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
  14 #include "super.h"
  15 #include "mds_client.h"
  16
  17 #include <linux/ceph/ceph_features.h>
  18 #include <linux/ceph/messenger.h>
  19 #include <linux/ceph/decode.h>
  20 #include <linux/ceph/pagelist.h>
  21 #include <linux/ceph/auth.h>
  22 #include <linux/ceph/debugfs.h>
  23
  24 #define RECONNECT_MAX_SIZE (INT_MAX - PAGE_SIZE)
  25
  26 /*
  27  * A cluster of MDS (metadata server) daemons is responsible for
  28  * managing the file system namespace (the directory hierarchy and
  29  * inodes) and for coordinating shared access to storage.  Metadata is
  30  * partitioning hierarchically across a number of servers, and that
  31  * partition varies over time as the cluster adjusts the distribution
  32  * in order to balance load.
  33  *
  34  * The MDS client is primarily responsible to managing synchronous
  35  * metadata requests for operations like open, unlink, and so forth.
  36  * If there is a MDS failure, we find out about it when we (possibly
  37  * request and) receive a new MDS map, and can resubmit affected
  38  * requests.
  39  *
  40  * For the most part, though, we take advantage of a lossless
  41  * communications channel to the MDS, and do not need to worry about
  42  * timing out or resubmitting requests.
  43  *
  44  * We maintain a stateful "session" with each MDS we interact with.
  45  * Within each session, we sent periodic heartbeat messages to ensure
  46  * any capabilities or leases we have been issues remain valid.  If
  47  * the session times out and goes stale, our leases and capabilities
  48  * are no longer valid.
  49  */
  50
  51 struct ceph_reconnect_state {
  52         struct ceph_mds_session *session;
  53         int nr_caps, nr_realms;
  54         struct ceph_pagelist *pagelist;
  55         unsigned msg_version;
  56         bool allow_multi;
  57 };
  58
  59 static void __wake_requests(struct ceph_mds_client *mdsc,
  60                             struct list_head *head);
  61 static void ceph_cap_release_work(struct work_struct *work);
  62 static void ceph_cap_reclaim_work(struct work_struct *work);
  63
  64 static const struct ceph_connection_operations mds_con_ops;
  65
  66
  67 /*
  68  * mds reply parsing
  69  */
  70
  71 static int parse_reply_info_quota(void **p, void *end,
  72                                   struct ceph_mds_reply_info_in *info)
  73 {
  74         u8 struct_v, struct_compat;
  75         u32 struct_len;
  76
  77         ceph_decode_8_safe(p, end, struct_v, bad);
  78         ceph_decode_8_safe(p, end, struct_compat, bad);
  79         /* struct_v is expected to be >= 1. we only
  80          * understand encoding with struct_compat == 1. */
  81         if (!struct_v || struct_compat != 1)
  82                 goto bad;
  83         ceph_decode_32_safe(p, end, struct_len, bad);
  84         ceph_decode_need(p, end, struct_len, bad);
  85         end = *p + struct_len;
  86         ceph_decode_64_safe(p, end, info->max_bytes, bad);
  87         ceph_decode_64_safe(p, end, info->max_files, bad);
  88         *p = end;
  89         return 0;
  90 bad:
  91         return -EIO;
  92 }
  93
  94 /*
  95  * parse individual inode info
  96  */
  97 static int parse_reply_info_in(void **p, void *end,
  98                                struct ceph_mds_reply_info_in *info,
  99                                u64 features)
 100 {
 101         int err = 0;
 102         u8 struct_v = 0;
 103
 104         if (features == (u64)-1) {
 105                 u32 struct_len;
 106                 u8 struct_compat;
 107                 ceph_decode_8_safe(p, end, struct_v, bad);
 108                 ceph_decode_8_safe(p, end, struct_compat, bad);
 109                 /* struct_v is expected to be >= 1. we only understand
 110                  * encoding with struct_compat == 1. */
 111                 if (!struct_v || struct_compat != 1)
 112                         goto bad;
 113                 ceph_decode_32_safe(p, end, struct_len, bad);
 114                 ceph_decode_need(p, end, struct_len, bad);
 115                 end = *p + struct_len;
 116         }
 117
 118         ceph_decode_need(p, end, sizeof(struct ceph_mds_reply_inode), bad);
 119         info->in = *p;
 120         *p += sizeof(struct ceph_mds_reply_inode) +
 121                 sizeof(*info->in->fragtree.splits) *
 122                 le32_to_cpu(info->in->fragtree.nsplits);
 123
 124         ceph_decode_32_safe(p, end, info->symlink_len, bad);
 125         ceph_decode_need(p, end, info->symlink_len, bad);
 126         info->symlink = *p;
 127         *p += info->symlink_len;
 128
 129         ceph_decode_copy_safe(p, end, &info->dir_layout,
 130                               sizeof(info->dir_layout), bad);
 131         ceph_decode_32_safe(p, end, info->xattr_len, bad);
 132         ceph_decode_need(p, end, info->xattr_len, bad);
 133         info->xattr_data = *p;
 134         *p += info->xattr_len;
 135
 136         if (features == (u64)-1) {
 137                 /* inline data */
 138                 ceph_decode_64_safe(p, end, info->inline_version, bad);
 139                 ceph_decode_32_safe(p, end, info->inline_len, bad);
 140                 ceph_decode_need(p, end, info->inline_len, bad);
 141                 info->inline_data = *p;
 142                 *p += info->inline_len;
 143                 /* quota */
 144                 err = parse_reply_info_quota(p, end, info);
 145                 if (err < 0)
 146                         goto out_bad;
 147                 /* pool namespace */
 148                 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
 149                 if (info->pool_ns_len > 0) {
 150                         ceph_decode_need(p, end, info->pool_ns_len, bad);
 151                         info->pool_ns_data = *p;
 152                         *p += info->pool_ns_len;
 153                 }
 154
 155                 /* btime */
 156                 ceph_decode_need(p, end, sizeof(info->btime), bad);
 157                 ceph_decode_copy(p, &info->btime, sizeof(info->btime));
 158
 159                 /* change attribute */
 160                 ceph_decode_64_safe(p, end, info->change_attr, bad);
 161
 162                 /* dir pin */
 163                 if (struct_v >= 2) {
 164                         ceph_decode_32_safe(p, end, info->dir_pin, bad);
 165                 } else {
 166                         info->dir_pin = -ENODATA;
 167                 }
 168
 169                 /* snapshot birth time, remains zero for v<=2 */
 170                 if (struct_v >= 3) {
 171                         ceph_decode_need(p, end, sizeof(info->snap_btime), bad);
 172                         ceph_decode_copy(p, &info->snap_btime,
 173                                          sizeof(info->snap_btime));
 174                 } else {
 175                         memset(&info->snap_btime, 0, sizeof(info->snap_btime));
 176                 }
 177
 178                 *p = end;
 179         } else {
 180                 if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
 181                         ceph_decode_64_safe(p, end, info->inline_version, bad);
 182                         ceph_decode_32_safe(p, end, info->inline_len, bad);
 183                         ceph_decode_need(p, end, info->inline_len, bad);
 184                         info->inline_data = *p;
 185                         *p += info->inline_len;
 186                 } else
 187                         info->inline_version = CEPH_INLINE_NONE;
 188
 189                 if (features & CEPH_FEATURE_MDS_QUOTA) {
 190                         err = parse_reply_info_quota(p, end, info);
 191                         if (err < 0)
 192                                 goto out_bad;
 193                 } else {
 194                         info->max_bytes = 0;
 195                         info->max_files = 0;
 196                 }
 197
 198                 info->pool_ns_len = 0;
 199                 info->pool_ns_data = NULL;
 200                 if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
 201                         ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
 202                         if (info->pool_ns_len > 0) {
 203                                 ceph_decode_need(p, end, info->pool_ns_len, bad);
 204                                 info->pool_ns_data = *p;
 205                                 *p += info->pool_ns_len;
 206                         }
 207                 }
 208
 209                 if (features & CEPH_FEATURE_FS_BTIME) {
 210                         ceph_decode_need(p, end, sizeof(info->btime), bad);
 211                         ceph_decode_copy(p, &info->btime, sizeof(info->btime));
 212                         ceph_decode_64_safe(p, end, info->change_attr, bad);
 213                 }
 214
 215                 info->dir_pin = -ENODATA;
 216                 /* info->snap_btime remains zero */
 217         }
 218         return 0;
 219 bad:
 220         err = -EIO;
 221 out_bad:
 222         return err;
 223 }
 224
 225 static int parse_reply_info_dir(void **p, void *end,
 226                                 struct ceph_mds_reply_dirfrag **dirfrag,
 227                                 u64 features)
 228 {
 229         if (features == (u64)-1) {
 230                 u8 struct_v, struct_compat;
 231                 u32 struct_len;
 232                 ceph_decode_8_safe(p, end, struct_v, bad);
 233                 ceph_decode_8_safe(p, end, struct_compat, bad);
 234                 /* struct_v is expected to be >= 1. we only understand
 235                  * encoding whose struct_compat == 1. */
 236                 if (!struct_v || struct_compat != 1)
 237                         goto bad;
 238                 ceph_decode_32_safe(p, end, struct_len, bad);
 239                 ceph_decode_need(p, end, struct_len, bad);
 240                 end = *p + struct_len;
 241         }
 242
 243         ceph_decode_need(p, end, sizeof(**dirfrag), bad);
 244         *dirfrag = *p;
 245         *p += sizeof(**dirfrag) + sizeof(u32) * le32_to_cpu((*dirfrag)->ndist);
 246         if (unlikely(*p > end))
 247                 goto bad;
 248         if (features == (u64)-1)
 249                 *p = end;
 250         return 0;
 251 bad:
 252         return -EIO;
 253 }
 254
 255 static int parse_reply_info_lease(void **p, void *end,
 256                                   struct ceph_mds_reply_lease **lease,
 257                                   u64 features)
 258 {
 259         if (features == (u64)-1) {
 260                 u8 struct_v, struct_compat;
 261                 u32 struct_len;
 262                 ceph_decode_8_safe(p, end, struct_v, bad);
 263                 ceph_decode_8_safe(p, end, struct_compat, bad);
 264                 /* struct_v is expected to be >= 1. we only understand
 265                  * encoding whose struct_compat == 1. */
 266                 if (!struct_v || struct_compat != 1)
 267                         goto bad;
 268                 ceph_decode_32_safe(p, end, struct_len, bad);
 269                 ceph_decode_need(p, end, struct_len, bad);
 270                 end = *p + struct_len;
 271         }
 272
 273         ceph_decode_need(p, end, sizeof(**lease), bad);
 274         *lease = *p;
 275         *p += sizeof(**lease);
 276         if (features == (u64)-1)
 277                 *p = end;
 278         return 0;
 279 bad:
 280         return -EIO;
 281 }
 282
 283 /*
 284  * parse a normal reply, which may contain a (dir+)dentry and/or a
 285  * target inode.
 286  */
 287 static int parse_reply_info_trace(void **p, void *end,
 288                                   struct ceph_mds_reply_info_parsed *info,
 289                                   u64 features)
 290 {
 291         int err;
 292
 293         if (info->head->is_dentry) {
 294                 err = parse_reply_info_in(p, end, &info->diri, features);
 295                 if (err < 0)
 296                         goto out_bad;
 297
 298                 err = parse_reply_info_dir(p, end, &info->dirfrag, features);
 299                 if (err < 0)
 300                         goto out_bad;
 301
 302                 ceph_decode_32_safe(p, end, info->dname_len, bad);
 303                 ceph_decode_need(p, end, info->dname_len, bad);
 304                 info->dname = *p;
 305                 *p += info->dname_len;
 306
 307                 err = parse_reply_info_lease(p, end, &info->dlease, features);
 308                 if (err < 0)
 309                         goto out_bad;
 310         }
 311
 312         if (info->head->is_target) {
 313                 err = parse_reply_info_in(p, end, &info->targeti, features);
 314                 if (err < 0)
 315                         goto out_bad;
 316         }
 317
 318         if (unlikely(*p != end))
 319                 goto bad;
 320         return 0;
 321
 322 bad:
 323         err = -EIO;
 324 out_bad:
 325         pr_err("problem parsing mds trace %d\n", err);
 326         return err;
 327 }
 328
 329 /*
 330  * parse readdir results
 331  */
 332 static int parse_reply_info_readdir(void **p, void *end,
 333                                 struct ceph_mds_reply_info_parsed *info,
 334                                 u64 features)
 335 {
 336         u32 num, i = 0;
 337         int err;
 338
 339         err = parse_reply_info_dir(p, end, &info->dir_dir, features);
 340         if (err < 0)
 341                 goto out_bad;
 342
 343         ceph_decode_need(p, end, sizeof(num) + 2, bad);
 344         num = ceph_decode_32(p);
 345         {
 346                 u16 flags = ceph_decode_16(p);
 347                 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
 348                 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
 349                 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
 350                 info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
 351         }
 352         if (num == 0)
 353                 goto done;
 354
 355         BUG_ON(!info->dir_entries);
 356         if ((unsigned long)(info->dir_entries + num) >
 357             (unsigned long)info->dir_entries + info->dir_buf_size) {
 358                 pr_err("dir contents are larger than expected\n");
 359                 WARN_ON(1);
 360                 goto bad;
 361         }
 362
 363         info->dir_nr = num;
 364         while (num) {
 365                 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
 366                 /* dentry */
 367                 ceph_decode_32_safe(p, end, rde->name_len, bad);
 368                 ceph_decode_need(p, end, rde->name_len, bad);
 369                 rde->name = *p;
 370                 *p += rde->name_len;
 371                 dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
 372
 373                 /* dentry lease */
 374                 err = parse_reply_info_lease(p, end, &rde->lease, features);
 375                 if (err)
 376                         goto out_bad;
 377                 /* inode */
 378                 err = parse_reply_info_in(p, end, &rde->inode, features);
 379                 if (err < 0)
 380                         goto out_bad;
 381                 /* ceph_readdir_prepopulate() will update it */
 382                 rde->offset = 0;
 383                 i++;
 384                 num--;
 385         }
 386
 387 done:
 388         /* Skip over any unrecognized fields */
 389         *p = end;
 390         return 0;
 391
 392 bad:
 393         err = -EIO;
 394 out_bad:
 395         pr_err("problem parsing dir contents %d\n", err);
 396         return err;
 397 }
 398
 399 /*
 400  * parse fcntl F_GETLK results
 401  */
 402 static int parse_reply_info_filelock(void **p, void *end,
 403                                      struct ceph_mds_reply_info_parsed *info,
 404                                      u64 features)
 405 {
 406         if (*p + sizeof(*info->filelock_reply) > end)
 407                 goto bad;
 408
 409         info->filelock_reply = *p;
 410
 411         /* Skip over any unrecognized fields */
 412         *p = end;
 413         return 0;
 414 bad:
 415         return -EIO;
 416 }
 417
 418 /*
 419  * parse create results
 420  */
 421 static int parse_reply_info_create(void **p, void *end,
 422                                   struct ceph_mds_reply_info_parsed *info,
 423                                   u64 features)
 424 {
 425         if (features == (u64)-1 ||
 426             (features & CEPH_FEATURE_REPLY_CREATE_INODE)) {
 427                 /* Malformed reply? */
 428                 if (*p == end) {
 429                         info->has_create_ino = false;
 430                 } else {
 431                         info->has_create_ino = true;
 432                         ceph_decode_64_safe(p, end, info->ino, bad);
 433                 }
 434         } else {
 435                 if (*p != end)
 436                         goto bad;
 437         }
 438
 439         /* Skip over any unrecognized fields */
 440         *p = end;
 441         return 0;
 442 bad:
 443         return -EIO;
 444 }
 445
 446 /*
 447  * parse extra results
 448  */
 449 static int parse_reply_info_extra(void **p, void *end,
 450                                   struct ceph_mds_reply_info_parsed *info,
 451                                   u64 features)
 452 {
 453         u32 op = le32_to_cpu(info->head->op);
 454
 455         if (op == CEPH_MDS_OP_GETFILELOCK)
 456                 return parse_reply_info_filelock(p, end, info, features);
 457         else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
 458                 return parse_reply_info_readdir(p, end, info, features);
 459         else if (op == CEPH_MDS_OP_CREATE)
 460                 return parse_reply_info_create(p, end, info, features);
 461         else
 462                 return -EIO;
 463 }
 464
 465 /*
 466  * parse entire mds reply
 467  */
 468 static int parse_reply_info(struct ceph_msg *msg,
 469                             struct ceph_mds_reply_info_parsed *info,
 470                             u64 features)
 471 {
 472         void *p, *end;
 473         u32 len;
 474         int err;
 475
 476         info->head = msg->front.iov_base;
 477         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 478         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 479
 480         /* trace */
 481         ceph_decode_32_safe(&p, end, len, bad);
 482         if (len > 0) {
 483                 ceph_decode_need(&p, end, len, bad);
 484                 err = parse_reply_info_trace(&p, p+len, info, features);
 485                 if (err < 0)
 486                         goto out_bad;
 487         }
 488
 489         /* extra */
 490         ceph_decode_32_safe(&p, end, len, bad);
 491         if (len > 0) {
 492                 ceph_decode_need(&p, end, len, bad);
 493                 err = parse_reply_info_extra(&p, p+len, info, features);
 494                 if (err < 0)
 495                         goto out_bad;
 496         }
 497
 498         /* snap blob */
 499         ceph_decode_32_safe(&p, end, len, bad);
 500         info->snapblob_len = len;
 501         info->snapblob = p;
 502         p += len;
 503
 504         if (p != end)
 505                 goto bad;
 506         return 0;
 507
 508 bad:
 509         err = -EIO;
 510 out_bad:
 511         pr_err("mds parse_reply err %d\n", err);
 512         return err;
 513 }
 514
 515 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
 516 {
 517         if (!info->dir_entries)
 518                 return;
 519         free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
 520 }
 521
 522
 523 /*
 524  * sessions
 525  */
 526 const char *ceph_session_state_name(int s)
 527 {
 528         switch (s) {
 529         case CEPH_MDS_SESSION_NEW: return "new";
 530         case CEPH_MDS_SESSION_OPENING: return "opening";
 531         case CEPH_MDS_SESSION_OPEN: return "open";
 532         case CEPH_MDS_SESSION_HUNG: return "hung";
 533         case CEPH_MDS_SESSION_CLOSING: return "closing";
 534         case CEPH_MDS_SESSION_CLOSED: return "closed";
 535         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
 536         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
 537         case CEPH_MDS_SESSION_REJECTED: return "rejected";
 538         default: return "???";
 539         }
 540 }
 541
 542 struct ceph_mds_session *ceph_get_mds_session(struct ceph_mds_session *s)
 543 {
 544         if (refcount_inc_not_zero(&s->s_ref)) {
 545                 dout("mdsc get_session %p %d -> %d\n", s,
 546                      refcount_read(&s->s_ref)-1, refcount_read(&s->s_ref));
 547                 return s;
 548         } else {
 549                 dout("mdsc get_session %p 0 -- FAIL\n", s);
 550                 return NULL;
 551         }
 552 }
 553
 554 void ceph_put_mds_session(struct ceph_mds_session *s)
 555 {
 556         dout("mdsc put_session %p %d -> %d\n", s,
 557              refcount_read(&s->s_ref), refcount_read(&s->s_ref)-1);
 558         if (refcount_dec_and_test(&s->s_ref)) {
 559                 if (s->s_auth.authorizer)
 560                         ceph_auth_destroy_authorizer(s->s_auth.authorizer);
 561                 kfree(s);
 562         }
 563 }
 564
 565 /*
 566  * called under mdsc->mutex
 567  */
 568 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
 569                                                    int mds)
 570 {
 571         if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
 572                 return NULL;
 573         return ceph_get_mds_session(mdsc->sessions[mds]);
 574 }
 575
 576 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
 577 {
 578         if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
 579                 return false;
 580         else
 581                 return true;
 582 }
 583
 584 static int __verify_registered_session(struct ceph_mds_client *mdsc,
 585                                        struct ceph_mds_session *s)
 586 {
 587         if (s->s_mds >= mdsc->max_sessions ||
 588             mdsc->sessions[s->s_mds] != s)
 589                 return -ENOENT;
 590         return 0;
 591 }
 592
 593 /*
 594  * create+register a new session for given mds.
 595  * called under mdsc->mutex.
