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  47  * are no longer valid.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