fs/ceph/mds_client.c

   1 #include <linux/ceph/ceph_debug.h>
   2
   3 #include <linux/fs.h>
   4 #include <linux/wait.h>
   5 #include <linux/slab.h>
   6 #include <linux/gfp.h>
   7 #include <linux/sched.h>
   8 #include <linux/debugfs.h>
   9 #include <linux/seq_file.h>
  10 #include <linux/utsname.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 /*
  24  * A cluster of MDS (metadata server) daemons is responsible for
  25  * managing the file system namespace (the directory hierarchy and
  26  * inodes) and for coordinating shared access to storage.  Metadata is
  27  * partitioning hierarchically across a number of servers, and that
  28  * partition varies over time as the cluster adjusts the distribution
  29  * in order to balance load.
  30  *
  31  * The MDS client is primarily responsible to managing synchronous
  32  * metadata requests for operations like open, unlink, and so forth.
  33  * If there is a MDS failure, we find out about it when we (possibly
  34  * request and) receive a new MDS map, and can resubmit affected
  35  * requests.
  36  *
  37  * For the most part, though, we take advantage of a lossless
  38  * communications channel to the MDS, and do not need to worry about
  39  * timing out or resubmitting requests.
  40  *
  41  * We maintain a stateful "session" with each MDS we interact with.
  42  * Within each session, we sent periodic heartbeat messages to ensure
  43  * any capabilities or leases we have been issues remain valid.  If
  44  * the session times out and goes stale, our leases and capabilities
  45  * are no longer valid.
  46  */
  47
  48 struct ceph_reconnect_state {
  49         int nr_caps;
  50         struct ceph_pagelist *pagelist;
  51         unsigned msg_version;
  52 };
  53
  54 static void __wake_requests(struct ceph_mds_client *mdsc,
  55                             struct list_head *head);
  56
  57 static const struct ceph_connection_operations mds_con_ops;
  58
  59
  60 /*
  61  * mds reply parsing
  62  */
  63
  64 /*
  65  * parse individual inode info
  66  */
  67 static int parse_reply_info_in(void **p, void *end,
  68                                struct ceph_mds_reply_info_in *info,
  69                                u64 features)
  70 {
  71         int err = -EIO;
  72
  73         info->in = *p;
  74         *p += sizeof(struct ceph_mds_reply_inode) +
  75                 sizeof(*info->in->fragtree.splits) *
  76                 le32_to_cpu(info->in->fragtree.nsplits);
  77
  78         ceph_decode_32_safe(p, end, info->symlink_len, bad);
  79         ceph_decode_need(p, end, info->symlink_len, bad);
  80         info->symlink = *p;
  81         *p += info->symlink_len;
  82
  83         if (features & CEPH_FEATURE_DIRLAYOUTHASH)
  84                 ceph_decode_copy_safe(p, end, &info->dir_layout,
  85                                       sizeof(info->dir_layout), bad);
  86         else
  87                 memset(&info->dir_layout, 0, sizeof(info->dir_layout));
  88
  89         ceph_decode_32_safe(p, end, info->xattr_len, bad);
  90         ceph_decode_need(p, end, info->xattr_len, bad);
  91         info->xattr_data = *p;
  92         *p += info->xattr_len;
  93
  94         if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
  95                 ceph_decode_64_safe(p, end, info->inline_version, bad);
  96                 ceph_decode_32_safe(p, end, info->inline_len, bad);
  97                 ceph_decode_need(p, end, info->inline_len, bad);
  98                 info->inline_data = *p;
  99                 *p += info->inline_len;
 100         } else
 101                 info->inline_version = CEPH_INLINE_NONE;
 102
 103         info->pool_ns_len = 0;
 104         info->pool_ns_data = NULL;
 105         if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
 106                 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
 107                 if (info->pool_ns_len > 0) {
 108                         ceph_decode_need(p, end, info->pool_ns_len, bad);
 109                         info->pool_ns_data = *p;
 110                         *p += info->pool_ns_len;
 111                 }
 112         }
 113
 114         return 0;
 115 bad:
 116         return err;
 117 }
 118
 119 /*
 120  * parse a normal reply, which may contain a (dir+)dentry and/or a
 121  * target inode.
 122  */
 123 static int parse_reply_info_trace(void **p, void *end,
 124                                   struct ceph_mds_reply_info_parsed *info,
 125                                   u64 features)
 126 {
 127         int err;
 128
 129         if (info->head->is_dentry) {
 130                 err = parse_reply_info_in(p, end, &info->diri, features);
 131                 if (err < 0)
 132                         goto out_bad;
 133
 134                 if (unlikely(*p + sizeof(*info->dirfrag) > end))
 135                         goto bad;
 136                 info->dirfrag = *p;
 137                 *p += sizeof(*info->dirfrag) +
 138                         sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
 139                 if (unlikely(*p > end))
 140                         goto bad;
 141
 142                 ceph_decode_32_safe(p, end, info->dname_len, bad);
 143                 ceph_decode_need(p, end, info->dname_len, bad);
 144                 info->dname = *p;
 145                 *p += info->dname_len;
 146                 info->dlease = *p;
 147                 *p += sizeof(*info->dlease);
 148         }
 149
 150         if (info->head->is_target) {
 151                 err = parse_reply_info_in(p, end, &info->targeti, features);
 152                 if (err < 0)
 153                         goto out_bad;
 154         }
 155
 156         if (unlikely(*p != end))
 157                 goto bad;
 158         return 0;
 159
 160 bad:
 161         err = -EIO;
 162 out_bad:
 163         pr_err("problem parsing mds trace %d\n", err);
 164         return err;
 165 }
 166
 167 /*
 168  * parse readdir results
 169  */
 170 static int parse_reply_info_dir(void **p, void *end,
 171                                 struct ceph_mds_reply_info_parsed *info,
 172                                 u64 features)
 173 {
 174         u32 num, i = 0;
 175         int err;
 176
 177         info->dir_dir = *p;
 178         if (*p + sizeof(*info->dir_dir) > end)
 179                 goto bad;
 180         *p += sizeof(*info->dir_dir) +
 181                 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
 182         if (*p > end)
 183                 goto bad;
 184
 185         ceph_decode_need(p, end, sizeof(num) + 2, bad);
 186         num = ceph_decode_32(p);
 187         {
 188                 u16 flags = ceph_decode_16(p);
 189                 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
 190                 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
 191                 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
 192                 info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
 193         }
 194         if (num == 0)
 195                 goto done;
 196
 197         BUG_ON(!info->dir_entries);
 198         if ((unsigned long)(info->dir_entries + num) >
 199             (unsigned long)info->dir_entries + info->dir_buf_size) {
 200                 pr_err("dir contents are larger than expected\n");
 201                 WARN_ON(1);
 202                 goto bad;
 203         }
 204
 205         info->dir_nr = num;
 206         while (num) {
 207                 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
 208                 /* dentry */
 209                 ceph_decode_need(p, end, sizeof(u32)*2, bad);
 210                 rde->name_len = ceph_decode_32(p);
 211                 ceph_decode_need(p, end, rde->name_len, bad);
 212                 rde->name = *p;
 213                 *p += rde->name_len;
 214                 dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
 215                 rde->lease = *p;
 216                 *p += sizeof(struct ceph_mds_reply_lease);
 217
 218                 /* inode */
 219                 err = parse_reply_info_in(p, end, &rde->inode, features);
 220                 if (err < 0)
 221                         goto out_bad;
 222                 /* ceph_readdir_prepopulate() will update it */
 223                 rde->offset = 0;
 224                 i++;
 225                 num--;
 226         }
 227
 228 done:
 229         if (*p != end)
 230                 goto bad;
 231         return 0;
 232
 233 bad:
 234         err = -EIO;
 235 out_bad:
 236         pr_err("problem parsing dir contents %d\n", err);
 237         return err;
 238 }
 239
 240 /*
 241  * parse fcntl F_GETLK results
 242  */
 243 static int parse_reply_info_filelock(void **p, void *end,
 244                                      struct ceph_mds_reply_info_parsed *info,
 245                                      u64 features)
 246 {
 247         if (*p + sizeof(*info->filelock_reply) > end)
 248                 goto bad;
 249
 250         info->filelock_reply = *p;
 251         *p += sizeof(*info->filelock_reply);
 252
 253         if (unlikely(*p != end))
 254                 goto bad;
 255         return 0;
 256
 257 bad:
 258         return -EIO;
 259 }
 260
 261 /*
 262  * parse create results
 263  */
 264 static int parse_reply_info_create(void **p, void *end,
 265                                   struct ceph_mds_reply_info_parsed *info,
 266                                   u64 features)
 267 {
 268         if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
 269                 if (*p == end) {
 270                         info->has_create_ino = false;
 271                 } else {
 272                         info->has_create_ino = true;
 273                         info->ino = ceph_decode_64(p);
 274                 }
 275         }
 276
 277         if (unlikely(*p != end))
 278                 goto bad;
 279         return 0;
 280
 281 bad:
 282         return -EIO;
 283 }
 284
 285 /*
 286  * parse extra results
 287  */
 288 static int parse_reply_info_extra(void **p, void *end,
 289                                   struct ceph_mds_reply_info_parsed *info,
 290                                   u64 features)
 291 {
 292         u32 op = le32_to_cpu(info->head->op);
 293
 294         if (op == CEPH_MDS_OP_GETFILELOCK)
 295                 return parse_reply_info_filelock(p, end, info, features);
 296         else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
 297                 return parse_reply_info_dir(p, end, info, features);
 298         else if (op == CEPH_MDS_OP_CREATE)
 299                 return parse_reply_info_create(p, end, info, features);
 300         else
 301                 return -EIO;
 302 }
 303
 304 /*
 305  * parse entire mds reply
 306  */
 307 static int parse_reply_info(struct ceph_msg *msg,
 308                             struct ceph_mds_reply_info_parsed *info,
 309                             u64 features)
 310 {
 311         void *p, *end;
 312         u32 len;
 313         int err;
 314
 315         info->head = msg->front.iov_base;
 316         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 317         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 318
 319         /* trace */
 320         ceph_decode_32_safe(&p, end, len, bad);
 321         if (len > 0) {
 322                 ceph_decode_need(&p, end, len, bad);
 323                 err = parse_reply_info_trace(&p, p+len, info, features);
 324                 if (err < 0)
 325                         goto out_bad;
 326         }
 327
 328         /* extra */
 329         ceph_decode_32_safe(&p, end, len, bad);
 330         if (len > 0) {
 331                 ceph_decode_need(&p, end, len, bad);
 332                 err = parse_reply_info_extra(&p, p+len, info, features);
 333                 if (err < 0)
 334                         goto out_bad;
 335         }
 336
 337         /* snap blob */
 338         ceph_decode_32_safe(&p, end, len, bad);
 339         info->snapblob_len = len;
 340         info->snapblob = p;
 341         p += len;
 342
 343         if (p != end)
 344                 goto bad;
 345         return 0;
 346
 347 bad:
 348         err = -EIO;
 349 out_bad:
 350         pr_err("mds parse_reply err %d\n", err);
 351         return err;
 352 }
 353
 354 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
 355 {
 356         if (!info->dir_entries)
 357                 return;
 358         free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
 359 }
 360
 361
 362 /*
 363  * sessions
 364  */
 365 const char *ceph_session_state_name(int s)
 366 {
 367         switch (s) {
 368         case CEPH_MDS_SESSION_NEW: return "new";
 369         case CEPH_MDS_SESSION_OPENING: return "opening";
 370         case CEPH_MDS_SESSION_OPEN: return "open";
 371         case CEPH_MDS_SESSION_HUNG: return "hung";
 372         case CEPH_MDS_SESSION_CLOSING: return "closing";
 373         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
 374         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
 375         case CEPH_MDS_SESSION_REJECTED: return "rejected";
 376         default: return "???";
 377         }
 378 }
 379
 380 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
 381 {
 382         if (refcount_inc_not_zero(&s->s_ref)) {
 383                 dout("mdsc get_session %p %d -> %d\n", s,
 384                      refcount_read(&s->s_ref)-1, refcount_read(&s->s_ref));
 385                 return s;
 386         } else {
 387                 dout("mdsc get_session %p 0 -- FAIL", s);
 388                 return NULL;
 389         }
 390 }
 391
 392 void ceph_put_mds_session(struct ceph_mds_session *s)
 393 {
 394         dout("mdsc put_session %p %d -> %d\n", s,
 395              refcount_read(&s->s_ref), refcount_read(&s->s_ref)-1);
 396         if (refcount_dec_and_test(&s->s_ref)) {
 397                 if (s->s_auth.authorizer)
 398                         ceph_auth_destroy_authorizer(s->s_auth.authorizer);
 399                 kfree(s);
 400         }
 401 }
 402
 403 /*
 404  * called under mdsc->mutex
 405  */
 406 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
 407                                                    int mds)
 408 {
 409         struct ceph_mds_session *session;
 410
 411         if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
 412                 return NULL;
 413         session = mdsc->sessions[mds];
 414         dout("lookup_mds_session %p %d\n", session,
 415              refcount_read(&session->s_ref));
 416         get_session(session);
 417         return session;
 418 }
 419
 420 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
 421 {
 422         if (mds >= mdsc->max_sessions)
 423                 return false;
 424         return mdsc->sessions[mds];
 425 }
 426
 427 static int __verify_registered_session(struct ceph_mds_client *mdsc,
 428                                        struct ceph_mds_session *s)
 429 {
 430         if (s->s_mds >= mdsc->max_sessions ||
 431             mdsc->sessions[s->s_mds] != s)
 432                 return -ENOENT;
 433         return 0;
 434 }
 435
 436 /*
 437  * create+register a new session for given mds.
 438  * called under mdsc->mutex.
 439  */
 440 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 441                                                  int mds)
 442 {
 443         struct ceph_mds_session *s;
 444
 445         if (mds >= mdsc->mdsmap->m_num_mds)
 446                 return ERR_PTR(-EINVAL);
 447
 448         s = kzalloc(sizeof(*s), GFP_NOFS);
 449         if (!s)
 450                 return ERR_PTR(-ENOMEM);
 451         s->s_mdsc = mdsc;
 452         s->s_mds = mds;
 453         s->s_state = CEPH_MDS_SESSION_NEW;
 454         s->s_ttl = 0;
 455         s->s_seq = 0;
 456         mutex_init(&s->s_mutex);
 457
 458         ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
 459
 460         spin_lock_init(&s->s_gen_ttl_lock);
 461         s->s_cap_gen = 0;
 462         s->s_cap_ttl = jiffies - 1;
 463
 464         spin_lock_init(&s->s_cap_lock);
 465         s->s_renew_requested = 0;
 466         s->s_renew_seq = 0;
 467         INIT_LIST_HEAD(&s->s_caps);
 468         s->s_nr_caps = 0;
 469         s->s_trim_caps = 0;
 470         refcount_set(&s->s_ref, 1);
 471         INIT_LIST_HEAD(&s->s_waiting);
 472         INIT_LIST_HEAD(&s->s_unsafe);
 473         s->s_num_cap_releases = 0;
 474         s->s_cap_reconnect = 0;
 475         s->s_cap_iterator = NULL;
 476         INIT_LIST_HEAD(&s->s_cap_releases);
 477         INIT_LIST_HEAD(&s->s_cap_flushing);
 478
 479         dout("register_session mds%d\n", mds);
 480         if (mds >= mdsc->max_sessions) {
 481                 int newmax = 1 << get_count_order(mds+1);
 482                 struct ceph_mds_session **sa;
 483
 484                 dout("register_session realloc to %d\n", newmax);
 485                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 486                 if (sa == NULL)
 487                         goto fail_realloc;
 488                 if (mdsc->sessions) {
 489                         memcpy(sa, mdsc->sessions,
 490                                mdsc->max_sessions * sizeof(void *));
 491                         kfree(mdsc->sessions);
 492                 }
 493                 mdsc->sessions = sa;
 494                 mdsc->max_sessions = newmax;
 495         }
 496         mdsc->sessions[mds] = s;
 497         atomic_inc(&mdsc->num_sessions);
 498         refcount_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
 499
 500         ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
 501                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
 502
 503         return s;
 504
 505 fail_realloc:
 506         kfree(s);
 507         return ERR_PTR(-ENOMEM);
 508 }
 509
 510 /*
 511  * called under mdsc->mutex
 512  */
 513 static void __unregister_session(struct ceph_mds_client *mdsc,
 514                                struct ceph_mds_session *s)
 515 {
 516         dout("__unregister_session mds%d %p\n", s->s_mds, s);
 517         BUG_ON(mdsc->sessions[s->s_mds] != s);
 518         mdsc->sessions[s->s_mds] = NULL;
 519         ceph_con_close(&s->s_con);
 520         ceph_put_mds_session(s);
 521         atomic_dec(&mdsc->num_sessions);
 522 }
 523
 524 /*
 525  * drop session refs in request.
