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/sched.h>
   7 #include <linux/debugfs.h>
   8 #include <linux/seq_file.h>
   9
  10 #include "super.h"
  11 #include "mds_client.h"
  12
  13 #include <linux/ceph/messenger.h>
  14 #include <linux/ceph/decode.h>
  15 #include <linux/ceph/pagelist.h>
  16 #include <linux/ceph/auth.h>
  17 #include <linux/ceph/debugfs.h>
  18
  19 /*
  20  * A cluster of MDS (metadata server) daemons is responsible for
  21  * managing the file system namespace (the directory hierarchy and
  22  * inodes) and for coordinating shared access to storage.  Metadata is
  23  * partitioning hierarchically across a number of servers, and that
  24  * partition varies over time as the cluster adjusts the distribution
  25  * in order to balance load.
  26  *
  27  * The MDS client is primarily responsible to managing synchronous
  28  * metadata requests for operations like open, unlink, and so forth.
  29  * If there is a MDS failure, we find out about it when we (possibly
  30  * request and) receive a new MDS map, and can resubmit affected
  31  * requests.
  32  *
  33  * For the most part, though, we take advantage of a lossless
  34  * communications channel to the MDS, and do not need to worry about
  35  * timing out or resubmitting requests.
  36  *
  37  * We maintain a stateful "session" with each MDS we interact with.
  38  * Within each session, we sent periodic heartbeat messages to ensure
  39  * any capabilities or leases we have been issues remain valid.  If
  40  * the session times out and goes stale, our leases and capabilities
  41  * are no longer valid.
  42  */
  43
  44 struct ceph_reconnect_state {
  45         struct ceph_pagelist *pagelist;
  46         bool flock;
  47 };
  48
  49 static void __wake_requests(struct ceph_mds_client *mdsc,
  50                             struct list_head *head);
  51
  52 static const struct ceph_connection_operations mds_con_ops;
  53
  54
  55 /*
  56  * mds reply parsing
  57  */
  58
  59 /*
  60  * parse individual inode info
  61  */
  62 static int parse_reply_info_in(void **p, void *end,
  63                                struct ceph_mds_reply_info_in *info,
  64                                int features)
  65 {
  66         int err = -EIO;
  67
  68         info->in = *p;
  69         *p += sizeof(struct ceph_mds_reply_inode) +
  70                 sizeof(*info->in->fragtree.splits) *
  71                 le32_to_cpu(info->in->fragtree.nsplits);
  72
  73         ceph_decode_32_safe(p, end, info->symlink_len, bad);
  74         ceph_decode_need(p, end, info->symlink_len, bad);
  75         info->symlink = *p;
  76         *p += info->symlink_len;
  77
  78         if (features & CEPH_FEATURE_DIRLAYOUTHASH)
  79                 ceph_decode_copy_safe(p, end, &info->dir_layout,
  80                                       sizeof(info->dir_layout), bad);
  81         else
  82                 memset(&info->dir_layout, 0, sizeof(info->dir_layout));
  83
  84         ceph_decode_32_safe(p, end, info->xattr_len, bad);
  85         ceph_decode_need(p, end, info->xattr_len, bad);
  86         info->xattr_data = *p;
  87         *p += info->xattr_len;
  88         return 0;
  89 bad:
  90         return err;
  91 }
  92
  93 /*
  94  * parse a normal reply, which may contain a (dir+)dentry and/or a
  95  * target inode.
  96  */
  97 static int parse_reply_info_trace(void **p, void *end,
  98                                   struct ceph_mds_reply_info_parsed *info,
  99                                   int features)
 100 {
 101         int err;
 102
 103         if (info->head->is_dentry) {
 104                 err = parse_reply_info_in(p, end, &info->diri, features);
 105                 if (err < 0)
 106                         goto out_bad;
 107
 108                 if (unlikely(*p + sizeof(*info->dirfrag) > end))
 109                         goto bad;
 110                 info->dirfrag = *p;
 111                 *p += sizeof(*info->dirfrag) +
 112                         sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
 113                 if (unlikely(*p > end))
 114                         goto bad;
 115
 116                 ceph_decode_32_safe(p, end, info->dname_len, bad);
 117                 ceph_decode_need(p, end, info->dname_len, bad);
 118                 info->dname = *p;
 119                 *p += info->dname_len;
 120                 info->dlease = *p;
 121                 *p += sizeof(*info->dlease);
 122         }
 123
 124         if (info->head->is_target) {
 125                 err = parse_reply_info_in(p, end, &info->targeti, features);
 126                 if (err < 0)
 127                         goto out_bad;
 128         }
 129
 130         if (unlikely(*p != end))
 131                 goto bad;
 132         return 0;
 133
 134 bad:
 135         err = -EIO;
 136 out_bad:
 137         pr_err("problem parsing mds trace %d\n", err);
 138         return err;
 139 }
 140
 141 /*
 142  * parse readdir results
 143  */
 144 static int parse_reply_info_dir(void **p, void *end,
 145                                 struct ceph_mds_reply_info_parsed *info,
 146                                 int features)
 147 {
 148         u32 num, i = 0;
 149         int err;
 150
 151         info->dir_dir = *p;
 152         if (*p + sizeof(*info->dir_dir) > end)
 153                 goto bad;
 154         *p += sizeof(*info->dir_dir) +
 155                 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
 156         if (*p > end)
 157                 goto bad;
 158
 159         ceph_decode_need(p, end, sizeof(num) + 2, bad);
 160         num = ceph_decode_32(p);
 161         info->dir_end = ceph_decode_8(p);
 162         info->dir_complete = ceph_decode_8(p);
 163         if (num == 0)
 164                 goto done;
 165
 166         /* alloc large array */
 167         info->dir_nr = num;
 168         info->dir_in = kcalloc(num, sizeof(*info->dir_in) +
 169                                sizeof(*info->dir_dname) +
 170                                sizeof(*info->dir_dname_len) +
 171                                sizeof(*info->dir_dlease),
 172                                GFP_NOFS);
 173         if (info->dir_in == NULL) {
 174                 err = -ENOMEM;
 175                 goto out_bad;
 176         }
 177         info->dir_dname = (void *)(info->dir_in + num);
 178         info->dir_dname_len = (void *)(info->dir_dname + num);
 179         info->dir_dlease = (void *)(info->dir_dname_len + num);
 180
 181         while (num) {
 182                 /* dentry */
 183                 ceph_decode_need(p, end, sizeof(u32)*2, bad);
 184                 info->dir_dname_len[i] = ceph_decode_32(p);
 185                 ceph_decode_need(p, end, info->dir_dname_len[i], bad);
 186                 info->dir_dname[i] = *p;
 187                 *p += info->dir_dname_len[i];
 188                 dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
 189                      info->dir_dname[i]);
 190                 info->dir_dlease[i] = *p;
 191                 *p += sizeof(struct ceph_mds_reply_lease);
 192
 193                 /* inode */
 194                 err = parse_reply_info_in(p, end, &info->dir_in[i], features);
 195                 if (err < 0)
 196                         goto out_bad;
 197                 i++;
 198                 num--;
 199         }
 200
 201 done:
 202         if (*p != end)
 203                 goto bad;
 204         return 0;
 205
 206 bad:
 207         err = -EIO;
 208 out_bad:
 209         pr_err("problem parsing dir contents %d\n", err);
 210         return err;
 211 }
 212
 213 /*
 214  * parse fcntl F_GETLK results
 215  */
 216 static int parse_reply_info_filelock(void **p, void *end,
 217                                      struct ceph_mds_reply_info_parsed *info,
 218                                      int features)
 219 {
 220         if (*p + sizeof(*info->filelock_reply) > end)
 221                 goto bad;
 222
 223         info->filelock_reply = *p;
 224         *p += sizeof(*info->filelock_reply);
 225
 226         if (unlikely(*p != end))
 227                 goto bad;
 228         return 0;
 229
 230 bad:
 231         return -EIO;
 232 }
 233
 234 /*
 235  * parse extra results
 236  */
 237 static int parse_reply_info_extra(void **p, void *end,
 238                                   struct ceph_mds_reply_info_parsed *info,
 239                                   int features)
 240 {
 241         if (info->head->op == CEPH_MDS_OP_GETFILELOCK)
 242                 return parse_reply_info_filelock(p, end, info, features);
 243         else
 244                 return parse_reply_info_dir(p, end, info, features);
 245 }
 246
 247 /*
 248  * parse entire mds reply
 249  */
 250 static int parse_reply_info(struct ceph_msg *msg,
 251                             struct ceph_mds_reply_info_parsed *info,
 252                             int features)
 253 {
 254         void *p, *end;
 255         u32 len;
 256         int err;
 257
 258         info->head = msg->front.iov_base;
 259         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 260         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 261
 262         /* trace */
 263         ceph_decode_32_safe(&p, end, len, bad);
 264         if (len > 0) {
 265                 err = parse_reply_info_trace(&p, p+len, info, features);
 266                 if (err < 0)
 267                         goto out_bad;
 268         }
 269
 270         /* extra */
 271         ceph_decode_32_safe(&p, end, len, bad);
 272         if (len > 0) {
 273                 err = parse_reply_info_extra(&p, p+len, info, features);
 274                 if (err < 0)
 275                         goto out_bad;
 276         }
 277
 278         /* snap blob */
 279         ceph_decode_32_safe(&p, end, len, bad);
 280         info->snapblob_len = len;
 281         info->snapblob = p;
 282         p += len;
 283
 284         if (p != end)
 285                 goto bad;
 286         return 0;
 287
 288 bad:
 289         err = -EIO;
 290 out_bad:
 291         pr_err("mds parse_reply err %d\n", err);
 292         return err;
 293 }
 294
 295 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
 296 {
 297         kfree(info->dir_in);
 298 }
 299
 300
 301 /*
 302  * sessions
 303  */
 304 static const char *session_state_name(int s)
 305 {
 306         switch (s) {
 307         case CEPH_MDS_SESSION_NEW: return "new";
 308         case CEPH_MDS_SESSION_OPENING: return "opening";
 309         case CEPH_MDS_SESSION_OPEN: return "open";
 310         case CEPH_MDS_SESSION_HUNG: return "hung";
 311         case CEPH_MDS_SESSION_CLOSING: return "closing";
 312         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
 313         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
 314         default: return "???";
 315         }
 316 }
 317
 318 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
 319 {
 320         if (atomic_inc_not_zero(&s->s_ref)) {
 321                 dout("mdsc get_session %p %d -> %d\n", s,
 322                      atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
 323                 return s;
 324         } else {
 325                 dout("mdsc get_session %p 0 -- FAIL", s);
 326                 return NULL;
 327         }
 328 }
 329
 330 void ceph_put_mds_session(struct ceph_mds_session *s)
 331 {
 332         dout("mdsc put_session %p %d -> %d\n", s,
 333              atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
 334         if (atomic_dec_and_test(&s->s_ref)) {
 335                 if (s->s_authorizer)
 336                      s->s_mdsc->fsc->client->monc.auth->ops->destroy_authorizer(
 337                              s->s_mdsc->fsc->client->monc.auth,
 338                              s->s_authorizer);
 339                 kfree(s);
 340         }
 341 }
 342
 343 /*
 344  * called under mdsc->mutex
 345  */
 346 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
 347                                                    int mds)
 348 {
 349         struct ceph_mds_session *session;
 350
 351         if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
 352                 return NULL;
 353         session = mdsc->sessions[mds];
 354         dout("lookup_mds_session %p %d\n", session,
 355              atomic_read(&session->s_ref));
 356         get_session(session);
 357         return session;
 358 }
 359
 360 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
 361 {
 362         if (mds >= mdsc->max_sessions)
 363                 return false;
 364         return mdsc->sessions[mds];
 365 }
 366
 367 static int __verify_registered_session(struct ceph_mds_client *mdsc,
 368                                        struct ceph_mds_session *s)
 369 {
 370         if (s->s_mds >= mdsc->max_sessions ||
 371             mdsc->sessions[s->s_mds] != s)
 372                 return -ENOENT;
 373         return 0;
 374 }
 375
 376 /*
 377  * create+register a new session for given mds.
 378  * called under mdsc->mutex.
 379  */
 380 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 381                                                  int mds)
 382 {
 383         struct ceph_mds_session *s;
 384
 385         s = kzalloc(sizeof(*s), GFP_NOFS);
 386         if (!s)
 387                 return ERR_PTR(-ENOMEM);
 388         s->s_mdsc = mdsc;
 389         s->s_mds = mds;
 390         s->s_state = CEPH_MDS_SESSION_NEW;
 391         s->s_ttl = 0;
 392         s->s_seq = 0;
 393         mutex_init(&s->s_mutex);
 394
 395         ceph_con_init(mdsc->fsc->client->msgr, &s->s_con);
 396         s->s_con.private = s;
 397         s->s_con.ops = &mds_con_ops;
 398         s->s_con.peer_name.type = CEPH_ENTITY_TYPE_MDS;
 399         s->s_con.peer_name.num = cpu_to_le64(mds);
 400
 401         spin_lock_init(&s->s_cap_lock);
 402         s->s_cap_gen = 0;
 403         s->s_cap_ttl = 0;
 404         s->s_renew_requested = 0;
 405         s->s_renew_seq = 0;
 406         INIT_LIST_HEAD(&s->s_caps);
 407         s->s_nr_caps = 0;
 408         s->s_trim_caps = 0;
 409         atomic_set(&s->s_ref, 1);
 410         INIT_LIST_HEAD(&s->s_waiting);
 411         INIT_LIST_HEAD(&s->s_unsafe);
 412         s->s_num_cap_releases = 0;
 413         s->s_cap_iterator = NULL;
 414         INIT_LIST_HEAD(&s->s_cap_releases);
 415         INIT_LIST_HEAD(&s->s_cap_releases_done);
 416         INIT_LIST_HEAD(&s->s_cap_flushing);
 417         INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
 418
 419         dout("register_session mds%d\n", mds);
 420         if (mds >= mdsc->max_sessions) {
 421                 int newmax = 1 << get_count_order(mds+1);
 422                 struct ceph_mds_session **sa;
 423
 424                 dout("register_session realloc to %d\n", newmax);
 425                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 426                 if (sa == NULL)
 427                         goto fail_realloc;
 428                 if (mdsc->sessions) {
 429                         memcpy(sa, mdsc->sessions,
 430                                mdsc->max_sessions * sizeof(void *));
 431                         kfree(mdsc->sessions);
 432                 }
 433                 mdsc->sessions = sa;
 434                 mdsc->max_sessions = newmax;
 435         }
 436         mdsc->sessions[mds] = s;
 437         atomic_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
 438
 439         ceph_con_open(&s->s_con, ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
 440
 441         return s;
 442
 443 fail_realloc:
 444         kfree(s);
 445         return ERR_PTR(-ENOMEM);
 446 }
 447
 448 /*
 449  * called under mdsc->mutex
 450  */
 451 static void __unregister_session(struct ceph_mds_client *mdsc,
 452                                struct ceph_mds_session *s)
 453 {
 454         dout("__unregister_session mds%d %p\n", s->s_mds, s);
 455         BUG_ON(mdsc->sessions[s->s_mds] != s);
 456         mdsc->sessions[s->s_mds] = NULL;
 457         ceph_con_close(&s->s_con);
 458         ceph_put_mds_session(s);
 459 }
 460
 461 /*
 462  * drop session refs in request.
