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 #include <linux/bitmap.h>
  15
  16 #include "super.h"
  17 #include "mds_client.h"
  18
  19 #include <linux/ceph/ceph_features.h>
  20 #include <linux/ceph/messenger.h>
  21 #include <linux/ceph/decode.h>
  22 #include <linux/ceph/pagelist.h>
  23 #include <linux/ceph/auth.h>
  24 #include <linux/ceph/debugfs.h>
  25
  26 #define RECONNECT_MAX_SIZE (INT_MAX - PAGE_SIZE)
  27
  28 /*
  29  * A cluster of MDS (metadata server) daemons is responsible for
  30  * managing the file system namespace (the directory hierarchy and
  31  * inodes) and for coordinating shared access to storage.  Metadata is
  32  * partitioning hierarchically across a number of servers, and that
  33  * partition varies over time as the cluster adjusts the distribution
  34  * in order to balance load.
  35  *
  36  * The MDS client is primarily responsible to managing synchronous
  37  * metadata requests for operations like open, unlink, and so forth.
  38  * If there is a MDS failure, we find out about it when we (possibly
  39  * request and) receive a new MDS map, and can resubmit affected
  40  * requests.
  41  *
  42  * For the most part, though, we take advantage of a lossless
  43  * communications channel to the MDS, and do not need to worry about
  44  * timing out or resubmitting requests.
  45  *
  46  * We maintain a stateful "session" with each MDS we interact with.
  47  * Within each session, we sent periodic heartbeat messages to ensure
  48  * any capabilities or leases we have been issues remain valid.  If
  49  * the session times out and goes stale, our leases and capabilities
  50  * are no longer valid.
  51  */
  52
  53 struct ceph_reconnect_state {
  54         struct ceph_mds_session *session;
  55         int nr_caps, nr_realms;
  56         struct ceph_pagelist *pagelist;
  57         unsigned msg_version;
  58         bool allow_multi;
  59 };
  60
  61 static void __wake_requests(struct ceph_mds_client *mdsc,
  62                             struct list_head *head);
  63 static void ceph_cap_release_work(struct work_struct *work);
  64 static void ceph_cap_reclaim_work(struct work_struct *work);
  65
  66 static const struct ceph_connection_operations mds_con_ops;
  67
  68
  69 /*
  70  * mds reply parsing
  71  */
  72
  73 static int parse_reply_info_quota(void **p, void *end,
  74                                   struct ceph_mds_reply_info_in *info)
  75 {
  76         u8 struct_v, struct_compat;
  77         u32 struct_len;
  78
  79         ceph_decode_8_safe(p, end, struct_v, bad);
  80         ceph_decode_8_safe(p, end, struct_compat, bad);
  81         /* struct_v is expected to be >= 1. we only
  82          * understand encoding with struct_compat == 1. */
  83         if (!struct_v || struct_compat != 1)
  84                 goto bad;
  85         ceph_decode_32_safe(p, end, struct_len, bad);
  86         ceph_decode_need(p, end, struct_len, bad);
  87         end = *p + struct_len;
  88         ceph_decode_64_safe(p, end, info->max_bytes, bad);
  89         ceph_decode_64_safe(p, end, info->max_files, bad);
  90         *p = end;
  91         return 0;
  92 bad:
  93         return -EIO;
  94 }
  95
  96 /*
  97  * parse individual inode info
  98  */
  99 static int parse_reply_info_in(void **p, void *end,
 100                                struct ceph_mds_reply_info_in *info,
 101                                u64 features)
 102 {
 103         int err = 0;
 104         u8 struct_v = 0;
 105
 106         if (features == (u64)-1) {
 107                 u32 struct_len;
 108                 u8 struct_compat;
 109                 ceph_decode_8_safe(p, end, struct_v, bad);
 110                 ceph_decode_8_safe(p, end, struct_compat, bad);
 111                 /* struct_v is expected to be >= 1. we only understand
 112                  * encoding with struct_compat == 1. */
 113                 if (!struct_v || struct_compat != 1)
 114                         goto bad;
 115                 ceph_decode_32_safe(p, end, struct_len, bad);
 116                 ceph_decode_need(p, end, struct_len, bad);
 117                 end = *p + struct_len;
 118         }
 119
 120         ceph_decode_need(p, end, sizeof(struct ceph_mds_reply_inode), bad);
 121         info->in = *p;
 122         *p += sizeof(struct ceph_mds_reply_inode) +
 123                 sizeof(*info->in->fragtree.splits) *
 124                 le32_to_cpu(info->in->fragtree.nsplits);
 125
 126         ceph_decode_32_safe(p, end, info->symlink_len, bad);
 127         ceph_decode_need(p, end, info->symlink_len, bad);
 128         info->symlink = *p;
 129         *p += info->symlink_len;
 130
 131         ceph_decode_copy_safe(p, end, &info->dir_layout,
 132                               sizeof(info->dir_layout), bad);
 133         ceph_decode_32_safe(p, end, info->xattr_len, bad);
 134         ceph_decode_need(p, end, info->xattr_len, bad);
 135         info->xattr_data = *p;
 136         *p += info->xattr_len;
 137
 138         if (features == (u64)-1) {
 139                 /* inline data */
 140                 ceph_decode_64_safe(p, end, info->inline_version, bad);
 141                 ceph_decode_32_safe(p, end, info->inline_len, bad);
 142                 ceph_decode_need(p, end, info->inline_len, bad);
 143                 info->inline_data = *p;
 144                 *p += info->inline_len;
 145                 /* quota */
 146                 err = parse_reply_info_quota(p, end, info);
 147                 if (err < 0)
 148                         goto out_bad;
 149                 /* pool namespace */
 150                 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
 151                 if (info->pool_ns_len > 0) {
 152                         ceph_decode_need(p, end, info->pool_ns_len, bad);
 153                         info->pool_ns_data = *p;
 154                         *p += info->pool_ns_len;
 155                 }
 156
 157                 /* btime */
 158                 ceph_decode_need(p, end, sizeof(info->btime), bad);
 159                 ceph_decode_copy(p, &info->btime, sizeof(info->btime));
 160
 161                 /* change attribute */
 162                 ceph_decode_64_safe(p, end, info->change_attr, bad);
 163
 164                 /* dir pin */
 165                 if (struct_v >= 2) {
 166                         ceph_decode_32_safe(p, end, info->dir_pin, bad);
 167                 } else {
 168                         info->dir_pin = -ENODATA;
 169                 }
 170
 171                 /* snapshot birth time, remains zero for v<=2 */
 172                 if (struct_v >= 3) {
 173                         ceph_decode_need(p, end, sizeof(info->snap_btime), bad);
 174                         ceph_decode_copy(p, &info->snap_btime,
 175                                          sizeof(info->snap_btime));
 176                 } else {
 177                         memset(&info->snap_btime, 0, sizeof(info->snap_btime));
 178                 }
 179
 180                 /* snapshot count, remains zero for v<=3 */
 181                 if (struct_v >= 4) {
 182                         ceph_decode_64_safe(p, end, info->rsnaps, bad);
 183                 } else {
 184                         info->rsnaps = 0;
 185                 }
 186
 187                 *p = end;
 188         } else {
 189                 if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
 190                         ceph_decode_64_safe(p, end, info->inline_version, bad);
 191                         ceph_decode_32_safe(p, end, info->inline_len, bad);
 192                         ceph_decode_need(p, end, info->inline_len, bad);
 193                         info->inline_data = *p;
 194                         *p += info->inline_len;
 195                 } else
 196                         info->inline_version = CEPH_INLINE_NONE;
 197
 198                 if (features & CEPH_FEATURE_MDS_QUOTA) {
 199                         err = parse_reply_info_quota(p, end, info);
 200                         if (err < 0)
 201                                 goto out_bad;
 202                 } else {
 203                         info->max_bytes = 0;
 204                         info->max_files = 0;
 205                 }
 206
 207                 info->pool_ns_len = 0;
 208                 info->pool_ns_data = NULL;
 209                 if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
 210                         ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
 211                         if (info->pool_ns_len > 0) {
 212                                 ceph_decode_need(p, end, info->pool_ns_len, bad);
 213                                 info->pool_ns_data = *p;
 214                                 *p += info->pool_ns_len;
 215                         }
 216                 }
 217
 218                 if (features & CEPH_FEATURE_FS_BTIME) {
 219                         ceph_decode_need(p, end, sizeof(info->btime), bad);
 220                         ceph_decode_copy(p, &info->btime, sizeof(info->btime));
 221                         ceph_decode_64_safe(p, end, info->change_attr, bad);
 222                 }
 223
 224                 info->dir_pin = -ENODATA;
 225                 /* info->snap_btime and info->rsnaps remain zero */
 226         }
 227         return 0;
 228 bad:
 229         err = -EIO;
 230 out_bad:
 231         return err;
 232 }
 233
 234 static int parse_reply_info_dir(void **p, void *end,
 235                                 struct ceph_mds_reply_dirfrag **dirfrag,
 236                                 u64 features)
 237 {
 238         if (features == (u64)-1) {
 239                 u8 struct_v, struct_compat;
 240                 u32 struct_len;
 241                 ceph_decode_8_safe(p, end, struct_v, bad);
 242                 ceph_decode_8_safe(p, end, struct_compat, bad);
 243                 /* struct_v is expected to be >= 1. we only understand
 244                  * encoding whose struct_compat == 1. */
 245                 if (!struct_v || struct_compat != 1)
 246                         goto bad;
 247                 ceph_decode_32_safe(p, end, struct_len, bad);
 248                 ceph_decode_need(p, end, struct_len, bad);
 249                 end = *p + struct_len;
 250         }
 251
 252         ceph_decode_need(p, end, sizeof(**dirfrag), bad);
 253         *dirfrag = *p;
 254         *p += sizeof(**dirfrag) + sizeof(u32) * le32_to_cpu((*dirfrag)->ndist);
 255         if (unlikely(*p > end))
 256                 goto bad;
 257         if (features == (u64)-1)
 258                 *p = end;
 259         return 0;
 260 bad:
 261         return -EIO;
 262 }
 263
 264 static int parse_reply_info_lease(void **p, void *end,
 265                                   struct ceph_mds_reply_lease **lease,
 266                                   u64 features)
 267 {
 268         if (features == (u64)-1) {
 269                 u8 struct_v, struct_compat;
 270                 u32 struct_len;
 271                 ceph_decode_8_safe(p, end, struct_v, bad);
 272                 ceph_decode_8_safe(p, end, struct_compat, bad);
 273                 /* struct_v is expected to be >= 1. we only understand
 274                  * encoding whose struct_compat == 1. */
 275                 if (!struct_v || struct_compat != 1)
 276                         goto bad;
 277                 ceph_decode_32_safe(p, end, struct_len, bad);
 278                 ceph_decode_need(p, end, struct_len, bad);
 279                 end = *p + struct_len;
 280         }
 281
 282         ceph_decode_need(p, end, sizeof(**lease), bad);
 283         *lease = *p;
 284         *p += sizeof(**lease);
 285         if (features == (u64)-1)
 286                 *p = end;
 287         return 0;
 288 bad:
 289         return -EIO;
 290 }
 291
 292 /*
 293  * parse a normal reply, which may contain a (dir+)dentry and/or a
 294  * target inode.
 295  */
 296 static int parse_reply_info_trace(void **p, void *end,
 297                                   struct ceph_mds_reply_info_parsed *info,
 298                                   u64 features)
 299 {
 300         int err;
 301
 302         if (info->head->is_dentry) {
 303                 err = parse_reply_info_in(p, end, &info->diri, features);
 304                 if (err < 0)
 305                         goto out_bad;
 306
 307                 err = parse_reply_info_dir(p, end, &info->dirfrag, features);
 308                 if (err < 0)
 309                         goto out_bad;
 310
 311                 ceph_decode_32_safe(p, end, info->dname_len, bad);
 312                 ceph_decode_need(p, end, info->dname_len, bad);
 313                 info->dname = *p;
 314                 *p += info->dname_len;
 315
 316                 err = parse_reply_info_lease(p, end, &info->dlease, features);
 317                 if (err < 0)
 318                         goto out_bad;
 319         }
 320
 321         if (info->head->is_target) {
 322                 err = parse_reply_info_in(p, end, &info->targeti, features);
 323                 if (err < 0)
 324                         goto out_bad;
 325         }
 326
 327         if (unlikely(*p != end))
 328                 goto bad;
 329         return 0;
 330
 331 bad:
 332         err = -EIO;
 333 out_bad:
 334         pr_err("problem parsing mds trace %d\n", err);
 335         return err;
 336 }
 337
 338 /*
 339  * parse readdir results
 340  */
 341 static int parse_reply_info_readdir(void **p, void *end,
 342                                 struct ceph_mds_reply_info_parsed *info,
 343                                 u64 features)
 344 {
 345         u32 num, i = 0;
 346         int err;
 347
 348         err = parse_reply_info_dir(p, end, &info->dir_dir, features);
 349         if (err < 0)
 350                 goto out_bad;
 351
 352         ceph_decode_need(p, end, sizeof(num) + 2, bad);
 353         num = ceph_decode_32(p);
 354         {
 355                 u16 flags = ceph_decode_16(p);
 356                 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
 357                 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
 358                 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
 359                 info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
 360         }
 361         if (num == 0)
 362                 goto done;
 363
 364         BUG_ON(!info->dir_entries);
 365         if ((unsigned long)(info->dir_entries + num) >
 366             (unsigned long)info->dir_entries + info->dir_buf_size) {
 367                 pr_err("dir contents are larger than expected\n");
 368                 WARN_ON(1);
 369                 goto bad;
 370         }
 371
 372         info->dir_nr = num;
 373         while (num) {
 374                 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
 375                 /* dentry */
 376                 ceph_decode_32_safe(p, end, rde->name_len, bad);
 377                 ceph_decode_need(p, end, rde->name_len, bad);
 378                 rde->name = *p;
 379                 *p += rde->name_len;
 380                 dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
 381
 382                 /* dentry lease */
 383                 err = parse_reply_info_lease(p, end, &rde->lease, features);
 384                 if (err)
 385                         goto out_bad;
 386                 /* inode */
 387                 err = parse_reply_info_in(p, end, &rde->inode, features);
 388                 if (err < 0)
 389                         goto out_bad;
 390                 /* ceph_readdir_prepopulate() will update it */
 391                 rde->offset = 0;
 392                 i++;
 393                 num--;
 394         }
 395
 396 done:
 397         /* Skip over any unrecognized fields */
 398         *p = end;
 399         return 0;
 400
 401 bad:
 402         err = -EIO;
 403 out_bad:
 404         pr_err("problem parsing dir contents %d\n", err);
 405         return err;
 406 }
 407
 408 /*
 409  * parse fcntl F_GETLK results
 410  */
 411 static int parse_reply_info_filelock(void **p, void *end,
 412                                      struct ceph_mds_reply_info_parsed *info,
 413                                      u64 features)
 414 {
 415         if (*p + sizeof(*info->filelock_reply) > end)
 416                 goto bad;
 417
 418         info->filelock_reply = *p;
 419
 420         /* Skip over any unrecognized fields */
 421         *p = end;
 422         return 0;
 423 bad:
 424         return -EIO;
 425 }
 426
 427
 428 #if BITS_PER_LONG == 64
 429
 430 #define DELEGATED_INO_AVAILABLE         xa_mk_value(1)
 431
 432 static int ceph_parse_deleg_inos(void **p, void *end,
 433                                  struct ceph_mds_session *s)
 434 {
 435         u32 sets;
 436
 437         ceph_decode_32_safe(p, end, sets, bad);
 438         dout("got %u sets of delegated inodes\n", sets);
 439         while (sets--) {
 440                 u64 start, len;
 441
 442                 ceph_decode_64_safe(p, end, start, bad);
 443                 ceph_decode_64_safe(p, end, len, bad);
 444
 445                 /* Don't accept a delegation of system inodes */
 446                 if (start < CEPH_INO_SYSTEM_BASE) {
 447                         pr_warn_ratelimited("ceph: ignoring reserved inode range delegation (start=0x%llx len=0x%llx)\n",
 448                                         start, len);
 449                         continue;
 450                 }
 451                 while (len--) {
 452                         int err = xa_insert(&s->s_delegated_inos, start++,
 453                                             DELEGATED_INO_AVAILABLE,
 454                                             GFP_KERNEL);
 455                         if (!err) {
 456                                 dout("added delegated inode 0x%llx\n",
 457                                      start - 1);
 458                         } else if (err == -EBUSY) {
 459                                 pr_warn("MDS delegated inode 0x%llx more than once.\n",
 460                                         start - 1);
 461                         } else {
 462                                 return err;
 463                         }
 464                 }
 465         }
 466         return 0;
 467 bad:
 468         return -EIO;
 469 }
 470
 471 u64 ceph_get_deleg_ino(struct ceph_mds_session *s)
 472 {
 473         unsigned long ino;
 474         void *val;
 475
 476         xa_for_each(&s->s_delegated_inos, ino, val) {
 477                 val = xa_erase(&s->s_delegated_inos, ino);
 478                 if (val == DELEGATED_INO_AVAILABLE)
 479                         return ino;
 480         }
 481         return 0;
 482 }
 483
 484 int ceph_restore_deleg_ino(struct ceph_mds_session *s, u64 ino)
 485 {
 486         return xa_insert(&s->s_delegated_inos, ino, DELEGATED_INO_AVAILABLE,
 487                          GFP_KERNEL);
 488 }
 489 #else /* BITS_PER_LONG == 64 */
 490 /*
 491  * FIXME: xarrays can't handle 64-bit indexes on a 32-bit arch. For now, just
 492  * ignore delegated_inos on 32 bit arch. Maybe eventually add xarrays for top
 493  * and bottom words?
 494  */
 495 static int ceph_parse_deleg_inos(void **p, void *end,
 496                                  struct ceph_mds_session *s)
 497 {
 498         u32 sets;
 499
 500         ceph_decode_32_safe(p, end, sets, bad);
 501         if (sets)
 502                 ceph_decode_skip_n(p, end, sets * 2 * sizeof(__le64), bad);
 503         return 0;
 504 bad:
 505         return -EIO;
 506 }
 507
 508 u64 ceph_get_deleg_ino(struct ceph_mds_session *s)
 509 {
 510         return 0;
 511 }
 512
 513 int ceph_restore_deleg_ino(struct ceph_mds_session *s, u64 ino)
 514 {
 515         return 0;
 516 }
 517 #endif /* BITS_PER_LONG == 64 */
 518
 519 /*
 520  * parse create results
 521  */
 522 static int parse_reply_info_create(void **p, void *end,
 523                                   struct ceph_mds_reply_info_parsed *info,
 524                                   u64 features, struct ceph_mds_session *s)
 525 {
 526         int ret;
 527
 528         if (features == (u64)-1 ||
 529             (features & CEPH_FEATURE_REPLY_CREATE_INODE)) {
 530                 if (*p == end) {
 531                         /* Malformed reply? */
 532                         info->has_create_ino = false;
 533                 } else if (test_bit(CEPHFS_FEATURE_DELEG_INO, &s->s_features)) {
 534                         info->has_create_ino = true;
 535                         /* struct_v, struct_compat, and len */
 536                         ceph_decode_skip_n(p, end, 2 + sizeof(u32), bad);
 537                         ceph_decode_64_safe(p, end, info->ino, bad);
 538                         ret = ceph_parse_deleg_inos(p, end, s);
 539                         if (ret)
 540                                 return ret;
 541                 } else {
 542                         /* legacy */
 543                         ceph_decode_64_safe(p, end, info->ino, bad);
 544                         info->has_create_ino = true;
 545                 }
 546         } else {
 547                 if (*p != end)
 548                         goto bad;
 549         }
 550
 551         /* Skip over any unrecognized fields */
 552         *p = end;
 553         return 0;
 554 bad:
 555         return -EIO;
 556 }
 557
 558 static int parse_reply_info_getvxattr(void **p, void *end,
 559                                       struct ceph_mds_reply_info_parsed *info,
 560                                       u64 features)
 561 {
 562         u32 value_len;
 563
 564         ceph_decode_skip_8(p, end, bad); /* skip current version: 1 */
 565         ceph_decode_skip_8(p, end, bad); /* skip first version: 1 */
 566         ceph_decode_skip_32(p, end, bad); /* skip payload length */
 567
 568         ceph_decode_32_safe(p, end, value_len, bad);
 569
 570         if (value_len == end - *p) {
 571           info->xattr_info.xattr_value = *p;
 572           info->xattr_info.xattr_value_len = value_len;
 573           *p = end;
 574           return value_len;
 575         }
 576 bad:
 577         return -EIO;
 578 }
 579
 580 /*
 581  * parse extra results
 582  */
 583 static int parse_reply_info_extra(void **p, void *end,
 584                                   struct ceph_mds_reply_info_parsed *info,
 585                                   u64 features, struct ceph_mds_session *s)
 586 {
 587         u32 op = le32_to_cpu(info->head->op);
 588
 589         if (op == CEPH_MDS_OP_GETFILELOCK)
 590                 return parse_reply_info_filelock(p, end, info, features);
 591         else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
 592                 return parse_reply_info_readdir(p, end, info, features);
 593         else if (op == CEPH_MDS_OP_CREATE)
 594                 return parse_reply_info_create(p, end, info, features, s);
 595         else if (op == CEPH_MDS_OP_GETVXATTR)
 596                 return parse_reply_info_getvxattr(p, end, info, features);
 597         else
 598                 return -EIO;
 599 }
 600
 601 /*
 602  * parse entire mds reply
 603  */
 604 static int parse_reply_info(struct ceph_mds_session *s, struct ceph_msg *msg,
 605                             struct ceph_mds_reply_info_parsed *info,
 606                             u64 features)
 607 {
 608         void *p, *end;
 609         u32 len;
 610         int err;
 611
 612         info->head = msg->front.iov_base;
 613         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 614         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 615
 616         /* trace */
 617         ceph_decode_32_safe(&p, end, len, bad);
 618         if (len > 0) {
 619                 ceph_decode_need(&p, end, len, bad);
 620                 err = parse_reply_info_trace(&p, p+len, info, features);
 621                 if (err < 0)
 622                         goto out_bad;
 623         }
 624
 625         /* extra */
 626         ceph_decode_32_safe(&p, end, len, bad);
 627         if (len > 0) {
 628                 ceph_decode_need(&p, end, len, bad);
 629                 err = parse_reply_info_extra(&p, p+len, info, features, s);
 630                 if (err < 0)
 631                         goto out_bad;
 632         }
 633
 634         /* snap blob */
 635         ceph_decode_32_safe(&p, end, len, bad);
 636         info->snapblob_len = len;
 637         info->snapblob = p;
 638         p += len;
 639
 640         if (p != end)
 641                 goto bad;
 642         return 0;
 643
 644 bad:
 645         err = -EIO;
 646 out_bad:
 647         pr_err("mds parse_reply err %d\n", err);
 648         ceph_msg_dump(msg);
 649         return err;
 650 }
 651
 652 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
 653 {
 654         if (!info->dir_entries)
 655                 return;
 656         free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
 657 }
 658
 659 /*
 660  * In async unlink case the kclient won't wait for the first reply
 661  * from MDS and just drop all the links and unhash the dentry and then
 662  * succeeds immediately.
 663  *
 664  * For any new create/link/rename,etc requests followed by using the
 665  * same file names we must wait for the first reply of the inflight
 666  * unlink request, or the MDS possibly will fail these following
 667  * requests with -EEXIST if the inflight async unlink request was
 668  * delayed for some reasons.
 669  *
 670  * And the worst case is that for the none async openc request it will
 671  * successfully open the file if the CDentry hasn't been unlinked yet,
 672  * but later the previous delayed async unlink request will remove the
 673  * CDenty. That means the just created file is possiblly deleted later
 674  * by accident.
 675  *
 676  * We need to wait for the inflight async unlink requests to finish
 677  * when creating new files/directories by using the same file names.