 526  *
 527  * should be last request ref, or hold mdsc->mutex
 528  */
 529 static void put_request_session(struct ceph_mds_request *req)
 530 {
 531         if (req->r_session) {
 532                 ceph_put_mds_session(req->r_session);
 533                 req->r_session = NULL;
 534         }
 535 }
 536
 537 void ceph_mdsc_release_request(struct kref *kref)
 538 {
 539         struct ceph_mds_request *req = container_of(kref,
 540                                                     struct ceph_mds_request,
 541                                                     r_kref);
 542         destroy_reply_info(&req->r_reply_info);
 543         if (req->r_request)
 544                 ceph_msg_put(req->r_request);
 545         if (req->r_reply)
 546                 ceph_msg_put(req->r_reply);
 547         if (req->r_inode) {
 548                 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
 549                 iput(req->r_inode);
 550         }
 551         if (req->r_parent)
 552                 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
 553         iput(req->r_target_inode);
 554         if (req->r_dentry)
 555                 dput(req->r_dentry);
 556         if (req->r_old_dentry)
 557                 dput(req->r_old_dentry);
 558         if (req->r_old_dentry_dir) {
 559                 /*
 560                  * track (and drop pins for) r_old_dentry_dir
 561                  * separately, since r_old_dentry's d_parent may have
 562                  * changed between the dir mutex being dropped and
 563                  * this request being freed.
 564                  */
 565                 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
 566                                   CEPH_CAP_PIN);
 567                 iput(req->r_old_dentry_dir);
 568         }
 569         kfree(req->r_path1);
 570         kfree(req->r_path2);
 571         if (req->r_pagelist)
 572                 ceph_pagelist_release(req->r_pagelist);
 573         put_request_session(req);
 574         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
 575         kfree(req);
 576 }
 577
 578 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
 579
 580 /*
 581  * lookup session, bump ref if found.
 582  *
 583  * called under mdsc->mutex.
 584  */
 585 static struct ceph_mds_request *
 586 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
 587 {
 588         struct ceph_mds_request *req;
 589
 590         req = lookup_request(&mdsc->request_tree, tid);
 591         if (req)
 592                 ceph_mdsc_get_request(req);
 593
 594         return req;
 595 }
 596
 597 /*
 598  * Register an in-flight request, and assign a tid.  Link to directory
 599  * are modifying (if any).
 600  *
 601  * Called under mdsc->mutex.
 602  */
 603 static void __register_request(struct ceph_mds_client *mdsc,
 604                                struct ceph_mds_request *req,
 605                                struct inode *dir)
 606 {
 607         req->r_tid = ++mdsc->last_tid;
 608         if (req->r_num_caps)
 609                 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
 610                                   req->r_num_caps);
 611         dout("__register_request %p tid %lld\n", req, req->r_tid);
 612         ceph_mdsc_get_request(req);
 613         insert_request(&mdsc->request_tree, req);
 614
 615         req->r_uid = current_fsuid();
 616         req->r_gid = current_fsgid();
 617
 618         if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
 619                 mdsc->oldest_tid = req->r_tid;
 620
 621         if (dir) {
 622                 ihold(dir);
 623                 req->r_unsafe_dir = dir;
 624         }
 625 }
 626
 627 static void __unregister_request(struct ceph_mds_client *mdsc,
 628                                  struct ceph_mds_request *req)
 629 {
 630         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 631
 632         /* Never leave an unregistered request on an unsafe list! */
 633         list_del_init(&req->r_unsafe_item);
 634
 635         if (req->r_tid == mdsc->oldest_tid) {
 636                 struct rb_node *p = rb_next(&req->r_node);
 637                 mdsc->oldest_tid = 0;
 638                 while (p) {
 639                         struct ceph_mds_request *next_req =
 640                                 rb_entry(p, struct ceph_mds_request, r_node);
 641                         if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
 642                                 mdsc->oldest_tid = next_req->r_tid;
 643                                 break;
 644                         }
 645                         p = rb_next(p);
 646                 }
 647         }
 648
 649         erase_request(&mdsc->request_tree, req);
 650
 651         if (req->r_unsafe_dir  &&
 652             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
 653                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 654                 spin_lock(&ci->i_unsafe_lock);
 655                 list_del_init(&req->r_unsafe_dir_item);
 656                 spin_unlock(&ci->i_unsafe_lock);
 657         }
 658         if (req->r_target_inode &&
 659             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
 660                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
 661                 spin_lock(&ci->i_unsafe_lock);
 662                 list_del_init(&req->r_unsafe_target_item);
 663                 spin_unlock(&ci->i_unsafe_lock);
 664         }
 665
 666         if (req->r_unsafe_dir) {
 667                 iput(req->r_unsafe_dir);
 668                 req->r_unsafe_dir = NULL;
 669         }
 670
 671         complete_all(&req->r_safe_completion);
 672
 673         ceph_mdsc_put_request(req);
 674 }
 675
 676 /*
 677  * Walk back up the dentry tree until we hit a dentry representing a
 678  * non-snapshot inode. We do this using the rcu_read_lock (which must be held
 679  * when calling this) to ensure that the objects won't disappear while we're
 680  * working with them. Once we hit a candidate dentry, we attempt to take a
 681  * reference to it, and return that as the result.
 682  */
 683 static struct inode *get_nonsnap_parent(struct dentry *dentry)
 684 {
 685         struct inode *inode = NULL;
 686
 687         while (dentry && !IS_ROOT(dentry)) {
 688                 inode = d_inode_rcu(dentry);
 689                 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
 690                         break;
 691                 dentry = dentry->d_parent;
 692         }
 693         if (inode)
 694                 inode = igrab(inode);
 695         return inode;
 696 }
 697
 698 /*
 699  * Choose mds to send request to next.  If there is a hint set in the
 700  * request (e.g., due to a prior forward hint from the mds), use that.
 701  * Otherwise, consult frag tree and/or caps to identify the
 702  * appropriate mds.  If all else fails, choose randomly.
 703  *
 704  * Called under mdsc->mutex.
 705  */
 706 static int __choose_mds(struct ceph_mds_client *mdsc,
 707                         struct ceph_mds_request *req)
 708 {
 709         struct inode *inode;
 710         struct ceph_inode_info *ci;
 711         struct ceph_cap *cap;
 712         int mode = req->r_direct_mode;
 713         int mds = -1;
 714         u32 hash = req->r_direct_hash;
 715         bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
 716
 717         /*
 718          * is there a specific mds we should try?  ignore hint if we have
 719          * no session and the mds is not up (active or recovering).
 720          */
 721         if (req->r_resend_mds >= 0 &&
 722             (__have_session(mdsc, req->r_resend_mds) ||
 723              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 724                 dout("choose_mds using resend_mds mds%d\n",
 725                      req->r_resend_mds);
 726                 return req->r_resend_mds;
 727         }
 728
 729         if (mode == USE_RANDOM_MDS)
 730                 goto random;
 731
 732         inode = NULL;
 733         if (req->r_inode) {
 734                 inode = req->r_inode;
 735                 ihold(inode);
 736         } else if (req->r_dentry) {
 737                 /* ignore race with rename; old or new d_parent is okay */
 738                 struct dentry *parent;
 739                 struct inode *dir;
 740
 741                 rcu_read_lock();
 742                 parent = req->r_dentry->d_parent;
 743                 dir = req->r_parent ? : d_inode_rcu(parent);
 744
 745                 if (!dir || dir->i_sb != mdsc->fsc->sb) {
 746                         /*  not this fs or parent went negative */
 747                         inode = d_inode(req->r_dentry);
 748                         if (inode)
 749                                 ihold(inode);
 750                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
 751                         /* direct snapped/virtual snapdir requests
 752                          * based on parent dir inode */
 753                         inode = get_nonsnap_parent(parent);
 754                         dout("__choose_mds using nonsnap parent %p\n", inode);
 755                 } else {
 756                         /* dentry target */
 757                         inode = d_inode(req->r_dentry);
 758                         if (!inode || mode == USE_AUTH_MDS) {
 759                                 /* dir + name */
 760                                 inode = igrab(dir);
 761                                 hash = ceph_dentry_hash(dir, req->r_dentry);
 762                                 is_hash = true;
 763                         } else {
 764                                 ihold(inode);
 765                         }
 766                 }
 767                 rcu_read_unlock();
 768         }
 769
 770         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 771              (int)hash, mode);
 772         if (!inode)
 773                 goto random;
 774         ci = ceph_inode(inode);
 775
 776         if (is_hash && S_ISDIR(inode->i_mode)) {
 777                 struct ceph_inode_frag frag;
 778                 int found;
 779
 780                 ceph_choose_frag(ci, hash, &frag, &found);
 781                 if (found) {
 782                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 783                                 u8 r;
 784
 785                                 /* choose a random replica */
 786                                 get_random_bytes(&r, 1);
 787                                 r %= frag.ndist;
 788                                 mds = frag.dist[r];
 789                                 dout("choose_mds %p %llx.%llx "
 790                                      "frag %u mds%d (%d/%d)\n",
 791                                      inode, ceph_vinop(inode),
 792                                      frag.frag, mds,
 793                                      (int)r, frag.ndist);
 794                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 795                                     CEPH_MDS_STATE_ACTIVE)
 796                                         goto out;
 797                         }
 798
 799                         /* since this file/dir wasn't known to be
 800                          * replicated, then we want to look for the
 801                          * authoritative mds. */
 802                         mode = USE_AUTH_MDS;
 803                         if (frag.mds >= 0) {
 804                                 /* choose auth mds */
 805                                 mds = frag.mds;
 806                                 dout("choose_mds %p %llx.%llx "
 807                                      "frag %u mds%d (auth)\n",
 808                                      inode, ceph_vinop(inode), frag.frag, mds);
 809                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 810                                     CEPH_MDS_STATE_ACTIVE)
 811                                         goto out;
 812                         }
 813                 }
 814         }
 815
 816         spin_lock(&ci->i_ceph_lock);
 817         cap = NULL;
 818         if (mode == USE_AUTH_MDS)
 819                 cap = ci->i_auth_cap;
 820         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
 821                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
 822         if (!cap) {
 823                 spin_unlock(&ci->i_ceph_lock);
 824                 iput(inode);
 825                 goto random;
 826         }
 827         mds = cap->session->s_mds;
 828         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
 829              inode, ceph_vinop(inode), mds,
 830              cap == ci->i_auth_cap ? "auth " : "", cap);
 831         spin_unlock(&ci->i_ceph_lock);
 832 out:
 833         iput(inode);
 834         return mds;
 835
 836 random:
 837         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
 838         dout("choose_mds chose random mds%d\n", mds);
 839         return mds;
 840 }
 841
 842
 843 /*
 844  * session messages
 845  */
 846 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
 847 {
 848         struct ceph_msg *msg;
 849         struct ceph_mds_session_head *h;
 850
 851         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
 852                            false);
 853         if (!msg) {
 854                 pr_err("create_session_msg ENOMEM creating msg\n");
 855                 return NULL;
 856         }
 857         h = msg->front.iov_base;
 858         h->op = cpu_to_le32(op);
 859         h->seq = cpu_to_le64(seq);
 860
 861         return msg;
 862 }
 863
 864 /*
 865  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
 866  * to include additional client metadata fields.
 867  */
 868 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
 869 {
 870         struct ceph_msg *msg;
 871         struct ceph_mds_session_head *h;
 872         int i = -1;
 873         int metadata_bytes = 0;
 874         int metadata_key_count = 0;
 875         struct ceph_options *opt = mdsc->fsc->client->options;
 876         struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
 877         void *p;
 878
 879         const char* metadata[][2] = {
 880                 {"hostname", utsname()->nodename},
 881                 {"kernel_version", utsname()->release},
 882                 {"entity_id", opt->name ? : ""},
 883                 {"root", fsopt->server_path ? : "/"},
 884                 {NULL, NULL}
 885         };
 886
 887         /* Calculate serialized length of metadata */
 888         metadata_bytes = 4;  /* map length */
 889         for (i = 0; metadata[i][0] != NULL; ++i) {
 890                 metadata_bytes += 8 + strlen(metadata[i][0]) +
 891                         strlen(metadata[i][1]);
 892                 metadata_key_count++;
 893         }
 894
 895         /* Allocate the message */
 896         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + metadata_bytes,
 897                            GFP_NOFS, false);
 898         if (!msg) {
 899                 pr_err("create_session_msg ENOMEM creating msg\n");
 900                 return NULL;
 901         }
 902         h = msg->front.iov_base;
 903         h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
 904         h->seq = cpu_to_le64(seq);
 905
 906         /*
 907          * Serialize client metadata into waiting buffer space, using
 908          * the format that userspace expects for map<string, string>
 909          *
 910          * ClientSession messages with metadata are v2
 911          */
 912         msg->hdr.version = cpu_to_le16(2);
 913         msg->hdr.compat_version = cpu_to_le16(1);
 914
 915         /* The write pointer, following the session_head structure */
 916         p = msg->front.iov_base + sizeof(*h);
 917
 918         /* Number of entries in the map */
 919         ceph_encode_32(&p, metadata_key_count);
 920
 921         /* Two length-prefixed strings for each entry in the map */
 922         for (i = 0; metadata[i][0] != NULL; ++i) {
 923                 size_t const key_len = strlen(metadata[i][0]);
 924                 size_t const val_len = strlen(metadata[i][1]);
 925
 926                 ceph_encode_32(&p, key_len);
 927                 memcpy(p, metadata[i][0], key_len);
 928                 p += key_len;
 929                 ceph_encode_32(&p, val_len);
 930                 memcpy(p, metadata[i][1], val_len);
 931                 p += val_len;
 932         }
 933
 934         return msg;
 935 }
 936
 937 /*
 938  * send session open request.
 939  *
 940  * called under mdsc->mutex
 941  */
 942 static int __open_session(struct ceph_mds_client *mdsc,
 943                           struct ceph_mds_session *session)
 944 {
 945         struct ceph_msg *msg;
 946         int mstate;
 947         int mds = session->s_mds;
 948
 949         /* wait for mds to go active? */
 950         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
 951         dout("open_session to mds%d (%s)\n", mds,
 952              ceph_mds_state_name(mstate));
 953         session->s_state = CEPH_MDS_SESSION_OPENING;
 954         session->s_renew_requested = jiffies;
 955
 956         /* send connect message */
 957         msg = create_session_open_msg(mdsc, session->s_seq);
 958         if (!msg)
 959                 return -ENOMEM;
 960         ceph_con_send(&session->s_con, msg);
 961         return 0;
 962 }
 963
 964 /*
 965  * open sessions for any export targets for the given mds
 966  *
 967  * called under mdsc->mutex
 968  */
 969 static struct ceph_mds_session *
 970 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
 971 {
 972         struct ceph_mds_session *session;
 973
 974         session = __ceph_lookup_mds_session(mdsc, target);
 975         if (!session) {
 976                 session = register_session(mdsc, target);
 977                 if (IS_ERR(session))
 978                         return session;
 979         }
 980         if (session->s_state == CEPH_MDS_SESSION_NEW ||
 981             session->s_state == CEPH_MDS_SESSION_CLOSING)
 982                 __open_session(mdsc, session);
 983
 984         return session;
 985 }
 986
 987 struct ceph_mds_session *
 988 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
 989 {
 990         struct ceph_mds_session *session;
 991
 992         dout("open_export_target_session to mds%d\n", target);
 993
 994         mutex_lock(&mdsc->mutex);
 995         session = __open_export_target_session(mdsc, target);
 996         mutex_unlock(&mdsc->mutex);
 997
 998         return session;
 999 }
1000
1001 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1002                                           struct ceph_mds_session *session)
1003 {
1004         struct ceph_mds_info *mi;
1005         struct ceph_mds_session *ts;
1006         int i, mds = session->s_mds;
1007
1008         if (mds >= mdsc->mdsmap->m_num_mds)
1009                 return;
1010
1011         mi = &mdsc->mdsmap->m_info[mds];
1012         dout("open_export_target_sessions for mds%d (%d targets)\n",
1013              session->s_mds, mi->num_export_targets);
1014
1015         for (i = 0; i < mi->num_export_targets; i++) {
1016                 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1017                 if (!IS_ERR(ts))
1018                         ceph_put_mds_session(ts);
1019         }
1020 }
1021
1022 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1023                                            struct ceph_mds_session *session)
1024 {
1025         mutex_lock(&mdsc->mutex);
1026         __open_export_target_sessions(mdsc, session);
1027         mutex_unlock(&mdsc->mutex);
1028 }
1029
1030 /*
1031  * session caps
1032  */
1033
1034 /* caller holds s_cap_lock, we drop it */
1035 static void cleanup_cap_releases(struct ceph_mds_client *mdsc,
1036                                  struct ceph_mds_session *session)
1037         __releases(session->s_cap_lock)
1038 {
1039         LIST_HEAD(tmp_list);
1040         list_splice_init(&session->s_cap_releases, &tmp_list);
1041         session->s_num_cap_releases = 0;
1042         spin_unlock(&session->s_cap_lock);
1043
1044         dout("cleanup_cap_releases mds%d\n", session->s_mds);
1045         while (!list_empty(&tmp_list)) {
1046                 struct ceph_cap *cap;
1047                 /* zero out the in-progress message */
1048                 cap = list_first_entry(&tmp_list,
1049                                         struct ceph_cap, session_caps);
1050                 list_del(&cap->session_caps);
1051                 ceph_put_cap(mdsc, cap);
1052         }
1053 }
1054
1055 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1056                                      struct ceph_mds_session *session)
1057 {
1058         struct ceph_mds_request *req;
1059         struct rb_node *p;
1060
1061         dout("cleanup_session_requests mds%d\n", session->s_mds);
1062         mutex_lock(&mdsc->mutex);
1063         while (!list_empty(&session->s_unsafe)) {
1064                 req = list_first_entry(&session->s_unsafe,
1065                                        struct ceph_mds_request, r_unsafe_item);
1066                 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1067                                     req->r_tid);
1068                 __unregister_request(mdsc, req);
1069         }
1070         /* zero r_attempts, so kick_requests() will re-send requests */
1071         p = rb_first(&mdsc->request_tree);
1072         while (p) {
1073                 req = rb_entry(p, struct ceph_mds_request, r_node);
1074                 p = rb_next(p);
1075                 if (req->r_session &&
1076                     req->r_session->s_mds == session->s_mds)
1077                         req->r_attempts = 0;
1078         }
1079         mutex_unlock(&mdsc->mutex);
1080 }
1081
1082 /*
1083  * Helper to safely iterate over all caps associated with a session, with
1084  * special care taken to handle a racing __ceph_remove_cap().
