session->s_cap_gen &&
4175                     di->lease_renew_from &&
4176                     di->lease_renew_after == 0) {
4177                         unsigned long duration =
4178                                 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
4179
4180                         di->lease_seq = seq;
4181                         di->time = di->lease_renew_from + duration;
4182                         di->lease_renew_after = di->lease_renew_from +
4183                                 (duration >> 1);
4184                         di->lease_renew_from = 0;
4185                 }
4186                 break;
4187         }
4188         spin_unlock(&dentry->d_lock);
4189         dput(dentry);
4190
4191         if (!release)
4192                 goto out;
4193
4194 release:
4195         /* let's just reuse the same message */
4196         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
4197         ceph_msg_get(msg);
4198         ceph_con_send(&session->s_con, msg);
4199
4200 out:
4201         mutex_unlock(&session->s_mutex);
4202         /* avoid calling iput_final() in mds dispatch threads */
4203         ceph_async_iput(inode);
4204         return;
4205
4206 bad:
4207         pr_err("corrupt lease message\n");
4208         ceph_msg_dump(msg);
4209 }
4210
4211 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
4212                               struct dentry *dentry, char action,
4213                               u32 seq)
4214 {
4215         struct ceph_msg *msg;
4216         struct ceph_mds_lease *lease;
4217         struct inode *dir;
4218         int len = sizeof(*lease) + sizeof(u32) + NAME_MAX;
4219
4220         dout("lease_send_msg identry %p %s to mds%d\n",
4221              dentry, ceph_lease_op_name(action), session->s_mds);
4222
4223         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
4224         if (!msg)
4225                 return;
4226         lease = msg->front.iov_base;
4227         lease->action = action;
4228         lease->seq = cpu_to_le32(seq);
4229
4230         spin_lock(&dentry->d_lock);
4231         dir = d_inode(dentry->d_parent);
4232         lease->ino = cpu_to_le64(ceph_ino(dir));
4233         lease->first = lease->last = cpu_to_le64(ceph_snap(dir));
4234
4235         put_unaligned_le32(dentry->d_name.len, lease + 1);
4236         memcpy((void *)(lease + 1) + 4,
4237                dentry->d_name.name, dentry->d_name.len);
4238         spin_unlock(&dentry->d_lock);
4239         /*
4240          * if this is a preemptive lease RELEASE, no need to
4241          * flush request stream, since the actual request will
4242          * soon follow.
4243          */
4244         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
4245
4246         ceph_con_send(&session->s_con, msg);
4247 }
4248
4249 /*
4250  * lock unlock sessions, to wait ongoing session activities
4251  */
4252 static void lock_unlock_sessions(struct ceph_mds_client *mdsc)
4253 {
4254         int i;
4255
4256         mutex_lock(&mdsc->mutex);
4257         for (i = 0; i < mdsc->max_sessions; i++) {
4258                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4259                 if (!s)
4260                         continue;
4261                 mutex_unlock(&mdsc->mutex);
4262                 mutex_lock(&s->s_mutex);
4263                 mutex_unlock(&s->s_mutex);
4264                 ceph_put_mds_session(s);
4265                 mutex_lock(&mdsc->mutex);
4266         }
4267         mutex_unlock(&mdsc->mutex);
4268 }
4269
4270 static void maybe_recover_session(struct ceph_mds_client *mdsc)
4271 {
4272         struct ceph_fs_client *fsc = mdsc->fsc;
4273
4274         if (!ceph_test_mount_opt(fsc, CLEANRECOVER))
4275                 return;
4276
4277         if (READ_ONCE(fsc->mount_state) != CEPH_MOUNT_MOUNTED)
4278                 return;
4279
4280         if (!READ_ONCE(fsc->blacklisted))
4281                 return;
4282
4283         if (fsc->last_auto_reconnect &&
4284             time_before(jiffies, fsc->last_auto_reconnect + HZ * 60 * 30))