 678  */
 679 int ceph_wait_on_conflict_unlink(struct dentry *dentry)
 680 {
 681         struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
 682         struct dentry *pdentry = dentry->d_parent;
 683         struct dentry *udentry, *found = NULL;
 684         struct ceph_dentry_info *di;
 685         struct qstr dname;
 686         u32 hash = dentry->d_name.hash;
 687         int err;
 688
 689         dname.name = dentry->d_name.name;
 690         dname.len = dentry->d_name.len;
 691
 692         rcu_read_lock();
 693         hash_for_each_possible_rcu(fsc->async_unlink_conflict, di,
 694                                    hnode, hash) {
 695                 udentry = di->dentry;
 696
 697                 spin_lock(&udentry->d_lock);
 698                 if (udentry->d_name.hash != hash)
 699                         goto next;
 700                 if (unlikely(udentry->d_parent != pdentry))
 701                         goto next;
 702                 if (!hash_hashed(&di->hnode))
 703                         goto next;
 704
 705                 if (!test_bit(CEPH_DENTRY_ASYNC_UNLINK_BIT, &di->flags))
 706                         pr_warn("%s dentry %p:%pd async unlink bit is not set\n",
 707                                 __func__, dentry, dentry);
 708
 709                 if (!d_same_name(udentry, pdentry, &dname))
 710                         goto next;
 711
 712                 found = dget_dlock(udentry);
 713                 spin_unlock(&udentry->d_lock);
 714                 break;
 715 next:
 716                 spin_unlock(&udentry->d_lock);
 717         }
 718         rcu_read_unlock();
 719
 720         if (likely(!found))
 721                 return 0;
 722
 723         dout("%s dentry %p:%pd conflict with old %p:%pd\n", __func__,
 724              dentry, dentry, found, found);
 725
 726         err = wait_on_bit(&di->flags, CEPH_DENTRY_ASYNC_UNLINK_BIT,
 727                           TASK_KILLABLE);
 728         dput(found);
 729         return err;
 730 }
 731
 732
 733 /*
 734  * sessions
 735  */
 736 const char *ceph_session_state_name(int s)
 737 {
 738         switch (s) {
 739         case CEPH_MDS_SESSION_NEW: return "new";
 740         case CEPH_MDS_SESSION_OPENING: return "opening";
 741         case CEPH_MDS_SESSION_OPEN: return "open";
 742         case CEPH_MDS_SESSION_HUNG: return "hung";
 743         case CEPH_MDS_SESSION_CLOSING: return "closing";
 744         case CEPH_MDS_SESSION_CLOSED: return "closed";
 745         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
 746         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
 747         case CEPH_MDS_SESSION_REJECTED: return "rejected";
 748         default: return "???";
 749         }
 750 }
 751
 752 struct ceph_mds_session *ceph_get_mds_session(struct ceph_mds_session *s)
 753 {
 754         if (refcount_inc_not_zero(&s->s_ref))
 755                 return s;
 756         return NULL;
 757 }
 758
 759 void ceph_put_mds_session(struct ceph_mds_session *s)
 760 {
 761         if (IS_ERR_OR_NULL(s))
 762                 return;
 763
 764         if (refcount_dec_and_test(&s->s_ref)) {
 765                 if (s->s_auth.authorizer)
 766                         ceph_auth_destroy_authorizer(s->s_auth.authorizer);
 767                 WARN_ON(mutex_is_locked(&s->s_mutex));
 768                 xa_destroy(&s->s_delegated_inos);
 769                 kfree(s);
 770         }
 771 }
 772
 773 /*
 774  * called under mdsc->mutex
 775  */
 776 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
 777                                                    int mds)
 778 {
 779         if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
 780                 return NULL;
 781         return ceph_get_mds_session(mdsc->sessions[mds]);
 782 }
 783
 784 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
 785 {
 786         if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
 787                 return false;
 788         else
 789                 return true;
 790 }
 791
 792 static int __verify_registered_session(struct ceph_mds_client *mdsc,
 793                                        struct ceph_mds_session *s)
 794 {
 795         if (s->s_mds >= mdsc->max_sessions ||
 796             mdsc->sessions[s->s_mds] != s)
 797                 return -ENOENT;
 798         return 0;
 799 }
 800
 801 /*
 802  * create+register a new session for given mds.
 803  * called under mdsc->mutex.
 804  */
 805 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 806                                                  int mds)
 807 {
 808         struct ceph_mds_session *s;
 809
 810         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_FENCE_IO)
 811                 return ERR_PTR(-EIO);
 812
 813         if (mds >= mdsc->mdsmap->possible_max_rank)
 814                 return ERR_PTR(-EINVAL);
 815
 816         s = kzalloc(sizeof(*s), GFP_NOFS);
 817         if (!s)
 818                 return ERR_PTR(-ENOMEM);
 819
 820         if (mds >= mdsc->max_sessions) {
 821                 int newmax = 1 << get_count_order(mds + 1);
 822                 struct ceph_mds_session **sa;
 823
 824                 dout("%s: realloc to %d\n", __func__, newmax);
 825                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 826                 if (!sa)
 827                         goto fail_realloc;
 828                 if (mdsc->sessions) {
 829                         memcpy(sa, mdsc->sessions,
 830                                mdsc->max_sessions * sizeof(void *));
 831                         kfree(mdsc->sessions);
 832                 }
 833                 mdsc->sessions = sa;
 834                 mdsc->max_sessions = newmax;
 835         }
 836
 837         dout("%s: mds%d\n", __func__, mds);
 838         s->s_mdsc = mdsc;
 839         s->s_mds = mds;
 840         s->s_state = CEPH_MDS_SESSION_NEW;
 841         mutex_init(&s->s_mutex);
 842
 843         ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
 844
 845         atomic_set(&s->s_cap_gen, 1);
 846         s->s_cap_ttl = jiffies - 1;
 847
 848         spin_lock_init(&s->s_cap_lock);
 849         INIT_LIST_HEAD(&s->s_caps);
 850         refcount_set(&s->s_ref, 1);
 851         INIT_LIST_HEAD(&s->s_waiting);
 852         INIT_LIST_HEAD(&s->s_unsafe);
 853         xa_init(&s->s_delegated_inos);
 854         INIT_LIST_HEAD(&s->s_cap_releases);
 855         INIT_WORK(&s->s_cap_release_work, ceph_cap_release_work);
 856
 857         INIT_LIST_HEAD(&s->s_cap_dirty);
 858         INIT_LIST_HEAD(&s->s_cap_flushing);
 859
 860         mdsc->sessions[mds] = s;
 861         atomic_inc(&mdsc->num_sessions);
 862         refcount_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
 863
 864         ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
 865                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
 866
 867         return s;
 868
 869 fail_realloc:
 870         kfree(s);
 871         return ERR_PTR(-ENOMEM);
 872 }
 873
 874 /*
 875  * called under mdsc->mutex
 876  */
 877 static void __unregister_session(struct ceph_mds_client *mdsc,
 878                                struct ceph_mds_session *s)
 879 {
 880         dout("__unregister_session mds%d %p\n", s->s_mds, s);
 881         BUG_ON(mdsc->sessions[s->s_mds] != s);
 882         mdsc->sessions[s->s_mds] = NULL;
 883         ceph_con_close(&s->s_con);
 884         ceph_put_mds_session(s);
 885         atomic_dec(&mdsc->num_sessions);
 886 }
 887
 888 /*
 889  * drop session refs in request.
 890  *
 891  * should be last request ref, or hold mdsc->mutex
 892  */
 893 static void put_request_session(struct ceph_mds_request *req)
 894 {
 895         if (req->r_session) {
 896                 ceph_put_mds_session(req->r_session);
 897                 req->r_session = NULL;
 898         }
 899 }
 900
 901 void ceph_mdsc_iterate_sessions(struct ceph_mds_client *mdsc,
 902                                 void (*cb)(struct ceph_mds_session *),
 903                                 bool check_state)
 904 {
 905         int mds;
 906
 907         mutex_lock(&mdsc->mutex);
 908         for (mds = 0; mds < mdsc->max_sessions; ++mds) {
 909                 struct ceph_mds_session *s;
 910
 911                 s = __ceph_lookup_mds_session(mdsc, mds);
 912                 if (!s)
 913                         continue;
 914
 915                 if (check_state && !check_session_state(s)) {
 916                         ceph_put_mds_session(s);
 917                         continue;
 918                 }
 919
 920                 mutex_unlock(&mdsc->mutex);
 921                 cb(s);
 922                 ceph_put_mds_session(s);
 923                 mutex_lock(&mdsc->mutex);
 924         }
 925         mutex_unlock(&mdsc->mutex);
 926 }
 927
 928 void ceph_mdsc_release_request(struct kref *kref)
 929 {
 930         struct ceph_mds_request *req = container_of(kref,
 931                                                     struct ceph_mds_request,
 932                                                     r_kref);
 933         ceph_mdsc_release_dir_caps_no_check(req);
 934         destroy_reply_info(&req->r_reply_info);
 935         if (req->r_request)
 936                 ceph_msg_put(req->r_request);
 937         if (req->r_reply)
 938                 ceph_msg_put(req->r_reply);
 939         if (req->r_inode) {
 940                 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
 941                 iput(req->r_inode);
 942         }
 943         if (req->r_parent) {
 944                 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
 945                 iput(req->r_parent);
 946         }
 947         iput(req->r_target_inode);
 948         if (req->r_dentry)
 949                 dput(req->r_dentry);
 950         if (req->r_old_dentry)
 951                 dput(req->r_old_dentry);
 952         if (req->r_old_dentry_dir) {
 953                 /*
 954                  * track (and drop pins for) r_old_dentry_dir
 955                  * separately, since r_old_dentry's d_parent may have
 956                  * changed between the dir mutex being dropped and
 957                  * this request being freed.
 958                  */
 959                 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
 960                                   CEPH_CAP_PIN);
 961                 iput(req->r_old_dentry_dir);
 962         }
 963         kfree(req->r_path1);
 964         kfree(req->r_path2);
 965         put_cred(req->r_cred);
 966         if (req->r_pagelist)
 967                 ceph_pagelist_release(req->r_pagelist);
 968         put_request_session(req);
 969         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
 970         WARN_ON_ONCE(!list_empty(&req->r_wait));
 971         kmem_cache_free(ceph_mds_request_cachep, req);
 972 }
 973
 974 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
 975
 976 /*
 977  * lookup session, bump ref if found.
 978  *
 979  * called under mdsc->mutex.
 980  */
 981 static struct ceph_mds_request *
 982 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
 983 {
 984         struct ceph_mds_request *req;
 985
 986         req = lookup_request(&mdsc->request_tree, tid);
 987         if (req)
 988                 ceph_mdsc_get_request(req);
 989
 990         return req;
 991 }
 992
 993 /*
 994  * Register an in-flight request, and assign a tid.  Link to directory
 995  * are modifying (if any).
 996  *
 997  * Called under mdsc->mutex.
 998  */
 999 static void __register_request(struct ceph_mds_client *mdsc,
1000                                struct ceph_mds_request *req,
1001                                struct inode *dir)
1002 {
1003         int ret = 0;
1004
1005         req->r_tid = ++mdsc->last_tid;
1006         if (req->r_num_caps) {
1007                 ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
1008                                         req->r_num_caps);
1009                 if (ret < 0) {
1010                         pr_err("__register_request %p "
1011                                "failed to reserve caps: %d\n", req, ret);
1012                         /* set req->r_err to fail early from __do_request */
1013                         req->r_err = ret;
1014                         return;
1015                 }
1016         }
1017         dout("__register_request %p tid %lld\n", req, req->r_tid);
1018         ceph_mdsc_get_request(req);
1019         insert_request(&mdsc->request_tree, req);
1020
1021         req->r_cred = get_current_cred();
1022
1023         if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
1024                 mdsc->oldest_tid = req->r_tid;
1025
1026         if (dir) {
1027                 struct ceph_inode_info *ci = ceph_inode(dir);
1028
1029                 ihold(dir);
1030                 req->r_unsafe_dir = dir;
1031                 spin_lock(&ci->i_unsafe_lock);
1032                 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
1033                 spin_unlock(&ci->i_unsafe_lock);
1034         }
1035 }
1036
1037 static void __unregister_request(struct ceph_mds_client *mdsc,
1038                                  struct ceph_mds_request *req)
1039 {
1040         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
1041
1042         /* Never leave an unregistered request on an unsafe list! */
1043         list_del_init(&req->r_unsafe_item);
1044
1045         if (req->r_tid == mdsc->oldest_tid) {
1046                 struct rb_node *p = rb_next(&req->r_node);
1047                 mdsc->oldest_tid = 0;
1048                 while (p) {
1049                         struct ceph_mds_request *next_req =
1050                                 rb_entry(p, struct ceph_mds_request, r_node);
1051                         if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
1052                                 mdsc->oldest_tid = next_req->r_tid;
1053                                 break;
1054                         }
1055                         p = rb_next(p);
1056                 }
1057         }
1058
1059         erase_request(&mdsc->request_tree, req);
1060
1061         if (req->r_unsafe_dir) {
1062                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
1063                 spin_lock(&ci->i_unsafe_lock);
1064                 list_del_init(&req->r_unsafe_dir_item);
1065                 spin_unlock(&ci->i_unsafe_lock);
1066         }
1067         if (req->r_target_inode &&
1068             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
1069                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
1070                 spin_lock(&ci->i_unsafe_lock);
1071                 list_del_init(&req->r_unsafe_target_item);
1072                 spin_unlock(&ci->i_unsafe_lock);
1073         }
1074
1075         if (req->r_unsafe_dir) {
1076                 iput(req->r_unsafe_dir);
1077                 req->r_unsafe_dir = NULL;
1078         }
1079
1080         complete_all(&req->r_safe_completion);
1081
1082         ceph_mdsc_put_request(req);
1083 }
1084
1085 /*
1086  * Walk back up the dentry tree until we hit a dentry representing a
1087  * non-snapshot inode. We do this using the rcu_read_lock (which must be held
1088  * when calling this) to ensure that the objects won't disappear while we're
1089  * working with them. Once we hit a candidate dentry, we attempt to take a
1090  * reference to it, and return that as the result.
1091  */
1092 static struct inode *get_nonsnap_parent(struct dentry *dentry)
1093 {
1094         struct inode *inode = NULL;
1095
1096         while (dentry && !IS_ROOT(dentry)) {
1097                 inode = d_inode_rcu(dentry);
1098                 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
1099                         break;
1100                 dentry = dentry->d_parent;
1101         }
1102         if (inode)
1103                 inode = igrab(inode);
1104         return inode;
1105 }
1106
1107 /*
1108  * Choose mds to send request to next.  If there is a hint set in the
1109  * request (e.g., due to a prior forward hint from the mds), use that.
1110  * Otherwise, consult frag tree and/or caps to identify the
1111  * appropriate mds.  If all else fails, choose randomly.
1112  *
1113  * Called under mdsc->mutex.
1114  */
1115 static int __choose_mds(struct ceph_mds_client *mdsc,
1116                         struct ceph_mds_request *req,
1117                         bool *random)
1118 {
1119         struct inode *inode;
1120         struct ceph_inode_info *ci;
1121         struct ceph_cap *cap;
1122         int mode = req->r_direct_mode;
1123         int mds = -1;
1124         u32 hash = req->r_direct_hash;
1125         bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
1126
1127         if (random)
1128                 *random = false;
1129
1130         /*
1131          * is there a specific mds we should try?  ignore hint if we have
1132          * no session and the mds is not up (active or recovering).
1133          */
1134         if (req->r_resend_mds >= 0 &&
1135             (__have_session(mdsc, req->r_resend_mds) ||
1136              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
1137                 dout("%s using resend_mds mds%d\n", __func__,
1138                      req->r_resend_mds);
1139                 return req->r_resend_mds;
1140         }
1141
1142         if (mode == USE_RANDOM_MDS)
1143                 goto random;
1144
1145         inode = NULL;
1146         if (req->r_inode) {
1147                 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
1148                         inode = req->r_inode;
1149                         ihold(inode);
1150                 } else {
1151                         /* req->r_dentry is non-null for LSSNAP request */
1152                         rcu_read_lock();
1153                         inode = get_nonsnap_parent(req->r_dentry);
1154                         rcu_read_unlock();
1155                         dout("%s using snapdir's parent %p\n", __func__, inode);
1156                 }
1157         } else if (req->r_dentry) {
1158                 /* ignore race with rename; old or new d_parent is okay */
1159                 struct dentry *parent;
1160                 struct inode *dir;
1161
1162                 rcu_read_lock();
1163                 parent = READ_ONCE(req->r_dentry->d_parent);
1164                 dir = req->r_parent ? : d_inode_rcu(parent);
1165
1166                 if (!dir || dir->i_sb != mdsc->fsc->sb) {
1167                         /*  not this fs or parent went negative */
1168                         inode = d_inode(req->r_dentry);
1169                         if (inode)
1170                                 ihold(inode);
1171                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
1172                         /* direct snapped/virtual snapdir requests
1173                          * based on parent dir inode */
1174                         inode = get_nonsnap_parent(parent);
1175                         dout("%s using nonsnap parent %p\n", __func__, inode);
1176                 } else {
1177                         /* dentry target */
1178                         inode = d_inode(req->r_dentry);
1179                         if (!inode || mode == USE_AUTH_MDS) {
1180                                 /* dir + name */
1181                                 inode = igrab(dir);
1182                                 hash = ceph_dentry_hash(dir, req->r_dentry);
1183                                 is_hash = true;
1184                         } else {
1185                                 ihold(inode);
1186                         }
1187                 }
1188                 rcu_read_unlock();
1189         }
1190
1191         dout("%s %p is_hash=%d (0x%x) mode %d\n", __func__, inode, (int)is_hash,
1192              hash, mode);
1193         if (!inode)
1194                 goto random;
1195         ci = ceph_inode(inode);
1196
1197         if (is_hash && S_ISDIR(inode->i_mode)) {
1198                 struct ceph_inode_frag frag;
1199                 int found;
1200
1201                 ceph_choose_frag(ci, hash, &frag, &found);
1202                 if (found) {
1203                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
1204                                 u8 r;
1205
1206                                 /* choose a random replica */
1207                                 get_random_bytes(&r, 1);
1208                                 r %= frag.ndist;
1209                                 mds = frag.dist[r];
1210                                 dout("%s %p %llx.%llx frag %u mds%d (%d/%d)\n",
1211                                      __func__, inode, ceph_vinop(inode),
1212                                      frag.frag, mds, (int)r, frag.ndist);
1213                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1214                                     CEPH_MDS_STATE_ACTIVE &&
1215                                     !ceph_mdsmap_is_laggy(mdsc->mdsmap, mds))
1216                                         goto out;
1217                         }
1218
1219                         /* since this file/dir wasn't known to be
1220                          * replicated, then we want to look for the
1221                          * authoritative mds. */
1222                         if (frag.mds >= 0) {
1223                                 /* choose auth mds */
1224                                 mds = frag.mds;
1225                                 dout("%s %p %llx.%llx frag %u mds%d (auth)\n",
1226                                      __func__, inode, ceph_vinop(inode),
1227                                      frag.frag, mds);
1228                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1229                                     CEPH_MDS_STATE_ACTIVE) {
1230                                         if (!ceph_mdsmap_is_laggy(mdsc->mdsmap,
1231                                                                   mds))
1232                                                 goto out;
1233                                 }
1234                         }
1235                         mode = USE_AUTH_MDS;
1236                 }
1237         }
1238
1239         spin_lock(&ci->i_ceph_lock);
1240         cap = NULL;
1241         if (mode == USE_AUTH_MDS)
1242                 cap = ci->i_auth_cap;
1243         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
1244                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
1245         if (!cap) {
1246                 spin_unlock(&ci->i_ceph_lock);
1247                 iput(inode);
1248                 goto random;
1249         }
1250         mds = cap->session->s_mds;
1251         dout("%s %p %llx.%llx mds%d (%scap %p)\n", __func__,
1252              inode, ceph_vinop(inode), mds,
1253              cap == ci->i_auth_cap ? "auth " : "", cap);
1254         spin_unlock(&ci->i_ceph_lock);
1255 out:
1256         iput(inode);
1257         return mds;
1258
1259 random:
1260         if (random)
1261                 *random = true;
1262
1263         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
1264         dout("%s chose random mds%d\n", __func__, mds);
1265         return mds;
1266 }
1267
1268
1269 /*
1270  * session messages
1271  */
1272 struct ceph_msg *ceph_create_session_msg(u32 op, u64 seq)
1273 {
1274         struct ceph_msg *msg;
1275         struct ceph_mds_session_head *h;
1276
1277         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
1278                            false);
1279         if (!msg) {
1280                 pr_err("ENOMEM creating session %s msg\n",
1281                        ceph_session_op_name(op));
1282                 return NULL;
1283         }
1284         h = msg->front.iov_base;
1285         h->op = cpu_to_le32(op);
1286         h->seq = cpu_to_le64(seq);
1287
1288         return msg;
1289 }
1290
1291 static const unsigned char feature_bits[] = CEPHFS_FEATURES_CLIENT_SUPPORTED;
1292 #define FEATURE_BYTES(c) (DIV_ROUND_UP((size_t)feature_bits[c - 1] + 1, 64) * 8)
1293 static int encode_supported_features(void **p, void *end)
1294 {
1295         static const size_t count = ARRAY_SIZE(feature_bits);
1296
1297         if (count > 0) {
1298                 size_t i;
1299                 size_t size = FEATURE_BYTES(count);
1300                 unsigned long bit;
1301
1302                 if (WARN_ON_ONCE(*p + 4 + size > end))
1303                         return -ERANGE;
1304
1305                 ceph_encode_32(p, size);
1306                 memset(*p, 0, size);
1307                 for (i = 0; i < count; i++) {
1308                         bit = feature_bits[i];
1309                         ((unsigned char *)(*p))[bit / 8] |= BIT(bit % 8);
1310                 }
1311                 *p += size;
1312         } else {
1313                 if (WARN_ON_ONCE(*p + 4 > end))
1314                         return -ERANGE;
1315
1316                 ceph_encode_32(p, 0);
1317         }
1318
1319         return 0;
1320 }
1321
1322 static const unsigned char metric_bits[] = CEPHFS_METRIC_SPEC_CLIENT_SUPPORTED;
1323 #define METRIC_BYTES(cnt) (DIV_ROUND_UP((size_t)metric_bits[cnt - 1] + 1, 64) * 8)
1324 static int encode_metric_spec(void **p, void *end)
1325 {
1326         static const size_t count = ARRAY_SIZE(metric_bits);
1327
1328         /* header */
1329         if (WARN_ON_ONCE(*p + 2 > end))
1330                 return -ERANGE;
1331
1332         ceph_encode_8(p, 1); /* version */
1333         ceph_encode_8(p, 1); /* compat */
1334
1335         if (count > 0) {
1336                 size_t i;
1337                 size_t size = METRIC_BYTES(count);
1338
1339                 if (WARN_ON_ONCE(*p + 4 + 4 + size > end))
1340                         return -ERANGE;
1341
1342                 /* metric spec info length */
1343                 ceph_encode_32(p, 4 + size);
1344
1345                 /* metric spec */
1346                 ceph_encode_32(p, size);
1347                 memset(*p, 0, size);
1348                 for (i = 0; i < count; i++)
1349                         ((unsigned char *)(*p))[i / 8] |= BIT(metric_bits[i] % 8);
1350                 *p += size;
1351         } else {
1352                 if (WARN_ON_ONCE(*p + 4 + 4 > end))
1353                         return -ERANGE;
1354
1355                 /* metric spec info length */
1356                 ceph_encode_32(p, 4);
1357                 /* metric spec */
1358                 ceph_encode_32(p, 0);
1359         }
1360
1361         return 0;
1362 }
1363
1364 /*
1365  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
1366  * to include additional client metadata fields.
1367  */
1368 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
1369 {
1370         struct ceph_msg *msg;
1371         struct ceph_mds_session_head *h;
1372         int i;
1373         int extra_bytes = 0;
1374         int metadata_key_count = 0;
1375         struct ceph_options *opt = mdsc->fsc->client->options;
1376         struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
1377         size_t size, count;
1378         void *p, *end;
1379         int ret;
1380
1381         const char* metadata[][2] = {
1382                 {"hostname", mdsc->nodename},
1383                 {"kernel_version", init_utsname()->release},
1384                 {"entity_id", opt->name ? : ""},
1385                 {"root", fsopt->server_path ? : "/"},
1386                 {NULL, NULL}
1387         };
1388
1389         /* Calculate serialized length of metadata */
1390         extra_bytes = 4;  /* map length */
1391         for (i = 0; metadata[i][0]; ++i) {
1392                 extra_bytes += 8 + strlen(metadata[i][0]) +
1393                         strlen(metadata[i][1]);
1394                 metadata_key_count++;
1395         }
1396
1397         /* supported feature */
1398         size = 0;
1399         count = ARRAY_SIZE(feature_bits);
1400         if (count > 0)
1401                 size = FEATURE_BYTES(count);
1402         extra_bytes += 4 + size;
1403
1404         /* metric spec */
1405         size = 0;
1406         count = ARRAY_SIZE(metric_bits);
1407         if (count > 0)
1408                 size = METRIC_BYTES(count);
1409         extra_bytes += 2 + 4 + 4 + size;
1410
1411         /* Allocate the message */
1412         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + extra_bytes,
1413                            GFP_NOFS, false);
1414         if (!msg) {
1415                 pr_err("ENOMEM creating session open msg\n");
1416                 return ERR_PTR(-ENOMEM);
1417         }
1418         p = msg->front.iov_base;
1419         end = p + msg->front.iov_len;
1420
1421         h = p;
1422         h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
1423         h->seq = cpu_to_le64(seq);
1424
1425         /*
1426          * Serialize client metadata into waiting buffer space, using
1427          * the format that userspace expects for map<string, string>
1428          *
1429          * ClientSession messages with metadata are v4
1430          */
1431         msg->hdr.version = cpu_to_le16(4);
1432         msg->hdr.compat_version = cpu_to_le16(1);
1433
1434         /* The write pointer, following the session_head structure */
1435         p += sizeof(*h);
1436
1437         /* Number of entries in the map */
1438         ceph_encode_32(&p, metadata_key_count);
1439
1440         /* Two length-prefixed strings for each entry in the map */
1441         for (i = 0; metadata[i][0]; ++i) {
1442                 size_t const key_len = strlen(metadata[i][0]);
1443                 size_t const val_len = strlen(metadata[i][1]);
1444
1445                 ceph_encode_32(&p, key_len);
1446                 memcpy(p, metadata[i][0], key_len);
1447                 p += key_len;
1448                 ceph_encode_32(&p, val_len);
1449                 memcpy(p, metadata[i][1], val_len);
1450                 p += val_len;
1451         }
1452
1453         ret = encode_supported_features(&p, end);
1454         if (ret) {
1455                 pr_err("encode_supported_features failed!\n");
1456                 ceph_msg_put(msg);
1457                 return ERR_PTR(ret);
1458         }
1459
1460         ret = encode_metric_spec(&p, end);
1461         if (ret) {
1462                 pr_err("encode_metric_spec failed!\n");
1463                 ceph_msg_put(msg);
1464                 return ERR_PTR(ret);
1465         }
1466
1467         msg->front.iov_len = p - msg->front.iov_base;
1468         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1469
1470         return msg;
1471 }
1472
1473 /*
1474  * send session open request.
