fs/ceph/mds_client.c

   1 #include <linux/ceph/ceph_debug.h>
   2
   3 #include <linux/fs.h>
   4 #include <linux/wait.h>
   5 #include <linux/slab.h>
   6 #include <linux/sched.h>
   7 #include <linux/debugfs.h>
   8 #include <linux/seq_file.h>
   9
  10 #include "super.h"
  11 #include "mds_client.h"
  12
  13 #include <linux/ceph/messenger.h>
  14 #include <linux/ceph/decode.h>
  15 #include <linux/ceph/pagelist.h>
  16 #include <linux/ceph/auth.h>
  17 #include <linux/ceph/debugfs.h>
  18
  19 /*
  20  * A cluster of MDS (metadata server) daemons is responsible for
  21  * managing the file system namespace (the directory hierarchy and
  22  * inodes) and for coordinating shared access to storage.  Metadata is
  23  * partitioning hierarchically across a number of servers, and that
  24  * partition varies over time as the cluster adjusts the distribution
  25  * in order to balance load.
  26  *
  27  * The MDS client is primarily responsible to managing synchronous
  28  * metadata requests for operations like open, unlink, and so forth.
  29  * If there is a MDS failure, we find out about it when we (possibly
  30  * request and) receive a new MDS map, and can resubmit affected
  31  * requests.
  32  *
  33  * For the most part, though, we take advantage of a lossless
  34  * communications channel to the MDS, and do not need to worry about
  35  * timing out or resubmitting requests.
  36  *
  37  * We maintain a stateful "session" with each MDS we interact with.
  38  * Within each session, we sent periodic heartbeat messages to ensure
  39  * any capabilities or leases we have been issues remain valid.  If
  40  * the session times out and goes stale, our leases and capabilities
  41  * are no longer valid.
  42  */
  43
  44 struct ceph_reconnect_state {
  45         struct ceph_pagelist *pagelist;
  46         bool flock;
  47 };
  48
  49 static void __wake_requests(struct ceph_mds_client *mdsc,
  50                             struct list_head *head);
  51
  52 static const struct ceph_connection_operations mds_con_ops;
  53
  54
  55 /*
  56  * mds reply parsing
  57  */
  58
  59 /*
  60  * parse individual inode info
  61  */
  62 static int parse_reply_info_in(void **p, void *end,
  63                                struct ceph_mds_reply_info_in *info,
  64                                int features)
  65 {
  66         int err = -EIO;
  67
  68         info->in = *p;
  69         *p += sizeof(struct ceph_mds_reply_inode) +
  70                 sizeof(*info->in->fragtree.splits) *
  71                 le32_to_cpu(info->in->fragtree.nsplits);
  72
  73         ceph_decode_32_safe(p, end, info->symlink_len, bad);
  74         ceph_decode_need(p, end, info->symlink_len, bad);
  75         info->symlink = *p;
  76         *p += info->symlink_len;
  77
  78         if (features & CEPH_FEATURE_DIRLAYOUTHASH)
  79                 ceph_decode_copy_safe(p, end, &info->dir_layout,
  80                                       sizeof(info->dir_layout), bad);
  81         else
  82                 memset(&info->dir_layout, 0, sizeof(info->dir_layout));
  83
  84         ceph_decode_32_safe(p, end, info->xattr_len, bad);
  85         ceph_decode_need(p, end, info->xattr_len, bad);
  86         info->xattr_data = *p;
  87         *p += info->xattr_len;
  88         return 0;
  89 bad:
  90         return err;
  91 }
  92
  93 /*
  94  * parse a normal reply, which may contain a (dir+)dentry and/or a
  95  * target inode.
  96  */
  97 static int parse_reply_info_trace(void **p, void *end,
  98                                   struct ceph_mds_reply_info_parsed *info,
  99                                   int features)
 100 {
 101         int err;
 102
 103         if (info->head->is_dentry) {
 104                 err = parse_reply_info_in(p, end, &info->diri, features);
 105                 if (err < 0)
 106                         goto out_bad;
 107
 108                 if (unlikely(*p + sizeof(*info->dirfrag) > end))
 109                         goto bad;
 110                 info->dirfrag = *p;
 111                 *p += sizeof(*info->dirfrag) +
 112                         sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
 113                 if (unlikely(*p > end))
 114                         goto bad;
 115
 116                 ceph_decode_32_safe(p, end, info->dname_len, bad);
 117                 ceph_decode_need(p, end, info->dname_len, bad);
 118                 info->dname = *p;
 119                 *p += info->dname_len;
 120                 info->dlease = *p;
 121                 *p += sizeof(*info->dlease);
 122         }
 123
 124         if (info->head->is_target) {
 125                 err = parse_reply_info_in(p, end, &info->targeti, features);
 126                 if (err < 0)
 127                         goto out_bad;
 128         }
 129
 130         if (unlikely(*p != end))
 131                 goto bad;
 132         return 0;
 133
 134 bad:
 135         err = -EIO;
 136 out_bad:
 137         pr_err("problem parsing mds trace %d\n", err);
 138         return err;
 139 }
 140
 141 /*
 142  * parse readdir results
 143  */
 144 static int parse_reply_info_dir(void **p, void *end,
 145                                 struct ceph_mds_reply_info_parsed *info,
 146                                 int features)
 147 {
 148         u32 num, i = 0;
 149         int err;
 150
 151         info->dir_dir = *p;
 152         if (*p + sizeof(*info->dir_dir) > end)
 153                 goto bad;
 154         *p += sizeof(*info->dir_dir) +
 155                 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
 156         if (*p > end)
 157                 goto bad;
 158
 159         ceph_decode_need(p, end, sizeof(num) + 2, bad);
 160         num = ceph_decode_32(p);
 161         info->dir_end = ceph_decode_8(p);
 162         info->dir_complete = ceph_decode_8(p);
 163         if (num == 0)
 164                 goto done;
 165
 166         /* alloc large array */
 167         info->dir_nr = num;
 168         info->dir_in = kcalloc(num, sizeof(*info->dir_in) +
 169                                sizeof(*info->dir_dname) +
 170                                sizeof(*info->dir_dname_len) +
 171                                sizeof(*info->dir_dlease),
 172                                GFP_NOFS);
 173         if (info->dir_in == NULL) {
 174                 err = -ENOMEM;
 175                 goto out_bad;
 176         }
 177         info->dir_dname = (void *)(info->dir_in + num);
 178         info->dir_dname_len = (void *)(info->dir_dname + num);
 179         info->dir_dlease = (void *)(info->dir_dname_len + num);
 180
 181         while (num) {
 182                 /* dentry */
 183                 ceph_decode_need(p, end, sizeof(u32)*2, bad);
 184                 info->dir_dname_len[i] = ceph_decode_32(p);
 185                 ceph_decode_need(p, end, info->dir_dname_len[i], bad);
 186                 info->dir_dname[i] = *p;
 187                 *p += info->dir_dname_len[i];
 188                 dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
 189                      info->dir_dname[i]);
 190                 info->dir_dlease[i] = *p;
 191                 *p += sizeof(struct ceph_mds_reply_lease);
 192
 193                 /* inode */
 194                 err = parse_reply_info_in(p, end, &info->dir_in[i], features);
 195                 if (err < 0)
 196                         goto out_bad;
 197                 i++;
 198                 num--;
 199         }
 200
 201 done:
 202         if (*p != end)
 203                 goto bad;
 204         return 0;
 205
 206 bad:
 207         err = -EIO;
 208 out_bad:
 209         pr_err("problem parsing dir contents %d\n", err);
 210         return err;
 211 }
 212
 213 /*
 214  * parse fcntl F_GETLK results
 215  */
 216 static int parse_reply_info_filelock(void **p, void *end,
 217                                      struct ceph_mds_reply_info_parsed *info,
 218                                      int features)
 219 {
 220         if (*p + sizeof(*info->filelock_reply) > end)
 221                 goto bad;
 222
 223         info->filelock_reply = *p;
 224         *p += sizeof(*info->filelock_reply);
 225
 226         if (unlikely(*p != end))
 227                 goto bad;
 228         return 0;
 229
 230 bad:
 231         return -EIO;
 232 }
 233
 234 /*
 235  * parse extra results
 236  */
 237 static int parse_reply_info_extra(void **p, void *end,
 238                                   struct ceph_mds_reply_info_parsed *info,
 239                                   int features)
 240 {
 241         if (info->head->op == CEPH_MDS_OP_GETFILELOCK)
 242                 return parse_reply_info_filelock(p, end, info, features);
 243         else
 244                 return parse_reply_info_dir(p, end, info, features);
 245 }
 246
 247 /*
 248  * parse entire mds reply
 249  */
 250 static int parse_reply_info(struct ceph_msg *msg,
 251                             struct ceph_mds_reply_info_parsed *info,
 252                             int features)
 253 {
 254         void *p, *end;
 255         u32 len;
 256         int err;
 257
 258         info->head = msg->front.iov_base;
 259         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 260         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 261
 262         /* trace */
 263         ceph_decode_32_safe(&p, end, len, bad);
 264         if (len > 0) {
 265                 err = parse_reply_info_trace(&p, p+len, info, features);
 266                 if (err < 0)
 267                         goto out_bad;
 268         }
 269
 270         /* extra */
 271         ceph_decode_32_safe(&p, end, len, bad);
 272         if (len > 0) {
 273                 err = parse_reply_info_extra(&p, p+len, info, features);
 274                 if (err < 0)
 275                         goto out_bad;
 276         }
 277
 278         /* snap blob */
 279         ceph_decode_32_safe(&p, end, len, bad);
 280         info->snapblob_len = len;
 281         info->snapblob = p;
 282         p += len;