 596  */
 597 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 598                                                  int mds)
 599 {
 600         struct ceph_mds_session *s;
 601
 602         if (mds >= mdsc->mdsmap->possible_max_rank)
 603                 return ERR_PTR(-EINVAL);
 604
 605         s = kzalloc(sizeof(*s), GFP_NOFS);
 606         if (!s)
 607                 return ERR_PTR(-ENOMEM);
 608
 609         if (mds >= mdsc->max_sessions) {
 610                 int newmax = 1 << get_count_order(mds + 1);
 611                 struct ceph_mds_session **sa;
 612
 613                 dout("%s: realloc to %d\n", __func__, newmax);
 614                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 615                 if (!sa)
 616                         goto fail_realloc;
 617                 if (mdsc->sessions) {
 618                         memcpy(sa, mdsc->sessions,
 619                                mdsc->max_sessions * sizeof(void *));
 620                         kfree(mdsc->sessions);
 621                 }
 622                 mdsc->sessions = sa;
 623                 mdsc->max_sessions = newmax;
 624         }
 625
 626         dout("%s: mds%d\n", __func__, mds);
 627         s->s_mdsc = mdsc;
 628         s->s_mds = mds;
 629         s->s_state = CEPH_MDS_SESSION_NEW;
 630         s->s_ttl = 0;
 631         s->s_seq = 0;
 632         mutex_init(&s->s_mutex);
 633
 634         ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
 635
 636         spin_lock_init(&s->s_gen_ttl_lock);
 637         s->s_cap_gen = 1;
 638         s->s_cap_ttl = jiffies - 1;
 639
 640         spin_lock_init(&s->s_cap_lock);
 641         s->s_renew_requested = 0;
 642         s->s_renew_seq = 0;
 643         INIT_LIST_HEAD(&s->s_caps);
 644         s->s_nr_caps = 0;
 645         refcount_set(&s->s_ref, 1);
 646         INIT_LIST_HEAD(&s->s_waiting);
 647         INIT_LIST_HEAD(&s->s_unsafe);
 648         s->s_num_cap_releases = 0;
 649         s->s_cap_reconnect = 0;
 650         s->s_cap_iterator = NULL;
 651         INIT_LIST_HEAD(&s->s_cap_releases);
 652         INIT_WORK(&s->s_cap_release_work, ceph_cap_release_work);
 653
 654         INIT_LIST_HEAD(&s->s_cap_flushing);
 655
 656         mdsc->sessions[mds] = s;
 657         atomic_inc(&mdsc->num_sessions);
 658         refcount_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
 659
 660         ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
 661                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
 662
 663         return s;
 664
 665 fail_realloc:
 666         kfree(s);
 667         return ERR_PTR(-ENOMEM);
 668 }
 669
 670 /*
 671  * called under mdsc->mutex
 672  */
 673 static void __unregister_session(struct ceph_mds_client *mdsc,
 674                                struct ceph_mds_session *s)
 675 {
 676         dout("__unregister_session mds%d %p\n", s->s_mds, s);
 677         BUG_ON(mdsc->sessions[s->s_mds] != s);
 678         mdsc->sessions[s->s_mds] = NULL;
 679         ceph_con_close(&s->s_con);
 680         ceph_put_mds_session(s);
 681         atomic_dec(&mdsc->num_sessions);
 682 }
 683
 684 /*
 685  * drop session refs in request.
 686  *
 687  * should be last request ref, or hold mdsc->mutex
 688  */
 689 static void put_request_session(struct ceph_mds_request *req)
 690 {
 691         if (req->r_session) {
 692                 ceph_put_mds_session(req->r_session);
 693                 req->r_session = NULL;
 694         }
 695 }
 696
 697 void ceph_mdsc_release_request(struct kref *kref)
 698 {
 699         struct ceph_mds_request *req = container_of(kref,
 700                                                     struct ceph_mds_request,
 701                                                     r_kref);
 702         destroy_reply_info(&req->r_reply_info);
 703         if (req->r_request)
 704                 ceph_msg_put(req->r_request);
 705         if (req->r_reply)
 706                 ceph_msg_put(req->r_reply);
 707         if (req->r_inode) {
 708                 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
 709                 /* avoid calling iput_final() in mds dispatch threads */
 710                 ceph_async_iput(req->r_inode);
 711         }
 712         if (req->r_parent) {
 713                 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
 714                 ceph_async_iput(req->r_parent);
 715         }
 716         ceph_async_iput(req->r_target_inode);
 717         if (req->r_dentry)
 718                 dput(req->r_dentry);
 719         if (req->r_old_dentry)
 720                 dput(req->r_old_dentry);
 721         if (req->r_old_dentry_dir) {
 722                 /*
 723                  * track (and drop pins for) r_old_dentry_dir
 724                  * separately, since r_old_dentry's d_parent may have
 725                  * changed between the dir mutex being dropped and
 726                  * this request being freed.
 727                  */
 728                 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
 729                                   CEPH_CAP_PIN);
 730                 ceph_async_iput(req->r_old_dentry_dir);
 731         }
 732         kfree(req->r_path1);
 733         kfree(req->r_path2);
 734         if (req->r_pagelist)
 735                 ceph_pagelist_release(req->r_pagelist);
 736         put_request_session(req);
 737         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
 738         WARN_ON_ONCE(!list_empty(&req->r_wait));
 739         kfree(req);
 740 }
 741
 742 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
 743
 744 /*
 745  * lookup session, bump ref if found.
 746  *
 747  * called under mdsc->mutex.
 748  */
 749 static struct ceph_mds_request *
 750 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
 751 {
 752         struct ceph_mds_request *req;
 753
 754         req = lookup_request(&mdsc->request_tree, tid);
 755         if (req)
 756                 ceph_mdsc_get_request(req);
 757
 758         return req;
 759 }
 760
 761 /*
 762  * Register an in-flight request, and assign a tid.  Link to directory
 763  * are modifying (if any).
 764  *
 765  * Called under mdsc->mutex.
 766  */
 767 static void __register_request(struct ceph_mds_client *mdsc,
 768                                struct ceph_mds_request *req,
 769                                struct inode *dir)
 770 {
 771         int ret = 0;
 772
 773         req->r_tid = ++mdsc->last_tid;
 774         if (req->r_num_caps) {
 775                 ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
 776                                         req->r_num_caps);
 777                 if (ret < 0) {
 778                         pr_err("__register_request %p "
 779                                "failed to reserve caps: %d\n", req, ret);
 780                         /* set req->r_err to fail early from __do_request */
 781                         req->r_err = ret;
 782                         return;
 783                 }
 784         }
 785         dout("__register_request %p tid %lld\n", req, req->r_tid);
 786         ceph_mdsc_get_request(req);
 787         insert_request(&mdsc->request_tree, req);
 788
 789         req->r_uid = current_fsuid();
 790         req->r_gid = current_fsgid();
 791
 792         if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
 793                 mdsc->oldest_tid = req->r_tid;
 794
 795         if (dir) {
 796                 ihold(dir);
 797                 req->r_unsafe_dir = dir;
 798         }
 799 }
 800
 801 static void __unregister_request(struct ceph_mds_client *mdsc,
 802                                  struct ceph_mds_request *req)
 803 {
 804         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 805
 806         /* Never leave an unregistered request on an unsafe list! */
 807         list_del_init(&req->r_unsafe_item);
 808
 809         if (req->r_tid == mdsc->oldest_tid) {
 810                 struct rb_node *p = rb_next(&req->r_node);
 811                 mdsc->oldest_tid = 0;
 812                 while (p) {
 813                         struct ceph_mds_request *next_req =
 814                                 rb_entry(p, struct ceph_mds_request, r_node);
 815                         if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
 816                                 mdsc->oldest_tid = next_req->r_tid;
 817                                 break;
 818                         }
 819                         p = rb_next(p);
 820                 }
 821         }
 822
 823         erase_request(&mdsc->request_tree, req);
 824
 825         if (req->r_unsafe_dir  &&
 826             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
 827                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 828                 spin_lock(&ci->i_unsafe_lock);
 829                 list_del_init(&req->r_unsafe_dir_item);
 830                 spin_unlock(&ci->i_unsafe_lock);
 831         }
 832         if (req->r_target_inode &&
 833             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
 834                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
 835                 spin_lock(&ci->i_unsafe_lock);
 836                 list_del_init(&req->r_unsafe_target_item);
 837                 spin_unlock(&ci->i_unsafe_lock);
 838         }
 839
 840         if (req->r_unsafe_dir) {
 841                 /* avoid calling iput_final() in mds dispatch threads */
 842                 ceph_async_iput(req->r_unsafe_dir);
 843                 req->r_unsafe_dir = NULL;
 844         }
 845
 846         complete_all(&req->r_safe_completion);
 847
 848         ceph_mdsc_put_request(req);
 849 }
 850
 851 /*
 852  * Walk back up the dentry tree until we hit a dentry representing a
 853  * non-snapshot inode. We do this using the rcu_read_lock (which must be held
 854  * when calling this) to ensure that the objects won't disappear while we're
 855  * working with them. Once we hit a candidate dentry, we attempt to take a
 856  * reference to it, and return that as the result.
 857  */
 858 static struct inode *get_nonsnap_parent(struct dentry *dentry)
 859 {
 860         struct inode *inode = NULL;
 861
 862         while (dentry && !IS_ROOT(dentry)) {
 863                 inode = d_inode_rcu(dentry);
 864                 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
 865                         break;
 866                 dentry = dentry->d_parent;
 867         }
 868         if (inode)
 869                 inode = igrab(inode);
 870         return inode;
 871 }
 872
 873 /*
 874  * Choose mds to send request to next.  If there is a hint set in the
 875  * request (e.g., due to a prior forward hint from the mds), use that.
 876  * Otherwise, consult frag tree and/or caps to identify the
 877  * appropriate mds.  If all else fails, choose randomly.
 878  *
 879  * Called under mdsc->mutex.
 880  */
 881 static int __choose_mds(struct ceph_mds_client *mdsc,
 882                         struct ceph_mds_request *req,
 883                         bool *random)
 884 {
 885         struct inode *inode;
 886         struct ceph_inode_info *ci;
 887         struct ceph_cap *cap;
 888         int mode = req->r_direct_mode;
 889         int mds = -1;
 890         u32 hash = req->r_direct_hash;
 891         bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
 892
 893         if (random)
 894                 *random = false;
 895
 896         /*
 897          * is there a specific mds we should try?  ignore hint if we have
 898          * no session and the mds is not up (active or recovering).
 899          */
 900         if (req->r_resend_mds >= 0 &&
 901             (__have_session(mdsc, req->r_resend_mds) ||
 902              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 903                 dout("%s using resend_mds mds%d\n", __func__,
 904                      req->r_resend_mds);
 905                 return req->r_resend_mds;
 906         }
 907
 908         if (mode == USE_RANDOM_MDS)
 909                 goto random;
 910
 911         inode = NULL;
 912         if (req->r_inode) {
 913                 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
 914                         inode = req->r_inode;
 915                         ihold(inode);
 916                 } else {
 917                         /* req->r_dentry is non-null for LSSNAP request */
 918                         rcu_read_lock();
 919                         inode = get_nonsnap_parent(req->r_dentry);
 920                         rcu_read_unlock();
 921                         dout("%s using snapdir's parent %p\n", __func__, inode);
 922                 }
 923         } else if (req->r_dentry) {
 924                 /* ignore race with rename; old or new d_parent is okay */
 925                 struct dentry *parent;
 926                 struct inode *dir;
 927
 928                 rcu_read_lock();
 929                 parent = READ_ONCE(req->r_dentry->d_parent);
 930                 dir = req->r_parent ? : d_inode_rcu(parent);
 931
 932                 if (!dir || dir->i_sb != mdsc->fsc->sb) {
 933                         /*  not this fs or parent went negative */
 934                         inode = d_inode(req->r_dentry);
 935                         if (inode)
 936                                 ihold(inode);
 937                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
 938                         /* direct snapped/virtual snapdir requests
 939                          * based on parent dir inode */
 940                         inode = get_nonsnap_parent(parent);
 941                         dout("%s using nonsnap parent %p\n", __func__, inode);
 942                 } else {
 943                         /* dentry target */
 944                         inode = d_inode(req->r_dentry);
 945                         if (!inode || mode == USE_AUTH_MDS) {
 946                                 /* dir + name */
 947                                 inode = igrab(dir);
 948                                 hash = ceph_dentry_hash(dir, req->r_dentry);
 949                                 is_hash = true;
 950                         } else {
 951                                 ihold(inode);
 952                         }
 953                 }
 954                 rcu_read_unlock();
 955         }
 956
 957         dout("%s %p is_hash=%d (0x%x) mode %d\n", __func__, inode, (int)is_hash,
 958              hash, mode);
 959         if (!inode)
 960                 goto random;
 961         ci = ceph_inode(inode);
 962
 963         if (is_hash && S_ISDIR(inode->i_mode)) {
 964                 struct ceph_inode_frag frag;
 965                 int found;
 966
 967                 ceph_choose_frag(ci, hash, &frag, &found);
 968                 if (found) {
 969                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 970                                 u8 r;
 971
 972                                 /* choose a random replica */
 973                                 get_random_bytes(&r, 1);
 974                                 r %= frag.ndist;
 975                                 mds = frag.dist[r];
 976                                 dout("%s %p %llx.%llx frag %u mds%d (%d/%d)\n",
 977                                      __func__, inode, ceph_vinop(inode),
 978                                      frag.frag, mds, (int)r, frag.ndist);
 979                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 980                                     CEPH_MDS_STATE_ACTIVE &&
 981                                     !ceph_mdsmap_is_laggy(mdsc->mdsmap, mds))
 982                                         goto out;
 983                         }
 984
 985                         /* since this file/dir wasn't known to be
 986                          * replicated, then we want to look for the
 987                          * authoritative mds. */
 988                         if (frag.mds >= 0) {
 989                                 /* choose auth mds */
 990                                 mds = frag.mds;
 991                                 dout("%s %p %llx.%llx frag %u mds%d (auth)\n",
 992                                      __func__, inode, ceph_vinop(inode),
 993                                      frag.frag, mds);
 994                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 995                                     CEPH_MDS_STATE_ACTIVE) {
 996                                         if (mode == USE_ANY_MDS &&
 997                                             !ceph_mdsmap_is_laggy(mdsc->mdsmap,
 998                                                                   mds))
 999                                                 goto out;
1000                                 }
1001                         }
1002                         mode = USE_AUTH_MDS;
1003                 }
1004         }
1005
1006         spin_lock(&ci->i_ceph_lock);
1007         cap = NULL;
1008         if (mode == USE_AUTH_MDS)
1009                 cap = ci->i_auth_cap;
1010         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
1011                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
1012         if (!cap) {
1013                 spin_unlock(&ci->i_ceph_lock);
1014                 ceph_async_iput(inode);
1015                 goto random;
1016         }
1017         mds = cap->session->s_mds;
1018         dout("%s %p %llx.%llx mds%d (%scap %p)\n", __func__,
1019              inode, ceph_vinop(inode), mds,
1020              cap == ci->i_auth_cap ? "auth " : "", cap);
1021         spin_unlock(&ci->i_ceph_lock);
1022 out:
1023         /* avoid calling iput_final() while holding mdsc->mutex or
1024          * in mds dispatch threads */
1025         ceph_async_iput(inode);
1026         return mds;
1027
1028 random:
1029         if (random)
1030                 *random = true;
1031
1032         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
1033         dout("%s chose random mds%d\n", __func__, mds);
1034         return mds;
1035 }
1036
1037
1038 /*
1039  * session messages
1040  */
1041 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
1042 {
1043         struct ceph_msg *msg;
1044         struct ceph_mds_session_head *h;
1045
1046         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
1047                            false);
1048         if (!msg) {
1049                 pr_err("create_session_msg ENOMEM creating msg\n");
1050                 return NULL;
1051         }
1052         h = msg->front.iov_base;
1053         h->op = cpu_to_le32(op);
1054         h->seq = cpu_to_le64(seq);
1055
1056         return msg;
1057 }
1058
1059 static const unsigned char feature_bits[] = CEPHFS_FEATURES_CLIENT_SUPPORTED;
1060 #define FEATURE_BYTES(c) (DIV_ROUND_UP((size_t)feature_bits[c - 1] + 1, 64) * 8)
1061 static void encode_supported_features(void **p, void *end)
1062 {
1063         static const size_t count = ARRAY_SIZE(feature_bits);
1064
1065         if (count > 0) {
1066                 size_t i;
1067                 size_t size = FEATURE_BYTES(count);
1068
1069                 BUG_ON(*p + 4 + size > end);
1070                 ceph_encode_32(p, size);
1071                 memset(*p, 0, size);
1072                 for (i = 0; i < count; i++)
1073                         ((unsigned char*)(*p))[i / 8] |= BIT(feature_bits[i] % 8);
1074                 *p += size;
1075         } else {
1076                 BUG_ON(*p + 4 > end);
1077                 ceph_encode_32(p, 0);
1078         }
1079 }
1080
1081 /*
1082  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
1083  * to include additional client metadata fields.
1084  */
1085 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
1086 {
1087         struct ceph_msg *msg;
1088         struct ceph_mds_session_head *h;
1089         int i = -1;
1090         int extra_bytes = 0;
1091         int metadata_key_count = 0;
1092         struct ceph_options *opt = mdsc->fsc->client->options;
1093         struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
1094         size_t size, count;
1095         void *p, *end;
1096
1097         const char* metadata[][2] = {
1098                 {"hostname", mdsc->nodename},
1099                 {"kernel_version", init_utsname()->release},
1100                 {"entity_id", opt->name ? : ""},
1101                 {"root", fsopt->server_path ? : "/"},
1102                 {NULL, NULL}
1103         };
1104
1105         /* Calculate serialized length of metadata */
1106         extra_bytes = 4;  /* map length */
1107         for (i = 0; metadata[i][0]; ++i) {
1108                 extra_bytes += 8 + strlen(metadata[i][0]) +
1109                         strlen(metadata[i][1]);
1110                 metadata_key_count++;
1111         }
1112
1113         /* supported feature */
1114         size = 0;
1115         count = ARRAY_SIZE(feature_bits);
1116         if (count > 0)
1117                 size = FEATURE_BYTES(count);
1118         extra_bytes += 4 + size;
1119
1120         /* Allocate the message */
1121         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + extra_bytes,
1122                            GFP_NOFS, false);
1123         if (!msg) {
1124                 pr_err("create_session_msg ENOMEM creating msg\n");
1125                 return NULL;
1126         }
1127         p = msg->front.iov_base;
1128         end = p + msg->front.iov_len;
1129
1130         h = p;
1131         h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
1132         h->seq = cpu_to_le64(seq);
1133
1134         /*
1135          * Serialize client metadata into waiting buffer space, using
1136          * the format that userspace expects for map<string, string>
1137          *
1138          * ClientSession messages with metadata are v3
1139          */
1140         msg->hdr.version = cpu_to_le16(3);
1141         msg->hdr.compat_version = cpu_to_le16(1);
1142
1143         /* The write pointer, following the session_head structure */
1144         p += sizeof(*h);
1145
1146         /* Number of entries in the map */
1147         ceph_encode_32(&p, metadata_key_count);
1148
1149         /* Two length-prefixed strings for each entry in the map */
1150         for (i = 0; metadata[i][0]; ++i) {
1151                 size_t const key_len = strlen(metadata[i][0]);
1152                 size_t const val_len = strlen(metadata[i][1]);
1153
1154                 ceph_encode_32(&p, key_len);
1155                 memcpy(p, metadata[i][0], key_len);
1156                 p += key_len;
1157                 ceph_encode_32(&p, val_len);
1158                 memcpy(p, metadata[i][1], val_len);
1159                 p += val_len;
1160         }
1161
1162         encode_supported_features(&p, end);
1163         msg->front.iov_len = p - msg->front.iov_base;
1164         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1165
1166         return msg;
1167 }
1168
1169 /*
1170  * send session open request.
