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  47  * are no longer valid.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_parent);
 714         }
 715         ceph_async_iput(req->r_target_inode);
 716         if (req->r_dentry)
 717                 dput(req->r_dentry);
 718         if (req->r_old_dentry)
 719                 dput(req->r_old_dentry);
 720         if (req->r_old_dentry_dir) {
 721                 /*
 722                  * track (and drop pins for) r_old_dentry_dir
 723                  * separately, since r_old_dentry's d_parent may have
 724                  * changed between the dir mutex being dropped and
 725                  * this request being freed.
 726                  */
 727                 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
 728                                   CEPH_CAP_PIN);
 729                 ceph_async_iput(req->r_old_dentry_dir);
 730         }
 731         kfree(req->r_path1);
 732         kfree(req->r_path2);
 733         if (req->r_pagelist)
 734                 ceph_pagelist_release(req->r_pagelist);
 735         put_request_session(req);
 736         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
 737         WARN_ON_ONCE(!list_empty(&req->r_wait));
 738         kfree(req);
 739 }
 740
 741 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
 742
 743 /*
 744  * lookup session, bump ref if found.
 745  *
 746  * called under mdsc->mutex.
 747  */
 748 static struct ceph_mds_request *
 749 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
 750 {
 751         struct ceph_mds_request *req;
 752
 753         req = lookup_request(&mdsc->request_tree, tid);
 754         if (req)
 755                 ceph_mdsc_get_request(req);
 756
 757         return req;
 758 }
 759
 760 /*
 761  * Register an in-flight request, and assign a tid.  Link to directory
 762  * are modifying (if any).
 763  *
 764  * Called under mdsc->mutex.
 765  */
 766 static void __register_request(struct ceph_mds_client *mdsc,
 767                                struct ceph_mds_request *req,
 768                                struct inode *dir)
 769 {
 770         int ret = 0;
 771
 772         req->r_tid = ++mdsc->last_tid;
 773         if (req->r_num_caps) {
 774                 ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
 775                                         req->r_num_caps);
 776                 if (ret < 0) {
 777                         pr_err("__register_request %p "
 778                                "failed to reserve caps: %d\n", req, ret);
 779                         /* set req->r_err to fail early from __do_request */
 780                         req->r_err = ret;
 781                         return;
 782                 }
 783         }
 784         dout("__register_request %p tid %lld\n", req, req->r_tid);
 785         ceph_mdsc_get_request(req);
 786         insert_request(&mdsc->request_tree, req);
 787
 788         req->r_uid = current_fsuid();
 789         req->r_gid = current_fsgid();
 790
 791         if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
 792                 mdsc->oldest_tid = req->r_tid;
 793
 794         if (dir) {
 795                 ihold(dir);
 796                 req->r_unsafe_dir = dir;
 797         }
 798 }
 799
 800 static void __unregister_request(struct ceph_mds_client *mdsc,
 801                                  struct ceph_mds_request *req)
 802 {
 803         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 804
 805         /* Never leave an unregistered request on an unsafe list! */
 806         list_del_init(&req->r_unsafe_item);
 807
 808         if (req->r_tid == mdsc->oldest_tid) {
 809                 struct rb_node *p = rb_next(&req->r_node);
 810                 mdsc->oldest_tid = 0;
 811                 while (p) {
 812                         struct ceph_mds_request *next_req =
 813                                 rb_entry(p, struct ceph_mds_request, r_node);
 814                         if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
 815                                 mdsc->oldest_tid = next_req->r_tid;
 816                                 break;
 817                         }
 818                         p = rb_next(p);
 819                 }
 820         }
 821
 822         erase_request(&mdsc->request_tree, req);
 823
 824         if (req->r_unsafe_dir  &&
 825             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
 826                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 827                 spin_lock(&ci->i_unsafe_lock);
 828                 list_del_init(&req->r_unsafe_dir_item);
 829                 spin_unlock(&ci->i_unsafe_lock);
 830         }
 831         if (req->r_target_inode &&
 832             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
 833                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
 834                 spin_lock(&ci->i_unsafe_lock);
 835                 list_del_init(&req->r_unsafe_target_item);
 836                 spin_unlock(&ci->i_unsafe_lock);
 837         }
 838
 839         if (req->r_unsafe_dir) {
 840                 /* avoid calling iput_final() in mds dispatch threads */
 841                 ceph_async_iput(req->r_unsafe_dir);
 842                 req->r_unsafe_dir = NULL;
 843         }
 844
 845         complete_all(&req->r_safe_completion);
 846
 847         ceph_mdsc_put_request(req);
 848 }
 849
 850 /*
 851  * Walk back up the dentry tree until we hit a dentry representing a
 852  * non-snapshot inode. We do this using the rcu_read_lock (which must be held
 853  * when calling this) to ensure that the objects won't disappear while we're
 854  * working with them. Once we hit a candidate dentry, we attempt to take a
 855  * reference to it, and return that as the result.
 856  */
 857 static struct inode *get_nonsnap_parent(struct dentry *dentry)
 858 {
 859         struct inode *inode = NULL;
 860
 861         while (dentry && !IS_ROOT(dentry)) {
 862                 inode = d_inode_rcu(dentry);
 863                 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
 864                         break;
 865                 dentry = dentry->d_parent;
 866         }
 867         if (inode)
 868                 inode = igrab(inode);
 869         return inode;
 870 }
 871
 872 /*
 873  * Choose mds to send request to next.  If there is a hint set in the
 874  * request (e.g., due to a prior forward hint from the mds), use that.
 875  * Otherwise, consult frag tree and/or caps to identify the
 876  * appropriate mds.  If all else fails, choose randomly.
 877  *
 878  * Called under mdsc->mutex.
 879  */
 880 static int __choose_mds(struct ceph_mds_client *mdsc,
 881                         struct ceph_mds_request *req)
 882 {
 883         struct inode *inode;
 884         struct ceph_inode_info *ci;
 885         struct ceph_cap *cap;
 886         int mode = req->r_direct_mode;
 887         int mds = -1;
 888         u32 hash = req->r_direct_hash;
 889         bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
 890
 891         /*
 892          * is there a specific mds we should try?  ignore hint if we have
 893          * no session and the mds is not up (active or recovering).
 894          */
 895         if (req->r_resend_mds >= 0 &&
 896             (__have_session(mdsc, req->r_resend_mds) ||
 897              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 898                 dout("choose_mds using resend_mds mds%d\n",
 899                      req->r_resend_mds);
 900                 return req->r_resend_mds;
 901         }
 902
 903         if (mode == USE_RANDOM_MDS)
 904                 goto random;
 905
 906         inode = NULL;
 907         if (req->r_inode) {
 908                 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
 909                         inode = req->r_inode;
 910                         ihold(inode);
 911                 } else {
 912                         /* req->r_dentry is non-null for LSSNAP request */
 913                         rcu_read_lock();
 914                         inode = get_nonsnap_parent(req->r_dentry);
 915                         rcu_read_unlock();
 916                         dout("__choose_mds using snapdir's parent %p\n", inode);
 917                 }
 918         } else if (req->r_dentry) {
 919                 /* ignore race with rename; old or new d_parent is okay */
 920                 struct dentry *parent;
 921                 struct inode *dir;
 922
 923                 rcu_read_lock();
 924                 parent = READ_ONCE(req->r_dentry->d_parent);
 925                 dir = req->r_parent ? : d_inode_rcu(parent);
 926
 927                 if (!dir || dir->i_sb != mdsc->fsc->sb) {
 928                         /*  not this fs or parent went negative */
 929                         inode = d_inode(req->r_dentry);
 930                         if (inode)
 931                                 ihold(inode);
 932                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
 933                         /* direct snapped/virtual snapdir requests
 934                          * based on parent dir inode */
 935                         inode = get_nonsnap_parent(parent);
 936                         dout("__choose_mds using nonsnap parent %p\n", inode);
 937                 } else {
 938                         /* dentry target */
 939                         inode = d_inode(req->r_dentry);
 940                         if (!inode || mode == USE_AUTH_MDS) {
 941                                 /* dir + name */
 942                                 inode = igrab(dir);
 943                                 hash = ceph_dentry_hash(dir, req->r_dentry);
 944                                 is_hash = true;
 945                         } else {
 946                                 ihold(inode);
 947                         }
 948                 }
 949                 rcu_read_unlock();
 950         }
 951
 952         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 953              (int)hash, mode);
 954         if (!inode)
 955                 goto random;
 956         ci = ceph_inode(inode);
 957
 958         if (is_hash && S_ISDIR(inode->i_mode)) {
 959                 struct ceph_inode_frag frag;
 960                 int found;
 961
 962                 ceph_choose_frag(ci, hash, &frag, &found);
 963                 if (found) {
 964                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 965                                 u8 r;
 966
 967                                 /* choose a random replica */
 968                                 get_random_bytes(&r, 1);
 969                                 r %= frag.ndist;
 970                                 mds = frag.dist[r];
 971                                 dout("choose_mds %p %llx.%llx "
 972                                      "frag %u mds%d (%d/%d)\n",
 973                                      inode, ceph_vinop(inode),
 974                                      frag.frag, mds,
 975                                      (int)r, frag.ndist);
 976                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 977                                     CEPH_MDS_STATE_ACTIVE)
 978                                         goto out;
 979                         }
 980
 981                         /* since this file/dir wasn't known to be
 982                          * replicated, then we want to look for the
 983                          * authoritative mds. */
 984                         mode = USE_AUTH_MDS;
 985                         if (frag.mds >= 0) {
 986                                 /* choose auth mds */
 987                                 mds = frag.mds;
 988                                 dout("choose_mds %p %llx.%llx "
 989                                      "frag %u mds%d (auth)\n",
 990                                      inode, ceph_vinop(inode), frag.frag, mds);
 991                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 992                                     CEPH_MDS_STATE_ACTIVE)
 993                                         goto out;
 994                         }
 995                 }
 996         }
 997
 998         spin_lock(&ci->i_ceph_lock);
 999         cap = NULL;
1000         if (mode == USE_AUTH_MDS)
1001                 cap = ci->i_auth_cap;
1002         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
1003                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
1004         if (!cap) {
1005                 spin_unlock(&ci->i_ceph_lock);
1006                 ceph_async_iput(inode);
1007                 goto random;
1008         }
1009         mds = cap->session->s_mds;
1010         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
1011              inode, ceph_vinop(inode), mds,
1012              cap == ci->i_auth_cap ? "auth " : "", cap);
1013         spin_unlock(&ci->i_ceph_lock);
1014 out:
1015         /* avoid calling iput_final() while holding mdsc->mutex or
1016          * in mds dispatch threads */
1017         ceph_async_iput(inode);
1018         return mds;
1019
1020 random:
1021         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
1022         dout("choose_mds chose random mds%d\n", mds);
1023         return mds;
1024 }
1025
1026
1027 /*
1028  * session messages
1029  */
1030 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
1031 {
1032         struct ceph_msg *msg;
1033         struct ceph_mds_session_head *h;
1034
1035         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
1036                            false);
1037         if (!msg) {
1038                 pr_err("create_session_msg ENOMEM creating msg\n");
1039                 return NULL;
1040         }
1041         h = msg->front.iov_base;
1042         h->op = cpu_to_le32(op);
1043         h->seq = cpu_to_le64(seq);
1044
1045         return msg;
1046 }
1047
1048 static void encode_supported_features(void **p, void *end)
1049 {
1050         static const unsigned char bits[] = CEPHFS_FEATURES_CLIENT_SUPPORTED;
1051         static const size_t count = ARRAY_SIZE(bits);
1052
1053         if (count > 0) {
1054                 size_t i;
1055                 size_t size = ((size_t)bits[count - 1] + 64) / 64 * 8;
1056
1057                 BUG_ON(*p + 4 + size > end);
1058                 ceph_encode_32(p, size);
1059                 memset(*p, 0, size);
1060                 for (i = 0; i < count; i++)
1061                         ((unsigned char*)(*p))[i / 8] |= 1 << (bits[i] % 8);
1062                 *p += size;
1063         } else {
1064                 BUG_ON(*p + 4 > end);
1065                 ceph_encode_32(p, 0);
1066         }
1067 }
1068
1069 /*
1070  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
1071  * to include additional client metadata fields.
1072  */
1073 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
1074 {
1075         struct ceph_msg *msg;
1076         struct ceph_mds_session_head *h;
1077         int i = -1;
1078         int extra_bytes = 0;
1079         int metadata_key_count = 0;
1080         struct ceph_options *opt = mdsc->fsc->client->options;
1081         struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
1082         void *p, *end;
1083
1084         const char* metadata[][2] = {
1085                 {"hostname", mdsc->nodename},
1086                 {"kernel_version", init_utsname()->release},
1087                 {"entity_id", opt->name ? : ""},
1088                 {"root", fsopt->server_path ? : "/"},
1089                 {NULL, NULL}
1090         };
1091
1092         /* Calculate serialized length of metadata */
1093         extra_bytes = 4;  /* map length */
1094         for (i = 0; metadata[i][0]; ++i) {
1095                 extra_bytes += 8 + strlen(metadata[i][0]) +
1096                         strlen(metadata[i][1]);
1097                 metadata_key_count++;
1098         }
1099         /* supported feature */
1100         extra_bytes += 4 + 8;
1101
1102         /* Allocate the message */
1103         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + extra_bytes,
1104                            GFP_NOFS, false);
1105         if (!msg) {
1106                 pr_err("create_session_msg ENOMEM creating msg\n");
1107                 return NULL;
1108         }
1109         p = msg->front.iov_base;
1110         end = p + msg->front.iov_len;
1111
1112         h = p;
1113         h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
1114         h->seq = cpu_to_le64(seq);
1115
1116         /*
1117          * Serialize client metadata into waiting buffer space, using
1118          * the format that userspace expects for map<string, string>
1119          *
1120          * ClientSession messages with metadata are v2
1121          */
1122         msg->hdr.version = cpu_to_le16(3);
1123         msg->hdr.compat_version = cpu_to_le16(1);
1124
1125         /* The write pointer, following the session_head structure */
1126         p += sizeof(*h);
1127
1128         /* Number of entries in the map */
1129         ceph_encode_32(&p, metadata_key_count);
1130
1131         /* Two length-prefixed strings for each entry in the map */
1132         for (i = 0; metadata[i][0]; ++i) {
1133                 size_t const key_len = strlen(metadata[i][0]);
1134                 size_t const val_len = strlen(metadata[i][1]);
1135
1136                 ceph_encode_32(&p, key_len);
1137                 memcpy(p, metadata[i][0], key_len);
1138                 p += key_len;
1139                 ceph_encode_32(&p, val_len);
1140                 memcpy(p, metadata[i][1], val_len);
1141                 p += val_len;
1142         }
1143
1144         encode_supported_features(&p, end);
1145         msg->front.iov_len = p - msg->front.iov_base;
1146         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1147
1148         return msg;
1149 }
1150
1151 /*
1152  * send session open request.
