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  41  * are no longer valid.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         complete_all(&req->r_safe_completion);
 612
 613         ceph_mdsc_put_request(req);
 614 }
 615
 616 /*
 617  * Choose mds to send request to next.  If there is a hint set in the
 618  * request (e.g., due to a prior forward hint from the mds), use that.
 619  * Otherwise, consult frag tree and/or caps to identify the
 620  * appropriate mds.  If all else fails, choose randomly.
 621  *
 622  * Called under mdsc->mutex.
 623  */
 624 static struct dentry *get_nonsnap_parent(struct dentry *dentry)
 625 {
 626         /*
 627          * we don't need to worry about protecting the d_parent access
 628          * here because we never renaming inside the snapped namespace
 629          * except to resplice to another snapdir, and either the old or new
 630          * result is a valid result.
 631          */
 632         while (!IS_ROOT(dentry) && ceph_snap(dentry->d_inode) != CEPH_NOSNAP)
 633                 dentry = dentry->d_parent;
 634         return dentry;
 635 }
 636
 637 static int __choose_mds(struct ceph_mds_client *mdsc,
 638                         struct ceph_mds_request *req)
 639 {
 640         struct inode *inode;
 641         struct ceph_inode_info *ci;
 642         struct ceph_cap *cap;
 643         int mode = req->r_direct_mode;
 644         int mds = -1;
 645         u32 hash = req->r_direct_hash;
 646         bool is_hash = req->r_direct_is_hash;
 647
 648         /*
 649          * is there a specific mds we should try?  ignore hint if we have
 650          * no session and the mds is not up (active or recovering).
 651          */
 652         if (req->r_resend_mds >= 0 &&
 653             (__have_session(mdsc, req->r_resend_mds) ||
 654              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 655                 dout("choose_mds using resend_mds mds%d\n",
 656                      req->r_resend_mds);
 657                 return req->r_resend_mds;
 658         }
 659
 660         if (mode == USE_RANDOM_MDS)
 661                 goto random;
 662
 663         inode = NULL;
 664         if (req->r_inode) {
 665                 inode = req->r_inode;
 666         } else if (req->r_dentry) {
 667                 /* ignore race with rename; old or new d_parent is okay */
 668                 struct dentry *parent = req->r_dentry->d_parent;
 669                 struct inode *dir = parent->d_inode;
 670
 671                 if (dir->i_sb != mdsc->fsc->sb) {
 672                         /* not this fs! */
 673                         inode = req->r_dentry->d_inode;
 674                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
 675                         /* direct snapped/virtual snapdir requests
 676                          * based on parent dir inode */
 677                         struct dentry *dn = get_nonsnap_parent(parent);
 678                         inode = dn->d_inode;
 679                         dout("__choose_mds using nonsnap parent %p\n", inode);
 680                 } else if (req->r_dentry->d_inode) {
 681                         /* dentry target */
 682                         inode = req->r_dentry->d_inode;
 683                 } else {
 684                         /* dir + name */
 685                         inode = dir;
 686                         hash = ceph_dentry_hash(dir, req->r_dentry);
 687                         is_hash = true;
 688                 }
 689         }
 690
 691         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 692              (int)hash, mode);
 693         if (!inode)
 694                 goto random;
 695         ci = ceph_inode(inode);
 696
 697         if (is_hash && S_ISDIR(inode->i_mode)) {
 698                 struct ceph_inode_frag frag;
 699                 int found;
 700
 701                 ceph_choose_frag(ci, hash, &frag, &found);
 702                 if (found) {
 703                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 704                                 u8 r;
 705
 706                                 /* choose a random replica */
 707                                 get_random_bytes(&r, 1);
 708                                 r %= frag.ndist;
 709                                 mds = frag.dist[r];
 710                                 dout("choose_mds %p %llx.%llx "
 711                                      "frag %u mds%d (%d/%d)\n",
 712                                      inode, ceph_vinop(inode),
 713                                      frag.frag, mds,
 714                                      (int)r, frag.ndist);
 715                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 716                                     CEPH_MDS_STATE_ACTIVE)
 717                                         return mds;
 718                         }
 719
 720                         /* since this file/dir wasn't known to be
 721                          * replicated, then we want to look for the
 722                          * authoritative mds. */
 723                         mode = USE_AUTH_MDS;
 724                         if (frag.mds >= 0) {
 725                                 /* choose auth mds */
 726                                 mds = frag.mds;
 727                                 dout("choose_mds %p %llx.%llx "
 728                                      "frag %u mds%d (auth)\n",
 729                                      inode, ceph_vinop(inode), frag.frag, mds);
 730                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 731                                     CEPH_MDS_STATE_ACTIVE)
 732                                         return mds;
 733                         }
 734                 }
 735         }
 736
 737         spin_lock(&ci->i_ceph_lock);
 738         cap = NULL;
 739         if (mode == USE_AUTH_MDS)
 740                 cap = ci->i_auth_cap;
 741         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
 742                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
 743         if (!cap) {
 744                 spin_unlock(&ci->i_ceph_lock);
 745                 goto random;
 746         }
 747         mds = cap->session->s_mds;
 748         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
 749              inode, ceph_vinop(inode), mds,
 750              cap == ci->i_auth_cap ? "auth " : "", cap);
 751         spin_unlock(&ci->i_ceph_lock);
 752         return mds;
 753
 754 random:
 755         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
 756         dout("choose_mds chose random mds%d\n", mds);
 757         return mds;
 758 }
 759
 760
 761 /*
 762  * session messages
 763  */
 764 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
 765 {
 766         struct ceph_msg *msg;
 767         struct ceph_mds_session_head *h;
 768
 769         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
 770                            false);
 771         if (!msg) {
 772                 pr_err("create_session_msg ENOMEM creating msg\n");
 773                 return NULL;
 774         }
 775         h = msg->front.iov_base;
 776         h->op = cpu_to_le32(op);
 777         h->seq = cpu_to_le64(seq);
 778         return msg;
 779 }
 780
 781 /*
 782  * send session open request.
 783  *
 784  * called under mdsc->mutex
 785  */
 786 static int __open_session(struct ceph_mds_client *mdsc,
 787                           struct ceph_mds_session *session)
 788 {
 789         struct ceph_msg *msg;
 790         int mstate;
 791         int mds = session->s_mds;
 792
 793         /* wait for mds to go active? */
 794         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
 795         dout("open_session to mds%d (%s)\n", mds,
 796              ceph_mds_state_name(mstate));
 797         session->s_state = CEPH_MDS_SESSION_OPENING;
 798         session->s_renew_requested = jiffies;
 799
 800         /* send connect message */
 801         msg = create_session_msg(CEPH_SESSION_REQUEST_OPEN, session->s_seq);
 802         if (!msg)
 803                 return -ENOMEM;
 804         ceph_con_send(&session->s_con, msg);
 805         return 0;
 806 }
 807
 808 /*
 809  * open sessions for any export targets for the given mds
 810  *
 811  * called under mdsc->mutex
 812  */
 813 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
 814                                           struct ceph_mds_session *session)
 815 {
 816         struct ceph_mds_info *mi;
 817         struct ceph_mds_session *ts;
 818         int i, mds = session->s_mds;
 819         int target;
 820
 821         if (mds >= mdsc->mdsmap->m_max_mds)
 822                 return;
 823         mi = &mdsc->mdsmap->m_info[mds];
 824         dout("open_export_target_sessions for mds%d (%d targets)\n",
 825              session->s_mds, mi->num_export_targets);
 826
 827         for (i = 0; i < mi->num_export_targets; i++) {
 828                 target = mi->export_targets[i];
 829                 ts = __ceph_lookup_mds_session(mdsc, target);
 830                 if (!ts) {
 831                         ts = register_session(mdsc, target);
 832                         if (IS_ERR(ts))
 833                                 return;
 834                 }
 835                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
 836                     session->s_state == CEPH_MDS_SESSION_CLOSING)
 837                         __open_session(mdsc, session);
 838                 else
 839                         dout(" mds%d target mds%d %p is %s\n", session->s_mds,
 840                              i, ts, session_state_name(ts->s_state));
 841                 ceph_put_mds_session(ts);
 842         }
 843 }
 844
 845 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
 846                                            struct ceph_mds_session *session)
 847 {
 848         mutex_lock(&mdsc->mutex);
 849         __open_export_target_sessions(mdsc, session);
 850         mutex_unlock(&mdsc->mutex);
 851 }
 852
 853 /*
 854  * session caps
 855  */
 856
 857 /*
 858  * Free preallocated cap messages assigned to this session
 859  */
 860 static void cleanup_cap_releases(struct ceph_mds_session *session)
 861 {
 862         struct ceph_msg *msg;
 863
 864         spin_lock(&session->s_cap_lock);
 865         while (!list_empty(&session->s_cap_releases)) {
 866                 msg = list_first_entry(&session->s_cap_releases,
 867                                        struct ceph_msg, list_head);
 868                 list_del_init(&msg->list_head);
 869                 ceph_msg_put(msg);
 870         }
 871         while (!list_empty(&session->s_cap_releases_done)) {
 872                 msg = list_first_entry(&session->s_cap_releases_done,
 873                                        struct ceph_msg, list_head);
 874                 list_del_init(&msg->list_head);
 875                 ceph_msg_put(msg);
 876         }
 877         spin_unlock(&session->s_cap_lock);
 878 }
 879
 880 /*
 881  * Helper to safely iterate over all caps associated with a session, with
 882  * special care taken to handle a racing __ceph_remove_cap().
 883  *
 884  * Caller must hold session s_mutex.
 885  */
 886 static int iterate_session_caps(struct ceph_mds_session *session,
 887                                  int (*cb)(struct inode *, struct ceph_cap *,
 888                                             void *), void *arg)
 889 {
 890         struct list_head *p;
 891         struct ceph_cap *cap;
 892         struct inode *inode, *last_inode = NULL;
 893         struct ceph_cap *old_cap = NULL;
 894         int ret;
 895
 896         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
 897         spin_lock(&session->s_cap_lock);
 898         p = session->s_caps.next;
 899         while (p != &session->s_caps) {
 900                 cap = list_entry(p, struct ceph_cap, session_caps);
 901                 inode = igrab(&cap->ci->vfs_inode);
 902                 if (!inode) {
 903                         p = p->next;
 904                         continue;
 905                 }
 906                 session->s_cap_iterator = cap;
 907                 spin_unlock(&session->s_cap_lock);
 908
 909                 if (last_inode) {
 910                         iput(last_inode);
 911                         last_inode = NULL;
 912                 }
 913                 if (old_cap) {
 914                         ceph_put_cap(session->s_mdsc, old_cap);
 915                         old_cap = NULL;
 916                 }
 917
 918                 ret = cb(inode, cap, arg);
 919                 last_inode = inode;
 920
 921                 spin_lock(&session->s_cap_lock);
 922                 p = p->next;
 923                 if (cap->ci == NULL) {
 924                         dout("iterate_session_caps  finishing cap %p removal\n",
 925                              cap);
 926                         BUG_ON(cap->session != session);
 927                         list_del_init(&cap->session_caps);
 928                         session->s_nr_caps--;
 929                         cap->session = NULL;
 930                         old_cap = cap;  /* put_cap it w/o locks held */
 931                 }
 932                 if (ret < 0)
 933                         goto out;
 934         }
 935         ret = 0;
 936 out:
 937         session->s_cap_iterator = NULL;
 938         spin_unlock(&session->s_cap_lock);
 939
 940         if (last_inode)
 941                 iput(last_inode);
 942         if (old_cap)
 943                 ceph_put_cap(session->s_mdsc, old_cap);
 944
 945         return ret;
 946 }
 947
 948 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
 949                                   void *arg)
 950 {
 951         struct ceph_inode_info *ci = ceph_inode(inode);
 952         int drop = 0;
 953
 954         dout("removing cap %p, ci is %p, inode is %p\n",
 955              cap, ci, &ci->vfs_inode);
 956         spin_lock(&ci->i_ceph_lock);
 957         __ceph_remove_cap(cap);
 958         if (!__ceph_is_any_real_caps(ci)) {
 959                 struct ceph_mds_client *mdsc =
 960                         ceph_sb_to_client(inode->i_sb)->mdsc;
 961
 962                 spin_lock(&mdsc->cap_dirty_lock);
 963                 if (!list_empty(&ci->i_dirty_item)) {
 964                         pr_info(" dropping dirty %s state for %p %lld\n",
 965                                 ceph_cap_string(ci->i_dirty_caps),
 966                                 inode, ceph_ino(inode));
 967                         ci->i_dirty_caps = 0;
 968                         list_del_init(&ci->i_dirty_item);
 969                         drop = 1;
 970                 }
 971                 if (!list_empty(&ci->i_flushing_item)) {
 972                         pr_info(" dropping dirty+flushing %s state for %p %lld\n",
 973                                 ceph_cap_string(ci->i_flushing_caps),
 974                                 inode, ceph_ino(inode));
 975                         ci->i_flushing_caps = 0;
 976                         list_del_init(&ci->i_flushing_item);
 977                         mdsc->num_cap_flushing--;
 978                         drop = 1;
 979                 }
 980                 if (drop && ci->i_wrbuffer_ref) {
 981                         pr_info(" dropping dirty data for %p %lld\n",
 982                                 inode, ceph_ino(inode));
 983                         ci->i_wrbuffer_ref = 0;
 984                         ci->i_wrbuffer_ref_head = 0;
 985                         drop++;
 986                 }
 987                 spin_unlock(&mdsc->cap_dirty_lock);
 988         }
 989         spin_unlock(&ci->i_ceph_lock);
 990         while (drop--)
 991                 iput(inode);
 992         return 0;
 993 }
 994
 995 /*
 996  * caller must hold session s_mutex
 997  */
 998 static void remove_session_caps(struct ceph_mds_session *session)
 999 {
1000         dout("remove_session_caps on %p\n", session);
1001         iterate_session_caps(session, remove_session_caps_cb, NULL);
1002         BUG_ON(session->s_nr_caps > 0);
1003         BUG_ON(!list_empty(&session->s_cap_flushing));
1004         cleanup_cap_releases(session);
1005 }
1006
1007 /*
1008  * wake up any threads waiting on this session's caps.  if the cap is
1009  * old (didn't get renewed on the client reconnect), remove it now.
