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  41  * are no longer valid.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         u32 op = le32_to_cpu(info->head->op);
 242
 243         if (op == CEPH_MDS_OP_GETFILELOCK)
 244                 return parse_reply_info_filelock(p, end, info, features);
 245         else
 246                 return parse_reply_info_dir(p, end, info, features);
 247 }
 248
 249 /*
 250  * parse entire mds reply
 251  */
 252 static int parse_reply_info(struct ceph_msg *msg,
 253                             struct ceph_mds_reply_info_parsed *info,
 254                             int features)
 255 {
 256         void *p, *end;
 257         u32 len;
 258         int err;
 259
 260         info->head = msg->front.iov_base;
 261         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 262         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 263
 264         /* trace */
 265         ceph_decode_32_safe(&p, end, len, bad);
 266         if (len > 0) {
 267                 err = parse_reply_info_trace(&p, p+len, info, features);
 268                 if (err < 0)
 269                         goto out_bad;
 270         }
 271
 272         /* extra */
 273         ceph_decode_32_safe(&p, end, len, bad);
 274         if (len > 0) {
 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_authorizer)
 338                      s->s_mdsc->fsc->client->monc.auth->ops->destroy_authorizer(
 339                              s->s_mdsc->fsc->client->monc.auth,
 340                              s->s_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(mdsc->fsc->client->msgr, &s->s_con);
 398         s->s_con.private = s;
 399         s->s_con.ops = &mds_con_ops;
 400         s->s_con.peer_name.type = CEPH_ENTITY_TYPE_MDS;
 401         s->s_con.peer_name.num = cpu_to_le64(mds);
 402
 403         spin_lock_init(&s->s_cap_lock);
 404         s->s_cap_gen = 0;
 405         s->s_cap_ttl = 0;
 406         s->s_renew_requested = 0;
 407         s->s_renew_seq = 0;
 408         INIT_LIST_HEAD(&s->s_caps);
 409         s->s_nr_caps = 0;
 410         s->s_trim_caps = 0;
 411         atomic_set(&s->s_ref, 1);
 412         INIT_LIST_HEAD(&s->s_waiting);
 413         INIT_LIST_HEAD(&s->s_unsafe);
 414         s->s_num_cap_releases = 0;
 415         s->s_cap_iterator = NULL;
 416         INIT_LIST_HEAD(&s->s_cap_releases);
 417         INIT_LIST_HEAD(&s->s_cap_releases_done);
 418         INIT_LIST_HEAD(&s->s_cap_flushing);
 419         INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
 420
 421         dout("register_session mds%d\n", mds);
 422         if (mds >= mdsc->max_sessions) {
 423                 int newmax = 1 << get_count_order(mds+1);
 424                 struct ceph_mds_session **sa;
 425
 426                 dout("register_session realloc to %d\n", newmax);
 427                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 428                 if (sa == NULL)
 429                         goto fail_realloc;
 430                 if (mdsc->sessions) {
 431                         memcpy(sa, mdsc->sessions,
 432                                mdsc->max_sessions * sizeof(void *));
 433                         kfree(mdsc->sessions);
 434                 }
 435                 mdsc->sessions = sa;
 436                 mdsc->max_sessions = newmax;
 437         }
 438         mdsc->sessions[mds] = s;
 439         atomic_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
 440
 441         ceph_con_open(&s->s_con, ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
 442
 443         return s;
 444
 445 fail_realloc:
 446         kfree(s);
 447         return ERR_PTR(-ENOMEM);
 448 }
 449
 450 /*
 451  * called under mdsc->mutex
 452  */
 453 static void __unregister_session(struct ceph_mds_client *mdsc,
 454                                struct ceph_mds_session *s)
 455 {
 456         dout("__unregister_session mds%d %p\n", s->s_mds, s);
 457         BUG_ON(mdsc->sessions[s->s_mds] != s);
 458         mdsc->sessions[s->s_mds] = NULL;
 459         ceph_con_close(&s->s_con);
 460         ceph_put_mds_session(s);
 461 }
 462
 463 /*
 464  * drop session refs in request.
 465  *
 466  * should be last request ref, or hold mdsc->mutex
 467  */
 468 static void put_request_session(struct ceph_mds_request *req)
 469 {
 470         if (req->r_session) {
 471                 ceph_put_mds_session(req->r_session);
 472                 req->r_session = NULL;
 473         }
 474 }
 475
 476 void ceph_mdsc_release_request(struct kref *kref)
 477 {
 478         struct ceph_mds_request *req = container_of(kref,
 479                                                     struct ceph_mds_request,
 480                                                     r_kref);
 481         if (req->r_request)
 482                 ceph_msg_put(req->r_request);
 483         if (req->r_reply) {
 484                 ceph_msg_put(req->r_reply);
 485                 destroy_reply_info(&req->r_reply_info);
 486         }
 487         if (req->r_inode) {
 488                 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
 489                 iput(req->r_inode);
 490         }
 491         if (req->r_locked_dir)
 492                 ceph_put_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
 493         if (req->r_target_inode)
 494                 iput(req->r_target_inode);
 495         if (req->r_dentry)
 496                 dput(req->r_dentry);
 497         if (req->r_old_dentry) {
 498                 /*
 499                  * track (and drop pins for) r_old_dentry_dir
 500                  * separately, since r_old_dentry's d_parent may have
 501                  * changed between the dir mutex being dropped and
 502                  * this request being freed.
 503                  */
 504                 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
 505                                   CEPH_CAP_PIN);
 506                 dput(req->r_old_dentry);
 507                 iput(req->r_old_dentry_dir);
 508         }
 509         kfree(req->r_path1);
 510         kfree(req->r_path2);
 511         put_request_session(req);
 512         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
 513         kfree(req);
 514 }
 515
 516 /*
 517  * lookup session, bump ref if found.
 518  *
 519  * called under mdsc->mutex.
 520  */
 521 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
 522                                              u64 tid)
 523 {
 524         struct ceph_mds_request *req;
 525         struct rb_node *n = mdsc->request_tree.rb_node;
 526
 527         while (n) {
 528                 req = rb_entry(n, struct ceph_mds_request, r_node);
 529                 if (tid < req->r_tid)
 530                         n = n->rb_left;
 531                 else if (tid > req->r_tid)
 532                         n = n->rb_right;
 533                 else {
 534                         ceph_mdsc_get_request(req);
 535                         return req;
 536                 }
 537         }
 538         return NULL;
 539 }
 540
 541 static void __insert_request(struct ceph_mds_client *mdsc,
 542                              struct ceph_mds_request *new)
 543 {
 544         struct rb_node **p = &mdsc->request_tree.rb_node;
 545         struct rb_node *parent = NULL;
 546         struct ceph_mds_request *req = NULL;
 547
 548         while (*p) {
 549                 parent = *p;
 550                 req = rb_entry(parent, struct ceph_mds_request, r_node);
 551                 if (new->r_tid < req->r_tid)
 552                         p = &(*p)->rb_left;
 553                 else if (new->r_tid > req->r_tid)
 554                         p = &(*p)->rb_right;
 555                 else
 556                         BUG();
 557         }
 558
 559         rb_link_node(&new->r_node, parent, p);
 560         rb_insert_color(&new->r_node, &mdsc->request_tree);
 561 }
 562
 563 /*
 564  * Register an in-flight request, and assign a tid.  Link to directory
 565  * are modifying (if any).
 566  *
 567  * Called under mdsc->mutex.
 568  */
 569 static void __register_request(struct ceph_mds_client *mdsc,
 570                                struct ceph_mds_request *req,
 571                                struct inode *dir)
 572 {
 573         req->r_tid = ++mdsc->last_tid;
 574         if (req->r_num_caps)
 575                 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
 576                                   req->r_num_caps);
 577         dout("__register_request %p tid %lld\n", req, req->r_tid);
 578         ceph_mdsc_get_request(req);
 579         __insert_request(mdsc, req);
 580
 581         req->r_uid = current_fsuid();
 582         req->r_gid = current_fsgid();
 583
 584         if (dir) {
 585                 struct ceph_inode_info *ci = ceph_inode(dir);
 586
 587                 ihold(dir);
 588                 spin_lock(&ci->i_unsafe_lock);
 589                 req->r_unsafe_dir = dir;
 590                 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
 591                 spin_unlock(&ci->i_unsafe_lock);
 592         }
 593 }
 594
 595 static void __unregister_request(struct ceph_mds_client *mdsc,
 596                                  struct ceph_mds_request *req)
 597 {
 598         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 599         rb_erase(&req->r_node, &mdsc->request_tree);
 600         RB_CLEAR_NODE(&req->r_node);
 601
 602         if (req->r_unsafe_dir) {
 603                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 604
 605                 spin_lock(&ci->i_unsafe_lock);
 606                 list_del_init(&req->r_unsafe_dir_item);
 607                 spin_unlock(&ci->i_unsafe_lock);
 608
 609                 iput(req->r_unsafe_dir);
 610                 req->r_unsafe_dir = NULL;
 611         }
 612
 613         complete_all(&req->r_safe_completion);
 614
 615         ceph_mdsc_put_request(req);
 616 }
 617
 618 /*
 619  * Choose mds to send request to next.  If there is a hint set in the
 620  * request (e.g., due to a prior forward hint from the mds), use that.
 621  * Otherwise, consult frag tree and/or caps to identify the
 622  * appropriate mds.  If all else fails, choose randomly.
 623  *
 624  * Called under mdsc->mutex.
 625  */
 626 static struct dentry *get_nonsnap_parent(struct dentry *dentry)
 627 {
 628         /*
 629          * we don't need to worry about protecting the d_parent access
 630          * here because we never renaming inside the snapped namespace
 631          * except to resplice to another snapdir, and either the old or new
 632          * result is a valid result.
 633          */
 634         while (!IS_ROOT(dentry) && ceph_snap(dentry->d_inode) != CEPH_NOSNAP)
 635                 dentry = dentry->d_parent;
 636         return dentry;
 637 }
 638
 639 static int __choose_mds(struct ceph_mds_client *mdsc,
 640                         struct ceph_mds_request *req)
 641 {
 642         struct inode *inode;
 643         struct ceph_inode_info *ci;
 644         struct ceph_cap *cap;
 645         int mode = req->r_direct_mode;
 646         int mds = -1;
 647         u32 hash = req->r_direct_hash;
 648         bool is_hash = req->r_direct_is_hash;
 649
 650         /*
 651          * is there a specific mds we should try?  ignore hint if we have
 652          * no session and the mds is not up (active or recovering).
 653          */
 654         if (req->r_resend_mds >= 0 &&
 655             (__have_session(mdsc, req->r_resend_mds) ||
 656              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 657                 dout("choose_mds using resend_mds mds%d\n",
 658                      req->r_resend_mds);
 659                 return req->r_resend_mds;
 660         }
 661
 662         if (mode == USE_RANDOM_MDS)
 663                 goto random;
 664
 665         inode = NULL;
 666         if (req->r_inode) {
 667                 inode = req->r_inode;
 668         } else if (req->r_dentry) {
 669                 /* ignore race with rename; old or new d_parent is okay */
 670                 struct dentry *parent = req->r_dentry->d_parent;
 671                 struct inode *dir = parent->d_inode;
 672
 673                 if (dir->i_sb != mdsc->fsc->sb) {
 674                         /* not this fs! */
 675                         inode = req->r_dentry->d_inode;
 676                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
 677                         /* direct snapped/virtual snapdir requests
 678                          * based on parent dir inode */
 679                         struct dentry *dn = get_nonsnap_parent(parent);
 680                         inode = dn->d_inode;
 681                         dout("__choose_mds using nonsnap parent %p\n", inode);
 682                 } else if (req->r_dentry->d_inode) {
 683                         /* dentry target */
 684                         inode = req->r_dentry->d_inode;
 685                 } else {
 686                         /* dir + name */
 687                         inode = dir;
 688                         hash = ceph_dentry_hash(dir, req->r_dentry);
 689                         is_hash = true;
 690                 }
 691         }
 692
 693         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 694              (int)hash, mode);
 695         if (!inode)
 696                 goto random;
 697         ci = ceph_inode(inode);
 698
 699         if (is_hash && S_ISDIR(inode->i_mode)) {
 700                 struct ceph_inode_frag frag;
 701                 int found;
 702
 703                 ceph_choose_frag(ci, hash, &frag, &found);
 704                 if (found) {
 705                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 706                                 u8 r;
 707
 708                                 /* choose a random replica */
 709                                 get_random_bytes(&r, 1);
 710                                 r %= frag.ndist;
 711                                 mds = frag.dist[r];
 712                                 dout("choose_mds %p %llx.%llx "
 713                                      "frag %u mds%d (%d/%d)\n",
 714                                      inode, ceph_vinop(inode),
 715                                      frag.frag, mds,
 716                                      (int)r, frag.ndist);
 717                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 718                                     CEPH_MDS_STATE_ACTIVE)
 719                                         return mds;
 720                         }
 721
 722                         /* since this file/dir wasn't known to be
 723                          * replicated, then we want to look for the
 724                          * authoritative mds. */
 725                         mode = USE_AUTH_MDS;
 726                         if (frag.mds >= 0) {
 727                                 /* choose auth mds */
 728                                 mds = frag.mds;
 729                                 dout("choose_mds %p %llx.%llx "
 730                                      "frag %u mds%d (auth)\n",
 731                                      inode, ceph_vinop(inode), frag.frag, mds);
 732                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 733                                     CEPH_MDS_STATE_ACTIVE)
 734                                         return mds;
 735                         }
 736                 }
 737         }
 738
 739         spin_lock(&ci->i_ceph_lock);
 740         cap = NULL;
 741         if (mode == USE_AUTH_MDS)
 742                 cap = ci->i_auth_cap;
 743         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
 744                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
 745         if (!cap) {
 746                 spin_unlock(&ci->i_ceph_lock);
 747                 goto random;
 748         }
 749         mds = cap->session->s_mds;
 750         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
 751              inode, ceph_vinop(inode), mds,
 752              cap == ci->i_auth_cap ? "auth " : "", cap);
 753         spin_unlock(&ci->i_ceph_lock);
 754         return mds;
 755
 756 random:
 757         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
 758         dout("choose_mds chose random mds%d\n", mds);
 759         return mds;
 760 }
 761
 762
 763 /*
 764  * session messages
 765  */
 766 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
 767 {
 768         struct ceph_msg *msg;
 769         struct ceph_mds_session_head *h;
 770
 771         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
 772                            false);
 773         if (!msg) {
 774                 pr_err("create_session_msg ENOMEM creating msg\n");
 775                 return NULL;
 776         }
 777         h = msg->front.iov_base;
 778         h->op = cpu_to_le32(op);
 779         h->seq = cpu_to_le64(seq);
 780         return msg;
 781 }
 782
 783 /*
 784  * send session open request.
