fs/ceph/mds_client.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/ceph/ceph_debug.h>
   3
   4 #include <linux/fs.h>
   5 #include <linux/wait.h>
   6 #include <linux/slab.h>
   7 #include <linux/gfp.h>
   8 #include <linux/sched.h>
   9 #include <linux/debugfs.h>
  10 #include <linux/seq_file.h>
  11 #include <linux/ratelimit.h>
  12 #include <linux/bits.h>
  13 #include <linux/ktime.h>
  14
  15 #include "super.h"
  16 #include "mds_client.h"
  17
  18 #include <linux/ceph/ceph_features.h>
  19 #include <linux/ceph/messenger.h>
  20 #include <linux/ceph/decode.h>
  21 #include <linux/ceph/pagelist.h>
  22 #include <linux/ceph/auth.h>
  23 #include <linux/ceph/debugfs.h>
  24
  25 #define RECONNECT_MAX_SIZE (INT_MAX - PAGE_SIZE)
  26
  27 /*
  28  * A cluster of MDS (metadata server) daemons is responsible for
  29  * managing the file system namespace (the directory hierarchy and
  30  * inodes) and for coordinating shared access to storage.  Metadata is
  31  * partitioning hierarchically across a number of servers, and that
  32  * partition varies over time as the cluster adjusts the distribution
  33  * in order to balance load.
  34  *
  35  * The MDS client is primarily responsible to managing synchronous
  36  * metadata requests for operations like open, unlink, and so forth.
  37  * If there is a MDS failure, we find out about it when we (possibly
  38  * request and) receive a new MDS map, and can resubmit affected
  39  * requests.
  40  *
  41  * For the most part, though, we take advantage of a lossless
  42  * communications channel to the MDS, and do not need to worry about
  43  * timing out or resubmitting requests.
  44  *
  45  * We maintain a stateful "session" with each MDS we interact with.
  46  * Within each session, we sent periodic heartbeat messages to ensure
  47  * any capabilities or leases we have been issues remain valid.  If
  48  * the session times out and goes stale, our leases and capabilities
  49  * are no longer valid.
  50  */
  51
  52 struct ceph_reconnect_state {
  53         struct ceph_mds_session *session;
  54         int nr_caps, nr_realms;
  55         struct ceph_pagelist *pagelist;
  56         unsigned msg_version;
  57         bool allow_multi;
  58 };
  59
  60 static void __wake_requests(struct ceph_mds_client *mdsc,
  61                             struct list_head *head);
  62 static void ceph_cap_release_work(struct work_struct *work);
  63 static void ceph_cap_reclaim_work(struct work_struct *work);
  64
  65 static const struct ceph_connection_operations mds_con_ops;
  66
  67
  68 /*
  69  * mds reply parsing
  70  */
  71
  72 static int parse_reply_info_quota(void **p, void *end,
  73                                   struct ceph_mds_reply_info_in *info)
  74 {
  75         u8 struct_v, struct_compat;
  76         u32 struct_len;
  77
  78         ceph_decode_8_safe(p, end, struct_v, bad);
  79         ceph_decode_8_safe(p, end, struct_compat, bad);
  80         /* struct_v is expected to be >= 1. we only
  81          * understand encoding with struct_compat == 1. */
  82         if (!struct_v || struct_compat != 1)
  83                 goto bad;
  84         ceph_decode_32_safe(p, end, struct_len, bad);
  85         ceph_decode_need(p, end, struct_len, bad);
  86         end = *p + struct_len;
  87         ceph_decode_64_safe(p, end, info->max_bytes, bad);
  88         ceph_decode_64_safe(p, end, info->max_files, bad);
  89         *p = end;
  90         return 0;
  91 bad:
  92         return -EIO;
  93 }
  94
  95 /*
  96  * parse individual inode info
  97  */
  98 static int parse_reply_info_in(void **p, void *end,
  99                                struct ceph_mds_reply_info_in *info,
 100                                u64 features)
 101 {
 102         int err = 0;
 103         u8 struct_v = 0;
 104
 105         if (features == (u64)-1) {
 106                 u32 struct_len;
 107                 u8 struct_compat;
 108                 ceph_decode_8_safe(p, end, struct_v, bad);
 109                 ceph_decode_8_safe(p, end, struct_compat, bad);
 110                 /* struct_v is expected to be >= 1. we only understand
 111                  * encoding with struct_compat == 1. */
 112                 if (!struct_v || struct_compat != 1)
 113                         goto bad;
 114                 ceph_decode_32_safe(p, end, struct_len, bad);
 115                 ceph_decode_need(p, end, struct_len, bad);
 116                 end = *p + struct_len;
 117         }
 118
 119         ceph_decode_need(p, end, sizeof(struct ceph_mds_reply_inode), bad);
 120         info->in = *p;
 121         *p += sizeof(struct ceph_mds_reply_inode) +
 122                 sizeof(*info->in->fragtree.splits) *
 123                 le32_to_cpu(info->in->fragtree.nsplits);
 124
 125         ceph_decode_32_safe(p, end, info->symlink_len, bad);
 126         ceph_decode_need(p, end, info->symlink_len, bad);
 127         info->symlink = *p;
 128         *p += info->symlink_len;
 129
 130         ceph_decode_copy_safe(p, end, &info->dir_layout,
 131                               sizeof(info->dir_layout), bad);
 132         ceph_decode_32_safe(p, end, info->xattr_len, bad);
 133         ceph_decode_need(p, end, info->xattr_len, bad);
 134         info->xattr_data = *p;
 135         *p += info->xattr_len;
 136
 137         if (features == (u64)-1) {
 138                 /* inline data */
 139                 ceph_decode_64_safe(p, end, info->inline_version, bad);
 140                 ceph_decode_32_safe(p, end, info->inline_len, bad);
 141                 ceph_decode_need(p, end, info->inline_len, bad);
 142                 info->inline_data = *p;
 143                 *p += info->inline_len;
 144                 /* quota */
 145                 err = parse_reply_info_quota(p, end, info);
 146                 if (err < 0)
 147                         goto out_bad;
 148                 /* pool namespace */
 149                 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
 150                 if (info->pool_ns_len > 0) {
 151                         ceph_decode_need(p, end, info->pool_ns_len, bad);
 152                         info->pool_ns_data = *p;
 153                         *p += info->pool_ns_len;
 154                 }
 155
 156                 /* btime */
 157                 ceph_decode_need(p, end, sizeof(info->btime), bad);
 158                 ceph_decode_copy(p, &info->btime, sizeof(info->btime));
 159
 160                 /* change attribute */
 161                 ceph_decode_64_safe(p, end, info->change_attr, bad);
 162
 163                 /* dir pin */
 164                 if (struct_v >= 2) {
 165                         ceph_decode_32_safe(p, end, info->dir_pin, bad);
 166                 } else {
 167                         info->dir_pin = -ENODATA;
 168                 }
 169
 170                 /* snapshot birth time, remains zero for v<=2 */
 171                 if (struct_v >= 3) {
 172                         ceph_decode_need(p, end, sizeof(info->snap_btime), bad);
 173                         ceph_decode_copy(p, &info->snap_btime,
 174                                          sizeof(info->snap_btime));
 175                 } else {
 176                         memset(&info->snap_btime, 0, sizeof(info->snap_btime));
 177                 }
 178
 179                 /* snapshot count, remains zero for v<=3 */
 180                 if (struct_v >= 4) {
 181                         ceph_decode_64_safe(p, end, info->rsnaps, bad);
 182                 } else {
 183                         info->rsnaps = 0;
 184                 }
 185
 186                 *p = end;
 187         } else {
 188                 if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
 189                         ceph_decode_64_safe(p, end, info->inline_version, bad);
 190                         ceph_decode_32_safe(p, end, info->inline_len, bad);
 191                         ceph_decode_need(p, end, info->inline_len, bad);
 192                         info->inline_data = *p;
 193                         *p += info->inline_len;
 194                 } else
 195                         info->inline_version = CEPH_INLINE_NONE;
 196
 197                 if (features & CEPH_FEATURE_MDS_QUOTA) {
 198                         err = parse_reply_info_quota(p, end, info);
 199                         if (err < 0)
 200                                 goto out_bad;
 201                 } else {
 202                         info->max_bytes = 0;
 203                         info->max_files = 0;
 204                 }
 205
 206                 info->pool_ns_len = 0;
 207                 info->pool_ns_data = NULL;
 208                 if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
 209                         ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
 210                         if (info->pool_ns_len > 0) {
 211                                 ceph_decode_need(p, end, info->pool_ns_len, bad);
 212                                 info->pool_ns_data = *p;
 213                                 *p += info->pool_ns_len;
 214                         }
 215                 }
 216
 217                 if (features & CEPH_FEATURE_FS_BTIME) {
 218                         ceph_decode_need(p, end, sizeof(info->btime), bad);
 219                         ceph_decode_copy(p, &info->btime, sizeof(info->btime));
 220                         ceph_decode_64_safe(p, end, info->change_attr, bad);
 221                 }
 222
 223                 info->dir_pin = -ENODATA;
 224                 /* info->snap_btime and info->rsnaps remain zero */
 225         }
 226         return 0;
 227 bad:
 228         err = -EIO;
 229 out_bad:
 230         return err;
 231 }
 232
 233 static int parse_reply_info_dir(void **p, void *end,
 234                                 struct ceph_mds_reply_dirfrag **dirfrag,
 235                                 u64 features)
 236 {
 237         if (features == (u64)-1) {
 238                 u8 struct_v, struct_compat;
 239                 u32 struct_len;
 240                 ceph_decode_8_safe(p, end, struct_v, bad);
 241                 ceph_decode_8_safe(p, end, struct_compat, bad);
 242                 /* struct_v is expected to be >= 1. we only understand
 243                  * encoding whose struct_compat == 1. */
 244                 if (!struct_v || struct_compat != 1)
 245                         goto bad;
 246                 ceph_decode_32_safe(p, end, struct_len, bad);
 247                 ceph_decode_need(p, end, struct_len, bad);
 248                 end = *p + struct_len;
 249         }
 250
 251         ceph_decode_need(p, end, sizeof(**dirfrag), bad);
 252         *dirfrag = *p;
 253         *p += sizeof(**dirfrag) + sizeof(u32) * le32_to_cpu((*dirfrag)->ndist);
 254         if (unlikely(*p > end))
 255                 goto bad;
 256         if (features == (u64)-1)
 257                 *p = end;
 258         return 0;
 259 bad:
 260         return -EIO;
 261 }
 262
 263 static int parse_reply_info_lease(void **p, void *end,
 264                                   struct ceph_mds_reply_lease **lease,
 265                                   u64 features)
 266 {
 267         if (features == (u64)-1) {
 268                 u8 struct_v, struct_compat;
 269                 u32 struct_len;
 270                 ceph_decode_8_safe(p, end, struct_v, bad);
 271                 ceph_decode_8_safe(p, end, struct_compat, bad);
 272                 /* struct_v is expected to be >= 1. we only understand
 273                  * encoding whose struct_compat == 1. */
 274                 if (!struct_v || struct_compat != 1)
 275                         goto bad;
 276                 ceph_decode_32_safe(p, end, struct_len, bad);
 277                 ceph_decode_need(p, end, struct_len, bad);
 278                 end = *p + struct_len;
 279         }
 280
 281         ceph_decode_need(p, end, sizeof(**lease), bad);
 282         *lease = *p;
 283         *p += sizeof(**lease);
 284         if (features == (u64)-1)
 285                 *p = end;
 286         return 0;
 287 bad:
 288         return -EIO;
 289 }
 290
 291 /*
 292  * parse a normal reply, which may contain a (dir+)dentry and/or a
 293  * target inode.
 294  */
 295 static int parse_reply_info_trace(void **p, void *end,
 296                                   struct ceph_mds_reply_info_parsed *info,
 297                                   u64 features)
 298 {
 299         int err;
 300
 301         if (info->head->is_dentry) {
 302                 err = parse_reply_info_in(p, end, &info->diri, features);
 303                 if (err < 0)
 304                         goto out_bad;
 305
 306                 err = parse_reply_info_dir(p, end, &info->dirfrag, features);
 307                 if (err < 0)
 308                         goto out_bad;
 309
 310                 ceph_decode_32_safe(p, end, info->dname_len, bad);
 311                 ceph_decode_need(p, end, info->dname_len, bad);
 312                 info->dname = *p;
 313                 *p += info->dname_len;
 314
 315                 err = parse_reply_info_lease(p, end, &info->dlease, features);
 316                 if (err < 0)
 317                         goto out_bad;
 318         }
 319
 320         if (info->head->is_target) {
 321                 err = parse_reply_info_in(p, end, &info->targeti, features);
 322                 if (err < 0)
 323                         goto out_bad;
 324         }
 325
 326         if (unlikely(*p != end))
 327                 goto bad;
 328         return 0;
 329
 330 bad:
 331         err = -EIO;
 332 out_bad:
 333         pr_err("problem parsing mds trace %d\n", err);
 334         return err;
 335 }
 336
 337 /*
 338  * parse readdir results
 339  */
 340 static int parse_reply_info_readdir(void **p, void *end,
 341                                 struct ceph_mds_reply_info_parsed *info,
 342                                 u64 features)
 343 {
 344         u32 num, i = 0;
 345         int err;
 346
 347         err = parse_reply_info_dir(p, end, &info->dir_dir, features);
 348         if (err < 0)
 349                 goto out_bad;
 350
 351         ceph_decode_need(p, end, sizeof(num) + 2, bad);
 352         num = ceph_decode_32(p);
 353         {
 354                 u16 flags = ceph_decode_16(p);
 355                 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
 356                 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
 357                 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
 358                 info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
 359         }
 360         if (num == 0)
 361                 goto done;
 362
 363         BUG_ON(!info->dir_entries);
 364         if ((unsigned long)(info->dir_entries + num) >
 365             (unsigned long)info->dir_entries + info->dir_buf_size) {
 366                 pr_err("dir contents are larger than expected\n");
 367                 WARN_ON(1);
 368                 goto bad;
 369         }
 370
 371         info->dir_nr = num;
 372         while (num) {
 373                 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
 374                 /* dentry */
 375                 ceph_decode_32_safe(p, end, rde->name_len, bad);
 376                 ceph_decode_need(p, end, rde->name_len, bad);
 377                 rde->name = *p;
 378                 *p += rde->name_len;
 379                 dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
 380
 381                 /* dentry lease */
 382                 err = parse_reply_info_lease(p, end, &rde->lease, features);
 383                 if (err)
 384                         goto out_bad;
 385                 /* inode */
 386                 err = parse_reply_info_in(p, end, &rde->inode, features);
 387                 if (err < 0)
 388                         goto out_bad;
 389                 /* ceph_readdir_prepopulate() will update it */
 390                 rde->offset = 0;
 391                 i++;
 392                 num--;
 393         }
 394
 395 done:
 396         /* Skip over any unrecognized fields */
 397         *p = end;
 398         return 0;
 399
 400 bad:
 401         err = -EIO;
 402 out_bad:
 403         pr_err("problem parsing dir contents %d\n", err);
 404         return err;
 405 }
 406
 407 /*
 408  * parse fcntl F_GETLK results
 409  */
 410 static int parse_reply_info_filelock(void **p, void *end,
 411                                      struct ceph_mds_reply_info_parsed *info,
 412                                      u64 features)
 413 {
 414         if (*p + sizeof(*info->filelock_reply) > end)
 415                 goto bad;
 416
 417         info->filelock_reply = *p;
 418
 419         /* Skip over any unrecognized fields */
 420         *p = end;
 421         return 0;
 422 bad:
 423         return -EIO;
 424 }
 425
 426
 427 #if BITS_PER_LONG == 64
 428
 429 #define DELEGATED_INO_AVAILABLE         xa_mk_value(1)
 430
 431 static int ceph_parse_deleg_inos(void **p, void *end,
 432                                  struct ceph_mds_session *s)
 433 {
 434         u32 sets;
 435
 436         ceph_decode_32_safe(p, end, sets, bad);
 437         dout("got %u sets of delegated inodes\n", sets);
 438         while (sets--) {
 439                 u64 start, len, ino;
 440
 441                 ceph_decode_64_safe(p, end, start, bad);
 442                 ceph_decode_64_safe(p, end, len, bad);
 443
 444                 /* Don't accept a delegation of system inodes */
 445                 if (start < CEPH_INO_SYSTEM_BASE) {
 446                         pr_warn_ratelimited("ceph: ignoring reserved inode range delegation (start=0x%llx len=0x%llx)\n",
 447                                         start, len);
 448                         continue;
 449                 }
 450                 while (len--) {
 451                         int err = xa_insert(&s->s_delegated_inos, ino = start++,
 452                                             DELEGATED_INO_AVAILABLE,
 453                                             GFP_KERNEL);
 454                         if (!err) {
 455                                 dout("added delegated inode 0x%llx\n",
 456                                      start - 1);
 457                         } else if (err == -EBUSY) {
 458                                 pr_warn("ceph: MDS delegated inode 0x%llx more than once.\n",
 459                                         start - 1);
 460                         } else {
 461                                 return err;
 462                         }
 463                 }
 464         }
 465         return 0;
 466 bad:
 467         return -EIO;
 468 }
 469
 470 u64 ceph_get_deleg_ino(struct ceph_mds_session *s)
 471 {
 472         unsigned long ino;
 473         void *val;
 474
 475         xa_for_each(&s->s_delegated_inos, ino, val) {
 476                 val = xa_erase(&s->s_delegated_inos, ino);
 477                 if (val == DELEGATED_INO_AVAILABLE)
 478                         return ino;
 479         }
 480         return 0;
 481 }
 482
 483 int ceph_restore_deleg_ino(struct ceph_mds_session *s, u64 ino)
 484 {
 485         return xa_insert(&s->s_delegated_inos, ino, DELEGATED_INO_AVAILABLE,
 486                          GFP_KERNEL);
 487 }
 488 #else /* BITS_PER_LONG == 64 */
 489 /*
 490  * FIXME: xarrays can't handle 64-bit indexes on a 32-bit arch. For now, just
 491  * ignore delegated_inos on 32 bit arch. Maybe eventually add xarrays for top
 492  * and bottom words?
 493  */
 494 static int ceph_parse_deleg_inos(void **p, void *end,
 495                                  struct ceph_mds_session *s)
 496 {
 497         u32 sets;
 498
 499         ceph_decode_32_safe(p, end, sets, bad);
 500         if (sets)
 501                 ceph_decode_skip_n(p, end, sets * 2 * sizeof(__le64), bad);
 502         return 0;
 503 bad:
 504         return -EIO;
 505 }
 506
 507 u64 ceph_get_deleg_ino(struct ceph_mds_session *s)
 508 {
 509         return 0;
 510 }
 511
 512 int ceph_restore_deleg_ino(struct ceph_mds_session *s, u64 ino)
 513 {
 514         return 0;
 515 }
 516 #endif /* BITS_PER_LONG == 64 */
 517
 518 /*
 519  * parse create results
 520  */
 521 static int parse_reply_info_create(void **p, void *end,
 522                                   struct ceph_mds_reply_info_parsed *info,
 523                                   u64 features, struct ceph_mds_session *s)
 524 {
 525         int ret;
 526
 527         if (features == (u64)-1 ||
 528             (features & CEPH_FEATURE_REPLY_CREATE_INODE)) {
 529                 if (*p == end) {
 530                         /* Malformed reply? */
 531                         info->has_create_ino = false;
 532                 } else if (test_bit(CEPHFS_FEATURE_DELEG_INO, &s->s_features)) {
 533                         info->has_create_ino = true;
 534                         /* struct_v, struct_compat, and len */
 535                         ceph_decode_skip_n(p, end, 2 + sizeof(u32), bad);
 536                         ceph_decode_64_safe(p, end, info->ino, bad);
 537                         ret = ceph_parse_deleg_inos(p, end, s);
 538                         if (ret)
 539                                 return ret;
 540                 } else {
 541                         /* legacy */
 542                         ceph_decode_64_safe(p, end, info->ino, bad);
 543                         info->has_create_ino = true;
 544                 }
 545         } else {
 546                 if (*p != end)
 547                         goto bad;
 548         }
 549
 550         /* Skip over any unrecognized fields */
 551         *p = end;
 552         return 0;
 553 bad:
 554         return -EIO;
 555 }
 556
 557 /*
 558  * parse extra results
 559  */
 560 static int parse_reply_info_extra(void **p, void *end,
 561                                   struct ceph_mds_reply_info_parsed *info,
 562                                   u64 features, struct ceph_mds_session *s)
 563 {
 564         u32 op = le32_to_cpu(info->head->op);
 565
 566         if (op == CEPH_MDS_OP_GETFILELOCK)
 567                 return parse_reply_info_filelock(p, end, info, features);
 568         else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
 569                 return parse_reply_info_readdir(p, end, info, features);
 570         else if (op == CEPH_MDS_OP_CREATE)
 571                 return parse_reply_info_create(p, end, info, features, s);
 572         else
 573                 return -EIO;
 574 }
 575
 576 /*
 577  * parse entire mds reply
 578  */
 579 static int parse_reply_info(struct ceph_mds_session *s, struct ceph_msg *msg,
 580                             struct ceph_mds_reply_info_parsed *info,
 581                             u64 features)
 582 {
 583         void *p, *end;
 584         u32 len;
 585         int err;
 586
 587         info->head = msg->front.iov_base;
 588         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 589         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 590
 591         /* trace */
 592         ceph_decode_32_safe(&p, end, len, bad);
 593         if (len > 0) {
 594                 ceph_decode_need(&p, end, len, bad);
 595                 err = parse_reply_info_trace(&p, p+len, info, features);
 596                 if (err < 0)
 597                         goto out_bad;
 598         }
 599
 600         /* extra */
 601         ceph_decode_32_safe(&p, end, len, bad);
 602         if (len > 0) {
 603                 ceph_decode_need(&p, end, len, bad);
 604                 err = parse_reply_info_extra(&p, p+len, info, features, s);
 605                 if (err < 0)
 606                         goto out_bad;
 607         }
 608
 609         /* snap blob */
 610         ceph_decode_32_safe(&p, end, len, bad);
 611         info->snapblob_len = len;
 612         info->snapblob = p;
 613         p += len;
 614
 615         if (p != end)
 616                 goto bad;
 617         return 0;
 618
 619 bad:
 620         err = -EIO;
 621 out_bad:
 622         pr_err("mds parse_reply err %d\n", err);
 623         return err;
 624 }
 625
 626 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
 627 {
 628         if (!info->dir_entries)
 629                 return;
 630         free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
 631 }
 632
 633
 634 /*
 635  * sessions
 636  */
 637 const char *ceph_session_state_name(int s)
 638 {
 639         switch (s) {
 640         case CEPH_MDS_SESSION_NEW: return "new";
 641         case CEPH_MDS_SESSION_OPENING: return "opening";
 642         case CEPH_MDS_SESSION_OPEN: return "open";
 643         case CEPH_MDS_SESSION_HUNG: return "hung";
 644         case CEPH_MDS_SESSION_CLOSING: return "closing";
 645         case CEPH_MDS_SESSION_CLOSED: return "closed";
 646         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
 647         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
 648         case CEPH_MDS_SESSION_REJECTED: return "rejected";
 649         default: return "???";
 650         }
 651 }
 652
 653 struct ceph_mds_session *ceph_get_mds_session(struct ceph_mds_session *s)
 654 {
 655         if (refcount_inc_not_zero(&s->s_ref)) {
 656                 dout("mdsc get_session %p %d -> %d\n", s,
 657                      refcount_read(&s->s_ref)-1, refcount_read(&s->s_ref));
 658                 return s;
 659         } else {
 660                 dout("mdsc get_session %p 0 -- FAIL\n", s);
 661                 return NULL;
 662         }
 663 }
 664
 665 void ceph_put_mds_session(struct ceph_mds_session *s)
 666 {
 667         dout("mdsc put_session %p %d -> %d\n", s,
 668              refcount_read(&s->s_ref), refcount_read(&s->s_ref)-1);
 669         if (refcount_dec_and_test(&s->s_ref)) {
 670                 if (s->s_auth.authorizer)
 671                         ceph_auth_destroy_authorizer(s->s_auth.authorizer);
 672                 WARN_ON(mutex_is_locked(&s->s_mutex));
 673                 xa_destroy(&s->s_delegated_inos);
 674                 kfree(s);
 675         }
 676 }
 677
 678 /*
 679  * called under mdsc->mutex
 680  */
 681 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
 682                                                    int mds)
 683 {
 684         if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
 685                 return NULL;
 686         return ceph_get_mds_session(mdsc->sessions[mds]);
 687 }
 688
 689 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
 690 {
 691         if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
 692                 return false;
 693         else
 694                 return true;
 695 }
 696
 697 static int __verify_registered_session(struct ceph_mds_client *mdsc,
 698                                        struct ceph_mds_session *s)
 699 {
 700         if (s->s_mds >= mdsc->max_sessions ||
 701             mdsc->sessions[s->s_mds] != s)
 702                 return -ENOENT;
 703         return 0;
 704 }
 705
 706 /*
 707  * create+register a new session for given mds.