1010  *
1011  * caller must hold s_mutex.
1012  */
1013 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1014                               void *arg)
1015 {
1016         struct ceph_inode_info *ci = ceph_inode(inode);
1017
1018         wake_up_all(&ci->i_cap_wq);
1019         if (arg) {
1020                 spin_lock(&ci->i_ceph_lock);
1021                 ci->i_wanted_max_size = 0;
1022                 ci->i_requested_max_size = 0;
1023                 spin_unlock(&ci->i_ceph_lock);
1024         }
1025         return 0;
1026 }
1027
1028 static void wake_up_session_caps(struct ceph_mds_session *session,
1029                                  int reconnect)
1030 {
1031         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1032         iterate_session_caps(session, wake_up_session_cb,
1033                              (void *)(unsigned long)reconnect);
1034 }
1035
1036 /*
1037  * Send periodic message to MDS renewing all currently held caps.  The
1038  * ack will reset the expiration for all caps from this session.
1039  *
1040  * caller holds s_mutex
1041  */
1042 static int send_renew_caps(struct ceph_mds_client *mdsc,
1043                            struct ceph_mds_session *session)
1044 {
1045         struct ceph_msg *msg;
1046         int state;
1047
1048         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1049             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1050                 pr_info("mds%d caps stale\n", session->s_mds);
1051         session->s_renew_requested = jiffies;
1052
1053         /* do not try to renew caps until a recovering mds has reconnected
1054          * with its clients. */
1055         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1056         if (state < CEPH_MDS_STATE_RECONNECT) {
1057                 dout("send_renew_caps ignoring mds%d (%s)\n",
1058                      session->s_mds, ceph_mds_state_name(state));
1059                 return 0;
1060         }
1061
1062         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1063                 ceph_mds_state_name(state));
1064         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1065                                  ++session->s_renew_seq);
1066         if (!msg)
1067                 return -ENOMEM;
1068         ceph_con_send(&session->s_con, msg);
1069         return 0;
1070 }
1071
1072 /*
1073  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1074  *
1075  * Called under session->s_mutex
1076  */
1077 static void renewed_caps(struct ceph_mds_client *mdsc,
1078                          struct ceph_mds_session *session, int is_renew)
1079 {
1080         int was_stale;
1081         int wake = 0;
1082
1083         spin_lock(&session->s_cap_lock);
1084         was_stale = is_renew && (session->s_cap_ttl == 0 ||
1085                                  time_after_eq(jiffies, session->s_cap_ttl));
1086
1087         session->s_cap_ttl = session->s_renew_requested +
1088                 mdsc->mdsmap->m_session_timeout*HZ;
1089
1090         if (was_stale) {
1091                 if (time_before(jiffies, session->s_cap_ttl)) {
1092                         pr_info("mds%d caps renewed\n", session->s_mds);
1093                         wake = 1;
1094                 } else {
1095                         pr_info("mds%d caps still stale\n", session->s_mds);
1096                 }
1097         }
1098         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1099              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1100              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1101         spin_unlock(&session->s_cap_lock);
1102
1103         if (wake)
1104                 wake_up_session_caps(session, 0);
1105 }
1106
1107 /*
1108  * send a session close request
1109  */
1110 static int request_close_session(struct ceph_mds_client *mdsc,
1111                                  struct ceph_mds_session *session)
1112 {
1113         struct ceph_msg *msg;
1114
1115         dout("request_close_session mds%d state %s seq %lld\n",
1116              session->s_mds, session_state_name(session->s_state),
1117              session->s_seq);
1118         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1119         if (!msg)
1120                 return -ENOMEM;
1121         ceph_con_send(&session->s_con, msg);
1122         return 0;
1123 }
1124
1125 /*
1126  * Called with s_mutex held.
1127  */
1128 static int __close_session(struct ceph_mds_client *mdsc,
1129                          struct ceph_mds_session *session)
1130 {
1131         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1132                 return 0;
1133         session->s_state = CEPH_MDS_SESSION_CLOSING;
1134         return request_close_session(mdsc, session);
1135 }
1136
1137 /*
1138  * Trim old(er) caps.
1139  *
1140  * Because we can't cache an inode without one or more caps, we do
1141  * this indirectly: if a cap is unused, we prune its aliases, at which
1142  * point the inode will hopefully get dropped to.
1143  *
1144  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1145  * memory pressure from the MDS, though, so it needn't be perfect.
1146  */
1147 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1148 {
1149         struct ceph_mds_session *session = arg;
1150         struct ceph_inode_info *ci = ceph_inode(inode);
1151         int used, oissued, mine;
1152
1153         if (session->s_trim_caps <= 0)
1154                 return -1;
1155
1156         spin_lock(&ci->i_ceph_lock);
1157         mine = cap->issued | cap->implemented;
1158         used = __ceph_caps_used(ci);
1159         oissued = __ceph_caps_issued_other(ci, cap);
1160
1161         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
1162              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1163              ceph_cap_string(used));
1164         if (ci->i_dirty_caps)
1165                 goto out;   /* dirty caps */
1166         if ((used & ~oissued) & mine)
1167                 goto out;   /* we need these caps */
1168
1169         session->s_trim_caps--;
1170         if (oissued) {
1171                 /* we aren't the only cap.. just remove us */
1172                 __ceph_remove_cap(cap);
1173         } else {
1174                 /* try to drop referring dentries */
1175                 spin_unlock(&ci->i_ceph_lock);
1176                 d_prune_aliases(inode);
1177                 dout("trim_caps_cb %p cap %p  pruned, count now %d\n",
1178                      inode, cap, atomic_read(&inode->i_count));
1179                 return 0;
1180         }
1181
1182 out:
1183         spin_unlock(&ci->i_ceph_lock);
1184         return 0;
1185 }
1186
1187 /*
1188  * Trim session cap count down to some max number.
1189  */
1190 static int trim_caps(struct ceph_mds_client *mdsc,
1191                      struct ceph_mds_session *session,
1192                      int max_caps)
1193 {
1194         int trim_caps = session->s_nr_caps - max_caps;
1195
1196         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1197              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1198         if (trim_caps > 0) {
1199                 session->s_trim_caps = trim_caps;
1200                 iterate_session_caps(session, trim_caps_cb, session);
1201                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1202                      session->s_mds, session->s_nr_caps, max_caps,
1203                         trim_caps - session->s_trim_caps);
1204                 session->s_trim_caps = 0;
1205         }
1206         return 0;
1207 }
1208
1209 /*
1210  * Allocate cap_release messages.  If there is a partially full message
1211  * in the queue, try to allocate enough to cover it's remainder, so that
1212  * we can send it immediately.
1213  *
1214  * Called under s_mutex.
1215  */
1216 int ceph_add_cap_releases(struct ceph_mds_client *mdsc,
1217                           struct ceph_mds_session *session)
1218 {
1219         struct ceph_msg *msg, *partial = NULL;
1220         struct ceph_mds_cap_release *head;
1221         int err = -ENOMEM;
1222         int extra = mdsc->fsc->mount_options->cap_release_safety;
1223         int num;
1224
1225         dout("add_cap_releases %p mds%d extra %d\n", session, session->s_mds,
1226              extra);
1227
1228         spin_lock(&session->s_cap_lock);
1229
1230         if (!list_empty(&session->s_cap_releases)) {
1231                 msg = list_first_entry(&session->s_cap_releases,
1232                                        struct ceph_msg,
1233                                  list_head);
1234                 head = msg->front.iov_base;
1235                 num = le32_to_cpu(head->num);
1236                 if (num) {
1237                         dout(" partial %p with (%d/%d)\n", msg, num,
1238                              (int)CEPH_CAPS_PER_RELEASE);
1239                         extra += CEPH_CAPS_PER_RELEASE - num;
1240                         partial = msg;
1241                 }
1242         }
1243         while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1244                 spin_unlock(&session->s_cap_lock);
1245                 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
1246                                    GFP_NOFS, false);
1247                 if (!msg)
1248                         goto out_unlocked;
1249                 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1250                      (int)msg->front.iov_len);
1251                 head = msg->front.iov_base;
1252                 head->num = cpu_to_le32(0);
1253                 msg->front.iov_len = sizeof(*head);
1254                 spin_lock(&session->s_cap_lock);
1255                 list_add(&msg->list_head, &session->s_cap_releases);
1256                 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1257         }
1258
1259         if (partial) {
1260                 head = partial->front.iov_base;
1261                 num = le32_to_cpu(head->num);
1262                 dout(" queueing partial %p with %d/%d\n", partial, num,
1263                      (int)CEPH_CAPS_PER_RELEASE);
1264                 list_move_tail(&partial->list_head,
1265                                &session->s_cap_releases_done);
1266                 session->s_num_cap_releases -= CEPH_CAPS_PER_RELEASE - num;
1267         }
1268         err = 0;
1269         spin_unlock(&session->s_cap_lock);
1270 out_unlocked:
1271         return err;
1272 }
1273
1274 /*
1275  * flush all dirty inode data to disk.