1153  *
1154  * called under mdsc->mutex
1155  */
1156 static int __open_session(struct ceph_mds_client *mdsc,
1157                           struct ceph_mds_session *session)
1158 {
1159         struct ceph_msg *msg;
1160         int mstate;
1161         int mds = session->s_mds;
1162
1163         /* wait for mds to go active? */
1164         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
1165         dout("open_session to mds%d (%s)\n", mds,
1166              ceph_mds_state_name(mstate));
1167         session->s_state = CEPH_MDS_SESSION_OPENING;
1168         session->s_renew_requested = jiffies;
1169
1170         /* send connect message */
1171         msg = create_session_open_msg(mdsc, session->s_seq);
1172         if (!msg)
1173                 return -ENOMEM;
1174         ceph_con_send(&session->s_con, msg);
1175         return 0;
1176 }
1177
1178 /*
1179  * open sessions for any export targets for the given mds
1180  *
1181  * called under mdsc->mutex
1182  */
1183 static struct ceph_mds_session *
1184 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
1185 {
1186         struct ceph_mds_session *session;
1187
1188         session = __ceph_lookup_mds_session(mdsc, target);
1189         if (!session) {
1190                 session = register_session(mdsc, target);
1191                 if (IS_ERR(session))
1192                         return session;
1193         }
1194         if (session->s_state == CEPH_MDS_SESSION_NEW ||
1195             session->s_state == CEPH_MDS_SESSION_CLOSING)
1196                 __open_session(mdsc, session);
1197
1198         return session;
1199 }
1200
1201 struct ceph_mds_session *
1202 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1203 {
1204         struct ceph_mds_session *session;
1205
1206         dout("open_export_target_session to mds%d\n", target);
1207
1208         mutex_lock(&mdsc->mutex);
1209         session = __open_export_target_session(mdsc, target);
1210         mutex_unlock(&mdsc->mutex);
1211
1212         return session;
1213 }
1214
1215 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1216                                           struct ceph_mds_session *session)
1217 {
1218         struct ceph_mds_info *mi;
1219         struct ceph_mds_session *ts;
1220         int i, mds = session->s_mds;
1221
1222         if (mds >= mdsc->mdsmap->m_num_mds)
1223                 return;
1224
1225         mi = &mdsc->mdsmap->m_info[mds];
1226         dout("open_export_target_sessions for mds%d (%d targets)\n",
1227              session->s_mds, mi->num_export_targets);
1228
1229         for (i = 0; i < mi->num_export_targets; i++) {
1230                 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1231                 if (!IS_ERR(ts))
1232                         ceph_put_mds_session(ts);
1233         }
1234 }
1235
1236 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1237                                            struct ceph_mds_session *session)
1238 {
1239         mutex_lock(&mdsc->mutex);
1240         __open_export_target_sessions(mdsc, session);
1241         mutex_unlock(&mdsc->mutex);
1242 }
1243
1244 /*
1245  * session caps
1246  */
1247
1248 static void detach_cap_releases(struct ceph_mds_session *session,
1249                                 struct list_head *target)
1250 {
1251         lockdep_assert_held(&session->s_cap_lock);
1252
1253         list_splice_init(&session->s_cap_releases, target);
1254         session->s_num_cap_releases = 0;
1255         dout("dispose_cap_releases mds%d\n", session->s_mds);
1256 }
1257
1258 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1259                                  struct list_head *dispose)
1260 {
1261         while (!list_empty(dispose)) {
1262                 struct ceph_cap *cap;
1263                 /* zero out the in-progress message */
1264                 cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1265                 list_del(&cap->session_caps);
1266                 ceph_put_cap(mdsc, cap);
1267         }
1268 }
1269
1270 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1271                                      struct ceph_mds_session *session)
1272 {
1273         struct ceph_mds_request *req;
1274         struct rb_node *p;
1275         struct ceph_inode_info *ci;
1276
1277         dout("cleanup_session_requests mds%d\n", session->s_mds);
1278         mutex_lock(&mdsc->mutex);
1279         while (!list_empty(&session->s_unsafe)) {
1280                 req = list_first_entry(&session->s_unsafe,
1281                                        struct ceph_mds_request, r_unsafe_item);
1282                 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1283                                     req->r_tid);
1284                 if (req->r_target_inode) {
1285                         /* dropping unsafe change of inode's attributes */
1286                         ci = ceph_inode(req->r_target_inode);
1287                         errseq_set(&ci->i_meta_err, -EIO);
1288                 }
1289                 if (req->r_unsafe_dir) {
1290                         /* dropping unsafe directory operation */
1291                         ci = ceph_inode(req->r_unsafe_dir);
1292                         errseq_set(&ci->i_meta_err, -EIO);
1293                 }
1294                 __unregister_request(mdsc, req);
1295         }
1296         /* zero r_attempts, so kick_requests() will re-send requests */
1297         p = rb_first(&mdsc->request_tree);
1298         while (p) {
1299                 req = rb_entry(p, struct ceph_mds_request, r_node);
1300                 p = rb_next(p);
1301                 if (req->r_session &&
1302                     req->r_session->s_mds == session->s_mds)
1303                         req->r_attempts = 0;
1304         }
1305         mutex_unlock(&mdsc->mutex);
1306 }
1307
1308 /*
1309  * Helper to safely iterate over all caps associated with a session, with
1310  * special care taken to handle a racing __ceph_remove_cap().
1311  *
1312  * Caller must hold session s_mutex.
1313  */
1314 int ceph_iterate_session_caps(struct ceph_mds_session *session,
1315                               int (*cb)(struct inode *, struct ceph_cap *,
1316                                         void *), void *arg)
1317 {
1318         struct list_head *p;
1319         struct ceph_cap *cap;
1320         struct inode *inode, *last_inode = NULL;
1321         struct ceph_cap *old_cap = NULL;
1322         int ret;
1323
1324         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1325         spin_lock(&session->s_cap_lock);
1326         p = session->s_caps.next;
1327         while (p != &session->s_caps) {
1328                 cap = list_entry(p, struct ceph_cap, session_caps);
1329                 inode = igrab(&cap->ci->vfs_inode);
1330                 if (!inode) {
1331                         p = p->next;
1332                         continue;
1333                 }
1334                 session->s_cap_iterator = cap;
1335                 spin_unlock(&session->s_cap_lock);
1336
1337                 if (last_inode) {
1338                         /* avoid calling iput_final() while holding
1339                          * s_mutex or in mds dispatch threads */
1340                         ceph_async_iput(last_inode);
1341                         last_inode = NULL;
1342                 }
1343                 if (old_cap) {
1344                         ceph_put_cap(session->s_mdsc, old_cap);
1345                         old_cap = NULL;
1346                 }
1347
1348                 ret = cb(inode, cap, arg);
1349                 last_inode = inode;
1350
1351                 spin_lock(&session->s_cap_lock);
1352                 p = p->next;
1353                 if (!cap->ci) {
1354                         dout("iterate_session_caps  finishing cap %p removal\n",
1355                              cap);
1356                         BUG_ON(cap->session != session);
1357                         cap->session = NULL;
1358                         list_del_init(&cap->session_caps);
1359                         session->s_nr_caps--;
1360                         if (cap->queue_release)
1361                                 __ceph_queue_cap_release(session, cap);
1362                         else
1363                                 old_cap = cap;  /* put_cap it w/o locks held */
1364                 }
1365                 if (ret < 0)
1366                         goto out;
1367         }
1368         ret = 0;
1369 out:
1370         session->s_cap_iterator = NULL;
1371         spin_unlock(&session->s_cap_lock);
1372
1373         ceph_async_iput(last_inode);
1374         if (old_cap)
1375                 ceph_put_cap(session->s_mdsc, old_cap);
1376
1377         return ret;
1378 }
1379
1380 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1381                                   void *arg)
1382 {
1383         struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1384         struct ceph_inode_info *ci = ceph_inode(inode);
1385         LIST_HEAD(to_remove);
1386         bool dirty_dropped = false;
1387         bool invalidate = false;
1388
1389         dout("removing cap %p, ci is %p, inode is %p\n",
1390              cap, ci, &ci->vfs_inode);
1391         spin_lock(&ci->i_ceph_lock);
1392         if (cap->mds_wanted | cap->issued)
1393                 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1394         __ceph_remove_cap(cap, false);
1395         if (!ci->i_auth_cap) {
1396                 struct ceph_cap_flush *cf;
1397                 struct ceph_mds_client *mdsc = fsc->mdsc;
1398
1399                 if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1400                         if (inode->i_data.nrpages > 0)
1401                                 invalidate = true;
1402                         if (ci->i_wrbuffer_ref > 0)
1403                                 mapping_set_error(&inode->i_data, -EIO);
1404                 }
1405
1406                 while (!list_empty(&ci->i_cap_flush_list)) {
1407                         cf = list_first_entry(&ci->i_cap_flush_list,
1408                                               struct ceph_cap_flush, i_list);
1409                         list_move(&cf->i_list, &to_remove);
1410                 }
1411
1412                 spin_lock(&mdsc->cap_dirty_lock);
1413
1414                 list_for_each_entry(cf, &to_remove, i_list)
1415                         list_del(&cf->g_list);
1416
1417                 if (!list_empty(&ci->i_dirty_item)) {
1418                         pr_warn_ratelimited(
1419                                 " dropping dirty %s state for %p %lld\n",
1420                                 ceph_cap_string(ci->i_dirty_caps),
1421                                 inode, ceph_ino(inode));
1422                         ci->i_dirty_caps = 0;
1423                         list_del_init(&ci->i_dirty_item);
1424                         dirty_dropped = true;
1425                 }
1426                 if (!list_empty(&ci->i_flushing_item)) {
1427                         pr_warn_ratelimited(
1428                                 " dropping dirty+flushing %s state for %p %lld\n",
1429                                 ceph_cap_string(ci->i_flushing_caps),
1430                                 inode, ceph_ino(inode));
1431                         ci->i_flushing_caps = 0;
1432                         list_del_init(&ci->i_flushing_item);
1433                         mdsc->num_cap_flushing--;
1434                         dirty_dropped = true;
1435                 }
1436                 spin_unlock(&mdsc->cap_dirty_lock);
1437
1438                 if (dirty_dropped) {
1439                         errseq_set(&ci->i_meta_err, -EIO);
1440
1441                         if (ci->i_wrbuffer_ref_head == 0 &&
1442                             ci->i_wr_ref == 0 &&
1443                             ci->i_dirty_caps == 0 &&
1444                             ci->i_flushing_caps == 0) {
1445                                 ceph_put_snap_context(ci->i_head_snapc);
1446                                 ci->i_head_snapc = NULL;
1447                         }
1448                 }
1449
1450                 if (atomic_read(&ci->i_filelock_ref) > 0) {
1451                         /* make further file lock syscall return -EIO */
1452                         ci->i_ceph_flags |= CEPH_I_ERROR_FILELOCK;
1453                         pr_warn_ratelimited(" dropping file locks for %p %lld\n",
1454                                             inode, ceph_ino(inode));
1455                 }
1456
1457                 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1458                         list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1459                         ci->i_prealloc_cap_flush = NULL;
1460                 }
1461         }
1462         spin_unlock(&ci->i_ceph_lock);
1463         while (!list_empty(&to_remove)) {
1464                 struct ceph_cap_flush *cf;
1465                 cf = list_first_entry(&to_remove,
1466                                       struct ceph_cap_flush, i_list);
1467                 list_del(&cf->i_list);
1468                 ceph_free_cap_flush(cf);
1469         }
1470
1471         wake_up_all(&ci->i_cap_wq);
1472         if (invalidate)
1473                 ceph_queue_invalidate(inode);
1474         if (dirty_dropped)
1475                 iput(inode);
1476         return 0;
1477 }
1478
1479 /*
1480  * caller must hold session s_mutex
1481  */
1482 static void remove_session_caps(struct ceph_mds_session *session)
1483 {
1484         struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1485         struct super_block *sb = fsc->sb;
1486         LIST_HEAD(dispose);
1487
1488         dout("remove_session_caps on %p\n", session);
1489         ceph_iterate_session_caps(session, remove_session_caps_cb, fsc);
1490
1491         wake_up_all(&fsc->mdsc->cap_flushing_wq);
1492
1493         spin_lock(&session->s_cap_lock);
1494         if (session->s_nr_caps > 0) {
1495                 struct inode *inode;
1496                 struct ceph_cap *cap, *prev = NULL;
1497                 struct ceph_vino vino;
1498                 /*
1499                  * iterate_session_caps() skips inodes that are being
1500                  * deleted, we need to wait until deletions are complete.
1501                  * __wait_on_freeing_inode() is designed for the job,
1502                  * but it is not exported, so use lookup inode function
1503                  * to access it.
1504                  */
1505                 while (!list_empty(&session->s_caps)) {
1506                         cap = list_entry(session->s_caps.next,
1507                                          struct ceph_cap, session_caps);
1508                         if (cap == prev)
1509                                 break;
1510                         prev = cap;
1511                         vino = cap->ci->i_vino;
1512                         spin_unlock(&session->s_cap_lock);
1513
1514                         inode = ceph_find_inode(sb, vino);
1515                          /* avoid calling iput_final() while holding s_mutex */
1516                         ceph_async_iput(inode);
1517
1518                         spin_lock(&session->s_cap_lock);
1519                 }
1520         }
1521
1522         // drop cap expires and unlock s_cap_lock
1523         detach_cap_releases(session, &dispose);
1524
1525         BUG_ON(session->s_nr_caps > 0);
1526         BUG_ON(!list_empty(&session->s_cap_flushing));
1527         spin_unlock(&session->s_cap_lock);
1528         dispose_cap_releases(session->s_mdsc, &dispose);
1529 }
1530
1531 enum {
1532         RECONNECT,
1533         RENEWCAPS,
1534         FORCE_RO,
1535 };
1536
1537 /*
1538  * wake up any threads waiting on this session's caps.  if the cap is
1539  * old (didn't get renewed on the client reconnect), remove it now.
1540  *
1541  * caller must hold s_mutex.
1542  */
1543 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1544                               void *arg)
1545 {
1546         struct ceph_inode_info *ci = ceph_inode(inode);
1547         unsigned long ev = (unsigned long)arg;
1548
1549         if (ev == RECONNECT) {
1550                 spin_lock(&ci->i_ceph_lock);
1551                 ci->i_wanted_max_size = 0;
1552                 ci->i_requested_max_size = 0;
1553                 spin_unlock(&ci->i_ceph_lock);
1554         } else if (ev == RENEWCAPS) {
1555                 if (cap->cap_gen < cap->session->s_cap_gen) {
1556                         /* mds did not re-issue stale cap */
1557                         spin_lock(&ci->i_ceph_lock);
1558                         cap->issued = cap->implemented = CEPH_CAP_PIN;
1559                         /* make sure mds knows what we want */
1560                         if (__ceph_caps_file_wanted(ci) & ~cap->mds_wanted)
1561                                 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1562                         spin_unlock(&ci->i_ceph_lock);
1563                 }
1564         } else if (ev == FORCE_RO) {
1565         }
1566         wake_up_all(&ci->i_cap_wq);
1567         return 0;
1568 }
1569
1570 static void wake_up_session_caps(struct ceph_mds_session *session, int ev)
1571 {
1572         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1573         ceph_iterate_session_caps(session, wake_up_session_cb,
1574                                   (void *)(unsigned long)ev);
1575 }
1576
1577 /*
1578  * Send periodic message to MDS renewing all currently held caps.  The
1579  * ack will reset the expiration for all caps from this session.