4285                 return;
4286
4287         pr_info("auto reconnect after blacklisted\n");
4288         fsc->last_auto_reconnect = jiffies;
4289         ceph_force_reconnect(fsc->sb);
4290 }
4291
4292 bool check_session_state(struct ceph_mds_session *s)
4293 {
4294         if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
4295                 dout("resending session close request for mds%d\n",
4296                                 s->s_mds);
4297                 request_close_session(s);
4298                 return false;
4299         }
4300         if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
4301                 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
4302                         s->s_state = CEPH_MDS_SESSION_HUNG;
4303                         pr_info("mds%d hung\n", s->s_mds);
4304                 }
4305         }
4306         if (s->s_state == CEPH_MDS_SESSION_NEW ||
4307             s->s_state == CEPH_MDS_SESSION_RESTARTING ||
4308             s->s_state == CEPH_MDS_SESSION_CLOSED ||
4309             s->s_state == CEPH_MDS_SESSION_REJECTED)
4310                 /* this mds is failed or recovering, just wait */
4311                 return false;
4312
4313         return true;
4314 }
4315
4316 /*
4317  * delayed work -- periodically trim expired leases, renew caps with mds
4318  */
4319 static void schedule_delayed(struct ceph_mds_client *mdsc)
4320 {
4321         int delay = 5;
4322         unsigned hz = round_jiffies_relative(HZ * delay);
4323         schedule_delayed_work(&mdsc->delayed_work, hz);
4324 }
4325
4326 static void delayed_work(struct work_struct *work)
4327 {
4328         int i;
4329         struct ceph_mds_client *mdsc =
4330                 container_of(work, struct ceph_mds_client, delayed_work.work);
4331         int renew_interval;
4332         int renew_caps;
4333
4334         dout("mdsc delayed_work\n");
4335
4336         if (mdsc->stopping)
4337                 return;
4338
4339         mutex_lock(&mdsc->mutex);
4340         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
4341         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
4342                                    mdsc->last_renew_caps);
4343         if (renew_caps)
4344                 mdsc->last_renew_caps = jiffies;
4345
4346         for (i = 0; i < mdsc->max_sessions; i++) {
4347                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4348                 if (!s)
4349                         continue;
4350
4351                 if (!check_session_state(s)) {
4352                         ceph_put_mds_session(s);
4353                         continue;
4354                 }
4355                 mutex_unlock(&mdsc->mutex);
4356
4357                 mutex_lock(&s->s_mutex);
4358                 if (renew_caps)
4359                         send_renew_caps(mdsc, s);
4360                 else
4361                         ceph_con_keepalive(&s->s_con);
4362                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4363                     s->s_state == CEPH_MDS_SESSION_HUNG)
4364                         ceph_send_cap_releases(mdsc, s);
4365                 mutex_unlock(&s->s_mutex);
4366                 ceph_put_mds_session(s);
4367
4368                 mutex_lock(&mdsc->mutex);
4369         }
4370         mutex_unlock(&mdsc->mutex);
4371
4372         ceph_check_delayed_caps(mdsc);
4373
4374         ceph_queue_cap_reclaim_work(mdsc);
4375
4376         ceph_trim_snapid_map(mdsc);
4377
4378         maybe_recover_session(mdsc);
4379
4380         schedule_delayed(mdsc);
4381 }
4382
4383 int ceph_mdsc_init(struct ceph_fs_client *fsc)
4384
4385 {
4386         struct ceph_mds_client *mdsc;
4387         int err;
4388
4389         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
4390         if (!mdsc)
4391                 return -ENOMEM;
4392         mdsc->fsc = fsc;
4393         mutex_init(&mdsc->mutex);
4394         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
4395         if (!mdsc->mdsmap) {