1276  *
1277  * returns true if we've flushed through want_flush_seq
1278  */
1279 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1280 {
1281         int mds, ret = 1;
1282
1283         dout("check_cap_flush want %lld\n", want_flush_seq);
1284         mutex_lock(&mdsc->mutex);
1285         for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1286                 struct ceph_mds_session *session = mdsc->sessions[mds];
1287
1288                 if (!session)
1289                         continue;
1290                 get_session(session);
1291                 mutex_unlock(&mdsc->mutex);
1292
1293                 mutex_lock(&session->s_mutex);
1294                 if (!list_empty(&session->s_cap_flushing)) {
1295                         struct ceph_inode_info *ci =
1296                                 list_entry(session->s_cap_flushing.next,
1297                                            struct ceph_inode_info,
1298                                            i_flushing_item);
1299                         struct inode *inode = &ci->vfs_inode;
1300
1301                         spin_lock(&ci->i_ceph_lock);
1302                         if (ci->i_cap_flush_seq <= want_flush_seq) {
1303                                 dout("check_cap_flush still flushing %p "
1304                                      "seq %lld <= %lld to mds%d\n", inode,
1305                                      ci->i_cap_flush_seq, want_flush_seq,
1306                                      session->s_mds);
1307                                 ret = 0;
1308                         }
1309                         spin_unlock(&ci->i_ceph_lock);
1310                 }
1311                 mutex_unlock(&session->s_mutex);
1312                 ceph_put_mds_session(session);
1313
1314                 if (!ret)
1315                         return ret;
1316                 mutex_lock(&mdsc->mutex);
1317         }
1318
1319         mutex_unlock(&mdsc->mutex);
1320         dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1321         return ret;
1322 }
1323
1324 /*
1325  * called under s_mutex
1326  */
1327 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1328                             struct ceph_mds_session *session)
1329 {
1330         struct ceph_msg *msg;
1331
1332         dout("send_cap_releases mds%d\n", session->s_mds);
1333         spin_lock(&session->s_cap_lock);
1334         while (!list_empty(&session->s_cap_releases_done)) {
1335                 msg = list_first_entry(&session->s_cap_releases_done,
1336                                  struct ceph_msg, list_head);
1337                 list_del_init(&msg->list_head);
1338                 spin_unlock(&session->s_cap_lock);
1339                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1340                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1341                 ceph_con_send(&session->s_con, msg);
1342                 spin_lock(&session->s_cap_lock);
1343         }
1344         spin_unlock(&session->s_cap_lock);
1345 }
1346
1347 static void discard_cap_releases(struct ceph_mds_client *mdsc,
1348                                  struct ceph_mds_session *session)
1349 {
1350         struct ceph_msg *msg;
1351         struct ceph_mds_cap_release *head;
1352         unsigned num;
1353
1354         dout("discard_cap_releases mds%d\n", session->s_mds);
1355         spin_lock(&session->s_cap_lock);
1356
1357         /* zero out the in-progress message */
1358         msg = list_first_entry(&session->s_cap_releases,
1359                                struct ceph_msg, list_head);
1360         head = msg->front.iov_base;
1361         num = le32_to_cpu(head->num);
1362         dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg, num);
1363         head->num = cpu_to_le32(0);
1364         session->s_num_cap_releases += num;
1365
1366         /* requeue completed messages */
1367         while (!list_empty(&session->s_cap_releases_done)) {
1368                 msg = list_first_entry(&session->s_cap_releases_done,
1369                                  struct ceph_msg, list_head);
1370                 list_del_init(&msg->list_head);
1371
1372                 head = msg->front.iov_base;
1373                 num = le32_to_cpu(head->num);
1374                 dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg,
1375                      num);
1376                 session->s_num_cap_releases += num;
1377                 head->num = cpu_to_le32(0);
1378                 msg->front.iov_len = sizeof(*head);
1379                 list_add(&msg->list_head, &session->s_cap_releases);
1380         }
1381
1382         spin_unlock(&session->s_cap_lock);
1383 }
1384
1385 /*
1386  * requests
1387  */
1388
1389 /*
1390  * Create an mds request.
1391  */
1392 struct ceph_mds_request *
1393 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1394 {
1395         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1396
1397         if (!req)
1398                 return ERR_PTR(-ENOMEM);
1399
1400         mutex_init(&req->r_fill_mutex);
1401         req->r_mdsc = mdsc;
1402         req->r_started = jiffies;
1403         req->r_resend_mds = -1;
1404         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1405         req->r_fmode = -1;
1406         kref_init(&req->r_kref);
1407         INIT_LIST_HEAD(&req->r_wait);
1408         init_completion(&req->r_completion);
1409         init_completion(&req->r_safe_completion);
1410         INIT_LIST_HEAD(&req->r_unsafe_item);
1411
1412         req->r_op = op;
1413         req->r_direct_mode = mode;
1414         return req;
1415 }
1416
1417 /*
1418  * return oldest (lowest) request, tid in request tree, 0 if none.
1419  *
1420  * called under mdsc->mutex.
1421  */
1422 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1423 {
1424         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1425                 return NULL;
1426         return rb_entry(rb_first(&mdsc->request_tree),
1427                         struct ceph_mds_request, r_node);
1428 }
1429
1430 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1431 {
1432         struct ceph_mds_request *req = __get_oldest_req(mdsc);
1433
1434         if (req)
1435                 return req->r_tid;
1436         return 0;
1437 }
1438
1439 /*
1440  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1441  * on build_path_from_dentry in fs/cifs/dir.c.
1442  *
1443  * If @stop_on_nosnap, generate path relative to the first non-snapped
1444  * inode.
1445  *
1446  * Encode hidden .snap dirs as a double /, i.e.
1447  *   foo/.snap/bar -> foo//bar
1448  */
1449 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1450                            int stop_on_nosnap)
1451 {
1452         struct dentry *temp;
1453         char *path;
1454         int len, pos;
1455         unsigned seq;
1456
1457         if (dentry == NULL)
1458                 return ERR_PTR(-EINVAL);
1459
1460 retry:
1461         len = 0;
1462         seq = read_seqbegin(&rename_lock);
1463         rcu_read_lock();
1464         for (temp = dentry; !IS_ROOT(temp);) {
1465                 struct inode *inode = temp->d_inode;
1466                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1467                         len++;  /* slash only */
1468                 else if (stop_on_nosnap && inode &&
1469                          ceph_snap(inode) == CEPH_NOSNAP)
1470                         break;
1471                 else
1472                         len += 1 + temp->d_name.len;
1473                 temp = temp->d_parent;
1474                 if (temp == NULL) {
1475                         rcu_read_unlock();
1476                         pr_err("build_path corrupt dentry %p\n", dentry);
1477                         return ERR_PTR(-EINVAL);
1478                 }
1479         }
1480         rcu_read_unlock();
1481         if (len)
1482                 len--;  /* no leading '/' */
1483
1484         path = kmalloc(len+1, GFP_NOFS);
1485         if (path == NULL)
1486                 return ERR_PTR(-ENOMEM);
1487         pos = len;
1488         path[pos] = 0;  /* trailing null */
1489         rcu_read_lock();
1490         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1491                 struct inode *inode;
1492
1493                 spin_lock(&temp->d_lock);
1494                 inode = temp->d_inode;
1495                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1496                         dout("build_path path+%d: %p SNAPDIR\n",
1497                              pos, temp);
1498                 } else if (stop_on_nosnap && inode &&
1499                            ceph_snap(inode) == CEPH_NOSNAP) {
1500                         spin_unlock(&temp->d_lock);
1501                         break;
1502                 } else {
1503                         pos -= temp->d_name.len;
1504                         if (pos < 0) {
1505                                 spin_unlock(&temp->d_lock);
1506                                 break;
1507                         }
1508                         strncpy(path + pos, temp->d_name.name,
1509                                 temp->d_name.len);
1510                 }
1511                 spin_unlock(&temp->d_lock);
1512                 if (pos)
1513                         path[--pos] = '/';
1514                 temp = temp->d_parent;
1515                 if (temp == NULL) {
1516                         rcu_read_unlock();
1517                         pr_err("build_path corrupt dentry\n");
1518                         kfree(path);
1519                         return ERR_PTR(-EINVAL);
1520                 }
1521         }
1522         rcu_read_unlock();
1523         if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1524                 pr_err("build_path did not end path lookup where "
1525                        "expected, namelen is %d, pos is %d\n", len, pos);
1526                 /* presumably this is only possible if racing with a
1527                    rename of one of the parent directories (we can not
1528                    lock the dentries above us to prevent this, but
1529                    retrying should be harmless) */
1530                 kfree(path);
1531                 goto retry;
1532         }
1533
1534         *base = ceph_ino(temp->d_inode);
1535         *plen = len;
1536         dout("build_path on %p %d built %llx '%.*s'\n",
1537              dentry, dentry->d_count, *base, len, path);
1538         return path;
1539 }
1540
1541 static int build_dentry_path(struct dentry *dentry,
1542                              const char **ppath, int *ppathlen, u64 *pino,
1543                              int *pfreepath)
1544 {
1545         char *path;
1546
1547         if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1548                 *pino = ceph_ino(dentry->d_parent->d_inode);
1549                 *ppath = dentry->d_name.name;
1550                 *ppathlen = dentry->d_name.len;
1551                 return 0;
1552         }
1553         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1554         if (IS_ERR(path))
1555                 return PTR_ERR(path);
1556         *ppath = path;
1557         *pfreepath = 1;
1558         return 0;
1559 }
1560
1561 static int build_inode_path(struct inode *inode,
1562                             const char **ppath, int *ppathlen, u64 *pino,
1563                             int *pfreepath)
1564 {
1565         struct dentry *dentry;
1566         char *path;
1567
1568         if (ceph_snap(inode) == CEPH_NOSNAP) {
1569                 *pino = ceph_ino(inode);
1570                 *ppathlen = 0;
1571                 return 0;
1572         }
1573         dentry = d_find_alias(inode);
1574         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1575         dput(dentry);
1576         if (IS_ERR(path))
1577                 return PTR_ERR(path);
1578         *ppath = path;
1579         *pfreepath = 1;
1580         return 0;
1581 }
1582
1583 /*
1584  * request arguments may be specified via an inode *, a dentry *, or
1585  * an explicit ino+path.
1586  */
1587 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1588                                   const char *rpath, u64 rino,
1589                                   const char **ppath, int *pathlen,
1590                                   u64 *ino, int *freepath)
1591 {
1592         int r = 0;
1593
1594         if (rinode) {
1595                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1596                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1597                      ceph_snap(rinode));
1598         } else if (rdentry) {
1599                 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1600                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1601                      *ppath);
1602         } else if (rpath || rino) {
1603                 *ino = rino;
1604                 *ppath = rpath;
1605                 *pathlen = strlen(rpath);
1606                 dout(" path %.*s\n", *pathlen, rpath);
1607         }
1608
1609         return r;
1610 }
1611
1612 /*
1613  * called under mdsc->mutex
1614  */
1615 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1616                                                struct ceph_mds_request *req,
1617                                                int mds)
1618 {
1619         struct ceph_msg *msg;
1620         struct ceph_mds_request_head *head;
1621         const char *path1 = NULL;
1622         const char *path2 = NULL;
1623         u64 ino1 = 0, ino2 = 0;
1624         int pathlen1 = 0, pathlen2 = 0;
1625         int freepath1 = 0, freepath2 = 0;
1626         int len;
1627         u16 releases;
1628         void *p, *end;
1629         int ret;
1630
1631         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1632                               req->r_path1, req->r_ino1.ino,
1633                               &path1, &pathlen1, &ino1, &freepath1);
1634         if (ret < 0) {
1635                 msg = ERR_PTR(ret);
1636                 goto out;
1637         }
1638
1639         ret = set_request_path_attr(NULL, req->r_old_dentry,
1640                               req->r_path2, req->r_ino2.ino,
1641                               &path2, &pathlen2, &ino2, &freepath2);
1642         if (ret < 0) {
1643                 msg = ERR_PTR(ret);
1644                 goto out_free1;
1645         }
1646
1647         len = sizeof(*head) +
1648                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64));
1649
1650         /* calculate (max) length for cap releases */
1651         len += sizeof(struct ceph_mds_request_release) *
1652                 (!!req->r_inode_drop + !!req->r_dentry_drop +
1653                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1654         if (req->r_dentry_drop)
1655                 len += req->r_dentry->d_name.len;
1656         if (req->r_old_dentry_drop)
1657                 len += req->r_old_dentry->d_name.len;
1658
1659         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1660         if (!msg) {
1661                 msg = ERR_PTR(-ENOMEM);
1662                 goto out_free2;
1663         }
1664
1665         msg->hdr.tid = cpu_to_le64(req->r_tid);
1666
1667         head = msg->front.iov_base;
1668         p = msg->front.iov_base + sizeof(*head);
1669         end = msg->front.iov_base + msg->front.iov_len;
1670
1671         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1672         head->op = cpu_to_le32(req->r_op);
1673         head->caller_uid = cpu_to_le32(req->r_uid);
1674         head->caller_gid = cpu_to_le32(req->r_gid);
1675         head->args = req->r_args;
1676
1677         ceph_encode_filepath(&p, end, ino1, path1);
1678         ceph_encode_filepath(&p, end, ino2, path2);
1679
1680         /* make note of release offset, in case we need to replay */
1681         req->r_request_release_offset = p - msg->front.iov_base;
1682
1683         /* cap releases */
1684         releases = 0;
1685         if (req->r_inode_drop)
1686                 releases += ceph_encode_inode_release(&p,
1687                       req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1688                       mds, req->r_inode_drop, req->r_inode_unless, 0);
1689         if (req->r_dentry_drop)
1690                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1691                        mds, req->r_dentry_drop, req->r_dentry_unless);
1692         if (req->r_old_dentry_drop)
1693                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1694                        mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1695         if (req->r_old_inode_drop)
1696                 releases += ceph_encode_inode_release(&p,
1697                       req->r_old_dentry->d_inode,
1698                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1699         head->num_releases = cpu_to_le16(releases);
1700
1701         BUG_ON(p > end);
1702         msg->front.iov_len = p - msg->front.iov_base;
1703         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1704
1705         msg->pages = req->r_pages;
1706         msg->nr_pages = req->r_num_pages;
1707         msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1708         msg->hdr.data_off = cpu_to_le16(0);
1709
1710 out_free2:
1711         if (freepath2)
1712                 kfree((char *)path2);
1713 out_free1:
1714         if (freepath1)
1715                 kfree((char *)path1);
1716 out:
1717         return msg;
1718 }
1719
1720 /*
1721  * called under mdsc->mutex if error, under no mutex if
1722  * success.
