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/ratelimit.h>
  11
  12 #include "super.h"
  13 #include "mds_client.h"
  14
  15 #include <linux/ceph/ceph_features.h>
  16 #include <linux/ceph/messenger.h>
  17 #include <linux/ceph/decode.h>
  18 #include <linux/ceph/pagelist.h>
  19 #include <linux/ceph/auth.h>
  20 #include <linux/ceph/debugfs.h>
  21
  22 /*
  23  * A cluster of MDS (metadata server) daemons is responsible for
  24  * managing the file system namespace (the directory hierarchy and
  25  * inodes) and for coordinating shared access to storage.  Metadata is
  26  * partitioning hierarchically across a number of servers, and that
  27  * partition varies over time as the cluster adjusts the distribution
  28  * in order to balance load.
  29  *
  30  * The MDS client is primarily responsible to managing synchronous
  31  * metadata requests for operations like open, unlink, and so forth.
  32  * If there is a MDS failure, we find out about it when we (possibly
  33  * request and) receive a new MDS map, and can resubmit affected
  34  * requests.
  35  *
  36  * For the most part, though, we take advantage of a lossless
  37  * communications channel to the MDS, and do not need to worry about
  38  * timing out or resubmitting requests.
  39  *
  40  * We maintain a stateful "session" with each MDS we interact with.
  41  * Within each session, we sent periodic heartbeat messages to ensure
  42  * any capabilities or leases we have been issues remain valid.  If
  43  * the session times out and goes stale, our leases and capabilities
  44  * are no longer valid.
  45  */
  46
  47 struct ceph_reconnect_state {
  48         int nr_caps;
  49         struct ceph_pagelist *pagelist;
  50         unsigned msg_version;
  51 };
  52
  53 static void __wake_requests(struct ceph_mds_client *mdsc,
  54                             struct list_head *head);
  55
  56 static const struct ceph_connection_operations mds_con_ops;
  57
  58
  59 /*
  60  * mds reply parsing
  61  */
  62
  63 /*
  64  * parse individual inode info
  65  */
  66 static int parse_reply_info_in(void **p, void *end,
  67                                struct ceph_mds_reply_info_in *info,
  68                                u64 features)
  69 {
  70         int err = -EIO;
  71
  72         info->in = *p;
  73         *p += sizeof(struct ceph_mds_reply_inode) +
  74                 sizeof(*info->in->fragtree.splits) *
  75                 le32_to_cpu(info->in->fragtree.nsplits);
  76
  77         ceph_decode_32_safe(p, end, info->symlink_len, bad);
  78         ceph_decode_need(p, end, info->symlink_len, bad);
  79         info->symlink = *p;
  80         *p += info->symlink_len;
  81
  82         if (features & CEPH_FEATURE_DIRLAYOUTHASH)
  83                 ceph_decode_copy_safe(p, end, &info->dir_layout,
  84                                       sizeof(info->dir_layout), bad);
  85         else
  86                 memset(&info->dir_layout, 0, sizeof(info->dir_layout));
  87
  88         ceph_decode_32_safe(p, end, info->xattr_len, bad);
  89         ceph_decode_need(p, end, info->xattr_len, bad);
  90         info->xattr_data = *p;
  91         *p += info->xattr_len;
  92
  93         if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
  94                 ceph_decode_64_safe(p, end, info->inline_version, bad);
  95                 ceph_decode_32_safe(p, end, info->inline_len, bad);
  96                 ceph_decode_need(p, end, info->inline_len, bad);
  97                 info->inline_data = *p;
  98                 *p += info->inline_len;
  99         } else
 100                 info->inline_version = CEPH_INLINE_NONE;
 101
 102         info->pool_ns_len = 0;
 103         info->pool_ns_data = NULL;
 104         if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
 105                 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
 106                 if (info->pool_ns_len > 0) {
 107                         ceph_decode_need(p, end, info->pool_ns_len, bad);
 108                         info->pool_ns_data = *p;
 109                         *p += info->pool_ns_len;
 110                 }
 111         }
 112
 113         return 0;
 114 bad:
 115         return err;
 116 }
 117
 118 /*
 119  * parse a normal reply, which may contain a (dir+)dentry and/or a
 120  * target inode.
 121  */
 122 static int parse_reply_info_trace(void **p, void *end,
 123                                   struct ceph_mds_reply_info_parsed *info,
 124                                   u64 features)
 125 {
 126         int err;
 127
 128         if (info->head->is_dentry) {
 129                 err = parse_reply_info_in(p, end, &info->diri, features);
 130                 if (err < 0)
 131                         goto out_bad;
 132
 133                 if (unlikely(*p + sizeof(*info->dirfrag) > end))
 134                         goto bad;
 135                 info->dirfrag = *p;
 136                 *p += sizeof(*info->dirfrag) +
 137                         sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
 138                 if (unlikely(*p > end))
 139                         goto bad;
 140
 141                 ceph_decode_32_safe(p, end, info->dname_len, bad);
 142                 ceph_decode_need(p, end, info->dname_len, bad);
 143                 info->dname = *p;
 144                 *p += info->dname_len;
 145                 info->dlease = *p;
 146                 *p += sizeof(*info->dlease);
 147         }
 148
 149         if (info->head->is_target) {
 150                 err = parse_reply_info_in(p, end, &info->targeti, features);
 151                 if (err < 0)
 152                         goto out_bad;
 153         }
 154
 155         if (unlikely(*p != end))
 156                 goto bad;
 157         return 0;
 158
 159 bad:
 160         err = -EIO;
 161 out_bad:
 162         pr_err("problem parsing mds trace %d\n", err);
 163         return err;
 164 }
 165
 166 /*
 167  * parse readdir results
 168  */
 169 static int parse_reply_info_dir(void **p, void *end,
 170                                 struct ceph_mds_reply_info_parsed *info,
 171                                 u64 features)
 172 {
 173         u32 num, i = 0;
 174         int err;
 175
 176         info->dir_dir = *p;
 177         if (*p + sizeof(*info->dir_dir) > end)
 178                 goto bad;
 179         *p += sizeof(*info->dir_dir) +
 180                 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
 181         if (*p > end)
 182                 goto bad;
 183
 184         ceph_decode_need(p, end, sizeof(num) + 2, bad);
 185         num = ceph_decode_32(p);
 186         {
 187                 u16 flags = ceph_decode_16(p);
 188                 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
 189                 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
 190                 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
 191                 info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
 192         }
 193         if (num == 0)
 194                 goto done;
 195
 196         BUG_ON(!info->dir_entries);
 197         if ((unsigned long)(info->dir_entries + num) >
 198             (unsigned long)info->dir_entries + info->dir_buf_size) {
 199                 pr_err("dir contents are larger than expected\n");
 200                 WARN_ON(1);
 201                 goto bad;
 202         }
 203
 204         info->dir_nr = num;
 205         while (num) {
 206                 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
 207                 /* dentry */
 208                 ceph_decode_need(p, end, sizeof(u32)*2, bad);
 209                 rde->name_len = ceph_decode_32(p);
 210                 ceph_decode_need(p, end, rde->name_len, bad);
 211                 rde->name = *p;
 212                 *p += rde->name_len;
 213                 dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
 214                 rde->lease = *p;
 215                 *p += sizeof(struct ceph_mds_reply_lease);
 216
 217                 /* inode */
 218                 err = parse_reply_info_in(p, end, &rde->inode, features);
 219                 if (err < 0)
 220                         goto out_bad;
 221                 /* ceph_readdir_prepopulate() will update it */
 222                 rde->offset = 0;
 223                 i++;
 224                 num--;
 225         }
 226
 227 done:
 228         if (*p != end)
 229                 goto bad;
 230         return 0;
 231
 232 bad:
 233         err = -EIO;
 234 out_bad:
 235         pr_err("problem parsing dir contents %d\n", err);
 236         return err;
 237 }
 238
 239 /*
 240  * parse fcntl F_GETLK results
 241  */
 242 static int parse_reply_info_filelock(void **p, void *end,
 243                                      struct ceph_mds_reply_info_parsed *info,
 244                                      u64 features)
 245 {
 246         if (*p + sizeof(*info->filelock_reply) > end)
 247                 goto bad;
 248
 249         info->filelock_reply = *p;
 250         *p += sizeof(*info->filelock_reply);
 251
 252         if (unlikely(*p != end))
 253                 goto bad;
 254         return 0;
 255
 256 bad:
 257         return -EIO;
 258 }
 259
 260 /*
 261  * parse create results
 262  */
 263 static int parse_reply_info_create(void **p, void *end,
 264                                   struct ceph_mds_reply_info_parsed *info,
 265                                   u64 features)
 266 {
 267         if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
 268                 if (*p == end) {
 269                         info->has_create_ino = false;
 270                 } else {
 271                         info->has_create_ino = true;
 272                         info->ino = ceph_decode_64(p);
 273                 }
 274         }
 275
 276         if (unlikely(*p != end))
 277                 goto bad;
 278         return 0;
 279
 280 bad:
 281         return -EIO;
 282 }
 283
 284 /*
 285  * parse extra results
 286  */
 287 static int parse_reply_info_extra(void **p, void *end,
 288                                   struct ceph_mds_reply_info_parsed *info,
 289                                   u64 features)
 290 {
 291         u32 op = le32_to_cpu(info->head->op);
 292
 293         if (op == CEPH_MDS_OP_GETFILELOCK)
 294                 return parse_reply_info_filelock(p, end, info, features);
 295         else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
 296                 return parse_reply_info_dir(p, end, info, features);
 297         else if (op == CEPH_MDS_OP_CREATE)
 298                 return parse_reply_info_create(p, end, info, features);
 299         else
 300                 return -EIO;
 301 }
 302
 303 /*
 304  * parse entire mds reply
 305  */
 306 static int parse_reply_info(struct ceph_msg *msg,
 307                             struct ceph_mds_reply_info_parsed *info,
 308                             u64 features)
 309 {
 310         void *p, *end;
 311         u32 len;
 312         int err;
 313
 314         info->head = msg->front.iov_base;
 315         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 316         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 317
 318         /* trace */
 319         ceph_decode_32_safe(&p, end, len, bad);
 320         if (len > 0) {
 321                 ceph_decode_need(&p, end, len, bad);
 322                 err = parse_reply_info_trace(&p, p+len, info, features);
 323                 if (err < 0)
 324                         goto out_bad;
 325         }
 326
 327         /* extra */
 328         ceph_decode_32_safe(&p, end, len, bad);
 329         if (len > 0) {
 330                 ceph_decode_need(&p, end, len, bad);
 331                 err = parse_reply_info_extra(&p, p+len, info, features);
 332                 if (err < 0)
 333                         goto out_bad;
 334         }
 335
 336         /* snap blob */
 337         ceph_decode_32_safe(&p, end, len, bad);
 338         info->snapblob_len = len;
 339         info->snapblob = p;
 340         p += len;
 341
 342         if (p != end)
 343                 goto bad;
 344         return 0;
 345
 346 bad:
 347         err = -EIO;
 348 out_bad:
 349         pr_err("mds parse_reply err %d\n", err);
 350         return err;
 351 }
 352
 353 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
 354 {
 355         if (!info->dir_entries)
 356                 return;
 357         free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
 358 }
 359
 360
 361 /*
 362  * sessions
 363  */
 364 const char *ceph_session_state_name(int s)
 365 {
 366         switch (s) {
 367         case CEPH_MDS_SESSION_NEW: return "new";
 368         case CEPH_MDS_SESSION_OPENING: return "opening";
 369         case CEPH_MDS_SESSION_OPEN: return "open";
 370         case CEPH_MDS_SESSION_HUNG: return "hung";
 371         case CEPH_MDS_SESSION_CLOSING: return "closing";
 372         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
 373         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
 374         case CEPH_MDS_SESSION_REJECTED: return "rejected";
 375         default: return "???";
 376         }
 377 }
 378
 379 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
 380 {
 381         if (refcount_inc_not_zero(&s->s_ref)) {
 382                 dout("mdsc get_session %p %d -> %d\n", s,
 383                      refcount_read(&s->s_ref)-1, refcount_read(&s->s_ref));
 384                 return s;
 385         } else {
 386                 dout("mdsc get_session %p 0 -- FAIL", s);
 387                 return NULL;
 388         }
 389 }
 390
 391 void ceph_put_mds_session(struct ceph_mds_session *s)
 392 {
 393         dout("mdsc put_session %p %d -> %d\n", s,
 394              refcount_read(&s->s_ref), refcount_read(&s->s_ref)-1);
 395         if (refcount_dec_and_test(&s->s_ref)) {
 396                 if (s->s_auth.authorizer)
 397                         ceph_auth_destroy_authorizer(s->s_auth.authorizer);
 398                 kfree(s);
 399         }
 400 }
 401
 402 /*
 403  * called under mdsc->mutex
 404  */
 405 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
 406                                                    int mds)
 407 {
 408         struct ceph_mds_session *session;
 409
 410         if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
 411                 return NULL;
 412         session = mdsc->sessions[mds];
 413         dout("lookup_mds_session %p %d\n", session,
 414              refcount_read(&session->s_ref));
 415         get_session(session);
 416         return session;
 417 }
 418
 419 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
 420 {
 421         if (mds >= mdsc->max_sessions)
 422                 return false;
 423         return mdsc->sessions[mds];
 424 }
 425
 426 static int __verify_registered_session(struct ceph_mds_client *mdsc,
 427                                        struct ceph_mds_session *s)
 428 {
 429         if (s->s_mds >= mdsc->max_sessions ||
 430             mdsc->sessions[s->s_mds] != s)
 431                 return -ENOENT;
 432         return 0;
 433 }
 434
 435 /*
 436  * create+register a new session for given mds.
 437  * called under mdsc->mutex.
 438  */
 439 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 440                                                  int mds)
 441 {
 442         struct ceph_mds_session *s;
 443
 444         if (mds >= mdsc->mdsmap->m_num_mds)
 445                 return ERR_PTR(-EINVAL);
 446
 447         s = kzalloc(sizeof(*s), GFP_NOFS);
 448         if (!s)
 449                 return ERR_PTR(-ENOMEM);
 450         s->s_mdsc = mdsc;
 451         s->s_mds = mds;
 452         s->s_state = CEPH_MDS_SESSION_NEW;
 453         s->s_ttl = 0;
 454         s->s_seq = 0;
 455         mutex_init(&s->s_mutex);
 456
 457         ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
 458
 459         spin_lock_init(&s->s_gen_ttl_lock);
 460         s->s_cap_gen = 0;
 461         s->s_cap_ttl = jiffies - 1;
 462
 463         spin_lock_init(&s->s_cap_lock);
 464         s->s_renew_requested = 0;
 465         s->s_renew_seq = 0;
 466         INIT_LIST_HEAD(&s->s_caps);
 467         s->s_nr_caps = 0;
 468         s->s_trim_caps = 0;
 469         refcount_set(&s->s_ref, 1);
 470         INIT_LIST_HEAD(&s->s_waiting);
 471         INIT_LIST_HEAD(&s->s_unsafe);
 472         s->s_num_cap_releases = 0;
 473         s->s_cap_reconnect = 0;
 474         s->s_cap_iterator = NULL;
 475         INIT_LIST_HEAD(&s->s_cap_releases);
 476         INIT_LIST_HEAD(&s->s_cap_flushing);
 477
 478         dout("register_session mds%d\n", mds);
 479         if (mds >= mdsc->max_sessions) {
 480                 int newmax = 1 << get_count_order(mds+1);
 481                 struct ceph_mds_session **sa;
 482
 483                 dout("register_session realloc to %d\n", newmax);
 484                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 485                 if (!sa)
 486                         goto fail_realloc;
 487                 if (mdsc->sessions) {
 488                         memcpy(sa, mdsc->sessions,
 489                                mdsc->max_sessions * sizeof(void *));
 490                         kfree(mdsc->sessions);
 491                 }
 492                 mdsc->sessions = sa;
 493                 mdsc->max_sessions = newmax;
 494         }
 495         mdsc->sessions[mds] = s;
 496         atomic_inc(&mdsc->num_sessions);
 497         refcount_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
 498
 499         ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
 500                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
 501
 502         return s;
 503
 504 fail_realloc:
 505         kfree(s);
 506         return ERR_PTR(-ENOMEM);
 507 }
 508
 509 /*
 510  * called under mdsc->mutex
 511  */
 512 static void __unregister_session(struct ceph_mds_client *mdsc,
 513                                struct ceph_mds_session *s)
 514 {
 515         dout("__unregister_session mds%d %p\n", s->s_mds, s);
 516         BUG_ON(mdsc->sessions[s->s_mds] != s);
 517         mdsc->sessions[s->s_mds] = NULL;
 518         ceph_con_close(&s->s_con);
 519         ceph_put_mds_session(s);
 520         atomic_dec(&mdsc->num_sessions);
 521 }
 522
 523 /*
 524  * drop session refs in request.