 283
 284         if (p != end)
 285                 goto bad;
 286         return 0;
 287
 288 bad:
 289         err = -EIO;
 290 out_bad:
 291         pr_err("mds parse_reply err %d\n", err);
 292         return err;
 293 }
 294
 295 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
 296 {
 297         kfree(info->dir_in);
 298 }
 299
 300
 301 /*
 302  * sessions
 303  */
 304 static const char *session_state_name(int s)
 305 {
 306         switch (s) {
 307         case CEPH_MDS_SESSION_NEW: return "new";
 308         case CEPH_MDS_SESSION_OPENING: return "opening";
 309         case CEPH_MDS_SESSION_OPEN: return "open";
 310         case CEPH_MDS_SESSION_HUNG: return "hung";
 311         case CEPH_MDS_SESSION_CLOSING: return "closing";
 312         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
 313         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
 314         default: return "???";
 315         }
 316 }
 317
 318 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
 319 {
 320         if (atomic_inc_not_zero(&s->s_ref)) {
 321                 dout("mdsc get_session %p %d -> %d\n", s,
 322                      atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
 323                 return s;
 324         } else {
 325                 dout("mdsc get_session %p 0 -- FAIL", s);
 326                 return NULL;
 327         }
 328 }
 329
 330 void ceph_put_mds_session(struct ceph_mds_session *s)
 331 {
 332         dout("mdsc put_session %p %d -> %d\n", s,
 333              atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
 334         if (atomic_dec_and_test(&s->s_ref)) {
 335                 if (s->s_authorizer)
 336                      s->s_mdsc->fsc->client->monc.auth->ops->destroy_authorizer(
 337                              s->s_mdsc->fsc->client->monc.auth,
 338                              s->s_authorizer);
 339                 kfree(s);
 340         }
 341 }
 342
 343 /*
 344  * called under mdsc->mutex
 345  */
 346 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
 347                                                    int mds)
 348 {
 349         struct ceph_mds_session *session;
 350
 351         if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
 352                 return NULL;
 353         session = mdsc->sessions[mds];
 354         dout("lookup_mds_session %p %d\n", session,
 355              atomic_read(&session->s_ref));
 356         get_session(session);
 357         return session;
 358 }
 359
 360 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
 361 {
 362         if (mds >= mdsc->max_sessions)
 363                 return false;
 364         return mdsc->sessions[mds];
 365 }
 366
 367 static int __verify_registered_session(struct ceph_mds_client *mdsc,
 368                                        struct ceph_mds_session *s)
 369 {
 370         if (s->s_mds >= mdsc->max_sessions ||
 371             mdsc->sessions[s->s_mds] != s)
 372                 return -ENOENT;
 373         return 0;
 374 }
 375
 376 /*
 377  * create+register a new session for given mds.
 378  * called under mdsc->mutex.
 379  */
 380 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 381                                                  int mds)
 382 {
 383         struct ceph_mds_session *s;
 384
 385         s = kzalloc(sizeof(*s), GFP_NOFS);
 386         if (!s)
 387                 return ERR_PTR(-ENOMEM);
 388         s->s_mdsc = mdsc;
 389         s->s_mds = mds;
 390         s->s_state = CEPH_MDS_SESSION_NEW;
 391         s->s_ttl = 0;
 392         s->s_seq = 0;
 393         mutex_init(&s->s_mutex);
 394
 395         ceph_con_init(mdsc->fsc->client->msgr, &s->s_con);
 396         s->s_con.private = s;
 397         s->s_con.ops = &mds_con_ops;
 398         s->s_con.peer_name.type = CEPH_ENTITY_TYPE_MDS;
 399         s->s_con.peer_name.num = cpu_to_le64(mds);
 400
 401         spin_lock_init(&s->s_cap_lock);
 402         s->s_cap_gen = 0;
 403         s->s_cap_ttl = 0;
 404         s->s_renew_requested = 0;
 405         s->s_renew_seq = 0;
 406         INIT_LIST_HEAD(&s->s_caps);
 407         s->s_nr_caps = 0;
 408         s->s_trim_caps = 0;
 409         atomic_set(&s->s_ref, 1);
 410         INIT_LIST_HEAD(&s->s_waiting);
 411         INIT_LIST_HEAD(&s->s_unsafe);
 412         s->s_num_cap_releases = 0;
 413         s->s_cap_iterator = NULL;
 414         INIT_LIST_HEAD(&s->s_cap_releases);
 415         INIT_LIST_HEAD(&s->s_cap_releases_done);
 416         INIT_LIST_HEAD(&s->s_cap_flushing);
 417         INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
 418
 419         dout("register_session mds%d\n", mds);
 420         if (mds >= mdsc->max_sessions) {
 421                 int newmax = 1 << get_count_order(mds+1);
 422                 struct ceph_mds_session **sa;
 423
 424                 dout("register_session realloc to %d\n", newmax);
 425                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 426                 if (sa == NULL)
 427                         goto fail_realloc;
 428                 if (mdsc->sessions) {
 429                         memcpy(sa, mdsc->sessions,
 430                                mdsc->max_sessions * sizeof(void *));
 431                         kfree(mdsc->sessions);
 432                 }
 433                 mdsc->sessions = sa;
 434                 mdsc->max_sessions = newmax;
 435         }
 436         mdsc->sessions[mds] = s;
 437         atomic_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
 438
 439         ceph_con_open(&s->s_con, ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
 440
 441         return s;
 442
 443 fail_realloc:
 444         kfree(s);
 445         return ERR_PTR(-ENOMEM);
 446 }
 447
 448 /*
 449  * called under mdsc->mutex
 450  */
 451 static void __unregister_session(struct ceph_mds_client *mdsc,
 452                                struct ceph_mds_session *s)
 453 {
 454         dout("__unregister_session mds%d %p\n", s->s_mds, s);
 455         BUG_ON(mdsc->sessions[s->s_mds] != s);
 456         mdsc->sessions[s->s_mds] = NULL;
 457         ceph_con_close(&s->s_con);
 458         ceph_put_mds_session(s);
 459 }
 460
 461 /*
 462  * drop session refs in request.
 463  *
 464  * should be last request ref, or hold mdsc->mutex
 465  */
 466 static void put_request_session(struct ceph_mds_request *req)
 467 {
 468         if (req->r_session) {
 469                 ceph_put_mds_session(req->r_session);
 470                 req->r_session = NULL;
 471         }
 472 }
 473
 474 void ceph_mdsc_release_request(struct kref *kref)
 475 {
 476         struct ceph_mds_request *req = container_of(kref,
 477                                                     struct ceph_mds_request,
 478                                                     r_kref);
 479         if (req->r_request)
 480                 ceph_msg_put(req->r_request);
 481         if (req->r_reply) {
 482                 ceph_msg_put(req->r_reply);
 483                 destroy_reply_info(&req->r_reply_info);
 484         }
 485         if (req->r_inode) {
 486                 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
 487                 iput(req->r_inode);
 488         }
 489         if (req->r_locked_dir)
 490                 ceph_put_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
 491         if (req->r_target_inode)
 492                 iput(req->r_target_inode);
 493         if (req->r_dentry)
 494                 dput(req->r_dentry);
 495         if (req->r_old_dentry) {
 496                 /*
 497                  * track (and drop pins for) r_old_dentry_dir
 498                  * separately, since r_old_dentry's d_parent may have
 499                  * changed between the dir mutex being dropped and
 500                  * this request being freed.
 501                  */
 502                 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
 503                                   CEPH_CAP_PIN);
 504                 dput(req->r_old_dentry);
 505                 iput(req->r_old_dentry_dir);
 506         }
 507         kfree(req->r_path1);
 508         kfree(req->r_path2);
 509         put_request_session(req);
 510         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
 511         kfree(req);
 512 }
 513
 514 /*
 515  * lookup session, bump ref if found.
 516  *
 517  * called under mdsc->mutex.
 518  */
 519 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
 520                                              u64 tid)
 521 {
 522         struct ceph_mds_request *req;
 523         struct rb_node *n = mdsc->request_tree.rb_node;
 524
 525         while (n) {
 526                 req = rb_entry(n, struct ceph_mds_request, r_node);
 527                 if (tid < req->r_tid)
 528                         n = n->rb_left;
 529                 else if (tid > req->r_tid)
 530                         n = n->rb_right;
 531                 else {
 532                         ceph_mdsc_get_request(req);
 533                         return req;
 534                 }
 535         }
 536         return NULL;
 537 }
 538
 539 static void __insert_request(struct ceph_mds_client *mdsc,
 540                              struct ceph_mds_request *new)
 541 {
 542         struct rb_node **p = &mdsc->request_tree.rb_node;
 543         struct rb_node *parent = NULL;
 544         struct ceph_mds_request *req = NULL;
 545
 546         while (*p) {
 547                 parent = *p;
 548                 req = rb_entry(parent, struct ceph_mds_request, r_node);
 549                 if (new->r_tid < req->r_tid)
 550                         p = &(*p)->rb_left;
 551                 else if (new->r_tid > req->r_tid)
 552                         p = &(*p)->rb_right;
 553                 else
 554                         BUG();
 555         }
 556
 557         rb_link_node(&new->r_node, parent, p);
 558         rb_insert_color(&new->r_node, &mdsc->request_tree);
 559 }
 560
 561 /*
 562  * Register an in-flight request, and assign a tid.  Link to directory
 563  * are modifying (if any).