1085  *
1086  * Caller must hold session s_mutex.
1087  */
1088 static int iterate_session_caps(struct ceph_mds_session *session,
1089                                  int (*cb)(struct inode *, struct ceph_cap *,
1090                                             void *), void *arg)
1091 {
1092         struct list_head *p;
1093         struct ceph_cap *cap;
1094         struct inode *inode, *last_inode = NULL;
1095         struct ceph_cap *old_cap = NULL;
1096         int ret;
1097
1098         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1099         spin_lock(&session->s_cap_lock);
1100         p = session->s_caps.next;
1101         while (p != &session->s_caps) {
1102                 cap = list_entry(p, struct ceph_cap, session_caps);
1103                 inode = igrab(&cap->ci->vfs_inode);
1104                 if (!inode) {
1105                         p = p->next;
1106                         continue;
1107                 }
1108                 session->s_cap_iterator = cap;
1109                 spin_unlock(&session->s_cap_lock);
1110
1111                 if (last_inode) {
1112                         iput(last_inode);
1113                         last_inode = NULL;
1114                 }
1115                 if (old_cap) {
1116                         ceph_put_cap(session->s_mdsc, old_cap);
1117                         old_cap = NULL;
1118                 }
1119
1120                 ret = cb(inode, cap, arg);
1121                 last_inode = inode;
1122
1123                 spin_lock(&session->s_cap_lock);
1124                 p = p->next;
1125                 if (cap->ci == NULL) {
1126                         dout("iterate_session_caps  finishing cap %p removal\n",
1127                              cap);
1128                         BUG_ON(cap->session != session);
1129                         cap->session = NULL;
1130                         list_del_init(&cap->session_caps);
1131                         session->s_nr_caps--;
1132                         if (cap->queue_release) {
1133                                 list_add_tail(&cap->session_caps,
1134                                               &session->s_cap_releases);
1135                                 session->s_num_cap_releases++;
1136                         } else {
1137                                 old_cap = cap;  /* put_cap it w/o locks held */
1138                         }
1139                 }
1140                 if (ret < 0)
1141                         goto out;
1142         }
1143         ret = 0;
1144 out:
1145         session->s_cap_iterator = NULL;
1146         spin_unlock(&session->s_cap_lock);
1147
1148         iput(last_inode);
1149         if (old_cap)
1150                 ceph_put_cap(session->s_mdsc, old_cap);
1151
1152         return ret;
1153 }
1154
1155 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1156                                   void *arg)
1157 {
1158         struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1159         struct ceph_inode_info *ci = ceph_inode(inode);
1160         LIST_HEAD(to_remove);
1161         bool drop = false;
1162         bool invalidate = false;
1163
1164         dout("removing cap %p, ci is %p, inode is %p\n",
1165              cap, ci, &ci->vfs_inode);
1166         spin_lock(&ci->i_ceph_lock);
1167         __ceph_remove_cap(cap, false);
1168         if (!ci->i_auth_cap) {
1169                 struct ceph_cap_flush *cf;
1170                 struct ceph_mds_client *mdsc = fsc->mdsc;
1171
1172                 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1173
1174                 if (ci->i_wrbuffer_ref > 0 &&
1175                     READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
1176                         invalidate = true;
1177
1178                 while (!list_empty(&ci->i_cap_flush_list)) {
1179                         cf = list_first_entry(&ci->i_cap_flush_list,
1180                                               struct ceph_cap_flush, i_list);
1181                         list_move(&cf->i_list, &to_remove);
1182                 }
1183
1184                 spin_lock(&mdsc->cap_dirty_lock);
1185
1186                 list_for_each_entry(cf, &to_remove, i_list)
1187                         list_del(&cf->g_list);
1188
1189                 if (!list_empty(&ci->i_dirty_item)) {
1190                         pr_warn_ratelimited(
1191                                 " dropping dirty %s state for %p %lld\n",
1192                                 ceph_cap_string(ci->i_dirty_caps),
1193                                 inode, ceph_ino(inode));
1194                         ci->i_dirty_caps = 0;
1195                         list_del_init(&ci->i_dirty_item);
1196                         drop = true;
1197                 }
1198                 if (!list_empty(&ci->i_flushing_item)) {
1199                         pr_warn_ratelimited(
1200                                 " dropping dirty+flushing %s state for %p %lld\n",
1201                                 ceph_cap_string(ci->i_flushing_caps),
1202                                 inode, ceph_ino(inode));
1203                         ci->i_flushing_caps = 0;
1204                         list_del_init(&ci->i_flushing_item);
1205                         mdsc->num_cap_flushing--;
1206                         drop = true;
1207                 }
1208                 spin_unlock(&mdsc->cap_dirty_lock);
1209
1210                 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1211                         list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1212                         ci->i_prealloc_cap_flush = NULL;
1213                 }
1214         }
1215         spin_unlock(&ci->i_ceph_lock);
1216         while (!list_empty(&to_remove)) {
1217                 struct ceph_cap_flush *cf;
1218                 cf = list_first_entry(&to_remove,
1219                                       struct ceph_cap_flush, i_list);
1220                 list_del(&cf->i_list);
1221                 ceph_free_cap_flush(cf);
1222         }
1223
1224         wake_up_all(&ci->i_cap_wq);
1225         if (invalidate)
1226                 ceph_queue_invalidate(inode);
1227         if (drop)
1228                 iput(inode);
1229         return 0;
1230 }
1231
1232 /*
1233  * caller must hold session s_mutex
1234  */
1235 static void remove_session_caps(struct ceph_mds_session *session)
1236 {
1237         struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1238         struct super_block *sb = fsc->sb;
1239         dout("remove_session_caps on %p\n", session);
1240         iterate_session_caps(session, remove_session_caps_cb, fsc);
1241
1242         wake_up_all(&fsc->mdsc->cap_flushing_wq);
1243
1244         spin_lock(&session->s_cap_lock);
1245         if (session->s_nr_caps > 0) {
1246                 struct inode *inode;
1247                 struct ceph_cap *cap, *prev = NULL;
1248                 struct ceph_vino vino;
1249                 /*
1250                  * iterate_session_caps() skips inodes that are being
1251                  * deleted, we need to wait until deletions are complete.
1252                  * __wait_on_freeing_inode() is designed for the job,
1253                  * but it is not exported, so use lookup inode function
1254                  * to access it.
1255                  */
1256                 while (!list_empty(&session->s_caps)) {
1257                         cap = list_entry(session->s_caps.next,
1258                                          struct ceph_cap, session_caps);
1259                         if (cap == prev)
1260                                 break;
1261                         prev = cap;
1262                         vino = cap->ci->i_vino;
1263                         spin_unlock(&session->s_cap_lock);
1264
1265                         inode = ceph_find_inode(sb, vino);
1266                         iput(inode);
1267
1268                         spin_lock(&session->s_cap_lock);
1269                 }
1270         }
1271
1272         // drop cap expires and unlock s_cap_lock
1273         cleanup_cap_releases(session->s_mdsc, session);
1274
1275         BUG_ON(session->s_nr_caps > 0);
1276         BUG_ON(!list_empty(&session->s_cap_flushing));
1277 }
1278
1279 /*
1280  * wake up any threads waiting on this session's caps.  if the cap is
1281  * old (didn't get renewed on the client reconnect), remove it now.
1282  *
1283  * caller must hold s_mutex.
1284  */
1285 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1286                               void *arg)
1287 {
1288         struct ceph_inode_info *ci = ceph_inode(inode);
1289
1290         if (arg) {
1291                 spin_lock(&ci->i_ceph_lock);
1292                 ci->i_wanted_max_size = 0;
1293                 ci->i_requested_max_size = 0;
1294                 spin_unlock(&ci->i_ceph_lock);
1295         }
1296         wake_up_all(&ci->i_cap_wq);
1297         return 0;
1298 }
1299
1300 static void wake_up_session_caps(struct ceph_mds_session *session,
1301                                  int reconnect)
1302 {
1303         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1304         iterate_session_caps(session, wake_up_session_cb,
1305                              (void *)(unsigned long)reconnect);
1306 }
1307
1308 /*
1309  * Send periodic message to MDS renewing all currently held caps.  The
1310  * ack will reset the expiration for all caps from this session.
1311  *
1312  * caller holds s_mutex
1313  */
1314 static int send_renew_caps(struct ceph_mds_client *mdsc,
1315                            struct ceph_mds_session *session)
1316 {
1317         struct ceph_msg *msg;
1318         int state;
1319
1320         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1321             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1322                 pr_info("mds%d caps stale\n", session->s_mds);
1323         session->s_renew_requested = jiffies;
1324
1325         /* do not try to renew caps until a recovering mds has reconnected
1326          * with its clients. */
1327         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1328         if (state < CEPH_MDS_STATE_RECONNECT) {
1329                 dout("send_renew_caps ignoring mds%d (%s)\n",
1330                      session->s_mds, ceph_mds_state_name(state));
1331                 return 0;
1332         }
1333
1334         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1335                 ceph_mds_state_name(state));
1336         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1337                                  ++session->s_renew_seq);
1338         if (!msg)
1339                 return -ENOMEM;
1340         ceph_con_send(&session->s_con, msg);
1341         return 0;
1342 }
1343
1344 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1345                              struct ceph_mds_session *session, u64 seq)
1346 {
1347         struct ceph_msg *msg;
1348
1349         dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1350              session->s_mds, ceph_session_state_name(session->s_state), seq);
1351         msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1352         if (!msg)
1353                 return -ENOMEM;
1354         ceph_con_send(&session->s_con, msg);
1355         return 0;
1356 }
1357
1358
1359 /*
1360  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1361  *
1362  * Called under session->s_mutex
1363  */
1364 static void renewed_caps(struct ceph_mds_client *mdsc,
1365                          struct ceph_mds_session *session, int is_renew)
1366 {
1367         int was_stale;
1368         int wake = 0;
1369
1370         spin_lock(&session->s_cap_lock);
1371         was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1372
1373         session->s_cap_ttl = session->s_renew_requested +
1374                 mdsc->mdsmap->m_session_timeout*HZ;
1375
1376         if (was_stale) {
1377                 if (time_before(jiffies, session->s_cap_ttl)) {
1378                         pr_info("mds%d caps renewed\n", session->s_mds);
1379                         wake = 1;
1380                 } else {
1381                         pr_info("mds%d caps still stale\n", session->s_mds);
1382                 }
1383         }
1384         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1385              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1386              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1387         spin_unlock(&session->s_cap_lock);
1388
1389         if (wake)
1390                 wake_up_session_caps(session, 0);
1391 }
1392
1393 /*
1394  * send a session close request
1395  */
1396 static int request_close_session(struct ceph_mds_client *mdsc,
1397                                  struct ceph_mds_session *session)
1398 {
1399         struct ceph_msg *msg;
1400
1401         dout("request_close_session mds%d state %s seq %lld\n",
1402              session->s_mds, ceph_session_state_name(session->s_state),
1403              session->s_seq);
1404         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1405         if (!msg)
1406                 return -ENOMEM;
1407         ceph_con_send(&session->s_con, msg);
1408         return 1;
1409 }
1410
1411 /*
1412  * Called with s_mutex held.
1413  */
1414 static int __close_session(struct ceph_mds_client *mdsc,
1415                          struct ceph_mds_session *session)
1416 {
1417         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1418                 return 0;
1419         session->s_state = CEPH_MDS_SESSION_CLOSING;
1420         return request_close_session(mdsc, session);
1421 }
1422
1423 /*
1424  * Trim old(er) caps.
1425  *
1426  * Because we can't cache an inode without one or more caps, we do
1427  * this indirectly: if a cap is unused, we prune its aliases, at which
1428  * point the inode will hopefully get dropped to.
1429  *
1430  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1431  * memory pressure from the MDS, though, so it needn't be perfect.
1432  */
1433 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1434 {
1435         struct ceph_mds_session *session = arg;
1436         struct ceph_inode_info *ci = ceph_inode(inode);
1437         int used, wanted, oissued, mine;
1438
1439         if (session->s_trim_caps <= 0)
1440                 return -1;
1441
1442         spin_lock(&ci->i_ceph_lock);
1443         mine = cap->issued | cap->implemented;
1444         used = __ceph_caps_used(ci);
1445         wanted = __ceph_caps_file_wanted(ci);
1446         oissued = __ceph_caps_issued_other(ci, cap);
1447
1448         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1449              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1450              ceph_cap_string(used), ceph_cap_string(wanted));
1451         if (cap == ci->i_auth_cap) {
1452                 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1453                     !list_empty(&ci->i_cap_snaps))
1454                         goto out;
1455                 if ((used | wanted) & CEPH_CAP_ANY_WR)
1456                         goto out;
1457         }
1458         /* The inode has cached pages, but it's no longer used.
1459          * we can safely drop it */
1460         if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1461             !(oissued & CEPH_CAP_FILE_CACHE)) {
1462           used = 0;
1463           oissued = 0;
1464         }
1465         if ((used | wanted) & ~oissued & mine)
1466                 goto out;   /* we need these caps */
1467
1468         session->s_trim_caps--;
1469         if (oissued) {
1470                 /* we aren't the only cap.. just remove us */
1471                 __ceph_remove_cap(cap, true);
1472         } else {
1473                 /* try dropping referring dentries */
1474                 spin_unlock(&ci->i_ceph_lock);
1475                 d_prune_aliases(inode);
1476                 dout("trim_caps_cb %p cap %p  pruned, count now %d\n",
1477                      inode, cap, atomic_read(&inode->i_count));
1478                 return 0;
1479         }
1480
1481 out:
1482         spin_unlock(&ci->i_ceph_lock);
1483         return 0;
1484 }
1485
1486 /*
1487  * Trim session cap count down to some max number.
1488  */
1489 static int trim_caps(struct ceph_mds_client *mdsc,
1490                      struct ceph_mds_session *session,
1491                      int max_caps)
1492 {
1493         int trim_caps = session->s_nr_caps - max_caps;
1494
1495         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1496              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1497         if (trim_caps > 0) {
1498                 session->s_trim_caps = trim_caps;
1499                 iterate_session_caps(session, trim_caps_cb, session);
1500                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1501                      session->s_mds, session->s_nr_caps, max_caps,
1502                         trim_caps - session->s_trim_caps);
1503                 session->s_trim_caps = 0;
1504         }
1505
1506         ceph_send_cap_releases(mdsc, session);
1507         return 0;
1508 }
1509
1510 static int check_caps_flush(struct ceph_mds_client *mdsc,
1511                             u64 want_flush_tid)
1512 {
1513         int ret = 1;
1514
1515         spin_lock(&mdsc->cap_dirty_lock);
1516         if (!list_empty(&mdsc->cap_flush_list)) {
1517                 struct ceph_cap_flush *cf =
1518                         list_first_entry(&mdsc->cap_flush_list,
1519                                          struct ceph_cap_flush, g_list);
1520                 if (cf->tid <= want_flush_tid) {
1521                         dout("check_caps_flush still flushing tid "
1522                              "%llu <= %llu\n", cf->tid, want_flush_tid);
1523                         ret = 0;
1524                 }
1525         }
1526         spin_unlock(&mdsc->cap_dirty_lock);
1527         return ret;
1528 }
1529
1530 /*
1531  * flush all dirty inode data to disk.
