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  44  * are no longer valid.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                          * fall-thru */
 739                         WARN_ON_ONCE(!req->r_dentry);
 740                 }
 741         }
 742         if (!inode && req->r_dentry) {
 743                 /* ignore race with rename; old or new d_parent is okay */
 744                 struct dentry *parent;
 745                 struct inode *dir;
 746
 747                 rcu_read_lock();
 748                 parent = req->r_dentry->d_parent;
 749                 dir = req->r_parent ? : d_inode_rcu(parent);
 750
 751                 if (!dir || dir->i_sb != mdsc->fsc->sb) {
 752                         /*  not this fs or parent went negative */
 753                         inode = d_inode(req->r_dentry);
 754                         if (inode)
 755                                 ihold(inode);
 756                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
 757                         /* direct snapped/virtual snapdir requests
 758                          * based on parent dir inode */
 759                         inode = get_nonsnap_parent(parent);
 760                         dout("__choose_mds using nonsnap parent %p\n", inode);
 761                 } else {
 762                         /* dentry target */
 763                         inode = d_inode(req->r_dentry);
 764                         if (!inode || mode == USE_AUTH_MDS) {
 765                                 /* dir + name */
 766                                 inode = igrab(dir);
 767                                 hash = ceph_dentry_hash(dir, req->r_dentry);
 768                                 is_hash = true;
 769                         } else {
 770                                 ihold(inode);
 771                         }
 772                 }
 773                 rcu_read_unlock();
 774         }
 775
 776         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 777              (int)hash, mode);
 778         if (!inode)
 779                 goto random;
 780         ci = ceph_inode(inode);
 781
 782         if (is_hash && S_ISDIR(inode->i_mode)) {
 783                 struct ceph_inode_frag frag;
 784                 int found;
 785
 786                 ceph_choose_frag(ci, hash, &frag, &found);
 787                 if (found) {
 788                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 789                                 u8 r;
 790
 791                                 /* choose a random replica */
 792                                 get_random_bytes(&r, 1);
 793                                 r %= frag.ndist;
 794                                 mds = frag.dist[r];
 795                                 dout("choose_mds %p %llx.%llx "
 796                                      "frag %u mds%d (%d/%d)\n",
 797                                      inode, ceph_vinop(inode),
 798                                      frag.frag, mds,
 799                                      (int)r, frag.ndist);
 800                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 801                                     CEPH_MDS_STATE_ACTIVE)
 802                                         goto out;
 803                         }
 804
 805                         /* since this file/dir wasn't known to be
 806                          * replicated, then we want to look for the
 807                          * authoritative mds. */
 808                         mode = USE_AUTH_MDS;
 809                         if (frag.mds >= 0) {
 810                                 /* choose auth mds */
 811                                 mds = frag.mds;
 812                                 dout("choose_mds %p %llx.%llx "
 813                                      "frag %u mds%d (auth)\n",
 814                                      inode, ceph_vinop(inode), frag.frag, mds);
 815                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 816                                     CEPH_MDS_STATE_ACTIVE)
 817                                         goto out;
 818                         }
 819                 }
 820         }
 821
 822         spin_lock(&ci->i_ceph_lock);
 823         cap = NULL;
 824         if (mode == USE_AUTH_MDS)
 825                 cap = ci->i_auth_cap;
 826         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
 827                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
 828         if (!cap) {
 829                 spin_unlock(&ci->i_ceph_lock);
 830                 iput(inode);
 831                 goto random;
 832         }
 833         mds = cap->session->s_mds;
 834         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
 835              inode, ceph_vinop(inode), mds,
 836              cap == ci->i_auth_cap ? "auth " : "", cap);
 837         spin_unlock(&ci->i_ceph_lock);
 838 out:
 839         iput(inode);
 840         return mds;
 841
 842 random:
 843         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
 844         dout("choose_mds chose random mds%d\n", mds);
 845         return mds;
 846 }
 847
 848
 849 /*
 850  * session messages
 851  */
 852 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
 853 {
 854         struct ceph_msg *msg;
 855         struct ceph_mds_session_head *h;
 856
 857         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
 858                            false);
 859         if (!msg) {
 860                 pr_err("create_session_msg ENOMEM creating msg\n");
 861                 return NULL;
 862         }
 863         h = msg->front.iov_base;
 864         h->op = cpu_to_le32(op);
 865         h->seq = cpu_to_le64(seq);
 866
 867         return msg;
 868 }
 869
 870 /*
 871  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
 872  * to include additional client metadata fields.
 873  */
 874 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
 875 {
 876         struct ceph_msg *msg;
 877         struct ceph_mds_session_head *h;
 878         int i = -1;
 879         int metadata_bytes = 0;
 880         int metadata_key_count = 0;
 881         struct ceph_options *opt = mdsc->fsc->client->options;
 882         struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
 883         void *p;
 884
 885         const char* metadata[][2] = {
 886                 {"hostname", mdsc->nodename},
 887                 {"kernel_version", init_utsname()->release},
 888                 {"entity_id", opt->name ? : ""},
 889                 {"root", fsopt->server_path ? : "/"},
 890                 {NULL, NULL}
 891         };
 892
 893         /* Calculate serialized length of metadata */
 894         metadata_bytes = 4;  /* map length */
 895         for (i = 0; metadata[i][0]; ++i) {
 896                 metadata_bytes += 8 + strlen(metadata[i][0]) +
 897                         strlen(metadata[i][1]);
 898                 metadata_key_count++;
 899         }
 900
 901         /* Allocate the message */
 902         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + metadata_bytes,
 903                            GFP_NOFS, false);
 904         if (!msg) {
 905                 pr_err("create_session_msg ENOMEM creating msg\n");
 906                 return NULL;
 907         }
 908         h = msg->front.iov_base;
 909         h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
 910         h->seq = cpu_to_le64(seq);
 911
 912         /*
 913          * Serialize client metadata into waiting buffer space, using
 914          * the format that userspace expects for map<string, string>
 915          *
 916          * ClientSession messages with metadata are v2
 917          */
 918         msg->hdr.version = cpu_to_le16(2);
 919         msg->hdr.compat_version = cpu_to_le16(1);
 920
 921         /* The write pointer, following the session_head structure */
 922         p = msg->front.iov_base + sizeof(*h);
 923
 924         /* Number of entries in the map */
 925         ceph_encode_32(&p, metadata_key_count);
 926
 927         /* Two length-prefixed strings for each entry in the map */
 928         for (i = 0; metadata[i][0]; ++i) {
 929                 size_t const key_len = strlen(metadata[i][0]);
 930                 size_t const val_len = strlen(metadata[i][1]);
 931
 932                 ceph_encode_32(&p, key_len);
 933                 memcpy(p, metadata[i][0], key_len);
 934                 p += key_len;
 935                 ceph_encode_32(&p, val_len);
 936                 memcpy(p, metadata[i][1], val_len);
 937                 p += val_len;
 938         }
 939
 940         return msg;
 941 }
 942
 943 /*
 944  * send session open request.
 945  *
 946  * called under mdsc->mutex
 947  */
 948 static int __open_session(struct ceph_mds_client *mdsc,
 949                           struct ceph_mds_session *session)
 950 {
 951         struct ceph_msg *msg;
 952         int mstate;
 953         int mds = session->s_mds;
 954
 955         /* wait for mds to go active? */
 956         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
 957         dout("open_session to mds%d (%s)\n", mds,
 958              ceph_mds_state_name(mstate));
 959         session->s_state = CEPH_MDS_SESSION_OPENING;
 960         session->s_renew_requested = jiffies;
 961
 962         /* send connect message */
 963         msg = create_session_open_msg(mdsc, session->s_seq);
 964         if (!msg)
 965                 return -ENOMEM;
 966         ceph_con_send(&session->s_con, msg);
 967         return 0;
 968 }
 969
 970 /*
 971  * open sessions for any export targets for the given mds
 972  *
 973  * called under mdsc->mutex
 974  */
 975 static struct ceph_mds_session *
 976 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
 977 {
 978         struct ceph_mds_session *session;
 979
 980         session = __ceph_lookup_mds_session(mdsc, target);
 981         if (!session) {
 982                 session = register_session(mdsc, target);
 983                 if (IS_ERR(session))
 984                         return session;
 985         }
 986         if (session->s_state == CEPH_MDS_SESSION_NEW ||
 987             session->s_state == CEPH_MDS_SESSION_CLOSING)
 988                 __open_session(mdsc, session);
 989
 990         return session;
 991 }
 992
 993 struct ceph_mds_session *
 994 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
 995 {
 996         struct ceph_mds_session *session;
 997
 998         dout("open_export_target_session to mds%d\n", target);
 999
1000         mutex_lock(&mdsc->mutex);
1001         session = __open_export_target_session(mdsc, target);
1002         mutex_unlock(&mdsc->mutex);
1003
1004         return session;
1005 }
1006
1007 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1008                                           struct ceph_mds_session *session)
1009 {
1010         struct ceph_mds_info *mi;
1011         struct ceph_mds_session *ts;
1012         int i, mds = session->s_mds;
1013
1014         if (mds >= mdsc->mdsmap->m_num_mds)
1015                 return;
1016
1017         mi = &mdsc->mdsmap->m_info[mds];
1018         dout("open_export_target_sessions for mds%d (%d targets)\n",
1019              session->s_mds, mi->num_export_targets);
1020
1021         for (i = 0; i < mi->num_export_targets; i++) {
1022                 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1023                 if (!IS_ERR(ts))
1024                         ceph_put_mds_session(ts);
1025         }
1026 }
1027
1028 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1029                                            struct ceph_mds_session *session)
1030 {
1031         mutex_lock(&mdsc->mutex);
1032         __open_export_target_sessions(mdsc, session);
1033         mutex_unlock(&mdsc->mutex);
1034 }
1035
1036 /*
1037  * session caps
1038  */
1039
1040 /* caller holds s_cap_lock, we drop it */
1041 static void cleanup_cap_releases(struct ceph_mds_client *mdsc,
1042                                  struct ceph_mds_session *session)
1043         __releases(session->s_cap_lock)
1044 {
1045         LIST_HEAD(tmp_list);
1046         list_splice_init(&session->s_cap_releases, &tmp_list);
1047         session->s_num_cap_releases = 0;
1048         spin_unlock(&session->s_cap_lock);
1049
1050         dout("cleanup_cap_releases mds%d\n", session->s_mds);
1051         while (!list_empty(&tmp_list)) {
1052                 struct ceph_cap *cap;
1053                 /* zero out the in-progress message */
1054                 cap = list_first_entry(&tmp_list,
1055                                         struct ceph_cap, session_caps);
1056                 list_del(&cap->session_caps);
1057                 ceph_put_cap(mdsc, cap);
1058         }
1059 }
1060
1061 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1062                                      struct ceph_mds_session *session)
1063 {
1064         struct ceph_mds_request *req;
1065         struct rb_node *p;
1066
1067         dout("cleanup_session_requests mds%d\n", session->s_mds);
1068         mutex_lock(&mdsc->mutex);
1069         while (!list_empty(&session->s_unsafe)) {
1070                 req = list_first_entry(&session->s_unsafe,
1071                                        struct ceph_mds_request, r_unsafe_item);
1072                 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1073                                     req->r_tid);
1074                 __unregister_request(mdsc, req);
1075         }
1076         /* zero r_attempts, so kick_requests() will re-send requests */
1077         p = rb_first(&mdsc->request_tree);
1078         while (p) {
1079                 req = rb_entry(p, struct ceph_mds_request, r_node);
1080                 p = rb_next(p);
1081                 if (req->r_session &&
1082                     req->r_session->s_mds == session->s_mds)
1083                         req->r_attempts = 0;
1084         }
1085         mutex_unlock(&mdsc->mutex);
1086 }
1087
1088 /*
1089  * Helper to safely iterate over all caps associated with a session, with
1090  * special care taken to handle a racing __ceph_remove_cap().
1091  *
1092  * Caller must hold session s_mutex.