 708  * called under mdsc->mutex.
 709  */
 710 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 711                                                  int mds)
 712 {
 713         struct ceph_mds_session *s;
 714
 715         if (mds >= mdsc->mdsmap->possible_max_rank)
 716                 return ERR_PTR(-EINVAL);
 717
 718         s = kzalloc(sizeof(*s), GFP_NOFS);
 719         if (!s)
 720                 return ERR_PTR(-ENOMEM);
 721
 722         if (mds >= mdsc->max_sessions) {
 723                 int newmax = 1 << get_count_order(mds + 1);
 724                 struct ceph_mds_session **sa;
 725
 726                 dout("%s: realloc to %d\n", __func__, newmax);
 727                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 728                 if (!sa)
 729                         goto fail_realloc;
 730                 if (mdsc->sessions) {
 731                         memcpy(sa, mdsc->sessions,
 732                                mdsc->max_sessions * sizeof(void *));
 733                         kfree(mdsc->sessions);
 734                 }
 735                 mdsc->sessions = sa;
 736                 mdsc->max_sessions = newmax;
 737         }
 738
 739         dout("%s: mds%d\n", __func__, mds);
 740         s->s_mdsc = mdsc;
 741         s->s_mds = mds;
 742         s->s_state = CEPH_MDS_SESSION_NEW;
 743         s->s_ttl = 0;
 744         s->s_seq = 0;
 745         mutex_init(&s->s_mutex);
 746
 747         ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
 748
 749         spin_lock_init(&s->s_gen_ttl_lock);
 750         s->s_cap_gen = 1;
 751         s->s_cap_ttl = jiffies - 1;
 752
 753         spin_lock_init(&s->s_cap_lock);
 754         s->s_renew_requested = 0;
 755         s->s_renew_seq = 0;
 756         INIT_LIST_HEAD(&s->s_caps);
 757         s->s_nr_caps = 0;
 758         refcount_set(&s->s_ref, 1);
 759         INIT_LIST_HEAD(&s->s_waiting);
 760         INIT_LIST_HEAD(&s->s_unsafe);
 761         xa_init(&s->s_delegated_inos);
 762         s->s_num_cap_releases = 0;
 763         s->s_cap_reconnect = 0;
 764         s->s_cap_iterator = NULL;
 765         INIT_LIST_HEAD(&s->s_cap_releases);
 766         INIT_WORK(&s->s_cap_release_work, ceph_cap_release_work);
 767
 768         INIT_LIST_HEAD(&s->s_cap_dirty);
 769         INIT_LIST_HEAD(&s->s_cap_flushing);
 770
 771         mdsc->sessions[mds] = s;
 772         atomic_inc(&mdsc->num_sessions);
 773         refcount_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
 774
 775         ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
 776                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
 777
 778         return s;
 779
 780 fail_realloc:
 781         kfree(s);
 782         return ERR_PTR(-ENOMEM);
 783 }
 784
 785 /*
 786  * called under mdsc->mutex
 787  */
 788 static void __unregister_session(struct ceph_mds_client *mdsc,
 789                                struct ceph_mds_session *s)
 790 {
 791         dout("__unregister_session mds%d %p\n", s->s_mds, s);
 792         BUG_ON(mdsc->sessions[s->s_mds] != s);
 793         mdsc->sessions[s->s_mds] = NULL;
 794         ceph_con_close(&s->s_con);
 795         ceph_put_mds_session(s);
 796         atomic_dec(&mdsc->num_sessions);
 797 }
 798
 799 /*
 800  * drop session refs in request.
 801  *
 802  * should be last request ref, or hold mdsc->mutex
 803  */
 804 static void put_request_session(struct ceph_mds_request *req)
 805 {
 806         if (req->r_session) {
 807                 ceph_put_mds_session(req->r_session);
 808                 req->r_session = NULL;
 809         }
 810 }
 811
 812 void ceph_mdsc_release_request(struct kref *kref)
 813 {
 814         struct ceph_mds_request *req = container_of(kref,
 815                                                     struct ceph_mds_request,
 816                                                     r_kref);
 817         ceph_mdsc_release_dir_caps_no_check(req);
 818         destroy_reply_info(&req->r_reply_info);
 819         if (req->r_request)
 820                 ceph_msg_put(req->r_request);
 821         if (req->r_reply)
 822                 ceph_msg_put(req->r_reply);
 823         if (req->r_inode) {
 824                 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
 825                 /* avoid calling iput_final() in mds dispatch threads */
 826                 ceph_async_iput(req->r_inode);
 827         }
 828         if (req->r_parent) {
 829                 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
 830                 ceph_async_iput(req->r_parent);
 831         }
 832         ceph_async_iput(req->r_target_inode);
 833         if (req->r_dentry)
 834                 dput(req->r_dentry);
 835         if (req->r_old_dentry)
 836                 dput(req->r_old_dentry);
 837         if (req->r_old_dentry_dir) {
 838                 /*
 839                  * track (and drop pins for) r_old_dentry_dir
 840                  * separately, since r_old_dentry's d_parent may have
 841                  * changed between the dir mutex being dropped and
 842                  * this request being freed.
 843                  */
 844                 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
 845                                   CEPH_CAP_PIN);
 846                 ceph_async_iput(req->r_old_dentry_dir);
 847         }
 848         kfree(req->r_path1);
 849         kfree(req->r_path2);
 850         put_cred(req->r_cred);
 851         if (req->r_pagelist)
 852                 ceph_pagelist_release(req->r_pagelist);
 853         put_request_session(req);
 854         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
 855         WARN_ON_ONCE(!list_empty(&req->r_wait));
 856         kmem_cache_free(ceph_mds_request_cachep, req);
 857 }
 858
 859 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
 860
 861 /*
 862  * lookup session, bump ref if found.
 863  *
 864  * called under mdsc->mutex.
 865  */
 866 static struct ceph_mds_request *
 867 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
 868 {
 869         struct ceph_mds_request *req;
 870
 871         req = lookup_request(&mdsc->request_tree, tid);
 872         if (req)
 873                 ceph_mdsc_get_request(req);
 874
 875         return req;
 876 }
 877
 878 /*
 879  * Register an in-flight request, and assign a tid.  Link to directory
 880  * are modifying (if any).
 881  *
 882  * Called under mdsc->mutex.
 883  */
 884 static void __register_request(struct ceph_mds_client *mdsc,
 885                                struct ceph_mds_request *req,
 886                                struct inode *dir)
 887 {
 888         int ret = 0;
 889
 890         req->r_tid = ++mdsc->last_tid;
 891         if (req->r_num_caps) {
 892                 ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
 893                                         req->r_num_caps);
 894                 if (ret < 0) {
 895                         pr_err("__register_request %p "
 896                                "failed to reserve caps: %d\n", req, ret);
 897                         /* set req->r_err to fail early from __do_request */
 898                         req->r_err = ret;
 899                         return;
 900                 }
 901         }
 902         dout("__register_request %p tid %lld\n", req, req->r_tid);
 903         ceph_mdsc_get_request(req);
 904         insert_request(&mdsc->request_tree, req);
 905
 906         req->r_cred = get_current_cred();
 907
 908         if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
 909                 mdsc->oldest_tid = req->r_tid;
 910
 911         if (dir) {
 912                 struct ceph_inode_info *ci = ceph_inode(dir);
 913
 914                 ihold(dir);
 915                 req->r_unsafe_dir = dir;
 916                 spin_lock(&ci->i_unsafe_lock);
 917                 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
 918                 spin_unlock(&ci->i_unsafe_lock);
 919         }
 920 }
 921
 922 static void __unregister_request(struct ceph_mds_client *mdsc,
 923                                  struct ceph_mds_request *req)
 924 {
 925         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 926
 927         /* Never leave an unregistered request on an unsafe list! */
 928         list_del_init(&req->r_unsafe_item);
 929
 930         if (req->r_tid == mdsc->oldest_tid) {
 931                 struct rb_node *p = rb_next(&req->r_node);
 932                 mdsc->oldest_tid = 0;
 933                 while (p) {
 934                         struct ceph_mds_request *next_req =
 935                                 rb_entry(p, struct ceph_mds_request, r_node);
 936                         if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
 937                                 mdsc->oldest_tid = next_req->r_tid;
 938                                 break;
 939                         }
 940                         p = rb_next(p);
 941                 }
 942         }
 943
 944         erase_request(&mdsc->request_tree, req);
 945
 946         if (req->r_unsafe_dir) {
 947                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 948                 spin_lock(&ci->i_unsafe_lock);
 949                 list_del_init(&req->r_unsafe_dir_item);
 950                 spin_unlock(&ci->i_unsafe_lock);
 951         }
 952         if (req->r_target_inode &&
 953             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
 954                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
 955                 spin_lock(&ci->i_unsafe_lock);
 956                 list_del_init(&req->r_unsafe_target_item);
 957                 spin_unlock(&ci->i_unsafe_lock);
 958         }
 959
 960         if (req->r_unsafe_dir) {
 961                 /* avoid calling iput_final() in mds dispatch threads */
 962                 ceph_async_iput(req->r_unsafe_dir);
 963                 req->r_unsafe_dir = NULL;
 964         }
 965
 966         complete_all(&req->r_safe_completion);
 967
 968         ceph_mdsc_put_request(req);
 969 }
 970
 971 /*
 972  * Walk back up the dentry tree until we hit a dentry representing a
 973  * non-snapshot inode. We do this using the rcu_read_lock (which must be held
 974  * when calling this) to ensure that the objects won't disappear while we're
 975  * working with them. Once we hit a candidate dentry, we attempt to take a
 976  * reference to it, and return that as the result.
 977  */
 978 static struct inode *get_nonsnap_parent(struct dentry *dentry)
 979 {
 980         struct inode *inode = NULL;
 981
 982         while (dentry && !IS_ROOT(dentry)) {
 983                 inode = d_inode_rcu(dentry);
 984                 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
 985                         break;
 986                 dentry = dentry->d_parent;
 987         }
 988         if (inode)
 989                 inode = igrab(inode);
 990         return inode;
 991 }
 992
 993 /*
 994  * Choose mds to send request to next.  If there is a hint set in the
 995  * request (e.g., due to a prior forward hint from the mds), use that.
 996  * Otherwise, consult frag tree and/or caps to identify the
 997  * appropriate mds.  If all else fails, choose randomly.
 998  *
 999  * Called under mdsc->mutex.
1000  */
1001 static int __choose_mds(struct ceph_mds_client *mdsc,
1002                         struct ceph_mds_request *req,
1003                         bool *random)
1004 {
1005         struct inode *inode;
1006         struct ceph_inode_info *ci;
1007         struct ceph_cap *cap;
1008         int mode = req->r_direct_mode;
1009         int mds = -1;
1010         u32 hash = req->r_direct_hash;
1011         bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
1012
1013         if (random)
1014                 *random = false;
1015
1016         /*
1017          * is there a specific mds we should try?  ignore hint if we have
1018          * no session and the mds is not up (active or recovering).
1019          */
1020         if (req->r_resend_mds >= 0 &&
1021             (__have_session(mdsc, req->r_resend_mds) ||
1022              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
1023                 dout("%s using resend_mds mds%d\n", __func__,
1024                      req->r_resend_mds);
1025                 return req->r_resend_mds;
1026         }
1027
1028         if (mode == USE_RANDOM_MDS)
1029                 goto random;
1030
1031         inode = NULL;
1032         if (req->r_inode) {
1033                 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
1034                         inode = req->r_inode;
1035                         ihold(inode);
1036                 } else {
1037                         /* req->r_dentry is non-null for LSSNAP request */
1038                         rcu_read_lock();
1039                         inode = get_nonsnap_parent(req->r_dentry);
1040                         rcu_read_unlock();
1041                         dout("%s using snapdir's parent %p\n", __func__, inode);
1042                 }
1043         } else if (req->r_dentry) {
1044                 /* ignore race with rename; old or new d_parent is okay */
1045                 struct dentry *parent;
1046                 struct inode *dir;
1047
1048                 rcu_read_lock();
1049                 parent = READ_ONCE(req->r_dentry->d_parent);
1050                 dir = req->r_parent ? : d_inode_rcu(parent);
1051
1052                 if (!dir || dir->i_sb != mdsc->fsc->sb) {
1053                         /*  not this fs or parent went negative */
1054                         inode = d_inode(req->r_dentry);
1055                         if (inode)
1056                                 ihold(inode);
1057                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
1058                         /* direct snapped/virtual snapdir requests
1059                          * based on parent dir inode */
1060                         inode = get_nonsnap_parent(parent);
1061                         dout("%s using nonsnap parent %p\n", __func__, inode);
1062                 } else {
1063                         /* dentry target */
1064                         inode = d_inode(req->r_dentry);
1065                         if (!inode || mode == USE_AUTH_MDS) {
1066                                 /* dir + name */
1067                                 inode = igrab(dir);
1068                                 hash = ceph_dentry_hash(dir, req->r_dentry);
1069                                 is_hash = true;
1070                         } else {
1071                                 ihold(inode);
1072                         }
1073                 }
1074                 rcu_read_unlock();
1075         }
1076
1077         dout("%s %p is_hash=%d (0x%x) mode %d\n", __func__, inode, (int)is_hash,
1078              hash, mode);
1079         if (!inode)
1080                 goto random;
1081         ci = ceph_inode(inode);
1082
1083         if (is_hash && S_ISDIR(inode->i_mode)) {
1084                 struct ceph_inode_frag frag;
1085                 int found;
1086
1087                 ceph_choose_frag(ci, hash, &frag, &found);
1088                 if (found) {
1089                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
1090                                 u8 r;
1091
1092                                 /* choose a random replica */
1093                                 get_random_bytes(&r, 1);
1094                                 r %= frag.ndist;
1095                                 mds = frag.dist[r];
1096                                 dout("%s %p %llx.%llx frag %u mds%d (%d/%d)\n",
1097                                      __func__, inode, ceph_vinop(inode),
1098                                      frag.frag, mds, (int)r, frag.ndist);
1099                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1100                                     CEPH_MDS_STATE_ACTIVE &&
1101                                     !ceph_mdsmap_is_laggy(mdsc->mdsmap, mds))
1102                                         goto out;
1103                         }
1104
1105                         /* since this file/dir wasn't known to be
1106                          * replicated, then we want to look for the
1107                          * authoritative mds. */
1108                         if (frag.mds >= 0) {
1109                                 /* choose auth mds */
1110                                 mds = frag.mds;
1111                                 dout("%s %p %llx.%llx frag %u mds%d (auth)\n",
1112                                      __func__, inode, ceph_vinop(inode),
1113                                      frag.frag, mds);
1114                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1115                                     CEPH_MDS_STATE_ACTIVE) {
1116                                         if (!ceph_mdsmap_is_laggy(mdsc->mdsmap,
1117                                                                   mds))
1118                                                 goto out;
1119                                 }
1120                         }
1121                         mode = USE_AUTH_MDS;
1122                 }
1123         }
1124
1125         spin_lock(&ci->i_ceph_lock);
1126         cap = NULL;
1127         if (mode == USE_AUTH_MDS)
1128                 cap = ci->i_auth_cap;
1129         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
1130                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
1131         if (!cap) {
1132                 spin_unlock(&ci->i_ceph_lock);
1133                 ceph_async_iput(inode);
1134                 goto random;
1135         }
1136         mds = cap->session->s_mds;
1137         dout("%s %p %llx.%llx mds%d (%scap %p)\n", __func__,
1138              inode, ceph_vinop(inode), mds,
1139              cap == ci->i_auth_cap ? "auth " : "", cap);
1140         spin_unlock(&ci->i_ceph_lock);
1141 out:
1142         /* avoid calling iput_final() while holding mdsc->mutex or
1143          * in mds dispatch threads */
1144         ceph_async_iput(inode);
1145         return mds;
1146
1147 random:
1148         if (random)
1149                 *random = true;
1150
1151         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
1152         dout("%s chose random mds%d\n", __func__, mds);
1153         return mds;
1154 }
1155
1156
1157 /*
1158  * session messages
1159  */
1160 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
1161 {
1162         struct ceph_msg *msg;
1163         struct ceph_mds_session_head *h;
1164
1165         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
1166                            false);
1167         if (!msg) {
1168                 pr_err("create_session_msg ENOMEM creating msg\n");
1169                 return NULL;
1170         }
1171         h = msg->front.iov_base;
1172         h->op = cpu_to_le32(op);
1173         h->seq = cpu_to_le64(seq);
1174
1175         return msg;
1176 }
1177
1178 static const unsigned char feature_bits[] = CEPHFS_FEATURES_CLIENT_SUPPORTED;
1179 #define FEATURE_BYTES(c) (DIV_ROUND_UP((size_t)feature_bits[c - 1] + 1, 64) * 8)
1180 static int encode_supported_features(void **p, void *end)
1181 {
1182         static const size_t count = ARRAY_SIZE(feature_bits);
1183
1184         if (count > 0) {
1185                 size_t i;
1186                 size_t size = FEATURE_BYTES(count);
1187
1188                 if (WARN_ON_ONCE(*p + 4 + size > end))
1189                         return -ERANGE;
1190
1191                 ceph_encode_32(p, size);
1192                 memset(*p, 0, size);
1193                 for (i = 0; i < count; i++)
1194                         ((unsigned char*)(*p))[i / 8] |= BIT(feature_bits[i] % 8);
1195                 *p += size;
1196         } else {
1197                 if (WARN_ON_ONCE(*p + 4 > end))
1198                         return -ERANGE;
1199
1200                 ceph_encode_32(p, 0);
1201         }
1202
1203         return 0;
1204 }
1205
1206 static const unsigned char metric_bits[] = CEPHFS_METRIC_SPEC_CLIENT_SUPPORTED;
1207 #define METRIC_BYTES(cnt) (DIV_ROUND_UP((size_t)metric_bits[cnt - 1] + 1, 64) * 8)
1208 static int encode_metric_spec(void **p, void *end)
1209 {
1210         static const size_t count = ARRAY_SIZE(metric_bits);
1211
1212         /* header */
1213         if (WARN_ON_ONCE(*p + 2 > end))
1214                 return -ERANGE;
1215
1216         ceph_encode_8(p, 1); /* version */
1217         ceph_encode_8(p, 1); /* compat */
1218
1219         if (count > 0) {
1220                 size_t i;
1221                 size_t size = METRIC_BYTES(count);
1222
1223                 if (WARN_ON_ONCE(*p + 4 + 4 + size > end))
1224                         return -ERANGE;
1225
1226                 /* metric spec info length */
1227                 ceph_encode_32(p, 4 + size);
1228
1229                 /* metric spec */
1230                 ceph_encode_32(p, size);
1231                 memset(*p, 0, size);
1232                 for (i = 0; i < count; i++)
1233                         ((unsigned char *)(*p))[i / 8] |= BIT(metric_bits[i] % 8);
1234                 *p += size;
1235         } else {
1236                 if (WARN_ON_ONCE(*p + 4 + 4 > end))
1237                         return -ERANGE;
1238
1239                 /* metric spec info length */
1240                 ceph_encode_32(p, 4);
1241                 /* metric spec */
1242                 ceph_encode_32(p, 0);
1243         }
1244
1245         return 0;
1246 }
1247
1248 /*
1249  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
1250  * to include additional client metadata fields.
1251  */
1252 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
1253 {
1254         struct ceph_msg *msg;
1255         struct ceph_mds_session_head *h;
1256         int i;
1257         int extra_bytes = 0;
1258         int metadata_key_count = 0;
1259         struct ceph_options *opt = mdsc->fsc->client->options;
1260         struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
1261         size_t size, count;
1262         void *p, *end;
1263         int ret;
1264
1265         const char* metadata[][2] = {
1266                 {"hostname", mdsc->nodename},
1267                 {"kernel_version", init_utsname()->release},
1268                 {"entity_id", opt->name ? : ""},
1269                 {"root", fsopt->server_path ? : "/"},
1270                 {NULL, NULL}
1271         };
1272
1273         /* Calculate serialized length of metadata */
1274         extra_bytes = 4;  /* map length */
1275         for (i = 0; metadata[i][0]; ++i) {
1276                 extra_bytes += 8 + strlen(metadata[i][0]) +
1277                         strlen(metadata[i][1]);
1278                 metadata_key_count++;
1279         }
1280
1281         /* supported feature */
1282         size = 0;
1283         count = ARRAY_SIZE(feature_bits);
1284         if (count > 0)
1285                 size = FEATURE_BYTES(count);
1286         extra_bytes += 4 + size;
1287
1288         /* metric spec */
1289         size = 0;
1290         count = ARRAY_SIZE(metric_bits);
1291         if (count > 0)
1292                 size = METRIC_BYTES(count);
1293         extra_bytes += 2 + 4 + 4 + size;
1294
1295         /* Allocate the message */
1296         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + extra_bytes,
1297                            GFP_NOFS, false);
1298         if (!msg) {
1299                 pr_err("create_session_msg ENOMEM creating msg\n");
1300                 return ERR_PTR(-ENOMEM);
1301         }
1302         p = msg->front.iov_base;
1303         end = p + msg->front.iov_len;
1304
1305         h = p;
1306         h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
1307         h->seq = cpu_to_le64(seq);
1308
1309         /*
1310          * Serialize client metadata into waiting buffer space, using
1311          * the format that userspace expects for map<string, string>
1312          *
1313          * ClientSession messages with metadata are v4
1314          */
1315         msg->hdr.version = cpu_to_le16(4);
1316         msg->hdr.compat_version = cpu_to_le16(1);
1317
1318         /* The write pointer, following the session_head structure */
1319         p += sizeof(*h);
1320
1321         /* Number of entries in the map */
1322         ceph_encode_32(&p, metadata_key_count);
1323
1324         /* Two length-prefixed strings for each entry in the map */
1325         for (i = 0; metadata[i][0]; ++i) {
1326                 size_t const key_len = strlen(metadata[i][0]);
1327                 size_t const val_len = strlen(metadata[i][1]);
1328
1329                 ceph_encode_32(&p, key_len);
1330                 memcpy(p, metadata[i][0], key_len);
1331                 p += key_len;
1332                 ceph_encode_32(&p, val_len);
1333                 memcpy(p, metadata[i][1], val_len);
1334                 p += val_len;
1335         }
1336
1337         ret = encode_supported_features(&p, end);
1338         if (ret) {
1339                 pr_err("encode_supported_features failed!\n");
1340                 ceph_msg_put(msg);
1341                 return ERR_PTR(ret);
1342         }
1343
1344         ret = encode_metric_spec(&p, end);
1345         if (ret) {
1346                 pr_err("encode_metric_spec failed!\n");
1347                 ceph_msg_put(msg);
1348                 return ERR_PTR(ret);
1349         }
1350
1351         msg->front.iov_len = p - msg->front.iov_base;
1352         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1353
1354         return msg;
1355 }
1356
1357 /*
1358  * send session open request.