1580  *
1581  * caller holds s_mutex
1582  */
1583 static int send_renew_caps(struct ceph_mds_client *mdsc,
1584                            struct ceph_mds_session *session)
1585 {
1586         struct ceph_msg *msg;
1587         int state;
1588
1589         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1590             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1591                 pr_info("mds%d caps stale\n", session->s_mds);
1592         session->s_renew_requested = jiffies;
1593
1594         /* do not try to renew caps until a recovering mds has reconnected
1595          * with its clients. */
1596         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1597         if (state < CEPH_MDS_STATE_RECONNECT) {
1598                 dout("send_renew_caps ignoring mds%d (%s)\n",
1599                      session->s_mds, ceph_mds_state_name(state));
1600                 return 0;
1601         }
1602
1603         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1604                 ceph_mds_state_name(state));
1605         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1606                                  ++session->s_renew_seq);
1607         if (!msg)
1608                 return -ENOMEM;
1609         ceph_con_send(&session->s_con, msg);
1610         return 0;
1611 }
1612
1613 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1614                              struct ceph_mds_session *session, u64 seq)
1615 {
1616         struct ceph_msg *msg;
1617
1618         dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1619              session->s_mds, ceph_session_state_name(session->s_state), seq);
1620         msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1621         if (!msg)
1622                 return -ENOMEM;
1623         ceph_con_send(&session->s_con, msg);
1624         return 0;
1625 }
1626
1627
1628 /*
1629  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1630  *
1631  * Called under session->s_mutex
1632  */
1633 static void renewed_caps(struct ceph_mds_client *mdsc,
1634                          struct ceph_mds_session *session, int is_renew)
1635 {
1636         int was_stale;
1637         int wake = 0;
1638
1639         spin_lock(&session->s_cap_lock);
1640         was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1641
1642         session->s_cap_ttl = session->s_renew_requested +
1643                 mdsc->mdsmap->m_session_timeout*HZ;
1644
1645         if (was_stale) {
1646                 if (time_before(jiffies, session->s_cap_ttl)) {
1647                         pr_info("mds%d caps renewed\n", session->s_mds);
1648                         wake = 1;
1649                 } else {
1650                         pr_info("mds%d caps still stale\n", session->s_mds);
1651                 }
1652         }
1653         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1654              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1655              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1656         spin_unlock(&session->s_cap_lock);
1657
1658         if (wake)
1659                 wake_up_session_caps(session, RENEWCAPS);
1660 }
1661
1662 /*
1663  * send a session close request
1664  */
1665 static int request_close_session(struct ceph_mds_client *mdsc,
1666                                  struct ceph_mds_session *session)
1667 {
1668         struct ceph_msg *msg;
1669
1670         dout("request_close_session mds%d state %s seq %lld\n",
1671              session->s_mds, ceph_session_state_name(session->s_state),
1672              session->s_seq);
1673         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1674         if (!msg)
1675                 return -ENOMEM;
1676         ceph_con_send(&session->s_con, msg);
1677         return 1;
1678 }
1679
1680 /*
1681  * Called with s_mutex held.
1682  */
1683 static int __close_session(struct ceph_mds_client *mdsc,
1684                          struct ceph_mds_session *session)
1685 {
1686         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1687                 return 0;
1688         session->s_state = CEPH_MDS_SESSION_CLOSING;
1689         return request_close_session(mdsc, session);
1690 }
1691
1692 static bool drop_negative_children(struct dentry *dentry)
1693 {
1694         struct dentry *child;
1695         bool all_negative = true;
1696
1697         if (!d_is_dir(dentry))
1698                 goto out;
1699
1700         spin_lock(&dentry->d_lock);
1701         list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1702                 if (d_really_is_positive(child)) {
1703                         all_negative = false;
1704                         break;
1705                 }
1706         }
1707         spin_unlock(&dentry->d_lock);
1708
1709         if (all_negative)
1710                 shrink_dcache_parent(dentry);
1711 out:
1712         return all_negative;
1713 }
1714
1715 /*
1716  * Trim old(er) caps.
1717  *
1718  * Because we can't cache an inode without one or more caps, we do
1719  * this indirectly: if a cap is unused, we prune its aliases, at which
1720  * point the inode will hopefully get dropped to.
1721  *
1722  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1723  * memory pressure from the MDS, though, so it needn't be perfect.
1724  */
1725 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1726 {
1727         int *remaining = arg;
1728         struct ceph_inode_info *ci = ceph_inode(inode);
1729         int used, wanted, oissued, mine;
1730
1731         if (*remaining <= 0)
1732                 return -1;
1733
1734         spin_lock(&ci->i_ceph_lock);
1735         mine = cap->issued | cap->implemented;
1736         used = __ceph_caps_used(ci);
1737         wanted = __ceph_caps_file_wanted(ci);
1738         oissued = __ceph_caps_issued_other(ci, cap);
1739
1740         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1741              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1742              ceph_cap_string(used), ceph_cap_string(wanted));
1743         if (cap == ci->i_auth_cap) {
1744                 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1745                     !list_empty(&ci->i_cap_snaps))
1746                         goto out;
1747                 if ((used | wanted) & CEPH_CAP_ANY_WR)
1748                         goto out;
1749                 /* Note: it's possible that i_filelock_ref becomes non-zero
1750                  * after dropping auth caps. It doesn't hurt because reply
1751                  * of lock mds request will re-add auth caps. */
1752                 if (atomic_read(&ci->i_filelock_ref) > 0)
1753                         goto out;
1754         }
1755         /* The inode has cached pages, but it's no longer used.
1756          * we can safely drop it */
1757         if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1758             !(oissued & CEPH_CAP_FILE_CACHE)) {
1759           used = 0;
1760           oissued = 0;
1761         }
1762         if ((used | wanted) & ~oissued & mine)
1763                 goto out;   /* we need these caps */
1764
1765         if (oissued) {
1766                 /* we aren't the only cap.. just remove us */
1767                 __ceph_remove_cap(cap, true);
1768                 (*remaining)--;
1769         } else {
1770                 struct dentry *dentry;
1771                 /* try dropping referring dentries */
1772                 spin_unlock(&ci->i_ceph_lock);
1773                 dentry = d_find_any_alias(inode);
1774                 if (dentry && drop_negative_children(dentry)) {
1775                         int count;
1776                         dput(dentry);
1777                         d_prune_aliases(inode);
1778                         count = atomic_read(&inode->i_count);
1779                         if (count == 1)
1780                                 (*remaining)--;
1781                         dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1782                              inode, cap, count);
1783                 } else {
1784                         dput(dentry);
1785                 }
1786                 return 0;
1787         }
1788
1789 out:
1790         spin_unlock(&ci->i_ceph_lock);
1791         return 0;
1792 }
1793
1794 /*
1795  * Trim session cap count down to some max number.
1796  */
1797 int ceph_trim_caps(struct ceph_mds_client *mdsc,
1798                    struct ceph_mds_session *session,
1799                    int max_caps)
1800 {
1801         int trim_caps = session->s_nr_caps - max_caps;
1802
1803         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1804              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1805         if (trim_caps > 0) {
1806                 int remaining = trim_caps;
1807
1808                 ceph_iterate_session_caps(session, trim_caps_cb, &remaining);
1809                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1810                      session->s_mds, session->s_nr_caps, max_caps,
1811                         trim_caps - remaining);
1812         }
1813
1814         ceph_flush_cap_releases(mdsc, session);
1815         return 0;
1816 }
1817
1818 static int check_caps_flush(struct ceph_mds_client *mdsc,
1819                             u64 want_flush_tid)
1820 {
1821         int ret = 1;
1822
1823         spin_lock(&mdsc->cap_dirty_lock);
1824         if (!list_empty(&mdsc->cap_flush_list)) {
1825                 struct ceph_cap_flush *cf =
1826                         list_first_entry(&mdsc->cap_flush_list,
1827                                          struct ceph_cap_flush, g_list);
1828                 if (cf->tid <= want_flush_tid) {
1829                         dout("check_caps_flush still flushing tid "
1830                              "%llu <= %llu\n", cf->tid, want_flush_tid);
1831                         ret = 0;
1832                 }
1833         }
1834         spin_unlock(&mdsc->cap_dirty_lock);
1835         return ret;
1836 }
1837
1838 /*
1839  * flush all dirty inode data to disk.
1840  *
1841  * returns true if we've flushed through want_flush_tid
1842  */
1843 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1844                             u64 want_flush_tid)
1845 {
1846         dout("check_caps_flush want %llu\n", want_flush_tid);
1847
1848         wait_event(mdsc->cap_flushing_wq,
1849                    check_caps_flush(mdsc, want_flush_tid));
1850
1851         dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1852 }
1853
1854 /*
1855  * called under s_mutex
1856  */
1857 static void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1858                                    struct ceph_mds_session *session)
1859 {
1860         struct ceph_msg *msg = NULL;
1861         struct ceph_mds_cap_release *head;
1862         struct ceph_mds_cap_item *item;
1863         struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
1864         struct ceph_cap *cap;
1865         LIST_HEAD(tmp_list);
1866         int num_cap_releases;
1867         __le32  barrier, *cap_barrier;
1868
1869         down_read(&osdc->lock);
1870         barrier = cpu_to_le32(osdc->epoch_barrier);
1871         up_read(&osdc->lock);
1872
1873         spin_lock(&session->s_cap_lock);
1874 again:
1875         list_splice_init(&session->s_cap_releases, &tmp_list);
1876         num_cap_releases = session->s_num_cap_releases;
1877         session->s_num_cap_releases = 0;
1878         spin_unlock(&session->s_cap_lock);
1879
1880         while (!list_empty(&tmp_list)) {
1881                 if (!msg) {
1882                         msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1883                                         PAGE_SIZE, GFP_NOFS, false);
1884                         if (!msg)
1885                                 goto out_err;
1886                         head = msg->front.iov_base;
1887                         head->num = cpu_to_le32(0);
1888                         msg->front.iov_len = sizeof(*head);
1889
1890                         msg->hdr.version = cpu_to_le16(2);
1891                         msg->hdr.compat_version = cpu_to_le16(1);
1892                 }
1893
1894                 cap = list_first_entry(&tmp_list, struct ceph_cap,
1895                                         session_caps);
1896                 list_del(&cap->session_caps);
1897                 num_cap_releases--;
1898
1899                 head = msg->front.iov_base;
1900                 put_unaligned_le32(get_unaligned_le32(&head->num) + 1,
1901                                    &head->num);
1902                 item = msg->front.iov_base + msg->front.iov_len;
1903                 item->ino = cpu_to_le64(cap->cap_ino);
1904                 item->cap_id = cpu_to_le64(cap->cap_id);
1905                 item->migrate_seq = cpu_to_le32(cap->mseq);
1906                 item->seq = cpu_to_le32(cap->issue_seq);
1907                 msg->front.iov_len += sizeof(*item);
1908
1909                 ceph_put_cap(mdsc, cap);
1910
1911                 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1912                         // Append cap_barrier field
1913                         cap_barrier = msg->front.iov_base + msg->front.iov_len;
1914                         *cap_barrier = barrier;
1915                         msg->front.iov_len += sizeof(*cap_barrier);
1916
1917                         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1918                         dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1919                         ceph_con_send(&session->s_con, msg);
1920                         msg = NULL;
1921                 }
1922         }
1923
1924         BUG_ON(num_cap_releases != 0);
1925
1926         spin_lock(&session->s_cap_lock);
1927         if (!list_empty(&session->s_cap_releases))
1928                 goto again;
1929         spin_unlock(&session->s_cap_lock);
1930
1931         if (msg) {
1932                 // Append cap_barrier field
1933                 cap_barrier = msg->front.iov_base + msg->front.iov_len;
1934                 *cap_barrier = barrier;
1935                 msg->front.iov_len += sizeof(*cap_barrier);
1936
1937                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1938                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1939                 ceph_con_send(&session->s_con, msg);
1940         }
1941         return;
1942 out_err:
1943         pr_err("send_cap_releases mds%d, failed to allocate message\n",
1944                 session->s_mds);
1945         spin_lock(&session->s_cap_lock);
1946         list_splice(&tmp_list, &session->s_cap_releases);
1947         session->s_num_cap_releases += num_cap_releases;
1948         spin_unlock(&session->s_cap_lock);
1949 }
1950
1951 static void ceph_cap_release_work(struct work_struct *work)
1952 {
1953         struct ceph_mds_session *session =
1954                 container_of(work, struct ceph_mds_session, s_cap_release_work);
1955
1956         mutex_lock(&session->s_mutex);
1957         if (session->s_state == CEPH_MDS_SESSION_OPEN ||
1958             session->s_state == CEPH_MDS_SESSION_HUNG)
1959                 ceph_send_cap_releases(session->s_mdsc, session);
1960         mutex_unlock(&session->s_mutex);
1961         ceph_put_mds_session(session);
1962 }
1963
1964 void ceph_flush_cap_releases(struct ceph_mds_client *mdsc,
1965                              struct ceph_mds_session *session)
1966 {
1967         if (mdsc->stopping)
1968                 return;
1969
1970         get_session(session);
1971         if (queue_work(mdsc->fsc->cap_wq,
1972                        &session->s_cap_release_work)) {
1973                 dout("cap release work queued\n");
1974         } else {
1975                 ceph_put_mds_session(session);
1976                 dout("failed to queue cap release work\n");
1977         }
1978 }
1979
1980 /*
1981  * caller holds session->s_cap_lock
1982  */
1983 void __ceph_queue_cap_release(struct ceph_mds_session *session,
1984                               struct ceph_cap *cap)
1985 {
1986         list_add_tail(&cap->session_caps, &session->s_cap_releases);
1987         session->s_num_cap_releases++;
1988
1989         if (!(session->s_num_cap_releases % CEPH_CAPS_PER_RELEASE))
1990                 ceph_flush_cap_releases(session->s_mdsc, session);
1991 }
1992
1993 static void ceph_cap_reclaim_work(struct work_struct *work)
1994 {
1995         struct ceph_mds_client *mdsc =
1996                 container_of(work, struct ceph_mds_client, cap_reclaim_work);
1997         int ret = ceph_trim_dentries(mdsc);
1998         if (ret == -EAGAIN)
1999                 ceph_queue_cap_reclaim_work(mdsc);
2000 }
2001
2002 void ceph_queue_cap_reclaim_work(struct ceph_mds_client *mdsc)
2003 {
2004         if (mdsc->stopping)
2005                 return;
2006
2007         if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_reclaim_work)) {
2008                 dout("caps reclaim work queued\n");
2009         } else {
2010                 dout("failed to queue caps release work\n");
2011         }
2012 }
2013
2014 void ceph_reclaim_caps_nr(struct ceph_mds_client *mdsc, int nr)
2015 {
2016         int val;
2017         if (!nr)
2018                 return;
2019         val = atomic_add_return(nr, &mdsc->cap_reclaim_pending);
2020         if ((val % CEPH_CAPS_PER_RELEASE) < nr) {
2021                 atomic_set(&mdsc->cap_reclaim_pending, 0);
2022                 ceph_queue_cap_reclaim_work(mdsc);
2023         }
2024 }
2025
2026 /*
2027  * requests
2028  */
2029
2030 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
2031                                     struct inode *dir)
2032 {
2033         struct ceph_inode_info *ci = ceph_inode(dir);
2034         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
2035         struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
2036         size_t size = sizeof(struct ceph_mds_reply_dir_entry);
2037         unsigned int num_entries;
2038         int order;
2039
2040         spin_lock(&ci->i_ceph_lock);
2041         num_entries = ci->i_files + ci->i_subdirs;
2042         spin_unlock(&ci->i_ceph_lock);
2043         num_entries = max(num_entries, 1U);
2044         num_entries = min(num_entries, opt->max_readdir);
2045
2046         order = get_order(size * num_entries);
2047         while (order >= 0) {
2048                 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
2049                                                              __GFP_NOWARN,
2050                                                              order);
2051                 if (rinfo->dir_entries)
2052                         break;
2053                 order--;
2054         }
2055         if (!rinfo->dir_entries)
2056                 return -ENOMEM;
2057
2058         num_entries = (PAGE_SIZE << order) / size;
2059         num_entries = min(num_entries, opt->max_readdir);
2060
2061         rinfo->dir_buf_size = PAGE_SIZE << order;
2062         req->r_num_caps = num_entries + 1;
2063         req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
2064         req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
2065         return 0;
2066 }
2067
2068 /*
2069  * Create an mds request.
