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  41  * are no longer valid.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