1475  *
1476  * called under mdsc->mutex
1477  */
1478 static int __open_session(struct ceph_mds_client *mdsc,
1479                           struct ceph_mds_session *session)
1480 {
1481         struct ceph_msg *msg;
1482         int mstate;
1483         int mds = session->s_mds;
1484
1485         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_FENCE_IO)
1486                 return -EIO;
1487
1488         /* wait for mds to go active? */
1489         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
1490         dout("open_session to mds%d (%s)\n", mds,
1491              ceph_mds_state_name(mstate));
1492         session->s_state = CEPH_MDS_SESSION_OPENING;
1493         session->s_renew_requested = jiffies;
1494
1495         /* send connect message */
1496         msg = create_session_open_msg(mdsc, session->s_seq);
1497         if (IS_ERR(msg))
1498                 return PTR_ERR(msg);
1499         ceph_con_send(&session->s_con, msg);
1500         return 0;
1501 }
1502
1503 /*
1504  * open sessions for any export targets for the given mds
1505  *
1506  * called under mdsc->mutex
1507  */
1508 static struct ceph_mds_session *
1509 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
1510 {
1511         struct ceph_mds_session *session;
1512         int ret;
1513
1514         session = __ceph_lookup_mds_session(mdsc, target);
1515         if (!session) {
1516                 session = register_session(mdsc, target);
1517                 if (IS_ERR(session))
1518                         return session;
1519         }
1520         if (session->s_state == CEPH_MDS_SESSION_NEW ||
1521             session->s_state == CEPH_MDS_SESSION_CLOSING) {
1522                 ret = __open_session(mdsc, session);
1523                 if (ret)
1524                         return ERR_PTR(ret);
1525         }
1526
1527         return session;
1528 }
1529
1530 struct ceph_mds_session *
1531 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1532 {
1533         struct ceph_mds_session *session;
1534
1535         dout("open_export_target_session to mds%d\n", target);
1536
1537         mutex_lock(&mdsc->mutex);
1538         session = __open_export_target_session(mdsc, target);
1539         mutex_unlock(&mdsc->mutex);
1540
1541         return session;
1542 }
1543
1544 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1545                                           struct ceph_mds_session *session)
1546 {
1547         struct ceph_mds_info *mi;
1548         struct ceph_mds_session *ts;
1549         int i, mds = session->s_mds;
1550
1551         if (mds >= mdsc->mdsmap->possible_max_rank)
1552                 return;
1553
1554         mi = &mdsc->mdsmap->m_info[mds];
1555         dout("open_export_target_sessions for mds%d (%d targets)\n",
1556              session->s_mds, mi->num_export_targets);
1557
1558         for (i = 0; i < mi->num_export_targets; i++) {
1559                 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1560                 ceph_put_mds_session(ts);
1561         }
1562 }
1563
1564 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1565                                            struct ceph_mds_session *session)
1566 {
1567         mutex_lock(&mdsc->mutex);
1568         __open_export_target_sessions(mdsc, session);
1569         mutex_unlock(&mdsc->mutex);
1570 }
1571
1572 /*
1573  * session caps
1574  */
1575
1576 static void detach_cap_releases(struct ceph_mds_session *session,
1577                                 struct list_head *target)
1578 {
1579         lockdep_assert_held(&session->s_cap_lock);
1580
1581         list_splice_init(&session->s_cap_releases, target);
1582         session->s_num_cap_releases = 0;
1583         dout("dispose_cap_releases mds%d\n", session->s_mds);
1584 }
1585
1586 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1587                                  struct list_head *dispose)
1588 {
1589         while (!list_empty(dispose)) {
1590                 struct ceph_cap *cap;
1591                 /* zero out the in-progress message */
1592                 cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1593                 list_del(&cap->session_caps);
1594                 ceph_put_cap(mdsc, cap);
1595         }
1596 }
1597
1598 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1599                                      struct ceph_mds_session *session)
1600 {
1601         struct ceph_mds_request *req;
1602         struct rb_node *p;
1603
1604         dout("cleanup_session_requests mds%d\n", session->s_mds);
1605         mutex_lock(&mdsc->mutex);
1606         while (!list_empty(&session->s_unsafe)) {
1607                 req = list_first_entry(&session->s_unsafe,
1608                                        struct ceph_mds_request, r_unsafe_item);
1609                 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1610                                     req->r_tid);
1611                 if (req->r_target_inode)
1612                         mapping_set_error(req->r_target_inode->i_mapping, -EIO);
1613                 if (req->r_unsafe_dir)
1614                         mapping_set_error(req->r_unsafe_dir->i_mapping, -EIO);
1615                 __unregister_request(mdsc, req);
1616         }
1617         /* zero r_attempts, so kick_requests() will re-send requests */
1618         p = rb_first(&mdsc->request_tree);
1619         while (p) {
1620                 req = rb_entry(p, struct ceph_mds_request, r_node);
1621                 p = rb_next(p);
1622                 if (req->r_session &&
1623                     req->r_session->s_mds == session->s_mds)
1624                         req->r_attempts = 0;
1625         }
1626         mutex_unlock(&mdsc->mutex);
1627 }
1628
1629 /*
1630  * Helper to safely iterate over all caps associated with a session, with
1631  * special care taken to handle a racing __ceph_remove_cap().
1632  *
1633  * Caller must hold session s_mutex.
1634  */
1635 int ceph_iterate_session_caps(struct ceph_mds_session *session,
1636                               int (*cb)(struct inode *, int mds, void *),
1637                               void *arg)
1638 {
1639         struct list_head *p;
1640         struct ceph_cap *cap;
1641         struct inode *inode, *last_inode = NULL;
1642         struct ceph_cap *old_cap = NULL;
1643         int ret;
1644
1645         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1646         spin_lock(&session->s_cap_lock);
1647         p = session->s_caps.next;
1648         while (p != &session->s_caps) {
1649                 int mds;
1650
1651                 cap = list_entry(p, struct ceph_cap, session_caps);
1652                 inode = igrab(&cap->ci->netfs.inode);
1653                 if (!inode) {
1654                         p = p->next;
1655                         continue;
1656                 }
1657                 session->s_cap_iterator = cap;
1658                 mds = cap->mds;
1659                 spin_unlock(&session->s_cap_lock);
1660
1661                 if (last_inode) {
1662                         iput(last_inode);
1663                         last_inode = NULL;
1664                 }
1665                 if (old_cap) {
1666                         ceph_put_cap(session->s_mdsc, old_cap);
1667                         old_cap = NULL;
1668                 }
1669
1670                 ret = cb(inode, mds, arg);
1671                 last_inode = inode;
1672
1673                 spin_lock(&session->s_cap_lock);
1674                 p = p->next;
1675                 if (!cap->ci) {
1676                         dout("iterate_session_caps  finishing cap %p removal\n",
1677                              cap);
1678                         BUG_ON(cap->session != session);
1679                         cap->session = NULL;
1680                         list_del_init(&cap->session_caps);
1681                         session->s_nr_caps--;
1682                         atomic64_dec(&session->s_mdsc->metric.total_caps);
1683                         if (cap->queue_release)
1684                                 __ceph_queue_cap_release(session, cap);
1685                         else
1686                                 old_cap = cap;  /* put_cap it w/o locks held */
1687                 }
1688                 if (ret < 0)
1689                         goto out;
1690         }
1691         ret = 0;
1692 out:
1693         session->s_cap_iterator = NULL;
1694         spin_unlock(&session->s_cap_lock);
1695
1696         iput(last_inode);
1697         if (old_cap)
1698                 ceph_put_cap(session->s_mdsc, old_cap);
1699
1700         return ret;
1701 }
1702
1703 static int remove_session_caps_cb(struct inode *inode, int mds, void *arg)
1704 {
1705         struct ceph_inode_info *ci = ceph_inode(inode);
1706         bool invalidate = false;
1707         struct ceph_cap *cap;
1708         int iputs = 0;
1709
1710         spin_lock(&ci->i_ceph_lock);
1711         cap = __get_cap_for_mds(ci, mds);
1712         if (cap) {
1713                 dout(" removing cap %p, ci is %p, inode is %p\n",
1714                      cap, ci, &ci->netfs.inode);
1715
1716                 iputs = ceph_purge_inode_cap(inode, cap, &invalidate);
1717         }
1718         spin_unlock(&ci->i_ceph_lock);
1719
1720         if (cap)
1721                 wake_up_all(&ci->i_cap_wq);
1722         if (invalidate)
1723                 ceph_queue_invalidate(inode);
1724         while (iputs--)
1725                 iput(inode);
1726         return 0;
1727 }
1728
1729 /*
1730  * caller must hold session s_mutex
1731  */
1732 static void remove_session_caps(struct ceph_mds_session *session)
1733 {
1734         struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1735         struct super_block *sb = fsc->sb;
1736         LIST_HEAD(dispose);
1737
1738         dout("remove_session_caps on %p\n", session);
1739         ceph_iterate_session_caps(session, remove_session_caps_cb, fsc);
1740
1741         wake_up_all(&fsc->mdsc->cap_flushing_wq);
1742
1743         spin_lock(&session->s_cap_lock);
1744         if (session->s_nr_caps > 0) {
1745                 struct inode *inode;
1746                 struct ceph_cap *cap, *prev = NULL;
1747                 struct ceph_vino vino;
1748                 /*
1749                  * iterate_session_caps() skips inodes that are being
1750                  * deleted, we need to wait until deletions are complete.
1751                  * __wait_on_freeing_inode() is designed for the job,
1752                  * but it is not exported, so use lookup inode function
1753                  * to access it.
1754                  */
1755                 while (!list_empty(&session->s_caps)) {
1756                         cap = list_entry(session->s_caps.next,
1757                                          struct ceph_cap, session_caps);
1758                         if (cap == prev)
1759                                 break;
1760                         prev = cap;
1761                         vino = cap->ci->i_vino;
1762                         spin_unlock(&session->s_cap_lock);
1763
1764                         inode = ceph_find_inode(sb, vino);
1765                         iput(inode);
1766
1767                         spin_lock(&session->s_cap_lock);
1768                 }
1769         }
1770
1771         // drop cap expires and unlock s_cap_lock
1772         detach_cap_releases(session, &dispose);
1773
1774         BUG_ON(session->s_nr_caps > 0);
1775         BUG_ON(!list_empty(&session->s_cap_flushing));
1776         spin_unlock(&session->s_cap_lock);
1777         dispose_cap_releases(session->s_mdsc, &dispose);
1778 }
1779
1780 enum {
1781         RECONNECT,
1782         RENEWCAPS,
1783         FORCE_RO,
1784 };
1785
1786 /*
1787  * wake up any threads waiting on this session's caps.  if the cap is
1788  * old (didn't get renewed on the client reconnect), remove it now.
1789  *
1790  * caller must hold s_mutex.
1791  */
1792 static int wake_up_session_cb(struct inode *inode, int mds, void *arg)
1793 {
1794         struct ceph_inode_info *ci = ceph_inode(inode);
1795         unsigned long ev = (unsigned long)arg;
1796
1797         if (ev == RECONNECT) {
1798                 spin_lock(&ci->i_ceph_lock);
1799                 ci->i_wanted_max_size = 0;
1800                 ci->i_requested_max_size = 0;
1801                 spin_unlock(&ci->i_ceph_lock);
1802         } else if (ev == RENEWCAPS) {
1803                 struct ceph_cap *cap;
1804
1805                 spin_lock(&ci->i_ceph_lock);
1806                 cap = __get_cap_for_mds(ci, mds);
1807                 /* mds did not re-issue stale cap */
1808                 if (cap && cap->cap_gen < atomic_read(&cap->session->s_cap_gen))
1809                         cap->issued = cap->implemented = CEPH_CAP_PIN;
1810                 spin_unlock(&ci->i_ceph_lock);
1811         } else if (ev == FORCE_RO) {
1812         }
1813         wake_up_all(&ci->i_cap_wq);
1814         return 0;
1815 }
1816
1817 static void wake_up_session_caps(struct ceph_mds_session *session, int ev)
1818 {
1819         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1820         ceph_iterate_session_caps(session, wake_up_session_cb,
1821                                   (void *)(unsigned long)ev);
1822 }
1823
1824 /*
1825  * Send periodic message to MDS renewing all currently held caps.  The
1826  * ack will reset the expiration for all caps from this session.
1827  *
1828  * caller holds s_mutex
1829  */
1830 static int send_renew_caps(struct ceph_mds_client *mdsc,
1831                            struct ceph_mds_session *session)
1832 {
1833         struct ceph_msg *msg;
1834         int state;
1835
1836         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1837             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1838                 pr_info("mds%d caps stale\n", session->s_mds);
1839         session->s_renew_requested = jiffies;
1840
1841         /* do not try to renew caps until a recovering mds has reconnected
1842          * with its clients. */
1843         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1844         if (state < CEPH_MDS_STATE_RECONNECT) {
1845                 dout("send_renew_caps ignoring mds%d (%s)\n",
1846                      session->s_mds, ceph_mds_state_name(state));
1847                 return 0;
1848         }
1849
1850         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1851                 ceph_mds_state_name(state));
1852         msg = ceph_create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1853                                       ++session->s_renew_seq);
1854         if (!msg)
1855                 return -ENOMEM;
1856         ceph_con_send(&session->s_con, msg);
1857         return 0;
1858 }
1859
1860 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1861                              struct ceph_mds_session *session, u64 seq)
1862 {
1863         struct ceph_msg *msg;
1864
1865         dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1866              session->s_mds, ceph_session_state_name(session->s_state), seq);
1867         msg = ceph_create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1868         if (!msg)
1869                 return -ENOMEM;
1870         ceph_con_send(&session->s_con, msg);
1871         return 0;
1872 }
1873
1874
1875 /*
1876  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1877  *
1878  * Called under session->s_mutex
1879  */
1880 static void renewed_caps(struct ceph_mds_client *mdsc,
1881                          struct ceph_mds_session *session, int is_renew)
1882 {
1883         int was_stale;
1884         int wake = 0;
1885
1886         spin_lock(&session->s_cap_lock);
1887         was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1888
1889         session->s_cap_ttl = session->s_renew_requested +
1890                 mdsc->mdsmap->m_session_timeout*HZ;
1891
1892         if (was_stale) {
1893                 if (time_before(jiffies, session->s_cap_ttl)) {
1894                         pr_info("mds%d caps renewed\n", session->s_mds);
1895                         wake = 1;
1896                 } else {
1897                         pr_info("mds%d caps still stale\n", session->s_mds);
1898                 }
1899         }
1900         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1901              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1902              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1903         spin_unlock(&session->s_cap_lock);
1904
1905         if (wake)
1906                 wake_up_session_caps(session, RENEWCAPS);
1907 }
1908
1909 /*
1910  * send a session close request
1911  */
1912 static int request_close_session(struct ceph_mds_session *session)
1913 {
1914         struct ceph_msg *msg;
1915
1916         dout("request_close_session mds%d state %s seq %lld\n",
1917              session->s_mds, ceph_session_state_name(session->s_state),
1918              session->s_seq);
1919         msg = ceph_create_session_msg(CEPH_SESSION_REQUEST_CLOSE,
1920                                       session->s_seq);
1921         if (!msg)
1922                 return -ENOMEM;
1923         ceph_con_send(&session->s_con, msg);
1924         return 1;
1925 }
1926
1927 /*
1928  * Called with s_mutex held.
1929  */
1930 static int __close_session(struct ceph_mds_client *mdsc,
1931                          struct ceph_mds_session *session)
1932 {
1933         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1934                 return 0;
1935         session->s_state = CEPH_MDS_SESSION_CLOSING;
1936         return request_close_session(session);
1937 }
1938
1939 static bool drop_negative_children(struct dentry *dentry)
1940 {
1941         struct dentry *child;
1942         bool all_negative = true;
1943
1944         if (!d_is_dir(dentry))
1945                 goto out;
1946
1947         spin_lock(&dentry->d_lock);
1948         list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1949                 if (d_really_is_positive(child)) {
1950                         all_negative = false;
1951                         break;
1952                 }
1953         }
1954         spin_unlock(&dentry->d_lock);
1955
1956         if (all_negative)
1957                 shrink_dcache_parent(dentry);
1958 out:
1959         return all_negative;
1960 }
1961
1962 /*
1963  * Trim old(er) caps.
1964  *
1965  * Because we can't cache an inode without one or more caps, we do
1966  * this indirectly: if a cap is unused, we prune its aliases, at which
1967  * point the inode will hopefully get dropped to.
1968  *
1969  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1970  * memory pressure from the MDS, though, so it needn't be perfect.
1971  */
1972 static int trim_caps_cb(struct inode *inode, int mds, void *arg)
1973 {
1974         int *remaining = arg;
1975         struct ceph_inode_info *ci = ceph_inode(inode);
1976         int used, wanted, oissued, mine;
1977         struct ceph_cap *cap;
1978
1979         if (*remaining <= 0)
1980                 return -1;
1981
1982         spin_lock(&ci->i_ceph_lock);
1983         cap = __get_cap_for_mds(ci, mds);
1984         if (!cap) {
1985                 spin_unlock(&ci->i_ceph_lock);
1986                 return 0;
1987         }
1988         mine = cap->issued | cap->implemented;
1989         used = __ceph_caps_used(ci);
1990         wanted = __ceph_caps_file_wanted(ci);
1991         oissued = __ceph_caps_issued_other(ci, cap);
1992
1993         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1994              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1995              ceph_cap_string(used), ceph_cap_string(wanted));
1996         if (cap == ci->i_auth_cap) {
1997                 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1998                     !list_empty(&ci->i_cap_snaps))
1999                         goto out;
2000                 if ((used | wanted) & CEPH_CAP_ANY_WR)
2001                         goto out;
2002                 /* Note: it's possible that i_filelock_ref becomes non-zero
2003                  * after dropping auth caps. It doesn't hurt because reply
2004                  * of lock mds request will re-add auth caps. */
2005                 if (atomic_read(&ci->i_filelock_ref) > 0)
2006                         goto out;
2007         }
2008         /* The inode has cached pages, but it's no longer used.
2009          * we can safely drop it */
2010         if (S_ISREG(inode->i_mode) &&
2011             wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
2012             !(oissued & CEPH_CAP_FILE_CACHE)) {
2013           used = 0;
2014           oissued = 0;
2015         }
2016         if ((used | wanted) & ~oissued & mine)
2017                 goto out;   /* we need these caps */
2018
2019         if (oissued) {
2020                 /* we aren't the only cap.. just remove us */
2021                 ceph_remove_cap(cap, true);
2022                 (*remaining)--;
2023         } else {
2024                 struct dentry *dentry;
2025                 /* try dropping referring dentries */
2026                 spin_unlock(&ci->i_ceph_lock);
2027                 dentry = d_find_any_alias(inode);
2028                 if (dentry && drop_negative_children(dentry)) {
2029                         int count;
2030                         dput(dentry);
2031                         d_prune_aliases(inode);
2032                         count = atomic_read(&inode->i_count);
2033                         if (count == 1)
2034                                 (*remaining)--;
2035                         dout("trim_caps_cb %p cap %p pruned, count now %d\n",
2036                              inode, cap, count);
2037                 } else {
2038                         dput(dentry);
2039                 }
2040                 return 0;
2041         }
2042
2043 out:
2044         spin_unlock(&ci->i_ceph_lock);
2045         return 0;
2046 }
2047
2048 /*
2049  * Trim session cap count down to some max number.
2050  */
2051 int ceph_trim_caps(struct ceph_mds_client *mdsc,
2052                    struct ceph_mds_session *session,
2053                    int max_caps)
2054 {
2055         int trim_caps = session->s_nr_caps - max_caps;
2056
2057         dout("trim_caps mds%d start: %d / %d, trim %d\n",
2058              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
2059         if (trim_caps > 0) {
2060                 int remaining = trim_caps;
2061
2062                 ceph_iterate_session_caps(session, trim_caps_cb, &remaining);
2063                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
2064                      session->s_mds, session->s_nr_caps, max_caps,
2065                         trim_caps - remaining);
2066         }
2067
2068         ceph_flush_cap_releases(mdsc, session);
2069         return 0;
2070 }
2071
2072 static int check_caps_flush(struct ceph_mds_client *mdsc,
2073                             u64 want_flush_tid)
2074 {
2075         int ret = 1;
2076
2077         spin_lock(&mdsc->cap_dirty_lock);
2078         if (!list_empty(&mdsc->cap_flush_list)) {
2079                 struct ceph_cap_flush *cf =
2080                         list_first_entry(&mdsc->cap_flush_list,
2081                                          struct ceph_cap_flush, g_list);
2082                 if (cf->tid <= want_flush_tid) {
2083                         dout("check_caps_flush still flushing tid "
2084                              "%llu <= %llu\n", cf->tid, want_flush_tid);
2085                         ret = 0;
2086                 }
2087         }
2088         spin_unlock(&mdsc->cap_dirty_lock);
2089         return ret;
2090 }
2091
2092 /*
2093  * flush all dirty inode data to disk.