2070  */
2071 struct ceph_mds_request *
2072 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
2073 {
2074         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
2075
2076         if (!req)
2077                 return ERR_PTR(-ENOMEM);
2078
2079         mutex_init(&req->r_fill_mutex);
2080         req->r_mdsc = mdsc;
2081         req->r_started = jiffies;
2082         req->r_resend_mds = -1;
2083         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
2084         INIT_LIST_HEAD(&req->r_unsafe_target_item);
2085         req->r_fmode = -1;
2086         kref_init(&req->r_kref);
2087         RB_CLEAR_NODE(&req->r_node);
2088         INIT_LIST_HEAD(&req->r_wait);
2089         init_completion(&req->r_completion);
2090         init_completion(&req->r_safe_completion);
2091         INIT_LIST_HEAD(&req->r_unsafe_item);
2092
2093         ktime_get_coarse_real_ts64(&req->r_stamp);
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                 ihold(req->r_parent);
2682         }
2683         if (req->r_old_dentry_dir)
2684                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2685                                   CEPH_CAP_PIN);
2686
2687         dout("submit_request on %p for inode %p\n", req, dir);
2688         mutex_lock(&mdsc->mutex);
2689         __register_request(mdsc, req, dir);
2690         __do_request(mdsc, req);
2691         err = req->r_err;
2692         mutex_unlock(&mdsc->mutex);
2693         return err;
2694 }
2695
2696 static int ceph_mdsc_wait_request(struct ceph_mds_client *mdsc,
2697                                   struct ceph_mds_request *req)
2698 {
2699         int err;
2700
2701         /* wait */
2702         dout("do_request waiting\n");
2703         if (!req->r_timeout && req->r_wait_for_completion) {
2704                 err = req->r_wait_for_completion(mdsc, req);
2705         } else {
2706                 long timeleft = wait_for_completion_killable_timeout(
2707                                         &req->r_completion,
2708                                         ceph_timeout_jiffies(req->r_timeout));
2709                 if (timeleft > 0)
2710                         err = 0;
2711                 else if (!timeleft)
2712                         err = -EIO;  /* timed out */
2713                 else
2714                         err = timeleft;  /* killed */
2715         }
2716         dout("do_request waited, got %d\n", err);
2717         mutex_lock(&mdsc->mutex);
2718
2719         /* only abort if we didn't race with a real reply */
2720         if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2721                 err = le32_to_cpu(req->r_reply_info.head->result);
2722         } else if (err < 0) {
2723                 dout("aborted request %lld with %d\n", req->r_tid, err);
2724
2725                 /*
2726                  * ensure we aren't running concurrently with
2727                  * ceph_fill_trace or ceph_readdir_prepopulate, which
2728                  * rely on locks (dir mutex) held by our caller.
2729                  */
2730                 mutex_lock(&req->r_fill_mutex);
2731                 req->r_err = err;
2732                 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2733                 mutex_unlock(&req->r_fill_mutex);
2734
2735                 if (req->r_parent &&
2736                     (req->r_op & CEPH_MDS_OP_WRITE))
2737                         ceph_invalidate_dir_request(req);
2738         } else {
2739                 err = req->r_err;
2740         }
2741
2742         mutex_unlock(&mdsc->mutex);
2743         return err;
2744 }
2745
2746 /*
2747  * Synchrously perform an mds request.  Take care of all of the
2748  * session setup, forwarding, retry details.
2749  */
2750 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2751                          struct inode *dir,
2752                          struct ceph_mds_request *req)
2753 {
2754         int err;
2755
2756         dout("do_request on %p\n", req);
2757
2758         /* issue */
2759         err = ceph_mdsc_submit_request(mdsc, dir, req);
2760         if (!err)
2761                 err = ceph_mdsc_wait_request(mdsc, req);
2762         dout("do_request %p done, result %d\n", req, err);
2763         return err;
2764 }
2765
2766 /*
2767  * Invalidate dir's completeness, dentry lease state on an aborted MDS
2768  * namespace request.
2769  */
2770 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2771 {
2772         struct inode *dir = req->r_parent;
2773         struct inode *old_dir = req->r_old_dentry_dir;
2774
2775         dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
2776
2777         ceph_dir_clear_complete(dir);
2778         if (old_dir)
2779                 ceph_dir_clear_complete(old_dir);
2780         if (req->r_dentry)
2781                 ceph_invalidate_dentry_lease(req->r_dentry);
2782         if (req->r_old_dentry)
2783                 ceph_invalidate_dentry_lease(req->r_old_dentry);
2784 }
2785
2786 /*
2787  * Handle mds reply.
2788  *
2789  * We take the session mutex and parse and process the reply immediately.
2790  * This preserves the logical ordering of replies, capabilities, etc., sent
2791  * by the MDS as they are applied to our local cache.
2792  */
2793 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2794 {
2795         struct ceph_mds_client *mdsc = session->s_mdsc;
2796         struct ceph_mds_request *req;
2797         struct ceph_mds_reply_head *head = msg->front.iov_base;
2798         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2799         struct ceph_snap_realm *realm;
2800         u64 tid;
2801         int err, result;
2802         int mds = session->s_mds;
2803
2804         if (msg->front.iov_len < sizeof(*head)) {
2805                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2806                 ceph_msg_dump(msg);
2807                 return;
2808         }
2809
2810         /* get request, session */
2811         tid = le64_to_cpu(msg->hdr.tid);
2812         mutex_lock(&mdsc->mutex);
2813         req = lookup_get_request(mdsc, tid);
2814         if (!req) {
2815                 dout("handle_reply on unknown tid %llu\n", tid);
2816                 mutex_unlock(&mdsc->mutex);
2817                 return;
2818         }
2819         dout("handle_reply %p\n", req);
2820
2821         /* correct session? */
2822         if (req->r_session != session) {
2823                 pr_err("mdsc_handle_reply got %llu on session mds%d"
2824                        " not mds%d\n", tid, session->s_mds,
2825                        req->r_session ? req->r_session->s_mds : -1);
2826                 mutex_unlock(&mdsc->mutex);
2827                 goto out;
2828         }
2829
2830         /* dup? */
2831         if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
2832             (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
2833                 pr_warn("got a dup %s reply on %llu from mds%d\n",
2834                            head->safe ? "safe" : "unsafe", tid, mds);
2835                 mutex_unlock(&mdsc->mutex);
2836                 goto out;
2837         }
2838         if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
2839                 pr_warn("got unsafe after safe on %llu from mds%d\n",
2840                            tid, mds);
2841                 mutex_unlock(&mdsc->mutex);
2842                 goto out;
2843         }
2844
2845         result = le32_to_cpu(head->result);
2846
2847         /*
2848          * Handle an ESTALE
2849          * if we're not talking to the authority, send to them
2850          * if the authority has changed while we weren't looking,
2851          * send to new authority
2852          * Otherwise we just have to return an ESTALE
2853          */
2854         if (result == -ESTALE) {
2855                 dout("got ESTALE on request %llu\n", req->r_tid);
2856                 req->r_resend_mds = -1;
2857                 if (req->r_direct_mode != USE_AUTH_MDS) {
2858                         dout("not using auth, setting for that now\n");
2859                         req->r_direct_mode = USE_AUTH_MDS;
2860                         __do_request(mdsc, req);
2861                         mutex_unlock(&mdsc->mutex);
2862                         goto out;
2863                 } else  {
2864                         int mds = __choose_mds(mdsc, req);
2865                         if (mds >= 0 && mds != req->r_session->s_mds) {
2866                                 dout("but auth changed, so resending\n");
2867                                 __do_request(mdsc, req);
2868                                 mutex_unlock(&mdsc->mutex);
2869                                 goto out;
2870                         }
2871                 }
2872                 dout("have to return ESTALE on request %llu\n", req->r_tid);
2873         }
2874
2875
2876         if (head->safe) {
2877                 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
2878                 __unregister_request(mdsc, req);
2879
2880                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2881                         /*
2882                          * We already handled the unsafe response, now do the
2883                          * cleanup.  No need to examine the response; the MDS
2884                          * doesn't include any result info in the safe
2885                          * response.  And even if it did, there is nothing
2886                          * useful we could do with a revised return value.
2887                          */
2888                         dout("got safe reply %llu, mds%d\n", tid, mds);
2889
2890                         /* last unsafe request during umount? */
2891                         if (mdsc->stopping && !__get_oldest_req(mdsc))
2892                                 complete_all(&mdsc->safe_umount_waiters);
2893                         mutex_unlock(&mdsc->mutex);
2894                         goto out;
2895                 }
2896         } else {
2897                 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
2898                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2899                 if (req->r_unsafe_dir) {
2900                         struct ceph_inode_info *ci =
2901                                         ceph_inode(req->r_unsafe_dir);
2902                         spin_lock(&ci->i_unsafe_lock);
2903                         list_add_tail(&req->r_unsafe_dir_item,
2904                                       &ci->i_unsafe_dirops);
2905                         spin_unlock(&ci->i_unsafe_lock);
2906                 }
2907         }
2908
2909         dout("handle_reply tid %lld result %d\n", tid, result);
2910         rinfo = &req->r_reply_info;
2911         if (test_bit(CEPHFS_FEATURE_REPLY_ENCODING, &session->s_features))
2912                 err = parse_reply_info(msg, rinfo, (u64)-1);
2913         else
2914                 err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2915         mutex_unlock(&mdsc->mutex);
2916
2917         mutex_lock(&session->s_mutex);
2918         if (err < 0) {
2919                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2920                 ceph_msg_dump(msg);
2921                 goto out_err;
2922         }
2923
2924         /* snap trace */
2925         realm = NULL;
2926         if (rinfo->snapblob_len) {
2927                 down_write(&mdsc->snap_rwsem);
2928                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2929                                 rinfo->snapblob + rinfo->snapblob_len,
2930                                 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2931                                 &realm);
2932                 downgrade_write(&mdsc->snap_rwsem);
2933         } else {
2934                 down_read(&mdsc->snap_rwsem);
2935         }
2936
2937         /* insert trace into our cache */
2938         mutex_lock(&req->r_fill_mutex);
2939         current->journal_info = req;
2940         err = ceph_fill_trace(mdsc->fsc->sb, req);
2941         if (err == 0) {
2942                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2943                                     req->r_op == CEPH_MDS_OP_LSSNAP))
2944                         ceph_readdir_prepopulate(req, req->r_session);
2945         }
2946         current->journal_info = NULL;
2947         mutex_unlock(&req->r_fill_mutex);
2948
2949         up_read(&mdsc->snap_rwsem);
2950         if (realm)
2951                 ceph_put_snap_realm(mdsc, realm);
2952
2953         if (err == 0) {
2954                 if (req->r_target_inode &&
2955                     test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2956                         struct ceph_inode_info *ci =
2957                                 ceph_inode(req->r_target_inode);
2958                         spin_lock(&ci->i_unsafe_lock);
2959                         list_add_tail(&req->r_unsafe_target_item,
2960                                       &ci->i_unsafe_iops);
2961                         spin_unlock(&ci->i_unsafe_lock);
2962                 }
2963
2964                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2965         }
2966 out_err:
2967         mutex_lock(&mdsc->mutex);
2968         if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2969                 if (err) {
2970                         req->r_err = err;
2971                 } else {
2972                         req->r_reply =  ceph_msg_get(msg);
2973                         set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
2974                 }
2975         } else {
2976                 dout("reply arrived after request %lld was aborted\n", tid);
2977         }
2978         mutex_unlock(&mdsc->mutex);
2979
2980         mutex_unlock(&session->s_mutex);
2981
2982         /* kick calling process */
2983         complete_request(mdsc, req);
2984 out:
2985         ceph_mdsc_put_request(req);
2986         return;
2987 }
2988
2989
2990
2991 /*
2992  * handle mds notification that our request has been forwarded.
