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  47  * are no longer valid.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