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
  13 #include "super.h"
  14 #include "mds_client.h"
  15
  16 #include <linux/ceph/ceph_features.h>
  17 #include <linux/ceph/messenger.h>
  18 #include <linux/ceph/decode.h>
  19 #include <linux/ceph/pagelist.h>
  20 #include <linux/ceph/auth.h>
  21 #include <linux/ceph/debugfs.h>
  22
  23 #define RECONNECT_MAX_SIZE (INT_MAX - PAGE_SIZE)
  24
  25 /*
  26  * A cluster of MDS (metadata server) daemons is responsible for
  27  * managing the file system namespace (the directory hierarchy and
  28  * inodes) and for coordinating shared access to storage.  Metadata is
  29  * partitioning hierarchically across a number of servers, and that
  30  * partition varies over time as the cluster adjusts the distribution
  31  * in order to balance load.
  32  *
  33  * The MDS client is primarily responsible to managing synchronous
  34  * metadata requests for operations like open, unlink, and so forth.
  35  * If there is a MDS failure, we find out about it when we (possibly
  36  * request and) receive a new MDS map, and can resubmit affected
  37  * requests.
  38  *
  39  * For the most part, though, we take advantage of a lossless
  40  * communications channel to the MDS, and do not need to worry about
  41  * timing out or resubmitting requests.
  42  *
  43  * We maintain a stateful "session" with each MDS we interact with.
  44  * Within each session, we sent periodic heartbeat messages to ensure
  45  * any capabilities or leases we have been issues remain valid.  If
  46  * the session times out and goes stale, our leases and capabilities
  47  * are no longer valid.
  48  */
  49
  50 struct ceph_reconnect_state {
  51         struct ceph_mds_session *session;
  52         int nr_caps, nr_realms;
  53         struct ceph_pagelist *pagelist;
  54         unsigned msg_version;
  55         bool allow_multi;
  56 };
  57
  58 static void __wake_requests(struct ceph_mds_client *mdsc,
  59                             struct list_head *head);
  60 static void ceph_cap_release_work(struct work_struct *work);
  61 static void ceph_cap_reclaim_work(struct work_struct *work);
  62
  63 static const struct ceph_connection_operations mds_con_ops;
  64
  65
  66 /*
  67  * mds reply parsing
  68  */
  69
  70 static int parse_reply_info_quota(void **p, void *end,
  71                                   struct ceph_mds_reply_info_in *info)
  72 {
  73         u8 struct_v, struct_compat;
  74         u32 struct_len;
  75
  76         ceph_decode_8_safe(p, end, struct_v, bad);
  77         ceph_decode_8_safe(p, end, struct_compat, bad);
  78         /* struct_v is expected to be >= 1. we only
  79          * understand encoding with struct_compat == 1. */
  80         if (!struct_v || struct_compat != 1)
  81                 goto bad;
  82         ceph_decode_32_safe(p, end, struct_len, bad);
  83         ceph_decode_need(p, end, struct_len, bad);
  84         end = *p + struct_len;
  85         ceph_decode_64_safe(p, end, info->max_bytes, bad);
  86         ceph_decode_64_safe(p, end, info->max_files, bad);
  87         *p = end;
  88         return 0;
  89 bad:
  90         return -EIO;
  91 }
  92
  93 /*
  94  * parse individual inode info
  95  */
  96 static int parse_reply_info_in(void **p, void *end,
  97                                struct ceph_mds_reply_info_in *info,
  98                                u64 features)
  99 {
 100         int err = 0;
 101         u8 struct_v = 0;
 102
 103         if (features == (u64)-1) {
 104                 u32 struct_len;
 105                 u8 struct_compat;
 106                 ceph_decode_8_safe(p, end, struct_v, bad);
 107                 ceph_decode_8_safe(p, end, struct_compat, bad);
 108                 /* struct_v is expected to be >= 1. we only understand
 109                  * encoding with struct_compat == 1. */
 110                 if (!struct_v || struct_compat != 1)
 111                         goto bad;
 112                 ceph_decode_32_safe(p, end, struct_len, bad);
 113                 ceph_decode_need(p, end, struct_len, bad);
 114                 end = *p + struct_len;
 115         }
 116
 117         ceph_decode_need(p, end, sizeof(struct ceph_mds_reply_inode), bad);
 118         info->in = *p;
 119         *p += sizeof(struct ceph_mds_reply_inode) +
 120                 sizeof(*info->in->fragtree.splits) *
 121                 le32_to_cpu(info->in->fragtree.nsplits);
 122
 123         ceph_decode_32_safe(p, end, info->symlink_len, bad);
 124         ceph_decode_need(p, end, info->symlink_len, bad);
 125         info->symlink = *p;
 126         *p += info->symlink_len;
 127
 128         ceph_decode_copy_safe(p, end, &info->dir_layout,
 129                               sizeof(info->dir_layout), bad);
 130         ceph_decode_32_safe(p, end, info->xattr_len, bad);
 131         ceph_decode_need(p, end, info->xattr_len, bad);
 132         info->xattr_data = *p;
 133         *p += info->xattr_len;
 134
 135         if (features == (u64)-1) {
 136                 /* inline data */
 137                 ceph_decode_64_safe(p, end, info->inline_version, bad);
 138                 ceph_decode_32_safe(p, end, info->inline_len, bad);
 139                 ceph_decode_need(p, end, info->inline_len, bad);
 140                 info->inline_data = *p;
 141                 *p += info->inline_len;
 142                 /* quota */
 143                 err = parse_reply_info_quota(p, end, info);
 144                 if (err < 0)
 145                         goto out_bad;
 146                 /* pool namespace */
 147                 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
 148                 if (info->pool_ns_len > 0) {
 149                         ceph_decode_need(p, end, info->pool_ns_len, bad);
 150                         info->pool_ns_data = *p;
 151                         *p += info->pool_ns_len;
 152                 }
 153
 154                 /* btime */
 155                 ceph_decode_need(p, end, sizeof(info->btime), bad);
 156                 ceph_decode_copy(p, &info->btime, sizeof(info->btime));
 157
 158                 /* change attribute */
 159                 ceph_decode_64_safe(p, end, info->change_attr, bad);
 160
 161                 /* dir pin */
 162                 if (struct_v >= 2) {
 163                         ceph_decode_32_safe(p, end, info->dir_pin, bad);
 164                 } else {
 165                         info->dir_pin = -ENODATA;
 166                 }
 167
 168                 /* snapshot birth time, remains zero for v<=2 */
 169                 if (struct_v >= 3) {
 170                         ceph_decode_need(p, end, sizeof(info->snap_btime), bad);
 171                         ceph_decode_copy(p, &info->snap_btime,
 172                                          sizeof(info->snap_btime));
 173                 } else {
 174                         memset(&info->snap_btime, 0, sizeof(info->snap_btime));
 175                 }
 176
 177                 *p = end;
 178         } else {
 179                 if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
 180                         ceph_decode_64_safe(p, end, info->inline_version, bad);
 181                         ceph_decode_32_safe(p, end, info->inline_len, bad);
 182                         ceph_decode_need(p, end, info->inline_len, bad);
 183                         info->inline_data = *p;
 184                         *p += info->inline_len;
 185                 } else
 186                         info->inline_version = CEPH_INLINE_NONE;
 187
 188                 if (features & CEPH_FEATURE_MDS_QUOTA) {
 189                         err = parse_reply_info_quota(p, end, info);
 190                         if (err < 0)
 191                                 goto out_bad;
 192                 } else {
 193                         info->max_bytes = 0;
 194                         info->max_files = 0;
 195                 }
 196
 197                 info->pool_ns_len = 0;
 198                 info->pool_ns_data = NULL;
 199                 if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
 200                         ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
 201                         if (info->pool_ns_len > 0) {
 202                                 ceph_decode_need(p, end, info->pool_ns_len, bad);
 203                                 info->pool_ns_data = *p;
 204                                 *p += info->pool_ns_len;
 205                         }
 206                 }
 207
 208                 if (features & CEPH_FEATURE_FS_BTIME) {
 209                         ceph_decode_need(p, end, sizeof(info->btime), bad);
 210                         ceph_decode_copy(p, &info->btime, sizeof(info->btime));
 211                         ceph_decode_64_safe(p, end, info->change_attr, bad);
 212                 }
 213
 214                 info->dir_pin = -ENODATA;
 215                 /* info->snap_btime remains zero */
 216         }
 217         return 0;
 218 bad:
 219         err = -EIO;
 220 out_bad:
 221         return err;
 222 }
 223
 224 static int parse_reply_info_dir(void **p, void *end,
 225                                 struct ceph_mds_reply_dirfrag **dirfrag,
 226                                 u64 features)
 227 {
 228         if (features == (u64)-1) {
 229                 u8 struct_v, struct_compat;
 230                 u32 struct_len;
 231                 ceph_decode_8_safe(p, end, struct_v, bad);
 232                 ceph_decode_8_safe(p, end, struct_compat, bad);
 233                 /* struct_v is expected to be >= 1. we only understand
 234                  * encoding whose struct_compat == 1. */
 235                 if (!struct_v || struct_compat != 1)
 236                         goto bad;
 237                 ceph_decode_32_safe(p, end, struct_len, bad);
 238                 ceph_decode_need(p, end, struct_len, bad);
 239                 end = *p + struct_len;
 240         }
 241
 242         ceph_decode_need(p, end, sizeof(**dirfrag), bad);
 243         *dirfrag = *p;
 244         *p += sizeof(**dirfrag) + sizeof(u32) * le32_to_cpu((*dirfrag)->ndist);
 245         if (unlikely(*p > end))
 246                 goto bad;
 247         if (features == (u64)-1)
 248                 *p = end;
 249         return 0;
 250 bad:
 251         return -EIO;
 252 }
 253
 254 static int parse_reply_info_lease(void **p, void *end,
 255                                   struct ceph_mds_reply_lease **lease,
 256                                   u64 features)
 257 {
 258         if (features == (u64)-1) {
 259                 u8 struct_v, struct_compat;
 260                 u32 struct_len;
 261                 ceph_decode_8_safe(p, end, struct_v, bad);
 262                 ceph_decode_8_safe(p, end, struct_compat, bad);
 263                 /* struct_v is expected to be >= 1. we only understand
 264                  * encoding whose struct_compat == 1. */
 265                 if (!struct_v || struct_compat != 1)
 266                         goto bad;
 267                 ceph_decode_32_safe(p, end, struct_len, bad);
 268                 ceph_decode_need(p, end, struct_len, bad);
 269                 end = *p + struct_len;
 270         }
 271
 272         ceph_decode_need(p, end, sizeof(**lease), bad);
 273         *lease = *p;
 274         *p += sizeof(**lease);
 275         if (features == (u64)-1)
 276                 *p = end;
 277         return 0;
 278 bad:
 279         return -EIO;
 280 }
 281
 282 /*
 283  * parse a normal reply, which may contain a (dir+)dentry and/or a
 284  * target inode.
 285  */
 286 static int parse_reply_info_trace(void **p, void *end,
 287                                   struct ceph_mds_reply_info_parsed *info,
 288                                   u64 features)
 289 {
 290         int err;
 291
 292         if (info->head->is_dentry) {
 293                 err = parse_reply_info_in(p, end, &info->diri, features);
 294                 if (err < 0)
 295                         goto out_bad;
 296
 297                 err = parse_reply_info_dir(p, end, &info->dirfrag, features);
 298                 if (err < 0)
 299                         goto out_bad;
 300
 301                 ceph_decode_32_safe(p, end, info->dname_len, bad);
 302                 ceph_decode_need(p, end, info->dname_len, bad);
 303                 info->dname = *p;
 304                 *p += info->dname_len;
 305
 306                 err = parse_reply_info_lease(p, end, &info->dlease, features);
 307                 if (err < 0)
 308                         goto out_bad;
 309         }
 310
 311         if (info->head->is_target) {
 312                 err = parse_reply_info_in(p, end, &info->targeti, features);
 313                 if (err < 0)
 314                         goto out_bad;
 315         }
 316
 317         if (unlikely(*p != end))
 318                 goto bad;
 319         return 0;
 320
 321 bad:
 322         err = -EIO;
 323 out_bad:
 324         pr_err("problem parsing mds trace %d\n", err);
 325         return err;
 326 }
 327
 328 /*
 329  * parse readdir results
 330  */
 331 static int parse_reply_info_readdir(void **p, void *end,
 332                                 struct ceph_mds_reply_info_parsed *info,
 333                                 u64 features)
 334 {
 335         u32 num, i = 0;
 336         int err;
 337
 338         err = parse_reply_info_dir(p, end, &info->dir_dir, features);
 339         if (err < 0)
 340                 goto out_bad;
 341
 342         ceph_decode_need(p, end, sizeof(num) + 2, bad);
 343         num = ceph_decode_32(p);
 344         {
 345                 u16 flags = ceph_decode_16(p);
 346                 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
 347                 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
 348                 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
 349                 info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
 350         }
 351         if (num == 0)
 352                 goto done;
 353
 354         BUG_ON(!info->dir_entries);
 355         if ((unsigned long)(info->dir_entries + num) >
 356             (unsigned long)info->dir_entries + info->dir_buf_size) {
 357                 pr_err("dir contents are larger than expected\n");
 358                 WARN_ON(1);
 359                 goto bad;
 360         }
 361
 362         info->dir_nr = num;
 363         while (num) {
 364                 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
 365                 /* dentry */
 366                 ceph_decode_32_safe(p, end, rde->name_len, bad);
 367                 ceph_decode_need(p, end, rde->name_len, bad);
 368                 rde->name = *p;
 369                 *p += rde->name_len;
 370                 dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
 371
 372                 /* dentry lease */
 373                 err = parse_reply_info_lease(p, end, &rde->lease, features);
 374                 if (err)
 375                         goto out_bad;
 376                 /* inode */
 377                 err = parse_reply_info_in(p, end, &rde->inode, features);
 378                 if (err < 0)
 379                         goto out_bad;
 380                 /* ceph_readdir_prepopulate() will update it */
 381                 rde->offset = 0;
 382                 i++;
 383                 num--;
 384         }
 385
 386 done:
 387         if (*p != end)
 388                 goto bad;
 389         return 0;
 390
 391 bad:
 392         err = -EIO;
 393 out_bad:
 394         pr_err("problem parsing dir contents %d\n", err);
 395         return err;
 396 }
 397
 398 /*
 399  * parse fcntl F_GETLK results
 400  */
 401 static int parse_reply_info_filelock(void **p, void *end,
 402                                      struct ceph_mds_reply_info_parsed *info,
 403                                      u64 features)
 404 {
 405         if (*p + sizeof(*info->filelock_reply) > end)
 406                 goto bad;
 407
 408         info->filelock_reply = *p;
 409         *p += sizeof(*info->filelock_reply);
 410
 411         if (unlikely(*p != end))
 412                 goto bad;
 413         return 0;
 414
 415 bad:
 416         return -EIO;
 417 }
 418
 419 /*
 420  * parse create results
 421  */
 422 static int parse_reply_info_create(void **p, void *end,
 423                                   struct ceph_mds_reply_info_parsed *info,
 424                                   u64 features)
 425 {
 426         if (features == (u64)-1 ||
 427             (features & CEPH_FEATURE_REPLY_CREATE_INODE)) {
 428                 if (*p == end) {
 429                         info->has_create_ino = false;
 430                 } else {
 431                         info->has_create_ino = true;
 432                         info->ino = ceph_decode_64(p);
 433                 }
 434         }
 435
 436         if (unlikely(*p != end))
 437                 goto bad;
 438         return 0;
 439
 440 bad:
 441         return -EIO;
 442 }
 443
 444 /*
 445  * parse extra results
 446  */
 447 static int parse_reply_info_extra(void **p, void *end,
 448                                   struct ceph_mds_reply_info_parsed *info,
 449                                   u64 features)
 450 {
 451         u32 op = le32_to_cpu(info->head->op);
 452
 453         if (op == CEPH_MDS_OP_GETFILELOCK)
 454                 return parse_reply_info_filelock(p, end, info, features);
 455         else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
 456                 return parse_reply_info_readdir(p, end, info, features);
 457         else if (op == CEPH_MDS_OP_CREATE)
 458                 return parse_reply_info_create(p, end, info, features);
 459         else
 460                 return -EIO;
 461 }
 462
 463 /*
 464  * parse entire mds reply
 465  */
 466 static int parse_reply_info(struct ceph_msg *msg,
 467                             struct ceph_mds_reply_info_parsed *info,
 468                             u64 features)
 469 {
 470         void *p, *end;
 471         u32 len;
 472         int err;
 473
 474         info->head = msg->front.iov_base;
 475         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 476         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 477
 478         /* trace */
 479         ceph_decode_32_safe(&p, end, len, bad);
 480         if (len > 0) {
 481                 ceph_decode_need(&p, end, len, bad);
 482                 err = parse_reply_info_trace(&p, p+len, info, features);
 483                 if (err < 0)
 484                         goto out_bad;
 485         }
 486
 487         /* extra */
 488         ceph_decode_32_safe(&p, end, len, bad);
 489         if (len > 0) {
 490                 ceph_decode_need(&p, end, len, bad);
 491                 err = parse_reply_info_extra(&p, p+len, info, features);
 492                 if (err < 0)
 493                         goto out_bad;
 494         }
 495
 496         /* snap blob */
 497         ceph_decode_32_safe(&p, end, len, bad);
 498         info->snapblob_len = len;
 499         info->snapblob = p;
 500         p += len;
 501
 502         if (p != end)
 503                 goto bad;
 504         return 0;
 505
 506 bad:
 507         err = -EIO;
 508 out_bad:
 509         pr_err("mds parse_reply err %d\n", err);
 510         return err;
 511 }
 512
 513 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
 514 {
 515         if (!info->dir_entries)
 516                 return;
 517         free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
 518 }
 519
 520
 521 /*
 522  * sessions
 523  */
 524 const char *ceph_session_state_name(int s)
 525 {
 526         switch (s) {
 527         case CEPH_MDS_SESSION_NEW: return "new";
 528         case CEPH_MDS_SESSION_OPENING: return "opening";
 529         case CEPH_MDS_SESSION_OPEN: return "open";
 530         case CEPH_MDS_SESSION_HUNG: return "hung";
 531         case CEPH_MDS_SESSION_CLOSING: return "closing";
 532         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
 533         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
 534         case CEPH_MDS_SESSION_REJECTED: return "rejected";
 535         default: return "???";
 536         }
 537 }
 538
 539 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
 540 {
 541         if (refcount_inc_not_zero(&s->s_ref)) {
 542                 dout("mdsc get_session %p %d -> %d\n", s,
 543                      refcount_read(&s->s_ref)-1, refcount_read(&s->s_ref));
 544                 return s;
 545         } else {
 546                 dout("mdsc get_session %p 0 -- FAIL\n", s);
 547                 return NULL;
 548         }
 549 }
 550
 551 void ceph_put_mds_session(struct ceph_mds_session *s)
 552 {
 553         dout("mdsc put_session %p %d -> %d\n", s,
 554              refcount_read(&s->s_ref), refcount_read(&s->s_ref)-1);
 555         if (refcount_dec_and_test(&s->s_ref)) {
 556                 if (s->s_auth.authorizer)
 557                         ceph_auth_destroy_authorizer(s->s_auth.authorizer);
 558                 kfree(s);
 559         }
 560 }
 561
 562 /*
 563  * called under mdsc->mutex
 564  */
 565 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
 566                                                    int mds)
 567 {
 568         if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
 569                 return NULL;
 570         return get_session(mdsc->sessions[mds]);
 571 }
 572
 573 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
 574 {
 575         if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
 576                 return false;
 577         else
 578                 return true;
 579 }
 580
 581 static int __verify_registered_session(struct ceph_mds_client *mdsc,
 582                                        struct ceph_mds_session *s)
 583 {
 584         if (s->s_mds >= mdsc->max_sessions ||
 585             mdsc->sessions[s->s_mds] != s)
 586                 return -ENOENT;
 587         return 0;
 588 }
 589
 590 /*
 591  * create+register a new session for given mds.
 592  * called under mdsc->mutex.