1093  */
1094 static int iterate_session_caps(struct ceph_mds_session *session,
1095                                  int (*cb)(struct inode *, struct ceph_cap *,
1096                                             void *), void *arg)
1097 {
1098         struct list_head *p;
1099         struct ceph_cap *cap;
1100         struct inode *inode, *last_inode = NULL;
1101         struct ceph_cap *old_cap = NULL;
1102         int ret;
1103
1104         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1105         spin_lock(&session->s_cap_lock);
1106         p = session->s_caps.next;
1107         while (p != &session->s_caps) {
1108                 cap = list_entry(p, struct ceph_cap, session_caps);
1109                 inode = igrab(&cap->ci->vfs_inode);
1110                 if (!inode) {
1111                         p = p->next;
1112                         continue;
1113                 }
1114                 session->s_cap_iterator = cap;
1115                 spin_unlock(&session->s_cap_lock);
1116
1117                 if (last_inode) {
1118                         iput(last_inode);
1119                         last_inode = NULL;
1120                 }
1121                 if (old_cap) {
1122                         ceph_put_cap(session->s_mdsc, old_cap);
1123                         old_cap = NULL;
1124                 }
1125
1126                 ret = cb(inode, cap, arg);
1127                 last_inode = inode;
1128
1129                 spin_lock(&session->s_cap_lock);
1130                 p = p->next;
1131                 if (!cap->ci) {
1132                         dout("iterate_session_caps  finishing cap %p removal\n",
1133                              cap);
1134                         BUG_ON(cap->session != session);
1135                         cap->session = NULL;
1136                         list_del_init(&cap->session_caps);
1137                         session->s_nr_caps--;
1138                         if (cap->queue_release) {
1139                                 list_add_tail(&cap->session_caps,
1140                                               &session->s_cap_releases);
1141                                 session->s_num_cap_releases++;
1142                         } else {
1143                                 old_cap = cap;  /* put_cap it w/o locks held */
1144                         }
1145                 }
1146                 if (ret < 0)
1147                         goto out;
1148         }
1149         ret = 0;
1150 out:
1151         session->s_cap_iterator = NULL;
1152         spin_unlock(&session->s_cap_lock);
1153
1154         iput(last_inode);
1155         if (old_cap)
1156                 ceph_put_cap(session->s_mdsc, old_cap);
1157
1158         return ret;
1159 }
1160
1161 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1162                                   void *arg)
1163 {
1164         struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1165         struct ceph_inode_info *ci = ceph_inode(inode);
1166         LIST_HEAD(to_remove);
1167         bool drop = false;
1168         bool invalidate = false;
1169
1170         dout("removing cap %p, ci is %p, inode is %p\n",
1171              cap, ci, &ci->vfs_inode);
1172         spin_lock(&ci->i_ceph_lock);
1173         __ceph_remove_cap(cap, false);
1174         if (!ci->i_auth_cap) {
1175                 struct ceph_cap_flush *cf;
1176                 struct ceph_mds_client *mdsc = fsc->mdsc;
1177
1178                 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1179
1180                 if (ci->i_wrbuffer_ref > 0 &&
1181                     READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
1182                         invalidate = true;
1183
1184                 while (!list_empty(&ci->i_cap_flush_list)) {
1185                         cf = list_first_entry(&ci->i_cap_flush_list,
1186                                               struct ceph_cap_flush, i_list);
1187                         list_move(&cf->i_list, &to_remove);
1188                 }
1189
1190                 spin_lock(&mdsc->cap_dirty_lock);
1191
1192                 list_for_each_entry(cf, &to_remove, i_list)
1193                         list_del(&cf->g_list);
1194
1195                 if (!list_empty(&ci->i_dirty_item)) {
1196                         pr_warn_ratelimited(
1197                                 " dropping dirty %s state for %p %lld\n",
1198                                 ceph_cap_string(ci->i_dirty_caps),
1199                                 inode, ceph_ino(inode));
1200                         ci->i_dirty_caps = 0;
1201                         list_del_init(&ci->i_dirty_item);
1202                         drop = true;
1203                 }
1204                 if (!list_empty(&ci->i_flushing_item)) {
1205                         pr_warn_ratelimited(
1206                                 " dropping dirty+flushing %s state for %p %lld\n",
1207                                 ceph_cap_string(ci->i_flushing_caps),
1208                                 inode, ceph_ino(inode));
1209                         ci->i_flushing_caps = 0;
1210                         list_del_init(&ci->i_flushing_item);
1211                         mdsc->num_cap_flushing--;
1212                         drop = true;
1213                 }
1214                 spin_unlock(&mdsc->cap_dirty_lock);
1215
1216                 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1217                         list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1218                         ci->i_prealloc_cap_flush = NULL;
1219                 }
1220         }
1221         spin_unlock(&ci->i_ceph_lock);
1222         while (!list_empty(&to_remove)) {
1223                 struct ceph_cap_flush *cf;
1224                 cf = list_first_entry(&to_remove,
1225                                       struct ceph_cap_flush, i_list);
1226                 list_del(&cf->i_list);
1227                 ceph_free_cap_flush(cf);
1228         }
1229
1230         wake_up_all(&ci->i_cap_wq);
1231         if (invalidate)
1232                 ceph_queue_invalidate(inode);
1233         if (drop)
1234                 iput(inode);
1235         return 0;
1236 }
1237
1238 /*
1239  * caller must hold session s_mutex
1240  */
1241 static void remove_session_caps(struct ceph_mds_session *session)
1242 {
1243         struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1244         struct super_block *sb = fsc->sb;
1245         dout("remove_session_caps on %p\n", session);
1246         iterate_session_caps(session, remove_session_caps_cb, fsc);
1247
1248         wake_up_all(&fsc->mdsc->cap_flushing_wq);
1249
1250         spin_lock(&session->s_cap_lock);
1251         if (session->s_nr_caps > 0) {
1252                 struct inode *inode;
1253                 struct ceph_cap *cap, *prev = NULL;
1254                 struct ceph_vino vino;
1255                 /*
1256                  * iterate_session_caps() skips inodes that are being
1257                  * deleted, we need to wait until deletions are complete.
1258                  * __wait_on_freeing_inode() is designed for the job,
1259                  * but it is not exported, so use lookup inode function
1260                  * to access it.
1261                  */
1262                 while (!list_empty(&session->s_caps)) {
1263                         cap = list_entry(session->s_caps.next,
1264                                          struct ceph_cap, session_caps);
1265                         if (cap == prev)
1266                                 break;
1267                         prev = cap;
1268                         vino = cap->ci->i_vino;
1269                         spin_unlock(&session->s_cap_lock);
1270
1271                         inode = ceph_find_inode(sb, vino);
1272                         iput(inode);
1273
1274                         spin_lock(&session->s_cap_lock);
1275                 }
1276         }
1277
1278         // drop cap expires and unlock s_cap_lock
1279         cleanup_cap_releases(session->s_mdsc, session);
1280
1281         BUG_ON(session->s_nr_caps > 0);
1282         BUG_ON(!list_empty(&session->s_cap_flushing));
1283 }
1284
1285 /*
1286  * wake up any threads waiting on this session's caps.  if the cap is
1287  * old (didn't get renewed on the client reconnect), remove it now.
1288  *
1289  * caller must hold s_mutex.
1290  */
1291 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1292                               void *arg)
1293 {
1294         struct ceph_inode_info *ci = ceph_inode(inode);
1295
1296         if (arg) {
1297                 spin_lock(&ci->i_ceph_lock);
1298                 ci->i_wanted_max_size = 0;
1299                 ci->i_requested_max_size = 0;
1300                 spin_unlock(&ci->i_ceph_lock);
1301         }
1302         wake_up_all(&ci->i_cap_wq);
1303         return 0;
1304 }
1305
1306 static void wake_up_session_caps(struct ceph_mds_session *session,
1307                                  int reconnect)
1308 {
1309         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1310         iterate_session_caps(session, wake_up_session_cb,
1311                              (void *)(unsigned long)reconnect);
1312 }
1313
1314 /*
1315  * Send periodic message to MDS renewing all currently held caps.  The
1316  * ack will reset the expiration for all caps from this session.
1317  *
1318  * caller holds s_mutex
1319  */
1320 static int send_renew_caps(struct ceph_mds_client *mdsc,
1321                            struct ceph_mds_session *session)
1322 {
1323         struct ceph_msg *msg;
1324         int state;
1325
1326         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1327             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1328                 pr_info("mds%d caps stale\n", session->s_mds);
1329         session->s_renew_requested = jiffies;
1330
1331         /* do not try to renew caps until a recovering mds has reconnected
1332          * with its clients. */
1333         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1334         if (state < CEPH_MDS_STATE_RECONNECT) {
1335                 dout("send_renew_caps ignoring mds%d (%s)\n",
1336                      session->s_mds, ceph_mds_state_name(state));
1337                 return 0;
1338         }
1339
1340         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1341                 ceph_mds_state_name(state));
1342         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1343                                  ++session->s_renew_seq);
1344         if (!msg)
1345                 return -ENOMEM;
1346         ceph_con_send(&session->s_con, msg);
1347         return 0;
1348 }
1349
1350 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1351                              struct ceph_mds_session *session, u64 seq)
1352 {
1353         struct ceph_msg *msg;
1354
1355         dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1356              session->s_mds, ceph_session_state_name(session->s_state), seq);
1357         msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1358         if (!msg)
1359                 return -ENOMEM;
1360         ceph_con_send(&session->s_con, msg);
1361         return 0;
1362 }
1363
1364
1365 /*
1366  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1367  *
1368  * Called under session->s_mutex
1369  */
1370 static void renewed_caps(struct ceph_mds_client *mdsc,
1371                          struct ceph_mds_session *session, int is_renew)
1372 {
1373         int was_stale;
1374         int wake = 0;
1375
1376         spin_lock(&session->s_cap_lock);
1377         was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1378
1379         session->s_cap_ttl = session->s_renew_requested +
1380                 mdsc->mdsmap->m_session_timeout*HZ;
1381
1382         if (was_stale) {
1383                 if (time_before(jiffies, session->s_cap_ttl)) {
1384                         pr_info("mds%d caps renewed\n", session->s_mds);
1385                         wake = 1;
1386                 } else {
1387                         pr_info("mds%d caps still stale\n", session->s_mds);
1388                 }
1389         }
1390         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1391              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1392              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1393         spin_unlock(&session->s_cap_lock);
1394
1395         if (wake)
1396                 wake_up_session_caps(session, 0);
1397 }
1398
1399 /*
1400  * send a session close request
1401  */
1402 static int request_close_session(struct ceph_mds_client *mdsc,
1403                                  struct ceph_mds_session *session)
1404 {
1405         struct ceph_msg *msg;
1406
1407         dout("request_close_session mds%d state %s seq %lld\n",
1408              session->s_mds, ceph_session_state_name(session->s_state),
1409              session->s_seq);
1410         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1411         if (!msg)
1412                 return -ENOMEM;
1413         ceph_con_send(&session->s_con, msg);
1414         return 1;
1415 }
1416
1417 /*
1418  * Called with s_mutex held.
1419  */
1420 static int __close_session(struct ceph_mds_client *mdsc,
1421                          struct ceph_mds_session *session)
1422 {
1423         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1424                 return 0;
1425         session->s_state = CEPH_MDS_SESSION_CLOSING;
1426         return request_close_session(mdsc, session);
1427 }
1428
1429 /*
1430  * Trim old(er) caps.
1431  *
1432  * Because we can't cache an inode without one or more caps, we do
1433  * this indirectly: if a cap is unused, we prune its aliases, at which
1434  * point the inode will hopefully get dropped to.
1435  *
1436  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1437  * memory pressure from the MDS, though, so it needn't be perfect.
1438  */
1439 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1440 {
1441         struct ceph_mds_session *session = arg;
1442         struct ceph_inode_info *ci = ceph_inode(inode);
1443         int used, wanted, oissued, mine;
1444
1445         if (session->s_trim_caps <= 0)
1446                 return -1;
1447
1448         spin_lock(&ci->i_ceph_lock);
1449         mine = cap->issued | cap->implemented;
1450         used = __ceph_caps_used(ci);
1451         wanted = __ceph_caps_file_wanted(ci);
1452         oissued = __ceph_caps_issued_other(ci, cap);
1453
1454         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1455              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1456              ceph_cap_string(used), ceph_cap_string(wanted));
1457         if (cap == ci->i_auth_cap) {
1458                 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1459                     !list_empty(&ci->i_cap_snaps))
1460                         goto out;
1461                 if ((used | wanted) & CEPH_CAP_ANY_WR)
1462                         goto out;
1463         }
1464         /* The inode has cached pages, but it's no longer used.
1465          * we can safely drop it */
1466         if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1467             !(oissued & CEPH_CAP_FILE_CACHE)) {
1468           used = 0;
1469           oissued = 0;
1470         }
1471         if ((used | wanted) & ~oissued & mine)
1472                 goto out;   /* we need these caps */
1473
1474         session->s_trim_caps--;
1475         if (oissued) {
1476                 /* we aren't the only cap.. just remove us */
1477                 __ceph_remove_cap(cap, true);
1478         } else {
1479                 /* try dropping referring dentries */
1480                 spin_unlock(&ci->i_ceph_lock);
1481                 d_prune_aliases(inode);
1482                 dout("trim_caps_cb %p cap %p  pruned, count now %d\n",
1483                      inode, cap, atomic_read(&inode->i_count));
1484                 return 0;
1485         }
1486
1487 out:
1488         spin_unlock(&ci->i_ceph_lock);
1489         return 0;
1490 }
1491
1492 /*
1493  * Trim session cap count down to some max number.
1494  */
1495 static int trim_caps(struct ceph_mds_client *mdsc,
1496                      struct ceph_mds_session *session,
1497                      int max_caps)
1498 {
1499         int trim_caps = session->s_nr_caps - max_caps;
1500
1501         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1502              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1503         if (trim_caps > 0) {
1504                 session->s_trim_caps = trim_caps;
1505                 iterate_session_caps(session, trim_caps_cb, session);
1506                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1507                      session->s_mds, session->s_nr_caps, max_caps,
1508                         trim_caps - session->s_trim_caps);
1509                 session->s_trim_caps = 0;
1510         }
1511
1512         ceph_send_cap_releases(mdsc, session);
1513         return 0;
1514 }
1515
1516 static int check_caps_flush(struct ceph_mds_client *mdsc,
1517                             u64 want_flush_tid)
1518 {
1519         int ret = 1;
1520
1521         spin_lock(&mdsc->cap_dirty_lock);
1522         if (!list_empty(&mdsc->cap_flush_list)) {
1523                 struct ceph_cap_flush *cf =
1524                         list_first_entry(&mdsc->cap_flush_list,
1525                                          struct ceph_cap_flush, g_list);
1526                 if (cf->tid <= want_flush_tid) {
1527                         dout("check_caps_flush still flushing tid "
1528                              "%llu <= %llu\n", cf->tid, want_flush_tid);
1529                         ret = 0;
1530                 }
1531         }
1532         spin_unlock(&mdsc->cap_dirty_lock);
1533         return ret;
1534 }
1535
1536 /*
1537  * flush all dirty inode data to disk.