1359  *
1360  * called under mdsc->mutex
1361  */
1362 static int __open_session(struct ceph_mds_client *mdsc,
1363                           struct ceph_mds_session *session)
1364 {
1365         struct ceph_msg *msg;
1366         int mstate;
1367         int mds = session->s_mds;
1368
1369         /* wait for mds to go active? */
1370         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
1371         dout("open_session to mds%d (%s)\n", mds,
1372              ceph_mds_state_name(mstate));
1373         session->s_state = CEPH_MDS_SESSION_OPENING;
1374         session->s_renew_requested = jiffies;
1375
1376         /* send connect message */
1377         msg = create_session_open_msg(mdsc, session->s_seq);
1378         if (IS_ERR(msg))
1379                 return PTR_ERR(msg);
1380         ceph_con_send(&session->s_con, msg);
1381         return 0;
1382 }
1383
1384 /*
1385  * open sessions for any export targets for the given mds
1386  *
1387  * called under mdsc->mutex
1388  */
1389 static struct ceph_mds_session *
1390 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
1391 {
1392         struct ceph_mds_session *session;
1393         int ret;
1394
1395         session = __ceph_lookup_mds_session(mdsc, target);
1396         if (!session) {
1397                 session = register_session(mdsc, target);
1398                 if (IS_ERR(session))
1399                         return session;
1400         }
1401         if (session->s_state == CEPH_MDS_SESSION_NEW ||
1402             session->s_state == CEPH_MDS_SESSION_CLOSING) {
1403                 ret = __open_session(mdsc, session);
1404                 if (ret)
1405                         return ERR_PTR(ret);
1406         }
1407
1408         return session;
1409 }
1410
1411 struct ceph_mds_session *
1412 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1413 {
1414         struct ceph_mds_session *session;
1415
1416         dout("open_export_target_session to mds%d\n", target);
1417
1418         mutex_lock(&mdsc->mutex);
1419         session = __open_export_target_session(mdsc, target);
1420         mutex_unlock(&mdsc->mutex);
1421
1422         return session;
1423 }
1424
1425 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1426                                           struct ceph_mds_session *session)
1427 {
1428         struct ceph_mds_info *mi;
1429         struct ceph_mds_session *ts;
1430         int i, mds = session->s_mds;
1431
1432         if (mds >= mdsc->mdsmap->possible_max_rank)
1433                 return;
1434
1435         mi = &mdsc->mdsmap->m_info[mds];
1436         dout("open_export_target_sessions for mds%d (%d targets)\n",
1437              session->s_mds, mi->num_export_targets);
1438
1439         for (i = 0; i < mi->num_export_targets; i++) {
1440                 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1441                 if (!IS_ERR(ts))
1442                         ceph_put_mds_session(ts);
1443         }
1444 }
1445
1446 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1447                                            struct ceph_mds_session *session)
1448 {
1449         mutex_lock(&mdsc->mutex);
1450         __open_export_target_sessions(mdsc, session);
1451         mutex_unlock(&mdsc->mutex);
1452 }
1453
1454 /*
1455  * session caps
1456  */
1457
1458 static void detach_cap_releases(struct ceph_mds_session *session,
1459                                 struct list_head *target)
1460 {
1461         lockdep_assert_held(&session->s_cap_lock);
1462
1463         list_splice_init(&session->s_cap_releases, target);
1464         session->s_num_cap_releases = 0;
1465         dout("dispose_cap_releases mds%d\n", session->s_mds);
1466 }
1467
1468 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1469                                  struct list_head *dispose)
1470 {
1471         while (!list_empty(dispose)) {
1472                 struct ceph_cap *cap;
1473                 /* zero out the in-progress message */
1474                 cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1475                 list_del(&cap->session_caps);
1476                 ceph_put_cap(mdsc, cap);
1477         }
1478 }
1479
1480 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1481                                      struct ceph_mds_session *session)
1482 {
1483         struct ceph_mds_request *req;
1484         struct rb_node *p;
1485         struct ceph_inode_info *ci;
1486
1487         dout("cleanup_session_requests mds%d\n", session->s_mds);
1488         mutex_lock(&mdsc->mutex);
1489         while (!list_empty(&session->s_unsafe)) {
1490                 req = list_first_entry(&session->s_unsafe,
1491                                        struct ceph_mds_request, r_unsafe_item);
1492                 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1493                                     req->r_tid);
1494                 if (req->r_target_inode) {
1495                         /* dropping unsafe change of inode's attributes */
1496                         ci = ceph_inode(req->r_target_inode);
1497                         errseq_set(&ci->i_meta_err, -EIO);
1498                 }
1499                 if (req->r_unsafe_dir) {
1500                         /* dropping unsafe directory operation */
1501                         ci = ceph_inode(req->r_unsafe_dir);
1502                         errseq_set(&ci->i_meta_err, -EIO);
1503                 }
1504                 __unregister_request(mdsc, req);
1505         }
1506         /* zero r_attempts, so kick_requests() will re-send requests */
1507         p = rb_first(&mdsc->request_tree);
1508         while (p) {
1509                 req = rb_entry(p, struct ceph_mds_request, r_node);
1510                 p = rb_next(p);
1511                 if (req->r_session &&
1512                     req->r_session->s_mds == session->s_mds)
1513                         req->r_attempts = 0;
1514         }
1515         mutex_unlock(&mdsc->mutex);
1516 }
1517
1518 /*
1519  * Helper to safely iterate over all caps associated with a session, with
1520  * special care taken to handle a racing __ceph_remove_cap().
1521  *
1522  * Caller must hold session s_mutex.
1523  */
1524 int ceph_iterate_session_caps(struct ceph_mds_session *session,
1525                               int (*cb)(struct inode *, struct ceph_cap *,
1526                                         void *), void *arg)
1527 {
1528         struct list_head *p;
1529         struct ceph_cap *cap;
1530         struct inode *inode, *last_inode = NULL;
1531         struct ceph_cap *old_cap = NULL;
1532         int ret;
1533
1534         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1535         spin_lock(&session->s_cap_lock);
1536         p = session->s_caps.next;
1537         while (p != &session->s_caps) {
1538                 cap = list_entry(p, struct ceph_cap, session_caps);
1539                 inode = igrab(&cap->ci->vfs_inode);
1540                 if (!inode) {
1541                         p = p->next;
1542                         continue;
1543                 }
1544                 session->s_cap_iterator = cap;
1545                 spin_unlock(&session->s_cap_lock);
1546
1547                 if (last_inode) {
1548                         /* avoid calling iput_final() while holding
1549                          * s_mutex or in mds dispatch threads */
1550                         ceph_async_iput(last_inode);
1551                         last_inode = NULL;
1552                 }
1553                 if (old_cap) {
1554                         ceph_put_cap(session->s_mdsc, old_cap);
1555                         old_cap = NULL;
1556                 }
1557
1558                 ret = cb(inode, cap, arg);
1559                 last_inode = inode;
1560
1561                 spin_lock(&session->s_cap_lock);
1562                 p = p->next;
1563                 if (!cap->ci) {
1564                         dout("iterate_session_caps  finishing cap %p removal\n",
1565                              cap);
1566                         BUG_ON(cap->session != session);
1567                         cap->session = NULL;
1568                         list_del_init(&cap->session_caps);
1569                         session->s_nr_caps--;
1570                         atomic64_dec(&session->s_mdsc->metric.total_caps);
1571                         if (cap->queue_release)
1572                                 __ceph_queue_cap_release(session, cap);
1573                         else
1574                                 old_cap = cap;  /* put_cap it w/o locks held */
1575                 }
1576                 if (ret < 0)
1577                         goto out;
1578         }
1579         ret = 0;
1580 out:
1581         session->s_cap_iterator = NULL;
1582         spin_unlock(&session->s_cap_lock);
1583
1584         ceph_async_iput(last_inode);
1585         if (old_cap)
1586                 ceph_put_cap(session->s_mdsc, old_cap);
1587
1588         return ret;
1589 }
1590
1591 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1592                                   void *arg)
1593 {
1594         struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1595         struct ceph_inode_info *ci = ceph_inode(inode);
1596         LIST_HEAD(to_remove);
1597         bool dirty_dropped = false;
1598         bool invalidate = false;
1599
1600         dout("removing cap %p, ci is %p, inode is %p\n",
1601              cap, ci, &ci->vfs_inode);
1602         spin_lock(&ci->i_ceph_lock);
1603         __ceph_remove_cap(cap, false);
1604         if (!ci->i_auth_cap) {
1605                 struct ceph_cap_flush *cf;
1606                 struct ceph_mds_client *mdsc = fsc->mdsc;
1607
1608                 if (READ_ONCE(fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN) {
1609                         if (inode->i_data.nrpages > 0)
1610                                 invalidate = true;
1611                         if (ci->i_wrbuffer_ref > 0)
1612                                 mapping_set_error(&inode->i_data, -EIO);
1613                 }
1614
1615                 while (!list_empty(&ci->i_cap_flush_list)) {
1616                         cf = list_first_entry(&ci->i_cap_flush_list,
1617                                               struct ceph_cap_flush, i_list);
1618                         list_move(&cf->i_list, &to_remove);
1619                 }
1620
1621                 spin_lock(&mdsc->cap_dirty_lock);
1622
1623                 list_for_each_entry(cf, &to_remove, i_list)
1624                         list_del(&cf->g_list);
1625
1626                 if (!list_empty(&ci->i_dirty_item)) {
1627                         pr_warn_ratelimited(
1628                                 " dropping dirty %s state for %p %lld\n",
1629                                 ceph_cap_string(ci->i_dirty_caps),
1630                                 inode, ceph_ino(inode));
1631                         ci->i_dirty_caps = 0;
1632                         list_del_init(&ci->i_dirty_item);
1633                         dirty_dropped = true;
1634                 }
1635                 if (!list_empty(&ci->i_flushing_item)) {
1636                         pr_warn_ratelimited(
1637                                 " dropping dirty+flushing %s state for %p %lld\n",
1638                                 ceph_cap_string(ci->i_flushing_caps),
1639                                 inode, ceph_ino(inode));
1640                         ci->i_flushing_caps = 0;
1641                         list_del_init(&ci->i_flushing_item);
1642                         mdsc->num_cap_flushing--;
1643                         dirty_dropped = true;
1644                 }
1645                 spin_unlock(&mdsc->cap_dirty_lock);
1646
1647                 if (dirty_dropped) {
1648                         errseq_set(&ci->i_meta_err, -EIO);
1649
1650                         if (ci->i_wrbuffer_ref_head == 0 &&
1651                             ci->i_wr_ref == 0 &&
1652                             ci->i_dirty_caps == 0 &&
1653                             ci->i_flushing_caps == 0) {
1654                                 ceph_put_snap_context(ci->i_head_snapc);
1655                                 ci->i_head_snapc = NULL;
1656                         }
1657                 }
1658
1659                 if (atomic_read(&ci->i_filelock_ref) > 0) {
1660                         /* make further file lock syscall return -EIO */
1661                         ci->i_ceph_flags |= CEPH_I_ERROR_FILELOCK;
1662                         pr_warn_ratelimited(" dropping file locks for %p %lld\n",
1663                                             inode, ceph_ino(inode));
1664                 }
1665
1666                 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1667                         list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1668                         ci->i_prealloc_cap_flush = NULL;
1669                 }
1670         }
1671         spin_unlock(&ci->i_ceph_lock);
1672         while (!list_empty(&to_remove)) {
1673                 struct ceph_cap_flush *cf;
1674                 cf = list_first_entry(&to_remove,
1675                                       struct ceph_cap_flush, i_list);
1676                 list_del(&cf->i_list);
1677                 ceph_free_cap_flush(cf);
1678         }
1679
1680         wake_up_all(&ci->i_cap_wq);
1681         if (invalidate)
1682                 ceph_queue_invalidate(inode);
1683         if (dirty_dropped)
1684                 iput(inode);
1685         return 0;
1686 }
1687
1688 /*
1689  * caller must hold session s_mutex
1690  */
1691 static void remove_session_caps(struct ceph_mds_session *session)
1692 {
1693         struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1694         struct super_block *sb = fsc->sb;
1695         LIST_HEAD(dispose);
1696
1697         dout("remove_session_caps on %p\n", session);
1698         ceph_iterate_session_caps(session, remove_session_caps_cb, fsc);
1699
1700         wake_up_all(&fsc->mdsc->cap_flushing_wq);
1701
1702         spin_lock(&session->s_cap_lock);
1703         if (session->s_nr_caps > 0) {
1704                 struct inode *inode;
1705                 struct ceph_cap *cap, *prev = NULL;
1706                 struct ceph_vino vino;
1707                 /*
1708                  * iterate_session_caps() skips inodes that are being
1709                  * deleted, we need to wait until deletions are complete.
1710                  * __wait_on_freeing_inode() is designed for the job,
1711                  * but it is not exported, so use lookup inode function
1712                  * to access it.
1713                  */
1714                 while (!list_empty(&session->s_caps)) {
1715                         cap = list_entry(session->s_caps.next,
1716                                          struct ceph_cap, session_caps);
1717                         if (cap == prev)
1718                                 break;
1719                         prev = cap;
1720                         vino = cap->ci->i_vino;
1721                         spin_unlock(&session->s_cap_lock);
1722
1723                         inode = ceph_find_inode(sb, vino);
1724                          /* avoid calling iput_final() while holding s_mutex */
1725                         ceph_async_iput(inode);
1726
1727                         spin_lock(&session->s_cap_lock);
1728                 }
1729         }
1730
1731         // drop cap expires and unlock s_cap_lock
1732         detach_cap_releases(session, &dispose);
1733
1734         BUG_ON(session->s_nr_caps > 0);
1735         BUG_ON(!list_empty(&session->s_cap_flushing));
1736         spin_unlock(&session->s_cap_lock);
1737         dispose_cap_releases(session->s_mdsc, &dispose);
1738 }
1739
1740 enum {
1741         RECONNECT,
1742         RENEWCAPS,
1743         FORCE_RO,
1744 };
1745
1746 /*
1747  * wake up any threads waiting on this session's caps.  if the cap is
1748  * old (didn't get renewed on the client reconnect), remove it now.
1749  *
1750  * caller must hold s_mutex.
1751  */
1752 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1753                               void *arg)
1754 {
1755         struct ceph_inode_info *ci = ceph_inode(inode);
1756         unsigned long ev = (unsigned long)arg;
1757
1758         if (ev == RECONNECT) {
1759                 spin_lock(&ci->i_ceph_lock);
1760                 ci->i_wanted_max_size = 0;
1761                 ci->i_requested_max_size = 0;
1762                 spin_unlock(&ci->i_ceph_lock);
1763         } else if (ev == RENEWCAPS) {
1764                 if (cap->cap_gen < cap->session->s_cap_gen) {
1765                         /* mds did not re-issue stale cap */
1766                         spin_lock(&ci->i_ceph_lock);
1767                         cap->issued = cap->implemented = CEPH_CAP_PIN;
1768                         spin_unlock(&ci->i_ceph_lock);
1769                 }
1770         } else if (ev == FORCE_RO) {
1771         }
1772         wake_up_all(&ci->i_cap_wq);
1773         return 0;
1774 }
1775
1776 static void wake_up_session_caps(struct ceph_mds_session *session, int ev)
1777 {
1778         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1779         ceph_iterate_session_caps(session, wake_up_session_cb,
1780                                   (void *)(unsigned long)ev);
1781 }
1782
1783 /*
1784  * Send periodic message to MDS renewing all currently held caps.  The
1785  * ack will reset the expiration for all caps from this session.
1786  *
1787  * caller holds s_mutex
1788  */
1789 static int send_renew_caps(struct ceph_mds_client *mdsc,
1790                            struct ceph_mds_session *session)
1791 {
1792         struct ceph_msg *msg;
1793         int state;
1794
1795         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1796             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1797                 pr_info("mds%d caps stale\n", session->s_mds);
1798         session->s_renew_requested = jiffies;
1799
1800         /* do not try to renew caps until a recovering mds has reconnected
1801          * with its clients. */
1802         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1803         if (state < CEPH_MDS_STATE_RECONNECT) {
1804                 dout("send_renew_caps ignoring mds%d (%s)\n",
1805                      session->s_mds, ceph_mds_state_name(state));
1806                 return 0;
1807         }
1808
1809         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1810                 ceph_mds_state_name(state));
1811         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1812                                  ++session->s_renew_seq);
1813         if (!msg)
1814                 return -ENOMEM;
1815         ceph_con_send(&session->s_con, msg);
1816         return 0;
1817 }
1818
1819 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1820                              struct ceph_mds_session *session, u64 seq)
1821 {
1822         struct ceph_msg *msg;
1823
1824         dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1825              session->s_mds, ceph_session_state_name(session->s_state), seq);
1826         msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1827         if (!msg)
1828                 return -ENOMEM;
1829         ceph_con_send(&session->s_con, msg);
1830         return 0;
1831 }
1832
1833
1834 /*
1835  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1836  *
1837  * Called under session->s_mutex
1838  */
1839 static void renewed_caps(struct ceph_mds_client *mdsc,
1840                          struct ceph_mds_session *session, int is_renew)
1841 {
1842         int was_stale;
1843         int wake = 0;
1844
1845         spin_lock(&session->s_cap_lock);
1846         was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1847
1848         session->s_cap_ttl = session->s_renew_requested +
1849                 mdsc->mdsmap->m_session_timeout*HZ;
1850
1851         if (was_stale) {
1852                 if (time_before(jiffies, session->s_cap_ttl)) {
1853                         pr_info("mds%d caps renewed\n", session->s_mds);
1854                         wake = 1;
1855                 } else {
1856                         pr_info("mds%d caps still stale\n", session->s_mds);
1857                 }
1858         }
1859         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1860              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1861              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1862         spin_unlock(&session->s_cap_lock);
1863
1864         if (wake)
1865                 wake_up_session_caps(session, RENEWCAPS);
1866 }
1867
1868 /*
1869  * send a session close request
1870  */
1871 static int request_close_session(struct ceph_mds_session *session)
1872 {
1873         struct ceph_msg *msg;
1874
1875         dout("request_close_session mds%d state %s seq %lld\n",
1876              session->s_mds, ceph_session_state_name(session->s_state),
1877              session->s_seq);
1878         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1879         if (!msg)
1880                 return -ENOMEM;
1881         ceph_con_send(&session->s_con, msg);
1882         return 1;
1883 }
1884
1885 /*
1886  * Called with s_mutex held.
1887  */
1888 static int __close_session(struct ceph_mds_client *mdsc,
1889                          struct ceph_mds_session *session)
1890 {
1891         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1892                 return 0;
1893         session->s_state = CEPH_MDS_SESSION_CLOSING;
1894         return request_close_session(session);
1895 }
1896
1897 static bool drop_negative_children(struct dentry *dentry)
1898 {
1899         struct dentry *child;
1900         bool all_negative = true;
1901
1902         if (!d_is_dir(dentry))
1903                 goto out;
1904
1905         spin_lock(&dentry->d_lock);
1906         list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1907                 if (d_really_is_positive(child)) {
1908                         all_negative = false;
1909                         break;
1910                 }
1911         }
1912         spin_unlock(&dentry->d_lock);
1913
1914         if (all_negative)
1915                 shrink_dcache_parent(dentry);
1916 out:
1917         return all_negative;
1918 }
1919
1920 /*
1921  * Trim old(er) caps.
1922  *
1923  * Because we can't cache an inode without one or more caps, we do
1924  * this indirectly: if a cap is unused, we prune its aliases, at which
1925  * point the inode will hopefully get dropped to.
1926  *
1927  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1928  * memory pressure from the MDS, though, so it needn't be perfect.
1929  */
1930 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1931 {
1932         int *remaining = arg;
1933         struct ceph_inode_info *ci = ceph_inode(inode);
1934         int used, wanted, oissued, mine;
1935
1936         if (*remaining <= 0)
1937                 return -1;
1938
1939         spin_lock(&ci->i_ceph_lock);
1940         mine = cap->issued | cap->implemented;
1941         used = __ceph_caps_used(ci);
1942         wanted = __ceph_caps_file_wanted(ci);
1943         oissued = __ceph_caps_issued_other(ci, cap);
1944
1945         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1946              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1947              ceph_cap_string(used), ceph_cap_string(wanted));
1948         if (cap == ci->i_auth_cap) {
1949                 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1950                     !list_empty(&ci->i_cap_snaps))
1951                         goto out;
1952                 if ((used | wanted) & CEPH_CAP_ANY_WR)
1953                         goto out;
1954                 /* Note: it's possible that i_filelock_ref becomes non-zero
1955                  * after dropping auth caps. It doesn't hurt because reply
1956                  * of lock mds request will re-add auth caps. */
1957                 if (atomic_read(&ci->i_filelock_ref) > 0)
1958                         goto out;
1959         }
1960         /* The inode has cached pages, but it's no longer used.
1961          * we can safely drop it */
1962         if (S_ISREG(inode->i_mode) &&
1963             wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1964             !(oissued & CEPH_CAP_FILE_CACHE)) {
1965           used = 0;
1966           oissued = 0;
1967         }
1968         if ((used | wanted) & ~oissued & mine)
1969                 goto out;   /* we need these caps */
1970
1971         if (oissued) {
1972                 /* we aren't the only cap.. just remove us */
1973                 __ceph_remove_cap(cap, true);
1974                 (*remaining)--;
1975         } else {
1976                 struct dentry *dentry;
1977                 /* try dropping referring dentries */
1978                 spin_unlock(&ci->i_ceph_lock);
1979                 dentry = d_find_any_alias(inode);
1980                 if (dentry && drop_negative_children(dentry)) {
1981                         int count;
1982                         dput(dentry);
1983                         d_prune_aliases(inode);
1984                         count = atomic_read(&inode->i_count);
1985                         if (count == 1)
1986                                 (*remaining)--;
1987                         dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1988                              inode, cap, count);
1989                 } else {
1990                         dput(dentry);
1991                 }
1992                 return 0;
1993         }
1994
1995 out:
1996         spin_unlock(&ci->i_ceph_lock);
1997         return 0;
1998 }
1999
2000 /*
2001  * Trim session cap count down to some max number.