 525  *
 526  * should be last request ref, or hold mdsc->mutex
 527  */
 528 static void put_request_session(struct ceph_mds_request *req)
 529 {
 530         if (req->r_session) {
 531                 ceph_put_mds_session(req->r_session);
 532                 req->r_session = NULL;
 533         }
 534 }
 535
 536 void ceph_mdsc_release_request(struct kref *kref)
 537 {
 538         struct ceph_mds_request *req = container_of(kref,
 539                                                     struct ceph_mds_request,
 540                                                     r_kref);
 541         destroy_reply_info(&req->r_reply_info);
 542         if (req->r_request)
 543                 ceph_msg_put(req->r_request);
 544         if (req->r_reply)
 545                 ceph_msg_put(req->r_reply);
 546         if (req->r_inode) {
 547                 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
 548                 iput(req->r_inode);
 549         }
 550         if (req->r_parent)
 551                 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
 552         iput(req->r_target_inode);
 553         if (req->r_dentry)
 554                 dput(req->r_dentry);
 555         if (req->r_old_dentry)
 556                 dput(req->r_old_dentry);
 557         if (req->r_old_dentry_dir) {
 558                 /*
 559                  * track (and drop pins for) r_old_dentry_dir
 560                  * separately, since r_old_dentry's d_parent may have
 561                  * changed between the dir mutex being dropped and
 562                  * this request being freed.
 563                  */
 564                 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
 565                                   CEPH_CAP_PIN);
 566                 iput(req->r_old_dentry_dir);
 567         }
 568         kfree(req->r_path1);
 569         kfree(req->r_path2);
 570         if (req->r_pagelist)
 571                 ceph_pagelist_release(req->r_pagelist);
 572         put_request_session(req);
 573         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
 574         kfree(req);
 575 }
 576
 577 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
 578
 579 /*
 580  * lookup session, bump ref if found.
 581  *
 582  * called under mdsc->mutex.
 583  */
 584 static struct ceph_mds_request *
 585 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
 586 {
 587         struct ceph_mds_request *req;
 588
 589         req = lookup_request(&mdsc->request_tree, tid);
 590         if (req)
 591                 ceph_mdsc_get_request(req);
 592
 593         return req;
 594 }
 595
 596 /*
 597  * Register an in-flight request, and assign a tid.  Link to directory
 598  * are modifying (if any).
 599  *
 600  * Called under mdsc->mutex.
 601  */
 602 static void __register_request(struct ceph_mds_client *mdsc,
 603                                struct ceph_mds_request *req,
 604                                struct inode *dir)
 605 {
 606         req->r_tid = ++mdsc->last_tid;
 607         if (req->r_num_caps)
 608                 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
 609                                   req->r_num_caps);
 610         dout("__register_request %p tid %lld\n", req, req->r_tid);
 611         ceph_mdsc_get_request(req);
 612         insert_request(&mdsc->request_tree, req);
 613
 614         req->r_uid = current_fsuid();
 615         req->r_gid = current_fsgid();
 616
 617         if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
 618                 mdsc->oldest_tid = req->r_tid;
 619
 620         if (dir) {
 621                 ihold(dir);
 622                 req->r_unsafe_dir = dir;
 623         }
 624 }
 625
 626 static void __unregister_request(struct ceph_mds_client *mdsc,
 627                                  struct ceph_mds_request *req)
 628 {
 629         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 630
 631         /* Never leave an unregistered request on an unsafe list! */
 632         list_del_init(&req->r_unsafe_item);
 633
 634         if (req->r_tid == mdsc->oldest_tid) {
 635                 struct rb_node *p = rb_next(&req->r_node);
 636                 mdsc->oldest_tid = 0;
 637                 while (p) {
 638                         struct ceph_mds_request *next_req =
 639                                 rb_entry(p, struct ceph_mds_request, r_node);
 640                         if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
 641                                 mdsc->oldest_tid = next_req->r_tid;
 642                                 break;
 643                         }
 644                         p = rb_next(p);
 645                 }
 646         }
 647
 648         erase_request(&mdsc->request_tree, req);
 649
 650         if (req->r_unsafe_dir  &&
 651             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
 652                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 653                 spin_lock(&ci->i_unsafe_lock);
 654                 list_del_init(&req->r_unsafe_dir_item);
 655                 spin_unlock(&ci->i_unsafe_lock);
 656         }
 657         if (req->r_target_inode &&
 658             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
 659                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
 660                 spin_lock(&ci->i_unsafe_lock);
 661                 list_del_init(&req->r_unsafe_target_item);
 662                 spin_unlock(&ci->i_unsafe_lock);
 663         }
 664
 665         if (req->r_unsafe_dir) {
 666                 iput(req->r_unsafe_dir);
 667                 req->r_unsafe_dir = NULL;
 668         }
 669
 670         complete_all(&req->r_safe_completion);
 671
 672         ceph_mdsc_put_request(req);
 673 }
 674
 675 /*
 676  * Walk back up the dentry tree until we hit a dentry representing a
 677  * non-snapshot inode. We do this using the rcu_read_lock (which must be held
 678  * when calling this) to ensure that the objects won't disappear while we're
 679  * working with them. Once we hit a candidate dentry, we attempt to take a
 680  * reference to it, and return that as the result.
 681  */
 682 static struct inode *get_nonsnap_parent(struct dentry *dentry)
 683 {
 684         struct inode *inode = NULL;
 685
 686         while (dentry && !IS_ROOT(dentry)) {
 687                 inode = d_inode_rcu(dentry);
 688                 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
 689                         break;
 690                 dentry = dentry->d_parent;
 691         }
 692         if (inode)
 693                 inode = igrab(inode);
 694         return inode;
 695 }
 696
 697 /*
 698  * Choose mds to send request to next.  If there is a hint set in the
 699  * request (e.g., due to a prior forward hint from the mds), use that.
 700  * Otherwise, consult frag tree and/or caps to identify the
 701  * appropriate mds.  If all else fails, choose randomly.
 702  *
 703  * Called under mdsc->mutex.
 704  */
 705 static int __choose_mds(struct ceph_mds_client *mdsc,
 706                         struct ceph_mds_request *req)
 707 {
 708         struct inode *inode;
 709         struct ceph_inode_info *ci;
 710         struct ceph_cap *cap;
 711         int mode = req->r_direct_mode;
 712         int mds = -1;
 713         u32 hash = req->r_direct_hash;
 714         bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
 715
 716         /*
 717          * is there a specific mds we should try?  ignore hint if we have
 718          * no session and the mds is not up (active or recovering).
 719          */
 720         if (req->r_resend_mds >= 0 &&
 721             (__have_session(mdsc, req->r_resend_mds) ||
 722              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 723                 dout("choose_mds using resend_mds mds%d\n",
 724                      req->r_resend_mds);
 725                 return req->r_resend_mds;
 726         }
 727
 728         if (mode == USE_RANDOM_MDS)
 729                 goto random;
 730
 731         inode = NULL;
 732         if (req->r_inode) {
 733                 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
 734                         inode = req->r_inode;
 735                         ihold(inode);
 736                 } else {
 737                         /* req->r_dentry is non-null for LSSNAP request */
 738                         rcu_read_lock();
 739                         inode = get_nonsnap_parent(req->r_dentry);
 740                         rcu_read_unlock();
 741                         dout("__choose_mds using snapdir's parent %p\n", inode);
 742                 }
 743         } else if (req->r_dentry) {
 744                 /* ignore race with rename; old or new d_parent is okay */
 745                 struct dentry *parent;
 746                 struct inode *dir;
 747
 748                 rcu_read_lock();
 749                 parent = req->r_dentry->d_parent;
 750                 dir = req->r_parent ? : d_inode_rcu(parent);
 751
 752                 if (!dir || dir->i_sb != mdsc->fsc->sb) {
 753                         /*  not this fs or parent went negative */
 754                         inode = d_inode(req->r_dentry);
 755                         if (inode)
 756                                 ihold(inode);
 757                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
 758                         /* direct snapped/virtual snapdir requests
 759                          * based on parent dir inode */
 760                         inode = get_nonsnap_parent(parent);
 761                         dout("__choose_mds using nonsnap parent %p\n", inode);
 762                 } else {
 763                         /* dentry target */
 764                         inode = d_inode(req->r_dentry);
 765                         if (!inode || mode == USE_AUTH_MDS) {
 766                                 /* dir + name */
 767                                 inode = igrab(dir);
 768                                 hash = ceph_dentry_hash(dir, req->r_dentry);
 769                                 is_hash = true;
 770                         } else {
 771                                 ihold(inode);
 772                         }
 773                 }
 774                 rcu_read_unlock();
 775         }
 776
 777         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 778              (int)hash, mode);
 779         if (!inode)
 780                 goto random;
 781         ci = ceph_inode(inode);
 782
 783         if (is_hash && S_ISDIR(inode->i_mode)) {
 784                 struct ceph_inode_frag frag;
 785                 int found;
 786
 787                 ceph_choose_frag(ci, hash, &frag, &found);
 788                 if (found) {
 789                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 790                                 u8 r;
 791
 792                                 /* choose a random replica */
 793                                 get_random_bytes(&r, 1);
 794                                 r %= frag.ndist;
 795                                 mds = frag.dist[r];
 796                                 dout("choose_mds %p %llx.%llx "
 797                                      "frag %u mds%d (%d/%d)\n",
 798                                      inode, ceph_vinop(inode),
 799                                      frag.frag, mds,
 800                                      (int)r, frag.ndist);
 801                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 802                                     CEPH_MDS_STATE_ACTIVE)
 803                                         goto out;
 804                         }
 805
 806                         /* since this file/dir wasn't known to be
 807                          * replicated, then we want to look for the
 808                          * authoritative mds. */
 809                         mode = USE_AUTH_MDS;
 810                         if (frag.mds >= 0) {
 811                                 /* choose auth mds */
 812                                 mds = frag.mds;
 813                                 dout("choose_mds %p %llx.%llx "
 814                                      "frag %u mds%d (auth)\n",
 815                                      inode, ceph_vinop(inode), frag.frag, mds);
 816                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 817                                     CEPH_MDS_STATE_ACTIVE)
 818                                         goto out;
 819                         }
 820                 }
 821         }
 822
 823         spin_lock(&ci->i_ceph_lock);
 824         cap = NULL;
 825         if (mode == USE_AUTH_MDS)
 826                 cap = ci->i_auth_cap;
 827         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
 828                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
 829         if (!cap) {
 830                 spin_unlock(&ci->i_ceph_lock);
 831                 iput(inode);
 832                 goto random;
 833         }
 834         mds = cap->session->s_mds;
 835         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
 836              inode, ceph_vinop(inode), mds,
 837              cap == ci->i_auth_cap ? "auth " : "", cap);
 838         spin_unlock(&ci->i_ceph_lock);
 839 out:
 840         iput(inode);
 841         return mds;
 842
 843 random:
 844         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
 845         dout("choose_mds chose random mds%d\n", mds);
 846         return mds;
 847 }
 848
 849
 850 /*
 851  * session messages
 852  */
 853 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
 854 {
 855         struct ceph_msg *msg;
 856         struct ceph_mds_session_head *h;
 857
 858         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
 859                            false);
 860         if (!msg) {
 861                 pr_err("create_session_msg ENOMEM creating msg\n");
 862                 return NULL;
 863         }
 864         h = msg->front.iov_base;
 865         h->op = cpu_to_le32(op);
 866         h->seq = cpu_to_le64(seq);
 867
 868         return msg;
 869 }
 870
 871 /*
 872  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
 873  * to include additional client metadata fields.
 874  */
 875 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
 876 {
 877         struct ceph_msg *msg;
 878         struct ceph_mds_session_head *h;
 879         int i = -1;
 880         int metadata_bytes = 0;
 881         int metadata_key_count = 0;
 882         struct ceph_options *opt = mdsc->fsc->client->options;
 883         struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
 884         void *p;
 885
 886         const char* metadata[][2] = {
 887                 {"hostname", mdsc->nodename},
 888                 {"kernel_version", init_utsname()->release},
 889                 {"entity_id", opt->name ? : ""},
 890                 {"root", fsopt->server_path ? : "/"},
 891                 {NULL, NULL}
 892         };
 893
 894         /* Calculate serialized length of metadata */
 895         metadata_bytes = 4;  /* map length */
 896         for (i = 0; metadata[i][0]; ++i) {
 897                 metadata_bytes += 8 + strlen(metadata[i][0]) +
 898                         strlen(metadata[i][1]);
 899                 metadata_key_count++;
 900         }
 901
 902         /* Allocate the message */
 903         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + metadata_bytes,
 904                            GFP_NOFS, false);
 905         if (!msg) {
 906                 pr_err("create_session_msg ENOMEM creating msg\n");
 907                 return NULL;
 908         }
 909         h = msg->front.iov_base;
 910         h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
 911         h->seq = cpu_to_le64(seq);
 912
 913         /*
 914          * Serialize client metadata into waiting buffer space, using
 915          * the format that userspace expects for map<string, string>
 916          *
 917          * ClientSession messages with metadata are v2
 918          */
 919         msg->hdr.version = cpu_to_le16(2);
 920         msg->hdr.compat_version = cpu_to_le16(1);
 921
 922         /* The write pointer, following the session_head structure */
 923         p = msg->front.iov_base + sizeof(*h);
 924
 925         /* Number of entries in the map */
 926         ceph_encode_32(&p, metadata_key_count);
 927
 928         /* Two length-prefixed strings for each entry in the map */
 929         for (i = 0; metadata[i][0]; ++i) {
 930                 size_t const key_len = strlen(metadata[i][0]);
 931                 size_t const val_len = strlen(metadata[i][1]);
 932
 933                 ceph_encode_32(&p, key_len);
 934                 memcpy(p, metadata[i][0], key_len);
 935                 p += key_len;
 936                 ceph_encode_32(&p, val_len);
 937                 memcpy(p, metadata[i][1], val_len);
 938                 p += val_len;
 939         }
 940
 941         return msg;
 942 }
 943
 944 /*
 945  * send session open request.
 946  *
 947  * called under mdsc->mutex
 948  */
 949 static int __open_session(struct ceph_mds_client *mdsc,
 950                           struct ceph_mds_session *session)
 951 {
 952         struct ceph_msg *msg;
 953         int mstate;
 954         int mds = session->s_mds;
 955
 956         /* wait for mds to go active? */
 957         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
 958         dout("open_session to mds%d (%s)\n", mds,
 959              ceph_mds_state_name(mstate));
 960         session->s_state = CEPH_MDS_SESSION_OPENING;
 961         session->s_renew_requested = jiffies;
 962
 963         /* send connect message */
 964         msg = create_session_open_msg(mdsc, session->s_seq);
 965         if (!msg)
 966                 return -ENOMEM;
 967         ceph_con_send(&session->s_con, msg);
 968         return 0;
 969 }
 970
 971 /*
 972  * open sessions for any export targets for the given mds
 973  *
 974  * called under mdsc->mutex
 975  */
 976 static struct ceph_mds_session *
 977 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
 978 {
 979         struct ceph_mds_session *session;
 980
 981         session = __ceph_lookup_mds_session(mdsc, target);
 982         if (!session) {
 983                 session = register_session(mdsc, target);
 984                 if (IS_ERR(session))
 985                         return session;
 986         }
 987         if (session->s_state == CEPH_MDS_SESSION_NEW ||
 988             session->s_state == CEPH_MDS_SESSION_CLOSING)
 989                 __open_session(mdsc, session);
 990
 991         return session;
 992 }
 993
 994 struct ceph_mds_session *
 995 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
 996 {
 997         struct ceph_mds_session *session;
 998
 999         dout("open_export_target_session to mds%d\n", target);
1000
1001         mutex_lock(&mdsc->mutex);
1002         session = __open_export_target_session(mdsc, target);
1003         mutex_unlock(&mdsc->mutex);
1004
1005         return session;
1006 }
1007
1008 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1009                                           struct ceph_mds_session *session)
1010 {
1011         struct ceph_mds_info *mi;
1012         struct ceph_mds_session *ts;
1013         int i, mds = session->s_mds;
1014
1015         if (mds >= mdsc->mdsmap->m_num_mds)
1016                 return;
1017
1018         mi = &mdsc->mdsmap->m_info[mds];
1019         dout("open_export_target_sessions for mds%d (%d targets)\n",
1020              session->s_mds, mi->num_export_targets);
1021
1022         for (i = 0; i < mi->num_export_targets; i++) {
1023                 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1024                 if (!IS_ERR(ts))
1025                         ceph_put_mds_session(ts);
1026         }
1027 }
1028
1029 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1030                                            struct ceph_mds_session *session)
1031 {
1032         mutex_lock(&mdsc->mutex);
1033         __open_export_target_sessions(mdsc, session);
1034         mutex_unlock(&mdsc->mutex);
1035 }
1036
1037 /*
1038  * session caps
1039  */
1040
1041 /* caller holds s_cap_lock, we drop it */
1042 static void cleanup_cap_releases(struct ceph_mds_client *mdsc,
1043                                  struct ceph_mds_session *session)
1044         __releases(session->s_cap_lock)
1045 {
1046         LIST_HEAD(tmp_list);
1047         list_splice_init(&session->s_cap_releases, &tmp_list);
1048         session->s_num_cap_releases = 0;
1049         spin_unlock(&session->s_cap_lock);
1050
1051         dout("cleanup_cap_releases mds%d\n", session->s_mds);
1052         while (!list_empty(&tmp_list)) {
1053                 struct ceph_cap *cap;
1054                 /* zero out the in-progress message */
1055                 cap = list_first_entry(&tmp_list,
1056                                         struct ceph_cap, session_caps);
1057                 list_del(&cap->session_caps);
1058                 ceph_put_cap(mdsc, cap);
1059         }
1060 }
1061
1062 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1063                                      struct ceph_mds_session *session)
1064 {
1065         struct ceph_mds_request *req;
1066         struct rb_node *p;
1067
1068         dout("cleanup_session_requests mds%d\n", session->s_mds);
1069         mutex_lock(&mdsc->mutex);
1070         while (!list_empty(&session->s_unsafe)) {
1071                 req = list_first_entry(&session->s_unsafe,
1072                                        struct ceph_mds_request, r_unsafe_item);
1073                 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1074                                     req->r_tid);
1075                 __unregister_request(mdsc, req);
1076         }
1077         /* zero r_attempts, so kick_requests() will re-send requests */
1078         p = rb_first(&mdsc->request_tree);
1079         while (p) {
1080                 req = rb_entry(p, struct ceph_mds_request, r_node);
1081                 p = rb_next(p);
1082                 if (req->r_session &&
1083                     req->r_session->s_mds == session->s_mds)
1084                         req->r_attempts = 0;
1085         }
1086         mutex_unlock(&mdsc->mutex);
1087 }
1088
1089 /*
1090  * Helper to safely iterate over all caps associated with a session, with
1091  * special care taken to handle a racing __ceph_remove_cap().