1538  *
1539  * returns true if we've flushed through want_flush_tid
1540  */
1541 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1542                             u64 want_flush_tid)
1543 {
1544         dout("check_caps_flush want %llu\n", want_flush_tid);
1545
1546         wait_event(mdsc->cap_flushing_wq,
1547                    check_caps_flush(mdsc, want_flush_tid));
1548
1549         dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1550 }
1551
1552 /*
1553  * called under s_mutex
1554  */
1555 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1556                             struct ceph_mds_session *session)
1557 {
1558         struct ceph_msg *msg = NULL;
1559         struct ceph_mds_cap_release *head;
1560         struct ceph_mds_cap_item *item;
1561         struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
1562         struct ceph_cap *cap;
1563         LIST_HEAD(tmp_list);
1564         int num_cap_releases;
1565         __le32  barrier, *cap_barrier;
1566
1567         down_read(&osdc->lock);
1568         barrier = cpu_to_le32(osdc->epoch_barrier);
1569         up_read(&osdc->lock);
1570
1571         spin_lock(&session->s_cap_lock);
1572 again:
1573         list_splice_init(&session->s_cap_releases, &tmp_list);
1574         num_cap_releases = session->s_num_cap_releases;
1575         session->s_num_cap_releases = 0;
1576         spin_unlock(&session->s_cap_lock);
1577
1578         while (!list_empty(&tmp_list)) {
1579                 if (!msg) {
1580                         msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1581                                         PAGE_SIZE, GFP_NOFS, false);
1582                         if (!msg)
1583                                 goto out_err;
1584                         head = msg->front.iov_base;
1585                         head->num = cpu_to_le32(0);
1586                         msg->front.iov_len = sizeof(*head);
1587
1588                         msg->hdr.version = cpu_to_le16(2);
1589                         msg->hdr.compat_version = cpu_to_le16(1);
1590                 }
1591
1592                 cap = list_first_entry(&tmp_list, struct ceph_cap,
1593                                         session_caps);
1594                 list_del(&cap->session_caps);
1595                 num_cap_releases--;
1596
1597                 head = msg->front.iov_base;
1598                 le32_add_cpu(&head->num, 1);
1599                 item = msg->front.iov_base + msg->front.iov_len;
1600                 item->ino = cpu_to_le64(cap->cap_ino);
1601                 item->cap_id = cpu_to_le64(cap->cap_id);
1602                 item->migrate_seq = cpu_to_le32(cap->mseq);
1603                 item->seq = cpu_to_le32(cap->issue_seq);
1604                 msg->front.iov_len += sizeof(*item);
1605
1606                 ceph_put_cap(mdsc, cap);
1607
1608                 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1609                         // Append cap_barrier field
1610                         cap_barrier = msg->front.iov_base + msg->front.iov_len;
1611                         *cap_barrier = barrier;
1612                         msg->front.iov_len += sizeof(*cap_barrier);
1613
1614                         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1615                         dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1616                         ceph_con_send(&session->s_con, msg);
1617                         msg = NULL;
1618                 }
1619         }
1620
1621         BUG_ON(num_cap_releases != 0);
1622
1623         spin_lock(&session->s_cap_lock);
1624         if (!list_empty(&session->s_cap_releases))
1625                 goto again;
1626         spin_unlock(&session->s_cap_lock);
1627
1628         if (msg) {
1629                 // Append cap_barrier field
1630                 cap_barrier = msg->front.iov_base + msg->front.iov_len;
1631                 *cap_barrier = barrier;
1632                 msg->front.iov_len += sizeof(*cap_barrier);
1633
1634                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1635                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1636                 ceph_con_send(&session->s_con, msg);
1637         }
1638         return;
1639 out_err:
1640         pr_err("send_cap_releases mds%d, failed to allocate message\n",
1641                 session->s_mds);
1642         spin_lock(&session->s_cap_lock);
1643         list_splice(&tmp_list, &session->s_cap_releases);
1644         session->s_num_cap_releases += num_cap_releases;
1645         spin_unlock(&session->s_cap_lock);
1646 }
1647
1648 /*
1649  * requests
1650  */
1651
1652 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1653                                     struct inode *dir)
1654 {
1655         struct ceph_inode_info *ci = ceph_inode(dir);
1656         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1657         struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1658         size_t size = sizeof(struct ceph_mds_reply_dir_entry);
1659         int order, num_entries;
1660
1661         spin_lock(&ci->i_ceph_lock);
1662         num_entries = ci->i_files + ci->i_subdirs;
1663         spin_unlock(&ci->i_ceph_lock);
1664         num_entries = max(num_entries, 1);
1665         num_entries = min(num_entries, opt->max_readdir);
1666
1667         order = get_order(size * num_entries);
1668         while (order >= 0) {
1669                 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
1670                                                              __GFP_NOWARN,
1671                                                              order);
1672                 if (rinfo->dir_entries)
1673                         break;
1674                 order--;
1675         }
1676         if (!rinfo->dir_entries)
1677                 return -ENOMEM;
1678
1679         num_entries = (PAGE_SIZE << order) / size;
1680         num_entries = min(num_entries, opt->max_readdir);
1681
1682         rinfo->dir_buf_size = PAGE_SIZE << order;
1683         req->r_num_caps = num_entries + 1;
1684         req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1685         req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1686         return 0;
1687 }
1688
1689 /*
1690  * Create an mds request.
1691  */
1692 struct ceph_mds_request *
1693 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1694 {
1695         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1696
1697         if (!req)
1698                 return ERR_PTR(-ENOMEM);
1699
1700         mutex_init(&req->r_fill_mutex);
1701         req->r_mdsc = mdsc;
1702         req->r_started = jiffies;
1703         req->r_resend_mds = -1;
1704         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1705         INIT_LIST_HEAD(&req->r_unsafe_target_item);
1706         req->r_fmode = -1;
1707         kref_init(&req->r_kref);
1708         RB_CLEAR_NODE(&req->r_node);
1709         INIT_LIST_HEAD(&req->r_wait);
1710         init_completion(&req->r_completion);
1711         init_completion(&req->r_safe_completion);
1712         INIT_LIST_HEAD(&req->r_unsafe_item);
1713
1714         req->r_stamp = timespec_trunc(current_kernel_time(), mdsc->fsc->sb->s_time_gran);
1715
1716         req->r_op = op;
1717         req->r_direct_mode = mode;
1718         return req;
1719 }
1720
1721 /*
1722  * return oldest (lowest) request, tid in request tree, 0 if none.
1723  *
1724  * called under mdsc->mutex.
1725  */
1726 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1727 {
1728         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1729                 return NULL;
1730         return rb_entry(rb_first(&mdsc->request_tree),
1731                         struct ceph_mds_request, r_node);
1732 }
1733
1734 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1735 {
1736         return mdsc->oldest_tid;
1737 }
1738
1739 /*
1740  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1741  * on build_path_from_dentry in fs/cifs/dir.c.
1742  *
1743  * If @stop_on_nosnap, generate path relative to the first non-snapped
1744  * inode.
1745  *
1746  * Encode hidden .snap dirs as a double /, i.e.
1747  *   foo/.snap/bar -> foo//bar
1748  */
1749 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1750                            int stop_on_nosnap)
1751 {
1752         struct dentry *temp;
1753         char *path;
1754         int len, pos;
1755         unsigned seq;
1756
1757         if (!dentry)
1758                 return ERR_PTR(-EINVAL);
1759
1760 retry:
1761         len = 0;
1762         seq = read_seqbegin(&rename_lock);
1763         rcu_read_lock();
1764         for (temp = dentry; !IS_ROOT(temp);) {
1765                 struct inode *inode = d_inode(temp);
1766                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1767                         len++;  /* slash only */
1768                 else if (stop_on_nosnap && inode &&
1769                          ceph_snap(inode) == CEPH_NOSNAP)
1770                         break;
1771                 else
1772                         len += 1 + temp->d_name.len;
1773                 temp = temp->d_parent;
1774         }
1775         rcu_read_unlock();
1776         if (len)
1777                 len--;  /* no leading '/' */
1778
1779         path = kmalloc(len+1, GFP_NOFS);
1780         if (!path)
1781                 return ERR_PTR(-ENOMEM);
1782         pos = len;
1783         path[pos] = 0;  /* trailing null */
1784         rcu_read_lock();
1785         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1786                 struct inode *inode;
1787
1788                 spin_lock(&temp->d_lock);
1789                 inode = d_inode(temp);
1790                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1791                         dout("build_path path+%d: %p SNAPDIR\n",
1792                              pos, temp);
1793                 } else if (stop_on_nosnap && inode &&
1794                            ceph_snap(inode) == CEPH_NOSNAP) {
1795                         spin_unlock(&temp->d_lock);
1796                         break;
1797                 } else {
1798                         pos -= temp->d_name.len;
1799                         if (pos < 0) {
1800                                 spin_unlock(&temp->d_lock);
1801                                 break;
1802                         }
1803                         strncpy(path + pos, temp->d_name.name,
1804                                 temp->d_name.len);
1805                 }
1806                 spin_unlock(&temp->d_lock);
1807                 if (pos)
1808                         path[--pos] = '/';
1809                 temp = temp->d_parent;
1810         }
1811         rcu_read_unlock();
1812         if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1813                 pr_err("build_path did not end path lookup where "
1814                        "expected, namelen is %d, pos is %d\n", len, pos);
1815                 /* presumably this is only possible if racing with a
1816                    rename of one of the parent directories (we can not
1817                    lock the dentries above us to prevent this, but
1818                    retrying should be harmless) */
1819                 kfree(path);
1820                 goto retry;
1821         }
1822
1823         *base = ceph_ino(d_inode(temp));
1824         *plen = len;
1825         dout("build_path on %p %d built %llx '%.*s'\n",
1826              dentry, d_count(dentry), *base, len, path);
1827         return path;
1828 }
1829
1830 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
1831                              const char **ppath, int *ppathlen, u64 *pino,
1832                              int *pfreepath)
1833 {
1834         char *path;
1835
1836         rcu_read_lock();
1837         if (!dir)
1838                 dir = d_inode_rcu(dentry->d_parent);
1839         if (dir && ceph_snap(dir) == CEPH_NOSNAP) {
1840                 *pino = ceph_ino(dir);
1841                 rcu_read_unlock();
1842                 *ppath = dentry->d_name.name;
1843                 *ppathlen = dentry->d_name.len;
1844                 return 0;
1845         }
1846         rcu_read_unlock();
1847         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1848         if (IS_ERR(path))
1849                 return PTR_ERR(path);
1850         *ppath = path;
1851         *pfreepath = 1;
1852         return 0;
1853 }
1854
1855 static int build_inode_path(struct inode *inode,
1856                             const char **ppath, int *ppathlen, u64 *pino,
1857                             int *pfreepath)
1858 {
1859         struct dentry *dentry;
1860         char *path;
1861
1862         if (ceph_snap(inode) == CEPH_NOSNAP) {
1863                 *pino = ceph_ino(inode);
1864                 *ppathlen = 0;
1865                 return 0;
1866         }
1867         dentry = d_find_alias(inode);
1868         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1869         dput(dentry);
1870         if (IS_ERR(path))
1871                 return PTR_ERR(path);
1872         *ppath = path;
1873         *pfreepath = 1;
1874         return 0;
1875 }
1876
1877 /*
1878  * request arguments may be specified via an inode *, a dentry *, or
1879  * an explicit ino+path.
1880  */
1881 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1882                                   struct inode *rdiri, const char *rpath,
1883                                   u64 rino, const char **ppath, int *pathlen,
1884                                   u64 *ino, int *freepath)
1885 {
1886         int r = 0;
1887
1888         if (rinode) {
1889                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1890                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1891                      ceph_snap(rinode));
1892         } else if (rdentry) {
1893                 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
1894                                         freepath);
1895                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1896                      *ppath);
1897         } else if (rpath || rino) {
1898                 *ino = rino;
1899                 *ppath = rpath;
1900                 *pathlen = rpath ? strlen(rpath) : 0;
1901                 dout(" path %.*s\n", *pathlen, rpath);
1902         }
1903
1904         return r;
1905 }
1906
1907 /*
1908  * called under mdsc->mutex
1909  */
1910 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1911                                                struct ceph_mds_request *req,
1912                                                int mds, bool drop_cap_releases)
1913 {
1914         struct ceph_msg *msg;
1915         struct ceph_mds_request_head *head;
1916         const char *path1 = NULL;
1917         const char *path2 = NULL;
1918         u64 ino1 = 0, ino2 = 0;
1919         int pathlen1 = 0, pathlen2 = 0;
1920         int freepath1 = 0, freepath2 = 0;
1921         int len;
1922         u16 releases;
1923         void *p, *end;
1924         int ret;
1925
1926         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1927                               req->r_parent, req->r_path1, req->r_ino1.ino,
1928                               &path1, &pathlen1, &ino1, &freepath1);
1929         if (ret < 0) {
1930                 msg = ERR_PTR(ret);
1931                 goto out;
1932         }
1933
1934         ret = set_request_path_attr(NULL, req->r_old_dentry,
1935                               req->r_old_dentry_dir,
1936                               req->r_path2, req->r_ino2.ino,
1937                               &path2, &pathlen2, &ino2, &freepath2);
1938         if (ret < 0) {
1939                 msg = ERR_PTR(ret);
1940                 goto out_free1;
1941         }
1942
1943         len = sizeof(*head) +
1944                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
1945                 sizeof(struct ceph_timespec);
1946
1947         /* calculate (max) length for cap releases */
1948         len += sizeof(struct ceph_mds_request_release) *
1949                 (!!req->r_inode_drop + !!req->r_dentry_drop +
1950                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1951         if (req->r_dentry_drop)
1952                 len += req->r_dentry->d_name.len;
1953         if (req->r_old_dentry_drop)
1954                 len += req->r_old_dentry->d_name.len;
1955
1956         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1957         if (!msg) {
1958                 msg = ERR_PTR(-ENOMEM);
1959                 goto out_free2;
1960         }
1961
1962         msg->hdr.version = cpu_to_le16(2);
1963         msg->hdr.tid = cpu_to_le64(req->r_tid);
1964
1965         head = msg->front.iov_base;
1966         p = msg->front.iov_base + sizeof(*head);
1967         end = msg->front.iov_base + msg->front.iov_len;
1968
1969         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1970         head->op = cpu_to_le32(req->r_op);
1971         head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1972         head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1973         head->args = req->r_args;
1974
1975         ceph_encode_filepath(&p, end, ino1, path1);
1976         ceph_encode_filepath(&p, end, ino2, path2);
1977
1978         /* make note of release offset, in case we need to replay */
1979         req->r_request_release_offset = p - msg->front.iov_base;
1980
1981         /* cap releases */
1982         releases = 0;
1983         if (req->r_inode_drop)
1984                 releases += ceph_encode_inode_release(&p,
1985                       req->r_inode ? req->r_inode : d_inode(req->r_dentry),
1986                       mds, req->r_inode_drop, req->r_inode_unless, 0);
1987         if (req->r_dentry_drop)
1988                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1989                                 req->r_parent, mds, req->r_dentry_drop,
1990                                 req->r_dentry_unless);
1991         if (req->r_old_dentry_drop)
1992                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1993                                 req->r_old_dentry_dir, mds,
1994                                 req->r_old_dentry_drop,
1995                                 req->r_old_dentry_unless);
1996         if (req->r_old_inode_drop)
1997                 releases += ceph_encode_inode_release(&p,
1998                       d_inode(req->r_old_dentry),
1999                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2000
2001         if (drop_cap_releases) {
2002                 releases = 0;
2003                 p = msg->front.iov_base + req->r_request_release_offset;
2004         }
2005
2006         head->num_releases = cpu_to_le16(releases);
2007
2008         /* time stamp */
2009         {
2010                 struct ceph_timespec ts;
2011                 ceph_encode_timespec(&ts, &req->r_stamp);
2012                 ceph_encode_copy(&p, &ts, sizeof(ts));
2013         }
2014
2015         BUG_ON(p > end);
2016         msg->front.iov_len = p - msg->front.iov_base;
2017         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2018
2019         if (req->r_pagelist) {
2020                 struct ceph_pagelist *pagelist = req->r_pagelist;
2021                 refcount_inc(&pagelist->refcnt);
2022                 ceph_msg_data_add_pagelist(msg, pagelist);
2023                 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2024         } else {
2025                 msg->hdr.data_len = 0;
2026         }
2027
2028         msg->hdr.data_off = cpu_to_le16(0);
2029
2030 out_free2:
2031         if (freepath2)
2032                 kfree((char *)path2);
2033 out_free1:
2034         if (freepath1)
2035                 kfree((char *)path1);
2036 out:
2037         return msg;
2038 }
2039
2040 /*
2041  * called under mdsc->mutex if error, under no mutex if
2042  * success.