1532  *
1533  * returns true if we've flushed through want_flush_tid
1534  */
1535 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1536                             u64 want_flush_tid)
1537 {
1538         dout("check_caps_flush want %llu\n", want_flush_tid);
1539
1540         wait_event(mdsc->cap_flushing_wq,
1541                    check_caps_flush(mdsc, want_flush_tid));
1542
1543         dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1544 }
1545
1546 /*
1547  * called under s_mutex
1548  */
1549 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1550                             struct ceph_mds_session *session)
1551 {
1552         struct ceph_msg *msg = NULL;
1553         struct ceph_mds_cap_release *head;
1554         struct ceph_mds_cap_item *item;
1555         struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
1556         struct ceph_cap *cap;
1557         LIST_HEAD(tmp_list);
1558         int num_cap_releases;
1559         __le32  barrier, *cap_barrier;
1560
1561         down_read(&osdc->lock);
1562         barrier = cpu_to_le32(osdc->epoch_barrier);
1563         up_read(&osdc->lock);
1564
1565         spin_lock(&session->s_cap_lock);
1566 again:
1567         list_splice_init(&session->s_cap_releases, &tmp_list);
1568         num_cap_releases = session->s_num_cap_releases;
1569         session->s_num_cap_releases = 0;
1570         spin_unlock(&session->s_cap_lock);
1571
1572         while (!list_empty(&tmp_list)) {
1573                 if (!msg) {
1574                         msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1575                                         PAGE_SIZE, GFP_NOFS, false);
1576                         if (!msg)
1577                                 goto out_err;
1578                         head = msg->front.iov_base;
1579                         head->num = cpu_to_le32(0);
1580                         msg->front.iov_len = sizeof(*head);
1581
1582                         msg->hdr.version = cpu_to_le16(2);
1583                         msg->hdr.compat_version = cpu_to_le16(1);
1584                 }
1585
1586                 cap = list_first_entry(&tmp_list, struct ceph_cap,
1587                                         session_caps);
1588                 list_del(&cap->session_caps);
1589                 num_cap_releases--;
1590
1591                 head = msg->front.iov_base;
1592                 le32_add_cpu(&head->num, 1);
1593                 item = msg->front.iov_base + msg->front.iov_len;
1594                 item->ino = cpu_to_le64(cap->cap_ino);
1595                 item->cap_id = cpu_to_le64(cap->cap_id);
1596                 item->migrate_seq = cpu_to_le32(cap->mseq);
1597                 item->seq = cpu_to_le32(cap->issue_seq);
1598                 msg->front.iov_len += sizeof(*item);
1599
1600                 ceph_put_cap(mdsc, cap);
1601
1602                 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1603                         // Append cap_barrier field
1604                         cap_barrier = msg->front.iov_base + msg->front.iov_len;
1605                         *cap_barrier = barrier;
1606                         msg->front.iov_len += sizeof(*cap_barrier);
1607
1608                         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1609                         dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1610                         ceph_con_send(&session->s_con, msg);
1611                         msg = NULL;
1612                 }
1613         }
1614
1615         BUG_ON(num_cap_releases != 0);
1616
1617         spin_lock(&session->s_cap_lock);
1618         if (!list_empty(&session->s_cap_releases))
1619                 goto again;
1620         spin_unlock(&session->s_cap_lock);
1621
1622         if (msg) {
1623                 // Append cap_barrier field
1624                 cap_barrier = msg->front.iov_base + msg->front.iov_len;
1625                 *cap_barrier = barrier;
1626                 msg->front.iov_len += sizeof(*cap_barrier);
1627
1628                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1629                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1630                 ceph_con_send(&session->s_con, msg);
1631         }
1632         return;
1633 out_err:
1634         pr_err("send_cap_releases mds%d, failed to allocate message\n",
1635                 session->s_mds);
1636         spin_lock(&session->s_cap_lock);
1637         list_splice(&tmp_list, &session->s_cap_releases);
1638         session->s_num_cap_releases += num_cap_releases;
1639         spin_unlock(&session->s_cap_lock);
1640 }
1641
1642 /*
1643  * requests
1644  */
1645
1646 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1647                                     struct inode *dir)
1648 {
1649         struct ceph_inode_info *ci = ceph_inode(dir);
1650         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1651         struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1652         size_t size = sizeof(struct ceph_mds_reply_dir_entry);
1653         int order, num_entries;
1654
1655         spin_lock(&ci->i_ceph_lock);
1656         num_entries = ci->i_files + ci->i_subdirs;
1657         spin_unlock(&ci->i_ceph_lock);
1658         num_entries = max(num_entries, 1);
1659         num_entries = min(num_entries, opt->max_readdir);
1660
1661         order = get_order(size * num_entries);
1662         while (order >= 0) {
1663                 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
1664                                                              __GFP_NOWARN,
1665                                                              order);
1666                 if (rinfo->dir_entries)
1667                         break;
1668                 order--;
1669         }
1670         if (!rinfo->dir_entries)
1671                 return -ENOMEM;
1672
1673         num_entries = (PAGE_SIZE << order) / size;
1674         num_entries = min(num_entries, opt->max_readdir);
1675
1676         rinfo->dir_buf_size = PAGE_SIZE << order;
1677         req->r_num_caps = num_entries + 1;
1678         req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1679         req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1680         return 0;
1681 }
1682
1683 /*
1684  * Create an mds request.
1685  */
1686 struct ceph_mds_request *
1687 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1688 {
1689         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1690
1691         if (!req)
1692                 return ERR_PTR(-ENOMEM);
1693
1694         mutex_init(&req->r_fill_mutex);
1695         req->r_mdsc = mdsc;
1696         req->r_started = jiffies;
1697         req->r_resend_mds = -1;
1698         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1699         INIT_LIST_HEAD(&req->r_unsafe_target_item);
1700         req->r_fmode = -1;
1701         kref_init(&req->r_kref);
1702         RB_CLEAR_NODE(&req->r_node);
1703         INIT_LIST_HEAD(&req->r_wait);
1704         init_completion(&req->r_completion);
1705         init_completion(&req->r_safe_completion);
1706         INIT_LIST_HEAD(&req->r_unsafe_item);
1707
1708         req->r_stamp = timespec_trunc(current_kernel_time(), mdsc->fsc->sb->s_time_gran);
1709
1710         req->r_op = op;
1711         req->r_direct_mode = mode;
1712         return req;
1713 }
1714
1715 /*
1716  * return oldest (lowest) request, tid in request tree, 0 if none.
1717  *
1718  * called under mdsc->mutex.
1719  */
1720 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1721 {
1722         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1723                 return NULL;
1724         return rb_entry(rb_first(&mdsc->request_tree),
1725                         struct ceph_mds_request, r_node);
1726 }
1727
1728 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1729 {
1730         return mdsc->oldest_tid;
1731 }
1732
1733 /*
1734  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1735  * on build_path_from_dentry in fs/cifs/dir.c.
1736  *
1737  * If @stop_on_nosnap, generate path relative to the first non-snapped
1738  * inode.
1739  *
1740  * Encode hidden .snap dirs as a double /, i.e.
1741  *   foo/.snap/bar -> foo//bar
1742  */
1743 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1744                            int stop_on_nosnap)
1745 {
1746         struct dentry *temp;
1747         char *path;
1748         int len, pos;
1749         unsigned seq;
1750
1751         if (dentry == NULL)
1752                 return ERR_PTR(-EINVAL);
1753
1754 retry:
1755         len = 0;
1756         seq = read_seqbegin(&rename_lock);
1757         rcu_read_lock();
1758         for (temp = dentry; !IS_ROOT(temp);) {
1759                 struct inode *inode = d_inode(temp);
1760                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1761                         len++;  /* slash only */
1762                 else if (stop_on_nosnap && inode &&
1763                          ceph_snap(inode) == CEPH_NOSNAP)
1764                         break;
1765                 else
1766                         len += 1 + temp->d_name.len;
1767                 temp = temp->d_parent;
1768         }
1769         rcu_read_unlock();
1770         if (len)
1771                 len--;  /* no leading '/' */
1772
1773         path = kmalloc(len+1, GFP_NOFS);
1774         if (path == NULL)
1775                 return ERR_PTR(-ENOMEM);
1776         pos = len;
1777         path[pos] = 0;  /* trailing null */
1778         rcu_read_lock();
1779         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1780                 struct inode *inode;
1781
1782                 spin_lock(&temp->d_lock);
1783                 inode = d_inode(temp);
1784                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1785                         dout("build_path path+%d: %p SNAPDIR\n",
1786                              pos, temp);
1787                 } else if (stop_on_nosnap && inode &&
1788                            ceph_snap(inode) == CEPH_NOSNAP) {
1789                         spin_unlock(&temp->d_lock);
1790                         break;
1791                 } else {
1792                         pos -= temp->d_name.len;
1793                         if (pos < 0) {
1794                                 spin_unlock(&temp->d_lock);
1795                                 break;
1796                         }
1797                         strncpy(path + pos, temp->d_name.name,
1798                                 temp->d_name.len);
1799                 }
1800                 spin_unlock(&temp->d_lock);
1801                 if (pos)
1802                         path[--pos] = '/';
1803                 temp = temp->d_parent;
1804         }
1805         rcu_read_unlock();
1806         if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1807                 pr_err("build_path did not end path lookup where "
1808                        "expected, namelen is %d, pos is %d\n", len, pos);
1809                 /* presumably this is only possible if racing with a
1810                    rename of one of the parent directories (we can not
1811                    lock the dentries above us to prevent this, but
1812                    retrying should be harmless) */
1813                 kfree(path);
1814                 goto retry;
1815         }
1816
1817         *base = ceph_ino(d_inode(temp));
1818         *plen = len;
1819         dout("build_path on %p %d built %llx '%.*s'\n",
1820              dentry, d_count(dentry), *base, len, path);
1821         return path;
1822 }
1823
1824 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
1825                              const char **ppath, int *ppathlen, u64 *pino,
1826                              int *pfreepath)
1827 {
1828         char *path;
1829
1830         rcu_read_lock();
1831         if (!dir)
1832                 dir = d_inode_rcu(dentry->d_parent);
1833         if (dir && ceph_snap(dir) == CEPH_NOSNAP) {
1834                 *pino = ceph_ino(dir);
1835                 rcu_read_unlock();
1836                 *ppath = dentry->d_name.name;
1837                 *ppathlen = dentry->d_name.len;
1838                 return 0;
1839         }
1840         rcu_read_unlock();
1841         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1842         if (IS_ERR(path))
1843                 return PTR_ERR(path);
1844         *ppath = path;
1845         *pfreepath = 1;
1846         return 0;
1847 }
1848
1849 static int build_inode_path(struct inode *inode,
1850                             const char **ppath, int *ppathlen, u64 *pino,
1851                             int *pfreepath)
1852 {
1853         struct dentry *dentry;
1854         char *path;
1855
1856         if (ceph_snap(inode) == CEPH_NOSNAP) {
1857                 *pino = ceph_ino(inode);
1858                 *ppathlen = 0;
1859                 return 0;
1860         }
1861         dentry = d_find_alias(inode);
1862         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1863         dput(dentry);
1864         if (IS_ERR(path))
1865                 return PTR_ERR(path);
1866         *ppath = path;
1867         *pfreepath = 1;
1868         return 0;
1869 }
1870
1871 /*
1872  * request arguments may be specified via an inode *, a dentry *, or
1873  * an explicit ino+path.
1874  */
1875 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1876                                   struct inode *rdiri, const char *rpath,
1877                                   u64 rino, const char **ppath, int *pathlen,
1878                                   u64 *ino, int *freepath)
1879 {
1880         int r = 0;
1881
1882         if (rinode) {
1883                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1884                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1885                      ceph_snap(rinode));
1886         } else if (rdentry) {
1887                 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
1888                                         freepath);
1889                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1890                      *ppath);
1891         } else if (rpath || rino) {
1892                 *ino = rino;
1893                 *ppath = rpath;
1894                 *pathlen = rpath ? strlen(rpath) : 0;
1895                 dout(" path %.*s\n", *pathlen, rpath);
1896         }
1897
1898         return r;
1899 }
1900
1901 /*
1902  * called under mdsc->mutex
1903  */
1904 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1905                                                struct ceph_mds_request *req,
1906                                                int mds, bool drop_cap_releases)
1907 {
1908         struct ceph_msg *msg;
1909         struct ceph_mds_request_head *head;
1910         const char *path1 = NULL;
1911         const char *path2 = NULL;
1912         u64 ino1 = 0, ino2 = 0;
1913         int pathlen1 = 0, pathlen2 = 0;
1914         int freepath1 = 0, freepath2 = 0;
1915         int len;
1916         u16 releases;
1917         void *p, *end;
1918         int ret;
1919
1920         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1921                               req->r_parent, req->r_path1, req->r_ino1.ino,
1922                               &path1, &pathlen1, &ino1, &freepath1);
1923         if (ret < 0) {
1924                 msg = ERR_PTR(ret);
1925                 goto out;
1926         }
1927
1928         ret = set_request_path_attr(NULL, req->r_old_dentry,
1929                               req->r_old_dentry_dir,
1930                               req->r_path2, req->r_ino2.ino,
1931                               &path2, &pathlen2, &ino2, &freepath2);
1932         if (ret < 0) {
1933                 msg = ERR_PTR(ret);
1934                 goto out_free1;
1935         }
1936
1937         len = sizeof(*head) +
1938                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
1939                 sizeof(struct ceph_timespec);
1940
1941         /* calculate (max) length for cap releases */
1942         len += sizeof(struct ceph_mds_request_release) *
1943                 (!!req->r_inode_drop + !!req->r_dentry_drop +
1944                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1945         if (req->r_dentry_drop)
1946                 len += req->r_dentry->d_name.len;
1947         if (req->r_old_dentry_drop)
1948                 len += req->r_old_dentry->d_name.len;
1949
1950         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1951         if (!msg) {
1952                 msg = ERR_PTR(-ENOMEM);
1953                 goto out_free2;
1954         }
1955
1956         msg->hdr.version = cpu_to_le16(2);
1957         msg->hdr.tid = cpu_to_le64(req->r_tid);
1958
1959         head = msg->front.iov_base;
1960         p = msg->front.iov_base + sizeof(*head);
1961         end = msg->front.iov_base + msg->front.iov_len;
1962
1963         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1964         head->op = cpu_to_le32(req->r_op);
1965         head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1966         head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1967         head->args = req->r_args;
1968
1969         ceph_encode_filepath(&p, end, ino1, path1);
1970         ceph_encode_filepath(&p, end, ino2, path2);
1971
1972         /* make note of release offset, in case we need to replay */
1973         req->r_request_release_offset = p - msg->front.iov_base;
1974
1975         /* cap releases */
1976         releases = 0;
1977         if (req->r_inode_drop)
1978                 releases += ceph_encode_inode_release(&p,
1979                       req->r_inode ? req->r_inode : d_inode(req->r_dentry),
1980                       mds, req->r_inode_drop, req->r_inode_unless, 0);
1981         if (req->r_dentry_drop)
1982                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1983                                 req->r_parent, mds, req->r_dentry_drop,
1984                                 req->r_dentry_unless);
1985         if (req->r_old_dentry_drop)
1986                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1987                                 req->r_old_dentry_dir, mds,
1988                                 req->r_old_dentry_drop,
1989                                 req->r_old_dentry_unless);
1990         if (req->r_old_inode_drop)
1991                 releases += ceph_encode_inode_release(&p,
1992                       d_inode(req->r_old_dentry),
1993                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1994
1995         if (drop_cap_releases) {
1996                 releases = 0;
1997                 p = msg->front.iov_base + req->r_request_release_offset;
1998         }
1999
2000         head->num_releases = cpu_to_le16(releases);
2001
2002         /* time stamp */
2003         {
2004                 struct ceph_timespec ts;
2005                 ceph_encode_timespec(&ts, &req->r_stamp);
2006                 ceph_encode_copy(&p, &ts, sizeof(ts));
2007         }
2008
2009         BUG_ON(p > end);
2010         msg->front.iov_len = p - msg->front.iov_base;
2011         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2012
2013         if (req->r_pagelist) {
2014                 struct ceph_pagelist *pagelist = req->r_pagelist;
2015                 refcount_inc(&pagelist->refcnt);
2016                 ceph_msg_data_add_pagelist(msg, pagelist);
2017                 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2018         } else {
2019                 msg->hdr.data_len = 0;
2020         }
2021
2022         msg->hdr.data_off = cpu_to_le16(0);
2023
2024 out_free2:
2025         if (freepath2)
2026                 kfree((char *)path2);
2027 out_free1:
2028         if (freepath1)
2029                 kfree((char *)path1);
2030 out:
2031         return msg;
2032 }
2033
2034 /*
2035  * called under mdsc->mutex if error, under no mutex if
2036  * success.
2037  */
2038 static void complete_request(struct ceph_mds_client *mdsc,
2039                              struct ceph_mds_request *req)
2040 {
2041         if (req->r_callback)
2042                 req->r_callback(mdsc, req);
2043         else
2044                 complete_all(&req->r_completion);
2045 }
2046
2047 /*
2048  * called under mdsc->mutex
2049  */
2050 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2051                                   struct ceph_mds_request *req,
2052                                   int mds, bool drop_cap_releases)
2053 {
2054         struct ceph_mds_request_head *rhead;
2055         struct ceph_msg *msg;
2056         int flags = 0;
2057
2058         req->r_attempts++;
2059         if (req->r_inode) {
2060                 struct ceph_cap *cap =
2061                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2062
2063                 if (cap)
2064                         req->r_sent_on_mseq = cap->mseq;
2065                 else
2066                         req->r_sent_on_mseq = -1;
2067         }
2068         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2069              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2070
2071         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2072                 void *p;
2073                 /*
2074                  * Replay.  Do not regenerate message (and rebuild
2075                  * paths, etc.); just use the original message.
2076                  * Rebuilding paths will break for renames because
2077                  * d_move mangles the src name.