 463  *
 464  * should be last request ref, or hold mdsc->mutex
 465  */
 466 static void put_request_session(struct ceph_mds_request *req)
 467 {
 468         if (req->r_session) {
 469                 ceph_put_mds_session(req->r_session);
 470                 req->r_session = NULL;
 471         }
 472 }
 473
 474 void ceph_mdsc_release_request(struct kref *kref)
 475 {
 476         struct ceph_mds_request *req = container_of(kref,
 477                                                     struct ceph_mds_request,
 478                                                     r_kref);
 479         if (req->r_request)
 480                 ceph_msg_put(req->r_request);
 481         if (req->r_reply) {
 482                 ceph_msg_put(req->r_reply);
 483                 destroy_reply_info(&req->r_reply_info);
 484         }
 485         if (req->r_inode) {
 486                 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
 487                 iput(req->r_inode);
 488         }
 489         if (req->r_locked_dir)
 490                 ceph_put_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
 491         if (req->r_target_inode)
 492                 iput(req->r_target_inode);
 493         if (req->r_dentry)
 494                 dput(req->r_dentry);
 495         if (req->r_old_dentry) {
 496                 /*
 497                  * track (and drop pins for) r_old_dentry_dir
 498                  * separately, since r_old_dentry's d_parent may have
 499                  * changed between the dir mutex being dropped and
 500                  * this request being freed.
 501                  */
 502                 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
 503                                   CEPH_CAP_PIN);
 504                 dput(req->r_old_dentry);
 505                 iput(req->r_old_dentry_dir);
 506         }
 507         kfree(req->r_path1);
 508         kfree(req->r_path2);
 509         put_request_session(req);
 510         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
 511         kfree(req);
 512 }
 513
 514 /*
 515  * lookup session, bump ref if found.
 516  *
 517  * called under mdsc->mutex.
 518  */
 519 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
 520                                              u64 tid)
 521 {
 522         struct ceph_mds_request *req;
 523         struct rb_node *n = mdsc->request_tree.rb_node;
 524
 525         while (n) {
 526                 req = rb_entry(n, struct ceph_mds_request, r_node);
 527                 if (tid < req->r_tid)
 528                         n = n->rb_left;
 529                 else if (tid > req->r_tid)
 530                         n = n->rb_right;
 531                 else {
 532                         ceph_mdsc_get_request(req);
 533                         return req;
 534                 }
 535         }
 536         return NULL;
 537 }
 538
 539 static void __insert_request(struct ceph_mds_client *mdsc,
 540                              struct ceph_mds_request *new)
 541 {
 542         struct rb_node **p = &mdsc->request_tree.rb_node;
 543         struct rb_node *parent = NULL;
 544         struct ceph_mds_request *req = NULL;
 545
 546         while (*p) {
 547                 parent = *p;
 548                 req = rb_entry(parent, struct ceph_mds_request, r_node);
 549                 if (new->r_tid < req->r_tid)
 550                         p = &(*p)->rb_left;
 551                 else if (new->r_tid > req->r_tid)
 552                         p = &(*p)->rb_right;
 553                 else
 554                         BUG();
 555         }
 556
 557         rb_link_node(&new->r_node, parent, p);
 558         rb_insert_color(&new->r_node, &mdsc->request_tree);
 559 }
 560
 561 /*
 562  * Register an in-flight request, and assign a tid.  Link to directory
 563  * are modifying (if any).
 564  *
 565  * Called under mdsc->mutex.
 566  */
 567 static void __register_request(struct ceph_mds_client *mdsc,
 568                                struct ceph_mds_request *req,
 569                                struct inode *dir)
 570 {
 571         req->r_tid = ++mdsc->last_tid;
 572         if (req->r_num_caps)
 573                 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
 574                                   req->r_num_caps);
 575         dout("__register_request %p tid %lld\n", req, req->r_tid);
 576         ceph_mdsc_get_request(req);
 577         __insert_request(mdsc, req);
 578
 579         req->r_uid = current_fsuid();
 580         req->r_gid = current_fsgid();
 581
 582         if (dir) {
 583                 struct ceph_inode_info *ci = ceph_inode(dir);
 584
 585                 ihold(dir);
 586                 spin_lock(&ci->i_unsafe_lock);
 587                 req->r_unsafe_dir = dir;
 588                 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
 589                 spin_unlock(&ci->i_unsafe_lock);
 590         }
 591 }
 592
 593 static void __unregister_request(struct ceph_mds_client *mdsc,
 594                                  struct ceph_mds_request *req)
 595 {
 596         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 597         rb_erase(&req->r_node, &mdsc->request_tree);
 598         RB_CLEAR_NODE(&req->r_node);
 599
 600         if (req->r_unsafe_dir) {
 601                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 602
 603                 spin_lock(&ci->i_unsafe_lock);
 604                 list_del_init(&req->r_unsafe_dir_item);
 605                 spin_unlock(&ci->i_unsafe_lock);
 606
 607                 iput(req->r_unsafe_dir);
 608                 req->r_unsafe_dir = NULL;
 609         }
 610
 611         ceph_mdsc_put_request(req);
 612 }
 613
 614 /*
 615  * Choose mds to send request to next.  If there is a hint set in the
 616  * request (e.g., due to a prior forward hint from the mds), use that.
 617  * Otherwise, consult frag tree and/or caps to identify the
 618  * appropriate mds.  If all else fails, choose randomly.
 619  *
 620  * Called under mdsc->mutex.
 621  */
 622 static struct dentry *get_nonsnap_parent(struct dentry *dentry)
 623 {
 624         /*
 625          * we don't need to worry about protecting the d_parent access
 626          * here because we never renaming inside the snapped namespace
 627          * except to resplice to another snapdir, and either the old or new
 628          * result is a valid result.
 629          */
 630         while (!IS_ROOT(dentry) && ceph_snap(dentry->d_inode) != CEPH_NOSNAP)
 631                 dentry = dentry->d_parent;
 632         return dentry;
 633 }
 634
 635 static int __choose_mds(struct ceph_mds_client *mdsc,
 636                         struct ceph_mds_request *req)
 637 {
 638         struct inode *inode;
 639         struct ceph_inode_info *ci;
 640         struct ceph_cap *cap;
 641         int mode = req->r_direct_mode;
 642         int mds = -1;
 643         u32 hash = req->r_direct_hash;
 644         bool is_hash = req->r_direct_is_hash;
 645
 646         /*
 647          * is there a specific mds we should try?  ignore hint if we have
 648          * no session and the mds is not up (active or recovering).
 649          */
 650         if (req->r_resend_mds >= 0 &&
 651             (__have_session(mdsc, req->r_resend_mds) ||
 652              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 653                 dout("choose_mds using resend_mds mds%d\n",
 654                      req->r_resend_mds);
 655                 return req->r_resend_mds;
 656         }
 657
 658         if (mode == USE_RANDOM_MDS)
 659                 goto random;
 660
 661         inode = NULL;
 662         if (req->r_inode) {
 663                 inode = req->r_inode;
 664         } else if (req->r_dentry) {
 665                 /* ignore race with rename; old or new d_parent is okay */
 666                 struct dentry *parent = req->r_dentry->d_parent;
 667                 struct inode *dir = parent->d_inode;
 668
 669                 if (dir->i_sb != mdsc->fsc->sb) {
 670                         /* not this fs! */
 671                         inode = req->r_dentry->d_inode;
 672                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
 673                         /* direct snapped/virtual snapdir requests
 674                          * based on parent dir inode */
 675                         struct dentry *dn = get_nonsnap_parent(parent);
 676                         inode = dn->d_inode;
 677                         dout("__choose_mds using nonsnap parent %p\n", inode);
 678                 } else if (req->r_dentry->d_inode) {
 679                         /* dentry target */
 680                         inode = req->r_dentry->d_inode;
 681                 } else {
 682                         /* dir + name */
 683                         inode = dir;
 684                         hash = ceph_dentry_hash(dir, req->r_dentry);
 685                         is_hash = true;
 686                 }
 687         }
 688
 689         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 690              (int)hash, mode);
 691         if (!inode)
 692                 goto random;
 693         ci = ceph_inode(inode);
 694
 695         if (is_hash && S_ISDIR(inode->i_mode)) {
 696                 struct ceph_inode_frag frag;
 697                 int found;
 698
 699                 ceph_choose_frag(ci, hash, &frag, &found);
 700                 if (found) {
 701                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 702                                 u8 r;
 703
 704                                 /* choose a random replica */
 705                                 get_random_bytes(&r, 1);
 706                                 r %= frag.ndist;
 707                                 mds = frag.dist[r];
 708                                 dout("choose_mds %p %llx.%llx "
 709                                      "frag %u mds%d (%d/%d)\n",
 710                                      inode, ceph_vinop(inode),
 711                                      frag.frag, mds,
 712                                      (int)r, frag.ndist);
 713                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 714                                     CEPH_MDS_STATE_ACTIVE)
 715                                         return mds;
 716                         }
 717
 718                         /* since this file/dir wasn't known to be
 719                          * replicated, then we want to look for the
 720                          * authoritative mds. */
 721                         mode = USE_AUTH_MDS;
 722                         if (frag.mds >= 0) {
 723                                 /* choose auth mds */
 724                                 mds = frag.mds;
 725                                 dout("choose_mds %p %llx.%llx "
 726                                      "frag %u mds%d (auth)\n",
 727                                      inode, ceph_vinop(inode), frag.frag, mds);
 728                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 729                                     CEPH_MDS_STATE_ACTIVE)
 730                                         return mds;
 731                         }
 732                 }
 733         }
 734
 735         spin_lock(&inode->i_lock);
 736         cap = NULL;
 737         if (mode == USE_AUTH_MDS)
 738                 cap = ci->i_auth_cap;
 739         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
 740                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
 741         if (!cap) {
 742                 spin_unlock(&inode->i_lock);
 743                 goto random;
 744         }
 745         mds = cap->session->s_mds;
 746         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
 747              inode, ceph_vinop(inode), mds,
 748              cap == ci->i_auth_cap ? "auth " : "", cap);
 749         spin_unlock(&inode->i_lock);
 750         return mds;
 751
 752 random:
 753         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
 754         dout("choose_mds chose random mds%d\n", mds);
 755         return mds;
 756 }
 757
 758
 759 /*
 760  * session messages
 761  */
 762 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
 763 {
 764         struct ceph_msg *msg;
 765         struct ceph_mds_session_head *h;
 766
 767         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
 768                            false);
 769         if (!msg) {
 770                 pr_err("create_session_msg ENOMEM creating msg\n");
 771                 return NULL;
 772         }
 773         h = msg->front.iov_base;
 774         h->op = cpu_to_le32(op);
 775         h->seq = cpu_to_le64(seq);
 776         return msg;
 777 }
 778
 779 /*
 780  * send session open request.
 781  *
 782  * called under mdsc->mutex
 783  */
 784 static int __open_session(struct ceph_mds_client *mdsc,
 785                           struct ceph_mds_session *session)
 786 {
 787         struct ceph_msg *msg;
 788         int mstate;
 789         int mds = session->s_mds;
 790
 791         /* wait for mds to go active? */
 792         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
 793         dout("open_session to mds%d (%s)\n", mds,
 794              ceph_mds_state_name(mstate));
 795         session->s_state = CEPH_MDS_SESSION_OPENING;
 796         session->s_renew_requested = jiffies;
 797
 798         /* send connect message */
 799         msg = create_session_msg(CEPH_SESSION_REQUEST_OPEN, session->s_seq);
 800         if (!msg)
 801                 return -ENOMEM;
 802         ceph_con_send(&session->s_con, msg);
 803         return 0;
 804 }
 805
 806 /*
 807  * open sessions for any export targets for the given mds
 808  *
 809  * called under mdsc->mutex
 810  */
 811 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
 812                                           struct ceph_mds_session *session)
 813 {
 814         struct ceph_mds_info *mi;
 815         struct ceph_mds_session *ts;
 816         int i, mds = session->s_mds;
 817         int target;
 818
 819         if (mds >= mdsc->mdsmap->m_max_mds)
 820                 return;
 821         mi = &mdsc->mdsmap->m_info[mds];
 822         dout("open_export_target_sessions for mds%d (%d targets)\n",
 823              session->s_mds, mi->num_export_targets);
 824
 825         for (i = 0; i < mi->num_export_targets; i++) {
 826                 target = mi->export_targets[i];
 827                 ts = __ceph_lookup_mds_session(mdsc, target);
 828                 if (!ts) {
 829                         ts = register_session(mdsc, target);
 830                         if (IS_ERR(ts))
 831                                 return;
 832                 }
 833                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
 834                     session->s_state == CEPH_MDS_SESSION_CLOSING)
 835                         __open_session(mdsc, session);
 836                 else
 837                         dout(" mds%d target mds%d %p is %s\n", session->s_mds,
 838                              i, ts, session_state_name(ts->s_state));
 839                 ceph_put_mds_session(ts);
 840         }
 841 }
 842
 843 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
 844                                            struct ceph_mds_session *session)
 845 {
 846         mutex_lock(&mdsc->mutex);
 847         __open_export_target_sessions(mdsc, session);
 848         mutex_unlock(&mdsc->mutex);
 849 }
 850
 851 /*
 852  * session caps
 853  */
 854
 855 /*
 856  * Free preallocated cap messages assigned to this session
 857  */
 858 static void cleanup_cap_releases(struct ceph_mds_session *session)
 859 {
 860         struct ceph_msg *msg;
 861
 862         spin_lock(&session->s_cap_lock);
 863         while (!list_empty(&session->s_cap_releases)) {
 864                 msg = list_first_entry(&session->s_cap_releases,
 865                                        struct ceph_msg, list_head);
 866                 list_del_init(&msg->list_head);
 867                 ceph_msg_put(msg);
 868         }
 869         while (!list_empty(&session->s_cap_releases_done)) {
 870                 msg = list_first_entry(&session->s_cap_releases_done,
 871                                        struct ceph_msg, list_head);
 872                 list_del_init(&msg->list_head);
 873                 ceph_msg_put(msg);
 874         }
 875         spin_unlock(&session->s_cap_lock);
 876 }
 877
 878 /*
 879  * Helper to safely iterate over all caps associated with a session, with
 880  * special care taken to handle a racing __ceph_remove_cap().