 593  */
 594 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 595                                                  int mds)
 596 {
 597         struct ceph_mds_session *s;
 598
 599         if (mds >= mdsc->mdsmap->m_num_mds)
 600                 return ERR_PTR(-EINVAL);
 601
 602         s = kzalloc(sizeof(*s), GFP_NOFS);
 603         if (!s)
 604                 return ERR_PTR(-ENOMEM);
 605
 606         if (mds >= mdsc->max_sessions) {
 607                 int newmax = 1 << get_count_order(mds + 1);
 608                 struct ceph_mds_session **sa;
 609
 610                 dout("%s: realloc to %d\n", __func__, newmax);
 611                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 612                 if (!sa)
 613                         goto fail_realloc;
 614                 if (mdsc->sessions) {
 615                         memcpy(sa, mdsc->sessions,
 616                                mdsc->max_sessions * sizeof(void *));
 617                         kfree(mdsc->sessions);
 618                 }
 619                 mdsc->sessions = sa;
 620                 mdsc->max_sessions = newmax;
 621         }
 622
 623         dout("%s: mds%d\n", __func__, mds);
 624         s->s_mdsc = mdsc;
 625         s->s_mds = mds;
 626         s->s_state = CEPH_MDS_SESSION_NEW;
 627         s->s_ttl = 0;
 628         s->s_seq = 0;
 629         mutex_init(&s->s_mutex);
 630
 631         ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
 632
 633         spin_lock_init(&s->s_gen_ttl_lock);
 634         s->s_cap_gen = 1;
 635         s->s_cap_ttl = jiffies - 1;
 636
 637         spin_lock_init(&s->s_cap_lock);
 638         s->s_renew_requested = 0;
 639         s->s_renew_seq = 0;
 640         INIT_LIST_HEAD(&s->s_caps);
 641         s->s_nr_caps = 0;
 642         refcount_set(&s->s_ref, 1);
 643         INIT_LIST_HEAD(&s->s_waiting);
 644         INIT_LIST_HEAD(&s->s_unsafe);
 645         s->s_num_cap_releases = 0;
 646         s->s_cap_reconnect = 0;
 647         s->s_cap_iterator = NULL;
 648         INIT_LIST_HEAD(&s->s_cap_releases);
 649         INIT_WORK(&s->s_cap_release_work, ceph_cap_release_work);
 650
 651         INIT_LIST_HEAD(&s->s_cap_flushing);
 652
 653         mdsc->sessions[mds] = s;
 654         atomic_inc(&mdsc->num_sessions);
 655         refcount_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
 656
 657         ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
 658                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
 659
 660         return s;
 661
 662 fail_realloc:
 663         kfree(s);
 664         return ERR_PTR(-ENOMEM);
 665 }
 666
 667 /*
 668  * called under mdsc->mutex
 669  */
 670 static void __unregister_session(struct ceph_mds_client *mdsc,
 671                                struct ceph_mds_session *s)
 672 {
 673         dout("__unregister_session mds%d %p\n", s->s_mds, s);
 674         BUG_ON(mdsc->sessions[s->s_mds] != s);
 675         mdsc->sessions[s->s_mds] = NULL;
 676         s->s_state = 0;
 677         ceph_con_close(&s->s_con);
 678         ceph_put_mds_session(s);
 679         atomic_dec(&mdsc->num_sessions);
 680 }
 681
 682 /*
 683  * drop session refs in request.
 684  *
 685  * should be last request ref, or hold mdsc->mutex
 686  */
 687 static void put_request_session(struct ceph_mds_request *req)
 688 {
 689         if (req->r_session) {
 690                 ceph_put_mds_session(req->r_session);
 691                 req->r_session = NULL;
 692         }
 693 }
 694
 695 void ceph_mdsc_release_request(struct kref *kref)
 696 {
 697         struct ceph_mds_request *req = container_of(kref,
 698                                                     struct ceph_mds_request,
 699                                                     r_kref);
 700         destroy_reply_info(&req->r_reply_info);
 701         if (req->r_request)
 702                 ceph_msg_put(req->r_request);
 703         if (req->r_reply)
 704                 ceph_msg_put(req->r_reply);
 705         if (req->r_inode) {
 706                 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
 707                 /* avoid calling iput_final() in mds dispatch threads */
 708                 ceph_async_iput(req->r_inode);
 709         }
 710         if (req->r_parent)
 711                 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
 712         ceph_async_iput(req->r_target_inode);
 713         if (req->r_dentry)
 714                 dput(req->r_dentry);
 715         if (req->r_old_dentry)
 716                 dput(req->r_old_dentry);
 717         if (req->r_old_dentry_dir) {
 718                 /*
 719                  * track (and drop pins for) r_old_dentry_dir
 720                  * separately, since r_old_dentry's d_parent may have
 721                  * changed between the dir mutex being dropped and
 722                  * this request being freed.
 723                  */
 724                 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
 725                                   CEPH_CAP_PIN);
 726                 ceph_async_iput(req->r_old_dentry_dir);
 727         }
 728         kfree(req->r_path1);
 729         kfree(req->r_path2);
 730         if (req->r_pagelist)
 731                 ceph_pagelist_release(req->r_pagelist);
 732         put_request_session(req);
 733         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
 734         WARN_ON_ONCE(!list_empty(&req->r_wait));
 735         kfree(req);
 736 }
 737
 738 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
 739
 740 /*
 741  * lookup session, bump ref if found.
 742  *
 743  * called under mdsc->mutex.
 744  */
 745 static struct ceph_mds_request *
 746 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
 747 {
 748         struct ceph_mds_request *req;
 749
 750         req = lookup_request(&mdsc->request_tree, tid);
 751         if (req)
 752                 ceph_mdsc_get_request(req);
 753
 754         return req;
 755 }
 756
 757 /*
 758  * Register an in-flight request, and assign a tid.  Link to directory
 759  * are modifying (if any).
 760  *
 761  * Called under mdsc->mutex.
 762  */
 763 static void __register_request(struct ceph_mds_client *mdsc,
 764                                struct ceph_mds_request *req,
 765                                struct inode *dir)
 766 {
 767         int ret = 0;
 768
 769         req->r_tid = ++mdsc->last_tid;
 770         if (req->r_num_caps) {
 771                 ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
 772                                         req->r_num_caps);
 773                 if (ret < 0) {
 774                         pr_err("__register_request %p "
 775                                "failed to reserve caps: %d\n", req, ret);
 776                         /* set req->r_err to fail early from __do_request */
 777                         req->r_err = ret;
 778                         return;
 779                 }
 780         }
 781         dout("__register_request %p tid %lld\n", req, req->r_tid);
 782         ceph_mdsc_get_request(req);
 783         insert_request(&mdsc->request_tree, req);
 784
 785         req->r_uid = current_fsuid();
 786         req->r_gid = current_fsgid();
 787
 788         if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
 789                 mdsc->oldest_tid = req->r_tid;
 790
 791         if (dir) {
 792                 ihold(dir);
 793                 req->r_unsafe_dir = dir;
 794         }
 795 }
 796
 797 static void __unregister_request(struct ceph_mds_client *mdsc,
 798                                  struct ceph_mds_request *req)
 799 {
 800         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 801
 802         /* Never leave an unregistered request on an unsafe list! */
 803         list_del_init(&req->r_unsafe_item);
 804
 805         if (req->r_tid == mdsc->oldest_tid) {
 806                 struct rb_node *p = rb_next(&req->r_node);
 807                 mdsc->oldest_tid = 0;
 808                 while (p) {
 809                         struct ceph_mds_request *next_req =
 810                                 rb_entry(p, struct ceph_mds_request, r_node);
 811                         if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
 812                                 mdsc->oldest_tid = next_req->r_tid;
 813                                 break;
 814                         }
 815                         p = rb_next(p);
 816                 }
 817         }
 818
 819         erase_request(&mdsc->request_tree, req);
 820
 821         if (req->r_unsafe_dir  &&
 822             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
 823                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 824                 spin_lock(&ci->i_unsafe_lock);
 825                 list_del_init(&req->r_unsafe_dir_item);
 826                 spin_unlock(&ci->i_unsafe_lock);
 827         }
 828         if (req->r_target_inode &&
 829             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
 830                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
 831                 spin_lock(&ci->i_unsafe_lock);
 832                 list_del_init(&req->r_unsafe_target_item);
 833                 spin_unlock(&ci->i_unsafe_lock);
 834         }
 835
 836         if (req->r_unsafe_dir) {
 837                 /* avoid calling iput_final() in mds dispatch threads */
 838                 ceph_async_iput(req->r_unsafe_dir);
 839                 req->r_unsafe_dir = NULL;
 840         }
 841
 842         complete_all(&req->r_safe_completion);
 843
 844         ceph_mdsc_put_request(req);
 845 }
 846
 847 /*
 848  * Walk back up the dentry tree until we hit a dentry representing a
 849  * non-snapshot inode. We do this using the rcu_read_lock (which must be held
 850  * when calling this) to ensure that the objects won't disappear while we're
 851  * working with them. Once we hit a candidate dentry, we attempt to take a
 852  * reference to it, and return that as the result.
 853  */
 854 static struct inode *get_nonsnap_parent(struct dentry *dentry)
 855 {
 856         struct inode *inode = NULL;
 857
 858         while (dentry && !IS_ROOT(dentry)) {
 859                 inode = d_inode_rcu(dentry);
 860                 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
 861                         break;
 862                 dentry = dentry->d_parent;
 863         }
 864         if (inode)
 865                 inode = igrab(inode);
 866         return inode;
 867 }
 868
 869 /*
 870  * Choose mds to send request to next.  If there is a hint set in the
 871  * request (e.g., due to a prior forward hint from the mds), use that.
 872  * Otherwise, consult frag tree and/or caps to identify the
 873  * appropriate mds.  If all else fails, choose randomly.
 874  *
 875  * Called under mdsc->mutex.
 876  */
 877 static int __choose_mds(struct ceph_mds_client *mdsc,
 878                         struct ceph_mds_request *req)
 879 {
 880         struct inode *inode;
 881         struct ceph_inode_info *ci;
 882         struct ceph_cap *cap;
 883         int mode = req->r_direct_mode;
 884         int mds = -1;
 885         u32 hash = req->r_direct_hash;
 886         bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
 887
 888         /*
 889          * is there a specific mds we should try?  ignore hint if we have
 890          * no session and the mds is not up (active or recovering).
 891          */
 892         if (req->r_resend_mds >= 0 &&
 893             (__have_session(mdsc, req->r_resend_mds) ||
 894              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 895                 dout("choose_mds using resend_mds mds%d\n",
 896                      req->r_resend_mds);
 897                 return req->r_resend_mds;
 898         }
 899
 900         if (mode == USE_RANDOM_MDS)
 901                 goto random;
 902
 903         inode = NULL;
 904         if (req->r_inode) {
 905                 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
 906                         inode = req->r_inode;
 907                         ihold(inode);
 908                 } else {
 909                         /* req->r_dentry is non-null for LSSNAP request */
 910                         rcu_read_lock();
 911                         inode = get_nonsnap_parent(req->r_dentry);
 912                         rcu_read_unlock();
 913                         dout("__choose_mds using snapdir's parent %p\n", inode);
 914                 }
 915         } else if (req->r_dentry) {
 916                 /* ignore race with rename; old or new d_parent is okay */
 917                 struct dentry *parent;
 918                 struct inode *dir;
 919
 920                 rcu_read_lock();
 921                 parent = READ_ONCE(req->r_dentry->d_parent);
 922                 dir = req->r_parent ? : d_inode_rcu(parent);
 923
 924                 if (!dir || dir->i_sb != mdsc->fsc->sb) {
 925                         /*  not this fs or parent went negative */
 926                         inode = d_inode(req->r_dentry);
 927                         if (inode)
 928                                 ihold(inode);
 929                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
 930                         /* direct snapped/virtual snapdir requests
 931                          * based on parent dir inode */
 932                         inode = get_nonsnap_parent(parent);
 933                         dout("__choose_mds using nonsnap parent %p\n", inode);
 934                 } else {
 935                         /* dentry target */
 936                         inode = d_inode(req->r_dentry);
 937                         if (!inode || mode == USE_AUTH_MDS) {
 938                                 /* dir + name */
 939                                 inode = igrab(dir);
 940                                 hash = ceph_dentry_hash(dir, req->r_dentry);
 941                                 is_hash = true;
 942                         } else {
 943                                 ihold(inode);
 944                         }
 945                 }
 946                 rcu_read_unlock();
 947         }
 948
 949         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 950              (int)hash, mode);
 951         if (!inode)
 952                 goto random;
 953         ci = ceph_inode(inode);
 954
 955         if (is_hash && S_ISDIR(inode->i_mode)) {
 956                 struct ceph_inode_frag frag;
 957                 int found;
 958
 959                 ceph_choose_frag(ci, hash, &frag, &found);
 960                 if (found) {
 961                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 962                                 u8 r;
 963
 964                                 /* choose a random replica */
 965                                 get_random_bytes(&r, 1);
 966                                 r %= frag.ndist;
 967                                 mds = frag.dist[r];
 968                                 dout("choose_mds %p %llx.%llx "
 969                                      "frag %u mds%d (%d/%d)\n",
 970                                      inode, ceph_vinop(inode),
 971                                      frag.frag, mds,
 972                                      (int)r, frag.ndist);
 973                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 974                                     CEPH_MDS_STATE_ACTIVE)
 975                                         goto out;
 976                         }
 977
 978                         /* since this file/dir wasn't known to be
 979                          * replicated, then we want to look for the
 980                          * authoritative mds. */
 981                         mode = USE_AUTH_MDS;
 982                         if (frag.mds >= 0) {
 983                                 /* choose auth mds */
 984                                 mds = frag.mds;
 985                                 dout("choose_mds %p %llx.%llx "
 986                                      "frag %u mds%d (auth)\n",
 987                                      inode, ceph_vinop(inode), frag.frag, mds);
 988                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 989                                     CEPH_MDS_STATE_ACTIVE)
 990                                         goto out;
 991                         }
 992                 }
 993         }
 994
 995         spin_lock(&ci->i_ceph_lock);
 996         cap = NULL;
 997         if (mode == USE_AUTH_MDS)
 998                 cap = ci->i_auth_cap;
 999         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
1000                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
1001         if (!cap) {
1002                 spin_unlock(&ci->i_ceph_lock);
1003                 ceph_async_iput(inode);
1004                 goto random;
1005         }
1006         mds = cap->session->s_mds;
1007         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
1008              inode, ceph_vinop(inode), mds,
1009              cap == ci->i_auth_cap ? "auth " : "", cap);
1010         spin_unlock(&ci->i_ceph_lock);
1011 out:
1012         /* avoid calling iput_final() while holding mdsc->mutex or
1013          * in mds dispatch threads */
1014         ceph_async_iput(inode);
1015         return mds;
1016
1017 random:
1018         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
1019         dout("choose_mds chose random mds%d\n", mds);
1020         return mds;
1021 }
1022
1023
1024 /*
1025  * session messages
1026  */
1027 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
1028 {
1029         struct ceph_msg *msg;
1030         struct ceph_mds_session_head *h;
1031
1032         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
1033                            false);
1034         if (!msg) {
1035                 pr_err("create_session_msg ENOMEM creating msg\n");
1036                 return NULL;
1037         }
1038         h = msg->front.iov_base;
1039         h->op = cpu_to_le32(op);
1040         h->seq = cpu_to_le64(seq);
1041
1042         return msg;
1043 }
1044
1045 static void encode_supported_features(void **p, void *end)
1046 {
1047         static const unsigned char bits[] = CEPHFS_FEATURES_CLIENT_SUPPORTED;
1048         static const size_t count = ARRAY_SIZE(bits);
1049
1050         if (count > 0) {
1051                 size_t i;
1052                 size_t size = ((size_t)bits[count - 1] + 64) / 64 * 8;
1053
1054                 BUG_ON(*p + 4 + size > end);
1055                 ceph_encode_32(p, size);
1056                 memset(*p, 0, size);
1057                 for (i = 0; i < count; i++)
1058                         ((unsigned char*)(*p))[i / 8] |= 1 << (bits[i] % 8);
1059                 *p += size;
1060         } else {
1061                 BUG_ON(*p + 4 > end);
1062                 ceph_encode_32(p, 0);
1063         }
1064 }
1065
1066 /*
1067  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
1068  * to include additional client metadata fields.
1069  */
1070 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
1071 {
1072         struct ceph_msg *msg;
1073         struct ceph_mds_session_head *h;
1074         int i = -1;
1075         int extra_bytes = 0;
1076         int metadata_key_count = 0;
1077         struct ceph_options *opt = mdsc->fsc->client->options;
1078         struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
1079         void *p, *end;
1080
1081         const char* metadata[][2] = {
1082                 {"hostname", mdsc->nodename},
1083                 {"kernel_version", init_utsname()->release},
1084                 {"entity_id", opt->name ? : ""},
1085                 {"root", fsopt->server_path ? : "/"},
1086                 {NULL, NULL}
1087         };
1088
1089         /* Calculate serialized length of metadata */
1090         extra_bytes = 4;  /* map length */
1091         for (i = 0; metadata[i][0]; ++i) {
1092                 extra_bytes += 8 + strlen(metadata[i][0]) +
1093                         strlen(metadata[i][1]);
1094                 metadata_key_count++;
1095         }
1096         /* supported feature */
1097         extra_bytes += 4 + 8;
1098
1099         /* Allocate the message */
1100         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + extra_bytes,
1101                            GFP_NOFS, false);
1102         if (!msg) {
1103                 pr_err("create_session_msg ENOMEM creating msg\n");
1104                 return NULL;
1105         }
1106         p = msg->front.iov_base;
1107         end = p + msg->front.iov_len;
1108
1109         h = p;
1110         h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
1111         h->seq = cpu_to_le64(seq);
1112
1113         /*
1114          * Serialize client metadata into waiting buffer space, using
1115          * the format that userspace expects for map<string, string>
1116          *
1117          * ClientSession messages with metadata are v2
1118          */
1119         msg->hdr.version = cpu_to_le16(3);
1120         msg->hdr.compat_version = cpu_to_le16(1);
1121
1122         /* The write pointer, following the session_head structure */
1123         p += sizeof(*h);
1124
1125         /* Number of entries in the map */
1126         ceph_encode_32(&p, metadata_key_count);
1127
1128         /* Two length-prefixed strings for each entry in the map */
1129         for (i = 0; metadata[i][0]; ++i) {
1130                 size_t const key_len = strlen(metadata[i][0]);
1131                 size_t const val_len = strlen(metadata[i][1]);
1132
1133                 ceph_encode_32(&p, key_len);
1134                 memcpy(p, metadata[i][0], key_len);
1135                 p += key_len;
1136                 ceph_encode_32(&p, val_len);
1137                 memcpy(p, metadata[i][1], val_len);
1138                 p += val_len;
1139         }
1140
1141         encode_supported_features(&p, end);
1142         msg->front.iov_len = p - msg->front.iov_base;
1143         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1144
1145         return msg;
1146 }
1147
1148 /*
1149  * send session open request.