1092  *
1093  * Caller must hold session s_mutex.
1094  */
1095 static int iterate_session_caps(struct ceph_mds_session *session,
1096                                  int (*cb)(struct inode *, struct ceph_cap *,
1097                                             void *), void *arg)
1098 {
1099         struct list_head *p;
1100         struct ceph_cap *cap;
1101         struct inode *inode, *last_inode = NULL;
1102         struct ceph_cap *old_cap = NULL;
1103         int ret;
1104
1105         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1106         spin_lock(&session->s_cap_lock);
1107         p = session->s_caps.next;
1108         while (p != &session->s_caps) {
1109                 cap = list_entry(p, struct ceph_cap, session_caps);
1110                 inode = igrab(&cap->ci->vfs_inode);
1111                 if (!inode) {
1112                         p = p->next;
1113                         continue;
1114                 }
1115                 session->s_cap_iterator = cap;
1116                 spin_unlock(&session->s_cap_lock);
1117
1118                 if (last_inode) {
1119                         iput(last_inode);
1120                         last_inode = NULL;
1121                 }
1122                 if (old_cap) {
1123                         ceph_put_cap(session->s_mdsc, old_cap);
1124                         old_cap = NULL;
1125                 }
1126
1127                 ret = cb(inode, cap, arg);
1128                 last_inode = inode;
1129
1130                 spin_lock(&session->s_cap_lock);
1131                 p = p->next;
1132                 if (!cap->ci) {
1133                         dout("iterate_session_caps  finishing cap %p removal\n",
1134                              cap);
1135                         BUG_ON(cap->session != session);
1136                         cap->session = NULL;
1137                         list_del_init(&cap->session_caps);
1138                         session->s_nr_caps--;
1139                         if (cap->queue_release) {
1140                                 list_add_tail(&cap->session_caps,
1141                                               &session->s_cap_releases);
1142                                 session->s_num_cap_releases++;
1143                         } else {
1144                                 old_cap = cap;  /* put_cap it w/o locks held */
1145                         }
1146                 }
1147                 if (ret < 0)
1148                         goto out;
1149         }
1150         ret = 0;
1151 out:
1152         session->s_cap_iterator = NULL;
1153         spin_unlock(&session->s_cap_lock);
1154
1155         iput(last_inode);
1156         if (old_cap)
1157                 ceph_put_cap(session->s_mdsc, old_cap);
1158
1159         return ret;
1160 }
1161
1162 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1163                                   void *arg)
1164 {
1165         struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1166         struct ceph_inode_info *ci = ceph_inode(inode);
1167         LIST_HEAD(to_remove);
1168         bool drop = false;
1169         bool invalidate = false;
1170
1171         dout("removing cap %p, ci is %p, inode is %p\n",
1172              cap, ci, &ci->vfs_inode);
1173         spin_lock(&ci->i_ceph_lock);
1174         __ceph_remove_cap(cap, false);
1175         if (!ci->i_auth_cap) {
1176                 struct ceph_cap_flush *cf;
1177                 struct ceph_mds_client *mdsc = fsc->mdsc;
1178
1179                 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1180
1181                 if (ci->i_wrbuffer_ref > 0 &&
1182                     READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
1183                         invalidate = true;
1184
1185                 while (!list_empty(&ci->i_cap_flush_list)) {
1186                         cf = list_first_entry(&ci->i_cap_flush_list,
1187                                               struct ceph_cap_flush, i_list);
1188                         list_move(&cf->i_list, &to_remove);
1189                 }
1190
1191                 spin_lock(&mdsc->cap_dirty_lock);
1192
1193                 list_for_each_entry(cf, &to_remove, i_list)
1194                         list_del(&cf->g_list);
1195
1196                 if (!list_empty(&ci->i_dirty_item)) {
1197                         pr_warn_ratelimited(
1198                                 " dropping dirty %s state for %p %lld\n",
1199                                 ceph_cap_string(ci->i_dirty_caps),
1200                                 inode, ceph_ino(inode));
1201                         ci->i_dirty_caps = 0;
1202                         list_del_init(&ci->i_dirty_item);
1203                         drop = true;
1204                 }
1205                 if (!list_empty(&ci->i_flushing_item)) {
1206                         pr_warn_ratelimited(
1207                                 " dropping dirty+flushing %s state for %p %lld\n",
1208                                 ceph_cap_string(ci->i_flushing_caps),
1209                                 inode, ceph_ino(inode));
1210                         ci->i_flushing_caps = 0;
1211                         list_del_init(&ci->i_flushing_item);
1212                         mdsc->num_cap_flushing--;
1213                         drop = true;
1214                 }
1215                 spin_unlock(&mdsc->cap_dirty_lock);
1216
1217                 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1218                         list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1219                         ci->i_prealloc_cap_flush = NULL;
1220                 }
1221         }
1222         spin_unlock(&ci->i_ceph_lock);
1223         while (!list_empty(&to_remove)) {
1224                 struct ceph_cap_flush *cf;
1225                 cf = list_first_entry(&to_remove,
1226                                       struct ceph_cap_flush, i_list);
1227                 list_del(&cf->i_list);
1228                 ceph_free_cap_flush(cf);
1229         }
1230
1231         wake_up_all(&ci->i_cap_wq);
1232         if (invalidate)
1233                 ceph_queue_invalidate(inode);
1234         if (drop)
1235                 iput(inode);
1236         return 0;
1237 }
1238
1239 /*
1240  * caller must hold session s_mutex
1241  */
1242 static void remove_session_caps(struct ceph_mds_session *session)
1243 {
1244         struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1245         struct super_block *sb = fsc->sb;
1246         dout("remove_session_caps on %p\n", session);
1247         iterate_session_caps(session, remove_session_caps_cb, fsc);
1248
1249         wake_up_all(&fsc->mdsc->cap_flushing_wq);
1250
1251         spin_lock(&session->s_cap_lock);
1252         if (session->s_nr_caps > 0) {
1253                 struct inode *inode;
1254                 struct ceph_cap *cap, *prev = NULL;
1255                 struct ceph_vino vino;
1256                 /*
1257                  * iterate_session_caps() skips inodes that are being
1258                  * deleted, we need to wait until deletions are complete.
1259                  * __wait_on_freeing_inode() is designed for the job,
1260                  * but it is not exported, so use lookup inode function
1261                  * to access it.
1262                  */
1263                 while (!list_empty(&session->s_caps)) {
1264                         cap = list_entry(session->s_caps.next,
1265                                          struct ceph_cap, session_caps);
1266                         if (cap == prev)
1267                                 break;
1268                         prev = cap;
1269                         vino = cap->ci->i_vino;
1270                         spin_unlock(&session->s_cap_lock);
1271
1272                         inode = ceph_find_inode(sb, vino);
1273                         iput(inode);
1274
1275                         spin_lock(&session->s_cap_lock);
1276                 }
1277         }
1278
1279         // drop cap expires and unlock s_cap_lock
1280         cleanup_cap_releases(session->s_mdsc, session);
1281
1282         BUG_ON(session->s_nr_caps > 0);
1283         BUG_ON(!list_empty(&session->s_cap_flushing));
1284 }
1285
1286 /*
1287  * wake up any threads waiting on this session's caps.  if the cap is
1288  * old (didn't get renewed on the client reconnect), remove it now.
1289  *
1290  * caller must hold s_mutex.
1291  */
1292 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1293                               void *arg)
1294 {
1295         struct ceph_inode_info *ci = ceph_inode(inode);
1296
1297         if (arg) {
1298                 spin_lock(&ci->i_ceph_lock);
1299                 ci->i_wanted_max_size = 0;
1300                 ci->i_requested_max_size = 0;
1301                 spin_unlock(&ci->i_ceph_lock);
1302         }
1303         wake_up_all(&ci->i_cap_wq);
1304         return 0;
1305 }
1306
1307 static void wake_up_session_caps(struct ceph_mds_session *session,
1308                                  int reconnect)
1309 {
1310         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1311         iterate_session_caps(session, wake_up_session_cb,
1312                              (void *)(unsigned long)reconnect);
1313 }
1314
1315 /*
1316  * Send periodic message to MDS renewing all currently held caps.  The
1317  * ack will reset the expiration for all caps from this session.
1318  *
1319  * caller holds s_mutex
1320  */
1321 static int send_renew_caps(struct ceph_mds_client *mdsc,
1322                            struct ceph_mds_session *session)
1323 {
1324         struct ceph_msg *msg;
1325         int state;
1326
1327         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1328             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1329                 pr_info("mds%d caps stale\n", session->s_mds);
1330         session->s_renew_requested = jiffies;
1331
1332         /* do not try to renew caps until a recovering mds has reconnected
1333          * with its clients. */
1334         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1335         if (state < CEPH_MDS_STATE_RECONNECT) {
1336                 dout("send_renew_caps ignoring mds%d (%s)\n",
1337                      session->s_mds, ceph_mds_state_name(state));
1338                 return 0;
1339         }
1340
1341         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1342                 ceph_mds_state_name(state));
1343         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1344                                  ++session->s_renew_seq);
1345         if (!msg)
1346                 return -ENOMEM;
1347         ceph_con_send(&session->s_con, msg);
1348         return 0;
1349 }
1350
1351 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1352                              struct ceph_mds_session *session, u64 seq)
1353 {
1354         struct ceph_msg *msg;
1355
1356         dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1357              session->s_mds, ceph_session_state_name(session->s_state), seq);
1358         msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1359         if (!msg)
1360                 return -ENOMEM;
1361         ceph_con_send(&session->s_con, msg);
1362         return 0;
1363 }
1364
1365
1366 /*
1367  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1368  *
1369  * Called under session->s_mutex
1370  */
1371 static void renewed_caps(struct ceph_mds_client *mdsc,
1372                          struct ceph_mds_session *session, int is_renew)
1373 {
1374         int was_stale;
1375         int wake = 0;
1376
1377         spin_lock(&session->s_cap_lock);
1378         was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1379
1380         session->s_cap_ttl = session->s_renew_requested +
1381                 mdsc->mdsmap->m_session_timeout*HZ;
1382
1383         if (was_stale) {
1384                 if (time_before(jiffies, session->s_cap_ttl)) {
1385                         pr_info("mds%d caps renewed\n", session->s_mds);
1386                         wake = 1;
1387                 } else {
1388                         pr_info("mds%d caps still stale\n", session->s_mds);
1389                 }
1390         }
1391         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1392              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1393              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1394         spin_unlock(&session->s_cap_lock);
1395
1396         if (wake)
1397                 wake_up_session_caps(session, 0);
1398 }
1399
1400 /*
1401  * send a session close request
1402  */
1403 static int request_close_session(struct ceph_mds_client *mdsc,
1404                                  struct ceph_mds_session *session)
1405 {
1406         struct ceph_msg *msg;
1407
1408         dout("request_close_session mds%d state %s seq %lld\n",
1409              session->s_mds, ceph_session_state_name(session->s_state),
1410              session->s_seq);
1411         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1412         if (!msg)
1413                 return -ENOMEM;
1414         ceph_con_send(&session->s_con, msg);
1415         return 1;
1416 }
1417
1418 /*
1419  * Called with s_mutex held.
1420  */
1421 static int __close_session(struct ceph_mds_client *mdsc,
1422                          struct ceph_mds_session *session)
1423 {
1424         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1425                 return 0;
1426         session->s_state = CEPH_MDS_SESSION_CLOSING;
1427         return request_close_session(mdsc, session);
1428 }
1429
1430 /*
1431  * Trim old(er) caps.
1432  *
1433  * Because we can't cache an inode without one or more caps, we do
1434  * this indirectly: if a cap is unused, we prune its aliases, at which
1435  * point the inode will hopefully get dropped to.
1436  *
1437  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1438  * memory pressure from the MDS, though, so it needn't be perfect.
1439  */
1440 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1441 {
1442         struct ceph_mds_session *session = arg;
1443         struct ceph_inode_info *ci = ceph_inode(inode);
1444         int used, wanted, oissued, mine;
1445
1446         if (session->s_trim_caps <= 0)
1447                 return -1;
1448
1449         spin_lock(&ci->i_ceph_lock);
1450         mine = cap->issued | cap->implemented;
1451         used = __ceph_caps_used(ci);
1452         wanted = __ceph_caps_file_wanted(ci);
1453         oissued = __ceph_caps_issued_other(ci, cap);
1454
1455         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1456              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1457              ceph_cap_string(used), ceph_cap_string(wanted));
1458         if (cap == ci->i_auth_cap) {
1459                 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1460                     !list_empty(&ci->i_cap_snaps))
1461                         goto out;
1462                 if ((used | wanted) & CEPH_CAP_ANY_WR)
1463                         goto out;
1464         }
1465         /* The inode has cached pages, but it's no longer used.
1466          * we can safely drop it */
1467         if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1468             !(oissued & CEPH_CAP_FILE_CACHE)) {
1469           used = 0;
1470           oissued = 0;
1471         }
1472         if ((used | wanted) & ~oissued & mine)
1473                 goto out;   /* we need these caps */
1474
1475         session->s_trim_caps--;
1476         if (oissued) {
1477                 /* we aren't the only cap.. just remove us */
1478                 __ceph_remove_cap(cap, true);
1479         } else {
1480                 /* try dropping referring dentries */
1481                 spin_unlock(&ci->i_ceph_lock);
1482                 d_prune_aliases(inode);
1483                 dout("trim_caps_cb %p cap %p  pruned, count now %d\n",
1484                      inode, cap, atomic_read(&inode->i_count));
1485                 return 0;
1486         }
1487
1488 out:
1489         spin_unlock(&ci->i_ceph_lock);
1490         return 0;
1491 }
1492
1493 /*
1494  * Trim session cap count down to some max number.
1495  */
1496 static int trim_caps(struct ceph_mds_client *mdsc,
1497                      struct ceph_mds_session *session,
1498                      int max_caps)
1499 {
1500         int trim_caps = session->s_nr_caps - max_caps;
1501
1502         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1503              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1504         if (trim_caps > 0) {
1505                 session->s_trim_caps = trim_caps;
1506                 iterate_session_caps(session, trim_caps_cb, session);
1507                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1508                      session->s_mds, session->s_nr_caps, max_caps,
1509                         trim_caps - session->s_trim_caps);
1510                 session->s_trim_caps = 0;
1511         }
1512
1513         ceph_send_cap_releases(mdsc, session);
1514         return 0;
1515 }
1516
1517 static int check_caps_flush(struct ceph_mds_client *mdsc,
1518                             u64 want_flush_tid)
1519 {
1520         int ret = 1;
1521
1522         spin_lock(&mdsc->cap_dirty_lock);
1523         if (!list_empty(&mdsc->cap_flush_list)) {
1524                 struct ceph_cap_flush *cf =
1525                         list_first_entry(&mdsc->cap_flush_list,
1526                                          struct ceph_cap_flush, g_list);
1527                 if (cf->tid <= want_flush_tid) {
1528                         dout("check_caps_flush still flushing tid "
1529                              "%llu <= %llu\n", cf->tid, want_flush_tid);
1530                         ret = 0;
1531                 }
1532         }
1533         spin_unlock(&mdsc->cap_dirty_lock);
1534         return ret;
1535 }
1536
1537 /*
1538  * flush all dirty inode data to disk.