 881  *
 882  * Caller must hold session s_mutex.
 883  */
 884 static int iterate_session_caps(struct ceph_mds_session *session,
 885                                  int (*cb)(struct inode *, struct ceph_cap *,
 886                                             void *), void *arg)
 887 {
 888         struct list_head *p;
 889         struct ceph_cap *cap;
 890         struct inode *inode, *last_inode = NULL;
 891         struct ceph_cap *old_cap = NULL;
 892         int ret;
 893
 894         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
 895         spin_lock(&session->s_cap_lock);
 896         p = session->s_caps.next;
 897         while (p != &session->s_caps) {
 898                 cap = list_entry(p, struct ceph_cap, session_caps);
 899                 inode = igrab(&cap->ci->vfs_inode);
 900                 if (!inode) {
 901                         p = p->next;
 902                         continue;
 903                 }
 904                 session->s_cap_iterator = cap;
 905                 spin_unlock(&session->s_cap_lock);
 906
 907                 if (last_inode) {
 908                         iput(last_inode);
 909                         last_inode = NULL;
 910                 }
 911                 if (old_cap) {
 912                         ceph_put_cap(session->s_mdsc, old_cap);
 913                         old_cap = NULL;
 914                 }
 915
 916                 ret = cb(inode, cap, arg);
 917                 last_inode = inode;
 918
 919                 spin_lock(&session->s_cap_lock);
 920                 p = p->next;
 921                 if (cap->ci == NULL) {
 922                         dout("iterate_session_caps  finishing cap %p removal\n",
 923                              cap);
 924                         BUG_ON(cap->session != session);
 925                         list_del_init(&cap->session_caps);
 926                         session->s_nr_caps--;
 927                         cap->session = NULL;
 928                         old_cap = cap;  /* put_cap it w/o locks held */
 929                 }
 930                 if (ret < 0)
 931                         goto out;
 932         }
 933         ret = 0;
 934 out:
 935         session->s_cap_iterator = NULL;
 936         spin_unlock(&session->s_cap_lock);
 937
 938         if (last_inode)
 939                 iput(last_inode);
 940         if (old_cap)
 941                 ceph_put_cap(session->s_mdsc, old_cap);
 942
 943         return ret;
 944 }
 945
 946 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
 947                                   void *arg)
 948 {
 949         struct ceph_inode_info *ci = ceph_inode(inode);
 950         int drop = 0;
 951
 952         dout("removing cap %p, ci is %p, inode is %p\n",
 953              cap, ci, &ci->vfs_inode);
 954         spin_lock(&inode->i_lock);
 955         __ceph_remove_cap(cap);
 956         if (!__ceph_is_any_real_caps(ci)) {
 957                 struct ceph_mds_client *mdsc =
 958                         ceph_sb_to_client(inode->i_sb)->mdsc;
 959
 960                 spin_lock(&mdsc->cap_dirty_lock);
 961                 if (!list_empty(&ci->i_dirty_item)) {
 962                         pr_info(" dropping dirty %s state for %p %lld\n",
 963                                 ceph_cap_string(ci->i_dirty_caps),
 964                                 inode, ceph_ino(inode));
 965                         ci->i_dirty_caps = 0;
 966                         list_del_init(&ci->i_dirty_item);
 967                         drop = 1;
 968                 }
 969                 if (!list_empty(&ci->i_flushing_item)) {
 970                         pr_info(" dropping dirty+flushing %s state for %p %lld\n",
 971                                 ceph_cap_string(ci->i_flushing_caps),
 972                                 inode, ceph_ino(inode));
 973                         ci->i_flushing_caps = 0;
 974                         list_del_init(&ci->i_flushing_item);
 975                         mdsc->num_cap_flushing--;
 976                         drop = 1;
 977                 }
 978                 if (drop && ci->i_wrbuffer_ref) {
 979                         pr_info(" dropping dirty data for %p %lld\n",
 980                                 inode, ceph_ino(inode));
 981                         ci->i_wrbuffer_ref = 0;
 982                         ci->i_wrbuffer_ref_head = 0;
 983                         drop++;
 984                 }
 985                 spin_unlock(&mdsc->cap_dirty_lock);
 986         }
 987         spin_unlock(&inode->i_lock);
 988         while (drop--)
 989                 iput(inode);
 990         return 0;
 991 }
 992
 993 /*
 994  * caller must hold session s_mutex
 995  */
 996 static void remove_session_caps(struct ceph_mds_session *session)
 997 {
 998         dout("remove_session_caps on %p\n", session);
 999         iterate_session_caps(session, remove_session_caps_cb, NULL);
1000         BUG_ON(session->s_nr_caps > 0);
1001         BUG_ON(!list_empty(&session->s_cap_flushing));
1002         cleanup_cap_releases(session);
1003 }
1004
1005 /*
1006  * wake up any threads waiting on this session's caps.  if the cap is
1007  * old (didn't get renewed on the client reconnect), remove it now.
1008  *
1009  * caller must hold s_mutex.
1010  */
1011 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1012                               void *arg)
1013 {
1014         struct ceph_inode_info *ci = ceph_inode(inode);
1015
1016         wake_up_all(&ci->i_cap_wq);
1017         if (arg) {
1018                 spin_lock(&inode->i_lock);
1019                 ci->i_wanted_max_size = 0;
1020                 ci->i_requested_max_size = 0;
1021                 spin_unlock(&inode->i_lock);
1022         }
1023         return 0;
1024 }
1025
1026 static void wake_up_session_caps(struct ceph_mds_session *session,
1027                                  int reconnect)
1028 {
1029         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1030         iterate_session_caps(session, wake_up_session_cb,
1031                              (void *)(unsigned long)reconnect);
1032 }
1033
1034 /*
1035  * Send periodic message to MDS renewing all currently held caps.  The
1036  * ack will reset the expiration for all caps from this session.
1037  *
1038  * caller holds s_mutex
1039  */
1040 static int send_renew_caps(struct ceph_mds_client *mdsc,
1041                            struct ceph_mds_session *session)
1042 {
1043         struct ceph_msg *msg;
1044         int state;
1045
1046         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1047             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1048                 pr_info("mds%d caps stale\n", session->s_mds);
1049         session->s_renew_requested = jiffies;
1050
1051         /* do not try to renew caps until a recovering mds has reconnected
1052          * with its clients. */
1053         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1054         if (state < CEPH_MDS_STATE_RECONNECT) {
1055                 dout("send_renew_caps ignoring mds%d (%s)\n",
1056                      session->s_mds, ceph_mds_state_name(state));
1057                 return 0;
1058         }
1059
1060         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1061                 ceph_mds_state_name(state));
1062         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1063                                  ++session->s_renew_seq);
1064         if (!msg)
1065                 return -ENOMEM;
1066         ceph_con_send(&session->s_con, msg);
1067         return 0;
1068 }
1069
1070 /*
1071  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1072  *
1073  * Called under session->s_mutex
1074  */
1075 static void renewed_caps(struct ceph_mds_client *mdsc,
1076                          struct ceph_mds_session *session, int is_renew)
1077 {
1078         int was_stale;
1079         int wake = 0;
1080
1081         spin_lock(&session->s_cap_lock);
1082         was_stale = is_renew && (session->s_cap_ttl == 0 ||
1083                                  time_after_eq(jiffies, session->s_cap_ttl));
1084
1085         session->s_cap_ttl = session->s_renew_requested +
1086                 mdsc->mdsmap->m_session_timeout*HZ;
1087
1088         if (was_stale) {
1089                 if (time_before(jiffies, session->s_cap_ttl)) {
1090                         pr_info("mds%d caps renewed\n", session->s_mds);
1091                         wake = 1;
1092                 } else {
1093                         pr_info("mds%d caps still stale\n", session->s_mds);
1094                 }
1095         }
1096         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1097              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1098              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1099         spin_unlock(&session->s_cap_lock);
1100
1101         if (wake)
1102                 wake_up_session_caps(session, 0);
1103 }
1104
1105 /*
1106  * send a session close request
1107  */
1108 static int request_close_session(struct ceph_mds_client *mdsc,
1109                                  struct ceph_mds_session *session)
1110 {
1111         struct ceph_msg *msg;
1112
1113         dout("request_close_session mds%d state %s seq %lld\n",
1114              session->s_mds, session_state_name(session->s_state),
1115              session->s_seq);
1116         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1117         if (!msg)
1118                 return -ENOMEM;
1119         ceph_con_send(&session->s_con, msg);
1120         return 0;
1121 }
1122
1123 /*
1124  * Called with s_mutex held.
1125  */
1126 static int __close_session(struct ceph_mds_client *mdsc,
1127                          struct ceph_mds_session *session)
1128 {
1129         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1130                 return 0;
1131         session->s_state = CEPH_MDS_SESSION_CLOSING;
1132         return request_close_session(mdsc, session);
1133 }
1134
1135 /*
1136  * Trim old(er) caps.
1137  *
1138  * Because we can't cache an inode without one or more caps, we do
1139  * this indirectly: if a cap is unused, we prune its aliases, at which
1140  * point the inode will hopefully get dropped to.
1141  *
1142  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1143  * memory pressure from the MDS, though, so it needn't be perfect.
1144  */
1145 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1146 {
1147         struct ceph_mds_session *session = arg;
1148         struct ceph_inode_info *ci = ceph_inode(inode);
1149         int used, oissued, mine;
1150
1151         if (session->s_trim_caps <= 0)
1152                 return -1;
1153
1154         spin_lock(&inode->i_lock);
1155         mine = cap->issued | cap->implemented;
1156         used = __ceph_caps_used(ci);
1157         oissued = __ceph_caps_issued_other(ci, cap);
1158
1159         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
1160              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1161              ceph_cap_string(used));
1162         if (ci->i_dirty_caps)
1163                 goto out;   /* dirty caps */
1164         if ((used & ~oissued) & mine)
1165                 goto out;   /* we need these caps */
1166
1167         session->s_trim_caps--;
1168         if (oissued) {
1169                 /* we aren't the only cap.. just remove us */
1170                 __ceph_remove_cap(cap);
1171         } else {
1172                 /* try to drop referring dentries */
1173                 spin_unlock(&inode->i_lock);
1174                 d_prune_aliases(inode);
1175                 dout("trim_caps_cb %p cap %p  pruned, count now %d\n",
1176                      inode, cap, atomic_read(&inode->i_count));
1177                 return 0;
1178         }
1179
1180 out:
1181         spin_unlock(&inode->i_lock);
1182         return 0;
1183 }
1184
1185 /*
1186  * Trim session cap count down to some max number.
1187  */
1188 static int trim_caps(struct ceph_mds_client *mdsc,
1189                      struct ceph_mds_session *session,
1190                      int max_caps)
1191 {
1192         int trim_caps = session->s_nr_caps - max_caps;
1193
1194         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1195              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1196         if (trim_caps > 0) {
1197                 session->s_trim_caps = trim_caps;
1198                 iterate_session_caps(session, trim_caps_cb, session);
1199                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1200                      session->s_mds, session->s_nr_caps, max_caps,
1201                         trim_caps - session->s_trim_caps);
1202                 session->s_trim_caps = 0;
1203         }
1204         return 0;
1205 }
1206
1207 /*
1208  * Allocate cap_release messages.  If there is a partially full message
1209  * in the queue, try to allocate enough to cover it's remainder, so that
1210  * we can send it immediately.
1211  *
1212  * Called under s_mutex.
1213  */
1214 int ceph_add_cap_releases(struct ceph_mds_client *mdsc,
1215                           struct ceph_mds_session *session)
1216 {
1217         struct ceph_msg *msg, *partial = NULL;
1218         struct ceph_mds_cap_release *head;
1219         int err = -ENOMEM;
1220         int extra = mdsc->fsc->mount_options->cap_release_safety;
1221         int num;
1222
1223         dout("add_cap_releases %p mds%d extra %d\n", session, session->s_mds,
1224              extra);
1225
1226         spin_lock(&session->s_cap_lock);
1227
1228         if (!list_empty(&session->s_cap_releases)) {
1229                 msg = list_first_entry(&session->s_cap_releases,
1230                                        struct ceph_msg,
1231                                  list_head);
1232                 head = msg->front.iov_base;
1233                 num = le32_to_cpu(head->num);
1234                 if (num) {
1235                         dout(" partial %p with (%d/%d)\n", msg, num,
1236                              (int)CEPH_CAPS_PER_RELEASE);
1237                         extra += CEPH_CAPS_PER_RELEASE - num;
1238                         partial = msg;
1239                 }
1240         }
1241         while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1242                 spin_unlock(&session->s_cap_lock);
1243                 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
1244                                    GFP_NOFS, false);
1245                 if (!msg)
1246                         goto out_unlocked;
1247                 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1248                      (int)msg->front.iov_len);
1249                 head = msg->front.iov_base;
1250                 head->num = cpu_to_le32(0);
1251                 msg->front.iov_len = sizeof(*head);
1252                 spin_lock(&session->s_cap_lock);
1253                 list_add(&msg->list_head, &session->s_cap_releases);
1254                 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1255         }
1256
1257         if (partial) {
1258                 head = partial->front.iov_base;
1259                 num = le32_to_cpu(head->num);
1260                 dout(" queueing partial %p with %d/%d\n", partial, num,
1261                      (int)CEPH_CAPS_PER_RELEASE);
1262                 list_move_tail(&partial->list_head,
1263                                &session->s_cap_releases_done);
1264                 session->s_num_cap_releases -= CEPH_CAPS_PER_RELEASE - num;
1265         }
1266         err = 0;
1267         spin_unlock(&session->s_cap_lock);
1268 out_unlocked:
1269         return err;
1270 }
1271
1272 /*
1273  * flush all dirty inode data to disk.