1150  *
1151  * called under mdsc->mutex
1152  */
1153 static int __open_session(struct ceph_mds_client *mdsc,
1154                           struct ceph_mds_session *session)
1155 {
1156         struct ceph_msg *msg;
1157         int mstate;
1158         int mds = session->s_mds;
1159
1160         /* wait for mds to go active? */
1161         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
1162         dout("open_session to mds%d (%s)\n", mds,
1163              ceph_mds_state_name(mstate));
1164         session->s_state = CEPH_MDS_SESSION_OPENING;
1165         session->s_renew_requested = jiffies;
1166
1167         /* send connect message */
1168         msg = create_session_open_msg(mdsc, session->s_seq);
1169         if (!msg)
1170                 return -ENOMEM;
1171         ceph_con_send(&session->s_con, msg);
1172         return 0;
1173 }
1174
1175 /*
1176  * open sessions for any export targets for the given mds
1177  *
1178  * called under mdsc->mutex
1179  */
1180 static struct ceph_mds_session *
1181 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
1182 {
1183         struct ceph_mds_session *session;
1184
1185         session = __ceph_lookup_mds_session(mdsc, target);
1186         if (!session) {
1187                 session = register_session(mdsc, target);
1188                 if (IS_ERR(session))
1189                         return session;
1190         }
1191         if (session->s_state == CEPH_MDS_SESSION_NEW ||
1192             session->s_state == CEPH_MDS_SESSION_CLOSING)
1193                 __open_session(mdsc, session);
1194
1195         return session;
1196 }
1197
1198 struct ceph_mds_session *
1199 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1200 {
1201         struct ceph_mds_session *session;
1202
1203         dout("open_export_target_session to mds%d\n", target);
1204
1205         mutex_lock(&mdsc->mutex);
1206         session = __open_export_target_session(mdsc, target);
1207         mutex_unlock(&mdsc->mutex);
1208
1209         return session;
1210 }
1211
1212 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1213                                           struct ceph_mds_session *session)
1214 {
1215         struct ceph_mds_info *mi;
1216         struct ceph_mds_session *ts;
1217         int i, mds = session->s_mds;
1218
1219         if (mds >= mdsc->mdsmap->m_num_mds)
1220                 return;
1221
1222         mi = &mdsc->mdsmap->m_info[mds];
1223         dout("open_export_target_sessions for mds%d (%d targets)\n",
1224              session->s_mds, mi->num_export_targets);
1225
1226         for (i = 0; i < mi->num_export_targets; i++) {
1227                 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1228                 if (!IS_ERR(ts))
1229                         ceph_put_mds_session(ts);
1230         }
1231 }
1232
1233 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1234                                            struct ceph_mds_session *session)
1235 {
1236         mutex_lock(&mdsc->mutex);
1237         __open_export_target_sessions(mdsc, session);
1238         mutex_unlock(&mdsc->mutex);
1239 }
1240
1241 /*
1242  * session caps
1243  */
1244
1245 static void detach_cap_releases(struct ceph_mds_session *session,
1246                                 struct list_head *target)
1247 {
1248         lockdep_assert_held(&session->s_cap_lock);
1249
1250         list_splice_init(&session->s_cap_releases, target);
1251         session->s_num_cap_releases = 0;
1252         dout("dispose_cap_releases mds%d\n", session->s_mds);
1253 }
1254
1255 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1256                                  struct list_head *dispose)
1257 {
1258         while (!list_empty(dispose)) {
1259                 struct ceph_cap *cap;
1260                 /* zero out the in-progress message */
1261                 cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1262                 list_del(&cap->session_caps);
1263                 ceph_put_cap(mdsc, cap);
1264         }
1265 }
1266
1267 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1268                                      struct ceph_mds_session *session)
1269 {
1270         struct ceph_mds_request *req;
1271         struct rb_node *p;
1272         struct ceph_inode_info *ci;
1273
1274         dout("cleanup_session_requests mds%d\n", session->s_mds);
1275         mutex_lock(&mdsc->mutex);
1276         while (!list_empty(&session->s_unsafe)) {
1277                 req = list_first_entry(&session->s_unsafe,
1278                                        struct ceph_mds_request, r_unsafe_item);
1279                 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1280                                     req->r_tid);
1281                 if (req->r_target_inode) {
1282                         /* dropping unsafe change of inode's attributes */
1283                         ci = ceph_inode(req->r_target_inode);
1284                         errseq_set(&ci->i_meta_err, -EIO);
1285                 }
1286                 if (req->r_unsafe_dir) {
1287                         /* dropping unsafe directory operation */
1288                         ci = ceph_inode(req->r_unsafe_dir);
1289                         errseq_set(&ci->i_meta_err, -EIO);
1290                 }
1291                 __unregister_request(mdsc, req);
1292         }
1293         /* zero r_attempts, so kick_requests() will re-send requests */
1294         p = rb_first(&mdsc->request_tree);
1295         while (p) {
1296                 req = rb_entry(p, struct ceph_mds_request, r_node);
1297                 p = rb_next(p);
1298                 if (req->r_session &&
1299                     req->r_session->s_mds == session->s_mds)
1300                         req->r_attempts = 0;
1301         }
1302         mutex_unlock(&mdsc->mutex);
1303 }
1304
1305 /*
1306  * Helper to safely iterate over all caps associated with a session, with
1307  * special care taken to handle a racing __ceph_remove_cap().
1308  *
1309  * Caller must hold session s_mutex.
1310  */
1311 int ceph_iterate_session_caps(struct ceph_mds_session *session,
1312                               int (*cb)(struct inode *, struct ceph_cap *,
1313                                         void *), void *arg)
1314 {
1315         struct list_head *p;
1316         struct ceph_cap *cap;
1317         struct inode *inode, *last_inode = NULL;
1318         struct ceph_cap *old_cap = NULL;
1319         int ret;
1320
1321         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1322         spin_lock(&session->s_cap_lock);
1323         p = session->s_caps.next;
1324         while (p != &session->s_caps) {
1325                 cap = list_entry(p, struct ceph_cap, session_caps);
1326                 inode = igrab(&cap->ci->vfs_inode);
1327                 if (!inode) {
1328                         p = p->next;
1329                         continue;
1330                 }
1331                 session->s_cap_iterator = cap;
1332                 spin_unlock(&session->s_cap_lock);
1333
1334                 if (last_inode) {
1335                         /* avoid calling iput_final() while holding
1336                          * s_mutex or in mds dispatch threads */
1337                         ceph_async_iput(last_inode);
1338                         last_inode = NULL;
1339                 }
1340                 if (old_cap) {
1341                         ceph_put_cap(session->s_mdsc, old_cap);
1342                         old_cap = NULL;
1343                 }
1344
1345                 ret = cb(inode, cap, arg);
1346                 last_inode = inode;
1347
1348                 spin_lock(&session->s_cap_lock);
1349                 p = p->next;
1350                 if (!cap->ci) {
1351                         dout("iterate_session_caps  finishing cap %p removal\n",
1352                              cap);
1353                         BUG_ON(cap->session != session);
1354                         cap->session = NULL;
1355                         list_del_init(&cap->session_caps);
1356                         session->s_nr_caps--;
1357                         if (cap->queue_release)
1358                                 __ceph_queue_cap_release(session, cap);
1359                         else
1360                                 old_cap = cap;  /* put_cap it w/o locks held */
1361                 }
1362                 if (ret < 0)
1363                         goto out;
1364         }
1365         ret = 0;
1366 out:
1367         session->s_cap_iterator = NULL;
1368         spin_unlock(&session->s_cap_lock);
1369
1370         ceph_async_iput(last_inode);
1371         if (old_cap)
1372                 ceph_put_cap(session->s_mdsc, old_cap);
1373
1374         return ret;
1375 }
1376
1377 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1378                                   void *arg)
1379 {
1380         struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1381         struct ceph_inode_info *ci = ceph_inode(inode);
1382         LIST_HEAD(to_remove);
1383         bool dirty_dropped = false;
1384         bool invalidate = false;
1385
1386         dout("removing cap %p, ci is %p, inode is %p\n",
1387              cap, ci, &ci->vfs_inode);
1388         spin_lock(&ci->i_ceph_lock);
1389         if (cap->mds_wanted | cap->issued)
1390                 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1391         __ceph_remove_cap(cap, false);
1392         if (!ci->i_auth_cap) {
1393                 struct ceph_cap_flush *cf;
1394                 struct ceph_mds_client *mdsc = fsc->mdsc;
1395
1396                 if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1397                         if (inode->i_data.nrpages > 0)
1398                                 invalidate = true;
1399                         if (ci->i_wrbuffer_ref > 0)
1400                                 mapping_set_error(&inode->i_data, -EIO);
1401                 }
1402
1403                 while (!list_empty(&ci->i_cap_flush_list)) {
1404                         cf = list_first_entry(&ci->i_cap_flush_list,
1405                                               struct ceph_cap_flush, i_list);
1406                         list_move(&cf->i_list, &to_remove);
1407                 }
1408
1409                 spin_lock(&mdsc->cap_dirty_lock);
1410
1411                 list_for_each_entry(cf, &to_remove, i_list)
1412                         list_del(&cf->g_list);
1413
1414                 if (!list_empty(&ci->i_dirty_item)) {
1415                         pr_warn_ratelimited(
1416                                 " dropping dirty %s state for %p %lld\n",
1417                                 ceph_cap_string(ci->i_dirty_caps),
1418                                 inode, ceph_ino(inode));
1419                         ci->i_dirty_caps = 0;
1420                         list_del_init(&ci->i_dirty_item);
1421                         dirty_dropped = true;
1422                 }
1423                 if (!list_empty(&ci->i_flushing_item)) {
1424                         pr_warn_ratelimited(
1425                                 " dropping dirty+flushing %s state for %p %lld\n",
1426                                 ceph_cap_string(ci->i_flushing_caps),
1427                                 inode, ceph_ino(inode));
1428                         ci->i_flushing_caps = 0;
1429                         list_del_init(&ci->i_flushing_item);
1430                         mdsc->num_cap_flushing--;
1431                         dirty_dropped = true;
1432                 }
1433                 spin_unlock(&mdsc->cap_dirty_lock);
1434
1435                 if (dirty_dropped) {
1436                         errseq_set(&ci->i_meta_err, -EIO);
1437
1438                         if (ci->i_wrbuffer_ref_head == 0 &&
1439                             ci->i_wr_ref == 0 &&
1440                             ci->i_dirty_caps == 0 &&
1441                             ci->i_flushing_caps == 0) {
1442                                 ceph_put_snap_context(ci->i_head_snapc);
1443                                 ci->i_head_snapc = NULL;
1444                         }
1445                 }
1446
1447                 if (atomic_read(&ci->i_filelock_ref) > 0) {
1448                         /* make further file lock syscall return -EIO */
1449                         ci->i_ceph_flags |= CEPH_I_ERROR_FILELOCK;
1450                         pr_warn_ratelimited(" dropping file locks for %p %lld\n",
1451                                             inode, ceph_ino(inode));
1452                 }
1453
1454                 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1455                         list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1456                         ci->i_prealloc_cap_flush = NULL;
1457                 }
1458         }
1459         spin_unlock(&ci->i_ceph_lock);
1460         while (!list_empty(&to_remove)) {
1461                 struct ceph_cap_flush *cf;
1462                 cf = list_first_entry(&to_remove,
1463                                       struct ceph_cap_flush, i_list);
1464                 list_del(&cf->i_list);
1465                 ceph_free_cap_flush(cf);
1466         }
1467
1468         wake_up_all(&ci->i_cap_wq);
1469         if (invalidate)
1470                 ceph_queue_invalidate(inode);
1471         if (dirty_dropped)
1472                 iput(inode);
1473         return 0;
1474 }
1475
1476 /*
1477  * caller must hold session s_mutex
1478  */
1479 static void remove_session_caps(struct ceph_mds_session *session)
1480 {
1481         struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1482         struct super_block *sb = fsc->sb;
1483         LIST_HEAD(dispose);
1484
1485         dout("remove_session_caps on %p\n", session);
1486         ceph_iterate_session_caps(session, remove_session_caps_cb, fsc);
1487
1488         wake_up_all(&fsc->mdsc->cap_flushing_wq);
1489
1490         spin_lock(&session->s_cap_lock);
1491         if (session->s_nr_caps > 0) {
1492                 struct inode *inode;
1493                 struct ceph_cap *cap, *prev = NULL;
1494                 struct ceph_vino vino;
1495                 /*
1496                  * iterate_session_caps() skips inodes that are being
1497                  * deleted, we need to wait until deletions are complete.
1498                  * __wait_on_freeing_inode() is designed for the job,
1499                  * but it is not exported, so use lookup inode function
1500                  * to access it.
1501                  */
1502                 while (!list_empty(&session->s_caps)) {
1503                         cap = list_entry(session->s_caps.next,
1504                                          struct ceph_cap, session_caps);
1505                         if (cap == prev)
1506                                 break;
1507                         prev = cap;
1508                         vino = cap->ci->i_vino;
1509                         spin_unlock(&session->s_cap_lock);
1510
1511                         inode = ceph_find_inode(sb, vino);
1512                          /* avoid calling iput_final() while holding s_mutex */
1513                         ceph_async_iput(inode);
1514
1515                         spin_lock(&session->s_cap_lock);
1516                 }
1517         }
1518
1519         // drop cap expires and unlock s_cap_lock
1520         detach_cap_releases(session, &dispose);
1521
1522         BUG_ON(session->s_nr_caps > 0);
1523         BUG_ON(!list_empty(&session->s_cap_flushing));
1524         spin_unlock(&session->s_cap_lock);
1525         dispose_cap_releases(session->s_mdsc, &dispose);
1526 }
1527
1528 enum {
1529         RECONNECT,
1530         RENEWCAPS,
1531         FORCE_RO,
1532 };
1533
1534 /*
1535  * wake up any threads waiting on this session's caps.  if the cap is
1536  * old (didn't get renewed on the client reconnect), remove it now.
1537  *
1538  * caller must hold s_mutex.
1539  */
1540 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1541                               void *arg)
1542 {
1543         struct ceph_inode_info *ci = ceph_inode(inode);
1544         unsigned long ev = (unsigned long)arg;
1545
1546         if (ev == RECONNECT) {
1547                 spin_lock(&ci->i_ceph_lock);
1548                 ci->i_wanted_max_size = 0;
1549                 ci->i_requested_max_size = 0;
1550                 spin_unlock(&ci->i_ceph_lock);
1551         } else if (ev == RENEWCAPS) {
1552                 if (cap->cap_gen < cap->session->s_cap_gen) {
1553                         /* mds did not re-issue stale cap */
1554                         spin_lock(&ci->i_ceph_lock);
1555                         cap->issued = cap->implemented = CEPH_CAP_PIN;
1556                         /* make sure mds knows what we want */
1557                         if (__ceph_caps_file_wanted(ci) & ~cap->mds_wanted)
1558                                 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1559                         spin_unlock(&ci->i_ceph_lock);
1560                 }
1561         } else if (ev == FORCE_RO) {
1562         }
1563         wake_up_all(&ci->i_cap_wq);
1564         return 0;
1565 }
1566
1567 static void wake_up_session_caps(struct ceph_mds_session *session, int ev)
1568 {
1569         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1570         ceph_iterate_session_caps(session, wake_up_session_cb,
1571                                   (void *)(unsigned long)ev);
1572 }
1573
1574 /*
1575  * Send periodic message to MDS renewing all currently held caps.  The
1576  * ack will reset the expiration for all caps from this session.
1577  *
1578  * caller holds s_mutex
1579  */
1580 static int send_renew_caps(struct ceph_mds_client *mdsc,
1581                            struct ceph_mds_session *session)
1582 {
1583         struct ceph_msg *msg;
1584         int state;
1585
1586         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1587             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1588                 pr_info("mds%d caps stale\n", session->s_mds);
1589         session->s_renew_requested = jiffies;
1590
1591         /* do not try to renew caps until a recovering mds has reconnected
1592          * with its clients. */
1593         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1594         if (state < CEPH_MDS_STATE_RECONNECT) {
1595                 dout("send_renew_caps ignoring mds%d (%s)\n",
1596                      session->s_mds, ceph_mds_state_name(state));
1597                 return 0;
1598         }
1599
1600         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1601                 ceph_mds_state_name(state));
1602         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1603                                  ++session->s_renew_seq);
1604         if (!msg)
1605                 return -ENOMEM;
1606         ceph_con_send(&session->s_con, msg);
1607         return 0;
1608 }
1609
1610 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1611                              struct ceph_mds_session *session, u64 seq)
1612 {
1613         struct ceph_msg *msg;
1614
1615         dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1616              session->s_mds, ceph_session_state_name(session->s_state), seq);
1617         msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1618         if (!msg)
1619                 return -ENOMEM;
1620         ceph_con_send(&session->s_con, msg);
1621         return 0;
1622 }
1623
1624
1625 /*
1626  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1627  *
1628  * Called under session->s_mutex
1629  */
1630 static void renewed_caps(struct ceph_mds_client *mdsc,
1631                          struct ceph_mds_session *session, int is_renew)
1632 {
1633         int was_stale;
1634         int wake = 0;
1635
1636         spin_lock(&session->s_cap_lock);
1637         was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1638
1639         session->s_cap_ttl = session->s_renew_requested +
1640                 mdsc->mdsmap->m_session_timeout*HZ;
1641
1642         if (was_stale) {
1643                 if (time_before(jiffies, session->s_cap_ttl)) {
1644                         pr_info("mds%d caps renewed\n", session->s_mds);
1645                         wake = 1;
1646                 } else {
1647                         pr_info("mds%d caps still stale\n", session->s_mds);
1648                 }
1649         }
1650         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1651              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1652              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1653         spin_unlock(&session->s_cap_lock);
1654
1655         if (wake)
1656                 wake_up_session_caps(session, RENEWCAPS);
1657 }
1658
1659 /*
1660  * send a session close request
1661  */
1662 static int request_close_session(struct ceph_mds_client *mdsc,
1663                                  struct ceph_mds_session *session)
1664 {
1665         struct ceph_msg *msg;
1666
1667         dout("request_close_session mds%d state %s seq %lld\n",
1668              session->s_mds, ceph_session_state_name(session->s_state),
1669              session->s_seq);
1670         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1671         if (!msg)
1672                 return -ENOMEM;
1673         ceph_con_send(&session->s_con, msg);
1674         return 1;
1675 }
1676
1677 /*
1678  * Called with s_mutex held.
1679  */
1680 static int __close_session(struct ceph_mds_client *mdsc,
1681                          struct ceph_mds_session *session)
1682 {
1683         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1684                 return 0;
1685         session->s_state = CEPH_MDS_SESSION_CLOSING;
1686         return request_close_session(mdsc, session);
1687 }
1688
1689 static bool drop_negative_children(struct dentry *dentry)
1690 {
1691         struct dentry *child;
1692         bool all_negative = true;
1693
1694         if (!d_is_dir(dentry))
1695                 goto out;
1696
1697         spin_lock(&dentry->d_lock);
1698         list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1699                 if (d_really_is_positive(child)) {
1700                         all_negative = false;
1701                         break;
1702                 }
1703         }
1704         spin_unlock(&dentry->d_lock);
1705
1706         if (all_negative)
1707                 shrink_dcache_parent(dentry);
1708 out:
1709         return all_negative;
1710 }
1711
1712 /*
1713  * Trim old(er) caps.
1714  *
1715  * Because we can't cache an inode without one or more caps, we do
1716  * this indirectly: if a cap is unused, we prune its aliases, at which
1717  * point the inode will hopefully get dropped to.
1718  *
1719  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1720  * memory pressure from the MDS, though, so it needn't be perfect.
1721  */
1722 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1723 {
1724         int *remaining = arg;
1725         struct ceph_inode_info *ci = ceph_inode(inode);
1726         int used, wanted, oissued, mine;
1727
1728         if (*remaining <= 0)
1729                 return -1;
1730
1731         spin_lock(&ci->i_ceph_lock);
1732         mine = cap->issued | cap->implemented;
1733         used = __ceph_caps_used(ci);
1734         wanted = __ceph_caps_file_wanted(ci);
1735         oissued = __ceph_caps_issued_other(ci, cap);
1736
1737         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1738              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1739              ceph_cap_string(used), ceph_cap_string(wanted));
1740         if (cap == ci->i_auth_cap) {
1741                 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1742                     !list_empty(&ci->i_cap_snaps))
1743                         goto out;
1744                 if ((used | wanted) & CEPH_CAP_ANY_WR)
1745                         goto out;
1746                 /* Note: it's possible that i_filelock_ref becomes non-zero
1747                  * after dropping auth caps. It doesn't hurt because reply
1748                  * of lock mds request will re-add auth caps. */
1749                 if (atomic_read(&ci->i_filelock_ref) > 0)
1750                         goto out;
1751         }
1752         /* The inode has cached pages, but it's no longer used.
1753          * we can safely drop it */
1754         if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1755             !(oissued & CEPH_CAP_FILE_CACHE)) {
1756           used = 0;
1757           oissued = 0;
1758         }
1759         if ((used | wanted) & ~oissued & mine)
1760                 goto out;   /* we need these caps */
1761
1762         if (oissued) {
1763                 /* we aren't the only cap.. just remove us */
1764                 __ceph_remove_cap(cap, true);
1765                 (*remaining)--;
1766         } else {
1767                 struct dentry *dentry;
1768                 /* try dropping referring dentries */
1769                 spin_unlock(&ci->i_ceph_lock);
1770                 dentry = d_find_any_alias(inode);
1771                 if (dentry && drop_negative_children(dentry)) {
1772                         int count;
1773                         dput(dentry);
1774                         d_prune_aliases(inode);
1775                         count = atomic_read(&inode->i_count);
1776                         if (count == 1)
1777                                 (*remaining)--;
1778                         dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1779                              inode, cap, count);
1780                 } else {
1781                         dput(dentry);
1782                 }
1783                 return 0;
1784         }
1785
1786 out:
1787         spin_unlock(&ci->i_ceph_lock);
1788         return 0;
1789 }
1790
1791 /*
1792  * Trim session cap count down to some max number.
1793  */
1794 int ceph_trim_caps(struct ceph_mds_client *mdsc,
1795                    struct ceph_mds_session *session,
1796                    int max_caps)
1797 {
1798         int trim_caps = session->s_nr_caps - max_caps;
1799
1800         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1801              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1802         if (trim_caps > 0) {
1803                 int remaining = trim_caps;
1804
1805                 ceph_iterate_session_caps(session, trim_caps_cb, &remaining);
1806                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1807                      session->s_mds, session->s_nr_caps, max_caps,
1808                         trim_caps - remaining);
1809         }
1810
1811         ceph_flush_cap_releases(mdsc, session);
1812         return 0;
1813 }
1814
1815 static int check_caps_flush(struct ceph_mds_client *mdsc,
1816                             u64 want_flush_tid)
1817 {
1818         int ret = 1;
1819
1820         spin_lock(&mdsc->cap_dirty_lock);
1821         if (!list_empty(&mdsc->cap_flush_list)) {
1822                 struct ceph_cap_flush *cf =
1823                         list_first_entry(&mdsc->cap_flush_list,
1824                                          struct ceph_cap_flush, g_list);
1825                 if (cf->tid <= want_flush_tid) {
1826                         dout("check_caps_flush still flushing tid "
1827                              "%llu <= %llu\n", cf->tid, want_flush_tid);
1828                         ret = 0;
1829                 }
1830         }
1831         spin_unlock(&mdsc->cap_dirty_lock);
1832         return ret;
1833 }
1834
1835 /*
1836  * flush all dirty inode data to disk.