2094  *
2095  * returns true if we've flushed through want_flush_tid
2096  */
2097 static void wait_caps_flush(struct ceph_mds_client *mdsc,
2098                             u64 want_flush_tid)
2099 {
2100         dout("check_caps_flush want %llu\n", want_flush_tid);
2101
2102         wait_event(mdsc->cap_flushing_wq,
2103                    check_caps_flush(mdsc, want_flush_tid));
2104
2105         dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
2106 }
2107
2108 /*
2109  * called under s_mutex
2110  */
2111 static void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
2112                                    struct ceph_mds_session *session)
2113 {
2114         struct ceph_msg *msg = NULL;
2115         struct ceph_mds_cap_release *head;
2116         struct ceph_mds_cap_item *item;
2117         struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
2118         struct ceph_cap *cap;
2119         LIST_HEAD(tmp_list);
2120         int num_cap_releases;
2121         __le32  barrier, *cap_barrier;
2122
2123         down_read(&osdc->lock);
2124         barrier = cpu_to_le32(osdc->epoch_barrier);
2125         up_read(&osdc->lock);
2126
2127         spin_lock(&session->s_cap_lock);
2128 again:
2129         list_splice_init(&session->s_cap_releases, &tmp_list);
2130         num_cap_releases = session->s_num_cap_releases;
2131         session->s_num_cap_releases = 0;
2132         spin_unlock(&session->s_cap_lock);
2133
2134         while (!list_empty(&tmp_list)) {
2135                 if (!msg) {
2136                         msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
2137                                         PAGE_SIZE, GFP_NOFS, false);
2138                         if (!msg)
2139                                 goto out_err;
2140                         head = msg->front.iov_base;
2141                         head->num = cpu_to_le32(0);
2142                         msg->front.iov_len = sizeof(*head);
2143
2144                         msg->hdr.version = cpu_to_le16(2);
2145                         msg->hdr.compat_version = cpu_to_le16(1);
2146                 }
2147
2148                 cap = list_first_entry(&tmp_list, struct ceph_cap,
2149                                         session_caps);
2150                 list_del(&cap->session_caps);
2151                 num_cap_releases--;
2152
2153                 head = msg->front.iov_base;
2154                 put_unaligned_le32(get_unaligned_le32(&head->num) + 1,
2155                                    &head->num);
2156                 item = msg->front.iov_base + msg->front.iov_len;
2157                 item->ino = cpu_to_le64(cap->cap_ino);
2158                 item->cap_id = cpu_to_le64(cap->cap_id);
2159                 item->migrate_seq = cpu_to_le32(cap->mseq);
2160                 item->seq = cpu_to_le32(cap->issue_seq);
2161                 msg->front.iov_len += sizeof(*item);
2162
2163                 ceph_put_cap(mdsc, cap);
2164
2165                 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
2166                         // Append cap_barrier field
2167                         cap_barrier = msg->front.iov_base + msg->front.iov_len;
2168                         *cap_barrier = barrier;
2169                         msg->front.iov_len += sizeof(*cap_barrier);
2170
2171                         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2172                         dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
2173                         ceph_con_send(&session->s_con, msg);
2174                         msg = NULL;
2175                 }
2176         }
2177
2178         BUG_ON(num_cap_releases != 0);
2179
2180         spin_lock(&session->s_cap_lock);
2181         if (!list_empty(&session->s_cap_releases))
2182                 goto again;
2183         spin_unlock(&session->s_cap_lock);
2184
2185         if (msg) {
2186                 // Append cap_barrier field
2187                 cap_barrier = msg->front.iov_base + msg->front.iov_len;
2188                 *cap_barrier = barrier;
2189                 msg->front.iov_len += sizeof(*cap_barrier);
2190
2191                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2192                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
2193                 ceph_con_send(&session->s_con, msg);
2194         }
2195         return;
2196 out_err:
2197         pr_err("send_cap_releases mds%d, failed to allocate message\n",
2198                 session->s_mds);
2199         spin_lock(&session->s_cap_lock);
2200         list_splice(&tmp_list, &session->s_cap_releases);
2201         session->s_num_cap_releases += num_cap_releases;
2202         spin_unlock(&session->s_cap_lock);
2203 }
2204
2205 static void ceph_cap_release_work(struct work_struct *work)
2206 {
2207         struct ceph_mds_session *session =
2208                 container_of(work, struct ceph_mds_session, s_cap_release_work);
2209
2210         mutex_lock(&session->s_mutex);
2211         if (session->s_state == CEPH_MDS_SESSION_OPEN ||
2212             session->s_state == CEPH_MDS_SESSION_HUNG)
2213                 ceph_send_cap_releases(session->s_mdsc, session);
2214         mutex_unlock(&session->s_mutex);
2215         ceph_put_mds_session(session);
2216 }
2217
2218 void ceph_flush_cap_releases(struct ceph_mds_client *mdsc,
2219                              struct ceph_mds_session *session)
2220 {
2221         if (mdsc->stopping)
2222                 return;
2223
2224         ceph_get_mds_session(session);
2225         if (queue_work(mdsc->fsc->cap_wq,
2226                        &session->s_cap_release_work)) {
2227                 dout("cap release work queued\n");
2228         } else {
2229                 ceph_put_mds_session(session);
2230                 dout("failed to queue cap release work\n");
2231         }
2232 }
2233
2234 /*
2235  * caller holds session->s_cap_lock
2236  */
2237 void __ceph_queue_cap_release(struct ceph_mds_session *session,
2238                               struct ceph_cap *cap)
2239 {
2240         list_add_tail(&cap->session_caps, &session->s_cap_releases);
2241         session->s_num_cap_releases++;
2242
2243         if (!(session->s_num_cap_releases % CEPH_CAPS_PER_RELEASE))
2244                 ceph_flush_cap_releases(session->s_mdsc, session);
2245 }
2246
2247 static void ceph_cap_reclaim_work(struct work_struct *work)
2248 {
2249         struct ceph_mds_client *mdsc =
2250                 container_of(work, struct ceph_mds_client, cap_reclaim_work);
2251         int ret = ceph_trim_dentries(mdsc);
2252         if (ret == -EAGAIN)
2253                 ceph_queue_cap_reclaim_work(mdsc);
2254 }
2255
2256 void ceph_queue_cap_reclaim_work(struct ceph_mds_client *mdsc)
2257 {
2258         if (mdsc->stopping)
2259                 return;
2260
2261         if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_reclaim_work)) {
2262                 dout("caps reclaim work queued\n");
2263         } else {
2264                 dout("failed to queue caps release work\n");
2265         }
2266 }
2267
2268 void ceph_reclaim_caps_nr(struct ceph_mds_client *mdsc, int nr)
2269 {
2270         int val;
2271         if (!nr)
2272                 return;
2273         val = atomic_add_return(nr, &mdsc->cap_reclaim_pending);
2274         if ((val % CEPH_CAPS_PER_RELEASE) < nr) {
2275                 atomic_set(&mdsc->cap_reclaim_pending, 0);
2276                 ceph_queue_cap_reclaim_work(mdsc);
2277         }
2278 }
2279
2280 /*
2281  * requests
2282  */
2283
2284 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
2285                                     struct inode *dir)
2286 {
2287         struct ceph_inode_info *ci = ceph_inode(dir);
2288         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
2289         struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
2290         size_t size = sizeof(struct ceph_mds_reply_dir_entry);
2291         unsigned int num_entries;
2292         int order;
2293
2294         spin_lock(&ci->i_ceph_lock);
2295         num_entries = ci->i_files + ci->i_subdirs;
2296         spin_unlock(&ci->i_ceph_lock);
2297         num_entries = max(num_entries, 1U);
2298         num_entries = min(num_entries, opt->max_readdir);
2299
2300         order = get_order(size * num_entries);
2301         while (order >= 0) {
2302                 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
2303                                                              __GFP_NOWARN |
2304                                                              __GFP_ZERO,
2305                                                              order);
2306                 if (rinfo->dir_entries)
2307                         break;
2308                 order--;
2309         }
2310         if (!rinfo->dir_entries)
2311                 return -ENOMEM;
2312
2313         num_entries = (PAGE_SIZE << order) / size;
2314         num_entries = min(num_entries, opt->max_readdir);
2315
2316         rinfo->dir_buf_size = PAGE_SIZE << order;
2317         req->r_num_caps = num_entries + 1;
2318         req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
2319         req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
2320         return 0;
2321 }
2322
2323 /*
2324  * Create an mds request.
2325  */
2326 struct ceph_mds_request *
2327 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
2328 {
2329         struct ceph_mds_request *req;
2330
2331         req = kmem_cache_zalloc(ceph_mds_request_cachep, GFP_NOFS);
2332         if (!req)
2333                 return ERR_PTR(-ENOMEM);
2334
2335         mutex_init(&req->r_fill_mutex);
2336         req->r_mdsc = mdsc;
2337         req->r_started = jiffies;
2338         req->r_start_latency = ktime_get();
2339         req->r_resend_mds = -1;
2340         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
2341         INIT_LIST_HEAD(&req->r_unsafe_target_item);
2342         req->r_fmode = -1;
2343         req->r_feature_needed = -1;
2344         kref_init(&req->r_kref);
2345         RB_CLEAR_NODE(&req->r_node);
2346         INIT_LIST_HEAD(&req->r_wait);
2347         init_completion(&req->r_completion);
2348         init_completion(&req->r_safe_completion);
2349         INIT_LIST_HEAD(&req->r_unsafe_item);
2350
2351         ktime_get_coarse_real_ts64(&req->r_stamp);
2352
2353         req->r_op = op;
2354         req->r_direct_mode = mode;
2355         return req;
2356 }
2357
2358 /*
2359  * return oldest (lowest) request, tid in request tree, 0 if none.
2360  *
2361  * called under mdsc->mutex.
2362  */
2363 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
2364 {
2365         if (RB_EMPTY_ROOT(&mdsc->request_tree))
2366                 return NULL;
2367         return rb_entry(rb_first(&mdsc->request_tree),
2368                         struct ceph_mds_request, r_node);
2369 }
2370
2371 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
2372 {
2373         return mdsc->oldest_tid;
2374 }
2375
2376 /*
2377  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
2378  * on build_path_from_dentry in fs/cifs/dir.c.
2379  *
2380  * If @stop_on_nosnap, generate path relative to the first non-snapped
2381  * inode.
2382  *
2383  * Encode hidden .snap dirs as a double /, i.e.
2384  *   foo/.snap/bar -> foo//bar
2385  */
2386 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *pbase,
2387                            int stop_on_nosnap)
2388 {
2389         struct dentry *temp;
2390         char *path;
2391         int pos;
2392         unsigned seq;
2393         u64 base;
2394
2395         if (!dentry)
2396                 return ERR_PTR(-EINVAL);
2397
2398         path = __getname();
2399         if (!path)
2400                 return ERR_PTR(-ENOMEM);
2401 retry:
2402         pos = PATH_MAX - 1;
2403         path[pos] = '\0';
2404
2405         seq = read_seqbegin(&rename_lock);
2406         rcu_read_lock();
2407         temp = dentry;
2408         for (;;) {
2409                 struct inode *inode;
2410
2411                 spin_lock(&temp->d_lock);
2412                 inode = d_inode(temp);
2413                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
2414                         dout("build_path path+%d: %p SNAPDIR\n",
2415                              pos, temp);
2416                 } else if (stop_on_nosnap && inode && dentry != temp &&
2417                            ceph_snap(inode) == CEPH_NOSNAP) {
2418                         spin_unlock(&temp->d_lock);
2419                         pos++; /* get rid of any prepended '/' */
2420                         break;
2421                 } else {
2422                         pos -= temp->d_name.len;
2423                         if (pos < 0) {
2424                                 spin_unlock(&temp->d_lock);
2425                                 break;
2426                         }
2427                         memcpy(path + pos, temp->d_name.name, temp->d_name.len);
2428                 }
2429                 spin_unlock(&temp->d_lock);
2430                 temp = READ_ONCE(temp->d_parent);
2431
2432                 /* Are we at the root? */
2433                 if (IS_ROOT(temp))
2434                         break;
2435
2436                 /* Are we out of buffer? */
2437                 if (--pos < 0)
2438                         break;
2439
2440                 path[pos] = '/';
2441         }
2442         base = ceph_ino(d_inode(temp));
2443         rcu_read_unlock();
2444
2445         if (read_seqretry(&rename_lock, seq))
2446                 goto retry;
2447
2448         if (pos < 0) {
2449                 /*
2450                  * A rename didn't occur, but somehow we didn't end up where
2451                  * we thought we would. Throw a warning and try again.
2452                  */
2453                 pr_warn("build_path did not end path lookup where "
2454                         "expected, pos is %d\n", pos);
2455                 goto retry;
2456         }
2457
2458         *pbase = base;
2459         *plen = PATH_MAX - 1 - pos;
2460         dout("build_path on %p %d built %llx '%.*s'\n",
2461              dentry, d_count(dentry), base, *plen, path + pos);
2462         return path + pos;
2463 }
2464
2465 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
2466                              const char **ppath, int *ppathlen, u64 *pino,
2467                              bool *pfreepath, bool parent_locked)
2468 {
2469         char *path;
2470
2471         rcu_read_lock();
2472         if (!dir)
2473                 dir = d_inode_rcu(dentry->d_parent);
2474         if (dir && parent_locked && ceph_snap(dir) == CEPH_NOSNAP) {
2475                 *pino = ceph_ino(dir);
2476                 rcu_read_unlock();
2477                 *ppath = dentry->d_name.name;
2478                 *ppathlen = dentry->d_name.len;
2479                 return 0;
2480         }
2481         rcu_read_unlock();
2482         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2483         if (IS_ERR(path))
2484                 return PTR_ERR(path);
2485         *ppath = path;
2486         *pfreepath = true;
2487         return 0;
2488 }
2489
2490 static int build_inode_path(struct inode *inode,
2491                             const char **ppath, int *ppathlen, u64 *pino,
2492                             bool *pfreepath)
2493 {
2494         struct dentry *dentry;
2495         char *path;
2496
2497         if (ceph_snap(inode) == CEPH_NOSNAP) {
2498                 *pino = ceph_ino(inode);
2499                 *ppathlen = 0;
2500                 return 0;
2501         }
2502         dentry = d_find_alias(inode);
2503         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2504         dput(dentry);
2505         if (IS_ERR(path))
2506                 return PTR_ERR(path);
2507         *ppath = path;
2508         *pfreepath = true;
2509         return 0;
2510 }
2511
2512 /*
2513  * request arguments may be specified via an inode *, a dentry *, or
2514  * an explicit ino+path.
2515  */
2516 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
2517                                   struct inode *rdiri, const char *rpath,
2518                                   u64 rino, const char **ppath, int *pathlen,
2519                                   u64 *ino, bool *freepath, bool parent_locked)
2520 {
2521         int r = 0;
2522
2523         if (rinode) {
2524                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2525                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2526                      ceph_snap(rinode));
2527         } else if (rdentry) {
2528                 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
2529                                         freepath, parent_locked);
2530                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
2531                      *ppath);
2532         } else if (rpath || rino) {
2533                 *ino = rino;
2534                 *ppath = rpath;
2535                 *pathlen = rpath ? strlen(rpath) : 0;
2536                 dout(" path %.*s\n", *pathlen, rpath);
2537         }
2538
2539         return r;
2540 }
2541
2542 static void encode_timestamp_and_gids(void **p,
2543                                       const struct ceph_mds_request *req)
2544 {
2545         struct ceph_timespec ts;
2546         int i;
2547
2548         ceph_encode_timespec64(&ts, &req->r_stamp);
2549         ceph_encode_copy(p, &ts, sizeof(ts));
2550
2551         /* gid_list */
2552         ceph_encode_32(p, req->r_cred->group_info->ngroups);
2553         for (i = 0; i < req->r_cred->group_info->ngroups; i++)
2554                 ceph_encode_64(p, from_kgid(&init_user_ns,
2555                                             req->r_cred->group_info->gid[i]));
2556 }
2557
2558 /*
2559  * called under mdsc->mutex
2560  */
2561 static struct ceph_msg *create_request_message(struct ceph_mds_session *session,
2562                                                struct ceph_mds_request *req,
2563                                                bool drop_cap_releases)
2564 {
2565         int mds = session->s_mds;
2566         struct ceph_mds_client *mdsc = session->s_mdsc;
2567         struct ceph_msg *msg;
2568         struct ceph_mds_request_head_old *head;
2569         const char *path1 = NULL;
2570         const char *path2 = NULL;
2571         u64 ino1 = 0, ino2 = 0;
2572         int pathlen1 = 0, pathlen2 = 0;
2573         bool freepath1 = false, freepath2 = false;
2574         int len;
2575         u16 releases;
2576         void *p, *end;
2577         int ret;
2578         bool legacy = !(session->s_con.peer_features & CEPH_FEATURE_FS_BTIME);
2579
2580         ret = set_request_path_attr(req->r_inode, req->r_dentry,
2581                               req->r_parent, req->r_path1, req->r_ino1.ino,
2582                               &path1, &pathlen1, &ino1, &freepath1,
2583                               test_bit(CEPH_MDS_R_PARENT_LOCKED,
2584                                         &req->r_req_flags));
2585         if (ret < 0) {
2586                 msg = ERR_PTR(ret);
2587                 goto out;
2588         }
2589
2590         /* If r_old_dentry is set, then assume that its parent is locked */
2591         ret = set_request_path_attr(NULL, req->r_old_dentry,
2592                               req->r_old_dentry_dir,
2593                               req->r_path2, req->r_ino2.ino,
2594                               &path2, &pathlen2, &ino2, &freepath2, true);
2595         if (ret < 0) {
2596                 msg = ERR_PTR(ret);
2597                 goto out_free1;
2598         }
2599
2600         len = legacy ? sizeof(*head) : sizeof(struct ceph_mds_request_head);
2601         len += pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2602                 sizeof(struct ceph_timespec);
2603         len += sizeof(u32) + (sizeof(u64) * req->r_cred->group_info->ngroups);
2604
2605         /* calculate (max) length for cap releases */
2606         len += sizeof(struct ceph_mds_request_release) *
2607                 (!!req->r_inode_drop + !!req->r_dentry_drop +
2608                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2609
2610         if (req->r_dentry_drop)
2611                 len += pathlen1;
2612         if (req->r_old_dentry_drop)
2613                 len += pathlen2;
2614
2615         msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
2616         if (!msg) {
2617                 msg = ERR_PTR(-ENOMEM);
2618                 goto out_free2;
2619         }
2620
2621         msg->hdr.tid = cpu_to_le64(req->r_tid);
2622
2623         /*
2624          * The old ceph_mds_request_head didn't contain a version field, and
2625          * one was added when we moved the message version from 3->4.
2626          */
2627         if (legacy) {
2628                 msg->hdr.version = cpu_to_le16(3);
2629                 head = msg->front.iov_base;
2630                 p = msg->front.iov_base + sizeof(*head);
2631         } else {
2632                 struct ceph_mds_request_head *new_head = msg->front.iov_base;
2633
2634                 msg->hdr.version = cpu_to_le16(4);
2635                 new_head->version = cpu_to_le16(CEPH_MDS_REQUEST_HEAD_VERSION);
2636                 head = (struct ceph_mds_request_head_old *)&new_head->oldest_client_tid;
2637                 p = msg->front.iov_base + sizeof(*new_head);
2638         }
2639
2640         end = msg->front.iov_base + msg->front.iov_len;
2641
2642         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2643         head->op = cpu_to_le32(req->r_op);
2644         head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns,
2645                                                  req->r_cred->fsuid));
2646         head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns,
2647                                                  req->r_cred->fsgid));
2648         head->ino = cpu_to_le64(req->r_deleg_ino);
2649         head->args = req->r_args;
2650
2651         ceph_encode_filepath(&p, end, ino1, path1);
2652         ceph_encode_filepath(&p, end, ino2, path2);
2653
2654         /* make note of release offset, in case we need to replay */
2655         req->r_request_release_offset = p - msg->front.iov_base;
2656
2657         /* cap releases */
2658         releases = 0;
2659         if (req->r_inode_drop)
2660                 releases += ceph_encode_inode_release(&p,
2661                       req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2662                       mds, req->r_inode_drop, req->r_inode_unless,
2663                       req->r_op == CEPH_MDS_OP_READDIR);
2664         if (req->r_dentry_drop)
2665                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
2666                                 req->r_parent, mds, req->r_dentry_drop,
2667                                 req->r_dentry_unless);
2668         if (req->r_old_dentry_drop)
2669                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2670                                 req->r_old_dentry_dir, mds,
2671                                 req->r_old_dentry_drop,
2672                                 req->r_old_dentry_unless);
2673         if (req->r_old_inode_drop)
2674                 releases += ceph_encode_inode_release(&p,
2675                       d_inode(req->r_old_dentry),
2676                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2677
2678         if (drop_cap_releases) {
2679                 releases = 0;
2680                 p = msg->front.iov_base + req->r_request_release_offset;
2681         }
2682
2683         head->num_releases = cpu_to_le16(releases);
2684
2685         encode_timestamp_and_gids(&p, req);
2686
2687         if (WARN_ON_ONCE(p > end)) {
2688                 ceph_msg_put(msg);
2689                 msg = ERR_PTR(-ERANGE);
2690                 goto out_free2;
2691         }
2692
2693         msg->front.iov_len = p - msg->front.iov_base;
2694         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2695
2696         if (req->r_pagelist) {
2697                 struct ceph_pagelist *pagelist = req->r_pagelist;
2698                 ceph_msg_data_add_pagelist(msg, pagelist);
2699                 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2700         } else {
2701                 msg->hdr.data_len = 0;
2702         }
2703
2704         msg->hdr.data_off = cpu_to_le16(0);
2705
2706 out_free2:
2707         if (freepath2)
2708                 ceph_mdsc_free_path((char *)path2, pathlen2);
2709 out_free1:
2710         if (freepath1)
2711                 ceph_mdsc_free_path((char *)path1, pathlen1);
2712 out:
2713         return msg;
2714 }
2715
2716 /*
2717  * called under mdsc->mutex if error, under no mutex if
2718  * success.
2719  */
2720 static void complete_request(struct ceph_mds_client *mdsc,
2721                              struct ceph_mds_request *req)
2722 {
2723         req->r_end_latency = ktime_get();
2724
2725         if (req->r_callback)
2726                 req->r_callback(mdsc, req);
2727         complete_all(&req->r_completion);
2728 }
2729
2730 static struct ceph_mds_request_head_old *
2731 find_old_request_head(void *p, u64 features)
2732 {
2733         bool legacy = !(features & CEPH_FEATURE_FS_BTIME);
2734         struct ceph_mds_request_head *new_head;
2735
2736         if (legacy)
2737                 return (struct ceph_mds_request_head_old *)p;
2738         new_head = (struct ceph_mds_request_head *)p;
2739         return (struct ceph_mds_request_head_old *)&new_head->oldest_client_tid;
2740 }
2741
2742 /*
2743  * called under mdsc->mutex
2744  */
2745 static int __prepare_send_request(struct ceph_mds_session *session,
2746                                   struct ceph_mds_request *req,
2747                                   bool drop_cap_releases)
2748 {
2749         int mds = session->s_mds;
2750         struct ceph_mds_client *mdsc = session->s_mdsc;
2751         struct ceph_mds_request_head_old *rhead;
2752         struct ceph_msg *msg;
2753         int flags = 0, max_retry;
2754
2755         /*
2756          * The type of 'r_attempts' in kernel 'ceph_mds_request'
2757          * is 'int', while in 'ceph_mds_request_head' the type of
2758          * 'num_retry' is '__u8'. So in case the request retries
2759          *  exceeding 256 times, the MDS will receive a incorrect
2760          *  retry seq.
2761          *
2762          * In this case it's ususally a bug in MDS and continue
2763          * retrying the request makes no sense.
2764          *
2765          * In future this could be fixed in ceph code, so avoid
2766          * using the hardcode here.
2767          */
2768         max_retry = sizeof_field(struct ceph_mds_request_head, num_retry);
2769         max_retry = 1 << (max_retry * BITS_PER_BYTE);
2770         if (req->r_attempts >= max_retry) {
2771                 pr_warn_ratelimited("%s request tid %llu seq overflow\n",
2772                                     __func__, req->r_tid);
2773                 return -EMULTIHOP;
2774         }
2775
2776         req->r_attempts++;
2777         if (req->r_inode) {
2778                 struct ceph_cap *cap =
2779                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2780
2781                 if (cap)
2782                         req->r_sent_on_mseq = cap->mseq;
2783                 else
2784                         req->r_sent_on_mseq = -1;
2785         }
2786         dout("%s %p tid %lld %s (attempt %d)\n", __func__, req,
2787              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2788
2789         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2790                 void *p;
2791
2792                 /*
2793                  * Replay.  Do not regenerate message (and rebuild
2794                  * paths, etc.); just use the original message.
2795                  * Rebuilding paths will break for renames because
2796                  * d_move mangles the src name.
2797                  */
2798                 msg = req->r_request;
2799                 rhead = find_old_request_head(msg->front.iov_base,
2800                                               session->s_con.peer_features);
2801
2802                 flags = le32_to_cpu(rhead->flags);
2803                 flags |= CEPH_MDS_FLAG_REPLAY;
2804                 rhead->flags = cpu_to_le32(flags);
2805
2806                 if (req->r_target_inode)
2807                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2808
2809                 rhead->num_retry = req->r_attempts - 1;
2810
2811                 /* remove cap/dentry releases from message */
2812                 rhead->num_releases = 0;
2813
2814                 p = msg->front.iov_base + req->r_request_release_offset;
2815                 encode_timestamp_and_gids(&p, req);
2816
2817                 msg->front.iov_len = p - msg->front.iov_base;
2818                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2819                 return 0;
2820         }
2821
2822         if (req->r_request) {
2823                 ceph_msg_put(req->r_request);
2824                 req->r_request = NULL;
2825         }
2826         msg = create_request_message(session, req, drop_cap_releases);
2827         if (IS_ERR(msg)) {
2828                 req->r_err = PTR_ERR(msg);
2829                 return PTR_ERR(msg);
2830         }
2831         req->r_request = msg;
2832
2833         rhead = find_old_request_head(msg->front.iov_base,
2834                                       session->s_con.peer_features);
2835         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2836         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2837                 flags |= CEPH_MDS_FLAG_REPLAY;
2838         if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags))
2839                 flags |= CEPH_MDS_FLAG_ASYNC;
2840         if (req->r_parent)
2841                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2842         rhead->flags = cpu_to_le32(flags);
2843         rhead->num_fwd = req->r_num_fwd;
2844         rhead->num_retry = req->r_attempts - 1;
2845
2846         dout(" r_parent = %p\n", req->r_parent);
2847         return 0;
2848 }
2849
2850 /*
2851  * called under mdsc->mutex
2852  */
2853 static int __send_request(struct ceph_mds_session *session,
2854                           struct ceph_mds_request *req,
2855                           bool drop_cap_releases)
2856 {
2857         int err;
2858
2859         err = __prepare_send_request(session, req, drop_cap_releases);
2860         if (!err) {
2861                 ceph_msg_get(req->r_request);
2862                 ceph_con_send(&session->s_con, req->r_request);
2863         }
2864
2865         return err;
2866 }
2867
2868 /*
2869  * send request, or put it on the appropriate wait list.