1539  *
1540  * returns true if we've flushed through want_flush_tid
1541  */
1542 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1543                             u64 want_flush_tid)
1544 {
1545         dout("check_caps_flush want %llu\n", want_flush_tid);
1546
1547         wait_event(mdsc->cap_flushing_wq,
1548                    check_caps_flush(mdsc, want_flush_tid));
1549
1550         dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1551 }
1552
1553 /*
1554  * called under s_mutex
1555  */
1556 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1557                             struct ceph_mds_session *session)
1558 {
1559         struct ceph_msg *msg = NULL;
1560         struct ceph_mds_cap_release *head;
1561         struct ceph_mds_cap_item *item;
1562         struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
1563         struct ceph_cap *cap;
1564         LIST_HEAD(tmp_list);
1565         int num_cap_releases;
1566         __le32  barrier, *cap_barrier;
1567
1568         down_read(&osdc->lock);
1569         barrier = cpu_to_le32(osdc->epoch_barrier);
1570         up_read(&osdc->lock);
1571
1572         spin_lock(&session->s_cap_lock);
1573 again:
1574         list_splice_init(&session->s_cap_releases, &tmp_list);
1575         num_cap_releases = session->s_num_cap_releases;
1576         session->s_num_cap_releases = 0;
1577         spin_unlock(&session->s_cap_lock);
1578
1579         while (!list_empty(&tmp_list)) {
1580                 if (!msg) {
1581                         msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1582                                         PAGE_SIZE, GFP_NOFS, false);
1583                         if (!msg)
1584                                 goto out_err;
1585                         head = msg->front.iov_base;
1586                         head->num = cpu_to_le32(0);
1587                         msg->front.iov_len = sizeof(*head);
1588
1589                         msg->hdr.version = cpu_to_le16(2);
1590                         msg->hdr.compat_version = cpu_to_le16(1);
1591                 }
1592
1593                 cap = list_first_entry(&tmp_list, struct ceph_cap,
1594                                         session_caps);
1595                 list_del(&cap->session_caps);
1596                 num_cap_releases--;
1597
1598                 head = msg->front.iov_base;
1599                 le32_add_cpu(&head->num, 1);
1600                 item = msg->front.iov_base + msg->front.iov_len;
1601                 item->ino = cpu_to_le64(cap->cap_ino);
1602                 item->cap_id = cpu_to_le64(cap->cap_id);
1603                 item->migrate_seq = cpu_to_le32(cap->mseq);
1604                 item->seq = cpu_to_le32(cap->issue_seq);
1605                 msg->front.iov_len += sizeof(*item);
1606
1607                 ceph_put_cap(mdsc, cap);
1608
1609                 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1610                         // Append cap_barrier field
1611                         cap_barrier = msg->front.iov_base + msg->front.iov_len;
1612                         *cap_barrier = barrier;
1613                         msg->front.iov_len += sizeof(*cap_barrier);
1614
1615                         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1616                         dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1617                         ceph_con_send(&session->s_con, msg);
1618                         msg = NULL;
1619                 }
1620         }
1621
1622         BUG_ON(num_cap_releases != 0);
1623
1624         spin_lock(&session->s_cap_lock);
1625         if (!list_empty(&session->s_cap_releases))
1626                 goto again;
1627         spin_unlock(&session->s_cap_lock);
1628
1629         if (msg) {
1630                 // Append cap_barrier field
1631                 cap_barrier = msg->front.iov_base + msg->front.iov_len;
1632                 *cap_barrier = barrier;
1633                 msg->front.iov_len += sizeof(*cap_barrier);
1634
1635                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1636                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1637                 ceph_con_send(&session->s_con, msg);
1638         }
1639         return;
1640 out_err:
1641         pr_err("send_cap_releases mds%d, failed to allocate message\n",
1642                 session->s_mds);
1643         spin_lock(&session->s_cap_lock);
1644         list_splice(&tmp_list, &session->s_cap_releases);
1645         session->s_num_cap_releases += num_cap_releases;
1646         spin_unlock(&session->s_cap_lock);
1647 }
1648
1649 /*
1650  * requests
1651  */
1652
1653 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1654                                     struct inode *dir)
1655 {
1656         struct ceph_inode_info *ci = ceph_inode(dir);
1657         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1658         struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1659         size_t size = sizeof(struct ceph_mds_reply_dir_entry);
1660         int order, num_entries;
1661
1662         spin_lock(&ci->i_ceph_lock);
1663         num_entries = ci->i_files + ci->i_subdirs;
1664         spin_unlock(&ci->i_ceph_lock);
1665         num_entries = max(num_entries, 1);
1666         num_entries = min(num_entries, opt->max_readdir);
1667
1668         order = get_order(size * num_entries);
1669         while (order >= 0) {
1670                 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
1671                                                              __GFP_NOWARN,
1672                                                              order);
1673                 if (rinfo->dir_entries)
1674                         break;
1675                 order--;
1676         }
1677         if (!rinfo->dir_entries)
1678                 return -ENOMEM;
1679
1680         num_entries = (PAGE_SIZE << order) / size;
1681         num_entries = min(num_entries, opt->max_readdir);
1682
1683         rinfo->dir_buf_size = PAGE_SIZE << order;
1684         req->r_num_caps = num_entries + 1;
1685         req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1686         req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1687         return 0;
1688 }
1689
1690 /*
1691  * Create an mds request.
1692  */
1693 struct ceph_mds_request *
1694 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1695 {
1696         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1697
1698         if (!req)
1699                 return ERR_PTR(-ENOMEM);
1700
1701         mutex_init(&req->r_fill_mutex);
1702         req->r_mdsc = mdsc;
1703         req->r_started = jiffies;
1704         req->r_resend_mds = -1;
1705         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1706         INIT_LIST_HEAD(&req->r_unsafe_target_item);
1707         req->r_fmode = -1;
1708         kref_init(&req->r_kref);
1709         RB_CLEAR_NODE(&req->r_node);
1710         INIT_LIST_HEAD(&req->r_wait);
1711         init_completion(&req->r_completion);
1712         init_completion(&req->r_safe_completion);
1713         INIT_LIST_HEAD(&req->r_unsafe_item);
1714
1715         req->r_stamp = timespec_trunc(current_kernel_time(), mdsc->fsc->sb->s_time_gran);
1716
1717         req->r_op = op;
1718         req->r_direct_mode = mode;
1719         return req;
1720 }
1721
1722 /*
1723  * return oldest (lowest) request, tid in request tree, 0 if none.
1724  *
1725  * called under mdsc->mutex.
1726  */
1727 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1728 {
1729         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1730                 return NULL;
1731         return rb_entry(rb_first(&mdsc->request_tree),
1732                         struct ceph_mds_request, r_node);
1733 }
1734
1735 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1736 {
1737         return mdsc->oldest_tid;
1738 }
1739
1740 /*
1741  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1742  * on build_path_from_dentry in fs/cifs/dir.c.
1743  *
1744  * If @stop_on_nosnap, generate path relative to the first non-snapped
1745  * inode.
1746  *
1747  * Encode hidden .snap dirs as a double /, i.e.
1748  *   foo/.snap/bar -> foo//bar
1749  */
1750 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1751                            int stop_on_nosnap)
1752 {
1753         struct dentry *temp;
1754         char *path;
1755         int len, pos;
1756         unsigned seq;
1757
1758         if (!dentry)
1759                 return ERR_PTR(-EINVAL);
1760
1761 retry:
1762         len = 0;
1763         seq = read_seqbegin(&rename_lock);
1764         rcu_read_lock();
1765         for (temp = dentry; !IS_ROOT(temp);) {
1766                 struct inode *inode = d_inode(temp);
1767                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1768                         len++;  /* slash only */
1769                 else if (stop_on_nosnap && inode &&
1770                          ceph_snap(inode) == CEPH_NOSNAP)
1771                         break;
1772                 else
1773                         len += 1 + temp->d_name.len;
1774                 temp = temp->d_parent;
1775         }
1776         rcu_read_unlock();
1777         if (len)
1778                 len--;  /* no leading '/' */
1779
1780         path = kmalloc(len+1, GFP_NOFS);
1781         if (!path)
1782                 return ERR_PTR(-ENOMEM);
1783         pos = len;
1784         path[pos] = 0;  /* trailing null */
1785         rcu_read_lock();
1786         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1787                 struct inode *inode;
1788
1789                 spin_lock(&temp->d_lock);
1790                 inode = d_inode(temp);
1791                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1792                         dout("build_path path+%d: %p SNAPDIR\n",
1793                              pos, temp);
1794                 } else if (stop_on_nosnap && inode &&
1795                            ceph_snap(inode) == CEPH_NOSNAP) {
1796                         spin_unlock(&temp->d_lock);
1797                         break;
1798                 } else {
1799                         pos -= temp->d_name.len;
1800                         if (pos < 0) {
1801                                 spin_unlock(&temp->d_lock);
1802                                 break;
1803                         }
1804                         strncpy(path + pos, temp->d_name.name,
1805                                 temp->d_name.len);
1806                 }
1807                 spin_unlock(&temp->d_lock);
1808                 if (pos)
1809                         path[--pos] = '/';
1810                 temp = temp->d_parent;
1811         }
1812         rcu_read_unlock();
1813         if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1814                 pr_err("build_path did not end path lookup where "
1815                        "expected, namelen is %d, pos is %d\n", len, pos);
1816                 /* presumably this is only possible if racing with a
1817                    rename of one of the parent directories (we can not
1818                    lock the dentries above us to prevent this, but
1819                    retrying should be harmless) */
1820                 kfree(path);
1821                 goto retry;
1822         }
1823
1824         *base = ceph_ino(d_inode(temp));
1825         *plen = len;
1826         dout("build_path on %p %d built %llx '%.*s'\n",
1827              dentry, d_count(dentry), *base, len, path);
1828         return path;
1829 }
1830
1831 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
1832                              const char **ppath, int *ppathlen, u64 *pino,
1833                              int *pfreepath)
1834 {
1835         char *path;
1836
1837         rcu_read_lock();
1838         if (!dir)
1839                 dir = d_inode_rcu(dentry->d_parent);
1840         if (dir && ceph_snap(dir) == CEPH_NOSNAP) {
1841                 *pino = ceph_ino(dir);
1842                 rcu_read_unlock();
1843                 *ppath = dentry->d_name.name;
1844                 *ppathlen = dentry->d_name.len;
1845                 return 0;
1846         }
1847         rcu_read_unlock();
1848         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1849         if (IS_ERR(path))
1850                 return PTR_ERR(path);
1851         *ppath = path;
1852         *pfreepath = 1;
1853         return 0;
1854 }
1855
1856 static int build_inode_path(struct inode *inode,
1857                             const char **ppath, int *ppathlen, u64 *pino,
1858                             int *pfreepath)
1859 {
1860         struct dentry *dentry;
1861         char *path;
1862
1863         if (ceph_snap(inode) == CEPH_NOSNAP) {
1864                 *pino = ceph_ino(inode);
1865                 *ppathlen = 0;
1866                 return 0;
1867         }
1868         dentry = d_find_alias(inode);
1869         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1870         dput(dentry);
1871         if (IS_ERR(path))
1872                 return PTR_ERR(path);
1873         *ppath = path;
1874         *pfreepath = 1;
1875         return 0;
1876 }
1877
1878 /*
1879  * request arguments may be specified via an inode *, a dentry *, or
1880  * an explicit ino+path.
1881  */
1882 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1883                                   struct inode *rdiri, const char *rpath,
1884                                   u64 rino, const char **ppath, int *pathlen,
1885                                   u64 *ino, int *freepath)
1886 {
1887         int r = 0;
1888
1889         if (rinode) {
1890                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1891                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1892                      ceph_snap(rinode));
1893         } else if (rdentry) {
1894                 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
1895                                         freepath);
1896                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1897                      *ppath);
1898         } else if (rpath || rino) {
1899                 *ino = rino;
1900                 *ppath = rpath;
1901                 *pathlen = rpath ? strlen(rpath) : 0;
1902                 dout(" path %.*s\n", *pathlen, rpath);
1903         }
1904
1905         return r;
1906 }
1907
1908 /*
1909  * called under mdsc->mutex
1910  */
1911 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1912                                                struct ceph_mds_request *req,
1913                                                int mds, bool drop_cap_releases)
1914 {
1915         struct ceph_msg *msg;
1916         struct ceph_mds_request_head *head;
1917         const char *path1 = NULL;
1918         const char *path2 = NULL;
1919         u64 ino1 = 0, ino2 = 0;
1920         int pathlen1 = 0, pathlen2 = 0;
1921         int freepath1 = 0, freepath2 = 0;
1922         int len;
1923         u16 releases;
1924         void *p, *end;
1925         int ret;
1926
1927         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1928                               req->r_parent, req->r_path1, req->r_ino1.ino,
1929                               &path1, &pathlen1, &ino1, &freepath1);
1930         if (ret < 0) {
1931                 msg = ERR_PTR(ret);
1932                 goto out;
1933         }
1934
1935         ret = set_request_path_attr(NULL, req->r_old_dentry,
1936                               req->r_old_dentry_dir,
1937                               req->r_path2, req->r_ino2.ino,
1938                               &path2, &pathlen2, &ino2, &freepath2);
1939         if (ret < 0) {
1940                 msg = ERR_PTR(ret);
1941                 goto out_free1;
1942         }
1943
1944         len = sizeof(*head) +
1945                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
1946                 sizeof(struct ceph_timespec);
1947
1948         /* calculate (max) length for cap releases */
1949         len += sizeof(struct ceph_mds_request_release) *
1950                 (!!req->r_inode_drop + !!req->r_dentry_drop +
1951                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1952         if (req->r_dentry_drop)
1953                 len += req->r_dentry->d_name.len;
1954         if (req->r_old_dentry_drop)
1955                 len += req->r_old_dentry->d_name.len;
1956
1957         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1958         if (!msg) {
1959                 msg = ERR_PTR(-ENOMEM);
1960                 goto out_free2;
1961         }
1962
1963         msg->hdr.version = cpu_to_le16(2);
1964         msg->hdr.tid = cpu_to_le64(req->r_tid);
1965
1966         head = msg->front.iov_base;
1967         p = msg->front.iov_base + sizeof(*head);
1968         end = msg->front.iov_base + msg->front.iov_len;
1969
1970         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1971         head->op = cpu_to_le32(req->r_op);
1972         head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1973         head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1974         head->args = req->r_args;
1975
1976         ceph_encode_filepath(&p, end, ino1, path1);
1977         ceph_encode_filepath(&p, end, ino2, path2);
1978
1979         /* make note of release offset, in case we need to replay */
1980         req->r_request_release_offset = p - msg->front.iov_base;
1981
1982         /* cap releases */
1983         releases = 0;
1984         if (req->r_inode_drop)
1985                 releases += ceph_encode_inode_release(&p,
1986                       req->r_inode ? req->r_inode : d_inode(req->r_dentry),
1987                       mds, req->r_inode_drop, req->r_inode_unless, 0);
1988         if (req->r_dentry_drop)
1989                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1990                                 req->r_parent, mds, req->r_dentry_drop,
1991                                 req->r_dentry_unless);
1992         if (req->r_old_dentry_drop)
1993                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1994                                 req->r_old_dentry_dir, mds,
1995                                 req->r_old_dentry_drop,
1996                                 req->r_old_dentry_unless);
1997         if (req->r_old_inode_drop)
1998                 releases += ceph_encode_inode_release(&p,
1999                       d_inode(req->r_old_dentry),
2000                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2001
2002         if (drop_cap_releases) {
2003                 releases = 0;
2004                 p = msg->front.iov_base + req->r_request_release_offset;
2005         }
2006
2007         head->num_releases = cpu_to_le16(releases);
2008
2009         /* time stamp */
2010         {
2011                 struct ceph_timespec ts;
2012                 ceph_encode_timespec(&ts, &req->r_stamp);
2013                 ceph_encode_copy(&p, &ts, sizeof(ts));
2014         }
2015
2016         BUG_ON(p > end);
2017         msg->front.iov_len = p - msg->front.iov_base;
2018         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2019
2020         if (req->r_pagelist) {
2021                 struct ceph_pagelist *pagelist = req->r_pagelist;
2022                 refcount_inc(&pagelist->refcnt);
2023                 ceph_msg_data_add_pagelist(msg, pagelist);
2024                 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2025         } else {
2026                 msg->hdr.data_len = 0;
2027         }
2028
2029         msg->hdr.data_off = cpu_to_le16(0);
2030
2031 out_free2:
2032         if (freepath2)
2033                 kfree((char *)path2);
2034 out_free1:
2035         if (freepath1)
2036                 kfree((char *)path1);
2037 out:
2038         return msg;
2039 }
2040
2041 /*
2042  * called under mdsc->mutex if error, under no mutex if
2043  * success.
2044  */
2045 static void complete_request(struct ceph_mds_client *mdsc,
2046                              struct ceph_mds_request *req)
2047 {
2048         if (req->r_callback)
2049                 req->r_callback(mdsc, req);
2050         else
2051                 complete_all(&req->r_completion);
2052 }
2053
2054 /*
2055  * called under mdsc->mutex
2056  */
2057 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2058                                   struct ceph_mds_request *req,
2059                                   int mds, bool drop_cap_releases)
2060 {
2061         struct ceph_mds_request_head *rhead;
2062         struct ceph_msg *msg;
2063         int flags = 0;
2064
2065         req->r_attempts++;
2066         if (req->r_inode) {
2067                 struct ceph_cap *cap =
2068                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2069
2070                 if (cap)
2071                         req->r_sent_on_mseq = cap->mseq;
2072                 else
2073                         req->r_sent_on_mseq = -1;
2074         }
2075         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2076              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2077
2078         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2079                 void *p;
2080                 /*
2081                  * Replay.  Do not regenerate message (and rebuild
2082                  * paths, etc.); just use the original message.
2083                  * Rebuilding paths will break for renames because
2084                  * d_move mangles the src name.