1274  *
1275  * returns true if we've flushed through want_flush_seq
1276  */
1277 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1278 {
1279         int mds, ret = 1;
1280
1281         dout("check_cap_flush want %lld\n", want_flush_seq);
1282         mutex_lock(&mdsc->mutex);
1283         for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1284                 struct ceph_mds_session *session = mdsc->sessions[mds];
1285
1286                 if (!session)
1287                         continue;
1288                 get_session(session);
1289                 mutex_unlock(&mdsc->mutex);
1290
1291                 mutex_lock(&session->s_mutex);
1292                 if (!list_empty(&session->s_cap_flushing)) {
1293                         struct ceph_inode_info *ci =
1294                                 list_entry(session->s_cap_flushing.next,
1295                                            struct ceph_inode_info,
1296                                            i_flushing_item);
1297                         struct inode *inode = &ci->vfs_inode;
1298
1299                         spin_lock(&inode->i_lock);
1300                         if (ci->i_cap_flush_seq <= want_flush_seq) {
1301                                 dout("check_cap_flush still flushing %p "
1302                                      "seq %lld <= %lld to mds%d\n", inode,
1303                                      ci->i_cap_flush_seq, want_flush_seq,
1304                                      session->s_mds);
1305                                 ret = 0;
1306                         }
1307                         spin_unlock(&inode->i_lock);
1308                 }
1309                 mutex_unlock(&session->s_mutex);
1310                 ceph_put_mds_session(session);
1311
1312                 if (!ret)
1313                         return ret;
1314                 mutex_lock(&mdsc->mutex);
1315         }
1316
1317         mutex_unlock(&mdsc->mutex);
1318         dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1319         return ret;
1320 }
1321
1322 /*
1323  * called under s_mutex
1324  */
1325 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1326                             struct ceph_mds_session *session)
1327 {
1328         struct ceph_msg *msg;
1329
1330         dout("send_cap_releases mds%d\n", session->s_mds);
1331         spin_lock(&session->s_cap_lock);
1332         while (!list_empty(&session->s_cap_releases_done)) {
1333                 msg = list_first_entry(&session->s_cap_releases_done,
1334                                  struct ceph_msg, list_head);
1335                 list_del_init(&msg->list_head);
1336                 spin_unlock(&session->s_cap_lock);
1337                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1338                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1339                 ceph_con_send(&session->s_con, msg);
1340                 spin_lock(&session->s_cap_lock);
1341         }
1342         spin_unlock(&session->s_cap_lock);
1343 }
1344
1345 static void discard_cap_releases(struct ceph_mds_client *mdsc,
1346                                  struct ceph_mds_session *session)
1347 {
1348         struct ceph_msg *msg;
1349         struct ceph_mds_cap_release *head;
1350         unsigned num;
1351
1352         dout("discard_cap_releases mds%d\n", session->s_mds);
1353         spin_lock(&session->s_cap_lock);
1354
1355         /* zero out the in-progress message */
1356         msg = list_first_entry(&session->s_cap_releases,
1357                                struct ceph_msg, list_head);
1358         head = msg->front.iov_base;
1359         num = le32_to_cpu(head->num);
1360         dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg, num);
1361         head->num = cpu_to_le32(0);
1362         session->s_num_cap_releases += num;
1363
1364         /* requeue completed messages */
1365         while (!list_empty(&session->s_cap_releases_done)) {
1366                 msg = list_first_entry(&session->s_cap_releases_done,
1367                                  struct ceph_msg, list_head);
1368                 list_del_init(&msg->list_head);
1369
1370                 head = msg->front.iov_base;
1371                 num = le32_to_cpu(head->num);
1372                 dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg,
1373                      num);
1374                 session->s_num_cap_releases += num;
1375                 head->num = cpu_to_le32(0);
1376                 msg->front.iov_len = sizeof(*head);
1377                 list_add(&msg->list_head, &session->s_cap_releases);
1378         }
1379
1380         spin_unlock(&session->s_cap_lock);
1381 }
1382
1383 /*
1384  * requests
1385  */
1386
1387 /*
1388  * Create an mds request.
1389  */
1390 struct ceph_mds_request *
1391 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1392 {
1393         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1394
1395         if (!req)
1396                 return ERR_PTR(-ENOMEM);
1397
1398         mutex_init(&req->r_fill_mutex);
1399         req->r_mdsc = mdsc;
1400         req->r_started = jiffies;
1401         req->r_resend_mds = -1;
1402         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1403         req->r_fmode = -1;
1404         kref_init(&req->r_kref);
1405         INIT_LIST_HEAD(&req->r_wait);
1406         init_completion(&req->r_completion);
1407         init_completion(&req->r_safe_completion);
1408         INIT_LIST_HEAD(&req->r_unsafe_item);
1409
1410         req->r_op = op;
1411         req->r_direct_mode = mode;
1412         return req;
1413 }
1414
1415 /*
1416  * return oldest (lowest) request, tid in request tree, 0 if none.
1417  *
1418  * called under mdsc->mutex.
1419  */
1420 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1421 {
1422         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1423                 return NULL;
1424         return rb_entry(rb_first(&mdsc->request_tree),
1425                         struct ceph_mds_request, r_node);
1426 }
1427
1428 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1429 {
1430         struct ceph_mds_request *req = __get_oldest_req(mdsc);
1431
1432         if (req)
1433                 return req->r_tid;
1434         return 0;
1435 }
1436
1437 /*
1438  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1439  * on build_path_from_dentry in fs/cifs/dir.c.
1440  *
1441  * If @stop_on_nosnap, generate path relative to the first non-snapped
1442  * inode.
1443  *
1444  * Encode hidden .snap dirs as a double /, i.e.
1445  *   foo/.snap/bar -> foo//bar
1446  */
1447 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1448                            int stop_on_nosnap)
1449 {
1450         struct dentry *temp;
1451         char *path;
1452         int len, pos;
1453         unsigned seq;
1454
1455         if (dentry == NULL)
1456                 return ERR_PTR(-EINVAL);
1457
1458 retry:
1459         len = 0;
1460         seq = read_seqbegin(&rename_lock);
1461         rcu_read_lock();
1462         for (temp = dentry; !IS_ROOT(temp);) {
1463                 struct inode *inode = temp->d_inode;
1464                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1465                         len++;  /* slash only */
1466                 else if (stop_on_nosnap && inode &&
1467                          ceph_snap(inode) == CEPH_NOSNAP)
1468                         break;
1469                 else
1470                         len += 1 + temp->d_name.len;
1471                 temp = temp->d_parent;
1472                 if (temp == NULL) {
1473                         rcu_read_unlock();
1474                         pr_err("build_path corrupt dentry %p\n", dentry);
1475                         return ERR_PTR(-EINVAL);
1476                 }
1477         }
1478         rcu_read_unlock();
1479         if (len)
1480                 len--;  /* no leading '/' */
1481
1482         path = kmalloc(len+1, GFP_NOFS);
1483         if (path == NULL)
1484                 return ERR_PTR(-ENOMEM);
1485         pos = len;
1486         path[pos] = 0;  /* trailing null */
1487         rcu_read_lock();
1488         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1489                 struct inode *inode;
1490
1491                 spin_lock(&temp->d_lock);
1492                 inode = temp->d_inode;
1493                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1494                         dout("build_path path+%d: %p SNAPDIR\n",
1495                              pos, temp);
1496                 } else if (stop_on_nosnap && inode &&
1497                            ceph_snap(inode) == CEPH_NOSNAP) {
1498                         break;
1499                 } else {
1500                         pos -= temp->d_name.len;
1501                         if (pos < 0) {
1502                                 spin_unlock(&temp->d_lock);
1503                                 break;
1504                         }
1505                         strncpy(path + pos, temp->d_name.name,
1506                                 temp->d_name.len);
1507                 }
1508                 spin_unlock(&temp->d_lock);
1509                 if (pos)
1510                         path[--pos] = '/';
1511                 temp = temp->d_parent;
1512                 if (temp == NULL) {
1513                         rcu_read_unlock();
1514                         pr_err("build_path corrupt dentry\n");
1515                         kfree(path);
1516                         return ERR_PTR(-EINVAL);
1517                 }
1518         }
1519         rcu_read_unlock();
1520         if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1521                 pr_err("build_path did not end path lookup where "
1522                        "expected, namelen is %d, pos is %d\n", len, pos);
1523                 /* presumably this is only possible if racing with a
1524                    rename of one of the parent directories (we can not
1525                    lock the dentries above us to prevent this, but
1526                    retrying should be harmless) */
1527                 kfree(path);
1528                 goto retry;
1529         }
1530
1531         *base = ceph_ino(temp->d_inode);
1532         *plen = len;
1533         dout("build_path on %p %d built %llx '%.*s'\n",
1534              dentry, dentry->d_count, *base, len, path);
1535         return path;
1536 }
1537
1538 static int build_dentry_path(struct dentry *dentry,
1539                              const char **ppath, int *ppathlen, u64 *pino,
1540                              int *pfreepath)
1541 {
1542         char *path;
1543
1544         if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1545                 *pino = ceph_ino(dentry->d_parent->d_inode);
1546                 *ppath = dentry->d_name.name;
1547                 *ppathlen = dentry->d_name.len;
1548                 return 0;
1549         }
1550         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1551         if (IS_ERR(path))
1552                 return PTR_ERR(path);
1553         *ppath = path;
1554         *pfreepath = 1;
1555         return 0;
1556 }
1557
1558 static int build_inode_path(struct inode *inode,
1559                             const char **ppath, int *ppathlen, u64 *pino,
1560                             int *pfreepath)
1561 {
1562         struct dentry *dentry;
1563         char *path;
1564
1565         if (ceph_snap(inode) == CEPH_NOSNAP) {
1566                 *pino = ceph_ino(inode);
1567                 *ppathlen = 0;
1568                 return 0;
1569         }
1570         dentry = d_find_alias(inode);
1571         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1572         dput(dentry);
1573         if (IS_ERR(path))
1574                 return PTR_ERR(path);
1575         *ppath = path;
1576         *pfreepath = 1;
1577         return 0;
1578 }
1579
1580 /*
1581  * request arguments may be specified via an inode *, a dentry *, or
1582  * an explicit ino+path.
1583  */
1584 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1585                                   const char *rpath, u64 rino,
1586                                   const char **ppath, int *pathlen,
1587                                   u64 *ino, int *freepath)
1588 {
1589         int r = 0;
1590
1591         if (rinode) {
1592                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1593                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1594                      ceph_snap(rinode));
1595         } else if (rdentry) {
1596                 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1597                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1598                      *ppath);
1599         } else if (rpath || rino) {
1600                 *ino = rino;
1601                 *ppath = rpath;
1602                 *pathlen = strlen(rpath);
1603                 dout(" path %.*s\n", *pathlen, rpath);
1604         }
1605
1606         return r;
1607 }
1608
1609 /*
1610  * called under mdsc->mutex
1611  */
1612 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1613                                                struct ceph_mds_request *req,
1614                                                int mds)
1615 {
1616         struct ceph_msg *msg;
1617         struct ceph_mds_request_head *head;
1618         const char *path1 = NULL;
1619         const char *path2 = NULL;
1620         u64 ino1 = 0, ino2 = 0;
1621         int pathlen1 = 0, pathlen2 = 0;
1622         int freepath1 = 0, freepath2 = 0;
1623         int len;
1624         u16 releases;
1625         void *p, *end;
1626         int ret;
1627
1628         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1629                               req->r_path1, req->r_ino1.ino,
1630                               &path1, &pathlen1, &ino1, &freepath1);
1631         if (ret < 0) {
1632                 msg = ERR_PTR(ret);
1633                 goto out;
1634         }
1635
1636         ret = set_request_path_attr(NULL, req->r_old_dentry,
1637                               req->r_path2, req->r_ino2.ino,
1638                               &path2, &pathlen2, &ino2, &freepath2);
1639         if (ret < 0) {
1640                 msg = ERR_PTR(ret);
1641                 goto out_free1;
1642         }
1643
1644         len = sizeof(*head) +
1645                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64));
1646
1647         /* calculate (max) length for cap releases */
1648         len += sizeof(struct ceph_mds_request_release) *
1649                 (!!req->r_inode_drop + !!req->r_dentry_drop +
1650                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1651         if (req->r_dentry_drop)
1652                 len += req->r_dentry->d_name.len;
1653         if (req->r_old_dentry_drop)
1654                 len += req->r_old_dentry->d_name.len;
1655
1656         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1657         if (!msg) {
1658                 msg = ERR_PTR(-ENOMEM);
1659                 goto out_free2;
1660         }
1661
1662         msg->hdr.tid = cpu_to_le64(req->r_tid);
1663
1664         head = msg->front.iov_base;
1665         p = msg->front.iov_base + sizeof(*head);
1666         end = msg->front.iov_base + msg->front.iov_len;
1667
1668         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1669         head->op = cpu_to_le32(req->r_op);
1670         head->caller_uid = cpu_to_le32(req->r_uid);
1671         head->caller_gid = cpu_to_le32(req->r_gid);
1672         head->args = req->r_args;
1673
1674         ceph_encode_filepath(&p, end, ino1, path1);
1675         ceph_encode_filepath(&p, end, ino2, path2);
1676
1677         /* make note of release offset, in case we need to replay */
1678         req->r_request_release_offset = p - msg->front.iov_base;
1679
1680         /* cap releases */
1681         releases = 0;
1682         if (req->r_inode_drop)
1683                 releases += ceph_encode_inode_release(&p,
1684                       req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1685                       mds, req->r_inode_drop, req->r_inode_unless, 0);
1686         if (req->r_dentry_drop)
1687                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1688                        mds, req->r_dentry_drop, req->r_dentry_unless);
1689         if (req->r_old_dentry_drop)
1690                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1691                        mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1692         if (req->r_old_inode_drop)
1693                 releases += ceph_encode_inode_release(&p,
1694                       req->r_old_dentry->d_inode,
1695                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1696         head->num_releases = cpu_to_le16(releases);
1697
1698         BUG_ON(p > end);
1699         msg->front.iov_len = p - msg->front.iov_base;
1700         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1701
1702         msg->pages = req->r_pages;
1703         msg->nr_pages = req->r_num_pages;
1704         msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1705         msg->hdr.data_off = cpu_to_le16(0);
1706
1707 out_free2:
1708         if (freepath2)
1709                 kfree((char *)path2);
1710 out_free1:
1711         if (freepath1)
1712                 kfree((char *)path1);
1713 out:
1714         return msg;
1715 }
1716
1717 /*
1718  * called under mdsc->mutex if error, under no mutex if
1719  * success.
1720  */
1721 static void complete_request(struct ceph_mds_client *mdsc,
1722                              struct ceph_mds_request *req)
1723 {
1724         if (req->r_callback)
1725                 req->r_callback(mdsc, req);
1726         else
1727                 complete_all(&req->r_completion);
1728 }
1729
1730 /*
1731  * called under mdsc->mutex
1732  */
1733 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1734                                   struct ceph_mds_request *req,
1735                                   int mds)
1736 {
1737         struct ceph_mds_request_head *rhead;
1738         struct ceph_msg *msg;
1739         int flags = 0;
1740
1741         req->r_attempts++;
1742         if (req->r_inode) {
1743                 struct ceph_cap *cap =
1744                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
1745
1746                 if (cap)
1747                         req->r_sent_on_mseq = cap->mseq;
1748                 else
1749                         req->r_sent_on_mseq = -1;
1750         }
1751         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1752              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1753
1754         if (req->r_got_unsafe) {
1755                 /*
1756                  * Replay.  Do not regenerate message (and rebuild
1757                  * paths, etc.); just use the original message.