1171  *
1172  * called under mdsc->mutex
1173  */
1174 static int __open_session(struct ceph_mds_client *mdsc,
1175                           struct ceph_mds_session *session)
1176 {
1177         struct ceph_msg *msg;
1178         int mstate;
1179         int mds = session->s_mds;
1180
1181         /* wait for mds to go active? */
1182         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
1183         dout("open_session to mds%d (%s)\n", mds,
1184              ceph_mds_state_name(mstate));
1185         session->s_state = CEPH_MDS_SESSION_OPENING;
1186         session->s_renew_requested = jiffies;
1187
1188         /* send connect message */
1189         msg = create_session_open_msg(mdsc, session->s_seq);
1190         if (!msg)
1191                 return -ENOMEM;
1192         ceph_con_send(&session->s_con, msg);
1193         return 0;
1194 }
1195
1196 /*
1197  * open sessions for any export targets for the given mds
1198  *
1199  * called under mdsc->mutex
1200  */
1201 static struct ceph_mds_session *
1202 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
1203 {
1204         struct ceph_mds_session *session;
1205
1206         session = __ceph_lookup_mds_session(mdsc, target);
1207         if (!session) {
1208                 session = register_session(mdsc, target);
1209                 if (IS_ERR(session))
1210                         return session;
1211         }
1212         if (session->s_state == CEPH_MDS_SESSION_NEW ||
1213             session->s_state == CEPH_MDS_SESSION_CLOSING)
1214                 __open_session(mdsc, session);
1215
1216         return session;
1217 }
1218
1219 struct ceph_mds_session *
1220 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1221 {
1222         struct ceph_mds_session *session;
1223
1224         dout("open_export_target_session to mds%d\n", target);
1225
1226         mutex_lock(&mdsc->mutex);
1227         session = __open_export_target_session(mdsc, target);
1228         mutex_unlock(&mdsc->mutex);
1229
1230         return session;
1231 }
1232
1233 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1234                                           struct ceph_mds_session *session)
1235 {
1236         struct ceph_mds_info *mi;
1237         struct ceph_mds_session *ts;
1238         int i, mds = session->s_mds;
1239
1240         if (mds >= mdsc->mdsmap->possible_max_rank)
1241                 return;
1242
1243         mi = &mdsc->mdsmap->m_info[mds];
1244         dout("open_export_target_sessions for mds%d (%d targets)\n",
1245              session->s_mds, mi->num_export_targets);
1246
1247         for (i = 0; i < mi->num_export_targets; i++) {
1248                 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1249                 if (!IS_ERR(ts))
1250                         ceph_put_mds_session(ts);
1251         }
1252 }
1253
1254 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1255                                            struct ceph_mds_session *session)
1256 {
1257         mutex_lock(&mdsc->mutex);
1258         __open_export_target_sessions(mdsc, session);
1259         mutex_unlock(&mdsc->mutex);
1260 }
1261
1262 /*
1263  * session caps
1264  */
1265
1266 static void detach_cap_releases(struct ceph_mds_session *session,
1267                                 struct list_head *target)
1268 {
1269         lockdep_assert_held(&session->s_cap_lock);
1270
1271         list_splice_init(&session->s_cap_releases, target);
1272         session->s_num_cap_releases = 0;
1273         dout("dispose_cap_releases mds%d\n", session->s_mds);
1274 }
1275
1276 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1277                                  struct list_head *dispose)
1278 {
1279         while (!list_empty(dispose)) {
1280                 struct ceph_cap *cap;
1281                 /* zero out the in-progress message */
1282                 cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1283                 list_del(&cap->session_caps);
1284                 ceph_put_cap(mdsc, cap);
1285         }
1286 }
1287
1288 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1289                                      struct ceph_mds_session *session)
1290 {
1291         struct ceph_mds_request *req;
1292         struct rb_node *p;
1293         struct ceph_inode_info *ci;
1294
1295         dout("cleanup_session_requests mds%d\n", session->s_mds);
1296         mutex_lock(&mdsc->mutex);
1297         while (!list_empty(&session->s_unsafe)) {
1298                 req = list_first_entry(&session->s_unsafe,
1299                                        struct ceph_mds_request, r_unsafe_item);
1300                 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1301                                     req->r_tid);
1302                 if (req->r_target_inode) {
1303                         /* dropping unsafe change of inode's attributes */
1304                         ci = ceph_inode(req->r_target_inode);
1305                         errseq_set(&ci->i_meta_err, -EIO);
1306                 }
1307                 if (req->r_unsafe_dir) {
1308                         /* dropping unsafe directory operation */
1309                         ci = ceph_inode(req->r_unsafe_dir);
1310                         errseq_set(&ci->i_meta_err, -EIO);
1311                 }
1312                 __unregister_request(mdsc, req);
1313         }
1314         /* zero r_attempts, so kick_requests() will re-send requests */
1315         p = rb_first(&mdsc->request_tree);
1316         while (p) {
1317                 req = rb_entry(p, struct ceph_mds_request, r_node);
1318                 p = rb_next(p);
1319                 if (req->r_session &&
1320                     req->r_session->s_mds == session->s_mds)
1321                         req->r_attempts = 0;
1322         }
1323         mutex_unlock(&mdsc->mutex);
1324 }
1325
1326 /*
1327  * Helper to safely iterate over all caps associated with a session, with
1328  * special care taken to handle a racing __ceph_remove_cap().
1329  *
1330  * Caller must hold session s_mutex.
1331  */
1332 int ceph_iterate_session_caps(struct ceph_mds_session *session,
1333                               int (*cb)(struct inode *, struct ceph_cap *,
1334                                         void *), void *arg)
1335 {
1336         struct list_head *p;
1337         struct ceph_cap *cap;
1338         struct inode *inode, *last_inode = NULL;
1339         struct ceph_cap *old_cap = NULL;
1340         int ret;
1341
1342         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1343         spin_lock(&session->s_cap_lock);
1344         p = session->s_caps.next;
1345         while (p != &session->s_caps) {
1346                 cap = list_entry(p, struct ceph_cap, session_caps);
1347                 inode = igrab(&cap->ci->vfs_inode);
1348                 if (!inode) {
1349                         p = p->next;
1350                         continue;
1351                 }
1352                 session->s_cap_iterator = cap;
1353                 spin_unlock(&session->s_cap_lock);
1354
1355                 if (last_inode) {
1356                         /* avoid calling iput_final() while holding
1357                          * s_mutex or in mds dispatch threads */
1358                         ceph_async_iput(last_inode);
1359                         last_inode = NULL;
1360                 }
1361                 if (old_cap) {
1362                         ceph_put_cap(session->s_mdsc, old_cap);
1363                         old_cap = NULL;
1364                 }
1365
1366                 ret = cb(inode, cap, arg);
1367                 last_inode = inode;
1368
1369                 spin_lock(&session->s_cap_lock);
1370                 p = p->next;
1371                 if (!cap->ci) {
1372                         dout("iterate_session_caps  finishing cap %p removal\n",
1373                              cap);
1374                         BUG_ON(cap->session != session);
1375                         cap->session = NULL;
1376                         list_del_init(&cap->session_caps);
1377                         session->s_nr_caps--;
1378                         if (cap->queue_release)
1379                                 __ceph_queue_cap_release(session, cap);
1380                         else
1381                                 old_cap = cap;  /* put_cap it w/o locks held */
1382                 }
1383                 if (ret < 0)
1384                         goto out;
1385         }
1386         ret = 0;
1387 out:
1388         session->s_cap_iterator = NULL;
1389         spin_unlock(&session->s_cap_lock);
1390
1391         ceph_async_iput(last_inode);
1392         if (old_cap)
1393                 ceph_put_cap(session->s_mdsc, old_cap);
1394
1395         return ret;
1396 }
1397
1398 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1399                                   void *arg)
1400 {
1401         struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1402         struct ceph_inode_info *ci = ceph_inode(inode);
1403         LIST_HEAD(to_remove);
1404         bool dirty_dropped = false;
1405         bool invalidate = false;
1406
1407         dout("removing cap %p, ci is %p, inode is %p\n",
1408              cap, ci, &ci->vfs_inode);
1409         spin_lock(&ci->i_ceph_lock);
1410         if (cap->mds_wanted | cap->issued)
1411                 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1412         __ceph_remove_cap(cap, false);
1413         if (!ci->i_auth_cap) {
1414                 struct ceph_cap_flush *cf;
1415                 struct ceph_mds_client *mdsc = fsc->mdsc;
1416
1417                 if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1418                         if (inode->i_data.nrpages > 0)
1419                                 invalidate = true;
1420                         if (ci->i_wrbuffer_ref > 0)
1421                                 mapping_set_error(&inode->i_data, -EIO);
1422                 }
1423
1424                 while (!list_empty(&ci->i_cap_flush_list)) {
1425                         cf = list_first_entry(&ci->i_cap_flush_list,
1426                                               struct ceph_cap_flush, i_list);
1427                         list_move(&cf->i_list, &to_remove);
1428                 }
1429
1430                 spin_lock(&mdsc->cap_dirty_lock);
1431
1432                 list_for_each_entry(cf, &to_remove, i_list)
1433                         list_del(&cf->g_list);
1434
1435                 if (!list_empty(&ci->i_dirty_item)) {
1436                         pr_warn_ratelimited(
1437                                 " dropping dirty %s state for %p %lld\n",
1438                                 ceph_cap_string(ci->i_dirty_caps),
1439                                 inode, ceph_ino(inode));
1440                         ci->i_dirty_caps = 0;
1441                         list_del_init(&ci->i_dirty_item);
1442                         dirty_dropped = true;
1443                 }
1444                 if (!list_empty(&ci->i_flushing_item)) {
1445                         pr_warn_ratelimited(
1446                                 " dropping dirty+flushing %s state for %p %lld\n",
1447                                 ceph_cap_string(ci->i_flushing_caps),
1448                                 inode, ceph_ino(inode));
1449                         ci->i_flushing_caps = 0;
1450                         list_del_init(&ci->i_flushing_item);
1451                         mdsc->num_cap_flushing--;
1452                         dirty_dropped = true;
1453                 }
1454                 spin_unlock(&mdsc->cap_dirty_lock);
1455
1456                 if (dirty_dropped) {
1457                         errseq_set(&ci->i_meta_err, -EIO);
1458
1459                         if (ci->i_wrbuffer_ref_head == 0 &&
1460                             ci->i_wr_ref == 0 &&
1461                             ci->i_dirty_caps == 0 &&
1462                             ci->i_flushing_caps == 0) {
1463                                 ceph_put_snap_context(ci->i_head_snapc);
1464                                 ci->i_head_snapc = NULL;
1465                         }
1466                 }
1467
1468                 if (atomic_read(&ci->i_filelock_ref) > 0) {
1469                         /* make further file lock syscall return -EIO */
1470                         ci->i_ceph_flags |= CEPH_I_ERROR_FILELOCK;
1471                         pr_warn_ratelimited(" dropping file locks for %p %lld\n",
1472                                             inode, ceph_ino(inode));
1473                 }
1474
1475                 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1476                         list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1477                         ci->i_prealloc_cap_flush = NULL;
1478                 }
1479         }
1480         spin_unlock(&ci->i_ceph_lock);
1481         while (!list_empty(&to_remove)) {
1482                 struct ceph_cap_flush *cf;
1483                 cf = list_first_entry(&to_remove,
1484                                       struct ceph_cap_flush, i_list);
1485                 list_del(&cf->i_list);
1486                 ceph_free_cap_flush(cf);
1487         }
1488
1489         wake_up_all(&ci->i_cap_wq);
1490         if (invalidate)
1491                 ceph_queue_invalidate(inode);
1492         if (dirty_dropped)
1493                 iput(inode);
1494         return 0;
1495 }
1496
1497 /*
1498  * caller must hold session s_mutex
1499  */
1500 static void remove_session_caps(struct ceph_mds_session *session)
1501 {
1502         struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1503         struct super_block *sb = fsc->sb;
1504         LIST_HEAD(dispose);
1505
1506         dout("remove_session_caps on %p\n", session);
1507         ceph_iterate_session_caps(session, remove_session_caps_cb, fsc);
1508
1509         wake_up_all(&fsc->mdsc->cap_flushing_wq);
1510
1511         spin_lock(&session->s_cap_lock);
1512         if (session->s_nr_caps > 0) {
1513                 struct inode *inode;
1514                 struct ceph_cap *cap, *prev = NULL;
1515                 struct ceph_vino vino;
1516                 /*
1517                  * iterate_session_caps() skips inodes that are being
1518                  * deleted, we need to wait until deletions are complete.
1519                  * __wait_on_freeing_inode() is designed for the job,
1520                  * but it is not exported, so use lookup inode function
1521                  * to access it.
1522                  */
1523                 while (!list_empty(&session->s_caps)) {
1524                         cap = list_entry(session->s_caps.next,
1525                                          struct ceph_cap, session_caps);
1526                         if (cap == prev)
1527                                 break;
1528                         prev = cap;
1529                         vino = cap->ci->i_vino;
1530                         spin_unlock(&session->s_cap_lock);
1531
1532                         inode = ceph_find_inode(sb, vino);
1533                          /* avoid calling iput_final() while holding s_mutex */
1534                         ceph_async_iput(inode);
1535
1536                         spin_lock(&session->s_cap_lock);
1537                 }
1538         }
1539
1540         // drop cap expires and unlock s_cap_lock
1541         detach_cap_releases(session, &dispose);
1542
1543         BUG_ON(session->s_nr_caps > 0);
1544         BUG_ON(!list_empty(&session->s_cap_flushing));
1545         spin_unlock(&session->s_cap_lock);
1546         dispose_cap_releases(session->s_mdsc, &dispose);
1547 }
1548
1549 enum {
1550         RECONNECT,
1551         RENEWCAPS,
1552         FORCE_RO,
1553 };
1554
1555 /*
1556  * wake up any threads waiting on this session's caps.  if the cap is
1557  * old (didn't get renewed on the client reconnect), remove it now.
1558  *
1559  * caller must hold s_mutex.
1560  */
1561 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1562                               void *arg)
1563 {
1564         struct ceph_inode_info *ci = ceph_inode(inode);
1565         unsigned long ev = (unsigned long)arg;
1566
1567         if (ev == RECONNECT) {
1568                 spin_lock(&ci->i_ceph_lock);
1569                 ci->i_wanted_max_size = 0;
1570                 ci->i_requested_max_size = 0;
1571                 spin_unlock(&ci->i_ceph_lock);
1572         } else if (ev == RENEWCAPS) {
1573                 if (cap->cap_gen < cap->session->s_cap_gen) {
1574                         /* mds did not re-issue stale cap */
1575                         spin_lock(&ci->i_ceph_lock);
1576                         cap->issued = cap->implemented = CEPH_CAP_PIN;
1577                         /* make sure mds knows what we want */
1578                         if (__ceph_caps_file_wanted(ci) & ~cap->mds_wanted)
1579                                 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1580                         spin_unlock(&ci->i_ceph_lock);
1581                 }
1582         } else if (ev == FORCE_RO) {
1583         }
1584         wake_up_all(&ci->i_cap_wq);
1585         return 0;
1586 }
1587
1588 static void wake_up_session_caps(struct ceph_mds_session *session, int ev)
1589 {
1590         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1591         ceph_iterate_session_caps(session, wake_up_session_cb,
1592                                   (void *)(unsigned long)ev);
1593 }
1594
1595 /*
1596  * Send periodic message to MDS renewing all currently held caps.  The
1597  * ack will reset the expiration for all caps from this session.
1598  *
1599  * caller holds s_mutex
1600  */
1601 static int send_renew_caps(struct ceph_mds_client *mdsc,
1602                            struct ceph_mds_session *session)
1603 {
1604         struct ceph_msg *msg;
1605         int state;
1606
1607         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1608             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1609                 pr_info("mds%d caps stale\n", session->s_mds);
1610         session->s_renew_requested = jiffies;
1611
1612         /* do not try to renew caps until a recovering mds has reconnected
1613          * with its clients. */
1614         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1615         if (state < CEPH_MDS_STATE_RECONNECT) {
1616                 dout("send_renew_caps ignoring mds%d (%s)\n",
1617                      session->s_mds, ceph_mds_state_name(state));
1618                 return 0;
1619         }
1620
1621         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1622                 ceph_mds_state_name(state));
1623         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1624                                  ++session->s_renew_seq);
1625         if (!msg)
1626                 return -ENOMEM;
1627         ceph_con_send(&session->s_con, msg);
1628         return 0;
1629 }
1630
1631 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1632                              struct ceph_mds_session *session, u64 seq)
1633 {
1634         struct ceph_msg *msg;
1635
1636         dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1637              session->s_mds, ceph_session_state_name(session->s_state), seq);
1638         msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1639         if (!msg)
1640                 return -ENOMEM;
1641         ceph_con_send(&session->s_con, msg);
1642         return 0;
1643 }
1644
1645
1646 /*
1647  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1648  *
1649  * Called under session->s_mutex
1650  */
1651 static void renewed_caps(struct ceph_mds_client *mdsc,
1652                          struct ceph_mds_session *session, int is_renew)
1653 {
1654         int was_stale;
1655         int wake = 0;
1656
1657         spin_lock(&session->s_cap_lock);
1658         was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1659
1660         session->s_cap_ttl = session->s_renew_requested +
1661                 mdsc->mdsmap->m_session_timeout*HZ;
1662
1663         if (was_stale) {
1664                 if (time_before(jiffies, session->s_cap_ttl)) {
1665                         pr_info("mds%d caps renewed\n", session->s_mds);
1666                         wake = 1;
1667                 } else {
1668                         pr_info("mds%d caps still stale\n", session->s_mds);
1669                 }
1670         }
1671         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1672              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1673              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1674         spin_unlock(&session->s_cap_lock);
1675
1676         if (wake)
1677                 wake_up_session_caps(session, RENEWCAPS);
1678 }
1679
1680 /*
1681  * send a session close request
1682  */
1683 static int request_close_session(struct ceph_mds_client *mdsc,
1684                                  struct ceph_mds_session *session)
1685 {
1686         struct ceph_msg *msg;
1687
1688         dout("request_close_session mds%d state %s seq %lld\n",
1689              session->s_mds, ceph_session_state_name(session->s_state),
1690              session->s_seq);
1691         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1692         if (!msg)
1693                 return -ENOMEM;
1694         ceph_con_send(&session->s_con, msg);
1695         return 1;
1696 }
1697
1698 /*
1699  * Called with s_mutex held.
1700  */
1701 static int __close_session(struct ceph_mds_client *mdsc,
1702                          struct ceph_mds_session *session)
1703 {
1704         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1705                 return 0;
1706         session->s_state = CEPH_MDS_SESSION_CLOSING;
1707         return request_close_session(mdsc, session);
1708 }
1709
1710 static bool drop_negative_children(struct dentry *dentry)
1711 {
1712         struct dentry *child;
1713         bool all_negative = true;
1714
1715         if (!d_is_dir(dentry))
1716                 goto out;
1717
1718         spin_lock(&dentry->d_lock);
1719         list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1720                 if (d_really_is_positive(child)) {
1721                         all_negative = false;
1722                         break;
1723                 }
1724         }
1725         spin_unlock(&dentry->d_lock);
1726
1727         if (all_negative)
1728                 shrink_dcache_parent(dentry);
1729 out:
1730         return all_negative;
1731 }
1732
1733 /*
1734  * Trim old(er) caps.
1735  *
1736  * Because we can't cache an inode without one or more caps, we do
1737  * this indirectly: if a cap is unused, we prune its aliases, at which
1738  * point the inode will hopefully get dropped to.
1739  *
1740  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1741  * memory pressure from the MDS, though, so it needn't be perfect.
1742  */
1743 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1744 {
1745         int *remaining = arg;
1746         struct ceph_inode_info *ci = ceph_inode(inode);
1747         int used, wanted, oissued, mine;
1748
1749         if (*remaining <= 0)
1750                 return -1;
1751
1752         spin_lock(&ci->i_ceph_lock);
1753         mine = cap->issued | cap->implemented;
1754         used = __ceph_caps_used(ci);
1755         wanted = __ceph_caps_file_wanted(ci);
1756         oissued = __ceph_caps_issued_other(ci, cap);
1757
1758         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1759              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1760              ceph_cap_string(used), ceph_cap_string(wanted));
1761         if (cap == ci->i_auth_cap) {
1762                 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1763                     !list_empty(&ci->i_cap_snaps))
1764                         goto out;
1765                 if ((used | wanted) & CEPH_CAP_ANY_WR)
1766                         goto out;
1767                 /* Note: it's possible that i_filelock_ref becomes non-zero
1768                  * after dropping auth caps. It doesn't hurt because reply
1769                  * of lock mds request will re-add auth caps. */
1770                 if (atomic_read(&ci->i_filelock_ref) > 0)
1771                         goto out;
1772         }
1773         /* The inode has cached pages, but it's no longer used.
1774          * we can safely drop it */
1775         if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1776             !(oissued & CEPH_CAP_FILE_CACHE)) {
1777           used = 0;
1778           oissued = 0;
1779         }
1780         if ((used | wanted) & ~oissued & mine)
1781                 goto out;   /* we need these caps */
1782
1783         if (oissued) {
1784                 /* we aren't the only cap.. just remove us */
1785                 __ceph_remove_cap(cap, true);
1786                 (*remaining)--;
1787         } else {
1788                 struct dentry *dentry;
1789                 /* try dropping referring dentries */
1790                 spin_unlock(&ci->i_ceph_lock);
1791                 dentry = d_find_any_alias(inode);
1792                 if (dentry && drop_negative_children(dentry)) {
1793                         int count;
1794                         dput(dentry);
1795                         d_prune_aliases(inode);
1796                         count = atomic_read(&inode->i_count);
1797                         if (count == 1)
1798                                 (*remaining)--;
1799                         dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1800                              inode, cap, count);
1801                 } else {
1802                         dput(dentry);
1803                 }
1804                 return 0;
1805         }
1806
1807 out:
1808         spin_unlock(&ci->i_ceph_lock);
1809         return 0;
1810 }
1811
1812 /*
1813  * Trim session cap count down to some max number.