 785  *
 786  * called under mdsc->mutex
 787  */
 788 static int __open_session(struct ceph_mds_client *mdsc,
 789                           struct ceph_mds_session *session)
 790 {
 791         struct ceph_msg *msg;
 792         int mstate;
 793         int mds = session->s_mds;
 794
 795         /* wait for mds to go active? */
 796         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
 797         dout("open_session to mds%d (%s)\n", mds,
 798              ceph_mds_state_name(mstate));
 799         session->s_state = CEPH_MDS_SESSION_OPENING;
 800         session->s_renew_requested = jiffies;
 801
 802         /* send connect message */
 803         msg = create_session_msg(CEPH_SESSION_REQUEST_OPEN, session->s_seq);
 804         if (!msg)
 805                 return -ENOMEM;
 806         ceph_con_send(&session->s_con, msg);
 807         return 0;
 808 }
 809
 810 /*
 811  * open sessions for any export targets for the given mds
 812  *
 813  * called under mdsc->mutex
 814  */
 815 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
 816                                           struct ceph_mds_session *session)
 817 {
 818         struct ceph_mds_info *mi;
 819         struct ceph_mds_session *ts;
 820         int i, mds = session->s_mds;
 821         int target;
 822
 823         if (mds >= mdsc->mdsmap->m_max_mds)
 824                 return;
 825         mi = &mdsc->mdsmap->m_info[mds];
 826         dout("open_export_target_sessions for mds%d (%d targets)\n",
 827              session->s_mds, mi->num_export_targets);
 828
 829         for (i = 0; i < mi->num_export_targets; i++) {
 830                 target = mi->export_targets[i];
 831                 ts = __ceph_lookup_mds_session(mdsc, target);
 832                 if (!ts) {
 833                         ts = register_session(mdsc, target);
 834                         if (IS_ERR(ts))
 835                                 return;
 836                 }
 837                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
 838                     session->s_state == CEPH_MDS_SESSION_CLOSING)
 839                         __open_session(mdsc, session);
 840                 else
 841                         dout(" mds%d target mds%d %p is %s\n", session->s_mds,
 842                              i, ts, session_state_name(ts->s_state));
 843                 ceph_put_mds_session(ts);
 844         }
 845 }
 846
 847 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
 848                                            struct ceph_mds_session *session)
 849 {
 850         mutex_lock(&mdsc->mutex);
 851         __open_export_target_sessions(mdsc, session);
 852         mutex_unlock(&mdsc->mutex);
 853 }
 854
 855 /*
 856  * session caps
 857  */
 858
 859 /*
 860  * Free preallocated cap messages assigned to this session
 861  */
 862 static void cleanup_cap_releases(struct ceph_mds_session *session)
 863 {
 864         struct ceph_msg *msg;
 865
 866         spin_lock(&session->s_cap_lock);
 867         while (!list_empty(&session->s_cap_releases)) {
 868                 msg = list_first_entry(&session->s_cap_releases,
 869                                        struct ceph_msg, list_head);
 870                 list_del_init(&msg->list_head);
 871                 ceph_msg_put(msg);
 872         }
 873         while (!list_empty(&session->s_cap_releases_done)) {
 874                 msg = list_first_entry(&session->s_cap_releases_done,
 875                                        struct ceph_msg, list_head);
 876                 list_del_init(&msg->list_head);
 877                 ceph_msg_put(msg);
 878         }
 879         spin_unlock(&session->s_cap_lock);
 880 }
 881
 882 /*
 883  * Helper to safely iterate over all caps associated with a session, with
 884  * special care taken to handle a racing __ceph_remove_cap().
 885  *
 886  * Caller must hold session s_mutex.
 887  */
 888 static int iterate_session_caps(struct ceph_mds_session *session,
 889                                  int (*cb)(struct inode *, struct ceph_cap *,
 890                                             void *), void *arg)
 891 {
 892         struct list_head *p;
 893         struct ceph_cap *cap;
 894         struct inode *inode, *last_inode = NULL;
 895         struct ceph_cap *old_cap = NULL;
 896         int ret;
 897
 898         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
 899         spin_lock(&session->s_cap_lock);
 900         p = session->s_caps.next;
 901         while (p != &session->s_caps) {
 902                 cap = list_entry(p, struct ceph_cap, session_caps);
 903                 inode = igrab(&cap->ci->vfs_inode);
 904                 if (!inode) {
 905                         p = p->next;
 906                         continue;
 907                 }
 908                 session->s_cap_iterator = cap;
 909                 spin_unlock(&session->s_cap_lock);
 910
 911                 if (last_inode) {
 912                         iput(last_inode);
 913                         last_inode = NULL;
 914                 }
 915                 if (old_cap) {
 916                         ceph_put_cap(session->s_mdsc, old_cap);
 917                         old_cap = NULL;
 918                 }
 919
 920                 ret = cb(inode, cap, arg);
 921                 last_inode = inode;
 922
 923                 spin_lock(&session->s_cap_lock);
 924                 p = p->next;
 925                 if (cap->ci == NULL) {
 926                         dout("iterate_session_caps  finishing cap %p removal\n",
 927                              cap);
 928                         BUG_ON(cap->session != session);
 929                         list_del_init(&cap->session_caps);
 930                         session->s_nr_caps--;
 931                         cap->session = NULL;
 932                         old_cap = cap;  /* put_cap it w/o locks held */
 933                 }
 934                 if (ret < 0)
 935                         goto out;
 936         }
 937         ret = 0;
 938 out:
 939         session->s_cap_iterator = NULL;
 940         spin_unlock(&session->s_cap_lock);
 941
 942         if (last_inode)
 943                 iput(last_inode);
 944         if (old_cap)
 945                 ceph_put_cap(session->s_mdsc, old_cap);
 946
 947         return ret;
 948 }
 949
 950 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
 951                                   void *arg)
 952 {
 953         struct ceph_inode_info *ci = ceph_inode(inode);
 954         int drop = 0;
 955
 956         dout("removing cap %p, ci is %p, inode is %p\n",
 957              cap, ci, &ci->vfs_inode);
 958         spin_lock(&ci->i_ceph_lock);
 959         __ceph_remove_cap(cap);
 960         if (!__ceph_is_any_real_caps(ci)) {
 961                 struct ceph_mds_client *mdsc =
 962                         ceph_sb_to_client(inode->i_sb)->mdsc;
 963
 964                 spin_lock(&mdsc->cap_dirty_lock);
 965                 if (!list_empty(&ci->i_dirty_item)) {
 966                         pr_info(" dropping dirty %s state for %p %lld\n",
 967                                 ceph_cap_string(ci->i_dirty_caps),
 968                                 inode, ceph_ino(inode));
 969                         ci->i_dirty_caps = 0;
 970                         list_del_init(&ci->i_dirty_item);
 971                         drop = 1;
 972                 }
 973                 if (!list_empty(&ci->i_flushing_item)) {
 974                         pr_info(" dropping dirty+flushing %s state for %p %lld\n",
 975                                 ceph_cap_string(ci->i_flushing_caps),
 976                                 inode, ceph_ino(inode));
 977                         ci->i_flushing_caps = 0;
 978                         list_del_init(&ci->i_flushing_item);
 979                         mdsc->num_cap_flushing--;
 980                         drop = 1;
 981                 }
 982                 if (drop && ci->i_wrbuffer_ref) {
 983                         pr_info(" dropping dirty data for %p %lld\n",
 984                                 inode, ceph_ino(inode));
 985                         ci->i_wrbuffer_ref = 0;
 986                         ci->i_wrbuffer_ref_head = 0;
 987                         drop++;
 988                 }
 989                 spin_unlock(&mdsc->cap_dirty_lock);
 990         }
 991         spin_unlock(&ci->i_ceph_lock);
 992         while (drop--)
 993                 iput(inode);
 994         return 0;
 995 }
 996
 997 /*
 998  * caller must hold session s_mutex
 999  */
1000 static void remove_session_caps(struct ceph_mds_session *session)
1001 {
1002         dout("remove_session_caps on %p\n", session);
1003         iterate_session_caps(session, remove_session_caps_cb, NULL);
1004         BUG_ON(session->s_nr_caps > 0);
1005         BUG_ON(!list_empty(&session->s_cap_flushing));
1006         cleanup_cap_releases(session);
1007 }
1008
1009 /*
1010  * wake up any threads waiting on this session's caps.  if the cap is
1011  * old (didn't get renewed on the client reconnect), remove it now.
1012  *
1013  * caller must hold s_mutex.
1014  */
1015 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1016                               void *arg)
1017 {
1018         struct ceph_inode_info *ci = ceph_inode(inode);
1019
1020         wake_up_all(&ci->i_cap_wq);
1021         if (arg) {
1022                 spin_lock(&ci->i_ceph_lock);
1023                 ci->i_wanted_max_size = 0;
1024                 ci->i_requested_max_size = 0;
1025                 spin_unlock(&ci->i_ceph_lock);
1026         }
1027         return 0;
1028 }
1029
1030 static void wake_up_session_caps(struct ceph_mds_session *session,
1031                                  int reconnect)
1032 {
1033         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1034         iterate_session_caps(session, wake_up_session_cb,
1035                              (void *)(unsigned long)reconnect);
1036 }
1037
1038 /*
1039  * Send periodic message to MDS renewing all currently held caps.  The
1040  * ack will reset the expiration for all caps from this session.
1041  *
1042  * caller holds s_mutex
1043  */
1044 static int send_renew_caps(struct ceph_mds_client *mdsc,
1045                            struct ceph_mds_session *session)
1046 {
1047         struct ceph_msg *msg;
1048         int state;
1049
1050         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1051             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1052                 pr_info("mds%d caps stale\n", session->s_mds);
1053         session->s_renew_requested = jiffies;
1054
1055         /* do not try to renew caps until a recovering mds has reconnected
1056          * with its clients. */
1057         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1058         if (state < CEPH_MDS_STATE_RECONNECT) {
1059                 dout("send_renew_caps ignoring mds%d (%s)\n",
1060                      session->s_mds, ceph_mds_state_name(state));
1061                 return 0;
1062         }
1063
1064         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1065                 ceph_mds_state_name(state));
1066         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1067                                  ++session->s_renew_seq);
1068         if (!msg)
1069                 return -ENOMEM;
1070         ceph_con_send(&session->s_con, msg);
1071         return 0;
1072 }
1073
1074 /*
1075  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1076  *
1077  * Called under session->s_mutex
1078  */
1079 static void renewed_caps(struct ceph_mds_client *mdsc,
1080                          struct ceph_mds_session *session, int is_renew)
1081 {
1082         int was_stale;
1083         int wake = 0;
1084
1085         spin_lock(&session->s_cap_lock);
1086         was_stale = is_renew && (session->s_cap_ttl == 0 ||
1087                                  time_after_eq(jiffies, session->s_cap_ttl));
1088
1089         session->s_cap_ttl = session->s_renew_requested +
1090                 mdsc->mdsmap->m_session_timeout*HZ;
1091
1092         if (was_stale) {
1093                 if (time_before(jiffies, session->s_cap_ttl)) {
1094                         pr_info("mds%d caps renewed\n", session->s_mds);
1095                         wake = 1;
1096                 } else {
1097                         pr_info("mds%d caps still stale\n", session->s_mds);
1098                 }
1099         }
1100         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1101              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1102              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1103         spin_unlock(&session->s_cap_lock);
1104
1105         if (wake)
1106                 wake_up_session_caps(session, 0);
1107 }
1108
1109 /*
1110  * send a session close request
1111  */
1112 static int request_close_session(struct ceph_mds_client *mdsc,
1113                                  struct ceph_mds_session *session)
1114 {
1115         struct ceph_msg *msg;
1116
1117         dout("request_close_session mds%d state %s seq %lld\n",
1118              session->s_mds, session_state_name(session->s_state),
1119              session->s_seq);
1120         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1121         if (!msg)
1122                 return -ENOMEM;
1123         ceph_con_send(&session->s_con, msg);
1124         return 0;
1125 }
1126
1127 /*
1128  * Called with s_mutex held.