1758                  * Rebuilding paths will break for renames because
1759                  * d_move mangles the src name.
1760                  */
1761                 msg = req->r_request;
1762                 rhead = msg->front.iov_base;
1763
1764                 flags = le32_to_cpu(rhead->flags);
1765                 flags |= CEPH_MDS_FLAG_REPLAY;
1766                 rhead->flags = cpu_to_le32(flags);
1767
1768                 if (req->r_target_inode)
1769                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1770
1771                 rhead->num_retry = req->r_attempts - 1;
1772
1773                 /* remove cap/dentry releases from message */
1774                 rhead->num_releases = 0;
1775                 msg->hdr.front_len = cpu_to_le32(req->r_request_release_offset);
1776                 msg->front.iov_len = req->r_request_release_offset;
1777                 return 0;
1778         }
1779
1780         if (req->r_request) {
1781                 ceph_msg_put(req->r_request);
1782                 req->r_request = NULL;
1783         }
1784         msg = create_request_message(mdsc, req, mds);
1785         if (IS_ERR(msg)) {
1786                 req->r_err = PTR_ERR(msg);
1787                 complete_request(mdsc, req);
1788                 return PTR_ERR(msg);
1789         }
1790         req->r_request = msg;
1791
1792         rhead = msg->front.iov_base;
1793         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1794         if (req->r_got_unsafe)
1795                 flags |= CEPH_MDS_FLAG_REPLAY;
1796         if (req->r_locked_dir)
1797                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1798         rhead->flags = cpu_to_le32(flags);
1799         rhead->num_fwd = req->r_num_fwd;
1800         rhead->num_retry = req->r_attempts - 1;
1801         rhead->ino = 0;
1802
1803         dout(" r_locked_dir = %p\n", req->r_locked_dir);
1804         return 0;
1805 }
1806
1807 /*
1808  * send request, or put it on the appropriate wait list.
1809  */
1810 static int __do_request(struct ceph_mds_client *mdsc,
1811                         struct ceph_mds_request *req)
1812 {
1813         struct ceph_mds_session *session = NULL;
1814         int mds = -1;
1815         int err = -EAGAIN;
1816
1817         if (req->r_err || req->r_got_result)
1818                 goto out;
1819
1820         if (req->r_timeout &&
1821             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1822                 dout("do_request timed out\n");
1823                 err = -EIO;
1824                 goto finish;
1825         }
1826
1827         put_request_session(req);
1828
1829         mds = __choose_mds(mdsc, req);
1830         if (mds < 0 ||
1831             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1832                 dout("do_request no mds or not active, waiting for map\n");
1833                 list_add(&req->r_wait, &mdsc->waiting_for_map);
1834                 goto out;
1835         }
1836
1837         /* get, open session */
1838         session = __ceph_lookup_mds_session(mdsc, mds);
1839         if (!session) {
1840                 session = register_session(mdsc, mds);
1841                 if (IS_ERR(session)) {
1842                         err = PTR_ERR(session);
1843                         goto finish;
1844                 }
1845         }
1846         req->r_session = get_session(session);
1847
1848         dout("do_request mds%d session %p state %s\n", mds, session,
1849              session_state_name(session->s_state));
1850         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1851             session->s_state != CEPH_MDS_SESSION_HUNG) {
1852                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1853                     session->s_state == CEPH_MDS_SESSION_CLOSING)
1854                         __open_session(mdsc, session);
1855                 list_add(&req->r_wait, &session->s_waiting);
1856                 goto out_session;
1857         }
1858
1859         /* send request */
1860         req->r_resend_mds = -1;   /* forget any previous mds hint */
1861
1862         if (req->r_request_started == 0)   /* note request start time */
1863                 req->r_request_started = jiffies;
1864
1865         err = __prepare_send_request(mdsc, req, mds);
1866         if (!err) {
1867                 ceph_msg_get(req->r_request);
1868                 ceph_con_send(&session->s_con, req->r_request);
1869         }
1870
1871 out_session:
1872         ceph_put_mds_session(session);
1873 out:
1874         return err;
1875
1876 finish:
1877         req->r_err = err;
1878         complete_request(mdsc, req);
1879         goto out;
1880 }
1881
1882 /*
1883  * called under mdsc->mutex
1884  */
1885 static void __wake_requests(struct ceph_mds_client *mdsc,
1886                             struct list_head *head)
1887 {
1888         struct ceph_mds_request *req, *nreq;
1889
1890         list_for_each_entry_safe(req, nreq, head, r_wait) {
1891                 list_del_init(&req->r_wait);
1892                 __do_request(mdsc, req);
1893         }
1894 }
1895
1896 /*
1897  * Wake up threads with requests pending for @mds, so that they can
1898  * resubmit their requests to a possibly different mds.
1899  */
1900 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
1901 {
1902         struct ceph_mds_request *req;
1903         struct rb_node *p;
1904
1905         dout("kick_requests mds%d\n", mds);
1906         for (p = rb_first(&mdsc->request_tree); p; p = rb_next(p)) {
1907                 req = rb_entry(p, struct ceph_mds_request, r_node);
1908                 if (req->r_got_unsafe)
1909                         continue;
1910                 if (req->r_session &&
1911                     req->r_session->s_mds == mds) {
1912                         dout(" kicking tid %llu\n", req->r_tid);
1913                         __do_request(mdsc, req);
1914                 }
1915         }
1916 }
1917
1918 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
1919                               struct ceph_mds_request *req)
1920 {
1921         dout("submit_request on %p\n", req);
1922         mutex_lock(&mdsc->mutex);
1923         __register_request(mdsc, req, NULL);
1924         __do_request(mdsc, req);
1925         mutex_unlock(&mdsc->mutex);
1926 }
1927
1928 /*
1929  * Synchrously perform an mds request.  Take care of all of the
1930  * session setup, forwarding, retry details.
1931  */
1932 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
1933                          struct inode *dir,
1934                          struct ceph_mds_request *req)
1935 {
1936         int err;
1937
1938         dout("do_request on %p\n", req);
1939
1940         /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
1941         if (req->r_inode)
1942                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
1943         if (req->r_locked_dir)
1944                 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
1945         if (req->r_old_dentry)
1946                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
1947                                   CEPH_CAP_PIN);
1948
1949         /* issue */
1950         mutex_lock(&mdsc->mutex);
1951         __register_request(mdsc, req, dir);
1952         __do_request(mdsc, req);
1953
1954         if (req->r_err) {
1955                 err = req->r_err;
1956                 __unregister_request(mdsc, req);
1957                 dout("do_request early error %d\n", err);
1958                 goto out;
1959         }
1960
1961         /* wait */
1962         mutex_unlock(&mdsc->mutex);
1963         dout("do_request waiting\n");
1964         if (req->r_timeout) {
1965                 err = (long)wait_for_completion_killable_timeout(
1966                         &req->r_completion, req->r_timeout);
1967                 if (err == 0)
1968                         err = -EIO;
1969         } else {
1970                 err = wait_for_completion_killable(&req->r_completion);
1971         }
1972         dout("do_request waited, got %d\n", err);
1973         mutex_lock(&mdsc->mutex);
1974
1975         /* only abort if we didn't race with a real reply */
1976         if (req->r_got_result) {
1977                 err = le32_to_cpu(req->r_reply_info.head->result);
1978         } else if (err < 0) {
1979                 dout("aborted request %lld with %d\n", req->r_tid, err);
1980
1981                 /*
1982                  * ensure we aren't running concurrently with
1983                  * ceph_fill_trace or ceph_readdir_prepopulate, which
1984                  * rely on locks (dir mutex) held by our caller.
1985                  */
1986                 mutex_lock(&req->r_fill_mutex);
1987                 req->r_err = err;
1988                 req->r_aborted = true;
1989                 mutex_unlock(&req->r_fill_mutex);
1990
1991                 if (req->r_locked_dir &&
1992                     (req->r_op & CEPH_MDS_OP_WRITE))
1993                         ceph_invalidate_dir_request(req);
1994         } else {
1995                 err = req->r_err;
1996         }
1997
1998 out:
1999         mutex_unlock(&mdsc->mutex);
2000         dout("do_request %p done, result %d\n", req, err);
2001         return err;
2002 }
2003
2004 /*
2005  * Invalidate dir D_COMPLETE, dentry lease state on an aborted MDS
2006  * namespace request.
2007  */
2008 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2009 {
2010         struct inode *inode = req->r_locked_dir;
2011         struct ceph_inode_info *ci = ceph_inode(inode);
2012
2013         dout("invalidate_dir_request %p (D_COMPLETE, lease(s))\n", inode);
2014         spin_lock(&inode->i_lock);
2015         ceph_dir_clear_complete(inode);
2016         ci->i_release_count++;
2017         spin_unlock(&inode->i_lock);
2018
2019         if (req->r_dentry)
2020                 ceph_invalidate_dentry_lease(req->r_dentry);
2021         if (req->r_old_dentry)
2022                 ceph_invalidate_dentry_lease(req->r_old_dentry);
2023 }
2024
2025 /*
2026  * Handle mds reply.
2027  *
2028  * We take the session mutex and parse and process the reply immediately.
2029  * This preserves the logical ordering of replies, capabilities, etc., sent
2030  * by the MDS as they are applied to our local cache.
2031  */
2032 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2033 {
2034         struct ceph_mds_client *mdsc = session->s_mdsc;
2035         struct ceph_mds_request *req;
2036         struct ceph_mds_reply_head *head = msg->front.iov_base;
2037         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2038         u64 tid;
2039         int err, result;
2040         int mds = session->s_mds;
2041
2042         if (msg->front.iov_len < sizeof(*head)) {
2043                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2044                 ceph_msg_dump(msg);
2045                 return;
2046         }
2047
2048         /* get request, session */
2049         tid = le64_to_cpu(msg->hdr.tid);
2050         mutex_lock(&mdsc->mutex);
2051         req = __lookup_request(mdsc, tid);
2052         if (!req) {
2053                 dout("handle_reply on unknown tid %llu\n", tid);
2054                 mutex_unlock(&mdsc->mutex);
2055                 return;
2056         }
2057         dout("handle_reply %p\n", req);
2058
2059         /* correct session? */
2060         if (req->r_session != session) {
2061                 pr_err("mdsc_handle_reply got %llu on session mds%d"
2062                        " not mds%d\n", tid, session->s_mds,
2063                        req->r_session ? req->r_session->s_mds : -1);
2064                 mutex_unlock(&mdsc->mutex);
2065                 goto out;
2066         }
2067
2068         /* dup? */
2069         if ((req->r_got_unsafe && !head->safe) ||
2070             (req->r_got_safe && head->safe)) {
2071                 pr_warning("got a dup %s reply on %llu from mds%d\n",
2072                            head->safe ? "safe" : "unsafe", tid, mds);
2073                 mutex_unlock(&mdsc->mutex);
2074                 goto out;
2075         }
2076         if (req->r_got_safe && !head->safe) {
2077                 pr_warning("got unsafe after safe on %llu from mds%d\n",
2078                            tid, mds);
2079                 mutex_unlock(&mdsc->mutex);
2080                 goto out;
2081         }
2082
2083         result = le32_to_cpu(head->result);
2084
2085         /*
2086          * Handle an ESTALE
2087          * if we're not talking to the authority, send to them
2088          * if the authority has changed while we weren't looking,
2089          * send to new authority
2090          * Otherwise we just have to return an ESTALE
2091          */
2092         if (result == -ESTALE) {
2093                 dout("got ESTALE on request %llu", req->r_tid);
2094                 if (!req->r_inode) {
2095                         /* do nothing; not an authority problem */
2096                 } else if (req->r_direct_mode != USE_AUTH_MDS) {
2097                         dout("not using auth, setting for that now");
2098                         req->r_direct_mode = USE_AUTH_MDS;
2099                         __do_request(mdsc, req);
2100                         mutex_unlock(&mdsc->mutex);
2101                         goto out;
2102                 } else  {
2103                         struct ceph_inode_info *ci = ceph_inode(req->r_inode);
2104                         struct ceph_cap *cap = NULL;
2105
2106                         if (req->r_session)
2107                                 cap = ceph_get_cap_for_mds(ci,
2108                                                    req->r_session->s_mds);
2109
2110                         dout("already using auth");
2111                         if ((!cap || cap != ci->i_auth_cap) ||
2112                             (cap->mseq != req->r_sent_on_mseq)) {
2113                                 dout("but cap changed, so resending");
2114                                 __do_request(mdsc, req);
2115                                 mutex_unlock(&mdsc->mutex);
2116                                 goto out;
2117                         }
2118                 }
2119                 dout("have to return ESTALE on request %llu", req->r_tid);
2120         }
2121
2122
2123         if (head->safe) {
2124                 req->r_got_safe = true;
2125                 __unregister_request(mdsc, req);
2126                 complete_all(&req->r_safe_completion);
2127
2128                 if (req->r_got_unsafe) {
2129                         /*
2130                          * We already handled the unsafe response, now do the
2131                          * cleanup.  No need to examine the response; the MDS
2132                          * doesn't include any result info in the safe
2133                          * response.  And even if it did, there is nothing
2134                          * useful we could do with a revised return value.