1814  */
1815 int ceph_trim_caps(struct ceph_mds_client *mdsc,
1816                    struct ceph_mds_session *session,
1817                    int max_caps)
1818 {
1819         int trim_caps = session->s_nr_caps - max_caps;
1820
1821         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1822              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1823         if (trim_caps > 0) {
1824                 int remaining = trim_caps;
1825
1826                 ceph_iterate_session_caps(session, trim_caps_cb, &remaining);
1827                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1828                      session->s_mds, session->s_nr_caps, max_caps,
1829                         trim_caps - remaining);
1830         }
1831
1832         ceph_flush_cap_releases(mdsc, session);
1833         return 0;
1834 }
1835
1836 static int check_caps_flush(struct ceph_mds_client *mdsc,
1837                             u64 want_flush_tid)
1838 {
1839         int ret = 1;
1840
1841         spin_lock(&mdsc->cap_dirty_lock);
1842         if (!list_empty(&mdsc->cap_flush_list)) {
1843                 struct ceph_cap_flush *cf =
1844                         list_first_entry(&mdsc->cap_flush_list,
1845                                          struct ceph_cap_flush, g_list);
1846                 if (cf->tid <= want_flush_tid) {
1847                         dout("check_caps_flush still flushing tid "
1848                              "%llu <= %llu\n", cf->tid, want_flush_tid);
1849                         ret = 0;
1850                 }
1851         }
1852         spin_unlock(&mdsc->cap_dirty_lock);
1853         return ret;
1854 }
1855
1856 /*
1857  * flush all dirty inode data to disk.
1858  *
1859  * returns true if we've flushed through want_flush_tid
1860  */
1861 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1862                             u64 want_flush_tid)
1863 {
1864         dout("check_caps_flush want %llu\n", want_flush_tid);
1865
1866         wait_event(mdsc->cap_flushing_wq,
1867                    check_caps_flush(mdsc, want_flush_tid));
1868
1869         dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1870 }
1871
1872 /*
1873  * called under s_mutex
1874  */
1875 static void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1876                                    struct ceph_mds_session *session)
1877 {
1878         struct ceph_msg *msg = NULL;
1879         struct ceph_mds_cap_release *head;
1880         struct ceph_mds_cap_item *item;
1881         struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
1882         struct ceph_cap *cap;
1883         LIST_HEAD(tmp_list);
1884         int num_cap_releases;
1885         __le32  barrier, *cap_barrier;
1886
1887         down_read(&osdc->lock);
1888         barrier = cpu_to_le32(osdc->epoch_barrier);
1889         up_read(&osdc->lock);
1890
1891         spin_lock(&session->s_cap_lock);
1892 again:
1893         list_splice_init(&session->s_cap_releases, &tmp_list);
1894         num_cap_releases = session->s_num_cap_releases;
1895         session->s_num_cap_releases = 0;
1896         spin_unlock(&session->s_cap_lock);
1897
1898         while (!list_empty(&tmp_list)) {
1899                 if (!msg) {
1900                         msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1901                                         PAGE_SIZE, GFP_NOFS, false);
1902                         if (!msg)
1903                                 goto out_err;
1904                         head = msg->front.iov_base;
1905                         head->num = cpu_to_le32(0);
1906                         msg->front.iov_len = sizeof(*head);
1907
1908                         msg->hdr.version = cpu_to_le16(2);
1909                         msg->hdr.compat_version = cpu_to_le16(1);
1910                 }
1911
1912                 cap = list_first_entry(&tmp_list, struct ceph_cap,
1913                                         session_caps);
1914                 list_del(&cap->session_caps);
1915                 num_cap_releases--;
1916
1917                 head = msg->front.iov_base;
1918                 put_unaligned_le32(get_unaligned_le32(&head->num) + 1,
1919                                    &head->num);
1920                 item = msg->front.iov_base + msg->front.iov_len;
1921                 item->ino = cpu_to_le64(cap->cap_ino);
1922                 item->cap_id = cpu_to_le64(cap->cap_id);
1923                 item->migrate_seq = cpu_to_le32(cap->mseq);
1924                 item->seq = cpu_to_le32(cap->issue_seq);
1925                 msg->front.iov_len += sizeof(*item);
1926
1927                 ceph_put_cap(mdsc, cap);
1928
1929                 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1930                         // Append cap_barrier field
1931                         cap_barrier = msg->front.iov_base + msg->front.iov_len;
1932                         *cap_barrier = barrier;
1933                         msg->front.iov_len += sizeof(*cap_barrier);
1934
1935                         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1936                         dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1937                         ceph_con_send(&session->s_con, msg);
1938                         msg = NULL;
1939                 }
1940         }
1941
1942         BUG_ON(num_cap_releases != 0);
1943
1944         spin_lock(&session->s_cap_lock);
1945         if (!list_empty(&session->s_cap_releases))
1946                 goto again;
1947         spin_unlock(&session->s_cap_lock);
1948
1949         if (msg) {
1950                 // Append cap_barrier field
1951                 cap_barrier = msg->front.iov_base + msg->front.iov_len;
1952                 *cap_barrier = barrier;
1953                 msg->front.iov_len += sizeof(*cap_barrier);
1954
1955                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1956                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1957                 ceph_con_send(&session->s_con, msg);
1958         }
1959         return;
1960 out_err:
1961         pr_err("send_cap_releases mds%d, failed to allocate message\n",
1962                 session->s_mds);
1963         spin_lock(&session->s_cap_lock);
1964         list_splice(&tmp_list, &session->s_cap_releases);
1965         session->s_num_cap_releases += num_cap_releases;
1966         spin_unlock(&session->s_cap_lock);
1967 }
1968
1969 static void ceph_cap_release_work(struct work_struct *work)
1970 {
1971         struct ceph_mds_session *session =
1972                 container_of(work, struct ceph_mds_session, s_cap_release_work);
1973
1974         mutex_lock(&session->s_mutex);
1975         if (session->s_state == CEPH_MDS_SESSION_OPEN ||
1976             session->s_state == CEPH_MDS_SESSION_HUNG)
1977                 ceph_send_cap_releases(session->s_mdsc, session);
1978         mutex_unlock(&session->s_mutex);
1979         ceph_put_mds_session(session);
1980 }
1981
1982 void ceph_flush_cap_releases(struct ceph_mds_client *mdsc,
1983                              struct ceph_mds_session *session)
1984 {
1985         if (mdsc->stopping)
1986                 return;
1987
1988         ceph_get_mds_session(session);
1989         if (queue_work(mdsc->fsc->cap_wq,
1990                        &session->s_cap_release_work)) {
1991                 dout("cap release work queued\n");
1992         } else {
1993                 ceph_put_mds_session(session);
1994                 dout("failed to queue cap release work\n");
1995         }
1996 }
1997
1998 /*
1999  * caller holds session->s_cap_lock
2000  */
2001 void __ceph_queue_cap_release(struct ceph_mds_session *session,
2002                               struct ceph_cap *cap)
2003 {
2004         list_add_tail(&cap->session_caps, &session->s_cap_releases);
2005         session->s_num_cap_releases++;
2006
2007         if (!(session->s_num_cap_releases % CEPH_CAPS_PER_RELEASE))
2008                 ceph_flush_cap_releases(session->s_mdsc, session);
2009 }
2010
2011 static void ceph_cap_reclaim_work(struct work_struct *work)
2012 {
2013         struct ceph_mds_client *mdsc =
2014                 container_of(work, struct ceph_mds_client, cap_reclaim_work);
2015         int ret = ceph_trim_dentries(mdsc);
2016         if (ret == -EAGAIN)
2017                 ceph_queue_cap_reclaim_work(mdsc);
2018 }
2019
2020 void ceph_queue_cap_reclaim_work(struct ceph_mds_client *mdsc)
2021 {
2022         if (mdsc->stopping)
2023                 return;
2024
2025         if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_reclaim_work)) {
2026                 dout("caps reclaim work queued\n");
2027         } else {
2028                 dout("failed to queue caps release work\n");
2029         }
2030 }
2031
2032 void ceph_reclaim_caps_nr(struct ceph_mds_client *mdsc, int nr)
2033 {
2034         int val;
2035         if (!nr)
2036                 return;
2037         val = atomic_add_return(nr, &mdsc->cap_reclaim_pending);
2038         if ((val % CEPH_CAPS_PER_RELEASE) < nr) {
2039                 atomic_set(&mdsc->cap_reclaim_pending, 0);
2040                 ceph_queue_cap_reclaim_work(mdsc);
2041         }
2042 }
2043
2044 /*
2045  * requests
2046  */
2047
2048 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
2049                                     struct inode *dir)
2050 {
2051         struct ceph_inode_info *ci = ceph_inode(dir);
2052         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
2053         struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
2054         size_t size = sizeof(struct ceph_mds_reply_dir_entry);
2055         unsigned int num_entries;
2056         int order;
2057
2058         spin_lock(&ci->i_ceph_lock);
2059         num_entries = ci->i_files + ci->i_subdirs;
2060         spin_unlock(&ci->i_ceph_lock);
2061         num_entries = max(num_entries, 1U);
2062         num_entries = min(num_entries, opt->max_readdir);
2063
2064         order = get_order(size * num_entries);
2065         while (order >= 0) {
2066                 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
2067                                                              __GFP_NOWARN,
2068                                                              order);
2069                 if (rinfo->dir_entries)
2070                         break;
2071                 order--;
2072         }
2073         if (!rinfo->dir_entries)
2074                 return -ENOMEM;
2075
2076         num_entries = (PAGE_SIZE << order) / size;
2077         num_entries = min(num_entries, opt->max_readdir);
2078
2079         rinfo->dir_buf_size = PAGE_SIZE << order;
2080         req->r_num_caps = num_entries + 1;
2081         req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
2082         req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
2083         return 0;
2084 }
2085
2086 /*
2087  * Create an mds request.
2088  */
2089 struct ceph_mds_request *
2090 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
2091 {
2092         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
2093         struct timespec64 ts;
2094
2095         if (!req)
2096                 return ERR_PTR(-ENOMEM);
2097
2098         mutex_init(&req->r_fill_mutex);
2099         req->r_mdsc = mdsc;
2100         req->r_started = jiffies;
2101         req->r_resend_mds = -1;
2102         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
2103         INIT_LIST_HEAD(&req->r_unsafe_target_item);
2104         req->r_fmode = -1;
2105         kref_init(&req->r_kref);
2106         RB_CLEAR_NODE(&req->r_node);
2107         INIT_LIST_HEAD(&req->r_wait);
2108         init_completion(&req->r_completion);
2109         init_completion(&req->r_safe_completion);
2110         INIT_LIST_HEAD(&req->r_unsafe_item);
2111
2112         ktime_get_coarse_real_ts64(&ts);
2113         req->r_stamp = timespec64_trunc(ts, mdsc->fsc->sb->s_time_gran);
2114
2115         req->r_op = op;
2116         req->r_direct_mode = mode;
2117         return req;
2118 }
2119
2120 /*
2121  * return oldest (lowest) request, tid in request tree, 0 if none.
2122  *
2123  * called under mdsc->mutex.
2124  */
2125 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
2126 {
2127         if (RB_EMPTY_ROOT(&mdsc->request_tree))
2128                 return NULL;
2129         return rb_entry(rb_first(&mdsc->request_tree),
2130                         struct ceph_mds_request, r_node);
2131 }
2132
2133 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
2134 {
2135         return mdsc->oldest_tid;
2136 }
2137
2138 /*
2139  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
2140  * on build_path_from_dentry in fs/cifs/dir.c.
2141  *
2142  * If @stop_on_nosnap, generate path relative to the first non-snapped
2143  * inode.
2144  *
2145  * Encode hidden .snap dirs as a double /, i.e.
2146  *   foo/.snap/bar -> foo//bar
2147  */
2148 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *pbase,
2149                            int stop_on_nosnap)
2150 {
2151         struct dentry *temp;
2152         char *path;
2153         int pos;
2154         unsigned seq;
2155         u64 base;
2156
2157         if (!dentry)
2158                 return ERR_PTR(-EINVAL);
2159
2160         path = __getname();
2161         if (!path)
2162                 return ERR_PTR(-ENOMEM);
2163 retry:
2164         pos = PATH_MAX - 1;
2165         path[pos] = '\0';
2166
2167         seq = read_seqbegin(&rename_lock);
2168         rcu_read_lock();
2169         temp = dentry;
2170         for (;;) {
2171                 struct inode *inode;
2172
2173                 spin_lock(&temp->d_lock);
2174                 inode = d_inode(temp);
2175                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
2176                         dout("build_path path+%d: %p SNAPDIR\n",
2177                              pos, temp);
2178                 } else if (stop_on_nosnap && inode && dentry != temp &&
2179                            ceph_snap(inode) == CEPH_NOSNAP) {
2180                         spin_unlock(&temp->d_lock);
2181                         pos++; /* get rid of any prepended '/' */
2182                         break;
2183                 } else {
2184                         pos -= temp->d_name.len;
2185                         if (pos < 0) {
2186                                 spin_unlock(&temp->d_lock);
2187                                 break;
2188                         }
2189                         memcpy(path + pos, temp->d_name.name, temp->d_name.len);
2190                 }
2191                 spin_unlock(&temp->d_lock);
2192                 temp = READ_ONCE(temp->d_parent);
2193
2194                 /* Are we at the root? */
2195                 if (IS_ROOT(temp))
2196                         break;
2197
2198                 /* Are we out of buffer? */
2199                 if (--pos < 0)
2200                         break;
2201
2202                 path[pos] = '/';
2203         }
2204         base = ceph_ino(d_inode(temp));
2205         rcu_read_unlock();
2206
2207         if (read_seqretry(&rename_lock, seq))
2208                 goto retry;
2209
2210         if (pos < 0) {
2211                 /*
2212                  * A rename didn't occur, but somehow we didn't end up where
2213                  * we thought we would. Throw a warning and try again.
2214                  */
2215                 pr_warn("build_path did not end path lookup where "
2216                         "expected, pos is %d\n", pos);
2217                 goto retry;
2218         }
2219
2220         *pbase = base;
2221         *plen = PATH_MAX - 1 - pos;
2222         dout("build_path on %p %d built %llx '%.*s'\n",
2223              dentry, d_count(dentry), base, *plen, path + pos);
2224         return path + pos;
2225 }
2226
2227 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
2228                              const char **ppath, int *ppathlen, u64 *pino,
2229                              bool *pfreepath, bool parent_locked)
2230 {
2231         char *path;
2232
2233         rcu_read_lock();
2234         if (!dir)
2235                 dir = d_inode_rcu(dentry->d_parent);
2236         if (dir && parent_locked && ceph_snap(dir) == CEPH_NOSNAP) {
2237                 *pino = ceph_ino(dir);
2238                 rcu_read_unlock();
2239                 *ppath = dentry->d_name.name;
2240                 *ppathlen = dentry->d_name.len;
2241                 return 0;
2242         }
2243         rcu_read_unlock();
2244         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2245         if (IS_ERR(path))
2246                 return PTR_ERR(path);
2247         *ppath = path;
2248         *pfreepath = true;
2249         return 0;
2250 }
2251
2252 static int build_inode_path(struct inode *inode,
2253                             const char **ppath, int *ppathlen, u64 *pino,
2254                             bool *pfreepath)
2255 {
2256         struct dentry *dentry;
2257         char *path;
2258
2259         if (ceph_snap(inode) == CEPH_NOSNAP) {
2260                 *pino = ceph_ino(inode);
2261                 *ppathlen = 0;
2262                 return 0;
2263         }
2264         dentry = d_find_alias(inode);
2265         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2266         dput(dentry);
2267         if (IS_ERR(path))
2268                 return PTR_ERR(path);
2269         *ppath = path;
2270         *pfreepath = true;
2271         return 0;
2272 }
2273
2274 /*
2275  * request arguments may be specified via an inode *, a dentry *, or
2276  * an explicit ino+path.
2277  */
2278 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
2279                                   struct inode *rdiri, const char *rpath,
2280                                   u64 rino, const char **ppath, int *pathlen,
2281                                   u64 *ino, bool *freepath, bool parent_locked)
2282 {
2283         int r = 0;
2284
2285         if (rinode) {
2286                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2287                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2288                      ceph_snap(rinode));
2289         } else if (rdentry) {
2290                 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
2291                                         freepath, parent_locked);
2292                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
2293                      *ppath);
2294         } else if (rpath || rino) {
2295                 *ino = rino;
2296                 *ppath = rpath;
2297                 *pathlen = rpath ? strlen(rpath) : 0;
2298                 dout(" path %.*s\n", *pathlen, rpath);
2299         }
2300
2301         return r;
2302 }
2303
2304 /*
2305  * called under mdsc->mutex
2306  */
2307 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
2308                                                struct ceph_mds_request *req,
2309                                                int mds, bool drop_cap_releases)
2310 {
2311         struct ceph_msg *msg;
2312         struct ceph_mds_request_head *head;
2313         const char *path1 = NULL;
2314         const char *path2 = NULL;
2315         u64 ino1 = 0, ino2 = 0;
2316         int pathlen1 = 0, pathlen2 = 0;
2317         bool freepath1 = false, freepath2 = false;
2318         int len;
2319         u16 releases;
2320         void *p, *end;
2321         int ret;
2322
2323         ret = set_request_path_attr(req->r_inode, req->r_dentry,
2324                               req->r_parent, req->r_path1, req->r_ino1.ino,
2325                               &path1, &pathlen1, &ino1, &freepath1,
2326                               test_bit(CEPH_MDS_R_PARENT_LOCKED,
2327                                         &req->r_req_flags));
2328         if (ret < 0) {
2329                 msg = ERR_PTR(ret);
2330                 goto out;
2331         }
2332
2333         /* If r_old_dentry is set, then assume that its parent is locked */
2334         ret = set_request_path_attr(NULL, req->r_old_dentry,
2335                               req->r_old_dentry_dir,
2336                               req->r_path2, req->r_ino2.ino,
2337                               &path2, &pathlen2, &ino2, &freepath2, true);
2338         if (ret < 0) {
2339                 msg = ERR_PTR(ret);
2340                 goto out_free1;
2341         }
2342
2343         len = sizeof(*head) +
2344                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2345                 sizeof(struct ceph_timespec);
2346
2347         /* calculate (max) length for cap releases */
2348         len += sizeof(struct ceph_mds_request_release) *
2349                 (!!req->r_inode_drop + !!req->r_dentry_drop +
2350                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2351         if (req->r_dentry_drop)
2352                 len += pathlen1;
2353         if (req->r_old_dentry_drop)
2354                 len += pathlen2;
2355
2356         msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
2357         if (!msg) {
2358                 msg = ERR_PTR(-ENOMEM);
2359                 goto out_free2;
2360         }
2361
2362         msg->hdr.version = cpu_to_le16(2);
2363         msg->hdr.tid = cpu_to_le64(req->r_tid);
2364
2365         head = msg->front.iov_base;
2366         p = msg->front.iov_base + sizeof(*head);
2367         end = msg->front.iov_base + msg->front.iov_len;
2368
2369         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2370         head->op = cpu_to_le32(req->r_op);
2371         head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
2372         head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
2373         head->ino = 0;
2374         head->args = req->r_args;
2375
2376         ceph_encode_filepath(&p, end, ino1, path1);
2377         ceph_encode_filepath(&p, end, ino2, path2);
2378
2379         /* make note of release offset, in case we need to replay */
2380         req->r_request_release_offset = p - msg->front.iov_base;
2381
2382         /* cap releases */
2383         releases = 0;
2384         if (req->r_inode_drop)
2385                 releases += ceph_encode_inode_release(&p,
2386                       req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2387                       mds, req->r_inode_drop, req->r_inode_unless, 0);
2388         if (req->r_dentry_drop)
2389                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
2390                                 req->r_parent, mds, req->r_dentry_drop,
2391                                 req->r_dentry_unless);
2392         if (req->r_old_dentry_drop)
2393                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2394                                 req->r_old_dentry_dir, mds,
2395                                 req->r_old_dentry_drop,
2396                                 req->r_old_dentry_unless);
2397         if (req->r_old_inode_drop)
2398                 releases += ceph_encode_inode_release(&p,
2399                       d_inode(req->r_old_dentry),
2400                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2401
2402         if (drop_cap_releases) {
2403                 releases = 0;
2404                 p = msg->front.iov_base + req->r_request_release_offset;
2405         }
2406
2407         head->num_releases = cpu_to_le16(releases);
2408
2409         /* time stamp */
2410         {
2411                 struct ceph_timespec ts;
2412                 ceph_encode_timespec64(&ts, &req->r_stamp);
2413                 ceph_encode_copy(&p, &ts, sizeof(ts));
2414         }
2415
2416         BUG_ON(p > end);
2417         msg->front.iov_len = p - msg->front.iov_base;
2418         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2419
2420         if (req->r_pagelist) {
2421                 struct ceph_pagelist *pagelist = req->r_pagelist;
2422                 ceph_msg_data_add_pagelist(msg, pagelist);
2423                 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2424         } else {
2425                 msg->hdr.data_len = 0;
2426         }
2427
2428         msg->hdr.data_off = cpu_to_le16(0);
2429
2430 out_free2:
2431         if (freepath2)
2432                 ceph_mdsc_free_path((char *)path2, pathlen2);
2433 out_free1:
2434         if (freepath1)
2435                 ceph_mdsc_free_path((char *)path1, pathlen1);
2436 out:
2437         return msg;
2438 }
2439
2440 /*
2441  * called under mdsc->mutex if error, under no mutex if
2442  * success.