2043  */
2044 static void complete_request(struct ceph_mds_client *mdsc,
2045                              struct ceph_mds_request *req)
2046 {
2047         if (req->r_callback)
2048                 req->r_callback(mdsc, req);
2049         else
2050                 complete_all(&req->r_completion);
2051 }
2052
2053 /*
2054  * called under mdsc->mutex
2055  */
2056 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2057                                   struct ceph_mds_request *req,
2058                                   int mds, bool drop_cap_releases)
2059 {
2060         struct ceph_mds_request_head *rhead;
2061         struct ceph_msg *msg;
2062         int flags = 0;
2063
2064         req->r_attempts++;
2065         if (req->r_inode) {
2066                 struct ceph_cap *cap =
2067                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2068
2069                 if (cap)
2070                         req->r_sent_on_mseq = cap->mseq;
2071                 else
2072                         req->r_sent_on_mseq = -1;
2073         }
2074         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2075              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2076
2077         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2078                 void *p;
2079                 /*
2080                  * Replay.  Do not regenerate message (and rebuild
2081                  * paths, etc.); just use the original message.
2082                  * Rebuilding paths will break for renames because
2083                  * d_move mangles the src name.
2084                  */
2085                 msg = req->r_request;
2086                 rhead = msg->front.iov_base;
2087
2088                 flags = le32_to_cpu(rhead->flags);
2089                 flags |= CEPH_MDS_FLAG_REPLAY;
2090                 rhead->flags = cpu_to_le32(flags);
2091
2092                 if (req->r_target_inode)
2093                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2094
2095                 rhead->num_retry = req->r_attempts - 1;
2096
2097                 /* remove cap/dentry releases from message */
2098                 rhead->num_releases = 0;
2099
2100                 /* time stamp */
2101                 p = msg->front.iov_base + req->r_request_release_offset;
2102                 {
2103                         struct ceph_timespec ts;
2104                         ceph_encode_timespec(&ts, &req->r_stamp);
2105                         ceph_encode_copy(&p, &ts, sizeof(ts));
2106                 }
2107
2108                 msg->front.iov_len = p - msg->front.iov_base;
2109                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2110                 return 0;
2111         }
2112
2113         if (req->r_request) {
2114                 ceph_msg_put(req->r_request);
2115                 req->r_request = NULL;
2116         }
2117         msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2118         if (IS_ERR(msg)) {
2119                 req->r_err = PTR_ERR(msg);
2120                 return PTR_ERR(msg);
2121         }
2122         req->r_request = msg;
2123
2124         rhead = msg->front.iov_base;
2125         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2126         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2127                 flags |= CEPH_MDS_FLAG_REPLAY;
2128         if (req->r_parent)
2129                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2130         rhead->flags = cpu_to_le32(flags);
2131         rhead->num_fwd = req->r_num_fwd;
2132         rhead->num_retry = req->r_attempts - 1;
2133         rhead->ino = 0;
2134
2135         dout(" r_parent = %p\n", req->r_parent);
2136         return 0;
2137 }
2138
2139 /*
2140  * send request, or put it on the appropriate wait list.
2141  */
2142 static int __do_request(struct ceph_mds_client *mdsc,
2143                         struct ceph_mds_request *req)
2144 {
2145         struct ceph_mds_session *session = NULL;
2146         int mds = -1;
2147         int err = 0;
2148
2149         if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2150                 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2151                         __unregister_request(mdsc, req);
2152                 goto out;
2153         }
2154
2155         if (req->r_timeout &&
2156             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2157                 dout("do_request timed out\n");
2158                 err = -EIO;
2159                 goto finish;
2160         }
2161         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2162                 dout("do_request forced umount\n");
2163                 err = -EIO;
2164                 goto finish;
2165         }
2166         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2167                 if (mdsc->mdsmap_err) {
2168                         err = mdsc->mdsmap_err;
2169                         dout("do_request mdsmap err %d\n", err);
2170                         goto finish;
2171                 }
2172                 if (mdsc->mdsmap->m_epoch == 0) {
2173                         dout("do_request no mdsmap, waiting for map\n");
2174                         list_add(&req->r_wait, &mdsc->waiting_for_map);
2175                         goto finish;
2176                 }
2177                 if (!(mdsc->fsc->mount_options->flags &
2178                       CEPH_MOUNT_OPT_MOUNTWAIT) &&
2179                     !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2180                         err = -ENOENT;
2181                         pr_info("probably no mds server is up\n");
2182                         goto finish;
2183                 }
2184         }
2185
2186         put_request_session(req);
2187
2188         mds = __choose_mds(mdsc, req);
2189         if (mds < 0 ||
2190             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2191                 dout("do_request no mds or not active, waiting for map\n");
2192                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2193                 goto out;
2194         }
2195
2196         /* get, open session */
2197         session = __ceph_lookup_mds_session(mdsc, mds);
2198         if (!session) {
2199                 session = register_session(mdsc, mds);
2200                 if (IS_ERR(session)) {
2201                         err = PTR_ERR(session);
2202                         goto finish;
2203                 }
2204         }
2205         req->r_session = get_session(session);
2206
2207         dout("do_request mds%d session %p state %s\n", mds, session,
2208              ceph_session_state_name(session->s_state));
2209         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2210             session->s_state != CEPH_MDS_SESSION_HUNG) {
2211                 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2212                         err = -EACCES;
2213                         goto out_session;
2214                 }
2215                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2216                     session->s_state == CEPH_MDS_SESSION_CLOSING)
2217                         __open_session(mdsc, session);
2218                 list_add(&req->r_wait, &session->s_waiting);
2219                 goto out_session;
2220         }
2221
2222         /* send request */
2223         req->r_resend_mds = -1;   /* forget any previous mds hint */
2224
2225         if (req->r_request_started == 0)   /* note request start time */
2226                 req->r_request_started = jiffies;
2227
2228         err = __prepare_send_request(mdsc, req, mds, false);
2229         if (!err) {
2230                 ceph_msg_get(req->r_request);
2231                 ceph_con_send(&session->s_con, req->r_request);
2232         }
2233
2234 out_session:
2235         ceph_put_mds_session(session);
2236 finish:
2237         if (err) {
2238                 dout("__do_request early error %d\n", err);
2239                 req->r_err = err;
2240                 complete_request(mdsc, req);
2241                 __unregister_request(mdsc, req);
2242         }
2243 out:
2244         return err;
2245 }
2246
2247 /*
2248  * called under mdsc->mutex
2249  */
2250 static void __wake_requests(struct ceph_mds_client *mdsc,
2251                             struct list_head *head)
2252 {
2253         struct ceph_mds_request *req;
2254         LIST_HEAD(tmp_list);
2255
2256         list_splice_init(head, &tmp_list);
2257
2258         while (!list_empty(&tmp_list)) {
2259                 req = list_entry(tmp_list.next,
2260                                  struct ceph_mds_request, r_wait);
2261                 list_del_init(&req->r_wait);
2262                 dout(" wake request %p tid %llu\n", req, req->r_tid);
2263                 __do_request(mdsc, req);
2264         }
2265 }
2266
2267 /*
2268  * Wake up threads with requests pending for @mds, so that they can
2269  * resubmit their requests to a possibly different mds.
2270  */
2271 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2272 {
2273         struct ceph_mds_request *req;
2274         struct rb_node *p = rb_first(&mdsc->request_tree);
2275
2276         dout("kick_requests mds%d\n", mds);
2277         while (p) {
2278                 req = rb_entry(p, struct ceph_mds_request, r_node);
2279                 p = rb_next(p);
2280                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2281                         continue;
2282                 if (req->r_attempts > 0)
2283                         continue; /* only new requests */
2284                 if (req->r_session &&
2285                     req->r_session->s_mds == mds) {
2286                         dout(" kicking tid %llu\n", req->r_tid);
2287                         list_del_init(&req->r_wait);
2288                         __do_request(mdsc, req);
2289                 }
2290         }
2291 }
2292
2293 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2294                               struct ceph_mds_request *req)
2295 {
2296         dout("submit_request on %p\n", req);
2297         mutex_lock(&mdsc->mutex);
2298         __register_request(mdsc, req, NULL);
2299         __do_request(mdsc, req);
2300         mutex_unlock(&mdsc->mutex);
2301 }
2302
2303 /*
2304  * Synchrously perform an mds request.  Take care of all of the
2305  * session setup, forwarding, retry details.
2306  */
2307 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2308                          struct inode *dir,
2309                          struct ceph_mds_request *req)
2310 {
2311         int err;
2312
2313         dout("do_request on %p\n", req);
2314
2315         /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2316         if (req->r_inode)
2317                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2318         if (req->r_parent)
2319                 ceph_get_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
2320         if (req->r_old_dentry_dir)
2321                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2322                                   CEPH_CAP_PIN);
2323
2324         /* issue */
2325         mutex_lock(&mdsc->mutex);
2326         __register_request(mdsc, req, dir);
2327         __do_request(mdsc, req);
2328
2329         if (req->r_err) {
2330                 err = req->r_err;
2331                 goto out;
2332         }
2333
2334         /* wait */
2335         mutex_unlock(&mdsc->mutex);
2336         dout("do_request waiting\n");
2337         if (!req->r_timeout && req->r_wait_for_completion) {
2338                 err = req->r_wait_for_completion(mdsc, req);
2339         } else {
2340                 long timeleft = wait_for_completion_killable_timeout(
2341                                         &req->r_completion,
2342                                         ceph_timeout_jiffies(req->r_timeout));
2343                 if (timeleft > 0)
2344                         err = 0;
2345                 else if (!timeleft)
2346                         err = -EIO;  /* timed out */
2347                 else
2348                         err = timeleft;  /* killed */
2349         }
2350         dout("do_request waited, got %d\n", err);
2351         mutex_lock(&mdsc->mutex);
2352
2353         /* only abort if we didn't race with a real reply */
2354         if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2355                 err = le32_to_cpu(req->r_reply_info.head->result);
2356         } else if (err < 0) {
2357                 dout("aborted request %lld with %d\n", req->r_tid, err);
2358
2359                 /*
2360                  * ensure we aren't running concurrently with
2361                  * ceph_fill_trace or ceph_readdir_prepopulate, which
2362                  * rely on locks (dir mutex) held by our caller.
2363                  */
2364                 mutex_lock(&req->r_fill_mutex);
2365                 req->r_err = err;
2366                 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2367                 mutex_unlock(&req->r_fill_mutex);
2368
2369                 if (req->r_parent &&
2370                     (req->r_op & CEPH_MDS_OP_WRITE))
2371                         ceph_invalidate_dir_request(req);
2372         } else {
2373                 err = req->r_err;
2374         }
2375
2376 out:
2377         mutex_unlock(&mdsc->mutex);
2378         dout("do_request %p done, result %d\n", req, err);
2379         return err;
2380 }
2381
2382 /*
2383  * Invalidate dir's completeness, dentry lease state on an aborted MDS
2384  * namespace request.
2385  */
2386 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2387 {
2388         struct inode *inode = req->r_parent;
2389
2390         dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2391
2392         ceph_dir_clear_complete(inode);
2393         if (req->r_dentry)
2394                 ceph_invalidate_dentry_lease(req->r_dentry);
2395         if (req->r_old_dentry)
2396                 ceph_invalidate_dentry_lease(req->r_old_dentry);
2397 }
2398
2399 /*
2400  * Handle mds reply.
2401  *
2402  * We take the session mutex and parse and process the reply immediately.
2403  * This preserves the logical ordering of replies, capabilities, etc., sent
2404  * by the MDS as they are applied to our local cache.
2405  */
2406 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2407 {
2408         struct ceph_mds_client *mdsc = session->s_mdsc;
2409         struct ceph_mds_request *req;
2410         struct ceph_mds_reply_head *head = msg->front.iov_base;
2411         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2412         struct ceph_snap_realm *realm;
2413         u64 tid;
2414         int err, result;
2415         int mds = session->s_mds;
2416
2417         if (msg->front.iov_len < sizeof(*head)) {
2418                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2419                 ceph_msg_dump(msg);
2420                 return;
2421         }
2422
2423         /* get request, session */
2424         tid = le64_to_cpu(msg->hdr.tid);
2425         mutex_lock(&mdsc->mutex);
2426         req = lookup_get_request(mdsc, tid);
2427         if (!req) {
2428                 dout("handle_reply on unknown tid %llu\n", tid);
2429                 mutex_unlock(&mdsc->mutex);
2430                 return;
2431         }
2432         dout("handle_reply %p\n", req);
2433
2434         /* correct session? */
2435         if (req->r_session != session) {
2436                 pr_err("mdsc_handle_reply got %llu on session mds%d"
2437                        " not mds%d\n", tid, session->s_mds,
2438                        req->r_session ? req->r_session->s_mds : -1);
2439                 mutex_unlock(&mdsc->mutex);
2440                 goto out;
2441         }
2442
2443         /* dup? */
2444         if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
2445             (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
2446                 pr_warn("got a dup %s reply on %llu from mds%d\n",
2447                            head->safe ? "safe" : "unsafe", tid, mds);
2448                 mutex_unlock(&mdsc->mutex);
2449                 goto out;
2450         }
2451         if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
2452                 pr_warn("got unsafe after safe on %llu from mds%d\n",
2453                            tid, mds);
2454                 mutex_unlock(&mdsc->mutex);
2455                 goto out;
2456         }
2457
2458         result = le32_to_cpu(head->result);
2459
2460         /*
2461          * Handle an ESTALE
2462          * if we're not talking to the authority, send to them
2463          * if the authority has changed while we weren't looking,
2464          * send to new authority
2465          * Otherwise we just have to return an ESTALE
2466          */
2467         if (result == -ESTALE) {
2468                 dout("got ESTALE on request %llu", req->r_tid);
2469                 req->r_resend_mds = -1;
2470                 if (req->r_direct_mode != USE_AUTH_MDS) {
2471                         dout("not using auth, setting for that now");
2472                         req->r_direct_mode = USE_AUTH_MDS;
2473                         __do_request(mdsc, req);
2474                         mutex_unlock(&mdsc->mutex);
2475                         goto out;
2476                 } else  {
2477                         int mds = __choose_mds(mdsc, req);
2478                         if (mds >= 0 && mds != req->r_session->s_mds) {
2479                                 dout("but auth changed, so resending");
2480                                 __do_request(mdsc, req);
2481                                 mutex_unlock(&mdsc->mutex);
2482                                 goto out;
2483                         }
2484                 }
2485                 dout("have to return ESTALE on request %llu", req->r_tid);
2486         }
2487
2488
2489         if (head->safe) {
2490                 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
2491                 __unregister_request(mdsc, req);
2492
2493                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2494                         /*
2495                          * We already handled the unsafe response, now do the
2496                          * cleanup.  No need to examine the response; the MDS
2497                          * doesn't include any result info in the safe
2498                          * response.  And even if it did, there is nothing
2499                          * useful we could do with a revised return value.