 564  *
 565  * Called under mdsc->mutex.
 566  */
 567 static void __register_request(struct ceph_mds_client *mdsc,
 568                                struct ceph_mds_request *req,
 569                                struct inode *dir)
 570 {
 571         req->r_tid = ++mdsc->last_tid;
 572         if (req->r_num_caps)
 573                 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
 574                                   req->r_num_caps);
 575         dout("__register_request %p tid %lld\n", req, req->r_tid);
 576         ceph_mdsc_get_request(req);
 577         __insert_request(mdsc, req);
 578
 579         req->r_uid = current_fsuid();
 580         req->r_gid = current_fsgid();
 581
 582         if (dir) {
 583                 struct ceph_inode_info *ci = ceph_inode(dir);
 584
 585                 ihold(dir);
 586                 spin_lock(&ci->i_unsafe_lock);
 587                 req->r_unsafe_dir = dir;
 588                 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
 589                 spin_unlock(&ci->i_unsafe_lock);
 590         }
 591 }
 592
 593 static void __unregister_request(struct ceph_mds_client *mdsc,
 594                                  struct ceph_mds_request *req)
 595 {
 596         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 597         rb_erase(&req->r_node, &mdsc->request_tree);
 598         RB_CLEAR_NODE(&req->r_node);
 599
 600         if (req->r_unsafe_dir) {
 601                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 602
 603                 spin_lock(&ci->i_unsafe_lock);
 604                 list_del_init(&req->r_unsafe_dir_item);
 605                 spin_unlock(&ci->i_unsafe_lock);
 606
 607                 iput(req->r_unsafe_dir);
 608                 req->r_unsafe_dir = NULL;
 609         }
 610
 611         ceph_mdsc_put_request(req);
 612 }
 613
 614 /*
 615  * Choose mds to send request to next.  If there is a hint set in the
 616  * request (e.g., due to a prior forward hint from the mds), use that.
 617  * Otherwise, consult frag tree and/or caps to identify the
 618  * appropriate mds.  If all else fails, choose randomly.
 619  *
 620  * Called under mdsc->mutex.
 621  */
 622 static struct dentry *get_nonsnap_parent(struct dentry *dentry)
 623 {
 624         /*
 625          * we don't need to worry about protecting the d_parent access
 626          * here because we never renaming inside the snapped namespace
 627          * except to resplice to another snapdir, and either the old or new
 628          * result is a valid result.
 629          */
 630         while (!IS_ROOT(dentry) && ceph_snap(dentry->d_inode) != CEPH_NOSNAP)
 631                 dentry = dentry->d_parent;
 632         return dentry;
 633 }
 634
 635 static int __choose_mds(struct ceph_mds_client *mdsc,
 636                         struct ceph_mds_request *req)
 637 {
 638         struct inode *inode;
 639         struct ceph_inode_info *ci;
 640         struct ceph_cap *cap;
 641         int mode = req->r_direct_mode;
 642         int mds = -1;
 643         u32 hash = req->r_direct_hash;
 644         bool is_hash = req->r_direct_is_hash;
 645
 646         /*
 647          * is there a specific mds we should try?  ignore hint if we have
 648          * no session and the mds is not up (active or recovering).
 649          */
 650         if (req->r_resend_mds >= 0 &&
 651             (__have_session(mdsc, req->r_resend_mds) ||
 652              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 653                 dout("choose_mds using resend_mds mds%d\n",
 654                      req->r_resend_mds);
 655                 return req->r_resend_mds;
 656         }
 657
 658         if (mode == USE_RANDOM_MDS)
 659                 goto random;
 660
 661         inode = NULL;
 662         if (req->r_inode) {
 663                 inode = req->r_inode;
 664         } else if (req->r_dentry) {
 665                 /* ignore race with rename; old or new d_parent is okay */
 666                 struct dentry *parent = req->r_dentry->d_parent;
 667                 struct inode *dir = parent->d_inode;
 668
 669                 if (dir->i_sb != mdsc->fsc->sb) {
 670                         /* not this fs! */
 671                         inode = req->r_dentry->d_inode;
 672                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
 673                         /* direct snapped/virtual snapdir requests
 674                          * based on parent dir inode */
 675                         struct dentry *dn = get_nonsnap_parent(parent);
 676                         inode = dn->d_inode;
 677                         dout("__choose_mds using nonsnap parent %p\n", inode);
 678                 } else if (req->r_dentry->d_inode) {
 679                         /* dentry target */
 680                         inode = req->r_dentry->d_inode;
 681                 } else {
 682                         /* dir + name */
 683                         inode = dir;
 684                         hash = ceph_dentry_hash(dir, req->r_dentry);
 685                         is_hash = true;
 686                 }
 687         }
 688
 689         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 690              (int)hash, mode);
 691         if (!inode)
 692                 goto random;
 693         ci = ceph_inode(inode);
 694
 695         if (is_hash && S_ISDIR(inode->i_mode)) {
 696                 struct ceph_inode_frag frag;
 697                 int found;
 698
 699                 ceph_choose_frag(ci, hash, &frag, &found);
 700                 if (found) {
 701                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 702                                 u8 r;
 703
 704                                 /* choose a random replica */
 705                                 get_random_bytes(&r, 1);
 706                                 r %= frag.ndist;
 707                                 mds = frag.dist[r];
 708                                 dout("choose_mds %p %llx.%llx "
 709                                      "frag %u mds%d (%d/%d)\n",
 710                                      inode, ceph_vinop(inode),
 711                                      frag.frag, mds,
 712                                      (int)r, frag.ndist);
 713                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 714                                     CEPH_MDS_STATE_ACTIVE)
 715                                         return mds;
 716                         }
 717
 718                         /* since this file/dir wasn't known to be
 719                          * replicated, then we want to look for the
 720                          * authoritative mds. */
 721                         mode = USE_AUTH_MDS;
 722                         if (frag.mds >= 0) {
 723                                 /* choose auth mds */
 724                                 mds = frag.mds;
 725                                 dout("choose_mds %p %llx.%llx "
 726                                      "frag %u mds%d (auth)\n",
 727                                      inode, ceph_vinop(inode), frag.frag, mds);
 728                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 729                                     CEPH_MDS_STATE_ACTIVE)
 730                                         return mds;
 731                         }
 732                 }
 733         }
 734
 735         spin_lock(&ci->i_ceph_lock);
 736         cap = NULL;
 737         if (mode == USE_AUTH_MDS)
 738                 cap = ci->i_auth_cap;
 739         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
 740                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
 741         if (!cap) {
 742                 spin_unlock(&ci->i_ceph_lock);
 743                 goto random;
 744         }
 745         mds = cap->session->s_mds;
 746         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
 747              inode, ceph_vinop(inode), mds,
 748              cap == ci->i_auth_cap ? "auth " : "", cap);
 749         spin_unlock(&ci->i_ceph_lock);
 750         return mds;
 751
 752 random:
 753         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
 754         dout("choose_mds chose random mds%d\n", mds);
 755         return mds;
 756 }
 757
 758
 759 /*
 760  * session messages
 761  */
 762 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
 763 {
 764         struct ceph_msg *msg;
 765         struct ceph_mds_session_head *h;
 766
 767         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
 768                            false);
 769         if (!msg) {
 770                 pr_err("create_session_msg ENOMEM creating msg\n");
 771                 return NULL;
 772         }
 773         h = msg->front.iov_base;
 774         h->op = cpu_to_le32(op);
 775         h->seq = cpu_to_le64(seq);
 776         return msg;
 777 }
 778
 779 /*
 780  * send session open request.
 781  *
 782  * called under mdsc->mutex
 783  */
 784 static int __open_session(struct ceph_mds_client *mdsc,
 785                           struct ceph_mds_session *session)
 786 {
 787         struct ceph_msg *msg;
 788         int mstate;
 789         int mds = session->s_mds;
 790
 791         /* wait for mds to go active? */
 792         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
 793         dout("open_session to mds%d (%s)\n", mds,
 794              ceph_mds_state_name(mstate));
 795         session->s_state = CEPH_MDS_SESSION_OPENING;
 796         session->s_renew_requested = jiffies;
 797
 798         /* send connect message */
 799         msg = create_session_msg(CEPH_SESSION_REQUEST_OPEN, session->s_seq);
 800         if (!msg)
 801                 return -ENOMEM;
 802         ceph_con_send(&session->s_con, msg);
 803         return 0;
 804 }
 805
 806 /*
 807  * open sessions for any export targets for the given mds
 808  *
 809  * called under mdsc->mutex
 810  */
 811 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
 812                                           struct ceph_mds_session *session)
 813 {
 814         struct ceph_mds_info *mi;
 815         struct ceph_mds_session *ts;
 816         int i, mds = session->s_mds;
 817         int target;
 818
 819         if (mds >= mdsc->mdsmap->m_max_mds)
 820                 return;
 821         mi = &mdsc->mdsmap->m_info[mds];
 822         dout("open_export_target_sessions for mds%d (%d targets)\n",
 823              session->s_mds, mi->num_export_targets);
 824
 825         for (i = 0; i < mi->num_export_targets; i++) {
 826                 target = mi->export_targets[i];
 827                 ts = __ceph_lookup_mds_session(mdsc, target);
 828                 if (!ts) {
 829                         ts = register_session(mdsc, target);
 830                         if (IS_ERR(ts))
 831                                 return;
 832                 }
 833                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
 834                     session->s_state == CEPH_MDS_SESSION_CLOSING)
 835                         __open_session(mdsc, session);
 836                 else
 837                         dout(" mds%d target mds%d %p is %s\n", session->s_mds,
 838                              i, ts, session_state_name(ts->s_state));
 839                 ceph_put_mds_session(ts);
 840         }
 841 }
 842
 843 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
 844                                            struct ceph_mds_session *session)
 845 {
 846         mutex_lock(&mdsc->mutex);
 847         __open_export_target_sessions(mdsc, session);
 848         mutex_unlock(&mdsc->mutex);
 849 }
 850
 851 /*
 852  * session caps
 853  */
 854
 855 /*
 856  * Free preallocated cap messages assigned to this session
 857  */
 858 static void cleanup_cap_releases(struct ceph_mds_session *session)
 859 {
 860         struct ceph_msg *msg;
 861
 862         spin_lock(&session->s_cap_lock);
 863         while (!list_empty(&session->s_cap_releases)) {
 864                 msg = list_first_entry(&session->s_cap_releases,
 865                                        struct ceph_msg, list_head);
 866                 list_del_init(&msg->list_head);
 867                 ceph_msg_put(msg);
 868         }
 869         while (!list_empty(&session->s_cap_releases_done)) {
 870                 msg = list_first_entry(&session->s_cap_releases_done,
 871                                        struct ceph_msg, list_head);
 872                 list_del_init(&msg->list_head);
 873                 ceph_msg_put(msg);
 874         }
 875         spin_unlock(&session->s_cap_lock);
 876 }
 877
 878 /*
 879  * Helper to safely iterate over all caps associated with a session, with
 880  * special care taken to handle a racing __ceph_remove_cap().