2443  */
2444 static void complete_request(struct ceph_mds_client *mdsc,
2445                              struct ceph_mds_request *req)
2446 {
2447         if (req->r_callback)
2448                 req->r_callback(mdsc, req);
2449         complete_all(&req->r_completion);
2450 }
2451
2452 /*
2453  * called under mdsc->mutex
2454  */
2455 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2456                                   struct ceph_mds_request *req,
2457                                   int mds, bool drop_cap_releases)
2458 {
2459         struct ceph_mds_request_head *rhead;
2460         struct ceph_msg *msg;
2461         int flags = 0;
2462
2463         req->r_attempts++;
2464         if (req->r_inode) {
2465                 struct ceph_cap *cap =
2466                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2467
2468                 if (cap)
2469                         req->r_sent_on_mseq = cap->mseq;
2470                 else
2471                         req->r_sent_on_mseq = -1;
2472         }
2473         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2474              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2475
2476         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2477                 void *p;
2478                 /*
2479                  * Replay.  Do not regenerate message (and rebuild
2480                  * paths, etc.); just use the original message.
2481                  * Rebuilding paths will break for renames because
2482                  * d_move mangles the src name.
2483                  */
2484                 msg = req->r_request;
2485                 rhead = msg->front.iov_base;
2486
2487                 flags = le32_to_cpu(rhead->flags);
2488                 flags |= CEPH_MDS_FLAG_REPLAY;
2489                 rhead->flags = cpu_to_le32(flags);
2490
2491                 if (req->r_target_inode)
2492                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2493
2494                 rhead->num_retry = req->r_attempts - 1;
2495
2496                 /* remove cap/dentry releases from message */
2497                 rhead->num_releases = 0;
2498
2499                 /* time stamp */
2500                 p = msg->front.iov_base + req->r_request_release_offset;
2501                 {
2502                         struct ceph_timespec ts;
2503                         ceph_encode_timespec64(&ts, &req->r_stamp);
2504                         ceph_encode_copy(&p, &ts, sizeof(ts));
2505                 }
2506
2507                 msg->front.iov_len = p - msg->front.iov_base;
2508                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2509                 return 0;
2510         }
2511
2512         if (req->r_request) {
2513                 ceph_msg_put(req->r_request);
2514                 req->r_request = NULL;
2515         }
2516         msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2517         if (IS_ERR(msg)) {
2518                 req->r_err = PTR_ERR(msg);
2519                 return PTR_ERR(msg);
2520         }
2521         req->r_request = msg;
2522
2523         rhead = msg->front.iov_base;
2524         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2525         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2526                 flags |= CEPH_MDS_FLAG_REPLAY;
2527         if (req->r_parent)
2528                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2529         rhead->flags = cpu_to_le32(flags);
2530         rhead->num_fwd = req->r_num_fwd;
2531         rhead->num_retry = req->r_attempts - 1;
2532         rhead->ino = 0;
2533
2534         dout(" r_parent = %p\n", req->r_parent);
2535         return 0;
2536 }
2537
2538 /*
2539  * called under mdsc->mutex
2540  */
2541 static int __send_request(struct ceph_mds_client *mdsc,
2542                           struct ceph_mds_session *session,
2543                           struct ceph_mds_request *req,
2544                           bool drop_cap_releases)
2545 {
2546         int err;
2547
2548         err = __prepare_send_request(mdsc, req, session->s_mds,
2549                                      drop_cap_releases);
2550         if (!err) {
2551                 ceph_msg_get(req->r_request);
2552                 ceph_con_send(&session->s_con, req->r_request);
2553         }
2554
2555         return err;
2556 }
2557
2558 /*
2559  * send request, or put it on the appropriate wait list.
2560  */
2561 static void __do_request(struct ceph_mds_client *mdsc,
2562                         struct ceph_mds_request *req)
2563 {
2564         struct ceph_mds_session *session = NULL;
2565         int mds = -1;
2566         int err = 0;
2567         bool random;
2568
2569         if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2570                 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2571                         __unregister_request(mdsc, req);
2572                 return;
2573         }
2574
2575         if (req->r_timeout &&
2576             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2577                 dout("do_request timed out\n");
2578                 err = -EIO;
2579                 goto finish;
2580         }
2581         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2582                 dout("do_request forced umount\n");
2583                 err = -EIO;
2584                 goto finish;
2585         }
2586         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2587                 if (mdsc->mdsmap_err) {
2588                         err = mdsc->mdsmap_err;
2589                         dout("do_request mdsmap err %d\n", err);
2590                         goto finish;
2591                 }
2592                 if (mdsc->mdsmap->m_epoch == 0) {
2593                         dout("do_request no mdsmap, waiting for map\n");
2594                         list_add(&req->r_wait, &mdsc->waiting_for_map);
2595                         return;
2596                 }
2597                 if (!(mdsc->fsc->mount_options->flags &
2598                       CEPH_MOUNT_OPT_MOUNTWAIT) &&
2599                     !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2600                         err = -EHOSTUNREACH;
2601                         goto finish;
2602                 }
2603         }
2604
2605         put_request_session(req);
2606
2607         mds = __choose_mds(mdsc, req, &random);
2608         if (mds < 0 ||
2609             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2610                 dout("do_request no mds or not active, waiting for map\n");
2611                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2612                 return;
2613         }
2614
2615         /* get, open session */
2616         session = __ceph_lookup_mds_session(mdsc, mds);
2617         if (!session) {
2618                 session = register_session(mdsc, mds);
2619                 if (IS_ERR(session)) {
2620                         err = PTR_ERR(session);
2621                         goto finish;
2622                 }
2623         }
2624         req->r_session = ceph_get_mds_session(session);
2625
2626         dout("do_request mds%d session %p state %s\n", mds, session,
2627              ceph_session_state_name(session->s_state));
2628         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2629             session->s_state != CEPH_MDS_SESSION_HUNG) {
2630                 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2631                         err = -EACCES;
2632                         goto out_session;
2633                 }
2634                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2635                     session->s_state == CEPH_MDS_SESSION_CLOSING) {
2636                         __open_session(mdsc, session);
2637                         /* retry the same mds later */
2638                         if (random)
2639                                 req->r_resend_mds = mds;
2640                 }
2641                 list_add(&req->r_wait, &session->s_waiting);
2642                 goto out_session;
2643         }
2644
2645         /* send request */
2646         req->r_resend_mds = -1;   /* forget any previous mds hint */
2647
2648         if (req->r_request_started == 0)   /* note request start time */
2649                 req->r_request_started = jiffies;
2650
2651         err = __send_request(mdsc, session, req, false);
2652
2653 out_session:
2654         ceph_put_mds_session(session);
2655 finish:
2656         if (err) {
2657                 dout("__do_request early error %d\n", err);
2658                 req->r_err = err;
2659                 complete_request(mdsc, req);
2660                 __unregister_request(mdsc, req);
2661         }
2662         return;
2663 }
2664
2665 /*
2666  * called under mdsc->mutex
2667  */
2668 static void __wake_requests(struct ceph_mds_client *mdsc,
2669                             struct list_head *head)
2670 {
2671         struct ceph_mds_request *req;
2672         LIST_HEAD(tmp_list);
2673
2674         list_splice_init(head, &tmp_list);
2675
2676         while (!list_empty(&tmp_list)) {
2677                 req = list_entry(tmp_list.next,
2678                                  struct ceph_mds_request, r_wait);
2679                 list_del_init(&req->r_wait);
2680                 dout(" wake request %p tid %llu\n", req, req->r_tid);
2681                 __do_request(mdsc, req);
2682         }
2683 }
2684
2685 /*
2686  * Wake up threads with requests pending for @mds, so that they can
2687  * resubmit their requests to a possibly different mds.
2688  */
2689 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2690 {
2691         struct ceph_mds_request *req;
2692         struct rb_node *p = rb_first(&mdsc->request_tree);
2693
2694         dout("kick_requests mds%d\n", mds);
2695         while (p) {
2696                 req = rb_entry(p, struct ceph_mds_request, r_node);
2697                 p = rb_next(p);
2698                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2699                         continue;
2700                 if (req->r_attempts > 0)
2701                         continue; /* only new requests */
2702                 if (req->r_session &&
2703                     req->r_session->s_mds == mds) {
2704                         dout(" kicking tid %llu\n", req->r_tid);
2705                         list_del_init(&req->r_wait);
2706                         __do_request(mdsc, req);
2707                 }
2708         }
2709 }
2710
2711 int ceph_mdsc_submit_request(struct ceph_mds_client *mdsc, struct inode *dir,
2712                               struct ceph_mds_request *req)
2713 {
2714         int err;
2715
2716         /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2717         if (req->r_inode)
2718                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2719         if (req->r_parent) {
2720                 ceph_get_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
2721                 ihold(req->r_parent);
2722         }
2723         if (req->r_old_dentry_dir)
2724                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2725                                   CEPH_CAP_PIN);
2726
2727         dout("submit_request on %p for inode %p\n", req, dir);
2728         mutex_lock(&mdsc->mutex);
2729         __register_request(mdsc, req, dir);
2730         __do_request(mdsc, req);
2731         err = req->r_err;
2732         mutex_unlock(&mdsc->mutex);
2733         return err;
2734 }
2735
2736 static int ceph_mdsc_wait_request(struct ceph_mds_client *mdsc,
2737                                   struct ceph_mds_request *req)
2738 {
2739         int err;
2740
2741         /* wait */
2742         dout("do_request waiting\n");
2743         if (!req->r_timeout && req->r_wait_for_completion) {
2744                 err = req->r_wait_for_completion(mdsc, req);
2745         } else {
2746                 long timeleft = wait_for_completion_killable_timeout(
2747                                         &req->r_completion,
2748                                         ceph_timeout_jiffies(req->r_timeout));
2749                 if (timeleft > 0)
2750                         err = 0;
2751                 else if (!timeleft)
2752                         err = -EIO;  /* timed out */
2753                 else
2754                         err = timeleft;  /* killed */
2755         }
2756         dout("do_request waited, got %d\n", err);
2757         mutex_lock(&mdsc->mutex);
2758
2759         /* only abort if we didn't race with a real reply */
2760         if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2761                 err = le32_to_cpu(req->r_reply_info.head->result);
2762         } else if (err < 0) {
2763                 dout("aborted request %lld with %d\n", req->r_tid, err);
2764
2765                 /*
2766                  * ensure we aren't running concurrently with
2767                  * ceph_fill_trace or ceph_readdir_prepopulate, which
2768                  * rely on locks (dir mutex) held by our caller.
2769                  */
2770                 mutex_lock(&req->r_fill_mutex);
2771                 req->r_err = err;
2772                 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2773                 mutex_unlock(&req->r_fill_mutex);
2774
2775                 if (req->r_parent &&
2776                     (req->r_op & CEPH_MDS_OP_WRITE))
2777                         ceph_invalidate_dir_request(req);
2778         } else {
2779                 err = req->r_err;
2780         }
2781
2782         mutex_unlock(&mdsc->mutex);
2783         return err;
2784 }
2785
2786 /*
2787  * Synchrously perform an mds request.  Take care of all of the
2788  * session setup, forwarding, retry details.
2789  */
2790 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2791                          struct inode *dir,
2792                          struct ceph_mds_request *req)
2793 {
2794         int err;
2795
2796         dout("do_request on %p\n", req);
2797
2798         /* issue */
2799         err = ceph_mdsc_submit_request(mdsc, dir, req);
2800         if (!err)
2801                 err = ceph_mdsc_wait_request(mdsc, req);
2802         dout("do_request %p done, result %d\n", req, err);
2803         return err;
2804 }
2805
2806 /*
2807  * Invalidate dir's completeness, dentry lease state on an aborted MDS
2808  * namespace request.
2809  */
2810 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2811 {
2812         struct inode *dir = req->r_parent;
2813         struct inode *old_dir = req->r_old_dentry_dir;
2814
2815         dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
2816
2817         ceph_dir_clear_complete(dir);
2818         if (old_dir)
2819                 ceph_dir_clear_complete(old_dir);
2820         if (req->r_dentry)
2821                 ceph_invalidate_dentry_lease(req->r_dentry);
2822         if (req->r_old_dentry)
2823                 ceph_invalidate_dentry_lease(req->r_old_dentry);
2824 }
2825
2826 /*
2827  * Handle mds reply.
2828  *
2829  * We take the session mutex and parse and process the reply immediately.
2830  * This preserves the logical ordering of replies, capabilities, etc., sent
2831  * by the MDS as they are applied to our local cache.
2832  */
2833 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2834 {
2835         struct ceph_mds_client *mdsc = session->s_mdsc;
2836         struct ceph_mds_request *req;
2837         struct ceph_mds_reply_head *head = msg->front.iov_base;
2838         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2839         struct ceph_snap_realm *realm;
2840         u64 tid;
2841         int err, result;
2842         int mds = session->s_mds;
2843
2844         if (msg->front.iov_len < sizeof(*head)) {
2845                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2846                 ceph_msg_dump(msg);
2847                 return;
2848         }
2849
2850         /* get request, session */
2851         tid = le64_to_cpu(msg->hdr.tid);
2852         mutex_lock(&mdsc->mutex);
2853         req = lookup_get_request(mdsc, tid);
2854         if (!req) {
2855                 dout("handle_reply on unknown tid %llu\n", tid);
2856                 mutex_unlock(&mdsc->mutex);
2857                 return;
2858         }
2859         dout("handle_reply %p\n", req);
2860
2861         /* correct session? */
2862         if (req->r_session != session) {
2863                 pr_err("mdsc_handle_reply got %llu on session mds%d"
2864                        " not mds%d\n", tid, session->s_mds,
2865                        req->r_session ? req->r_session->s_mds : -1);
2866                 mutex_unlock(&mdsc->mutex);
2867                 goto out;
2868         }
2869
2870         /* dup? */
2871         if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
2872             (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
2873                 pr_warn("got a dup %s reply on %llu from mds%d\n",
2874                            head->safe ? "safe" : "unsafe", tid, mds);
2875                 mutex_unlock(&mdsc->mutex);
2876                 goto out;
2877         }
2878         if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
2879                 pr_warn("got unsafe after safe on %llu from mds%d\n",
2880                            tid, mds);
2881                 mutex_unlock(&mdsc->mutex);
2882                 goto out;
2883         }
2884
2885         result = le32_to_cpu(head->result);
2886
2887         /*
2888          * Handle an ESTALE
2889          * if we're not talking to the authority, send to them
2890          * if the authority has changed while we weren't looking,
2891          * send to new authority
2892          * Otherwise we just have to return an ESTALE
2893          */
2894         if (result == -ESTALE) {
2895                 dout("got ESTALE on request %llu\n", req->r_tid);
2896                 req->r_resend_mds = -1;
2897                 if (req->r_direct_mode != USE_AUTH_MDS) {
2898                         dout("not using auth, setting for that now\n");
2899                         req->r_direct_mode = USE_AUTH_MDS;
2900                         __do_request(mdsc, req);
2901                         mutex_unlock(&mdsc->mutex);
2902                         goto out;
2903                 } else  {
2904                         int mds = __choose_mds(mdsc, req, NULL);
2905                         if (mds >= 0 && mds != req->r_session->s_mds) {
2906                                 dout("but auth changed, so resending\n");
2907                                 __do_request(mdsc, req);
2908                                 mutex_unlock(&mdsc->mutex);
2909                                 goto out;
2910                         }
2911                 }
2912                 dout("have to return ESTALE on request %llu\n", req->r_tid);
2913         }
2914
2915
2916         if (head->safe) {
2917                 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
2918                 __unregister_request(mdsc, req);
2919
2920                 /* last request during umount? */
2921                 if (mdsc->stopping && !__get_oldest_req(mdsc))
2922                         complete_all(&mdsc->safe_umount_waiters);
2923
2924                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2925                         /*
2926                          * We already handled the unsafe response, now do the
2927                          * cleanup.  No need to examine the response; the MDS
2928                          * doesn't include any result info in the safe
2929                          * response.  And even if it did, there is nothing
2930                          * useful we could do with a revised return value.
2931                          */
2932                         dout("got safe reply %llu, mds%d\n", tid, mds);
2933
2934                         mutex_unlock(&mdsc->mutex);
2935                         goto out;
2936                 }
2937         } else {
2938                 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
2939                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2940                 if (req->r_unsafe_dir) {
2941                         struct ceph_inode_info *ci =
2942                                         ceph_inode(req->r_unsafe_dir);
2943                         spin_lock(&ci->i_unsafe_lock);
2944                         list_add_tail(&req->r_unsafe_dir_item,
2945                                       &ci->i_unsafe_dirops);
2946                         spin_unlock(&ci->i_unsafe_lock);
2947                 }
2948         }
2949
2950         dout("handle_reply tid %lld result %d\n", tid, result);
2951         rinfo = &req->r_reply_info;
2952         if (test_bit(CEPHFS_FEATURE_REPLY_ENCODING, &session->s_features))
2953                 err = parse_reply_info(msg, rinfo, (u64)-1);
2954         else
2955                 err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2956         mutex_unlock(&mdsc->mutex);
2957
2958         mutex_lock(&session->s_mutex);
2959         if (err < 0) {
2960                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2961                 ceph_msg_dump(msg);
2962                 goto out_err;
2963         }
2964
2965         /* snap trace */
2966         realm = NULL;
2967         if (rinfo->snapblob_len) {
2968                 down_write(&mdsc->snap_rwsem);
2969                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2970                                 rinfo->snapblob + rinfo->snapblob_len,
2971                                 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2972                                 &realm);
2973                 downgrade_write(&mdsc->snap_rwsem);
2974         } else {
2975                 down_read(&mdsc->snap_rwsem);
2976         }
2977
2978         /* insert trace into our cache */
2979         mutex_lock(&req->r_fill_mutex);
2980         current->journal_info = req;
2981         err = ceph_fill_trace(mdsc->fsc->sb, req);
2982         if (err == 0) {
2983                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2984                                     req->r_op == CEPH_MDS_OP_LSSNAP))
2985                         ceph_readdir_prepopulate(req, req->r_session);
2986         }
2987         current->journal_info = NULL;
2988         mutex_unlock(&req->r_fill_mutex);
2989
2990         up_read(&mdsc->snap_rwsem);
2991         if (realm)
2992                 ceph_put_snap_realm(mdsc, realm);
2993
2994         if (err == 0) {
2995                 if (req->r_target_inode &&
2996                     test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2997                         struct ceph_inode_info *ci =
2998                                 ceph_inode(req->r_target_inode);
2999                         spin_lock(&ci->i_unsafe_lock);
3000                         list_add_tail(&req->r_unsafe_target_item,
3001                                       &ci->i_unsafe_iops);
3002                         spin_unlock(&ci->i_unsafe_lock);
3003                 }
3004
3005                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
3006         }
3007 out_err:
3008         mutex_lock(&mdsc->mutex);
3009         if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3010                 if (err) {
3011                         req->r_err = err;
3012                 } else {
3013                         req->r_reply =  ceph_msg_get(msg);
3014                         set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
3015                 }
3016         } else {
3017                 dout("reply arrived after request %lld was aborted\n", tid);
3018         }
3019         mutex_unlock(&mdsc->mutex);
3020
3021         mutex_unlock(&session->s_mutex);
3022
3023         /* kick calling process */
3024         complete_request(mdsc, req);
3025 out:
3026         ceph_mdsc_put_request(req);
3027         return;
3028 }
3029
3030
3031
3032 /*
3033  * handle mds notification that our request has been forwarded.