2085                  */
2086                 msg = req->r_request;
2087                 rhead = msg->front.iov_base;
2088
2089                 flags = le32_to_cpu(rhead->flags);
2090                 flags |= CEPH_MDS_FLAG_REPLAY;
2091                 rhead->flags = cpu_to_le32(flags);
2092
2093                 if (req->r_target_inode)
2094                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2095
2096                 rhead->num_retry = req->r_attempts - 1;
2097
2098                 /* remove cap/dentry releases from message */
2099                 rhead->num_releases = 0;
2100
2101                 /* time stamp */
2102                 p = msg->front.iov_base + req->r_request_release_offset;
2103                 {
2104                         struct ceph_timespec ts;
2105                         ceph_encode_timespec(&ts, &req->r_stamp);
2106                         ceph_encode_copy(&p, &ts, sizeof(ts));
2107                 }
2108
2109                 msg->front.iov_len = p - msg->front.iov_base;
2110                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2111                 return 0;
2112         }
2113
2114         if (req->r_request) {
2115                 ceph_msg_put(req->r_request);
2116                 req->r_request = NULL;
2117         }
2118         msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2119         if (IS_ERR(msg)) {
2120                 req->r_err = PTR_ERR(msg);
2121                 return PTR_ERR(msg);
2122         }
2123         req->r_request = msg;
2124
2125         rhead = msg->front.iov_base;
2126         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2127         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2128                 flags |= CEPH_MDS_FLAG_REPLAY;
2129         if (req->r_parent)
2130                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2131         rhead->flags = cpu_to_le32(flags);
2132         rhead->num_fwd = req->r_num_fwd;
2133         rhead->num_retry = req->r_attempts - 1;
2134         rhead->ino = 0;
2135
2136         dout(" r_parent = %p\n", req->r_parent);
2137         return 0;
2138 }
2139
2140 /*
2141  * send request, or put it on the appropriate wait list.
2142  */
2143 static int __do_request(struct ceph_mds_client *mdsc,
2144                         struct ceph_mds_request *req)
2145 {
2146         struct ceph_mds_session *session = NULL;
2147         int mds = -1;
2148         int err = 0;
2149
2150         if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2151                 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2152                         __unregister_request(mdsc, req);
2153                 goto out;
2154         }
2155
2156         if (req->r_timeout &&
2157             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2158                 dout("do_request timed out\n");
2159                 err = -EIO;
2160                 goto finish;
2161         }
2162         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2163                 dout("do_request forced umount\n");
2164                 err = -EIO;
2165                 goto finish;
2166         }
2167         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2168                 if (mdsc->mdsmap_err) {
2169                         err = mdsc->mdsmap_err;
2170                         dout("do_request mdsmap err %d\n", err);
2171                         goto finish;
2172                 }
2173                 if (mdsc->mdsmap->m_epoch == 0) {
2174                         dout("do_request no mdsmap, waiting for map\n");
2175                         list_add(&req->r_wait, &mdsc->waiting_for_map);
2176                         goto finish;
2177                 }
2178                 if (!(mdsc->fsc->mount_options->flags &
2179                       CEPH_MOUNT_OPT_MOUNTWAIT) &&
2180                     !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2181                         err = -ENOENT;
2182                         pr_info("probably no mds server is up\n");
2183                         goto finish;
2184                 }
2185         }
2186
2187         put_request_session(req);
2188
2189         mds = __choose_mds(mdsc, req);
2190         if (mds < 0 ||
2191             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2192                 dout("do_request no mds or not active, waiting for map\n");
2193                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2194                 goto out;
2195         }
2196
2197         /* get, open session */
2198         session = __ceph_lookup_mds_session(mdsc, mds);
2199         if (!session) {
2200                 session = register_session(mdsc, mds);
2201                 if (IS_ERR(session)) {
2202                         err = PTR_ERR(session);
2203                         goto finish;
2204                 }
2205         }
2206         req->r_session = get_session(session);
2207
2208         dout("do_request mds%d session %p state %s\n", mds, session,
2209              ceph_session_state_name(session->s_state));
2210         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2211             session->s_state != CEPH_MDS_SESSION_HUNG) {
2212                 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2213                         err = -EACCES;
2214                         goto out_session;
2215                 }
2216                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2217                     session->s_state == CEPH_MDS_SESSION_CLOSING)
2218                         __open_session(mdsc, session);
2219                 list_add(&req->r_wait, &session->s_waiting);
2220                 goto out_session;
2221         }
2222
2223         /* send request */
2224         req->r_resend_mds = -1;   /* forget any previous mds hint */
2225
2226         if (req->r_request_started == 0)   /* note request start time */
2227                 req->r_request_started = jiffies;
2228
2229         err = __prepare_send_request(mdsc, req, mds, false);
2230         if (!err) {
2231                 ceph_msg_get(req->r_request);
2232                 ceph_con_send(&session->s_con, req->r_request);
2233         }
2234
2235 out_session:
2236         ceph_put_mds_session(session);
2237 finish:
2238         if (err) {
2239                 dout("__do_request early error %d\n", err);
2240                 req->r_err = err;
2241                 complete_request(mdsc, req);
2242                 __unregister_request(mdsc, req);
2243         }
2244 out:
2245         return err;
2246 }
2247
2248 /*
2249  * called under mdsc->mutex
2250  */
2251 static void __wake_requests(struct ceph_mds_client *mdsc,
2252                             struct list_head *head)
2253 {
2254         struct ceph_mds_request *req;
2255         LIST_HEAD(tmp_list);
2256
2257         list_splice_init(head, &tmp_list);
2258
2259         while (!list_empty(&tmp_list)) {
2260                 req = list_entry(tmp_list.next,
2261                                  struct ceph_mds_request, r_wait);
2262                 list_del_init(&req->r_wait);
2263                 dout(" wake request %p tid %llu\n", req, req->r_tid);
2264                 __do_request(mdsc, req);
2265         }
2266 }
2267
2268 /*
2269  * Wake up threads with requests pending for @mds, so that they can
2270  * resubmit their requests to a possibly different mds.
2271  */
2272 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2273 {
2274         struct ceph_mds_request *req;
2275         struct rb_node *p = rb_first(&mdsc->request_tree);
2276
2277         dout("kick_requests mds%d\n", mds);
2278         while (p) {
2279                 req = rb_entry(p, struct ceph_mds_request, r_node);
2280                 p = rb_next(p);
2281                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2282                         continue;
2283                 if (req->r_attempts > 0)
2284                         continue; /* only new requests */
2285                 if (req->r_session &&
2286                     req->r_session->s_mds == mds) {
2287                         dout(" kicking tid %llu\n", req->r_tid);
2288                         list_del_init(&req->r_wait);
2289                         __do_request(mdsc, req);
2290                 }
2291         }
2292 }
2293
2294 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2295                               struct ceph_mds_request *req)
2296 {
2297         dout("submit_request on %p\n", req);
2298         mutex_lock(&mdsc->mutex);
2299         __register_request(mdsc, req, NULL);
2300         __do_request(mdsc, req);
2301         mutex_unlock(&mdsc->mutex);
2302 }
2303
2304 /*
2305  * Synchrously perform an mds request.  Take care of all of the
2306  * session setup, forwarding, retry details.
2307  */
2308 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2309                          struct inode *dir,
2310                          struct ceph_mds_request *req)
2311 {
2312         int err;
2313
2314         dout("do_request on %p\n", req);
2315
2316         /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2317         if (req->r_inode)
2318                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2319         if (req->r_parent)
2320                 ceph_get_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
2321         if (req->r_old_dentry_dir)
2322                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2323                                   CEPH_CAP_PIN);
2324
2325         /* issue */
2326         mutex_lock(&mdsc->mutex);
2327         __register_request(mdsc, req, dir);
2328         __do_request(mdsc, req);
2329
2330         if (req->r_err) {
2331                 err = req->r_err;
2332                 goto out;
2333         }
2334
2335         /* wait */
2336         mutex_unlock(&mdsc->mutex);
2337         dout("do_request waiting\n");
2338         if (!req->r_timeout && req->r_wait_for_completion) {
2339                 err = req->r_wait_for_completion(mdsc, req);
2340         } else {
2341                 long timeleft = wait_for_completion_killable_timeout(
2342                                         &req->r_completion,
2343                                         ceph_timeout_jiffies(req->r_timeout));
2344                 if (timeleft > 0)
2345                         err = 0;
2346                 else if (!timeleft)
2347                         err = -EIO;  /* timed out */
2348                 else
2349                         err = timeleft;  /* killed */
2350         }
2351         dout("do_request waited, got %d\n", err);
2352         mutex_lock(&mdsc->mutex);
2353
2354         /* only abort if we didn't race with a real reply */
2355         if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2356                 err = le32_to_cpu(req->r_reply_info.head->result);
2357         } else if (err < 0) {
2358                 dout("aborted request %lld with %d\n", req->r_tid, err);
2359
2360                 /*
2361                  * ensure we aren't running concurrently with
2362                  * ceph_fill_trace or ceph_readdir_prepopulate, which
2363                  * rely on locks (dir mutex) held by our caller.
2364                  */
2365                 mutex_lock(&req->r_fill_mutex);
2366                 req->r_err = err;
2367                 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2368                 mutex_unlock(&req->r_fill_mutex);
2369
2370                 if (req->r_parent &&
2371                     (req->r_op & CEPH_MDS_OP_WRITE))
2372                         ceph_invalidate_dir_request(req);
2373         } else {
2374                 err = req->r_err;
2375         }
2376
2377 out:
2378         mutex_unlock(&mdsc->mutex);
2379         dout("do_request %p done, result %d\n", req, err);
2380         return err;
2381 }
2382
2383 /*
2384  * Invalidate dir's completeness, dentry lease state on an aborted MDS
2385  * namespace request.
2386  */
2387 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2388 {
2389         struct inode *inode = req->r_parent;
2390
2391         dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2392
2393         ceph_dir_clear_complete(inode);
2394         if (req->r_dentry)
2395                 ceph_invalidate_dentry_lease(req->r_dentry);
2396         if (req->r_old_dentry)
2397                 ceph_invalidate_dentry_lease(req->r_old_dentry);
2398 }
2399
2400 /*
2401  * Handle mds reply.
2402  *
2403  * We take the session mutex and parse and process the reply immediately.
2404  * This preserves the logical ordering of replies, capabilities, etc., sent
2405  * by the MDS as they are applied to our local cache.
2406  */
2407 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2408 {
2409         struct ceph_mds_client *mdsc = session->s_mdsc;
2410         struct ceph_mds_request *req;
2411         struct ceph_mds_reply_head *head = msg->front.iov_base;
2412         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2413         struct ceph_snap_realm *realm;
2414         u64 tid;
2415         int err, result;
2416         int mds = session->s_mds;
2417
2418         if (msg->front.iov_len < sizeof(*head)) {
2419                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2420                 ceph_msg_dump(msg);
2421                 return;
2422         }
2423
2424         /* get request, session */
2425         tid = le64_to_cpu(msg->hdr.tid);
2426         mutex_lock(&mdsc->mutex);
2427         req = lookup_get_request(mdsc, tid);
2428         if (!req) {
2429                 dout("handle_reply on unknown tid %llu\n", tid);
2430                 mutex_unlock(&mdsc->mutex);
2431                 return;
2432         }
2433         dout("handle_reply %p\n", req);
2434
2435         /* correct session? */
2436         if (req->r_session != session) {
2437                 pr_err("mdsc_handle_reply got %llu on session mds%d"
2438                        " not mds%d\n", tid, session->s_mds,
2439                        req->r_session ? req->r_session->s_mds : -1);
2440                 mutex_unlock(&mdsc->mutex);
2441                 goto out;
2442         }
2443
2444         /* dup? */
2445         if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
2446             (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
2447                 pr_warn("got a dup %s reply on %llu from mds%d\n",
2448                            head->safe ? "safe" : "unsafe", tid, mds);
2449                 mutex_unlock(&mdsc->mutex);
2450                 goto out;
2451         }
2452         if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
2453                 pr_warn("got unsafe after safe on %llu from mds%d\n",
2454                            tid, mds);
2455                 mutex_unlock(&mdsc->mutex);
2456                 goto out;
2457         }
2458
2459         result = le32_to_cpu(head->result);
2460
2461         /*
2462          * Handle an ESTALE
2463          * if we're not talking to the authority, send to them
2464          * if the authority has changed while we weren't looking,
2465          * send to new authority
2466          * Otherwise we just have to return an ESTALE
2467          */
2468         if (result == -ESTALE) {
2469                 dout("got ESTALE on request %llu", req->r_tid);
2470                 req->r_resend_mds = -1;
2471                 if (req->r_direct_mode != USE_AUTH_MDS) {
2472                         dout("not using auth, setting for that now");
2473                         req->r_direct_mode = USE_AUTH_MDS;
2474                         __do_request(mdsc, req);
2475                         mutex_unlock(&mdsc->mutex);
2476                         goto out;
2477                 } else  {
2478                         int mds = __choose_mds(mdsc, req);
2479                         if (mds >= 0 && mds != req->r_session->s_mds) {
2480                                 dout("but auth changed, so resending");
2481                                 __do_request(mdsc, req);
2482                                 mutex_unlock(&mdsc->mutex);
2483                                 goto out;
2484                         }
2485                 }
2486                 dout("have to return ESTALE on request %llu", req->r_tid);
2487         }
2488
2489
2490         if (head->safe) {
2491                 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
2492                 __unregister_request(mdsc, req);
2493
2494                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2495                         /*
2496                          * We already handled the unsafe response, now do the
2497                          * cleanup.  No need to examine the response; the MDS
2498                          * doesn't include any result info in the safe
2499                          * response.  And even if it did, there is nothing
2500                          * useful we could do with a revised return value.
2501                          */
2502                         dout("got safe reply %llu, mds%d\n", tid, mds);
2503
2504                         /* last unsafe request during umount? */
2505                         if (mdsc->stopping && !__get_oldest_req(mdsc))
2506                                 complete_all(&mdsc->safe_umount_waiters);
2507                         mutex_unlock(&mdsc->mutex);
2508                         goto out;
2509                 }
2510         } else {
2511                 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
2512                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2513                 if (req->r_unsafe_dir) {
2514                         struct ceph_inode_info *ci =
2515                                         ceph_inode(req->r_unsafe_dir);
2516                         spin_lock(&ci->i_unsafe_lock);
2517                         list_add_tail(&req->r_unsafe_dir_item,
2518                                       &ci->i_unsafe_dirops);
2519                         spin_unlock(&ci->i_unsafe_lock);
2520                 }
2521         }
2522
2523         dout("handle_reply tid %lld result %d\n", tid, result);
2524         rinfo = &req->r_reply_info;
2525         err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2526         mutex_unlock(&mdsc->mutex);
2527
2528         mutex_lock(&session->s_mutex);
2529         if (err < 0) {
2530                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2531                 ceph_msg_dump(msg);
2532                 goto out_err;
2533         }
2534
2535         /* snap trace */
2536         realm = NULL;
2537         if (rinfo->snapblob_len) {
2538                 down_write(&mdsc->snap_rwsem);
2539                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2540                                 rinfo->snapblob + rinfo->snapblob_len,
2541                                 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2542                                 &realm);
2543                 downgrade_write(&mdsc->snap_rwsem);
2544         } else {
2545                 down_read(&mdsc->snap_rwsem);
2546         }
2547
2548         /* insert trace into our cache */
2549         mutex_lock(&req->r_fill_mutex);
2550         current->journal_info = req;
2551         err = ceph_fill_trace(mdsc->fsc->sb, req);
2552         if (err == 0) {
2553                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2554                                     req->r_op == CEPH_MDS_OP_LSSNAP))
2555                         ceph_readdir_prepopulate(req, req->r_session);
2556                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2557         }
2558         current->journal_info = NULL;
2559         mutex_unlock(&req->r_fill_mutex);
2560
2561         up_read(&mdsc->snap_rwsem);
2562         if (realm)
2563                 ceph_put_snap_realm(mdsc, realm);
2564
2565         if (err == 0 && req->r_target_inode &&
2566             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2567                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
2568                 spin_lock(&ci->i_unsafe_lock);
2569                 list_add_tail(&req->r_unsafe_target_item, &ci->i_unsafe_iops);
2570                 spin_unlock(&ci->i_unsafe_lock);
2571         }
2572 out_err:
2573         mutex_lock(&mdsc->mutex);
2574         if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2575                 if (err) {
2576                         req->r_err = err;
2577                 } else {
2578                         req->r_reply =  ceph_msg_get(msg);
2579                         set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
2580                 }
2581         } else {
2582                 dout("reply arrived after request %lld was aborted\n", tid);
2583         }
2584         mutex_unlock(&mdsc->mutex);
2585
2586         mutex_unlock(&session->s_mutex);
2587
2588         /* kick calling process */
2589         complete_request(mdsc, req);
2590 out:
2591         ceph_mdsc_put_request(req);
2592         return;
2593 }
2594
2595
2596
2597 /*
2598  * handle mds notification that our request has been forwarded.