1723  */
1724 static void complete_request(struct ceph_mds_client *mdsc,
1725                              struct ceph_mds_request *req)
1726 {
1727         if (req->r_callback)
1728                 req->r_callback(mdsc, req);
1729         else
1730                 complete_all(&req->r_completion);
1731 }
1732
1733 /*
1734  * called under mdsc->mutex
1735  */
1736 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1737                                   struct ceph_mds_request *req,
1738                                   int mds)
1739 {
1740         struct ceph_mds_request_head *rhead;
1741         struct ceph_msg *msg;
1742         int flags = 0;
1743
1744         req->r_attempts++;
1745         if (req->r_inode) {
1746                 struct ceph_cap *cap =
1747                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
1748
1749                 if (cap)
1750                         req->r_sent_on_mseq = cap->mseq;
1751                 else
1752                         req->r_sent_on_mseq = -1;
1753         }
1754         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1755              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1756
1757         if (req->r_got_unsafe) {
1758                 /*
1759                  * Replay.  Do not regenerate message (and rebuild
1760                  * paths, etc.); just use the original message.
1761                  * Rebuilding paths will break for renames because
1762                  * d_move mangles the src name.
1763                  */
1764                 msg = req->r_request;
1765                 rhead = msg->front.iov_base;
1766
1767                 flags = le32_to_cpu(rhead->flags);
1768                 flags |= CEPH_MDS_FLAG_REPLAY;
1769                 rhead->flags = cpu_to_le32(flags);
1770
1771                 if (req->r_target_inode)
1772                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1773
1774                 rhead->num_retry = req->r_attempts - 1;
1775
1776                 /* remove cap/dentry releases from message */
1777                 rhead->num_releases = 0;
1778                 msg->hdr.front_len = cpu_to_le32(req->r_request_release_offset);
1779                 msg->front.iov_len = req->r_request_release_offset;
1780                 return 0;
1781         }
1782
1783         if (req->r_request) {
1784                 ceph_msg_put(req->r_request);
1785                 req->r_request = NULL;
1786         }
1787         msg = create_request_message(mdsc, req, mds);
1788         if (IS_ERR(msg)) {
1789                 req->r_err = PTR_ERR(msg);
1790                 complete_request(mdsc, req);
1791                 return PTR_ERR(msg);
1792         }
1793         req->r_request = msg;
1794
1795         rhead = msg->front.iov_base;
1796         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1797         if (req->r_got_unsafe)
1798                 flags |= CEPH_MDS_FLAG_REPLAY;
1799         if (req->r_locked_dir)
1800                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1801         rhead->flags = cpu_to_le32(flags);
1802         rhead->num_fwd = req->r_num_fwd;
1803         rhead->num_retry = req->r_attempts - 1;
1804         rhead->ino = 0;
1805
1806         dout(" r_locked_dir = %p\n", req->r_locked_dir);
1807         return 0;
1808 }
1809
1810 /*
1811  * send request, or put it on the appropriate wait list.
1812  */
1813 static int __do_request(struct ceph_mds_client *mdsc,
1814                         struct ceph_mds_request *req)
1815 {
1816         struct ceph_mds_session *session = NULL;
1817         int mds = -1;
1818         int err = -EAGAIN;
1819
1820         if (req->r_err || req->r_got_result) {
1821                 if (req->r_aborted)
1822                         __unregister_request(mdsc, req);
1823                 goto out;
1824         }
1825
1826         if (req->r_timeout &&
1827             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1828                 dout("do_request timed out\n");
1829                 err = -EIO;
1830                 goto finish;
1831         }
1832
1833         put_request_session(req);
1834
1835         mds = __choose_mds(mdsc, req);
1836         if (mds < 0 ||
1837             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1838                 dout("do_request no mds or not active, waiting for map\n");
1839                 list_add(&req->r_wait, &mdsc->waiting_for_map);
1840                 goto out;
1841         }
1842
1843         /* get, open session */
1844         session = __ceph_lookup_mds_session(mdsc, mds);
1845         if (!session) {
1846                 session = register_session(mdsc, mds);
1847                 if (IS_ERR(session)) {
1848                         err = PTR_ERR(session);
1849                         goto finish;
1850                 }
1851         }
1852         req->r_session = get_session(session);
1853
1854         dout("do_request mds%d session %p state %s\n", mds, session,
1855              session_state_name(session->s_state));
1856         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1857             session->s_state != CEPH_MDS_SESSION_HUNG) {
1858                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1859                     session->s_state == CEPH_MDS_SESSION_CLOSING)
1860                         __open_session(mdsc, session);
1861                 list_add(&req->r_wait, &session->s_waiting);
1862                 goto out_session;
1863         }
1864
1865         /* send request */
1866         req->r_resend_mds = -1;   /* forget any previous mds hint */
1867
1868         if (req->r_request_started == 0)   /* note request start time */
1869                 req->r_request_started = jiffies;
1870
1871         err = __prepare_send_request(mdsc, req, mds);
1872         if (!err) {
1873                 ceph_msg_get(req->r_request);
1874                 ceph_con_send(&session->s_con, req->r_request);
1875         }
1876
1877 out_session:
1878         ceph_put_mds_session(session);
1879 out:
1880         return err;
1881
1882 finish:
1883         req->r_err = err;
1884         complete_request(mdsc, req);
1885         goto out;
1886 }
1887
1888 /*
1889  * called under mdsc->mutex
1890  */
1891 static void __wake_requests(struct ceph_mds_client *mdsc,
1892                             struct list_head *head)
1893 {
1894         struct ceph_mds_request *req, *nreq;
1895
1896         list_for_each_entry_safe(req, nreq, head, r_wait) {
1897                 list_del_init(&req->r_wait);
1898                 __do_request(mdsc, req);
1899         }
1900 }
1901
1902 /*
1903  * Wake up threads with requests pending for @mds, so that they can
1904  * resubmit their requests to a possibly different mds.
1905  */
1906 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
1907 {
1908         struct ceph_mds_request *req;
1909         struct rb_node *p;
1910
1911         dout("kick_requests mds%d\n", mds);
1912         for (p = rb_first(&mdsc->request_tree); p; p = rb_next(p)) {
1913                 req = rb_entry(p, struct ceph_mds_request, r_node);
1914                 if (req->r_got_unsafe)
1915                         continue;
1916                 if (req->r_session &&
1917                     req->r_session->s_mds == mds) {
1918                         dout(" kicking tid %llu\n", req->r_tid);
1919                         __do_request(mdsc, req);
1920                 }
1921         }
1922 }
1923
1924 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
1925                               struct ceph_mds_request *req)
1926 {
1927         dout("submit_request on %p\n", req);
1928         mutex_lock(&mdsc->mutex);
1929         __register_request(mdsc, req, NULL);
1930         __do_request(mdsc, req);
1931         mutex_unlock(&mdsc->mutex);
1932 }
1933
1934 /*
1935  * Synchrously perform an mds request.  Take care of all of the
1936  * session setup, forwarding, retry details.
1937  */
1938 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
1939                          struct inode *dir,
1940                          struct ceph_mds_request *req)
1941 {
1942         int err;
1943
1944         dout("do_request on %p\n", req);
1945
1946         /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
1947         if (req->r_inode)
1948                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
1949         if (req->r_locked_dir)
1950                 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
1951         if (req->r_old_dentry)
1952                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
1953                                   CEPH_CAP_PIN);
1954
1955         /* issue */
1956         mutex_lock(&mdsc->mutex);
1957         __register_request(mdsc, req, dir);
1958         __do_request(mdsc, req);
1959
1960         if (req->r_err) {
1961                 err = req->r_err;
1962                 __unregister_request(mdsc, req);
1963                 dout("do_request early error %d\n", err);
1964                 goto out;
1965         }
1966
1967         /* wait */
1968         mutex_unlock(&mdsc->mutex);
1969         dout("do_request waiting\n");
1970         if (req->r_timeout) {
1971                 err = (long)wait_for_completion_killable_timeout(
1972                         &req->r_completion, req->r_timeout);
1973                 if (err == 0)
1974                         err = -EIO;
1975         } else {
1976                 err = wait_for_completion_killable(&req->r_completion);
1977         }
1978         dout("do_request waited, got %d\n", err);
1979         mutex_lock(&mdsc->mutex);
1980
1981         /* only abort if we didn't race with a real reply */
1982         if (req->r_got_result) {
1983                 err = le32_to_cpu(req->r_reply_info.head->result);
1984         } else if (err < 0) {
1985                 dout("aborted request %lld with %d\n", req->r_tid, err);
1986
1987                 /*
1988                  * ensure we aren't running concurrently with
1989                  * ceph_fill_trace or ceph_readdir_prepopulate, which
1990                  * rely on locks (dir mutex) held by our caller.
1991                  */
1992                 mutex_lock(&req->r_fill_mutex);
1993                 req->r_err = err;
1994                 req->r_aborted = true;
1995                 mutex_unlock(&req->r_fill_mutex);
1996
1997                 if (req->r_locked_dir &&
1998                     (req->r_op & CEPH_MDS_OP_WRITE))
1999                         ceph_invalidate_dir_request(req);
2000         } else {
2001                 err = req->r_err;
2002         }
2003
2004 out:
2005         mutex_unlock(&mdsc->mutex);
2006         dout("do_request %p done, result %d\n", req, err);
2007         return err;
2008 }
2009
2010 /*
2011  * Invalidate dir D_COMPLETE, dentry lease state on an aborted MDS
2012  * namespace request.
2013  */
2014 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2015 {
2016         struct inode *inode = req->r_locked_dir;
2017         struct ceph_inode_info *ci = ceph_inode(inode);
2018
2019         dout("invalidate_dir_request %p (D_COMPLETE, lease(s))\n", inode);
2020         spin_lock(&ci->i_ceph_lock);
2021         ceph_dir_clear_complete(inode);
2022         ci->i_release_count++;
2023         spin_unlock(&ci->i_ceph_lock);
2024
2025         if (req->r_dentry)
2026                 ceph_invalidate_dentry_lease(req->r_dentry);
2027         if (req->r_old_dentry)
2028                 ceph_invalidate_dentry_lease(req->r_old_dentry);
2029 }
2030
2031 /*
2032  * Handle mds reply.
2033  *
2034  * We take the session mutex and parse and process the reply immediately.
2035  * This preserves the logical ordering of replies, capabilities, etc., sent
2036  * by the MDS as they are applied to our local cache.