2500                          */
2501                         dout("got safe reply %llu, mds%d\n", tid, mds);
2502
2503                         /* last unsafe request during umount? */
2504                         if (mdsc->stopping && !__get_oldest_req(mdsc))
2505                                 complete_all(&mdsc->safe_umount_waiters);
2506                         mutex_unlock(&mdsc->mutex);
2507                         goto out;
2508                 }
2509         } else {
2510                 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
2511                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2512                 if (req->r_unsafe_dir) {
2513                         struct ceph_inode_info *ci =
2514                                         ceph_inode(req->r_unsafe_dir);
2515                         spin_lock(&ci->i_unsafe_lock);
2516                         list_add_tail(&req->r_unsafe_dir_item,
2517                                       &ci->i_unsafe_dirops);
2518                         spin_unlock(&ci->i_unsafe_lock);
2519                 }
2520         }
2521
2522         dout("handle_reply tid %lld result %d\n", tid, result);
2523         rinfo = &req->r_reply_info;
2524         err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2525         mutex_unlock(&mdsc->mutex);
2526
2527         mutex_lock(&session->s_mutex);
2528         if (err < 0) {
2529                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2530                 ceph_msg_dump(msg);
2531                 goto out_err;
2532         }
2533
2534         /* snap trace */
2535         realm = NULL;
2536         if (rinfo->snapblob_len) {
2537                 down_write(&mdsc->snap_rwsem);
2538                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2539                                 rinfo->snapblob + rinfo->snapblob_len,
2540                                 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2541                                 &realm);
2542                 downgrade_write(&mdsc->snap_rwsem);
2543         } else {
2544                 down_read(&mdsc->snap_rwsem);
2545         }
2546
2547         /* insert trace into our cache */
2548         mutex_lock(&req->r_fill_mutex);
2549         current->journal_info = req;
2550         err = ceph_fill_trace(mdsc->fsc->sb, req);
2551         if (err == 0) {
2552                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2553                                     req->r_op == CEPH_MDS_OP_LSSNAP))
2554                         ceph_readdir_prepopulate(req, req->r_session);
2555                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2556         }
2557         current->journal_info = NULL;
2558         mutex_unlock(&req->r_fill_mutex);
2559
2560         up_read(&mdsc->snap_rwsem);
2561         if (realm)
2562                 ceph_put_snap_realm(mdsc, realm);
2563
2564         if (err == 0 && req->r_target_inode &&
2565             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2566                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
2567                 spin_lock(&ci->i_unsafe_lock);
2568                 list_add_tail(&req->r_unsafe_target_item, &ci->i_unsafe_iops);
2569                 spin_unlock(&ci->i_unsafe_lock);
2570         }
2571 out_err:
2572         mutex_lock(&mdsc->mutex);
2573         if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2574                 if (err) {
2575                         req->r_err = err;
2576                 } else {
2577                         req->r_reply =  ceph_msg_get(msg);
2578                         set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
2579                 }
2580         } else {
2581                 dout("reply arrived after request %lld was aborted\n", tid);
2582         }
2583         mutex_unlock(&mdsc->mutex);
2584
2585         mutex_unlock(&session->s_mutex);
2586
2587         /* kick calling process */
2588         complete_request(mdsc, req);
2589 out:
2590         ceph_mdsc_put_request(req);
2591         return;
2592 }
2593
2594
2595
2596 /*
2597  * handle mds notification that our request has been forwarded.
2598  */
2599 static void handle_forward(struct ceph_mds_client *mdsc,
2600                            struct ceph_mds_session *session,
2601                            struct ceph_msg *msg)
2602 {
2603         struct ceph_mds_request *req;
2604         u64 tid = le64_to_cpu(msg->hdr.tid);
2605         u32 next_mds;
2606         u32 fwd_seq;
2607         int err = -EINVAL;
2608         void *p = msg->front.iov_base;
2609         void *end = p + msg->front.iov_len;
2610
2611         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2612         next_mds = ceph_decode_32(&p);
2613         fwd_seq = ceph_decode_32(&p);
2614
2615         mutex_lock(&mdsc->mutex);
2616         req = lookup_get_request(mdsc, tid);
2617         if (!req) {
2618                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2619                 goto out;  /* dup reply? */
2620         }
2621
2622         if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2623                 dout("forward tid %llu aborted, unregistering\n", tid);
2624                 __unregister_request(mdsc, req);
2625         } else if (fwd_seq <= req->r_num_fwd) {
2626                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2627                      tid, next_mds, req->r_num_fwd, fwd_seq);
2628         } else {
2629                 /* resend. forward race not possible; mds would drop */
2630                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2631                 BUG_ON(req->r_err);
2632                 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
2633                 req->r_attempts = 0;
2634                 req->r_num_fwd = fwd_seq;
2635                 req->r_resend_mds = next_mds;
2636                 put_request_session(req);
2637                 __do_request(mdsc, req);
2638         }
2639         ceph_mdsc_put_request(req);
2640 out:
2641         mutex_unlock(&mdsc->mutex);
2642         return;
2643
2644 bad:
2645         pr_err("mdsc_handle_forward decode error err=%d\n", err);
2646 }
2647
2648 /*
2649  * handle a mds session control message
2650  */
2651 static void handle_session(struct ceph_mds_session *session,
2652                            struct ceph_msg *msg)
2653 {
2654         struct ceph_mds_client *mdsc = session->s_mdsc;
2655         u32 op;
2656         u64 seq;
2657         int mds = session->s_mds;
2658         struct ceph_mds_session_head *h = msg->front.iov_base;
2659         int wake = 0;
2660
2661         /* decode */
2662         if (msg->front.iov_len != sizeof(*h))
2663                 goto bad;
2664         op = le32_to_cpu(h->op);
2665         seq = le64_to_cpu(h->seq);
2666
2667         mutex_lock(&mdsc->mutex);
2668         if (op == CEPH_SESSION_CLOSE) {
2669                 get_session(session);
2670                 __unregister_session(mdsc, session);
2671         }
2672         /* FIXME: this ttl calculation is generous */
2673         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2674         mutex_unlock(&mdsc->mutex);
2675
2676         mutex_lock(&session->s_mutex);
2677
2678         dout("handle_session mds%d %s %p state %s seq %llu\n",
2679              mds, ceph_session_op_name(op), session,
2680              ceph_session_state_name(session->s_state), seq);
2681
2682         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2683                 session->s_state = CEPH_MDS_SESSION_OPEN;
2684                 pr_info("mds%d came back\n", session->s_mds);
2685         }
2686
2687         switch (op) {
2688         case CEPH_SESSION_OPEN:
2689                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2690                         pr_info("mds%d reconnect success\n", session->s_mds);
2691                 session->s_state = CEPH_MDS_SESSION_OPEN;
2692                 renewed_caps(mdsc, session, 0);
2693                 wake = 1;
2694                 if (mdsc->stopping)
2695                         __close_session(mdsc, session);
2696                 break;
2697
2698         case CEPH_SESSION_RENEWCAPS:
2699                 if (session->s_renew_seq == seq)
2700                         renewed_caps(mdsc, session, 1);
2701                 break;
2702
2703         case CEPH_SESSION_CLOSE:
2704                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2705                         pr_info("mds%d reconnect denied\n", session->s_mds);
2706                 cleanup_session_requests(mdsc, session);
2707                 remove_session_caps(session);
2708                 wake = 2; /* for good measure */
2709                 wake_up_all(&mdsc->session_close_wq);
2710                 break;
2711
2712         case CEPH_SESSION_STALE:
2713                 pr_info("mds%d caps went stale, renewing\n",
2714                         session->s_mds);
2715                 spin_lock(&session->s_gen_ttl_lock);
2716                 session->s_cap_gen++;
2717                 session->s_cap_ttl = jiffies - 1;
2718                 spin_unlock(&session->s_gen_ttl_lock);
2719                 send_renew_caps(mdsc, session);
2720                 break;
2721
2722         case CEPH_SESSION_RECALL_STATE:
2723                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2724                 break;
2725
2726         case CEPH_SESSION_FLUSHMSG:
2727                 send_flushmsg_ack(mdsc, session, seq);
2728                 break;
2729
2730         case CEPH_SESSION_FORCE_RO:
2731                 dout("force_session_readonly %p\n", session);
2732                 spin_lock(&session->s_cap_lock);
2733                 session->s_readonly = true;
2734                 spin_unlock(&session->s_cap_lock);
2735                 wake_up_session_caps(session, 0);
2736                 break;
2737
2738         case CEPH_SESSION_REJECT:
2739                 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
2740                 pr_info("mds%d rejected session\n", session->s_mds);
2741                 session->s_state = CEPH_MDS_SESSION_REJECTED;
2742                 cleanup_session_requests(mdsc, session);
2743                 remove_session_caps(session);
2744                 wake = 2; /* for good measure */
2745                 break;
2746
2747         default:
2748                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2749                 WARN_ON(1);
2750         }
2751
2752         mutex_unlock(&session->s_mutex);
2753         if (wake) {
2754                 mutex_lock(&mdsc->mutex);
2755                 __wake_requests(mdsc, &session->s_waiting);
2756                 if (wake == 2)
2757                         kick_requests(mdsc, mds);
2758                 mutex_unlock(&mdsc->mutex);
2759         }
2760         if (op == CEPH_SESSION_CLOSE)
2761                 ceph_put_mds_session(session);
2762         return;
2763
2764 bad:
2765         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2766                (int)msg->front.iov_len);
2767         ceph_msg_dump(msg);
2768         return;
2769 }
2770
2771
2772 /*
2773  * called under session->mutex.
2774  */
2775 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2776                                    struct ceph_mds_session *session)
2777 {
2778         struct ceph_mds_request *req, *nreq;
2779         struct rb_node *p;
2780         int err;
2781
2782         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2783
2784         mutex_lock(&mdsc->mutex);
2785         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2786                 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2787                 if (!err) {
2788                         ceph_msg_get(req->r_request);
2789                         ceph_con_send(&session->s_con, req->r_request);
2790                 }
2791         }
2792
2793         /*
2794          * also re-send old requests when MDS enters reconnect stage. So that MDS
2795          * can process completed request in clientreplay stage.
2796          */
2797         p = rb_first(&mdsc->request_tree);
2798         while (p) {
2799                 req = rb_entry(p, struct ceph_mds_request, r_node);
2800                 p = rb_next(p);
2801                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2802                         continue;
2803                 if (req->r_attempts == 0)
2804                         continue; /* only old requests */
2805                 if (req->r_session &&
2806                     req->r_session->s_mds == session->s_mds) {
2807                         err = __prepare_send_request(mdsc, req,
2808                                                      session->s_mds, true);
2809                         if (!err) {
2810                                 ceph_msg_get(req->r_request);
2811                                 ceph_con_send(&session->s_con, req->r_request);
2812                         }
2813                 }
2814         }
2815         mutex_unlock(&mdsc->mutex);
2816 }
2817
2818 /*
2819  * Encode information about a cap for a reconnect with the MDS.