2078                  */
2079                 msg = req->r_request;
2080                 rhead = msg->front.iov_base;
2081
2082                 flags = le32_to_cpu(rhead->flags);
2083                 flags |= CEPH_MDS_FLAG_REPLAY;
2084                 rhead->flags = cpu_to_le32(flags);
2085
2086                 if (req->r_target_inode)
2087                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2088
2089                 rhead->num_retry = req->r_attempts - 1;
2090
2091                 /* remove cap/dentry releases from message */
2092                 rhead->num_releases = 0;
2093
2094                 /* time stamp */
2095                 p = msg->front.iov_base + req->r_request_release_offset;
2096                 {
2097                         struct ceph_timespec ts;
2098                         ceph_encode_timespec(&ts, &req->r_stamp);
2099                         ceph_encode_copy(&p, &ts, sizeof(ts));
2100                 }
2101
2102                 msg->front.iov_len = p - msg->front.iov_base;
2103                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2104                 return 0;
2105         }
2106
2107         if (req->r_request) {
2108                 ceph_msg_put(req->r_request);
2109                 req->r_request = NULL;
2110         }
2111         msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2112         if (IS_ERR(msg)) {
2113                 req->r_err = PTR_ERR(msg);
2114                 return PTR_ERR(msg);
2115         }
2116         req->r_request = msg;
2117
2118         rhead = msg->front.iov_base;
2119         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2120         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2121                 flags |= CEPH_MDS_FLAG_REPLAY;
2122         if (req->r_parent)
2123                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2124         rhead->flags = cpu_to_le32(flags);
2125         rhead->num_fwd = req->r_num_fwd;
2126         rhead->num_retry = req->r_attempts - 1;
2127         rhead->ino = 0;
2128
2129         dout(" r_parent = %p\n", req->r_parent);
2130         return 0;
2131 }
2132
2133 /*
2134  * send request, or put it on the appropriate wait list.
2135  */
2136 static int __do_request(struct ceph_mds_client *mdsc,
2137                         struct ceph_mds_request *req)
2138 {
2139         struct ceph_mds_session *session = NULL;
2140         int mds = -1;
2141         int err = 0;
2142
2143         if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2144                 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2145                         __unregister_request(mdsc, req);
2146                 goto out;
2147         }
2148
2149         if (req->r_timeout &&
2150             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2151                 dout("do_request timed out\n");
2152                 err = -EIO;
2153                 goto finish;
2154         }
2155         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2156                 dout("do_request forced umount\n");
2157                 err = -EIO;
2158                 goto finish;
2159         }
2160         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2161                 if (mdsc->mdsmap_err) {
2162                         err = mdsc->mdsmap_err;
2163                         dout("do_request mdsmap err %d\n", err);
2164                         goto finish;
2165                 }
2166                 if (mdsc->mdsmap->m_epoch == 0) {
2167                         dout("do_request no mdsmap, waiting for map\n");
2168                         list_add(&req->r_wait, &mdsc->waiting_for_map);
2169                         goto finish;
2170                 }
2171                 if (!(mdsc->fsc->mount_options->flags &
2172                       CEPH_MOUNT_OPT_MOUNTWAIT) &&
2173                     !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2174                         err = -ENOENT;
2175                         pr_info("probably no mds server is up\n");
2176                         goto finish;
2177                 }
2178         }
2179
2180         put_request_session(req);
2181
2182         mds = __choose_mds(mdsc, req);
2183         if (mds < 0 ||
2184             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2185                 dout("do_request no mds or not active, waiting for map\n");
2186                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2187                 goto out;
2188         }
2189
2190         /* get, open session */
2191         session = __ceph_lookup_mds_session(mdsc, mds);
2192         if (!session) {
2193                 session = register_session(mdsc, mds);
2194                 if (IS_ERR(session)) {
2195                         err = PTR_ERR(session);
2196                         goto finish;
2197                 }
2198         }
2199         req->r_session = get_session(session);
2200
2201         dout("do_request mds%d session %p state %s\n", mds, session,
2202              ceph_session_state_name(session->s_state));
2203         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2204             session->s_state != CEPH_MDS_SESSION_HUNG) {
2205                 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2206                         err = -EACCES;
2207                         goto out_session;
2208                 }
2209                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2210                     session->s_state == CEPH_MDS_SESSION_CLOSING)
2211                         __open_session(mdsc, session);
2212                 list_add(&req->r_wait, &session->s_waiting);
2213                 goto out_session;
2214         }
2215
2216         /* send request */
2217         req->r_resend_mds = -1;   /* forget any previous mds hint */
2218
2219         if (req->r_request_started == 0)   /* note request start time */
2220                 req->r_request_started = jiffies;
2221
2222         err = __prepare_send_request(mdsc, req, mds, false);
2223         if (!err) {
2224                 ceph_msg_get(req->r_request);
2225                 ceph_con_send(&session->s_con, req->r_request);
2226         }
2227
2228 out_session:
2229         ceph_put_mds_session(session);
2230 finish:
2231         if (err) {
2232                 dout("__do_request early error %d\n", err);
2233                 req->r_err = err;
2234                 complete_request(mdsc, req);
2235                 __unregister_request(mdsc, req);
2236         }
2237 out:
2238         return err;
2239 }
2240
2241 /*
2242  * called under mdsc->mutex
2243  */
2244 static void __wake_requests(struct ceph_mds_client *mdsc,
2245                             struct list_head *head)
2246 {
2247         struct ceph_mds_request *req;
2248         LIST_HEAD(tmp_list);
2249
2250         list_splice_init(head, &tmp_list);
2251
2252         while (!list_empty(&tmp_list)) {
2253                 req = list_entry(tmp_list.next,
2254                                  struct ceph_mds_request, r_wait);
2255                 list_del_init(&req->r_wait);
2256                 dout(" wake request %p tid %llu\n", req, req->r_tid);
2257                 __do_request(mdsc, req);
2258         }
2259 }
2260
2261 /*
2262  * Wake up threads with requests pending for @mds, so that they can
2263  * resubmit their requests to a possibly different mds.
2264  */
2265 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2266 {
2267         struct ceph_mds_request *req;
2268         struct rb_node *p = rb_first(&mdsc->request_tree);
2269
2270         dout("kick_requests mds%d\n", mds);
2271         while (p) {
2272                 req = rb_entry(p, struct ceph_mds_request, r_node);
2273                 p = rb_next(p);
2274                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2275                         continue;
2276                 if (req->r_attempts > 0)
2277                         continue; /* only new requests */
2278                 if (req->r_session &&
2279                     req->r_session->s_mds == mds) {
2280                         dout(" kicking tid %llu\n", req->r_tid);
2281                         list_del_init(&req->r_wait);
2282                         __do_request(mdsc, req);
2283                 }
2284         }
2285 }
2286
2287 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2288                               struct ceph_mds_request *req)
2289 {
2290         dout("submit_request on %p\n", req);
2291         mutex_lock(&mdsc->mutex);
2292         __register_request(mdsc, req, NULL);
2293         __do_request(mdsc, req);
2294         mutex_unlock(&mdsc->mutex);
2295 }
2296
2297 /*
2298  * Synchrously perform an mds request.  Take care of all of the
2299  * session setup, forwarding, retry details.
2300  */
2301 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2302                          struct inode *dir,
2303                          struct ceph_mds_request *req)
2304 {
2305         int err;
2306
2307         dout("do_request on %p\n", req);
2308
2309         /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2310         if (req->r_inode)
2311                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2312         if (req->r_parent)
2313                 ceph_get_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
2314         if (req->r_old_dentry_dir)
2315                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2316                                   CEPH_CAP_PIN);
2317
2318         /* issue */
2319         mutex_lock(&mdsc->mutex);
2320         __register_request(mdsc, req, dir);
2321         __do_request(mdsc, req);
2322
2323         if (req->r_err) {
2324                 err = req->r_err;
2325                 goto out;
2326         }
2327
2328         /* wait */
2329         mutex_unlock(&mdsc->mutex);
2330         dout("do_request waiting\n");
2331         if (!req->r_timeout && req->r_wait_for_completion) {
2332                 err = req->r_wait_for_completion(mdsc, req);
2333         } else {
2334                 long timeleft = wait_for_completion_killable_timeout(
2335                                         &req->r_completion,
2336                                         ceph_timeout_jiffies(req->r_timeout));
2337                 if (timeleft > 0)
2338                         err = 0;
2339                 else if (!timeleft)
2340                         err = -EIO;  /* timed out */
2341                 else
2342                         err = timeleft;  /* killed */
2343         }
2344         dout("do_request waited, got %d\n", err);
2345         mutex_lock(&mdsc->mutex);
2346
2347         /* only abort if we didn't race with a real reply */
2348         if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2349                 err = le32_to_cpu(req->r_reply_info.head->result);
2350         } else if (err < 0) {
2351                 dout("aborted request %lld with %d\n", req->r_tid, err);
2352
2353                 /*
2354                  * ensure we aren't running concurrently with
2355                  * ceph_fill_trace or ceph_readdir_prepopulate, which
2356                  * rely on locks (dir mutex) held by our caller.
2357                  */
2358                 mutex_lock(&req->r_fill_mutex);
2359                 req->r_err = err;
2360                 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2361                 mutex_unlock(&req->r_fill_mutex);
2362
2363                 if (req->r_parent &&
2364                     (req->r_op & CEPH_MDS_OP_WRITE))
2365                         ceph_invalidate_dir_request(req);
2366         } else {
2367                 err = req->r_err;
2368         }
2369
2370 out:
2371         mutex_unlock(&mdsc->mutex);
2372         dout("do_request %p done, result %d\n", req, err);
2373         return err;
2374 }
2375
2376 /*
2377  * Invalidate dir's completeness, dentry lease state on an aborted MDS
2378  * namespace request.
2379  */
2380 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2381 {
2382         struct inode *inode = req->r_parent;
2383
2384         dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2385
2386         ceph_dir_clear_complete(inode);
2387         if (req->r_dentry)
2388                 ceph_invalidate_dentry_lease(req->r_dentry);
2389         if (req->r_old_dentry)
2390                 ceph_invalidate_dentry_lease(req->r_old_dentry);
2391 }
2392
2393 /*
2394  * Handle mds reply.
2395  *
2396  * We take the session mutex and parse and process the reply immediately.
2397  * This preserves the logical ordering of replies, capabilities, etc., sent
2398  * by the MDS as they are applied to our local cache.
2399  */
2400 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2401 {
2402         struct ceph_mds_client *mdsc = session->s_mdsc;
2403         struct ceph_mds_request *req;
2404         struct ceph_mds_reply_head *head = msg->front.iov_base;
2405         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2406         struct ceph_snap_realm *realm;
2407         u64 tid;
2408         int err, result;
2409         int mds = session->s_mds;
2410
2411         if (msg->front.iov_len < sizeof(*head)) {
2412                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2413                 ceph_msg_dump(msg);
2414                 return;
2415         }
2416
2417         /* get request, session */
2418         tid = le64_to_cpu(msg->hdr.tid);
2419         mutex_lock(&mdsc->mutex);
2420         req = lookup_get_request(mdsc, tid);
2421         if (!req) {
2422                 dout("handle_reply on unknown tid %llu\n", tid);
2423                 mutex_unlock(&mdsc->mutex);
2424                 return;
2425         }
2426         dout("handle_reply %p\n", req);
2427
2428         /* correct session? */
2429         if (req->r_session != session) {
2430                 pr_err("mdsc_handle_reply got %llu on session mds%d"
2431                        " not mds%d\n", tid, session->s_mds,
2432                        req->r_session ? req->r_session->s_mds : -1);
2433                 mutex_unlock(&mdsc->mutex);
2434                 goto out;
2435         }
2436
2437         /* dup? */
2438         if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
2439             (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
2440                 pr_warn("got a dup %s reply on %llu from mds%d\n",
2441                            head->safe ? "safe" : "unsafe", tid, mds);
2442                 mutex_unlock(&mdsc->mutex);
2443                 goto out;
2444         }
2445         if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
2446                 pr_warn("got unsafe after safe on %llu from mds%d\n",
2447                            tid, mds);
2448                 mutex_unlock(&mdsc->mutex);
2449                 goto out;
2450         }
2451
2452         result = le32_to_cpu(head->result);
2453
2454         /*
2455          * Handle an ESTALE
2456          * if we're not talking to the authority, send to them
2457          * if the authority has changed while we weren't looking,
2458          * send to new authority
2459          * Otherwise we just have to return an ESTALE
2460          */
2461         if (result == -ESTALE) {
2462                 dout("got ESTALE on request %llu", req->r_tid);
2463                 req->r_resend_mds = -1;
2464                 if (req->r_direct_mode != USE_AUTH_MDS) {
2465                         dout("not using auth, setting for that now");
2466                         req->r_direct_mode = USE_AUTH_MDS;
2467                         __do_request(mdsc, req);
2468                         mutex_unlock(&mdsc->mutex);
2469                         goto out;
2470                 } else  {
2471                         int mds = __choose_mds(mdsc, req);
2472                         if (mds >= 0 && mds != req->r_session->s_mds) {
2473                                 dout("but auth changed, so resending");
2474                                 __do_request(mdsc, req);
2475                                 mutex_unlock(&mdsc->mutex);
2476                                 goto out;
2477                         }
2478                 }
2479                 dout("have to return ESTALE on request %llu", req->r_tid);
2480         }
2481
2482
2483         if (head->safe) {
2484                 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
2485                 __unregister_request(mdsc, req);
2486
2487                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2488                         /*
2489                          * We already handled the unsafe response, now do the
2490                          * cleanup.  No need to examine the response; the MDS
2491                          * doesn't include any result info in the safe
2492                          * response.  And even if it did, there is nothing
2493                          * useful we could do with a revised return value.
2494                          */
2495                         dout("got safe reply %llu, mds%d\n", tid, mds);
2496
2497                         /* last unsafe request during umount? */
2498                         if (mdsc->stopping && !__get_oldest_req(mdsc))
2499                                 complete_all(&mdsc->safe_umount_waiters);
2500                         mutex_unlock(&mdsc->mutex);
2501                         goto out;
2502                 }
2503         } else {
2504                 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
2505                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2506                 if (req->r_unsafe_dir) {
2507                         struct ceph_inode_info *ci =
2508                                         ceph_inode(req->r_unsafe_dir);
2509                         spin_lock(&ci->i_unsafe_lock);
2510                         list_add_tail(&req->r_unsafe_dir_item,
2511                                       &ci->i_unsafe_dirops);
2512                         spin_unlock(&ci->i_unsafe_lock);
2513                 }
2514         }
2515
2516         dout("handle_reply tid %lld result %d\n", tid, result);
2517         rinfo = &req->r_reply_info;
2518         err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2519         mutex_unlock(&mdsc->mutex);
2520
2521         mutex_lock(&session->s_mutex);
2522         if (err < 0) {
2523                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2524                 ceph_msg_dump(msg);
2525                 goto out_err;
2526         }
2527
2528         /* snap trace */
2529         realm = NULL;
2530         if (rinfo->snapblob_len) {
2531                 down_write(&mdsc->snap_rwsem);
2532                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2533                                 rinfo->snapblob + rinfo->snapblob_len,
2534                                 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2535                                 &realm);
2536                 downgrade_write(&mdsc->snap_rwsem);
2537         } else {
2538                 down_read(&mdsc->snap_rwsem);
2539         }
2540
2541         /* insert trace into our cache */
2542         mutex_lock(&req->r_fill_mutex);
2543         current->journal_info = req;
2544         err = ceph_fill_trace(mdsc->fsc->sb, req);
2545         if (err == 0) {
2546                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2547                                     req->r_op == CEPH_MDS_OP_LSSNAP))
2548                         ceph_readdir_prepopulate(req, req->r_session);
2549                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2550         }
2551         current->journal_info = NULL;
2552         mutex_unlock(&req->r_fill_mutex);
2553
2554         up_read(&mdsc->snap_rwsem);
2555         if (realm)
2556                 ceph_put_snap_realm(mdsc, realm);
2557
2558         if (err == 0 && req->r_target_inode &&
2559             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2560                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
2561                 spin_lock(&ci->i_unsafe_lock);
2562                 list_add_tail(&req->r_unsafe_target_item, &ci->i_unsafe_iops);
2563                 spin_unlock(&ci->i_unsafe_lock);
2564         }
2565 out_err:
2566         mutex_lock(&mdsc->mutex);
2567         if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2568                 if (err) {
2569                         req->r_err = err;
2570                 } else {
2571                         req->r_reply =  ceph_msg_get(msg);
2572                         set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
2573                 }
2574         } else {
2575                 dout("reply arrived after request %lld was aborted\n", tid);
2576         }
2577         mutex_unlock(&mdsc->mutex);
2578
2579         mutex_unlock(&session->s_mutex);
2580
2581         /* kick calling process */
2582         complete_request(mdsc, req);
2583 out:
2584         ceph_mdsc_put_request(req);
2585         return;
2586 }
2587
2588
2589
2590 /*
2591  * handle mds notification that our request has been forwarded.