1129  */
1130 static int __close_session(struct ceph_mds_client *mdsc,
1131                          struct ceph_mds_session *session)
1132 {
1133         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1134                 return 0;
1135         session->s_state = CEPH_MDS_SESSION_CLOSING;
1136         return request_close_session(mdsc, session);
1137 }
1138
1139 /*
1140  * Trim old(er) caps.
1141  *
1142  * Because we can't cache an inode without one or more caps, we do
1143  * this indirectly: if a cap is unused, we prune its aliases, at which
1144  * point the inode will hopefully get dropped to.
1145  *
1146  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1147  * memory pressure from the MDS, though, so it needn't be perfect.
1148  */
1149 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1150 {
1151         struct ceph_mds_session *session = arg;
1152         struct ceph_inode_info *ci = ceph_inode(inode);
1153         int used, oissued, mine;
1154
1155         if (session->s_trim_caps <= 0)
1156                 return -1;
1157
1158         spin_lock(&ci->i_ceph_lock);
1159         mine = cap->issued | cap->implemented;
1160         used = __ceph_caps_used(ci);
1161         oissued = __ceph_caps_issued_other(ci, cap);
1162
1163         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
1164              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1165              ceph_cap_string(used));
1166         if (ci->i_dirty_caps)
1167                 goto out;   /* dirty caps */
1168         if ((used & ~oissued) & mine)
1169                 goto out;   /* we need these caps */
1170
1171         session->s_trim_caps--;
1172         if (oissued) {
1173                 /* we aren't the only cap.. just remove us */
1174                 __ceph_remove_cap(cap);
1175         } else {
1176                 /* try to drop referring dentries */
1177                 spin_unlock(&ci->i_ceph_lock);
1178                 d_prune_aliases(inode);
1179                 dout("trim_caps_cb %p cap %p  pruned, count now %d\n",
1180                      inode, cap, atomic_read(&inode->i_count));
1181                 return 0;
1182         }
1183
1184 out:
1185         spin_unlock(&ci->i_ceph_lock);
1186         return 0;
1187 }
1188
1189 /*
1190  * Trim session cap count down to some max number.
1191  */
1192 static int trim_caps(struct ceph_mds_client *mdsc,
1193                      struct ceph_mds_session *session,
1194                      int max_caps)
1195 {
1196         int trim_caps = session->s_nr_caps - max_caps;
1197
1198         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1199              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1200         if (trim_caps > 0) {
1201                 session->s_trim_caps = trim_caps;
1202                 iterate_session_caps(session, trim_caps_cb, session);
1203                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1204                      session->s_mds, session->s_nr_caps, max_caps,
1205                         trim_caps - session->s_trim_caps);
1206                 session->s_trim_caps = 0;
1207         }
1208         return 0;
1209 }
1210
1211 /*
1212  * Allocate cap_release messages.  If there is a partially full message
1213  * in the queue, try to allocate enough to cover it's remainder, so that
1214  * we can send it immediately.
1215  *
1216  * Called under s_mutex.
1217  */
1218 int ceph_add_cap_releases(struct ceph_mds_client *mdsc,
1219                           struct ceph_mds_session *session)
1220 {
1221         struct ceph_msg *msg, *partial = NULL;
1222         struct ceph_mds_cap_release *head;
1223         int err = -ENOMEM;
1224         int extra = mdsc->fsc->mount_options->cap_release_safety;
1225         int num;
1226
1227         dout("add_cap_releases %p mds%d extra %d\n", session, session->s_mds,
1228              extra);
1229
1230         spin_lock(&session->s_cap_lock);
1231
1232         if (!list_empty(&session->s_cap_releases)) {
1233                 msg = list_first_entry(&session->s_cap_releases,
1234                                        struct ceph_msg,
1235                                  list_head);
1236                 head = msg->front.iov_base;
1237                 num = le32_to_cpu(head->num);
1238                 if (num) {
1239                         dout(" partial %p with (%d/%d)\n", msg, num,
1240                              (int)CEPH_CAPS_PER_RELEASE);
1241                         extra += CEPH_CAPS_PER_RELEASE - num;
1242                         partial = msg;
1243                 }
1244         }
1245         while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1246                 spin_unlock(&session->s_cap_lock);
1247                 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
1248                                    GFP_NOFS, false);
1249                 if (!msg)
1250                         goto out_unlocked;
1251                 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1252                      (int)msg->front.iov_len);
1253                 head = msg->front.iov_base;
1254                 head->num = cpu_to_le32(0);
1255                 msg->front.iov_len = sizeof(*head);
1256                 spin_lock(&session->s_cap_lock);
1257                 list_add(&msg->list_head, &session->s_cap_releases);
1258                 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1259         }
1260
1261         if (partial) {
1262                 head = partial->front.iov_base;
1263                 num = le32_to_cpu(head->num);
1264                 dout(" queueing partial %p with %d/%d\n", partial, num,
1265                      (int)CEPH_CAPS_PER_RELEASE);
1266                 list_move_tail(&partial->list_head,
1267                                &session->s_cap_releases_done);
1268                 session->s_num_cap_releases -= CEPH_CAPS_PER_RELEASE - num;
1269         }
1270         err = 0;
1271         spin_unlock(&session->s_cap_lock);
1272 out_unlocked:
1273         return err;
1274 }
1275
1276 /*
1277  * flush all dirty inode data to disk.
1278  *
1279  * returns true if we've flushed through want_flush_seq
1280  */
1281 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1282 {
1283         int mds, ret = 1;
1284
1285         dout("check_cap_flush want %lld\n", want_flush_seq);
1286         mutex_lock(&mdsc->mutex);
1287         for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1288                 struct ceph_mds_session *session = mdsc->sessions[mds];
1289
1290                 if (!session)
1291                         continue;
1292                 get_session(session);
1293                 mutex_unlock(&mdsc->mutex);
1294
1295                 mutex_lock(&session->s_mutex);
1296                 if (!list_empty(&session->s_cap_flushing)) {
1297                         struct ceph_inode_info *ci =
1298                                 list_entry(session->s_cap_flushing.next,
1299                                            struct ceph_inode_info,
1300                                            i_flushing_item);
1301                         struct inode *inode = &ci->vfs_inode;
1302
1303                         spin_lock(&ci->i_ceph_lock);
1304                         if (ci->i_cap_flush_seq <= want_flush_seq) {
1305                                 dout("check_cap_flush still flushing %p "
1306                                      "seq %lld <= %lld to mds%d\n", inode,
1307                                      ci->i_cap_flush_seq, want_flush_seq,
1308                                      session->s_mds);
1309                                 ret = 0;
1310                         }
1311                         spin_unlock(&ci->i_ceph_lock);
1312                 }
1313                 mutex_unlock(&session->s_mutex);
1314                 ceph_put_mds_session(session);
1315
1316                 if (!ret)
1317                         return ret;
1318                 mutex_lock(&mdsc->mutex);
1319         }
1320
1321         mutex_unlock(&mdsc->mutex);
1322         dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1323         return ret;
1324 }
1325
1326 /*
1327  * called under s_mutex
1328  */
1329 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1330                             struct ceph_mds_session *session)
1331 {
1332         struct ceph_msg *msg;
1333
1334         dout("send_cap_releases mds%d\n", session->s_mds);
1335         spin_lock(&session->s_cap_lock);
1336         while (!list_empty(&session->s_cap_releases_done)) {
1337                 msg = list_first_entry(&session->s_cap_releases_done,
1338                                  struct ceph_msg, list_head);
1339                 list_del_init(&msg->list_head);
1340                 spin_unlock(&session->s_cap_lock);
1341                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1342                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1343                 ceph_con_send(&session->s_con, msg);
1344                 spin_lock(&session->s_cap_lock);
1345         }
1346         spin_unlock(&session->s_cap_lock);
1347 }
1348
1349 static void discard_cap_releases(struct ceph_mds_client *mdsc,
1350                                  struct ceph_mds_session *session)
1351 {
1352         struct ceph_msg *msg;
1353         struct ceph_mds_cap_release *head;
1354         unsigned num;
1355
1356         dout("discard_cap_releases mds%d\n", session->s_mds);
1357         spin_lock(&session->s_cap_lock);
1358
1359         /* zero out the in-progress message */
1360         msg = list_first_entry(&session->s_cap_releases,
1361                                struct ceph_msg, list_head);
1362         head = msg->front.iov_base;
1363         num = le32_to_cpu(head->num);
1364         dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg, num);
1365         head->num = cpu_to_le32(0);
1366         session->s_num_cap_releases += num;
1367
1368         /* requeue completed messages */
1369         while (!list_empty(&session->s_cap_releases_done)) {
1370                 msg = list_first_entry(&session->s_cap_releases_done,
1371                                  struct ceph_msg, list_head);
1372                 list_del_init(&msg->list_head);
1373
1374                 head = msg->front.iov_base;
1375                 num = le32_to_cpu(head->num);
1376                 dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg,
1377                      num);
1378                 session->s_num_cap_releases += num;
1379                 head->num = cpu_to_le32(0);
1380                 msg->front.iov_len = sizeof(*head);
1381                 list_add(&msg->list_head, &session->s_cap_releases);
1382         }
1383
1384         spin_unlock(&session->s_cap_lock);
1385 }
1386
1387 /*
1388  * requests
1389  */
1390
1391 /*
1392  * Create an mds request.
1393  */
1394 struct ceph_mds_request *
1395 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1396 {
1397         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1398
1399         if (!req)
1400                 return ERR_PTR(-ENOMEM);
1401
1402         mutex_init(&req->r_fill_mutex);
1403         req->r_mdsc = mdsc;
1404         req->r_started = jiffies;
1405         req->r_resend_mds = -1;
1406         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1407         req->r_fmode = -1;
1408         kref_init(&req->r_kref);
1409         INIT_LIST_HEAD(&req->r_wait);
1410         init_completion(&req->r_completion);
1411         init_completion(&req->r_safe_completion);
1412         INIT_LIST_HEAD(&req->r_unsafe_item);
1413
1414         req->r_op = op;
1415         req->r_direct_mode = mode;
1416         return req;
1417 }
1418
1419 /*
1420  * return oldest (lowest) request, tid in request tree, 0 if none.
1421  *
1422  * called under mdsc->mutex.
1423  */
1424 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1425 {
1426         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1427                 return NULL;
1428         return rb_entry(rb_first(&mdsc->request_tree),
1429                         struct ceph_mds_request, r_node);
1430 }
1431
1432 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1433 {
1434         struct ceph_mds_request *req = __get_oldest_req(mdsc);
1435
1436         if (req)
1437                 return req->r_tid;
1438         return 0;
1439 }
1440
1441 /*
1442  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1443  * on build_path_from_dentry in fs/cifs/dir.c.
1444  *
1445  * If @stop_on_nosnap, generate path relative to the first non-snapped
1446  * inode.
1447  *
1448  * Encode hidden .snap dirs as a double /, i.e.