2993  */
2994 static void handle_forward(struct ceph_mds_client *mdsc,
2995                            struct ceph_mds_session *session,
2996                            struct ceph_msg *msg)
2997 {
2998         struct ceph_mds_request *req;
2999         u64 tid = le64_to_cpu(msg->hdr.tid);
3000         u32 next_mds;
3001         u32 fwd_seq;
3002         int err = -EINVAL;
3003         void *p = msg->front.iov_base;
3004         void *end = p + msg->front.iov_len;
3005
3006         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
3007         next_mds = ceph_decode_32(&p);
3008         fwd_seq = ceph_decode_32(&p);
3009
3010         mutex_lock(&mdsc->mutex);
3011         req = lookup_get_request(mdsc, tid);
3012         if (!req) {
3013                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
3014                 goto out;  /* dup reply? */
3015         }
3016
3017         if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3018                 dout("forward tid %llu aborted, unregistering\n", tid);
3019                 __unregister_request(mdsc, req);
3020         } else if (fwd_seq <= req->r_num_fwd) {
3021                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
3022                      tid, next_mds, req->r_num_fwd, fwd_seq);
3023         } else {
3024                 /* resend. forward race not possible; mds would drop */
3025                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
3026                 BUG_ON(req->r_err);
3027                 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
3028                 req->r_attempts = 0;
3029                 req->r_num_fwd = fwd_seq;
3030                 req->r_resend_mds = next_mds;
3031                 put_request_session(req);
3032                 __do_request(mdsc, req);
3033         }
3034         ceph_mdsc_put_request(req);
3035 out:
3036         mutex_unlock(&mdsc->mutex);
3037         return;
3038
3039 bad:
3040         pr_err("mdsc_handle_forward decode error err=%d\n", err);
3041 }
3042
3043 static int __decode_session_metadata(void **p, void *end,
3044                                      bool *blacklisted)
3045 {
3046         /* map<string,string> */
3047         u32 n;
3048         bool err_str;
3049         ceph_decode_32_safe(p, end, n, bad);
3050         while (n-- > 0) {
3051                 u32 len;
3052                 ceph_decode_32_safe(p, end, len, bad);
3053                 ceph_decode_need(p, end, len, bad);
3054                 err_str = !strncmp(*p, "error_string", len);
3055                 *p += len;
3056                 ceph_decode_32_safe(p, end, len, bad);
3057                 ceph_decode_need(p, end, len, bad);
3058                 if (err_str && strnstr(*p, "blacklisted", len))
3059                         *blacklisted = true;
3060                 *p += len;
3061         }
3062         return 0;
3063 bad:
3064         return -1;
3065 }
3066
3067 /*
3068  * handle a mds session control message
3069  */
3070 static void handle_session(struct ceph_mds_session *session,
3071                            struct ceph_msg *msg)
3072 {
3073         struct ceph_mds_client *mdsc = session->s_mdsc;
3074         int mds = session->s_mds;
3075         int msg_version = le16_to_cpu(msg->hdr.version);
3076         void *p = msg->front.iov_base;
3077         void *end = p + msg->front.iov_len;
3078         struct ceph_mds_session_head *h;
3079         u32 op;
3080         u64 seq;
3081         unsigned long features = 0;
3082         int wake = 0;
3083         bool blacklisted = false;
3084
3085         /* decode */
3086         ceph_decode_need(&p, end, sizeof(*h), bad);
3087         h = p;
3088         p += sizeof(*h);
3089
3090         op = le32_to_cpu(h->op);
3091         seq = le64_to_cpu(h->seq);
3092
3093         if (msg_version >= 3) {
3094                 u32 len;
3095                 /* version >= 2, metadata */
3096                 if (__decode_session_metadata(&p, end, &blacklisted) < 0)
3097                         goto bad;
3098                 /* version >= 3, feature bits */
3099                 ceph_decode_32_safe(&p, end, len, bad);
3100                 ceph_decode_need(&p, end, len, bad);
3101                 memcpy(&features, p, min_t(size_t, len, sizeof(features)));
3102                 p += len;
3103         }
3104
3105         mutex_lock(&mdsc->mutex);
3106         if (op == CEPH_SESSION_CLOSE) {
3107                 get_session(session);
3108                 __unregister_session(mdsc, session);
3109         }
3110         /* FIXME: this ttl calculation is generous */
3111         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
3112         mutex_unlock(&mdsc->mutex);
3113
3114         mutex_lock(&session->s_mutex);
3115
3116         dout("handle_session mds%d %s %p state %s seq %llu\n",
3117              mds, ceph_session_op_name(op), session,
3118              ceph_session_state_name(session->s_state), seq);
3119
3120         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
3121                 session->s_state = CEPH_MDS_SESSION_OPEN;
3122                 pr_info("mds%d came back\n", session->s_mds);
3123         }
3124
3125         switch (op) {
3126         case CEPH_SESSION_OPEN:
3127                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3128                         pr_info("mds%d reconnect success\n", session->s_mds);
3129                 session->s_state = CEPH_MDS_SESSION_OPEN;
3130                 session->s_features = features;
3131                 renewed_caps(mdsc, session, 0);
3132                 wake = 1;
3133                 if (mdsc->stopping)
3134                         __close_session(mdsc, session);
3135                 break;
3136
3137         case CEPH_SESSION_RENEWCAPS:
3138                 if (session->s_renew_seq == seq)
3139                         renewed_caps(mdsc, session, 1);
3140                 break;
3141
3142         case CEPH_SESSION_CLOSE:
3143                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3144                         pr_info("mds%d reconnect denied\n", session->s_mds);
3145                 cleanup_session_requests(mdsc, session);
3146                 remove_session_caps(session);
3147                 wake = 2; /* for good measure */
3148                 wake_up_all(&mdsc->session_close_wq);
3149                 break;
3150
3151         case CEPH_SESSION_STALE:
3152                 pr_info("mds%d caps went stale, renewing\n",
3153                         session->s_mds);
3154                 spin_lock(&session->s_gen_ttl_lock);
3155                 session->s_cap_gen++;
3156                 session->s_cap_ttl = jiffies - 1;
3157                 spin_unlock(&session->s_gen_ttl_lock);
3158                 send_renew_caps(mdsc, session);
3159                 break;
3160
3161         case CEPH_SESSION_RECALL_STATE:
3162                 ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
3163                 break;
3164
3165         case CEPH_SESSION_FLUSHMSG:
3166                 send_flushmsg_ack(mdsc, session, seq);
3167                 break;
3168
3169         case CEPH_SESSION_FORCE_RO:
3170                 dout("force_session_readonly %p\n", session);
3171                 spin_lock(&session->s_cap_lock);
3172                 session->s_readonly = true;
3173                 spin_unlock(&session->s_cap_lock);
3174                 wake_up_session_caps(session, FORCE_RO);
3175                 break;
3176
3177         case CEPH_SESSION_REJECT:
3178                 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
3179                 pr_info("mds%d rejected session\n", session->s_mds);
3180                 session->s_state = CEPH_MDS_SESSION_REJECTED;
3181                 cleanup_session_requests(mdsc, session);
3182                 remove_session_caps(session);
3183                 if (blacklisted)
3184                         mdsc->fsc->blacklisted = true;
3185                 wake = 2; /* for good measure */
3186                 break;
3187
3188         default:
3189                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
3190                 WARN_ON(1);
3191         }
3192
3193         mutex_unlock(&session->s_mutex);
3194         if (wake) {
3195                 mutex_lock(&mdsc->mutex);
3196                 __wake_requests(mdsc, &session->s_waiting);
3197                 if (wake == 2)
3198                         kick_requests(mdsc, mds);
3199                 mutex_unlock(&mdsc->mutex);
3200         }
3201         if (op == CEPH_SESSION_CLOSE)
3202                 ceph_put_mds_session(session);
3203         return;
3204
3205 bad:
3206         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
3207                (int)msg->front.iov_len);
3208         ceph_msg_dump(msg);
3209         return;
3210 }
3211
3212
3213 /*
3214  * called under session->mutex.
3215  */
3216 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
3217                                    struct ceph_mds_session *session)
3218 {
3219         struct ceph_mds_request *req, *nreq;
3220         struct rb_node *p;
3221         int err;
3222
3223         dout("replay_unsafe_requests mds%d\n", session->s_mds);
3224
3225         mutex_lock(&mdsc->mutex);
3226         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
3227                 err = __prepare_send_request(mdsc, req, session->s_mds, true);
3228                 if (!err) {
3229                         ceph_msg_get(req->r_request);
3230                         ceph_con_send(&session->s_con, req->r_request);
3231                 }
3232         }
3233
3234         /*
3235          * also re-send old requests when MDS enters reconnect stage. So that MDS
3236          * can process completed request in clientreplay stage.
3237          */
3238         p = rb_first(&mdsc->request_tree);
3239         while (p) {
3240                 req = rb_entry(p, struct ceph_mds_request, r_node);
3241                 p = rb_next(p);
3242                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3243                         continue;
3244                 if (req->r_attempts == 0)
3245                         continue; /* only old requests */
3246                 if (req->r_session &&
3247                     req->r_session->s_mds == session->s_mds) {
3248                         err = __prepare_send_request(mdsc, req,
3249                                                      session->s_mds, true);
3250                         if (!err) {
3251                                 ceph_msg_get(req->r_request);
3252                                 ceph_con_send(&session->s_con, req->r_request);
3253                         }
3254                 }
3255         }
3256         mutex_unlock(&mdsc->mutex);
3257 }
3258
3259 static int send_reconnect_partial(struct ceph_reconnect_state *recon_state)
3260 {
3261         struct ceph_msg *reply;
3262         struct ceph_pagelist *_pagelist;
3263         struct page *page;
3264         __le32 *addr;
3265         int err = -ENOMEM;
3266
3267         if (!recon_state->allow_multi)
3268                 return -ENOSPC;
3269
3270         /* can't handle message that contains both caps and realm */
3271         BUG_ON(!recon_state->nr_caps == !recon_state->nr_realms);
3272
3273         /* pre-allocate new pagelist */
3274         _pagelist = ceph_pagelist_alloc(GFP_NOFS);
3275         if (!_pagelist)
3276                 return -ENOMEM;
3277
3278         reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3279         if (!reply)
3280                 goto fail_msg;
3281
3282         /* placeholder for nr_caps */
3283         err = ceph_pagelist_encode_32(_pagelist, 0);
3284         if (err < 0)
3285                 goto fail;
3286
3287         if (recon_state->nr_caps) {
3288                 /* currently encoding caps */
3289                 err = ceph_pagelist_encode_32(recon_state->pagelist, 0);
3290                 if (err)
3291                         goto fail;
3292         } else {
3293                 /* placeholder for nr_realms (currently encoding relams) */
3294                 err = ceph_pagelist_encode_32(_pagelist, 0);
3295                 if (err < 0)
3296                         goto fail;
3297         }
3298
3299         err = ceph_pagelist_encode_8(recon_state->pagelist, 1);
3300         if (err)
3301                 goto fail;
3302
3303         page = list_first_entry(&recon_state->pagelist->head, struct page, lru);
3304         addr = kmap_atomic(page);
3305         if (recon_state->nr_caps) {
3306                 /* currently encoding caps */
3307                 *addr = cpu_to_le32(recon_state->nr_caps);
3308         } else {
3309                 /* currently encoding relams */
3310                 *(addr + 1) = cpu_to_le32(recon_state->nr_realms);
3311         }
3312         kunmap_atomic(addr);
3313
3314         reply->hdr.version = cpu_to_le16(5);
3315         reply->hdr.compat_version = cpu_to_le16(4);
3316
3317         reply->hdr.data_len = cpu_to_le32(recon_state->pagelist->length);
3318         ceph_msg_data_add_pagelist(reply, recon_state->pagelist);
3319
3320         ceph_con_send(&recon_state->session->s_con, reply);
3321         ceph_pagelist_release(recon_state->pagelist);
3322
3323         recon_state->pagelist = _pagelist;
3324         recon_state->nr_caps = 0;
3325         recon_state->nr_realms = 0;
3326         recon_state->msg_version = 5;
3327         return 0;
3328 fail:
3329         ceph_msg_put(reply);
3330 fail_msg:
3331         ceph_pagelist_release(_pagelist);
3332         return err;
3333 }
3334
3335 /*
3336  * Encode information about a cap for a reconnect with the MDS.
3337  */
3338 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
3339                           void *arg)
3340 {
3341         union {
3342                 struct ceph_mds_cap_reconnect v2;
3343                 struct ceph_mds_cap_reconnect_v1 v1;
3344         } rec;
3345         struct ceph_inode_info *ci = cap->ci;
3346         struct ceph_reconnect_state *recon_state = arg;
3347         struct ceph_pagelist *pagelist = recon_state->pagelist;
3348         int err;
3349         u64 snap_follows;
3350
3351         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
3352              inode, ceph_vinop(inode), cap, cap->cap_id,
3353              ceph_cap_string(cap->issued));
3354
3355         spin_lock(&ci->i_ceph_lock);
3356         cap->seq = 0;        /* reset cap seq */
3357         cap->issue_seq = 0;  /* and issue_seq */
3358         cap->mseq = 0;       /* and migrate_seq */
3359         cap->cap_gen = cap->session->s_cap_gen;
3360
3361         if (recon_state->msg_version >= 2) {
3362                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
3363                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3364                 rec.v2.issued = cpu_to_le32(cap->issued);
3365                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3366                 rec.v2.pathbase = 0;
3367                 rec.v2.flock_len = (__force __le32)
3368                         ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
3369         } else {
3370                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
3371                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3372                 rec.v1.issued = cpu_to_le32(cap->issued);
3373                 rec.v1.size = cpu_to_le64(inode->i_size);
3374                 ceph_encode_timespec64(&rec.v1.mtime, &inode->i_mtime);
3375                 ceph_encode_timespec64(&rec.v1.atime, &inode->i_atime);
3376                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3377                 rec.v1.pathbase = 0;
3378         }
3379
3380         if (list_empty(&ci->i_cap_snaps)) {
3381                 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
3382         } else {
3383                 struct ceph_cap_snap *capsnap =
3384                         list_first_entry(&ci->i_cap_snaps,
3385                                          struct ceph_cap_snap, ci_item);
3386                 snap_follows = capsnap->follows;
3387         }
3388         spin_unlock(&ci->i_ceph_lock);
3389
3390         if (recon_state->msg_version >= 2) {
3391                 int num_fcntl_locks, num_flock_locks;
3392                 struct ceph_filelock *flocks = NULL;
3393                 size_t struct_len, total_len = sizeof(u64);
3394                 u8 struct_v = 0;
3395
3396 encode_again:
3397                 if (rec.v2.flock_len) {
3398                         ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
3399                 } else {
3400                         num_fcntl_locks = 0;
3401                         num_flock_locks = 0;
3402                 }
3403                 if (num_fcntl_locks + num_flock_locks > 0) {
3404                         flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
3405                                                sizeof(struct ceph_filelock),
3406                                                GFP_NOFS);
3407                         if (!flocks) {
3408                                 err = -ENOMEM;
3409                                 goto out_err;
3410                         }
3411                         err = ceph_encode_locks_to_buffer(inode, flocks,
3412                                                           num_fcntl_locks,
3413                                                           num_flock_locks);
3414                         if (err) {
3415                                 kfree(flocks);
3416                                 flocks = NULL;
3417                                 if (err == -ENOSPC)
3418                                         goto encode_again;
3419                                 goto out_err;
3420                         }
3421                 } else {
3422                         kfree(flocks);
3423                         flocks = NULL;
3424                 }
3425
3426                 if (recon_state->msg_version >= 3) {
3427                         /* version, compat_version and struct_len */
3428                         total_len += 2 * sizeof(u8) + sizeof(u32);
3429                         struct_v = 2;
3430                 }
3431                 /*
3432                  * number of encoded locks is stable, so copy to pagelist
3433                  */
3434                 struct_len = 2 * sizeof(u32) +
3435                             (num_fcntl_locks + num_flock_locks) *
3436                             sizeof(struct ceph_filelock);
3437                 rec.v2.flock_len = cpu_to_le32(struct_len);
3438
3439                 struct_len += sizeof(u32) + sizeof(rec.v2);
3440
3441                 if (struct_v >= 2)
3442                         struct_len += sizeof(u64); /* snap_follows */
3443
3444                 total_len += struct_len;
3445
3446                 if (pagelist->length + total_len > RECONNECT_MAX_SIZE) {
3447                         err = send_reconnect_partial(recon_state);
3448                         if (err)
3449                                 goto out_freeflocks;
3450                         pagelist = recon_state->pagelist;
3451                 }
3452
3453                 err = ceph_pagelist_reserve(pagelist, total_len);
3454                 if (err)
3455                         goto out_freeflocks;
3456
3457                 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3458                 if (recon_state->msg_version >= 3) {
3459                         ceph_pagelist_encode_8(pagelist, struct_v);
3460                         ceph_pagelist_encode_8(pagelist, 1);
3461                         ceph_pagelist_encode_32(pagelist, struct_len);
3462                 }
3463                 ceph_pagelist_encode_string(pagelist, NULL, 0);
3464                 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
3465                 ceph_locks_to_pagelist(flocks, pagelist,
3466                                        num_fcntl_locks, num_flock_locks);
3467                 if (struct_v >= 2)
3468                         ceph_pagelist_encode_64(pagelist, snap_follows);
3469 out_freeflocks:
3470                 kfree(flocks);
3471         } else {
3472                 u64 pathbase = 0;
3473                 int pathlen = 0;
3474                 char *path = NULL;
3475                 struct dentry *dentry;
3476
3477                 dentry = d_find_alias(inode);
3478                 if (dentry) {
3479                         path = ceph_mdsc_build_path(dentry,
3480                                                 &pathlen, &pathbase, 0);
3481                         dput(dentry);
3482                         if (IS_ERR(path)) {
3483                                 err = PTR_ERR(path);
3484                                 goto out_err;
3485                         }
3486                         rec.v1.pathbase = cpu_to_le64(pathbase);
3487                 }
3488
3489                 err = ceph_pagelist_reserve(pagelist,
3490                                             sizeof(u64) + sizeof(u32) +
3491                                             pathlen + sizeof(rec.v1));
3492                 if (err) {
3493                         goto out_freepath;
3494                 }
3495
3496                 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3497                 ceph_pagelist_encode_string(pagelist, path, pathlen);
3498                 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
3499 out_freepath:
3500                 ceph_mdsc_free_path(path, pathlen);
3501         }
3502
3503 out_err:
3504         if (err >= 0)
3505                 recon_state->nr_caps++;
3506         return err;
3507 }
3508
3509 static int encode_snap_realms(struct ceph_mds_client *mdsc,
3510                               struct ceph_reconnect_state *recon_state)
3511 {
3512         struct rb_node *p;
3513         struct ceph_pagelist *pagelist = recon_state->pagelist;
3514         int err = 0;
3515
3516         if (recon_state->msg_version >= 4) {
3517                 err = ceph_pagelist_encode_32(pagelist, mdsc->num_snap_realms);
3518                 if (err < 0)
3519                         goto fail;
3520         }
3521
3522         /*
3523          * snaprealms.  we provide mds with the ino, seq (version), and
3524          * parent for all of our realms.  If the mds has any newer info,
3525          * it will tell us.