1837  *
1838  * returns true if we've flushed through want_flush_tid
1839  */
1840 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1841                             u64 want_flush_tid)
1842 {
1843         dout("check_caps_flush want %llu\n", want_flush_tid);
1844
1845         wait_event(mdsc->cap_flushing_wq,
1846                    check_caps_flush(mdsc, want_flush_tid));
1847
1848         dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1849 }
1850
1851 /*
1852  * called under s_mutex
1853  */
1854 static void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1855                                    struct ceph_mds_session *session)
1856 {
1857         struct ceph_msg *msg = NULL;
1858         struct ceph_mds_cap_release *head;
1859         struct ceph_mds_cap_item *item;
1860         struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
1861         struct ceph_cap *cap;
1862         LIST_HEAD(tmp_list);
1863         int num_cap_releases;
1864         __le32  barrier, *cap_barrier;
1865
1866         down_read(&osdc->lock);
1867         barrier = cpu_to_le32(osdc->epoch_barrier);
1868         up_read(&osdc->lock);
1869
1870         spin_lock(&session->s_cap_lock);
1871 again:
1872         list_splice_init(&session->s_cap_releases, &tmp_list);
1873         num_cap_releases = session->s_num_cap_releases;
1874         session->s_num_cap_releases = 0;
1875         spin_unlock(&session->s_cap_lock);
1876
1877         while (!list_empty(&tmp_list)) {
1878                 if (!msg) {
1879                         msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1880                                         PAGE_SIZE, GFP_NOFS, false);
1881                         if (!msg)
1882                                 goto out_err;
1883                         head = msg->front.iov_base;
1884                         head->num = cpu_to_le32(0);
1885                         msg->front.iov_len = sizeof(*head);
1886
1887                         msg->hdr.version = cpu_to_le16(2);
1888                         msg->hdr.compat_version = cpu_to_le16(1);
1889                 }
1890
1891                 cap = list_first_entry(&tmp_list, struct ceph_cap,
1892                                         session_caps);
1893                 list_del(&cap->session_caps);
1894                 num_cap_releases--;
1895
1896                 head = msg->front.iov_base;
1897                 put_unaligned_le32(get_unaligned_le32(&head->num) + 1,
1898                                    &head->num);
1899                 item = msg->front.iov_base + msg->front.iov_len;
1900                 item->ino = cpu_to_le64(cap->cap_ino);
1901                 item->cap_id = cpu_to_le64(cap->cap_id);
1902                 item->migrate_seq = cpu_to_le32(cap->mseq);
1903                 item->seq = cpu_to_le32(cap->issue_seq);
1904                 msg->front.iov_len += sizeof(*item);
1905
1906                 ceph_put_cap(mdsc, cap);
1907
1908                 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1909                         // Append cap_barrier field
1910                         cap_barrier = msg->front.iov_base + msg->front.iov_len;
1911                         *cap_barrier = barrier;
1912                         msg->front.iov_len += sizeof(*cap_barrier);
1913
1914                         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1915                         dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1916                         ceph_con_send(&session->s_con, msg);
1917                         msg = NULL;
1918                 }
1919         }
1920
1921         BUG_ON(num_cap_releases != 0);
1922
1923         spin_lock(&session->s_cap_lock);
1924         if (!list_empty(&session->s_cap_releases))
1925                 goto again;
1926         spin_unlock(&session->s_cap_lock);
1927
1928         if (msg) {
1929                 // Append cap_barrier field
1930                 cap_barrier = msg->front.iov_base + msg->front.iov_len;
1931                 *cap_barrier = barrier;
1932                 msg->front.iov_len += sizeof(*cap_barrier);
1933
1934                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1935                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1936                 ceph_con_send(&session->s_con, msg);
1937         }
1938         return;
1939 out_err:
1940         pr_err("send_cap_releases mds%d, failed to allocate message\n",
1941                 session->s_mds);
1942         spin_lock(&session->s_cap_lock);
1943         list_splice(&tmp_list, &session->s_cap_releases);
1944         session->s_num_cap_releases += num_cap_releases;
1945         spin_unlock(&session->s_cap_lock);
1946 }
1947
1948 static void ceph_cap_release_work(struct work_struct *work)
1949 {
1950         struct ceph_mds_session *session =
1951                 container_of(work, struct ceph_mds_session, s_cap_release_work);
1952
1953         mutex_lock(&session->s_mutex);
1954         if (session->s_state == CEPH_MDS_SESSION_OPEN ||
1955             session->s_state == CEPH_MDS_SESSION_HUNG)
1956                 ceph_send_cap_releases(session->s_mdsc, session);
1957         mutex_unlock(&session->s_mutex);
1958         ceph_put_mds_session(session);
1959 }
1960
1961 void ceph_flush_cap_releases(struct ceph_mds_client *mdsc,
1962                              struct ceph_mds_session *session)
1963 {
1964         if (mdsc->stopping)
1965                 return;
1966
1967         get_session(session);
1968         if (queue_work(mdsc->fsc->cap_wq,
1969                        &session->s_cap_release_work)) {
1970                 dout("cap release work queued\n");
1971         } else {
1972                 ceph_put_mds_session(session);
1973                 dout("failed to queue cap release work\n");
1974         }
1975 }
1976
1977 /*
1978  * caller holds session->s_cap_lock
1979  */
1980 void __ceph_queue_cap_release(struct ceph_mds_session *session,
1981                               struct ceph_cap *cap)
1982 {
1983         list_add_tail(&cap->session_caps, &session->s_cap_releases);
1984         session->s_num_cap_releases++;
1985
1986         if (!(session->s_num_cap_releases % CEPH_CAPS_PER_RELEASE))
1987                 ceph_flush_cap_releases(session->s_mdsc, session);
1988 }
1989
1990 static void ceph_cap_reclaim_work(struct work_struct *work)
1991 {
1992         struct ceph_mds_client *mdsc =
1993                 container_of(work, struct ceph_mds_client, cap_reclaim_work);
1994         int ret = ceph_trim_dentries(mdsc);
1995         if (ret == -EAGAIN)
1996                 ceph_queue_cap_reclaim_work(mdsc);
1997 }
1998
1999 void ceph_queue_cap_reclaim_work(struct ceph_mds_client *mdsc)
2000 {
2001         if (mdsc->stopping)
2002                 return;
2003
2004         if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_reclaim_work)) {
2005                 dout("caps reclaim work queued\n");
2006         } else {
2007                 dout("failed to queue caps release work\n");
2008         }
2009 }
2010
2011 void ceph_reclaim_caps_nr(struct ceph_mds_client *mdsc, int nr)
2012 {
2013         int val;
2014         if (!nr)
2015                 return;
2016         val = atomic_add_return(nr, &mdsc->cap_reclaim_pending);
2017         if (!(val % CEPH_CAPS_PER_RELEASE)) {
2018                 atomic_set(&mdsc->cap_reclaim_pending, 0);
2019                 ceph_queue_cap_reclaim_work(mdsc);
2020         }
2021 }
2022
2023 /*
2024  * requests
2025  */
2026
2027 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
2028                                     struct inode *dir)
2029 {
2030         struct ceph_inode_info *ci = ceph_inode(dir);
2031         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
2032         struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
2033         size_t size = sizeof(struct ceph_mds_reply_dir_entry);
2034         int order, num_entries;
2035
2036         spin_lock(&ci->i_ceph_lock);
2037         num_entries = ci->i_files + ci->i_subdirs;
2038         spin_unlock(&ci->i_ceph_lock);
2039         num_entries = max(num_entries, 1);
2040         num_entries = min(num_entries, opt->max_readdir);
2041
2042         order = get_order(size * num_entries);
2043         while (order >= 0) {
2044                 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
2045                                                              __GFP_NOWARN,
2046                                                              order);
2047                 if (rinfo->dir_entries)
2048                         break;
2049                 order--;
2050         }
2051         if (!rinfo->dir_entries)
2052                 return -ENOMEM;
2053
2054         num_entries = (PAGE_SIZE << order) / size;
2055         num_entries = min(num_entries, opt->max_readdir);
2056
2057         rinfo->dir_buf_size = PAGE_SIZE << order;
2058         req->r_num_caps = num_entries + 1;
2059         req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
2060         req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
2061         return 0;
2062 }
2063
2064 /*
2065  * Create an mds request.
2066  */
2067 struct ceph_mds_request *
2068 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
2069 {
2070         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
2071         struct timespec64 ts;
2072
2073         if (!req)
2074                 return ERR_PTR(-ENOMEM);
2075
2076         mutex_init(&req->r_fill_mutex);
2077         req->r_mdsc = mdsc;
2078         req->r_started = jiffies;
2079         req->r_resend_mds = -1;
2080         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
2081         INIT_LIST_HEAD(&req->r_unsafe_target_item);
2082         req->r_fmode = -1;
2083         kref_init(&req->r_kref);
2084         RB_CLEAR_NODE(&req->r_node);
2085         INIT_LIST_HEAD(&req->r_wait);
2086         init_completion(&req->r_completion);
2087         init_completion(&req->r_safe_completion);
2088         INIT_LIST_HEAD(&req->r_unsafe_item);
2089
2090         ktime_get_coarse_real_ts64(&ts);
2091         req->r_stamp = timespec64_trunc(ts, mdsc->fsc->sb->s_time_gran);
2092
2093         req->r_op = op;
2094         req->r_direct_mode = mode;
2095         return req;
2096 }
2097
2098 /*
2099  * return oldest (lowest) request, tid in request tree, 0 if none.
2100  *
2101  * called under mdsc->mutex.
2102  */
2103 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
2104 {
2105         if (RB_EMPTY_ROOT(&mdsc->request_tree))
2106                 return NULL;
2107         return rb_entry(rb_first(&mdsc->request_tree),
2108                         struct ceph_mds_request, r_node);
2109 }
2110
2111 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
2112 {
2113         return mdsc->oldest_tid;
2114 }
2115
2116 /*
2117  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
2118  * on build_path_from_dentry in fs/cifs/dir.c.
2119  *
2120  * If @stop_on_nosnap, generate path relative to the first non-snapped
2121  * inode.
2122  *
2123  * Encode hidden .snap dirs as a double /, i.e.
2124  *   foo/.snap/bar -> foo//bar
2125  */
2126 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *pbase,
2127                            int stop_on_nosnap)
2128 {
2129         struct dentry *temp;
2130         char *path;
2131         int pos;
2132         unsigned seq;
2133         u64 base;
2134
2135         if (!dentry)
2136                 return ERR_PTR(-EINVAL);
2137
2138         path = __getname();
2139         if (!path)
2140                 return ERR_PTR(-ENOMEM);
2141 retry:
2142         pos = PATH_MAX - 1;
2143         path[pos] = '\0';
2144
2145         seq = read_seqbegin(&rename_lock);
2146         rcu_read_lock();
2147         temp = dentry;
2148         for (;;) {
2149                 struct inode *inode;
2150
2151                 spin_lock(&temp->d_lock);
2152                 inode = d_inode(temp);
2153                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
2154                         dout("build_path path+%d: %p SNAPDIR\n",
2155                              pos, temp);
2156                 } else if (stop_on_nosnap && inode && dentry != temp &&
2157                            ceph_snap(inode) == CEPH_NOSNAP) {
2158                         spin_unlock(&temp->d_lock);
2159                         pos++; /* get rid of any prepended '/' */
2160                         break;
2161                 } else {
2162                         pos -= temp->d_name.len;
2163                         if (pos < 0) {
2164                                 spin_unlock(&temp->d_lock);
2165                                 break;
2166                         }
2167                         memcpy(path + pos, temp->d_name.name, temp->d_name.len);
2168                 }
2169                 spin_unlock(&temp->d_lock);
2170                 temp = READ_ONCE(temp->d_parent);
2171
2172                 /* Are we at the root? */
2173                 if (IS_ROOT(temp))
2174                         break;
2175
2176                 /* Are we out of buffer? */
2177                 if (--pos < 0)
2178                         break;
2179
2180                 path[pos] = '/';
2181         }
2182         base = ceph_ino(d_inode(temp));
2183         rcu_read_unlock();
2184         if (pos < 0 || read_seqretry(&rename_lock, seq)) {
2185                 pr_err("build_path did not end path lookup where "
2186                        "expected, pos is %d\n", pos);
2187                 /* presumably this is only possible if racing with a
2188                    rename of one of the parent directories (we can not
2189                    lock the dentries above us to prevent this, but
2190                    retrying should be harmless) */
2191                 goto retry;
2192         }
2193
2194         *pbase = base;
2195         *plen = PATH_MAX - 1 - pos;
2196         dout("build_path on %p %d built %llx '%.*s'\n",
2197              dentry, d_count(dentry), base, *plen, path + pos);
2198         return path + pos;
2199 }
2200
2201 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
2202                              const char **ppath, int *ppathlen, u64 *pino,
2203                              bool *pfreepath, bool parent_locked)
2204 {
2205         char *path;
2206
2207         rcu_read_lock();
2208         if (!dir)
2209                 dir = d_inode_rcu(dentry->d_parent);
2210         if (dir && parent_locked && ceph_snap(dir) == CEPH_NOSNAP) {
2211                 *pino = ceph_ino(dir);
2212                 rcu_read_unlock();
2213                 *ppath = dentry->d_name.name;
2214                 *ppathlen = dentry->d_name.len;
2215                 return 0;
2216         }
2217         rcu_read_unlock();
2218         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2219         if (IS_ERR(path))
2220                 return PTR_ERR(path);
2221         *ppath = path;
2222         *pfreepath = true;
2223         return 0;
2224 }
2225
2226 static int build_inode_path(struct inode *inode,
2227                             const char **ppath, int *ppathlen, u64 *pino,
2228                             bool *pfreepath)
2229 {
2230         struct dentry *dentry;
2231         char *path;
2232
2233         if (ceph_snap(inode) == CEPH_NOSNAP) {
2234                 *pino = ceph_ino(inode);
2235                 *ppathlen = 0;
2236                 return 0;
2237         }
2238         dentry = d_find_alias(inode);
2239         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2240         dput(dentry);
2241         if (IS_ERR(path))
2242                 return PTR_ERR(path);
2243         *ppath = path;
2244         *pfreepath = true;
2245         return 0;
2246 }
2247
2248 /*
2249  * request arguments may be specified via an inode *, a dentry *, or
2250  * an explicit ino+path.
2251  */
2252 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
2253                                   struct inode *rdiri, const char *rpath,
2254                                   u64 rino, const char **ppath, int *pathlen,
2255                                   u64 *ino, bool *freepath, bool parent_locked)
2256 {
2257         int r = 0;
2258
2259         if (rinode) {
2260                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2261                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2262                      ceph_snap(rinode));
2263         } else if (rdentry) {
2264                 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
2265                                         freepath, parent_locked);
2266                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
2267                      *ppath);
2268         } else if (rpath || rino) {
2269                 *ino = rino;
2270                 *ppath = rpath;
2271                 *pathlen = rpath ? strlen(rpath) : 0;
2272                 dout(" path %.*s\n", *pathlen, rpath);
2273         }
2274
2275         return r;
2276 }
2277
2278 /*
2279  * called under mdsc->mutex
2280  */
2281 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
2282                                                struct ceph_mds_request *req,
2283                                                int mds, bool drop_cap_releases)
2284 {
2285         struct ceph_msg *msg;
2286         struct ceph_mds_request_head *head;
2287         const char *path1 = NULL;
2288         const char *path2 = NULL;
2289         u64 ino1 = 0, ino2 = 0;
2290         int pathlen1 = 0, pathlen2 = 0;
2291         bool freepath1 = false, freepath2 = false;
2292         int len;
2293         u16 releases;
2294         void *p, *end;
2295         int ret;
2296
2297         ret = set_request_path_attr(req->r_inode, req->r_dentry,
2298                               req->r_parent, req->r_path1, req->r_ino1.ino,
2299                               &path1, &pathlen1, &ino1, &freepath1,
2300                               test_bit(CEPH_MDS_R_PARENT_LOCKED,
2301                                         &req->r_req_flags));
2302         if (ret < 0) {
2303                 msg = ERR_PTR(ret);
2304                 goto out;
2305         }
2306
2307         /* If r_old_dentry is set, then assume that its parent is locked */
2308         ret = set_request_path_attr(NULL, req->r_old_dentry,
2309                               req->r_old_dentry_dir,
2310                               req->r_path2, req->r_ino2.ino,
2311                               &path2, &pathlen2, &ino2, &freepath2, true);
2312         if (ret < 0) {
2313                 msg = ERR_PTR(ret);
2314                 goto out_free1;
2315         }
2316
2317         len = sizeof(*head) +
2318                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2319                 sizeof(struct ceph_timespec);
2320
2321         /* calculate (max) length for cap releases */
2322         len += sizeof(struct ceph_mds_request_release) *
2323                 (!!req->r_inode_drop + !!req->r_dentry_drop +
2324                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2325         if (req->r_dentry_drop)
2326                 len += pathlen1;
2327         if (req->r_old_dentry_drop)
2328                 len += pathlen2;
2329
2330         msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
2331         if (!msg) {
2332                 msg = ERR_PTR(-ENOMEM);
2333                 goto out_free2;
2334         }
2335
2336         msg->hdr.version = cpu_to_le16(2);
2337         msg->hdr.tid = cpu_to_le64(req->r_tid);
2338
2339         head = msg->front.iov_base;
2340         p = msg->front.iov_base + sizeof(*head);
2341         end = msg->front.iov_base + msg->front.iov_len;
2342
2343         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2344         head->op = cpu_to_le32(req->r_op);
2345         head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
2346         head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
2347         head->args = req->r_args;
2348
2349         ceph_encode_filepath(&p, end, ino1, path1);
2350         ceph_encode_filepath(&p, end, ino2, path2);
2351
2352         /* make note of release offset, in case we need to replay */
2353         req->r_request_release_offset = p - msg->front.iov_base;
2354
2355         /* cap releases */
2356         releases = 0;
2357         if (req->r_inode_drop)
2358                 releases += ceph_encode_inode_release(&p,
2359                       req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2360                       mds, req->r_inode_drop, req->r_inode_unless, 0);
2361         if (req->r_dentry_drop)
2362                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
2363                                 req->r_parent, mds, req->r_dentry_drop,
2364                                 req->r_dentry_unless);
2365         if (req->r_old_dentry_drop)
2366                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2367                                 req->r_old_dentry_dir, mds,
2368                                 req->r_old_dentry_drop,
2369                                 req->r_old_dentry_unless);
2370         if (req->r_old_inode_drop)
2371                 releases += ceph_encode_inode_release(&p,
2372                       d_inode(req->r_old_dentry),
2373                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2374
2375         if (drop_cap_releases) {
2376                 releases = 0;
2377                 p = msg->front.iov_base + req->r_request_release_offset;
2378         }
2379
2380         head->num_releases = cpu_to_le16(releases);
2381
2382         /* time stamp */
2383         {
2384                 struct ceph_timespec ts;
2385                 ceph_encode_timespec64(&ts, &req->r_stamp);
2386                 ceph_encode_copy(&p, &ts, sizeof(ts));
2387         }
2388
2389         BUG_ON(p > end);
2390         msg->front.iov_len = p - msg->front.iov_base;
2391         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2392
2393         if (req->r_pagelist) {
2394                 struct ceph_pagelist *pagelist = req->r_pagelist;
2395                 ceph_msg_data_add_pagelist(msg, pagelist);
2396                 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2397         } else {
2398                 msg->hdr.data_len = 0;
2399         }
2400
2401         msg->hdr.data_off = cpu_to_le16(0);
2402
2403 out_free2:
2404         if (freepath2)
2405                 ceph_mdsc_free_path((char *)path2, pathlen2);
2406 out_free1:
2407         if (freepath1)
2408                 ceph_mdsc_free_path((char *)path1, pathlen1);
2409 out:
2410         return msg;
2411 }
2412
2413 /*
2414  * called under mdsc->mutex if error, under no mutex if
2415  * success.