 881  *
 882  * Caller must hold session s_mutex.
 883  */
 884 static int iterate_session_caps(struct ceph_mds_session *session,
 885                                  int (*cb)(struct inode *, struct ceph_cap *,
 886                                             void *), void *arg)
 887 {
 888         struct list_head *p;
 889         struct ceph_cap *cap;
 890         struct inode *inode, *last_inode = NULL;
 891         struct ceph_cap *old_cap = NULL;
 892         int ret;
 893
 894         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
 895         spin_lock(&session->s_cap_lock);
 896         p = session->s_caps.next;
 897         while (p != &session->s_caps) {
 898                 cap = list_entry(p, struct ceph_cap, session_caps);
 899                 inode = igrab(&cap->ci->vfs_inode);
 900                 if (!inode) {
 901                         p = p->next;
 902                         continue;
 903                 }
 904                 session->s_cap_iterator = cap;
 905                 spin_unlock(&session->s_cap_lock);
 906
 907                 if (last_inode) {
 908                         iput(last_inode);
 909                         last_inode = NULL;
 910                 }
 911                 if (old_cap) {
 912                         ceph_put_cap(session->s_mdsc, old_cap);
 913                         old_cap = NULL;
 914                 }
 915
 916                 ret = cb(inode, cap, arg);
 917                 last_inode = inode;
 918
 919                 spin_lock(&session->s_cap_lock);
 920                 p = p->next;
 921                 if (cap->ci == NULL) {
 922                         dout("iterate_session_caps  finishing cap %p removal\n",
 923                              cap);
 924                         BUG_ON(cap->session != session);
 925                         list_del_init(&cap->session_caps);
 926                         session->s_nr_caps--;
 927                         cap->session = NULL;
 928                         old_cap = cap;  /* put_cap it w/o locks held */
 929                 }
 930                 if (ret < 0)
 931                         goto out;
 932         }
 933         ret = 0;
 934 out:
 935         session->s_cap_iterator = NULL;
 936         spin_unlock(&session->s_cap_lock);
 937
 938         if (last_inode)
 939                 iput(last_inode);
 940         if (old_cap)
 941                 ceph_put_cap(session->s_mdsc, old_cap);
 942
 943         return ret;
 944 }
 945
 946 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
 947                                   void *arg)
 948 {
 949         struct ceph_inode_info *ci = ceph_inode(inode);
 950         int drop = 0;
 951
 952         dout("removing cap %p, ci is %p, inode is %p\n",
 953              cap, ci, &ci->vfs_inode);
 954         spin_lock(&ci->i_ceph_lock);
 955         __ceph_remove_cap(cap);
 956         if (!__ceph_is_any_real_caps(ci)) {
 957                 struct ceph_mds_client *mdsc =
 958                         ceph_sb_to_client(inode->i_sb)->mdsc;
 959
 960                 spin_lock(&mdsc->cap_dirty_lock);
 961                 if (!list_empty(&ci->i_dirty_item)) {
 962                         pr_info(" dropping dirty %s state for %p %lld\n",
 963                                 ceph_cap_string(ci->i_dirty_caps),
 964                                 inode, ceph_ino(inode));
 965                         ci->i_dirty_caps = 0;
 966                         list_del_init(&ci->i_dirty_item);
 967                         drop = 1;
 968                 }
 969                 if (!list_empty(&ci->i_flushing_item)) {
 970                         pr_info(" dropping dirty+flushing %s state for %p %lld\n",
 971                                 ceph_cap_string(ci->i_flushing_caps),
 972                                 inode, ceph_ino(inode));
 973                         ci->i_flushing_caps = 0;
 974                         list_del_init(&ci->i_flushing_item);
 975                         mdsc->num_cap_flushing--;
 976                         drop = 1;
 977                 }
 978                 if (drop && ci->i_wrbuffer_ref) {
 979                         pr_info(" dropping dirty data for %p %lld\n",
 980                                 inode, ceph_ino(inode));
 981                         ci->i_wrbuffer_ref = 0;
 982                         ci->i_wrbuffer_ref_head = 0;
 983                         drop++;
 984                 }
 985                 spin_unlock(&mdsc->cap_dirty_lock);
 986         }
 987         spin_unlock(&ci->i_ceph_lock);
 988         while (drop--)
 989                 iput(inode);
 990         return 0;
 991 }
 992
 993 /*
 994  * caller must hold session s_mutex
 995  */
 996 static void remove_session_caps(struct ceph_mds_session *session)
 997 {
 998         dout("remove_session_caps on %p\n", session);
 999         iterate_session_caps(session, remove_session_caps_cb, NULL);
1000         BUG_ON(session->s_nr_caps > 0);
1001         BUG_ON(!list_empty(&session->s_cap_flushing));
1002         cleanup_cap_releases(session);
1003 }
1004
1005 /*
1006  * wake up any threads waiting on this session's caps.  if the cap is
1007  * old (didn't get renewed on the client reconnect), remove it now.
1008  *
1009  * caller must hold s_mutex.
1010  */
1011 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1012                               void *arg)
1013 {
1014         struct ceph_inode_info *ci = ceph_inode(inode);
1015
1016         wake_up_all(&ci->i_cap_wq);
1017         if (arg) {
1018                 spin_lock(&ci->i_ceph_lock);
1019                 ci->i_wanted_max_size = 0;
1020                 ci->i_requested_max_size = 0;
1021                 spin_unlock(&ci->i_ceph_lock);
1022         }
1023         return 0;
1024 }
1025
1026 static void wake_up_session_caps(struct ceph_mds_session *session,
1027                                  int reconnect)
1028 {
1029         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1030         iterate_session_caps(session, wake_up_session_cb,
1031                              (void *)(unsigned long)reconnect);
1032 }
1033
1034 /*
1035  * Send periodic message to MDS renewing all currently held caps.  The
1036  * ack will reset the expiration for all caps from this session.
1037  *
1038  * caller holds s_mutex
1039  */
1040 static int send_renew_caps(struct ceph_mds_client *mdsc,
1041                            struct ceph_mds_session *session)
1042 {
1043         struct ceph_msg *msg;
1044         int state;
1045
1046         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1047             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1048                 pr_info("mds%d caps stale\n", session->s_mds);
1049         session->s_renew_requested = jiffies;
1050
1051         /* do not try to renew caps until a recovering mds has reconnected
1052          * with its clients. */
1053         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1054         if (state < CEPH_MDS_STATE_RECONNECT) {
1055                 dout("send_renew_caps ignoring mds%d (%s)\n",
1056                      session->s_mds, ceph_mds_state_name(state));
1057                 return 0;
1058         }
1059
1060         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1061                 ceph_mds_state_name(state));
1062         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1063                                  ++session->s_renew_seq);
1064         if (!msg)
1065                 return -ENOMEM;
1066         ceph_con_send(&session->s_con, msg);
1067         return 0;
1068 }
1069
1070 /*
1071  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1072  *
1073  * Called under session->s_mutex
1074  */
1075 static void renewed_caps(struct ceph_mds_client *mdsc,
1076                          struct ceph_mds_session *session, int is_renew)
1077 {
1078         int was_stale;
1079         int wake = 0;
1080
1081         spin_lock(&session->s_cap_lock);
1082         was_stale = is_renew && (session->s_cap_ttl == 0 ||
1083                                  time_after_eq(jiffies, session->s_cap_ttl));
1084
1085         session->s_cap_ttl = session->s_renew_requested +
1086                 mdsc->mdsmap->m_session_timeout*HZ;
1087
1088         if (was_stale) {
1089                 if (time_before(jiffies, session->s_cap_ttl)) {
1090                         pr_info("mds%d caps renewed\n", session->s_mds);
1091                         wake = 1;
1092                 } else {
1093                         pr_info("mds%d caps still stale\n", session->s_mds);
1094                 }
1095         }
1096         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1097              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1098              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1099         spin_unlock(&session->s_cap_lock);
1100
1101         if (wake)
1102                 wake_up_session_caps(session, 0);
1103 }
1104
1105 /*
1106  * send a session close request
1107  */
1108 static int request_close_session(struct ceph_mds_client *mdsc,
1109                                  struct ceph_mds_session *session)
1110 {
1111         struct ceph_msg *msg;
1112
1113         dout("request_close_session mds%d state %s seq %lld\n",
1114              session->s_mds, session_state_name(session->s_state),
1115              session->s_seq);
1116         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1117         if (!msg)
1118                 return -ENOMEM;
1119         ceph_con_send(&session->s_con, msg);
1120         return 0;
1121 }
1122
1123 /*
1124  * Called with s_mutex held.