2592  */
2593 static void handle_forward(struct ceph_mds_client *mdsc,
2594                            struct ceph_mds_session *session,
2595                            struct ceph_msg *msg)
2596 {
2597         struct ceph_mds_request *req;
2598         u64 tid = le64_to_cpu(msg->hdr.tid);
2599         u32 next_mds;
2600         u32 fwd_seq;
2601         int err = -EINVAL;
2602         void *p = msg->front.iov_base;
2603         void *end = p + msg->front.iov_len;
2604
2605         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2606         next_mds = ceph_decode_32(&p);
2607         fwd_seq = ceph_decode_32(&p);
2608
2609         mutex_lock(&mdsc->mutex);
2610         req = lookup_get_request(mdsc, tid);
2611         if (!req) {
2612                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2613                 goto out;  /* dup reply? */
2614         }
2615
2616         if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2617                 dout("forward tid %llu aborted, unregistering\n", tid);
2618                 __unregister_request(mdsc, req);
2619         } else if (fwd_seq <= req->r_num_fwd) {
2620                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2621                      tid, next_mds, req->r_num_fwd, fwd_seq);
2622         } else {
2623                 /* resend. forward race not possible; mds would drop */
2624                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2625                 BUG_ON(req->r_err);
2626                 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
2627                 req->r_attempts = 0;
2628                 req->r_num_fwd = fwd_seq;
2629                 req->r_resend_mds = next_mds;
2630                 put_request_session(req);
2631                 __do_request(mdsc, req);
2632         }
2633         ceph_mdsc_put_request(req);
2634 out:
2635         mutex_unlock(&mdsc->mutex);
2636         return;
2637
2638 bad:
2639         pr_err("mdsc_handle_forward decode error err=%d\n", err);
2640 }
2641
2642 /*
2643  * handle a mds session control message
2644  */
2645 static void handle_session(struct ceph_mds_session *session,
2646                            struct ceph_msg *msg)
2647 {
2648         struct ceph_mds_client *mdsc = session->s_mdsc;
2649         u32 op;
2650         u64 seq;
2651         int mds = session->s_mds;
2652         struct ceph_mds_session_head *h = msg->front.iov_base;
2653         int wake = 0;
2654
2655         /* decode */
2656         if (msg->front.iov_len != sizeof(*h))
2657                 goto bad;
2658         op = le32_to_cpu(h->op);
2659         seq = le64_to_cpu(h->seq);
2660
2661         mutex_lock(&mdsc->mutex);
2662         if (op == CEPH_SESSION_CLOSE) {
2663                 get_session(session);
2664                 __unregister_session(mdsc, session);
2665         }
2666         /* FIXME: this ttl calculation is generous */
2667         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2668         mutex_unlock(&mdsc->mutex);
2669
2670         mutex_lock(&session->s_mutex);
2671
2672         dout("handle_session mds%d %s %p state %s seq %llu\n",
2673              mds, ceph_session_op_name(op), session,
2674              ceph_session_state_name(session->s_state), seq);
2675
2676         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2677                 session->s_state = CEPH_MDS_SESSION_OPEN;
2678                 pr_info("mds%d came back\n", session->s_mds);
2679         }
2680
2681         switch (op) {
2682         case CEPH_SESSION_OPEN:
2683                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2684                         pr_info("mds%d reconnect success\n", session->s_mds);
2685                 session->s_state = CEPH_MDS_SESSION_OPEN;
2686                 renewed_caps(mdsc, session, 0);
2687                 wake = 1;
2688                 if (mdsc->stopping)
2689                         __close_session(mdsc, session);
2690                 break;
2691
2692         case CEPH_SESSION_RENEWCAPS:
2693                 if (session->s_renew_seq == seq)
2694                         renewed_caps(mdsc, session, 1);
2695                 break;
2696
2697         case CEPH_SESSION_CLOSE:
2698                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2699                         pr_info("mds%d reconnect denied\n", session->s_mds);
2700                 cleanup_session_requests(mdsc, session);
2701                 remove_session_caps(session);
2702                 wake = 2; /* for good measure */
2703                 wake_up_all(&mdsc->session_close_wq);
2704                 break;
2705
2706         case CEPH_SESSION_STALE:
2707                 pr_info("mds%d caps went stale, renewing\n",
2708                         session->s_mds);
2709                 spin_lock(&session->s_gen_ttl_lock);
2710                 session->s_cap_gen++;
2711                 session->s_cap_ttl = jiffies - 1;
2712                 spin_unlock(&session->s_gen_ttl_lock);
2713                 send_renew_caps(mdsc, session);
2714                 break;
2715
2716         case CEPH_SESSION_RECALL_STATE:
2717                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2718                 break;
2719
2720         case CEPH_SESSION_FLUSHMSG:
2721                 send_flushmsg_ack(mdsc, session, seq);
2722                 break;
2723
2724         case CEPH_SESSION_FORCE_RO:
2725                 dout("force_session_readonly %p\n", session);
2726                 spin_lock(&session->s_cap_lock);
2727                 session->s_readonly = true;
2728                 spin_unlock(&session->s_cap_lock);
2729                 wake_up_session_caps(session, 0);
2730                 break;
2731
2732         case CEPH_SESSION_REJECT:
2733                 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
2734                 pr_info("mds%d rejected session\n", session->s_mds);
2735                 session->s_state = CEPH_MDS_SESSION_REJECTED;
2736                 cleanup_session_requests(mdsc, session);
2737                 remove_session_caps(session);
2738                 wake = 2; /* for good measure */
2739                 break;
2740
2741         default:
2742                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2743                 WARN_ON(1);
2744         }
2745
2746         mutex_unlock(&session->s_mutex);
2747         if (wake) {
2748                 mutex_lock(&mdsc->mutex);
2749                 __wake_requests(mdsc, &session->s_waiting);
2750                 if (wake == 2)
2751                         kick_requests(mdsc, mds);
2752                 mutex_unlock(&mdsc->mutex);
2753         }
2754         if (op == CEPH_SESSION_CLOSE)
2755                 ceph_put_mds_session(session);
2756         return;
2757
2758 bad:
2759         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2760                (int)msg->front.iov_len);
2761         ceph_msg_dump(msg);
2762         return;
2763 }
2764
2765
2766 /*
2767  * called under session->mutex.
2768  */
2769 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2770                                    struct ceph_mds_session *session)
2771 {
2772         struct ceph_mds_request *req, *nreq;
2773         struct rb_node *p;
2774         int err;
2775
2776         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2777
2778         mutex_lock(&mdsc->mutex);
2779         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2780                 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2781                 if (!err) {
2782                         ceph_msg_get(req->r_request);
2783                         ceph_con_send(&session->s_con, req->r_request);
2784                 }
2785         }
2786
2787         /*
2788          * also re-send old requests when MDS enters reconnect stage. So that MDS
2789          * can process completed request in clientreplay stage.
2790          */
2791         p = rb_first(&mdsc->request_tree);
2792         while (p) {
2793                 req = rb_entry(p, struct ceph_mds_request, r_node);
2794                 p = rb_next(p);
2795                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2796                         continue;
2797                 if (req->r_attempts == 0)
2798                         continue; /* only old requests */
2799                 if (req->r_session &&
2800                     req->r_session->s_mds == session->s_mds) {
2801                         err = __prepare_send_request(mdsc, req,
2802                                                      session->s_mds, true);
2803                         if (!err) {
2804                                 ceph_msg_get(req->r_request);
2805                                 ceph_con_send(&session->s_con, req->r_request);
2806                         }
2807                 }
2808         }
2809         mutex_unlock(&mdsc->mutex);
2810 }
2811
2812 /*
2813  * Encode information about a cap for a reconnect with the MDS.
2814  */
2815 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2816                           void *arg)
2817 {
2818         union {
2819                 struct ceph_mds_cap_reconnect v2;
2820                 struct ceph_mds_cap_reconnect_v1 v1;
2821         } rec;
2822         struct ceph_inode_info *ci;
2823         struct ceph_reconnect_state *recon_state = arg;
2824         struct ceph_pagelist *pagelist = recon_state->pagelist;
2825         char *path;
2826         int pathlen, err;
2827         u64 pathbase;
2828         u64 snap_follows;
2829         struct dentry *dentry;
2830
2831         ci = cap->ci;
2832
2833         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2834              inode, ceph_vinop(inode), cap, cap->cap_id,
2835              ceph_cap_string(cap->issued));
2836         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2837         if (err)
2838                 return err;
2839
2840         dentry = d_find_alias(inode);
2841         if (dentry) {
2842                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2843                 if (IS_ERR(path)) {
2844                         err = PTR_ERR(path);
2845                         goto out_dput;
2846                 }
2847         } else {
2848                 path = NULL;
2849                 pathlen = 0;
2850                 pathbase = 0;
2851         }
2852
2853         spin_lock(&ci->i_ceph_lock);
2854         cap->seq = 0;        /* reset cap seq */
2855         cap->issue_seq = 0;  /* and issue_seq */
2856         cap->mseq = 0;       /* and migrate_seq */
2857         cap->cap_gen = cap->session->s_cap_gen;
2858
2859         if (recon_state->msg_version >= 2) {
2860                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2861                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2862                 rec.v2.issued = cpu_to_le32(cap->issued);
2863                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2864                 rec.v2.pathbase = cpu_to_le64(pathbase);
2865                 rec.v2.flock_len = 0;
2866         } else {
2867                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2868                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2869                 rec.v1.issued = cpu_to_le32(cap->issued);
2870                 rec.v1.size = cpu_to_le64(inode->i_size);
2871                 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2872                 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2873                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2874                 rec.v1.pathbase = cpu_to_le64(pathbase);
2875         }
2876
2877         if (list_empty(&ci->i_cap_snaps)) {
2878                 snap_follows = 0;
2879         } else {
2880                 struct ceph_cap_snap *capsnap =
2881                         list_first_entry(&ci->i_cap_snaps,
2882                                          struct ceph_cap_snap, ci_item);
2883                 snap_follows = capsnap->follows;
2884         }
2885         spin_unlock(&ci->i_ceph_lock);
2886
2887         if (recon_state->msg_version >= 2) {
2888                 int num_fcntl_locks, num_flock_locks;
2889                 struct ceph_filelock *flocks;
2890                 size_t struct_len, total_len = 0;
2891                 u8 struct_v = 0;
2892
2893 encode_again:
2894                 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2895                 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2896                                  sizeof(struct ceph_filelock), GFP_NOFS);
2897                 if (!flocks) {
2898                         err = -ENOMEM;
2899                         goto out_free;
2900                 }
2901                 err = ceph_encode_locks_to_buffer(inode, flocks,
2902                                                   num_fcntl_locks,
2903                                                   num_flock_locks);
2904                 if (err) {
2905                         kfree(flocks);
2906                         if (err == -ENOSPC)
2907                                 goto encode_again;
2908                         goto out_free;
2909                 }
2910
2911                 if (recon_state->msg_version >= 3) {
2912                         /* version, compat_version and struct_len */
2913                         total_len = 2 * sizeof(u8) + sizeof(u32);
2914                         struct_v = 2;
2915                 }
2916                 /*
2917                  * number of encoded locks is stable, so copy to pagelist
2918                  */
2919                 struct_len = 2 * sizeof(u32) +
2920                             (num_fcntl_locks + num_flock_locks) *
2921                             sizeof(struct ceph_filelock);
2922                 rec.v2.flock_len = cpu_to_le32(struct_len);
2923
2924                 struct_len += sizeof(rec.v2);
2925                 struct_len += sizeof(u32) + pathlen;
2926
2927                 if (struct_v >= 2)
2928                         struct_len += sizeof(u64); /* snap_follows */
2929
2930                 total_len += struct_len;
2931                 err = ceph_pagelist_reserve(pagelist, total_len);
2932
2933                 if (!err) {
2934                         if (recon_state->msg_version >= 3) {
2935                                 ceph_pagelist_encode_8(pagelist, struct_v);
2936                                 ceph_pagelist_encode_8(pagelist, 1);
2937                                 ceph_pagelist_encode_32(pagelist, struct_len);
2938                         }
2939                         ceph_pagelist_encode_string(pagelist, path, pathlen);
2940                         ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
2941                         ceph_locks_to_pagelist(flocks, pagelist,
2942                                                num_fcntl_locks,
2943                                                num_flock_locks);
2944                         if (struct_v >= 2)
2945                                 ceph_pagelist_encode_64(pagelist, snap_follows);
2946                 }
2947                 kfree(flocks);
2948         } else {
2949                 size_t size = sizeof(u32) + pathlen + sizeof(rec.v1);
2950                 err = ceph_pagelist_reserve(pagelist, size);
2951                 if (!err) {
2952                         ceph_pagelist_encode_string(pagelist, path, pathlen);
2953                         ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
2954                 }
2955         }
2956
2957         recon_state->nr_caps++;
2958 out_free:
2959         kfree(path);
2960 out_dput:
2961         dput(dentry);
2962         return err;
2963 }
2964
2965
2966 /*
2967  * If an MDS fails and recovers, clients need to reconnect in order to
2968  * reestablish shared state.  This includes all caps issued through
2969  * this session _and_ the snap_realm hierarchy.  Because it's not
2970  * clear which snap realms the mds cares about, we send everything we
2971  * know about.. that ensures we'll then get any new info the
2972  * recovering MDS might have.
2973  *
2974  * This is a relatively heavyweight operation, but it's rare.
2975  *
2976  * called with mdsc->mutex held.
2977  */
2978 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2979                                struct ceph_mds_session *session)
2980 {
2981         struct ceph_msg *reply;
2982         struct rb_node *p;
2983         int mds = session->s_mds;
2984         int err = -ENOMEM;
2985         int s_nr_caps;
2986         struct ceph_pagelist *pagelist;
2987         struct ceph_reconnect_state recon_state;
2988
2989         pr_info("mds%d reconnect start\n", mds);
2990
2991         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2992         if (!pagelist)
2993                 goto fail_nopagelist;
2994         ceph_pagelist_init(pagelist);
2995
2996         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2997         if (!reply)
2998                 goto fail_nomsg;
2999
3000         mutex_lock(&session->s_mutex);
3001         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3002         session->s_seq = 0;
3003
3004         dout("session %p state %s\n", session,
3005              ceph_session_state_name(session->s_state));
3006
3007         spin_lock(&session->s_gen_ttl_lock);
3008         session->s_cap_gen++;
3009         spin_unlock(&session->s_gen_ttl_lock);
3010
3011         spin_lock(&session->s_cap_lock);
3012         /* don't know if session is readonly */
3013         session->s_readonly = 0;
3014         /*
3015          * notify __ceph_remove_cap() that we are composing cap reconnect.
3016          * If a cap get released before being added to the cap reconnect,
3017          * __ceph_remove_cap() should skip queuing cap release.
3018          */
3019         session->s_cap_reconnect = 1;
3020         /* drop old cap expires; we're about to reestablish that state */
3021         cleanup_cap_releases(mdsc, session);
3022
3023         /* trim unused caps to reduce MDS's cache rejoin time */
3024         if (mdsc->fsc->sb->s_root)
3025                 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3026
3027         ceph_con_close(&session->s_con);
3028         ceph_con_open(&session->s_con,
3029                       CEPH_ENTITY_TYPE_MDS, mds,
3030                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3031
3032         /* replay unsafe requests */
3033         replay_unsafe_requests(mdsc, session);
3034
3035         down_read(&mdsc->snap_rwsem);
3036
3037         /* traverse this session's caps */
3038         s_nr_caps = session->s_nr_caps;
3039         err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
3040         if (err)
3041                 goto fail;
3042
3043         recon_state.nr_caps = 0;
3044         recon_state.pagelist = pagelist;
3045         if (session->s_con.peer_features & CEPH_FEATURE_MDSENC)
3046                 recon_state.msg_version = 3;
3047         else if (session->s_con.peer_features & CEPH_FEATURE_FLOCK)
3048                 recon_state.msg_version = 2;
3049         else
3050                 recon_state.msg_version = 1;
3051         err = iterate_session_caps(session, encode_caps_cb, &recon_state);
3052         if (err < 0)
3053                 goto fail;
3054
3055         spin_lock(&session->s_cap_lock);
3056         session->s_cap_reconnect = 0;
3057         spin_unlock(&session->s_cap_lock);
3058
3059         /*
3060          * snaprealms.  we provide mds with the ino, seq (version), and
3061          * parent for all of our realms.  If the mds has any newer info,
3062          * it will tell us.
3063          */
3064         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3065                 struct ceph_snap_realm *realm =
3066                         rb_entry(p, struct ceph_snap_realm, node);
3067                 struct ceph_mds_snaprealm_reconnect sr_rec;
3068
3069                 dout(" adding snap realm %llx seq %lld parent %llx\n",
3070                      realm->ino, realm->seq, realm->parent_ino);
3071                 sr_rec.ino = cpu_to_le64(realm->ino);
3072                 sr_rec.seq = cpu_to_le64(realm->seq);
3073                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3074                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3075                 if (err)
3076                         goto fail;
3077         }
3078
3079         reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3080
3081         /* raced with cap release? */
3082         if (s_nr_caps != recon_state.nr_caps) {
3083                 struct page *page = list_first_entry(&pagelist->head,
3084                                                      struct page, lru);
3085                 __le32 *addr = kmap_atomic(page);
3086                 *addr = cpu_to_le32(recon_state.nr_caps);
3087                 kunmap_atomic(addr);
3088         }
3089
3090         reply->hdr.data_len = cpu_to_le32(pagelist->length);
3091         ceph_msg_data_add_pagelist(reply, pagelist);
3092
3093         ceph_early_kick_flushing_caps(mdsc, session);
3094
3095         ceph_con_send(&session->s_con, reply);
3096
3097         mutex_unlock(&session->s_mutex);
3098
3099         mutex_lock(&mdsc->mutex);
3100         __wake_requests(mdsc, &session->s_waiting);
3101         mutex_unlock(&mdsc->mutex);
3102
3103         up_read(&mdsc->snap_rwsem);
3104         return;
3105
3106 fail:
3107         ceph_msg_put(reply);
3108         up_read(&mdsc->snap_rwsem);
3109         mutex_unlock(&session->s_mutex);
3110 fail_nomsg:
3111         ceph_pagelist_release(pagelist);
3112 fail_nopagelist:
3113         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3114         return;
3115 }
3116
3117
3118 /*
3119  * compare old and new mdsmaps, kicking requests
3120  * and closing out old connections as necessary
3121  *
3122  * called under mdsc->mutex.