1449  *   foo/.snap/bar -> foo//bar
1450  */
1451 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1452                            int stop_on_nosnap)
1453 {
1454         struct dentry *temp;
1455         char *path;
1456         int len, pos;
1457         unsigned seq;
1458
1459         if (dentry == NULL)
1460                 return ERR_PTR(-EINVAL);
1461
1462 retry:
1463         len = 0;
1464         seq = read_seqbegin(&rename_lock);
1465         rcu_read_lock();
1466         for (temp = dentry; !IS_ROOT(temp);) {
1467                 struct inode *inode = temp->d_inode;
1468                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1469                         len++;  /* slash only */
1470                 else if (stop_on_nosnap && inode &&
1471                          ceph_snap(inode) == CEPH_NOSNAP)
1472                         break;
1473                 else
1474                         len += 1 + temp->d_name.len;
1475                 temp = temp->d_parent;
1476                 if (temp == NULL) {
1477                         rcu_read_unlock();
1478                         pr_err("build_path corrupt dentry %p\n", dentry);
1479                         return ERR_PTR(-EINVAL);
1480                 }
1481         }
1482         rcu_read_unlock();
1483         if (len)
1484                 len--;  /* no leading '/' */
1485
1486         path = kmalloc(len+1, GFP_NOFS);
1487         if (path == NULL)
1488                 return ERR_PTR(-ENOMEM);
1489         pos = len;
1490         path[pos] = 0;  /* trailing null */
1491         rcu_read_lock();
1492         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1493                 struct inode *inode;
1494
1495                 spin_lock(&temp->d_lock);
1496                 inode = temp->d_inode;
1497                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1498                         dout("build_path path+%d: %p SNAPDIR\n",
1499                              pos, temp);
1500                 } else if (stop_on_nosnap && inode &&
1501                            ceph_snap(inode) == CEPH_NOSNAP) {
1502                         spin_unlock(&temp->d_lock);
1503                         break;
1504                 } else {
1505                         pos -= temp->d_name.len;
1506                         if (pos < 0) {
1507                                 spin_unlock(&temp->d_lock);
1508                                 break;
1509                         }
1510                         strncpy(path + pos, temp->d_name.name,
1511                                 temp->d_name.len);
1512                 }
1513                 spin_unlock(&temp->d_lock);
1514                 if (pos)
1515                         path[--pos] = '/';
1516                 temp = temp->d_parent;
1517                 if (temp == NULL) {
1518                         rcu_read_unlock();
1519                         pr_err("build_path corrupt dentry\n");
1520                         kfree(path);
1521                         return ERR_PTR(-EINVAL);
1522                 }
1523         }
1524         rcu_read_unlock();
1525         if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1526                 pr_err("build_path did not end path lookup where "
1527                        "expected, namelen is %d, pos is %d\n", len, pos);
1528                 /* presumably this is only possible if racing with a
1529                    rename of one of the parent directories (we can not
1530                    lock the dentries above us to prevent this, but
1531                    retrying should be harmless) */
1532                 kfree(path);
1533                 goto retry;
1534         }
1535
1536         *base = ceph_ino(temp->d_inode);
1537         *plen = len;
1538         dout("build_path on %p %d built %llx '%.*s'\n",
1539              dentry, dentry->d_count, *base, len, path);
1540         return path;
1541 }
1542
1543 static int build_dentry_path(struct dentry *dentry,
1544                              const char **ppath, int *ppathlen, u64 *pino,
1545                              int *pfreepath)
1546 {
1547         char *path;
1548
1549         if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1550                 *pino = ceph_ino(dentry->d_parent->d_inode);
1551                 *ppath = dentry->d_name.name;
1552                 *ppathlen = dentry->d_name.len;
1553                 return 0;
1554         }
1555         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1556         if (IS_ERR(path))
1557                 return PTR_ERR(path);
1558         *ppath = path;
1559         *pfreepath = 1;
1560         return 0;
1561 }
1562
1563 static int build_inode_path(struct inode *inode,
1564                             const char **ppath, int *ppathlen, u64 *pino,
1565                             int *pfreepath)
1566 {
1567         struct dentry *dentry;
1568         char *path;
1569
1570         if (ceph_snap(inode) == CEPH_NOSNAP) {
1571                 *pino = ceph_ino(inode);
1572                 *ppathlen = 0;
1573                 return 0;
1574         }
1575         dentry = d_find_alias(inode);
1576         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1577         dput(dentry);
1578         if (IS_ERR(path))
1579                 return PTR_ERR(path);
1580         *ppath = path;
1581         *pfreepath = 1;
1582         return 0;
1583 }
1584
1585 /*
1586  * request arguments may be specified via an inode *, a dentry *, or
1587  * an explicit ino+path.
1588  */
1589 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1590                                   const char *rpath, u64 rino,
1591                                   const char **ppath, int *pathlen,
1592                                   u64 *ino, int *freepath)
1593 {
1594         int r = 0;
1595
1596         if (rinode) {
1597                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1598                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1599                      ceph_snap(rinode));
1600         } else if (rdentry) {
1601                 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1602                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1603                      *ppath);
1604         } else if (rpath || rino) {
1605                 *ino = rino;
1606                 *ppath = rpath;
1607                 *pathlen = strlen(rpath);
1608                 dout(" path %.*s\n", *pathlen, rpath);
1609         }
1610
1611         return r;
1612 }
1613
1614 /*
1615  * called under mdsc->mutex
1616  */
1617 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1618                                                struct ceph_mds_request *req,
1619                                                int mds)
1620 {
1621         struct ceph_msg *msg;
1622         struct ceph_mds_request_head *head;
1623         const char *path1 = NULL;
1624         const char *path2 = NULL;
1625         u64 ino1 = 0, ino2 = 0;
1626         int pathlen1 = 0, pathlen2 = 0;
1627         int freepath1 = 0, freepath2 = 0;
1628         int len;
1629         u16 releases;
1630         void *p, *end;
1631         int ret;
1632
1633         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1634                               req->r_path1, req->r_ino1.ino,
1635                               &path1, &pathlen1, &ino1, &freepath1);
1636         if (ret < 0) {
1637                 msg = ERR_PTR(ret);
1638                 goto out;
1639         }
1640
1641         ret = set_request_path_attr(NULL, req->r_old_dentry,
1642                               req->r_path2, req->r_ino2.ino,
1643                               &path2, &pathlen2, &ino2, &freepath2);
1644         if (ret < 0) {
1645                 msg = ERR_PTR(ret);
1646                 goto out_free1;
1647         }
1648
1649         len = sizeof(*head) +
1650                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64));
1651
1652         /* calculate (max) length for cap releases */
1653         len += sizeof(struct ceph_mds_request_release) *
1654                 (!!req->r_inode_drop + !!req->r_dentry_drop +
1655                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1656         if (req->r_dentry_drop)
1657                 len += req->r_dentry->d_name.len;
1658         if (req->r_old_dentry_drop)
1659                 len += req->r_old_dentry->d_name.len;
1660
1661         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1662         if (!msg) {
1663                 msg = ERR_PTR(-ENOMEM);
1664                 goto out_free2;
1665         }
1666
1667         msg->hdr.tid = cpu_to_le64(req->r_tid);
1668
1669         head = msg->front.iov_base;
1670         p = msg->front.iov_base + sizeof(*head);
1671         end = msg->front.iov_base + msg->front.iov_len;
1672
1673         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1674         head->op = cpu_to_le32(req->r_op);
1675         head->caller_uid = cpu_to_le32(req->r_uid);
1676         head->caller_gid = cpu_to_le32(req->r_gid);
1677         head->args = req->r_args;
1678
1679         ceph_encode_filepath(&p, end, ino1, path1);
1680         ceph_encode_filepath(&p, end, ino2, path2);
1681
1682         /* make note of release offset, in case we need to replay */
1683         req->r_request_release_offset = p - msg->front.iov_base;
1684
1685         /* cap releases */
1686         releases = 0;
1687         if (req->r_inode_drop)
1688                 releases += ceph_encode_inode_release(&p,
1689                       req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1690                       mds, req->r_inode_drop, req->r_inode_unless, 0);
1691         if (req->r_dentry_drop)
1692                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1693                        mds, req->r_dentry_drop, req->r_dentry_unless);
1694         if (req->r_old_dentry_drop)
1695                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1696                        mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1697         if (req->r_old_inode_drop)
1698                 releases += ceph_encode_inode_release(&p,
1699                       req->r_old_dentry->d_inode,
1700                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1701         head->num_releases = cpu_to_le16(releases);
1702
1703         BUG_ON(p > end);
1704         msg->front.iov_len = p - msg->front.iov_base;
1705         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1706
1707         msg->pages = req->r_pages;
1708         msg->nr_pages = req->r_num_pages;
1709         msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1710         msg->hdr.data_off = cpu_to_le16(0);
1711
1712 out_free2:
1713         if (freepath2)
1714                 kfree((char *)path2);
1715 out_free1:
1716         if (freepath1)
1717                 kfree((char *)path1);
1718 out:
1719         return msg;
1720 }
1721
1722 /*
1723  * called under mdsc->mutex if error, under no mutex if
1724  * success.
1725  */
1726 static void complete_request(struct ceph_mds_client *mdsc,
1727                              struct ceph_mds_request *req)
1728 {
1729         if (req->r_callback)
1730                 req->r_callback(mdsc, req);
1731         else
1732                 complete_all(&req->r_completion);
1733 }
1734
1735 /*
1736  * called under mdsc->mutex
1737  */
1738 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1739                                   struct ceph_mds_request *req,
1740                                   int mds)
1741 {
1742         struct ceph_mds_request_head *rhead;
1743         struct ceph_msg *msg;
1744         int flags = 0;
1745
1746         req->r_attempts++;
1747         if (req->r_inode) {
1748                 struct ceph_cap *cap =
1749                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
1750
1751                 if (cap)
1752                         req->r_sent_on_mseq = cap->mseq;
1753                 else
1754                         req->r_sent_on_mseq = -1;
1755         }
1756         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1757              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1758
1759         if (req->r_got_unsafe) {
1760                 /*
1761                  * Replay.  Do not regenerate message (and rebuild
1762                  * paths, etc.); just use the original message.
1763                  * Rebuilding paths will break for renames because
1764                  * d_move mangles the src name.
1765                  */
1766                 msg = req->r_request;
1767                 rhead = msg->front.iov_base;
1768
1769                 flags = le32_to_cpu(rhead->flags);
1770                 flags |= CEPH_MDS_FLAG_REPLAY;
1771                 rhead->flags = cpu_to_le32(flags);
1772
1773                 if (req->r_target_inode)
1774                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1775
1776                 rhead->num_retry = req->r_attempts - 1;
1777
1778                 /* remove cap/dentry releases from message */
1779                 rhead->num_releases = 0;
1780                 msg->hdr.front_len = cpu_to_le32(req->r_request_release_offset);
1781                 msg->front.iov_len = req->r_request_release_offset;
1782                 return 0;
1783         }
1784
1785         if (req->r_request) {
1786                 ceph_msg_put(req->r_request);
1787                 req->r_request = NULL;
1788         }
1789         msg = create_request_message(mdsc, req, mds);
1790         if (IS_ERR(msg)) {
1791                 req->r_err = PTR_ERR(msg);
1792                 complete_request(mdsc, req);
1793                 return PTR_ERR(msg);
1794         }
1795         req->r_request = msg;
1796
1797         rhead = msg->front.iov_base;
1798         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1799         if (req->r_got_unsafe)
1800                 flags |= CEPH_MDS_FLAG_REPLAY;
1801         if (req->r_locked_dir)
1802                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1803         rhead->flags = cpu_to_le32(flags);
1804         rhead->num_fwd = req->r_num_fwd;
1805         rhead->num_retry = req->r_attempts - 1;
1806         rhead->ino = 0;
1807
1808         dout(" r_locked_dir = %p\n", req->r_locked_dir);
1809         return 0;
1810 }
1811
1812 /*
1813  * send request, or put it on the appropriate wait list.
1814  */
1815 static int __do_request(struct ceph_mds_client *mdsc,
1816                         struct ceph_mds_request *req)
1817 {
1818         struct ceph_mds_session *session = NULL;
1819         int mds = -1;
1820         int err = -EAGAIN;
1821
1822         if (req->r_err || req->r_got_result) {
1823                 if (req->r_aborted)
1824                         __unregister_request(mdsc, req);
1825                 goto out;
1826         }
1827
1828         if (req->r_timeout &&
1829             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1830                 dout("do_request timed out\n");
1831                 err = -EIO;
1832                 goto finish;
1833         }
1834
1835         put_request_session(req);
1836
1837         mds = __choose_mds(mdsc, req);
1838         if (mds < 0 ||
1839             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1840                 dout("do_request no mds or not active, waiting for map\n");
1841                 list_add(&req->r_wait, &mdsc->waiting_for_map);
1842                 goto out;
1843         }
1844
1845         /* get, open session */
1846         session = __ceph_lookup_mds_session(mdsc, mds);
1847         if (!session) {
1848                 session = register_session(mdsc, mds);
1849                 if (IS_ERR(session)) {
1850                         err = PTR_ERR(session);
1851                         goto finish;
1852                 }
1853         }
1854         req->r_session = get_session(session);
1855
1856         dout("do_request mds%d session %p state %s\n", mds, session,
1857              session_state_name(session->s_state));
1858         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1859             session->s_state != CEPH_MDS_SESSION_HUNG) {
1860                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1861                     session->s_state == CEPH_MDS_SESSION_CLOSING)
1862                         __open_session(mdsc, session);
1863                 list_add(&req->r_wait, &session->s_waiting);
1864                 goto out_session;
1865         }
1866
1867         /* send request */
1868         req->r_resend_mds = -1;   /* forget any previous mds hint */
1869
1870         if (req->r_request_started == 0)   /* note request start time */
1871                 req->r_request_started = jiffies;
1872
1873         err = __prepare_send_request(mdsc, req, mds);
1874         if (!err) {
1875                 ceph_msg_get(req->r_request);
1876                 ceph_con_send(&session->s_con, req->r_request);
1877         }
1878
1879 out_session:
1880         ceph_put_mds_session(session);
1881 out:
1882         return err;
1883
1884 finish:
1885         req->r_err = err;
1886         complete_request(mdsc, req);
1887         goto out;
1888 }
1889
1890 /*
1891  * called under mdsc->mutex
1892  */
1893 static void __wake_requests(struct ceph_mds_client *mdsc,
1894                             struct list_head *head)
1895 {
1896         struct ceph_mds_request *req, *nreq;
1897
1898         list_for_each_entry_safe(req, nreq, head, r_wait) {
1899                 list_del_init(&req->r_wait);
1900                 __do_request(mdsc, req);
1901         }
1902 }
1903
1904 /*
1905  * Wake up threads with requests pending for @mds, so that they can
1906  * resubmit their requests to a possibly different mds.