2416  */
2417 static void complete_request(struct ceph_mds_client *mdsc,
2418                              struct ceph_mds_request *req)
2419 {
2420         if (req->r_callback)
2421                 req->r_callback(mdsc, req);
2422         complete_all(&req->r_completion);
2423 }
2424
2425 /*
2426  * called under mdsc->mutex
2427  */
2428 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2429                                   struct ceph_mds_request *req,
2430                                   int mds, bool drop_cap_releases)
2431 {
2432         struct ceph_mds_request_head *rhead;
2433         struct ceph_msg *msg;
2434         int flags = 0;
2435
2436         req->r_attempts++;
2437         if (req->r_inode) {
2438                 struct ceph_cap *cap =
2439                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2440
2441                 if (cap)
2442                         req->r_sent_on_mseq = cap->mseq;
2443                 else
2444                         req->r_sent_on_mseq = -1;
2445         }
2446         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2447              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2448
2449         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2450                 void *p;
2451                 /*
2452                  * Replay.  Do not regenerate message (and rebuild
2453                  * paths, etc.); just use the original message.
2454                  * Rebuilding paths will break for renames because
2455                  * d_move mangles the src name.
2456                  */
2457                 msg = req->r_request;
2458                 rhead = msg->front.iov_base;
2459
2460                 flags = le32_to_cpu(rhead->flags);
2461                 flags |= CEPH_MDS_FLAG_REPLAY;
2462                 rhead->flags = cpu_to_le32(flags);
2463
2464                 if (req->r_target_inode)
2465                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2466
2467                 rhead->num_retry = req->r_attempts - 1;
2468
2469                 /* remove cap/dentry releases from message */
2470                 rhead->num_releases = 0;
2471
2472                 /* time stamp */
2473                 p = msg->front.iov_base + req->r_request_release_offset;
2474                 {
2475                         struct ceph_timespec ts;
2476                         ceph_encode_timespec64(&ts, &req->r_stamp);
2477                         ceph_encode_copy(&p, &ts, sizeof(ts));
2478                 }
2479
2480                 msg->front.iov_len = p - msg->front.iov_base;
2481                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2482                 return 0;
2483         }
2484
2485         if (req->r_request) {
2486                 ceph_msg_put(req->r_request);
2487                 req->r_request = NULL;
2488         }
2489         msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2490         if (IS_ERR(msg)) {
2491                 req->r_err = PTR_ERR(msg);
2492                 return PTR_ERR(msg);
2493         }
2494         req->r_request = msg;
2495
2496         rhead = msg->front.iov_base;
2497         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2498         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2499                 flags |= CEPH_MDS_FLAG_REPLAY;
2500         if (req->r_parent)
2501                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2502         rhead->flags = cpu_to_le32(flags);
2503         rhead->num_fwd = req->r_num_fwd;
2504         rhead->num_retry = req->r_attempts - 1;
2505         rhead->ino = 0;
2506
2507         dout(" r_parent = %p\n", req->r_parent);
2508         return 0;
2509 }
2510
2511 /*
2512  * send request, or put it on the appropriate wait list.
2513  */
2514 static void __do_request(struct ceph_mds_client *mdsc,
2515                         struct ceph_mds_request *req)
2516 {
2517         struct ceph_mds_session *session = NULL;
2518         int mds = -1;
2519         int err = 0;
2520
2521         if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2522                 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2523                         __unregister_request(mdsc, req);
2524                 return;
2525         }
2526
2527         if (req->r_timeout &&
2528             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2529                 dout("do_request timed out\n");
2530                 err = -EIO;
2531                 goto finish;
2532         }
2533         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2534                 dout("do_request forced umount\n");
2535                 err = -EIO;
2536                 goto finish;
2537         }
2538         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2539                 if (mdsc->mdsmap_err) {
2540                         err = mdsc->mdsmap_err;
2541                         dout("do_request mdsmap err %d\n", err);
2542                         goto finish;
2543                 }
2544                 if (mdsc->mdsmap->m_epoch == 0) {
2545                         dout("do_request no mdsmap, waiting for map\n");
2546                         list_add(&req->r_wait, &mdsc->waiting_for_map);
2547                         return;
2548                 }
2549                 if (!(mdsc->fsc->mount_options->flags &
2550                       CEPH_MOUNT_OPT_MOUNTWAIT) &&
2551                     !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2552                         err = -ENOENT;
2553                         pr_info("probably no mds server is up\n");
2554                         goto finish;
2555                 }
2556         }
2557
2558         put_request_session(req);
2559
2560         mds = __choose_mds(mdsc, req);
2561         if (mds < 0 ||
2562             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2563                 dout("do_request no mds or not active, waiting for map\n");
2564                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2565                 return;
2566         }
2567
2568         /* get, open session */
2569         session = __ceph_lookup_mds_session(mdsc, mds);
2570         if (!session) {
2571                 session = register_session(mdsc, mds);
2572                 if (IS_ERR(session)) {
2573                         err = PTR_ERR(session);
2574                         goto finish;
2575                 }
2576         }
2577         req->r_session = get_session(session);
2578
2579         dout("do_request mds%d session %p state %s\n", mds, session,
2580              ceph_session_state_name(session->s_state));
2581         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2582             session->s_state != CEPH_MDS_SESSION_HUNG) {
2583                 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2584                         err = -EACCES;
2585                         goto out_session;
2586                 }
2587                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2588                     session->s_state == CEPH_MDS_SESSION_CLOSING)
2589                         __open_session(mdsc, session);
2590                 list_add(&req->r_wait, &session->s_waiting);
2591                 goto out_session;
2592         }
2593
2594         /* send request */
2595         req->r_resend_mds = -1;   /* forget any previous mds hint */
2596
2597         if (req->r_request_started == 0)   /* note request start time */
2598                 req->r_request_started = jiffies;
2599
2600         err = __prepare_send_request(mdsc, req, mds, false);
2601         if (!err) {
2602                 ceph_msg_get(req->r_request);
2603                 ceph_con_send(&session->s_con, req->r_request);
2604         }
2605
2606 out_session:
2607         ceph_put_mds_session(session);
2608 finish:
2609         if (err) {
2610                 dout("__do_request early error %d\n", err);
2611                 req->r_err = err;
2612                 complete_request(mdsc, req);
2613                 __unregister_request(mdsc, req);
2614         }
2615         return;
2616 }
2617
2618 /*
2619  * called under mdsc->mutex
2620  */
2621 static void __wake_requests(struct ceph_mds_client *mdsc,
2622                             struct list_head *head)
2623 {
2624         struct ceph_mds_request *req;
2625         LIST_HEAD(tmp_list);
2626
2627         list_splice_init(head, &tmp_list);
2628
2629         while (!list_empty(&tmp_list)) {
2630                 req = list_entry(tmp_list.next,
2631                                  struct ceph_mds_request, r_wait);
2632                 list_del_init(&req->r_wait);
2633                 dout(" wake request %p tid %llu\n", req, req->r_tid);
2634                 __do_request(mdsc, req);
2635         }
2636 }
2637
2638 /*
2639  * Wake up threads with requests pending for @mds, so that they can
2640  * resubmit their requests to a possibly different mds.
2641  */
2642 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2643 {
2644         struct ceph_mds_request *req;
2645         struct rb_node *p = rb_first(&mdsc->request_tree);
2646
2647         dout("kick_requests mds%d\n", mds);
2648         while (p) {
2649                 req = rb_entry(p, struct ceph_mds_request, r_node);
2650                 p = rb_next(p);
2651                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2652                         continue;
2653                 if (req->r_attempts > 0)
2654                         continue; /* only new requests */
2655                 if (req->r_session &&
2656                     req->r_session->s_mds == mds) {
2657                         dout(" kicking tid %llu\n", req->r_tid);
2658                         list_del_init(&req->r_wait);
2659                         __do_request(mdsc, req);
2660                 }
2661         }
2662 }
2663
2664 int ceph_mdsc_submit_request(struct ceph_mds_client *mdsc, struct inode *dir,
2665                               struct ceph_mds_request *req)
2666 {
2667         int err;
2668
2669         /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2670         if (req->r_inode)
2671                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2672         if (req->r_parent)
2673                 ceph_get_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
2674         if (req->r_old_dentry_dir)
2675                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2676                                   CEPH_CAP_PIN);
2677
2678         dout("submit_request on %p for inode %p\n", req, dir);
2679         mutex_lock(&mdsc->mutex);
2680         __register_request(mdsc, req, dir);
2681         __do_request(mdsc, req);
2682         err = req->r_err;
2683         mutex_unlock(&mdsc->mutex);
2684         return err;
2685 }
2686
2687 static int ceph_mdsc_wait_request(struct ceph_mds_client *mdsc,
2688                                   struct ceph_mds_request *req)
2689 {
2690         int err;
2691
2692         /* wait */
2693         dout("do_request waiting\n");
2694         if (!req->r_timeout && req->r_wait_for_completion) {
2695                 err = req->r_wait_for_completion(mdsc, req);
2696         } else {
2697                 long timeleft = wait_for_completion_killable_timeout(
2698                                         &req->r_completion,
2699                                         ceph_timeout_jiffies(req->r_timeout));
2700                 if (timeleft > 0)
2701                         err = 0;
2702                 else if (!timeleft)
2703                         err = -EIO;  /* timed out */
2704                 else
2705                         err = timeleft;  /* killed */
2706         }
2707         dout("do_request waited, got %d\n", err);
2708         mutex_lock(&mdsc->mutex);
2709
2710         /* only abort if we didn't race with a real reply */
2711         if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2712                 err = le32_to_cpu(req->r_reply_info.head->result);
2713         } else if (err < 0) {
2714                 dout("aborted request %lld with %d\n", req->r_tid, err);
2715
2716                 /*
2717                  * ensure we aren't running concurrently with
2718                  * ceph_fill_trace or ceph_readdir_prepopulate, which
2719                  * rely on locks (dir mutex) held by our caller.
2720                  */
2721                 mutex_lock(&req->r_fill_mutex);
2722                 req->r_err = err;
2723                 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2724                 mutex_unlock(&req->r_fill_mutex);
2725
2726                 if (req->r_parent &&
2727                     (req->r_op & CEPH_MDS_OP_WRITE))
2728                         ceph_invalidate_dir_request(req);
2729         } else {
2730                 err = req->r_err;
2731         }
2732
2733         mutex_unlock(&mdsc->mutex);
2734         return err;
2735 }
2736
2737 /*
2738  * Synchrously perform an mds request.  Take care of all of the
2739  * session setup, forwarding, retry details.
2740  */
2741 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2742                          struct inode *dir,
2743                          struct ceph_mds_request *req)
2744 {
2745         int err;
2746
2747         dout("do_request on %p\n", req);
2748
2749         /* issue */
2750         err = ceph_mdsc_submit_request(mdsc, dir, req);
2751         if (!err)
2752                 err = ceph_mdsc_wait_request(mdsc, req);
2753         dout("do_request %p done, result %d\n", req, err);
2754         return err;
2755 }
2756
2757 /*
2758  * Invalidate dir's completeness, dentry lease state on an aborted MDS
2759  * namespace request.
2760  */
2761 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2762 {
2763         struct inode *dir = req->r_parent;
2764         struct inode *old_dir = req->r_old_dentry_dir;
2765
2766         dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
2767
2768         ceph_dir_clear_complete(dir);
2769         if (old_dir)
2770                 ceph_dir_clear_complete(old_dir);
2771         if (req->r_dentry)
2772                 ceph_invalidate_dentry_lease(req->r_dentry);
2773         if (req->r_old_dentry)
2774                 ceph_invalidate_dentry_lease(req->r_old_dentry);
2775 }
2776
2777 /*
2778  * Handle mds reply.
2779  *
2780  * We take the session mutex and parse and process the reply immediately.
2781  * This preserves the logical ordering of replies, capabilities, etc., sent
2782  * by the MDS as they are applied to our local cache.
2783  */
2784 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2785 {
2786         struct ceph_mds_client *mdsc = session->s_mdsc;
2787         struct ceph_mds_request *req;
2788         struct ceph_mds_reply_head *head = msg->front.iov_base;
2789         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2790         struct ceph_snap_realm *realm;
2791         u64 tid;
2792         int err, result;
2793         int mds = session->s_mds;
2794
2795         if (msg->front.iov_len < sizeof(*head)) {
2796                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2797                 ceph_msg_dump(msg);
2798                 return;
2799         }
2800
2801         /* get request, session */
2802         tid = le64_to_cpu(msg->hdr.tid);
2803         mutex_lock(&mdsc->mutex);
2804         req = lookup_get_request(mdsc, tid);
2805         if (!req) {
2806                 dout("handle_reply on unknown tid %llu\n", tid);
2807                 mutex_unlock(&mdsc->mutex);
2808                 return;
2809         }
2810         dout("handle_reply %p\n", req);
2811
2812         /* correct session? */
2813         if (req->r_session != session) {
2814                 pr_err("mdsc_handle_reply got %llu on session mds%d"
2815                        " not mds%d\n", tid, session->s_mds,
2816                        req->r_session ? req->r_session->s_mds : -1);
2817                 mutex_unlock(&mdsc->mutex);
2818                 goto out;
2819         }
2820
2821         /* dup? */
2822         if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
2823             (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
2824                 pr_warn("got a dup %s reply on %llu from mds%d\n",
2825                            head->safe ? "safe" : "unsafe", tid, mds);
2826                 mutex_unlock(&mdsc->mutex);
2827                 goto out;
2828         }
2829         if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
2830                 pr_warn("got unsafe after safe on %llu from mds%d\n",
2831                            tid, mds);
2832                 mutex_unlock(&mdsc->mutex);
2833                 goto out;
2834         }
2835
2836         result = le32_to_cpu(head->result);
2837
2838         /*
2839          * Handle an ESTALE
2840          * if we're not talking to the authority, send to them
2841          * if the authority has changed while we weren't looking,
2842          * send to new authority
2843          * Otherwise we just have to return an ESTALE
2844          */
2845         if (result == -ESTALE) {
2846                 dout("got ESTALE on request %llu\n", req->r_tid);
2847                 req->r_resend_mds = -1;
2848                 if (req->r_direct_mode != USE_AUTH_MDS) {
2849                         dout("not using auth, setting for that now\n");
2850                         req->r_direct_mode = USE_AUTH_MDS;
2851                         __do_request(mdsc, req);
2852                         mutex_unlock(&mdsc->mutex);
2853                         goto out;
2854                 } else  {
2855                         int mds = __choose_mds(mdsc, req);
2856                         if (mds >= 0 && mds != req->r_session->s_mds) {
2857                                 dout("but auth changed, so resending\n");
2858                                 __do_request(mdsc, req);
2859                                 mutex_unlock(&mdsc->mutex);
2860                                 goto out;
2861                         }
2862                 }
2863                 dout("have to return ESTALE on request %llu\n", req->r_tid);
2864         }
2865
2866
2867         if (head->safe) {
2868                 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
2869                 __unregister_request(mdsc, req);
2870
2871                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2872                         /*
2873                          * We already handled the unsafe response, now do the
2874                          * cleanup.  No need to examine the response; the MDS
2875                          * doesn't include any result info in the safe
2876                          * response.  And even if it did, there is nothing
2877                          * useful we could do with a revised return value.
2878                          */
2879                         dout("got safe reply %llu, mds%d\n", tid, mds);
2880
2881                         /* last unsafe request during umount? */
2882                         if (mdsc->stopping && !__get_oldest_req(mdsc))
2883                                 complete_all(&mdsc->safe_umount_waiters);
2884                         mutex_unlock(&mdsc->mutex);
2885                         goto out;
2886                 }
2887         } else {
2888                 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
2889                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2890                 if (req->r_unsafe_dir) {
2891                         struct ceph_inode_info *ci =
2892                                         ceph_inode(req->r_unsafe_dir);
2893                         spin_lock(&ci->i_unsafe_lock);
2894                         list_add_tail(&req->r_unsafe_dir_item,
2895                                       &ci->i_unsafe_dirops);
2896                         spin_unlock(&ci->i_unsafe_lock);
2897                 }
2898         }
2899
2900         dout("handle_reply tid %lld result %d\n", tid, result);
2901         rinfo = &req->r_reply_info;
2902         if (test_bit(CEPHFS_FEATURE_REPLY_ENCODING, &session->s_features))
2903                 err = parse_reply_info(msg, rinfo, (u64)-1);
2904         else
2905                 err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2906         mutex_unlock(&mdsc->mutex);
2907
2908         mutex_lock(&session->s_mutex);
2909         if (err < 0) {
2910                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2911                 ceph_msg_dump(msg);
2912                 goto out_err;
2913         }
2914
2915         /* snap trace */
2916         realm = NULL;
2917         if (rinfo->snapblob_len) {
2918                 down_write(&mdsc->snap_rwsem);
2919                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2920                                 rinfo->snapblob + rinfo->snapblob_len,
2921                                 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2922                                 &realm);
2923                 downgrade_write(&mdsc->snap_rwsem);
2924         } else {
2925                 down_read(&mdsc->snap_rwsem);
2926         }
2927
2928         /* insert trace into our cache */
2929         mutex_lock(&req->r_fill_mutex);
2930         current->journal_info = req;
2931         err = ceph_fill_trace(mdsc->fsc->sb, req);
2932         if (err == 0) {
2933                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2934                                     req->r_op == CEPH_MDS_OP_LSSNAP))
2935                         ceph_readdir_prepopulate(req, req->r_session);
2936         }
2937         current->journal_info = NULL;
2938         mutex_unlock(&req->r_fill_mutex);
2939
2940         up_read(&mdsc->snap_rwsem);
2941         if (realm)
2942                 ceph_put_snap_realm(mdsc, realm);
2943
2944         if (err == 0) {
2945                 if (req->r_target_inode &&
2946                     test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2947                         struct ceph_inode_info *ci =
2948                                 ceph_inode(req->r_target_inode);
2949                         spin_lock(&ci->i_unsafe_lock);
2950                         list_add_tail(&req->r_unsafe_target_item,
2951                                       &ci->i_unsafe_iops);
2952                         spin_unlock(&ci->i_unsafe_lock);
2953                 }
2954
2955                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2956         }
2957 out_err:
2958         mutex_lock(&mdsc->mutex);
2959         if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2960                 if (err) {
2961                         req->r_err = err;
2962                 } else {
2963                         req->r_reply =  ceph_msg_get(msg);
2964                         set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
2965                 }
2966         } else {
2967                 dout("reply arrived after request %lld was aborted\n", tid);
2968         }
2969         mutex_unlock(&mdsc->mutex);
2970
2971         mutex_unlock(&session->s_mutex);
2972
2973         /* kick calling process */
2974         complete_request(mdsc, req);
2975 out:
2976         ceph_mdsc_put_request(req);
2977         return;
2978 }
2979
2980
2981
2982 /*
2983  * handle mds notification that our request has been forwarded.