2870  */
2871 static void __do_request(struct ceph_mds_client *mdsc,
2872                         struct ceph_mds_request *req)
2873 {
2874         struct ceph_mds_session *session = NULL;
2875         int mds = -1;
2876         int err = 0;
2877         bool random;
2878
2879         if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2880                 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2881                         __unregister_request(mdsc, req);
2882                 return;
2883         }
2884
2885         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_FENCE_IO) {
2886                 dout("do_request metadata corrupted\n");
2887                 err = -EIO;
2888                 goto finish;
2889         }
2890         if (req->r_timeout &&
2891             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2892                 dout("do_request timed out\n");
2893                 err = -ETIMEDOUT;
2894                 goto finish;
2895         }
2896         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2897                 dout("do_request forced umount\n");
2898                 err = -EIO;
2899                 goto finish;
2900         }
2901         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2902                 if (mdsc->mdsmap_err) {
2903                         err = mdsc->mdsmap_err;
2904                         dout("do_request mdsmap err %d\n", err);
2905                         goto finish;
2906                 }
2907                 if (mdsc->mdsmap->m_epoch == 0) {
2908                         dout("do_request no mdsmap, waiting for map\n");
2909                         list_add(&req->r_wait, &mdsc->waiting_for_map);
2910                         return;
2911                 }
2912                 if (!(mdsc->fsc->mount_options->flags &
2913                       CEPH_MOUNT_OPT_MOUNTWAIT) &&
2914                     !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2915                         err = -EHOSTUNREACH;
2916                         goto finish;
2917                 }
2918         }
2919
2920         put_request_session(req);
2921
2922         mds = __choose_mds(mdsc, req, &random);
2923         if (mds < 0 ||
2924             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2925                 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2926                         err = -EJUKEBOX;
2927                         goto finish;
2928                 }
2929                 dout("do_request no mds or not active, waiting for map\n");
2930                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2931                 return;
2932         }
2933
2934         /* get, open session */
2935         session = __ceph_lookup_mds_session(mdsc, mds);
2936         if (!session) {
2937                 session = register_session(mdsc, mds);
2938                 if (IS_ERR(session)) {
2939                         err = PTR_ERR(session);
2940                         goto finish;
2941                 }
2942         }
2943         req->r_session = ceph_get_mds_session(session);
2944
2945         dout("do_request mds%d session %p state %s\n", mds, session,
2946              ceph_session_state_name(session->s_state));
2947
2948         /*
2949          * The old ceph will crash the MDSs when see unknown OPs
2950          */
2951         if (req->r_feature_needed > 0 &&
2952             !test_bit(req->r_feature_needed, &session->s_features)) {
2953                 err = -EOPNOTSUPP;
2954                 goto out_session;
2955         }
2956
2957         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2958             session->s_state != CEPH_MDS_SESSION_HUNG) {
2959                 /*
2960                  * We cannot queue async requests since the caps and delegated
2961                  * inodes are bound to the session. Just return -EJUKEBOX and
2962                  * let the caller retry a sync request in that case.
2963                  */
2964                 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2965                         err = -EJUKEBOX;
2966                         goto out_session;
2967                 }
2968
2969                 /*
2970                  * If the session has been REJECTED, then return a hard error,
2971                  * unless it's a CLEANRECOVER mount, in which case we'll queue
2972                  * it to the mdsc queue.
2973                  */
2974                 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2975                         if (ceph_test_mount_opt(mdsc->fsc, CLEANRECOVER))
2976                                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2977                         else
2978                                 err = -EACCES;
2979                         goto out_session;
2980                 }
2981
2982                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2983                     session->s_state == CEPH_MDS_SESSION_CLOSING) {
2984                         err = __open_session(mdsc, session);
2985                         if (err)
2986                                 goto out_session;
2987                         /* retry the same mds later */
2988                         if (random)
2989                                 req->r_resend_mds = mds;
2990                 }
2991                 list_add(&req->r_wait, &session->s_waiting);
2992                 goto out_session;
2993         }
2994
2995         /* send request */
2996         req->r_resend_mds = -1;   /* forget any previous mds hint */
2997
2998         if (req->r_request_started == 0)   /* note request start time */
2999                 req->r_request_started = jiffies;
3000
3001         /*
3002          * For async create we will choose the auth MDS of frag in parent
3003          * directory to send the request and ususally this works fine, but
3004          * if the migrated the dirtory to another MDS before it could handle
3005          * it the request will be forwarded.
3006          *
3007          * And then the auth cap will be changed.
3008          */
3009         if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags) && req->r_num_fwd) {
3010                 struct ceph_dentry_info *di = ceph_dentry(req->r_dentry);
3011                 struct ceph_inode_info *ci;
3012                 struct ceph_cap *cap;
3013
3014                 /*
3015                  * The request maybe handled very fast and the new inode
3016                  * hasn't been linked to the dentry yet. We need to wait
3017                  * for the ceph_finish_async_create(), which shouldn't be
3018                  * stuck too long or fail in thoery, to finish when forwarding
3019                  * the request.
3020                  */
3021                 if (!d_inode(req->r_dentry)) {
3022                         err = wait_on_bit(&di->flags, CEPH_DENTRY_ASYNC_CREATE_BIT,
3023                                           TASK_KILLABLE);
3024                         if (err) {
3025                                 mutex_lock(&req->r_fill_mutex);
3026                                 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
3027                                 mutex_unlock(&req->r_fill_mutex);
3028                                 goto out_session;
3029                         }
3030                 }
3031
3032                 ci = ceph_inode(d_inode(req->r_dentry));
3033
3034                 spin_lock(&ci->i_ceph_lock);
3035                 cap = ci->i_auth_cap;
3036                 if (ci->i_ceph_flags & CEPH_I_ASYNC_CREATE && mds != cap->mds) {
3037                         dout("do_request session changed for auth cap %d -> %d\n",
3038                              cap->session->s_mds, session->s_mds);
3039
3040                         /* Remove the auth cap from old session */
3041                         spin_lock(&cap->session->s_cap_lock);
3042                         cap->session->s_nr_caps--;
3043                         list_del_init(&cap->session_caps);
3044                         spin_unlock(&cap->session->s_cap_lock);
3045
3046                         /* Add the auth cap to the new session */
3047                         cap->mds = mds;
3048                         cap->session = session;
3049                         spin_lock(&session->s_cap_lock);
3050                         session->s_nr_caps++;
3051                         list_add_tail(&cap->session_caps, &session->s_caps);
3052                         spin_unlock(&session->s_cap_lock);
3053
3054                         change_auth_cap_ses(ci, session);
3055                 }
3056                 spin_unlock(&ci->i_ceph_lock);
3057         }
3058
3059         err = __send_request(session, req, false);
3060
3061 out_session:
3062         ceph_put_mds_session(session);
3063 finish:
3064         if (err) {
3065                 dout("__do_request early error %d\n", err);
3066                 req->r_err = err;
3067                 complete_request(mdsc, req);
3068                 __unregister_request(mdsc, req);
3069         }
3070         return;
3071 }
3072
3073 /*
3074  * called under mdsc->mutex
3075  */
3076 static void __wake_requests(struct ceph_mds_client *mdsc,
3077                             struct list_head *head)
3078 {
3079         struct ceph_mds_request *req;
3080         LIST_HEAD(tmp_list);
3081
3082         list_splice_init(head, &tmp_list);
3083
3084         while (!list_empty(&tmp_list)) {
3085                 req = list_entry(tmp_list.next,
3086                                  struct ceph_mds_request, r_wait);
3087                 list_del_init(&req->r_wait);
3088                 dout(" wake request %p tid %llu\n", req, req->r_tid);
3089                 __do_request(mdsc, req);
3090         }
3091 }
3092
3093 /*
3094  * Wake up threads with requests pending for @mds, so that they can
3095  * resubmit their requests to a possibly different mds.
3096  */
3097 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
3098 {
3099         struct ceph_mds_request *req;
3100         struct rb_node *p = rb_first(&mdsc->request_tree);
3101
3102         dout("kick_requests mds%d\n", mds);
3103         while (p) {
3104                 req = rb_entry(p, struct ceph_mds_request, r_node);
3105                 p = rb_next(p);
3106                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3107                         continue;
3108                 if (req->r_attempts > 0)
3109                         continue; /* only new requests */
3110                 if (req->r_session &&
3111                     req->r_session->s_mds == mds) {
3112                         dout(" kicking tid %llu\n", req->r_tid);
3113                         list_del_init(&req->r_wait);
3114                         __do_request(mdsc, req);
3115                 }
3116         }
3117 }
3118
3119 int ceph_mdsc_submit_request(struct ceph_mds_client *mdsc, struct inode *dir,
3120                               struct ceph_mds_request *req)
3121 {
3122         int err = 0;
3123
3124         /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
3125         if (req->r_inode)
3126                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
3127         if (req->r_parent) {
3128                 struct ceph_inode_info *ci = ceph_inode(req->r_parent);
3129                 int fmode = (req->r_op & CEPH_MDS_OP_WRITE) ?
3130                             CEPH_FILE_MODE_WR : CEPH_FILE_MODE_RD;
3131                 spin_lock(&ci->i_ceph_lock);
3132                 ceph_take_cap_refs(ci, CEPH_CAP_PIN, false);
3133                 __ceph_touch_fmode(ci, mdsc, fmode);
3134                 spin_unlock(&ci->i_ceph_lock);
3135         }
3136         if (req->r_old_dentry_dir)
3137                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
3138                                   CEPH_CAP_PIN);
3139
3140         if (req->r_inode) {
3141                 err = ceph_wait_on_async_create(req->r_inode);
3142                 if (err) {
3143                         dout("%s: wait for async create returned: %d\n",
3144                              __func__, err);
3145                         return err;
3146                 }
3147         }
3148
3149         if (!err && req->r_old_inode) {
3150                 err = ceph_wait_on_async_create(req->r_old_inode);
3151                 if (err) {
3152                         dout("%s: wait for async create returned: %d\n",
3153                              __func__, err);
3154                         return err;
3155                 }
3156         }
3157
3158         dout("submit_request on %p for inode %p\n", req, dir);
3159         mutex_lock(&mdsc->mutex);
3160         __register_request(mdsc, req, dir);
3161         __do_request(mdsc, req);
3162         err = req->r_err;
3163         mutex_unlock(&mdsc->mutex);
3164         return err;
3165 }
3166
3167 int ceph_mdsc_wait_request(struct ceph_mds_client *mdsc,
3168                            struct ceph_mds_request *req,
3169                            ceph_mds_request_wait_callback_t wait_func)
3170 {
3171         int err;
3172
3173         /* wait */
3174         dout("do_request waiting\n");
3175         if (wait_func) {
3176                 err = wait_func(mdsc, req);
3177         } else {
3178                 long timeleft = wait_for_completion_killable_timeout(
3179                                         &req->r_completion,
3180                                         ceph_timeout_jiffies(req->r_timeout));
3181                 if (timeleft > 0)
3182                         err = 0;
3183                 else if (!timeleft)
3184                         err = -ETIMEDOUT;  /* timed out */
3185                 else
3186                         err = timeleft;  /* killed */
3187         }
3188         dout("do_request waited, got %d\n", err);
3189         mutex_lock(&mdsc->mutex);
3190
3191         /* only abort if we didn't race with a real reply */
3192         if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
3193                 err = le32_to_cpu(req->r_reply_info.head->result);
3194         } else if (err < 0) {
3195                 dout("aborted request %lld with %d\n", req->r_tid, err);
3196
3197                 /*
3198                  * ensure we aren't running concurrently with
3199                  * ceph_fill_trace or ceph_readdir_prepopulate, which
3200                  * rely on locks (dir mutex) held by our caller.
3201                  */
3202                 mutex_lock(&req->r_fill_mutex);
3203                 req->r_err = err;
3204                 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
3205                 mutex_unlock(&req->r_fill_mutex);
3206
3207                 if (req->r_parent &&
3208                     (req->r_op & CEPH_MDS_OP_WRITE))
3209                         ceph_invalidate_dir_request(req);
3210         } else {
3211                 err = req->r_err;
3212         }
3213
3214         mutex_unlock(&mdsc->mutex);
3215         return err;
3216 }
3217
3218 /*
3219  * Synchrously perform an mds request.  Take care of all of the
3220  * session setup, forwarding, retry details.
3221  */
3222 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
3223                          struct inode *dir,
3224                          struct ceph_mds_request *req)
3225 {
3226         int err;
3227
3228         dout("do_request on %p\n", req);
3229
3230         /* issue */
3231         err = ceph_mdsc_submit_request(mdsc, dir, req);
3232         if (!err)
3233                 err = ceph_mdsc_wait_request(mdsc, req, NULL);
3234         dout("do_request %p done, result %d\n", req, err);
3235         return err;
3236 }
3237
3238 /*
3239  * Invalidate dir's completeness, dentry lease state on an aborted MDS
3240  * namespace request.
3241  */
3242 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
3243 {
3244         struct inode *dir = req->r_parent;
3245         struct inode *old_dir = req->r_old_dentry_dir;
3246
3247         dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
3248
3249         ceph_dir_clear_complete(dir);
3250         if (old_dir)
3251                 ceph_dir_clear_complete(old_dir);
3252         if (req->r_dentry)
3253                 ceph_invalidate_dentry_lease(req->r_dentry);
3254         if (req->r_old_dentry)
3255                 ceph_invalidate_dentry_lease(req->r_old_dentry);
3256 }
3257
3258 /*
3259  * Handle mds reply.
3260  *
3261  * We take the session mutex and parse and process the reply immediately.
3262  * This preserves the logical ordering of replies, capabilities, etc., sent
3263  * by the MDS as they are applied to our local cache.
3264  */
3265 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
3266 {
3267         struct ceph_mds_client *mdsc = session->s_mdsc;
3268         struct ceph_mds_request *req;
3269         struct ceph_mds_reply_head *head = msg->front.iov_base;
3270         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
3271         struct ceph_snap_realm *realm;
3272         u64 tid;
3273         int err, result;
3274         int mds = session->s_mds;
3275         bool close_sessions = false;
3276
3277         if (msg->front.iov_len < sizeof(*head)) {
3278                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
3279                 ceph_msg_dump(msg);
3280                 return;
3281         }
3282
3283         /* get request, session */
3284         tid = le64_to_cpu(msg->hdr.tid);
3285         mutex_lock(&mdsc->mutex);
3286         req = lookup_get_request(mdsc, tid);
3287         if (!req) {
3288                 dout("handle_reply on unknown tid %llu\n", tid);
3289                 mutex_unlock(&mdsc->mutex);
3290                 return;
3291         }
3292         dout("handle_reply %p\n", req);
3293
3294         /* correct session? */
3295         if (req->r_session != session) {
3296                 pr_err("mdsc_handle_reply got %llu on session mds%d"
3297                        " not mds%d\n", tid, session->s_mds,
3298                        req->r_session ? req->r_session->s_mds : -1);
3299                 mutex_unlock(&mdsc->mutex);
3300                 goto out;
3301         }
3302
3303         /* dup? */
3304         if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
3305             (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
3306                 pr_warn("got a dup %s reply on %llu from mds%d\n",
3307                            head->safe ? "safe" : "unsafe", tid, mds);
3308                 mutex_unlock(&mdsc->mutex);
3309                 goto out;
3310         }
3311         if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
3312                 pr_warn("got unsafe after safe on %llu from mds%d\n",
3313                            tid, mds);
3314                 mutex_unlock(&mdsc->mutex);
3315                 goto out;
3316         }
3317
3318         result = le32_to_cpu(head->result);
3319
3320         if (head->safe) {
3321                 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
3322                 __unregister_request(mdsc, req);
3323
3324                 /* last request during umount? */
3325                 if (mdsc->stopping && !__get_oldest_req(mdsc))
3326                         complete_all(&mdsc->safe_umount_waiters);
3327
3328                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3329                         /*
3330                          * We already handled the unsafe response, now do the
3331                          * cleanup.  No need to examine the response; the MDS
3332                          * doesn't include any result info in the safe
3333                          * response.  And even if it did, there is nothing
3334                          * useful we could do with a revised return value.
3335                          */
3336                         dout("got safe reply %llu, mds%d\n", tid, mds);
3337
3338                         mutex_unlock(&mdsc->mutex);
3339                         goto out;
3340                 }
3341         } else {
3342                 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
3343                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
3344         }
3345
3346         dout("handle_reply tid %lld result %d\n", tid, result);
3347         rinfo = &req->r_reply_info;
3348         if (test_bit(CEPHFS_FEATURE_REPLY_ENCODING, &session->s_features))
3349                 err = parse_reply_info(session, msg, rinfo, (u64)-1);
3350         else
3351                 err = parse_reply_info(session, msg, rinfo, session->s_con.peer_features);
3352         mutex_unlock(&mdsc->mutex);
3353
3354         /* Must find target inode outside of mutexes to avoid deadlocks */
3355         if ((err >= 0) && rinfo->head->is_target) {
3356                 struct inode *in;
3357                 struct ceph_vino tvino = {
3358                         .ino  = le64_to_cpu(rinfo->targeti.in->ino),
3359                         .snap = le64_to_cpu(rinfo->targeti.in->snapid)
3360                 };
3361
3362                 in = ceph_get_inode(mdsc->fsc->sb, tvino);
3363                 if (IS_ERR(in)) {
3364                         err = PTR_ERR(in);
3365                         mutex_lock(&session->s_mutex);
3366                         goto out_err;
3367                 }
3368                 req->r_target_inode = in;
3369         }
3370
3371         mutex_lock(&session->s_mutex);
3372         if (err < 0) {
3373                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
3374                 ceph_msg_dump(msg);
3375                 goto out_err;
3376         }
3377
3378         /* snap trace */
3379         realm = NULL;
3380         if (rinfo->snapblob_len) {
3381                 down_write(&mdsc->snap_rwsem);
3382                 err = ceph_update_snap_trace(mdsc, rinfo->snapblob,
3383                                 rinfo->snapblob + rinfo->snapblob_len,
3384                                 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
3385                                 &realm);
3386                 if (err) {
3387                         up_write(&mdsc->snap_rwsem);
3388                         close_sessions = true;
3389                         if (err == -EIO)
3390                                 ceph_msg_dump(msg);
3391                         goto out_err;
3392                 }
3393                 downgrade_write(&mdsc->snap_rwsem);
3394         } else {
3395                 down_read(&mdsc->snap_rwsem);
3396         }
3397
3398         /* insert trace into our cache */
3399         mutex_lock(&req->r_fill_mutex);
3400         current->journal_info = req;
3401         err = ceph_fill_trace(mdsc->fsc->sb, req);
3402         if (err == 0) {
3403                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
3404                                     req->r_op == CEPH_MDS_OP_LSSNAP))
3405                         ceph_readdir_prepopulate(req, req->r_session);
3406         }
3407         current->journal_info = NULL;
3408         mutex_unlock(&req->r_fill_mutex);
3409
3410         up_read(&mdsc->snap_rwsem);
3411         if (realm)
3412                 ceph_put_snap_realm(mdsc, realm);
3413
3414         if (err == 0) {
3415                 if (req->r_target_inode &&
3416                     test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3417                         struct ceph_inode_info *ci =
3418                                 ceph_inode(req->r_target_inode);
3419                         spin_lock(&ci->i_unsafe_lock);
3420                         list_add_tail(&req->r_unsafe_target_item,
3421                                       &ci->i_unsafe_iops);
3422                         spin_unlock(&ci->i_unsafe_lock);
3423                 }
3424
3425                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
3426         }
3427 out_err:
3428         mutex_lock(&mdsc->mutex);
3429         if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3430                 if (err) {
3431                         req->r_err = err;
3432                 } else {
3433                         req->r_reply =  ceph_msg_get(msg);
3434                         set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
3435                 }
3436         } else {
3437                 dout("reply arrived after request %lld was aborted\n", tid);
3438         }
3439         mutex_unlock(&mdsc->mutex);
3440
3441         mutex_unlock(&session->s_mutex);
3442
3443         /* kick calling process */
3444         complete_request(mdsc, req);
3445
3446         ceph_update_metadata_metrics(&mdsc->metric, req->r_start_latency,
3447                                      req->r_end_latency, err);
3448 out:
3449         ceph_mdsc_put_request(req);
3450
3451         /* Defer closing the sessions after s_mutex lock being released */
3452         if (close_sessions)
3453                 ceph_mdsc_close_sessions(mdsc);
3454         return;
3455 }
3456
3457
3458
3459 /*
3460  * handle mds notification that our request has been forwarded.
3461  */
3462 static void handle_forward(struct ceph_mds_client *mdsc,
3463                            struct ceph_mds_session *session,
3464                            struct ceph_msg *msg)
3465 {
3466         struct ceph_mds_request *req;
3467         u64 tid = le64_to_cpu(msg->hdr.tid);
3468         u32 next_mds;
3469         u32 fwd_seq;
3470         int err = -EINVAL;
3471         void *p = msg->front.iov_base;
3472         void *end = p + msg->front.iov_len;
3473         bool aborted = false;
3474
3475         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
3476         next_mds = ceph_decode_32(&p);
3477         fwd_seq = ceph_decode_32(&p);
3478
3479         mutex_lock(&mdsc->mutex);
3480         req = lookup_get_request(mdsc, tid);
3481         if (!req) {
3482                 mutex_unlock(&mdsc->mutex);
3483                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
3484                 return;  /* dup reply? */
3485         }
3486
3487         if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3488                 dout("forward tid %llu aborted, unregistering\n", tid);
3489                 __unregister_request(mdsc, req);
3490         } else if (fwd_seq <= req->r_num_fwd) {
3491                 /*
3492                  * The type of 'num_fwd' in ceph 'MClientRequestForward'
3493                  * is 'int32_t', while in 'ceph_mds_request_head' the
3494                  * type is '__u8'. So in case the request bounces between
3495                  * MDSes exceeding 256 times, the client will get stuck.
3496                  *
3497                  * In this case it's ususally a bug in MDS and continue
3498                  * bouncing the request makes no sense.
3499                  *
3500                  * In future this could be fixed in ceph code, so avoid
3501                  * using the hardcode here.
3502                  */
3503                 int max = sizeof_field(struct ceph_mds_request_head, num_fwd);
3504                 max = 1 << (max * BITS_PER_BYTE);
3505                 if (req->r_num_fwd >= max) {
3506                         mutex_lock(&req->r_fill_mutex);
3507                         req->r_err = -EMULTIHOP;
3508                         set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
3509                         mutex_unlock(&req->r_fill_mutex);
3510                         aborted = true;
3511                         pr_warn_ratelimited("forward tid %llu seq overflow\n",
3512                                             tid);
3513                 } else {
3514                         dout("forward tid %llu to mds%d - old seq %d <= %d\n",
3515                              tid, next_mds, req->r_num_fwd, fwd_seq);
3516                 }
3517         } else {
3518                 /* resend. forward race not possible; mds would drop */
3519                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
3520                 BUG_ON(req->r_err);
3521                 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
3522                 req->r_attempts = 0;
3523                 req->r_num_fwd = fwd_seq;
3524                 req->r_resend_mds = next_mds;
3525                 put_request_session(req);
3526                 __do_request(mdsc, req);
3527         }
3528         mutex_unlock(&mdsc->mutex);
3529
3530         /* kick calling process */
3531         if (aborted)
3532                 complete_request(mdsc, req);
3533         ceph_mdsc_put_request(req);
3534         return;
3535
3536 bad:
3537         pr_err("mdsc_handle_forward decode error err=%d\n", err);
3538         ceph_msg_dump(msg);
3539 }
3540
3541 static int __decode_session_metadata(void **p, void *end,
3542                                      bool *blocklisted)
3543 {
3544         /* map<string,string> */
3545         u32 n;
3546         bool err_str;
3547         ceph_decode_32_safe(p, end, n, bad);
3548         while (n-- > 0) {
3549                 u32 len;
3550                 ceph_decode_32_safe(p, end, len, bad);
3551                 ceph_decode_need(p, end, len, bad);
3552                 err_str = !strncmp(*p, "error_string", len);
3553                 *p += len;
3554                 ceph_decode_32_safe(p, end, len, bad);
3555                 ceph_decode_need(p, end, len, bad);
3556                 /*
3557                  * Match "blocklisted (blacklisted)" from newer MDSes,
3558                  * or "blacklisted" from older MDSes.