1125  */
1126 static int __close_session(struct ceph_mds_client *mdsc,
1127                          struct ceph_mds_session *session)
1128 {
1129         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1130                 return 0;
1131         session->s_state = CEPH_MDS_SESSION_CLOSING;
1132         return request_close_session(mdsc, session);
1133 }
1134
1135 /*
1136  * Trim old(er) caps.
1137  *
1138  * Because we can't cache an inode without one or more caps, we do
1139  * this indirectly: if a cap is unused, we prune its aliases, at which
1140  * point the inode will hopefully get dropped to.
1141  *
1142  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1143  * memory pressure from the MDS, though, so it needn't be perfect.
1144  */
1145 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1146 {
1147         struct ceph_mds_session *session = arg;
1148         struct ceph_inode_info *ci = ceph_inode(inode);
1149         int used, oissued, mine;
1150
1151         if (session->s_trim_caps <= 0)
1152                 return -1;
1153
1154         spin_lock(&ci->i_ceph_lock);
1155         mine = cap->issued | cap->implemented;
1156         used = __ceph_caps_used(ci);
1157         oissued = __ceph_caps_issued_other(ci, cap);
1158
1159         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
1160              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1161              ceph_cap_string(used));
1162         if (ci->i_dirty_caps)
1163                 goto out;   /* dirty caps */
1164         if ((used & ~oissued) & mine)
1165                 goto out;   /* we need these caps */
1166
1167         session->s_trim_caps--;
1168         if (oissued) {
1169                 /* we aren't the only cap.. just remove us */
1170                 __ceph_remove_cap(cap);
1171         } else {
1172                 /* try to drop referring dentries */
1173                 spin_unlock(&ci->i_ceph_lock);
1174                 d_prune_aliases(inode);
1175                 dout("trim_caps_cb %p cap %p  pruned, count now %d\n",
1176                      inode, cap, atomic_read(&inode->i_count));
1177                 return 0;
1178         }
1179
1180 out:
1181         spin_unlock(&ci->i_ceph_lock);
1182         return 0;
1183 }
1184
1185 /*
1186  * Trim session cap count down to some max number.
1187  */
1188 static int trim_caps(struct ceph_mds_client *mdsc,
1189                      struct ceph_mds_session *session,
1190                      int max_caps)
1191 {
1192         int trim_caps = session->s_nr_caps - max_caps;
1193
1194         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1195              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1196         if (trim_caps > 0) {
1197                 session->s_trim_caps = trim_caps;
1198                 iterate_session_caps(session, trim_caps_cb, session);
1199                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1200                      session->s_mds, session->s_nr_caps, max_caps,
1201                         trim_caps - session->s_trim_caps);
1202                 session->s_trim_caps = 0;
1203         }
1204         return 0;
1205 }
1206
1207 /*
1208  * Allocate cap_release messages.  If there is a partially full message
1209  * in the queue, try to allocate enough to cover it's remainder, so that
1210  * we can send it immediately.
1211  *
1212  * Called under s_mutex.
1213  */
1214 int ceph_add_cap_releases(struct ceph_mds_client *mdsc,
1215                           struct ceph_mds_session *session)
1216 {
1217         struct ceph_msg *msg, *partial = NULL;
1218         struct ceph_mds_cap_release *head;
1219         int err = -ENOMEM;
1220         int extra = mdsc->fsc->mount_options->cap_release_safety;
1221         int num;
1222
1223         dout("add_cap_releases %p mds%d extra %d\n", session, session->s_mds,
1224              extra);
1225
1226         spin_lock(&session->s_cap_lock);
1227
1228         if (!list_empty(&session->s_cap_releases)) {
1229                 msg = list_first_entry(&session->s_cap_releases,
1230                                        struct ceph_msg,
1231                                  list_head);
1232                 head = msg->front.iov_base;
1233                 num = le32_to_cpu(head->num);
1234                 if (num) {
1235                         dout(" partial %p with (%d/%d)\n", msg, num,
1236                              (int)CEPH_CAPS_PER_RELEASE);
1237                         extra += CEPH_CAPS_PER_RELEASE - num;
1238                         partial = msg;
1239                 }
1240         }
1241         while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1242                 spin_unlock(&session->s_cap_lock);
1243                 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
1244                                    GFP_NOFS, false);
1245                 if (!msg)
1246                         goto out_unlocked;
1247                 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1248                      (int)msg->front.iov_len);
1249                 head = msg->front.iov_base;
1250                 head->num = cpu_to_le32(0);
1251                 msg->front.iov_len = sizeof(*head);
1252                 spin_lock(&session->s_cap_lock);
1253                 list_add(&msg->list_head, &session->s_cap_releases);
1254                 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1255         }
1256
1257         if (partial) {
1258                 head = partial->front.iov_base;
1259                 num = le32_to_cpu(head->num);
1260                 dout(" queueing partial %p with %d/%d\n", partial, num,
1261                      (int)CEPH_CAPS_PER_RELEASE);
1262                 list_move_tail(&partial->list_head,
1263                                &session->s_cap_releases_done);
1264                 session->s_num_cap_releases -= CEPH_CAPS_PER_RELEASE - num;
1265         }
1266         err = 0;
1267         spin_unlock(&session->s_cap_lock);
1268 out_unlocked:
1269         return err;
1270 }
1271
1272 /*
1273  * flush all dirty inode data to disk.
1274  *
1275  * returns true if we've flushed through want_flush_seq
1276  */
1277 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1278 {
1279         int mds, ret = 1;
1280
1281         dout("check_cap_flush want %lld\n", want_flush_seq);
1282         mutex_lock(&mdsc->mutex);
1283         for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1284                 struct ceph_mds_session *session = mdsc->sessions[mds];
1285
1286                 if (!session)
1287                         continue;
1288                 get_session(session);
1289                 mutex_unlock(&mdsc->mutex);
1290
1291                 mutex_lock(&session->s_mutex);
1292                 if (!list_empty(&session->s_cap_flushing)) {
1293                         struct ceph_inode_info *ci =
1294                                 list_entry(session->s_cap_flushing.next,
1295                                            struct ceph_inode_info,
1296                                            i_flushing_item);
1297                         struct inode *inode = &ci->vfs_inode;
1298
1299                         spin_lock(&ci->i_ceph_lock);
1300                         if (ci->i_cap_flush_seq <= want_flush_seq) {
1301                                 dout("check_cap_flush still flushing %p "
1302                                      "seq %lld <= %lld to mds%d\n", inode,
1303                                      ci->i_cap_flush_seq, want_flush_seq,
1304                                      session->s_mds);
1305                                 ret = 0;
1306                         }
1307                         spin_unlock(&ci->i_ceph_lock);
1308                 }
1309                 mutex_unlock(&session->s_mutex);
1310                 ceph_put_mds_session(session);
1311
1312                 if (!ret)
1313                         return ret;
1314                 mutex_lock(&mdsc->mutex);
1315         }
1316
1317         mutex_unlock(&mdsc->mutex);
1318         dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1319         return ret;
1320 }
1321
1322 /*
1323  * called under s_mutex
1324  */
1325 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1326                             struct ceph_mds_session *session)
1327 {
1328         struct ceph_msg *msg;
1329
1330         dout("send_cap_releases mds%d\n", session->s_mds);
1331         spin_lock(&session->s_cap_lock);
1332         while (!list_empty(&session->s_cap_releases_done)) {
1333                 msg = list_first_entry(&session->s_cap_releases_done,
1334                                  struct ceph_msg, list_head);
1335                 list_del_init(&msg->list_head);
1336                 spin_unlock(&session->s_cap_lock);
1337                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1338                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1339                 ceph_con_send(&session->s_con, msg);
1340                 spin_lock(&session->s_cap_lock);
1341         }
1342         spin_unlock(&session->s_cap_lock);
1343 }
1344
1345 static void discard_cap_releases(struct ceph_mds_client *mdsc,
1346                                  struct ceph_mds_session *session)
1347 {
1348         struct ceph_msg *msg;
1349         struct ceph_mds_cap_release *head;
1350         unsigned num;
1351
1352         dout("discard_cap_releases mds%d\n", session->s_mds);
1353         spin_lock(&session->s_cap_lock);
1354
1355         /* zero out the in-progress message */
1356         msg = list_first_entry(&session->s_cap_releases,
1357                                struct ceph_msg, list_head);
1358         head = msg->front.iov_base;
1359         num = le32_to_cpu(head->num);
1360         dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg, num);
1361         head->num = cpu_to_le32(0);
1362         session->s_num_cap_releases += num;
1363
1364         /* requeue completed messages */
1365         while (!list_empty(&session->s_cap_releases_done)) {
1366                 msg = list_first_entry(&session->s_cap_releases_done,
1367                                  struct ceph_msg, list_head);
1368                 list_del_init(&msg->list_head);
1369
1370                 head = msg->front.iov_base;
1371                 num = le32_to_cpu(head->num);
1372                 dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg,
1373                      num);
1374                 session->s_num_cap_releases += num;
1375                 head->num = cpu_to_le32(0);
1376                 msg->front.iov_len = sizeof(*head);
1377                 list_add(&msg->list_head, &session->s_cap_releases);
1378         }
1379
1380         spin_unlock(&session->s_cap_lock);
1381 }
1382
1383 /*
1384  * requests
1385  */
1386
1387 /*
1388  * Create an mds request.