4396                 err = -ENOMEM;
4397                 goto err_mdsc;
4398         }
4399
4400         fsc->mdsc = mdsc;
4401         init_completion(&mdsc->safe_umount_waiters);
4402         init_waitqueue_head(&mdsc->session_close_wq);
4403         INIT_LIST_HEAD(&mdsc->waiting_for_map);
4404         mdsc->sessions = NULL;
4405         atomic_set(&mdsc->num_sessions, 0);
4406         mdsc->max_sessions = 0;
4407         mdsc->stopping = 0;
4408         atomic64_set(&mdsc->quotarealms_count, 0);
4409         mdsc->quotarealms_inodes = RB_ROOT;
4410         mutex_init(&mdsc->quotarealms_inodes_mutex);
4411         mdsc->last_snap_seq = 0;
4412         init_rwsem(&mdsc->snap_rwsem);
4413         mdsc->snap_realms = RB_ROOT;
4414         INIT_LIST_HEAD(&mdsc->snap_empty);
4415         mdsc->num_snap_realms = 0;
4416         spin_lock_init(&mdsc->snap_empty_lock);
4417         mdsc->last_tid = 0;
4418         mdsc->oldest_tid = 0;
4419         mdsc->request_tree = RB_ROOT;
4420         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
4421         mdsc->last_renew_caps = jiffies;
4422         INIT_LIST_HEAD(&mdsc->cap_delay_list);
4423         INIT_LIST_HEAD(&mdsc->cap_wait_list);
4424         spin_lock_init(&mdsc->cap_delay_lock);
4425         INIT_LIST_HEAD(&mdsc->snap_flush_list);
4426         spin_lock_init(&mdsc->snap_flush_lock);
4427         mdsc->last_cap_flush_tid = 1;
4428         INIT_LIST_HEAD(&mdsc->cap_flush_list);
4429         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
4430         mdsc->num_cap_flushing = 0;
4431         spin_lock_init(&mdsc->cap_dirty_lock);
4432         init_waitqueue_head(&mdsc->cap_flushing_wq);
4433         INIT_WORK(&mdsc->cap_reclaim_work, ceph_cap_reclaim_work);
4434         atomic_set(&mdsc->cap_reclaim_pending, 0);
4435         err = ceph_metric_init(&mdsc->metric);
4436         if (err)
4437                 goto err_mdsmap;
4438
4439         spin_lock_init(&mdsc->dentry_list_lock);
4440         INIT_LIST_HEAD(&mdsc->dentry_leases);
4441         INIT_LIST_HEAD(&mdsc->dentry_dir_leases);
4442
4443         ceph_caps_init(mdsc);
4444         ceph_adjust_caps_max_min(mdsc, fsc->mount_options);
4445
4446         spin_lock_init(&mdsc->snapid_map_lock);
4447         mdsc->snapid_map_tree = RB_ROOT;
4448         INIT_LIST_HEAD(&mdsc->snapid_map_lru);
4449
4450         init_rwsem(&mdsc->pool_perm_rwsem);
4451         mdsc->pool_perm_tree = RB_ROOT;
4452
4453         strscpy(mdsc->nodename, utsname()->nodename,
4454                 sizeof(mdsc->nodename));
4455         return 0;
4456
4457 err_mdsmap:
4458         kfree(mdsc->mdsmap);
4459 err_mdsc:
4460         kfree(mdsc);
4461         return err;
4462 }
4463
4464 /*
4465  * Wait for safe replies on open mds requests.  If we time out, drop
4466  * all requests from the tree to avoid dangling dentry refs.
4467  */
4468 static void wait_requests(struct ceph_mds_client *mdsc)
4469 {
4470         struct ceph_options *opts = mdsc->fsc->client->options;
4471         struct ceph_mds_request *req;
4472
4473         mutex_lock(&mdsc->mutex);
4474         if (__get_oldest_req(mdsc)) {
4475                 mutex_unlock(&mdsc->mutex);
4476
4477                 dout("wait_requests waiting for requests\n");
4478                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
4479                                     ceph_timeout_jiffies(opts->mount_timeout));
4480
4481                 /* tear down remaining requests */
4482                 mutex_lock(&mdsc->mutex);
4483                 while ((req = __get_oldest_req(mdsc))) {
4484                         dout("wait_requests timed out on tid %llu\n",
4485                              req->r_tid);
4486                         list_del_init(&req->r_wait);
4487                         __unregister_request(mdsc, req);
4488                 }
4489         }
4490         mutex_unlock(&mdsc->mutex);
4491         dout("wait_requests done\n");
4492 }
4493
4494 /*
4495  * called before mount is ro, and before dentries are torn down.