1907  */
1908 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
1909 {
1910         struct ceph_mds_request *req;
1911         struct rb_node *p;
1912
1913         dout("kick_requests mds%d\n", mds);
1914         for (p = rb_first(&mdsc->request_tree); p; p = rb_next(p)) {
1915                 req = rb_entry(p, struct ceph_mds_request, r_node);
1916                 if (req->r_got_unsafe)
1917                         continue;
1918                 if (req->r_session &&
1919                     req->r_session->s_mds == mds) {
1920                         dout(" kicking tid %llu\n", req->r_tid);
1921                         __do_request(mdsc, req);
1922                 }
1923         }
1924 }
1925
1926 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
1927                               struct ceph_mds_request *req)
1928 {
1929         dout("submit_request on %p\n", req);
1930         mutex_lock(&mdsc->mutex);
1931         __register_request(mdsc, req, NULL);
1932         __do_request(mdsc, req);
1933         mutex_unlock(&mdsc->mutex);
1934 }
1935
1936 /*
1937  * Synchrously perform an mds request.  Take care of all of the
1938  * session setup, forwarding, retry details.
1939  */
1940 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
1941                          struct inode *dir,
1942                          struct ceph_mds_request *req)
1943 {
1944         int err;
1945
1946         dout("do_request on %p\n", req);
1947
1948         /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
1949         if (req->r_inode)
1950                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
1951         if (req->r_locked_dir)
1952                 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
1953         if (req->r_old_dentry)
1954                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
1955                                   CEPH_CAP_PIN);
1956
1957         /* issue */
1958         mutex_lock(&mdsc->mutex);
1959         __register_request(mdsc, req, dir);
1960         __do_request(mdsc, req);
1961
1962         if (req->r_err) {
1963                 err = req->r_err;
1964                 __unregister_request(mdsc, req);
1965                 dout("do_request early error %d\n", err);
1966                 goto out;
1967         }
1968
1969         /* wait */
1970         mutex_unlock(&mdsc->mutex);
1971         dout("do_request waiting\n");
1972         if (req->r_timeout) {
1973                 err = (long)wait_for_completion_killable_timeout(
1974                         &req->r_completion, req->r_timeout);
1975                 if (err == 0)
1976                         err = -EIO;
1977         } else {
1978                 err = wait_for_completion_killable(&req->r_completion);
1979         }
1980         dout("do_request waited, got %d\n", err);
1981         mutex_lock(&mdsc->mutex);
1982
1983         /* only abort if we didn't race with a real reply */
1984         if (req->r_got_result) {
1985                 err = le32_to_cpu(req->r_reply_info.head->result);
1986         } else if (err < 0) {
1987                 dout("aborted request %lld with %d\n", req->r_tid, err);
1988
1989                 /*
1990                  * ensure we aren't running concurrently with
1991                  * ceph_fill_trace or ceph_readdir_prepopulate, which
1992                  * rely on locks (dir mutex) held by our caller.
1993                  */
1994                 mutex_lock(&req->r_fill_mutex);
1995                 req->r_err = err;
1996                 req->r_aborted = true;
1997                 mutex_unlock(&req->r_fill_mutex);
1998
1999                 if (req->r_locked_dir &&
2000                     (req->r_op & CEPH_MDS_OP_WRITE))
2001                         ceph_invalidate_dir_request(req);
2002         } else {
2003                 err = req->r_err;
2004         }
2005
2006 out:
2007         mutex_unlock(&mdsc->mutex);
2008         dout("do_request %p done, result %d\n", req, err);
2009         return err;
2010 }
2011
2012 /*
2013  * Invalidate dir D_COMPLETE, dentry lease state on an aborted MDS
2014  * namespace request.
2015  */
2016 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2017 {
2018         struct inode *inode = req->r_locked_dir;
2019         struct ceph_inode_info *ci = ceph_inode(inode);
2020
2021         dout("invalidate_dir_request %p (D_COMPLETE, lease(s))\n", inode);
2022         spin_lock(&ci->i_ceph_lock);
2023         ceph_dir_clear_complete(inode);
2024         ci->i_release_count++;
2025         spin_unlock(&ci->i_ceph_lock);
2026
2027         if (req->r_dentry)
2028                 ceph_invalidate_dentry_lease(req->r_dentry);
2029         if (req->r_old_dentry)
2030                 ceph_invalidate_dentry_lease(req->r_old_dentry);
2031 }
2032
2033 /*
2034  * Handle mds reply.
2035  *
2036  * We take the session mutex and parse and process the reply immediately.
2037  * This preserves the logical ordering of replies, capabilities, etc., sent
2038  * by the MDS as they are applied to our local cache.
2039  */
2040 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2041 {
2042         struct ceph_mds_client *mdsc = session->s_mdsc;
2043         struct ceph_mds_request *req;
2044         struct ceph_mds_reply_head *head = msg->front.iov_base;
2045         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2046         u64 tid;
2047         int err, result;
2048         int mds = session->s_mds;
2049
2050         if (msg->front.iov_len < sizeof(*head)) {
2051                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2052                 ceph_msg_dump(msg);
2053                 return;
2054         }
2055
2056         /* get request, session */
2057         tid = le64_to_cpu(msg->hdr.tid);
2058         mutex_lock(&mdsc->mutex);
2059         req = __lookup_request(mdsc, tid);
2060         if (!req) {
2061                 dout("handle_reply on unknown tid %llu\n", tid);
2062                 mutex_unlock(&mdsc->mutex);
2063                 return;
2064         }
2065         dout("handle_reply %p\n", req);
2066
2067         /* correct session? */
2068         if (req->r_session != session) {
2069                 pr_err("mdsc_handle_reply got %llu on session mds%d"
2070                        " not mds%d\n", tid, session->s_mds,
2071                        req->r_session ? req->r_session->s_mds : -1);
2072                 mutex_unlock(&mdsc->mutex);
2073                 goto out;
2074         }
2075
2076         /* dup? */
2077         if ((req->r_got_unsafe && !head->safe) ||
2078             (req->r_got_safe && head->safe)) {
2079                 pr_warning("got a dup %s reply on %llu from mds%d\n",
2080                            head->safe ? "safe" : "unsafe", tid, mds);
2081                 mutex_unlock(&mdsc->mutex);
2082                 goto out;
2083         }
2084         if (req->r_got_safe && !head->safe) {
2085                 pr_warning("got unsafe after safe on %llu from mds%d\n",
2086                            tid, mds);
2087                 mutex_unlock(&mdsc->mutex);
2088                 goto out;
2089         }
2090
2091         result = le32_to_cpu(head->result);
2092
2093         /*
2094          * Handle an ESTALE
2095          * if we're not talking to the authority, send to them
2096          * if the authority has changed while we weren't looking,
2097          * send to new authority
2098          * Otherwise we just have to return an ESTALE
2099          */
2100         if (result == -ESTALE) {
2101                 dout("got ESTALE on request %llu", req->r_tid);
2102                 if (!req->r_inode) {
2103                         /* do nothing; not an authority problem */
2104                 } else if (req->r_direct_mode != USE_AUTH_MDS) {
2105                         dout("not using auth, setting for that now");
2106                         req->r_direct_mode = USE_AUTH_MDS;
2107                         __do_request(mdsc, req);
2108                         mutex_unlock(&mdsc->mutex);
2109                         goto out;
2110                 } else  {
2111                         struct ceph_inode_info *ci = ceph_inode(req->r_inode);
2112                         struct ceph_cap *cap = NULL;
2113
2114                         if (req->r_session)
2115                                 cap = ceph_get_cap_for_mds(ci,
2116                                                    req->r_session->s_mds);
2117
2118                         dout("already using auth");
2119                         if ((!cap || cap != ci->i_auth_cap) ||
2120                             (cap->mseq != req->r_sent_on_mseq)) {
2121                                 dout("but cap changed, so resending");
2122                                 __do_request(mdsc, req);
2123                                 mutex_unlock(&mdsc->mutex);
2124                                 goto out;
2125                         }
2126                 }
2127                 dout("have to return ESTALE on request %llu", req->r_tid);
2128         }
2129
2130
2131         if (head->safe) {
2132                 req->r_got_safe = true;
2133                 __unregister_request(mdsc, req);
2134
2135                 if (req->r_got_unsafe) {
2136                         /*
2137                          * We already handled the unsafe response, now do the
2138                          * cleanup.  No need to examine the response; the MDS
2139                          * doesn't include any result info in the safe
2140                          * response.  And even if it did, there is nothing
2141                          * useful we could do with a revised return value.
2142                          */
2143                         dout("got safe reply %llu, mds%d\n", tid, mds);
2144                         list_del_init(&req->r_unsafe_item);
2145
2146                         /* last unsafe request during umount? */
2147                         if (mdsc->stopping && !__get_oldest_req(mdsc))
2148                                 complete_all(&mdsc->safe_umount_waiters);
2149                         mutex_unlock(&mdsc->mutex);
2150                         goto out;
2151                 }
2152         } else {
2153                 req->r_got_unsafe = true;
2154                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2155         }
2156
2157         dout("handle_reply tid %lld result %d\n", tid, result);
2158         rinfo = &req->r_reply_info;
2159         err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2160         mutex_unlock(&mdsc->mutex);
2161
2162         mutex_lock(&session->s_mutex);
2163         if (err < 0) {
2164                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2165                 ceph_msg_dump(msg);
2166                 goto out_err;
2167         }
2168
2169         /* snap trace */
2170         if (rinfo->snapblob_len) {
2171                 down_write(&mdsc->snap_rwsem);
2172                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2173                                rinfo->snapblob + rinfo->snapblob_len,
2174                                le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
2175                 downgrade_write(&mdsc->snap_rwsem);
2176         } else {
2177                 down_read(&mdsc->snap_rwsem);
2178         }
2179
2180         /* insert trace into our cache */
2181         mutex_lock(&req->r_fill_mutex);
2182         err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2183         if (err == 0) {
2184                 if (result == 0 && req->r_op != CEPH_MDS_OP_GETFILELOCK &&
2185                     rinfo->dir_nr)
2186                         ceph_readdir_prepopulate(req, req->r_session);
2187                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2188         }
2189         mutex_unlock(&req->r_fill_mutex);
2190
2191         up_read(&mdsc->snap_rwsem);
2192 out_err:
2193         mutex_lock(&mdsc->mutex);
2194         if (!req->r_aborted) {
2195                 if (err) {
2196                         req->r_err = err;
2197                 } else {
2198                         req->r_reply = msg;
2199                         ceph_msg_get(msg);
2200                         req->r_got_result = true;
2201                 }
2202         } else {
2203                 dout("reply arrived after request %lld was aborted\n", tid);
2204         }
2205         mutex_unlock(&mdsc->mutex);
2206
2207         ceph_add_cap_releases(mdsc, req->r_session);
2208         mutex_unlock(&session->s_mutex);
2209
2210         /* kick calling process */
2211         complete_request(mdsc, req);
2212 out:
2213         ceph_mdsc_put_request(req);
2214         return;
2215 }
2216
2217
2218
2219 /*
2220  * handle mds notification that our request has been forwarded.
2221  */
2222 static void handle_forward(struct ceph_mds_client *mdsc,
2223                            struct ceph_mds_session *session,
2224                            struct ceph_msg *msg)
2225 {
2226         struct ceph_mds_request *req;
2227         u64 tid = le64_to_cpu(msg->hdr.tid);
2228         u32 next_mds;
2229         u32 fwd_seq;
2230         int err = -EINVAL;
2231         void *p = msg->front.iov_base;
2232         void *end = p + msg->front.iov_len;
2233
2234         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2235         next_mds = ceph_decode_32(&p);
2236         fwd_seq = ceph_decode_32(&p);
2237
2238         mutex_lock(&mdsc->mutex);
2239         req = __lookup_request(mdsc, tid);
2240         if (!req) {
2241                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2242                 goto out;  /* dup reply? */
2243         }
2244
2245         if (req->r_aborted) {
2246                 dout("forward tid %llu aborted, unregistering\n", tid);
2247                 __unregister_request(mdsc, req);
2248         } else if (fwd_seq <= req->r_num_fwd) {
2249                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2250                      tid, next_mds, req->r_num_fwd, fwd_seq);
2251         } else {
2252                 /* resend. forward race not possible; mds would drop */
2253                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2254                 BUG_ON(req->r_err);
2255                 BUG_ON(req->r_got_result);
2256                 req->r_num_fwd = fwd_seq;
2257                 req->r_resend_mds = next_mds;
2258                 put_request_session(req);
2259                 __do_request(mdsc, req);
2260         }
2261         ceph_mdsc_put_request(req);
2262 out:
2263         mutex_unlock(&mdsc->mutex);
2264         return;
2265
2266 bad:
2267         pr_err("mdsc_handle_forward decode error err=%d\n", err);
2268 }
2269
2270 /*
2271  * handle a mds session control message
2272  */
2273 static void handle_session(struct ceph_mds_session *session,
2274                            struct ceph_msg *msg)
2275 {
2276         struct ceph_mds_client *mdsc = session->s_mdsc;
2277         u32 op;
2278         u64 seq;
2279         int mds = session->s_mds;
2280         struct ceph_mds_session_head *h = msg->front.iov_base;
2281         int wake = 0;
2282
2283         /* decode */
2284         if (msg->front.iov_len != sizeof(*h))
2285                 goto bad;
2286         op = le32_to_cpu(h->op);
2287         seq = le64_to_cpu(h->seq);
2288
2289         mutex_lock(&mdsc->mutex);
2290         if (op == CEPH_SESSION_CLOSE)
2291                 __unregister_session(mdsc, session);
2292         /* FIXME: this ttl calculation is generous */
2293         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2294         mutex_unlock(&mdsc->mutex);
2295
2296         mutex_lock(&session->s_mutex);
2297
2298         dout("handle_session mds%d %s %p state %s seq %llu\n",
2299              mds, ceph_session_op_name(op), session,
2300              session_state_name(session->s_state), seq);
2301
2302         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2303                 session->s_state = CEPH_MDS_SESSION_OPEN;
2304                 pr_info("mds%d came back\n", session->s_mds);
2305         }
2306
2307         switch (op) {
2308         case CEPH_SESSION_OPEN:
2309                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2310                         pr_info("mds%d reconnect success\n", session->s_mds);
2311                 session->s_state = CEPH_MDS_SESSION_OPEN;
2312                 renewed_caps(mdsc, session, 0);
2313                 wake = 1;
2314                 if (mdsc->stopping)
2315                         __close_session(mdsc, session);
2316                 break;
2317
2318         case CEPH_SESSION_RENEWCAPS:
2319                 if (session->s_renew_seq == seq)
2320                         renewed_caps(mdsc, session, 1);
2321                 break;
2322
2323         case CEPH_SESSION_CLOSE:
2324                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2325                         pr_info("mds%d reconnect denied\n", session->s_mds);
2326                 remove_session_caps(session);
2327                 wake = 1; /* for good measure */
2328                 wake_up_all(&mdsc->session_close_wq);
2329                 kick_requests(mdsc, mds);
2330                 break;
2331
2332         case CEPH_SESSION_STALE:
2333                 pr_info("mds%d caps went stale, renewing\n",
2334                         session->s_mds);
2335                 spin_lock(&session->s_cap_lock);
2336                 session->s_cap_gen++;
2337                 session->s_cap_ttl = 0;
2338                 spin_unlock(&session->s_cap_lock);
2339                 send_renew_caps(mdsc, session);
2340                 break;
2341
2342         case CEPH_SESSION_RECALL_STATE:
2343                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2344                 break;
2345
2346         default:
2347                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2348                 WARN_ON(1);
2349         }
2350
2351         mutex_unlock(&session->s_mutex);
2352         if (wake) {
2353                 mutex_lock(&mdsc->mutex);
2354                 __wake_requests(mdsc, &session->s_waiting);
2355                 mutex_unlock(&mdsc->mutex);
2356         }
2357         return;
2358
2359 bad:
2360         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2361                (int)msg->front.iov_len);
2362         ceph_msg_dump(msg);
2363         return;
2364 }
2365
2366
2367 /*
2368  * called under session->mutex.