2037  */
2038 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2039 {
2040         struct ceph_mds_client *mdsc = session->s_mdsc;
2041         struct ceph_mds_request *req;
2042         struct ceph_mds_reply_head *head = msg->front.iov_base;
2043         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2044         u64 tid;
2045         int err, result;
2046         int mds = session->s_mds;
2047
2048         if (msg->front.iov_len < sizeof(*head)) {
2049                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2050                 ceph_msg_dump(msg);
2051                 return;
2052         }
2053
2054         /* get request, session */
2055         tid = le64_to_cpu(msg->hdr.tid);
2056         mutex_lock(&mdsc->mutex);
2057         req = __lookup_request(mdsc, tid);
2058         if (!req) {
2059                 dout("handle_reply on unknown tid %llu\n", tid);
2060                 mutex_unlock(&mdsc->mutex);
2061                 return;
2062         }
2063         dout("handle_reply %p\n", req);
2064
2065         /* correct session? */
2066         if (req->r_session != session) {
2067                 pr_err("mdsc_handle_reply got %llu on session mds%d"
2068                        " not mds%d\n", tid, session->s_mds,
2069                        req->r_session ? req->r_session->s_mds : -1);
2070                 mutex_unlock(&mdsc->mutex);
2071                 goto out;
2072         }
2073
2074         /* dup? */
2075         if ((req->r_got_unsafe && !head->safe) ||
2076             (req->r_got_safe && head->safe)) {
2077                 pr_warning("got a dup %s reply on %llu from mds%d\n",
2078                            head->safe ? "safe" : "unsafe", tid, mds);
2079                 mutex_unlock(&mdsc->mutex);
2080                 goto out;
2081         }
2082         if (req->r_got_safe && !head->safe) {
2083                 pr_warning("got unsafe after safe on %llu from mds%d\n",
2084                            tid, mds);
2085                 mutex_unlock(&mdsc->mutex);
2086                 goto out;
2087         }
2088
2089         result = le32_to_cpu(head->result);
2090
2091         /*
2092          * Handle an ESTALE
2093          * if we're not talking to the authority, send to them
2094          * if the authority has changed while we weren't looking,
2095          * send to new authority
2096          * Otherwise we just have to return an ESTALE
2097          */
2098         if (result == -ESTALE) {
2099                 dout("got ESTALE on request %llu", req->r_tid);
2100                 if (!req->r_inode) {
2101                         /* do nothing; not an authority problem */
2102                 } else if (req->r_direct_mode != USE_AUTH_MDS) {
2103                         dout("not using auth, setting for that now");
2104                         req->r_direct_mode = USE_AUTH_MDS;
2105                         __do_request(mdsc, req);
2106                         mutex_unlock(&mdsc->mutex);
2107                         goto out;
2108                 } else  {
2109                         struct ceph_inode_info *ci = ceph_inode(req->r_inode);
2110                         struct ceph_cap *cap = NULL;
2111
2112                         if (req->r_session)
2113                                 cap = ceph_get_cap_for_mds(ci,
2114                                                    req->r_session->s_mds);
2115
2116                         dout("already using auth");
2117                         if ((!cap || cap != ci->i_auth_cap) ||
2118                             (cap->mseq != req->r_sent_on_mseq)) {
2119                                 dout("but cap changed, so resending");
2120                                 __do_request(mdsc, req);
2121                                 mutex_unlock(&mdsc->mutex);
2122                                 goto out;
2123                         }
2124                 }
2125                 dout("have to return ESTALE on request %llu", req->r_tid);
2126         }
2127
2128
2129         if (head->safe) {
2130                 req->r_got_safe = true;
2131                 __unregister_request(mdsc, req);
2132
2133                 if (req->r_got_unsafe) {
2134                         /*
2135                          * We already handled the unsafe response, now do the
2136                          * cleanup.  No need to examine the response; the MDS
2137                          * doesn't include any result info in the safe
2138                          * response.  And even if it did, there is nothing
2139                          * useful we could do with a revised return value.
2140                          */
2141                         dout("got safe reply %llu, mds%d\n", tid, mds);
2142                         list_del_init(&req->r_unsafe_item);
2143
2144                         /* last unsafe request during umount? */
2145                         if (mdsc->stopping && !__get_oldest_req(mdsc))
2146                                 complete_all(&mdsc->safe_umount_waiters);
2147                         mutex_unlock(&mdsc->mutex);
2148                         goto out;
2149                 }
2150         } else {
2151                 req->r_got_unsafe = true;
2152                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2153         }
2154
2155         dout("handle_reply tid %lld result %d\n", tid, result);
2156         rinfo = &req->r_reply_info;
2157         err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2158         mutex_unlock(&mdsc->mutex);
2159
2160         mutex_lock(&session->s_mutex);
2161         if (err < 0) {
2162                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2163                 ceph_msg_dump(msg);
2164                 goto out_err;
2165         }
2166
2167         /* snap trace */
2168         if (rinfo->snapblob_len) {
2169                 down_write(&mdsc->snap_rwsem);
2170                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2171                                rinfo->snapblob + rinfo->snapblob_len,
2172                                le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
2173                 downgrade_write(&mdsc->snap_rwsem);
2174         } else {
2175                 down_read(&mdsc->snap_rwsem);
2176         }
2177
2178         /* insert trace into our cache */
2179         mutex_lock(&req->r_fill_mutex);
2180         err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2181         if (err == 0) {
2182                 if (result == 0 && req->r_op != CEPH_MDS_OP_GETFILELOCK &&
2183                     rinfo->dir_nr)
2184                         ceph_readdir_prepopulate(req, req->r_session);
2185                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2186         }
2187         mutex_unlock(&req->r_fill_mutex);
2188
2189         up_read(&mdsc->snap_rwsem);
2190 out_err:
2191         mutex_lock(&mdsc->mutex);
2192         if (!req->r_aborted) {
2193                 if (err) {
2194                         req->r_err = err;
2195                 } else {
2196                         req->r_reply = msg;
2197                         ceph_msg_get(msg);
2198                         req->r_got_result = true;
2199                 }
2200         } else {
2201                 dout("reply arrived after request %lld was aborted\n", tid);
2202         }
2203         mutex_unlock(&mdsc->mutex);
2204
2205         ceph_add_cap_releases(mdsc, req->r_session);
2206         mutex_unlock(&session->s_mutex);
2207
2208         /* kick calling process */
2209         complete_request(mdsc, req);
2210 out:
2211         ceph_mdsc_put_request(req);
2212         return;
2213 }
2214
2215
2216
2217 /*
2218  * handle mds notification that our request has been forwarded.
2219  */
2220 static void handle_forward(struct ceph_mds_client *mdsc,
2221                            struct ceph_mds_session *session,
2222                            struct ceph_msg *msg)
2223 {
2224         struct ceph_mds_request *req;
2225         u64 tid = le64_to_cpu(msg->hdr.tid);
2226         u32 next_mds;
2227         u32 fwd_seq;
2228         int err = -EINVAL;
2229         void *p = msg->front.iov_base;
2230         void *end = p + msg->front.iov_len;
2231
2232         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2233         next_mds = ceph_decode_32(&p);
2234         fwd_seq = ceph_decode_32(&p);
2235
2236         mutex_lock(&mdsc->mutex);
2237         req = __lookup_request(mdsc, tid);
2238         if (!req) {
2239                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2240                 goto out;  /* dup reply? */
2241         }
2242
2243         if (req->r_aborted) {
2244                 dout("forward tid %llu aborted, unregistering\n", tid);
2245                 __unregister_request(mdsc, req);
2246         } else if (fwd_seq <= req->r_num_fwd) {
2247                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2248                      tid, next_mds, req->r_num_fwd, fwd_seq);
2249         } else {
2250                 /* resend. forward race not possible; mds would drop */
2251                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2252                 BUG_ON(req->r_err);
2253                 BUG_ON(req->r_got_result);
2254                 req->r_num_fwd = fwd_seq;
2255                 req->r_resend_mds = next_mds;
2256                 put_request_session(req);
2257                 __do_request(mdsc, req);
2258         }
2259         ceph_mdsc_put_request(req);
2260 out:
2261         mutex_unlock(&mdsc->mutex);
2262         return;
2263
2264 bad:
2265         pr_err("mdsc_handle_forward decode error err=%d\n", err);
2266 }
2267
2268 /*
2269  * handle a mds session control message
2270  */
2271 static void handle_session(struct ceph_mds_session *session,
2272                            struct ceph_msg *msg)
2273 {
2274         struct ceph_mds_client *mdsc = session->s_mdsc;
2275         u32 op;
2276         u64 seq;
2277         int mds = session->s_mds;
2278         struct ceph_mds_session_head *h = msg->front.iov_base;
2279         int wake = 0;
2280
2281         /* decode */
2282         if (msg->front.iov_len != sizeof(*h))
2283                 goto bad;
2284         op = le32_to_cpu(h->op);
2285         seq = le64_to_cpu(h->seq);
2286
2287         mutex_lock(&mdsc->mutex);
2288         if (op == CEPH_SESSION_CLOSE)
2289                 __unregister_session(mdsc, session);
2290         /* FIXME: this ttl calculation is generous */
2291         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2292         mutex_unlock(&mdsc->mutex);
2293
2294         mutex_lock(&session->s_mutex);
2295
2296         dout("handle_session mds%d %s %p state %s seq %llu\n",
2297              mds, ceph_session_op_name(op), session,
2298              session_state_name(session->s_state), seq);
2299
2300         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2301                 session->s_state = CEPH_MDS_SESSION_OPEN;
2302                 pr_info("mds%d came back\n", session->s_mds);
2303         }
2304
2305         switch (op) {
2306         case CEPH_SESSION_OPEN:
2307                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2308                         pr_info("mds%d reconnect success\n", session->s_mds);
2309                 session->s_state = CEPH_MDS_SESSION_OPEN;
2310                 renewed_caps(mdsc, session, 0);
2311                 wake = 1;
2312                 if (mdsc->stopping)
2313                         __close_session(mdsc, session);
2314                 break;
2315
2316         case CEPH_SESSION_RENEWCAPS:
2317                 if (session->s_renew_seq == seq)
2318                         renewed_caps(mdsc, session, 1);
2319                 break;
2320
2321         case CEPH_SESSION_CLOSE:
2322                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2323                         pr_info("mds%d reconnect denied\n", session->s_mds);
2324                 remove_session_caps(session);
2325                 wake = 1; /* for good measure */
2326                 wake_up_all(&mdsc->session_close_wq);
2327                 kick_requests(mdsc, mds);
2328                 break;
2329
2330         case CEPH_SESSION_STALE:
2331                 pr_info("mds%d caps went stale, renewing\n",
2332                         session->s_mds);
2333                 spin_lock(&session->s_cap_lock);
2334                 session->s_cap_gen++;
2335                 session->s_cap_ttl = 0;
2336                 spin_unlock(&session->s_cap_lock);
2337                 send_renew_caps(mdsc, session);
2338                 break;
2339
2340         case CEPH_SESSION_RECALL_STATE:
2341                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2342                 break;
2343
2344         default:
2345                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2346                 WARN_ON(1);
2347         }
2348
2349         mutex_unlock(&session->s_mutex);
2350         if (wake) {
2351                 mutex_lock(&mdsc->mutex);
2352                 __wake_requests(mdsc, &session->s_waiting);
2353                 mutex_unlock(&mdsc->mutex);
2354         }
2355         return;
2356
2357 bad:
2358         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2359                (int)msg->front.iov_len);
2360         ceph_msg_dump(msg);
2361         return;
2362 }
2363
2364
2365 /*
2366  * called under session->mutex.
2367  */
2368 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2369                                    struct ceph_mds_session *session)
2370 {
2371         struct ceph_mds_request *req, *nreq;
2372         int err;
2373
2374         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2375
2376         mutex_lock(&mdsc->mutex);
2377         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2378                 err = __prepare_send_request(mdsc, req, session->s_mds);
2379                 if (!err) {
2380                         ceph_msg_get(req->r_request);
2381                         ceph_con_send(&session->s_con, req->r_request);
2382                 }
2383         }
2384         mutex_unlock(&mdsc->mutex);
2385 }
2386
2387 /*
2388  * Encode information about a cap for a reconnect with the MDS.