3526          */
3527         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3528                 struct ceph_snap_realm *realm =
3529                        rb_entry(p, struct ceph_snap_realm, node);
3530                 struct ceph_mds_snaprealm_reconnect sr_rec;
3531
3532                 if (recon_state->msg_version >= 4) {
3533                         size_t need = sizeof(u8) * 2 + sizeof(u32) +
3534                                       sizeof(sr_rec);
3535
3536                         if (pagelist->length + need > RECONNECT_MAX_SIZE) {
3537                                 err = send_reconnect_partial(recon_state);
3538                                 if (err)
3539                                         goto fail;
3540                                 pagelist = recon_state->pagelist;
3541                         }
3542
3543                         err = ceph_pagelist_reserve(pagelist, need);
3544                         if (err)
3545                                 goto fail;
3546
3547                         ceph_pagelist_encode_8(pagelist, 1);
3548                         ceph_pagelist_encode_8(pagelist, 1);
3549                         ceph_pagelist_encode_32(pagelist, sizeof(sr_rec));
3550                 }
3551
3552                 dout(" adding snap realm %llx seq %lld parent %llx\n",
3553                      realm->ino, realm->seq, realm->parent_ino);
3554                 sr_rec.ino = cpu_to_le64(realm->ino);
3555                 sr_rec.seq = cpu_to_le64(realm->seq);
3556                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3557
3558                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3559                 if (err)
3560                         goto fail;
3561
3562                 recon_state->nr_realms++;
3563         }
3564 fail:
3565         return err;
3566 }
3567
3568
3569 /*
3570  * If an MDS fails and recovers, clients need to reconnect in order to
3571  * reestablish shared state.  This includes all caps issued through
3572  * this session _and_ the snap_realm hierarchy.  Because it's not
3573  * clear which snap realms the mds cares about, we send everything we
3574  * know about.. that ensures we'll then get any new info the
3575  * recovering MDS might have.
3576  *
3577  * This is a relatively heavyweight operation, but it's rare.
3578  *
3579  * called with mdsc->mutex held.
3580  */
3581 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3582                                struct ceph_mds_session *session)
3583 {
3584         struct ceph_msg *reply;
3585         int mds = session->s_mds;
3586         int err = -ENOMEM;
3587         struct ceph_reconnect_state recon_state = {
3588                 .session = session,
3589         };
3590         LIST_HEAD(dispose);
3591
3592         pr_info("mds%d reconnect start\n", mds);
3593
3594         recon_state.pagelist = ceph_pagelist_alloc(GFP_NOFS);
3595         if (!recon_state.pagelist)
3596                 goto fail_nopagelist;
3597
3598         reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3599         if (!reply)
3600                 goto fail_nomsg;
3601
3602         mutex_lock(&session->s_mutex);
3603         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3604         session->s_seq = 0;
3605
3606         dout("session %p state %s\n", session,
3607              ceph_session_state_name(session->s_state));
3608
3609         spin_lock(&session->s_gen_ttl_lock);
3610         session->s_cap_gen++;
3611         spin_unlock(&session->s_gen_ttl_lock);
3612
3613         spin_lock(&session->s_cap_lock);
3614         /* don't know if session is readonly */
3615         session->s_readonly = 0;
3616         /*
3617          * notify __ceph_remove_cap() that we are composing cap reconnect.
3618          * If a cap get released before being added to the cap reconnect,
3619          * __ceph_remove_cap() should skip queuing cap release.
3620          */
3621         session->s_cap_reconnect = 1;
3622         /* drop old cap expires; we're about to reestablish that state */
3623         detach_cap_releases(session, &dispose);
3624         spin_unlock(&session->s_cap_lock);
3625         dispose_cap_releases(mdsc, &dispose);
3626
3627         /* trim unused caps to reduce MDS's cache rejoin time */
3628         if (mdsc->fsc->sb->s_root)
3629                 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3630
3631         ceph_con_close(&session->s_con);
3632         ceph_con_open(&session->s_con,
3633                       CEPH_ENTITY_TYPE_MDS, mds,
3634                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3635
3636         /* replay unsafe requests */
3637         replay_unsafe_requests(mdsc, session);
3638
3639         ceph_early_kick_flushing_caps(mdsc, session);
3640
3641         down_read(&mdsc->snap_rwsem);
3642
3643         /* placeholder for nr_caps */
3644         err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
3645         if (err)
3646                 goto fail;
3647
3648         if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
3649                 recon_state.msg_version = 3;
3650                 recon_state.allow_multi = true;
3651         } else if (session->s_con.peer_features & CEPH_FEATURE_MDSENC) {
3652                 recon_state.msg_version = 3;
3653         } else {
3654                 recon_state.msg_version = 2;
3655         }
3656         /* trsaverse this session's caps */
3657         err = ceph_iterate_session_caps(session, encode_caps_cb, &recon_state);
3658
3659         spin_lock(&session->s_cap_lock);
3660         session->s_cap_reconnect = 0;
3661         spin_unlock(&session->s_cap_lock);
3662
3663         if (err < 0)
3664                 goto fail;
3665
3666         /* check if all realms can be encoded into current message */
3667         if (mdsc->num_snap_realms) {
3668                 size_t total_len =
3669                         recon_state.pagelist->length +
3670                         mdsc->num_snap_realms *
3671                         sizeof(struct ceph_mds_snaprealm_reconnect);
3672                 if (recon_state.msg_version >= 4) {
3673                         /* number of realms */
3674                         total_len += sizeof(u32);
3675                         /* version, compat_version and struct_len */
3676                         total_len += mdsc->num_snap_realms *
3677                                      (2 * sizeof(u8) + sizeof(u32));
3678                 }
3679                 if (total_len > RECONNECT_MAX_SIZE) {
3680                         if (!recon_state.allow_multi) {
3681                                 err = -ENOSPC;
3682                                 goto fail;
3683                         }
3684                         if (recon_state.nr_caps) {
3685                                 err = send_reconnect_partial(&recon_state);
3686                                 if (err)
3687                                         goto fail;
3688                         }
3689                         recon_state.msg_version = 5;
3690                 }
3691         }
3692
3693         err = encode_snap_realms(mdsc, &recon_state);
3694         if (err < 0)
3695                 goto fail;
3696
3697         if (recon_state.msg_version >= 5) {
3698                 err = ceph_pagelist_encode_8(recon_state.pagelist, 0);
3699                 if (err < 0)
3700                         goto fail;
3701         }
3702
3703         if (recon_state.nr_caps || recon_state.nr_realms) {
3704                 struct page *page =
3705                         list_first_entry(&recon_state.pagelist->head,
3706                                         struct page, lru);
3707                 __le32 *addr = kmap_atomic(page);
3708                 if (recon_state.nr_caps) {
3709                         WARN_ON(recon_state.nr_realms != mdsc->num_snap_realms);
3710                         *addr = cpu_to_le32(recon_state.nr_caps);
3711                 } else if (recon_state.msg_version >= 4) {
3712                         *(addr + 1) = cpu_to_le32(recon_state.nr_realms);
3713                 }
3714                 kunmap_atomic(addr);
3715         }
3716
3717         reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3718         if (recon_state.msg_version >= 4)
3719                 reply->hdr.compat_version = cpu_to_le16(4);
3720
3721         reply->hdr.data_len = cpu_to_le32(recon_state.pagelist->length);
3722         ceph_msg_data_add_pagelist(reply, recon_state.pagelist);
3723
3724         ceph_con_send(&session->s_con, reply);
3725
3726         mutex_unlock(&session->s_mutex);
3727
3728         mutex_lock(&mdsc->mutex);
3729         __wake_requests(mdsc, &session->s_waiting);
3730         mutex_unlock(&mdsc->mutex);
3731
3732         up_read(&mdsc->snap_rwsem);
3733         ceph_pagelist_release(recon_state.pagelist);
3734         return;
3735
3736 fail:
3737         ceph_msg_put(reply);
3738         up_read(&mdsc->snap_rwsem);
3739         mutex_unlock(&session->s_mutex);
3740 fail_nomsg:
3741         ceph_pagelist_release(recon_state.pagelist);
3742 fail_nopagelist:
3743         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3744         return;
3745 }
3746
3747
3748 /*
3749  * compare old and new mdsmaps, kicking requests
3750  * and closing out old connections as necessary
3751  *
3752  * called under mdsc->mutex.
3753  */
3754 static void check_new_map(struct ceph_mds_client *mdsc,
3755                           struct ceph_mdsmap *newmap,
3756                           struct ceph_mdsmap *oldmap)
3757 {
3758         int i;
3759         int oldstate, newstate;
3760         struct ceph_mds_session *s;
3761
3762         dout("check_new_map new %u old %u\n",
3763              newmap->m_epoch, oldmap->m_epoch);
3764
3765         for (i = 0; i < oldmap->m_num_mds && i < mdsc->max_sessions; i++) {
3766                 if (!mdsc->sessions[i])
3767                         continue;
3768                 s = mdsc->sessions[i];
3769                 oldstate = ceph_mdsmap_get_state(oldmap, i);
3770                 newstate = ceph_mdsmap_get_state(newmap, i);
3771
3772                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3773                      i, ceph_mds_state_name(oldstate),
3774                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3775                      ceph_mds_state_name(newstate),
3776                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3777                      ceph_session_state_name(s->s_state));
3778
3779                 if (i >= newmap->m_num_mds) {
3780                         /* force close session for stopped mds */
3781                         get_session(s);
3782                         __unregister_session(mdsc, s);
3783                         __wake_requests(mdsc, &s->s_waiting);
3784                         mutex_unlock(&mdsc->mutex);
3785
3786                         mutex_lock(&s->s_mutex);
3787                         cleanup_session_requests(mdsc, s);
3788                         remove_session_caps(s);
3789                         mutex_unlock(&s->s_mutex);
3790
3791                         ceph_put_mds_session(s);
3792
3793                         mutex_lock(&mdsc->mutex);
3794                         kick_requests(mdsc, i);
3795                         continue;
3796                 }
3797
3798                 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
3799                            ceph_mdsmap_get_addr(newmap, i),
3800                            sizeof(struct ceph_entity_addr))) {
3801                         /* just close it */
3802                         mutex_unlock(&mdsc->mutex);
3803                         mutex_lock(&s->s_mutex);
3804                         mutex_lock(&mdsc->mutex);
3805                         ceph_con_close(&s->s_con);
3806                         mutex_unlock(&s->s_mutex);
3807                         s->s_state = CEPH_MDS_SESSION_RESTARTING;
3808                 } else if (oldstate == newstate) {
3809                         continue;  /* nothing new with this mds */
3810                 }
3811
3812                 /*
3813                  * send reconnect?
3814                  */
3815                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3816                     newstate >= CEPH_MDS_STATE_RECONNECT) {
3817                         mutex_unlock(&mdsc->mutex);
3818                         send_mds_reconnect(mdsc, s);
3819                         mutex_lock(&mdsc->mutex);
3820                 }
3821
3822                 /*
3823                  * kick request on any mds that has gone active.
3824                  */
3825                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3826                     newstate >= CEPH_MDS_STATE_ACTIVE) {
3827                         if (oldstate != CEPH_MDS_STATE_CREATING &&
3828                             oldstate != CEPH_MDS_STATE_STARTING)
3829                                 pr_info("mds%d recovery completed\n", s->s_mds);
3830                         kick_requests(mdsc, i);
3831                         ceph_kick_flushing_caps(mdsc, s);
3832                         wake_up_session_caps(s, RECONNECT);
3833                 }
3834         }
3835
3836         for (i = 0; i < newmap->m_num_mds && i < mdsc->max_sessions; i++) {
3837                 s = mdsc->sessions[i];
3838                 if (!s)
3839                         continue;
3840                 if (!ceph_mdsmap_is_laggy(newmap, i))
3841                         continue;
3842                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3843                     s->s_state == CEPH_MDS_SESSION_HUNG ||
3844                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
3845                         dout(" connecting to export targets of laggy mds%d\n",
3846                              i);
3847                         __open_export_target_sessions(mdsc, s);
3848                 }
3849         }
3850 }
3851
3852
3853
3854 /*
3855  * leases
3856  */
3857
3858 /*
3859  * caller must hold session s_mutex, dentry->d_lock
3860  */
3861 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3862 {
3863         struct ceph_dentry_info *di = ceph_dentry(dentry);
3864
3865         ceph_put_mds_session(di->lease_session);
3866         di->lease_session = NULL;
3867 }
3868
3869 static void handle_lease(struct ceph_mds_client *mdsc,
3870                          struct ceph_mds_session *session,
3871                          struct ceph_msg *msg)
3872 {
3873         struct super_block *sb = mdsc->fsc->sb;
3874         struct inode *inode;
3875         struct dentry *parent, *dentry;
3876         struct ceph_dentry_info *di;
3877         int mds = session->s_mds;
3878         struct ceph_mds_lease *h = msg->front.iov_base;
3879         u32 seq;
3880         struct ceph_vino vino;
3881         struct qstr dname;
3882         int release = 0;
3883
3884         dout("handle_lease from mds%d\n", mds);
3885
3886         /* decode */
3887         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3888                 goto bad;
3889         vino.ino = le64_to_cpu(h->ino);
3890         vino.snap = CEPH_NOSNAP;
3891         seq = le32_to_cpu(h->seq);
3892         dname.len = get_unaligned_le32(h + 1);
3893         if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
3894                 goto bad;
3895         dname.name = (void *)(h + 1) + sizeof(u32);
3896
3897         /* lookup inode */
3898         inode = ceph_find_inode(sb, vino);
3899         dout("handle_lease %s, ino %llx %p %.*s\n",
3900              ceph_lease_op_name(h->action), vino.ino, inode,
3901              dname.len, dname.name);
3902
3903         mutex_lock(&session->s_mutex);
3904         session->s_seq++;
3905
3906         if (!inode) {
3907                 dout("handle_lease no inode %llx\n", vino.ino);
3908                 goto release;
3909         }
3910
3911         /* dentry */
3912         parent = d_find_alias(inode);
3913         if (!parent) {
3914                 dout("no parent dentry on inode %p\n", inode);
3915                 WARN_ON(1);
3916                 goto release;  /* hrm... */
3917         }
3918         dname.hash = full_name_hash(parent, dname.name, dname.len);
3919         dentry = d_lookup(parent, &dname);
3920         dput(parent);
3921         if (!dentry)
3922                 goto release;
3923
3924         spin_lock(&dentry->d_lock);
3925         di = ceph_dentry(dentry);
3926         switch (h->action) {
3927         case CEPH_MDS_LEASE_REVOKE:
3928                 if (di->lease_session == session) {
3929                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3930                                 h->seq = cpu_to_le32(di->lease_seq);
3931                         __ceph_mdsc_drop_dentry_lease(dentry);
3932                 }
3933                 release = 1;
3934                 break;
3935
3936         case CEPH_MDS_LEASE_RENEW:
3937                 if (di->lease_session == session &&
3938                     di->lease_gen == session->s_cap_gen &&
3939                     di->lease_renew_from &&
3940                     di->lease_renew_after == 0) {
3941                         unsigned long duration =
3942                                 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3943
3944                         di->lease_seq = seq;
3945                         di->time = di->lease_renew_from + duration;
3946                         di->lease_renew_after = di->lease_renew_from +
3947                                 (duration >> 1);
3948                         di->lease_renew_from = 0;
3949                 }
3950                 break;
3951         }
3952         spin_unlock(&dentry->d_lock);
3953         dput(dentry);
3954
3955         if (!release)
3956                 goto out;
3957
3958 release:
3959         /* let's just reuse the same message */
3960         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3961         ceph_msg_get(msg);
3962         ceph_con_send(&session->s_con, msg);
3963
3964 out:
3965         mutex_unlock(&session->s_mutex);
3966         /* avoid calling iput_final() in mds dispatch threads */
3967         ceph_async_iput(inode);
3968         return;
3969
3970 bad:
3971         pr_err("corrupt lease message\n");
3972         ceph_msg_dump(msg);
3973 }
3974
3975 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3976                               struct dentry *dentry, char action,
3977                               u32 seq)
3978 {
3979         struct ceph_msg *msg;
3980         struct ceph_mds_lease *lease;
3981         struct inode *dir;
3982         int len = sizeof(*lease) + sizeof(u32) + NAME_MAX;
3983
3984         dout("lease_send_msg identry %p %s to mds%d\n",
3985              dentry, ceph_lease_op_name(action), session->s_mds);
3986
3987         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3988         if (!msg)
3989                 return;
3990         lease = msg->front.iov_base;
3991         lease->action = action;
3992         lease->seq = cpu_to_le32(seq);
3993
3994         spin_lock(&dentry->d_lock);
3995         dir = d_inode(dentry->d_parent);
3996         lease->ino = cpu_to_le64(ceph_ino(dir));
3997         lease->first = lease->last = cpu_to_le64(ceph_snap(dir));
3998
3999         put_unaligned_le32(dentry->d_name.len, lease + 1);
4000         memcpy((void *)(lease + 1) + 4,
4001                dentry->d_name.name, dentry->d_name.len);
4002         spin_unlock(&dentry->d_lock);
4003         /*
4004          * if this is a preemptive lease RELEASE, no need to
4005          * flush request stream, since the actual request will
4006          * soon follow.