2820  */
2821 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2822                           void *arg)
2823 {
2824         union {
2825                 struct ceph_mds_cap_reconnect v2;
2826                 struct ceph_mds_cap_reconnect_v1 v1;
2827         } rec;
2828         struct ceph_inode_info *ci;
2829         struct ceph_reconnect_state *recon_state = arg;
2830         struct ceph_pagelist *pagelist = recon_state->pagelist;
2831         char *path;
2832         int pathlen, err;
2833         u64 pathbase;
2834         u64 snap_follows;
2835         struct dentry *dentry;
2836
2837         ci = cap->ci;
2838
2839         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2840              inode, ceph_vinop(inode), cap, cap->cap_id,
2841              ceph_cap_string(cap->issued));
2842         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2843         if (err)
2844                 return err;
2845
2846         dentry = d_find_alias(inode);
2847         if (dentry) {
2848                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2849                 if (IS_ERR(path)) {
2850                         err = PTR_ERR(path);
2851                         goto out_dput;
2852                 }
2853         } else {
2854                 path = NULL;
2855                 pathlen = 0;
2856                 pathbase = 0;
2857         }
2858
2859         spin_lock(&ci->i_ceph_lock);
2860         cap->seq = 0;        /* reset cap seq */
2861         cap->issue_seq = 0;  /* and issue_seq */
2862         cap->mseq = 0;       /* and migrate_seq */
2863         cap->cap_gen = cap->session->s_cap_gen;
2864
2865         if (recon_state->msg_version >= 2) {
2866                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2867                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2868                 rec.v2.issued = cpu_to_le32(cap->issued);
2869                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2870                 rec.v2.pathbase = cpu_to_le64(pathbase);
2871                 rec.v2.flock_len = 0;
2872         } else {
2873                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2874                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2875                 rec.v1.issued = cpu_to_le32(cap->issued);
2876                 rec.v1.size = cpu_to_le64(inode->i_size);
2877                 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2878                 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2879                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2880                 rec.v1.pathbase = cpu_to_le64(pathbase);
2881         }
2882
2883         if (list_empty(&ci->i_cap_snaps)) {
2884                 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
2885         } else {
2886                 struct ceph_cap_snap *capsnap =
2887                         list_first_entry(&ci->i_cap_snaps,
2888                                          struct ceph_cap_snap, ci_item);
2889                 snap_follows = capsnap->follows;
2890         }
2891         spin_unlock(&ci->i_ceph_lock);
2892
2893         if (recon_state->msg_version >= 2) {
2894                 int num_fcntl_locks, num_flock_locks;
2895                 struct ceph_filelock *flocks;
2896                 size_t struct_len, total_len = 0;
2897                 u8 struct_v = 0;
2898
2899 encode_again:
2900                 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2901                 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2902                                  sizeof(struct ceph_filelock), GFP_NOFS);
2903                 if (!flocks) {
2904                         err = -ENOMEM;
2905                         goto out_free;
2906                 }
2907                 err = ceph_encode_locks_to_buffer(inode, flocks,
2908                                                   num_fcntl_locks,
2909                                                   num_flock_locks);
2910                 if (err) {
2911                         kfree(flocks);
2912                         if (err == -ENOSPC)
2913                                 goto encode_again;
2914                         goto out_free;
2915                 }
2916
2917                 if (recon_state->msg_version >= 3) {
2918                         /* version, compat_version and struct_len */
2919                         total_len = 2 * sizeof(u8) + sizeof(u32);
2920                         struct_v = 2;
2921                 }
2922                 /*
2923                  * number of encoded locks is stable, so copy to pagelist
2924                  */
2925                 struct_len = 2 * sizeof(u32) +
2926                             (num_fcntl_locks + num_flock_locks) *
2927                             sizeof(struct ceph_filelock);
2928                 rec.v2.flock_len = cpu_to_le32(struct_len);
2929
2930                 struct_len += sizeof(rec.v2);
2931                 struct_len += sizeof(u32) + pathlen;
2932
2933                 if (struct_v >= 2)
2934                         struct_len += sizeof(u64); /* snap_follows */
2935
2936                 total_len += struct_len;
2937                 err = ceph_pagelist_reserve(pagelist, total_len);
2938
2939                 if (!err) {
2940                         if (recon_state->msg_version >= 3) {
2941                                 ceph_pagelist_encode_8(pagelist, struct_v);
2942                                 ceph_pagelist_encode_8(pagelist, 1);
2943                                 ceph_pagelist_encode_32(pagelist, struct_len);
2944                         }
2945                         ceph_pagelist_encode_string(pagelist, path, pathlen);
2946                         ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
2947                         ceph_locks_to_pagelist(flocks, pagelist,
2948                                                num_fcntl_locks,
2949                                                num_flock_locks);
2950                         if (struct_v >= 2)
2951                                 ceph_pagelist_encode_64(pagelist, snap_follows);
2952                 }
2953                 kfree(flocks);
2954         } else {
2955                 size_t size = sizeof(u32) + pathlen + sizeof(rec.v1);
2956                 err = ceph_pagelist_reserve(pagelist, size);
2957                 if (!err) {
2958                         ceph_pagelist_encode_string(pagelist, path, pathlen);
2959                         ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
2960                 }
2961         }
2962
2963         recon_state->nr_caps++;
2964 out_free:
2965         kfree(path);
2966 out_dput:
2967         dput(dentry);
2968         return err;
2969 }
2970
2971
2972 /*
2973  * If an MDS fails and recovers, clients need to reconnect in order to
2974  * reestablish shared state.  This includes all caps issued through
2975  * this session _and_ the snap_realm hierarchy.  Because it's not
2976  * clear which snap realms the mds cares about, we send everything we
2977  * know about.. that ensures we'll then get any new info the
2978  * recovering MDS might have.
2979  *
2980  * This is a relatively heavyweight operation, but it's rare.
2981  *
2982  * called with mdsc->mutex held.
2983  */
2984 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2985                                struct ceph_mds_session *session)
2986 {
2987         struct ceph_msg *reply;
2988         struct rb_node *p;
2989         int mds = session->s_mds;
2990         int err = -ENOMEM;
2991         int s_nr_caps;
2992         struct ceph_pagelist *pagelist;
2993         struct ceph_reconnect_state recon_state;
2994
2995         pr_info("mds%d reconnect start\n", mds);
2996
2997         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2998         if (!pagelist)
2999                 goto fail_nopagelist;
3000         ceph_pagelist_init(pagelist);
3001
3002         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
3003         if (!reply)
3004                 goto fail_nomsg;
3005
3006         mutex_lock(&session->s_mutex);
3007         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3008         session->s_seq = 0;
3009
3010         dout("session %p state %s\n", session,
3011              ceph_session_state_name(session->s_state));
3012
3013         spin_lock(&session->s_gen_ttl_lock);
3014         session->s_cap_gen++;
3015         spin_unlock(&session->s_gen_ttl_lock);
3016
3017         spin_lock(&session->s_cap_lock);
3018         /* don't know if session is readonly */
3019         session->s_readonly = 0;
3020         /*
3021          * notify __ceph_remove_cap() that we are composing cap reconnect.
3022          * If a cap get released before being added to the cap reconnect,
3023          * __ceph_remove_cap() should skip queuing cap release.
3024          */
3025         session->s_cap_reconnect = 1;
3026         /* drop old cap expires; we're about to reestablish that state */
3027         cleanup_cap_releases(mdsc, session);
3028
3029         /* trim unused caps to reduce MDS's cache rejoin time */
3030         if (mdsc->fsc->sb->s_root)
3031                 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3032
3033         ceph_con_close(&session->s_con);
3034         ceph_con_open(&session->s_con,
3035                       CEPH_ENTITY_TYPE_MDS, mds,
3036                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3037
3038         /* replay unsafe requests */
3039         replay_unsafe_requests(mdsc, session);
3040
3041         down_read(&mdsc->snap_rwsem);
3042
3043         /* traverse this session's caps */
3044         s_nr_caps = session->s_nr_caps;
3045         err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
3046         if (err)
3047                 goto fail;
3048
3049         recon_state.nr_caps = 0;
3050         recon_state.pagelist = pagelist;
3051         if (session->s_con.peer_features & CEPH_FEATURE_MDSENC)
3052                 recon_state.msg_version = 3;
3053         else if (session->s_con.peer_features & CEPH_FEATURE_FLOCK)
3054                 recon_state.msg_version = 2;
3055         else
3056                 recon_state.msg_version = 1;
3057         err = iterate_session_caps(session, encode_caps_cb, &recon_state);
3058         if (err < 0)
3059                 goto fail;
3060
3061         spin_lock(&session->s_cap_lock);
3062         session->s_cap_reconnect = 0;
3063         spin_unlock(&session->s_cap_lock);
3064
3065         /*
3066          * snaprealms.  we provide mds with the ino, seq (version), and
3067          * parent for all of our realms.  If the mds has any newer info,
3068          * it will tell us.
3069          */
3070         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3071                 struct ceph_snap_realm *realm =
3072                         rb_entry(p, struct ceph_snap_realm, node);
3073                 struct ceph_mds_snaprealm_reconnect sr_rec;
3074
3075                 dout(" adding snap realm %llx seq %lld parent %llx\n",
3076                      realm->ino, realm->seq, realm->parent_ino);
3077                 sr_rec.ino = cpu_to_le64(realm->ino);
3078                 sr_rec.seq = cpu_to_le64(realm->seq);
3079                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3080                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3081                 if (err)
3082                         goto fail;
3083         }
3084
3085         reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3086
3087         /* raced with cap release? */
3088         if (s_nr_caps != recon_state.nr_caps) {
3089                 struct page *page = list_first_entry(&pagelist->head,
3090                                                      struct page, lru);
3091                 __le32 *addr = kmap_atomic(page);
3092                 *addr = cpu_to_le32(recon_state.nr_caps);
3093                 kunmap_atomic(addr);
3094         }
3095
3096         reply->hdr.data_len = cpu_to_le32(pagelist->length);
3097         ceph_msg_data_add_pagelist(reply, pagelist);
3098
3099         ceph_early_kick_flushing_caps(mdsc, session);
3100
3101         ceph_con_send(&session->s_con, reply);
3102
3103         mutex_unlock(&session->s_mutex);
3104
3105         mutex_lock(&mdsc->mutex);
3106         __wake_requests(mdsc, &session->s_waiting);
3107         mutex_unlock(&mdsc->mutex);
3108
3109         up_read(&mdsc->snap_rwsem);
3110         return;
3111
3112 fail:
3113         ceph_msg_put(reply);
3114         up_read(&mdsc->snap_rwsem);
3115         mutex_unlock(&session->s_mutex);
3116 fail_nomsg:
3117         ceph_pagelist_release(pagelist);
3118 fail_nopagelist:
3119         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3120         return;
3121 }
3122
3123
3124 /*
3125  * compare old and new mdsmaps, kicking requests
3126  * and closing out old connections as necessary
3127  *
3128  * called under mdsc->mutex.
3129  */
3130 static void check_new_map(struct ceph_mds_client *mdsc,
3131                           struct ceph_mdsmap *newmap,
3132                           struct ceph_mdsmap *oldmap)
3133 {
3134         int i;
3135         int oldstate, newstate;
3136         struct ceph_mds_session *s;
3137
3138         dout("check_new_map new %u old %u\n",
3139              newmap->m_epoch, oldmap->m_epoch);
3140
3141         for (i = 0; i < oldmap->m_num_mds && i < mdsc->max_sessions; i++) {
3142                 if (!mdsc->sessions[i])
3143                         continue;
3144                 s = mdsc->sessions[i];
3145                 oldstate = ceph_mdsmap_get_state(oldmap, i);
3146                 newstate = ceph_mdsmap_get_state(newmap, i);
3147
3148                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3149                      i, ceph_mds_state_name(oldstate),
3150                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3151                      ceph_mds_state_name(newstate),
3152                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3153                      ceph_session_state_name(s->s_state));
3154
3155                 if (i >= newmap->m_num_mds ||
3156                     memcmp(ceph_mdsmap_get_addr(oldmap, i),
3157                            ceph_mdsmap_get_addr(newmap, i),
3158                            sizeof(struct ceph_entity_addr))) {
3159                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3160                                 /* the session never opened, just close it
3161                                  * out now */
3162                                 get_session(s);
3163                                 __unregister_session(mdsc, s);
3164                                 __wake_requests(mdsc, &s->s_waiting);
3165                                 ceph_put_mds_session(s);
3166                         } else if (i >= newmap->m_num_mds) {
3167                                 /* force close session for stopped mds */
3168                                 get_session(s);
3169                                 __unregister_session(mdsc, s);
3170                                 __wake_requests(mdsc, &s->s_waiting);
3171                                 kick_requests(mdsc, i);
3172                                 mutex_unlock(&mdsc->mutex);
3173
3174                                 mutex_lock(&s->s_mutex);
3175                                 cleanup_session_requests(mdsc, s);
3176                                 remove_session_caps(s);
3177                                 mutex_unlock(&s->s_mutex);
3178
3179                                 ceph_put_mds_session(s);
3180
3181                                 mutex_lock(&mdsc->mutex);
3182                         } else {
3183                                 /* just close it */
3184                                 mutex_unlock(&mdsc->mutex);
3185                                 mutex_lock(&s->s_mutex);
3186                                 mutex_lock(&mdsc->mutex);
3187                                 ceph_con_close(&s->s_con);
3188                                 mutex_unlock(&s->s_mutex);
3189                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3190                         }
3191                 } else if (oldstate == newstate) {
3192                         continue;  /* nothing new with this mds */
3193                 }
3194
3195                 /*
3196                  * send reconnect?
3197                  */
3198                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3199                     newstate >= CEPH_MDS_STATE_RECONNECT) {
3200                         mutex_unlock(&mdsc->mutex);
3201                         send_mds_reconnect(mdsc, s);
3202                         mutex_lock(&mdsc->mutex);
3203                 }
3204
3205                 /*
3206                  * kick request on any mds that has gone active.