2002  */
2003 int ceph_trim_caps(struct ceph_mds_client *mdsc,
2004                    struct ceph_mds_session *session,
2005                    int max_caps)
2006 {
2007         int trim_caps = session->s_nr_caps - max_caps;
2008
2009         dout("trim_caps mds%d start: %d / %d, trim %d\n",
2010              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
2011         if (trim_caps > 0) {
2012                 int remaining = trim_caps;
2013
2014                 ceph_iterate_session_caps(session, trim_caps_cb, &remaining);
2015                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
2016                      session->s_mds, session->s_nr_caps, max_caps,
2017                         trim_caps - remaining);
2018         }
2019
2020         ceph_flush_cap_releases(mdsc, session);
2021         return 0;
2022 }
2023
2024 static int check_caps_flush(struct ceph_mds_client *mdsc,
2025                             u64 want_flush_tid)
2026 {
2027         int ret = 1;
2028
2029         spin_lock(&mdsc->cap_dirty_lock);
2030         if (!list_empty(&mdsc->cap_flush_list)) {
2031                 struct ceph_cap_flush *cf =
2032                         list_first_entry(&mdsc->cap_flush_list,
2033                                          struct ceph_cap_flush, g_list);
2034                 if (cf->tid <= want_flush_tid) {
2035                         dout("check_caps_flush still flushing tid "
2036                              "%llu <= %llu\n", cf->tid, want_flush_tid);
2037                         ret = 0;
2038                 }
2039         }
2040         spin_unlock(&mdsc->cap_dirty_lock);
2041         return ret;
2042 }
2043
2044 /*
2045  * flush all dirty inode data to disk.
2046  *
2047  * returns true if we've flushed through want_flush_tid
2048  */
2049 static void wait_caps_flush(struct ceph_mds_client *mdsc,
2050                             u64 want_flush_tid)
2051 {
2052         dout("check_caps_flush want %llu\n", want_flush_tid);
2053
2054         wait_event(mdsc->cap_flushing_wq,
2055                    check_caps_flush(mdsc, want_flush_tid));
2056
2057         dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
2058 }
2059
2060 /*
2061  * called under s_mutex
2062  */
2063 static void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
2064                                    struct ceph_mds_session *session)
2065 {
2066         struct ceph_msg *msg = NULL;
2067         struct ceph_mds_cap_release *head;
2068         struct ceph_mds_cap_item *item;
2069         struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
2070         struct ceph_cap *cap;
2071         LIST_HEAD(tmp_list);
2072         int num_cap_releases;
2073         __le32  barrier, *cap_barrier;
2074
2075         down_read(&osdc->lock);
2076         barrier = cpu_to_le32(osdc->epoch_barrier);
2077         up_read(&osdc->lock);
2078
2079         spin_lock(&session->s_cap_lock);
2080 again:
2081         list_splice_init(&session->s_cap_releases, &tmp_list);
2082         num_cap_releases = session->s_num_cap_releases;
2083         session->s_num_cap_releases = 0;
2084         spin_unlock(&session->s_cap_lock);
2085
2086         while (!list_empty(&tmp_list)) {
2087                 if (!msg) {
2088                         msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
2089                                         PAGE_SIZE, GFP_NOFS, false);
2090                         if (!msg)
2091                                 goto out_err;
2092                         head = msg->front.iov_base;
2093                         head->num = cpu_to_le32(0);
2094                         msg->front.iov_len = sizeof(*head);
2095
2096                         msg->hdr.version = cpu_to_le16(2);
2097                         msg->hdr.compat_version = cpu_to_le16(1);
2098                 }
2099
2100                 cap = list_first_entry(&tmp_list, struct ceph_cap,
2101                                         session_caps);
2102                 list_del(&cap->session_caps);
2103                 num_cap_releases--;
2104
2105                 head = msg->front.iov_base;
2106                 put_unaligned_le32(get_unaligned_le32(&head->num) + 1,
2107                                    &head->num);
2108                 item = msg->front.iov_base + msg->front.iov_len;
2109                 item->ino = cpu_to_le64(cap->cap_ino);
2110                 item->cap_id = cpu_to_le64(cap->cap_id);
2111                 item->migrate_seq = cpu_to_le32(cap->mseq);
2112                 item->seq = cpu_to_le32(cap->issue_seq);
2113                 msg->front.iov_len += sizeof(*item);
2114
2115                 ceph_put_cap(mdsc, cap);
2116
2117                 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
2118                         // Append cap_barrier field
2119                         cap_barrier = msg->front.iov_base + msg->front.iov_len;
2120                         *cap_barrier = barrier;
2121                         msg->front.iov_len += sizeof(*cap_barrier);
2122
2123                         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2124                         dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
2125                         ceph_con_send(&session->s_con, msg);
2126                         msg = NULL;
2127                 }
2128         }
2129
2130         BUG_ON(num_cap_releases != 0);
2131
2132         spin_lock(&session->s_cap_lock);
2133         if (!list_empty(&session->s_cap_releases))
2134                 goto again;
2135         spin_unlock(&session->s_cap_lock);
2136
2137         if (msg) {
2138                 // Append cap_barrier field
2139                 cap_barrier = msg->front.iov_base + msg->front.iov_len;
2140                 *cap_barrier = barrier;
2141                 msg->front.iov_len += sizeof(*cap_barrier);
2142
2143                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2144                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
2145                 ceph_con_send(&session->s_con, msg);
2146         }
2147         return;
2148 out_err:
2149         pr_err("send_cap_releases mds%d, failed to allocate message\n",
2150                 session->s_mds);
2151         spin_lock(&session->s_cap_lock);
2152         list_splice(&tmp_list, &session->s_cap_releases);
2153         session->s_num_cap_releases += num_cap_releases;
2154         spin_unlock(&session->s_cap_lock);
2155 }
2156
2157 static void ceph_cap_release_work(struct work_struct *work)
2158 {
2159         struct ceph_mds_session *session =
2160                 container_of(work, struct ceph_mds_session, s_cap_release_work);
2161
2162         mutex_lock(&session->s_mutex);
2163         if (session->s_state == CEPH_MDS_SESSION_OPEN ||
2164             session->s_state == CEPH_MDS_SESSION_HUNG)
2165                 ceph_send_cap_releases(session->s_mdsc, session);
2166         mutex_unlock(&session->s_mutex);
2167         ceph_put_mds_session(session);
2168 }
2169
2170 void ceph_flush_cap_releases(struct ceph_mds_client *mdsc,
2171                              struct ceph_mds_session *session)
2172 {
2173         if (mdsc->stopping)
2174                 return;
2175
2176         ceph_get_mds_session(session);
2177         if (queue_work(mdsc->fsc->cap_wq,
2178                        &session->s_cap_release_work)) {
2179                 dout("cap release work queued\n");
2180         } else {
2181                 ceph_put_mds_session(session);
2182                 dout("failed to queue cap release work\n");
2183         }
2184 }
2185
2186 /*
2187  * caller holds session->s_cap_lock
2188  */
2189 void __ceph_queue_cap_release(struct ceph_mds_session *session,
2190                               struct ceph_cap *cap)
2191 {
2192         list_add_tail(&cap->session_caps, &session->s_cap_releases);
2193         session->s_num_cap_releases++;
2194
2195         if (!(session->s_num_cap_releases % CEPH_CAPS_PER_RELEASE))
2196                 ceph_flush_cap_releases(session->s_mdsc, session);
2197 }
2198
2199 static void ceph_cap_reclaim_work(struct work_struct *work)
2200 {
2201         struct ceph_mds_client *mdsc =
2202                 container_of(work, struct ceph_mds_client, cap_reclaim_work);
2203         int ret = ceph_trim_dentries(mdsc);
2204         if (ret == -EAGAIN)
2205                 ceph_queue_cap_reclaim_work(mdsc);
2206 }
2207
2208 void ceph_queue_cap_reclaim_work(struct ceph_mds_client *mdsc)
2209 {
2210         if (mdsc->stopping)
2211                 return;
2212
2213         if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_reclaim_work)) {
2214                 dout("caps reclaim work queued\n");
2215         } else {
2216                 dout("failed to queue caps release work\n");
2217         }
2218 }
2219
2220 void ceph_reclaim_caps_nr(struct ceph_mds_client *mdsc, int nr)
2221 {
2222         int val;
2223         if (!nr)
2224                 return;
2225         val = atomic_add_return(nr, &mdsc->cap_reclaim_pending);
2226         if ((val % CEPH_CAPS_PER_RELEASE) < nr) {
2227                 atomic_set(&mdsc->cap_reclaim_pending, 0);
2228                 ceph_queue_cap_reclaim_work(mdsc);
2229         }
2230 }
2231
2232 /*
2233  * requests
2234  */
2235
2236 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
2237                                     struct inode *dir)
2238 {
2239         struct ceph_inode_info *ci = ceph_inode(dir);
2240         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
2241         struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
2242         size_t size = sizeof(struct ceph_mds_reply_dir_entry);
2243         unsigned int num_entries;
2244         int order;
2245
2246         spin_lock(&ci->i_ceph_lock);
2247         num_entries = ci->i_files + ci->i_subdirs;
2248         spin_unlock(&ci->i_ceph_lock);
2249         num_entries = max(num_entries, 1U);
2250         num_entries = min(num_entries, opt->max_readdir);
2251
2252         order = get_order(size * num_entries);
2253         while (order >= 0) {
2254                 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
2255                                                              __GFP_NOWARN,
2256                                                              order);
2257                 if (rinfo->dir_entries)
2258                         break;
2259                 order--;
2260         }
2261         if (!rinfo->dir_entries)
2262                 return -ENOMEM;
2263
2264         num_entries = (PAGE_SIZE << order) / size;
2265         num_entries = min(num_entries, opt->max_readdir);
2266
2267         rinfo->dir_buf_size = PAGE_SIZE << order;
2268         req->r_num_caps = num_entries + 1;
2269         req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
2270         req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
2271         return 0;
2272 }
2273
2274 /*
2275  * Create an mds request.
2276  */
2277 struct ceph_mds_request *
2278 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
2279 {
2280         struct ceph_mds_request *req;
2281
2282         req = kmem_cache_zalloc(ceph_mds_request_cachep, GFP_NOFS);
2283         if (!req)
2284                 return ERR_PTR(-ENOMEM);
2285
2286         mutex_init(&req->r_fill_mutex);
2287         req->r_mdsc = mdsc;
2288         req->r_started = jiffies;
2289         req->r_start_latency = ktime_get();
2290         req->r_resend_mds = -1;
2291         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
2292         INIT_LIST_HEAD(&req->r_unsafe_target_item);
2293         req->r_fmode = -1;
2294         kref_init(&req->r_kref);
2295         RB_CLEAR_NODE(&req->r_node);
2296         INIT_LIST_HEAD(&req->r_wait);
2297         init_completion(&req->r_completion);
2298         init_completion(&req->r_safe_completion);
2299         INIT_LIST_HEAD(&req->r_unsafe_item);
2300
2301         ktime_get_coarse_real_ts64(&req->r_stamp);
2302
2303         req->r_op = op;
2304         req->r_direct_mode = mode;
2305         return req;
2306 }
2307
2308 /*
2309  * return oldest (lowest) request, tid in request tree, 0 if none.
2310  *
2311  * called under mdsc->mutex.
2312  */
2313 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
2314 {
2315         if (RB_EMPTY_ROOT(&mdsc->request_tree))
2316                 return NULL;
2317         return rb_entry(rb_first(&mdsc->request_tree),
2318                         struct ceph_mds_request, r_node);
2319 }
2320
2321 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
2322 {
2323         return mdsc->oldest_tid;
2324 }
2325
2326 /*
2327  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
2328  * on build_path_from_dentry in fs/cifs/dir.c.
2329  *
2330  * If @stop_on_nosnap, generate path relative to the first non-snapped
2331  * inode.
2332  *
2333  * Encode hidden .snap dirs as a double /, i.e.
2334  *   foo/.snap/bar -> foo//bar
2335  */
2336 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *pbase,
2337                            int stop_on_nosnap)
2338 {
2339         struct dentry *temp;
2340         char *path;
2341         int pos;
2342         unsigned seq;
2343         u64 base;
2344
2345         if (!dentry)
2346                 return ERR_PTR(-EINVAL);
2347
2348         path = __getname();
2349         if (!path)
2350                 return ERR_PTR(-ENOMEM);
2351 retry:
2352         pos = PATH_MAX - 1;
2353         path[pos] = '\0';
2354
2355         seq = read_seqbegin(&rename_lock);
2356         rcu_read_lock();
2357         temp = dentry;
2358         for (;;) {
2359                 struct inode *inode;
2360
2361                 spin_lock(&temp->d_lock);
2362                 inode = d_inode(temp);
2363                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
2364                         dout("build_path path+%d: %p SNAPDIR\n",
2365                              pos, temp);
2366                 } else if (stop_on_nosnap && inode && dentry != temp &&
2367                            ceph_snap(inode) == CEPH_NOSNAP) {
2368                         spin_unlock(&temp->d_lock);
2369                         pos++; /* get rid of any prepended '/' */
2370                         break;
2371                 } else {
2372                         pos -= temp->d_name.len;
2373                         if (pos < 0) {
2374                                 spin_unlock(&temp->d_lock);
2375                                 break;
2376                         }
2377                         memcpy(path + pos, temp->d_name.name, temp->d_name.len);
2378                 }
2379                 spin_unlock(&temp->d_lock);
2380                 temp = READ_ONCE(temp->d_parent);
2381
2382                 /* Are we at the root? */
2383                 if (IS_ROOT(temp))
2384                         break;
2385
2386                 /* Are we out of buffer? */
2387                 if (--pos < 0)
2388                         break;
2389
2390                 path[pos] = '/';
2391         }
2392         base = ceph_ino(d_inode(temp));
2393         rcu_read_unlock();
2394
2395         if (read_seqretry(&rename_lock, seq))
2396                 goto retry;
2397
2398         if (pos < 0) {
2399                 /*
2400                  * A rename didn't occur, but somehow we didn't end up where
2401                  * we thought we would. Throw a warning and try again.
2402                  */
2403                 pr_warn("build_path did not end path lookup where "
2404                         "expected, pos is %d\n", pos);
2405                 goto retry;
2406         }
2407
2408         *pbase = base;
2409         *plen = PATH_MAX - 1 - pos;
2410         dout("build_path on %p %d built %llx '%.*s'\n",
2411              dentry, d_count(dentry), base, *plen, path + pos);
2412         return path + pos;
2413 }
2414
2415 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
2416                              const char **ppath, int *ppathlen, u64 *pino,
2417                              bool *pfreepath, bool parent_locked)
2418 {
2419         char *path;
2420
2421         rcu_read_lock();
2422         if (!dir)
2423                 dir = d_inode_rcu(dentry->d_parent);
2424         if (dir && parent_locked && ceph_snap(dir) == CEPH_NOSNAP) {
2425                 *pino = ceph_ino(dir);
2426                 rcu_read_unlock();
2427                 *ppath = dentry->d_name.name;
2428                 *ppathlen = dentry->d_name.len;
2429                 return 0;
2430         }
2431         rcu_read_unlock();
2432         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2433         if (IS_ERR(path))
2434                 return PTR_ERR(path);
2435         *ppath = path;
2436         *pfreepath = true;
2437         return 0;
2438 }
2439
2440 static int build_inode_path(struct inode *inode,
2441                             const char **ppath, int *ppathlen, u64 *pino,
2442                             bool *pfreepath)
2443 {
2444         struct dentry *dentry;
2445         char *path;
2446
2447         if (ceph_snap(inode) == CEPH_NOSNAP) {
2448                 *pino = ceph_ino(inode);
2449                 *ppathlen = 0;
2450                 return 0;
2451         }
2452         dentry = d_find_alias(inode);
2453         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2454         dput(dentry);
2455         if (IS_ERR(path))
2456                 return PTR_ERR(path);
2457         *ppath = path;
2458         *pfreepath = true;
2459         return 0;
2460 }
2461
2462 /*
2463  * request arguments may be specified via an inode *, a dentry *, or
2464  * an explicit ino+path.
2465  */
2466 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
2467                                   struct inode *rdiri, const char *rpath,
2468                                   u64 rino, const char **ppath, int *pathlen,
2469                                   u64 *ino, bool *freepath, bool parent_locked)
2470 {
2471         int r = 0;
2472
2473         if (rinode) {
2474                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2475                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2476                      ceph_snap(rinode));
2477         } else if (rdentry) {
2478                 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
2479                                         freepath, parent_locked);
2480                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
2481                      *ppath);
2482         } else if (rpath || rino) {
2483                 *ino = rino;
2484                 *ppath = rpath;
2485                 *pathlen = rpath ? strlen(rpath) : 0;
2486                 dout(" path %.*s\n", *pathlen, rpath);
2487         }
2488
2489         return r;
2490 }
2491
2492 static void encode_timestamp_and_gids(void **p,
2493                                       const struct ceph_mds_request *req)
2494 {
2495         struct ceph_timespec ts;
2496         int i;
2497
2498         ceph_encode_timespec64(&ts, &req->r_stamp);
2499         ceph_encode_copy(p, &ts, sizeof(ts));
2500
2501         /* gid_list */
2502         ceph_encode_32(p, req->r_cred->group_info->ngroups);
2503         for (i = 0; i < req->r_cred->group_info->ngroups; i++)
2504                 ceph_encode_64(p, from_kgid(&init_user_ns,
2505                                             req->r_cred->group_info->gid[i]));
2506 }
2507
2508 /*
2509  * called under mdsc->mutex
2510  */
2511 static struct ceph_msg *create_request_message(struct ceph_mds_session *session,
2512                                                struct ceph_mds_request *req,
2513                                                bool drop_cap_releases)
2514 {
2515         int mds = session->s_mds;
2516         struct ceph_mds_client *mdsc = session->s_mdsc;
2517         struct ceph_msg *msg;
2518         struct ceph_mds_request_head_old *head;
2519         const char *path1 = NULL;
2520         const char *path2 = NULL;
2521         u64 ino1 = 0, ino2 = 0;
2522         int pathlen1 = 0, pathlen2 = 0;
2523         bool freepath1 = false, freepath2 = false;
2524         int len;
2525         u16 releases;
2526         void *p, *end;
2527         int ret;
2528         bool legacy = !(session->s_con.peer_features & CEPH_FEATURE_FS_BTIME);
2529
2530         ret = set_request_path_attr(req->r_inode, req->r_dentry,
2531                               req->r_parent, req->r_path1, req->r_ino1.ino,
2532                               &path1, &pathlen1, &ino1, &freepath1,
2533                               test_bit(CEPH_MDS_R_PARENT_LOCKED,
2534                                         &req->r_req_flags));
2535         if (ret < 0) {
2536                 msg = ERR_PTR(ret);
2537                 goto out;
2538         }
2539
2540         /* If r_old_dentry is set, then assume that its parent is locked */
2541         ret = set_request_path_attr(NULL, req->r_old_dentry,
2542                               req->r_old_dentry_dir,
2543                               req->r_path2, req->r_ino2.ino,
2544                               &path2, &pathlen2, &ino2, &freepath2, true);
2545         if (ret < 0) {
2546                 msg = ERR_PTR(ret);
2547                 goto out_free1;
2548         }
2549
2550         len = legacy ? sizeof(*head) : sizeof(struct ceph_mds_request_head);
2551         len += pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2552                 sizeof(struct ceph_timespec);
2553         len += sizeof(u32) + (sizeof(u64) * req->r_cred->group_info->ngroups);
2554
2555         /* calculate (max) length for cap releases */
2556         len += sizeof(struct ceph_mds_request_release) *
2557                 (!!req->r_inode_drop + !!req->r_dentry_drop +
2558                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2559
2560         if (req->r_dentry_drop)
2561                 len += pathlen1;
2562         if (req->r_old_dentry_drop)
2563                 len += pathlen2;
2564
2565         msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
2566         if (!msg) {
2567                 msg = ERR_PTR(-ENOMEM);
2568                 goto out_free2;
2569         }
2570
2571         msg->hdr.tid = cpu_to_le64(req->r_tid);
2572
2573         /*
2574          * The old ceph_mds_request_head didn't contain a version field, and
2575          * one was added when we moved the message version from 3->4.
2576          */
2577         if (legacy) {
2578                 msg->hdr.version = cpu_to_le16(3);
2579                 head = msg->front.iov_base;
2580                 p = msg->front.iov_base + sizeof(*head);
2581         } else {
2582                 struct ceph_mds_request_head *new_head = msg->front.iov_base;
2583
2584                 msg->hdr.version = cpu_to_le16(4);
2585                 new_head->version = cpu_to_le16(CEPH_MDS_REQUEST_HEAD_VERSION);
2586                 head = (struct ceph_mds_request_head_old *)&new_head->oldest_client_tid;
2587                 p = msg->front.iov_base + sizeof(*new_head);
2588         }
2589
2590         end = msg->front.iov_base + msg->front.iov_len;
2591
2592         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2593         head->op = cpu_to_le32(req->r_op);
2594         head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns,
2595                                                  req->r_cred->fsuid));
2596         head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns,
2597                                                  req->r_cred->fsgid));
2598         head->ino = cpu_to_le64(req->r_deleg_ino);
2599         head->args = req->r_args;
2600
2601         ceph_encode_filepath(&p, end, ino1, path1);
2602         ceph_encode_filepath(&p, end, ino2, path2);
2603
2604         /* make note of release offset, in case we need to replay */
2605         req->r_request_release_offset = p - msg->front.iov_base;
2606
2607         /* cap releases */
2608         releases = 0;
2609         if (req->r_inode_drop)
2610                 releases += ceph_encode_inode_release(&p,
2611                       req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2612                       mds, req->r_inode_drop, req->r_inode_unless,
2613                       req->r_op == CEPH_MDS_OP_READDIR);
2614         if (req->r_dentry_drop)
2615                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
2616                                 req->r_parent, mds, req->r_dentry_drop,
2617                                 req->r_dentry_unless);
2618         if (req->r_old_dentry_drop)
2619                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2620                                 req->r_old_dentry_dir, mds,
2621                                 req->r_old_dentry_drop,
2622                                 req->r_old_dentry_unless);
2623         if (req->r_old_inode_drop)
2624                 releases += ceph_encode_inode_release(&p,
2625                       d_inode(req->r_old_dentry),
2626                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2627
2628         if (drop_cap_releases) {
2629                 releases = 0;
2630                 p = msg->front.iov_base + req->r_request_release_offset;
2631         }
2632
2633         head->num_releases = cpu_to_le16(releases);
2634
2635         encode_timestamp_and_gids(&p, req);
2636
2637         if (WARN_ON_ONCE(p > end)) {
2638                 ceph_msg_put(msg);
2639                 msg = ERR_PTR(-ERANGE);
2640                 goto out_free2;
2641         }
2642
2643         msg->front.iov_len = p - msg->front.iov_base;
2644         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2645
2646         if (req->r_pagelist) {
2647                 struct ceph_pagelist *pagelist = req->r_pagelist;
2648                 ceph_msg_data_add_pagelist(msg, pagelist);
2649                 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2650         } else {
2651                 msg->hdr.data_len = 0;
2652         }
2653
2654         msg->hdr.data_off = cpu_to_le16(0);
2655
2656 out_free2:
2657         if (freepath2)
2658                 ceph_mdsc_free_path((char *)path2, pathlen2);
2659 out_free1:
2660         if (freepath1)
2661                 ceph_mdsc_free_path((char *)path1, pathlen1);
2662 out:
2663         return msg;
2664 }
2665
2666 /*
2667  * called under mdsc->mutex if error, under no mutex if
2668  * success.