3034  */
3035 static void handle_forward(struct ceph_mds_client *mdsc,
3036                            struct ceph_mds_session *session,
3037                            struct ceph_msg *msg)
3038 {
3039         struct ceph_mds_request *req;
3040         u64 tid = le64_to_cpu(msg->hdr.tid);
3041         u32 next_mds;
3042         u32 fwd_seq;
3043         int err = -EINVAL;
3044         void *p = msg->front.iov_base;
3045         void *end = p + msg->front.iov_len;
3046
3047         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
3048         next_mds = ceph_decode_32(&p);
3049         fwd_seq = ceph_decode_32(&p);
3050
3051         mutex_lock(&mdsc->mutex);
3052         req = lookup_get_request(mdsc, tid);
3053         if (!req) {
3054                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
3055                 goto out;  /* dup reply? */
3056         }
3057
3058         if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3059                 dout("forward tid %llu aborted, unregistering\n", tid);
3060                 __unregister_request(mdsc, req);
3061         } else if (fwd_seq <= req->r_num_fwd) {
3062                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
3063                      tid, next_mds, req->r_num_fwd, fwd_seq);
3064         } else {
3065                 /* resend. forward race not possible; mds would drop */
3066                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
3067                 BUG_ON(req->r_err);
3068                 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
3069                 req->r_attempts = 0;
3070                 req->r_num_fwd = fwd_seq;
3071                 req->r_resend_mds = next_mds;
3072                 put_request_session(req);
3073                 __do_request(mdsc, req);
3074         }
3075         ceph_mdsc_put_request(req);
3076 out:
3077         mutex_unlock(&mdsc->mutex);
3078         return;
3079
3080 bad:
3081         pr_err("mdsc_handle_forward decode error err=%d\n", err);
3082 }
3083
3084 static int __decode_session_metadata(void **p, void *end,
3085                                      bool *blacklisted)
3086 {
3087         /* map<string,string> */
3088         u32 n;
3089         bool err_str;
3090         ceph_decode_32_safe(p, end, n, bad);
3091         while (n-- > 0) {
3092                 u32 len;
3093                 ceph_decode_32_safe(p, end, len, bad);
3094                 ceph_decode_need(p, end, len, bad);
3095                 err_str = !strncmp(*p, "error_string", len);
3096                 *p += len;
3097                 ceph_decode_32_safe(p, end, len, bad);
3098                 ceph_decode_need(p, end, len, bad);
3099                 if (err_str && strnstr(*p, "blacklisted", len))
3100                         *blacklisted = true;
3101                 *p += len;
3102         }
3103         return 0;
3104 bad:
3105         return -1;
3106 }
3107
3108 /*
3109  * handle a mds session control message
3110  */
3111 static void handle_session(struct ceph_mds_session *session,
3112                            struct ceph_msg *msg)
3113 {
3114         struct ceph_mds_client *mdsc = session->s_mdsc;
3115         int mds = session->s_mds;
3116         int msg_version = le16_to_cpu(msg->hdr.version);
3117         void *p = msg->front.iov_base;
3118         void *end = p + msg->front.iov_len;
3119         struct ceph_mds_session_head *h;
3120         u32 op;
3121         u64 seq;
3122         unsigned long features = 0;
3123         int wake = 0;
3124         bool blacklisted = false;
3125
3126         /* decode */
3127         ceph_decode_need(&p, end, sizeof(*h), bad);
3128         h = p;
3129         p += sizeof(*h);
3130
3131         op = le32_to_cpu(h->op);
3132         seq = le64_to_cpu(h->seq);
3133
3134         if (msg_version >= 3) {
3135                 u32 len;
3136                 /* version >= 2, metadata */
3137                 if (__decode_session_metadata(&p, end, &blacklisted) < 0)
3138                         goto bad;
3139                 /* version >= 3, feature bits */
3140                 ceph_decode_32_safe(&p, end, len, bad);
3141                 ceph_decode_need(&p, end, len, bad);
3142                 memcpy(&features, p, min_t(size_t, len, sizeof(features)));
3143                 p += len;
3144         }
3145
3146         mutex_lock(&mdsc->mutex);
3147         if (op == CEPH_SESSION_CLOSE) {
3148                 ceph_get_mds_session(session);
3149                 __unregister_session(mdsc, session);
3150         }
3151         /* FIXME: this ttl calculation is generous */
3152         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
3153         mutex_unlock(&mdsc->mutex);
3154
3155         mutex_lock(&session->s_mutex);
3156
3157         dout("handle_session mds%d %s %p state %s seq %llu\n",
3158              mds, ceph_session_op_name(op), session,
3159              ceph_session_state_name(session->s_state), seq);
3160
3161         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
3162                 session->s_state = CEPH_MDS_SESSION_OPEN;
3163                 pr_info("mds%d came back\n", session->s_mds);
3164         }
3165
3166         switch (op) {
3167         case CEPH_SESSION_OPEN:
3168                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3169                         pr_info("mds%d reconnect success\n", session->s_mds);
3170                 session->s_state = CEPH_MDS_SESSION_OPEN;
3171                 session->s_features = features;
3172                 renewed_caps(mdsc, session, 0);
3173                 wake = 1;
3174                 if (mdsc->stopping)
3175                         __close_session(mdsc, session);
3176                 break;
3177
3178         case CEPH_SESSION_RENEWCAPS:
3179                 if (session->s_renew_seq == seq)
3180                         renewed_caps(mdsc, session, 1);
3181                 break;
3182
3183         case CEPH_SESSION_CLOSE:
3184                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3185                         pr_info("mds%d reconnect denied\n", session->s_mds);
3186                 session->s_state = CEPH_MDS_SESSION_CLOSED;
3187                 cleanup_session_requests(mdsc, session);
3188                 remove_session_caps(session);
3189                 wake = 2; /* for good measure */
3190                 wake_up_all(&mdsc->session_close_wq);
3191                 break;
3192
3193         case CEPH_SESSION_STALE:
3194                 pr_info("mds%d caps went stale, renewing\n",
3195                         session->s_mds);
3196                 spin_lock(&session->s_gen_ttl_lock);
3197                 session->s_cap_gen++;
3198                 session->s_cap_ttl = jiffies - 1;
3199                 spin_unlock(&session->s_gen_ttl_lock);
3200                 send_renew_caps(mdsc, session);
3201                 break;
3202
3203         case CEPH_SESSION_RECALL_STATE:
3204                 ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
3205                 break;
3206
3207         case CEPH_SESSION_FLUSHMSG:
3208                 send_flushmsg_ack(mdsc, session, seq);
3209                 break;
3210
3211         case CEPH_SESSION_FORCE_RO:
3212                 dout("force_session_readonly %p\n", session);
3213                 spin_lock(&session->s_cap_lock);
3214                 session->s_readonly = true;
3215                 spin_unlock(&session->s_cap_lock);
3216                 wake_up_session_caps(session, FORCE_RO);
3217                 break;
3218
3219         case CEPH_SESSION_REJECT:
3220                 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
3221                 pr_info("mds%d rejected session\n", session->s_mds);
3222                 session->s_state = CEPH_MDS_SESSION_REJECTED;
3223                 cleanup_session_requests(mdsc, session);
3224                 remove_session_caps(session);
3225                 if (blacklisted)
3226                         mdsc->fsc->blacklisted = true;
3227                 wake = 2; /* for good measure */
3228                 break;
3229
3230         default:
3231                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
3232                 WARN_ON(1);
3233         }
3234
3235         mutex_unlock(&session->s_mutex);
3236         if (wake) {
3237                 mutex_lock(&mdsc->mutex);
3238                 __wake_requests(mdsc, &session->s_waiting);
3239                 if (wake == 2)
3240                         kick_requests(mdsc, mds);
3241                 mutex_unlock(&mdsc->mutex);
3242         }
3243         if (op == CEPH_SESSION_CLOSE)
3244                 ceph_put_mds_session(session);
3245         return;
3246
3247 bad:
3248         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
3249                (int)msg->front.iov_len);
3250         ceph_msg_dump(msg);
3251         return;
3252 }
3253
3254 /*
3255  * called under session->mutex.
3256  */
3257 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
3258                                    struct ceph_mds_session *session)
3259 {
3260         struct ceph_mds_request *req, *nreq;
3261         struct rb_node *p;
3262
3263         dout("replay_unsafe_requests mds%d\n", session->s_mds);
3264
3265         mutex_lock(&mdsc->mutex);
3266         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item)
3267                 __send_request(mdsc, session, req, true);
3268
3269         /*
3270          * also re-send old requests when MDS enters reconnect stage. So that MDS
3271          * can process completed request in clientreplay stage.
3272          */
3273         p = rb_first(&mdsc->request_tree);
3274         while (p) {
3275                 req = rb_entry(p, struct ceph_mds_request, r_node);
3276                 p = rb_next(p);
3277                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3278                         continue;
3279                 if (req->r_attempts == 0)
3280                         continue; /* only old requests */
3281                 if (req->r_session &&
3282                     req->r_session->s_mds == session->s_mds)
3283                         __send_request(mdsc, session, req, true);
3284         }
3285         mutex_unlock(&mdsc->mutex);
3286 }
3287
3288 static int send_reconnect_partial(struct ceph_reconnect_state *recon_state)
3289 {
3290         struct ceph_msg *reply;
3291         struct ceph_pagelist *_pagelist;
3292         struct page *page;
3293         __le32 *addr;
3294         int err = -ENOMEM;
3295
3296         if (!recon_state->allow_multi)
3297                 return -ENOSPC;
3298
3299         /* can't handle message that contains both caps and realm */
3300         BUG_ON(!recon_state->nr_caps == !recon_state->nr_realms);
3301
3302         /* pre-allocate new pagelist */
3303         _pagelist = ceph_pagelist_alloc(GFP_NOFS);
3304         if (!_pagelist)
3305                 return -ENOMEM;
3306
3307         reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3308         if (!reply)
3309                 goto fail_msg;
3310
3311         /* placeholder for nr_caps */
3312         err = ceph_pagelist_encode_32(_pagelist, 0);
3313         if (err < 0)
3314                 goto fail;
3315
3316         if (recon_state->nr_caps) {
3317                 /* currently encoding caps */
3318                 err = ceph_pagelist_encode_32(recon_state->pagelist, 0);
3319                 if (err)
3320                         goto fail;
3321         } else {
3322                 /* placeholder for nr_realms (currently encoding relams) */
3323                 err = ceph_pagelist_encode_32(_pagelist, 0);
3324                 if (err < 0)
3325                         goto fail;
3326         }
3327
3328         err = ceph_pagelist_encode_8(recon_state->pagelist, 1);
3329         if (err)
3330                 goto fail;
3331
3332         page = list_first_entry(&recon_state->pagelist->head, struct page, lru);
3333         addr = kmap_atomic(page);
3334         if (recon_state->nr_caps) {
3335                 /* currently encoding caps */
3336                 *addr = cpu_to_le32(recon_state->nr_caps);
3337         } else {
3338                 /* currently encoding relams */
3339                 *(addr + 1) = cpu_to_le32(recon_state->nr_realms);
3340         }
3341         kunmap_atomic(addr);
3342
3343         reply->hdr.version = cpu_to_le16(5);
3344         reply->hdr.compat_version = cpu_to_le16(4);
3345
3346         reply->hdr.data_len = cpu_to_le32(recon_state->pagelist->length);
3347         ceph_msg_data_add_pagelist(reply, recon_state->pagelist);
3348
3349         ceph_con_send(&recon_state->session->s_con, reply);
3350         ceph_pagelist_release(recon_state->pagelist);
3351
3352         recon_state->pagelist = _pagelist;
3353         recon_state->nr_caps = 0;
3354         recon_state->nr_realms = 0;
3355         recon_state->msg_version = 5;
3356         return 0;
3357 fail:
3358         ceph_msg_put(reply);
3359 fail_msg:
3360         ceph_pagelist_release(_pagelist);
3361         return err;
3362 }
3363
3364 /*
3365  * Encode information about a cap for a reconnect with the MDS.
3366  */
3367 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
3368                           void *arg)
3369 {
3370         union {
3371                 struct ceph_mds_cap_reconnect v2;
3372                 struct ceph_mds_cap_reconnect_v1 v1;
3373         } rec;
3374         struct ceph_inode_info *ci = cap->ci;
3375         struct ceph_reconnect_state *recon_state = arg;
3376         struct ceph_pagelist *pagelist = recon_state->pagelist;
3377         int err;
3378         u64 snap_follows;
3379
3380         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
3381              inode, ceph_vinop(inode), cap, cap->cap_id,
3382              ceph_cap_string(cap->issued));
3383
3384         spin_lock(&ci->i_ceph_lock);
3385         cap->seq = 0;        /* reset cap seq */
3386         cap->issue_seq = 0;  /* and issue_seq */
3387         cap->mseq = 0;       /* and migrate_seq */
3388         cap->cap_gen = cap->session->s_cap_gen;
3389
3390         if (recon_state->msg_version >= 2) {
3391                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
3392                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3393                 rec.v2.issued = cpu_to_le32(cap->issued);
3394                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3395                 rec.v2.pathbase = 0;
3396                 rec.v2.flock_len = (__force __le32)
3397                         ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
3398         } else {
3399                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
3400                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3401                 rec.v1.issued = cpu_to_le32(cap->issued);
3402                 rec.v1.size = cpu_to_le64(inode->i_size);
3403                 ceph_encode_timespec64(&rec.v1.mtime, &inode->i_mtime);
3404                 ceph_encode_timespec64(&rec.v1.atime, &inode->i_atime);
3405                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3406                 rec.v1.pathbase = 0;
3407         }
3408
3409         if (list_empty(&ci->i_cap_snaps)) {
3410                 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
3411         } else {
3412                 struct ceph_cap_snap *capsnap =
3413                         list_first_entry(&ci->i_cap_snaps,
3414                                          struct ceph_cap_snap, ci_item);
3415                 snap_follows = capsnap->follows;
3416         }
3417         spin_unlock(&ci->i_ceph_lock);
3418
3419         if (recon_state->msg_version >= 2) {
3420                 int num_fcntl_locks, num_flock_locks;
3421                 struct ceph_filelock *flocks = NULL;
3422                 size_t struct_len, total_len = sizeof(u64);
3423                 u8 struct_v = 0;
3424
3425 encode_again:
3426                 if (rec.v2.flock_len) {
3427                         ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
3428                 } else {
3429                         num_fcntl_locks = 0;
3430                         num_flock_locks = 0;
3431                 }
3432                 if (num_fcntl_locks + num_flock_locks > 0) {
3433                         flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
3434                                                sizeof(struct ceph_filelock),
3435                                                GFP_NOFS);
3436                         if (!flocks) {
3437                                 err = -ENOMEM;
3438                                 goto out_err;
3439                         }
3440                         err = ceph_encode_locks_to_buffer(inode, flocks,
3441                                                           num_fcntl_locks,
3442                                                           num_flock_locks);
3443                         if (err) {
3444                                 kfree(flocks);
3445                                 flocks = NULL;
3446                                 if (err == -ENOSPC)
3447                                         goto encode_again;
3448                                 goto out_err;
3449                         }
3450                 } else {
3451                         kfree(flocks);
3452                         flocks = NULL;
3453                 }
3454
3455                 if (recon_state->msg_version >= 3) {
3456                         /* version, compat_version and struct_len */
3457                         total_len += 2 * sizeof(u8) + sizeof(u32);
3458                         struct_v = 2;
3459                 }
3460                 /*
3461                  * number of encoded locks is stable, so copy to pagelist
3462                  */
3463                 struct_len = 2 * sizeof(u32) +
3464                             (num_fcntl_locks + num_flock_locks) *
3465                             sizeof(struct ceph_filelock);
3466                 rec.v2.flock_len = cpu_to_le32(struct_len);
3467
3468                 struct_len += sizeof(u32) + sizeof(rec.v2);
3469
3470                 if (struct_v >= 2)
3471                         struct_len += sizeof(u64); /* snap_follows */
3472
3473                 total_len += struct_len;
3474
3475                 if (pagelist->length + total_len > RECONNECT_MAX_SIZE) {
3476                         err = send_reconnect_partial(recon_state);
3477                         if (err)
3478                                 goto out_freeflocks;
3479                         pagelist = recon_state->pagelist;
3480                 }
3481
3482                 err = ceph_pagelist_reserve(pagelist, total_len);
3483                 if (err)
3484                         goto out_freeflocks;
3485
3486                 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3487                 if (recon_state->msg_version >= 3) {
3488                         ceph_pagelist_encode_8(pagelist, struct_v);
3489                         ceph_pagelist_encode_8(pagelist, 1);
3490                         ceph_pagelist_encode_32(pagelist, struct_len);
3491                 }
3492                 ceph_pagelist_encode_string(pagelist, NULL, 0);
3493                 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
3494                 ceph_locks_to_pagelist(flocks, pagelist,
3495                                        num_fcntl_locks, num_flock_locks);
3496                 if (struct_v >= 2)
3497                         ceph_pagelist_encode_64(pagelist, snap_follows);
3498 out_freeflocks:
3499                 kfree(flocks);
3500         } else {
3501                 u64 pathbase = 0;
3502                 int pathlen = 0;
3503                 char *path = NULL;
3504                 struct dentry *dentry;
3505
3506                 dentry = d_find_alias(inode);
3507                 if (dentry) {
3508                         path = ceph_mdsc_build_path(dentry,
3509                                                 &pathlen, &pathbase, 0);
3510                         dput(dentry);
3511                         if (IS_ERR(path)) {
3512                                 err = PTR_ERR(path);
3513                                 goto out_err;
3514                         }
3515                         rec.v1.pathbase = cpu_to_le64(pathbase);
3516                 }
3517
3518                 err = ceph_pagelist_reserve(pagelist,
3519                                             sizeof(u64) + sizeof(u32) +
3520                                             pathlen + sizeof(rec.v1));
3521                 if (err) {
3522                         goto out_freepath;
3523                 }
3524
3525                 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3526                 ceph_pagelist_encode_string(pagelist, path, pathlen);
3527                 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
3528 out_freepath:
3529                 ceph_mdsc_free_path(path, pathlen);
3530         }
3531
3532 out_err:
3533         if (err >= 0)
3534                 recon_state->nr_caps++;
3535         return err;
3536 }
3537
3538 static int encode_snap_realms(struct ceph_mds_client *mdsc,
3539                               struct ceph_reconnect_state *recon_state)
3540 {
3541         struct rb_node *p;
3542         struct ceph_pagelist *pagelist = recon_state->pagelist;
3543         int err = 0;
3544
3545         if (recon_state->msg_version >= 4) {
3546                 err = ceph_pagelist_encode_32(pagelist, mdsc->num_snap_realms);
3547                 if (err < 0)
3548                         goto fail;
3549         }
3550
3551         /*
3552          * snaprealms.  we provide mds with the ino, seq (version), and
3553          * parent for all of our realms.  If the mds has any newer info,
3554          * it will tell us.
3555          */
3556         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3557                 struct ceph_snap_realm *realm =
3558                        rb_entry(p, struct ceph_snap_realm, node);
3559                 struct ceph_mds_snaprealm_reconnect sr_rec;
3560
3561                 if (recon_state->msg_version >= 4) {
3562                         size_t need = sizeof(u8) * 2 + sizeof(u32) +
3563                                       sizeof(sr_rec);
3564
3565                         if (pagelist->length + need > RECONNECT_MAX_SIZE) {
3566                                 err = send_reconnect_partial(recon_state);
3567                                 if (err)
3568                                         goto fail;
3569                                 pagelist = recon_state->pagelist;
3570                         }
3571
3572                         err = ceph_pagelist_reserve(pagelist, need);
3573                         if (err)
3574                                 goto fail;
3575
3576                         ceph_pagelist_encode_8(pagelist, 1);
3577                         ceph_pagelist_encode_8(pagelist, 1);
3578                         ceph_pagelist_encode_32(pagelist, sizeof(sr_rec));
3579                 }
3580
3581                 dout(" adding snap realm %llx seq %lld parent %llx\n",
3582                      realm->ino, realm->seq, realm->parent_ino);
3583                 sr_rec.ino = cpu_to_le64(realm->ino);
3584                 sr_rec.seq = cpu_to_le64(realm->seq);
3585                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3586
3587                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3588                 if (err)
3589                         goto fail;
3590
3591                 recon_state->nr_realms++;
3592         }
3593 fail:
3594         return err;
3595 }
3596
3597
3598 /*
3599  * If an MDS fails and recovers, clients need to reconnect in order to
3600  * reestablish shared state.  This includes all caps issued through
3601  * this session _and_ the snap_realm hierarchy.  Because it's not
3602  * clear which snap realms the mds cares about, we send everything we
3603  * know about.. that ensures we'll then get any new info the
3604  * recovering MDS might have.
3605  *
3606  * This is a relatively heavyweight operation, but it's rare.
3607  *
3608  * called with mdsc->mutex held.
3609  */
3610 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3611                                struct ceph_mds_session *session)
3612 {
3613         struct ceph_msg *reply;
3614         int mds = session->s_mds;
3615         int err = -ENOMEM;
3616         struct ceph_reconnect_state recon_state = {
3617                 .session = session,
3618         };
3619         LIST_HEAD(dispose);
3620
3621         pr_info("mds%d reconnect start\n", mds);
3622
3623         recon_state.pagelist = ceph_pagelist_alloc(GFP_NOFS);
3624         if (!recon_state.pagelist)
3625                 goto fail_nopagelist;
3626
3627         reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3628         if (!reply)
3629                 goto fail_nomsg;
3630
3631         mutex_lock(&session->s_mutex);
3632         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3633         session->s_seq = 0;
3634
3635         dout("session %p state %s\n", session,
3636              ceph_session_state_name(session->s_state));
3637
3638         spin_lock(&session->s_gen_ttl_lock);
3639         session->s_cap_gen++;
3640         spin_unlock(&session->s_gen_ttl_lock);
3641
3642         spin_lock(&session->s_cap_lock);
3643         /* don't know if session is readonly */
3644         session->s_readonly = 0;
3645         /*
3646          * notify __ceph_remove_cap() that we are composing cap reconnect.
3647          * If a cap get released before being added to the cap reconnect,
3648          * __ceph_remove_cap() should skip queuing cap release.