3207                  */
3208                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3209                     newstate >= CEPH_MDS_STATE_ACTIVE) {
3210                         if (oldstate != CEPH_MDS_STATE_CREATING &&
3211                             oldstate != CEPH_MDS_STATE_STARTING)
3212                                 pr_info("mds%d recovery completed\n", s->s_mds);
3213                         kick_requests(mdsc, i);
3214                         ceph_kick_flushing_caps(mdsc, s);
3215                         wake_up_session_caps(s, 1);
3216                 }
3217         }
3218
3219         for (i = 0; i < newmap->m_num_mds && i < mdsc->max_sessions; i++) {
3220                 s = mdsc->sessions[i];
3221                 if (!s)
3222                         continue;
3223                 if (!ceph_mdsmap_is_laggy(newmap, i))
3224                         continue;
3225                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3226                     s->s_state == CEPH_MDS_SESSION_HUNG ||
3227                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
3228                         dout(" connecting to export targets of laggy mds%d\n",
3229                              i);
3230                         __open_export_target_sessions(mdsc, s);
3231                 }
3232         }
3233 }
3234
3235
3236
3237 /*
3238  * leases
3239  */
3240
3241 /*
3242  * caller must hold session s_mutex, dentry->d_lock
3243  */
3244 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3245 {
3246         struct ceph_dentry_info *di = ceph_dentry(dentry);
3247
3248         ceph_put_mds_session(di->lease_session);
3249         di->lease_session = NULL;
3250 }
3251
3252 static void handle_lease(struct ceph_mds_client *mdsc,
3253                          struct ceph_mds_session *session,
3254                          struct ceph_msg *msg)
3255 {
3256         struct super_block *sb = mdsc->fsc->sb;
3257         struct inode *inode;
3258         struct dentry *parent, *dentry;
3259         struct ceph_dentry_info *di;
3260         int mds = session->s_mds;
3261         struct ceph_mds_lease *h = msg->front.iov_base;
3262         u32 seq;
3263         struct ceph_vino vino;
3264         struct qstr dname;
3265         int release = 0;
3266
3267         dout("handle_lease from mds%d\n", mds);
3268
3269         /* decode */
3270         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3271                 goto bad;
3272         vino.ino = le64_to_cpu(h->ino);
3273         vino.snap = CEPH_NOSNAP;
3274         seq = le32_to_cpu(h->seq);
3275         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3276         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3277         if (dname.len != get_unaligned_le32(h+1))
3278                 goto bad;
3279
3280         /* lookup inode */
3281         inode = ceph_find_inode(sb, vino);
3282         dout("handle_lease %s, ino %llx %p %.*s\n",
3283              ceph_lease_op_name(h->action), vino.ino, inode,
3284              dname.len, dname.name);
3285
3286         mutex_lock(&session->s_mutex);
3287         session->s_seq++;
3288
3289         if (!inode) {
3290                 dout("handle_lease no inode %llx\n", vino.ino);
3291                 goto release;
3292         }
3293
3294         /* dentry */
3295         parent = d_find_alias(inode);
3296         if (!parent) {
3297                 dout("no parent dentry on inode %p\n", inode);
3298                 WARN_ON(1);
3299                 goto release;  /* hrm... */
3300         }
3301         dname.hash = full_name_hash(parent, dname.name, dname.len);
3302         dentry = d_lookup(parent, &dname);
3303         dput(parent);
3304         if (!dentry)
3305                 goto release;
3306
3307         spin_lock(&dentry->d_lock);
3308         di = ceph_dentry(dentry);
3309         switch (h->action) {
3310         case CEPH_MDS_LEASE_REVOKE:
3311                 if (di->lease_session == session) {
3312                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3313                                 h->seq = cpu_to_le32(di->lease_seq);
3314                         __ceph_mdsc_drop_dentry_lease(dentry);
3315                 }
3316                 release = 1;
3317                 break;
3318
3319         case CEPH_MDS_LEASE_RENEW:
3320                 if (di->lease_session == session &&
3321                     di->lease_gen == session->s_cap_gen &&
3322                     di->lease_renew_from &&
3323                     di->lease_renew_after == 0) {
3324                         unsigned long duration =
3325                                 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3326
3327                         di->lease_seq = seq;
3328                         di->time = di->lease_renew_from + duration;
3329                         di->lease_renew_after = di->lease_renew_from +
3330                                 (duration >> 1);
3331                         di->lease_renew_from = 0;
3332                 }
3333                 break;
3334         }
3335         spin_unlock(&dentry->d_lock);
3336         dput(dentry);
3337
3338         if (!release)
3339                 goto out;
3340
3341 release:
3342         /* let's just reuse the same message */
3343         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3344         ceph_msg_get(msg);
3345         ceph_con_send(&session->s_con, msg);
3346
3347 out:
3348         iput(inode);
3349         mutex_unlock(&session->s_mutex);
3350         return;
3351
3352 bad:
3353         pr_err("corrupt lease message\n");
3354         ceph_msg_dump(msg);
3355 }
3356
3357 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3358                               struct inode *inode,
3359                               struct dentry *dentry, char action,
3360                               u32 seq)
3361 {
3362         struct ceph_msg *msg;
3363         struct ceph_mds_lease *lease;
3364         int len = sizeof(*lease) + sizeof(u32);
3365         int dnamelen = 0;
3366
3367         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3368              inode, dentry, ceph_lease_op_name(action), session->s_mds);
3369         dnamelen = dentry->d_name.len;
3370         len += dnamelen;
3371
3372         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3373         if (!msg)
3374                 return;
3375         lease = msg->front.iov_base;
3376         lease->action = action;
3377         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3378         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3379         lease->seq = cpu_to_le32(seq);
3380         put_unaligned_le32(dnamelen, lease + 1);
3381         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3382
3383         /*
3384          * if this is a preemptive lease RELEASE, no need to
3385          * flush request stream, since the actual request will
3386          * soon follow.
3387          */
3388         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3389
3390         ceph_con_send(&session->s_con, msg);
3391 }
3392
3393 /*
3394  * drop all leases (and dentry refs) in preparation for umount
3395  */
3396 static void drop_leases(struct ceph_mds_client *mdsc)
3397 {
3398         int i;
3399
3400         dout("drop_leases\n");
3401         mutex_lock(&mdsc->mutex);
3402         for (i = 0; i < mdsc->max_sessions; i++) {
3403                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3404                 if (!s)
3405                         continue;
3406                 mutex_unlock(&mdsc->mutex);
3407                 mutex_lock(&s->s_mutex);
3408                 mutex_unlock(&s->s_mutex);
3409                 ceph_put_mds_session(s);
3410                 mutex_lock(&mdsc->mutex);
3411         }
3412         mutex_unlock(&mdsc->mutex);
3413 }
3414
3415
3416
3417 /*
3418  * delayed work -- periodically trim expired leases, renew caps with mds
3419  */
3420 static void schedule_delayed(struct ceph_mds_client *mdsc)
3421 {
3422         int delay = 5;
3423         unsigned hz = round_jiffies_relative(HZ * delay);
3424         schedule_delayed_work(&mdsc->delayed_work, hz);
3425 }
3426
3427 static void delayed_work(struct work_struct *work)
3428 {
3429         int i;
3430         struct ceph_mds_client *mdsc =
3431                 container_of(work, struct ceph_mds_client, delayed_work.work);
3432         int renew_interval;
3433         int renew_caps;
3434
3435         dout("mdsc delayed_work\n");
3436         ceph_check_delayed_caps(mdsc);
3437
3438         mutex_lock(&mdsc->mutex);
3439         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3440         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3441                                    mdsc->last_renew_caps);
3442         if (renew_caps)
3443                 mdsc->last_renew_caps = jiffies;
3444
3445         for (i = 0; i < mdsc->max_sessions; i++) {
3446                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3447                 if (!s)
3448                         continue;
3449                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3450                         dout("resending session close request for mds%d\n",
3451                              s->s_mds);
3452                         request_close_session(mdsc, s);
3453                         ceph_put_mds_session(s);
3454                         continue;
3455                 }
3456                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3457                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3458                                 s->s_state = CEPH_MDS_SESSION_HUNG;
3459                                 pr_info("mds%d hung\n", s->s_mds);
3460                         }
3461                 }
3462                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3463                         /* this mds is failed or recovering, just wait */
3464                         ceph_put_mds_session(s);
3465                         continue;
3466                 }
3467                 mutex_unlock(&mdsc->mutex);
3468
3469                 mutex_lock(&s->s_mutex);
3470                 if (renew_caps)
3471                         send_renew_caps(mdsc, s);
3472                 else
3473                         ceph_con_keepalive(&s->s_con);
3474                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3475                     s->s_state == CEPH_MDS_SESSION_HUNG)
3476                         ceph_send_cap_releases(mdsc, s);
3477                 mutex_unlock(&s->s_mutex);
3478                 ceph_put_mds_session(s);
3479
3480                 mutex_lock(&mdsc->mutex);
3481         }
3482         mutex_unlock(&mdsc->mutex);
3483
3484         schedule_delayed(mdsc);
3485 }
3486
3487 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3488
3489 {
3490         struct ceph_mds_client *mdsc;
3491
3492         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3493         if (!mdsc)
3494                 return -ENOMEM;
3495         mdsc->fsc = fsc;
3496         fsc->mdsc = mdsc;
3497         mutex_init(&mdsc->mutex);
3498         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3499         if (!mdsc->mdsmap) {
3500                 kfree(mdsc);
3501                 return -ENOMEM;
3502         }
3503
3504         init_completion(&mdsc->safe_umount_waiters);
3505         init_waitqueue_head(&mdsc->session_close_wq);
3506         INIT_LIST_HEAD(&mdsc->waiting_for_map);
3507         mdsc->sessions = NULL;
3508         atomic_set(&mdsc->num_sessions, 0);
3509         mdsc->max_sessions = 0;
3510         mdsc->stopping = 0;
3511         mdsc->last_snap_seq = 0;
3512         init_rwsem(&mdsc->snap_rwsem);
3513         mdsc->snap_realms = RB_ROOT;
3514         INIT_LIST_HEAD(&mdsc->snap_empty);
3515         spin_lock_init(&mdsc->snap_empty_lock);
3516         mdsc->last_tid = 0;
3517         mdsc->oldest_tid = 0;
3518         mdsc->request_tree = RB_ROOT;
3519         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3520         mdsc->last_renew_caps = jiffies;
3521         INIT_LIST_HEAD(&mdsc->cap_delay_list);
3522         spin_lock_init(&mdsc->cap_delay_lock);
3523         INIT_LIST_HEAD(&mdsc->snap_flush_list);
3524         spin_lock_init(&mdsc->snap_flush_lock);
3525         mdsc->last_cap_flush_tid = 1;
3526         INIT_LIST_HEAD(&mdsc->cap_flush_list);
3527         INIT_LIST_HEAD(&mdsc->cap_dirty);
3528         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3529         mdsc->num_cap_flushing = 0;
3530         spin_lock_init(&mdsc->cap_dirty_lock);
3531         init_waitqueue_head(&mdsc->cap_flushing_wq);
3532         spin_lock_init(&mdsc->dentry_lru_lock);
3533         INIT_LIST_HEAD(&mdsc->dentry_lru);
3534
3535         ceph_caps_init(mdsc);
3536         ceph_adjust_min_caps(mdsc, fsc->min_caps);
3537
3538         init_rwsem(&mdsc->pool_perm_rwsem);
3539         mdsc->pool_perm_tree = RB_ROOT;
3540
3541         strncpy(mdsc->nodename, utsname()->nodename,
3542                 sizeof(mdsc->nodename) - 1);
3543         return 0;
3544 }
3545
3546 /*
3547  * Wait for safe replies on open mds requests.  If we time out, drop
3548  * all requests from the tree to avoid dangling dentry refs.
3549  */
3550 static void wait_requests(struct ceph_mds_client *mdsc)
3551 {
3552         struct ceph_options *opts = mdsc->fsc->client->options;
3553         struct ceph_mds_request *req;
3554
3555         mutex_lock(&mdsc->mutex);
3556         if (__get_oldest_req(mdsc)) {
3557                 mutex_unlock(&mdsc->mutex);
3558
3559                 dout("wait_requests waiting for requests\n");
3560                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3561                                     ceph_timeout_jiffies(opts->mount_timeout));
3562
3563                 /* tear down remaining requests */
3564                 mutex_lock(&mdsc->mutex);
3565                 while ((req = __get_oldest_req(mdsc))) {
3566                         dout("wait_requests timed out on tid %llu\n",
3567                              req->r_tid);
3568                         __unregister_request(mdsc, req);
3569                 }
3570         }
3571         mutex_unlock(&mdsc->mutex);
3572         dout("wait_requests done\n");
3573 }
3574
3575 /*
3576  * called before mount is ro, and before dentries are torn down.
3577  * (hmm, does this still race with new lookups?)
3578  */
3579 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3580 {
3581         dout("pre_umount\n");
3582         mdsc->stopping = 1;
3583
3584         drop_leases(mdsc);
3585         ceph_flush_dirty_caps(mdsc);
3586         wait_requests(mdsc);
3587
3588         /*
3589          * wait for reply handlers to drop their request refs and
3590          * their inode/dcache refs
3591          */
3592         ceph_msgr_flush();
3593 }
3594
3595 /*
3596  * wait for all write mds requests to flush.
3597  */
3598 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3599 {
3600         struct ceph_mds_request *req = NULL, *nextreq;
3601         struct rb_node *n;
3602
3603         mutex_lock(&mdsc->mutex);
3604         dout("wait_unsafe_requests want %lld\n", want_tid);
3605 restart:
3606         req = __get_oldest_req(mdsc);
3607         while (req && req->r_tid <= want_tid) {
3608                 /* find next request */
3609                 n = rb_next(&req->r_node);
3610                 if (n)
3611                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3612                 else
3613                         nextreq = NULL;
3614                 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3615                     (req->r_op & CEPH_MDS_OP_WRITE)) {
3616                         /* write op */
3617                         ceph_mdsc_get_request(req);
3618                         if (nextreq)
3619                                 ceph_mdsc_get_request(nextreq);
3620                         mutex_unlock(&mdsc->mutex);
3621                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
3622                              req->r_tid, want_tid);
3623                         wait_for_completion(&req->r_safe_completion);
3624                         mutex_lock(&mdsc->mutex);
3625                         ceph_mdsc_put_request(req);
3626                         if (!nextreq)
3627                                 break;  /* next dne before, so we're done! */
3628                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
3629                                 /* next request was removed from tree */
3630                                 ceph_mdsc_put_request(nextreq);
3631                                 goto restart;
3632                         }
3633                         ceph_mdsc_put_request(nextreq);  /* won't go away */
3634                 }
3635                 req = nextreq;
3636         }
3637         mutex_unlock(&mdsc->mutex);
3638         dout("wait_unsafe_requests done\n");
3639 }
3640
3641 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3642 {
3643         u64 want_tid, want_flush;
3644
3645         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3646                 return;
3647
3648         dout("sync\n");
3649         mutex_lock(&mdsc->mutex);
3650         want_tid = mdsc->last_tid;
3651         mutex_unlock(&mdsc->mutex);
3652
3653         ceph_flush_dirty_caps(mdsc);
3654         spin_lock(&mdsc->cap_dirty_lock);
3655         want_flush = mdsc->last_cap_flush_tid;
3656         if (!list_empty(&mdsc->cap_flush_list)) {
3657                 struct ceph_cap_flush *cf =
3658                         list_last_entry(&mdsc->cap_flush_list,
3659                                         struct ceph_cap_flush, g_list);
3660                 cf->wake = true;
3661         }
3662         spin_unlock(&mdsc->cap_dirty_lock);
3663
3664         dout("sync want tid %lld flush_seq %lld\n",
3665              want_tid, want_flush);
3666
3667         wait_unsafe_requests(mdsc, want_tid);
3668         wait_caps_flush(mdsc, want_flush);
3669 }
3670
3671 /*
3672  * true if all sessions are closed, or we force unmount
3673  */
3674 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
3675 {
3676         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3677                 return true;
3678         return atomic_read(&mdsc->num_sessions) <= skipped;
3679 }
3680
3681 /*
3682  * called after sb is ro.