3559                  */
3560                 if (err_str && strnstr(*p, "blacklisted", len))
3561                         *blocklisted = true;
3562                 *p += len;
3563         }
3564         return 0;
3565 bad:
3566         return -1;
3567 }
3568
3569 /*
3570  * handle a mds session control message
3571  */
3572 static void handle_session(struct ceph_mds_session *session,
3573                            struct ceph_msg *msg)
3574 {
3575         struct ceph_mds_client *mdsc = session->s_mdsc;
3576         int mds = session->s_mds;
3577         int msg_version = le16_to_cpu(msg->hdr.version);
3578         void *p = msg->front.iov_base;
3579         void *end = p + msg->front.iov_len;
3580         struct ceph_mds_session_head *h;
3581         u32 op;
3582         u64 seq, features = 0;
3583         int wake = 0;
3584         bool blocklisted = false;
3585
3586         /* decode */
3587         ceph_decode_need(&p, end, sizeof(*h), bad);
3588         h = p;
3589         p += sizeof(*h);
3590
3591         op = le32_to_cpu(h->op);
3592         seq = le64_to_cpu(h->seq);
3593
3594         if (msg_version >= 3) {
3595                 u32 len;
3596                 /* version >= 2 and < 5, decode metadata, skip otherwise
3597                  * as it's handled via flags.
3598                  */
3599                 if (msg_version >= 5)
3600                         ceph_decode_skip_map(&p, end, string, string, bad);
3601                 else if (__decode_session_metadata(&p, end, &blocklisted) < 0)
3602                         goto bad;
3603
3604                 /* version >= 3, feature bits */
3605                 ceph_decode_32_safe(&p, end, len, bad);
3606                 if (len) {
3607                         ceph_decode_64_safe(&p, end, features, bad);
3608                         p += len - sizeof(features);
3609                 }
3610         }
3611
3612         if (msg_version >= 5) {
3613                 u32 flags, len;
3614
3615                 /* version >= 4 */
3616                 ceph_decode_skip_16(&p, end, bad); /* struct_v, struct_cv */
3617                 ceph_decode_32_safe(&p, end, len, bad); /* len */
3618                 ceph_decode_skip_n(&p, end, len, bad); /* metric_spec */
3619
3620                 /* version >= 5, flags   */
3621                 ceph_decode_32_safe(&p, end, flags, bad);
3622                 if (flags & CEPH_SESSION_BLOCKLISTED) {
3623                         pr_warn("mds%d session blocklisted\n", session->s_mds);
3624                         blocklisted = true;
3625                 }
3626         }
3627
3628         mutex_lock(&mdsc->mutex);
3629         if (op == CEPH_SESSION_CLOSE) {
3630                 ceph_get_mds_session(session);
3631                 __unregister_session(mdsc, session);
3632         }
3633         /* FIXME: this ttl calculation is generous */
3634         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
3635         mutex_unlock(&mdsc->mutex);
3636
3637         mutex_lock(&session->s_mutex);
3638
3639         dout("handle_session mds%d %s %p state %s seq %llu\n",
3640              mds, ceph_session_op_name(op), session,
3641              ceph_session_state_name(session->s_state), seq);
3642
3643         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
3644                 session->s_state = CEPH_MDS_SESSION_OPEN;
3645                 pr_info("mds%d came back\n", session->s_mds);
3646         }
3647
3648         switch (op) {
3649         case CEPH_SESSION_OPEN:
3650                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3651                         pr_info("mds%d reconnect success\n", session->s_mds);
3652
3653                 session->s_features = features;
3654                 if (session->s_state == CEPH_MDS_SESSION_OPEN) {
3655                         pr_notice("mds%d is already opened\n", session->s_mds);
3656                 } else {
3657                         session->s_state = CEPH_MDS_SESSION_OPEN;
3658                         renewed_caps(mdsc, session, 0);
3659                         if (test_bit(CEPHFS_FEATURE_METRIC_COLLECT,
3660                                      &session->s_features))
3661                                 metric_schedule_delayed(&mdsc->metric);
3662                 }
3663
3664                 /*
3665                  * The connection maybe broken and the session in client
3666                  * side has been reinitialized, need to update the seq
3667                  * anyway.
3668                  */
3669                 if (!session->s_seq && seq)
3670                         session->s_seq = seq;
3671
3672                 wake = 1;
3673                 if (mdsc->stopping)
3674                         __close_session(mdsc, session);
3675                 break;
3676
3677         case CEPH_SESSION_RENEWCAPS:
3678                 if (session->s_renew_seq == seq)
3679                         renewed_caps(mdsc, session, 1);
3680                 break;
3681
3682         case CEPH_SESSION_CLOSE:
3683                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3684                         pr_info("mds%d reconnect denied\n", session->s_mds);
3685                 session->s_state = CEPH_MDS_SESSION_CLOSED;
3686                 cleanup_session_requests(mdsc, session);
3687                 remove_session_caps(session);
3688                 wake = 2; /* for good measure */
3689                 wake_up_all(&mdsc->session_close_wq);
3690                 break;
3691
3692         case CEPH_SESSION_STALE:
3693                 pr_info("mds%d caps went stale, renewing\n",
3694                         session->s_mds);
3695                 atomic_inc(&session->s_cap_gen);
3696                 session->s_cap_ttl = jiffies - 1;
3697                 send_renew_caps(mdsc, session);
3698                 break;
3699
3700         case CEPH_SESSION_RECALL_STATE:
3701                 ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
3702                 break;
3703
3704         case CEPH_SESSION_FLUSHMSG:
3705                 /* flush cap releases */
3706                 spin_lock(&session->s_cap_lock);
3707                 if (session->s_num_cap_releases)
3708                         ceph_flush_cap_releases(mdsc, session);
3709                 spin_unlock(&session->s_cap_lock);
3710
3711                 send_flushmsg_ack(mdsc, session, seq);
3712                 break;
3713
3714         case CEPH_SESSION_FORCE_RO:
3715                 dout("force_session_readonly %p\n", session);
3716                 spin_lock(&session->s_cap_lock);
3717                 session->s_readonly = true;
3718                 spin_unlock(&session->s_cap_lock);
3719                 wake_up_session_caps(session, FORCE_RO);
3720                 break;
3721
3722         case CEPH_SESSION_REJECT:
3723                 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
3724                 pr_info("mds%d rejected session\n", session->s_mds);
3725                 session->s_state = CEPH_MDS_SESSION_REJECTED;
3726                 cleanup_session_requests(mdsc, session);
3727                 remove_session_caps(session);
3728                 if (blocklisted)
3729                         mdsc->fsc->blocklisted = true;
3730                 wake = 2; /* for good measure */
3731                 break;
3732
3733         default:
3734                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
3735                 WARN_ON(1);
3736         }
3737
3738         mutex_unlock(&session->s_mutex);
3739         if (wake) {
3740                 mutex_lock(&mdsc->mutex);
3741                 __wake_requests(mdsc, &session->s_waiting);
3742                 if (wake == 2)
3743                         kick_requests(mdsc, mds);
3744                 mutex_unlock(&mdsc->mutex);
3745         }
3746         if (op == CEPH_SESSION_CLOSE)
3747                 ceph_put_mds_session(session);
3748         return;
3749
3750 bad:
3751         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
3752                (int)msg->front.iov_len);
3753         ceph_msg_dump(msg);
3754         return;
3755 }
3756
3757 void ceph_mdsc_release_dir_caps(struct ceph_mds_request *req)
3758 {
3759         int dcaps;
3760
3761         dcaps = xchg(&req->r_dir_caps, 0);
3762         if (dcaps) {
3763                 dout("releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
3764                 ceph_put_cap_refs(ceph_inode(req->r_parent), dcaps);
3765         }
3766 }
3767
3768 void ceph_mdsc_release_dir_caps_no_check(struct ceph_mds_request *req)
3769 {
3770         int dcaps;
3771
3772         dcaps = xchg(&req->r_dir_caps, 0);
3773         if (dcaps) {
3774                 dout("releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
3775                 ceph_put_cap_refs_no_check_caps(ceph_inode(req->r_parent),
3776                                                 dcaps);
3777         }
3778 }
3779
3780 /*
3781  * called under session->mutex.
3782  */
3783 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
3784                                    struct ceph_mds_session *session)
3785 {
3786         struct ceph_mds_request *req, *nreq;
3787         struct rb_node *p;
3788
3789         dout("replay_unsafe_requests mds%d\n", session->s_mds);
3790
3791         mutex_lock(&mdsc->mutex);
3792         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item)
3793                 __send_request(session, req, true);
3794
3795         /*
3796          * also re-send old requests when MDS enters reconnect stage. So that MDS
3797          * can process completed request in clientreplay stage.
3798          */
3799         p = rb_first(&mdsc->request_tree);
3800         while (p) {
3801                 req = rb_entry(p, struct ceph_mds_request, r_node);
3802                 p = rb_next(p);
3803                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3804                         continue;
3805                 if (req->r_attempts == 0)
3806                         continue; /* only old requests */
3807                 if (!req->r_session)
3808                         continue;
3809                 if (req->r_session->s_mds != session->s_mds)
3810                         continue;
3811
3812                 ceph_mdsc_release_dir_caps_no_check(req);
3813
3814                 __send_request(session, req, true);
3815         }
3816         mutex_unlock(&mdsc->mutex);
3817 }
3818
3819 static int send_reconnect_partial(struct ceph_reconnect_state *recon_state)
3820 {
3821         struct ceph_msg *reply;
3822         struct ceph_pagelist *_pagelist;
3823         struct page *page;
3824         __le32 *addr;
3825         int err = -ENOMEM;
3826
3827         if (!recon_state->allow_multi)
3828                 return -ENOSPC;
3829
3830         /* can't handle message that contains both caps and realm */
3831         BUG_ON(!recon_state->nr_caps == !recon_state->nr_realms);
3832
3833         /* pre-allocate new pagelist */
3834         _pagelist = ceph_pagelist_alloc(GFP_NOFS);
3835         if (!_pagelist)
3836                 return -ENOMEM;
3837
3838         reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3839         if (!reply)
3840                 goto fail_msg;
3841
3842         /* placeholder for nr_caps */
3843         err = ceph_pagelist_encode_32(_pagelist, 0);
3844         if (err < 0)
3845                 goto fail;
3846
3847         if (recon_state->nr_caps) {
3848                 /* currently encoding caps */
3849                 err = ceph_pagelist_encode_32(recon_state->pagelist, 0);
3850                 if (err)
3851                         goto fail;
3852         } else {
3853                 /* placeholder for nr_realms (currently encoding relams) */
3854                 err = ceph_pagelist_encode_32(_pagelist, 0);
3855                 if (err < 0)
3856                         goto fail;
3857         }
3858
3859         err = ceph_pagelist_encode_8(recon_state->pagelist, 1);
3860         if (err)
3861                 goto fail;
3862
3863         page = list_first_entry(&recon_state->pagelist->head, struct page, lru);
3864         addr = kmap_atomic(page);
3865         if (recon_state->nr_caps) {
3866                 /* currently encoding caps */
3867                 *addr = cpu_to_le32(recon_state->nr_caps);
3868         } else {
3869                 /* currently encoding relams */
3870                 *(addr + 1) = cpu_to_le32(recon_state->nr_realms);
3871         }
3872         kunmap_atomic(addr);
3873
3874         reply->hdr.version = cpu_to_le16(5);
3875         reply->hdr.compat_version = cpu_to_le16(4);
3876
3877         reply->hdr.data_len = cpu_to_le32(recon_state->pagelist->length);
3878         ceph_msg_data_add_pagelist(reply, recon_state->pagelist);
3879
3880         ceph_con_send(&recon_state->session->s_con, reply);
3881         ceph_pagelist_release(recon_state->pagelist);
3882
3883         recon_state->pagelist = _pagelist;
3884         recon_state->nr_caps = 0;
3885         recon_state->nr_realms = 0;
3886         recon_state->msg_version = 5;
3887         return 0;
3888 fail:
3889         ceph_msg_put(reply);
3890 fail_msg:
3891         ceph_pagelist_release(_pagelist);
3892         return err;
3893 }
3894
3895 static struct dentry* d_find_primary(struct inode *inode)
3896 {
3897         struct dentry *alias, *dn = NULL;
3898
3899         if (hlist_empty(&inode->i_dentry))
3900                 return NULL;
3901
3902         spin_lock(&inode->i_lock);
3903         if (hlist_empty(&inode->i_dentry))
3904                 goto out_unlock;
3905
3906         if (S_ISDIR(inode->i_mode)) {
3907                 alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias);
3908                 if (!IS_ROOT(alias))
3909                         dn = dget(alias);
3910                 goto out_unlock;
3911         }
3912
3913         hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
3914                 spin_lock(&alias->d_lock);
3915                 if (!d_unhashed(alias) &&
3916                     (ceph_dentry(alias)->flags & CEPH_DENTRY_PRIMARY_LINK)) {
3917                         dn = dget_dlock(alias);
3918                 }
3919                 spin_unlock(&alias->d_lock);
3920                 if (dn)
3921                         break;
3922         }
3923 out_unlock:
3924         spin_unlock(&inode->i_lock);
3925         return dn;
3926 }
3927
3928 /*
3929  * Encode information about a cap for a reconnect with the MDS.
3930  */
3931 static int reconnect_caps_cb(struct inode *inode, int mds, void *arg)
3932 {
3933         union {
3934                 struct ceph_mds_cap_reconnect v2;
3935                 struct ceph_mds_cap_reconnect_v1 v1;
3936         } rec;
3937         struct ceph_inode_info *ci = ceph_inode(inode);
3938         struct ceph_reconnect_state *recon_state = arg;
3939         struct ceph_pagelist *pagelist = recon_state->pagelist;
3940         struct dentry *dentry;
3941         struct ceph_cap *cap;
3942         char *path;
3943         int pathlen = 0, err;
3944         u64 pathbase;
3945         u64 snap_follows;
3946
3947         dentry = d_find_primary(inode);
3948         if (dentry) {
3949                 /* set pathbase to parent dir when msg_version >= 2 */
3950                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase,
3951                                             recon_state->msg_version >= 2);
3952                 dput(dentry);
3953                 if (IS_ERR(path)) {
3954                         err = PTR_ERR(path);
3955                         goto out_err;
3956                 }
3957         } else {
3958                 path = NULL;
3959                 pathbase = 0;
3960         }
3961
3962         spin_lock(&ci->i_ceph_lock);
3963         cap = __get_cap_for_mds(ci, mds);
3964         if (!cap) {
3965                 spin_unlock(&ci->i_ceph_lock);
3966                 err = 0;
3967                 goto out_err;
3968         }
3969         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
3970              inode, ceph_vinop(inode), cap, cap->cap_id,
3971              ceph_cap_string(cap->issued));
3972
3973         cap->seq = 0;        /* reset cap seq */
3974         cap->issue_seq = 0;  /* and issue_seq */
3975         cap->mseq = 0;       /* and migrate_seq */
3976         cap->cap_gen = atomic_read(&cap->session->s_cap_gen);
3977
3978         /* These are lost when the session goes away */
3979         if (S_ISDIR(inode->i_mode)) {
3980                 if (cap->issued & CEPH_CAP_DIR_CREATE) {
3981                         ceph_put_string(rcu_dereference_raw(ci->i_cached_layout.pool_ns));
3982                         memset(&ci->i_cached_layout, 0, sizeof(ci->i_cached_layout));
3983                 }
3984                 cap->issued &= ~CEPH_CAP_ANY_DIR_OPS;
3985         }
3986
3987         if (recon_state->msg_version >= 2) {
3988                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
3989                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3990                 rec.v2.issued = cpu_to_le32(cap->issued);
3991                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3992                 rec.v2.pathbase = cpu_to_le64(pathbase);
3993                 rec.v2.flock_len = (__force __le32)
3994                         ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
3995         } else {
3996                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
3997                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3998                 rec.v1.issued = cpu_to_le32(cap->issued);
3999                 rec.v1.size = cpu_to_le64(i_size_read(inode));
4000                 ceph_encode_timespec64(&rec.v1.mtime, &inode->i_mtime);
4001                 ceph_encode_timespec64(&rec.v1.atime, &inode->i_atime);
4002                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
4003                 rec.v1.pathbase = cpu_to_le64(pathbase);
4004         }
4005
4006         if (list_empty(&ci->i_cap_snaps)) {
4007                 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
4008         } else {
4009                 struct ceph_cap_snap *capsnap =
4010                         list_first_entry(&ci->i_cap_snaps,
4011                                          struct ceph_cap_snap, ci_item);
4012                 snap_follows = capsnap->follows;
4013         }
4014         spin_unlock(&ci->i_ceph_lock);
4015
4016         if (recon_state->msg_version >= 2) {
4017                 int num_fcntl_locks, num_flock_locks;
4018                 struct ceph_filelock *flocks = NULL;
4019                 size_t struct_len, total_len = sizeof(u64);
4020                 u8 struct_v = 0;
4021
4022 encode_again:
4023                 if (rec.v2.flock_len) {
4024                         ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
4025                 } else {
4026                         num_fcntl_locks = 0;
4027                         num_flock_locks = 0;
4028                 }
4029                 if (num_fcntl_locks + num_flock_locks > 0) {
4030                         flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
4031                                                sizeof(struct ceph_filelock),
4032                                                GFP_NOFS);
4033                         if (!flocks) {
4034                                 err = -ENOMEM;
4035                                 goto out_err;
4036                         }
4037                         err = ceph_encode_locks_to_buffer(inode, flocks,
4038                                                           num_fcntl_locks,
4039                                                           num_flock_locks);
4040                         if (err) {
4041                                 kfree(flocks);
4042                                 flocks = NULL;
4043                                 if (err == -ENOSPC)
4044                                         goto encode_again;
4045                                 goto out_err;
4046                         }
4047                 } else {
4048                         kfree(flocks);
4049                         flocks = NULL;
4050                 }
4051
4052                 if (recon_state->msg_version >= 3) {
4053                         /* version, compat_version and struct_len */
4054                         total_len += 2 * sizeof(u8) + sizeof(u32);
4055                         struct_v = 2;
4056                 }
4057                 /*
4058                  * number of encoded locks is stable, so copy to pagelist
4059                  */
4060                 struct_len = 2 * sizeof(u32) +
4061                             (num_fcntl_locks + num_flock_locks) *
4062                             sizeof(struct ceph_filelock);
4063                 rec.v2.flock_len = cpu_to_le32(struct_len);
4064
4065                 struct_len += sizeof(u32) + pathlen + sizeof(rec.v2);
4066
4067                 if (struct_v >= 2)
4068                         struct_len += sizeof(u64); /* snap_follows */
4069
4070                 total_len += struct_len;
4071
4072                 if (pagelist->length + total_len > RECONNECT_MAX_SIZE) {
4073                         err = send_reconnect_partial(recon_state);
4074                         if (err)
4075                                 goto out_freeflocks;
4076                         pagelist = recon_state->pagelist;
4077                 }
4078
4079                 err = ceph_pagelist_reserve(pagelist, total_len);
4080                 if (err)
4081                         goto out_freeflocks;
4082
4083                 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
4084                 if (recon_state->msg_version >= 3) {
4085                         ceph_pagelist_encode_8(pagelist, struct_v);
4086                         ceph_pagelist_encode_8(pagelist, 1);
4087                         ceph_pagelist_encode_32(pagelist, struct_len);
4088                 }
4089                 ceph_pagelist_encode_string(pagelist, path, pathlen);
4090                 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
4091                 ceph_locks_to_pagelist(flocks, pagelist,
4092                                        num_fcntl_locks, num_flock_locks);
4093                 if (struct_v >= 2)
4094                         ceph_pagelist_encode_64(pagelist, snap_follows);
4095 out_freeflocks:
4096                 kfree(flocks);
4097         } else {
4098                 err = ceph_pagelist_reserve(pagelist,
4099                                             sizeof(u64) + sizeof(u32) +
4100                                             pathlen + sizeof(rec.v1));
4101                 if (err)
4102                         goto out_err;
4103
4104                 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
4105                 ceph_pagelist_encode_string(pagelist, path, pathlen);
4106                 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
4107         }
4108
4109 out_err:
4110         ceph_mdsc_free_path(path, pathlen);
4111         if (!err)
4112                 recon_state->nr_caps++;
4113         return err;
4114 }
4115
4116 static int encode_snap_realms(struct ceph_mds_client *mdsc,
4117                               struct ceph_reconnect_state *recon_state)
4118 {
4119         struct rb_node *p;
4120         struct ceph_pagelist *pagelist = recon_state->pagelist;
4121         int err = 0;
4122
4123         if (recon_state->msg_version >= 4) {
4124                 err = ceph_pagelist_encode_32(pagelist, mdsc->num_snap_realms);
4125                 if (err < 0)
4126                         goto fail;
4127         }
4128
4129         /*
4130          * snaprealms.  we provide mds with the ino, seq (version), and
4131          * parent for all of our realms.  If the mds has any newer info,
4132          * it will tell us.
4133          */
4134         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
4135                 struct ceph_snap_realm *realm =
4136                        rb_entry(p, struct ceph_snap_realm, node);
4137                 struct ceph_mds_snaprealm_reconnect sr_rec;
4138
4139                 if (recon_state->msg_version >= 4) {
4140                         size_t need = sizeof(u8) * 2 + sizeof(u32) +
4141                                       sizeof(sr_rec);
4142
4143                         if (pagelist->length + need > RECONNECT_MAX_SIZE) {
4144                                 err = send_reconnect_partial(recon_state);
4145                                 if (err)
4146                                         goto fail;
4147                                 pagelist = recon_state->pagelist;
4148                         }
4149
4150                         err = ceph_pagelist_reserve(pagelist, need);
4151                         if (err)
4152                                 goto fail;
4153
4154                         ceph_pagelist_encode_8(pagelist, 1);
4155                         ceph_pagelist_encode_8(pagelist, 1);
4156                         ceph_pagelist_encode_32(pagelist, sizeof(sr_rec));
4157                 }
4158
4159                 dout(" adding snap realm %llx seq %lld parent %llx\n",
4160                      realm->ino, realm->seq, realm->parent_ino);
4161                 sr_rec.ino = cpu_to_le64(realm->ino);
4162                 sr_rec.seq = cpu_to_le64(realm->seq);
4163                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
4164
4165                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
4166                 if (err)
4167                         goto fail;
4168
4169                 recon_state->nr_realms++;
4170         }
4171 fail:
4172         return err;
4173 }
4174
4175
4176 /*
4177  * If an MDS fails and recovers, clients need to reconnect in order to
4178  * reestablish shared state.  This includes all caps issued through
4179  * this session _and_ the snap_realm hierarchy.  Because it's not
4180  * clear which snap realms the mds cares about, we send everything we
4181  * know about.. that ensures we'll then get any new info the
4182  * recovering MDS might have.
4183  *
4184  * This is a relatively heavyweight operation, but it's rare.
4185  */
4186 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
4187                                struct ceph_mds_session *session)
4188 {
4189         struct ceph_msg *reply;
4190         int mds = session->s_mds;
4191         int err = -ENOMEM;
4192         struct ceph_reconnect_state recon_state = {
4193                 .session = session,
4194         };
4195         LIST_HEAD(dispose);
4196
4197         pr_info("mds%d reconnect start\n", mds);
4198
4199         recon_state.pagelist = ceph_pagelist_alloc(GFP_NOFS);
4200         if (!recon_state.pagelist)
4201                 goto fail_nopagelist;
4202
4203         reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
4204         if (!reply)
4205                 goto fail_nomsg;
4206
4207         xa_destroy(&session->s_delegated_inos);
4208
4209         mutex_lock(&session->s_mutex);
4210         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
4211         session->s_seq = 0;
4212
4213         dout("session %p state %s\n", session,
4214              ceph_session_state_name(session->s_state));
4215
4216         atomic_inc(&session->s_cap_gen);
4217
4218         spin_lock(&session->s_cap_lock);
4219         /* don't know if session is readonly */
4220         session->s_readonly = 0;
4221         /*
4222          * notify __ceph_remove_cap() that we are composing cap reconnect.
4223          * If a cap get released before being added to the cap reconnect,
4224          * __ceph_remove_cap() should skip queuing cap release.