4496  * (hmm, does this still race with new lookups?)
4497  */
4498 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
4499 {
4500         dout("pre_umount\n");
4501         mdsc->stopping = 1;
4502
4503         lock_unlock_sessions(mdsc);
4504         ceph_flush_dirty_caps(mdsc);
4505         wait_requests(mdsc);
4506
4507         /*
4508          * wait for reply handlers to drop their request refs and
4509          * their inode/dcache refs
4510          */
4511         ceph_msgr_flush();
4512
4513         ceph_cleanup_quotarealms_inodes(mdsc);
4514 }
4515
4516 /*
4517  * wait for all write mds requests to flush.
4518  */
4519 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
4520 {
4521         struct ceph_mds_request *req = NULL, *nextreq;
4522         struct rb_node *n;
4523
4524         mutex_lock(&mdsc->mutex);
4525         dout("wait_unsafe_requests want %lld\n", want_tid);
4526 restart:
4527         req = __get_oldest_req(mdsc);
4528         while (req && req->r_tid <= want_tid) {
4529                 /* find next request */
4530                 n = rb_next(&req->r_node);
4531                 if (n)
4532                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
4533                 else
4534                         nextreq = NULL;
4535                 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
4536                     (req->r_op & CEPH_MDS_OP_WRITE)) {
4537                         /* write op */
4538                         ceph_mdsc_get_request(req);
4539                         if (nextreq)
4540                                 ceph_mdsc_get_request(nextreq);
4541                         mutex_unlock(&mdsc->mutex);
4542                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
4543                              req->r_tid, want_tid);
4544                         wait_for_completion(&req->r_safe_completion);
4545                         mutex_lock(&mdsc->mutex);
4546                         ceph_mdsc_put_request(req);
4547                         if (!nextreq)
4548                                 break;  /* next dne before, so we're done! */
4549                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
4550                                 /* next request was removed from tree */
4551                                 ceph_mdsc_put_request(nextreq);
4552                                 goto restart;
4553                         }
4554                         ceph_mdsc_put_request(nextreq);  /* won't go away */
4555                 }
4556                 req = nextreq;
4557         }
4558         mutex_unlock(&mdsc->mutex);
4559         dout("wait_unsafe_requests done\n");
4560 }
4561
4562 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
4563 {
4564         u64 want_tid, want_flush;
4565
4566         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4567                 return;
4568
4569         dout("sync\n");
4570         mutex_lock(&mdsc->mutex);
4571         want_tid = mdsc->last_tid;
4572         mutex_unlock(&mdsc->mutex);
4573
4574         ceph_flush_dirty_caps(mdsc);
4575         spin_lock(&mdsc->cap_dirty_lock);
4576         want_flush = mdsc->last_cap_flush_tid;
4577         if (!list_empty(&mdsc->cap_flush_list)) {
4578                 struct ceph_cap_flush *cf =
4579                         list_last_entry(&mdsc->cap_flush_list,
4580                                         struct ceph_cap_flush, g_list);
4581                 cf->wake = true;
4582         }
4583         spin_unlock(&mdsc->cap_dirty_lock);
4584
4585         dout("sync want tid %lld flush_seq %lld\n",
4586              want_tid, want_flush);
4587
4588         wait_unsafe_requests(mdsc, want_tid);
4589         wait_caps_flush(mdsc, want_flush);
4590 }
4591
4592 /*
4593  * true if all sessions are closed, or we force unmount
4594  */
4595 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
4596 {
4597         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4598                 return true;
4599         return atomic_read(&mdsc->num_sessions) <= skipped;
4600 }
4601
4602 /*
4603  * called after sb is ro.