2369  */
2370 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2371                                    struct ceph_mds_session *session)
2372 {
2373         struct ceph_mds_request *req, *nreq;
2374         int err;
2375
2376         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2377
2378         mutex_lock(&mdsc->mutex);
2379         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2380                 err = __prepare_send_request(mdsc, req, session->s_mds);
2381                 if (!err) {
2382                         ceph_msg_get(req->r_request);
2383                         ceph_con_send(&session->s_con, req->r_request);
2384                 }
2385         }
2386         mutex_unlock(&mdsc->mutex);
2387 }
2388
2389 /*
2390  * Encode information about a cap for a reconnect with the MDS.
2391  */
2392 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2393                           void *arg)
2394 {
2395         union {
2396                 struct ceph_mds_cap_reconnect v2;
2397                 struct ceph_mds_cap_reconnect_v1 v1;
2398         } rec;
2399         size_t reclen;
2400         struct ceph_inode_info *ci;
2401         struct ceph_reconnect_state *recon_state = arg;
2402         struct ceph_pagelist *pagelist = recon_state->pagelist;
2403         char *path;
2404         int pathlen, err;
2405         u64 pathbase;
2406         struct dentry *dentry;
2407
2408         ci = cap->ci;
2409
2410         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2411              inode, ceph_vinop(inode), cap, cap->cap_id,
2412              ceph_cap_string(cap->issued));
2413         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2414         if (err)
2415                 return err;
2416
2417         dentry = d_find_alias(inode);
2418         if (dentry) {
2419                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2420                 if (IS_ERR(path)) {
2421                         err = PTR_ERR(path);
2422                         goto out_dput;
2423                 }
2424         } else {
2425                 path = NULL;
2426                 pathlen = 0;
2427         }
2428         err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2429         if (err)
2430                 goto out_free;
2431
2432         spin_lock(&ci->i_ceph_lock);
2433         cap->seq = 0;        /* reset cap seq */
2434         cap->issue_seq = 0;  /* and issue_seq */
2435
2436         if (recon_state->flock) {
2437                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2438                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2439                 rec.v2.issued = cpu_to_le32(cap->issued);
2440                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2441                 rec.v2.pathbase = cpu_to_le64(pathbase);
2442                 rec.v2.flock_len = 0;
2443                 reclen = sizeof(rec.v2);
2444         } else {
2445                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2446                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2447                 rec.v1.issued = cpu_to_le32(cap->issued);
2448                 rec.v1.size = cpu_to_le64(inode->i_size);
2449                 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2450                 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2451                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2452                 rec.v1.pathbase = cpu_to_le64(pathbase);
2453                 reclen = sizeof(rec.v1);
2454         }
2455         spin_unlock(&ci->i_ceph_lock);
2456
2457         if (recon_state->flock) {
2458                 int num_fcntl_locks, num_flock_locks;
2459                 struct ceph_pagelist_cursor trunc_point;
2460
2461                 ceph_pagelist_set_cursor(pagelist, &trunc_point);
2462                 do {
2463                         lock_flocks();
2464                         ceph_count_locks(inode, &num_fcntl_locks,
2465                                          &num_flock_locks);
2466                         rec.v2.flock_len = (2*sizeof(u32) +
2467                                             (num_fcntl_locks+num_flock_locks) *
2468                                             sizeof(struct ceph_filelock));
2469                         unlock_flocks();
2470
2471                         /* pre-alloc pagelist */
2472                         ceph_pagelist_truncate(pagelist, &trunc_point);
2473                         err = ceph_pagelist_append(pagelist, &rec, reclen);
2474                         if (!err)
2475                                 err = ceph_pagelist_reserve(pagelist,
2476                                                             rec.v2.flock_len);
2477
2478                         /* encode locks */
2479                         if (!err) {
2480                                 lock_flocks();
2481                                 err = ceph_encode_locks(inode,
2482                                                         pagelist,
2483                                                         num_fcntl_locks,
2484                                                         num_flock_locks);
2485                                 unlock_flocks();
2486                         }
2487                 } while (err == -ENOSPC);
2488         } else {
2489                 err = ceph_pagelist_append(pagelist, &rec, reclen);
2490         }
2491
2492 out_free:
2493         kfree(path);
2494 out_dput:
2495         dput(dentry);
2496         return err;
2497 }
2498
2499
2500 /*
2501  * If an MDS fails and recovers, clients need to reconnect in order to
2502  * reestablish shared state.  This includes all caps issued through
2503  * this session _and_ the snap_realm hierarchy.  Because it's not
2504  * clear which snap realms the mds cares about, we send everything we
2505  * know about.. that ensures we'll then get any new info the
2506  * recovering MDS might have.
2507  *
2508  * This is a relatively heavyweight operation, but it's rare.
2509  *
2510  * called with mdsc->mutex held.
2511  */
2512 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2513                                struct ceph_mds_session *session)
2514 {
2515         struct ceph_msg *reply;
2516         struct rb_node *p;
2517         int mds = session->s_mds;
2518         int err = -ENOMEM;
2519         struct ceph_pagelist *pagelist;
2520         struct ceph_reconnect_state recon_state;
2521
2522         pr_info("mds%d reconnect start\n", mds);
2523
2524         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2525         if (!pagelist)
2526                 goto fail_nopagelist;
2527         ceph_pagelist_init(pagelist);
2528
2529         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2530         if (!reply)
2531                 goto fail_nomsg;
2532
2533         mutex_lock(&session->s_mutex);
2534         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2535         session->s_seq = 0;
2536
2537         ceph_con_open(&session->s_con,
2538                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2539
2540         /* replay unsafe requests */
2541         replay_unsafe_requests(mdsc, session);
2542
2543         down_read(&mdsc->snap_rwsem);
2544
2545         dout("session %p state %s\n", session,
2546              session_state_name(session->s_state));
2547
2548         /* drop old cap expires; we're about to reestablish that state */
2549         discard_cap_releases(mdsc, session);
2550
2551         /* traverse this session's caps */
2552         err = ceph_pagelist_encode_32(pagelist, session->s_nr_caps);
2553         if (err)
2554                 goto fail;
2555
2556         recon_state.pagelist = pagelist;
2557         recon_state.flock = session->s_con.peer_features & CEPH_FEATURE_FLOCK;
2558         err = iterate_session_caps(session, encode_caps_cb, &recon_state);
2559         if (err < 0)
2560                 goto fail;
2561
2562         /*
2563          * snaprealms.  we provide mds with the ino, seq (version), and
2564          * parent for all of our realms.  If the mds has any newer info,
2565          * it will tell us.
2566          */
2567         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2568                 struct ceph_snap_realm *realm =
2569                         rb_entry(p, struct ceph_snap_realm, node);
2570                 struct ceph_mds_snaprealm_reconnect sr_rec;
2571
2572                 dout(" adding snap realm %llx seq %lld parent %llx\n",
2573                      realm->ino, realm->seq, realm->parent_ino);
2574                 sr_rec.ino = cpu_to_le64(realm->ino);
2575                 sr_rec.seq = cpu_to_le64(realm->seq);
2576                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2577                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2578                 if (err)
2579                         goto fail;
2580         }
2581
2582         reply->pagelist = pagelist;
2583         if (recon_state.flock)
2584                 reply->hdr.version = cpu_to_le16(2);
2585         reply->hdr.data_len = cpu_to_le32(pagelist->length);
2586         reply->nr_pages = calc_pages_for(0, pagelist->length);
2587         ceph_con_send(&session->s_con, reply);
2588
2589         mutex_unlock(&session->s_mutex);
2590
2591         mutex_lock(&mdsc->mutex);
2592         __wake_requests(mdsc, &session->s_waiting);
2593         mutex_unlock(&mdsc->mutex);
2594
2595         up_read(&mdsc->snap_rwsem);
2596         return;
2597
2598 fail:
2599         ceph_msg_put(reply);
2600         up_read(&mdsc->snap_rwsem);
2601         mutex_unlock(&session->s_mutex);
2602 fail_nomsg:
2603         ceph_pagelist_release(pagelist);
2604         kfree(pagelist);
2605 fail_nopagelist:
2606         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
2607         return;
2608 }
2609
2610
2611 /*
2612  * compare old and new mdsmaps, kicking requests
2613  * and closing out old connections as necessary
2614  *
2615  * called under mdsc->mutex.
2616  */
2617 static void check_new_map(struct ceph_mds_client *mdsc,
2618                           struct ceph_mdsmap *newmap,
2619                           struct ceph_mdsmap *oldmap)
2620 {
2621         int i;
2622         int oldstate, newstate;
2623         struct ceph_mds_session *s;
2624
2625         dout("check_new_map new %u old %u\n",
2626              newmap->m_epoch, oldmap->m_epoch);
2627
2628         for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2629                 if (mdsc->sessions[i] == NULL)
2630                         continue;
2631                 s = mdsc->sessions[i];
2632                 oldstate = ceph_mdsmap_get_state(oldmap, i);
2633                 newstate = ceph_mdsmap_get_state(newmap, i);
2634
2635                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
2636                      i, ceph_mds_state_name(oldstate),
2637                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
2638                      ceph_mds_state_name(newstate),
2639                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
2640                      session_state_name(s->s_state));
2641
2642                 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
2643                            ceph_mdsmap_get_addr(newmap, i),
2644                            sizeof(struct ceph_entity_addr))) {
2645                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2646                                 /* the session never opened, just close it
2647                                  * out now */
2648                                 __wake_requests(mdsc, &s->s_waiting);
2649                                 __unregister_session(mdsc, s);
2650                         } else {
2651                                 /* just close it */
2652                                 mutex_unlock(&mdsc->mutex);
2653                                 mutex_lock(&s->s_mutex);
2654                                 mutex_lock(&mdsc->mutex);
2655                                 ceph_con_close(&s->s_con);
2656                                 mutex_unlock(&s->s_mutex);
2657                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2658                         }
2659
2660                         /* kick any requests waiting on the recovering mds */
2661                         kick_requests(mdsc, i);
2662                 } else if (oldstate == newstate) {
2663                         continue;  /* nothing new with this mds */
2664                 }
2665
2666                 /*
2667                  * send reconnect?
2668                  */
2669                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2670                     newstate >= CEPH_MDS_STATE_RECONNECT) {
2671                         mutex_unlock(&mdsc->mutex);
2672                         send_mds_reconnect(mdsc, s);
2673                         mutex_lock(&mdsc->mutex);
2674                 }
2675
2676                 /*
2677                  * kick request on any mds that has gone active.