2669  */
2670 static void complete_request(struct ceph_mds_client *mdsc,
2671                              struct ceph_mds_request *req)
2672 {
2673         req->r_end_latency = ktime_get();
2674
2675         if (req->r_callback)
2676                 req->r_callback(mdsc, req);
2677         complete_all(&req->r_completion);
2678 }
2679
2680 static struct ceph_mds_request_head_old *
2681 find_old_request_head(void *p, u64 features)
2682 {
2683         bool legacy = !(features & CEPH_FEATURE_FS_BTIME);
2684         struct ceph_mds_request_head *new_head;
2685
2686         if (legacy)
2687                 return (struct ceph_mds_request_head_old *)p;
2688         new_head = (struct ceph_mds_request_head *)p;
2689         return (struct ceph_mds_request_head_old *)&new_head->oldest_client_tid;
2690 }
2691
2692 /*
2693  * called under mdsc->mutex
2694  */
2695 static int __prepare_send_request(struct ceph_mds_session *session,
2696                                   struct ceph_mds_request *req,
2697                                   bool drop_cap_releases)
2698 {
2699         int mds = session->s_mds;
2700         struct ceph_mds_client *mdsc = session->s_mdsc;
2701         struct ceph_mds_request_head_old *rhead;
2702         struct ceph_msg *msg;
2703         int flags = 0;
2704
2705         req->r_attempts++;
2706         if (req->r_inode) {
2707                 struct ceph_cap *cap =
2708                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2709
2710                 if (cap)
2711                         req->r_sent_on_mseq = cap->mseq;
2712                 else
2713                         req->r_sent_on_mseq = -1;
2714         }
2715         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2716              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2717
2718         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2719                 void *p;
2720
2721                 /*
2722                  * Replay.  Do not regenerate message (and rebuild
2723                  * paths, etc.); just use the original message.
2724                  * Rebuilding paths will break for renames because
2725                  * d_move mangles the src name.
2726                  */
2727                 msg = req->r_request;
2728                 rhead = find_old_request_head(msg->front.iov_base,
2729                                               session->s_con.peer_features);
2730
2731                 flags = le32_to_cpu(rhead->flags);
2732                 flags |= CEPH_MDS_FLAG_REPLAY;
2733                 rhead->flags = cpu_to_le32(flags);
2734
2735                 if (req->r_target_inode)
2736                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2737
2738                 rhead->num_retry = req->r_attempts - 1;
2739
2740                 /* remove cap/dentry releases from message */
2741                 rhead->num_releases = 0;
2742
2743                 p = msg->front.iov_base + req->r_request_release_offset;
2744                 encode_timestamp_and_gids(&p, req);
2745
2746                 msg->front.iov_len = p - msg->front.iov_base;
2747                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2748                 return 0;
2749         }
2750
2751         if (req->r_request) {
2752                 ceph_msg_put(req->r_request);
2753                 req->r_request = NULL;
2754         }
2755         msg = create_request_message(session, req, drop_cap_releases);
2756         if (IS_ERR(msg)) {
2757                 req->r_err = PTR_ERR(msg);
2758                 return PTR_ERR(msg);
2759         }
2760         req->r_request = msg;
2761
2762         rhead = find_old_request_head(msg->front.iov_base,
2763                                       session->s_con.peer_features);
2764         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2765         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2766                 flags |= CEPH_MDS_FLAG_REPLAY;
2767         if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags))
2768                 flags |= CEPH_MDS_FLAG_ASYNC;
2769         if (req->r_parent)
2770                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2771         rhead->flags = cpu_to_le32(flags);
2772         rhead->num_fwd = req->r_num_fwd;
2773         rhead->num_retry = req->r_attempts - 1;
2774
2775         dout(" r_parent = %p\n", req->r_parent);
2776         return 0;
2777 }
2778
2779 /*
2780  * called under mdsc->mutex
2781  */
2782 static int __send_request(struct ceph_mds_session *session,
2783                           struct ceph_mds_request *req,
2784                           bool drop_cap_releases)
2785 {
2786         int err;
2787
2788         err = __prepare_send_request(session, req, drop_cap_releases);
2789         if (!err) {
2790                 ceph_msg_get(req->r_request);
2791                 ceph_con_send(&session->s_con, req->r_request);
2792         }
2793
2794         return err;
2795 }
2796
2797 /*
2798  * send request, or put it on the appropriate wait list.
2799  */
2800 static void __do_request(struct ceph_mds_client *mdsc,
2801                         struct ceph_mds_request *req)
2802 {
2803         struct ceph_mds_session *session = NULL;
2804         int mds = -1;
2805         int err = 0;
2806         bool random;
2807
2808         if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2809                 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2810                         __unregister_request(mdsc, req);
2811                 return;
2812         }
2813
2814         if (req->r_timeout &&
2815             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2816                 dout("do_request timed out\n");
2817                 err = -ETIMEDOUT;
2818                 goto finish;
2819         }
2820         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2821                 dout("do_request forced umount\n");
2822                 err = -EIO;
2823                 goto finish;
2824         }
2825         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2826                 if (mdsc->mdsmap_err) {
2827                         err = mdsc->mdsmap_err;
2828                         dout("do_request mdsmap err %d\n", err);
2829                         goto finish;
2830                 }
2831                 if (mdsc->mdsmap->m_epoch == 0) {
2832                         dout("do_request no mdsmap, waiting for map\n");
2833                         list_add(&req->r_wait, &mdsc->waiting_for_map);
2834                         return;
2835                 }
2836                 if (!(mdsc->fsc->mount_options->flags &
2837                       CEPH_MOUNT_OPT_MOUNTWAIT) &&
2838                     !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2839                         err = -EHOSTUNREACH;
2840                         goto finish;
2841                 }
2842         }
2843
2844         put_request_session(req);
2845
2846         mds = __choose_mds(mdsc, req, &random);
2847         if (mds < 0 ||
2848             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2849                 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2850                         err = -EJUKEBOX;
2851                         goto finish;
2852                 }
2853                 dout("do_request no mds or not active, waiting for map\n");
2854                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2855                 return;
2856         }
2857
2858         /* get, open session */
2859         session = __ceph_lookup_mds_session(mdsc, mds);
2860         if (!session) {
2861                 session = register_session(mdsc, mds);
2862                 if (IS_ERR(session)) {
2863                         err = PTR_ERR(session);
2864                         goto finish;
2865                 }
2866         }
2867         req->r_session = ceph_get_mds_session(session);
2868
2869         dout("do_request mds%d session %p state %s\n", mds, session,
2870              ceph_session_state_name(session->s_state));
2871         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2872             session->s_state != CEPH_MDS_SESSION_HUNG) {
2873                 /*
2874                  * We cannot queue async requests since the caps and delegated
2875                  * inodes are bound to the session. Just return -EJUKEBOX and
2876                  * let the caller retry a sync request in that case.
2877                  */
2878                 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2879                         err = -EJUKEBOX;
2880                         goto out_session;
2881                 }
2882
2883                 /*
2884                  * If the session has been REJECTED, then return a hard error,
2885                  * unless it's a CLEANRECOVER mount, in which case we'll queue
2886                  * it to the mdsc queue.
2887                  */
2888                 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2889                         if (ceph_test_mount_opt(mdsc->fsc, CLEANRECOVER))
2890                                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2891                         else
2892                                 err = -EACCES;
2893                         goto out_session;
2894                 }
2895
2896                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2897                     session->s_state == CEPH_MDS_SESSION_CLOSING) {
2898                         err = __open_session(mdsc, session);
2899                         if (err)
2900                                 goto out_session;
2901                         /* retry the same mds later */
2902                         if (random)
2903                                 req->r_resend_mds = mds;
2904                 }
2905                 list_add(&req->r_wait, &session->s_waiting);
2906                 goto out_session;
2907         }
2908
2909         /* send request */
2910         req->r_resend_mds = -1;   /* forget any previous mds hint */
2911
2912         if (req->r_request_started == 0)   /* note request start time */
2913                 req->r_request_started = jiffies;
2914
2915         err = __send_request(session, req, false);
2916
2917 out_session:
2918         ceph_put_mds_session(session);
2919 finish:
2920         if (err) {
2921                 dout("__do_request early error %d\n", err);
2922                 req->r_err = err;
2923                 complete_request(mdsc, req);
2924                 __unregister_request(mdsc, req);
2925         }
2926         return;
2927 }
2928
2929 /*
2930  * called under mdsc->mutex
2931  */
2932 static void __wake_requests(struct ceph_mds_client *mdsc,
2933                             struct list_head *head)
2934 {
2935         struct ceph_mds_request *req;
2936         LIST_HEAD(tmp_list);
2937
2938         list_splice_init(head, &tmp_list);
2939
2940         while (!list_empty(&tmp_list)) {
2941                 req = list_entry(tmp_list.next,
2942                                  struct ceph_mds_request, r_wait);
2943                 list_del_init(&req->r_wait);
2944                 dout(" wake request %p tid %llu\n", req, req->r_tid);
2945                 __do_request(mdsc, req);
2946         }
2947 }
2948
2949 /*
2950  * Wake up threads with requests pending for @mds, so that they can
2951  * resubmit their requests to a possibly different mds.
2952  */
2953 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2954 {
2955         struct ceph_mds_request *req;
2956         struct rb_node *p = rb_first(&mdsc->request_tree);
2957
2958         dout("kick_requests mds%d\n", mds);
2959         while (p) {
2960                 req = rb_entry(p, struct ceph_mds_request, r_node);
2961                 p = rb_next(p);
2962                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2963                         continue;
2964                 if (req->r_attempts > 0)
2965                         continue; /* only new requests */
2966                 if (req->r_session &&
2967                     req->r_session->s_mds == mds) {
2968                         dout(" kicking tid %llu\n", req->r_tid);
2969                         list_del_init(&req->r_wait);
2970                         __do_request(mdsc, req);
2971                 }
2972         }
2973 }
2974
2975 int ceph_mdsc_submit_request(struct ceph_mds_client *mdsc, struct inode *dir,
2976                               struct ceph_mds_request *req)
2977 {
2978         int err = 0;
2979
2980         /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2981         if (req->r_inode)
2982                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2983         if (req->r_parent) {
2984                 struct ceph_inode_info *ci = ceph_inode(req->r_parent);
2985                 int fmode = (req->r_op & CEPH_MDS_OP_WRITE) ?
2986                             CEPH_FILE_MODE_WR : CEPH_FILE_MODE_RD;
2987                 spin_lock(&ci->i_ceph_lock);
2988                 ceph_take_cap_refs(ci, CEPH_CAP_PIN, false);
2989                 __ceph_touch_fmode(ci, mdsc, fmode);
2990                 spin_unlock(&ci->i_ceph_lock);
2991                 ihold(req->r_parent);
2992         }
2993         if (req->r_old_dentry_dir)
2994                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2995                                   CEPH_CAP_PIN);
2996
2997         if (req->r_inode) {
2998                 err = ceph_wait_on_async_create(req->r_inode);
2999                 if (err) {
3000                         dout("%s: wait for async create returned: %d\n",
3001                              __func__, err);
3002                         return err;
3003                 }
3004         }
3005
3006         if (!err && req->r_old_inode) {
3007                 err = ceph_wait_on_async_create(req->r_old_inode);
3008                 if (err) {
3009                         dout("%s: wait for async create returned: %d\n",
3010                              __func__, err);
3011                         return err;
3012                 }
3013         }
3014
3015         dout("submit_request on %p for inode %p\n", req, dir);
3016         mutex_lock(&mdsc->mutex);
3017         __register_request(mdsc, req, dir);
3018         __do_request(mdsc, req);
3019         err = req->r_err;
3020         mutex_unlock(&mdsc->mutex);
3021         return err;
3022 }
3023
3024 static int ceph_mdsc_wait_request(struct ceph_mds_client *mdsc,
3025                                   struct ceph_mds_request *req)
3026 {
3027         int err;
3028
3029         /* wait */
3030         dout("do_request waiting\n");
3031         if (!req->r_timeout && req->r_wait_for_completion) {
3032                 err = req->r_wait_for_completion(mdsc, req);
3033         } else {
3034                 long timeleft = wait_for_completion_killable_timeout(
3035                                         &req->r_completion,
3036                                         ceph_timeout_jiffies(req->r_timeout));
3037                 if (timeleft > 0)
3038                         err = 0;
3039                 else if (!timeleft)
3040                         err = -ETIMEDOUT;  /* timed out */
3041                 else
3042                         err = timeleft;  /* killed */
3043         }
3044         dout("do_request waited, got %d\n", err);
3045         mutex_lock(&mdsc->mutex);
3046
3047         /* only abort if we didn't race with a real reply */
3048         if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
3049                 err = le32_to_cpu(req->r_reply_info.head->result);
3050         } else if (err < 0) {
3051                 dout("aborted request %lld with %d\n", req->r_tid, err);
3052
3053                 /*
3054                  * ensure we aren't running concurrently with
3055                  * ceph_fill_trace or ceph_readdir_prepopulate, which
3056                  * rely on locks (dir mutex) held by our caller.
3057                  */
3058                 mutex_lock(&req->r_fill_mutex);
3059                 req->r_err = err;
3060                 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
3061                 mutex_unlock(&req->r_fill_mutex);
3062
3063                 if (req->r_parent &&
3064                     (req->r_op & CEPH_MDS_OP_WRITE))
3065                         ceph_invalidate_dir_request(req);
3066         } else {
3067                 err = req->r_err;
3068         }
3069
3070         mutex_unlock(&mdsc->mutex);
3071         return err;
3072 }
3073
3074 /*
3075  * Synchrously perform an mds request.  Take care of all of the
3076  * session setup, forwarding, retry details.
3077  */
3078 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
3079                          struct inode *dir,
3080                          struct ceph_mds_request *req)
3081 {
3082         int err;
3083
3084         dout("do_request on %p\n", req);
3085
3086         /* issue */
3087         err = ceph_mdsc_submit_request(mdsc, dir, req);
3088         if (!err)
3089                 err = ceph_mdsc_wait_request(mdsc, req);
3090         dout("do_request %p done, result %d\n", req, err);
3091         return err;
3092 }
3093
3094 /*
3095  * Invalidate dir's completeness, dentry lease state on an aborted MDS
3096  * namespace request.
3097  */
3098 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
3099 {
3100         struct inode *dir = req->r_parent;
3101         struct inode *old_dir = req->r_old_dentry_dir;
3102
3103         dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
3104
3105         ceph_dir_clear_complete(dir);
3106         if (old_dir)
3107                 ceph_dir_clear_complete(old_dir);
3108         if (req->r_dentry)
3109                 ceph_invalidate_dentry_lease(req->r_dentry);
3110         if (req->r_old_dentry)
3111                 ceph_invalidate_dentry_lease(req->r_old_dentry);
3112 }
3113
3114 /*
3115  * Handle mds reply.
3116  *
3117  * We take the session mutex and parse and process the reply immediately.
3118  * This preserves the logical ordering of replies, capabilities, etc., sent
3119  * by the MDS as they are applied to our local cache.
3120  */
3121 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
3122 {
3123         struct ceph_mds_client *mdsc = session->s_mdsc;
3124         struct ceph_mds_request *req;
3125         struct ceph_mds_reply_head *head = msg->front.iov_base;
3126         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
3127         struct ceph_snap_realm *realm;
3128         u64 tid;
3129         int err, result;
3130         int mds = session->s_mds;
3131
3132         if (msg->front.iov_len < sizeof(*head)) {
3133                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
3134                 ceph_msg_dump(msg);
3135                 return;
3136         }
3137
3138         /* get request, session */
3139         tid = le64_to_cpu(msg->hdr.tid);
3140         mutex_lock(&mdsc->mutex);
3141         req = lookup_get_request(mdsc, tid);
3142         if (!req) {
3143                 dout("handle_reply on unknown tid %llu\n", tid);
3144                 mutex_unlock(&mdsc->mutex);
3145                 return;
3146         }
3147         dout("handle_reply %p\n", req);
3148
3149         /* correct session? */
3150         if (req->r_session != session) {
3151                 pr_err("mdsc_handle_reply got %llu on session mds%d"
3152                        " not mds%d\n", tid, session->s_mds,
3153                        req->r_session ? req->r_session->s_mds : -1);
3154                 mutex_unlock(&mdsc->mutex);
3155                 goto out;
3156         }
3157
3158         /* dup? */
3159         if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
3160             (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
3161                 pr_warn("got a dup %s reply on %llu from mds%d\n",
3162                            head->safe ? "safe" : "unsafe", tid, mds);
3163                 mutex_unlock(&mdsc->mutex);
3164                 goto out;
3165         }
3166         if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
3167                 pr_warn("got unsafe after safe on %llu from mds%d\n",
3168                            tid, mds);
3169                 mutex_unlock(&mdsc->mutex);
3170                 goto out;
3171         }
3172
3173         result = le32_to_cpu(head->result);
3174
3175         /*
3176          * Handle an ESTALE
3177          * if we're not talking to the authority, send to them
3178          * if the authority has changed while we weren't looking,
3179          * send to new authority
3180          * Otherwise we just have to return an ESTALE
3181          */
3182         if (result == -ESTALE) {
3183                 dout("got ESTALE on request %llu\n", req->r_tid);
3184                 req->r_resend_mds = -1;
3185                 if (req->r_direct_mode != USE_AUTH_MDS) {
3186                         dout("not using auth, setting for that now\n");
3187                         req->r_direct_mode = USE_AUTH_MDS;
3188                         __do_request(mdsc, req);
3189                         mutex_unlock(&mdsc->mutex);
3190                         goto out;
3191                 } else  {
3192                         int mds = __choose_mds(mdsc, req, NULL);
3193                         if (mds >= 0 && mds != req->r_session->s_mds) {
3194                                 dout("but auth changed, so resending\n");
3195                                 __do_request(mdsc, req);
3196                                 mutex_unlock(&mdsc->mutex);
3197                                 goto out;
3198                         }
3199                 }
3200                 dout("have to return ESTALE on request %llu\n", req->r_tid);
3201         }
3202
3203
3204         if (head->safe) {
3205                 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
3206                 __unregister_request(mdsc, req);
3207
3208                 /* last request during umount? */
3209                 if (mdsc->stopping && !__get_oldest_req(mdsc))
3210                         complete_all(&mdsc->safe_umount_waiters);
3211
3212                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3213                         /*
3214                          * We already handled the unsafe response, now do the
3215                          * cleanup.  No need to examine the response; the MDS
3216                          * doesn't include any result info in the safe
3217                          * response.  And even if it did, there is nothing
3218                          * useful we could do with a revised return value.
3219                          */
3220                         dout("got safe reply %llu, mds%d\n", tid, mds);
3221
3222                         mutex_unlock(&mdsc->mutex);
3223                         goto out;
3224                 }
3225         } else {
3226                 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
3227                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
3228         }
3229
3230         dout("handle_reply tid %lld result %d\n", tid, result);
3231         rinfo = &req->r_reply_info;
3232         if (test_bit(CEPHFS_FEATURE_REPLY_ENCODING, &session->s_features))
3233                 err = parse_reply_info(session, msg, rinfo, (u64)-1);
3234         else
3235                 err = parse_reply_info(session, msg, rinfo, session->s_con.peer_features);
3236         mutex_unlock(&mdsc->mutex);
3237
3238         /* Must find target inode outside of mutexes to avoid deadlocks */
3239         if ((err >= 0) && rinfo->head->is_target) {
3240                 struct inode *in;
3241                 struct ceph_vino tvino = {
3242                         .ino  = le64_to_cpu(rinfo->targeti.in->ino),
3243                         .snap = le64_to_cpu(rinfo->targeti.in->snapid)
3244                 };
3245
3246                 in = ceph_get_inode(mdsc->fsc->sb, tvino);
3247                 if (IS_ERR(in)) {
3248                         err = PTR_ERR(in);
3249                         mutex_lock(&session->s_mutex);
3250                         goto out_err;
3251                 }
3252                 req->r_target_inode = in;
3253         }
3254
3255         mutex_lock(&session->s_mutex);
3256         if (err < 0) {
3257                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
3258                 ceph_msg_dump(msg);
3259                 goto out_err;
3260         }
3261
3262         /* snap trace */
3263         realm = NULL;
3264         if (rinfo->snapblob_len) {
3265                 down_write(&mdsc->snap_rwsem);
3266                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
3267                                 rinfo->snapblob + rinfo->snapblob_len,
3268                                 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
3269                                 &realm);
3270                 downgrade_write(&mdsc->snap_rwsem);
3271         } else {
3272                 down_read(&mdsc->snap_rwsem);
3273         }
3274
3275         /* insert trace into our cache */
3276         mutex_lock(&req->r_fill_mutex);
3277         current->journal_info = req;
3278         err = ceph_fill_trace(mdsc->fsc->sb, req);
3279         if (err == 0) {
3280                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
3281                                     req->r_op == CEPH_MDS_OP_LSSNAP))
3282                         ceph_readdir_prepopulate(req, req->r_session);
3283         }
3284         current->journal_info = NULL;
3285         mutex_unlock(&req->r_fill_mutex);
3286
3287         up_read(&mdsc->snap_rwsem);
3288         if (realm)
3289                 ceph_put_snap_realm(mdsc, realm);
3290
3291         if (err == 0) {
3292                 if (req->r_target_inode &&
3293                     test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3294                         struct ceph_inode_info *ci =
3295                                 ceph_inode(req->r_target_inode);
3296                         spin_lock(&ci->i_unsafe_lock);
3297                         list_add_tail(&req->r_unsafe_target_item,
3298                                       &ci->i_unsafe_iops);
3299                         spin_unlock(&ci->i_unsafe_lock);
3300                 }
3301
3302                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
3303         }
3304 out_err:
3305         mutex_lock(&mdsc->mutex);
3306         if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3307                 if (err) {
3308                         req->r_err = err;
3309                 } else {
3310                         req->r_reply =  ceph_msg_get(msg);
3311                         set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
3312                 }
3313         } else {
3314                 dout("reply arrived after request %lld was aborted\n", tid);
3315         }
3316         mutex_unlock(&mdsc->mutex);
3317
3318         mutex_unlock(&session->s_mutex);
3319
3320         /* kick calling process */
3321         complete_request(mdsc, req);
3322
3323         ceph_update_metadata_metrics(&mdsc->metric, req->r_start_latency,
3324                                      req->r_end_latency, err);
3325 out:
3326         ceph_mdsc_put_request(req);
3327         return;
3328 }
3329
3330
3331
3332 /*
3333  * handle mds notification that our request has been forwarded.
3334  */
3335 static void handle_forward(struct ceph_mds_client *mdsc,
3336                            struct ceph_mds_session *session,
3337                            struct ceph_msg *msg)
3338 {
3339         struct ceph_mds_request *req;
3340         u64 tid = le64_to_cpu(msg->hdr.tid);
3341         u32 next_mds;
3342         u32 fwd_seq;
3343         int err = -EINVAL;
3344         void *p = msg->front.iov_base;
3345         void *end = p + msg->front.iov_len;
3346
3347         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
3348         next_mds = ceph_decode_32(&p);
3349         fwd_seq = ceph_decode_32(&p);
3350
3351         mutex_lock(&mdsc->mutex);
3352         req = lookup_get_request(mdsc, tid);
3353         if (!req) {
3354                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
3355                 goto out;  /* dup reply? */
3356         }
3357
3358         if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3359                 dout("forward tid %llu aborted, unregistering\n", tid);
3360                 __unregister_request(mdsc, req);
3361         } else if (fwd_seq <= req->r_num_fwd) {
3362                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
3363                      tid, next_mds, req->r_num_fwd, fwd_seq);
3364         } else {
3365                 /* resend. forward race not possible; mds would drop */
3366                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
3367                 BUG_ON(req->r_err);
3368                 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
3369                 req->r_attempts = 0;
3370                 req->r_num_fwd = fwd_seq;
3371                 req->r_resend_mds = next_mds;
3372                 put_request_session(req);
3373                 __do_request(mdsc, req);
3374         }
3375         ceph_mdsc_put_request(req);
3376 out:
3377         mutex_unlock(&mdsc->mutex);
3378         return;
3379
3380 bad:
3381         pr_err("mdsc_handle_forward decode error err=%d\n", err);
3382 }
3383
3384 static int __decode_session_metadata(void **p, void *end,
3385                                      bool *blocklisted)
3386 {
3387         /* map<string,string> */
3388         u32 n;
3389         bool err_str;
3390         ceph_decode_32_safe(p, end, n, bad);
3391         while (n-- > 0) {
3392                 u32 len;
3393                 ceph_decode_32_safe(p, end, len, bad);
3394                 ceph_decode_need(p, end, len, bad);
3395                 err_str = !strncmp(*p, "error_string", len);
3396                 *p += len;
3397                 ceph_decode_32_safe(p, end, len, bad);
3398                 ceph_decode_need(p, end, len, bad);
3399                 /*
3400                  * Match "blocklisted (blacklisted)" from newer MDSes,
3401                  * or "blacklisted" from older MDSes.