1389  */
1390 struct ceph_mds_request *
1391 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1392 {
1393         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1394
1395         if (!req)
1396                 return ERR_PTR(-ENOMEM);
1397
1398         mutex_init(&req->r_fill_mutex);
1399         req->r_mdsc = mdsc;
1400         req->r_started = jiffies;
1401         req->r_resend_mds = -1;
1402         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1403         req->r_fmode = -1;
1404         kref_init(&req->r_kref);
1405         INIT_LIST_HEAD(&req->r_wait);
1406         init_completion(&req->r_completion);
1407         init_completion(&req->r_safe_completion);
1408         INIT_LIST_HEAD(&req->r_unsafe_item);
1409
1410         req->r_op = op;
1411         req->r_direct_mode = mode;
1412         return req;
1413 }
1414
1415 /*
1416  * return oldest (lowest) request, tid in request tree, 0 if none.
1417  *
1418  * called under mdsc->mutex.
1419  */
1420 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1421 {
1422         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1423                 return NULL;
1424         return rb_entry(rb_first(&mdsc->request_tree),
1425                         struct ceph_mds_request, r_node);
1426 }
1427
1428 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1429 {
1430         struct ceph_mds_request *req = __get_oldest_req(mdsc);
1431
1432         if (req)
1433                 return req->r_tid;
1434         return 0;
1435 }
1436
1437 /*
1438  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1439  * on build_path_from_dentry in fs/cifs/dir.c.
1440  *
1441  * If @stop_on_nosnap, generate path relative to the first non-snapped
1442  * inode.
1443  *
1444  * Encode hidden .snap dirs as a double /, i.e.
1445  *   foo/.snap/bar -> foo//bar
1446  */
1447 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1448                            int stop_on_nosnap)
1449 {
1450         struct dentry *temp;
1451         char *path;
1452         int len, pos;
1453         unsigned seq;
1454
1455         if (dentry == NULL)
1456                 return ERR_PTR(-EINVAL);
1457
1458 retry:
1459         len = 0;
1460         seq = read_seqbegin(&rename_lock);
1461         rcu_read_lock();
1462         for (temp = dentry; !IS_ROOT(temp);) {
1463                 struct inode *inode = temp->d_inode;
1464                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1465                         len++;  /* slash only */
1466                 else if (stop_on_nosnap && inode &&
1467                          ceph_snap(inode) == CEPH_NOSNAP)
1468                         break;
1469                 else
1470                         len += 1 + temp->d_name.len;
1471                 temp = temp->d_parent;
1472                 if (temp == NULL) {
1473                         rcu_read_unlock();
1474                         pr_err("build_path corrupt dentry %p\n", dentry);
1475                         return ERR_PTR(-EINVAL);
1476                 }
1477         }
1478         rcu_read_unlock();
1479         if (len)
1480                 len--;  /* no leading '/' */
1481
1482         path = kmalloc(len+1, GFP_NOFS);
1483         if (path == NULL)
1484                 return ERR_PTR(-ENOMEM);
1485         pos = len;
1486         path[pos] = 0;  /* trailing null */
1487         rcu_read_lock();
1488         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1489                 struct inode *inode;
1490
1491                 spin_lock(&temp->d_lock);
1492                 inode = temp->d_inode;
1493                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1494                         dout("build_path path+%d: %p SNAPDIR\n",
1495                              pos, temp);
1496                 } else if (stop_on_nosnap && inode &&
1497                            ceph_snap(inode) == CEPH_NOSNAP) {
1498                         spin_unlock(&temp->d_lock);
1499                         break;
1500                 } else {
1501                         pos -= temp->d_name.len;
1502                         if (pos < 0) {
1503                                 spin_unlock(&temp->d_lock);
1504                                 break;
1505                         }
1506                         strncpy(path + pos, temp->d_name.name,
1507                                 temp->d_name.len);
1508                 }
1509                 spin_unlock(&temp->d_lock);
1510                 if (pos)
1511                         path[--pos] = '/';
1512                 temp = temp->d_parent;
1513                 if (temp == NULL) {
1514                         rcu_read_unlock();
1515                         pr_err("build_path corrupt dentry\n");
1516                         kfree(path);
1517                         return ERR_PTR(-EINVAL);
1518                 }
1519         }
1520         rcu_read_unlock();
1521         if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1522                 pr_err("build_path did not end path lookup where "
1523                        "expected, namelen is %d, pos is %d\n", len, pos);
1524                 /* presumably this is only possible if racing with a
1525                    rename of one of the parent directories (we can not
1526                    lock the dentries above us to prevent this, but
1527                    retrying should be harmless) */
1528                 kfree(path);
1529                 goto retry;
1530         }
1531
1532         *base = ceph_ino(temp->d_inode);
1533         *plen = len;
1534         dout("build_path on %p %d built %llx '%.*s'\n",
1535              dentry, dentry->d_count, *base, len, path);
1536         return path;
1537 }
1538
1539 static int build_dentry_path(struct dentry *dentry,
1540                              const char **ppath, int *ppathlen, u64 *pino,
1541                              int *pfreepath)
1542 {
1543         char *path;
1544
1545         if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1546                 *pino = ceph_ino(dentry->d_parent->d_inode);
1547                 *ppath = dentry->d_name.name;
1548                 *ppathlen = dentry->d_name.len;
1549                 return 0;
1550         }
1551         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1552         if (IS_ERR(path))
1553                 return PTR_ERR(path);
1554         *ppath = path;
1555         *pfreepath = 1;
1556         return 0;
1557 }
1558
1559 static int build_inode_path(struct inode *inode,
1560                             const char **ppath, int *ppathlen, u64 *pino,
1561                             int *pfreepath)
1562 {
1563         struct dentry *dentry;
1564         char *path;
1565
1566         if (ceph_snap(inode) == CEPH_NOSNAP) {
1567                 *pino = ceph_ino(inode);
1568                 *ppathlen = 0;
1569                 return 0;
1570         }
1571         dentry = d_find_alias(inode);
1572         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1573         dput(dentry);
1574         if (IS_ERR(path))
1575                 return PTR_ERR(path);
1576         *ppath = path;
1577         *pfreepath = 1;
1578         return 0;
1579 }
1580
1581 /*
1582  * request arguments may be specified via an inode *, a dentry *, or
1583  * an explicit ino+path.
1584  */
1585 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1586                                   const char *rpath, u64 rino,
1587                                   const char **ppath, int *pathlen,
1588                                   u64 *ino, int *freepath)
1589 {
1590         int r = 0;
1591
1592         if (rinode) {
1593                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1594                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1595                      ceph_snap(rinode));
1596         } else if (rdentry) {
1597                 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1598                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1599                      *ppath);
1600         } else if (rpath || rino) {
1601                 *ino = rino;
1602                 *ppath = rpath;
1603                 *pathlen = strlen(rpath);
1604                 dout(" path %.*s\n", *pathlen, rpath);
1605         }
1606
1607         return r;
1608 }
1609
1610 /*
1611  * called under mdsc->mutex
1612  */
1613 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1614                                                struct ceph_mds_request *req,
1615                                                int mds)
1616 {
1617         struct ceph_msg *msg;
1618         struct ceph_mds_request_head *head;
1619         const char *path1 = NULL;
1620         const char *path2 = NULL;
1621         u64 ino1 = 0, ino2 = 0;
1622         int pathlen1 = 0, pathlen2 = 0;
1623         int freepath1 = 0, freepath2 = 0;
1624         int len;
1625         u16 releases;
1626         void *p, *end;
1627         int ret;
1628
1629         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1630                               req->r_path1, req->r_ino1.ino,
1631                               &path1, &pathlen1, &ino1, &freepath1);
1632         if (ret < 0) {
1633                 msg = ERR_PTR(ret);
1634                 goto out;
1635         }
1636
1637         ret = set_request_path_attr(NULL, req->r_old_dentry,
1638                               req->r_path2, req->r_ino2.ino,
1639                               &path2, &pathlen2, &ino2, &freepath2);
1640         if (ret < 0) {
1641                 msg = ERR_PTR(ret);
1642                 goto out_free1;
1643         }
1644
1645         len = sizeof(*head) +
1646                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64));
1647
1648         /* calculate (max) length for cap releases */
1649         len += sizeof(struct ceph_mds_request_release) *
1650                 (!!req->r_inode_drop + !!req->r_dentry_drop +
1651                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1652         if (req->r_dentry_drop)
1653                 len += req->r_dentry->d_name.len;
1654         if (req->r_old_dentry_drop)
1655                 len += req->r_old_dentry->d_name.len;
1656
1657         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1658         if (!msg) {
1659                 msg = ERR_PTR(-ENOMEM);
1660                 goto out_free2;
1661         }
1662
1663         msg->hdr.tid = cpu_to_le64(req->r_tid);
1664
1665         head = msg->front.iov_base;
1666         p = msg->front.iov_base + sizeof(*head);
1667         end = msg->front.iov_base + msg->front.iov_len;
1668
1669         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1670         head->op = cpu_to_le32(req->r_op);
1671         head->caller_uid = cpu_to_le32(req->r_uid);
1672         head->caller_gid = cpu_to_le32(req->r_gid);
1673         head->args = req->r_args;
1674
1675         ceph_encode_filepath(&p, end, ino1, path1);
1676         ceph_encode_filepath(&p, end, ino2, path2);
1677
1678         /* make note of release offset, in case we need to replay */
1679         req->r_request_release_offset = p - msg->front.iov_base;
1680
1681         /* cap releases */
1682         releases = 0;
1683         if (req->r_inode_drop)
1684                 releases += ceph_encode_inode_release(&p,
1685                       req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1686                       mds, req->r_inode_drop, req->r_inode_unless, 0);
1687         if (req->r_dentry_drop)
1688                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1689                        mds, req->r_dentry_drop, req->r_dentry_unless);
1690         if (req->r_old_dentry_drop)
1691                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1692                        mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1693         if (req->r_old_inode_drop)
1694                 releases += ceph_encode_inode_release(&p,
1695                       req->r_old_dentry->d_inode,
1696                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1697         head->num_releases = cpu_to_le16(releases);
1698
1699         BUG_ON(p > end);
1700         msg->front.iov_len = p - msg->front.iov_base;
1701         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1702
1703         msg->pages = req->r_pages;
1704         msg->nr_pages = req->r_num_pages;
1705         msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1706         msg->hdr.data_off = cpu_to_le16(0);
1707
1708 out_free2:
1709         if (freepath2)
1710                 kfree((char *)path2);
1711 out_free1:
1712         if (freepath1)
1713                 kfree((char *)path1);
1714 out:
1715         return msg;
1716 }
1717
1718 /*
1719  * called under mdsc->mutex if error, under no mutex if
1720  * success.