4007          */
4008         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
4009
4010         ceph_con_send(&session->s_con, msg);
4011 }
4012
4013 /*
4014  * lock unlock sessions, to wait ongoing session activities
4015  */
4016 static void lock_unlock_sessions(struct ceph_mds_client *mdsc)
4017 {
4018         int i;
4019
4020         mutex_lock(&mdsc->mutex);
4021         for (i = 0; i < mdsc->max_sessions; i++) {
4022                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4023                 if (!s)
4024                         continue;
4025                 mutex_unlock(&mdsc->mutex);
4026                 mutex_lock(&s->s_mutex);
4027                 mutex_unlock(&s->s_mutex);
4028                 ceph_put_mds_session(s);
4029                 mutex_lock(&mdsc->mutex);
4030         }
4031         mutex_unlock(&mdsc->mutex);
4032 }
4033
4034 static void maybe_recover_session(struct ceph_mds_client *mdsc)
4035 {
4036         struct ceph_fs_client *fsc = mdsc->fsc;
4037
4038         if (!ceph_test_mount_opt(fsc, CLEANRECOVER))
4039                 return;
4040
4041         if (READ_ONCE(fsc->mount_state) != CEPH_MOUNT_MOUNTED)
4042                 return;
4043
4044         if (!READ_ONCE(fsc->blacklisted))
4045                 return;
4046
4047         if (fsc->last_auto_reconnect &&
4048             time_before(jiffies, fsc->last_auto_reconnect + HZ * 60 * 30))
4049                 return;
4050
4051         pr_info("auto reconnect after blacklisted\n");
4052         fsc->last_auto_reconnect = jiffies;
4053         ceph_force_reconnect(fsc->sb);
4054 }
4055
4056 /*
4057  * delayed work -- periodically trim expired leases, renew caps with mds
4058  */
4059 static void schedule_delayed(struct ceph_mds_client *mdsc)
4060 {
4061         int delay = 5;
4062         unsigned hz = round_jiffies_relative(HZ * delay);
4063         schedule_delayed_work(&mdsc->delayed_work, hz);
4064 }
4065
4066 static void delayed_work(struct work_struct *work)
4067 {
4068         int i;
4069         struct ceph_mds_client *mdsc =
4070                 container_of(work, struct ceph_mds_client, delayed_work.work);
4071         int renew_interval;
4072         int renew_caps;
4073
4074         dout("mdsc delayed_work\n");
4075
4076         mutex_lock(&mdsc->mutex);
4077         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
4078         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
4079                                    mdsc->last_renew_caps);
4080         if (renew_caps)
4081                 mdsc->last_renew_caps = jiffies;
4082
4083         for (i = 0; i < mdsc->max_sessions; i++) {
4084                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4085                 if (!s)
4086                         continue;
4087                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
4088                         dout("resending session close request for mds%d\n",
4089                              s->s_mds);
4090                         request_close_session(mdsc, s);
4091                         ceph_put_mds_session(s);
4092                         continue;
4093                 }
4094                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
4095                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
4096                                 s->s_state = CEPH_MDS_SESSION_HUNG;
4097                                 pr_info("mds%d hung\n", s->s_mds);
4098                         }
4099                 }
4100                 if (s->s_state == CEPH_MDS_SESSION_NEW ||
4101                     s->s_state == CEPH_MDS_SESSION_RESTARTING ||
4102                     s->s_state == CEPH_MDS_SESSION_REJECTED) {
4103                         /* this mds is failed or recovering, just wait */
4104                         ceph_put_mds_session(s);
4105                         continue;
4106                 }
4107                 mutex_unlock(&mdsc->mutex);
4108
4109                 mutex_lock(&s->s_mutex);
4110                 if (renew_caps)
4111                         send_renew_caps(mdsc, s);
4112                 else
4113                         ceph_con_keepalive(&s->s_con);
4114                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4115                     s->s_state == CEPH_MDS_SESSION_HUNG)
4116                         ceph_send_cap_releases(mdsc, s);
4117                 mutex_unlock(&s->s_mutex);
4118                 ceph_put_mds_session(s);
4119
4120                 mutex_lock(&mdsc->mutex);
4121         }
4122         mutex_unlock(&mdsc->mutex);
4123
4124         ceph_check_delayed_caps(mdsc);
4125
4126         ceph_queue_cap_reclaim_work(mdsc);
4127
4128         ceph_trim_snapid_map(mdsc);
4129
4130         maybe_recover_session(mdsc);
4131
4132         schedule_delayed(mdsc);
4133 }
4134
4135 int ceph_mdsc_init(struct ceph_fs_client *fsc)
4136
4137 {
4138         struct ceph_mds_client *mdsc;
4139
4140         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
4141         if (!mdsc)
4142                 return -ENOMEM;
4143         mdsc->fsc = fsc;
4144         mutex_init(&mdsc->mutex);
4145         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
4146         if (!mdsc->mdsmap) {
4147                 kfree(mdsc);
4148                 return -ENOMEM;
4149         }
4150
4151         fsc->mdsc = mdsc;
4152         init_completion(&mdsc->safe_umount_waiters);
4153         init_waitqueue_head(&mdsc->session_close_wq);
4154         INIT_LIST_HEAD(&mdsc->waiting_for_map);
4155         mdsc->sessions = NULL;
4156         atomic_set(&mdsc->num_sessions, 0);
4157         mdsc->max_sessions = 0;
4158         mdsc->stopping = 0;
4159         atomic64_set(&mdsc->quotarealms_count, 0);
4160         mdsc->quotarealms_inodes = RB_ROOT;
4161         mutex_init(&mdsc->quotarealms_inodes_mutex);
4162         mdsc->last_snap_seq = 0;
4163         init_rwsem(&mdsc->snap_rwsem);
4164         mdsc->snap_realms = RB_ROOT;
4165         INIT_LIST_HEAD(&mdsc->snap_empty);
4166         mdsc->num_snap_realms = 0;
4167         spin_lock_init(&mdsc->snap_empty_lock);
4168         mdsc->last_tid = 0;
4169         mdsc->oldest_tid = 0;
4170         mdsc->request_tree = RB_ROOT;
4171         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
4172         mdsc->last_renew_caps = jiffies;
4173         INIT_LIST_HEAD(&mdsc->cap_delay_list);
4174         INIT_LIST_HEAD(&mdsc->cap_wait_list);
4175         spin_lock_init(&mdsc->cap_delay_lock);
4176         INIT_LIST_HEAD(&mdsc->snap_flush_list);
4177         spin_lock_init(&mdsc->snap_flush_lock);
4178         mdsc->last_cap_flush_tid = 1;
4179         INIT_LIST_HEAD(&mdsc->cap_flush_list);
4180         INIT_LIST_HEAD(&mdsc->cap_dirty);
4181         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
4182         mdsc->num_cap_flushing = 0;
4183         spin_lock_init(&mdsc->cap_dirty_lock);
4184         init_waitqueue_head(&mdsc->cap_flushing_wq);
4185         INIT_WORK(&mdsc->cap_reclaim_work, ceph_cap_reclaim_work);
4186         atomic_set(&mdsc->cap_reclaim_pending, 0);
4187
4188         spin_lock_init(&mdsc->dentry_list_lock);
4189         INIT_LIST_HEAD(&mdsc->dentry_leases);
4190         INIT_LIST_HEAD(&mdsc->dentry_dir_leases);
4191
4192         ceph_caps_init(mdsc);
4193         ceph_adjust_caps_max_min(mdsc, fsc->mount_options);
4194
4195         spin_lock_init(&mdsc->snapid_map_lock);
4196         mdsc->snapid_map_tree = RB_ROOT;
4197         INIT_LIST_HEAD(&mdsc->snapid_map_lru);
4198
4199         init_rwsem(&mdsc->pool_perm_rwsem);
4200         mdsc->pool_perm_tree = RB_ROOT;
4201
4202         strscpy(mdsc->nodename, utsname()->nodename,
4203                 sizeof(mdsc->nodename));
4204         return 0;
4205 }
4206
4207 /*
4208  * Wait for safe replies on open mds requests.  If we time out, drop
4209  * all requests from the tree to avoid dangling dentry refs.
4210  */
4211 static void wait_requests(struct ceph_mds_client *mdsc)
4212 {
4213         struct ceph_options *opts = mdsc->fsc->client->options;
4214         struct ceph_mds_request *req;
4215
4216         mutex_lock(&mdsc->mutex);
4217         if (__get_oldest_req(mdsc)) {
4218                 mutex_unlock(&mdsc->mutex);
4219
4220                 dout("wait_requests waiting for requests\n");
4221                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
4222                                     ceph_timeout_jiffies(opts->mount_timeout));
4223
4224                 /* tear down remaining requests */
4225                 mutex_lock(&mdsc->mutex);
4226                 while ((req = __get_oldest_req(mdsc))) {
4227                         dout("wait_requests timed out on tid %llu\n",
4228                              req->r_tid);
4229                         list_del_init(&req->r_wait);
4230                         __unregister_request(mdsc, req);
4231                 }
4232         }
4233         mutex_unlock(&mdsc->mutex);
4234         dout("wait_requests done\n");
4235 }
4236
4237 /*
4238  * called before mount is ro, and before dentries are torn down.
4239  * (hmm, does this still race with new lookups?)
4240  */
4241 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
4242 {
4243         dout("pre_umount\n");
4244         mdsc->stopping = 1;
4245
4246         lock_unlock_sessions(mdsc);
4247         ceph_flush_dirty_caps(mdsc);
4248         wait_requests(mdsc);
4249
4250         /*
4251          * wait for reply handlers to drop their request refs and
4252          * their inode/dcache refs
4253          */
4254         ceph_msgr_flush();
4255
4256         ceph_cleanup_quotarealms_inodes(mdsc);
4257 }
4258
4259 /*
4260  * wait for all write mds requests to flush.
4261  */
4262 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
4263 {
4264         struct ceph_mds_request *req = NULL, *nextreq;
4265         struct rb_node *n;
4266
4267         mutex_lock(&mdsc->mutex);
4268         dout("wait_unsafe_requests want %lld\n", want_tid);
4269 restart:
4270         req = __get_oldest_req(mdsc);
4271         while (req && req->r_tid <= want_tid) {
4272                 /* find next request */
4273                 n = rb_next(&req->r_node);
4274                 if (n)
4275                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
4276                 else
4277                         nextreq = NULL;
4278                 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
4279                     (req->r_op & CEPH_MDS_OP_WRITE)) {
4280                         /* write op */
4281                         ceph_mdsc_get_request(req);
4282                         if (nextreq)
4283                                 ceph_mdsc_get_request(nextreq);
4284                         mutex_unlock(&mdsc->mutex);
4285                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
4286                              req->r_tid, want_tid);
4287                         wait_for_completion(&req->r_safe_completion);
4288                         mutex_lock(&mdsc->mutex);
4289                         ceph_mdsc_put_request(req);
4290                         if (!nextreq)
4291                                 break;  /* next dne before, so we're done! */
4292                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
4293                                 /* next request was removed from tree */
4294                                 ceph_mdsc_put_request(nextreq);
4295                                 goto restart;
4296                         }
4297                         ceph_mdsc_put_request(nextreq);  /* won't go away */
4298                 }
4299                 req = nextreq;
4300         }
4301         mutex_unlock(&mdsc->mutex);
4302         dout("wait_unsafe_requests done\n");
4303 }
4304
4305 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
4306 {
4307         u64 want_tid, want_flush;
4308
4309         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4310                 return;
4311
4312         dout("sync\n");
4313         mutex_lock(&mdsc->mutex);
4314         want_tid = mdsc->last_tid;
4315         mutex_unlock(&mdsc->mutex);
4316
4317         ceph_flush_dirty_caps(mdsc);
4318         spin_lock(&mdsc->cap_dirty_lock);
4319         want_flush = mdsc->last_cap_flush_tid;
4320         if (!list_empty(&mdsc->cap_flush_list)) {
4321                 struct ceph_cap_flush *cf =
4322                         list_last_entry(&mdsc->cap_flush_list,
4323                                         struct ceph_cap_flush, g_list);
4324                 cf->wake = true;
4325         }
4326         spin_unlock(&mdsc->cap_dirty_lock);
4327
4328         dout("sync want tid %lld flush_seq %lld\n",
4329              want_tid, want_flush);
4330
4331         wait_unsafe_requests(mdsc, want_tid);
4332         wait_caps_flush(mdsc, want_flush);
4333 }
4334
4335 /*
4336  * true if all sessions are closed, or we force unmount
4337  */
4338 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
4339 {
4340         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4341                 return true;
4342         return atomic_read(&mdsc->num_sessions) <= skipped;
4343 }
4344
4345 /*
4346  * called after sb is ro.