2678                  */
2679                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2680                     newstate >= CEPH_MDS_STATE_ACTIVE) {
2681                         if (oldstate != CEPH_MDS_STATE_CREATING &&
2682                             oldstate != CEPH_MDS_STATE_STARTING)
2683                                 pr_info("mds%d recovery completed\n", s->s_mds);
2684                         kick_requests(mdsc, i);
2685                         ceph_kick_flushing_caps(mdsc, s);
2686                         wake_up_session_caps(s, 1);
2687                 }
2688         }
2689
2690         for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
2691                 s = mdsc->sessions[i];
2692                 if (!s)
2693                         continue;
2694                 if (!ceph_mdsmap_is_laggy(newmap, i))
2695                         continue;
2696                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2697                     s->s_state == CEPH_MDS_SESSION_HUNG ||
2698                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
2699                         dout(" connecting to export targets of laggy mds%d\n",
2700                              i);
2701                         __open_export_target_sessions(mdsc, s);
2702                 }
2703         }
2704 }
2705
2706
2707
2708 /*
2709  * leases
2710  */
2711
2712 /*
2713  * caller must hold session s_mutex, dentry->d_lock
2714  */
2715 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2716 {
2717         struct ceph_dentry_info *di = ceph_dentry(dentry);
2718
2719         ceph_put_mds_session(di->lease_session);
2720         di->lease_session = NULL;
2721 }
2722
2723 static void handle_lease(struct ceph_mds_client *mdsc,
2724                          struct ceph_mds_session *session,
2725                          struct ceph_msg *msg)
2726 {
2727         struct super_block *sb = mdsc->fsc->sb;
2728         struct inode *inode;
2729         struct dentry *parent, *dentry;
2730         struct ceph_dentry_info *di;
2731         int mds = session->s_mds;
2732         struct ceph_mds_lease *h = msg->front.iov_base;
2733         u32 seq;
2734         struct ceph_vino vino;
2735         struct qstr dname;
2736         int release = 0;
2737
2738         dout("handle_lease from mds%d\n", mds);
2739
2740         /* decode */
2741         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2742                 goto bad;
2743         vino.ino = le64_to_cpu(h->ino);
2744         vino.snap = CEPH_NOSNAP;
2745         seq = le32_to_cpu(h->seq);
2746         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2747         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2748         if (dname.len != get_unaligned_le32(h+1))
2749                 goto bad;
2750
2751         mutex_lock(&session->s_mutex);
2752         session->s_seq++;
2753
2754         /* lookup inode */
2755         inode = ceph_find_inode(sb, vino);
2756         dout("handle_lease %s, ino %llx %p %.*s\n",
2757              ceph_lease_op_name(h->action), vino.ino, inode,
2758              dname.len, dname.name);
2759         if (inode == NULL) {
2760                 dout("handle_lease no inode %llx\n", vino.ino);
2761                 goto release;
2762         }
2763
2764         /* dentry */
2765         parent = d_find_alias(inode);
2766         if (!parent) {
2767                 dout("no parent dentry on inode %p\n", inode);
2768                 WARN_ON(1);
2769                 goto release;  /* hrm... */
2770         }
2771         dname.hash = full_name_hash(dname.name, dname.len);
2772         dentry = d_lookup(parent, &dname);
2773         dput(parent);
2774         if (!dentry)
2775                 goto release;
2776
2777         spin_lock(&dentry->d_lock);
2778         di = ceph_dentry(dentry);
2779         switch (h->action) {
2780         case CEPH_MDS_LEASE_REVOKE:
2781                 if (di && di->lease_session == session) {
2782                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
2783                                 h->seq = cpu_to_le32(di->lease_seq);
2784                         __ceph_mdsc_drop_dentry_lease(dentry);
2785                 }
2786                 release = 1;
2787                 break;
2788
2789         case CEPH_MDS_LEASE_RENEW:
2790                 if (di && di->lease_session == session &&
2791                     di->lease_gen == session->s_cap_gen &&
2792                     di->lease_renew_from &&
2793                     di->lease_renew_after == 0) {
2794                         unsigned long duration =
2795                                 le32_to_cpu(h->duration_ms) * HZ / 1000;
2796
2797                         di->lease_seq = seq;
2798                         dentry->d_time = di->lease_renew_from + duration;
2799                         di->lease_renew_after = di->lease_renew_from +
2800                                 (duration >> 1);
2801                         di->lease_renew_from = 0;
2802                 }
2803                 break;
2804         }
2805         spin_unlock(&dentry->d_lock);
2806         dput(dentry);
2807
2808         if (!release)
2809                 goto out;
2810
2811 release:
2812         /* let's just reuse the same message */
2813         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2814         ceph_msg_get(msg);
2815         ceph_con_send(&session->s_con, msg);
2816
2817 out:
2818         iput(inode);
2819         mutex_unlock(&session->s_mutex);
2820         return;
2821
2822 bad:
2823         pr_err("corrupt lease message\n");
2824         ceph_msg_dump(msg);
2825 }
2826
2827 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
2828                               struct inode *inode,
2829                               struct dentry *dentry, char action,
2830                               u32 seq)
2831 {
2832         struct ceph_msg *msg;
2833         struct ceph_mds_lease *lease;
2834         int len = sizeof(*lease) + sizeof(u32);
2835         int dnamelen = 0;
2836
2837         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2838              inode, dentry, ceph_lease_op_name(action), session->s_mds);
2839         dnamelen = dentry->d_name.len;
2840         len += dnamelen;
2841
2842         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
2843         if (!msg)
2844                 return;
2845         lease = msg->front.iov_base;
2846         lease->action = action;
2847         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
2848         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
2849         lease->seq = cpu_to_le32(seq);
2850         put_unaligned_le32(dnamelen, lease + 1);
2851         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
2852
2853         /*
2854          * if this is a preemptive lease RELEASE, no need to
2855          * flush request stream, since the actual request will
2856          * soon follow.
2857          */
2858         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
2859
2860         ceph_con_send(&session->s_con, msg);
2861 }
2862
2863 /*
2864  * Preemptively release a lease we expect to invalidate anyway.
2865  * Pass @inode always, @dentry is optional.
2866  */
2867 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2868                              struct dentry *dentry)
2869 {
2870         struct ceph_dentry_info *di;
2871         struct ceph_mds_session *session;
2872         u32 seq;
2873
2874         BUG_ON(inode == NULL);
2875         BUG_ON(dentry == NULL);
2876
2877         /* is dentry lease valid? */
2878         spin_lock(&dentry->d_lock);
2879         di = ceph_dentry(dentry);
2880         if (!di || !di->lease_session ||
2881             di->lease_session->s_mds < 0 ||
2882             di->lease_gen != di->lease_session->s_cap_gen ||
2883             !time_before(jiffies, dentry->d_time)) {
2884                 dout("lease_release inode %p dentry %p -- "
2885                      "no lease\n",
2886                      inode, dentry);
2887                 spin_unlock(&dentry->d_lock);
2888                 return;
2889         }
2890
2891         /* we do have a lease on this dentry; note mds and seq */
2892         session = ceph_get_mds_session(di->lease_session);
2893         seq = di->lease_seq;
2894         __ceph_mdsc_drop_dentry_lease(dentry);
2895         spin_unlock(&dentry->d_lock);
2896
2897         dout("lease_release inode %p dentry %p to mds%d\n",
2898              inode, dentry, session->s_mds);
2899         ceph_mdsc_lease_send_msg(session, inode, dentry,
2900                                  CEPH_MDS_LEASE_RELEASE, seq);
2901         ceph_put_mds_session(session);
2902 }
2903
2904 /*
2905  * drop all leases (and dentry refs) in preparation for umount
2906  */
2907 static void drop_leases(struct ceph_mds_client *mdsc)
2908 {
2909         int i;
2910
2911         dout("drop_leases\n");
2912         mutex_lock(&mdsc->mutex);
2913         for (i = 0; i < mdsc->max_sessions; i++) {
2914                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2915                 if (!s)
2916                         continue;
2917                 mutex_unlock(&mdsc->mutex);
2918                 mutex_lock(&s->s_mutex);
2919                 mutex_unlock(&s->s_mutex);
2920                 ceph_put_mds_session(s);
2921                 mutex_lock(&mdsc->mutex);
2922         }
2923         mutex_unlock(&mdsc->mutex);
2924 }
2925
2926
2927
2928 /*
2929  * delayed work -- periodically trim expired leases, renew caps with mds
2930  */
2931 static void schedule_delayed(struct ceph_mds_client *mdsc)
2932 {
2933         int delay = 5;
2934         unsigned hz = round_jiffies_relative(HZ * delay);
2935         schedule_delayed_work(&mdsc->delayed_work, hz);
2936 }
2937
2938 static void delayed_work(struct work_struct *work)
2939 {
2940         int i;
2941         struct ceph_mds_client *mdsc =
2942                 container_of(work, struct ceph_mds_client, delayed_work.work);
2943         int renew_interval;
2944         int renew_caps;
2945
2946         dout("mdsc delayed_work\n");
2947         ceph_check_delayed_caps(mdsc);
2948
2949         mutex_lock(&mdsc->mutex);
2950         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
2951         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
2952                                    mdsc->last_renew_caps);
2953         if (renew_caps)
2954                 mdsc->last_renew_caps = jiffies;
2955
2956         for (i = 0; i < mdsc->max_sessions; i++) {
2957                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2958                 if (s == NULL)
2959                         continue;
2960                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
2961                         dout("resending session close request for mds%d\n",
2962                              s->s_mds);
2963                         request_close_session(mdsc, s);
2964                         ceph_put_mds_session(s);
2965                         continue;
2966                 }
2967                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
2968                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
2969                                 s->s_state = CEPH_MDS_SESSION_HUNG;
2970                                 pr_info("mds%d hung\n", s->s_mds);
2971                         }
2972                 }
2973                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
2974                         /* this mds is failed or recovering, just wait */
2975                         ceph_put_mds_session(s);
2976                         continue;
2977                 }
2978                 mutex_unlock(&mdsc->mutex);
2979
2980                 mutex_lock(&s->s_mutex);
2981                 if (renew_caps)
2982                         send_renew_caps(mdsc, s);
2983                 else
2984                         ceph_con_keepalive(&s->s_con);
2985                 ceph_add_cap_releases(mdsc, s);
2986                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2987                     s->s_state == CEPH_MDS_SESSION_HUNG)
2988                         ceph_send_cap_releases(mdsc, s);
2989                 mutex_unlock(&s->s_mutex);
2990                 ceph_put_mds_session(s);
2991
2992                 mutex_lock(&mdsc->mutex);
2993         }
2994         mutex_unlock(&mdsc->mutex);
2995
2996         schedule_delayed(mdsc);
2997 }
2998
2999 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3000
3001 {
3002         struct ceph_mds_client *mdsc;
3003
3004         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3005         if (!mdsc)
3006                 return -ENOMEM;
3007         mdsc->fsc = fsc;
3008         fsc->mdsc = mdsc;
3009         mutex_init(&mdsc->mutex);
3010         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3011         if (mdsc->mdsmap == NULL)
3012                 return -ENOMEM;
3013
3014         init_completion(&mdsc->safe_umount_waiters);
3015         init_waitqueue_head(&mdsc->session_close_wq);
3016         INIT_LIST_HEAD(&mdsc->waiting_for_map);
3017         mdsc->sessions = NULL;
3018         mdsc->max_sessions = 0;
3019         mdsc->stopping = 0;
3020         init_rwsem(&mdsc->snap_rwsem);
3021         mdsc->snap_realms = RB_ROOT;
3022         INIT_LIST_HEAD(&mdsc->snap_empty);
3023         spin_lock_init(&mdsc->snap_empty_lock);
3024         mdsc->last_tid = 0;
3025         mdsc->request_tree = RB_ROOT;
3026         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3027         mdsc->last_renew_caps = jiffies;
3028         INIT_LIST_HEAD(&mdsc->cap_delay_list);
3029         spin_lock_init(&mdsc->cap_delay_lock);
3030         INIT_LIST_HEAD(&mdsc->snap_flush_list);
3031         spin_lock_init(&mdsc->snap_flush_lock);
3032         mdsc->cap_flush_seq = 0;
3033         INIT_LIST_HEAD(&mdsc->cap_dirty);
3034         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3035         mdsc->num_cap_flushing = 0;
3036         spin_lock_init(&mdsc->cap_dirty_lock);
3037         init_waitqueue_head(&mdsc->cap_flushing_wq);
3038         spin_lock_init(&mdsc->dentry_lru_lock);
3039         INIT_LIST_HEAD(&mdsc->dentry_lru);
3040
3041         ceph_caps_init(mdsc);
3042         ceph_adjust_min_caps(mdsc, fsc->min_caps);
3043
3044         return 0;
3045 }
3046
3047 /*
3048  * Wait for safe replies on open mds requests.  If we time out, drop
3049  * all requests from the tree to avoid dangling dentry refs.
3050  */
3051 static void wait_requests(struct ceph_mds_client *mdsc)
3052 {
3053         struct ceph_mds_request *req;
3054         struct ceph_fs_client *fsc = mdsc->fsc;
3055
3056         mutex_lock(&mdsc->mutex);
3057         if (__get_oldest_req(mdsc)) {
3058                 mutex_unlock(&mdsc->mutex);
3059
3060                 dout("wait_requests waiting for requests\n");
3061                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3062                                     fsc->client->options->mount_timeout * HZ);
3063
3064                 /* tear down remaining requests */
3065                 mutex_lock(&mdsc->mutex);
3066                 while ((req = __get_oldest_req(mdsc))) {
3067                         dout("wait_requests timed out on tid %llu\n",
3068                              req->r_tid);
3069                         __unregister_request(mdsc, req);
3070                 }
3071         }
3072         mutex_unlock(&mdsc->mutex);
3073         dout("wait_requests done\n");
3074 }
3075
3076 /*
3077  * called before mount is ro, and before dentries are torn down.