3649          */
3650         session->s_cap_reconnect = 1;
3651         /* drop old cap expires; we're about to reestablish that state */
3652         detach_cap_releases(session, &dispose);
3653         spin_unlock(&session->s_cap_lock);
3654         dispose_cap_releases(mdsc, &dispose);
3655
3656         /* trim unused caps to reduce MDS's cache rejoin time */
3657         if (mdsc->fsc->sb->s_root)
3658                 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3659
3660         ceph_con_close(&session->s_con);
3661         ceph_con_open(&session->s_con,
3662                       CEPH_ENTITY_TYPE_MDS, mds,
3663                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3664
3665         /* replay unsafe requests */
3666         replay_unsafe_requests(mdsc, session);
3667
3668         ceph_early_kick_flushing_caps(mdsc, session);
3669
3670         down_read(&mdsc->snap_rwsem);
3671
3672         /* placeholder for nr_caps */
3673         err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
3674         if (err)
3675                 goto fail;
3676
3677         if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
3678                 recon_state.msg_version = 3;
3679                 recon_state.allow_multi = true;
3680         } else if (session->s_con.peer_features & CEPH_FEATURE_MDSENC) {
3681                 recon_state.msg_version = 3;
3682         } else {
3683                 recon_state.msg_version = 2;
3684         }
3685         /* trsaverse this session's caps */
3686         err = ceph_iterate_session_caps(session, encode_caps_cb, &recon_state);
3687
3688         spin_lock(&session->s_cap_lock);
3689         session->s_cap_reconnect = 0;
3690         spin_unlock(&session->s_cap_lock);
3691
3692         if (err < 0)
3693                 goto fail;
3694
3695         /* check if all realms can be encoded into current message */
3696         if (mdsc->num_snap_realms) {
3697                 size_t total_len =
3698                         recon_state.pagelist->length +
3699                         mdsc->num_snap_realms *
3700                         sizeof(struct ceph_mds_snaprealm_reconnect);
3701                 if (recon_state.msg_version >= 4) {
3702                         /* number of realms */
3703                         total_len += sizeof(u32);
3704                         /* version, compat_version and struct_len */
3705                         total_len += mdsc->num_snap_realms *
3706                                      (2 * sizeof(u8) + sizeof(u32));
3707                 }
3708                 if (total_len > RECONNECT_MAX_SIZE) {
3709                         if (!recon_state.allow_multi) {
3710                                 err = -ENOSPC;
3711                                 goto fail;
3712                         }
3713                         if (recon_state.nr_caps) {
3714                                 err = send_reconnect_partial(&recon_state);
3715                                 if (err)
3716                                         goto fail;
3717                         }
3718                         recon_state.msg_version = 5;
3719                 }
3720         }
3721
3722         err = encode_snap_realms(mdsc, &recon_state);
3723         if (err < 0)
3724                 goto fail;
3725
3726         if (recon_state.msg_version >= 5) {
3727                 err = ceph_pagelist_encode_8(recon_state.pagelist, 0);
3728                 if (err < 0)
3729                         goto fail;
3730         }
3731
3732         if (recon_state.nr_caps || recon_state.nr_realms) {
3733                 struct page *page =
3734                         list_first_entry(&recon_state.pagelist->head,
3735                                         struct page, lru);
3736                 __le32 *addr = kmap_atomic(page);
3737                 if (recon_state.nr_caps) {
3738                         WARN_ON(recon_state.nr_realms != mdsc->num_snap_realms);
3739                         *addr = cpu_to_le32(recon_state.nr_caps);
3740                 } else if (recon_state.msg_version >= 4) {
3741                         *(addr + 1) = cpu_to_le32(recon_state.nr_realms);
3742                 }
3743                 kunmap_atomic(addr);
3744         }
3745
3746         reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3747         if (recon_state.msg_version >= 4)
3748                 reply->hdr.compat_version = cpu_to_le16(4);
3749
3750         reply->hdr.data_len = cpu_to_le32(recon_state.pagelist->length);
3751         ceph_msg_data_add_pagelist(reply, recon_state.pagelist);
3752
3753         ceph_con_send(&session->s_con, reply);
3754
3755         mutex_unlock(&session->s_mutex);
3756
3757         mutex_lock(&mdsc->mutex);
3758         __wake_requests(mdsc, &session->s_waiting);
3759         mutex_unlock(&mdsc->mutex);
3760
3761         up_read(&mdsc->snap_rwsem);
3762         ceph_pagelist_release(recon_state.pagelist);
3763         return;
3764
3765 fail:
3766         ceph_msg_put(reply);
3767         up_read(&mdsc->snap_rwsem);
3768         mutex_unlock(&session->s_mutex);
3769 fail_nomsg:
3770         ceph_pagelist_release(recon_state.pagelist);
3771 fail_nopagelist:
3772         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3773         return;
3774 }
3775
3776
3777 /*
3778  * compare old and new mdsmaps, kicking requests
3779  * and closing out old connections as necessary
3780  *
3781  * called under mdsc->mutex.
3782  */
3783 static void check_new_map(struct ceph_mds_client *mdsc,
3784                           struct ceph_mdsmap *newmap,
3785                           struct ceph_mdsmap *oldmap)
3786 {
3787         int i;
3788         int oldstate, newstate;
3789         struct ceph_mds_session *s;
3790
3791         dout("check_new_map new %u old %u\n",
3792              newmap->m_epoch, oldmap->m_epoch);
3793
3794         for (i = 0; i < oldmap->possible_max_rank && i < mdsc->max_sessions; i++) {
3795                 if (!mdsc->sessions[i])
3796                         continue;
3797                 s = mdsc->sessions[i];
3798                 oldstate = ceph_mdsmap_get_state(oldmap, i);
3799                 newstate = ceph_mdsmap_get_state(newmap, i);
3800
3801                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3802                      i, ceph_mds_state_name(oldstate),
3803                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3804                      ceph_mds_state_name(newstate),
3805                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3806                      ceph_session_state_name(s->s_state));
3807
3808                 if (i >= newmap->possible_max_rank) {
3809                         /* force close session for stopped mds */
3810                         ceph_get_mds_session(s);
3811                         __unregister_session(mdsc, s);
3812                         __wake_requests(mdsc, &s->s_waiting);
3813                         mutex_unlock(&mdsc->mutex);
3814
3815                         mutex_lock(&s->s_mutex);
3816                         cleanup_session_requests(mdsc, s);
3817                         remove_session_caps(s);
3818                         mutex_unlock(&s->s_mutex);
3819
3820                         ceph_put_mds_session(s);
3821
3822                         mutex_lock(&mdsc->mutex);
3823                         kick_requests(mdsc, i);
3824                         continue;
3825                 }
3826
3827                 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
3828                            ceph_mdsmap_get_addr(newmap, i),
3829                            sizeof(struct ceph_entity_addr))) {
3830                         /* just close it */
3831                         mutex_unlock(&mdsc->mutex);
3832                         mutex_lock(&s->s_mutex);
3833                         mutex_lock(&mdsc->mutex);
3834                         ceph_con_close(&s->s_con);
3835                         mutex_unlock(&s->s_mutex);
3836                         s->s_state = CEPH_MDS_SESSION_RESTARTING;
3837                 } else if (oldstate == newstate) {
3838                         continue;  /* nothing new with this mds */
3839                 }
3840
3841                 /*
3842                  * send reconnect?
3843                  */
3844                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3845                     newstate >= CEPH_MDS_STATE_RECONNECT) {
3846                         mutex_unlock(&mdsc->mutex);
3847                         send_mds_reconnect(mdsc, s);
3848                         mutex_lock(&mdsc->mutex);
3849                 }
3850
3851                 /*
3852                  * kick request on any mds that has gone active.
3853                  */
3854                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3855                     newstate >= CEPH_MDS_STATE_ACTIVE) {
3856                         if (oldstate != CEPH_MDS_STATE_CREATING &&
3857                             oldstate != CEPH_MDS_STATE_STARTING)
3858                                 pr_info("mds%d recovery completed\n", s->s_mds);
3859                         kick_requests(mdsc, i);
3860                         ceph_kick_flushing_caps(mdsc, s);
3861                         wake_up_session_caps(s, RECONNECT);
3862                 }
3863         }
3864
3865         for (i = 0; i < newmap->possible_max_rank && i < mdsc->max_sessions; i++) {
3866                 s = mdsc->sessions[i];
3867                 if (!s)
3868                         continue;
3869                 if (!ceph_mdsmap_is_laggy(newmap, i))
3870                         continue;
3871                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3872                     s->s_state == CEPH_MDS_SESSION_HUNG ||
3873                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
3874                         dout(" connecting to export targets of laggy mds%d\n",
3875                              i);
3876                         __open_export_target_sessions(mdsc, s);
3877                 }
3878         }
3879 }
3880
3881
3882
3883 /*
3884  * leases
3885  */
3886
3887 /*
3888  * caller must hold session s_mutex, dentry->d_lock
3889  */
3890 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3891 {
3892         struct ceph_dentry_info *di = ceph_dentry(dentry);
3893
3894         ceph_put_mds_session(di->lease_session);
3895         di->lease_session = NULL;
3896 }
3897
3898 static void handle_lease(struct ceph_mds_client *mdsc,
3899                          struct ceph_mds_session *session,
3900                          struct ceph_msg *msg)
3901 {
3902         struct super_block *sb = mdsc->fsc->sb;
3903         struct inode *inode;
3904         struct dentry *parent, *dentry;
3905         struct ceph_dentry_info *di;
3906         int mds = session->s_mds;
3907         struct ceph_mds_lease *h = msg->front.iov_base;
3908         u32 seq;
3909         struct ceph_vino vino;
3910         struct qstr dname;
3911         int release = 0;
3912
3913         dout("handle_lease from mds%d\n", mds);
3914
3915         /* decode */
3916         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3917                 goto bad;
3918         vino.ino = le64_to_cpu(h->ino);
3919         vino.snap = CEPH_NOSNAP;
3920         seq = le32_to_cpu(h->seq);
3921         dname.len = get_unaligned_le32(h + 1);
3922         if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
3923                 goto bad;
3924         dname.name = (void *)(h + 1) + sizeof(u32);
3925
3926         /* lookup inode */
3927         inode = ceph_find_inode(sb, vino);
3928         dout("handle_lease %s, ino %llx %p %.*s\n",
3929              ceph_lease_op_name(h->action), vino.ino, inode,
3930              dname.len, dname.name);
3931
3932         mutex_lock(&session->s_mutex);
3933         session->s_seq++;
3934
3935         if (!inode) {
3936                 dout("handle_lease no inode %llx\n", vino.ino);
3937                 goto release;
3938         }
3939
3940         /* dentry */
3941         parent = d_find_alias(inode);
3942         if (!parent) {
3943                 dout("no parent dentry on inode %p\n", inode);
3944                 WARN_ON(1);
3945                 goto release;  /* hrm... */
3946         }
3947         dname.hash = full_name_hash(parent, dname.name, dname.len);
3948         dentry = d_lookup(parent, &dname);
3949         dput(parent);
3950         if (!dentry)
3951                 goto release;
3952
3953         spin_lock(&dentry->d_lock);
3954         di = ceph_dentry(dentry);
3955         switch (h->action) {
3956         case CEPH_MDS_LEASE_REVOKE:
3957                 if (di->lease_session == session) {
3958                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3959                                 h->seq = cpu_to_le32(di->lease_seq);
3960                         __ceph_mdsc_drop_dentry_lease(dentry);
3961                 }
3962                 release = 1;
3963                 break;
3964
3965         case CEPH_MDS_LEASE_RENEW:
3966                 if (di->lease_session == session &&
3967                     di->lease_gen == session->s_cap_gen &&
3968                     di->lease_renew_from &&
3969                     di->lease_renew_after == 0) {
3970                         unsigned long duration =
3971                                 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3972
3973                         di->lease_seq = seq;
3974                         di->time = di->lease_renew_from + duration;
3975                         di->lease_renew_after = di->lease_renew_from +
3976                                 (duration >> 1);
3977                         di->lease_renew_from = 0;
3978                 }
3979                 break;
3980         }
3981         spin_unlock(&dentry->d_lock);
3982         dput(dentry);
3983
3984         if (!release)
3985                 goto out;
3986
3987 release:
3988         /* let's just reuse the same message */
3989         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3990         ceph_msg_get(msg);
3991         ceph_con_send(&session->s_con, msg);
3992
3993 out:
3994         mutex_unlock(&session->s_mutex);
3995         /* avoid calling iput_final() in mds dispatch threads */
3996         ceph_async_iput(inode);
3997         return;
3998
3999 bad:
4000         pr_err("corrupt lease message\n");
4001         ceph_msg_dump(msg);
4002 }
4003
4004 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
4005                               struct dentry *dentry, char action,
4006                               u32 seq)
4007 {
4008         struct ceph_msg *msg;
4009         struct ceph_mds_lease *lease;
4010         struct inode *dir;
4011         int len = sizeof(*lease) + sizeof(u32) + NAME_MAX;
4012
4013         dout("lease_send_msg identry %p %s to mds%d\n",
4014              dentry, ceph_lease_op_name(action), session->s_mds);
4015
4016         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
4017         if (!msg)
4018                 return;
4019         lease = msg->front.iov_base;
4020         lease->action = action;
4021         lease->seq = cpu_to_le32(seq);
4022
4023         spin_lock(&dentry->d_lock);
4024         dir = d_inode(dentry->d_parent);
4025         lease->ino = cpu_to_le64(ceph_ino(dir));
4026         lease->first = lease->last = cpu_to_le64(ceph_snap(dir));
4027
4028         put_unaligned_le32(dentry->d_name.len, lease + 1);
4029         memcpy((void *)(lease + 1) + 4,
4030                dentry->d_name.name, dentry->d_name.len);
4031         spin_unlock(&dentry->d_lock);
4032         /*
4033          * if this is a preemptive lease RELEASE, no need to
4034          * flush request stream, since the actual request will
4035          * soon follow.
4036          */
4037         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
4038
4039         ceph_con_send(&session->s_con, msg);
4040 }
4041
4042 /*
4043  * lock unlock sessions, to wait ongoing session activities
4044  */
4045 static void lock_unlock_sessions(struct ceph_mds_client *mdsc)
4046 {
4047         int i;
4048
4049         mutex_lock(&mdsc->mutex);
4050         for (i = 0; i < mdsc->max_sessions; i++) {
4051                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4052                 if (!s)
4053                         continue;
4054                 mutex_unlock(&mdsc->mutex);
4055                 mutex_lock(&s->s_mutex);
4056                 mutex_unlock(&s->s_mutex);
4057                 ceph_put_mds_session(s);
4058                 mutex_lock(&mdsc->mutex);
4059         }
4060         mutex_unlock(&mdsc->mutex);
4061 }
4062
4063 static void maybe_recover_session(struct ceph_mds_client *mdsc)
4064 {
4065         struct ceph_fs_client *fsc = mdsc->fsc;
4066
4067         if (!ceph_test_mount_opt(fsc, CLEANRECOVER))
4068                 return;
4069
4070         if (READ_ONCE(fsc->mount_state) != CEPH_MOUNT_MOUNTED)
4071                 return;
4072
4073         if (!READ_ONCE(fsc->blacklisted))
4074                 return;
4075
4076         if (fsc->last_auto_reconnect &&
4077             time_before(jiffies, fsc->last_auto_reconnect + HZ * 60 * 30))
4078                 return;
4079
4080         pr_info("auto reconnect after blacklisted\n");
4081         fsc->last_auto_reconnect = jiffies;
4082         ceph_force_reconnect(fsc->sb);
4083 }
4084
4085 /*
4086  * delayed work -- periodically trim expired leases, renew caps with mds
4087  */
4088 static void schedule_delayed(struct ceph_mds_client *mdsc)
4089 {
4090         int delay = 5;
4091         unsigned hz = round_jiffies_relative(HZ * delay);
4092         schedule_delayed_work(&mdsc->delayed_work, hz);
4093 }
4094
4095 static void delayed_work(struct work_struct *work)
4096 {
4097         int i;
4098         struct ceph_mds_client *mdsc =
4099                 container_of(work, struct ceph_mds_client, delayed_work.work);
4100         int renew_interval;
4101         int renew_caps;
4102
4103         dout("mdsc delayed_work\n");
4104
4105         mutex_lock(&mdsc->mutex);
4106         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
4107         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
4108                                    mdsc->last_renew_caps);
4109         if (renew_caps)
4110                 mdsc->last_renew_caps = jiffies;
4111
4112         for (i = 0; i < mdsc->max_sessions; i++) {
4113                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4114                 if (!s)
4115                         continue;
4116                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
4117                         dout("resending session close request for mds%d\n",
4118                              s->s_mds);
4119                         request_close_session(mdsc, s);
4120                         ceph_put_mds_session(s);
4121                         continue;
4122                 }
4123                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
4124                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
4125                                 s->s_state = CEPH_MDS_SESSION_HUNG;
4126                                 pr_info("mds%d hung\n", s->s_mds);
4127                         }
4128                 }
4129                 if (s->s_state == CEPH_MDS_SESSION_NEW ||
4130                     s->s_state == CEPH_MDS_SESSION_RESTARTING ||
4131                     s->s_state == CEPH_MDS_SESSION_REJECTED) {
4132                         /* this mds is failed or recovering, just wait */
4133                         ceph_put_mds_session(s);
4134                         continue;
4135                 }
4136                 mutex_unlock(&mdsc->mutex);
4137
4138                 mutex_lock(&s->s_mutex);
4139                 if (renew_caps)
4140                         send_renew_caps(mdsc, s);
4141                 else
4142                         ceph_con_keepalive(&s->s_con);
4143                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4144                     s->s_state == CEPH_MDS_SESSION_HUNG)
4145                         ceph_send_cap_releases(mdsc, s);
4146                 mutex_unlock(&s->s_mutex);
4147                 ceph_put_mds_session(s);
4148
4149                 mutex_lock(&mdsc->mutex);
4150         }
4151         mutex_unlock(&mdsc->mutex);
4152
4153         ceph_check_delayed_caps(mdsc);
4154
4155         ceph_queue_cap_reclaim_work(mdsc);
4156
4157         ceph_trim_snapid_map(mdsc);
4158
4159         maybe_recover_session(mdsc);
4160
4161         schedule_delayed(mdsc);
4162 }
4163
4164 int ceph_mdsc_init(struct ceph_fs_client *fsc)
4165
4166 {
4167         struct ceph_mds_client *mdsc;
4168
4169         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
4170         if (!mdsc)
4171                 return -ENOMEM;
4172         mdsc->fsc = fsc;
4173         mutex_init(&mdsc->mutex);
4174         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
4175         if (!mdsc->mdsmap) {
4176                 kfree(mdsc);
4177                 return -ENOMEM;
4178         }
4179
4180         fsc->mdsc = mdsc;
4181         init_completion(&mdsc->safe_umount_waiters);
4182         init_waitqueue_head(&mdsc->session_close_wq);
4183         INIT_LIST_HEAD(&mdsc->waiting_for_map);
4184         mdsc->sessions = NULL;
4185         atomic_set(&mdsc->num_sessions, 0);
4186         mdsc->max_sessions = 0;
4187         mdsc->stopping = 0;
4188         atomic64_set(&mdsc->quotarealms_count, 0);
4189         mdsc->quotarealms_inodes = RB_ROOT;
4190         mutex_init(&mdsc->quotarealms_inodes_mutex);
4191         mdsc->last_snap_seq = 0;
4192         init_rwsem(&mdsc->snap_rwsem);
4193         mdsc->snap_realms = RB_ROOT;
4194         INIT_LIST_HEAD(&mdsc->snap_empty);
4195         mdsc->num_snap_realms = 0;
4196         spin_lock_init(&mdsc->snap_empty_lock);
4197         mdsc->last_tid = 0;
4198         mdsc->oldest_tid = 0;
4199         mdsc->request_tree = RB_ROOT;
4200         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
4201         mdsc->last_renew_caps = jiffies;
4202         INIT_LIST_HEAD(&mdsc->cap_delay_list);
4203         INIT_LIST_HEAD(&mdsc->cap_wait_list);
4204         spin_lock_init(&mdsc->cap_delay_lock);
4205         INIT_LIST_HEAD(&mdsc->snap_flush_list);
4206         spin_lock_init(&mdsc->snap_flush_lock);
4207         mdsc->last_cap_flush_tid = 1;
4208         INIT_LIST_HEAD(&mdsc->cap_flush_list);
4209         INIT_LIST_HEAD(&mdsc->cap_dirty);
4210         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
4211         mdsc->num_cap_flushing = 0;
4212         spin_lock_init(&mdsc->cap_dirty_lock);
4213         init_waitqueue_head(&mdsc->cap_flushing_wq);
4214         INIT_WORK(&mdsc->cap_reclaim_work, ceph_cap_reclaim_work);
4215         atomic_set(&mdsc->cap_reclaim_pending, 0);
4216
4217         spin_lock_init(&mdsc->dentry_list_lock);
4218         INIT_LIST_HEAD(&mdsc->dentry_leases);
4219         INIT_LIST_HEAD(&mdsc->dentry_dir_leases);
4220
4221         ceph_caps_init(mdsc);
4222         ceph_adjust_caps_max_min(mdsc, fsc->mount_options);
4223
4224         spin_lock_init(&mdsc->snapid_map_lock);
4225         mdsc->snapid_map_tree = RB_ROOT;
4226         INIT_LIST_HEAD(&mdsc->snapid_map_lru);
4227
4228         init_rwsem(&mdsc->pool_perm_rwsem);
4229         mdsc->pool_perm_tree = RB_ROOT;
4230
4231         strscpy(mdsc->nodename, utsname()->nodename,
4232                 sizeof(mdsc->nodename));
4233         return 0;
4234 }
4235
4236 /*
4237  * Wait for safe replies on open mds requests.  If we time out, drop
4238  * all requests from the tree to avoid dangling dentry refs.