2135                          */
2136                         dout("got safe reply %llu, mds%d\n", tid, mds);
2137                         list_del_init(&req->r_unsafe_item);
2138
2139                         /* last unsafe request during umount? */
2140                         if (mdsc->stopping && !__get_oldest_req(mdsc))
2141                                 complete_all(&mdsc->safe_umount_waiters);
2142                         mutex_unlock(&mdsc->mutex);
2143                         goto out;
2144                 }
2145         } else {
2146                 req->r_got_unsafe = true;
2147                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2148         }
2149
2150         dout("handle_reply tid %lld result %d\n", tid, result);
2151         rinfo = &req->r_reply_info;
2152         err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2153         mutex_unlock(&mdsc->mutex);
2154
2155         mutex_lock(&session->s_mutex);
2156         if (err < 0) {
2157                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2158                 ceph_msg_dump(msg);
2159                 goto out_err;
2160         }
2161
2162         /* snap trace */
2163         if (rinfo->snapblob_len) {
2164                 down_write(&mdsc->snap_rwsem);
2165                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2166                                rinfo->snapblob + rinfo->snapblob_len,
2167                                le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
2168                 downgrade_write(&mdsc->snap_rwsem);
2169         } else {
2170                 down_read(&mdsc->snap_rwsem);
2171         }
2172
2173         /* insert trace into our cache */
2174         mutex_lock(&req->r_fill_mutex);
2175         err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2176         if (err == 0) {
2177                 if (result == 0 && req->r_op != CEPH_MDS_OP_GETFILELOCK &&
2178                     rinfo->dir_nr)
2179                         ceph_readdir_prepopulate(req, req->r_session);
2180                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2181         }
2182         mutex_unlock(&req->r_fill_mutex);
2183
2184         up_read(&mdsc->snap_rwsem);
2185 out_err:
2186         mutex_lock(&mdsc->mutex);
2187         if (!req->r_aborted) {
2188                 if (err) {
2189                         req->r_err = err;
2190                 } else {
2191                         req->r_reply = msg;
2192                         ceph_msg_get(msg);
2193                         req->r_got_result = true;
2194                 }
2195         } else {
2196                 dout("reply arrived after request %lld was aborted\n", tid);
2197         }
2198         mutex_unlock(&mdsc->mutex);
2199
2200         ceph_add_cap_releases(mdsc, req->r_session);
2201         mutex_unlock(&session->s_mutex);
2202
2203         /* kick calling process */
2204         complete_request(mdsc, req);
2205 out:
2206         ceph_mdsc_put_request(req);
2207         return;
2208 }
2209
2210
2211
2212 /*
2213  * handle mds notification that our request has been forwarded.
2214  */
2215 static void handle_forward(struct ceph_mds_client *mdsc,
2216                            struct ceph_mds_session *session,
2217                            struct ceph_msg *msg)
2218 {
2219         struct ceph_mds_request *req;
2220         u64 tid = le64_to_cpu(msg->hdr.tid);
2221         u32 next_mds;
2222         u32 fwd_seq;
2223         int err = -EINVAL;
2224         void *p = msg->front.iov_base;
2225         void *end = p + msg->front.iov_len;
2226
2227         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2228         next_mds = ceph_decode_32(&p);
2229         fwd_seq = ceph_decode_32(&p);
2230
2231         mutex_lock(&mdsc->mutex);
2232         req = __lookup_request(mdsc, tid);
2233         if (!req) {
2234                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2235                 goto out;  /* dup reply? */
2236         }
2237
2238         if (req->r_aborted) {
2239                 dout("forward tid %llu aborted, unregistering\n", tid);
2240                 __unregister_request(mdsc, req);
2241         } else if (fwd_seq <= req->r_num_fwd) {
2242                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2243                      tid, next_mds, req->r_num_fwd, fwd_seq);
2244         } else {
2245                 /* resend. forward race not possible; mds would drop */
2246                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2247                 BUG_ON(req->r_err);
2248                 BUG_ON(req->r_got_result);
2249                 req->r_num_fwd = fwd_seq;
2250                 req->r_resend_mds = next_mds;
2251                 put_request_session(req);
2252                 __do_request(mdsc, req);
2253         }
2254         ceph_mdsc_put_request(req);
2255 out:
2256         mutex_unlock(&mdsc->mutex);
2257         return;
2258
2259 bad:
2260         pr_err("mdsc_handle_forward decode error err=%d\n", err);
2261 }
2262
2263 /*
2264  * handle a mds session control message
2265  */
2266 static void handle_session(struct ceph_mds_session *session,
2267                            struct ceph_msg *msg)
2268 {
2269         struct ceph_mds_client *mdsc = session->s_mdsc;
2270         u32 op;
2271         u64 seq;
2272         int mds = session->s_mds;
2273         struct ceph_mds_session_head *h = msg->front.iov_base;
2274         int wake = 0;
2275
2276         /* decode */
2277         if (msg->front.iov_len != sizeof(*h))
2278                 goto bad;
2279         op = le32_to_cpu(h->op);
2280         seq = le64_to_cpu(h->seq);
2281
2282         mutex_lock(&mdsc->mutex);
2283         if (op == CEPH_SESSION_CLOSE)
2284                 __unregister_session(mdsc, session);
2285         /* FIXME: this ttl calculation is generous */
2286         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2287         mutex_unlock(&mdsc->mutex);
2288
2289         mutex_lock(&session->s_mutex);
2290
2291         dout("handle_session mds%d %s %p state %s seq %llu\n",
2292              mds, ceph_session_op_name(op), session,
2293              session_state_name(session->s_state), seq);
2294
2295         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2296                 session->s_state = CEPH_MDS_SESSION_OPEN;
2297                 pr_info("mds%d came back\n", session->s_mds);
2298         }
2299
2300         switch (op) {
2301         case CEPH_SESSION_OPEN:
2302                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2303                         pr_info("mds%d reconnect success\n", session->s_mds);
2304                 session->s_state = CEPH_MDS_SESSION_OPEN;
2305                 renewed_caps(mdsc, session, 0);
2306                 wake = 1;
2307                 if (mdsc->stopping)
2308                         __close_session(mdsc, session);
2309                 break;
2310
2311         case CEPH_SESSION_RENEWCAPS:
2312                 if (session->s_renew_seq == seq)
2313                         renewed_caps(mdsc, session, 1);
2314                 break;
2315
2316         case CEPH_SESSION_CLOSE:
2317                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2318                         pr_info("mds%d reconnect denied\n", session->s_mds);
2319                 remove_session_caps(session);
2320                 wake = 1; /* for good measure */
2321                 wake_up_all(&mdsc->session_close_wq);
2322                 kick_requests(mdsc, mds);
2323                 break;
2324
2325         case CEPH_SESSION_STALE:
2326                 pr_info("mds%d caps went stale, renewing\n",
2327                         session->s_mds);
2328                 spin_lock(&session->s_cap_lock);
2329                 session->s_cap_gen++;
2330                 session->s_cap_ttl = 0;
2331                 spin_unlock(&session->s_cap_lock);
2332                 send_renew_caps(mdsc, session);
2333                 break;
2334
2335         case CEPH_SESSION_RECALL_STATE:
2336                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2337                 break;
2338
2339         default:
2340                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2341                 WARN_ON(1);
2342         }
2343
2344         mutex_unlock(&session->s_mutex);
2345         if (wake) {
2346                 mutex_lock(&mdsc->mutex);
2347                 __wake_requests(mdsc, &session->s_waiting);
2348                 mutex_unlock(&mdsc->mutex);
2349         }
2350         return;
2351
2352 bad:
2353         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2354                (int)msg->front.iov_len);
2355         ceph_msg_dump(msg);
2356         return;
2357 }
2358
2359
2360 /*
2361  * called under session->mutex.
2362  */
2363 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2364                                    struct ceph_mds_session *session)
2365 {
2366         struct ceph_mds_request *req, *nreq;
2367         int err;
2368
2369         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2370
2371         mutex_lock(&mdsc->mutex);
2372         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2373                 err = __prepare_send_request(mdsc, req, session->s_mds);
2374                 if (!err) {
2375                         ceph_msg_get(req->r_request);
2376                         ceph_con_send(&session->s_con, req->r_request);
2377                 }
2378         }
2379         mutex_unlock(&mdsc->mutex);
2380 }
2381
2382 /*
2383  * Encode information about a cap for a reconnect with the MDS.
2384  */
2385 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2386                           void *arg)
2387 {
2388         union {
2389                 struct ceph_mds_cap_reconnect v2;
2390                 struct ceph_mds_cap_reconnect_v1 v1;
2391         } rec;
2392         size_t reclen;
2393         struct ceph_inode_info *ci;
2394         struct ceph_reconnect_state *recon_state = arg;
2395         struct ceph_pagelist *pagelist = recon_state->pagelist;
2396         char *path;
2397         int pathlen, err;
2398         u64 pathbase;
2399         struct dentry *dentry;
2400
2401         ci = cap->ci;
2402
2403         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2404              inode, ceph_vinop(inode), cap, cap->cap_id,
2405              ceph_cap_string(cap->issued));
2406         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2407         if (err)
2408                 return err;
2409
2410         dentry = d_find_alias(inode);
2411         if (dentry) {
2412                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2413                 if (IS_ERR(path)) {
2414                         err = PTR_ERR(path);
2415                         goto out_dput;
2416                 }
2417         } else {
2418                 path = NULL;
2419                 pathlen = 0;
2420         }
2421         err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2422         if (err)
2423                 goto out_free;
2424
2425         spin_lock(&inode->i_lock);
2426         cap->seq = 0;        /* reset cap seq */
2427         cap->issue_seq = 0;  /* and issue_seq */
2428
2429         if (recon_state->flock) {
2430                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2431                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2432                 rec.v2.issued = cpu_to_le32(cap->issued);
2433                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2434                 rec.v2.pathbase = cpu_to_le64(pathbase);
2435                 rec.v2.flock_len = 0;
2436                 reclen = sizeof(rec.v2);
2437         } else {
2438                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2439                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2440                 rec.v1.issued = cpu_to_le32(cap->issued);
2441                 rec.v1.size = cpu_to_le64(inode->i_size);
2442                 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2443                 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2444                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2445                 rec.v1.pathbase = cpu_to_le64(pathbase);
2446                 reclen = sizeof(rec.v1);
2447         }
2448         spin_unlock(&inode->i_lock);
2449
2450         if (recon_state->flock) {
2451                 int num_fcntl_locks, num_flock_locks;
2452                 struct ceph_pagelist_cursor trunc_point;
2453
2454                 ceph_pagelist_set_cursor(pagelist, &trunc_point);
2455                 do {
2456                         lock_flocks();
2457                         ceph_count_locks(inode, &num_fcntl_locks,
2458                                          &num_flock_locks);
2459                         rec.v2.flock_len = (2*sizeof(u32) +
2460                                             (num_fcntl_locks+num_flock_locks) *
2461                                             sizeof(struct ceph_filelock));
2462                         unlock_flocks();
2463
2464                         /* pre-alloc pagelist */
2465                         ceph_pagelist_truncate(pagelist, &trunc_point);
2466                         err = ceph_pagelist_append(pagelist, &rec, reclen);
2467                         if (!err)
2468                                 err = ceph_pagelist_reserve(pagelist,
2469                                                             rec.v2.flock_len);
2470
2471                         /* encode locks */
2472                         if (!err) {
2473                                 lock_flocks();
2474                                 err = ceph_encode_locks(inode,
2475                                                         pagelist,
2476                                                         num_fcntl_locks,
2477                                                         num_flock_locks);
2478                                 unlock_flocks();
2479                         }
2480                 } while (err == -ENOSPC);
2481         } else {
2482                 err = ceph_pagelist_append(pagelist, &rec, reclen);
2483         }
2484
2485 out_free:
2486         kfree(path);
2487 out_dput:
2488         dput(dentry);
2489         return err;
2490 }
2491
2492
2493 /*
2494  * If an MDS fails and recovers, clients need to reconnect in order to
2495  * reestablish shared state.  This includes all caps issued through
2496  * this session _and_ the snap_realm hierarchy.  Because it's not
2497  * clear which snap realms the mds cares about, we send everything we
2498  * know about.. that ensures we'll then get any new info the
2499  * recovering MDS might have.
2500  *
2501  * This is a relatively heavyweight operation, but it's rare.
2502  *
2503  * called with mdsc->mutex held.
2504  */
2505 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2506                                struct ceph_mds_session *session)
2507 {
2508         struct ceph_msg *reply;
2509         struct rb_node *p;
2510         int mds = session->s_mds;
2511         int err = -ENOMEM;
2512         struct ceph_pagelist *pagelist;
2513         struct ceph_reconnect_state recon_state;
2514
2515         pr_info("mds%d reconnect start\n", mds);
2516
2517         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2518         if (!pagelist)
2519                 goto fail_nopagelist;
2520         ceph_pagelist_init(pagelist);
2521
2522         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2523         if (!reply)
2524                 goto fail_nomsg;
2525
2526         mutex_lock(&session->s_mutex);
2527         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2528         session->s_seq = 0;
2529
2530         ceph_con_open(&session->s_con,
2531                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2532
2533         /* replay unsafe requests */
2534         replay_unsafe_requests(mdsc, session);
2535
2536         down_read(&mdsc->snap_rwsem);
2537
2538         dout("session %p state %s\n", session,
2539              session_state_name(session->s_state));
2540
2541         /* drop old cap expires; we're about to reestablish that state */
2542         discard_cap_releases(mdsc, session);
2543
2544         /* traverse this session's caps */
2545         err = ceph_pagelist_encode_32(pagelist, session->s_nr_caps);
2546         if (err)
2547                 goto fail;
2548
2549         recon_state.pagelist = pagelist;
2550         recon_state.flock = session->s_con.peer_features & CEPH_FEATURE_FLOCK;
2551         err = iterate_session_caps(session, encode_caps_cb, &recon_state);
2552         if (err < 0)
2553                 goto fail;
2554
2555         /*
2556          * snaprealms.  we provide mds with the ino, seq (version), and
2557          * parent for all of our realms.  If the mds has any newer info,
2558          * it will tell us.
2559          */
2560         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2561                 struct ceph_snap_realm *realm =
2562                         rb_entry(p, struct ceph_snap_realm, node);
2563                 struct ceph_mds_snaprealm_reconnect sr_rec;
2564
2565                 dout(" adding snap realm %llx seq %lld parent %llx\n",
2566                      realm->ino, realm->seq, realm->parent_ino);
2567                 sr_rec.ino = cpu_to_le64(realm->ino);
2568                 sr_rec.seq = cpu_to_le64(realm->seq);
2569                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2570                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2571                 if (err)
2572                         goto fail;
2573         }
2574
2575         reply->pagelist = pagelist;
2576         if (recon_state.flock)
2577                 reply->hdr.version = cpu_to_le16(2);
2578         reply->hdr.data_len = cpu_to_le32(pagelist->length);
2579         reply->nr_pages = calc_pages_for(0, pagelist->length);
2580         ceph_con_send(&session->s_con, reply);
2581
2582         mutex_unlock(&session->s_mutex);
2583
2584         mutex_lock(&mdsc->mutex);
2585         __wake_requests(mdsc, &session->s_waiting);
2586         mutex_unlock(&mdsc->mutex);
2587
2588         up_read(&mdsc->snap_rwsem);
2589         return;
2590
2591 fail:
2592         ceph_msg_put(reply);
2593         up_read(&mdsc->snap_rwsem);
2594         mutex_unlock(&session->s_mutex);
2595 fail_nomsg:
2596         ceph_pagelist_release(pagelist);
2597         kfree(pagelist);
2598 fail_nopagelist:
2599         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
2600         return;
2601 }
2602
2603
2604 /*
2605  * compare old and new mdsmaps, kicking requests
2606  * and closing out old connections as necessary
2607  *
2608  * called under mdsc->mutex.