4604  */
4605 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
4606 {
4607         struct ceph_options *opts = mdsc->fsc->client->options;
4608         struct ceph_mds_session *session;
4609         int i;
4610         int skipped = 0;
4611
4612         dout("close_sessions\n");
4613
4614         /* close sessions */
4615         mutex_lock(&mdsc->mutex);
4616         for (i = 0; i < mdsc->max_sessions; i++) {
4617                 session = __ceph_lookup_mds_session(mdsc, i);
4618                 if (!session)
4619                         continue;
4620                 mutex_unlock(&mdsc->mutex);
4621                 mutex_lock(&session->s_mutex);
4622                 if (__close_session(mdsc, session) <= 0)
4623                         skipped++;
4624                 mutex_unlock(&session->s_mutex);
4625                 ceph_put_mds_session(session);
4626                 mutex_lock(&mdsc->mutex);
4627         }
4628         mutex_unlock(&mdsc->mutex);
4629
4630         dout("waiting for sessions to close\n");
4631         wait_event_timeout(mdsc->session_close_wq,
4632                            done_closing_sessions(mdsc, skipped),
4633                            ceph_timeout_jiffies(opts->mount_timeout));
4634
4635         /* tear down remaining sessions */
4636         mutex_lock(&mdsc->mutex);
4637         for (i = 0; i < mdsc->max_sessions; i++) {
4638                 if (mdsc->sessions[i]) {
4639                         session = ceph_get_mds_session(mdsc->sessions[i]);
4640                         __unregister_session(mdsc, session);
4641                         mutex_unlock(&mdsc->mutex);
4642                         mutex_lock(&session->s_mutex);
4643                         remove_session_caps(session);
4644                         mutex_unlock(&session->s_mutex);
4645                         ceph_put_mds_session(session);
4646                         mutex_lock(&mdsc->mutex);
4647                 }
4648         }
4649         WARN_ON(!list_empty(&mdsc->cap_delay_list));
4650         mutex_unlock(&mdsc->mutex);
4651
4652         ceph_cleanup_snapid_map(mdsc);
4653         ceph_cleanup_empty_realms(mdsc);
4654
4655         cancel_work_sync(&mdsc->cap_reclaim_work);
4656         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4657
4658         dout("stopped\n");
4659 }
4660
4661 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
4662 {
4663         struct ceph_mds_session *session;
4664         int mds;
4665
4666         dout("force umount\n");
4667
4668         mutex_lock(&mdsc->mutex);
4669         for (mds = 0; mds < mdsc->max_sessions; mds++) {
4670                 session = __ceph_lookup_mds_session(mdsc, mds);
4671                 if (!session)
4672                         continue;
4673
4674                 if (session->s_state == CEPH_MDS_SESSION_REJECTED)
4675                         __unregister_session(mdsc, session);
4676                 __wake_requests(mdsc, &session->s_waiting);
4677                 mutex_unlock(&mdsc->mutex);
4678
4679                 mutex_lock(&session->s_mutex);
4680                 __close_session(mdsc, session);
4681                 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
4682                         cleanup_session_requests(mdsc, session);
4683                         remove_session_caps(session);
4684                 }
4685                 mutex_unlock(&session->s_mutex);
4686                 ceph_put_mds_session(session);
4687
4688                 mutex_lock(&mdsc->mutex);
4689                 kick_requests(mdsc, mds);
4690         }
4691         __wake_requests(mdsc, &mdsc->waiting_for_map);
4692         mutex_unlock(&mdsc->mutex);
4693 }
4694
4695 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
4696 {
4697         dout("stop\n");
4698         /*
4699          * Make sure the delayed work stopped before releasing
4700          * the resources.
4701          *
4702          * Because the cancel_delayed_work_sync() will only
4703          * guarantee that the work finishes executing. But the
4704          * delayed work will re-arm itself again after that.