2599  */
2600 static void handle_forward(struct ceph_mds_client *mdsc,
2601                            struct ceph_mds_session *session,
2602                            struct ceph_msg *msg)
2603 {
2604         struct ceph_mds_request *req;
2605         u64 tid = le64_to_cpu(msg->hdr.tid);
2606         u32 next_mds;
2607         u32 fwd_seq;
2608         int err = -EINVAL;
2609         void *p = msg->front.iov_base;
2610         void *end = p + msg->front.iov_len;
2611
2612         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2613         next_mds = ceph_decode_32(&p);
2614         fwd_seq = ceph_decode_32(&p);
2615
2616         mutex_lock(&mdsc->mutex);
2617         req = lookup_get_request(mdsc, tid);
2618         if (!req) {
2619                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2620                 goto out;  /* dup reply? */
2621         }
2622
2623         if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2624                 dout("forward tid %llu aborted, unregistering\n", tid);
2625                 __unregister_request(mdsc, req);
2626         } else if (fwd_seq <= req->r_num_fwd) {
2627                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2628                      tid, next_mds, req->r_num_fwd, fwd_seq);
2629         } else {
2630                 /* resend. forward race not possible; mds would drop */
2631                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2632                 BUG_ON(req->r_err);
2633                 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
2634                 req->r_attempts = 0;
2635                 req->r_num_fwd = fwd_seq;
2636                 req->r_resend_mds = next_mds;
2637                 put_request_session(req);
2638                 __do_request(mdsc, req);
2639         }
2640         ceph_mdsc_put_request(req);
2641 out:
2642         mutex_unlock(&mdsc->mutex);
2643         return;
2644
2645 bad:
2646         pr_err("mdsc_handle_forward decode error err=%d\n", err);
2647 }
2648
2649 /*
2650  * handle a mds session control message
2651  */
2652 static void handle_session(struct ceph_mds_session *session,
2653                            struct ceph_msg *msg)
2654 {
2655         struct ceph_mds_client *mdsc = session->s_mdsc;
2656         u32 op;
2657         u64 seq;
2658         int mds = session->s_mds;
2659         struct ceph_mds_session_head *h = msg->front.iov_base;
2660         int wake = 0;
2661
2662         /* decode */
2663         if (msg->front.iov_len != sizeof(*h))
2664                 goto bad;
2665         op = le32_to_cpu(h->op);
2666         seq = le64_to_cpu(h->seq);
2667
2668         mutex_lock(&mdsc->mutex);
2669         if (op == CEPH_SESSION_CLOSE) {
2670                 get_session(session);
2671                 __unregister_session(mdsc, session);
2672         }
2673         /* FIXME: this ttl calculation is generous */
2674         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2675         mutex_unlock(&mdsc->mutex);
2676
2677         mutex_lock(&session->s_mutex);
2678
2679         dout("handle_session mds%d %s %p state %s seq %llu\n",
2680              mds, ceph_session_op_name(op), session,
2681              ceph_session_state_name(session->s_state), seq);
2682
2683         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2684                 session->s_state = CEPH_MDS_SESSION_OPEN;
2685                 pr_info("mds%d came back\n", session->s_mds);
2686         }
2687
2688         switch (op) {
2689         case CEPH_SESSION_OPEN:
2690                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2691                         pr_info("mds%d reconnect success\n", session->s_mds);
2692                 session->s_state = CEPH_MDS_SESSION_OPEN;
2693                 renewed_caps(mdsc, session, 0);
2694                 wake = 1;
2695                 if (mdsc->stopping)
2696                         __close_session(mdsc, session);
2697                 break;
2698
2699         case CEPH_SESSION_RENEWCAPS:
2700                 if (session->s_renew_seq == seq)
2701                         renewed_caps(mdsc, session, 1);
2702                 break;
2703
2704         case CEPH_SESSION_CLOSE:
2705                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2706                         pr_info("mds%d reconnect denied\n", session->s_mds);
2707                 cleanup_session_requests(mdsc, session);
2708                 remove_session_caps(session);
2709                 wake = 2; /* for good measure */
2710                 wake_up_all(&mdsc->session_close_wq);
2711                 break;
2712
2713         case CEPH_SESSION_STALE:
2714                 pr_info("mds%d caps went stale, renewing\n",
2715                         session->s_mds);
2716                 spin_lock(&session->s_gen_ttl_lock);
2717                 session->s_cap_gen++;
2718                 session->s_cap_ttl = jiffies - 1;
2719                 spin_unlock(&session->s_gen_ttl_lock);
2720                 send_renew_caps(mdsc, session);
2721                 break;
2722
2723         case CEPH_SESSION_RECALL_STATE:
2724                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2725                 break;
2726
2727         case CEPH_SESSION_FLUSHMSG:
2728                 send_flushmsg_ack(mdsc, session, seq);
2729                 break;
2730
2731         case CEPH_SESSION_FORCE_RO:
2732                 dout("force_session_readonly %p\n", session);
2733                 spin_lock(&session->s_cap_lock);
2734                 session->s_readonly = true;
2735                 spin_unlock(&session->s_cap_lock);
2736                 wake_up_session_caps(session, 0);
2737                 break;
2738
2739         case CEPH_SESSION_REJECT:
2740                 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
2741                 pr_info("mds%d rejected session\n", session->s_mds);
2742                 session->s_state = CEPH_MDS_SESSION_REJECTED;
2743                 cleanup_session_requests(mdsc, session);
2744                 remove_session_caps(session);
2745                 wake = 2; /* for good measure */
2746                 break;
2747
2748         default:
2749                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2750                 WARN_ON(1);
2751         }
2752
2753         mutex_unlock(&session->s_mutex);
2754         if (wake) {
2755                 mutex_lock(&mdsc->mutex);
2756                 __wake_requests(mdsc, &session->s_waiting);
2757                 if (wake == 2)
2758                         kick_requests(mdsc, mds);
2759                 mutex_unlock(&mdsc->mutex);
2760         }
2761         if (op == CEPH_SESSION_CLOSE)
2762                 ceph_put_mds_session(session);
2763         return;
2764
2765 bad:
2766         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2767                (int)msg->front.iov_len);
2768         ceph_msg_dump(msg);
2769         return;
2770 }
2771
2772
2773 /*
2774  * called under session->mutex.
2775  */
2776 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2777                                    struct ceph_mds_session *session)
2778 {
2779         struct ceph_mds_request *req, *nreq;
2780         struct rb_node *p;
2781         int err;
2782
2783         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2784
2785         mutex_lock(&mdsc->mutex);
2786         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2787                 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2788                 if (!err) {
2789                         ceph_msg_get(req->r_request);
2790                         ceph_con_send(&session->s_con, req->r_request);
2791                 }
2792         }
2793
2794         /*
2795          * also re-send old requests when MDS enters reconnect stage. So that MDS
2796          * can process completed request in clientreplay stage.
2797          */
2798         p = rb_first(&mdsc->request_tree);
2799         while (p) {
2800                 req = rb_entry(p, struct ceph_mds_request, r_node);
2801                 p = rb_next(p);
2802                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2803                         continue;
2804                 if (req->r_attempts == 0)
2805                         continue; /* only old requests */
2806                 if (req->r_session &&
2807                     req->r_session->s_mds == session->s_mds) {
2808                         err = __prepare_send_request(mdsc, req,
2809                                                      session->s_mds, true);
2810                         if (!err) {
2811                                 ceph_msg_get(req->r_request);
2812                                 ceph_con_send(&session->s_con, req->r_request);
2813                         }
2814                 }
2815         }
2816         mutex_unlock(&mdsc->mutex);
2817 }
2818
2819 /*
2820  * Encode information about a cap for a reconnect with the MDS.
2821  */
2822 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2823                           void *arg)
2824 {
2825         union {
2826                 struct ceph_mds_cap_reconnect v2;
2827                 struct ceph_mds_cap_reconnect_v1 v1;
2828         } rec;
2829         struct ceph_inode_info *ci;
2830         struct ceph_reconnect_state *recon_state = arg;
2831         struct ceph_pagelist *pagelist = recon_state->pagelist;
2832         char *path;
2833         int pathlen, err;
2834         u64 pathbase;
2835         u64 snap_follows;
2836         struct dentry *dentry;
2837
2838         ci = cap->ci;
2839
2840         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2841              inode, ceph_vinop(inode), cap, cap->cap_id,
2842              ceph_cap_string(cap->issued));
2843         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2844         if (err)
2845                 return err;
2846
2847         dentry = d_find_alias(inode);
2848         if (dentry) {
2849                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2850                 if (IS_ERR(path)) {
2851                         err = PTR_ERR(path);
2852                         goto out_dput;
2853                 }
2854         } else {
2855                 path = NULL;
2856                 pathlen = 0;
2857                 pathbase = 0;
2858         }
2859
2860         spin_lock(&ci->i_ceph_lock);
2861         cap->seq = 0;        /* reset cap seq */
2862         cap->issue_seq = 0;  /* and issue_seq */
2863         cap->mseq = 0;       /* and migrate_seq */
2864         cap->cap_gen = cap->session->s_cap_gen;
2865
2866         if (recon_state->msg_version >= 2) {
2867                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2868                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2869                 rec.v2.issued = cpu_to_le32(cap->issued);
2870                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2871                 rec.v2.pathbase = cpu_to_le64(pathbase);
2872                 rec.v2.flock_len = 0;
2873         } else {
2874                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2875                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2876                 rec.v1.issued = cpu_to_le32(cap->issued);
2877                 rec.v1.size = cpu_to_le64(inode->i_size);
2878                 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2879                 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2880                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2881                 rec.v1.pathbase = cpu_to_le64(pathbase);
2882         }
2883
2884         if (list_empty(&ci->i_cap_snaps)) {
2885                 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
2886         } else {
2887                 struct ceph_cap_snap *capsnap =
2888                         list_first_entry(&ci->i_cap_snaps,
2889                                          struct ceph_cap_snap, ci_item);
2890                 snap_follows = capsnap->follows;
2891         }
2892         spin_unlock(&ci->i_ceph_lock);
2893
2894         if (recon_state->msg_version >= 2) {
2895                 int num_fcntl_locks, num_flock_locks;
2896                 struct ceph_filelock *flocks;
2897                 size_t struct_len, total_len = 0;
2898                 u8 struct_v = 0;
2899
2900 encode_again:
2901                 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2902                 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2903                                  sizeof(struct ceph_filelock), GFP_NOFS);
2904                 if (!flocks) {
2905                         err = -ENOMEM;
2906                         goto out_free;
2907                 }
2908                 err = ceph_encode_locks_to_buffer(inode, flocks,
2909                                                   num_fcntl_locks,
2910                                                   num_flock_locks);
2911                 if (err) {
2912                         kfree(flocks);
2913                         if (err == -ENOSPC)
2914                                 goto encode_again;
2915                         goto out_free;
2916                 }
2917
2918                 if (recon_state->msg_version >= 3) {
2919                         /* version, compat_version and struct_len */
2920                         total_len = 2 * sizeof(u8) + sizeof(u32);
2921                         struct_v = 2;
2922                 }
2923                 /*
2924                  * number of encoded locks is stable, so copy to pagelist
2925                  */
2926                 struct_len = 2 * sizeof(u32) +
2927                             (num_fcntl_locks + num_flock_locks) *
2928                             sizeof(struct ceph_filelock);
2929                 rec.v2.flock_len = cpu_to_le32(struct_len);
2930
2931                 struct_len += sizeof(rec.v2);
2932                 struct_len += sizeof(u32) + pathlen;
2933
2934                 if (struct_v >= 2)
2935                         struct_len += sizeof(u64); /* snap_follows */
2936
2937                 total_len += struct_len;
2938                 err = ceph_pagelist_reserve(pagelist, total_len);
2939
2940                 if (!err) {
2941                         if (recon_state->msg_version >= 3) {
2942                                 ceph_pagelist_encode_8(pagelist, struct_v);
2943                                 ceph_pagelist_encode_8(pagelist, 1);
2944                                 ceph_pagelist_encode_32(pagelist, struct_len);
2945                         }
2946                         ceph_pagelist_encode_string(pagelist, path, pathlen);
2947                         ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
2948                         ceph_locks_to_pagelist(flocks, pagelist,
2949                                                num_fcntl_locks,
2950                                                num_flock_locks);
2951                         if (struct_v >= 2)
2952                                 ceph_pagelist_encode_64(pagelist, snap_follows);
2953                 }
2954                 kfree(flocks);
2955         } else {
2956                 size_t size = sizeof(u32) + pathlen + sizeof(rec.v1);
2957                 err = ceph_pagelist_reserve(pagelist, size);
2958                 if (!err) {
2959                         ceph_pagelist_encode_string(pagelist, path, pathlen);
2960                         ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
2961                 }
2962         }
2963
2964         recon_state->nr_caps++;
2965 out_free:
2966         kfree(path);
2967 out_dput:
2968         dput(dentry);
2969         return err;
2970 }
2971
2972
2973 /*
2974  * If an MDS fails and recovers, clients need to reconnect in order to
2975  * reestablish shared state.  This includes all caps issued through
2976  * this session _and_ the snap_realm hierarchy.  Because it's not
2977  * clear which snap realms the mds cares about, we send everything we
2978  * know about.. that ensures we'll then get any new info the
2979  * recovering MDS might have.
2980  *
2981  * This is a relatively heavyweight operation, but it's rare.
2982  *
2983  * called with mdsc->mutex held.
2984  */
2985 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2986                                struct ceph_mds_session *session)
2987 {
2988         struct ceph_msg *reply;
2989         struct rb_node *p;
2990         int mds = session->s_mds;
2991         int err = -ENOMEM;
2992         int s_nr_caps;
2993         struct ceph_pagelist *pagelist;
2994         struct ceph_reconnect_state recon_state;
2995
2996         pr_info("mds%d reconnect start\n", mds);
2997
2998         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2999         if (!pagelist)
3000                 goto fail_nopagelist;
3001         ceph_pagelist_init(pagelist);
3002
3003         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
3004         if (!reply)
3005                 goto fail_nomsg;
3006
3007         mutex_lock(&session->s_mutex);
3008         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3009         session->s_seq = 0;
3010
3011         dout("session %p state %s\n", session,
3012              ceph_session_state_name(session->s_state));
3013
3014         spin_lock(&session->s_gen_ttl_lock);
3015         session->s_cap_gen++;
3016         spin_unlock(&session->s_gen_ttl_lock);
3017
3018         spin_lock(&session->s_cap_lock);
3019         /* don't know if session is readonly */
3020         session->s_readonly = 0;
3021         /*
3022          * notify __ceph_remove_cap() that we are composing cap reconnect.
3023          * If a cap get released before being added to the cap reconnect,
3024          * __ceph_remove_cap() should skip queuing cap release.
3025          */
3026         session->s_cap_reconnect = 1;
3027         /* drop old cap expires; we're about to reestablish that state */
3028         cleanup_cap_releases(mdsc, session);
3029
3030         /* trim unused caps to reduce MDS's cache rejoin time */
3031         if (mdsc->fsc->sb->s_root)
3032                 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3033
3034         ceph_con_close(&session->s_con);
3035         ceph_con_open(&session->s_con,
3036                       CEPH_ENTITY_TYPE_MDS, mds,
3037                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3038
3039         /* replay unsafe requests */
3040         replay_unsafe_requests(mdsc, session);
3041
3042         down_read(&mdsc->snap_rwsem);
3043
3044         /* traverse this session's caps */
3045         s_nr_caps = session->s_nr_caps;
3046         err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
3047         if (err)
3048                 goto fail;
3049
3050         recon_state.nr_caps = 0;
3051         recon_state.pagelist = pagelist;
3052         if (session->s_con.peer_features & CEPH_FEATURE_MDSENC)
3053                 recon_state.msg_version = 3;
3054         else if (session->s_con.peer_features & CEPH_FEATURE_FLOCK)
3055                 recon_state.msg_version = 2;
3056         else
3057                 recon_state.msg_version = 1;
3058         err = iterate_session_caps(session, encode_caps_cb, &recon_state);
3059         if (err < 0)
3060                 goto fail;
3061
3062         spin_lock(&session->s_cap_lock);
3063         session->s_cap_reconnect = 0;
3064         spin_unlock(&session->s_cap_lock);
3065
3066         /*
3067          * snaprealms.  we provide mds with the ino, seq (version), and
3068          * parent for all of our realms.  If the mds has any newer info,
3069          * it will tell us.
3070          */
3071         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3072                 struct ceph_snap_realm *realm =
3073                         rb_entry(p, struct ceph_snap_realm, node);
3074                 struct ceph_mds_snaprealm_reconnect sr_rec;
3075
3076                 dout(" adding snap realm %llx seq %lld parent %llx\n",
3077                      realm->ino, realm->seq, realm->parent_ino);
3078                 sr_rec.ino = cpu_to_le64(realm->ino);
3079                 sr_rec.seq = cpu_to_le64(realm->seq);
3080                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3081                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3082                 if (err)
3083                         goto fail;
3084         }
3085
3086         reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3087
3088         /* raced with cap release? */
3089         if (s_nr_caps != recon_state.nr_caps) {
3090                 struct page *page = list_first_entry(&pagelist->head,
3091                                                      struct page, lru);
3092                 __le32 *addr = kmap_atomic(page);
3093                 *addr = cpu_to_le32(recon_state.nr_caps);
3094                 kunmap_atomic(addr);
3095         }
3096
3097         reply->hdr.data_len = cpu_to_le32(pagelist->length);
3098         ceph_msg_data_add_pagelist(reply, pagelist);
3099
3100         ceph_early_kick_flushing_caps(mdsc, session);
3101
3102         ceph_con_send(&session->s_con, reply);
3103
3104         mutex_unlock(&session->s_mutex);
3105
3106         mutex_lock(&mdsc->mutex);
3107         __wake_requests(mdsc, &session->s_waiting);
3108         mutex_unlock(&mdsc->mutex);
3109
3110         up_read(&mdsc->snap_rwsem);
3111         return;
3112
3113 fail:
3114         ceph_msg_put(reply);
3115         up_read(&mdsc->snap_rwsem);
3116         mutex_unlock(&session->s_mutex);
3117 fail_nomsg:
3118         ceph_pagelist_release(pagelist);
3119 fail_nopagelist:
3120         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3121         return;
3122 }
3123
3124
3125 /*
3126  * compare old and new mdsmaps, kicking requests
3127  * and closing out old connections as necessary
3128  *
3129  * called under mdsc->mutex.