3078  * (hmm, does this still race with new lookups?)
3079  */
3080 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3081 {
3082         dout("pre_umount\n");
3083         mdsc->stopping = 1;
3084
3085         drop_leases(mdsc);
3086         ceph_flush_dirty_caps(mdsc);
3087         wait_requests(mdsc);
3088
3089         /*
3090          * wait for reply handlers to drop their request refs and
3091          * their inode/dcache refs
3092          */
3093         ceph_msgr_flush();
3094 }
3095
3096 /*
3097  * wait for all write mds requests to flush.
3098  */
3099 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3100 {
3101         struct ceph_mds_request *req = NULL, *nextreq;
3102         struct rb_node *n;
3103
3104         mutex_lock(&mdsc->mutex);
3105         dout("wait_unsafe_requests want %lld\n", want_tid);
3106 restart:
3107         req = __get_oldest_req(mdsc);
3108         while (req && req->r_tid <= want_tid) {
3109                 /* find next request */
3110                 n = rb_next(&req->r_node);
3111                 if (n)
3112                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3113                 else
3114                         nextreq = NULL;
3115                 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
3116                         /* write op */
3117                         ceph_mdsc_get_request(req);
3118                         if (nextreq)
3119                                 ceph_mdsc_get_request(nextreq);
3120                         mutex_unlock(&mdsc->mutex);
3121                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
3122                              req->r_tid, want_tid);
3123                         wait_for_completion(&req->r_safe_completion);
3124                         mutex_lock(&mdsc->mutex);
3125                         ceph_mdsc_put_request(req);
3126                         if (!nextreq)
3127                                 break;  /* next dne before, so we're done! */
3128                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
3129                                 /* next request was removed from tree */
3130                                 ceph_mdsc_put_request(nextreq);
3131                                 goto restart;
3132                         }
3133                         ceph_mdsc_put_request(nextreq);  /* won't go away */
3134                 }
3135                 req = nextreq;
3136         }
3137         mutex_unlock(&mdsc->mutex);
3138         dout("wait_unsafe_requests done\n");
3139 }
3140
3141 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3142 {
3143         u64 want_tid, want_flush;
3144
3145         if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3146                 return;
3147
3148         dout("sync\n");
3149         mutex_lock(&mdsc->mutex);
3150         want_tid = mdsc->last_tid;
3151         want_flush = mdsc->cap_flush_seq;
3152         mutex_unlock(&mdsc->mutex);
3153         dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
3154
3155         ceph_flush_dirty_caps(mdsc);
3156
3157         wait_unsafe_requests(mdsc, want_tid);
3158         wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
3159 }
3160
3161 /*
3162  * true if all sessions are closed, or we force unmount
3163  */
3164 static bool done_closing_sessions(struct ceph_mds_client *mdsc)
3165 {
3166         int i, n = 0;
3167
3168         if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3169                 return true;
3170
3171         mutex_lock(&mdsc->mutex);
3172         for (i = 0; i < mdsc->max_sessions; i++)
3173                 if (mdsc->sessions[i])
3174                         n++;
3175         mutex_unlock(&mdsc->mutex);
3176         return n == 0;
3177 }
3178
3179 /*
3180  * called after sb is ro.
3181  */
3182 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3183 {
3184         struct ceph_mds_session *session;
3185         int i;
3186         struct ceph_fs_client *fsc = mdsc->fsc;
3187         unsigned long timeout = fsc->client->options->mount_timeout * HZ;
3188
3189         dout("close_sessions\n");
3190
3191         /* close sessions */
3192         mutex_lock(&mdsc->mutex);
3193         for (i = 0; i < mdsc->max_sessions; i++) {
3194                 session = __ceph_lookup_mds_session(mdsc, i);
3195                 if (!session)
3196                         continue;
3197                 mutex_unlock(&mdsc->mutex);
3198                 mutex_lock(&session->s_mutex);
3199                 __close_session(mdsc, session);
3200                 mutex_unlock(&session->s_mutex);
3201                 ceph_put_mds_session(session);
3202                 mutex_lock(&mdsc->mutex);
3203         }
3204         mutex_unlock(&mdsc->mutex);
3205
3206         dout("waiting for sessions to close\n");
3207         wait_event_timeout(mdsc->session_close_wq, done_closing_sessions(mdsc),
3208                            timeout);
3209
3210         /* tear down remaining sessions */
3211         mutex_lock(&mdsc->mutex);
3212         for (i = 0; i < mdsc->max_sessions; i++) {
3213                 if (mdsc->sessions[i]) {
3214                         session = get_session(mdsc->sessions[i]);
3215                         __unregister_session(mdsc, session);
3216                         mutex_unlock(&mdsc->mutex);
3217                         mutex_lock(&session->s_mutex);
3218                         remove_session_caps(session);
3219                         mutex_unlock(&session->s_mutex);
3220                         ceph_put_mds_session(session);
3221                         mutex_lock(&mdsc->mutex);
3222                 }
3223         }
3224         WARN_ON(!list_empty(&mdsc->cap_delay_list));
3225         mutex_unlock(&mdsc->mutex);
3226
3227         ceph_cleanup_empty_realms(mdsc);
3228
3229         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3230
3231         dout("stopped\n");
3232 }
3233
3234 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3235 {
3236         dout("stop\n");
3237         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3238         if (mdsc->mdsmap)
3239                 ceph_mdsmap_destroy(mdsc->mdsmap);
3240         kfree(mdsc->sessions);
3241         ceph_caps_finalize(mdsc);
3242 }
3243
3244 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3245 {
3246         struct ceph_mds_client *mdsc = fsc->mdsc;
3247
3248         dout("mdsc_destroy %p\n", mdsc);
3249         ceph_mdsc_stop(mdsc);
3250
3251         /* flush out any connection work with references to us */
3252         ceph_msgr_flush();
3253
3254         fsc->mdsc = NULL;
3255         kfree(mdsc);
3256         dout("mdsc_destroy %p done\n", mdsc);
3257 }
3258
3259
3260 /*
3261  * handle mds map update.
3262  */
3263 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3264 {
3265         u32 epoch;
3266         u32 maplen;
3267         void *p = msg->front.iov_base;
3268         void *end = p + msg->front.iov_len;
3269         struct ceph_mdsmap *newmap, *oldmap;
3270         struct ceph_fsid fsid;
3271         int err = -EINVAL;
3272
3273         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3274         ceph_decode_copy(&p, &fsid, sizeof(fsid));
3275         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3276                 return;
3277         epoch = ceph_decode_32(&p);
3278         maplen = ceph_decode_32(&p);
3279         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3280
3281         /* do we need it? */
3282         ceph_monc_got_mdsmap(&mdsc->fsc->client->monc, epoch);
3283         mutex_lock(&mdsc->mutex);
3284         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3285                 dout("handle_map epoch %u <= our %u\n",
3286                      epoch, mdsc->mdsmap->m_epoch);
3287                 mutex_unlock(&mdsc->mutex);
3288                 return;
3289         }
3290
3291         newmap = ceph_mdsmap_decode(&p, end);
3292         if (IS_ERR(newmap)) {
3293                 err = PTR_ERR(newmap);
3294                 goto bad_unlock;
3295         }
3296
3297         /* swap into place */
3298         if (mdsc->mdsmap) {
3299                 oldmap = mdsc->mdsmap;
3300                 mdsc->mdsmap = newmap;
3301                 check_new_map(mdsc, newmap, oldmap);
3302                 ceph_mdsmap_destroy(oldmap);
3303         } else {
3304                 mdsc->mdsmap = newmap;  /* first mds map */
3305         }
3306         mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3307
3308         __wake_requests(mdsc, &mdsc->waiting_for_map);
3309
3310         mutex_unlock(&mdsc->mutex);
3311         schedule_delayed(mdsc);
3312         return;
3313
3314 bad_unlock:
3315         mutex_unlock(&mdsc->mutex);
3316 bad:
3317         pr_err("error decoding mdsmap %d\n", err);
3318         return;
3319 }
3320
3321 static struct ceph_connection *con_get(struct ceph_connection *con)
3322 {
3323         struct ceph_mds_session *s = con->private;
3324
3325         if (get_session(s)) {
3326                 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3327                 return con;
3328         }
3329         dout("mdsc con_get %p FAIL\n", s);
3330         return NULL;
3331 }
3332
3333 static void con_put(struct ceph_connection *con)
3334 {
3335         struct ceph_mds_session *s = con->private;
3336
3337         dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3338         ceph_put_mds_session(s);
3339 }
3340
3341 /*
3342  * if the client is unresponsive for long enough, the mds will kill
3343  * the session entirely.
3344  */
3345 static void peer_reset(struct ceph_connection *con)
3346 {
3347         struct ceph_mds_session *s = con->private;
3348         struct ceph_mds_client *mdsc = s->s_mdsc;
3349
3350         pr_warning("mds%d closed our session\n", s->s_mds);
3351         send_mds_reconnect(mdsc, s);
3352 }
3353
3354 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3355 {
3356         struct ceph_mds_session *s = con->private;
3357         struct ceph_mds_client *mdsc = s->s_mdsc;
3358         int type = le16_to_cpu(msg->hdr.type);
3359
3360         mutex_lock(&mdsc->mutex);
3361         if (__verify_registered_session(mdsc, s) < 0) {
3362                 mutex_unlock(&mdsc->mutex);
3363                 goto out;
3364         }
3365         mutex_unlock(&mdsc->mutex);
3366
3367         switch (type) {
3368         case CEPH_MSG_MDS_MAP:
3369                 ceph_mdsc_handle_map(mdsc, msg);
3370                 break;
3371         case CEPH_MSG_CLIENT_SESSION:
3372                 handle_session(s, msg);
3373                 break;
3374         case CEPH_MSG_CLIENT_REPLY:
3375                 handle_reply(s, msg);
3376                 break;
3377         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3378                 handle_forward(mdsc, s, msg);
3379                 break;
3380         case CEPH_MSG_CLIENT_CAPS:
3381                 ceph_handle_caps(s, msg);
3382                 break;
3383         case CEPH_MSG_CLIENT_SNAP:
3384                 ceph_handle_snap(mdsc, s, msg);
3385                 break;
3386         case CEPH_MSG_CLIENT_LEASE:
3387                 handle_lease(mdsc, s, msg);
3388                 break;
3389
3390         default:
3391                 pr_err("received unknown message type %d %s\n", type,
3392                        ceph_msg_type_name(type));
3393         }
3394 out:
3395         ceph_msg_put(msg);
3396 }
3397
3398 /*
3399  * authentication
3400  */
3401 static int get_authorizer(struct ceph_connection *con,
3402                           void **buf, int *len, int *proto,
3403                           void **reply_buf, int *reply_len, int force_new)
3404 {
3405         struct ceph_mds_session *s = con->private;
3406         struct ceph_mds_client *mdsc = s->s_mdsc;
3407         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3408         int ret = 0;
3409
3410         if (force_new && s->s_authorizer) {
3411                 ac->ops->destroy_authorizer(ac, s->s_authorizer);
3412                 s->s_authorizer = NULL;
3413         }
3414         if (s->s_authorizer == NULL) {
3415                 if (ac->ops->create_authorizer) {
3416                         ret = ac->ops->create_authorizer(
3417                                 ac, CEPH_ENTITY_TYPE_MDS,
3418                                 &s->s_authorizer,
3419                                 &s->s_authorizer_buf,
3420                                 &s->s_authorizer_buf_len,
3421                                 &s->s_authorizer_reply_buf,
3422                                 &s->s_authorizer_reply_buf_len);
3423                         if (ret)
3424                                 return ret;
3425                 }
3426         }
3427
3428         *proto = ac->protocol;
3429         *buf = s->s_authorizer_buf;
3430         *len = s->s_authorizer_buf_len;
3431         *reply_buf = s->s_authorizer_reply_buf;
3432         *reply_len = s->s_authorizer_reply_buf_len;
3433         return 0;
3434 }
3435
3436
3437 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3438 {
3439         struct ceph_mds_session *s = con->private;
3440         struct ceph_mds_client *mdsc = s->s_mdsc;
3441         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3442
3443         return ac->ops->verify_authorizer_reply(ac, s->s_authorizer, len);
3444 }
3445
3446 static int invalidate_authorizer(struct ceph_connection *con)
3447 {
3448         struct ceph_mds_session *s = con->private;
3449         struct ceph_mds_client *mdsc = s->s_mdsc;
3450         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3451
3452         if (ac->ops->invalidate_authorizer)
3453                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3454
3455         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
3456 }
3457
3458 static const struct ceph_connection_operations mds_con_ops = {
3459         .get = con_get,
3460         .put = con_put,
3461         .dispatch = dispatch,
3462         .get_authorizer = get_authorizer,
3463         .verify_authorizer_reply = verify_authorizer_reply,
3464         .invalidate_authorizer = invalidate_authorizer,
3465         .peer_reset = peer_reset,
3466 };
3467
3468 /* eof */