2609  */
2610 static void check_new_map(struct ceph_mds_client *mdsc,
2611                           struct ceph_mdsmap *newmap,
2612                           struct ceph_mdsmap *oldmap)
2613 {
2614         int i;
2615         int oldstate, newstate;
2616         struct ceph_mds_session *s;
2617
2618         dout("check_new_map new %u old %u\n",
2619              newmap->m_epoch, oldmap->m_epoch);
2620
2621         for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2622                 if (mdsc->sessions[i] == NULL)
2623                         continue;
2624                 s = mdsc->sessions[i];
2625                 oldstate = ceph_mdsmap_get_state(oldmap, i);
2626                 newstate = ceph_mdsmap_get_state(newmap, i);
2627
2628                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
2629                      i, ceph_mds_state_name(oldstate),
2630                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
2631                      ceph_mds_state_name(newstate),
2632                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
2633                      session_state_name(s->s_state));
2634
2635                 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
2636                            ceph_mdsmap_get_addr(newmap, i),
2637                            sizeof(struct ceph_entity_addr))) {
2638                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2639                                 /* the session never opened, just close it
2640                                  * out now */
2641                                 __wake_requests(mdsc, &s->s_waiting);
2642                                 __unregister_session(mdsc, s);
2643                         } else {
2644                                 /* just close it */
2645                                 mutex_unlock(&mdsc->mutex);
2646                                 mutex_lock(&s->s_mutex);
2647                                 mutex_lock(&mdsc->mutex);
2648                                 ceph_con_close(&s->s_con);
2649                                 mutex_unlock(&s->s_mutex);
2650                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2651                         }
2652
2653                         /* kick any requests waiting on the recovering mds */
2654                         kick_requests(mdsc, i);
2655                 } else if (oldstate == newstate) {
2656                         continue;  /* nothing new with this mds */
2657                 }
2658
2659                 /*
2660                  * send reconnect?
2661                  */
2662                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2663                     newstate >= CEPH_MDS_STATE_RECONNECT) {
2664                         mutex_unlock(&mdsc->mutex);
2665                         send_mds_reconnect(mdsc, s);
2666                         mutex_lock(&mdsc->mutex);
2667                 }
2668
2669                 /*
2670                  * kick request on any mds that has gone active.
2671                  */
2672                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2673                     newstate >= CEPH_MDS_STATE_ACTIVE) {
2674                         if (oldstate != CEPH_MDS_STATE_CREATING &&
2675                             oldstate != CEPH_MDS_STATE_STARTING)
2676                                 pr_info("mds%d recovery completed\n", s->s_mds);
2677                         kick_requests(mdsc, i);
2678                         ceph_kick_flushing_caps(mdsc, s);
2679                         wake_up_session_caps(s, 1);
2680                 }
2681         }
2682
2683         for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
2684                 s = mdsc->sessions[i];
2685                 if (!s)
2686                         continue;
2687                 if (!ceph_mdsmap_is_laggy(newmap, i))
2688                         continue;
2689                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2690                     s->s_state == CEPH_MDS_SESSION_HUNG ||
2691                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
2692                         dout(" connecting to export targets of laggy mds%d\n",
2693                              i);
2694                         __open_export_target_sessions(mdsc, s);
2695                 }
2696         }
2697 }
2698
2699
2700
2701 /*
2702  * leases
2703  */
2704
2705 /*
2706  * caller must hold session s_mutex, dentry->d_lock
2707  */
2708 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2709 {
2710         struct ceph_dentry_info *di = ceph_dentry(dentry);
2711
2712         ceph_put_mds_session(di->lease_session);
2713         di->lease_session = NULL;
2714 }
2715
2716 static void handle_lease(struct ceph_mds_client *mdsc,
2717                          struct ceph_mds_session *session,
2718                          struct ceph_msg *msg)
2719 {
2720         struct super_block *sb = mdsc->fsc->sb;
2721         struct inode *inode;
2722         struct dentry *parent, *dentry;
2723         struct ceph_dentry_info *di;
2724         int mds = session->s_mds;
2725         struct ceph_mds_lease *h = msg->front.iov_base;
2726         u32 seq;
2727         struct ceph_vino vino;
2728         struct qstr dname;
2729         int release = 0;
2730
2731         dout("handle_lease from mds%d\n", mds);
2732
2733         /* decode */
2734         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2735                 goto bad;
2736         vino.ino = le64_to_cpu(h->ino);
2737         vino.snap = CEPH_NOSNAP;
2738         seq = le32_to_cpu(h->seq);
2739         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2740         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2741         if (dname.len != get_unaligned_le32(h+1))
2742                 goto bad;
2743
2744         mutex_lock(&session->s_mutex);
2745         session->s_seq++;
2746
2747         /* lookup inode */
2748         inode = ceph_find_inode(sb, vino);
2749         dout("handle_lease %s, ino %llx %p %.*s\n",
2750              ceph_lease_op_name(h->action), vino.ino, inode,
2751              dname.len, dname.name);
2752         if (inode == NULL) {
2753                 dout("handle_lease no inode %llx\n", vino.ino);
2754                 goto release;
2755         }
2756
2757         /* dentry */
2758         parent = d_find_alias(inode);
2759         if (!parent) {
2760                 dout("no parent dentry on inode %p\n", inode);
2761                 WARN_ON(1);
2762                 goto release;  /* hrm... */
2763         }
2764         dname.hash = full_name_hash(dname.name, dname.len);
2765         dentry = d_lookup(parent, &dname);
2766         dput(parent);
2767         if (!dentry)
2768                 goto release;
2769
2770         spin_lock(&dentry->d_lock);
2771         di = ceph_dentry(dentry);
2772         switch (h->action) {
2773         case CEPH_MDS_LEASE_REVOKE:
2774                 if (di && di->lease_session == session) {
2775                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
2776                                 h->seq = cpu_to_le32(di->lease_seq);
2777                         __ceph_mdsc_drop_dentry_lease(dentry);
2778                 }
2779                 release = 1;
2780                 break;
2781
2782         case CEPH_MDS_LEASE_RENEW:
2783                 if (di && di->lease_session == session &&
2784                     di->lease_gen == session->s_cap_gen &&
2785                     di->lease_renew_from &&
2786                     di->lease_renew_after == 0) {
2787                         unsigned long duration =
2788                                 le32_to_cpu(h->duration_ms) * HZ / 1000;
2789
2790                         di->lease_seq = seq;
2791                         dentry->d_time = di->lease_renew_from + duration;
2792                         di->lease_renew_after = di->lease_renew_from +
2793                                 (duration >> 1);
2794                         di->lease_renew_from = 0;
2795                 }
2796                 break;
2797         }
2798         spin_unlock(&dentry->d_lock);
2799         dput(dentry);
2800
2801         if (!release)
2802                 goto out;
2803
2804 release:
2805         /* let's just reuse the same message */
2806         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2807         ceph_msg_get(msg);
2808         ceph_con_send(&session->s_con, msg);
2809
2810 out:
2811         iput(inode);
2812         mutex_unlock(&session->s_mutex);
2813         return;
2814
2815 bad:
2816         pr_err("corrupt lease message\n");
2817         ceph_msg_dump(msg);
2818 }
2819
2820 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
2821                               struct inode *inode,
2822                               struct dentry *dentry, char action,
2823                               u32 seq)
2824 {
2825         struct ceph_msg *msg;
2826         struct ceph_mds_lease *lease;
2827         int len = sizeof(*lease) + sizeof(u32);
2828         int dnamelen = 0;
2829
2830         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2831              inode, dentry, ceph_lease_op_name(action), session->s_mds);
2832         dnamelen = dentry->d_name.len;
2833         len += dnamelen;
2834
2835         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
2836         if (!msg)
2837                 return;
2838         lease = msg->front.iov_base;
2839         lease->action = action;
2840         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
2841         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
2842         lease->seq = cpu_to_le32(seq);
2843         put_unaligned_le32(dnamelen, lease + 1);
2844         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
2845
2846         /*
2847          * if this is a preemptive lease RELEASE, no need to
2848          * flush request stream, since the actual request will
2849          * soon follow.
2850          */
2851         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
2852
2853         ceph_con_send(&session->s_con, msg);
2854 }
2855
2856 /*
2857  * Preemptively release a lease we expect to invalidate anyway.
2858  * Pass @inode always, @dentry is optional.
2859  */
2860 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2861                              struct dentry *dentry)
2862 {
2863         struct ceph_dentry_info *di;
2864         struct ceph_mds_session *session;
2865         u32 seq;
2866
2867         BUG_ON(inode == NULL);
2868         BUG_ON(dentry == NULL);
2869
2870         /* is dentry lease valid? */
2871         spin_lock(&dentry->d_lock);
2872         di = ceph_dentry(dentry);
2873         if (!di || !di->lease_session ||
2874             di->lease_session->s_mds < 0 ||
2875             di->lease_gen != di->lease_session->s_cap_gen ||
2876             !time_before(jiffies, dentry->d_time)) {
2877                 dout("lease_release inode %p dentry %p -- "
2878                      "no lease\n",
2879                      inode, dentry);
2880                 spin_unlock(&dentry->d_lock);
2881                 return;
2882         }
2883
2884         /* we do have a lease on this dentry; note mds and seq */
2885         session = ceph_get_mds_session(di->lease_session);
2886         seq = di->lease_seq;
2887         __ceph_mdsc_drop_dentry_lease(dentry);
2888         spin_unlock(&dentry->d_lock);
2889
2890         dout("lease_release inode %p dentry %p to mds%d\n",
2891              inode, dentry, session->s_mds);
2892         ceph_mdsc_lease_send_msg(session, inode, dentry,
2893                                  CEPH_MDS_LEASE_RELEASE, seq);
2894         ceph_put_mds_session(session);
2895 }
2896
2897 /*
2898  * drop all leases (and dentry refs) in preparation for umount
2899  */
2900 static void drop_leases(struct ceph_mds_client *mdsc)
2901 {
2902         int i;
2903
2904         dout("drop_leases\n");
2905         mutex_lock(&mdsc->mutex);
2906         for (i = 0; i < mdsc->max_sessions; i++) {
2907                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2908                 if (!s)
2909                         continue;
2910                 mutex_unlock(&mdsc->mutex);
2911                 mutex_lock(&s->s_mutex);
2912                 mutex_unlock(&s->s_mutex);
2913                 ceph_put_mds_session(s);
2914                 mutex_lock(&mdsc->mutex);
2915         }
2916         mutex_unlock(&mdsc->mutex);
2917 }
2918
2919
2920
2921 /*
2922  * delayed work -- periodically trim expired leases, renew caps with mds
2923  */
2924 static void schedule_delayed(struct ceph_mds_client *mdsc)
2925 {
2926         int delay = 5;
2927         unsigned hz = round_jiffies_relative(HZ * delay);
2928         schedule_delayed_work(&mdsc->delayed_work, hz);
2929 }
2930
2931 static void delayed_work(struct work_struct *work)
2932 {
2933         int i;
2934         struct ceph_mds_client *mdsc =
2935                 container_of(work, struct ceph_mds_client, delayed_work.work);
2936         int renew_interval;
2937         int renew_caps;
2938
2939         dout("mdsc delayed_work\n");
2940         ceph_check_delayed_caps(mdsc);
2941
2942         mutex_lock(&mdsc->mutex);
2943         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
2944         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
2945                                    mdsc->last_renew_caps);
2946         if (renew_caps)
2947                 mdsc->last_renew_caps = jiffies;
2948
2949         for (i = 0; i < mdsc->max_sessions; i++) {
2950                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2951                 if (s == NULL)
2952                         continue;
2953                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
2954                         dout("resending session close request for mds%d\n",
2955                              s->s_mds);
2956                         request_close_session(mdsc, s);
2957                         ceph_put_mds_session(s);
2958                         continue;
2959                 }
2960                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
2961                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
2962                                 s->s_state = CEPH_MDS_SESSION_HUNG;
2963                                 pr_info("mds%d hung\n", s->s_mds);
2964                         }
2965                 }
2966                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
2967                         /* this mds is failed or recovering, just wait */
2968                         ceph_put_mds_session(s);
2969                         continue;
2970                 }
2971                 mutex_unlock(&mdsc->mutex);
2972
2973                 mutex_lock(&s->s_mutex);
2974                 if (renew_caps)
2975                         send_renew_caps(mdsc, s);
2976                 else
2977                         ceph_con_keepalive(&s->s_con);
2978                 ceph_add_cap_releases(mdsc, s);
2979                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2980                     s->s_state == CEPH_MDS_SESSION_HUNG)
2981                         ceph_send_cap_releases(mdsc, s);
2982                 mutex_unlock(&s->s_mutex);
2983                 ceph_put_mds_session(s);
2984
2985                 mutex_lock(&mdsc->mutex);
2986         }
2987         mutex_unlock(&mdsc->mutex);
2988
2989         schedule_delayed(mdsc);
2990 }
2991
2992 int ceph_mdsc_init(struct ceph_fs_client *fsc)
2993
2994 {
2995         struct ceph_mds_client *mdsc;
2996
2997         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
2998         if (!mdsc)
2999                 return -ENOMEM;
3000         mdsc->fsc = fsc;
3001         fsc->mdsc = mdsc;
3002         mutex_init(&mdsc->mutex);
3003         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3004         if (mdsc->mdsmap == NULL)
3005                 return -ENOMEM;
3006
3007         init_completion(&mdsc->safe_umount_waiters);
3008         init_waitqueue_head(&mdsc->session_close_wq);
3009         INIT_LIST_HEAD(&mdsc->waiting_for_map);
3010         mdsc->sessions = NULL;
3011         mdsc->max_sessions = 0;
3012         mdsc->stopping = 0;
3013         init_rwsem(&mdsc->snap_rwsem);
3014         mdsc->snap_realms = RB_ROOT;
3015         INIT_LIST_HEAD(&mdsc->snap_empty);
3016         spin_lock_init(&mdsc->snap_empty_lock);
3017         mdsc->last_tid = 0;
3018         mdsc->request_tree = RB_ROOT;
3019         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3020         mdsc->last_renew_caps = jiffies;
3021         INIT_LIST_HEAD(&mdsc->cap_delay_list);
3022         spin_lock_init(&mdsc->cap_delay_lock);
3023         INIT_LIST_HEAD(&mdsc->snap_flush_list);
3024         spin_lock_init(&mdsc->snap_flush_lock);
3025         mdsc->cap_flush_seq = 0;
3026         INIT_LIST_HEAD(&mdsc->cap_dirty);
3027         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3028         mdsc->num_cap_flushing = 0;
3029         spin_lock_init(&mdsc->cap_dirty_lock);
3030         init_waitqueue_head(&mdsc->cap_flushing_wq);
3031         spin_lock_init(&mdsc->dentry_lru_lock);
3032         INIT_LIST_HEAD(&mdsc->dentry_lru);
3033
3034         ceph_caps_init(mdsc);
3035         ceph_adjust_min_caps(mdsc, fsc->min_caps);
3036
3037         return 0;
3038 }
3039
3040 /*
3041  * Wait for safe replies on open mds requests.  If we time out, drop
3042  * all requests from the tree to avoid dangling dentry refs.