3130  */
3131 static void check_new_map(struct ceph_mds_client *mdsc,
3132                           struct ceph_mdsmap *newmap,
3133                           struct ceph_mdsmap *oldmap)
3134 {
3135         int i;
3136         int oldstate, newstate;
3137         struct ceph_mds_session *s;
3138
3139         dout("check_new_map new %u old %u\n",
3140              newmap->m_epoch, oldmap->m_epoch);
3141
3142         for (i = 0; i < oldmap->m_num_mds && i < mdsc->max_sessions; i++) {
3143                 if (!mdsc->sessions[i])
3144                         continue;
3145                 s = mdsc->sessions[i];
3146                 oldstate = ceph_mdsmap_get_state(oldmap, i);
3147                 newstate = ceph_mdsmap_get_state(newmap, i);
3148
3149                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3150                      i, ceph_mds_state_name(oldstate),
3151                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3152                      ceph_mds_state_name(newstate),
3153                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3154                      ceph_session_state_name(s->s_state));
3155
3156                 if (i >= newmap->m_num_mds ||
3157                     memcmp(ceph_mdsmap_get_addr(oldmap, i),
3158                            ceph_mdsmap_get_addr(newmap, i),
3159                            sizeof(struct ceph_entity_addr))) {
3160                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3161                                 /* the session never opened, just close it
3162                                  * out now */
3163                                 get_session(s);
3164                                 __unregister_session(mdsc, s);
3165                                 __wake_requests(mdsc, &s->s_waiting);
3166                                 ceph_put_mds_session(s);
3167                         } else if (i >= newmap->m_num_mds) {
3168                                 /* force close session for stopped mds */
3169                                 get_session(s);
3170                                 __unregister_session(mdsc, s);
3171                                 __wake_requests(mdsc, &s->s_waiting);
3172                                 kick_requests(mdsc, i);
3173                                 mutex_unlock(&mdsc->mutex);
3174
3175                                 mutex_lock(&s->s_mutex);
3176                                 cleanup_session_requests(mdsc, s);
3177                                 remove_session_caps(s);
3178                                 mutex_unlock(&s->s_mutex);
3179
3180                                 ceph_put_mds_session(s);
3181
3182                                 mutex_lock(&mdsc->mutex);
3183                         } else {
3184                                 /* just close it */
3185                                 mutex_unlock(&mdsc->mutex);
3186                                 mutex_lock(&s->s_mutex);
3187                                 mutex_lock(&mdsc->mutex);
3188                                 ceph_con_close(&s->s_con);
3189                                 mutex_unlock(&s->s_mutex);
3190                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3191                         }
3192                 } else if (oldstate == newstate) {
3193                         continue;  /* nothing new with this mds */
3194                 }
3195
3196                 /*
3197                  * send reconnect?
3198                  */
3199                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3200                     newstate >= CEPH_MDS_STATE_RECONNECT) {
3201                         mutex_unlock(&mdsc->mutex);
3202                         send_mds_reconnect(mdsc, s);
3203                         mutex_lock(&mdsc->mutex);
3204                 }
3205
3206                 /*
3207                  * kick request on any mds that has gone active.
3208                  */
3209                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3210                     newstate >= CEPH_MDS_STATE_ACTIVE) {
3211                         if (oldstate != CEPH_MDS_STATE_CREATING &&
3212                             oldstate != CEPH_MDS_STATE_STARTING)
3213                                 pr_info("mds%d recovery completed\n", s->s_mds);
3214                         kick_requests(mdsc, i);
3215                         ceph_kick_flushing_caps(mdsc, s);
3216                         wake_up_session_caps(s, 1);
3217                 }
3218         }
3219
3220         for (i = 0; i < newmap->m_num_mds && i < mdsc->max_sessions; i++) {
3221                 s = mdsc->sessions[i];
3222                 if (!s)
3223                         continue;
3224                 if (!ceph_mdsmap_is_laggy(newmap, i))
3225                         continue;
3226                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3227                     s->s_state == CEPH_MDS_SESSION_HUNG ||
3228                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
3229                         dout(" connecting to export targets of laggy mds%d\n",
3230                              i);
3231                         __open_export_target_sessions(mdsc, s);
3232                 }
3233         }
3234 }
3235
3236
3237
3238 /*
3239  * leases
3240  */
3241
3242 /*
3243  * caller must hold session s_mutex, dentry->d_lock
3244  */
3245 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3246 {
3247         struct ceph_dentry_info *di = ceph_dentry(dentry);
3248
3249         ceph_put_mds_session(di->lease_session);
3250         di->lease_session = NULL;
3251 }
3252
3253 static void handle_lease(struct ceph_mds_client *mdsc,
3254                          struct ceph_mds_session *session,
3255                          struct ceph_msg *msg)
3256 {
3257         struct super_block *sb = mdsc->fsc->sb;
3258         struct inode *inode;
3259         struct dentry *parent, *dentry;
3260         struct ceph_dentry_info *di;
3261         int mds = session->s_mds;
3262         struct ceph_mds_lease *h = msg->front.iov_base;
3263         u32 seq;
3264         struct ceph_vino vino;
3265         struct qstr dname;
3266         int release = 0;
3267
3268         dout("handle_lease from mds%d\n", mds);
3269
3270         /* decode */
3271         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3272                 goto bad;
3273         vino.ino = le64_to_cpu(h->ino);
3274         vino.snap = CEPH_NOSNAP;
3275         seq = le32_to_cpu(h->seq);
3276         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3277         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3278         if (dname.len != get_unaligned_le32(h+1))
3279                 goto bad;
3280
3281         /* lookup inode */
3282         inode = ceph_find_inode(sb, vino);
3283         dout("handle_lease %s, ino %llx %p %.*s\n",
3284              ceph_lease_op_name(h->action), vino.ino, inode,
3285              dname.len, dname.name);
3286
3287         mutex_lock(&session->s_mutex);
3288         session->s_seq++;
3289
3290         if (!inode) {
3291                 dout("handle_lease no inode %llx\n", vino.ino);
3292                 goto release;
3293         }
3294
3295         /* dentry */
3296         parent = d_find_alias(inode);
3297         if (!parent) {
3298                 dout("no parent dentry on inode %p\n", inode);
3299                 WARN_ON(1);
3300                 goto release;  /* hrm... */
3301         }
3302         dname.hash = full_name_hash(parent, dname.name, dname.len);
3303         dentry = d_lookup(parent, &dname);
3304         dput(parent);
3305         if (!dentry)
3306                 goto release;
3307
3308         spin_lock(&dentry->d_lock);
3309         di = ceph_dentry(dentry);
3310         switch (h->action) {
3311         case CEPH_MDS_LEASE_REVOKE:
3312                 if (di->lease_session == session) {
3313                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3314                                 h->seq = cpu_to_le32(di->lease_seq);
3315                         __ceph_mdsc_drop_dentry_lease(dentry);
3316                 }
3317                 release = 1;
3318                 break;
3319
3320         case CEPH_MDS_LEASE_RENEW:
3321                 if (di->lease_session == session &&
3322                     di->lease_gen == session->s_cap_gen &&
3323                     di->lease_renew_from &&
3324                     di->lease_renew_after == 0) {
3325                         unsigned long duration =
3326                                 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3327
3328                         di->lease_seq = seq;
3329                         di->time = di->lease_renew_from + duration;
3330                         di->lease_renew_after = di->lease_renew_from +
3331                                 (duration >> 1);
3332                         di->lease_renew_from = 0;
3333                 }
3334                 break;
3335         }
3336         spin_unlock(&dentry->d_lock);
3337         dput(dentry);
3338
3339         if (!release)
3340                 goto out;
3341
3342 release:
3343         /* let's just reuse the same message */
3344         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3345         ceph_msg_get(msg);
3346         ceph_con_send(&session->s_con, msg);
3347
3348 out:
3349         iput(inode);
3350         mutex_unlock(&session->s_mutex);
3351         return;
3352
3353 bad:
3354         pr_err("corrupt lease message\n");
3355         ceph_msg_dump(msg);
3356 }
3357
3358 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3359                               struct inode *inode,
3360                               struct dentry *dentry, char action,
3361                               u32 seq)
3362 {
3363         struct ceph_msg *msg;
3364         struct ceph_mds_lease *lease;
3365         int len = sizeof(*lease) + sizeof(u32);
3366         int dnamelen = 0;
3367
3368         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3369              inode, dentry, ceph_lease_op_name(action), session->s_mds);
3370         dnamelen = dentry->d_name.len;
3371         len += dnamelen;
3372
3373         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3374         if (!msg)
3375                 return;
3376         lease = msg->front.iov_base;
3377         lease->action = action;
3378         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3379         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3380         lease->seq = cpu_to_le32(seq);
3381         put_unaligned_le32(dnamelen, lease + 1);
3382         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3383
3384         /*
3385          * if this is a preemptive lease RELEASE, no need to
3386          * flush request stream, since the actual request will
3387          * soon follow.
3388          */
3389         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3390
3391         ceph_con_send(&session->s_con, msg);
3392 }
3393
3394 /*
3395  * drop all leases (and dentry refs) in preparation for umount
3396  */
3397 static void drop_leases(struct ceph_mds_client *mdsc)
3398 {
3399         int i;
3400
3401         dout("drop_leases\n");
3402         mutex_lock(&mdsc->mutex);
3403         for (i = 0; i < mdsc->max_sessions; i++) {
3404                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3405                 if (!s)
3406                         continue;
3407                 mutex_unlock(&mdsc->mutex);
3408                 mutex_lock(&s->s_mutex);
3409                 mutex_unlock(&s->s_mutex);
3410                 ceph_put_mds_session(s);
3411                 mutex_lock(&mdsc->mutex);
3412         }
3413         mutex_unlock(&mdsc->mutex);
3414 }
3415
3416
3417
3418 /*
3419  * delayed work -- periodically trim expired leases, renew caps with mds
3420  */
3421 static void schedule_delayed(struct ceph_mds_client *mdsc)
3422 {
3423         int delay = 5;
3424         unsigned hz = round_jiffies_relative(HZ * delay);
3425         schedule_delayed_work(&mdsc->delayed_work, hz);
3426 }
3427
3428 static void delayed_work(struct work_struct *work)
3429 {
3430         int i;
3431         struct ceph_mds_client *mdsc =
3432                 container_of(work, struct ceph_mds_client, delayed_work.work);
3433         int renew_interval;
3434         int renew_caps;
3435
3436         dout("mdsc delayed_work\n");
3437         ceph_check_delayed_caps(mdsc);
3438
3439         mutex_lock(&mdsc->mutex);
3440         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3441         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3442                                    mdsc->last_renew_caps);
3443         if (renew_caps)
3444                 mdsc->last_renew_caps = jiffies;
3445
3446         for (i = 0; i < mdsc->max_sessions; i++) {
3447                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3448                 if (!s)
3449                         continue;
3450                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3451                         dout("resending session close request for mds%d\n",
3452                              s->s_mds);
3453                         request_close_session(mdsc, s);
3454                         ceph_put_mds_session(s);
3455                         continue;
3456                 }
3457                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3458                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3459                                 s->s_state = CEPH_MDS_SESSION_HUNG;
3460                                 pr_info("mds%d hung\n", s->s_mds);
3461                         }
3462                 }
3463                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3464                         /* this mds is failed or recovering, just wait */
3465                         ceph_put_mds_session(s);
3466                         continue;
3467                 }
3468                 mutex_unlock(&mdsc->mutex);
3469
3470                 mutex_lock(&s->s_mutex);
3471                 if (renew_caps)
3472                         send_renew_caps(mdsc, s);
3473                 else
3474                         ceph_con_keepalive(&s->s_con);
3475                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3476                     s->s_state == CEPH_MDS_SESSION_HUNG)
3477                         ceph_send_cap_releases(mdsc, s);
3478                 mutex_unlock(&s->s_mutex);
3479                 ceph_put_mds_session(s);
3480
3481                 mutex_lock(&mdsc->mutex);
3482         }
3483         mutex_unlock(&mdsc->mutex);
3484
3485         schedule_delayed(mdsc);
3486 }
3487
3488 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3489
3490 {
3491         struct ceph_mds_client *mdsc;
3492
3493         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3494         if (!mdsc)
3495                 return -ENOMEM;
3496         mdsc->fsc = fsc;
3497         fsc->mdsc = mdsc;
3498         mutex_init(&mdsc->mutex);
3499         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3500         if (!mdsc->mdsmap) {
3501                 kfree(mdsc);
3502                 return -ENOMEM;
3503         }
3504
3505         init_completion(&mdsc->safe_umount_waiters);
3506         init_waitqueue_head(&mdsc->session_close_wq);
3507         INIT_LIST_HEAD(&mdsc->waiting_for_map);
3508         mdsc->sessions = NULL;
3509         atomic_set(&mdsc->num_sessions, 0);
3510         mdsc->max_sessions = 0;
3511         mdsc->stopping = 0;
3512         mdsc->last_snap_seq = 0;
3513         init_rwsem(&mdsc->snap_rwsem);
3514         mdsc->snap_realms = RB_ROOT;
3515         INIT_LIST_HEAD(&mdsc->snap_empty);
3516         spin_lock_init(&mdsc->snap_empty_lock);
3517         mdsc->last_tid = 0;
3518         mdsc->oldest_tid = 0;
3519         mdsc->request_tree = RB_ROOT;
3520         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3521         mdsc->last_renew_caps = jiffies;
3522         INIT_LIST_HEAD(&mdsc->cap_delay_list);
3523         spin_lock_init(&mdsc->cap_delay_lock);
3524         INIT_LIST_HEAD(&mdsc->snap_flush_list);
3525         spin_lock_init(&mdsc->snap_flush_lock);
3526         mdsc->last_cap_flush_tid = 1;
3527         INIT_LIST_HEAD(&mdsc->cap_flush_list);
3528         INIT_LIST_HEAD(&mdsc->cap_dirty);
3529         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3530         mdsc->num_cap_flushing = 0;
3531         spin_lock_init(&mdsc->cap_dirty_lock);
3532         init_waitqueue_head(&mdsc->cap_flushing_wq);
3533         spin_lock_init(&mdsc->dentry_lru_lock);
3534         INIT_LIST_HEAD(&mdsc->dentry_lru);
3535
3536         ceph_caps_init(mdsc);
3537         ceph_adjust_min_caps(mdsc, fsc->min_caps);
3538
3539         init_rwsem(&mdsc->pool_perm_rwsem);
3540         mdsc->pool_perm_tree = RB_ROOT;
3541
3542         strncpy(mdsc->nodename, utsname()->nodename,
3543                 sizeof(mdsc->nodename) - 1);
3544         return 0;
3545 }
3546
3547 /*
3548  * Wait for safe replies on open mds requests.  If we time out, drop
3549  * all requests from the tree to avoid dangling dentry refs.
3550  */
3551 static void wait_requests(struct ceph_mds_client *mdsc)
3552 {
3553         struct ceph_options *opts = mdsc->fsc->client->options;
3554         struct ceph_mds_request *req;
3555
3556         mutex_lock(&mdsc->mutex);
3557         if (__get_oldest_req(mdsc)) {
3558                 mutex_unlock(&mdsc->mutex);
3559
3560                 dout("wait_requests waiting for requests\n");
3561                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3562                                     ceph_timeout_jiffies(opts->mount_timeout));
3563
3564                 /* tear down remaining requests */
3565                 mutex_lock(&mdsc->mutex);
3566                 while ((req = __get_oldest_req(mdsc))) {
3567                         dout("wait_requests timed out on tid %llu\n",
3568                              req->r_tid);
3569                         __unregister_request(mdsc, req);
3570                 }
3571         }
3572         mutex_unlock(&mdsc->mutex);
3573         dout("wait_requests done\n");
3574 }
3575
3576 /*
3577  * called before mount is ro, and before dentries are torn down.
3578  * (hmm, does this still race with new lookups?)
3579  */
3580 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3581 {
3582         dout("pre_umount\n");
3583         mdsc->stopping = 1;
3584
3585         drop_leases(mdsc);
3586         ceph_flush_dirty_caps(mdsc);
3587         wait_requests(mdsc);
3588
3589         /*
3590          * wait for reply handlers to drop their request refs and
3591          * their inode/dcache refs
3592          */
3593         ceph_msgr_flush();
3594 }
3595
3596 /*
3597  * wait for all write mds requests to flush.