2389  */
2390 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2391                           void *arg)
2392 {
2393         union {
2394                 struct ceph_mds_cap_reconnect v2;
2395                 struct ceph_mds_cap_reconnect_v1 v1;
2396         } rec;
2397         size_t reclen;
2398         struct ceph_inode_info *ci;
2399         struct ceph_reconnect_state *recon_state = arg;
2400         struct ceph_pagelist *pagelist = recon_state->pagelist;
2401         char *path;
2402         int pathlen, err;
2403         u64 pathbase;
2404         struct dentry *dentry;
2405
2406         ci = cap->ci;
2407
2408         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2409              inode, ceph_vinop(inode), cap, cap->cap_id,
2410              ceph_cap_string(cap->issued));
2411         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2412         if (err)
2413                 return err;
2414
2415         dentry = d_find_alias(inode);
2416         if (dentry) {
2417                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2418                 if (IS_ERR(path)) {
2419                         err = PTR_ERR(path);
2420                         goto out_dput;
2421                 }
2422         } else {
2423                 path = NULL;
2424                 pathlen = 0;
2425         }
2426         err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2427         if (err)
2428                 goto out_free;
2429
2430         spin_lock(&ci->i_ceph_lock);
2431         cap->seq = 0;        /* reset cap seq */
2432         cap->issue_seq = 0;  /* and issue_seq */
2433
2434         if (recon_state->flock) {
2435                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2436                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2437                 rec.v2.issued = cpu_to_le32(cap->issued);
2438                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2439                 rec.v2.pathbase = cpu_to_le64(pathbase);
2440                 rec.v2.flock_len = 0;
2441                 reclen = sizeof(rec.v2);
2442         } else {
2443                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2444                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2445                 rec.v1.issued = cpu_to_le32(cap->issued);
2446                 rec.v1.size = cpu_to_le64(inode->i_size);
2447                 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2448                 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2449                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2450                 rec.v1.pathbase = cpu_to_le64(pathbase);
2451                 reclen = sizeof(rec.v1);
2452         }
2453         spin_unlock(&ci->i_ceph_lock);
2454
2455         if (recon_state->flock) {
2456                 int num_fcntl_locks, num_flock_locks;
2457                 struct ceph_pagelist_cursor trunc_point;
2458
2459                 ceph_pagelist_set_cursor(pagelist, &trunc_point);
2460                 do {
2461                         lock_flocks();
2462                         ceph_count_locks(inode, &num_fcntl_locks,
2463                                          &num_flock_locks);
2464                         rec.v2.flock_len = (2*sizeof(u32) +
2465                                             (num_fcntl_locks+num_flock_locks) *
2466                                             sizeof(struct ceph_filelock));
2467                         unlock_flocks();
2468
2469                         /* pre-alloc pagelist */
2470                         ceph_pagelist_truncate(pagelist, &trunc_point);
2471                         err = ceph_pagelist_append(pagelist, &rec, reclen);
2472                         if (!err)
2473                                 err = ceph_pagelist_reserve(pagelist,
2474                                                             rec.v2.flock_len);
2475
2476                         /* encode locks */
2477                         if (!err) {
2478                                 lock_flocks();
2479                                 err = ceph_encode_locks(inode,
2480                                                         pagelist,
2481                                                         num_fcntl_locks,
2482                                                         num_flock_locks);
2483                                 unlock_flocks();
2484                         }
2485                 } while (err == -ENOSPC);
2486         } else {
2487                 err = ceph_pagelist_append(pagelist, &rec, reclen);
2488         }
2489
2490 out_free:
2491         kfree(path);
2492 out_dput:
2493         dput(dentry);
2494         return err;
2495 }
2496
2497
2498 /*
2499  * If an MDS fails and recovers, clients need to reconnect in order to
2500  * reestablish shared state.  This includes all caps issued through
2501  * this session _and_ the snap_realm hierarchy.  Because it's not
2502  * clear which snap realms the mds cares about, we send everything we
2503  * know about.. that ensures we'll then get any new info the
2504  * recovering MDS might have.
2505  *
2506  * This is a relatively heavyweight operation, but it's rare.
2507  *
2508  * called with mdsc->mutex held.
2509  */
2510 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2511                                struct ceph_mds_session *session)
2512 {
2513         struct ceph_msg *reply;
2514         struct rb_node *p;
2515         int mds = session->s_mds;
2516         int err = -ENOMEM;
2517         struct ceph_pagelist *pagelist;
2518         struct ceph_reconnect_state recon_state;
2519
2520         pr_info("mds%d reconnect start\n", mds);
2521
2522         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2523         if (!pagelist)
2524                 goto fail_nopagelist;
2525         ceph_pagelist_init(pagelist);
2526
2527         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2528         if (!reply)
2529                 goto fail_nomsg;
2530
2531         mutex_lock(&session->s_mutex);
2532         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2533         session->s_seq = 0;
2534
2535         ceph_con_open(&session->s_con,
2536                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2537
2538         /* replay unsafe requests */
2539         replay_unsafe_requests(mdsc, session);
2540
2541         down_read(&mdsc->snap_rwsem);
2542
2543         dout("session %p state %s\n", session,
2544              session_state_name(session->s_state));
2545
2546         /* drop old cap expires; we're about to reestablish that state */
2547         discard_cap_releases(mdsc, session);
2548
2549         /* traverse this session's caps */
2550         err = ceph_pagelist_encode_32(pagelist, session->s_nr_caps);
2551         if (err)
2552                 goto fail;
2553
2554         recon_state.pagelist = pagelist;
2555         recon_state.flock = session->s_con.peer_features & CEPH_FEATURE_FLOCK;
2556         err = iterate_session_caps(session, encode_caps_cb, &recon_state);
2557         if (err < 0)
2558                 goto fail;
2559
2560         /*
2561          * snaprealms.  we provide mds with the ino, seq (version), and
2562          * parent for all of our realms.  If the mds has any newer info,
2563          * it will tell us.
2564          */
2565         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2566                 struct ceph_snap_realm *realm =
2567                         rb_entry(p, struct ceph_snap_realm, node);
2568                 struct ceph_mds_snaprealm_reconnect sr_rec;
2569
2570                 dout(" adding snap realm %llx seq %lld parent %llx\n",
2571                      realm->ino, realm->seq, realm->parent_ino);
2572                 sr_rec.ino = cpu_to_le64(realm->ino);
2573                 sr_rec.seq = cpu_to_le64(realm->seq);
2574                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2575                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2576                 if (err)
2577                         goto fail;
2578         }
2579
2580         reply->pagelist = pagelist;
2581         if (recon_state.flock)
2582                 reply->hdr.version = cpu_to_le16(2);
2583         reply->hdr.data_len = cpu_to_le32(pagelist->length);
2584         reply->nr_pages = calc_pages_for(0, pagelist->length);
2585         ceph_con_send(&session->s_con, reply);
2586
2587         mutex_unlock(&session->s_mutex);
2588
2589         mutex_lock(&mdsc->mutex);
2590         __wake_requests(mdsc, &session->s_waiting);
2591         mutex_unlock(&mdsc->mutex);
2592
2593         up_read(&mdsc->snap_rwsem);
2594         return;
2595
2596 fail:
2597         ceph_msg_put(reply);
2598         up_read(&mdsc->snap_rwsem);
2599         mutex_unlock(&session->s_mutex);
2600 fail_nomsg:
2601         ceph_pagelist_release(pagelist);
2602         kfree(pagelist);
2603 fail_nopagelist:
2604         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
2605         return;
2606 }
2607
2608
2609 /*
2610  * compare old and new mdsmaps, kicking requests
2611  * and closing out old connections as necessary
2612  *
2613  * called under mdsc->mutex.
2614  */
2615 static void check_new_map(struct ceph_mds_client *mdsc,
2616                           struct ceph_mdsmap *newmap,
2617                           struct ceph_mdsmap *oldmap)
2618 {
2619         int i;
2620         int oldstate, newstate;
2621         struct ceph_mds_session *s;
2622
2623         dout("check_new_map new %u old %u\n",
2624              newmap->m_epoch, oldmap->m_epoch);
2625
2626         for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2627                 if (mdsc->sessions[i] == NULL)
2628                         continue;
2629                 s = mdsc->sessions[i];
2630                 oldstate = ceph_mdsmap_get_state(oldmap, i);
2631                 newstate = ceph_mdsmap_get_state(newmap, i);
2632
2633                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
2634                      i, ceph_mds_state_name(oldstate),
2635                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
2636                      ceph_mds_state_name(newstate),
2637                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
2638                      session_state_name(s->s_state));
2639
2640                 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
2641                            ceph_mdsmap_get_addr(newmap, i),
2642                            sizeof(struct ceph_entity_addr))) {
2643                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2644                                 /* the session never opened, just close it
2645                                  * out now */
2646                                 __wake_requests(mdsc, &s->s_waiting);
2647                                 __unregister_session(mdsc, s);
2648                         } else {
2649                                 /* just close it */
2650                                 mutex_unlock(&mdsc->mutex);
2651                                 mutex_lock(&s->s_mutex);
2652                                 mutex_lock(&mdsc->mutex);
2653                                 ceph_con_close(&s->s_con);
2654                                 mutex_unlock(&s->s_mutex);
2655                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2656                         }
2657
2658                         /* kick any requests waiting on the recovering mds */
2659                         kick_requests(mdsc, i);
2660                 } else if (oldstate == newstate) {
2661                         continue;  /* nothing new with this mds */
2662                 }
2663
2664                 /*
2665                  * send reconnect?
2666                  */
2667                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2668                     newstate >= CEPH_MDS_STATE_RECONNECT) {
2669                         mutex_unlock(&mdsc->mutex);
2670                         send_mds_reconnect(mdsc, s);
2671                         mutex_lock(&mdsc->mutex);
2672                 }
2673
2674                 /*
2675                  * kick request on any mds that has gone active.