4705          */
4706         flush_delayed_work(&mdsc->delayed_work);
4707
4708         if (mdsc->mdsmap)
4709                 ceph_mdsmap_destroy(mdsc->mdsmap);
4710         kfree(mdsc->sessions);
4711         ceph_caps_finalize(mdsc);
4712         ceph_pool_perm_destroy(mdsc);
4713 }
4714
4715 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
4716 {
4717         struct ceph_mds_client *mdsc = fsc->mdsc;
4718         dout("mdsc_destroy %p\n", mdsc);
4719
4720         if (!mdsc)
4721                 return;
4722
4723         /* flush out any connection work with references to us */
4724         ceph_msgr_flush();
4725
4726         ceph_mdsc_stop(mdsc);
4727
4728         ceph_metric_destroy(&mdsc->metric);
4729
4730         flush_delayed_work(&mdsc->metric.delayed_work);
4731         fsc->mdsc = NULL;
4732         kfree(mdsc);
4733         dout("mdsc_destroy %p done\n", mdsc);
4734 }
4735
4736 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4737 {
4738         struct ceph_fs_client *fsc = mdsc->fsc;
4739         const char *mds_namespace = fsc->mount_options->mds_namespace;
4740         void *p = msg->front.iov_base;
4741         void *end = p + msg->front.iov_len;
4742         u32 epoch;
4743         u32 map_len;
4744         u32 num_fs;
4745         u32 mount_fscid = (u32)-1;
4746         u8 struct_v, struct_cv;
4747         int err = -EINVAL;
4748
4749         ceph_decode_need(&p, end, sizeof(u32), bad);
4750         epoch = ceph_decode_32(&p);
4751
4752         dout("handle_fsmap epoch %u\n", epoch);
4753
4754         ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4755         struct_v = ceph_decode_8(&p);
4756         struct_cv = ceph_decode_8(&p);
4757         map_len = ceph_decode_32(&p);
4758
4759         ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
4760         p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
4761
4762         num_fs = ceph_decode_32(&p);
4763         while (num_fs-- > 0) {
4764                 void *info_p, *info_end;
4765                 u32 info_len;
4766                 u8 info_v, info_cv;
4767                 u32 fscid, namelen;
4768
4769                 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4770                 info_v = ceph_decode_8(&p);
4771                 info_cv = ceph_decode_8(&p);
4772                 info_len = ceph_decode_32(&p);
4773                 ceph_decode_need(&p, end, info_len, bad);
4774                 info_p = p;
4775                 info_end = p + info_len;
4776                 p = info_end;
4777
4778                 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
4779                 fscid = ceph_decode_32(&info_p);
4780                 namelen = ceph_decode_32(&info_p);
4781                 ceph_decode_need(&info_p, info_end, namelen, bad);
4782
4783                 if (mds_namespace &&
4784                     strlen(mds_namespace) == namelen &&
4785                     !strncmp(mds_namespace, (char *)info_p, namelen)) {
4786                         mount_fscid = fscid;
4787                         break;
4788                 }
4789         }
4790
4791         ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
4792         if (mount_fscid != (u32)-1) {
4793                 fsc->client->monc.fs_cluster_id = mount_fscid;
4794                 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
4795                                    0, true);
4796                 ceph_monc_renew_subs(&fsc->client->monc);
4797         } else {
4798                 err = -ENOENT;
4799                 goto err_out;
4800         }
4801         return;
4802
4803 bad:
4804         pr_err("error decoding fsmap\n");
4805 err_out:
4806         mutex_lock(&mdsc->mutex);
4807         mdsc->mdsmap_err = err;
4808         __wake_requests(mdsc, &mdsc->waiting_for_map);
4809         mutex_unlock(&mdsc->mutex);
4810 }
4811
4812 /*
4813  * handle mds map update.