4225          */
4226         session->s_cap_reconnect = 1;
4227         /* drop old cap expires; we're about to reestablish that state */
4228         detach_cap_releases(session, &dispose);
4229         spin_unlock(&session->s_cap_lock);
4230         dispose_cap_releases(mdsc, &dispose);
4231
4232         /* trim unused caps to reduce MDS's cache rejoin time */
4233         if (mdsc->fsc->sb->s_root)
4234                 shrink_dcache_parent(mdsc->fsc->sb->s_root);
4235
4236         ceph_con_close(&session->s_con);
4237         ceph_con_open(&session->s_con,
4238                       CEPH_ENTITY_TYPE_MDS, mds,
4239                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
4240
4241         /* replay unsafe requests */
4242         replay_unsafe_requests(mdsc, session);
4243
4244         ceph_early_kick_flushing_caps(mdsc, session);
4245
4246         down_read(&mdsc->snap_rwsem);
4247
4248         /* placeholder for nr_caps */
4249         err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
4250         if (err)
4251                 goto fail;
4252
4253         if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
4254                 recon_state.msg_version = 3;
4255                 recon_state.allow_multi = true;
4256         } else if (session->s_con.peer_features & CEPH_FEATURE_MDSENC) {
4257                 recon_state.msg_version = 3;
4258         } else {
4259                 recon_state.msg_version = 2;
4260         }
4261         /* trsaverse this session's caps */
4262         err = ceph_iterate_session_caps(session, reconnect_caps_cb, &recon_state);
4263
4264         spin_lock(&session->s_cap_lock);
4265         session->s_cap_reconnect = 0;
4266         spin_unlock(&session->s_cap_lock);
4267
4268         if (err < 0)
4269                 goto fail;
4270
4271         /* check if all realms can be encoded into current message */
4272         if (mdsc->num_snap_realms) {
4273                 size_t total_len =
4274                         recon_state.pagelist->length +
4275                         mdsc->num_snap_realms *
4276                         sizeof(struct ceph_mds_snaprealm_reconnect);
4277                 if (recon_state.msg_version >= 4) {
4278                         /* number of realms */
4279                         total_len += sizeof(u32);
4280                         /* version, compat_version and struct_len */
4281                         total_len += mdsc->num_snap_realms *
4282                                      (2 * sizeof(u8) + sizeof(u32));
4283                 }
4284                 if (total_len > RECONNECT_MAX_SIZE) {
4285                         if (!recon_state.allow_multi) {
4286                                 err = -ENOSPC;
4287                                 goto fail;
4288                         }
4289                         if (recon_state.nr_caps) {
4290                                 err = send_reconnect_partial(&recon_state);
4291                                 if (err)
4292                                         goto fail;
4293                         }
4294                         recon_state.msg_version = 5;
4295                 }
4296         }
4297
4298         err = encode_snap_realms(mdsc, &recon_state);
4299         if (err < 0)
4300                 goto fail;
4301
4302         if (recon_state.msg_version >= 5) {
4303                 err = ceph_pagelist_encode_8(recon_state.pagelist, 0);
4304                 if (err < 0)
4305                         goto fail;
4306         }
4307
4308         if (recon_state.nr_caps || recon_state.nr_realms) {
4309                 struct page *page =
4310                         list_first_entry(&recon_state.pagelist->head,
4311                                         struct page, lru);
4312                 __le32 *addr = kmap_atomic(page);
4313                 if (recon_state.nr_caps) {
4314                         WARN_ON(recon_state.nr_realms != mdsc->num_snap_realms);
4315                         *addr = cpu_to_le32(recon_state.nr_caps);
4316                 } else if (recon_state.msg_version >= 4) {
4317                         *(addr + 1) = cpu_to_le32(recon_state.nr_realms);
4318                 }
4319                 kunmap_atomic(addr);
4320         }
4321
4322         reply->hdr.version = cpu_to_le16(recon_state.msg_version);
4323         if (recon_state.msg_version >= 4)
4324                 reply->hdr.compat_version = cpu_to_le16(4);
4325
4326         reply->hdr.data_len = cpu_to_le32(recon_state.pagelist->length);
4327         ceph_msg_data_add_pagelist(reply, recon_state.pagelist);
4328
4329         ceph_con_send(&session->s_con, reply);
4330
4331         mutex_unlock(&session->s_mutex);
4332
4333         mutex_lock(&mdsc->mutex);
4334         __wake_requests(mdsc, &session->s_waiting);
4335         mutex_unlock(&mdsc->mutex);
4336
4337         up_read(&mdsc->snap_rwsem);
4338         ceph_pagelist_release(recon_state.pagelist);
4339         return;
4340
4341 fail:
4342         ceph_msg_put(reply);
4343         up_read(&mdsc->snap_rwsem);
4344         mutex_unlock(&session->s_mutex);
4345 fail_nomsg:
4346         ceph_pagelist_release(recon_state.pagelist);
4347 fail_nopagelist:
4348         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
4349         return;
4350 }
4351
4352
4353 /*
4354  * compare old and new mdsmaps, kicking requests
4355  * and closing out old connections as necessary
4356  *
4357  * called under mdsc->mutex.
4358  */
4359 static void check_new_map(struct ceph_mds_client *mdsc,
4360                           struct ceph_mdsmap *newmap,
4361                           struct ceph_mdsmap *oldmap)
4362 {
4363         int i, j, err;
4364         int oldstate, newstate;
4365         struct ceph_mds_session *s;
4366         unsigned long targets[DIV_ROUND_UP(CEPH_MAX_MDS, sizeof(unsigned long))] = {0};
4367
4368         dout("check_new_map new %u old %u\n",
4369              newmap->m_epoch, oldmap->m_epoch);
4370
4371         if (newmap->m_info) {
4372                 for (i = 0; i < newmap->possible_max_rank; i++) {
4373                         for (j = 0; j < newmap->m_info[i].num_export_targets; j++)
4374                                 set_bit(newmap->m_info[i].export_targets[j], targets);
4375                 }
4376         }
4377
4378         for (i = 0; i < oldmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4379                 if (!mdsc->sessions[i])
4380                         continue;
4381                 s = mdsc->sessions[i];
4382                 oldstate = ceph_mdsmap_get_state(oldmap, i);
4383                 newstate = ceph_mdsmap_get_state(newmap, i);
4384
4385                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
4386                      i, ceph_mds_state_name(oldstate),
4387                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
4388                      ceph_mds_state_name(newstate),
4389                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
4390                      ceph_session_state_name(s->s_state));
4391
4392                 if (i >= newmap->possible_max_rank) {
4393                         /* force close session for stopped mds */
4394                         ceph_get_mds_session(s);
4395                         __unregister_session(mdsc, s);
4396                         __wake_requests(mdsc, &s->s_waiting);
4397                         mutex_unlock(&mdsc->mutex);
4398
4399                         mutex_lock(&s->s_mutex);
4400                         cleanup_session_requests(mdsc, s);
4401                         remove_session_caps(s);
4402                         mutex_unlock(&s->s_mutex);
4403
4404                         ceph_put_mds_session(s);
4405
4406                         mutex_lock(&mdsc->mutex);
4407                         kick_requests(mdsc, i);
4408                         continue;
4409                 }
4410
4411                 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
4412                            ceph_mdsmap_get_addr(newmap, i),
4413                            sizeof(struct ceph_entity_addr))) {
4414                         /* just close it */
4415                         mutex_unlock(&mdsc->mutex);
4416                         mutex_lock(&s->s_mutex);
4417                         mutex_lock(&mdsc->mutex);
4418                         ceph_con_close(&s->s_con);
4419                         mutex_unlock(&s->s_mutex);
4420                         s->s_state = CEPH_MDS_SESSION_RESTARTING;
4421                 } else if (oldstate == newstate) {
4422                         continue;  /* nothing new with this mds */
4423                 }
4424
4425                 /*
4426                  * send reconnect?
4427                  */
4428                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
4429                     newstate >= CEPH_MDS_STATE_RECONNECT) {
4430                         mutex_unlock(&mdsc->mutex);
4431                         clear_bit(i, targets);
4432                         send_mds_reconnect(mdsc, s);
4433                         mutex_lock(&mdsc->mutex);
4434                 }
4435
4436                 /*
4437                  * kick request on any mds that has gone active.
4438                  */
4439                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
4440                     newstate >= CEPH_MDS_STATE_ACTIVE) {
4441                         if (oldstate != CEPH_MDS_STATE_CREATING &&
4442                             oldstate != CEPH_MDS_STATE_STARTING)
4443                                 pr_info("mds%d recovery completed\n", s->s_mds);
4444                         kick_requests(mdsc, i);
4445                         mutex_unlock(&mdsc->mutex);
4446                         mutex_lock(&s->s_mutex);
4447                         mutex_lock(&mdsc->mutex);
4448                         ceph_kick_flushing_caps(mdsc, s);
4449                         mutex_unlock(&s->s_mutex);
4450                         wake_up_session_caps(s, RECONNECT);
4451                 }
4452         }
4453
4454         /*
4455          * Only open and reconnect sessions that don't exist yet.
4456          */
4457         for (i = 0; i < newmap->possible_max_rank; i++) {
4458                 /*
4459                  * In case the import MDS is crashed just after
4460                  * the EImportStart journal is flushed, so when
4461                  * a standby MDS takes over it and is replaying
4462                  * the EImportStart journal the new MDS daemon
4463                  * will wait the client to reconnect it, but the
4464                  * client may never register/open the session yet.
4465                  *
4466                  * Will try to reconnect that MDS daemon if the
4467                  * rank number is in the export targets array and
4468                  * is the up:reconnect state.
4469                  */
4470                 newstate = ceph_mdsmap_get_state(newmap, i);
4471                 if (!test_bit(i, targets) || newstate != CEPH_MDS_STATE_RECONNECT)
4472                         continue;
4473
4474                 /*
4475                  * The session maybe registered and opened by some
4476                  * requests which were choosing random MDSes during
4477                  * the mdsc->mutex's unlock/lock gap below in rare
4478                  * case. But the related MDS daemon will just queue
4479                  * that requests and be still waiting for the client's
4480                  * reconnection request in up:reconnect state.
4481                  */
4482                 s = __ceph_lookup_mds_session(mdsc, i);
4483                 if (likely(!s)) {
4484                         s = __open_export_target_session(mdsc, i);
4485                         if (IS_ERR(s)) {
4486                                 err = PTR_ERR(s);
4487                                 pr_err("failed to open export target session, err %d\n",
4488                                        err);
4489                                 continue;
4490                         }
4491                 }
4492                 dout("send reconnect to export target mds.%d\n", i);
4493                 mutex_unlock(&mdsc->mutex);
4494                 send_mds_reconnect(mdsc, s);
4495                 ceph_put_mds_session(s);
4496                 mutex_lock(&mdsc->mutex);
4497         }
4498
4499         for (i = 0; i < newmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4500                 s = mdsc->sessions[i];
4501                 if (!s)
4502                         continue;
4503                 if (!ceph_mdsmap_is_laggy(newmap, i))
4504                         continue;
4505                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4506                     s->s_state == CEPH_MDS_SESSION_HUNG ||
4507                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
4508                         dout(" connecting to export targets of laggy mds%d\n",
4509                              i);
4510                         __open_export_target_sessions(mdsc, s);
4511                 }
4512         }
4513 }
4514
4515
4516
4517 /*
4518  * leases
4519  */
4520
4521 /*
4522  * caller must hold session s_mutex, dentry->d_lock
4523  */
4524 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
4525 {
4526         struct ceph_dentry_info *di = ceph_dentry(dentry);
4527
4528         ceph_put_mds_session(di->lease_session);
4529         di->lease_session = NULL;
4530 }
4531
4532 static void handle_lease(struct ceph_mds_client *mdsc,
4533                          struct ceph_mds_session *session,
4534                          struct ceph_msg *msg)
4535 {
4536         struct super_block *sb = mdsc->fsc->sb;
4537         struct inode *inode;
4538         struct dentry *parent, *dentry;
4539         struct ceph_dentry_info *di;
4540         int mds = session->s_mds;
4541         struct ceph_mds_lease *h = msg->front.iov_base;
4542         u32 seq;
4543         struct ceph_vino vino;
4544         struct qstr dname;
4545         int release = 0;
4546
4547         dout("handle_lease from mds%d\n", mds);
4548
4549         if (!ceph_inc_mds_stopping_blocker(mdsc, session))
4550                 return;
4551
4552         /* decode */
4553         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
4554                 goto bad;
4555         vino.ino = le64_to_cpu(h->ino);
4556         vino.snap = CEPH_NOSNAP;
4557         seq = le32_to_cpu(h->seq);
4558         dname.len = get_unaligned_le32(h + 1);
4559         if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
4560                 goto bad;
4561         dname.name = (void *)(h + 1) + sizeof(u32);
4562
4563         /* lookup inode */
4564         inode = ceph_find_inode(sb, vino);
4565         dout("handle_lease %s, ino %llx %p %.*s\n",
4566              ceph_lease_op_name(h->action), vino.ino, inode,
4567              dname.len, dname.name);
4568
4569         mutex_lock(&session->s_mutex);
4570         if (!inode) {
4571                 dout("handle_lease no inode %llx\n", vino.ino);
4572                 goto release;
4573         }
4574
4575         /* dentry */
4576         parent = d_find_alias(inode);
4577         if (!parent) {
4578                 dout("no parent dentry on inode %p\n", inode);
4579                 WARN_ON(1);
4580                 goto release;  /* hrm... */
4581         }
4582         dname.hash = full_name_hash(parent, dname.name, dname.len);
4583         dentry = d_lookup(parent, &dname);
4584         dput(parent);
4585         if (!dentry)
4586                 goto release;
4587
4588         spin_lock(&dentry->d_lock);
4589         di = ceph_dentry(dentry);
4590         switch (h->action) {
4591         case CEPH_MDS_LEASE_REVOKE:
4592                 if (di->lease_session == session) {
4593                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
4594                                 h->seq = cpu_to_le32(di->lease_seq);
4595                         __ceph_mdsc_drop_dentry_lease(dentry);
4596                 }
4597                 release = 1;
4598                 break;
4599
4600         case CEPH_MDS_LEASE_RENEW:
4601                 if (di->lease_session == session &&
4602                     di->lease_gen == atomic_read(&session->s_cap_gen) &&
4603                     di->lease_renew_from &&
4604                     di->lease_renew_after == 0) {
4605                         unsigned long duration =
4606                                 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
4607
4608                         di->lease_seq = seq;
4609                         di->time = di->lease_renew_from + duration;
4610                         di->lease_renew_after = di->lease_renew_from +
4611                                 (duration >> 1);
4612                         di->lease_renew_from = 0;
4613                 }
4614                 break;
4615         }
4616         spin_unlock(&dentry->d_lock);
4617         dput(dentry);
4618
4619         if (!release)
4620                 goto out;
4621
4622 release:
4623         /* let's just reuse the same message */
4624         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
4625         ceph_msg_get(msg);
4626         ceph_con_send(&session->s_con, msg);
4627
4628 out:
4629         mutex_unlock(&session->s_mutex);
4630         iput(inode);
4631
4632         ceph_dec_mds_stopping_blocker(mdsc);
4633         return;
4634
4635 bad:
4636         ceph_dec_mds_stopping_blocker(mdsc);
4637
4638         pr_err("corrupt lease message\n");
4639         ceph_msg_dump(msg);
4640 }
4641
4642 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
4643                               struct dentry *dentry, char action,
4644                               u32 seq)
4645 {
4646         struct ceph_msg *msg;
4647         struct ceph_mds_lease *lease;
4648         struct inode *dir;
4649         int len = sizeof(*lease) + sizeof(u32) + NAME_MAX;
4650
4651         dout("lease_send_msg identry %p %s to mds%d\n",
4652              dentry, ceph_lease_op_name(action), session->s_mds);
4653
4654         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
4655         if (!msg)
4656                 return;
4657         lease = msg->front.iov_base;
4658         lease->action = action;
4659         lease->seq = cpu_to_le32(seq);
4660
4661         spin_lock(&dentry->d_lock);
4662         dir = d_inode(dentry->d_parent);
4663         lease->ino = cpu_to_le64(ceph_ino(dir));
4664         lease->first = lease->last = cpu_to_le64(ceph_snap(dir));
4665
4666         put_unaligned_le32(dentry->d_name.len, lease + 1);
4667         memcpy((void *)(lease + 1) + 4,
4668                dentry->d_name.name, dentry->d_name.len);
4669         spin_unlock(&dentry->d_lock);
4670
4671         ceph_con_send(&session->s_con, msg);
4672 }
4673
4674 /*
4675  * lock unlock the session, to wait ongoing session activities
4676  */
4677 static void lock_unlock_session(struct ceph_mds_session *s)
4678 {
4679         mutex_lock(&s->s_mutex);
4680         mutex_unlock(&s->s_mutex);
4681 }
4682
4683 static void maybe_recover_session(struct ceph_mds_client *mdsc)
4684 {
4685         struct ceph_fs_client *fsc = mdsc->fsc;
4686
4687         if (!ceph_test_mount_opt(fsc, CLEANRECOVER))
4688                 return;
4689
4690         if (READ_ONCE(fsc->mount_state) != CEPH_MOUNT_MOUNTED)
4691                 return;
4692
4693         if (!READ_ONCE(fsc->blocklisted))
4694                 return;
4695
4696         pr_info("auto reconnect after blocklisted\n");
4697         ceph_force_reconnect(fsc->sb);
4698 }
4699
4700 bool check_session_state(struct ceph_mds_session *s)
4701 {
4702         switch (s->s_state) {
4703         case CEPH_MDS_SESSION_OPEN:
4704                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
4705                         s->s_state = CEPH_MDS_SESSION_HUNG;
4706                         pr_info("mds%d hung\n", s->s_mds);
4707                 }
4708                 break;
4709         case CEPH_MDS_SESSION_CLOSING:
4710         case CEPH_MDS_SESSION_NEW:
4711         case CEPH_MDS_SESSION_RESTARTING:
4712         case CEPH_MDS_SESSION_CLOSED:
4713         case CEPH_MDS_SESSION_REJECTED:
4714                 return false;
4715         }
4716
4717         return true;
4718 }
4719
4720 /*
4721  * If the sequence is incremented while we're waiting on a REQUEST_CLOSE reply,
4722  * then we need to retransmit that request.
4723  */
4724 void inc_session_sequence(struct ceph_mds_session *s)
4725 {
4726         lockdep_assert_held(&s->s_mutex);
4727
4728         s->s_seq++;
4729
4730         if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
4731                 int ret;
4732
4733                 dout("resending session close request for mds%d\n", s->s_mds);
4734                 ret = request_close_session(s);
4735                 if (ret < 0)
4736                         pr_err("unable to close session to mds%d: %d\n",
4737                                s->s_mds, ret);
4738         }
4739 }
4740
4741 /*
4742  * delayed work -- periodically trim expired leases, renew caps with mds.  If
4743  * the @delay parameter is set to 0 or if it's more than 5 secs, the default
4744  * workqueue delay value of 5 secs will be used.
4745  */
4746 static void schedule_delayed(struct ceph_mds_client *mdsc, unsigned long delay)
4747 {
4748         unsigned long max_delay = HZ * 5;
4749
4750         /* 5 secs default delay */
4751         if (!delay || (delay > max_delay))
4752                 delay = max_delay;
4753         schedule_delayed_work(&mdsc->delayed_work,
4754                               round_jiffies_relative(delay));
4755 }
4756
4757 static void delayed_work(struct work_struct *work)
4758 {
4759         struct ceph_mds_client *mdsc =
4760                 container_of(work, struct ceph_mds_client, delayed_work.work);
4761         unsigned long delay;
4762         int renew_interval;
4763         int renew_caps;
4764         int i;
4765
4766         dout("mdsc delayed_work\n");
4767
4768         if (mdsc->stopping >= CEPH_MDSC_STOPPING_FLUSHED)
4769                 return;
4770
4771         mutex_lock(&mdsc->mutex);
4772         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
4773         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
4774                                    mdsc->last_renew_caps);
4775         if (renew_caps)
4776                 mdsc->last_renew_caps = jiffies;
4777
4778         for (i = 0; i < mdsc->max_sessions; i++) {
4779                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4780                 if (!s)
4781                         continue;
4782
4783                 if (!check_session_state(s)) {
4784                         ceph_put_mds_session(s);
4785                         continue;
4786                 }
4787                 mutex_unlock(&mdsc->mutex);
4788
4789                 mutex_lock(&s->s_mutex);
4790                 if (renew_caps)
4791                         send_renew_caps(mdsc, s);
4792                 else
4793                         ceph_con_keepalive(&s->s_con);
4794                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4795                     s->s_state == CEPH_MDS_SESSION_HUNG)
4796                         ceph_send_cap_releases(mdsc, s);
4797                 mutex_unlock(&s->s_mutex);
4798                 ceph_put_mds_session(s);
4799
4800                 mutex_lock(&mdsc->mutex);
4801         }
4802         mutex_unlock(&mdsc->mutex);
4803
4804         delay = ceph_check_delayed_caps(mdsc);
4805
4806         ceph_queue_cap_reclaim_work(mdsc);
4807
4808         ceph_trim_snapid_map(mdsc);
4809
4810         maybe_recover_session(mdsc);
4811
4812         schedule_delayed(mdsc, delay);
4813 }
4814
4815 int ceph_mdsc_init(struct ceph_fs_client *fsc)
4816
4817 {
4818         struct ceph_mds_client *mdsc;
4819         int err;
4820
4821         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
4822         if (!mdsc)
4823                 return -ENOMEM;
4824         mdsc->fsc = fsc;
4825         mutex_init(&mdsc->mutex);
4826         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
4827         if (!mdsc->mdsmap) {
4828                 err = -ENOMEM;
4829                 goto err_mdsc;
4830         }
4831
4832         init_completion(&mdsc->safe_umount_waiters);
4833         spin_lock_init(&mdsc->stopping_lock);
4834         atomic_set(&mdsc->stopping_blockers, 0);
4835         init_completion(&mdsc->stopping_waiter);
4836         init_waitqueue_head(&mdsc->session_close_wq);
4837         INIT_LIST_HEAD(&mdsc->waiting_for_map);
4838         mdsc->quotarealms_inodes = RB_ROOT;
4839         mutex_init(&mdsc->quotarealms_inodes_mutex);
4840         init_rwsem(&mdsc->snap_rwsem);
4841         mdsc->snap_realms = RB_ROOT;
4842         INIT_LIST_HEAD(&mdsc->snap_empty);
4843         spin_lock_init(&mdsc->snap_empty_lock);
4844         mdsc->request_tree = RB_ROOT;
4845         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
4846         mdsc->last_renew_caps = jiffies;
4847         INIT_LIST_HEAD(&mdsc->cap_delay_list);
4848         INIT_LIST_HEAD(&mdsc->cap_wait_list);
4849         spin_lock_init(&mdsc->cap_delay_lock);
4850         INIT_LIST_HEAD(&mdsc->snap_flush_list);
4851         spin_lock_init(&mdsc->snap_flush_lock);
4852         mdsc->last_cap_flush_tid = 1;
4853         INIT_LIST_HEAD(&mdsc->cap_flush_list);
4854         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
4855         spin_lock_init(&mdsc->cap_dirty_lock);
4856         init_waitqueue_head(&mdsc->cap_flushing_wq);
4857         INIT_WORK(&mdsc->cap_reclaim_work, ceph_cap_reclaim_work);
4858         err = ceph_metric_init(&mdsc->metric);
4859         if (err)
4860                 goto err_mdsmap;
4861
4862         spin_lock_init(&mdsc->dentry_list_lock);
4863         INIT_LIST_HEAD(&mdsc->dentry_leases);
4864         INIT_LIST_HEAD(&mdsc->dentry_dir_leases);
4865
4866         ceph_caps_init(mdsc);
4867         ceph_adjust_caps_max_min(mdsc, fsc->mount_options);
4868
4869         spin_lock_init(&mdsc->snapid_map_lock);
4870         mdsc->snapid_map_tree = RB_ROOT;
4871         INIT_LIST_HEAD(&mdsc->snapid_map_lru);
4872
4873         init_rwsem(&mdsc->pool_perm_rwsem);
4874         mdsc->pool_perm_tree = RB_ROOT;
4875
4876         strscpy(mdsc->nodename, utsname()->nodename,
4877                 sizeof(mdsc->nodename));
4878
4879         fsc->mdsc = mdsc;
4880         return 0;
4881
4882 err_mdsmap:
4883         kfree(mdsc->mdsmap);
4884 err_mdsc:
4885         kfree(mdsc);
4886         return err;
4887 }
4888
4889 /*
4890  * Wait for safe replies on open mds requests.  If we time out, drop
4891  * all requests from the tree to avoid dangling dentry refs.
4892  */
4893 static void wait_requests(struct ceph_mds_client *mdsc)
4894 {
4895         struct ceph_options *opts = mdsc->fsc->client->options;
4896         struct ceph_mds_request *req;
4897
4898         mutex_lock(&mdsc->mutex);
4899         if (__get_oldest_req(mdsc)) {
4900                 mutex_unlock(&mdsc->mutex);
4901
4902                 dout("wait_requests waiting for requests\n");
4903                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
4904                                     ceph_timeout_jiffies(opts->mount_timeout));
4905
4906                 /* tear down remaining requests */
4907                 mutex_lock(&mdsc->mutex);
4908                 while ((req = __get_oldest_req(mdsc))) {
4909                         dout("wait_requests timed out on tid %llu\n",
4910                              req->r_tid);
4911                         list_del_init(&req->r_wait);
4912                         __unregister_request(mdsc, req);
4913                 }
4914         }
4915         mutex_unlock(&mdsc->mutex);
4916         dout("wait_requests done\n");
4917 }
4918
4919 void send_flush_mdlog(struct ceph_mds_session *s)
4920 {
4921         struct ceph_msg *msg;
4922
4923         /*
4924          * Pre-luminous MDS crashes when it sees an unknown session request
4925          */
4926         if (!CEPH_HAVE_FEATURE(s->s_con.peer_features, SERVER_LUMINOUS))
4927                 return;
4928
4929         mutex_lock(&s->s_mutex);
4930         dout("request mdlog flush to mds%d (%s)s seq %lld\n", s->s_mds,
4931              ceph_session_state_name(s->s_state), s->s_seq);
4932         msg = ceph_create_session_msg(CEPH_SESSION_REQUEST_FLUSH_MDLOG,
4933                                       s->s_seq);
4934         if (!msg) {
4935                 pr_err("failed to request mdlog flush to mds%d (%s) seq %lld\n",
4936                        s->s_mds, ceph_session_state_name(s->s_state), s->s_seq);
4937         } else {
4938                 ceph_con_send(&s->s_con, msg);
4939         }
4940         mutex_unlock(&s->s_mutex);
4941 }
4942
4943 /*
4944  * called before mount is ro, and before dentries are torn down.