3683  */
3684 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3685 {
3686         struct ceph_options *opts = mdsc->fsc->client->options;
3687         struct ceph_mds_session *session;
3688         int i;
3689         int skipped = 0;
3690
3691         dout("close_sessions\n");
3692
3693         /* close sessions */
3694         mutex_lock(&mdsc->mutex);
3695         for (i = 0; i < mdsc->max_sessions; i++) {
3696                 session = __ceph_lookup_mds_session(mdsc, i);
3697                 if (!session)
3698                         continue;
3699                 mutex_unlock(&mdsc->mutex);
3700                 mutex_lock(&session->s_mutex);
3701                 if (__close_session(mdsc, session) <= 0)
3702                         skipped++;
3703                 mutex_unlock(&session->s_mutex);
3704                 ceph_put_mds_session(session);
3705                 mutex_lock(&mdsc->mutex);
3706         }
3707         mutex_unlock(&mdsc->mutex);
3708
3709         dout("waiting for sessions to close\n");
3710         wait_event_timeout(mdsc->session_close_wq,
3711                            done_closing_sessions(mdsc, skipped),
3712                            ceph_timeout_jiffies(opts->mount_timeout));
3713
3714         /* tear down remaining sessions */
3715         mutex_lock(&mdsc->mutex);
3716         for (i = 0; i < mdsc->max_sessions; i++) {
3717                 if (mdsc->sessions[i]) {
3718                         session = get_session(mdsc->sessions[i]);
3719                         __unregister_session(mdsc, session);
3720                         mutex_unlock(&mdsc->mutex);
3721                         mutex_lock(&session->s_mutex);
3722                         remove_session_caps(session);
3723                         mutex_unlock(&session->s_mutex);
3724                         ceph_put_mds_session(session);
3725                         mutex_lock(&mdsc->mutex);
3726                 }
3727         }
3728         WARN_ON(!list_empty(&mdsc->cap_delay_list));
3729         mutex_unlock(&mdsc->mutex);
3730
3731         ceph_cleanup_empty_realms(mdsc);
3732
3733         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3734
3735         dout("stopped\n");
3736 }
3737
3738 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
3739 {
3740         struct ceph_mds_session *session;
3741         int mds;
3742
3743         dout("force umount\n");
3744
3745         mutex_lock(&mdsc->mutex);
3746         for (mds = 0; mds < mdsc->max_sessions; mds++) {
3747                 session = __ceph_lookup_mds_session(mdsc, mds);
3748                 if (!session)
3749                         continue;
3750                 mutex_unlock(&mdsc->mutex);
3751                 mutex_lock(&session->s_mutex);
3752                 __close_session(mdsc, session);
3753                 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
3754                         cleanup_session_requests(mdsc, session);
3755                         remove_session_caps(session);
3756                 }
3757                 mutex_unlock(&session->s_mutex);
3758                 ceph_put_mds_session(session);
3759                 mutex_lock(&mdsc->mutex);
3760                 kick_requests(mdsc, mds);
3761         }
3762         __wake_requests(mdsc, &mdsc->waiting_for_map);
3763         mutex_unlock(&mdsc->mutex);
3764 }
3765
3766 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3767 {
3768         dout("stop\n");
3769         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3770         if (mdsc->mdsmap)
3771                 ceph_mdsmap_destroy(mdsc->mdsmap);
3772         kfree(mdsc->sessions);
3773         ceph_caps_finalize(mdsc);
3774         ceph_pool_perm_destroy(mdsc);
3775 }
3776
3777 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3778 {
3779         struct ceph_mds_client *mdsc = fsc->mdsc;
3780         dout("mdsc_destroy %p\n", mdsc);
3781
3782         /* flush out any connection work with references to us */
3783         ceph_msgr_flush();
3784
3785         ceph_mdsc_stop(mdsc);
3786
3787         fsc->mdsc = NULL;
3788         kfree(mdsc);
3789         dout("mdsc_destroy %p done\n", mdsc);
3790 }
3791
3792 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3793 {
3794         struct ceph_fs_client *fsc = mdsc->fsc;
3795         const char *mds_namespace = fsc->mount_options->mds_namespace;
3796         void *p = msg->front.iov_base;
3797         void *end = p + msg->front.iov_len;
3798         u32 epoch;
3799         u32 map_len;
3800         u32 num_fs;
3801         u32 mount_fscid = (u32)-1;
3802         u8 struct_v, struct_cv;
3803         int err = -EINVAL;
3804
3805         ceph_decode_need(&p, end, sizeof(u32), bad);
3806         epoch = ceph_decode_32(&p);
3807
3808         dout("handle_fsmap epoch %u\n", epoch);
3809
3810         ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3811         struct_v = ceph_decode_8(&p);
3812         struct_cv = ceph_decode_8(&p);
3813         map_len = ceph_decode_32(&p);
3814
3815         ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
3816         p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
3817
3818         num_fs = ceph_decode_32(&p);
3819         while (num_fs-- > 0) {
3820                 void *info_p, *info_end;
3821                 u32 info_len;
3822                 u8 info_v, info_cv;
3823                 u32 fscid, namelen;
3824
3825                 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3826                 info_v = ceph_decode_8(&p);
3827                 info_cv = ceph_decode_8(&p);
3828                 info_len = ceph_decode_32(&p);
3829                 ceph_decode_need(&p, end, info_len, bad);
3830                 info_p = p;
3831                 info_end = p + info_len;
3832                 p = info_end;
3833
3834                 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
3835                 fscid = ceph_decode_32(&info_p);
3836                 namelen = ceph_decode_32(&info_p);
3837                 ceph_decode_need(&info_p, info_end, namelen, bad);
3838
3839                 if (mds_namespace &&
3840                     strlen(mds_namespace) == namelen &&
3841                     !strncmp(mds_namespace, (char *)info_p, namelen)) {
3842                         mount_fscid = fscid;
3843                         break;
3844                 }
3845         }
3846
3847         ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
3848         if (mount_fscid != (u32)-1) {
3849                 fsc->client->monc.fs_cluster_id = mount_fscid;
3850                 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
3851                                    0, true);
3852                 ceph_monc_renew_subs(&fsc->client->monc);
3853         } else {
3854                 err = -ENOENT;
3855                 goto err_out;
3856         }
3857         return;
3858 bad:
3859         pr_err("error decoding fsmap\n");
3860 err_out:
3861         mutex_lock(&mdsc->mutex);
3862         mdsc->mdsmap_err = -ENOENT;
3863         __wake_requests(mdsc, &mdsc->waiting_for_map);
3864         mutex_unlock(&mdsc->mutex);
3865         return;
3866 }
3867
3868 /*
3869  * handle mds map update.
3870  */
3871 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3872 {
3873         u32 epoch;
3874         u32 maplen;
3875         void *p = msg->front.iov_base;
3876         void *end = p + msg->front.iov_len;
3877         struct ceph_mdsmap *newmap, *oldmap;
3878         struct ceph_fsid fsid;
3879         int err = -EINVAL;
3880
3881         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3882         ceph_decode_copy(&p, &fsid, sizeof(fsid));
3883         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3884                 return;
3885         epoch = ceph_decode_32(&p);
3886         maplen = ceph_decode_32(&p);
3887         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3888
3889         /* do we need it? */
3890         mutex_lock(&mdsc->mutex);
3891         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3892                 dout("handle_map epoch %u <= our %u\n",
3893                      epoch, mdsc->mdsmap->m_epoch);
3894                 mutex_unlock(&mdsc->mutex);
3895                 return;
3896         }
3897
3898         newmap = ceph_mdsmap_decode(&p, end);
3899         if (IS_ERR(newmap)) {
3900                 err = PTR_ERR(newmap);
3901                 goto bad_unlock;
3902         }
3903
3904         /* swap into place */
3905         if (mdsc->mdsmap) {
3906                 oldmap = mdsc->mdsmap;
3907                 mdsc->mdsmap = newmap;
3908                 check_new_map(mdsc, newmap, oldmap);
3909                 ceph_mdsmap_destroy(oldmap);
3910         } else {
3911                 mdsc->mdsmap = newmap;  /* first mds map */
3912         }
3913         mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3914
3915         __wake_requests(mdsc, &mdsc->waiting_for_map);
3916         ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
3917                           mdsc->mdsmap->m_epoch);
3918
3919         mutex_unlock(&mdsc->mutex);
3920         schedule_delayed(mdsc);
3921         return;
3922
3923 bad_unlock:
3924         mutex_unlock(&mdsc->mutex);
3925 bad:
3926         pr_err("error decoding mdsmap %d\n", err);
3927         return;
3928 }
3929
3930 static struct ceph_connection *con_get(struct ceph_connection *con)
3931 {
3932         struct ceph_mds_session *s = con->private;
3933
3934         if (get_session(s)) {
3935                 dout("mdsc con_get %p ok (%d)\n", s, refcount_read(&s->s_ref));
3936                 return con;
3937         }
3938         dout("mdsc con_get %p FAIL\n", s);
3939         return NULL;
3940 }
3941
3942 static void con_put(struct ceph_connection *con)
3943 {
3944         struct ceph_mds_session *s = con->private;
3945
3946         dout("mdsc con_put %p (%d)\n", s, refcount_read(&s->s_ref) - 1);
3947         ceph_put_mds_session(s);
3948 }
3949
3950 /*
3951  * if the client is unresponsive for long enough, the mds will kill
3952  * the session entirely.
3953  */
3954 static void peer_reset(struct ceph_connection *con)
3955 {
3956         struct ceph_mds_session *s = con->private;
3957         struct ceph_mds_client *mdsc = s->s_mdsc;
3958
3959         pr_warn("mds%d closed our session\n", s->s_mds);
3960         send_mds_reconnect(mdsc, s);
3961 }
3962
3963 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3964 {
3965         struct ceph_mds_session *s = con->private;
3966         struct ceph_mds_client *mdsc = s->s_mdsc;
3967         int type = le16_to_cpu(msg->hdr.type);
3968
3969         mutex_lock(&mdsc->mutex);
3970         if (__verify_registered_session(mdsc, s) < 0) {
3971                 mutex_unlock(&mdsc->mutex);
3972                 goto out;
3973         }
3974         mutex_unlock(&mdsc->mutex);
3975
3976         switch (type) {
3977         case CEPH_MSG_MDS_MAP:
3978                 ceph_mdsc_handle_mdsmap(mdsc, msg);
3979                 break;
3980         case CEPH_MSG_FS_MAP_USER:
3981                 ceph_mdsc_handle_fsmap(mdsc, msg);
3982                 break;
3983         case CEPH_MSG_CLIENT_SESSION:
3984                 handle_session(s, msg);
3985                 break;
3986         case CEPH_MSG_CLIENT_REPLY:
3987                 handle_reply(s, msg);
3988                 break;
3989         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3990                 handle_forward(mdsc, s, msg);
3991                 break;
3992         case CEPH_MSG_CLIENT_CAPS:
3993                 ceph_handle_caps(s, msg);
3994                 break;
3995         case CEPH_MSG_CLIENT_SNAP:
3996                 ceph_handle_snap(mdsc, s, msg);
3997                 break;
3998         case CEPH_MSG_CLIENT_LEASE:
3999                 handle_lease(mdsc, s, msg);
4000                 break;
4001
4002         default:
4003                 pr_err("received unknown message type %d %s\n", type,
4004                        ceph_msg_type_name(type));
4005         }
4006 out:
4007         ceph_msg_put(msg);
4008 }
4009
4010 /*
4011  * authentication
4012  */
4013
4014 /*
4015  * Note: returned pointer is the address of a structure that's
4016  * managed separately.  Caller must *not* attempt to free it.
4017  */
4018 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
4019                                         int *proto, int force_new)
4020 {
4021         struct ceph_mds_session *s = con->private;
4022         struct ceph_mds_client *mdsc = s->s_mdsc;
4023         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4024         struct ceph_auth_handshake *auth = &s->s_auth;
4025
4026         if (force_new && auth->authorizer) {
4027                 ceph_auth_destroy_authorizer(auth->authorizer);
4028                 auth->authorizer = NULL;
4029         }
4030         if (!auth->authorizer) {
4031                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4032                                                       auth);
4033                 if (ret)
4034                         return ERR_PTR(ret);
4035         } else {
4036                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4037                                                       auth);
4038                 if (ret)
4039                         return ERR_PTR(ret);
4040         }
4041         *proto = ac->protocol;
4042
4043         return auth;
4044 }
4045
4046
4047 static int verify_authorizer_reply(struct ceph_connection *con)
4048 {
4049         struct ceph_mds_session *s = con->private;
4050         struct ceph_mds_client *mdsc = s->s_mdsc;
4051         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4052
4053         return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
4054 }
4055
4056 static int invalidate_authorizer(struct ceph_connection *con)
4057 {
4058         struct ceph_mds_session *s = con->private;
4059         struct ceph_mds_client *mdsc = s->s_mdsc;
4060         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4061
4062         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
4063
4064         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
4065 }
4066
4067 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
4068                                 struct ceph_msg_header *hdr, int *skip)
4069 {
4070         struct ceph_msg *msg;
4071         int type = (int) le16_to_cpu(hdr->type);
4072         int front_len = (int) le32_to_cpu(hdr->front_len);
4073
4074         if (con->in_msg)
4075                 return con->in_msg;
4076
4077         *skip = 0;
4078         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
4079         if (!msg) {
4080                 pr_err("unable to allocate msg type %d len %d\n",
4081                        type, front_len);
4082                 return NULL;
4083         }
4084
4085         return msg;
4086 }
4087
4088 static int mds_sign_message(struct ceph_msg *msg)
4089 {
4090        struct ceph_mds_session *s = msg->con->private;
4091        struct ceph_auth_handshake *auth = &s->s_auth;
4092
4093        return ceph_auth_sign_message(auth, msg);
4094 }
4095
4096 static int mds_check_message_signature(struct ceph_msg *msg)
4097 {
4098        struct ceph_mds_session *s = msg->con->private;
4099        struct ceph_auth_handshake *auth = &s->s_auth;
4100
4101        return ceph_auth_check_message_signature(auth, msg);
4102 }
4103
4104 static const struct ceph_connection_operations mds_con_ops = {
4105         .get = con_get,
4106         .put = con_put,
4107         .dispatch = dispatch,
4108         .get_authorizer = get_authorizer,
4109         .verify_authorizer_reply = verify_authorizer_reply,
4110         .invalidate_authorizer = invalidate_authorizer,
4111         .peer_reset = peer_reset,
4112         .alloc_msg = mds_alloc_msg,
4113         .sign_message = mds_sign_message,
4114         .check_message_signature = mds_check_message_signature,
4115 };
4116
4117 /* eof */