4814  */
4815 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4816 {
4817         u32 epoch;
4818         u32 maplen;
4819         void *p = msg->front.iov_base;
4820         void *end = p + msg->front.iov_len;
4821         struct ceph_mdsmap *newmap, *oldmap;
4822         struct ceph_fsid fsid;
4823         int err = -EINVAL;
4824
4825         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
4826         ceph_decode_copy(&p, &fsid, sizeof(fsid));
4827         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
4828                 return;
4829         epoch = ceph_decode_32(&p);
4830         maplen = ceph_decode_32(&p);
4831         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
4832
4833         /* do we need it? */
4834         mutex_lock(&mdsc->mutex);
4835         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
4836                 dout("handle_map epoch %u <= our %u\n",
4837                      epoch, mdsc->mdsmap->m_epoch);
4838                 mutex_unlock(&mdsc->mutex);
4839                 return;
4840         }
4841
4842         newmap = ceph_mdsmap_decode(&p, end);
4843         if (IS_ERR(newmap)) {
4844                 err = PTR_ERR(newmap);
4845                 goto bad_unlock;
4846         }
4847
4848         /* swap into place */
4849         if (mdsc->mdsmap) {
4850                 oldmap = mdsc->mdsmap;
4851                 mdsc->mdsmap = newmap;
4852                 check_new_map(mdsc, newmap, oldmap);
4853                 ceph_mdsmap_destroy(oldmap);
4854         } else {
4855                 mdsc->mdsmap = newmap;  /* first mds map */
4856         }
4857         mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
4858                                         MAX_LFS_FILESIZE);
4859
4860         __wake_requests(mdsc, &mdsc->waiting_for_map);
4861         ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
4862                           mdsc->mdsmap->m_epoch);
4863
4864         mutex_unlock(&mdsc->mutex);
4865         schedule_delayed(mdsc);
4866         return;
4867
4868 bad_unlock:
4869         mutex_unlock(&mdsc->mutex);
4870 bad:
4871         pr_err("error decoding mdsmap %d\n", err);
4872         return;
4873 }
4874
4875 static struct ceph_connection *con_get(struct ceph_connection *con)
4876 {
4877         struct ceph_mds_session *s = con->private;
4878
4879         if (ceph_get_mds_session(s))
4880                 return con;
4881         return NULL;
4882 }
4883
4884 static void con_put(struct ceph_connection *con)
4885 {
4886         struct ceph_mds_session *s = con->private;
4887
4888         ceph_put_mds_session(s);
4889 }
4890
4891 /*
4892  * if the client is unresponsive for long enough, the mds will kill
4893  * the session entirely.
4894  */
4895 static void peer_reset(struct ceph_connection *con)
4896 {
4897         struct ceph_mds_session *s = con->private;
4898         struct ceph_mds_client *mdsc = s->s_mdsc;
4899
4900         pr_warn("mds%d closed our session\n", s->s_mds);
4901         send_mds_reconnect(mdsc, s);
4902 }
4903
4904 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
4905 {
4906         struct ceph_mds_session *s = con->private;
4907         struct ceph_mds_client *mdsc = s->s_mdsc;
4908         int type = le16_to_cpu(msg->hdr.type);
4909
4910         mutex_lock(&mdsc->mutex);
4911         if (__verify_registered_session(mdsc, s) < 0) {
4912                 mutex_unlock(&mdsc->mutex);
4913                 goto out;
4914         }
4915         mutex_unlock(&mdsc->mutex);
4916
4917         switch (type) {
4918         case CEPH_MSG_MDS_MAP:
4919                 ceph_mdsc_handle_mdsmap(mdsc, msg);
4920                 break;
4921         case CEPH_MSG_FS_MAP_USER:
4922                 ceph_mdsc_handle_fsmap(mdsc, msg);
4923                 break;
4924         case CEPH_MSG_CLIENT_SESSION:
4925                 handle_session(s, msg);
4926                 break;
4927         case CEPH_MSG_CLIENT_REPLY:
4928                 handle_reply(s, msg);
4929                 break;
4930         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
4931                 handle_forward(mdsc, s, msg);
4932                 break;
4933         case CEPH_MSG_CLIENT_CAPS:
4934                 ceph_handle_caps(s, msg);
4935                 break;
4936         case CEPH_MSG_CLIENT_SNAP:
4937                 ceph_handle_snap(mdsc, s, msg);
4938                 break;
4939         case CEPH_MSG_CLIENT_LEASE:
4940                 handle_lease(mdsc, s, msg);
4941                 break;
4942         case CEPH_MSG_CLIENT_QUOTA:
4943                 ceph_handle_quota(mdsc, s, msg);
4944                 break;
4945
4946         default:
4947                 pr_err("received unknown message type %d %s\n", type,
4948                        ceph_msg_type_name(type));
4949         }
4950 out:
4951         ceph_msg_put(msg);
4952 }
4953
4954 /*
4955  * authentication
4956  */
4957
4958 /*
4959  * Note: returned pointer is the address of a structure that's
4960  * managed separately.  Caller must *not* attempt to free it.