3402                  */
3403                 if (err_str && strnstr(*p, "blacklisted", len))
3404                         *blocklisted = true;
3405                 *p += len;
3406         }
3407         return 0;
3408 bad:
3409         return -1;
3410 }
3411
3412 /*
3413  * handle a mds session control message
3414  */
3415 static void handle_session(struct ceph_mds_session *session,
3416                            struct ceph_msg *msg)
3417 {
3418         struct ceph_mds_client *mdsc = session->s_mdsc;
3419         int mds = session->s_mds;
3420         int msg_version = le16_to_cpu(msg->hdr.version);
3421         void *p = msg->front.iov_base;
3422         void *end = p + msg->front.iov_len;
3423         struct ceph_mds_session_head *h;
3424         u32 op;
3425         u64 seq, features = 0;
3426         int wake = 0;
3427         bool blocklisted = false;
3428
3429         /* decode */
3430         ceph_decode_need(&p, end, sizeof(*h), bad);
3431         h = p;
3432         p += sizeof(*h);
3433
3434         op = le32_to_cpu(h->op);
3435         seq = le64_to_cpu(h->seq);
3436
3437         if (msg_version >= 3) {
3438                 u32 len;
3439                 /* version >= 2, metadata */
3440                 if (__decode_session_metadata(&p, end, &blocklisted) < 0)
3441                         goto bad;
3442                 /* version >= 3, feature bits */
3443                 ceph_decode_32_safe(&p, end, len, bad);
3444                 if (len) {
3445                         ceph_decode_64_safe(&p, end, features, bad);
3446                         p += len - sizeof(features);
3447                 }
3448         }
3449
3450         mutex_lock(&mdsc->mutex);
3451         if (op == CEPH_SESSION_CLOSE) {
3452                 ceph_get_mds_session(session);
3453                 __unregister_session(mdsc, session);
3454         }
3455         /* FIXME: this ttl calculation is generous */
3456         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
3457         mutex_unlock(&mdsc->mutex);
3458
3459         mutex_lock(&session->s_mutex);
3460
3461         dout("handle_session mds%d %s %p state %s seq %llu\n",
3462              mds, ceph_session_op_name(op), session,
3463              ceph_session_state_name(session->s_state), seq);
3464
3465         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
3466                 session->s_state = CEPH_MDS_SESSION_OPEN;
3467                 pr_info("mds%d came back\n", session->s_mds);
3468         }
3469
3470         switch (op) {
3471         case CEPH_SESSION_OPEN:
3472                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3473                         pr_info("mds%d reconnect success\n", session->s_mds);
3474                 session->s_state = CEPH_MDS_SESSION_OPEN;
3475                 session->s_features = features;
3476                 renewed_caps(mdsc, session, 0);
3477                 if (test_bit(CEPHFS_FEATURE_METRIC_COLLECT, &session->s_features))
3478                         metric_schedule_delayed(&mdsc->metric);
3479                 wake = 1;
3480                 if (mdsc->stopping)
3481                         __close_session(mdsc, session);
3482                 break;
3483
3484         case CEPH_SESSION_RENEWCAPS:
3485                 if (session->s_renew_seq == seq)
3486                         renewed_caps(mdsc, session, 1);
3487                 break;
3488
3489         case CEPH_SESSION_CLOSE:
3490                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3491                         pr_info("mds%d reconnect denied\n", session->s_mds);
3492                 session->s_state = CEPH_MDS_SESSION_CLOSED;
3493                 cleanup_session_requests(mdsc, session);
3494                 remove_session_caps(session);
3495                 wake = 2; /* for good measure */
3496                 wake_up_all(&mdsc->session_close_wq);
3497                 break;
3498
3499         case CEPH_SESSION_STALE:
3500                 pr_info("mds%d caps went stale, renewing\n",
3501                         session->s_mds);
3502                 spin_lock(&session->s_gen_ttl_lock);
3503                 session->s_cap_gen++;
3504                 session->s_cap_ttl = jiffies - 1;
3505                 spin_unlock(&session->s_gen_ttl_lock);
3506                 send_renew_caps(mdsc, session);
3507                 break;
3508
3509         case CEPH_SESSION_RECALL_STATE:
3510                 ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
3511                 break;
3512
3513         case CEPH_SESSION_FLUSHMSG:
3514                 send_flushmsg_ack(mdsc, session, seq);
3515                 break;
3516
3517         case CEPH_SESSION_FORCE_RO:
3518                 dout("force_session_readonly %p\n", session);
3519                 spin_lock(&session->s_cap_lock);
3520                 session->s_readonly = true;
3521                 spin_unlock(&session->s_cap_lock);
3522                 wake_up_session_caps(session, FORCE_RO);
3523                 break;
3524
3525         case CEPH_SESSION_REJECT:
3526                 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
3527                 pr_info("mds%d rejected session\n", session->s_mds);
3528                 session->s_state = CEPH_MDS_SESSION_REJECTED;
3529                 cleanup_session_requests(mdsc, session);
3530                 remove_session_caps(session);
3531                 if (blocklisted)
3532                         mdsc->fsc->blocklisted = true;
3533                 wake = 2; /* for good measure */
3534                 break;
3535
3536         default:
3537                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
3538                 WARN_ON(1);
3539         }
3540
3541         mutex_unlock(&session->s_mutex);
3542         if (wake) {
3543                 mutex_lock(&mdsc->mutex);
3544                 __wake_requests(mdsc, &session->s_waiting);
3545                 if (wake == 2)
3546                         kick_requests(mdsc, mds);
3547                 mutex_unlock(&mdsc->mutex);
3548         }
3549         if (op == CEPH_SESSION_CLOSE)
3550                 ceph_put_mds_session(session);
3551         return;
3552
3553 bad:
3554         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
3555                (int)msg->front.iov_len);
3556         ceph_msg_dump(msg);
3557         return;
3558 }
3559
3560 void ceph_mdsc_release_dir_caps(struct ceph_mds_request *req)
3561 {
3562         int dcaps;
3563
3564         dcaps = xchg(&req->r_dir_caps, 0);
3565         if (dcaps) {
3566                 dout("releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
3567                 ceph_put_cap_refs(ceph_inode(req->r_parent), dcaps);
3568         }
3569 }
3570
3571 void ceph_mdsc_release_dir_caps_no_check(struct ceph_mds_request *req)
3572 {
3573         int dcaps;
3574
3575         dcaps = xchg(&req->r_dir_caps, 0);
3576         if (dcaps) {
3577                 dout("releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
3578                 ceph_put_cap_refs_no_check_caps(ceph_inode(req->r_parent),
3579                                                 dcaps);
3580         }
3581 }
3582
3583 /*
3584  * called under session->mutex.
3585  */
3586 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
3587                                    struct ceph_mds_session *session)
3588 {
3589         struct ceph_mds_request *req, *nreq;
3590         struct rb_node *p;
3591
3592         dout("replay_unsafe_requests mds%d\n", session->s_mds);
3593
3594         mutex_lock(&mdsc->mutex);
3595         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item)
3596                 __send_request(session, req, true);
3597
3598         /*
3599          * also re-send old requests when MDS enters reconnect stage. So that MDS
3600          * can process completed request in clientreplay stage.
3601          */
3602         p = rb_first(&mdsc->request_tree);
3603         while (p) {
3604                 req = rb_entry(p, struct ceph_mds_request, r_node);
3605                 p = rb_next(p);
3606                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3607                         continue;
3608                 if (req->r_attempts == 0)
3609                         continue; /* only old requests */
3610                 if (!req->r_session)
3611                         continue;
3612                 if (req->r_session->s_mds != session->s_mds)
3613                         continue;
3614
3615                 ceph_mdsc_release_dir_caps_no_check(req);
3616
3617                 __send_request(session, req, true);
3618         }
3619         mutex_unlock(&mdsc->mutex);
3620 }
3621
3622 static int send_reconnect_partial(struct ceph_reconnect_state *recon_state)
3623 {
3624         struct ceph_msg *reply;
3625         struct ceph_pagelist *_pagelist;
3626         struct page *page;
3627         __le32 *addr;
3628         int err = -ENOMEM;
3629
3630         if (!recon_state->allow_multi)
3631                 return -ENOSPC;
3632
3633         /* can't handle message that contains both caps and realm */
3634         BUG_ON(!recon_state->nr_caps == !recon_state->nr_realms);
3635
3636         /* pre-allocate new pagelist */
3637         _pagelist = ceph_pagelist_alloc(GFP_NOFS);
3638         if (!_pagelist)
3639                 return -ENOMEM;
3640
3641         reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3642         if (!reply)
3643                 goto fail_msg;
3644
3645         /* placeholder for nr_caps */
3646         err = ceph_pagelist_encode_32(_pagelist, 0);
3647         if (err < 0)
3648                 goto fail;
3649
3650         if (recon_state->nr_caps) {
3651                 /* currently encoding caps */
3652                 err = ceph_pagelist_encode_32(recon_state->pagelist, 0);
3653                 if (err)
3654                         goto fail;
3655         } else {
3656                 /* placeholder for nr_realms (currently encoding relams) */
3657                 err = ceph_pagelist_encode_32(_pagelist, 0);
3658                 if (err < 0)
3659                         goto fail;
3660         }
3661
3662         err = ceph_pagelist_encode_8(recon_state->pagelist, 1);
3663         if (err)
3664                 goto fail;
3665
3666         page = list_first_entry(&recon_state->pagelist->head, struct page, lru);
3667         addr = kmap_atomic(page);
3668         if (recon_state->nr_caps) {
3669                 /* currently encoding caps */
3670                 *addr = cpu_to_le32(recon_state->nr_caps);
3671         } else {
3672                 /* currently encoding relams */
3673                 *(addr + 1) = cpu_to_le32(recon_state->nr_realms);
3674         }
3675         kunmap_atomic(addr);
3676
3677         reply->hdr.version = cpu_to_le16(5);
3678         reply->hdr.compat_version = cpu_to_le16(4);
3679
3680         reply->hdr.data_len = cpu_to_le32(recon_state->pagelist->length);
3681         ceph_msg_data_add_pagelist(reply, recon_state->pagelist);
3682
3683         ceph_con_send(&recon_state->session->s_con, reply);
3684         ceph_pagelist_release(recon_state->pagelist);
3685
3686         recon_state->pagelist = _pagelist;
3687         recon_state->nr_caps = 0;
3688         recon_state->nr_realms = 0;
3689         recon_state->msg_version = 5;
3690         return 0;
3691 fail:
3692         ceph_msg_put(reply);
3693 fail_msg:
3694         ceph_pagelist_release(_pagelist);
3695         return err;
3696 }
3697
3698 static struct dentry* d_find_primary(struct inode *inode)
3699 {
3700         struct dentry *alias, *dn = NULL;
3701
3702         if (hlist_empty(&inode->i_dentry))
3703                 return NULL;
3704
3705         spin_lock(&inode->i_lock);
3706         if (hlist_empty(&inode->i_dentry))
3707                 goto out_unlock;
3708
3709         if (S_ISDIR(inode->i_mode)) {
3710                 alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias);
3711                 if (!IS_ROOT(alias))
3712                         dn = dget(alias);
3713                 goto out_unlock;
3714         }
3715
3716         hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
3717                 spin_lock(&alias->d_lock);
3718                 if (!d_unhashed(alias) &&
3719                     (ceph_dentry(alias)->flags & CEPH_DENTRY_PRIMARY_LINK)) {
3720                         dn = dget_dlock(alias);
3721                 }
3722                 spin_unlock(&alias->d_lock);
3723                 if (dn)
3724                         break;
3725         }
3726 out_unlock:
3727         spin_unlock(&inode->i_lock);
3728         return dn;
3729 }
3730
3731 /*
3732  * Encode information about a cap for a reconnect with the MDS.
3733  */
3734 static int reconnect_caps_cb(struct inode *inode, struct ceph_cap *cap,
3735                           void *arg)
3736 {
3737         union {
3738                 struct ceph_mds_cap_reconnect v2;
3739                 struct ceph_mds_cap_reconnect_v1 v1;
3740         } rec;
3741         struct ceph_inode_info *ci = cap->ci;
3742         struct ceph_reconnect_state *recon_state = arg;
3743         struct ceph_pagelist *pagelist = recon_state->pagelist;
3744         struct dentry *dentry;
3745         char *path;
3746         int pathlen, err;
3747         u64 pathbase;
3748         u64 snap_follows;
3749
3750         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
3751              inode, ceph_vinop(inode), cap, cap->cap_id,
3752              ceph_cap_string(cap->issued));
3753
3754         dentry = d_find_primary(inode);
3755         if (dentry) {
3756                 /* set pathbase to parent dir when msg_version >= 2 */
3757                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase,
3758                                             recon_state->msg_version >= 2);
3759                 dput(dentry);
3760                 if (IS_ERR(path)) {
3761                         err = PTR_ERR(path);
3762                         goto out_err;
3763                 }
3764         } else {
3765                 path = NULL;
3766                 pathlen = 0;
3767                 pathbase = 0;
3768         }
3769
3770         spin_lock(&ci->i_ceph_lock);
3771         cap->seq = 0;        /* reset cap seq */
3772         cap->issue_seq = 0;  /* and issue_seq */
3773         cap->mseq = 0;       /* and migrate_seq */
3774         cap->cap_gen = cap->session->s_cap_gen;
3775
3776         /* These are lost when the session goes away */
3777         if (S_ISDIR(inode->i_mode)) {
3778                 if (cap->issued & CEPH_CAP_DIR_CREATE) {
3779                         ceph_put_string(rcu_dereference_raw(ci->i_cached_layout.pool_ns));
3780                         memset(&ci->i_cached_layout, 0, sizeof(ci->i_cached_layout));
3781                 }
3782                 cap->issued &= ~CEPH_CAP_ANY_DIR_OPS;
3783         }
3784
3785         if (recon_state->msg_version >= 2) {
3786                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
3787                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3788                 rec.v2.issued = cpu_to_le32(cap->issued);
3789                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3790                 rec.v2.pathbase = cpu_to_le64(pathbase);
3791                 rec.v2.flock_len = (__force __le32)
3792                         ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
3793         } else {
3794                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
3795                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3796                 rec.v1.issued = cpu_to_le32(cap->issued);
3797                 rec.v1.size = cpu_to_le64(i_size_read(inode));
3798                 ceph_encode_timespec64(&rec.v1.mtime, &inode->i_mtime);
3799                 ceph_encode_timespec64(&rec.v1.atime, &inode->i_atime);
3800                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3801                 rec.v1.pathbase = cpu_to_le64(pathbase);
3802         }
3803
3804         if (list_empty(&ci->i_cap_snaps)) {
3805                 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
3806         } else {
3807                 struct ceph_cap_snap *capsnap =
3808                         list_first_entry(&ci->i_cap_snaps,
3809                                          struct ceph_cap_snap, ci_item);
3810                 snap_follows = capsnap->follows;
3811         }
3812         spin_unlock(&ci->i_ceph_lock);
3813
3814         if (recon_state->msg_version >= 2) {
3815                 int num_fcntl_locks, num_flock_locks;
3816                 struct ceph_filelock *flocks = NULL;
3817                 size_t struct_len, total_len = sizeof(u64);
3818                 u8 struct_v = 0;
3819
3820 encode_again:
3821                 if (rec.v2.flock_len) {
3822                         ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
3823                 } else {
3824                         num_fcntl_locks = 0;
3825                         num_flock_locks = 0;
3826                 }
3827                 if (num_fcntl_locks + num_flock_locks > 0) {
3828                         flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
3829                                                sizeof(struct ceph_filelock),
3830                                                GFP_NOFS);
3831                         if (!flocks) {
3832                                 err = -ENOMEM;
3833                                 goto out_err;
3834                         }
3835                         err = ceph_encode_locks_to_buffer(inode, flocks,
3836                                                           num_fcntl_locks,
3837                                                           num_flock_locks);
3838                         if (err) {
3839                                 kfree(flocks);
3840                                 flocks = NULL;
3841                                 if (err == -ENOSPC)
3842                                         goto encode_again;
3843                                 goto out_err;
3844                         }
3845                 } else {
3846                         kfree(flocks);
3847                         flocks = NULL;
3848                 }
3849
3850                 if (recon_state->msg_version >= 3) {
3851                         /* version, compat_version and struct_len */
3852                         total_len += 2 * sizeof(u8) + sizeof(u32);
3853                         struct_v = 2;
3854                 }
3855                 /*
3856                  * number of encoded locks is stable, so copy to pagelist
3857                  */
3858                 struct_len = 2 * sizeof(u32) +
3859                             (num_fcntl_locks + num_flock_locks) *
3860                             sizeof(struct ceph_filelock);
3861                 rec.v2.flock_len = cpu_to_le32(struct_len);
3862
3863                 struct_len += sizeof(u32) + pathlen + sizeof(rec.v2);
3864
3865                 if (struct_v >= 2)
3866                         struct_len += sizeof(u64); /* snap_follows */
3867
3868                 total_len += struct_len;
3869
3870                 if (pagelist->length + total_len > RECONNECT_MAX_SIZE) {
3871                         err = send_reconnect_partial(recon_state);
3872                         if (err)
3873                                 goto out_freeflocks;
3874                         pagelist = recon_state->pagelist;
3875                 }
3876
3877                 err = ceph_pagelist_reserve(pagelist, total_len);
3878                 if (err)
3879                         goto out_freeflocks;
3880
3881                 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3882                 if (recon_state->msg_version >= 3) {
3883                         ceph_pagelist_encode_8(pagelist, struct_v);
3884                         ceph_pagelist_encode_8(pagelist, 1);
3885                         ceph_pagelist_encode_32(pagelist, struct_len);
3886                 }
3887                 ceph_pagelist_encode_string(pagelist, path, pathlen);
3888                 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
3889                 ceph_locks_to_pagelist(flocks, pagelist,
3890                                        num_fcntl_locks, num_flock_locks);
3891                 if (struct_v >= 2)
3892                         ceph_pagelist_encode_64(pagelist, snap_follows);
3893 out_freeflocks:
3894                 kfree(flocks);
3895         } else {
3896                 err = ceph_pagelist_reserve(pagelist,
3897                                             sizeof(u64) + sizeof(u32) +
3898                                             pathlen + sizeof(rec.v1));
3899                 if (err)
3900                         goto out_err;
3901
3902                 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3903                 ceph_pagelist_encode_string(pagelist, path, pathlen);
3904                 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
3905         }
3906
3907 out_err:
3908         ceph_mdsc_free_path(path, pathlen);
3909         if (!err)
3910                 recon_state->nr_caps++;
3911         return err;
3912 }
3913
3914 static int encode_snap_realms(struct ceph_mds_client *mdsc,
3915                               struct ceph_reconnect_state *recon_state)
3916 {
3917         struct rb_node *p;
3918         struct ceph_pagelist *pagelist = recon_state->pagelist;
3919         int err = 0;
3920
3921         if (recon_state->msg_version >= 4) {
3922                 err = ceph_pagelist_encode_32(pagelist, mdsc->num_snap_realms);
3923                 if (err < 0)
3924                         goto fail;
3925         }
3926
3927         /*
3928          * snaprealms.  we provide mds with the ino, seq (version), and
3929          * parent for all of our realms.  If the mds has any newer info,
3930          * it will tell us.
3931          */
3932         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3933                 struct ceph_snap_realm *realm =
3934                        rb_entry(p, struct ceph_snap_realm, node);
3935                 struct ceph_mds_snaprealm_reconnect sr_rec;
3936
3937                 if (recon_state->msg_version >= 4) {
3938                         size_t need = sizeof(u8) * 2 + sizeof(u32) +
3939                                       sizeof(sr_rec);
3940
3941                         if (pagelist->length + need > RECONNECT_MAX_SIZE) {
3942                                 err = send_reconnect_partial(recon_state);
3943                                 if (err)
3944                                         goto fail;
3945                                 pagelist = recon_state->pagelist;
3946                         }
3947
3948                         err = ceph_pagelist_reserve(pagelist, need);
3949                         if (err)
3950                                 goto fail;
3951
3952                         ceph_pagelist_encode_8(pagelist, 1);
3953                         ceph_pagelist_encode_8(pagelist, 1);
3954                         ceph_pagelist_encode_32(pagelist, sizeof(sr_rec));
3955                 }
3956
3957                 dout(" adding snap realm %llx seq %lld parent %llx\n",
3958                      realm->ino, realm->seq, realm->parent_ino);
3959                 sr_rec.ino = cpu_to_le64(realm->ino);
3960                 sr_rec.seq = cpu_to_le64(realm->seq);
3961                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3962
3963                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3964                 if (err)
3965                         goto fail;
3966
3967                 recon_state->nr_realms++;
3968         }
3969 fail:
3970         return err;
3971 }
3972
3973
3974 /*
3975  * If an MDS fails and recovers, clients need to reconnect in order to
3976  * reestablish shared state.  This includes all caps issued through
3977  * this session _and_ the snap_realm hierarchy.  Because it's not
3978  * clear which snap realms the mds cares about, we send everything we
3979  * know about.. that ensures we'll then get any new info the
3980  * recovering MDS might have.
3981  *
3982  * This is a relatively heavyweight operation, but it's rare.
3983  */
3984 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3985                                struct ceph_mds_session *session)
3986 {
3987         struct ceph_msg *reply;
3988         int mds = session->s_mds;
3989         int err = -ENOMEM;
3990         struct ceph_reconnect_state recon_state = {
3991                 .session = session,
3992         };
3993         LIST_HEAD(dispose);
3994
3995         pr_info("mds%d reconnect start\n", mds);
3996
3997         recon_state.pagelist = ceph_pagelist_alloc(GFP_NOFS);
3998         if (!recon_state.pagelist)
3999                 goto fail_nopagelist;
4000
4001         reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
4002         if (!reply)
4003                 goto fail_nomsg;
4004
4005         xa_destroy(&session->s_delegated_inos);
4006
4007         mutex_lock(&session->s_mutex);
4008         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
4009         session->s_seq = 0;
4010
4011         dout("session %p state %s\n", session,
4012              ceph_session_state_name(session->s_state));
4013
4014         spin_lock(&session->s_gen_ttl_lock);
4015         session->s_cap_gen++;
4016         spin_unlock(&session->s_gen_ttl_lock);
4017
4018         spin_lock(&session->s_cap_lock);
4019         /* don't know if session is readonly */
4020         session->s_readonly = 0;
4021         /*
4022          * notify __ceph_remove_cap() that we are composing cap reconnect.