3123  */
3124 static void check_new_map(struct ceph_mds_client *mdsc,
3125                           struct ceph_mdsmap *newmap,
3126                           struct ceph_mdsmap *oldmap)
3127 {
3128         int i;
3129         int oldstate, newstate;
3130         struct ceph_mds_session *s;
3131
3132         dout("check_new_map new %u old %u\n",
3133              newmap->m_epoch, oldmap->m_epoch);
3134
3135         for (i = 0; i < oldmap->m_num_mds && i < mdsc->max_sessions; i++) {
3136                 if (mdsc->sessions[i] == NULL)
3137                         continue;
3138                 s = mdsc->sessions[i];
3139                 oldstate = ceph_mdsmap_get_state(oldmap, i);
3140                 newstate = ceph_mdsmap_get_state(newmap, i);
3141
3142                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3143                      i, ceph_mds_state_name(oldstate),
3144                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3145                      ceph_mds_state_name(newstate),
3146                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3147                      ceph_session_state_name(s->s_state));
3148
3149                 if (i >= newmap->m_num_mds ||
3150                     memcmp(ceph_mdsmap_get_addr(oldmap, i),
3151                            ceph_mdsmap_get_addr(newmap, i),
3152                            sizeof(struct ceph_entity_addr))) {
3153                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3154                                 /* the session never opened, just close it
3155                                  * out now */
3156                                 get_session(s);
3157                                 __unregister_session(mdsc, s);
3158                                 __wake_requests(mdsc, &s->s_waiting);
3159                                 ceph_put_mds_session(s);
3160                         } else if (i >= newmap->m_num_mds) {
3161                                 /* force close session for stopped mds */
3162                                 get_session(s);
3163                                 __unregister_session(mdsc, s);
3164                                 __wake_requests(mdsc, &s->s_waiting);
3165                                 kick_requests(mdsc, i);
3166                                 mutex_unlock(&mdsc->mutex);
3167
3168                                 mutex_lock(&s->s_mutex);
3169                                 cleanup_session_requests(mdsc, s);
3170                                 remove_session_caps(s);
3171                                 mutex_unlock(&s->s_mutex);
3172
3173                                 ceph_put_mds_session(s);
3174
3175                                 mutex_lock(&mdsc->mutex);
3176                         } else {
3177                                 /* just close it */
3178                                 mutex_unlock(&mdsc->mutex);
3179                                 mutex_lock(&s->s_mutex);
3180                                 mutex_lock(&mdsc->mutex);
3181                                 ceph_con_close(&s->s_con);
3182                                 mutex_unlock(&s->s_mutex);
3183                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3184                         }
3185                 } else if (oldstate == newstate) {
3186                         continue;  /* nothing new with this mds */
3187                 }
3188
3189                 /*
3190                  * send reconnect?
3191                  */
3192                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3193                     newstate >= CEPH_MDS_STATE_RECONNECT) {
3194                         mutex_unlock(&mdsc->mutex);
3195                         send_mds_reconnect(mdsc, s);
3196                         mutex_lock(&mdsc->mutex);
3197                 }
3198
3199                 /*
3200                  * kick request on any mds that has gone active.
3201                  */
3202                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3203                     newstate >= CEPH_MDS_STATE_ACTIVE) {
3204                         if (oldstate != CEPH_MDS_STATE_CREATING &&
3205                             oldstate != CEPH_MDS_STATE_STARTING)
3206                                 pr_info("mds%d recovery completed\n", s->s_mds);
3207                         kick_requests(mdsc, i);
3208                         ceph_kick_flushing_caps(mdsc, s);
3209                         wake_up_session_caps(s, 1);
3210                 }
3211         }
3212
3213         for (i = 0; i < newmap->m_num_mds && i < mdsc->max_sessions; i++) {
3214                 s = mdsc->sessions[i];
3215                 if (!s)
3216                         continue;
3217                 if (!ceph_mdsmap_is_laggy(newmap, i))
3218                         continue;
3219                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3220                     s->s_state == CEPH_MDS_SESSION_HUNG ||
3221                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
3222                         dout(" connecting to export targets of laggy mds%d\n",
3223                              i);
3224                         __open_export_target_sessions(mdsc, s);
3225                 }
3226         }
3227 }
3228
3229
3230
3231 /*
3232  * leases
3233  */
3234
3235 /*
3236  * caller must hold session s_mutex, dentry->d_lock
3237  */
3238 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3239 {
3240         struct ceph_dentry_info *di = ceph_dentry(dentry);
3241
3242         ceph_put_mds_session(di->lease_session);
3243         di->lease_session = NULL;
3244 }
3245
3246 static void handle_lease(struct ceph_mds_client *mdsc,
3247                          struct ceph_mds_session *session,
3248                          struct ceph_msg *msg)
3249 {
3250         struct super_block *sb = mdsc->fsc->sb;
3251         struct inode *inode;
3252         struct dentry *parent, *dentry;
3253         struct ceph_dentry_info *di;
3254         int mds = session->s_mds;
3255         struct ceph_mds_lease *h = msg->front.iov_base;
3256         u32 seq;
3257         struct ceph_vino vino;
3258         struct qstr dname;
3259         int release = 0;
3260
3261         dout("handle_lease from mds%d\n", mds);
3262
3263         /* decode */
3264         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3265                 goto bad;
3266         vino.ino = le64_to_cpu(h->ino);
3267         vino.snap = CEPH_NOSNAP;
3268         seq = le32_to_cpu(h->seq);
3269         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3270         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3271         if (dname.len != get_unaligned_le32(h+1))
3272                 goto bad;
3273
3274         /* lookup inode */
3275         inode = ceph_find_inode(sb, vino);
3276         dout("handle_lease %s, ino %llx %p %.*s\n",
3277              ceph_lease_op_name(h->action), vino.ino, inode,
3278              dname.len, dname.name);
3279
3280         mutex_lock(&session->s_mutex);
3281         session->s_seq++;
3282
3283         if (inode == NULL) {
3284                 dout("handle_lease no inode %llx\n", vino.ino);
3285                 goto release;
3286         }
3287
3288         /* dentry */
3289         parent = d_find_alias(inode);
3290         if (!parent) {
3291                 dout("no parent dentry on inode %p\n", inode);
3292                 WARN_ON(1);
3293                 goto release;  /* hrm... */
3294         }
3295         dname.hash = full_name_hash(parent, dname.name, dname.len);
3296         dentry = d_lookup(parent, &dname);
3297         dput(parent);
3298         if (!dentry)
3299                 goto release;
3300
3301         spin_lock(&dentry->d_lock);
3302         di = ceph_dentry(dentry);
3303         switch (h->action) {
3304         case CEPH_MDS_LEASE_REVOKE:
3305                 if (di->lease_session == session) {
3306                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3307                                 h->seq = cpu_to_le32(di->lease_seq);
3308                         __ceph_mdsc_drop_dentry_lease(dentry);
3309                 }
3310                 release = 1;
3311                 break;
3312
3313         case CEPH_MDS_LEASE_RENEW:
3314                 if (di->lease_session == session &&
3315                     di->lease_gen == session->s_cap_gen &&
3316                     di->lease_renew_from &&
3317                     di->lease_renew_after == 0) {
3318                         unsigned long duration =
3319                                 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3320
3321                         di->lease_seq = seq;
3322                         di->time = di->lease_renew_from + duration;
3323                         di->lease_renew_after = di->lease_renew_from +
3324                                 (duration >> 1);
3325                         di->lease_renew_from = 0;
3326                 }
3327                 break;
3328         }
3329         spin_unlock(&dentry->d_lock);
3330         dput(dentry);
3331
3332         if (!release)
3333                 goto out;
3334
3335 release:
3336         /* let's just reuse the same message */
3337         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3338         ceph_msg_get(msg);
3339         ceph_con_send(&session->s_con, msg);
3340
3341 out:
3342         iput(inode);
3343         mutex_unlock(&session->s_mutex);
3344         return;
3345
3346 bad:
3347         pr_err("corrupt lease message\n");
3348         ceph_msg_dump(msg);
3349 }
3350
3351 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3352                               struct inode *inode,
3353                               struct dentry *dentry, char action,
3354                               u32 seq)
3355 {
3356         struct ceph_msg *msg;
3357         struct ceph_mds_lease *lease;
3358         int len = sizeof(*lease) + sizeof(u32);
3359         int dnamelen = 0;
3360
3361         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3362              inode, dentry, ceph_lease_op_name(action), session->s_mds);
3363         dnamelen = dentry->d_name.len;
3364         len += dnamelen;
3365
3366         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3367         if (!msg)
3368                 return;
3369         lease = msg->front.iov_base;
3370         lease->action = action;
3371         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3372         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3373         lease->seq = cpu_to_le32(seq);
3374         put_unaligned_le32(dnamelen, lease + 1);
3375         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3376
3377         /*
3378          * if this is a preemptive lease RELEASE, no need to
3379          * flush request stream, since the actual request will
3380          * soon follow.
3381          */
3382         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3383
3384         ceph_con_send(&session->s_con, msg);
3385 }
3386
3387 /*
3388  * drop all leases (and dentry refs) in preparation for umount
3389  */
3390 static void drop_leases(struct ceph_mds_client *mdsc)
3391 {
3392         int i;
3393
3394         dout("drop_leases\n");
3395         mutex_lock(&mdsc->mutex);
3396         for (i = 0; i < mdsc->max_sessions; i++) {
3397                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3398                 if (!s)
3399                         continue;
3400                 mutex_unlock(&mdsc->mutex);
3401                 mutex_lock(&s->s_mutex);
3402                 mutex_unlock(&s->s_mutex);
3403                 ceph_put_mds_session(s);
3404                 mutex_lock(&mdsc->mutex);
3405         }
3406         mutex_unlock(&mdsc->mutex);
3407 }
3408
3409
3410
3411 /*
3412  * delayed work -- periodically trim expired leases, renew caps with mds
3413  */
3414 static void schedule_delayed(struct ceph_mds_client *mdsc)
3415 {
3416         int delay = 5;
3417         unsigned hz = round_jiffies_relative(HZ * delay);
3418         schedule_delayed_work(&mdsc->delayed_work, hz);
3419 }
3420
3421 static void delayed_work(struct work_struct *work)
3422 {
3423         int i;
3424         struct ceph_mds_client *mdsc =
3425                 container_of(work, struct ceph_mds_client, delayed_work.work);
3426         int renew_interval;
3427         int renew_caps;
3428
3429         dout("mdsc delayed_work\n");
3430         ceph_check_delayed_caps(mdsc);
3431
3432         mutex_lock(&mdsc->mutex);
3433         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3434         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3435                                    mdsc->last_renew_caps);
3436         if (renew_caps)
3437                 mdsc->last_renew_caps = jiffies;
3438
3439         for (i = 0; i < mdsc->max_sessions; i++) {
3440                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3441                 if (s == NULL)
3442                         continue;
3443                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3444                         dout("resending session close request for mds%d\n",
3445                              s->s_mds);
3446                         request_close_session(mdsc, s);
3447                         ceph_put_mds_session(s);
3448                         continue;
3449                 }
3450                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3451                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3452                                 s->s_state = CEPH_MDS_SESSION_HUNG;
3453                                 pr_info("mds%d hung\n", s->s_mds);
3454                         }
3455                 }
3456                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3457                         /* this mds is failed or recovering, just wait */
3458                         ceph_put_mds_session(s);
3459                         continue;
3460                 }
3461                 mutex_unlock(&mdsc->mutex);
3462
3463                 mutex_lock(&s->s_mutex);
3464                 if (renew_caps)
3465                         send_renew_caps(mdsc, s);
3466                 else
3467                         ceph_con_keepalive(&s->s_con);
3468                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3469                     s->s_state == CEPH_MDS_SESSION_HUNG)
3470                         ceph_send_cap_releases(mdsc, s);
3471                 mutex_unlock(&s->s_mutex);
3472                 ceph_put_mds_session(s);
3473
3474                 mutex_lock(&mdsc->mutex);
3475         }
3476         mutex_unlock(&mdsc->mutex);
3477
3478         schedule_delayed(mdsc);
3479 }
3480
3481 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3482
3483 {
3484         struct ceph_mds_client *mdsc;
3485
3486         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3487         if (!mdsc)
3488                 return -ENOMEM;
3489         mdsc->fsc = fsc;
3490         fsc->mdsc = mdsc;
3491         mutex_init(&mdsc->mutex);
3492         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3493         if (mdsc->mdsmap == NULL) {
3494                 kfree(mdsc);
3495                 return -ENOMEM;
3496         }
3497
3498         init_completion(&mdsc->safe_umount_waiters);
3499         init_waitqueue_head(&mdsc->session_close_wq);
3500         INIT_LIST_HEAD(&mdsc->waiting_for_map);
3501         mdsc->sessions = NULL;
3502         atomic_set(&mdsc->num_sessions, 0);
3503         mdsc->max_sessions = 0;
3504         mdsc->stopping = 0;
3505         mdsc->last_snap_seq = 0;
3506         init_rwsem(&mdsc->snap_rwsem);
3507         mdsc->snap_realms = RB_ROOT;
3508         INIT_LIST_HEAD(&mdsc->snap_empty);
3509         spin_lock_init(&mdsc->snap_empty_lock);
3510         mdsc->last_tid = 0;
3511         mdsc->oldest_tid = 0;
3512         mdsc->request_tree = RB_ROOT;
3513         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3514         mdsc->last_renew_caps = jiffies;
3515         INIT_LIST_HEAD(&mdsc->cap_delay_list);
3516         spin_lock_init(&mdsc->cap_delay_lock);
3517         INIT_LIST_HEAD(&mdsc->snap_flush_list);
3518         spin_lock_init(&mdsc->snap_flush_lock);
3519         mdsc->last_cap_flush_tid = 1;
3520         INIT_LIST_HEAD(&mdsc->cap_flush_list);
3521         INIT_LIST_HEAD(&mdsc->cap_dirty);
3522         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3523         mdsc->num_cap_flushing = 0;
3524         spin_lock_init(&mdsc->cap_dirty_lock);
3525         init_waitqueue_head(&mdsc->cap_flushing_wq);
3526         spin_lock_init(&mdsc->dentry_lru_lock);
3527         INIT_LIST_HEAD(&mdsc->dentry_lru);
3528
3529         ceph_caps_init(mdsc);
3530         ceph_adjust_min_caps(mdsc, fsc->min_caps);
3531
3532         init_rwsem(&mdsc->pool_perm_rwsem);
3533         mdsc->pool_perm_tree = RB_ROOT;
3534
3535         return 0;
3536 }
3537
3538 /*
3539  * Wait for safe replies on open mds requests.  If we time out, drop
3540  * all requests from the tree to avoid dangling dentry refs.
3541  */
3542 static void wait_requests(struct ceph_mds_client *mdsc)
3543 {
3544         struct ceph_options *opts = mdsc->fsc->client->options;
3545         struct ceph_mds_request *req;
3546
3547         mutex_lock(&mdsc->mutex);
3548         if (__get_oldest_req(mdsc)) {
3549                 mutex_unlock(&mdsc->mutex);
3550
3551                 dout("wait_requests waiting for requests\n");
3552                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3553                                     ceph_timeout_jiffies(opts->mount_timeout));
3554
3555                 /* tear down remaining requests */
3556                 mutex_lock(&mdsc->mutex);
3557                 while ((req = __get_oldest_req(mdsc))) {
3558                         dout("wait_requests timed out on tid %llu\n",
3559                              req->r_tid);
3560                         __unregister_request(mdsc, req);
3561                 }
3562         }
3563         mutex_unlock(&mdsc->mutex);
3564         dout("wait_requests done\n");
3565 }
3566
3567 /*
3568  * called before mount is ro, and before dentries are torn down.
3569  * (hmm, does this still race with new lookups?)
3570  */
3571 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3572 {
3573         dout("pre_umount\n");
3574         mdsc->stopping = 1;
3575
3576         drop_leases(mdsc);
3577         ceph_flush_dirty_caps(mdsc);
3578         wait_requests(mdsc);
3579
3580         /*
3581          * wait for reply handlers to drop their request refs and
3582          * their inode/dcache refs
3583          */
3584         ceph_msgr_flush();
3585 }
3586
3587 /*
3588  * wait for all write mds requests to flush.