4347  */
4348 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
4349 {
4350         struct ceph_options *opts = mdsc->fsc->client->options;
4351         struct ceph_mds_session *session;
4352         int i;
4353         int skipped = 0;
4354
4355         dout("close_sessions\n");
4356
4357         /* close sessions */
4358         mutex_lock(&mdsc->mutex);
4359         for (i = 0; i < mdsc->max_sessions; i++) {
4360                 session = __ceph_lookup_mds_session(mdsc, i);
4361                 if (!session)
4362                         continue;
4363                 mutex_unlock(&mdsc->mutex);
4364                 mutex_lock(&session->s_mutex);
4365                 if (__close_session(mdsc, session) <= 0)
4366                         skipped++;
4367                 mutex_unlock(&session->s_mutex);
4368                 ceph_put_mds_session(session);
4369                 mutex_lock(&mdsc->mutex);
4370         }
4371         mutex_unlock(&mdsc->mutex);
4372
4373         dout("waiting for sessions to close\n");
4374         wait_event_timeout(mdsc->session_close_wq,
4375                            done_closing_sessions(mdsc, skipped),
4376                            ceph_timeout_jiffies(opts->mount_timeout));
4377
4378         /* tear down remaining sessions */
4379         mutex_lock(&mdsc->mutex);
4380         for (i = 0; i < mdsc->max_sessions; i++) {
4381                 if (mdsc->sessions[i]) {
4382                         session = get_session(mdsc->sessions[i]);
4383                         __unregister_session(mdsc, session);
4384                         mutex_unlock(&mdsc->mutex);
4385                         mutex_lock(&session->s_mutex);
4386                         remove_session_caps(session);
4387                         mutex_unlock(&session->s_mutex);
4388                         ceph_put_mds_session(session);
4389                         mutex_lock(&mdsc->mutex);
4390                 }
4391         }
4392         WARN_ON(!list_empty(&mdsc->cap_delay_list));
4393         mutex_unlock(&mdsc->mutex);
4394
4395         ceph_cleanup_snapid_map(mdsc);
4396         ceph_cleanup_empty_realms(mdsc);
4397
4398         cancel_work_sync(&mdsc->cap_reclaim_work);
4399         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4400
4401         dout("stopped\n");
4402 }
4403
4404 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
4405 {
4406         struct ceph_mds_session *session;
4407         int mds;
4408
4409         dout("force umount\n");
4410
4411         mutex_lock(&mdsc->mutex);
4412         for (mds = 0; mds < mdsc->max_sessions; mds++) {
4413                 session = __ceph_lookup_mds_session(mdsc, mds);
4414                 if (!session)
4415                         continue;
4416
4417                 if (session->s_state == CEPH_MDS_SESSION_REJECTED)
4418                         __unregister_session(mdsc, session);
4419                 __wake_requests(mdsc, &session->s_waiting);
4420                 mutex_unlock(&mdsc->mutex);
4421
4422                 mutex_lock(&session->s_mutex);
4423                 __close_session(mdsc, session);
4424                 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
4425                         cleanup_session_requests(mdsc, session);
4426                         remove_session_caps(session);
4427                 }
4428                 mutex_unlock(&session->s_mutex);
4429                 ceph_put_mds_session(session);
4430
4431                 mutex_lock(&mdsc->mutex);
4432                 kick_requests(mdsc, mds);
4433         }
4434         __wake_requests(mdsc, &mdsc->waiting_for_map);
4435         mutex_unlock(&mdsc->mutex);
4436 }
4437
4438 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
4439 {
4440         dout("stop\n");
4441         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4442         if (mdsc->mdsmap)
4443                 ceph_mdsmap_destroy(mdsc->mdsmap);
4444         kfree(mdsc->sessions);
4445         ceph_caps_finalize(mdsc);
4446         ceph_pool_perm_destroy(mdsc);
4447 }
4448
4449 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
4450 {
4451         struct ceph_mds_client *mdsc = fsc->mdsc;
4452         dout("mdsc_destroy %p\n", mdsc);
4453
4454         if (!mdsc)
4455                 return;
4456
4457         /* flush out any connection work with references to us */
4458         ceph_msgr_flush();
4459
4460         ceph_mdsc_stop(mdsc);
4461
4462         fsc->mdsc = NULL;
4463         kfree(mdsc);
4464         dout("mdsc_destroy %p done\n", mdsc);
4465 }
4466
4467 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4468 {
4469         struct ceph_fs_client *fsc = mdsc->fsc;
4470         const char *mds_namespace = fsc->mount_options->mds_namespace;
4471         void *p = msg->front.iov_base;
4472         void *end = p + msg->front.iov_len;
4473         u32 epoch;
4474         u32 map_len;
4475         u32 num_fs;
4476         u32 mount_fscid = (u32)-1;
4477         u8 struct_v, struct_cv;
4478         int err = -EINVAL;
4479
4480         ceph_decode_need(&p, end, sizeof(u32), bad);
4481         epoch = ceph_decode_32(&p);
4482
4483         dout("handle_fsmap epoch %u\n", epoch);
4484
4485         ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4486         struct_v = ceph_decode_8(&p);
4487         struct_cv = ceph_decode_8(&p);
4488         map_len = ceph_decode_32(&p);
4489
4490         ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
4491         p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
4492
4493         num_fs = ceph_decode_32(&p);
4494         while (num_fs-- > 0) {
4495                 void *info_p, *info_end;
4496                 u32 info_len;
4497                 u8 info_v, info_cv;
4498                 u32 fscid, namelen;
4499
4500                 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4501                 info_v = ceph_decode_8(&p);
4502                 info_cv = ceph_decode_8(&p);
4503                 info_len = ceph_decode_32(&p);
4504                 ceph_decode_need(&p, end, info_len, bad);
4505                 info_p = p;
4506                 info_end = p + info_len;
4507                 p = info_end;
4508
4509                 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
4510                 fscid = ceph_decode_32(&info_p);
4511                 namelen = ceph_decode_32(&info_p);
4512                 ceph_decode_need(&info_p, info_end, namelen, bad);
4513
4514                 if (mds_namespace &&
4515                     strlen(mds_namespace) == namelen &&
4516                     !strncmp(mds_namespace, (char *)info_p, namelen)) {
4517                         mount_fscid = fscid;
4518                         break;
4519                 }
4520         }
4521
4522         ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
4523         if (mount_fscid != (u32)-1) {
4524                 fsc->client->monc.fs_cluster_id = mount_fscid;
4525                 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
4526                                    0, true);
4527                 ceph_monc_renew_subs(&fsc->client->monc);
4528         } else {
4529                 err = -ENOENT;
4530                 goto err_out;
4531         }
4532         return;
4533
4534 bad:
4535         pr_err("error decoding fsmap\n");
4536 err_out:
4537         mutex_lock(&mdsc->mutex);
4538         mdsc->mdsmap_err = err;
4539         __wake_requests(mdsc, &mdsc->waiting_for_map);
4540         mutex_unlock(&mdsc->mutex);
4541 }
4542
4543 /*
4544  * handle mds map update.
4545  */
4546 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4547 {
4548         u32 epoch;
4549         u32 maplen;
4550         void *p = msg->front.iov_base;
4551         void *end = p + msg->front.iov_len;
4552         struct ceph_mdsmap *newmap, *oldmap;
4553         struct ceph_fsid fsid;
4554         int err = -EINVAL;
4555
4556         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
4557         ceph_decode_copy(&p, &fsid, sizeof(fsid));
4558         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
4559                 return;
4560         epoch = ceph_decode_32(&p);
4561         maplen = ceph_decode_32(&p);
4562         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
4563
4564         /* do we need it? */
4565         mutex_lock(&mdsc->mutex);
4566         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
4567                 dout("handle_map epoch %u <= our %u\n",
4568                      epoch, mdsc->mdsmap->m_epoch);
4569                 mutex_unlock(&mdsc->mutex);
4570                 return;
4571         }
4572
4573         newmap = ceph_mdsmap_decode(&p, end);
4574         if (IS_ERR(newmap)) {
4575                 err = PTR_ERR(newmap);
4576                 goto bad_unlock;
4577         }
4578
4579         /* swap into place */
4580         if (mdsc->mdsmap) {
4581                 oldmap = mdsc->mdsmap;
4582                 mdsc->mdsmap = newmap;
4583                 check_new_map(mdsc, newmap, oldmap);
4584                 ceph_mdsmap_destroy(oldmap);
4585         } else {
4586                 mdsc->mdsmap = newmap;  /* first mds map */
4587         }
4588         mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
4589                                         MAX_LFS_FILESIZE);
4590
4591         __wake_requests(mdsc, &mdsc->waiting_for_map);
4592         ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
4593                           mdsc->mdsmap->m_epoch);
4594
4595         mutex_unlock(&mdsc->mutex);
4596         schedule_delayed(mdsc);
4597         return;
4598
4599 bad_unlock:
4600         mutex_unlock(&mdsc->mutex);
4601 bad:
4602         pr_err("error decoding mdsmap %d\n", err);
4603         return;
4604 }
4605
4606 static struct ceph_connection *con_get(struct ceph_connection *con)
4607 {
4608         struct ceph_mds_session *s = con->private;
4609
4610         if (get_session(s)) {
4611                 dout("mdsc con_get %p ok (%d)\n", s, refcount_read(&s->s_ref));
4612                 return con;
4613         }
4614         dout("mdsc con_get %p FAIL\n", s);
4615         return NULL;
4616 }
4617
4618 static void con_put(struct ceph_connection *con)
4619 {
4620         struct ceph_mds_session *s = con->private;
4621
4622         dout("mdsc con_put %p (%d)\n", s, refcount_read(&s->s_ref) - 1);
4623         ceph_put_mds_session(s);
4624 }
4625
4626 /*
4627  * if the client is unresponsive for long enough, the mds will kill
4628  * the session entirely.
4629  */
4630 static void peer_reset(struct ceph_connection *con)
4631 {
4632         struct ceph_mds_session *s = con->private;
4633         struct ceph_mds_client *mdsc = s->s_mdsc;
4634
4635         pr_warn("mds%d closed our session\n", s->s_mds);
4636         send_mds_reconnect(mdsc, s);
4637 }
4638
4639 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
4640 {
4641         struct ceph_mds_session *s = con->private;
4642         struct ceph_mds_client *mdsc = s->s_mdsc;
4643         int type = le16_to_cpu(msg->hdr.type);
4644
4645         mutex_lock(&mdsc->mutex);
4646         if (__verify_registered_session(mdsc, s) < 0) {
4647                 mutex_unlock(&mdsc->mutex);
4648                 goto out;
4649         }
4650         mutex_unlock(&mdsc->mutex);
4651
4652         switch (type) {
4653         case CEPH_MSG_MDS_MAP:
4654                 ceph_mdsc_handle_mdsmap(mdsc, msg);
4655                 break;
4656         case CEPH_MSG_FS_MAP_USER:
4657                 ceph_mdsc_handle_fsmap(mdsc, msg);
4658                 break;
4659         case CEPH_MSG_CLIENT_SESSION:
4660                 handle_session(s, msg);
4661                 break;
4662         case CEPH_MSG_CLIENT_REPLY:
4663                 handle_reply(s, msg);
4664                 break;
4665         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
4666                 handle_forward(mdsc, s, msg);
4667                 break;
4668         case CEPH_MSG_CLIENT_CAPS:
4669                 ceph_handle_caps(s, msg);
4670                 break;
4671         case CEPH_MSG_CLIENT_SNAP:
4672                 ceph_handle_snap(mdsc, s, msg);
4673                 break;
4674         case CEPH_MSG_CLIENT_LEASE:
4675                 handle_lease(mdsc, s, msg);
4676                 break;
4677         case CEPH_MSG_CLIENT_QUOTA:
4678                 ceph_handle_quota(mdsc, s, msg);
4679                 break;
4680
4681         default:
4682                 pr_err("received unknown message type %d %s\n", type,
4683                        ceph_msg_type_name(type));
4684         }
4685 out:
4686         ceph_msg_put(msg);
4687 }
4688
4689 /*
4690  * authentication
4691  */
4692
4693 /*
4694  * Note: returned pointer is the address of a structure that's
4695  * managed separately.  Caller must *not* attempt to free it.
4696  */
4697 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
4698                                         int *proto, int force_new)
4699 {
4700         struct ceph_mds_session *s = con->private;
4701         struct ceph_mds_client *mdsc = s->s_mdsc;
4702         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4703         struct ceph_auth_handshake *auth = &s->s_auth;
4704
4705         if (force_new && auth->authorizer) {
4706                 ceph_auth_destroy_authorizer(auth->authorizer);
4707                 auth->authorizer = NULL;
4708         }
4709         if (!auth->authorizer) {
4710                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4711                                                       auth);
4712                 if (ret)
4713                         return ERR_PTR(ret);
4714         } else {
4715                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4716                                                       auth);
4717                 if (ret)
4718                         return ERR_PTR(ret);
4719         }
4720         *proto = ac->protocol;
4721
4722         return auth;
4723 }
4724
4725 static int add_authorizer_challenge(struct ceph_connection *con,
4726                                     void *challenge_buf, int challenge_buf_len)
4727 {
4728         struct ceph_mds_session *s = con->private;
4729         struct ceph_mds_client *mdsc = s->s_mdsc;
4730         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4731
4732         return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
4733                                             challenge_buf, challenge_buf_len);
4734 }
4735
4736 static int verify_authorizer_reply(struct ceph_connection *con)
4737 {
4738         struct ceph_mds_session *s = con->private;
4739         struct ceph_mds_client *mdsc = s->s_mdsc;
4740         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4741
4742         return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
4743 }
4744
4745 static int invalidate_authorizer(struct ceph_connection *con)
4746 {
4747         struct ceph_mds_session *s = con->private;
4748         struct ceph_mds_client *mdsc = s->s_mdsc;
4749         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4750
4751         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
4752
4753         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
4754 }
4755
4756 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
4757                                 struct ceph_msg_header *hdr, int *skip)
4758 {
4759         struct ceph_msg *msg;
4760         int type = (int) le16_to_cpu(hdr->type);
4761         int front_len = (int) le32_to_cpu(hdr->front_len);
4762
4763         if (con->in_msg)
4764                 return con->in_msg;
4765
4766         *skip = 0;
4767         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
4768         if (!msg) {
4769                 pr_err("unable to allocate msg type %d len %d\n",
4770                        type, front_len);
4771                 return NULL;
4772         }
4773
4774         return msg;
4775 }
4776
4777 static int mds_sign_message(struct ceph_msg *msg)
4778 {
4779        struct ceph_mds_session *s = msg->con->private;
4780        struct ceph_auth_handshake *auth = &s->s_auth;
4781
4782        return ceph_auth_sign_message(auth, msg);
4783 }
4784
4785 static int mds_check_message_signature(struct ceph_msg *msg)
4786 {
4787        struct ceph_mds_session *s = msg->con->private;
4788        struct ceph_auth_handshake *auth = &s->s_auth;
4789
4790        return ceph_auth_check_message_signature(auth, msg);
4791 }
4792
4793 static const struct ceph_connection_operations mds_con_ops = {
4794         .get = con_get,
4795         .put = con_put,
4796         .dispatch = dispatch,
4797         .get_authorizer = get_authorizer,
4798         .add_authorizer_challenge = add_authorizer_challenge,
4799         .verify_authorizer_reply = verify_authorizer_reply,
4800         .invalidate_authorizer = invalidate_authorizer,
4801         .peer_reset = peer_reset,
4802         .alloc_msg = mds_alloc_msg,
4803         .sign_message = mds_sign_message,
4804         .check_message_signature = mds_check_message_signature,
4805 };
4806
4807 /* eof */