4023          * If a cap get released before being added to the cap reconnect,
4024          * __ceph_remove_cap() should skip queuing cap release.
4025          */
4026         session->s_cap_reconnect = 1;
4027         /* drop old cap expires; we're about to reestablish that state */
4028         detach_cap_releases(session, &dispose);
4029         spin_unlock(&session->s_cap_lock);
4030         dispose_cap_releases(mdsc, &dispose);
4031
4032         /* trim unused caps to reduce MDS's cache rejoin time */
4033         if (mdsc->fsc->sb->s_root)
4034                 shrink_dcache_parent(mdsc->fsc->sb->s_root);
4035
4036         ceph_con_close(&session->s_con);
4037         ceph_con_open(&session->s_con,
4038                       CEPH_ENTITY_TYPE_MDS, mds,
4039                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
4040
4041         /* replay unsafe requests */
4042         replay_unsafe_requests(mdsc, session);
4043
4044         ceph_early_kick_flushing_caps(mdsc, session);
4045
4046         down_read(&mdsc->snap_rwsem);
4047
4048         /* placeholder for nr_caps */
4049         err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
4050         if (err)
4051                 goto fail;
4052
4053         if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
4054                 recon_state.msg_version = 3;
4055                 recon_state.allow_multi = true;
4056         } else if (session->s_con.peer_features & CEPH_FEATURE_MDSENC) {
4057                 recon_state.msg_version = 3;
4058         } else {
4059                 recon_state.msg_version = 2;
4060         }
4061         /* trsaverse this session's caps */
4062         err = ceph_iterate_session_caps(session, reconnect_caps_cb, &recon_state);
4063
4064         spin_lock(&session->s_cap_lock);
4065         session->s_cap_reconnect = 0;
4066         spin_unlock(&session->s_cap_lock);
4067
4068         if (err < 0)
4069                 goto fail;
4070
4071         /* check if all realms can be encoded into current message */
4072         if (mdsc->num_snap_realms) {
4073                 size_t total_len =
4074                         recon_state.pagelist->length +
4075                         mdsc->num_snap_realms *
4076                         sizeof(struct ceph_mds_snaprealm_reconnect);
4077                 if (recon_state.msg_version >= 4) {
4078                         /* number of realms */
4079                         total_len += sizeof(u32);
4080                         /* version, compat_version and struct_len */
4081                         total_len += mdsc->num_snap_realms *
4082                                      (2 * sizeof(u8) + sizeof(u32));
4083                 }
4084                 if (total_len > RECONNECT_MAX_SIZE) {
4085                         if (!recon_state.allow_multi) {
4086                                 err = -ENOSPC;
4087                                 goto fail;
4088                         }
4089                         if (recon_state.nr_caps) {
4090                                 err = send_reconnect_partial(&recon_state);
4091                                 if (err)
4092                                         goto fail;
4093                         }
4094                         recon_state.msg_version = 5;
4095                 }
4096         }
4097
4098         err = encode_snap_realms(mdsc, &recon_state);
4099         if (err < 0)
4100                 goto fail;
4101
4102         if (recon_state.msg_version >= 5) {
4103                 err = ceph_pagelist_encode_8(recon_state.pagelist, 0);
4104                 if (err < 0)
4105                         goto fail;
4106         }
4107
4108         if (recon_state.nr_caps || recon_state.nr_realms) {
4109                 struct page *page =
4110                         list_first_entry(&recon_state.pagelist->head,
4111                                         struct page, lru);
4112                 __le32 *addr = kmap_atomic(page);
4113                 if (recon_state.nr_caps) {
4114                         WARN_ON(recon_state.nr_realms != mdsc->num_snap_realms);
4115                         *addr = cpu_to_le32(recon_state.nr_caps);
4116                 } else if (recon_state.msg_version >= 4) {
4117                         *(addr + 1) = cpu_to_le32(recon_state.nr_realms);
4118                 }
4119                 kunmap_atomic(addr);
4120         }
4121
4122         reply->hdr.version = cpu_to_le16(recon_state.msg_version);
4123         if (recon_state.msg_version >= 4)
4124                 reply->hdr.compat_version = cpu_to_le16(4);
4125
4126         reply->hdr.data_len = cpu_to_le32(recon_state.pagelist->length);
4127         ceph_msg_data_add_pagelist(reply, recon_state.pagelist);
4128
4129         ceph_con_send(&session->s_con, reply);
4130
4131         mutex_unlock(&session->s_mutex);
4132
4133         mutex_lock(&mdsc->mutex);
4134         __wake_requests(mdsc, &session->s_waiting);
4135         mutex_unlock(&mdsc->mutex);
4136
4137         up_read(&mdsc->snap_rwsem);
4138         ceph_pagelist_release(recon_state.pagelist);
4139         return;
4140
4141 fail:
4142         ceph_msg_put(reply);
4143         up_read(&mdsc->snap_rwsem);
4144         mutex_unlock(&session->s_mutex);
4145 fail_nomsg:
4146         ceph_pagelist_release(recon_state.pagelist);
4147 fail_nopagelist:
4148         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
4149         return;
4150 }
4151
4152
4153 /*
4154  * compare old and new mdsmaps, kicking requests
4155  * and closing out old connections as necessary
4156  *
4157  * called under mdsc->mutex.
4158  */
4159 static void check_new_map(struct ceph_mds_client *mdsc,
4160                           struct ceph_mdsmap *newmap,
4161                           struct ceph_mdsmap *oldmap)
4162 {
4163         int i;
4164         int oldstate, newstate;
4165         struct ceph_mds_session *s;
4166
4167         dout("check_new_map new %u old %u\n",
4168              newmap->m_epoch, oldmap->m_epoch);
4169
4170         for (i = 0; i < oldmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4171                 if (!mdsc->sessions[i])
4172                         continue;
4173                 s = mdsc->sessions[i];
4174                 oldstate = ceph_mdsmap_get_state(oldmap, i);
4175                 newstate = ceph_mdsmap_get_state(newmap, i);
4176
4177                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
4178                      i, ceph_mds_state_name(oldstate),
4179                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
4180                      ceph_mds_state_name(newstate),
4181                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
4182                      ceph_session_state_name(s->s_state));
4183
4184                 if (i >= newmap->possible_max_rank) {
4185                         /* force close session for stopped mds */
4186                         ceph_get_mds_session(s);
4187                         __unregister_session(mdsc, s);
4188                         __wake_requests(mdsc, &s->s_waiting);
4189                         mutex_unlock(&mdsc->mutex);
4190
4191                         mutex_lock(&s->s_mutex);
4192                         cleanup_session_requests(mdsc, s);
4193                         remove_session_caps(s);
4194                         mutex_unlock(&s->s_mutex);
4195
4196                         ceph_put_mds_session(s);
4197
4198                         mutex_lock(&mdsc->mutex);
4199                         kick_requests(mdsc, i);
4200                         continue;
4201                 }
4202
4203                 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
4204                            ceph_mdsmap_get_addr(newmap, i),
4205                            sizeof(struct ceph_entity_addr))) {
4206                         /* just close it */
4207                         mutex_unlock(&mdsc->mutex);
4208                         mutex_lock(&s->s_mutex);
4209                         mutex_lock(&mdsc->mutex);
4210                         ceph_con_close(&s->s_con);
4211                         mutex_unlock(&s->s_mutex);
4212                         s->s_state = CEPH_MDS_SESSION_RESTARTING;
4213                 } else if (oldstate == newstate) {
4214                         continue;  /* nothing new with this mds */
4215                 }
4216
4217                 /*
4218                  * send reconnect?
4219                  */
4220                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
4221                     newstate >= CEPH_MDS_STATE_RECONNECT) {
4222                         mutex_unlock(&mdsc->mutex);
4223                         send_mds_reconnect(mdsc, s);
4224                         mutex_lock(&mdsc->mutex);
4225                 }
4226
4227                 /*
4228                  * kick request on any mds that has gone active.
4229                  */
4230                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
4231                     newstate >= CEPH_MDS_STATE_ACTIVE) {
4232                         if (oldstate != CEPH_MDS_STATE_CREATING &&
4233                             oldstate != CEPH_MDS_STATE_STARTING)
4234                                 pr_info("mds%d recovery completed\n", s->s_mds);
4235                         kick_requests(mdsc, i);
4236                         mutex_unlock(&mdsc->mutex);
4237                         mutex_lock(&s->s_mutex);
4238                         mutex_lock(&mdsc->mutex);
4239                         ceph_kick_flushing_caps(mdsc, s);
4240                         mutex_unlock(&s->s_mutex);
4241                         wake_up_session_caps(s, RECONNECT);
4242                 }
4243         }
4244
4245         for (i = 0; i < newmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4246                 s = mdsc->sessions[i];
4247                 if (!s)
4248                         continue;
4249                 if (!ceph_mdsmap_is_laggy(newmap, i))
4250                         continue;
4251                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4252                     s->s_state == CEPH_MDS_SESSION_HUNG ||
4253                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
4254                         dout(" connecting to export targets of laggy mds%d\n",
4255                              i);
4256                         __open_export_target_sessions(mdsc, s);
4257                 }
4258         }
4259 }
4260
4261
4262
4263 /*
4264  * leases
4265  */
4266
4267 /*
4268  * caller must hold session s_mutex, dentry->d_lock
4269  */
4270 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
4271 {
4272         struct ceph_dentry_info *di = ceph_dentry(dentry);
4273
4274         ceph_put_mds_session(di->lease_session);
4275         di->lease_session = NULL;
4276 }
4277
4278 static void handle_lease(struct ceph_mds_client *mdsc,
4279                          struct ceph_mds_session *session,
4280                          struct ceph_msg *msg)
4281 {
4282         struct super_block *sb = mdsc->fsc->sb;
4283         struct inode *inode;
4284         struct dentry *parent, *dentry;
4285         struct ceph_dentry_info *di;
4286         int mds = session->s_mds;
4287         struct ceph_mds_lease *h = msg->front.iov_base;
4288         u32 seq;
4289         struct ceph_vino vino;
4290         struct qstr dname;
4291         int release = 0;
4292
4293         dout("handle_lease from mds%d\n", mds);
4294
4295         /* decode */
4296         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
4297                 goto bad;
4298         vino.ino = le64_to_cpu(h->ino);
4299         vino.snap = CEPH_NOSNAP;
4300         seq = le32_to_cpu(h->seq);
4301         dname.len = get_unaligned_le32(h + 1);
4302         if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
4303                 goto bad;
4304         dname.name = (void *)(h + 1) + sizeof(u32);
4305
4306         /* lookup inode */
4307         inode = ceph_find_inode(sb, vino);
4308         dout("handle_lease %s, ino %llx %p %.*s\n",
4309              ceph_lease_op_name(h->action), vino.ino, inode,
4310              dname.len, dname.name);
4311
4312         mutex_lock(&session->s_mutex);
4313         inc_session_sequence(session);
4314
4315         if (!inode) {
4316                 dout("handle_lease no inode %llx\n", vino.ino);
4317                 goto release;
4318         }
4319
4320         /* dentry */
4321         parent = d_find_alias(inode);
4322         if (!parent) {
4323                 dout("no parent dentry on inode %p\n", inode);
4324                 WARN_ON(1);
4325                 goto release;  /* hrm... */
4326         }
4327         dname.hash = full_name_hash(parent, dname.name, dname.len);
4328         dentry = d_lookup(parent, &dname);
4329         dput(parent);
4330         if (!dentry)
4331                 goto release;
4332
4333         spin_lock(&dentry->d_lock);
4334         di = ceph_dentry(dentry);
4335         switch (h->action) {
4336         case CEPH_MDS_LEASE_REVOKE:
4337                 if (di->lease_session == session) {
4338                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
4339                                 h->seq = cpu_to_le32(di->lease_seq);
4340                         __ceph_mdsc_drop_dentry_lease(dentry);
4341                 }
4342                 release = 1;
4343                 break;
4344
4345         case CEPH_MDS_LEASE_RENEW:
4346                 if (di->lease_session == session &&
4347                     di->lease_gen == session->s_cap_gen &&
4348                     di->lease_renew_from &&
4349                     di->lease_renew_after == 0) {
4350                         unsigned long duration =
4351                                 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
4352
4353                         di->lease_seq = seq;
4354                         di->time = di->lease_renew_from + duration;
4355                         di->lease_renew_after = di->lease_renew_from +
4356                                 (duration >> 1);
4357                         di->lease_renew_from = 0;
4358                 }
4359                 break;
4360         }
4361         spin_unlock(&dentry->d_lock);
4362         dput(dentry);
4363
4364         if (!release)
4365                 goto out;
4366
4367 release:
4368         /* let's just reuse the same message */
4369         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
4370         ceph_msg_get(msg);
4371         ceph_con_send(&session->s_con, msg);
4372
4373 out:
4374         mutex_unlock(&session->s_mutex);
4375         /* avoid calling iput_final() in mds dispatch threads */
4376         ceph_async_iput(inode);
4377         return;
4378
4379 bad:
4380         pr_err("corrupt lease message\n");
4381         ceph_msg_dump(msg);
4382 }
4383
4384 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
4385                               struct dentry *dentry, char action,
4386                               u32 seq)
4387 {
4388         struct ceph_msg *msg;
4389         struct ceph_mds_lease *lease;
4390         struct inode *dir;
4391         int len = sizeof(*lease) + sizeof(u32) + NAME_MAX;
4392
4393         dout("lease_send_msg identry %p %s to mds%d\n",
4394              dentry, ceph_lease_op_name(action), session->s_mds);
4395
4396         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
4397         if (!msg)
4398                 return;
4399         lease = msg->front.iov_base;
4400         lease->action = action;
4401         lease->seq = cpu_to_le32(seq);
4402
4403         spin_lock(&dentry->d_lock);
4404         dir = d_inode(dentry->d_parent);
4405         lease->ino = cpu_to_le64(ceph_ino(dir));
4406         lease->first = lease->last = cpu_to_le64(ceph_snap(dir));
4407
4408         put_unaligned_le32(dentry->d_name.len, lease + 1);
4409         memcpy((void *)(lease + 1) + 4,
4410                dentry->d_name.name, dentry->d_name.len);
4411         spin_unlock(&dentry->d_lock);
4412         /*
4413          * if this is a preemptive lease RELEASE, no need to
4414          * flush request stream, since the actual request will
4415          * soon follow.
4416          */
4417         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
4418
4419         ceph_con_send(&session->s_con, msg);
4420 }
4421
4422 /*
4423  * lock unlock sessions, to wait ongoing session activities
4424  */
4425 static void lock_unlock_sessions(struct ceph_mds_client *mdsc)
4426 {
4427         int i;
4428
4429         mutex_lock(&mdsc->mutex);
4430         for (i = 0; i < mdsc->max_sessions; i++) {
4431                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4432                 if (!s)
4433                         continue;
4434                 mutex_unlock(&mdsc->mutex);
4435                 mutex_lock(&s->s_mutex);
4436                 mutex_unlock(&s->s_mutex);
4437                 ceph_put_mds_session(s);
4438                 mutex_lock(&mdsc->mutex);
4439         }
4440         mutex_unlock(&mdsc->mutex);
4441 }
4442
4443 static void maybe_recover_session(struct ceph_mds_client *mdsc)
4444 {
4445         struct ceph_fs_client *fsc = mdsc->fsc;
4446
4447         if (!ceph_test_mount_opt(fsc, CLEANRECOVER))
4448                 return;
4449
4450         if (READ_ONCE(fsc->mount_state) != CEPH_MOUNT_MOUNTED)
4451                 return;
4452
4453         if (!READ_ONCE(fsc->blocklisted))
4454                 return;
4455
4456         pr_info("auto reconnect after blocklisted\n");
4457         ceph_force_reconnect(fsc->sb);
4458 }
4459
4460 bool check_session_state(struct ceph_mds_session *s)
4461 {
4462         switch (s->s_state) {
4463         case CEPH_MDS_SESSION_OPEN:
4464                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
4465                         s->s_state = CEPH_MDS_SESSION_HUNG;
4466                         pr_info("mds%d hung\n", s->s_mds);
4467                 }
4468                 break;
4469         case CEPH_MDS_SESSION_CLOSING:
4470                 /* Should never reach this when we're unmounting */
4471                 WARN_ON_ONCE(true);
4472                 fallthrough;
4473         case CEPH_MDS_SESSION_NEW:
4474         case CEPH_MDS_SESSION_RESTARTING:
4475         case CEPH_MDS_SESSION_CLOSED:
4476         case CEPH_MDS_SESSION_REJECTED:
4477                 return false;
4478         }
4479
4480         return true;
4481 }
4482
4483 /*
4484  * If the sequence is incremented while we're waiting on a REQUEST_CLOSE reply,
4485  * then we need to retransmit that request.
4486  */
4487 void inc_session_sequence(struct ceph_mds_session *s)
4488 {
4489         lockdep_assert_held(&s->s_mutex);
4490
4491         s->s_seq++;
4492
4493         if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
4494                 int ret;
4495
4496                 dout("resending session close request for mds%d\n", s->s_mds);
4497                 ret = request_close_session(s);
4498                 if (ret < 0)
4499                         pr_err("unable to close session to mds%d: %d\n",
4500                                s->s_mds, ret);
4501         }
4502 }
4503
4504 /*
4505  * delayed work -- periodically trim expired leases, renew caps with mds
4506  */
4507 static void schedule_delayed(struct ceph_mds_client *mdsc)
4508 {
4509         int delay = 5;
4510         unsigned hz = round_jiffies_relative(HZ * delay);
4511         schedule_delayed_work(&mdsc->delayed_work, hz);
4512 }
4513
4514 static void delayed_work(struct work_struct *work)
4515 {
4516         int i;
4517         struct ceph_mds_client *mdsc =
4518                 container_of(work, struct ceph_mds_client, delayed_work.work);
4519         int renew_interval;
4520         int renew_caps;
4521
4522         dout("mdsc delayed_work\n");
4523
4524         if (mdsc->stopping)
4525                 return;
4526
4527         mutex_lock(&mdsc->mutex);
4528         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
4529         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
4530                                    mdsc->last_renew_caps);
4531         if (renew_caps)
4532                 mdsc->last_renew_caps = jiffies;
4533
4534         for (i = 0; i < mdsc->max_sessions; i++) {
4535                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4536                 if (!s)
4537                         continue;
4538
4539                 if (!check_session_state(s)) {
4540                         ceph_put_mds_session(s);
4541                         continue;
4542                 }
4543                 mutex_unlock(&mdsc->mutex);
4544
4545                 mutex_lock(&s->s_mutex);
4546                 if (renew_caps)
4547                         send_renew_caps(mdsc, s);
4548                 else
4549                         ceph_con_keepalive(&s->s_con);
4550                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4551                     s->s_state == CEPH_MDS_SESSION_HUNG)
4552                         ceph_send_cap_releases(mdsc, s);
4553                 mutex_unlock(&s->s_mutex);
4554                 ceph_put_mds_session(s);
4555
4556                 mutex_lock(&mdsc->mutex);
4557         }
4558         mutex_unlock(&mdsc->mutex);
4559
4560         ceph_check_delayed_caps(mdsc);
4561
4562         ceph_queue_cap_reclaim_work(mdsc);
4563
4564         ceph_trim_snapid_map(mdsc);
4565
4566         maybe_recover_session(mdsc);
4567
4568         schedule_delayed(mdsc);
4569 }
4570
4571 int ceph_mdsc_init(struct ceph_fs_client *fsc)
4572
4573 {
4574         struct ceph_mds_client *mdsc;
4575         int err;
4576
4577         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
4578         if (!mdsc)
4579                 return -ENOMEM;
4580         mdsc->fsc = fsc;
4581         mutex_init(&mdsc->mutex);
4582         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
4583         if (!mdsc->mdsmap) {
4584                 err = -ENOMEM;
4585                 goto err_mdsc;
4586         }
4587
4588         init_completion(&mdsc->safe_umount_waiters);
4589         init_waitqueue_head(&mdsc->session_close_wq);
4590         INIT_LIST_HEAD(&mdsc->waiting_for_map);
4591         mdsc->sessions = NULL;
4592         atomic_set(&mdsc->num_sessions, 0);
4593         mdsc->max_sessions = 0;
4594         mdsc->stopping = 0;
4595         atomic64_set(&mdsc->quotarealms_count, 0);
4596         mdsc->quotarealms_inodes = RB_ROOT;
4597         mutex_init(&mdsc->quotarealms_inodes_mutex);
4598         mdsc->last_snap_seq = 0;
4599         init_rwsem(&mdsc->snap_rwsem);
4600         mdsc->snap_realms = RB_ROOT;
4601         INIT_LIST_HEAD(&mdsc->snap_empty);
4602         mdsc->num_snap_realms = 0;
4603         spin_lock_init(&mdsc->snap_empty_lock);
4604         mdsc->last_tid = 0;
4605         mdsc->oldest_tid = 0;
4606         mdsc->request_tree = RB_ROOT;
4607         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
4608         mdsc->last_renew_caps = jiffies;
4609         INIT_LIST_HEAD(&mdsc->cap_delay_list);
4610         INIT_LIST_HEAD(&mdsc->cap_wait_list);
4611         spin_lock_init(&mdsc->cap_delay_lock);
4612         INIT_LIST_HEAD(&mdsc->snap_flush_list);
4613         spin_lock_init(&mdsc->snap_flush_lock);
4614         mdsc->last_cap_flush_tid = 1;
4615         INIT_LIST_HEAD(&mdsc->cap_flush_list);
4616         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
4617         mdsc->num_cap_flushing = 0;
4618         spin_lock_init(&mdsc->cap_dirty_lock);
4619         init_waitqueue_head(&mdsc->cap_flushing_wq);
4620         INIT_WORK(&mdsc->cap_reclaim_work, ceph_cap_reclaim_work);
4621         atomic_set(&mdsc->cap_reclaim_pending, 0);
4622         err = ceph_metric_init(&mdsc->metric);
4623         if (err)
4624                 goto err_mdsmap;
4625
4626         spin_lock_init(&mdsc->dentry_list_lock);
4627         INIT_LIST_HEAD(&mdsc->dentry_leases);
4628         INIT_LIST_HEAD(&mdsc->dentry_dir_leases);
4629
4630         ceph_caps_init(mdsc);
4631         ceph_adjust_caps_max_min(mdsc, fsc->mount_options);
4632
4633         spin_lock_init(&mdsc->snapid_map_lock);
4634         mdsc->snapid_map_tree = RB_ROOT;
4635         INIT_LIST_HEAD(&mdsc->snapid_map_lru);
4636
4637         init_rwsem(&mdsc->pool_perm_rwsem);
4638         mdsc->pool_perm_tree = RB_ROOT;
4639
4640         strscpy(mdsc->nodename, utsname()->nodename,
4641                 sizeof(mdsc->nodename));
4642
4643         fsc->mdsc = mdsc;
4644         return 0;
4645
4646 err_mdsmap:
4647         kfree(mdsc->mdsmap);
4648 err_mdsc:
4649         kfree(mdsc);
4650         return err;
4651 }
4652
4653 /*
4654  * Wait for safe replies on open mds requests.  If we time out, drop
4655  * all requests from the tree to avoid dangling dentry refs.