3043  */
3044 static void wait_requests(struct ceph_mds_client *mdsc)
3045 {
3046         struct ceph_mds_request *req;
3047         struct ceph_fs_client *fsc = mdsc->fsc;
3048
3049         mutex_lock(&mdsc->mutex);
3050         if (__get_oldest_req(mdsc)) {
3051                 mutex_unlock(&mdsc->mutex);
3052
3053                 dout("wait_requests waiting for requests\n");
3054                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3055                                     fsc->client->options->mount_timeout * HZ);
3056
3057                 /* tear down remaining requests */
3058                 mutex_lock(&mdsc->mutex);
3059                 while ((req = __get_oldest_req(mdsc))) {
3060                         dout("wait_requests timed out on tid %llu\n",
3061                              req->r_tid);
3062                         __unregister_request(mdsc, req);
3063                 }
3064         }
3065         mutex_unlock(&mdsc->mutex);
3066         dout("wait_requests done\n");
3067 }
3068
3069 /*
3070  * called before mount is ro, and before dentries are torn down.
3071  * (hmm, does this still race with new lookups?)
3072  */
3073 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3074 {
3075         dout("pre_umount\n");
3076         mdsc->stopping = 1;
3077
3078         drop_leases(mdsc);
3079         ceph_flush_dirty_caps(mdsc);
3080         wait_requests(mdsc);
3081
3082         /*
3083          * wait for reply handlers to drop their request refs and
3084          * their inode/dcache refs
3085          */
3086         ceph_msgr_flush();
3087 }
3088
3089 /*
3090  * wait for all write mds requests to flush.
3091  */
3092 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3093 {
3094         struct ceph_mds_request *req = NULL, *nextreq;
3095         struct rb_node *n;
3096
3097         mutex_lock(&mdsc->mutex);
3098         dout("wait_unsafe_requests want %lld\n", want_tid);
3099 restart:
3100         req = __get_oldest_req(mdsc);
3101         while (req && req->r_tid <= want_tid) {
3102                 /* find next request */
3103                 n = rb_next(&req->r_node);
3104                 if (n)
3105                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3106                 else
3107                         nextreq = NULL;
3108                 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
3109                         /* write op */
3110                         ceph_mdsc_get_request(req);
3111                         if (nextreq)
3112                                 ceph_mdsc_get_request(nextreq);
3113                         mutex_unlock(&mdsc->mutex);
3114                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
3115                              req->r_tid, want_tid);
3116                         wait_for_completion(&req->r_safe_completion);
3117                         mutex_lock(&mdsc->mutex);
3118                         ceph_mdsc_put_request(req);
3119                         if (!nextreq)
3120                                 break;  /* next dne before, so we're done! */
3121                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
3122                                 /* next request was removed from tree */
3123                                 ceph_mdsc_put_request(nextreq);
3124                                 goto restart;
3125                         }
3126                         ceph_mdsc_put_request(nextreq);  /* won't go away */
3127                 }
3128                 req = nextreq;
3129         }
3130         mutex_unlock(&mdsc->mutex);
3131         dout("wait_unsafe_requests done\n");
3132 }
3133
3134 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3135 {
3136         u64 want_tid, want_flush;
3137
3138         if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3139                 return;
3140
3141         dout("sync\n");
3142         mutex_lock(&mdsc->mutex);
3143         want_tid = mdsc->last_tid;
3144         want_flush = mdsc->cap_flush_seq;
3145         mutex_unlock(&mdsc->mutex);
3146         dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
3147
3148         ceph_flush_dirty_caps(mdsc);
3149
3150         wait_unsafe_requests(mdsc, want_tid);
3151         wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
3152 }
3153
3154 /*
3155  * true if all sessions are closed, or we force unmount
3156  */
3157 static bool done_closing_sessions(struct ceph_mds_client *mdsc)
3158 {
3159         int i, n = 0;
3160
3161         if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3162                 return true;
3163
3164         mutex_lock(&mdsc->mutex);
3165         for (i = 0; i < mdsc->max_sessions; i++)
3166                 if (mdsc->sessions[i])
3167                         n++;
3168         mutex_unlock(&mdsc->mutex);
3169         return n == 0;
3170 }
3171
3172 /*
3173  * called after sb is ro.
3174  */
3175 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3176 {
3177         struct ceph_mds_session *session;
3178         int i;
3179         struct ceph_fs_client *fsc = mdsc->fsc;
3180         unsigned long timeout = fsc->client->options->mount_timeout * HZ;
3181
3182         dout("close_sessions\n");
3183
3184         /* close sessions */
3185         mutex_lock(&mdsc->mutex);
3186         for (i = 0; i < mdsc->max_sessions; i++) {
3187                 session = __ceph_lookup_mds_session(mdsc, i);
3188                 if (!session)
3189                         continue;
3190                 mutex_unlock(&mdsc->mutex);
3191                 mutex_lock(&session->s_mutex);
3192                 __close_session(mdsc, session);
3193                 mutex_unlock(&session->s_mutex);
3194                 ceph_put_mds_session(session);
3195                 mutex_lock(&mdsc->mutex);
3196         }
3197         mutex_unlock(&mdsc->mutex);
3198
3199         dout("waiting for sessions to close\n");
3200         wait_event_timeout(mdsc->session_close_wq, done_closing_sessions(mdsc),
3201                            timeout);
3202
3203         /* tear down remaining sessions */
3204         mutex_lock(&mdsc->mutex);
3205         for (i = 0; i < mdsc->max_sessions; i++) {
3206                 if (mdsc->sessions[i]) {
3207                         session = get_session(mdsc->sessions[i]);
3208                         __unregister_session(mdsc, session);
3209                         mutex_unlock(&mdsc->mutex);
3210                         mutex_lock(&session->s_mutex);
3211                         remove_session_caps(session);
3212                         mutex_unlock(&session->s_mutex);
3213                         ceph_put_mds_session(session);
3214                         mutex_lock(&mdsc->mutex);
3215                 }
3216         }
3217         WARN_ON(!list_empty(&mdsc->cap_delay_list));
3218         mutex_unlock(&mdsc->mutex);
3219
3220         ceph_cleanup_empty_realms(mdsc);
3221
3222         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3223
3224         dout("stopped\n");
3225 }
3226
3227 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3228 {
3229         dout("stop\n");
3230         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3231         if (mdsc->mdsmap)
3232                 ceph_mdsmap_destroy(mdsc->mdsmap);
3233         kfree(mdsc->sessions);
3234         ceph_caps_finalize(mdsc);
3235 }
3236
3237 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3238 {
3239         struct ceph_mds_client *mdsc = fsc->mdsc;
3240
3241         dout("mdsc_destroy %p\n", mdsc);
3242         ceph_mdsc_stop(mdsc);
3243
3244         /* flush out any connection work with references to us */
3245         ceph_msgr_flush();
3246
3247         fsc->mdsc = NULL;
3248         kfree(mdsc);
3249         dout("mdsc_destroy %p done\n", mdsc);
3250 }
3251
3252
3253 /*
3254  * handle mds map update.
3255  */
3256 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3257 {
3258         u32 epoch;
3259         u32 maplen;
3260         void *p = msg->front.iov_base;
3261         void *end = p + msg->front.iov_len;
3262         struct ceph_mdsmap *newmap, *oldmap;
3263         struct ceph_fsid fsid;
3264         int err = -EINVAL;
3265
3266         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3267         ceph_decode_copy(&p, &fsid, sizeof(fsid));
3268         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3269                 return;
3270         epoch = ceph_decode_32(&p);
3271         maplen = ceph_decode_32(&p);
3272         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3273
3274         /* do we need it? */
3275         ceph_monc_got_mdsmap(&mdsc->fsc->client->monc, epoch);
3276         mutex_lock(&mdsc->mutex);
3277         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3278                 dout("handle_map epoch %u <= our %u\n",
3279                      epoch, mdsc->mdsmap->m_epoch);
3280                 mutex_unlock(&mdsc->mutex);
3281                 return;
3282         }
3283
3284         newmap = ceph_mdsmap_decode(&p, end);
3285         if (IS_ERR(newmap)) {
3286                 err = PTR_ERR(newmap);
3287                 goto bad_unlock;
3288         }
3289
3290         /* swap into place */
3291         if (mdsc->mdsmap) {
3292                 oldmap = mdsc->mdsmap;
3293                 mdsc->mdsmap = newmap;
3294                 check_new_map(mdsc, newmap, oldmap);
3295                 ceph_mdsmap_destroy(oldmap);
3296         } else {
3297                 mdsc->mdsmap = newmap;  /* first mds map */
3298         }
3299         mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3300
3301         __wake_requests(mdsc, &mdsc->waiting_for_map);
3302
3303         mutex_unlock(&mdsc->mutex);
3304         schedule_delayed(mdsc);
3305         return;
3306
3307 bad_unlock:
3308         mutex_unlock(&mdsc->mutex);
3309 bad:
3310         pr_err("error decoding mdsmap %d\n", err);
3311         return;
3312 }
3313
3314 static struct ceph_connection *con_get(struct ceph_connection *con)
3315 {
3316         struct ceph_mds_session *s = con->private;
3317
3318         if (get_session(s)) {
3319                 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3320                 return con;
3321         }
3322         dout("mdsc con_get %p FAIL\n", s);
3323         return NULL;
3324 }
3325
3326 static void con_put(struct ceph_connection *con)
3327 {
3328         struct ceph_mds_session *s = con->private;
3329
3330         dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3331         ceph_put_mds_session(s);
3332 }
3333
3334 /*
3335  * if the client is unresponsive for long enough, the mds will kill
3336  * the session entirely.
3337  */
3338 static void peer_reset(struct ceph_connection *con)
3339 {
3340         struct ceph_mds_session *s = con->private;
3341         struct ceph_mds_client *mdsc = s->s_mdsc;
3342
3343         pr_warning("mds%d closed our session\n", s->s_mds);
3344         send_mds_reconnect(mdsc, s);
3345 }
3346
3347 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3348 {
3349         struct ceph_mds_session *s = con->private;
3350         struct ceph_mds_client *mdsc = s->s_mdsc;
3351         int type = le16_to_cpu(msg->hdr.type);
3352
3353         mutex_lock(&mdsc->mutex);
3354         if (__verify_registered_session(mdsc, s) < 0) {
3355                 mutex_unlock(&mdsc->mutex);
3356                 goto out;
3357         }
3358         mutex_unlock(&mdsc->mutex);
3359
3360         switch (type) {
3361         case CEPH_MSG_MDS_MAP:
3362                 ceph_mdsc_handle_map(mdsc, msg);
3363                 break;
3364         case CEPH_MSG_CLIENT_SESSION:
3365                 handle_session(s, msg);
3366                 break;
3367         case CEPH_MSG_CLIENT_REPLY:
3368                 handle_reply(s, msg);
3369                 break;
3370         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3371                 handle_forward(mdsc, s, msg);
3372                 break;
3373         case CEPH_MSG_CLIENT_CAPS:
3374                 ceph_handle_caps(s, msg);
3375                 break;
3376         case CEPH_MSG_CLIENT_SNAP:
3377                 ceph_handle_snap(mdsc, s, msg);
3378                 break;
3379         case CEPH_MSG_CLIENT_LEASE:
3380                 handle_lease(mdsc, s, msg);
3381                 break;
3382
3383         default:
3384                 pr_err("received unknown message type %d %s\n", type,
3385                        ceph_msg_type_name(type));
3386         }
3387 out:
3388         ceph_msg_put(msg);
3389 }
3390
3391 /*
3392  * authentication
3393  */
3394 static int get_authorizer(struct ceph_connection *con,
3395                           void **buf, int *len, int *proto,
3396                           void **reply_buf, int *reply_len, int force_new)
3397 {
3398         struct ceph_mds_session *s = con->private;
3399         struct ceph_mds_client *mdsc = s->s_mdsc;
3400         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3401         int ret = 0;
3402
3403         if (force_new && s->s_authorizer) {
3404                 ac->ops->destroy_authorizer(ac, s->s_authorizer);
3405                 s->s_authorizer = NULL;
3406         }
3407         if (s->s_authorizer == NULL) {
3408                 if (ac->ops->create_authorizer) {
3409                         ret = ac->ops->create_authorizer(
3410                                 ac, CEPH_ENTITY_TYPE_MDS,
3411                                 &s->s_authorizer,
3412                                 &s->s_authorizer_buf,
3413                                 &s->s_authorizer_buf_len,
3414                                 &s->s_authorizer_reply_buf,
3415                                 &s->s_authorizer_reply_buf_len);
3416                         if (ret)
3417                                 return ret;
3418                 }
3419         }
3420
3421         *proto = ac->protocol;
3422         *buf = s->s_authorizer_buf;
3423         *len = s->s_authorizer_buf_len;
3424         *reply_buf = s->s_authorizer_reply_buf;
3425         *reply_len = s->s_authorizer_reply_buf_len;
3426         return 0;
3427 }
3428
3429
3430 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3431 {
3432         struct ceph_mds_session *s = con->private;
3433         struct ceph_mds_client *mdsc = s->s_mdsc;
3434         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3435
3436         return ac->ops->verify_authorizer_reply(ac, s->s_authorizer, len);
3437 }
3438
3439 static int invalidate_authorizer(struct ceph_connection *con)
3440 {
3441         struct ceph_mds_session *s = con->private;
3442         struct ceph_mds_client *mdsc = s->s_mdsc;
3443         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3444
3445         if (ac->ops->invalidate_authorizer)
3446                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3447
3448         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
3449 }
3450
3451 static const struct ceph_connection_operations mds_con_ops = {
3452         .get = con_get,
3453         .put = con_put,
3454         .dispatch = dispatch,
3455         .get_authorizer = get_authorizer,
3456         .verify_authorizer_reply = verify_authorizer_reply,
3457         .invalidate_authorizer = invalidate_authorizer,
3458         .peer_reset = peer_reset,
3459 };
3460
3461 /* eof */