1721  */
1722 static void complete_request(struct ceph_mds_client *mdsc,
1723                              struct ceph_mds_request *req)
1724 {
1725         if (req->r_callback)
1726                 req->r_callback(mdsc, req);
1727         else
1728                 complete_all(&req->r_completion);
1729 }
1730
1731 /*
1732  * called under mdsc->mutex
1733  */
1734 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1735                                   struct ceph_mds_request *req,
1736                                   int mds)
1737 {
1738         struct ceph_mds_request_head *rhead;
1739         struct ceph_msg *msg;
1740         int flags = 0;
1741
1742         req->r_attempts++;
1743         if (req->r_inode) {
1744                 struct ceph_cap *cap =
1745                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
1746
1747                 if (cap)
1748                         req->r_sent_on_mseq = cap->mseq;
1749                 else
1750                         req->r_sent_on_mseq = -1;
1751         }
1752         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1753              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1754
1755         if (req->r_got_unsafe) {
1756                 /*
1757                  * Replay.  Do not regenerate message (and rebuild
1758                  * paths, etc.); just use the original message.
1759                  * Rebuilding paths will break for renames because
1760                  * d_move mangles the src name.
1761                  */
1762                 msg = req->r_request;
1763                 rhead = msg->front.iov_base;
1764
1765                 flags = le32_to_cpu(rhead->flags);
1766                 flags |= CEPH_MDS_FLAG_REPLAY;
1767                 rhead->flags = cpu_to_le32(flags);
1768
1769                 if (req->r_target_inode)
1770                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1771
1772                 rhead->num_retry = req->r_attempts - 1;
1773
1774                 /* remove cap/dentry releases from message */
1775                 rhead->num_releases = 0;
1776                 msg->hdr.front_len = cpu_to_le32(req->r_request_release_offset);
1777                 msg->front.iov_len = req->r_request_release_offset;
1778                 return 0;
1779         }
1780
1781         if (req->r_request) {
1782                 ceph_msg_put(req->r_request);
1783                 req->r_request = NULL;
1784         }
1785         msg = create_request_message(mdsc, req, mds);
1786         if (IS_ERR(msg)) {
1787                 req->r_err = PTR_ERR(msg);
1788                 complete_request(mdsc, req);
1789                 return PTR_ERR(msg);
1790         }
1791         req->r_request = msg;
1792
1793         rhead = msg->front.iov_base;
1794         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1795         if (req->r_got_unsafe)
1796                 flags |= CEPH_MDS_FLAG_REPLAY;
1797         if (req->r_locked_dir)
1798                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1799         rhead->flags = cpu_to_le32(flags);
1800         rhead->num_fwd = req->r_num_fwd;
1801         rhead->num_retry = req->r_attempts - 1;
1802         rhead->ino = 0;
1803
1804         dout(" r_locked_dir = %p\n", req->r_locked_dir);
1805         return 0;
1806 }
1807
1808 /*
1809  * send request, or put it on the appropriate wait list.
1810  */
1811 static int __do_request(struct ceph_mds_client *mdsc,
1812                         struct ceph_mds_request *req)
1813 {
1814         struct ceph_mds_session *session = NULL;
1815         int mds = -1;
1816         int err = -EAGAIN;
1817
1818         if (req->r_err || req->r_got_result)
1819                 goto out;
1820
1821         if (req->r_timeout &&
1822             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1823                 dout("do_request timed out\n");
1824                 err = -EIO;
1825                 goto finish;
1826         }
1827
1828         put_request_session(req);
1829
1830         mds = __choose_mds(mdsc, req);
1831         if (mds < 0 ||
1832             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1833                 dout("do_request no mds or not active, waiting for map\n");
1834                 list_add(&req->r_wait, &mdsc->waiting_for_map);
1835                 goto out;
1836         }
1837
1838         /* get, open session */
1839         session = __ceph_lookup_mds_session(mdsc, mds);
1840         if (!session) {
1841                 session = register_session(mdsc, mds);
1842                 if (IS_ERR(session)) {
1843                         err = PTR_ERR(session);
1844                         goto finish;
1845                 }
1846         }
1847         req->r_session = get_session(session);
1848
1849         dout("do_request mds%d session %p state %s\n", mds, session,
1850              session_state_name(session->s_state));
1851         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1852             session->s_state != CEPH_MDS_SESSION_HUNG) {
1853                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1854                     session->s_state == CEPH_MDS_SESSION_CLOSING)
1855                         __open_session(mdsc, session);
1856                 list_add(&req->r_wait, &session->s_waiting);
1857                 goto out_session;
1858         }
1859
1860         /* send request */
1861         req->r_resend_mds = -1;   /* forget any previous mds hint */
1862
1863         if (req->r_request_started == 0)   /* note request start time */
1864                 req->r_request_started = jiffies;
1865
1866         err = __prepare_send_request(mdsc, req, mds);
1867         if (!err) {
1868                 ceph_msg_get(req->r_request);
1869                 ceph_con_send(&session->s_con, req->r_request);
1870         }
1871
1872 out_session:
1873         ceph_put_mds_session(session);
1874 out:
1875         return err;
1876
1877 finish:
1878         req->r_err = err;
1879         complete_request(mdsc, req);
1880         goto out;
1881 }
1882
1883 /*
1884  * called under mdsc->mutex
1885  */
1886 static void __wake_requests(struct ceph_mds_client *mdsc,
1887                             struct list_head *head)
1888 {
1889         struct ceph_mds_request *req, *nreq;
1890
1891         list_for_each_entry_safe(req, nreq, head, r_wait) {
1892                 list_del_init(&req->r_wait);
1893                 __do_request(mdsc, req);
1894         }
1895 }
1896
1897 /*
1898  * Wake up threads with requests pending for @mds, so that they can
1899  * resubmit their requests to a possibly different mds.
1900  */
1901 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
1902 {
1903         struct ceph_mds_request *req;
1904         struct rb_node *p;
1905
1906         dout("kick_requests mds%d\n", mds);
1907         for (p = rb_first(&mdsc->request_tree); p; p = rb_next(p)) {
1908                 req = rb_entry(p, struct ceph_mds_request, r_node);
1909                 if (req->r_got_unsafe)
1910                         continue;
1911                 if (req->r_session &&
1912                     req->r_session->s_mds == mds) {
1913                         dout(" kicking tid %llu\n", req->r_tid);
1914                         __do_request(mdsc, req);
1915                 }
1916         }
1917 }
1918
1919 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
1920                               struct ceph_mds_request *req)
1921 {
1922         dout("submit_request on %p\n", req);
1923         mutex_lock(&mdsc->mutex);
1924         __register_request(mdsc, req, NULL);
1925         __do_request(mdsc, req);
1926         mutex_unlock(&mdsc->mutex);
1927 }
1928
1929 /*
1930  * Synchrously perform an mds request.  Take care of all of the
1931  * session setup, forwarding, retry details.