2984  */
2985 static void handle_forward(struct ceph_mds_client *mdsc,
2986                            struct ceph_mds_session *session,
2987                            struct ceph_msg *msg)
2988 {
2989         struct ceph_mds_request *req;
2990         u64 tid = le64_to_cpu(msg->hdr.tid);
2991         u32 next_mds;
2992         u32 fwd_seq;
2993         int err = -EINVAL;
2994         void *p = msg->front.iov_base;
2995         void *end = p + msg->front.iov_len;
2996
2997         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2998         next_mds = ceph_decode_32(&p);
2999         fwd_seq = ceph_decode_32(&p);
3000
3001         mutex_lock(&mdsc->mutex);
3002         req = lookup_get_request(mdsc, tid);
3003         if (!req) {
3004                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
3005                 goto out;  /* dup reply? */
3006         }
3007
3008         if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3009                 dout("forward tid %llu aborted, unregistering\n", tid);
3010                 __unregister_request(mdsc, req);
3011         } else if (fwd_seq <= req->r_num_fwd) {
3012                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
3013                      tid, next_mds, req->r_num_fwd, fwd_seq);
3014         } else {
3015                 /* resend. forward race not possible; mds would drop */
3016                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
3017                 BUG_ON(req->r_err);
3018                 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
3019                 req->r_attempts = 0;
3020                 req->r_num_fwd = fwd_seq;
3021                 req->r_resend_mds = next_mds;
3022                 put_request_session(req);
3023                 __do_request(mdsc, req);
3024         }
3025         ceph_mdsc_put_request(req);
3026 out:
3027         mutex_unlock(&mdsc->mutex);
3028         return;
3029
3030 bad:
3031         pr_err("mdsc_handle_forward decode error err=%d\n", err);
3032 }
3033
3034 static int __decode_session_metadata(void **p, void *end,
3035                                      bool *blacklisted)
3036 {
3037         /* map<string,string> */
3038         u32 n;
3039         bool err_str;
3040         ceph_decode_32_safe(p, end, n, bad);
3041         while (n-- > 0) {
3042                 u32 len;
3043                 ceph_decode_32_safe(p, end, len, bad);
3044                 ceph_decode_need(p, end, len, bad);
3045                 err_str = !strncmp(*p, "error_string", len);
3046                 *p += len;
3047                 ceph_decode_32_safe(p, end, len, bad);
3048                 ceph_decode_need(p, end, len, bad);
3049                 if (err_str && strnstr(*p, "blacklisted", len))
3050                         *blacklisted = true;
3051                 *p += len;
3052         }
3053         return 0;
3054 bad:
3055         return -1;
3056 }
3057
3058 /*
3059  * handle a mds session control message
3060  */
3061 static void handle_session(struct ceph_mds_session *session,
3062                            struct ceph_msg *msg)
3063 {
3064         struct ceph_mds_client *mdsc = session->s_mdsc;
3065         int mds = session->s_mds;
3066         int msg_version = le16_to_cpu(msg->hdr.version);
3067         void *p = msg->front.iov_base;
3068         void *end = p + msg->front.iov_len;
3069         struct ceph_mds_session_head *h;
3070         u32 op;
3071         u64 seq;
3072         unsigned long features = 0;
3073         int wake = 0;
3074         bool blacklisted = false;
3075
3076         /* decode */
3077         ceph_decode_need(&p, end, sizeof(*h), bad);
3078         h = p;
3079         p += sizeof(*h);
3080
3081         op = le32_to_cpu(h->op);
3082         seq = le64_to_cpu(h->seq);
3083
3084         if (msg_version >= 3) {
3085                 u32 len;
3086                 /* version >= 2, metadata */
3087                 if (__decode_session_metadata(&p, end, &blacklisted) < 0)
3088                         goto bad;
3089                 /* version >= 3, feature bits */
3090                 ceph_decode_32_safe(&p, end, len, bad);
3091                 ceph_decode_need(&p, end, len, bad);
3092                 memcpy(&features, p, min_t(size_t, len, sizeof(features)));
3093                 p += len;
3094         }
3095
3096         mutex_lock(&mdsc->mutex);
3097         if (op == CEPH_SESSION_CLOSE) {
3098                 get_session(session);
3099                 __unregister_session(mdsc, session);
3100         }
3101         /* FIXME: this ttl calculation is generous */
3102         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
3103         mutex_unlock(&mdsc->mutex);
3104
3105         mutex_lock(&session->s_mutex);
3106
3107         dout("handle_session mds%d %s %p state %s seq %llu\n",
3108              mds, ceph_session_op_name(op), session,
3109              ceph_session_state_name(session->s_state), seq);
3110
3111         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
3112                 session->s_state = CEPH_MDS_SESSION_OPEN;
3113                 pr_info("mds%d came back\n", session->s_mds);
3114         }
3115
3116         switch (op) {
3117         case CEPH_SESSION_OPEN:
3118                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3119                         pr_info("mds%d reconnect success\n", session->s_mds);
3120                 session->s_state = CEPH_MDS_SESSION_OPEN;
3121                 session->s_features = features;
3122                 renewed_caps(mdsc, session, 0);
3123                 wake = 1;
3124                 if (mdsc->stopping)
3125                         __close_session(mdsc, session);
3126                 break;
3127
3128         case CEPH_SESSION_RENEWCAPS:
3129                 if (session->s_renew_seq == seq)
3130                         renewed_caps(mdsc, session, 1);
3131                 break;
3132
3133         case CEPH_SESSION_CLOSE:
3134                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3135                         pr_info("mds%d reconnect denied\n", session->s_mds);
3136                 cleanup_session_requests(mdsc, session);
3137                 remove_session_caps(session);
3138                 wake = 2; /* for good measure */
3139                 wake_up_all(&mdsc->session_close_wq);
3140                 break;
3141
3142         case CEPH_SESSION_STALE:
3143                 pr_info("mds%d caps went stale, renewing\n",
3144                         session->s_mds);
3145                 spin_lock(&session->s_gen_ttl_lock);
3146                 session->s_cap_gen++;
3147                 session->s_cap_ttl = jiffies - 1;
3148                 spin_unlock(&session->s_gen_ttl_lock);
3149                 send_renew_caps(mdsc, session);
3150                 break;
3151
3152         case CEPH_SESSION_RECALL_STATE:
3153                 ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
3154                 break;
3155
3156         case CEPH_SESSION_FLUSHMSG:
3157                 send_flushmsg_ack(mdsc, session, seq);
3158                 break;
3159
3160         case CEPH_SESSION_FORCE_RO:
3161                 dout("force_session_readonly %p\n", session);
3162                 spin_lock(&session->s_cap_lock);
3163                 session->s_readonly = true;
3164                 spin_unlock(&session->s_cap_lock);
3165                 wake_up_session_caps(session, FORCE_RO);
3166                 break;
3167
3168         case CEPH_SESSION_REJECT:
3169                 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
3170                 pr_info("mds%d rejected session\n", session->s_mds);
3171                 session->s_state = CEPH_MDS_SESSION_REJECTED;
3172                 cleanup_session_requests(mdsc, session);
3173                 remove_session_caps(session);
3174                 if (blacklisted)
3175                         mdsc->fsc->blacklisted = true;
3176                 wake = 2; /* for good measure */
3177                 break;
3178
3179         default:
3180                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
3181                 WARN_ON(1);
3182         }
3183
3184         mutex_unlock(&session->s_mutex);
3185         if (wake) {
3186                 mutex_lock(&mdsc->mutex);
3187                 __wake_requests(mdsc, &session->s_waiting);
3188                 if (wake == 2)
3189                         kick_requests(mdsc, mds);
3190                 mutex_unlock(&mdsc->mutex);
3191         }
3192         if (op == CEPH_SESSION_CLOSE)
3193                 ceph_put_mds_session(session);
3194         return;
3195
3196 bad:
3197         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
3198                (int)msg->front.iov_len);
3199         ceph_msg_dump(msg);
3200         return;
3201 }
3202
3203
3204 /*
3205  * called under session->mutex.
3206  */
3207 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
3208                                    struct ceph_mds_session *session)
3209 {
3210         struct ceph_mds_request *req, *nreq;
3211         struct rb_node *p;
3212         int err;
3213
3214         dout("replay_unsafe_requests mds%d\n", session->s_mds);
3215
3216         mutex_lock(&mdsc->mutex);
3217         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
3218                 err = __prepare_send_request(mdsc, req, session->s_mds, true);
3219                 if (!err) {
3220                         ceph_msg_get(req->r_request);
3221                         ceph_con_send(&session->s_con, req->r_request);
3222                 }
3223         }
3224
3225         /*
3226          * also re-send old requests when MDS enters reconnect stage. So that MDS
3227          * can process completed request in clientreplay stage.
3228          */
3229         p = rb_first(&mdsc->request_tree);
3230         while (p) {
3231                 req = rb_entry(p, struct ceph_mds_request, r_node);
3232                 p = rb_next(p);
3233                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3234                         continue;
3235                 if (req->r_attempts == 0)
3236                         continue; /* only old requests */
3237                 if (req->r_session &&
3238                     req->r_session->s_mds == session->s_mds) {
3239                         err = __prepare_send_request(mdsc, req,
3240                                                      session->s_mds, true);
3241                         if (!err) {
3242                                 ceph_msg_get(req->r_request);
3243                                 ceph_con_send(&session->s_con, req->r_request);
3244                         }
3245                 }
3246         }
3247         mutex_unlock(&mdsc->mutex);
3248 }
3249
3250 static int send_reconnect_partial(struct ceph_reconnect_state *recon_state)
3251 {
3252         struct ceph_msg *reply;
3253         struct ceph_pagelist *_pagelist;
3254         struct page *page;
3255         __le32 *addr;
3256         int err = -ENOMEM;
3257
3258         if (!recon_state->allow_multi)
3259                 return -ENOSPC;
3260
3261         /* can't handle message that contains both caps and realm */
3262         BUG_ON(!recon_state->nr_caps == !recon_state->nr_realms);
3263
3264         /* pre-allocate new pagelist */
3265         _pagelist = ceph_pagelist_alloc(GFP_NOFS);
3266         if (!_pagelist)
3267                 return -ENOMEM;
3268
3269         reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3270         if (!reply)
3271                 goto fail_msg;
3272
3273         /* placeholder for nr_caps */
3274         err = ceph_pagelist_encode_32(_pagelist, 0);
3275         if (err < 0)
3276                 goto fail;
3277
3278         if (recon_state->nr_caps) {
3279                 /* currently encoding caps */
3280                 err = ceph_pagelist_encode_32(recon_state->pagelist, 0);
3281                 if (err)
3282                         goto fail;
3283         } else {
3284                 /* placeholder for nr_realms (currently encoding relams) */
3285                 err = ceph_pagelist_encode_32(_pagelist, 0);
3286                 if (err < 0)
3287                         goto fail;
3288         }
3289
3290         err = ceph_pagelist_encode_8(recon_state->pagelist, 1);
3291         if (err)
3292                 goto fail;
3293
3294         page = list_first_entry(&recon_state->pagelist->head, struct page, lru);
3295         addr = kmap_atomic(page);
3296         if (recon_state->nr_caps) {
3297                 /* currently encoding caps */
3298                 *addr = cpu_to_le32(recon_state->nr_caps);
3299         } else {
3300                 /* currently encoding relams */
3301                 *(addr + 1) = cpu_to_le32(recon_state->nr_realms);
3302         }
3303         kunmap_atomic(addr);
3304
3305         reply->hdr.version = cpu_to_le16(5);
3306         reply->hdr.compat_version = cpu_to_le16(4);
3307
3308         reply->hdr.data_len = cpu_to_le32(recon_state->pagelist->length);
3309         ceph_msg_data_add_pagelist(reply, recon_state->pagelist);
3310
3311         ceph_con_send(&recon_state->session->s_con, reply);
3312         ceph_pagelist_release(recon_state->pagelist);
3313
3314         recon_state->pagelist = _pagelist;
3315         recon_state->nr_caps = 0;
3316         recon_state->nr_realms = 0;
3317         recon_state->msg_version = 5;
3318         return 0;
3319 fail:
3320         ceph_msg_put(reply);
3321 fail_msg:
3322         ceph_pagelist_release(_pagelist);
3323         return err;
3324 }
3325
3326 /*
3327  * Encode information about a cap for a reconnect with the MDS.
3328  */
3329 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
3330                           void *arg)
3331 {
3332         union {
3333                 struct ceph_mds_cap_reconnect v2;
3334                 struct ceph_mds_cap_reconnect_v1 v1;
3335         } rec;
3336         struct ceph_inode_info *ci = cap->ci;
3337         struct ceph_reconnect_state *recon_state = arg;
3338         struct ceph_pagelist *pagelist = recon_state->pagelist;
3339         int err;
3340         u64 snap_follows;
3341
3342         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
3343              inode, ceph_vinop(inode), cap, cap->cap_id,
3344              ceph_cap_string(cap->issued));
3345
3346         spin_lock(&ci->i_ceph_lock);
3347         cap->seq = 0;        /* reset cap seq */
3348         cap->issue_seq = 0;  /* and issue_seq */
3349         cap->mseq = 0;       /* and migrate_seq */
3350         cap->cap_gen = cap->session->s_cap_gen;
3351
3352         if (recon_state->msg_version >= 2) {
3353                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
3354                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3355                 rec.v2.issued = cpu_to_le32(cap->issued);
3356                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3357                 rec.v2.pathbase = 0;
3358                 rec.v2.flock_len = (__force __le32)
3359                         ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
3360         } else {
3361                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
3362                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3363                 rec.v1.issued = cpu_to_le32(cap->issued);
3364                 rec.v1.size = cpu_to_le64(inode->i_size);
3365                 ceph_encode_timespec64(&rec.v1.mtime, &inode->i_mtime);
3366                 ceph_encode_timespec64(&rec.v1.atime, &inode->i_atime);
3367                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3368                 rec.v1.pathbase = 0;
3369         }
3370
3371         if (list_empty(&ci->i_cap_snaps)) {
3372                 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
3373         } else {
3374                 struct ceph_cap_snap *capsnap =
3375                         list_first_entry(&ci->i_cap_snaps,
3376                                          struct ceph_cap_snap, ci_item);
3377                 snap_follows = capsnap->follows;
3378         }
3379         spin_unlock(&ci->i_ceph_lock);
3380
3381         if (recon_state->msg_version >= 2) {
3382                 int num_fcntl_locks, num_flock_locks;
3383                 struct ceph_filelock *flocks = NULL;
3384                 size_t struct_len, total_len = sizeof(u64);
3385                 u8 struct_v = 0;
3386
3387 encode_again:
3388                 if (rec.v2.flock_len) {
3389                         ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
3390                 } else {
3391                         num_fcntl_locks = 0;
3392                         num_flock_locks = 0;
3393                 }
3394                 if (num_fcntl_locks + num_flock_locks > 0) {
3395                         flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
3396                                                sizeof(struct ceph_filelock),
3397                                                GFP_NOFS);
3398                         if (!flocks) {
3399                                 err = -ENOMEM;
3400                                 goto out_err;
3401                         }
3402                         err = ceph_encode_locks_to_buffer(inode, flocks,
3403                                                           num_fcntl_locks,
3404                                                           num_flock_locks);
3405                         if (err) {
3406                                 kfree(flocks);
3407                                 flocks = NULL;
3408                                 if (err == -ENOSPC)
3409                                         goto encode_again;
3410                                 goto out_err;
3411                         }
3412                 } else {
3413                         kfree(flocks);
3414                         flocks = NULL;
3415                 }
3416
3417                 if (recon_state->msg_version >= 3) {
3418                         /* version, compat_version and struct_len */
3419                         total_len += 2 * sizeof(u8) + sizeof(u32);
3420                         struct_v = 2;
3421                 }
3422                 /*
3423                  * number of encoded locks is stable, so copy to pagelist
3424                  */
3425                 struct_len = 2 * sizeof(u32) +
3426                             (num_fcntl_locks + num_flock_locks) *
3427                             sizeof(struct ceph_filelock);
3428                 rec.v2.flock_len = cpu_to_le32(struct_len);
3429
3430                 struct_len += sizeof(u32) + sizeof(rec.v2);
3431
3432                 if (struct_v >= 2)
3433                         struct_len += sizeof(u64); /* snap_follows */
3434
3435                 total_len += struct_len;
3436
3437                 if (pagelist->length + total_len > RECONNECT_MAX_SIZE) {
3438                         err = send_reconnect_partial(recon_state);
3439                         if (err)
3440                                 goto out_freeflocks;
3441                         pagelist = recon_state->pagelist;
3442                 }
3443
3444                 err = ceph_pagelist_reserve(pagelist, total_len);
3445                 if (err)
3446                         goto out_freeflocks;
3447
3448                 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3449                 if (recon_state->msg_version >= 3) {
3450                         ceph_pagelist_encode_8(pagelist, struct_v);
3451                         ceph_pagelist_encode_8(pagelist, 1);
3452                         ceph_pagelist_encode_32(pagelist, struct_len);
3453                 }
3454                 ceph_pagelist_encode_string(pagelist, NULL, 0);
3455                 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
3456                 ceph_locks_to_pagelist(flocks, pagelist,
3457                                        num_fcntl_locks, num_flock_locks);
3458                 if (struct_v >= 2)
3459                         ceph_pagelist_encode_64(pagelist, snap_follows);
3460 out_freeflocks:
3461                 kfree(flocks);
3462         } else {
3463                 u64 pathbase = 0;
3464                 int pathlen = 0;
3465                 char *path = NULL;
3466                 struct dentry *dentry;
3467
3468                 dentry = d_find_alias(inode);
3469                 if (dentry) {
3470                         path = ceph_mdsc_build_path(dentry,
3471                                                 &pathlen, &pathbase, 0);
3472                         dput(dentry);
3473                         if (IS_ERR(path)) {
3474                                 err = PTR_ERR(path);
3475                                 goto out_err;
3476                         }
3477                         rec.v1.pathbase = cpu_to_le64(pathbase);
3478                 }
3479
3480                 err = ceph_pagelist_reserve(pagelist,
3481                                             sizeof(u64) + sizeof(u32) +
3482                                             pathlen + sizeof(rec.v1));
3483                 if (err) {
3484                         goto out_freepath;
3485                 }
3486
3487                 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3488                 ceph_pagelist_encode_string(pagelist, path, pathlen);
3489                 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
3490 out_freepath:
3491                 ceph_mdsc_free_path(path, pathlen);
3492         }
3493
3494 out_err:
3495         if (err >= 0)
3496                 recon_state->nr_caps++;
3497         return err;
3498 }
3499
3500 static int encode_snap_realms(struct ceph_mds_client *mdsc,
3501                               struct ceph_reconnect_state *recon_state)
3502 {
3503         struct rb_node *p;
3504         struct ceph_pagelist *pagelist = recon_state->pagelist;
3505         int err = 0;
3506
3507         if (recon_state->msg_version >= 4) {
3508                 err = ceph_pagelist_encode_32(pagelist, mdsc->num_snap_realms);
3509                 if (err < 0)
3510                         goto fail;
3511         }
3512
3513         /*
3514          * snaprealms.  we provide mds with the ino, seq (version), and
3515          * parent for all of our realms.  If the mds has any newer info,
3516          * it will tell us.
3517          */
3518         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3519                 struct ceph_snap_realm *realm =
3520                        rb_entry(p, struct ceph_snap_realm, node);
3521                 struct ceph_mds_snaprealm_reconnect sr_rec;
3522
3523                 if (recon_state->msg_version >= 4) {
3524                         size_t need = sizeof(u8) * 2 + sizeof(u32) +
3525                                       sizeof(sr_rec);
3526
3527                         if (pagelist->length + need > RECONNECT_MAX_SIZE) {
3528                                 err = send_reconnect_partial(recon_state);
3529                                 if (err)
3530                                         goto fail;
3531                                 pagelist = recon_state->pagelist;
3532                         }
3533
3534                         err = ceph_pagelist_reserve(pagelist, need);
3535                         if (err)
3536                                 goto fail;
3537
3538                         ceph_pagelist_encode_8(pagelist, 1);
3539                         ceph_pagelist_encode_8(pagelist, 1);
3540                         ceph_pagelist_encode_32(pagelist, sizeof(sr_rec));
3541                 }
3542
3543                 dout(" adding snap realm %llx seq %lld parent %llx\n",
3544                      realm->ino, realm->seq, realm->parent_ino);
3545                 sr_rec.ino = cpu_to_le64(realm->ino);
3546                 sr_rec.seq = cpu_to_le64(realm->seq);
3547                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3548
3549                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3550                 if (err)
3551                         goto fail;
3552
3553                 recon_state->nr_realms++;
3554         }
3555 fail:
3556         return err;
3557 }
3558
3559
3560 /*
3561  * If an MDS fails and recovers, clients need to reconnect in order to
3562  * reestablish shared state.  This includes all caps issued through
3563  * this session _and_ the snap_realm hierarchy.  Because it's not
3564  * clear which snap realms the mds cares about, we send everything we
3565  * know about.. that ensures we'll then get any new info the
3566  * recovering MDS might have.
3567  *
3568  * This is a relatively heavyweight operation, but it's rare.
3569  *
3570  * called with mdsc->mutex held.
3571  */
3572 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3573                                struct ceph_mds_session *session)
3574 {
3575         struct ceph_msg *reply;
3576         int mds = session->s_mds;
3577         int err = -ENOMEM;
3578         struct ceph_reconnect_state recon_state = {
3579                 .session = session,
3580         };
3581         LIST_HEAD(dispose);
3582
3583         pr_info("mds%d reconnect start\n", mds);
3584
3585         recon_state.pagelist = ceph_pagelist_alloc(GFP_NOFS);
3586         if (!recon_state.pagelist)
3587                 goto fail_nopagelist;
3588
3589         reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3590         if (!reply)
3591                 goto fail_nomsg;
3592
3593         mutex_lock(&session->s_mutex);
3594         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3595         session->s_seq = 0;
3596
3597         dout("session %p state %s\n", session,
3598              ceph_session_state_name(session->s_state));
3599
3600         spin_lock(&session->s_gen_ttl_lock);
3601         session->s_cap_gen++;
3602         spin_unlock(&session->s_gen_ttl_lock);
3603
3604         spin_lock(&session->s_cap_lock);
3605         /* don't know if session is readonly */
3606         session->s_readonly = 0;
3607         /*
3608          * notify __ceph_remove_cap() that we are composing cap reconnect.
3609          * If a cap get released before being added to the cap reconnect,
3610          * __ceph_remove_cap() should skip queuing cap release.