3589  */
3590 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3591 {
3592         struct ceph_mds_request *req = NULL, *nextreq;
3593         struct rb_node *n;
3594
3595         mutex_lock(&mdsc->mutex);
3596         dout("wait_unsafe_requests want %lld\n", want_tid);
3597 restart:
3598         req = __get_oldest_req(mdsc);
3599         while (req && req->r_tid <= want_tid) {
3600                 /* find next request */
3601                 n = rb_next(&req->r_node);
3602                 if (n)
3603                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3604                 else
3605                         nextreq = NULL;
3606                 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3607                     (req->r_op & CEPH_MDS_OP_WRITE)) {
3608                         /* write op */
3609                         ceph_mdsc_get_request(req);
3610                         if (nextreq)
3611                                 ceph_mdsc_get_request(nextreq);
3612                         mutex_unlock(&mdsc->mutex);
3613                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
3614                              req->r_tid, want_tid);
3615                         wait_for_completion(&req->r_safe_completion);
3616                         mutex_lock(&mdsc->mutex);
3617                         ceph_mdsc_put_request(req);
3618                         if (!nextreq)
3619                                 break;  /* next dne before, so we're done! */
3620                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
3621                                 /* next request was removed from tree */
3622                                 ceph_mdsc_put_request(nextreq);
3623                                 goto restart;
3624                         }
3625                         ceph_mdsc_put_request(nextreq);  /* won't go away */
3626                 }
3627                 req = nextreq;
3628         }
3629         mutex_unlock(&mdsc->mutex);
3630         dout("wait_unsafe_requests done\n");
3631 }
3632
3633 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3634 {
3635         u64 want_tid, want_flush;
3636
3637         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3638                 return;
3639
3640         dout("sync\n");
3641         mutex_lock(&mdsc->mutex);
3642         want_tid = mdsc->last_tid;
3643         mutex_unlock(&mdsc->mutex);
3644
3645         ceph_flush_dirty_caps(mdsc);
3646         spin_lock(&mdsc->cap_dirty_lock);
3647         want_flush = mdsc->last_cap_flush_tid;
3648         if (!list_empty(&mdsc->cap_flush_list)) {
3649                 struct ceph_cap_flush *cf =
3650                         list_last_entry(&mdsc->cap_flush_list,
3651                                         struct ceph_cap_flush, g_list);
3652                 cf->wake = true;
3653         }
3654         spin_unlock(&mdsc->cap_dirty_lock);
3655
3656         dout("sync want tid %lld flush_seq %lld\n",
3657              want_tid, want_flush);
3658
3659         wait_unsafe_requests(mdsc, want_tid);
3660         wait_caps_flush(mdsc, want_flush);
3661 }
3662
3663 /*
3664  * true if all sessions are closed, or we force unmount
3665  */
3666 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
3667 {
3668         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3669                 return true;
3670         return atomic_read(&mdsc->num_sessions) <= skipped;
3671 }
3672
3673 /*
3674  * called after sb is ro.
3675  */
3676 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3677 {
3678         struct ceph_options *opts = mdsc->fsc->client->options;
3679         struct ceph_mds_session *session;
3680         int i;
3681         int skipped = 0;
3682
3683         dout("close_sessions\n");
3684
3685         /* close sessions */
3686         mutex_lock(&mdsc->mutex);
3687         for (i = 0; i < mdsc->max_sessions; i++) {
3688                 session = __ceph_lookup_mds_session(mdsc, i);
3689                 if (!session)
3690                         continue;
3691                 mutex_unlock(&mdsc->mutex);
3692                 mutex_lock(&session->s_mutex);
3693                 if (__close_session(mdsc, session) <= 0)
3694                         skipped++;
3695                 mutex_unlock(&session->s_mutex);
3696                 ceph_put_mds_session(session);
3697                 mutex_lock(&mdsc->mutex);
3698         }
3699         mutex_unlock(&mdsc->mutex);
3700
3701         dout("waiting for sessions to close\n");
3702         wait_event_timeout(mdsc->session_close_wq,
3703                            done_closing_sessions(mdsc, skipped),
3704                            ceph_timeout_jiffies(opts->mount_timeout));
3705
3706         /* tear down remaining sessions */
3707         mutex_lock(&mdsc->mutex);
3708         for (i = 0; i < mdsc->max_sessions; i++) {
3709                 if (mdsc->sessions[i]) {
3710                         session = get_session(mdsc->sessions[i]);
3711                         __unregister_session(mdsc, session);
3712                         mutex_unlock(&mdsc->mutex);
3713                         mutex_lock(&session->s_mutex);
3714                         remove_session_caps(session);
3715                         mutex_unlock(&session->s_mutex);
3716                         ceph_put_mds_session(session);
3717                         mutex_lock(&mdsc->mutex);
3718                 }
3719         }
3720         WARN_ON(!list_empty(&mdsc->cap_delay_list));
3721         mutex_unlock(&mdsc->mutex);
3722
3723         ceph_cleanup_empty_realms(mdsc);
3724
3725         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3726
3727         dout("stopped\n");
3728 }
3729
3730 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
3731 {
3732         struct ceph_mds_session *session;
3733         int mds;
3734
3735         dout("force umount\n");
3736
3737         mutex_lock(&mdsc->mutex);
3738         for (mds = 0; mds < mdsc->max_sessions; mds++) {
3739                 session = __ceph_lookup_mds_session(mdsc, mds);
3740                 if (!session)
3741                         continue;
3742                 mutex_unlock(&mdsc->mutex);
3743                 mutex_lock(&session->s_mutex);
3744                 __close_session(mdsc, session);
3745                 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
3746                         cleanup_session_requests(mdsc, session);
3747                         remove_session_caps(session);
3748                 }
3749                 mutex_unlock(&session->s_mutex);
3750                 ceph_put_mds_session(session);
3751                 mutex_lock(&mdsc->mutex);
3752                 kick_requests(mdsc, mds);
3753         }
3754         __wake_requests(mdsc, &mdsc->waiting_for_map);
3755         mutex_unlock(&mdsc->mutex);
3756 }
3757
3758 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3759 {
3760         dout("stop\n");
3761         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3762         if (mdsc->mdsmap)
3763                 ceph_mdsmap_destroy(mdsc->mdsmap);
3764         kfree(mdsc->sessions);
3765         ceph_caps_finalize(mdsc);
3766         ceph_pool_perm_destroy(mdsc);
3767 }
3768
3769 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3770 {
3771         struct ceph_mds_client *mdsc = fsc->mdsc;
3772
3773         dout("mdsc_destroy %p\n", mdsc);
3774         ceph_mdsc_stop(mdsc);
3775
3776         /* flush out any connection work with references to us */
3777         ceph_msgr_flush();
3778
3779         fsc->mdsc = NULL;
3780         kfree(mdsc);
3781         dout("mdsc_destroy %p done\n", mdsc);
3782 }
3783
3784 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3785 {
3786         struct ceph_fs_client *fsc = mdsc->fsc;
3787         const char *mds_namespace = fsc->mount_options->mds_namespace;
3788         void *p = msg->front.iov_base;
3789         void *end = p + msg->front.iov_len;
3790         u32 epoch;
3791         u32 map_len;
3792         u32 num_fs;
3793         u32 mount_fscid = (u32)-1;
3794         u8 struct_v, struct_cv;
3795         int err = -EINVAL;
3796
3797         ceph_decode_need(&p, end, sizeof(u32), bad);
3798         epoch = ceph_decode_32(&p);
3799
3800         dout("handle_fsmap epoch %u\n", epoch);
3801
3802         ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3803         struct_v = ceph_decode_8(&p);
3804         struct_cv = ceph_decode_8(&p);
3805         map_len = ceph_decode_32(&p);
3806
3807         ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
3808         p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
3809
3810         num_fs = ceph_decode_32(&p);
3811         while (num_fs-- > 0) {
3812                 void *info_p, *info_end;
3813                 u32 info_len;
3814                 u8 info_v, info_cv;
3815                 u32 fscid, namelen;
3816
3817                 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3818                 info_v = ceph_decode_8(&p);
3819                 info_cv = ceph_decode_8(&p);
3820                 info_len = ceph_decode_32(&p);
3821                 ceph_decode_need(&p, end, info_len, bad);
3822                 info_p = p;
3823                 info_end = p + info_len;
3824                 p = info_end;
3825
3826                 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
3827                 fscid = ceph_decode_32(&info_p);
3828                 namelen = ceph_decode_32(&info_p);
3829                 ceph_decode_need(&info_p, info_end, namelen, bad);
3830
3831                 if (mds_namespace &&
3832                     strlen(mds_namespace) == namelen &&
3833                     !strncmp(mds_namespace, (char *)info_p, namelen)) {
3834                         mount_fscid = fscid;
3835                         break;
3836                 }
3837         }
3838
3839         ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
3840         if (mount_fscid != (u32)-1) {
3841                 fsc->client->monc.fs_cluster_id = mount_fscid;
3842                 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
3843                                    0, true);
3844                 ceph_monc_renew_subs(&fsc->client->monc);
3845         } else {
3846                 err = -ENOENT;
3847                 goto err_out;
3848         }
3849         return;
3850 bad:
3851         pr_err("error decoding fsmap\n");
3852 err_out:
3853         mutex_lock(&mdsc->mutex);
3854         mdsc->mdsmap_err = -ENOENT;
3855         __wake_requests(mdsc, &mdsc->waiting_for_map);
3856         mutex_unlock(&mdsc->mutex);
3857         return;
3858 }
3859
3860 /*
3861  * handle mds map update.
3862  */
3863 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3864 {
3865         u32 epoch;
3866         u32 maplen;
3867         void *p = msg->front.iov_base;
3868         void *end = p + msg->front.iov_len;
3869         struct ceph_mdsmap *newmap, *oldmap;
3870         struct ceph_fsid fsid;
3871         int err = -EINVAL;
3872
3873         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3874         ceph_decode_copy(&p, &fsid, sizeof(fsid));
3875         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3876                 return;
3877         epoch = ceph_decode_32(&p);
3878         maplen = ceph_decode_32(&p);
3879         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3880
3881         /* do we need it? */
3882         mutex_lock(&mdsc->mutex);
3883         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3884                 dout("handle_map epoch %u <= our %u\n",
3885                      epoch, mdsc->mdsmap->m_epoch);
3886                 mutex_unlock(&mdsc->mutex);
3887                 return;
3888         }
3889
3890         newmap = ceph_mdsmap_decode(&p, end);
3891         if (IS_ERR(newmap)) {
3892                 err = PTR_ERR(newmap);
3893                 goto bad_unlock;
3894         }
3895
3896         /* swap into place */
3897         if (mdsc->mdsmap) {
3898                 oldmap = mdsc->mdsmap;
3899                 mdsc->mdsmap = newmap;
3900                 check_new_map(mdsc, newmap, oldmap);
3901                 ceph_mdsmap_destroy(oldmap);
3902         } else {
3903                 mdsc->mdsmap = newmap;  /* first mds map */
3904         }
3905         mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3906
3907         __wake_requests(mdsc, &mdsc->waiting_for_map);
3908         ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
3909                           mdsc->mdsmap->m_epoch);
3910
3911         mutex_unlock(&mdsc->mutex);
3912         schedule_delayed(mdsc);
3913         return;
3914
3915 bad_unlock:
3916         mutex_unlock(&mdsc->mutex);
3917 bad:
3918         pr_err("error decoding mdsmap %d\n", err);
3919         return;
3920 }
3921
3922 static struct ceph_connection *con_get(struct ceph_connection *con)
3923 {
3924         struct ceph_mds_session *s = con->private;
3925
3926         if (get_session(s)) {
3927                 dout("mdsc con_get %p ok (%d)\n", s, refcount_read(&s->s_ref));
3928                 return con;
3929         }
3930         dout("mdsc con_get %p FAIL\n", s);
3931         return NULL;
3932 }
3933
3934 static void con_put(struct ceph_connection *con)
3935 {
3936         struct ceph_mds_session *s = con->private;
3937
3938         dout("mdsc con_put %p (%d)\n", s, refcount_read(&s->s_ref) - 1);
3939         ceph_put_mds_session(s);
3940 }
3941
3942 /*
3943  * if the client is unresponsive for long enough, the mds will kill
3944  * the session entirely.
3945  */
3946 static void peer_reset(struct ceph_connection *con)
3947 {
3948         struct ceph_mds_session *s = con->private;
3949         struct ceph_mds_client *mdsc = s->s_mdsc;
3950
3951         pr_warn("mds%d closed our session\n", s->s_mds);
3952         send_mds_reconnect(mdsc, s);
3953 }
3954
3955 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3956 {
3957         struct ceph_mds_session *s = con->private;
3958         struct ceph_mds_client *mdsc = s->s_mdsc;
3959         int type = le16_to_cpu(msg->hdr.type);
3960
3961         mutex_lock(&mdsc->mutex);
3962         if (__verify_registered_session(mdsc, s) < 0) {
3963                 mutex_unlock(&mdsc->mutex);
3964                 goto out;
3965         }
3966         mutex_unlock(&mdsc->mutex);
3967
3968         switch (type) {
3969         case CEPH_MSG_MDS_MAP:
3970                 ceph_mdsc_handle_mdsmap(mdsc, msg);
3971                 break;
3972         case CEPH_MSG_FS_MAP_USER:
3973                 ceph_mdsc_handle_fsmap(mdsc, msg);
3974                 break;
3975         case CEPH_MSG_CLIENT_SESSION:
3976                 handle_session(s, msg);
3977                 break;
3978         case CEPH_MSG_CLIENT_REPLY:
3979                 handle_reply(s, msg);
3980                 break;
3981         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3982                 handle_forward(mdsc, s, msg);
3983                 break;
3984         case CEPH_MSG_CLIENT_CAPS:
3985                 ceph_handle_caps(s, msg);
3986                 break;
3987         case CEPH_MSG_CLIENT_SNAP:
3988                 ceph_handle_snap(mdsc, s, msg);
3989                 break;
3990         case CEPH_MSG_CLIENT_LEASE:
3991                 handle_lease(mdsc, s, msg);
3992                 break;
3993
3994         default:
3995                 pr_err("received unknown message type %d %s\n", type,
3996                        ceph_msg_type_name(type));
3997         }
3998 out:
3999         ceph_msg_put(msg);
4000 }
4001
4002 /*
4003  * authentication
4004  */
4005
4006 /*
4007  * Note: returned pointer is the address of a structure that's
4008  * managed separately.  Caller must *not* attempt to free it.
4009  */
4010 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
4011                                         int *proto, int force_new)
4012 {
4013         struct ceph_mds_session *s = con->private;
4014         struct ceph_mds_client *mdsc = s->s_mdsc;
4015         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4016         struct ceph_auth_handshake *auth = &s->s_auth;
4017
4018         if (force_new && auth->authorizer) {
4019                 ceph_auth_destroy_authorizer(auth->authorizer);
4020                 auth->authorizer = NULL;
4021         }
4022         if (!auth->authorizer) {
4023                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4024                                                       auth);
4025                 if (ret)
4026                         return ERR_PTR(ret);
4027         } else {
4028                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4029                                                       auth);
4030                 if (ret)
4031                         return ERR_PTR(ret);
4032         }
4033         *proto = ac->protocol;
4034
4035         return auth;
4036 }
4037
4038
4039 static int verify_authorizer_reply(struct ceph_connection *con)
4040 {
4041         struct ceph_mds_session *s = con->private;
4042         struct ceph_mds_client *mdsc = s->s_mdsc;
4043         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4044
4045         return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
4046 }
4047
4048 static int invalidate_authorizer(struct ceph_connection *con)
4049 {
4050         struct ceph_mds_session *s = con->private;
4051         struct ceph_mds_client *mdsc = s->s_mdsc;
4052         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4053
4054         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
4055
4056         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
4057 }
4058
4059 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
4060                                 struct ceph_msg_header *hdr, int *skip)
4061 {
4062         struct ceph_msg *msg;
4063         int type = (int) le16_to_cpu(hdr->type);
4064         int front_len = (int) le32_to_cpu(hdr->front_len);
4065
4066         if (con->in_msg)
4067                 return con->in_msg;
4068
4069         *skip = 0;
4070         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
4071         if (!msg) {
4072                 pr_err("unable to allocate msg type %d len %d\n",
4073                        type, front_len);
4074                 return NULL;
4075         }
4076
4077         return msg;
4078 }
4079
4080 static int mds_sign_message(struct ceph_msg *msg)
4081 {
4082        struct ceph_mds_session *s = msg->con->private;
4083        struct ceph_auth_handshake *auth = &s->s_auth;
4084
4085        return ceph_auth_sign_message(auth, msg);
4086 }
4087
4088 static int mds_check_message_signature(struct ceph_msg *msg)
4089 {
4090        struct ceph_mds_session *s = msg->con->private;
4091        struct ceph_auth_handshake *auth = &s->s_auth;
4092
4093        return ceph_auth_check_message_signature(auth, msg);
4094 }
4095
4096 static const struct ceph_connection_operations mds_con_ops = {
4097         .get = con_get,
4098         .put = con_put,
4099         .dispatch = dispatch,
4100         .get_authorizer = get_authorizer,
4101         .verify_authorizer_reply = verify_authorizer_reply,
4102         .invalidate_authorizer = invalidate_authorizer,
4103         .peer_reset = peer_reset,
4104         .alloc_msg = mds_alloc_msg,
4105         .sign_message = mds_sign_message,
4106         .check_message_signature = mds_check_message_signature,
4107 };
4108
4109 /* eof */