4239  */
4240 static void wait_requests(struct ceph_mds_client *mdsc)
4241 {
4242         struct ceph_options *opts = mdsc->fsc->client->options;
4243         struct ceph_mds_request *req;
4244
4245         mutex_lock(&mdsc->mutex);
4246         if (__get_oldest_req(mdsc)) {
4247                 mutex_unlock(&mdsc->mutex);
4248
4249                 dout("wait_requests waiting for requests\n");
4250                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
4251                                     ceph_timeout_jiffies(opts->mount_timeout));
4252
4253                 /* tear down remaining requests */
4254                 mutex_lock(&mdsc->mutex);
4255                 while ((req = __get_oldest_req(mdsc))) {
4256                         dout("wait_requests timed out on tid %llu\n",
4257                              req->r_tid);
4258                         list_del_init(&req->r_wait);
4259                         __unregister_request(mdsc, req);
4260                 }
4261         }
4262         mutex_unlock(&mdsc->mutex);
4263         dout("wait_requests done\n");
4264 }
4265
4266 /*
4267  * called before mount is ro, and before dentries are torn down.
4268  * (hmm, does this still race with new lookups?)
4269  */
4270 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
4271 {
4272         dout("pre_umount\n");
4273         mdsc->stopping = 1;
4274
4275         lock_unlock_sessions(mdsc);
4276         ceph_flush_dirty_caps(mdsc);
4277         wait_requests(mdsc);
4278
4279         /*
4280          * wait for reply handlers to drop their request refs and
4281          * their inode/dcache refs
4282          */
4283         ceph_msgr_flush();
4284
4285         ceph_cleanup_quotarealms_inodes(mdsc);
4286 }
4287
4288 /*
4289  * wait for all write mds requests to flush.
4290  */
4291 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
4292 {
4293         struct ceph_mds_request *req = NULL, *nextreq;
4294         struct rb_node *n;
4295
4296         mutex_lock(&mdsc->mutex);
4297         dout("wait_unsafe_requests want %lld\n", want_tid);
4298 restart:
4299         req = __get_oldest_req(mdsc);
4300         while (req && req->r_tid <= want_tid) {
4301                 /* find next request */
4302                 n = rb_next(&req->r_node);
4303                 if (n)
4304                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
4305                 else
4306                         nextreq = NULL;
4307                 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
4308                     (req->r_op & CEPH_MDS_OP_WRITE)) {
4309                         /* write op */
4310                         ceph_mdsc_get_request(req);
4311                         if (nextreq)
4312                                 ceph_mdsc_get_request(nextreq);
4313                         mutex_unlock(&mdsc->mutex);
4314                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
4315                              req->r_tid, want_tid);
4316                         wait_for_completion(&req->r_safe_completion);
4317                         mutex_lock(&mdsc->mutex);
4318                         ceph_mdsc_put_request(req);
4319                         if (!nextreq)
4320                                 break;  /* next dne before, so we're done! */
4321                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
4322                                 /* next request was removed from tree */
4323                                 ceph_mdsc_put_request(nextreq);
4324                                 goto restart;
4325                         }
4326                         ceph_mdsc_put_request(nextreq);  /* won't go away */
4327                 }
4328                 req = nextreq;
4329         }
4330         mutex_unlock(&mdsc->mutex);
4331         dout("wait_unsafe_requests done\n");
4332 }
4333
4334 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
4335 {
4336         u64 want_tid, want_flush;
4337
4338         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4339                 return;
4340
4341         dout("sync\n");
4342         mutex_lock(&mdsc->mutex);
4343         want_tid = mdsc->last_tid;
4344         mutex_unlock(&mdsc->mutex);
4345
4346         ceph_flush_dirty_caps(mdsc);
4347         spin_lock(&mdsc->cap_dirty_lock);
4348         want_flush = mdsc->last_cap_flush_tid;
4349         if (!list_empty(&mdsc->cap_flush_list)) {
4350                 struct ceph_cap_flush *cf =
4351                         list_last_entry(&mdsc->cap_flush_list,
4352                                         struct ceph_cap_flush, g_list);
4353                 cf->wake = true;
4354         }
4355         spin_unlock(&mdsc->cap_dirty_lock);
4356
4357         dout("sync want tid %lld flush_seq %lld\n",
4358              want_tid, want_flush);
4359
4360         wait_unsafe_requests(mdsc, want_tid);
4361         wait_caps_flush(mdsc, want_flush);
4362 }
4363
4364 /*
4365  * true if all sessions are closed, or we force unmount
4366  */
4367 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
4368 {
4369         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4370                 return true;
4371         return atomic_read(&mdsc->num_sessions) <= skipped;
4372 }
4373
4374 /*
4375  * called after sb is ro.
4376  */
4377 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
4378 {
4379         struct ceph_options *opts = mdsc->fsc->client->options;
4380         struct ceph_mds_session *session;
4381         int i;
4382         int skipped = 0;
4383
4384         dout("close_sessions\n");
4385
4386         /* close sessions */
4387         mutex_lock(&mdsc->mutex);
4388         for (i = 0; i < mdsc->max_sessions; i++) {
4389                 session = __ceph_lookup_mds_session(mdsc, i);
4390                 if (!session)
4391                         continue;
4392                 mutex_unlock(&mdsc->mutex);
4393                 mutex_lock(&session->s_mutex);
4394                 if (__close_session(mdsc, session) <= 0)
4395                         skipped++;
4396                 mutex_unlock(&session->s_mutex);
4397                 ceph_put_mds_session(session);
4398                 mutex_lock(&mdsc->mutex);
4399         }
4400         mutex_unlock(&mdsc->mutex);
4401
4402         dout("waiting for sessions to close\n");
4403         wait_event_timeout(mdsc->session_close_wq,
4404                            done_closing_sessions(mdsc, skipped),
4405                            ceph_timeout_jiffies(opts->mount_timeout));
4406
4407         /* tear down remaining sessions */
4408         mutex_lock(&mdsc->mutex);
4409         for (i = 0; i < mdsc->max_sessions; i++) {
4410                 if (mdsc->sessions[i]) {
4411                         session = ceph_get_mds_session(mdsc->sessions[i]);
4412                         __unregister_session(mdsc, session);
4413                         mutex_unlock(&mdsc->mutex);
4414                         mutex_lock(&session->s_mutex);
4415                         remove_session_caps(session);
4416                         mutex_unlock(&session->s_mutex);
4417                         ceph_put_mds_session(session);
4418                         mutex_lock(&mdsc->mutex);
4419                 }
4420         }
4421         WARN_ON(!list_empty(&mdsc->cap_delay_list));
4422         mutex_unlock(&mdsc->mutex);
4423
4424         ceph_cleanup_snapid_map(mdsc);
4425         ceph_cleanup_empty_realms(mdsc);
4426
4427         cancel_work_sync(&mdsc->cap_reclaim_work);
4428         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4429
4430         dout("stopped\n");
4431 }
4432
4433 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
4434 {
4435         struct ceph_mds_session *session;
4436         int mds;
4437
4438         dout("force umount\n");
4439
4440         mutex_lock(&mdsc->mutex);
4441         for (mds = 0; mds < mdsc->max_sessions; mds++) {
4442                 session = __ceph_lookup_mds_session(mdsc, mds);
4443                 if (!session)
4444                         continue;
4445
4446                 if (session->s_state == CEPH_MDS_SESSION_REJECTED)
4447                         __unregister_session(mdsc, session);
4448                 __wake_requests(mdsc, &session->s_waiting);
4449                 mutex_unlock(&mdsc->mutex);
4450
4451                 mutex_lock(&session->s_mutex);
4452                 __close_session(mdsc, session);
4453                 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
4454                         cleanup_session_requests(mdsc, session);
4455                         remove_session_caps(session);
4456                 }
4457                 mutex_unlock(&session->s_mutex);
4458                 ceph_put_mds_session(session);
4459
4460                 mutex_lock(&mdsc->mutex);
4461                 kick_requests(mdsc, mds);
4462         }
4463         __wake_requests(mdsc, &mdsc->waiting_for_map);
4464         mutex_unlock(&mdsc->mutex);
4465 }
4466
4467 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
4468 {
4469         dout("stop\n");
4470         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4471         if (mdsc->mdsmap)
4472                 ceph_mdsmap_destroy(mdsc->mdsmap);
4473         kfree(mdsc->sessions);
4474         ceph_caps_finalize(mdsc);
4475         ceph_pool_perm_destroy(mdsc);
4476 }
4477
4478 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
4479 {
4480         struct ceph_mds_client *mdsc = fsc->mdsc;
4481         dout("mdsc_destroy %p\n", mdsc);
4482
4483         if (!mdsc)
4484                 return;
4485
4486         /* flush out any connection work with references to us */
4487         ceph_msgr_flush();
4488
4489         ceph_mdsc_stop(mdsc);
4490
4491         fsc->mdsc = NULL;
4492         kfree(mdsc);
4493         dout("mdsc_destroy %p done\n", mdsc);
4494 }
4495
4496 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4497 {
4498         struct ceph_fs_client *fsc = mdsc->fsc;
4499         const char *mds_namespace = fsc->mount_options->mds_namespace;
4500         void *p = msg->front.iov_base;
4501         void *end = p + msg->front.iov_len;
4502         u32 epoch;
4503         u32 map_len;
4504         u32 num_fs;
4505         u32 mount_fscid = (u32)-1;
4506         u8 struct_v, struct_cv;
4507         int err = -EINVAL;
4508
4509         ceph_decode_need(&p, end, sizeof(u32), bad);
4510         epoch = ceph_decode_32(&p);
4511
4512         dout("handle_fsmap epoch %u\n", epoch);
4513
4514         ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4515         struct_v = ceph_decode_8(&p);
4516         struct_cv = ceph_decode_8(&p);
4517         map_len = ceph_decode_32(&p);
4518
4519         ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
4520         p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
4521
4522         num_fs = ceph_decode_32(&p);
4523         while (num_fs-- > 0) {
4524                 void *info_p, *info_end;
4525                 u32 info_len;
4526                 u8 info_v, info_cv;
4527                 u32 fscid, namelen;
4528
4529                 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4530                 info_v = ceph_decode_8(&p);
4531                 info_cv = ceph_decode_8(&p);
4532                 info_len = ceph_decode_32(&p);
4533                 ceph_decode_need(&p, end, info_len, bad);
4534                 info_p = p;
4535                 info_end = p + info_len;
4536                 p = info_end;
4537
4538                 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
4539                 fscid = ceph_decode_32(&info_p);
4540                 namelen = ceph_decode_32(&info_p);
4541                 ceph_decode_need(&info_p, info_end, namelen, bad);
4542
4543                 if (mds_namespace &&
4544                     strlen(mds_namespace) == namelen &&
4545                     !strncmp(mds_namespace, (char *)info_p, namelen)) {
4546                         mount_fscid = fscid;
4547                         break;
4548                 }
4549         }
4550
4551         ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
4552         if (mount_fscid != (u32)-1) {
4553                 fsc->client->monc.fs_cluster_id = mount_fscid;
4554                 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
4555                                    0, true);
4556                 ceph_monc_renew_subs(&fsc->client->monc);
4557         } else {
4558                 err = -ENOENT;
4559                 goto err_out;
4560         }
4561         return;
4562
4563 bad:
4564         pr_err("error decoding fsmap\n");
4565 err_out:
4566         mutex_lock(&mdsc->mutex);
4567         mdsc->mdsmap_err = err;
4568         __wake_requests(mdsc, &mdsc->waiting_for_map);
4569         mutex_unlock(&mdsc->mutex);
4570 }
4571
4572 /*
4573  * handle mds map update.
4574  */
4575 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4576 {
4577         u32 epoch;
4578         u32 maplen;
4579         void *p = msg->front.iov_base;
4580         void *end = p + msg->front.iov_len;
4581         struct ceph_mdsmap *newmap, *oldmap;
4582         struct ceph_fsid fsid;
4583         int err = -EINVAL;
4584
4585         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
4586         ceph_decode_copy(&p, &fsid, sizeof(fsid));
4587         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
4588                 return;
4589         epoch = ceph_decode_32(&p);
4590         maplen = ceph_decode_32(&p);
4591         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
4592
4593         /* do we need it? */
4594         mutex_lock(&mdsc->mutex);
4595         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
4596                 dout("handle_map epoch %u <= our %u\n",
4597                      epoch, mdsc->mdsmap->m_epoch);
4598                 mutex_unlock(&mdsc->mutex);
4599                 return;
4600         }
4601
4602         newmap = ceph_mdsmap_decode(&p, end);
4603         if (IS_ERR(newmap)) {
4604                 err = PTR_ERR(newmap);
4605                 goto bad_unlock;
4606         }
4607
4608         /* swap into place */
4609         if (mdsc->mdsmap) {
4610                 oldmap = mdsc->mdsmap;
4611                 mdsc->mdsmap = newmap;
4612                 check_new_map(mdsc, newmap, oldmap);
4613                 ceph_mdsmap_destroy(oldmap);
4614         } else {
4615                 mdsc->mdsmap = newmap;  /* first mds map */
4616         }
4617         mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
4618                                         MAX_LFS_FILESIZE);
4619
4620         __wake_requests(mdsc, &mdsc->waiting_for_map);
4621         ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
4622                           mdsc->mdsmap->m_epoch);
4623
4624         mutex_unlock(&mdsc->mutex);
4625         schedule_delayed(mdsc);
4626         return;
4627
4628 bad_unlock:
4629         mutex_unlock(&mdsc->mutex);
4630 bad:
4631         pr_err("error decoding mdsmap %d\n", err);
4632         return;
4633 }
4634
4635 static struct ceph_connection *con_get(struct ceph_connection *con)
4636 {
4637         struct ceph_mds_session *s = con->private;
4638
4639         if (ceph_get_mds_session(s))
4640                 return con;
4641         return NULL;
4642 }
4643
4644 static void con_put(struct ceph_connection *con)
4645 {
4646         struct ceph_mds_session *s = con->private;
4647
4648         ceph_put_mds_session(s);
4649 }
4650
4651 /*
4652  * if the client is unresponsive for long enough, the mds will kill
4653  * the session entirely.
4654  */
4655 static void peer_reset(struct ceph_connection *con)
4656 {
4657         struct ceph_mds_session *s = con->private;
4658         struct ceph_mds_client *mdsc = s->s_mdsc;
4659
4660         pr_warn("mds%d closed our session\n", s->s_mds);
4661         send_mds_reconnect(mdsc, s);
4662 }
4663
4664 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
4665 {
4666         struct ceph_mds_session *s = con->private;
4667         struct ceph_mds_client *mdsc = s->s_mdsc;
4668         int type = le16_to_cpu(msg->hdr.type);
4669
4670         mutex_lock(&mdsc->mutex);
4671         if (__verify_registered_session(mdsc, s) < 0) {
4672                 mutex_unlock(&mdsc->mutex);
4673                 goto out;
4674         }
4675         mutex_unlock(&mdsc->mutex);
4676
4677         switch (type) {
4678         case CEPH_MSG_MDS_MAP:
4679                 ceph_mdsc_handle_mdsmap(mdsc, msg);
4680                 break;
4681         case CEPH_MSG_FS_MAP_USER:
4682                 ceph_mdsc_handle_fsmap(mdsc, msg);
4683                 break;
4684         case CEPH_MSG_CLIENT_SESSION:
4685                 handle_session(s, msg);
4686                 break;
4687         case CEPH_MSG_CLIENT_REPLY:
4688                 handle_reply(s, msg);
4689                 break;
4690         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
4691                 handle_forward(mdsc, s, msg);
4692                 break;
4693         case CEPH_MSG_CLIENT_CAPS:
4694                 ceph_handle_caps(s, msg);
4695                 break;
4696         case CEPH_MSG_CLIENT_SNAP:
4697                 ceph_handle_snap(mdsc, s, msg);
4698                 break;
4699         case CEPH_MSG_CLIENT_LEASE:
4700                 handle_lease(mdsc, s, msg);
4701                 break;
4702         case CEPH_MSG_CLIENT_QUOTA:
4703                 ceph_handle_quota(mdsc, s, msg);
4704                 break;
4705
4706         default:
4707                 pr_err("received unknown message type %d %s\n", type,
4708                        ceph_msg_type_name(type));
4709         }
4710 out:
4711         ceph_msg_put(msg);
4712 }
4713
4714 /*
4715  * authentication
4716  */
4717
4718 /*
4719  * Note: returned pointer is the address of a structure that's
4720  * managed separately.  Caller must *not* attempt to free it.
4721  */
4722 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
4723                                         int *proto, int force_new)
4724 {
4725         struct ceph_mds_session *s = con->private;
4726         struct ceph_mds_client *mdsc = s->s_mdsc;
4727         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4728         struct ceph_auth_handshake *auth = &s->s_auth;
4729
4730         if (force_new && auth->authorizer) {
4731                 ceph_auth_destroy_authorizer(auth->authorizer);
4732                 auth->authorizer = NULL;
4733         }
4734         if (!auth->authorizer) {
4735                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4736                                                       auth);
4737                 if (ret)
4738                         return ERR_PTR(ret);
4739         } else {
4740                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4741                                                       auth);
4742                 if (ret)
4743                         return ERR_PTR(ret);
4744         }
4745         *proto = ac->protocol;
4746
4747         return auth;
4748 }
4749
4750 static int add_authorizer_challenge(struct ceph_connection *con,
4751                                     void *challenge_buf, int challenge_buf_len)
4752 {
4753         struct ceph_mds_session *s = con->private;
4754         struct ceph_mds_client *mdsc = s->s_mdsc;
4755         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4756
4757         return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
4758                                             challenge_buf, challenge_buf_len);
4759 }
4760
4761 static int verify_authorizer_reply(struct ceph_connection *con)
4762 {
4763         struct ceph_mds_session *s = con->private;
4764         struct ceph_mds_client *mdsc = s->s_mdsc;
4765         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4766
4767         return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
4768 }
4769
4770 static int invalidate_authorizer(struct ceph_connection *con)
4771 {
4772         struct ceph_mds_session *s = con->private;
4773         struct ceph_mds_client *mdsc = s->s_mdsc;
4774         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4775
4776         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
4777
4778         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
4779 }
4780
4781 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
4782                                 struct ceph_msg_header *hdr, int *skip)
4783 {
4784         struct ceph_msg *msg;
4785         int type = (int) le16_to_cpu(hdr->type);
4786         int front_len = (int) le32_to_cpu(hdr->front_len);
4787
4788         if (con->in_msg)
4789                 return con->in_msg;
4790
4791         *skip = 0;
4792         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
4793         if (!msg) {
4794                 pr_err("unable to allocate msg type %d len %d\n",
4795                        type, front_len);
4796                 return NULL;
4797         }
4798
4799         return msg;
4800 }
4801
4802 static int mds_sign_message(struct ceph_msg *msg)
4803 {
4804        struct ceph_mds_session *s = msg->con->private;
4805        struct ceph_auth_handshake *auth = &s->s_auth;
4806
4807        return ceph_auth_sign_message(auth, msg);
4808 }
4809
4810 static int mds_check_message_signature(struct ceph_msg *msg)
4811 {
4812        struct ceph_mds_session *s = msg->con->private;
4813        struct ceph_auth_handshake *auth = &s->s_auth;
4814
4815        return ceph_auth_check_message_signature(auth, msg);
4816 }
4817
4818 static const struct ceph_connection_operations mds_con_ops = {
4819         .get = con_get,
4820         .put = con_put,
4821         .dispatch = dispatch,
4822         .get_authorizer = get_authorizer,
4823         .add_authorizer_challenge = add_authorizer_challenge,
4824         .verify_authorizer_reply = verify_authorizer_reply,
4825         .invalidate_authorizer = invalidate_authorizer,
4826         .peer_reset = peer_reset,
4827         .alloc_msg = mds_alloc_msg,
4828         .sign_message = mds_sign_message,
4829         .check_message_signature = mds_check_message_signature,
4830 };
4831
4832 /* eof */