3611          */
3612         session->s_cap_reconnect = 1;
3613         /* drop old cap expires; we're about to reestablish that state */
3614         detach_cap_releases(session, &dispose);
3615         spin_unlock(&session->s_cap_lock);
3616         dispose_cap_releases(mdsc, &dispose);
3617
3618         /* trim unused caps to reduce MDS's cache rejoin time */
3619         if (mdsc->fsc->sb->s_root)
3620                 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3621
3622         ceph_con_close(&session->s_con);
3623         ceph_con_open(&session->s_con,
3624                       CEPH_ENTITY_TYPE_MDS, mds,
3625                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3626
3627         /* replay unsafe requests */
3628         replay_unsafe_requests(mdsc, session);
3629
3630         ceph_early_kick_flushing_caps(mdsc, session);
3631
3632         down_read(&mdsc->snap_rwsem);
3633
3634         /* placeholder for nr_caps */
3635         err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
3636         if (err)
3637                 goto fail;
3638
3639         if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
3640                 recon_state.msg_version = 3;
3641                 recon_state.allow_multi = true;
3642         } else if (session->s_con.peer_features & CEPH_FEATURE_MDSENC) {
3643                 recon_state.msg_version = 3;
3644         } else {
3645                 recon_state.msg_version = 2;
3646         }
3647         /* trsaverse this session's caps */
3648         err = ceph_iterate_session_caps(session, encode_caps_cb, &recon_state);
3649
3650         spin_lock(&session->s_cap_lock);
3651         session->s_cap_reconnect = 0;
3652         spin_unlock(&session->s_cap_lock);
3653
3654         if (err < 0)
3655                 goto fail;
3656
3657         /* check if all realms can be encoded into current message */
3658         if (mdsc->num_snap_realms) {
3659                 size_t total_len =
3660                         recon_state.pagelist->length +
3661                         mdsc->num_snap_realms *
3662                         sizeof(struct ceph_mds_snaprealm_reconnect);
3663                 if (recon_state.msg_version >= 4) {
3664                         /* number of realms */
3665                         total_len += sizeof(u32);
3666                         /* version, compat_version and struct_len */
3667                         total_len += mdsc->num_snap_realms *
3668                                      (2 * sizeof(u8) + sizeof(u32));
3669                 }
3670                 if (total_len > RECONNECT_MAX_SIZE) {
3671                         if (!recon_state.allow_multi) {
3672                                 err = -ENOSPC;
3673                                 goto fail;
3674                         }
3675                         if (recon_state.nr_caps) {
3676                                 err = send_reconnect_partial(&recon_state);
3677                                 if (err)
3678                                         goto fail;
3679                         }
3680                         recon_state.msg_version = 5;
3681                 }
3682         }
3683
3684         err = encode_snap_realms(mdsc, &recon_state);
3685         if (err < 0)
3686                 goto fail;
3687
3688         if (recon_state.msg_version >= 5) {
3689                 err = ceph_pagelist_encode_8(recon_state.pagelist, 0);
3690                 if (err < 0)
3691                         goto fail;
3692         }
3693
3694         if (recon_state.nr_caps || recon_state.nr_realms) {
3695                 struct page *page =
3696                         list_first_entry(&recon_state.pagelist->head,
3697                                         struct page, lru);
3698                 __le32 *addr = kmap_atomic(page);
3699                 if (recon_state.nr_caps) {
3700                         WARN_ON(recon_state.nr_realms != mdsc->num_snap_realms);
3701                         *addr = cpu_to_le32(recon_state.nr_caps);
3702                 } else if (recon_state.msg_version >= 4) {
3703                         *(addr + 1) = cpu_to_le32(recon_state.nr_realms);
3704                 }
3705                 kunmap_atomic(addr);
3706         }
3707
3708         reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3709         if (recon_state.msg_version >= 4)
3710                 reply->hdr.compat_version = cpu_to_le16(4);
3711
3712         reply->hdr.data_len = cpu_to_le32(recon_state.pagelist->length);
3713         ceph_msg_data_add_pagelist(reply, recon_state.pagelist);
3714
3715         ceph_con_send(&session->s_con, reply);
3716
3717         mutex_unlock(&session->s_mutex);
3718
3719         mutex_lock(&mdsc->mutex);
3720         __wake_requests(mdsc, &session->s_waiting);
3721         mutex_unlock(&mdsc->mutex);
3722
3723         up_read(&mdsc->snap_rwsem);
3724         ceph_pagelist_release(recon_state.pagelist);
3725         return;
3726
3727 fail:
3728         ceph_msg_put(reply);
3729         up_read(&mdsc->snap_rwsem);
3730         mutex_unlock(&session->s_mutex);
3731 fail_nomsg:
3732         ceph_pagelist_release(recon_state.pagelist);
3733 fail_nopagelist:
3734         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3735         return;
3736 }
3737
3738
3739 /*
3740  * compare old and new mdsmaps, kicking requests
3741  * and closing out old connections as necessary
3742  *
3743  * called under mdsc->mutex.
3744  */
3745 static void check_new_map(struct ceph_mds_client *mdsc,
3746                           struct ceph_mdsmap *newmap,
3747                           struct ceph_mdsmap *oldmap)
3748 {
3749         int i;
3750         int oldstate, newstate;
3751         struct ceph_mds_session *s;
3752
3753         dout("check_new_map new %u old %u\n",
3754              newmap->m_epoch, oldmap->m_epoch);
3755
3756         for (i = 0; i < oldmap->m_num_mds && i < mdsc->max_sessions; i++) {
3757                 if (!mdsc->sessions[i])
3758                         continue;
3759                 s = mdsc->sessions[i];
3760                 oldstate = ceph_mdsmap_get_state(oldmap, i);
3761                 newstate = ceph_mdsmap_get_state(newmap, i);
3762
3763                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3764                      i, ceph_mds_state_name(oldstate),
3765                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3766                      ceph_mds_state_name(newstate),
3767                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3768                      ceph_session_state_name(s->s_state));
3769
3770                 if (i >= newmap->m_num_mds) {
3771                         /* force close session for stopped mds */
3772                         get_session(s);
3773                         __unregister_session(mdsc, s);
3774                         __wake_requests(mdsc, &s->s_waiting);
3775                         mutex_unlock(&mdsc->mutex);
3776
3777                         mutex_lock(&s->s_mutex);
3778                         cleanup_session_requests(mdsc, s);
3779                         remove_session_caps(s);
3780                         mutex_unlock(&s->s_mutex);
3781
3782                         ceph_put_mds_session(s);
3783
3784                         mutex_lock(&mdsc->mutex);
3785                         kick_requests(mdsc, i);
3786                         continue;
3787                 }
3788
3789                 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
3790                            ceph_mdsmap_get_addr(newmap, i),
3791                            sizeof(struct ceph_entity_addr))) {
3792                         /* just close it */
3793                         mutex_unlock(&mdsc->mutex);
3794                         mutex_lock(&s->s_mutex);
3795                         mutex_lock(&mdsc->mutex);
3796                         ceph_con_close(&s->s_con);
3797                         mutex_unlock(&s->s_mutex);
3798                         s->s_state = CEPH_MDS_SESSION_RESTARTING;
3799                 } else if (oldstate == newstate) {
3800                         continue;  /* nothing new with this mds */
3801                 }
3802
3803                 /*
3804                  * send reconnect?
3805                  */
3806                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3807                     newstate >= CEPH_MDS_STATE_RECONNECT) {
3808                         mutex_unlock(&mdsc->mutex);
3809                         send_mds_reconnect(mdsc, s);
3810                         mutex_lock(&mdsc->mutex);
3811                 }
3812
3813                 /*
3814                  * kick request on any mds that has gone active.
3815                  */
3816                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3817                     newstate >= CEPH_MDS_STATE_ACTIVE) {
3818                         if (oldstate != CEPH_MDS_STATE_CREATING &&
3819                             oldstate != CEPH_MDS_STATE_STARTING)
3820                                 pr_info("mds%d recovery completed\n", s->s_mds);
3821                         kick_requests(mdsc, i);
3822                         ceph_kick_flushing_caps(mdsc, s);
3823                         wake_up_session_caps(s, RECONNECT);
3824                 }
3825         }
3826
3827         for (i = 0; i < newmap->m_num_mds && i < mdsc->max_sessions; i++) {
3828                 s = mdsc->sessions[i];
3829                 if (!s)
3830                         continue;
3831                 if (!ceph_mdsmap_is_laggy(newmap, i))
3832                         continue;
3833                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3834                     s->s_state == CEPH_MDS_SESSION_HUNG ||
3835                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
3836                         dout(" connecting to export targets of laggy mds%d\n",
3837                              i);
3838                         __open_export_target_sessions(mdsc, s);
3839                 }
3840         }
3841 }
3842
3843
3844
3845 /*
3846  * leases
3847  */
3848
3849 /*
3850  * caller must hold session s_mutex, dentry->d_lock
3851  */
3852 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3853 {
3854         struct ceph_dentry_info *di = ceph_dentry(dentry);
3855
3856         ceph_put_mds_session(di->lease_session);
3857         di->lease_session = NULL;
3858 }
3859
3860 static void handle_lease(struct ceph_mds_client *mdsc,
3861                          struct ceph_mds_session *session,
3862                          struct ceph_msg *msg)
3863 {
3864         struct super_block *sb = mdsc->fsc->sb;
3865         struct inode *inode;
3866         struct dentry *parent, *dentry;
3867         struct ceph_dentry_info *di;
3868         int mds = session->s_mds;
3869         struct ceph_mds_lease *h = msg->front.iov_base;
3870         u32 seq;
3871         struct ceph_vino vino;
3872         struct qstr dname;
3873         int release = 0;
3874
3875         dout("handle_lease from mds%d\n", mds);
3876
3877         /* decode */
3878         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3879                 goto bad;
3880         vino.ino = le64_to_cpu(h->ino);
3881         vino.snap = CEPH_NOSNAP;
3882         seq = le32_to_cpu(h->seq);
3883         dname.len = get_unaligned_le32(h + 1);
3884         if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
3885                 goto bad;
3886         dname.name = (void *)(h + 1) + sizeof(u32);
3887
3888         /* lookup inode */
3889         inode = ceph_find_inode(sb, vino);
3890         dout("handle_lease %s, ino %llx %p %.*s\n",
3891              ceph_lease_op_name(h->action), vino.ino, inode,
3892              dname.len, dname.name);
3893
3894         mutex_lock(&session->s_mutex);
3895         session->s_seq++;
3896
3897         if (!inode) {
3898                 dout("handle_lease no inode %llx\n", vino.ino);
3899                 goto release;
3900         }
3901
3902         /* dentry */
3903         parent = d_find_alias(inode);
3904         if (!parent) {
3905                 dout("no parent dentry on inode %p\n", inode);
3906                 WARN_ON(1);
3907                 goto release;  /* hrm... */
3908         }
3909         dname.hash = full_name_hash(parent, dname.name, dname.len);
3910         dentry = d_lookup(parent, &dname);
3911         dput(parent);
3912         if (!dentry)
3913                 goto release;
3914
3915         spin_lock(&dentry->d_lock);
3916         di = ceph_dentry(dentry);
3917         switch (h->action) {
3918         case CEPH_MDS_LEASE_REVOKE:
3919                 if (di->lease_session == session) {
3920                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3921                                 h->seq = cpu_to_le32(di->lease_seq);
3922                         __ceph_mdsc_drop_dentry_lease(dentry);
3923                 }
3924                 release = 1;
3925                 break;
3926
3927         case CEPH_MDS_LEASE_RENEW:
3928                 if (di->lease_session == session &&
3929                     di->lease_gen == session->s_cap_gen &&
3930                     di->lease_renew_from &&
3931                     di->lease_renew_after == 0) {
3932                         unsigned long duration =
3933                                 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3934
3935                         di->lease_seq = seq;
3936                         di->time = di->lease_renew_from + duration;
3937                         di->lease_renew_after = di->lease_renew_from +
3938                                 (duration >> 1);
3939                         di->lease_renew_from = 0;
3940                 }
3941                 break;
3942         }
3943         spin_unlock(&dentry->d_lock);
3944         dput(dentry);
3945
3946         if (!release)
3947                 goto out;
3948
3949 release:
3950         /* let's just reuse the same message */
3951         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3952         ceph_msg_get(msg);
3953         ceph_con_send(&session->s_con, msg);
3954
3955 out:
3956         mutex_unlock(&session->s_mutex);
3957         /* avoid calling iput_final() in mds dispatch threads */
3958         ceph_async_iput(inode);
3959         return;
3960
3961 bad:
3962         pr_err("corrupt lease message\n");
3963         ceph_msg_dump(msg);
3964 }
3965
3966 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3967                               struct dentry *dentry, char action,
3968                               u32 seq)
3969 {
3970         struct ceph_msg *msg;
3971         struct ceph_mds_lease *lease;
3972         struct inode *dir;
3973         int len = sizeof(*lease) + sizeof(u32) + NAME_MAX;
3974
3975         dout("lease_send_msg identry %p %s to mds%d\n",
3976              dentry, ceph_lease_op_name(action), session->s_mds);
3977
3978         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3979         if (!msg)
3980                 return;
3981         lease = msg->front.iov_base;
3982         lease->action = action;
3983         lease->seq = cpu_to_le32(seq);
3984
3985         spin_lock(&dentry->d_lock);
3986         dir = d_inode(dentry->d_parent);
3987         lease->ino = cpu_to_le64(ceph_ino(dir));
3988         lease->first = lease->last = cpu_to_le64(ceph_snap(dir));
3989
3990         put_unaligned_le32(dentry->d_name.len, lease + 1);
3991         memcpy((void *)(lease + 1) + 4,
3992                dentry->d_name.name, dentry->d_name.len);
3993         spin_unlock(&dentry->d_lock);
3994         /*
3995          * if this is a preemptive lease RELEASE, no need to
3996          * flush request stream, since the actual request will
3997          * soon follow.
3998          */
3999         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
4000
4001         ceph_con_send(&session->s_con, msg);
4002 }
4003
4004 /*
4005  * lock unlock sessions, to wait ongoing session activities
4006  */
4007 static void lock_unlock_sessions(struct ceph_mds_client *mdsc)
4008 {
4009         int i;
4010
4011         mutex_lock(&mdsc->mutex);
4012         for (i = 0; i < mdsc->max_sessions; i++) {
4013                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4014                 if (!s)
4015                         continue;
4016                 mutex_unlock(&mdsc->mutex);
4017                 mutex_lock(&s->s_mutex);
4018                 mutex_unlock(&s->s_mutex);
4019                 ceph_put_mds_session(s);
4020                 mutex_lock(&mdsc->mutex);
4021         }
4022         mutex_unlock(&mdsc->mutex);
4023 }
4024
4025 static void maybe_recover_session(struct ceph_mds_client *mdsc)
4026 {
4027         struct ceph_fs_client *fsc = mdsc->fsc;
4028
4029         if (!ceph_test_mount_opt(fsc, CLEANRECOVER))
4030                 return;
4031
4032         if (READ_ONCE(fsc->mount_state) != CEPH_MOUNT_MOUNTED)
4033                 return;
4034
4035         if (!READ_ONCE(fsc->blacklisted))
4036                 return;
4037
4038         if (fsc->last_auto_reconnect &&
4039             time_before(jiffies, fsc->last_auto_reconnect + HZ * 60 * 30))
4040                 return;
4041
4042         pr_info("auto reconnect after blacklisted\n");
4043         fsc->last_auto_reconnect = jiffies;
4044         ceph_force_reconnect(fsc->sb);
4045 }
4046
4047 /*
4048  * delayed work -- periodically trim expired leases, renew caps with mds
4049  */
4050 static void schedule_delayed(struct ceph_mds_client *mdsc)
4051 {
4052         int delay = 5;
4053         unsigned hz = round_jiffies_relative(HZ * delay);
4054         schedule_delayed_work(&mdsc->delayed_work, hz);
4055 }
4056
4057 static void delayed_work(struct work_struct *work)
4058 {
4059         int i;
4060         struct ceph_mds_client *mdsc =
4061                 container_of(work, struct ceph_mds_client, delayed_work.work);
4062         int renew_interval;
4063         int renew_caps;
4064
4065         dout("mdsc delayed_work\n");
4066
4067         mutex_lock(&mdsc->mutex);
4068         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
4069         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
4070                                    mdsc->last_renew_caps);
4071         if (renew_caps)
4072                 mdsc->last_renew_caps = jiffies;
4073
4074         for (i = 0; i < mdsc->max_sessions; i++) {
4075                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4076                 if (!s)
4077                         continue;
4078                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
4079                         dout("resending session close request for mds%d\n",
4080                              s->s_mds);
4081                         request_close_session(mdsc, s);
4082                         ceph_put_mds_session(s);
4083                         continue;
4084                 }
4085                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
4086                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
4087                                 s->s_state = CEPH_MDS_SESSION_HUNG;
4088                                 pr_info("mds%d hung\n", s->s_mds);
4089                         }
4090                 }
4091                 if (s->s_state == CEPH_MDS_SESSION_NEW ||
4092                     s->s_state == CEPH_MDS_SESSION_RESTARTING ||
4093                     s->s_state == CEPH_MDS_SESSION_REJECTED) {
4094                         /* this mds is failed or recovering, just wait */
4095                         ceph_put_mds_session(s);
4096                         continue;
4097                 }
4098                 mutex_unlock(&mdsc->mutex);
4099
4100                 mutex_lock(&s->s_mutex);
4101                 if (renew_caps)
4102                         send_renew_caps(mdsc, s);
4103                 else
4104                         ceph_con_keepalive(&s->s_con);
4105                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4106                     s->s_state == CEPH_MDS_SESSION_HUNG)
4107                         ceph_send_cap_releases(mdsc, s);
4108                 mutex_unlock(&s->s_mutex);
4109                 ceph_put_mds_session(s);
4110
4111                 mutex_lock(&mdsc->mutex);
4112         }
4113         mutex_unlock(&mdsc->mutex);
4114
4115         ceph_check_delayed_caps(mdsc);
4116
4117         ceph_queue_cap_reclaim_work(mdsc);
4118
4119         ceph_trim_snapid_map(mdsc);
4120
4121         maybe_recover_session(mdsc);
4122
4123         schedule_delayed(mdsc);
4124 }
4125
4126 int ceph_mdsc_init(struct ceph_fs_client *fsc)
4127
4128 {
4129         struct ceph_mds_client *mdsc;
4130
4131         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
4132         if (!mdsc)
4133                 return -ENOMEM;
4134         mdsc->fsc = fsc;
4135         mutex_init(&mdsc->mutex);
4136         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
4137         if (!mdsc->mdsmap) {
4138                 kfree(mdsc);
4139                 return -ENOMEM;
4140         }
4141
4142         fsc->mdsc = mdsc;
4143         init_completion(&mdsc->safe_umount_waiters);
4144         init_waitqueue_head(&mdsc->session_close_wq);
4145         INIT_LIST_HEAD(&mdsc->waiting_for_map);
4146         mdsc->sessions = NULL;
4147         atomic_set(&mdsc->num_sessions, 0);
4148         mdsc->max_sessions = 0;
4149         mdsc->stopping = 0;
4150         atomic64_set(&mdsc->quotarealms_count, 0);
4151         mdsc->quotarealms_inodes = RB_ROOT;
4152         mutex_init(&mdsc->quotarealms_inodes_mutex);
4153         mdsc->last_snap_seq = 0;
4154         init_rwsem(&mdsc->snap_rwsem);
4155         mdsc->snap_realms = RB_ROOT;
4156         INIT_LIST_HEAD(&mdsc->snap_empty);
4157         mdsc->num_snap_realms = 0;
4158         spin_lock_init(&mdsc->snap_empty_lock);
4159         mdsc->last_tid = 0;
4160         mdsc->oldest_tid = 0;
4161         mdsc->request_tree = RB_ROOT;
4162         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
4163         mdsc->last_renew_caps = jiffies;
4164         INIT_LIST_HEAD(&mdsc->cap_delay_list);
4165         spin_lock_init(&mdsc->cap_delay_lock);
4166         INIT_LIST_HEAD(&mdsc->snap_flush_list);
4167         spin_lock_init(&mdsc->snap_flush_lock);
4168         mdsc->last_cap_flush_tid = 1;
4169         INIT_LIST_HEAD(&mdsc->cap_flush_list);
4170         INIT_LIST_HEAD(&mdsc->cap_dirty);
4171         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
4172         mdsc->num_cap_flushing = 0;
4173         spin_lock_init(&mdsc->cap_dirty_lock);
4174         init_waitqueue_head(&mdsc->cap_flushing_wq);
4175         INIT_WORK(&mdsc->cap_reclaim_work, ceph_cap_reclaim_work);
4176         atomic_set(&mdsc->cap_reclaim_pending, 0);
4177
4178         spin_lock_init(&mdsc->dentry_list_lock);
4179         INIT_LIST_HEAD(&mdsc->dentry_leases);
4180         INIT_LIST_HEAD(&mdsc->dentry_dir_leases);
4181
4182         ceph_caps_init(mdsc);
4183         ceph_adjust_caps_max_min(mdsc, fsc->mount_options);
4184
4185         spin_lock_init(&mdsc->snapid_map_lock);
4186         mdsc->snapid_map_tree = RB_ROOT;
4187         INIT_LIST_HEAD(&mdsc->snapid_map_lru);
4188
4189         init_rwsem(&mdsc->pool_perm_rwsem);
4190         mdsc->pool_perm_tree = RB_ROOT;
4191
4192         strscpy(mdsc->nodename, utsname()->nodename,
4193                 sizeof(mdsc->nodename));
4194         return 0;
4195 }
4196
4197 /*
4198  * Wait for safe replies on open mds requests.  If we time out, drop
4199  * all requests from the tree to avoid dangling dentry refs.