2676                  */
2677                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2678                     newstate >= CEPH_MDS_STATE_ACTIVE) {
2679                         if (oldstate != CEPH_MDS_STATE_CREATING &&
2680                             oldstate != CEPH_MDS_STATE_STARTING)
2681                                 pr_info("mds%d recovery completed\n", s->s_mds);
2682                         kick_requests(mdsc, i);
2683                         ceph_kick_flushing_caps(mdsc, s);
2684                         wake_up_session_caps(s, 1);
2685                 }
2686         }
2687
2688         for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
2689                 s = mdsc->sessions[i];
2690                 if (!s)
2691                         continue;
2692                 if (!ceph_mdsmap_is_laggy(newmap, i))
2693                         continue;
2694                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2695                     s->s_state == CEPH_MDS_SESSION_HUNG ||
2696                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
2697                         dout(" connecting to export targets of laggy mds%d\n",
2698                              i);
2699                         __open_export_target_sessions(mdsc, s);
2700                 }
2701         }
2702 }
2703
2704
2705
2706 /*
2707  * leases
2708  */
2709
2710 /*
2711  * caller must hold session s_mutex, dentry->d_lock
2712  */
2713 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2714 {
2715         struct ceph_dentry_info *di = ceph_dentry(dentry);
2716
2717         ceph_put_mds_session(di->lease_session);
2718         di->lease_session = NULL;
2719 }
2720
2721 static void handle_lease(struct ceph_mds_client *mdsc,
2722                          struct ceph_mds_session *session,
2723                          struct ceph_msg *msg)
2724 {
2725         struct super_block *sb = mdsc->fsc->sb;
2726         struct inode *inode;
2727         struct dentry *parent, *dentry;
2728         struct ceph_dentry_info *di;
2729         int mds = session->s_mds;
2730         struct ceph_mds_lease *h = msg->front.iov_base;
2731         u32 seq;
2732         struct ceph_vino vino;
2733         struct qstr dname;
2734         int release = 0;
2735
2736         dout("handle_lease from mds%d\n", mds);
2737
2738         /* decode */
2739         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2740                 goto bad;
2741         vino.ino = le64_to_cpu(h->ino);
2742         vino.snap = CEPH_NOSNAP;
2743         seq = le32_to_cpu(h->seq);
2744         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2745         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2746         if (dname.len != get_unaligned_le32(h+1))
2747                 goto bad;
2748
2749         mutex_lock(&session->s_mutex);
2750         session->s_seq++;
2751
2752         /* lookup inode */
2753         inode = ceph_find_inode(sb, vino);
2754         dout("handle_lease %s, ino %llx %p %.*s\n",
2755              ceph_lease_op_name(h->action), vino.ino, inode,
2756              dname.len, dname.name);
2757         if (inode == NULL) {
2758                 dout("handle_lease no inode %llx\n", vino.ino);
2759                 goto release;
2760         }
2761
2762         /* dentry */
2763         parent = d_find_alias(inode);
2764         if (!parent) {
2765                 dout("no parent dentry on inode %p\n", inode);
2766                 WARN_ON(1);
2767                 goto release;  /* hrm... */
2768         }
2769         dname.hash = full_name_hash(dname.name, dname.len);
2770         dentry = d_lookup(parent, &dname);
2771         dput(parent);
2772         if (!dentry)
2773                 goto release;
2774
2775         spin_lock(&dentry->d_lock);
2776         di = ceph_dentry(dentry);
2777         switch (h->action) {
2778         case CEPH_MDS_LEASE_REVOKE:
2779                 if (di && di->lease_session == session) {
2780                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
2781                                 h->seq = cpu_to_le32(di->lease_seq);
2782                         __ceph_mdsc_drop_dentry_lease(dentry);
2783                 }
2784                 release = 1;
2785                 break;
2786
2787         case CEPH_MDS_LEASE_RENEW:
2788                 if (di && di->lease_session == session &&
2789                     di->lease_gen == session->s_cap_gen &&
2790                     di->lease_renew_from &&
2791                     di->lease_renew_after == 0) {
2792                         unsigned long duration =
2793                                 le32_to_cpu(h->duration_ms) * HZ / 1000;
2794
2795                         di->lease_seq = seq;
2796                         dentry->d_time = di->lease_renew_from + duration;
2797                         di->lease_renew_after = di->lease_renew_from +
2798                                 (duration >> 1);
2799                         di->lease_renew_from = 0;
2800                 }
2801                 break;
2802         }
2803         spin_unlock(&dentry->d_lock);
2804         dput(dentry);
2805
2806         if (!release)
2807                 goto out;
2808
2809 release:
2810         /* let's just reuse the same message */
2811         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2812         ceph_msg_get(msg);
2813         ceph_con_send(&session->s_con, msg);
2814
2815 out:
2816         iput(inode);
2817         mutex_unlock(&session->s_mutex);
2818         return;
2819
2820 bad:
2821         pr_err("corrupt lease message\n");
2822         ceph_msg_dump(msg);
2823 }
2824
2825 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
2826                               struct inode *inode,
2827                               struct dentry *dentry, char action,
2828                               u32 seq)
2829 {
2830         struct ceph_msg *msg;
2831         struct ceph_mds_lease *lease;
2832         int len = sizeof(*lease) + sizeof(u32);
2833         int dnamelen = 0;
2834
2835         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2836              inode, dentry, ceph_lease_op_name(action), session->s_mds);
2837         dnamelen = dentry->d_name.len;
2838         len += dnamelen;
2839
2840         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
2841         if (!msg)
2842                 return;
2843         lease = msg->front.iov_base;
2844         lease->action = action;
2845         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
2846         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
2847         lease->seq = cpu_to_le32(seq);
2848         put_unaligned_le32(dnamelen, lease + 1);
2849         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
2850
2851         /*
2852          * if this is a preemptive lease RELEASE, no need to
2853          * flush request stream, since the actual request will
2854          * soon follow.
2855          */
2856         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
2857
2858         ceph_con_send(&session->s_con, msg);
2859 }
2860
2861 /*
2862  * Preemptively release a lease we expect to invalidate anyway.
2863  * Pass @inode always, @dentry is optional.
2864  */
2865 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2866                              struct dentry *dentry)
2867 {
2868         struct ceph_dentry_info *di;
2869         struct ceph_mds_session *session;
2870         u32 seq;
2871
2872         BUG_ON(inode == NULL);
2873         BUG_ON(dentry == NULL);
2874
2875         /* is dentry lease valid? */
2876         spin_lock(&dentry->d_lock);
2877         di = ceph_dentry(dentry);
2878         if (!di || !di->lease_session ||
2879             di->lease_session->s_mds < 0 ||
2880             di->lease_gen != di->lease_session->s_cap_gen ||
2881             !time_before(jiffies, dentry->d_time)) {
2882                 dout("lease_release inode %p dentry %p -- "
2883                      "no lease\n",
2884                      inode, dentry);
2885                 spin_unlock(&dentry->d_lock);
2886                 return;
2887         }
2888
2889         /* we do have a lease on this dentry; note mds and seq */
2890         session = ceph_get_mds_session(di->lease_session);
2891         seq = di->lease_seq;
2892         __ceph_mdsc_drop_dentry_lease(dentry);
2893         spin_unlock(&dentry->d_lock);
2894
2895         dout("lease_release inode %p dentry %p to mds%d\n",
2896              inode, dentry, session->s_mds);
2897         ceph_mdsc_lease_send_msg(session, inode, dentry,
2898                                  CEPH_MDS_LEASE_RELEASE, seq);
2899         ceph_put_mds_session(session);
2900 }
2901
2902 /*
2903  * drop all leases (and dentry refs) in preparation for umount
2904  */
2905 static void drop_leases(struct ceph_mds_client *mdsc)
2906 {
2907         int i;
2908
2909         dout("drop_leases\n");
2910         mutex_lock(&mdsc->mutex);
2911         for (i = 0; i < mdsc->max_sessions; i++) {
2912                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2913                 if (!s)
2914                         continue;
2915                 mutex_unlock(&mdsc->mutex);
2916                 mutex_lock(&s->s_mutex);
2917                 mutex_unlock(&s->s_mutex);
2918                 ceph_put_mds_session(s);
2919                 mutex_lock(&mdsc->mutex);
2920         }
2921         mutex_unlock(&mdsc->mutex);
2922 }
2923
2924
2925
2926 /*
2927  * delayed work -- periodically trim expired leases, renew caps with mds
2928  */
2929 static void schedule_delayed(struct ceph_mds_client *mdsc)
2930 {
2931         int delay = 5;
2932         unsigned hz = round_jiffies_relative(HZ * delay);
2933         schedule_delayed_work(&mdsc->delayed_work, hz);
2934 }
2935
2936 static void delayed_work(struct work_struct *work)
2937 {
2938         int i;
2939         struct ceph_mds_client *mdsc =
2940                 container_of(work, struct ceph_mds_client, delayed_work.work);
2941         int renew_interval;
2942         int renew_caps;
2943
2944         dout("mdsc delayed_work\n");
2945         ceph_check_delayed_caps(mdsc);
2946
2947         mutex_lock(&mdsc->mutex);
2948         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
2949         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
2950                                    mdsc->last_renew_caps);
2951         if (renew_caps)
2952                 mdsc->last_renew_caps = jiffies;
2953
2954         for (i = 0; i < mdsc->max_sessions; i++) {
2955                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2956                 if (s == NULL)
2957                         continue;
2958                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
2959                         dout("resending session close request for mds%d\n",
2960                              s->s_mds);
2961                         request_close_session(mdsc, s);
2962                         ceph_put_mds_session(s);
2963                         continue;
2964                 }
2965                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
2966                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
2967                                 s->s_state = CEPH_MDS_SESSION_HUNG;
2968                                 pr_info("mds%d hung\n", s->s_mds);
2969                         }
2970                 }
2971                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
2972                         /* this mds is failed or recovering, just wait */
2973                         ceph_put_mds_session(s);
2974                         continue;
2975                 }
2976                 mutex_unlock(&mdsc->mutex);
2977
2978                 mutex_lock(&s->s_mutex);
2979                 if (renew_caps)
2980                         send_renew_caps(mdsc, s);
2981                 else
2982                         ceph_con_keepalive(&s->s_con);
2983                 ceph_add_cap_releases(mdsc, s);
2984                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2985                     s->s_state == CEPH_MDS_SESSION_HUNG)
2986                         ceph_send_cap_releases(mdsc, s);
2987                 mutex_unlock(&s->s_mutex);
2988                 ceph_put_mds_session(s);
2989
2990                 mutex_lock(&mdsc->mutex);
2991         }
2992         mutex_unlock(&mdsc->mutex);
2993
2994         schedule_delayed(mdsc);
2995 }
2996
2997 int ceph_mdsc_init(struct ceph_fs_client *fsc)
2998
2999 {
3000         struct ceph_mds_client *mdsc;
3001
3002         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3003         if (!mdsc)
3004                 return -ENOMEM;
3005         mdsc->fsc = fsc;
3006         fsc->mdsc = mdsc;
3007         mutex_init(&mdsc->mutex);
3008         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3009         if (mdsc->mdsmap == NULL)
3010                 return -ENOMEM;
3011
3012         init_completion(&mdsc->safe_umount_waiters);
3013         init_waitqueue_head(&mdsc->session_close_wq);
3014         INIT_LIST_HEAD(&mdsc->waiting_for_map);
3015         mdsc->sessions = NULL;
3016         mdsc->max_sessions = 0;
3017         mdsc->stopping = 0;
3018         init_rwsem(&mdsc->snap_rwsem);
3019         mdsc->snap_realms = RB_ROOT;
3020         INIT_LIST_HEAD(&mdsc->snap_empty);
3021         spin_lock_init(&mdsc->snap_empty_lock);
3022         mdsc->last_tid = 0;
3023         mdsc->request_tree = RB_ROOT;
3024         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3025         mdsc->last_renew_caps = jiffies;
3026         INIT_LIST_HEAD(&mdsc->cap_delay_list);
3027         spin_lock_init(&mdsc->cap_delay_lock);
3028         INIT_LIST_HEAD(&mdsc->snap_flush_list);
3029         spin_lock_init(&mdsc->snap_flush_lock);
3030         mdsc->cap_flush_seq = 0;
3031         INIT_LIST_HEAD(&mdsc->cap_dirty);
3032         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3033         mdsc->num_cap_flushing = 0;
3034         spin_lock_init(&mdsc->cap_dirty_lock);
3035         init_waitqueue_head(&mdsc->cap_flushing_wq);
3036         spin_lock_init(&mdsc->dentry_lru_lock);
3037         INIT_LIST_HEAD(&mdsc->dentry_lru);
3038
3039         ceph_caps_init(mdsc);
3040         ceph_adjust_min_caps(mdsc, fsc->min_caps);
3041
3042         return 0;
3043 }
3044
3045 /*
3046  * Wait for safe replies on open mds requests.  If we time out, drop
3047  * all requests from the tree to avoid dangling dentry refs.
3048  */
3049 static void wait_requests(struct ceph_mds_client *mdsc)
3050 {
3051         struct ceph_mds_request *req;
3052         struct ceph_fs_client *fsc = mdsc->fsc;
3053
3054         mutex_lock(&mdsc->mutex);
3055         if (__get_oldest_req(mdsc)) {
3056                 mutex_unlock(&mdsc->mutex);
3057
3058                 dout("wait_requests waiting for requests\n");
3059                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3060                                     fsc->client->options->mount_timeout * HZ);
3061
3062                 /* tear down remaining requests */
3063                 mutex_lock(&mdsc->mutex);
3064                 while ((req = __get_oldest_req(mdsc))) {
3065                         dout("wait_requests timed out on tid %llu\n",
3066                              req->r_tid);
3067                         __unregister_request(mdsc, req);
3068                 }
3069         }
3070         mutex_unlock(&mdsc->mutex);
3071         dout("wait_requests done\n");
3072 }
3073
3074 /*
3075  * called before mount is ro, and before dentries are torn down.
3076  * (hmm, does this still race with new lookups?)
3077  */
3078 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3079 {
3080         dout("pre_umount\n");
3081         mdsc->stopping = 1;
3082
3083         drop_leases(mdsc);
3084         ceph_flush_dirty_caps(mdsc);
3085         wait_requests(mdsc);
3086
3087         /*
3088          * wait for reply handlers to drop their request refs and
3089          * their inode/dcache refs
3090          */
3091         ceph_msgr_flush();
3092 }
3093
3094 /*
3095  * wait for all write mds requests to flush.
