fs/ceph/mds_client.c

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  41  * are no longer valid.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