4961  */
4962 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
4963                                         int *proto, int force_new)
4964 {
4965         struct ceph_mds_session *s = con->private;
4966         struct ceph_mds_client *mdsc = s->s_mdsc;
4967         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4968         struct ceph_auth_handshake *auth = &s->s_auth;
4969
4970         if (force_new && auth->authorizer) {
4971                 ceph_auth_destroy_authorizer(auth->authorizer);
4972                 auth->authorizer = NULL;
4973         }
4974         if (!auth->authorizer) {
4975                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4976                                                       auth);
4977                 if (ret)
4978                         return ERR_PTR(ret);
4979         } else {
4980                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4981                                                       auth);
4982                 if (ret)
4983                         return ERR_PTR(ret);
4984         }
4985         *proto = ac->protocol;
4986
4987         return auth;
4988 }
4989
4990 static int add_authorizer_challenge(struct ceph_connection *con,
4991                                     void *challenge_buf, int challenge_buf_len)
4992 {
4993         struct ceph_mds_session *s = con->private;
4994         struct ceph_mds_client *mdsc = s->s_mdsc;
4995         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4996
4997         return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
4998                                             challenge_buf, challenge_buf_len);
4999 }
5000
5001 static int verify_authorizer_reply(struct ceph_connection *con)
5002 {
5003         struct ceph_mds_session *s = con->private;
5004         struct ceph_mds_client *mdsc = s->s_mdsc;
5005         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5006
5007         return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
5008 }
5009
5010 static int invalidate_authorizer(struct ceph_connection *con)
5011 {
5012         struct ceph_mds_session *s = con->private;
5013         struct ceph_mds_client *mdsc = s->s_mdsc;
5014         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5015
5016         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
5017
5018         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
5019 }
5020
5021 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
5022                                 struct ceph_msg_header *hdr, int *skip)
5023 {
5024         struct ceph_msg *msg;
5025         int type = (int) le16_to_cpu(hdr->type);
5026         int front_len = (int) le32_to_cpu(hdr->front_len);
5027
5028         if (con->in_msg)
5029                 return con->in_msg;
5030
5031         *skip = 0;
5032         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
5033         if (!msg) {
5034                 pr_err("unable to allocate msg type %d len %d\n",
5035                        type, front_len);
5036                 return NULL;
5037         }
5038
5039         return msg;
5040 }
5041
5042 static int mds_sign_message(struct ceph_msg *msg)
5043 {
5044        struct ceph_mds_session *s = msg->con->private;
5045        struct ceph_auth_handshake *auth = &s->s_auth;
5046
5047        return ceph_auth_sign_message(auth, msg);
5048 }
5049
5050 static int mds_check_message_signature(struct ceph_msg *msg)
5051 {
5052        struct ceph_mds_session *s = msg->con->private;
5053        struct ceph_auth_handshake *auth = &s->s_auth;
5054
5055        return ceph_auth_check_message_signature(auth, msg);
5056 }
5057
5058 static const struct ceph_connection_operations mds_con_ops = {
5059         .get = con_get,
5060         .put = con_put,
5061         .dispatch = dispatch,
5062         .get_authorizer = get_authorizer,
5063         .add_authorizer_challenge = add_authorizer_challenge,
5064         .verify_authorizer_reply = verify_authorizer_reply,
5065         .invalidate_authorizer = invalidate_authorizer,
5066         .peer_reset = peer_reset,
5067         .alloc_msg = mds_alloc_msg,
5068         .sign_message = mds_sign_message,
5069         .check_message_signature = mds_check_message_signature,
5070 };
5071
5072 /* eof */