3096  */
3097 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3098 {
3099         struct ceph_mds_request *req = NULL, *nextreq;
3100         struct rb_node *n;
3101
3102         mutex_lock(&mdsc->mutex);
3103         dout("wait_unsafe_requests want %lld\n", want_tid);
3104 restart:
3105         req = __get_oldest_req(mdsc);
3106         while (req && req->r_tid <= want_tid) {
3107                 /* find next request */
3108                 n = rb_next(&req->r_node);
3109                 if (n)
3110                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3111                 else
3112                         nextreq = NULL;
3113                 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
3114                         /* write op */
3115                         ceph_mdsc_get_request(req);
3116                         if (nextreq)
3117                                 ceph_mdsc_get_request(nextreq);
3118                         mutex_unlock(&mdsc->mutex);
3119                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
3120                              req->r_tid, want_tid);
3121                         wait_for_completion(&req->r_safe_completion);
3122                         mutex_lock(&mdsc->mutex);
3123                         ceph_mdsc_put_request(req);
3124                         if (!nextreq)
3125                                 break;  /* next dne before, so we're done! */
3126                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
3127                                 /* next request was removed from tree */
3128                                 ceph_mdsc_put_request(nextreq);
3129                                 goto restart;
3130                         }
3131                         ceph_mdsc_put_request(nextreq);  /* won't go away */
3132                 }
3133                 req = nextreq;
3134         }
3135         mutex_unlock(&mdsc->mutex);
3136         dout("wait_unsafe_requests done\n");
3137 }
3138
3139 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3140 {
3141         u64 want_tid, want_flush;
3142
3143         if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3144                 return;
3145
3146         dout("sync\n");
3147         mutex_lock(&mdsc->mutex);
3148         want_tid = mdsc->last_tid;
3149         want_flush = mdsc->cap_flush_seq;
3150         mutex_unlock(&mdsc->mutex);
3151         dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
3152
3153         ceph_flush_dirty_caps(mdsc);
3154
3155         wait_unsafe_requests(mdsc, want_tid);
3156         wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
3157 }
3158
3159 /*
3160  * true if all sessions are closed, or we force unmount
3161  */
3162 static bool done_closing_sessions(struct ceph_mds_client *mdsc)
3163 {
3164         int i, n = 0;
3165
3166         if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3167                 return true;
3168
3169         mutex_lock(&mdsc->mutex);
3170         for (i = 0; i < mdsc->max_sessions; i++)
3171                 if (mdsc->sessions[i])
3172                         n++;
3173         mutex_unlock(&mdsc->mutex);
3174         return n == 0;
3175 }
3176
3177 /*
3178  * called after sb is ro.
3179  */
3180 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3181 {
3182         struct ceph_mds_session *session;
3183         int i;
3184         struct ceph_fs_client *fsc = mdsc->fsc;
3185         unsigned long timeout = fsc->client->options->mount_timeout * HZ;
3186
3187         dout("close_sessions\n");
3188
3189         /* close sessions */
3190         mutex_lock(&mdsc->mutex);
3191         for (i = 0; i < mdsc->max_sessions; i++) {
3192                 session = __ceph_lookup_mds_session(mdsc, i);
3193                 if (!session)
3194                         continue;
3195                 mutex_unlock(&mdsc->mutex);
3196                 mutex_lock(&session->s_mutex);
3197                 __close_session(mdsc, session);
3198                 mutex_unlock(&session->s_mutex);
3199                 ceph_put_mds_session(session);
3200                 mutex_lock(&mdsc->mutex);
3201         }
3202         mutex_unlock(&mdsc->mutex);
3203
3204         dout("waiting for sessions to close\n");
3205         wait_event_timeout(mdsc->session_close_wq, done_closing_sessions(mdsc),
3206                            timeout);
3207
3208         /* tear down remaining sessions */
3209         mutex_lock(&mdsc->mutex);
3210         for (i = 0; i < mdsc->max_sessions; i++) {
3211                 if (mdsc->sessions[i]) {
3212                         session = get_session(mdsc->sessions[i]);
3213                         __unregister_session(mdsc, session);
3214                         mutex_unlock(&mdsc->mutex);
3215                         mutex_lock(&session->s_mutex);
3216                         remove_session_caps(session);
3217                         mutex_unlock(&session->s_mutex);
3218                         ceph_put_mds_session(session);
3219                         mutex_lock(&mdsc->mutex);
3220                 }
3221         }
3222         WARN_ON(!list_empty(&mdsc->cap_delay_list));
3223         mutex_unlock(&mdsc->mutex);
3224
3225         ceph_cleanup_empty_realms(mdsc);
3226
3227         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3228
3229         dout("stopped\n");
3230 }
3231
3232 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3233 {
3234         dout("stop\n");
3235         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3236         if (mdsc->mdsmap)
3237                 ceph_mdsmap_destroy(mdsc->mdsmap);
3238         kfree(mdsc->sessions);
3239         ceph_caps_finalize(mdsc);
3240 }
3241
3242 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3243 {
3244         struct ceph_mds_client *mdsc = fsc->mdsc;
3245
3246         dout("mdsc_destroy %p\n", mdsc);
3247         ceph_mdsc_stop(mdsc);
3248
3249         /* flush out any connection work with references to us */
3250         ceph_msgr_flush();
3251
3252         fsc->mdsc = NULL;
3253         kfree(mdsc);
3254         dout("mdsc_destroy %p done\n", mdsc);
3255 }
3256
3257
3258 /*
3259  * handle mds map update.
3260  */
3261 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3262 {
3263         u32 epoch;
3264         u32 maplen;
3265         void *p = msg->front.iov_base;
3266         void *end = p + msg->front.iov_len;
3267         struct ceph_mdsmap *newmap, *oldmap;
3268         struct ceph_fsid fsid;
3269         int err = -EINVAL;
3270
3271         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3272         ceph_decode_copy(&p, &fsid, sizeof(fsid));
3273         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3274                 return;
3275         epoch = ceph_decode_32(&p);
3276         maplen = ceph_decode_32(&p);
3277         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3278
3279         /* do we need it? */
3280         ceph_monc_got_mdsmap(&mdsc->fsc->client->monc, epoch);
3281         mutex_lock(&mdsc->mutex);
3282         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3283                 dout("handle_map epoch %u <= our %u\n",
3284                      epoch, mdsc->mdsmap->m_epoch);
3285                 mutex_unlock(&mdsc->mutex);
3286                 return;
3287         }
3288
3289         newmap = ceph_mdsmap_decode(&p, end);
3290         if (IS_ERR(newmap)) {
3291                 err = PTR_ERR(newmap);
3292                 goto bad_unlock;
3293         }
3294
3295         /* swap into place */
3296         if (mdsc->mdsmap) {
3297                 oldmap = mdsc->mdsmap;
3298                 mdsc->mdsmap = newmap;
3299                 check_new_map(mdsc, newmap, oldmap);
3300                 ceph_mdsmap_destroy(oldmap);
3301         } else {
3302                 mdsc->mdsmap = newmap;  /* first mds map */
3303         }
3304         mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3305
3306         __wake_requests(mdsc, &mdsc->waiting_for_map);
3307
3308         mutex_unlock(&mdsc->mutex);
3309         schedule_delayed(mdsc);
3310         return;
3311
3312 bad_unlock:
3313         mutex_unlock(&mdsc->mutex);
3314 bad:
3315         pr_err("error decoding mdsmap %d\n", err);
3316         return;
3317 }
3318
3319 static struct ceph_connection *con_get(struct ceph_connection *con)
3320 {
3321         struct ceph_mds_session *s = con->private;
3322
3323         if (get_session(s)) {
3324                 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3325                 return con;
3326         }
3327         dout("mdsc con_get %p FAIL\n", s);
3328         return NULL;
3329 }
3330
3331 static void con_put(struct ceph_connection *con)
3332 {
3333         struct ceph_mds_session *s = con->private;
3334
3335         dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3336         ceph_put_mds_session(s);
3337 }
3338
3339 /*
3340  * if the client is unresponsive for long enough, the mds will kill
3341  * the session entirely.
3342  */
3343 static void peer_reset(struct ceph_connection *con)
3344 {
3345         struct ceph_mds_session *s = con->private;
3346         struct ceph_mds_client *mdsc = s->s_mdsc;
3347
3348         pr_warning("mds%d closed our session\n", s->s_mds);
3349         send_mds_reconnect(mdsc, s);
3350 }
3351
3352 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3353 {
3354         struct ceph_mds_session *s = con->private;
3355         struct ceph_mds_client *mdsc = s->s_mdsc;
3356         int type = le16_to_cpu(msg->hdr.type);
3357
3358         mutex_lock(&mdsc->mutex);
3359         if (__verify_registered_session(mdsc, s) < 0) {
3360                 mutex_unlock(&mdsc->mutex);
3361                 goto out;
3362         }
3363         mutex_unlock(&mdsc->mutex);
3364
3365         switch (type) {
3366         case CEPH_MSG_MDS_MAP:
3367                 ceph_mdsc_handle_map(mdsc, msg);
3368                 break;
3369         case CEPH_MSG_CLIENT_SESSION:
3370                 handle_session(s, msg);
3371                 break;
3372         case CEPH_MSG_CLIENT_REPLY:
3373                 handle_reply(s, msg);
3374                 break;
3375         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3376                 handle_forward(mdsc, s, msg);
3377                 break;
3378         case CEPH_MSG_CLIENT_CAPS:
3379                 ceph_handle_caps(s, msg);
3380                 break;
3381         case CEPH_MSG_CLIENT_SNAP:
3382                 ceph_handle_snap(mdsc, s, msg);
3383                 break;
3384         case CEPH_MSG_CLIENT_LEASE:
3385                 handle_lease(mdsc, s, msg);
3386                 break;
3387
3388         default:
3389                 pr_err("received unknown message type %d %s\n", type,
3390                        ceph_msg_type_name(type));
3391         }
3392 out:
3393         ceph_msg_put(msg);
3394 }
3395
3396 /*
3397  * authentication
3398  */
3399 static int get_authorizer(struct ceph_connection *con,
3400                           void **buf, int *len, int *proto,
3401                           void **reply_buf, int *reply_len, int force_new)
3402 {
3403         struct ceph_mds_session *s = con->private;
3404         struct ceph_mds_client *mdsc = s->s_mdsc;
3405         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3406         int ret = 0;
3407
3408         if (force_new && s->s_authorizer) {
3409                 ac->ops->destroy_authorizer(ac, s->s_authorizer);
3410                 s->s_authorizer = NULL;
3411         }
3412         if (s->s_authorizer == NULL) {
3413                 if (ac->ops->create_authorizer) {
3414                         ret = ac->ops->create_authorizer(
3415                                 ac, CEPH_ENTITY_TYPE_MDS,
3416                                 &s->s_authorizer,
3417                                 &s->s_authorizer_buf,
3418                                 &s->s_authorizer_buf_len,
3419                                 &s->s_authorizer_reply_buf,
3420                                 &s->s_authorizer_reply_buf_len);
3421                         if (ret)
3422                                 return ret;
3423                 }
3424         }
3425
3426         *proto = ac->protocol;
3427         *buf = s->s_authorizer_buf;
3428         *len = s->s_authorizer_buf_len;
3429         *reply_buf = s->s_authorizer_reply_buf;
3430         *reply_len = s->s_authorizer_reply_buf_len;
3431         return 0;
3432 }
3433
3434
3435 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3436 {
3437         struct ceph_mds_session *s = con->private;
3438         struct ceph_mds_client *mdsc = s->s_mdsc;
3439         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3440
3441         return ac->ops->verify_authorizer_reply(ac, s->s_authorizer, len);
3442 }
3443
3444 static int invalidate_authorizer(struct ceph_connection *con)
3445 {
3446         struct ceph_mds_session *s = con->private;
3447         struct ceph_mds_client *mdsc = s->s_mdsc;
3448         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3449
3450         if (ac->ops->invalidate_authorizer)
3451                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3452
3453         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
3454 }
3455
3456 static const struct ceph_connection_operations mds_con_ops = {
3457         .get = con_get,
3458         .put = con_put,
3459         .dispatch = dispatch,
3460         .get_authorizer = get_authorizer,
3461         .verify_authorizer_reply = verify_authorizer_reply,
3462         .invalidate_authorizer = invalidate_authorizer,
3463         .peer_reset = peer_reset,
3464 };
3465
3466 /* eof */