4945  * (hmm, does this still race with new lookups?)
4946  */
4947 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
4948 {
4949         dout("pre_umount\n");
4950         mdsc->stopping = CEPH_MDSC_STOPPING_BEGIN;
4951
4952         ceph_mdsc_iterate_sessions(mdsc, send_flush_mdlog, true);
4953         ceph_mdsc_iterate_sessions(mdsc, lock_unlock_session, false);
4954         ceph_flush_dirty_caps(mdsc);
4955         wait_requests(mdsc);
4956
4957         /*
4958          * wait for reply handlers to drop their request refs and
4959          * their inode/dcache refs
4960          */
4961         ceph_msgr_flush();
4962
4963         ceph_cleanup_quotarealms_inodes(mdsc);
4964 }
4965
4966 /*
4967  * flush the mdlog and wait for all write mds requests to flush.
4968  */
4969 static void flush_mdlog_and_wait_mdsc_unsafe_requests(struct ceph_mds_client *mdsc,
4970                                                  u64 want_tid)
4971 {
4972         struct ceph_mds_request *req = NULL, *nextreq;
4973         struct ceph_mds_session *last_session = NULL;
4974         struct rb_node *n;
4975
4976         mutex_lock(&mdsc->mutex);
4977         dout("%s want %lld\n", __func__, want_tid);
4978 restart:
4979         req = __get_oldest_req(mdsc);
4980         while (req && req->r_tid <= want_tid) {
4981                 /* find next request */
4982                 n = rb_next(&req->r_node);
4983                 if (n)
4984                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
4985                 else
4986                         nextreq = NULL;
4987                 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
4988                     (req->r_op & CEPH_MDS_OP_WRITE)) {
4989                         struct ceph_mds_session *s = req->r_session;
4990
4991                         if (!s) {
4992                                 req = nextreq;
4993                                 continue;
4994                         }
4995
4996                         /* write op */
4997                         ceph_mdsc_get_request(req);
4998                         if (nextreq)
4999                                 ceph_mdsc_get_request(nextreq);
5000                         s = ceph_get_mds_session(s);
5001                         mutex_unlock(&mdsc->mutex);
5002
5003                         /* send flush mdlog request to MDS */
5004                         if (last_session != s) {
5005                                 send_flush_mdlog(s);
5006                                 ceph_put_mds_session(last_session);
5007                                 last_session = s;
5008                         } else {
5009                                 ceph_put_mds_session(s);
5010                         }
5011                         dout("%s wait on %llu (want %llu)\n", __func__,
5012                              req->r_tid, want_tid);
5013                         wait_for_completion(&req->r_safe_completion);
5014
5015                         mutex_lock(&mdsc->mutex);
5016                         ceph_mdsc_put_request(req);
5017                         if (!nextreq)
5018                                 break;  /* next dne before, so we're done! */
5019                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
5020                                 /* next request was removed from tree */
5021                                 ceph_mdsc_put_request(nextreq);
5022                                 goto restart;
5023                         }
5024                         ceph_mdsc_put_request(nextreq);  /* won't go away */
5025                 }
5026                 req = nextreq;
5027         }
5028         mutex_unlock(&mdsc->mutex);
5029         ceph_put_mds_session(last_session);
5030         dout("%s done\n", __func__);
5031 }
5032
5033 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
5034 {
5035         u64 want_tid, want_flush;
5036
5037         if (READ_ONCE(mdsc->fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN)
5038                 return;
5039
5040         dout("sync\n");
5041         mutex_lock(&mdsc->mutex);
5042         want_tid = mdsc->last_tid;
5043         mutex_unlock(&mdsc->mutex);
5044
5045         ceph_flush_dirty_caps(mdsc);
5046         spin_lock(&mdsc->cap_dirty_lock);
5047         want_flush = mdsc->last_cap_flush_tid;
5048         if (!list_empty(&mdsc->cap_flush_list)) {
5049                 struct ceph_cap_flush *cf =
5050                         list_last_entry(&mdsc->cap_flush_list,
5051                                         struct ceph_cap_flush, g_list);
5052                 cf->wake = true;
5053         }
5054         spin_unlock(&mdsc->cap_dirty_lock);
5055
5056         dout("sync want tid %lld flush_seq %lld\n",
5057              want_tid, want_flush);
5058
5059         flush_mdlog_and_wait_mdsc_unsafe_requests(mdsc, want_tid);
5060         wait_caps_flush(mdsc, want_flush);
5061 }
5062
5063 /*
5064  * true if all sessions are closed, or we force unmount
5065  */
5066 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
5067 {
5068         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
5069                 return true;
5070         return atomic_read(&mdsc->num_sessions) <= skipped;
5071 }
5072
5073 /*
5074  * called after sb is ro or when metadata corrupted.
5075  */
5076 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
5077 {
5078         struct ceph_options *opts = mdsc->fsc->client->options;
5079         struct ceph_mds_session *session;
5080         int i;
5081         int skipped = 0;
5082
5083         dout("close_sessions\n");
5084
5085         /* close sessions */
5086         mutex_lock(&mdsc->mutex);
5087         for (i = 0; i < mdsc->max_sessions; i++) {
5088                 session = __ceph_lookup_mds_session(mdsc, i);
5089                 if (!session)
5090                         continue;
5091                 mutex_unlock(&mdsc->mutex);
5092                 mutex_lock(&session->s_mutex);
5093                 if (__close_session(mdsc, session) <= 0)
5094                         skipped++;
5095                 mutex_unlock(&session->s_mutex);
5096                 ceph_put_mds_session(session);
5097                 mutex_lock(&mdsc->mutex);
5098         }
5099         mutex_unlock(&mdsc->mutex);
5100
5101         dout("waiting for sessions to close\n");
5102         wait_event_timeout(mdsc->session_close_wq,
5103                            done_closing_sessions(mdsc, skipped),
5104                            ceph_timeout_jiffies(opts->mount_timeout));
5105
5106         /* tear down remaining sessions */
5107         mutex_lock(&mdsc->mutex);
5108         for (i = 0; i < mdsc->max_sessions; i++) {
5109                 if (mdsc->sessions[i]) {
5110                         session = ceph_get_mds_session(mdsc->sessions[i]);
5111                         __unregister_session(mdsc, session);
5112                         mutex_unlock(&mdsc->mutex);
5113                         mutex_lock(&session->s_mutex);
5114                         remove_session_caps(session);
5115                         mutex_unlock(&session->s_mutex);
5116                         ceph_put_mds_session(session);
5117                         mutex_lock(&mdsc->mutex);
5118                 }
5119         }
5120         WARN_ON(!list_empty(&mdsc->cap_delay_list));
5121         mutex_unlock(&mdsc->mutex);
5122
5123         ceph_cleanup_snapid_map(mdsc);
5124         ceph_cleanup_global_and_empty_realms(mdsc);
5125
5126         cancel_work_sync(&mdsc->cap_reclaim_work);
5127         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
5128
5129         dout("stopped\n");
5130 }
5131
5132 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
5133 {
5134         struct ceph_mds_session *session;
5135         int mds;
5136
5137         dout("force umount\n");
5138
5139         mutex_lock(&mdsc->mutex);
5140         for (mds = 0; mds < mdsc->max_sessions; mds++) {
5141                 session = __ceph_lookup_mds_session(mdsc, mds);
5142                 if (!session)
5143                         continue;
5144
5145                 if (session->s_state == CEPH_MDS_SESSION_REJECTED)
5146                         __unregister_session(mdsc, session);
5147                 __wake_requests(mdsc, &session->s_waiting);
5148                 mutex_unlock(&mdsc->mutex);
5149
5150                 mutex_lock(&session->s_mutex);
5151                 __close_session(mdsc, session);
5152                 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
5153                         cleanup_session_requests(mdsc, session);
5154                         remove_session_caps(session);
5155                 }
5156                 mutex_unlock(&session->s_mutex);
5157                 ceph_put_mds_session(session);
5158
5159                 mutex_lock(&mdsc->mutex);
5160                 kick_requests(mdsc, mds);
5161         }
5162         __wake_requests(mdsc, &mdsc->waiting_for_map);
5163         mutex_unlock(&mdsc->mutex);
5164 }
5165
5166 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
5167 {
5168         dout("stop\n");
5169         /*
5170          * Make sure the delayed work stopped before releasing
5171          * the resources.
5172          *
5173          * Because the cancel_delayed_work_sync() will only
5174          * guarantee that the work finishes executing. But the
5175          * delayed work will re-arm itself again after that.
5176          */
5177         flush_delayed_work(&mdsc->delayed_work);
5178
5179         if (mdsc->mdsmap)
5180                 ceph_mdsmap_destroy(mdsc->mdsmap);
5181         kfree(mdsc->sessions);
5182         ceph_caps_finalize(mdsc);
5183         ceph_pool_perm_destroy(mdsc);
5184 }
5185
5186 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
5187 {
5188         struct ceph_mds_client *mdsc = fsc->mdsc;
5189         dout("mdsc_destroy %p\n", mdsc);
5190
5191         if (!mdsc)
5192                 return;
5193
5194         /* flush out any connection work with references to us */
5195         ceph_msgr_flush();
5196
5197         ceph_mdsc_stop(mdsc);
5198
5199         ceph_metric_destroy(&mdsc->metric);
5200
5201         fsc->mdsc = NULL;
5202         kfree(mdsc);
5203         dout("mdsc_destroy %p done\n", mdsc);
5204 }
5205
5206 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
5207 {
5208         struct ceph_fs_client *fsc = mdsc->fsc;
5209         const char *mds_namespace = fsc->mount_options->mds_namespace;
5210         void *p = msg->front.iov_base;
5211         void *end = p + msg->front.iov_len;
5212         u32 epoch;
5213         u32 num_fs;
5214         u32 mount_fscid = (u32)-1;
5215         int err = -EINVAL;
5216
5217         ceph_decode_need(&p, end, sizeof(u32), bad);
5218         epoch = ceph_decode_32(&p);
5219
5220         dout("handle_fsmap epoch %u\n", epoch);
5221
5222         /* struct_v, struct_cv, map_len, epoch, legacy_client_fscid */
5223         ceph_decode_skip_n(&p, end, 2 + sizeof(u32) * 3, bad);
5224
5225         ceph_decode_32_safe(&p, end, num_fs, bad);
5226         while (num_fs-- > 0) {
5227                 void *info_p, *info_end;
5228                 u32 info_len;
5229                 u32 fscid, namelen;
5230
5231                 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
5232                 p += 2;         // info_v, info_cv
5233                 info_len = ceph_decode_32(&p);
5234                 ceph_decode_need(&p, end, info_len, bad);
5235                 info_p = p;
5236                 info_end = p + info_len;
5237                 p = info_end;
5238
5239                 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
5240                 fscid = ceph_decode_32(&info_p);
5241                 namelen = ceph_decode_32(&info_p);
5242                 ceph_decode_need(&info_p, info_end, namelen, bad);
5243
5244                 if (mds_namespace &&
5245                     strlen(mds_namespace) == namelen &&
5246                     !strncmp(mds_namespace, (char *)info_p, namelen)) {
5247                         mount_fscid = fscid;
5248                         break;
5249                 }
5250         }
5251
5252         ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
5253         if (mount_fscid != (u32)-1) {
5254                 fsc->client->monc.fs_cluster_id = mount_fscid;
5255                 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
5256                                    0, true);
5257                 ceph_monc_renew_subs(&fsc->client->monc);
5258         } else {
5259                 err = -ENOENT;
5260                 goto err_out;
5261         }
5262         return;
5263
5264 bad:
5265         pr_err("error decoding fsmap %d. Shutting down mount.\n", err);
5266         ceph_umount_begin(mdsc->fsc->sb);
5267         ceph_msg_dump(msg);
5268 err_out:
5269         mutex_lock(&mdsc->mutex);
5270         mdsc->mdsmap_err = err;
5271         __wake_requests(mdsc, &mdsc->waiting_for_map);
5272         mutex_unlock(&mdsc->mutex);
5273 }
5274
5275 /*
5276  * handle mds map update.
5277  */
5278 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
5279 {
5280         u32 epoch;
5281         u32 maplen;
5282         void *p = msg->front.iov_base;
5283         void *end = p + msg->front.iov_len;
5284         struct ceph_mdsmap *newmap, *oldmap;
5285         struct ceph_fsid fsid;
5286         int err = -EINVAL;
5287
5288         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
5289         ceph_decode_copy(&p, &fsid, sizeof(fsid));
5290         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
5291                 return;
5292         epoch = ceph_decode_32(&p);
5293         maplen = ceph_decode_32(&p);
5294         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
5295
5296         /* do we need it? */
5297         mutex_lock(&mdsc->mutex);
5298         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
5299                 dout("handle_map epoch %u <= our %u\n",
5300                      epoch, mdsc->mdsmap->m_epoch);
5301                 mutex_unlock(&mdsc->mutex);
5302                 return;
5303         }
5304
5305         newmap = ceph_mdsmap_decode(&p, end, ceph_msgr2(mdsc->fsc->client));
5306         if (IS_ERR(newmap)) {
5307                 err = PTR_ERR(newmap);
5308                 goto bad_unlock;
5309         }
5310
5311         /* swap into place */
5312         if (mdsc->mdsmap) {
5313                 oldmap = mdsc->mdsmap;
5314                 mdsc->mdsmap = newmap;
5315                 check_new_map(mdsc, newmap, oldmap);
5316                 ceph_mdsmap_destroy(oldmap);
5317         } else {
5318                 mdsc->mdsmap = newmap;  /* first mds map */
5319         }
5320         mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
5321                                         MAX_LFS_FILESIZE);
5322
5323         __wake_requests(mdsc, &mdsc->waiting_for_map);
5324         ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
5325                           mdsc->mdsmap->m_epoch);
5326
5327         mutex_unlock(&mdsc->mutex);
5328         schedule_delayed(mdsc, 0);
5329         return;
5330
5331 bad_unlock:
5332         mutex_unlock(&mdsc->mutex);
5333 bad:
5334         pr_err("error decoding mdsmap %d. Shutting down mount.\n", err);
5335         ceph_umount_begin(mdsc->fsc->sb);
5336         ceph_msg_dump(msg);
5337         return;
5338 }
5339
5340 static struct ceph_connection *mds_get_con(struct ceph_connection *con)
5341 {
5342         struct ceph_mds_session *s = con->private;
5343
5344         if (ceph_get_mds_session(s))
5345                 return con;
5346         return NULL;
5347 }
5348
5349 static void mds_put_con(struct ceph_connection *con)
5350 {
5351         struct ceph_mds_session *s = con->private;
5352
5353         ceph_put_mds_session(s);
5354 }
5355
5356 /*
5357  * if the client is unresponsive for long enough, the mds will kill
5358  * the session entirely.
5359  */
5360 static void mds_peer_reset(struct ceph_connection *con)
5361 {
5362         struct ceph_mds_session *s = con->private;
5363         struct ceph_mds_client *mdsc = s->s_mdsc;
5364
5365         pr_warn("mds%d closed our session\n", s->s_mds);
5366         if (READ_ONCE(mdsc->fsc->mount_state) != CEPH_MOUNT_FENCE_IO)
5367                 send_mds_reconnect(mdsc, s);
5368 }
5369
5370 static void mds_dispatch(struct ceph_connection *con, struct ceph_msg *msg)
5371 {
5372         struct ceph_mds_session *s = con->private;
5373         struct ceph_mds_client *mdsc = s->s_mdsc;
5374         int type = le16_to_cpu(msg->hdr.type);
5375
5376         mutex_lock(&mdsc->mutex);
5377         if (__verify_registered_session(mdsc, s) < 0) {
5378                 mutex_unlock(&mdsc->mutex);
5379                 goto out;
5380         }
5381         mutex_unlock(&mdsc->mutex);
5382
5383         switch (type) {
5384         case CEPH_MSG_MDS_MAP:
5385                 ceph_mdsc_handle_mdsmap(mdsc, msg);
5386                 break;
5387         case CEPH_MSG_FS_MAP_USER:
5388                 ceph_mdsc_handle_fsmap(mdsc, msg);
5389                 break;
5390         case CEPH_MSG_CLIENT_SESSION:
5391                 handle_session(s, msg);
5392                 break;
5393         case CEPH_MSG_CLIENT_REPLY:
5394                 handle_reply(s, msg);
5395                 break;
5396         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
5397                 handle_forward(mdsc, s, msg);
5398                 break;
5399         case CEPH_MSG_CLIENT_CAPS:
5400                 ceph_handle_caps(s, msg);
5401                 break;
5402         case CEPH_MSG_CLIENT_SNAP:
5403                 ceph_handle_snap(mdsc, s, msg);
5404                 break;
5405         case CEPH_MSG_CLIENT_LEASE:
5406                 handle_lease(mdsc, s, msg);
5407                 break;
5408         case CEPH_MSG_CLIENT_QUOTA:
5409                 ceph_handle_quota(mdsc, s, msg);
5410                 break;
5411
5412         default:
5413                 pr_err("received unknown message type %d %s\n", type,
5414                        ceph_msg_type_name(type));
5415         }
5416 out:
5417         ceph_msg_put(msg);
5418 }
5419
5420 /*
5421  * authentication
5422  */
5423
5424 /*
5425  * Note: returned pointer is the address of a structure that's
5426  * managed separately.  Caller must *not* attempt to free it.
5427  */
5428 static struct ceph_auth_handshake *
5429 mds_get_authorizer(struct ceph_connection *con, int *proto, int force_new)
5430 {
5431         struct ceph_mds_session *s = con->private;
5432         struct ceph_mds_client *mdsc = s->s_mdsc;
5433         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5434         struct ceph_auth_handshake *auth = &s->s_auth;
5435         int ret;
5436
5437         ret = __ceph_auth_get_authorizer(ac, auth, CEPH_ENTITY_TYPE_MDS,
5438                                          force_new, proto, NULL, NULL);
5439         if (ret)
5440                 return ERR_PTR(ret);
5441
5442         return auth;
5443 }
5444
5445 static int mds_add_authorizer_challenge(struct ceph_connection *con,
5446                                     void *challenge_buf, int challenge_buf_len)
5447 {
5448         struct ceph_mds_session *s = con->private;
5449         struct ceph_mds_client *mdsc = s->s_mdsc;
5450         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5451
5452         return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
5453                                             challenge_buf, challenge_buf_len);
5454 }
5455
5456 static int mds_verify_authorizer_reply(struct ceph_connection *con)
5457 {
5458         struct ceph_mds_session *s = con->private;
5459         struct ceph_mds_client *mdsc = s->s_mdsc;
5460         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5461         struct ceph_auth_handshake *auth = &s->s_auth;
5462
5463         return ceph_auth_verify_authorizer_reply(ac, auth->authorizer,
5464                 auth->authorizer_reply_buf, auth->authorizer_reply_buf_len,
5465                 NULL, NULL, NULL, NULL);
5466 }
5467
5468 static int mds_invalidate_authorizer(struct ceph_connection *con)
5469 {
5470         struct ceph_mds_session *s = con->private;
5471         struct ceph_mds_client *mdsc = s->s_mdsc;
5472         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5473
5474         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
5475
5476         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
5477 }
5478
5479 static int mds_get_auth_request(struct ceph_connection *con,
5480                                 void *buf, int *buf_len,
5481                                 void **authorizer, int *authorizer_len)
5482 {
5483         struct ceph_mds_session *s = con->private;
5484         struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
5485         struct ceph_auth_handshake *auth = &s->s_auth;
5486         int ret;
5487
5488         ret = ceph_auth_get_authorizer(ac, auth, CEPH_ENTITY_TYPE_MDS,
5489                                        buf, buf_len);
5490         if (ret)
5491                 return ret;
5492
5493         *authorizer = auth->authorizer_buf;
5494         *authorizer_len = auth->authorizer_buf_len;
5495         return 0;
5496 }
5497
5498 static int mds_handle_auth_reply_more(struct ceph_connection *con,
5499                                       void *reply, int reply_len,
5500                                       void *buf, int *buf_len,
5501                                       void **authorizer, int *authorizer_len)
5502 {
5503         struct ceph_mds_session *s = con->private;
5504         struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
5505         struct ceph_auth_handshake *auth = &s->s_auth;
5506         int ret;
5507
5508         ret = ceph_auth_handle_svc_reply_more(ac, auth, reply, reply_len,
5509                                               buf, buf_len);
5510         if (ret)
5511                 return ret;
5512
5513         *authorizer = auth->authorizer_buf;
5514         *authorizer_len = auth->authorizer_buf_len;
5515         return 0;
5516 }
5517
5518 static int mds_handle_auth_done(struct ceph_connection *con,
5519                                 u64 global_id, void *reply, int reply_len,
5520                                 u8 *session_key, int *session_key_len,
5521                                 u8 *con_secret, int *con_secret_len)
5522 {
5523         struct ceph_mds_session *s = con->private;
5524         struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
5525         struct ceph_auth_handshake *auth = &s->s_auth;
5526
5527         return ceph_auth_handle_svc_reply_done(ac, auth, reply, reply_len,
5528                                                session_key, session_key_len,
5529                                                con_secret, con_secret_len);
5530 }
5531
5532 static int mds_handle_auth_bad_method(struct ceph_connection *con,
5533                                       int used_proto, int result,
5534                                       const int *allowed_protos, int proto_cnt,
5535                                       const int *allowed_modes, int mode_cnt)
5536 {
5537         struct ceph_mds_session *s = con->private;
5538         struct ceph_mon_client *monc = &s->s_mdsc->fsc->client->monc;
5539         int ret;
5540
5541         if (ceph_auth_handle_bad_authorizer(monc->auth, CEPH_ENTITY_TYPE_MDS,
5542                                             used_proto, result,
5543                                             allowed_protos, proto_cnt,
5544                                             allowed_modes, mode_cnt)) {
5545                 ret = ceph_monc_validate_auth(monc);
5546                 if (ret)
5547                         return ret;
5548         }
5549
5550         return -EACCES;
5551 }
5552
5553 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
5554                                 struct ceph_msg_header *hdr, int *skip)
5555 {
5556         struct ceph_msg *msg;
5557         int type = (int) le16_to_cpu(hdr->type);
5558         int front_len = (int) le32_to_cpu(hdr->front_len);
5559
5560         if (con->in_msg)
5561                 return con->in_msg;
5562
5563         *skip = 0;
5564         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
5565         if (!msg) {
5566                 pr_err("unable to allocate msg type %d len %d\n",
5567                        type, front_len);
5568                 return NULL;
5569         }
5570
5571         return msg;
5572 }
5573
5574 static int mds_sign_message(struct ceph_msg *msg)
5575 {
5576        struct ceph_mds_session *s = msg->con->private;
5577        struct ceph_auth_handshake *auth = &s->s_auth;
5578
5579        return ceph_auth_sign_message(auth, msg);
5580 }
5581
5582 static int mds_check_message_signature(struct ceph_msg *msg)
5583 {
5584        struct ceph_mds_session *s = msg->con->private;
5585        struct ceph_auth_handshake *auth = &s->s_auth;
5586
5587        return ceph_auth_check_message_signature(auth, msg);
5588 }
5589
5590 static const struct ceph_connection_operations mds_con_ops = {
5591         .get = mds_get_con,
5592         .put = mds_put_con,
5593         .alloc_msg = mds_alloc_msg,
5594         .dispatch = mds_dispatch,
5595         .peer_reset = mds_peer_reset,
5596         .get_authorizer = mds_get_authorizer,
5597         .add_authorizer_challenge = mds_add_authorizer_challenge,
5598         .verify_authorizer_reply = mds_verify_authorizer_reply,
5599         .invalidate_authorizer = mds_invalidate_authorizer,
5600         .sign_message = mds_sign_message,
5601         .check_message_signature = mds_check_message_signature,
5602         .get_auth_request = mds_get_auth_request,
5603         .handle_auth_reply_more = mds_handle_auth_reply_more,
5604         .handle_auth_done = mds_handle_auth_done,
5605         .handle_auth_bad_method = mds_handle_auth_bad_method,
5606 };
5607
5608 /* eof */