4656  */
4657 static void wait_requests(struct ceph_mds_client *mdsc)
4658 {
4659         struct ceph_options *opts = mdsc->fsc->client->options;
4660         struct ceph_mds_request *req;
4661
4662         mutex_lock(&mdsc->mutex);
4663         if (__get_oldest_req(mdsc)) {
4664                 mutex_unlock(&mdsc->mutex);
4665
4666                 dout("wait_requests waiting for requests\n");
4667                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
4668                                     ceph_timeout_jiffies(opts->mount_timeout));
4669
4670                 /* tear down remaining requests */
4671                 mutex_lock(&mdsc->mutex);
4672                 while ((req = __get_oldest_req(mdsc))) {
4673                         dout("wait_requests timed out on tid %llu\n",
4674                              req->r_tid);
4675                         list_del_init(&req->r_wait);
4676                         __unregister_request(mdsc, req);
4677                 }
4678         }
4679         mutex_unlock(&mdsc->mutex);
4680         dout("wait_requests done\n");
4681 }
4682
4683 /*
4684  * called before mount is ro, and before dentries are torn down.
4685  * (hmm, does this still race with new lookups?)
4686  */
4687 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
4688 {
4689         dout("pre_umount\n");
4690         mdsc->stopping = 1;
4691
4692         lock_unlock_sessions(mdsc);
4693         ceph_flush_dirty_caps(mdsc);
4694         wait_requests(mdsc);
4695
4696         /*
4697          * wait for reply handlers to drop their request refs and
4698          * their inode/dcache refs
4699          */
4700         ceph_msgr_flush();
4701
4702         ceph_cleanup_quotarealms_inodes(mdsc);
4703 }
4704
4705 /*
4706  * wait for all write mds requests to flush.
4707  */
4708 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
4709 {
4710         struct ceph_mds_request *req = NULL, *nextreq;
4711         struct rb_node *n;
4712
4713         mutex_lock(&mdsc->mutex);
4714         dout("wait_unsafe_requests want %lld\n", want_tid);
4715 restart:
4716         req = __get_oldest_req(mdsc);
4717         while (req && req->r_tid <= want_tid) {
4718                 /* find next request */
4719                 n = rb_next(&req->r_node);
4720                 if (n)
4721                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
4722                 else
4723                         nextreq = NULL;
4724                 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
4725                     (req->r_op & CEPH_MDS_OP_WRITE)) {
4726                         /* write op */
4727                         ceph_mdsc_get_request(req);
4728                         if (nextreq)
4729                                 ceph_mdsc_get_request(nextreq);
4730                         mutex_unlock(&mdsc->mutex);
4731                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
4732                              req->r_tid, want_tid);
4733                         wait_for_completion(&req->r_safe_completion);
4734                         mutex_lock(&mdsc->mutex);
4735                         ceph_mdsc_put_request(req);
4736                         if (!nextreq)
4737                                 break;  /* next dne before, so we're done! */
4738                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
4739                                 /* next request was removed from tree */
4740                                 ceph_mdsc_put_request(nextreq);
4741                                 goto restart;
4742                         }
4743                         ceph_mdsc_put_request(nextreq);  /* won't go away */
4744                 }
4745                 req = nextreq;
4746         }
4747         mutex_unlock(&mdsc->mutex);
4748         dout("wait_unsafe_requests done\n");
4749 }
4750
4751 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
4752 {
4753         u64 want_tid, want_flush;
4754
4755         if (READ_ONCE(mdsc->fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN)
4756                 return;
4757
4758         dout("sync\n");
4759         mutex_lock(&mdsc->mutex);
4760         want_tid = mdsc->last_tid;
4761         mutex_unlock(&mdsc->mutex);
4762
4763         ceph_flush_dirty_caps(mdsc);
4764         spin_lock(&mdsc->cap_dirty_lock);
4765         want_flush = mdsc->last_cap_flush_tid;
4766         if (!list_empty(&mdsc->cap_flush_list)) {
4767                 struct ceph_cap_flush *cf =
4768                         list_last_entry(&mdsc->cap_flush_list,
4769                                         struct ceph_cap_flush, g_list);
4770                 cf->wake = true;
4771         }
4772         spin_unlock(&mdsc->cap_dirty_lock);
4773
4774         dout("sync want tid %lld flush_seq %lld\n",
4775              want_tid, want_flush);
4776
4777         wait_unsafe_requests(mdsc, want_tid);
4778         wait_caps_flush(mdsc, want_flush);
4779 }
4780
4781 /*
4782  * true if all sessions are closed, or we force unmount
4783  */
4784 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
4785 {
4786         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4787                 return true;
4788         return atomic_read(&mdsc->num_sessions) <= skipped;
4789 }
4790
4791 /*
4792  * called after sb is ro.
4793  */
4794 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
4795 {
4796         struct ceph_options *opts = mdsc->fsc->client->options;
4797         struct ceph_mds_session *session;
4798         int i;
4799         int skipped = 0;
4800
4801         dout("close_sessions\n");
4802
4803         /* close sessions */
4804         mutex_lock(&mdsc->mutex);
4805         for (i = 0; i < mdsc->max_sessions; i++) {
4806                 session = __ceph_lookup_mds_session(mdsc, i);
4807                 if (!session)
4808                         continue;
4809                 mutex_unlock(&mdsc->mutex);
4810                 mutex_lock(&session->s_mutex);
4811                 if (__close_session(mdsc, session) <= 0)
4812                         skipped++;
4813                 mutex_unlock(&session->s_mutex);
4814                 ceph_put_mds_session(session);
4815                 mutex_lock(&mdsc->mutex);
4816         }
4817         mutex_unlock(&mdsc->mutex);
4818
4819         dout("waiting for sessions to close\n");
4820         wait_event_timeout(mdsc->session_close_wq,
4821                            done_closing_sessions(mdsc, skipped),
4822                            ceph_timeout_jiffies(opts->mount_timeout));
4823
4824         /* tear down remaining sessions */
4825         mutex_lock(&mdsc->mutex);
4826         for (i = 0; i < mdsc->max_sessions; i++) {
4827                 if (mdsc->sessions[i]) {
4828                         session = ceph_get_mds_session(mdsc->sessions[i]);
4829                         __unregister_session(mdsc, session);
4830                         mutex_unlock(&mdsc->mutex);
4831                         mutex_lock(&session->s_mutex);
4832                         remove_session_caps(session);
4833                         mutex_unlock(&session->s_mutex);
4834                         ceph_put_mds_session(session);
4835                         mutex_lock(&mdsc->mutex);
4836                 }
4837         }
4838         WARN_ON(!list_empty(&mdsc->cap_delay_list));
4839         mutex_unlock(&mdsc->mutex);
4840
4841         ceph_cleanup_snapid_map(mdsc);
4842         ceph_cleanup_empty_realms(mdsc);
4843
4844         cancel_work_sync(&mdsc->cap_reclaim_work);
4845         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4846
4847         dout("stopped\n");
4848 }
4849
4850 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
4851 {
4852         struct ceph_mds_session *session;
4853         int mds;
4854
4855         dout("force umount\n");
4856
4857         mutex_lock(&mdsc->mutex);
4858         for (mds = 0; mds < mdsc->max_sessions; mds++) {
4859                 session = __ceph_lookup_mds_session(mdsc, mds);
4860                 if (!session)
4861                         continue;
4862
4863                 if (session->s_state == CEPH_MDS_SESSION_REJECTED)
4864                         __unregister_session(mdsc, session);
4865                 __wake_requests(mdsc, &session->s_waiting);
4866                 mutex_unlock(&mdsc->mutex);
4867
4868                 mutex_lock(&session->s_mutex);
4869                 __close_session(mdsc, session);
4870                 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
4871                         cleanup_session_requests(mdsc, session);
4872                         remove_session_caps(session);
4873                 }
4874                 mutex_unlock(&session->s_mutex);
4875                 ceph_put_mds_session(session);
4876
4877                 mutex_lock(&mdsc->mutex);
4878                 kick_requests(mdsc, mds);
4879         }
4880         __wake_requests(mdsc, &mdsc->waiting_for_map);
4881         mutex_unlock(&mdsc->mutex);
4882 }
4883
4884 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
4885 {
4886         dout("stop\n");
4887         /*
4888          * Make sure the delayed work stopped before releasing
4889          * the resources.
4890          *
4891          * Because the cancel_delayed_work_sync() will only
4892          * guarantee that the work finishes executing. But the
4893          * delayed work will re-arm itself again after that.
4894          */
4895         flush_delayed_work(&mdsc->delayed_work);
4896
4897         if (mdsc->mdsmap)
4898                 ceph_mdsmap_destroy(mdsc->mdsmap);
4899         kfree(mdsc->sessions);
4900         ceph_caps_finalize(mdsc);
4901         ceph_pool_perm_destroy(mdsc);
4902 }
4903
4904 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
4905 {
4906         struct ceph_mds_client *mdsc = fsc->mdsc;
4907         dout("mdsc_destroy %p\n", mdsc);
4908
4909         if (!mdsc)
4910                 return;
4911
4912         /* flush out any connection work with references to us */
4913         ceph_msgr_flush();
4914
4915         ceph_mdsc_stop(mdsc);
4916
4917         ceph_metric_destroy(&mdsc->metric);
4918
4919         flush_delayed_work(&mdsc->metric.delayed_work);
4920         fsc->mdsc = NULL;
4921         kfree(mdsc);
4922         dout("mdsc_destroy %p done\n", mdsc);
4923 }
4924
4925 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4926 {
4927         struct ceph_fs_client *fsc = mdsc->fsc;
4928         const char *mds_namespace = fsc->mount_options->mds_namespace;
4929         void *p = msg->front.iov_base;
4930         void *end = p + msg->front.iov_len;
4931         u32 epoch;
4932         u32 num_fs;
4933         u32 mount_fscid = (u32)-1;
4934         int err = -EINVAL;
4935
4936         ceph_decode_need(&p, end, sizeof(u32), bad);
4937         epoch = ceph_decode_32(&p);
4938
4939         dout("handle_fsmap epoch %u\n", epoch);
4940
4941         /* struct_v, struct_cv, map_len, epoch, legacy_client_fscid */
4942         ceph_decode_skip_n(&p, end, 2 + sizeof(u32) * 3, bad);
4943
4944         ceph_decode_32_safe(&p, end, num_fs, bad);
4945         while (num_fs-- > 0) {
4946                 void *info_p, *info_end;
4947                 u32 info_len;
4948                 u32 fscid, namelen;
4949
4950                 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4951                 p += 2;         // info_v, info_cv
4952                 info_len = ceph_decode_32(&p);
4953                 ceph_decode_need(&p, end, info_len, bad);
4954                 info_p = p;
4955                 info_end = p + info_len;
4956                 p = info_end;
4957
4958                 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
4959                 fscid = ceph_decode_32(&info_p);
4960                 namelen = ceph_decode_32(&info_p);
4961                 ceph_decode_need(&info_p, info_end, namelen, bad);
4962
4963                 if (mds_namespace &&
4964                     strlen(mds_namespace) == namelen &&
4965                     !strncmp(mds_namespace, (char *)info_p, namelen)) {
4966                         mount_fscid = fscid;
4967                         break;
4968                 }
4969         }
4970
4971         ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
4972         if (mount_fscid != (u32)-1) {
4973                 fsc->client->monc.fs_cluster_id = mount_fscid;
4974                 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
4975                                    0, true);
4976                 ceph_monc_renew_subs(&fsc->client->monc);
4977         } else {
4978                 err = -ENOENT;
4979                 goto err_out;
4980         }
4981         return;
4982
4983 bad:
4984         pr_err("error decoding fsmap\n");
4985 err_out:
4986         mutex_lock(&mdsc->mutex);
4987         mdsc->mdsmap_err = err;
4988         __wake_requests(mdsc, &mdsc->waiting_for_map);
4989         mutex_unlock(&mdsc->mutex);
4990 }
4991
4992 /*
4993  * handle mds map update.
4994  */
4995 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4996 {
4997         u32 epoch;
4998         u32 maplen;
4999         void *p = msg->front.iov_base;
5000         void *end = p + msg->front.iov_len;
5001         struct ceph_mdsmap *newmap, *oldmap;
5002         struct ceph_fsid fsid;
5003         int err = -EINVAL;
5004
5005         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
5006         ceph_decode_copy(&p, &fsid, sizeof(fsid));
5007         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
5008                 return;
5009         epoch = ceph_decode_32(&p);
5010         maplen = ceph_decode_32(&p);
5011         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
5012
5013         /* do we need it? */
5014         mutex_lock(&mdsc->mutex);
5015         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
5016                 dout("handle_map epoch %u <= our %u\n",
5017                      epoch, mdsc->mdsmap->m_epoch);
5018                 mutex_unlock(&mdsc->mutex);
5019                 return;
5020         }
5021
5022         newmap = ceph_mdsmap_decode(&p, end, ceph_msgr2(mdsc->fsc->client));
5023         if (IS_ERR(newmap)) {
5024                 err = PTR_ERR(newmap);
5025                 goto bad_unlock;
5026         }
5027
5028         /* swap into place */
5029         if (mdsc->mdsmap) {
5030                 oldmap = mdsc->mdsmap;
5031                 mdsc->mdsmap = newmap;
5032                 check_new_map(mdsc, newmap, oldmap);
5033                 ceph_mdsmap_destroy(oldmap);
5034         } else {
5035                 mdsc->mdsmap = newmap;  /* first mds map */
5036         }
5037         mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
5038                                         MAX_LFS_FILESIZE);
5039
5040         __wake_requests(mdsc, &mdsc->waiting_for_map);
5041         ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
5042                           mdsc->mdsmap->m_epoch);
5043
5044         mutex_unlock(&mdsc->mutex);
5045         schedule_delayed(mdsc);
5046         return;
5047
5048 bad_unlock:
5049         mutex_unlock(&mdsc->mutex);
5050 bad:
5051         pr_err("error decoding mdsmap %d\n", err);
5052         return;
5053 }
5054
5055 static struct ceph_connection *mds_get_con(struct ceph_connection *con)
5056 {
5057         struct ceph_mds_session *s = con->private;
5058
5059         if (ceph_get_mds_session(s))
5060                 return con;
5061         return NULL;
5062 }
5063
5064 static void mds_put_con(struct ceph_connection *con)
5065 {
5066         struct ceph_mds_session *s = con->private;
5067
5068         ceph_put_mds_session(s);
5069 }
5070
5071 /*
5072  * if the client is unresponsive for long enough, the mds will kill
5073  * the session entirely.
5074  */
5075 static void mds_peer_reset(struct ceph_connection *con)
5076 {
5077         struct ceph_mds_session *s = con->private;
5078         struct ceph_mds_client *mdsc = s->s_mdsc;
5079
5080         pr_warn("mds%d closed our session\n", s->s_mds);
5081         send_mds_reconnect(mdsc, s);
5082 }
5083
5084 static void mds_dispatch(struct ceph_connection *con, struct ceph_msg *msg)
5085 {
5086         struct ceph_mds_session *s = con->private;
5087         struct ceph_mds_client *mdsc = s->s_mdsc;
5088         int type = le16_to_cpu(msg->hdr.type);
5089
5090         mutex_lock(&mdsc->mutex);
5091         if (__verify_registered_session(mdsc, s) < 0) {
5092                 mutex_unlock(&mdsc->mutex);
5093                 goto out;
5094         }
5095         mutex_unlock(&mdsc->mutex);
5096
5097         switch (type) {
5098         case CEPH_MSG_MDS_MAP:
5099                 ceph_mdsc_handle_mdsmap(mdsc, msg);
5100                 break;
5101         case CEPH_MSG_FS_MAP_USER:
5102                 ceph_mdsc_handle_fsmap(mdsc, msg);
5103                 break;
5104         case CEPH_MSG_CLIENT_SESSION:
5105                 handle_session(s, msg);
5106                 break;
5107         case CEPH_MSG_CLIENT_REPLY:
5108                 handle_reply(s, msg);
5109                 break;
5110         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
5111                 handle_forward(mdsc, s, msg);
5112                 break;
5113         case CEPH_MSG_CLIENT_CAPS:
5114                 ceph_handle_caps(s, msg);
5115                 break;
5116         case CEPH_MSG_CLIENT_SNAP:
5117                 ceph_handle_snap(mdsc, s, msg);
5118                 break;
5119         case CEPH_MSG_CLIENT_LEASE:
5120                 handle_lease(mdsc, s, msg);
5121                 break;
5122         case CEPH_MSG_CLIENT_QUOTA:
5123                 ceph_handle_quota(mdsc, s, msg);
5124                 break;
5125
5126         default:
5127                 pr_err("received unknown message type %d %s\n", type,
5128                        ceph_msg_type_name(type));
5129         }
5130 out:
5131         ceph_msg_put(msg);
5132 }
5133
5134 /*
5135  * authentication
5136  */
5137
5138 /*
5139  * Note: returned pointer is the address of a structure that's
5140  * managed separately.  Caller must *not* attempt to free it.
5141  */
5142 static struct ceph_auth_handshake *
5143 mds_get_authorizer(struct ceph_connection *con, int *proto, int force_new)
5144 {
5145         struct ceph_mds_session *s = con->private;
5146         struct ceph_mds_client *mdsc = s->s_mdsc;
5147         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5148         struct ceph_auth_handshake *auth = &s->s_auth;
5149         int ret;
5150
5151         ret = __ceph_auth_get_authorizer(ac, auth, CEPH_ENTITY_TYPE_MDS,
5152                                          force_new, proto, NULL, NULL);
5153         if (ret)
5154                 return ERR_PTR(ret);
5155
5156         return auth;
5157 }
5158
5159 static int mds_add_authorizer_challenge(struct ceph_connection *con,
5160                                     void *challenge_buf, int challenge_buf_len)
5161 {
5162         struct ceph_mds_session *s = con->private;
5163         struct ceph_mds_client *mdsc = s->s_mdsc;
5164         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5165
5166         return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
5167                                             challenge_buf, challenge_buf_len);
5168 }
5169
5170 static int mds_verify_authorizer_reply(struct ceph_connection *con)
5171 {
5172         struct ceph_mds_session *s = con->private;
5173         struct ceph_mds_client *mdsc = s->s_mdsc;
5174         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5175         struct ceph_auth_handshake *auth = &s->s_auth;
5176
5177         return ceph_auth_verify_authorizer_reply(ac, auth->authorizer,
5178                 auth->authorizer_reply_buf, auth->authorizer_reply_buf_len,
5179                 NULL, NULL, NULL, NULL);
5180 }
5181
5182 static int mds_invalidate_authorizer(struct ceph_connection *con)
5183 {
5184         struct ceph_mds_session *s = con->private;
5185         struct ceph_mds_client *mdsc = s->s_mdsc;
5186         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5187
5188         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
5189
5190         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
5191 }
5192
5193 static int mds_get_auth_request(struct ceph_connection *con,
5194                                 void *buf, int *buf_len,
5195                                 void **authorizer, int *authorizer_len)
5196 {
5197         struct ceph_mds_session *s = con->private;
5198         struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
5199         struct ceph_auth_handshake *auth = &s->s_auth;
5200         int ret;
5201
5202         ret = ceph_auth_get_authorizer(ac, auth, CEPH_ENTITY_TYPE_MDS,
5203                                        buf, buf_len);
5204         if (ret)
5205                 return ret;
5206
5207         *authorizer = auth->authorizer_buf;
5208         *authorizer_len = auth->authorizer_buf_len;
5209         return 0;
5210 }
5211
5212 static int mds_handle_auth_reply_more(struct ceph_connection *con,
5213                                       void *reply, int reply_len,
5214                                       void *buf, int *buf_len,
5215                                       void **authorizer, int *authorizer_len)
5216 {
5217         struct ceph_mds_session *s = con->private;
5218         struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
5219         struct ceph_auth_handshake *auth = &s->s_auth;
5220         int ret;
5221
5222         ret = ceph_auth_handle_svc_reply_more(ac, auth, reply, reply_len,
5223                                               buf, buf_len);
5224         if (ret)
5225                 return ret;
5226
5227         *authorizer = auth->authorizer_buf;
5228         *authorizer_len = auth->authorizer_buf_len;
5229         return 0;
5230 }
5231
5232 static int mds_handle_auth_done(struct ceph_connection *con,
5233                                 u64 global_id, void *reply, int reply_len,
5234                                 u8 *session_key, int *session_key_len,
5235                                 u8 *con_secret, int *con_secret_len)
5236 {
5237         struct ceph_mds_session *s = con->private;
5238         struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
5239         struct ceph_auth_handshake *auth = &s->s_auth;
5240
5241         return ceph_auth_handle_svc_reply_done(ac, auth, reply, reply_len,
5242                                                session_key, session_key_len,
5243                                                con_secret, con_secret_len);
5244 }
5245
5246 static int mds_handle_auth_bad_method(struct ceph_connection *con,
5247                                       int used_proto, int result,
5248                                       const int *allowed_protos, int proto_cnt,
5249                                       const int *allowed_modes, int mode_cnt)
5250 {
5251         struct ceph_mds_session *s = con->private;
5252         struct ceph_mon_client *monc = &s->s_mdsc->fsc->client->monc;
5253         int ret;
5254
5255         if (ceph_auth_handle_bad_authorizer(monc->auth, CEPH_ENTITY_TYPE_MDS,
5256                                             used_proto, result,
5257                                             allowed_protos, proto_cnt,
5258                                             allowed_modes, mode_cnt)) {
5259                 ret = ceph_monc_validate_auth(monc);
5260                 if (ret)
5261                         return ret;
5262         }
5263
5264         return -EACCES;
5265 }
5266
5267 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
5268                                 struct ceph_msg_header *hdr, int *skip)
5269 {
5270         struct ceph_msg *msg;
5271         int type = (int) le16_to_cpu(hdr->type);
5272         int front_len = (int) le32_to_cpu(hdr->front_len);
5273
5274         if (con->in_msg)
5275                 return con->in_msg;
5276
5277         *skip = 0;
5278         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
5279         if (!msg) {
5280                 pr_err("unable to allocate msg type %d len %d\n",
5281                        type, front_len);
5282                 return NULL;
5283         }
5284
5285         return msg;
5286 }
5287
5288 static int mds_sign_message(struct ceph_msg *msg)
5289 {
5290        struct ceph_mds_session *s = msg->con->private;
5291        struct ceph_auth_handshake *auth = &s->s_auth;
5292
5293        return ceph_auth_sign_message(auth, msg);
5294 }
5295
5296 static int mds_check_message_signature(struct ceph_msg *msg)
5297 {
5298        struct ceph_mds_session *s = msg->con->private;
5299        struct ceph_auth_handshake *auth = &s->s_auth;
5300
5301        return ceph_auth_check_message_signature(auth, msg);
5302 }
5303
5304 static const struct ceph_connection_operations mds_con_ops = {
5305         .get = mds_get_con,
5306         .put = mds_put_con,
5307         .alloc_msg = mds_alloc_msg,
5308         .dispatch = mds_dispatch,
5309         .peer_reset = mds_peer_reset,
5310         .get_authorizer = mds_get_authorizer,
5311         .add_authorizer_challenge = mds_add_authorizer_challenge,
5312         .verify_authorizer_reply = mds_verify_authorizer_reply,
5313         .invalidate_authorizer = mds_invalidate_authorizer,
5314         .sign_message = mds_sign_message,
5315         .check_message_signature = mds_check_message_signature,
5316         .get_auth_request = mds_get_auth_request,
5317         .handle_auth_reply_more = mds_handle_auth_reply_more,
5318         .handle_auth_done = mds_handle_auth_done,
5319         .handle_auth_bad_method = mds_handle_auth_bad_method,
5320 };
5321
5322 /* eof */