4200  */
4201 static void wait_requests(struct ceph_mds_client *mdsc)
4202 {
4203         struct ceph_options *opts = mdsc->fsc->client->options;
4204         struct ceph_mds_request *req;
4205
4206         mutex_lock(&mdsc->mutex);
4207         if (__get_oldest_req(mdsc)) {
4208                 mutex_unlock(&mdsc->mutex);
4209
4210                 dout("wait_requests waiting for requests\n");
4211                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
4212                                     ceph_timeout_jiffies(opts->mount_timeout));
4213
4214                 /* tear down remaining requests */
4215                 mutex_lock(&mdsc->mutex);
4216                 while ((req = __get_oldest_req(mdsc))) {
4217                         dout("wait_requests timed out on tid %llu\n",
4218                              req->r_tid);
4219                         list_del_init(&req->r_wait);
4220                         __unregister_request(mdsc, req);
4221                 }
4222         }
4223         mutex_unlock(&mdsc->mutex);
4224         dout("wait_requests done\n");
4225 }
4226
4227 /*
4228  * called before mount is ro, and before dentries are torn down.
4229  * (hmm, does this still race with new lookups?)
4230  */
4231 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
4232 {
4233         dout("pre_umount\n");
4234         mdsc->stopping = 1;
4235
4236         lock_unlock_sessions(mdsc);
4237         ceph_flush_dirty_caps(mdsc);
4238         wait_requests(mdsc);
4239
4240         /*
4241          * wait for reply handlers to drop their request refs and
4242          * their inode/dcache refs
4243          */
4244         ceph_msgr_flush();
4245
4246         ceph_cleanup_quotarealms_inodes(mdsc);
4247 }
4248
4249 /*
4250  * wait for all write mds requests to flush.
4251  */
4252 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
4253 {
4254         struct ceph_mds_request *req = NULL, *nextreq;
4255         struct rb_node *n;
4256
4257         mutex_lock(&mdsc->mutex);
4258         dout("wait_unsafe_requests want %lld\n", want_tid);
4259 restart:
4260         req = __get_oldest_req(mdsc);
4261         while (req && req->r_tid <= want_tid) {
4262                 /* find next request */
4263                 n = rb_next(&req->r_node);
4264                 if (n)
4265                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
4266                 else
4267                         nextreq = NULL;
4268                 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
4269                     (req->r_op & CEPH_MDS_OP_WRITE)) {
4270                         /* write op */
4271                         ceph_mdsc_get_request(req);
4272                         if (nextreq)
4273                                 ceph_mdsc_get_request(nextreq);
4274                         mutex_unlock(&mdsc->mutex);
4275                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
4276                              req->r_tid, want_tid);
4277                         wait_for_completion(&req->r_safe_completion);
4278                         mutex_lock(&mdsc->mutex);
4279                         ceph_mdsc_put_request(req);
4280                         if (!nextreq)
4281                                 break;  /* next dne before, so we're done! */
4282                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
4283                                 /* next request was removed from tree */
4284                                 ceph_mdsc_put_request(nextreq);
4285                                 goto restart;
4286                         }
4287                         ceph_mdsc_put_request(nextreq);  /* won't go away */
4288                 }
4289                 req = nextreq;
4290         }
4291         mutex_unlock(&mdsc->mutex);
4292         dout("wait_unsafe_requests done\n");
4293 }
4294
4295 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
4296 {
4297         u64 want_tid, want_flush;
4298
4299         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4300                 return;
4301
4302         dout("sync\n");
4303         mutex_lock(&mdsc->mutex);
4304         want_tid = mdsc->last_tid;
4305         mutex_unlock(&mdsc->mutex);
4306
4307         ceph_flush_dirty_caps(mdsc);
4308         spin_lock(&mdsc->cap_dirty_lock);
4309         want_flush = mdsc->last_cap_flush_tid;
4310         if (!list_empty(&mdsc->cap_flush_list)) {
4311                 struct ceph_cap_flush *cf =
4312                         list_last_entry(&mdsc->cap_flush_list,
4313                                         struct ceph_cap_flush, g_list);
4314                 cf->wake = true;
4315         }
4316         spin_unlock(&mdsc->cap_dirty_lock);
4317
4318         dout("sync want tid %lld flush_seq %lld\n",
4319              want_tid, want_flush);
4320
4321         wait_unsafe_requests(mdsc, want_tid);
4322         wait_caps_flush(mdsc, want_flush);
4323 }
4324
4325 /*
4326  * true if all sessions are closed, or we force unmount
4327  */
4328 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
4329 {
4330         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4331                 return true;
4332         return atomic_read(&mdsc->num_sessions) <= skipped;
4333 }
4334
4335 /*
4336  * called after sb is ro.
4337  */
4338 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
4339 {
4340         struct ceph_options *opts = mdsc->fsc->client->options;
4341         struct ceph_mds_session *session;
4342         int i;
4343         int skipped = 0;
4344
4345         dout("close_sessions\n");
4346
4347         /* close sessions */
4348         mutex_lock(&mdsc->mutex);
4349         for (i = 0; i < mdsc->max_sessions; i++) {
4350                 session = __ceph_lookup_mds_session(mdsc, i);
4351                 if (!session)
4352                         continue;
4353                 mutex_unlock(&mdsc->mutex);
4354                 mutex_lock(&session->s_mutex);
4355                 if (__close_session(mdsc, session) <= 0)
4356                         skipped++;
4357                 mutex_unlock(&session->s_mutex);
4358                 ceph_put_mds_session(session);
4359                 mutex_lock(&mdsc->mutex);
4360         }
4361         mutex_unlock(&mdsc->mutex);
4362
4363         dout("waiting for sessions to close\n");
4364         wait_event_timeout(mdsc->session_close_wq,
4365                            done_closing_sessions(mdsc, skipped),
4366                            ceph_timeout_jiffies(opts->mount_timeout));
4367
4368         /* tear down remaining sessions */
4369         mutex_lock(&mdsc->mutex);
4370         for (i = 0; i < mdsc->max_sessions; i++) {
4371                 if (mdsc->sessions[i]) {
4372                         session = get_session(mdsc->sessions[i]);
4373                         __unregister_session(mdsc, session);
4374                         mutex_unlock(&mdsc->mutex);
4375                         mutex_lock(&session->s_mutex);
4376                         remove_session_caps(session);
4377                         mutex_unlock(&session->s_mutex);
4378                         ceph_put_mds_session(session);
4379                         mutex_lock(&mdsc->mutex);
4380                 }
4381         }
4382         WARN_ON(!list_empty(&mdsc->cap_delay_list));
4383         mutex_unlock(&mdsc->mutex);
4384
4385         ceph_cleanup_snapid_map(mdsc);
4386         ceph_cleanup_empty_realms(mdsc);
4387
4388         cancel_work_sync(&mdsc->cap_reclaim_work);
4389         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4390
4391         dout("stopped\n");
4392 }
4393
4394 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
4395 {
4396         struct ceph_mds_session *session;
4397         int mds;
4398
4399         dout("force umount\n");
4400
4401         mutex_lock(&mdsc->mutex);
4402         for (mds = 0; mds < mdsc->max_sessions; mds++) {
4403                 session = __ceph_lookup_mds_session(mdsc, mds);
4404                 if (!session)
4405                         continue;
4406
4407                 if (session->s_state == CEPH_MDS_SESSION_REJECTED)
4408                         __unregister_session(mdsc, session);
4409                 __wake_requests(mdsc, &session->s_waiting);
4410                 mutex_unlock(&mdsc->mutex);
4411
4412                 mutex_lock(&session->s_mutex);
4413                 __close_session(mdsc, session);
4414                 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
4415                         cleanup_session_requests(mdsc, session);
4416                         remove_session_caps(session);
4417                 }
4418                 mutex_unlock(&session->s_mutex);
4419                 ceph_put_mds_session(session);
4420
4421                 mutex_lock(&mdsc->mutex);
4422                 kick_requests(mdsc, mds);
4423         }
4424         __wake_requests(mdsc, &mdsc->waiting_for_map);
4425         mutex_unlock(&mdsc->mutex);
4426 }
4427
4428 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
4429 {
4430         dout("stop\n");
4431         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4432         if (mdsc->mdsmap)
4433                 ceph_mdsmap_destroy(mdsc->mdsmap);
4434         kfree(mdsc->sessions);
4435         ceph_caps_finalize(mdsc);
4436         ceph_pool_perm_destroy(mdsc);
4437 }
4438
4439 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
4440 {
4441         struct ceph_mds_client *mdsc = fsc->mdsc;
4442         dout("mdsc_destroy %p\n", mdsc);
4443
4444         if (!mdsc)
4445                 return;
4446
4447         /* flush out any connection work with references to us */
4448         ceph_msgr_flush();
4449
4450         ceph_mdsc_stop(mdsc);
4451
4452         fsc->mdsc = NULL;
4453         kfree(mdsc);
4454         dout("mdsc_destroy %p done\n", mdsc);
4455 }
4456
4457 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4458 {
4459         struct ceph_fs_client *fsc = mdsc->fsc;
4460         const char *mds_namespace = fsc->mount_options->mds_namespace;
4461         void *p = msg->front.iov_base;
4462         void *end = p + msg->front.iov_len;
4463         u32 epoch;
4464         u32 map_len;
4465         u32 num_fs;
4466         u32 mount_fscid = (u32)-1;
4467         u8 struct_v, struct_cv;
4468         int err = -EINVAL;
4469
4470         ceph_decode_need(&p, end, sizeof(u32), bad);
4471         epoch = ceph_decode_32(&p);
4472
4473         dout("handle_fsmap epoch %u\n", epoch);
4474
4475         ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4476         struct_v = ceph_decode_8(&p);
4477         struct_cv = ceph_decode_8(&p);
4478         map_len = ceph_decode_32(&p);
4479
4480         ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
4481         p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
4482
4483         num_fs = ceph_decode_32(&p);
4484         while (num_fs-- > 0) {
4485                 void *info_p, *info_end;
4486                 u32 info_len;
4487                 u8 info_v, info_cv;
4488                 u32 fscid, namelen;
4489
4490                 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4491                 info_v = ceph_decode_8(&p);
4492                 info_cv = ceph_decode_8(&p);
4493                 info_len = ceph_decode_32(&p);
4494                 ceph_decode_need(&p, end, info_len, bad);
4495                 info_p = p;
4496                 info_end = p + info_len;
4497                 p = info_end;
4498
4499                 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
4500                 fscid = ceph_decode_32(&info_p);
4501                 namelen = ceph_decode_32(&info_p);
4502                 ceph_decode_need(&info_p, info_end, namelen, bad);
4503
4504                 if (mds_namespace &&
4505                     strlen(mds_namespace) == namelen &&
4506                     !strncmp(mds_namespace, (char *)info_p, namelen)) {
4507                         mount_fscid = fscid;
4508                         break;
4509                 }
4510         }
4511
4512         ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
4513         if (mount_fscid != (u32)-1) {
4514                 fsc->client->monc.fs_cluster_id = mount_fscid;
4515                 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
4516                                    0, true);
4517                 ceph_monc_renew_subs(&fsc->client->monc);
4518         } else {
4519                 err = -ENOENT;
4520                 goto err_out;
4521         }
4522         return;
4523
4524 bad:
4525         pr_err("error decoding fsmap\n");
4526 err_out:
4527         mutex_lock(&mdsc->mutex);
4528         mdsc->mdsmap_err = err;
4529         __wake_requests(mdsc, &mdsc->waiting_for_map);
4530         mutex_unlock(&mdsc->mutex);
4531 }
4532
4533 /*
4534  * handle mds map update.
4535  */
4536 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4537 {
4538         u32 epoch;
4539         u32 maplen;
4540         void *p = msg->front.iov_base;
4541         void *end = p + msg->front.iov_len;
4542         struct ceph_mdsmap *newmap, *oldmap;
4543         struct ceph_fsid fsid;
4544         int err = -EINVAL;
4545
4546         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
4547         ceph_decode_copy(&p, &fsid, sizeof(fsid));
4548         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
4549                 return;
4550         epoch = ceph_decode_32(&p);
4551         maplen = ceph_decode_32(&p);
4552         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
4553
4554         /* do we need it? */
4555         mutex_lock(&mdsc->mutex);
4556         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
4557                 dout("handle_map epoch %u <= our %u\n",
4558                      epoch, mdsc->mdsmap->m_epoch);
4559                 mutex_unlock(&mdsc->mutex);
4560                 return;
4561         }
4562
4563         newmap = ceph_mdsmap_decode(&p, end);
4564         if (IS_ERR(newmap)) {
4565                 err = PTR_ERR(newmap);
4566                 goto bad_unlock;
4567         }
4568
4569         /* swap into place */
4570         if (mdsc->mdsmap) {
4571                 oldmap = mdsc->mdsmap;
4572                 mdsc->mdsmap = newmap;
4573                 check_new_map(mdsc, newmap, oldmap);
4574                 ceph_mdsmap_destroy(oldmap);
4575         } else {
4576                 mdsc->mdsmap = newmap;  /* first mds map */
4577         }
4578         mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
4579                                         MAX_LFS_FILESIZE);
4580
4581         __wake_requests(mdsc, &mdsc->waiting_for_map);
4582         ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
4583                           mdsc->mdsmap->m_epoch);
4584
4585         mutex_unlock(&mdsc->mutex);
4586         schedule_delayed(mdsc);
4587         return;
4588
4589 bad_unlock:
4590         mutex_unlock(&mdsc->mutex);
4591 bad:
4592         pr_err("error decoding mdsmap %d\n", err);
4593         return;
4594 }
4595
4596 static struct ceph_connection *con_get(struct ceph_connection *con)
4597 {
4598         struct ceph_mds_session *s = con->private;
4599
4600         if (get_session(s)) {
4601                 dout("mdsc con_get %p ok (%d)\n", s, refcount_read(&s->s_ref));
4602                 return con;
4603         }
4604         dout("mdsc con_get %p FAIL\n", s);
4605         return NULL;
4606 }
4607
4608 static void con_put(struct ceph_connection *con)
4609 {
4610         struct ceph_mds_session *s = con->private;
4611
4612         dout("mdsc con_put %p (%d)\n", s, refcount_read(&s->s_ref) - 1);
4613         ceph_put_mds_session(s);
4614 }
4615
4616 /*
4617  * if the client is unresponsive for long enough, the mds will kill
4618  * the session entirely.
4619  */
4620 static void peer_reset(struct ceph_connection *con)
4621 {
4622         struct ceph_mds_session *s = con->private;
4623         struct ceph_mds_client *mdsc = s->s_mdsc;
4624
4625         pr_warn("mds%d closed our session\n", s->s_mds);
4626         send_mds_reconnect(mdsc, s);
4627 }
4628
4629 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
4630 {
4631         struct ceph_mds_session *s = con->private;
4632         struct ceph_mds_client *mdsc = s->s_mdsc;
4633         int type = le16_to_cpu(msg->hdr.type);
4634
4635         mutex_lock(&mdsc->mutex);
4636         if (__verify_registered_session(mdsc, s) < 0) {
4637                 mutex_unlock(&mdsc->mutex);
4638                 goto out;
4639         }
4640         mutex_unlock(&mdsc->mutex);
4641
4642         switch (type) {
4643         case CEPH_MSG_MDS_MAP:
4644                 ceph_mdsc_handle_mdsmap(mdsc, msg);
4645                 break;
4646         case CEPH_MSG_FS_MAP_USER:
4647                 ceph_mdsc_handle_fsmap(mdsc, msg);
4648                 break;
4649         case CEPH_MSG_CLIENT_SESSION:
4650                 handle_session(s, msg);
4651                 break;
4652         case CEPH_MSG_CLIENT_REPLY:
4653                 handle_reply(s, msg);
4654                 break;
4655         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
4656                 handle_forward(mdsc, s, msg);
4657                 break;
4658         case CEPH_MSG_CLIENT_CAPS:
4659                 ceph_handle_caps(s, msg);
4660                 break;
4661         case CEPH_MSG_CLIENT_SNAP:
4662                 ceph_handle_snap(mdsc, s, msg);
4663                 break;
4664         case CEPH_MSG_CLIENT_LEASE:
4665                 handle_lease(mdsc, s, msg);
4666                 break;
4667         case CEPH_MSG_CLIENT_QUOTA:
4668                 ceph_handle_quota(mdsc, s, msg);
4669                 break;
4670
4671         default:
4672                 pr_err("received unknown message type %d %s\n", type,
4673                        ceph_msg_type_name(type));
4674         }
4675 out:
4676         ceph_msg_put(msg);
4677 }
4678
4679 /*
4680  * authentication
4681  */
4682
4683 /*
4684  * Note: returned pointer is the address of a structure that's
4685  * managed separately.  Caller must *not* attempt to free it.
4686  */
4687 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
4688                                         int *proto, int force_new)
4689 {
4690         struct ceph_mds_session *s = con->private;
4691         struct ceph_mds_client *mdsc = s->s_mdsc;
4692         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4693         struct ceph_auth_handshake *auth = &s->s_auth;
4694
4695         if (force_new && auth->authorizer) {
4696                 ceph_auth_destroy_authorizer(auth->authorizer);
4697                 auth->authorizer = NULL;
4698         }
4699         if (!auth->authorizer) {
4700                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4701                                                       auth);
4702                 if (ret)
4703                         return ERR_PTR(ret);
4704         } else {
4705                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4706                                                       auth);
4707                 if (ret)
4708                         return ERR_PTR(ret);
4709         }
4710         *proto = ac->protocol;
4711
4712         return auth;
4713 }
4714
4715 static int add_authorizer_challenge(struct ceph_connection *con,
4716                                     void *challenge_buf, int challenge_buf_len)
4717 {
4718         struct ceph_mds_session *s = con->private;
4719         struct ceph_mds_client *mdsc = s->s_mdsc;
4720         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4721
4722         return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
4723                                             challenge_buf, challenge_buf_len);
4724 }
4725
4726 static int verify_authorizer_reply(struct ceph_connection *con)
4727 {
4728         struct ceph_mds_session *s = con->private;
4729         struct ceph_mds_client *mdsc = s->s_mdsc;
4730         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4731
4732         return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
4733 }
4734
4735 static int invalidate_authorizer(struct ceph_connection *con)
4736 {
4737         struct ceph_mds_session *s = con->private;
4738         struct ceph_mds_client *mdsc = s->s_mdsc;
4739         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4740
4741         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
4742
4743         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
4744 }
4745
4746 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
4747                                 struct ceph_msg_header *hdr, int *skip)
4748 {
4749         struct ceph_msg *msg;
4750         int type = (int) le16_to_cpu(hdr->type);
4751         int front_len = (int) le32_to_cpu(hdr->front_len);
4752
4753         if (con->in_msg)
4754                 return con->in_msg;
4755
4756         *skip = 0;
4757         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
4758         if (!msg) {
4759                 pr_err("unable to allocate msg type %d len %d\n",
4760                        type, front_len);
4761                 return NULL;
4762         }
4763
4764         return msg;
4765 }
4766
4767 static int mds_sign_message(struct ceph_msg *msg)
4768 {
4769        struct ceph_mds_session *s = msg->con->private;
4770        struct ceph_auth_handshake *auth = &s->s_auth;
4771
4772        return ceph_auth_sign_message(auth, msg);
4773 }
4774
4775 static int mds_check_message_signature(struct ceph_msg *msg)
4776 {
4777        struct ceph_mds_session *s = msg->con->private;
4778        struct ceph_auth_handshake *auth = &s->s_auth;
4779
4780        return ceph_auth_check_message_signature(auth, msg);
4781 }
4782
4783 static const struct ceph_connection_operations mds_con_ops = {
4784         .get = con_get,
4785         .put = con_put,
4786         .dispatch = dispatch,
4787         .get_authorizer = get_authorizer,
4788         .add_authorizer_challenge = add_authorizer_challenge,
4789         .verify_authorizer_reply = verify_authorizer_reply,
4790         .invalidate_authorizer = invalidate_authorizer,
4791         .peer_reset = peer_reset,
4792         .alloc_msg = mds_alloc_msg,
4793         .sign_message = mds_sign_message,
4794         .check_message_signature = mds_check_message_signature,
4795 };
4796
4797 /* eof */