3598  */
3599 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3600 {
3601         struct ceph_mds_request *req = NULL, *nextreq;
3602         struct rb_node *n;
3603
3604         mutex_lock(&mdsc->mutex);
3605         dout("wait_unsafe_requests want %lld\n", want_tid);
3606 restart:
3607         req = __get_oldest_req(mdsc);
3608         while (req && req->r_tid <= want_tid) {
3609                 /* find next request */
3610                 n = rb_next(&req->r_node);
3611                 if (n)
3612                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3613                 else
3614                         nextreq = NULL;
3615                 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3616                     (req->r_op & CEPH_MDS_OP_WRITE)) {
3617                         /* write op */
3618                         ceph_mdsc_get_request(req);
3619                         if (nextreq)
3620                                 ceph_mdsc_get_request(nextreq);
3621                         mutex_unlock(&mdsc->mutex);
3622                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
3623                              req->r_tid, want_tid);
3624                         wait_for_completion(&req->r_safe_completion);
3625                         mutex_lock(&mdsc->mutex);
3626                         ceph_mdsc_put_request(req);
3627                         if (!nextreq)
3628                                 break;  /* next dne before, so we're done! */
3629                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
3630                                 /* next request was removed from tree */
3631                                 ceph_mdsc_put_request(nextreq);
3632                                 goto restart;
3633                         }
3634                         ceph_mdsc_put_request(nextreq);  /* won't go away */
3635                 }
3636                 req = nextreq;
3637         }
3638         mutex_unlock(&mdsc->mutex);
3639         dout("wait_unsafe_requests done\n");
3640 }
3641
3642 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3643 {
3644         u64 want_tid, want_flush;
3645
3646         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3647                 return;
3648
3649         dout("sync\n");
3650         mutex_lock(&mdsc->mutex);
3651         want_tid = mdsc->last_tid;
3652         mutex_unlock(&mdsc->mutex);
3653
3654         ceph_flush_dirty_caps(mdsc);
3655         spin_lock(&mdsc->cap_dirty_lock);
3656         want_flush = mdsc->last_cap_flush_tid;
3657         if (!list_empty(&mdsc->cap_flush_list)) {
3658                 struct ceph_cap_flush *cf =
3659                         list_last_entry(&mdsc->cap_flush_list,
3660                                         struct ceph_cap_flush, g_list);
3661                 cf->wake = true;
3662         }
3663         spin_unlock(&mdsc->cap_dirty_lock);
3664
3665         dout("sync want tid %lld flush_seq %lld\n",
3666              want_tid, want_flush);
3667
3668         wait_unsafe_requests(mdsc, want_tid);
3669         wait_caps_flush(mdsc, want_flush);
3670 }
3671
3672 /*
3673  * true if all sessions are closed, or we force unmount
3674  */
3675 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
3676 {
3677         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3678                 return true;
3679         return atomic_read(&mdsc->num_sessions) <= skipped;
3680 }
3681
3682 /*
3683  * called after sb is ro.
3684  */
3685 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3686 {
3687         struct ceph_options *opts = mdsc->fsc->client->options;
3688         struct ceph_mds_session *session;
3689         int i;
3690         int skipped = 0;
3691
3692         dout("close_sessions\n");
3693
3694         /* close sessions */
3695         mutex_lock(&mdsc->mutex);
3696         for (i = 0; i < mdsc->max_sessions; i++) {
3697                 session = __ceph_lookup_mds_session(mdsc, i);
3698                 if (!session)
3699                         continue;
3700                 mutex_unlock(&mdsc->mutex);
3701                 mutex_lock(&session->s_mutex);
3702                 if (__close_session(mdsc, session) <= 0)
3703                         skipped++;
3704                 mutex_unlock(&session->s_mutex);
3705                 ceph_put_mds_session(session);
3706                 mutex_lock(&mdsc->mutex);
3707         }
3708         mutex_unlock(&mdsc->mutex);
3709
3710         dout("waiting for sessions to close\n");
3711         wait_event_timeout(mdsc->session_close_wq,
3712                            done_closing_sessions(mdsc, skipped),
3713                            ceph_timeout_jiffies(opts->mount_timeout));
3714
3715         /* tear down remaining sessions */
3716         mutex_lock(&mdsc->mutex);
3717         for (i = 0; i < mdsc->max_sessions; i++) {
3718                 if (mdsc->sessions[i]) {
3719                         session = get_session(mdsc->sessions[i]);
3720                         __unregister_session(mdsc, session);
3721                         mutex_unlock(&mdsc->mutex);
3722                         mutex_lock(&session->s_mutex);
3723                         remove_session_caps(session);
3724                         mutex_unlock(&session->s_mutex);
3725                         ceph_put_mds_session(session);
3726                         mutex_lock(&mdsc->mutex);
3727                 }
3728         }
3729         WARN_ON(!list_empty(&mdsc->cap_delay_list));
3730         mutex_unlock(&mdsc->mutex);
3731
3732         ceph_cleanup_empty_realms(mdsc);
3733
3734         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3735
3736         dout("stopped\n");
3737 }
3738
3739 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
3740 {
3741         struct ceph_mds_session *session;
3742         int mds;
3743
3744         dout("force umount\n");
3745
3746         mutex_lock(&mdsc->mutex);
3747         for (mds = 0; mds < mdsc->max_sessions; mds++) {
3748                 session = __ceph_lookup_mds_session(mdsc, mds);
3749                 if (!session)
3750                         continue;
3751                 mutex_unlock(&mdsc->mutex);
3752                 mutex_lock(&session->s_mutex);
3753                 __close_session(mdsc, session);
3754                 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
3755                         cleanup_session_requests(mdsc, session);
3756                         remove_session_caps(session);
3757                 }
3758                 mutex_unlock(&session->s_mutex);
3759                 ceph_put_mds_session(session);
3760                 mutex_lock(&mdsc->mutex);
3761                 kick_requests(mdsc, mds);
3762         }
3763         __wake_requests(mdsc, &mdsc->waiting_for_map);
3764         mutex_unlock(&mdsc->mutex);
3765 }
3766
3767 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3768 {
3769         dout("stop\n");
3770         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3771         if (mdsc->mdsmap)
3772                 ceph_mdsmap_destroy(mdsc->mdsmap);
3773         kfree(mdsc->sessions);
3774         ceph_caps_finalize(mdsc);
3775         ceph_pool_perm_destroy(mdsc);
3776 }
3777
3778 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3779 {
3780         struct ceph_mds_client *mdsc = fsc->mdsc;
3781         dout("mdsc_destroy %p\n", mdsc);
3782
3783         /* flush out any connection work with references to us */
3784         ceph_msgr_flush();
3785
3786         ceph_mdsc_stop(mdsc);
3787
3788         fsc->mdsc = NULL;
3789         kfree(mdsc);
3790         dout("mdsc_destroy %p done\n", mdsc);
3791 }
3792
3793 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3794 {
3795         struct ceph_fs_client *fsc = mdsc->fsc;
3796         const char *mds_namespace = fsc->mount_options->mds_namespace;
3797         void *p = msg->front.iov_base;
3798         void *end = p + msg->front.iov_len;
3799         u32 epoch;
3800         u32 map_len;
3801         u32 num_fs;
3802         u32 mount_fscid = (u32)-1;
3803         u8 struct_v, struct_cv;
3804         int err = -EINVAL;
3805
3806         ceph_decode_need(&p, end, sizeof(u32), bad);
3807         epoch = ceph_decode_32(&p);
3808
3809         dout("handle_fsmap epoch %u\n", epoch);
3810
3811         ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3812         struct_v = ceph_decode_8(&p);
3813         struct_cv = ceph_decode_8(&p);
3814         map_len = ceph_decode_32(&p);
3815
3816         ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
3817         p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
3818
3819         num_fs = ceph_decode_32(&p);
3820         while (num_fs-- > 0) {
3821                 void *info_p, *info_end;
3822                 u32 info_len;
3823                 u8 info_v, info_cv;
3824                 u32 fscid, namelen;
3825
3826                 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3827                 info_v = ceph_decode_8(&p);
3828                 info_cv = ceph_decode_8(&p);
3829                 info_len = ceph_decode_32(&p);
3830                 ceph_decode_need(&p, end, info_len, bad);
3831                 info_p = p;
3832                 info_end = p + info_len;
3833                 p = info_end;
3834
3835                 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
3836                 fscid = ceph_decode_32(&info_p);
3837                 namelen = ceph_decode_32(&info_p);
3838                 ceph_decode_need(&info_p, info_end, namelen, bad);
3839
3840                 if (mds_namespace &&
3841                     strlen(mds_namespace) == namelen &&
3842                     !strncmp(mds_namespace, (char *)info_p, namelen)) {
3843                         mount_fscid = fscid;
3844                         break;
3845                 }
3846         }
3847
3848         ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
3849         if (mount_fscid != (u32)-1) {
3850                 fsc->client->monc.fs_cluster_id = mount_fscid;
3851                 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
3852                                    0, true);
3853                 ceph_monc_renew_subs(&fsc->client->monc);
3854         } else {
3855                 err = -ENOENT;
3856                 goto err_out;
3857         }
3858         return;
3859 bad:
3860         pr_err("error decoding fsmap\n");
3861 err_out:
3862         mutex_lock(&mdsc->mutex);
3863         mdsc->mdsmap_err = -ENOENT;
3864         __wake_requests(mdsc, &mdsc->waiting_for_map);
3865         mutex_unlock(&mdsc->mutex);
3866         return;
3867 }
3868
3869 /*
3870  * handle mds map update.
3871  */
3872 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3873 {
3874         u32 epoch;
3875         u32 maplen;
3876         void *p = msg->front.iov_base;
3877         void *end = p + msg->front.iov_len;
3878         struct ceph_mdsmap *newmap, *oldmap;
3879         struct ceph_fsid fsid;
3880         int err = -EINVAL;
3881
3882         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3883         ceph_decode_copy(&p, &fsid, sizeof(fsid));
3884         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3885                 return;
3886         epoch = ceph_decode_32(&p);
3887         maplen = ceph_decode_32(&p);
3888         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3889
3890         /* do we need it? */
3891         mutex_lock(&mdsc->mutex);
3892         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3893                 dout("handle_map epoch %u <= our %u\n",
3894                      epoch, mdsc->mdsmap->m_epoch);
3895                 mutex_unlock(&mdsc->mutex);
3896                 return;
3897         }
3898
3899         newmap = ceph_mdsmap_decode(&p, end);
3900         if (IS_ERR(newmap)) {
3901                 err = PTR_ERR(newmap);
3902                 goto bad_unlock;
3903         }
3904
3905         /* swap into place */
3906         if (mdsc->mdsmap) {
3907                 oldmap = mdsc->mdsmap;
3908                 mdsc->mdsmap = newmap;
3909                 check_new_map(mdsc, newmap, oldmap);
3910                 ceph_mdsmap_destroy(oldmap);
3911         } else {
3912                 mdsc->mdsmap = newmap;  /* first mds map */
3913         }
3914         mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3915
3916         __wake_requests(mdsc, &mdsc->waiting_for_map);
3917         ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
3918                           mdsc->mdsmap->m_epoch);
3919
3920         mutex_unlock(&mdsc->mutex);
3921         schedule_delayed(mdsc);
3922         return;
3923
3924 bad_unlock:
3925         mutex_unlock(&mdsc->mutex);
3926 bad:
3927         pr_err("error decoding mdsmap %d\n", err);
3928         return;
3929 }
3930
3931 static struct ceph_connection *con_get(struct ceph_connection *con)
3932 {
3933         struct ceph_mds_session *s = con->private;
3934
3935         if (get_session(s)) {
3936                 dout("mdsc con_get %p ok (%d)\n", s, refcount_read(&s->s_ref));
3937                 return con;
3938         }
3939         dout("mdsc con_get %p FAIL\n", s);
3940         return NULL;
3941 }
3942
3943 static void con_put(struct ceph_connection *con)
3944 {
3945         struct ceph_mds_session *s = con->private;
3946
3947         dout("mdsc con_put %p (%d)\n", s, refcount_read(&s->s_ref) - 1);
3948         ceph_put_mds_session(s);
3949 }
3950
3951 /*
3952  * if the client is unresponsive for long enough, the mds will kill
3953  * the session entirely.
3954  */
3955 static void peer_reset(struct ceph_connection *con)
3956 {
3957         struct ceph_mds_session *s = con->private;
3958         struct ceph_mds_client *mdsc = s->s_mdsc;
3959
3960         pr_warn("mds%d closed our session\n", s->s_mds);
3961         send_mds_reconnect(mdsc, s);
3962 }
3963
3964 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3965 {
3966         struct ceph_mds_session *s = con->private;
3967         struct ceph_mds_client *mdsc = s->s_mdsc;
3968         int type = le16_to_cpu(msg->hdr.type);
3969
3970         mutex_lock(&mdsc->mutex);
3971         if (__verify_registered_session(mdsc, s) < 0) {
3972                 mutex_unlock(&mdsc->mutex);
3973                 goto out;
3974         }
3975         mutex_unlock(&mdsc->mutex);
3976
3977         switch (type) {
3978         case CEPH_MSG_MDS_MAP:
3979                 ceph_mdsc_handle_mdsmap(mdsc, msg);
3980                 break;
3981         case CEPH_MSG_FS_MAP_USER:
3982                 ceph_mdsc_handle_fsmap(mdsc, msg);
3983                 break;
3984         case CEPH_MSG_CLIENT_SESSION:
3985                 handle_session(s, msg);
3986                 break;
3987         case CEPH_MSG_CLIENT_REPLY:
3988                 handle_reply(s, msg);
3989                 break;
3990         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3991                 handle_forward(mdsc, s, msg);
3992                 break;
3993         case CEPH_MSG_CLIENT_CAPS:
3994                 ceph_handle_caps(s, msg);
3995                 break;
3996         case CEPH_MSG_CLIENT_SNAP:
3997                 ceph_handle_snap(mdsc, s, msg);
3998                 break;
3999         case CEPH_MSG_CLIENT_LEASE:
4000                 handle_lease(mdsc, s, msg);
4001                 break;
4002
4003         default:
4004                 pr_err("received unknown message type %d %s\n", type,
4005                        ceph_msg_type_name(type));
4006         }
4007 out:
4008         ceph_msg_put(msg);
4009 }
4010
4011 /*
4012  * authentication
4013  */
4014
4015 /*
4016  * Note: returned pointer is the address of a structure that's
4017  * managed separately.  Caller must *not* attempt to free it.
4018  */
4019 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
4020                                         int *proto, int force_new)
4021 {
4022         struct ceph_mds_session *s = con->private;
4023         struct ceph_mds_client *mdsc = s->s_mdsc;
4024         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4025         struct ceph_auth_handshake *auth = &s->s_auth;
4026
4027         if (force_new && auth->authorizer) {
4028                 ceph_auth_destroy_authorizer(auth->authorizer);
4029                 auth->authorizer = NULL;
4030         }
4031         if (!auth->authorizer) {
4032                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4033                                                       auth);
4034                 if (ret)
4035                         return ERR_PTR(ret);
4036         } else {
4037                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4038                                                       auth);
4039                 if (ret)
4040                         return ERR_PTR(ret);
4041         }
4042         *proto = ac->protocol;
4043
4044         return auth;
4045 }
4046
4047
4048 static int verify_authorizer_reply(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         return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
4055 }
4056
4057 static int invalidate_authorizer(struct ceph_connection *con)
4058 {
4059         struct ceph_mds_session *s = con->private;
4060         struct ceph_mds_client *mdsc = s->s_mdsc;
4061         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4062
4063         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
4064
4065         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
4066 }
4067
4068 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
4069                                 struct ceph_msg_header *hdr, int *skip)
4070 {
4071         struct ceph_msg *msg;
4072         int type = (int) le16_to_cpu(hdr->type);
4073         int front_len = (int) le32_to_cpu(hdr->front_len);
4074
4075         if (con->in_msg)
4076                 return con->in_msg;
4077
4078         *skip = 0;
4079         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
4080         if (!msg) {
4081                 pr_err("unable to allocate msg type %d len %d\n",
4082                        type, front_len);
4083                 return NULL;
4084         }
4085
4086         return msg;
4087 }
4088
4089 static int mds_sign_message(struct ceph_msg *msg)
4090 {
4091        struct ceph_mds_session *s = msg->con->private;
4092        struct ceph_auth_handshake *auth = &s->s_auth;
4093
4094        return ceph_auth_sign_message(auth, msg);
4095 }
4096
4097 static int mds_check_message_signature(struct ceph_msg *msg)
4098 {
4099        struct ceph_mds_session *s = msg->con->private;
4100        struct ceph_auth_handshake *auth = &s->s_auth;
4101
4102        return ceph_auth_check_message_signature(auth, msg);
4103 }
4104
4105 static const struct ceph_connection_operations mds_con_ops = {
4106         .get = con_get,
4107         .put = con_put,
4108         .dispatch = dispatch,
4109         .get_authorizer = get_authorizer,
4110         .verify_authorizer_reply = verify_authorizer_reply,
4111         .invalidate_authorizer = invalidate_authorizer,
4112         .peer_reset = peer_reset,
4113         .alloc_msg = mds_alloc_msg,
4114         .sign_message = mds_sign_message,
4115         .check_message_signature = mds_check_message_signature,
4116 };
4117
4118 /* eof */