1932  */
1933 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
1934                          struct inode *dir,
1935                          struct ceph_mds_request *req)
1936 {
1937         int err;
1938
1939         dout("do_request on %p\n", req);
1940
1941         /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
1942         if (req->r_inode)
1943                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
1944         if (req->r_locked_dir)
1945                 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
1946         if (req->r_old_dentry)
1947                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
1948                                   CEPH_CAP_PIN);
1949
1950         /* issue */
1951         mutex_lock(&mdsc->mutex);
1952         __register_request(mdsc, req, dir);
1953         __do_request(mdsc, req);
1954
1955         if (req->r_err) {
1956                 err = req->r_err;
1957                 __unregister_request(mdsc, req);
1958                 dout("do_request early error %d\n", err);
1959                 goto out;
1960         }
1961
1962         /* wait */
1963         mutex_unlock(&mdsc->mutex);
1964         dout("do_request waiting\n");
1965         if (req->r_timeout) {
1966                 err = (long)wait_for_completion_killable_timeout(
1967                         &req->r_completion, req->r_timeout);
1968                 if (err == 0)
1969                         err = -EIO;
1970         } else {
1971                 err = wait_for_completion_killable(&req->r_completion);
1972         }
1973         dout("do_request waited, got %d\n", err);
1974         mutex_lock(&mdsc->mutex);
1975
1976         /* only abort if we didn't race with a real reply */
1977         if (req->r_got_result) {
1978                 err = le32_to_cpu(req->r_reply_info.head->result);
1979         } else if (err < 0) {
1980                 dout("aborted request %lld with %d\n", req->r_tid, err);
1981
1982                 /*
1983                  * ensure we aren't running concurrently with
1984                  * ceph_fill_trace or ceph_readdir_prepopulate, which
1985                  * rely on locks (dir mutex) held by our caller.
1986                  */
1987                 mutex_lock(&req->r_fill_mutex);
1988                 req->r_err = err;
1989                 req->r_aborted = true;
1990                 mutex_unlock(&req->r_fill_mutex);
1991
1992                 if (req->r_locked_dir &&
1993                     (req->r_op & CEPH_MDS_OP_WRITE))
1994                         ceph_invalidate_dir_request(req);
1995         } else {
1996                 err = req->r_err;
1997         }
1998
1999 out:
2000         mutex_unlock(&mdsc->mutex);
2001         dout("do_request %p done, result %d\n", req, err);
2002         return err;
2003 }
2004
2005 /*
2006  * Invalidate dir D_COMPLETE, dentry lease state on an aborted MDS
2007  * namespace request.
2008  */
2009 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2010 {
2011         struct inode *inode = req->r_locked_dir;
2012         struct ceph_inode_info *ci = ceph_inode(inode);
2013
2014         dout("invalidate_dir_request %p (D_COMPLETE, lease(s))\n", inode);
2015         spin_lock(&ci->i_ceph_lock);
2016         ceph_dir_clear_complete(inode);
2017         ci->i_release_count++;
2018         spin_unlock(&ci->i_ceph_lock);
2019
2020         if (req->r_dentry)
2021                 ceph_invalidate_dentry_lease(req->r_dentry);
2022         if (req->r_old_dentry)
2023                 ceph_invalidate_dentry_lease(req->r_old_dentry);
2024 }
2025
2026 /*
2027  * Handle mds reply.
2028  *
2029  * We take the session mutex and parse and process the reply immediately.
2030  * This preserves the logical ordering of replies, capabilities, etc., sent
2031  * by the MDS as they are applied to our local cache.
2032  */
2033 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2034 {
2035         struct ceph_mds_client *mdsc = session->s_mdsc;
2036         struct ceph_mds_request *req;
2037         struct ceph_mds_reply_head *head = msg->front.iov_base;
2038         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2039         u64 tid;
2040         int err, result;
2041         int mds = session->s_mds;
2042
2043         if (msg->front.iov_len < sizeof(*head)) {
2044                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2045                 ceph_msg_dump(msg);
2046                 return;
2047         }
2048
2049         /* get request, session */
2050         tid = le64_to_cpu(msg->hdr.tid);
2051         mutex_lock(&mdsc->mutex);
2052         req = __lookup_request(mdsc, tid);
2053         if (!req) {
2054                 dout("handle_reply on unknown tid %llu\n", tid);
2055                 mutex_unlock(&mdsc->mutex);
2056                 return;
2057         }
2058         dout("handle_reply %p\n", req);
2059
2060         /* correct session? */
2061         if (req->r_session != session) {
2062                 pr_err("mdsc_handle_reply got %llu on session mds%d"
2063                        " not mds%d\n", tid, session->s_mds,
2064                        req->r_session ? req->r_session->s_mds : -1);
2065                 mutex_unlock(&mdsc->mutex);
2066                 goto out;
2067         }
2068
2069         /* dup? */
2070         if ((req->r_got_unsafe && !head->safe) ||
2071             (req->r_got_safe && head->safe)) {
2072                 pr_warning("got a dup %s reply on %llu from mds%d\n",
2073                            head->safe ? "safe" : "unsafe", tid, mds);
2074                 mutex_unlock(&mdsc->mutex);
2075                 goto out;
2076         }
2077         if (req->r_got_safe && !head->safe) {
2078                 pr_warning("got unsafe after safe on %llu from mds%d\n",
2079                            tid, mds);
2080                 mutex_unlock(&mdsc->mutex);
2081                 goto out;
2082         }
2083
2084         result = le32_to_cpu(head->result);
2085
2086         /*
2087          * Handle an ESTALE
2088          * if we're not talking to the authority, send to them
2089          * if the authority has changed while we weren't looking,
2090          * send to new authority
2091          * Otherwise we just have to return an ESTALE
2092          */
2093         if (result == -ESTALE) {
2094                 dout("got ESTALE on request %llu", req->r_tid);
2095                 if (!req->r_inode) {
2096                         /* do nothing; not an authority problem */
2097                 } else if (req->r_direct_mode != USE_AUTH_MDS) {
2098                         dout("not using auth, setting for that now");
2099                         req->r_direct_mode = USE_AUTH_MDS;
2100                         __do_request(mdsc, req);
2101                         mutex_unlock(&mdsc->mutex);
2102                         goto out;
2103                 } else  {
2104                         struct ceph_inode_info *ci = ceph_inode(req->r_inode);
2105                         struct ceph_cap *cap = NULL;
2106
2107                         if (req->r_session)
2108                                 cap = ceph_get_cap_for_mds(ci,
2109                                                    req->r_session->s_mds);
2110
2111                         dout("already using auth");
2112                         if ((!cap || cap != ci->i_auth_cap) ||
2113                             (cap->mseq != req->r_sent_on_mseq)) {
2114                                 dout("but cap changed, so resending");
2115                                 __do_request(mdsc, req);
2116                                 mutex_unlock(&mdsc->mutex);
2117                                 goto out;
2118                         }
2119                 }
2120                 dout("have to return ESTALE on request %llu", req->r_tid);
2121         }
2122
2123
2124         if (head->safe) {
2125                 req->r_got_safe = true;
2126                 __unregister_request(mdsc, req);
2127                 complete_all(&req->r_safe_completion);
2128
2129                 if (req->r_got_unsafe) {
2130                         /*
2131                          * We already handled the unsafe response, now do the
2132                          * cleanup.  No need to examine the response; the MDS
2133                          * doesn't include any result info in the safe
2134                          * response.  And even if it did, there is nothing
2135                          * useful we could do with a revised return value.
2136                          */
2137                         dout("got safe reply %llu, mds%d\n", tid, mds);
2138                         list_del_init(&req->r_unsafe_item);
2139
2140                         /* last unsafe request during umount? */
2141                         if (mdsc->stopping && !__get_oldest_req(mdsc))
2142                                 complete_all(&mdsc->safe_umount_waiters);
2143                         mutex_unlock(&mdsc->mutex);
2144                         goto out;
2145                 }
2146         } else {
2147                 req->r_got_unsafe = true;
2148                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2149         }
2150
2151         dout("handle_reply tid %lld result %d\n", tid, result);
2152         rinfo = &req->r_reply_info;
2153         err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2154         mutex_unlock(&mdsc->mutex);
2155
2156         mutex_lock(&session->s_mutex);
2157         if (err < 0) {
2158                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2159                 ceph_msg_dump(msg);
2160                 goto out_err;
2161         }
2162
2163         /* snap trace */
2164         if (rinfo->snapblob_len) {
2165                 down_write(&mdsc->snap_rwsem);
2166                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2167                                rinfo->snapblob + rinfo->snapblob_len,
2168                                le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
2169                 downgrade_write(&mdsc->snap_rwsem);
2170         } else {
2171                 down_read(&mdsc->snap_rwsem);
2172         }
2173
2174         /* insert trace into our cache */
2175         mutex_lock(&req->r_fill_mutex);
2176         err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2177         if (err == 0) {
2178                 if (result == 0 && req->r_op != CEPH_MDS_OP_GETFILELOCK &&
2179                     rinfo->dir_nr)
2180                         ceph_readdir_prepopulate(req, req->r_session);
2181                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2182         }
2183         mutex_unlock(&req->r_fill_mutex);
2184
2185         up_read(&mdsc->snap_rwsem);
2186 out_err:
2187         mutex_lock(&mdsc->mutex);
2188         if (!req->r_aborted) {
2189                 if (err) {
2190                         req->r_err = err;
2191                 } else {
2192                         req->r_reply = msg;
2193                         ceph_msg_get(msg);
2194                         req->r_got_result = true;
2195                 }
2196         } else {
2197                 dout("reply arrived after request %lld was aborted\n", tid);
2198         }
2199         mutex_unlock(&mdsc->mutex);
2200
2201         ceph_add_cap_releases(mdsc, req->r_session);
2202         mutex_unlock(&session->s_mutex);
2203
2204         /* kick calling process */
2205         complete_request(mdsc, req);
2206 out:
2207         ceph_mdsc_put_request(req);
2208         return;
2209 }
2210
2211
2212
2213 /*
2214  * handle mds notification that our request has been forwarded.
2215  */
2216 static void handle_forward(struct ceph_mds_client *mdsc,
2217                            struct ceph_mds_session *session,
2218                            struct ceph_msg *msg)
2219 {
2220         struct ceph_mds_request *req;
2221         u64 tid = le64_to_cpu(msg->hdr.tid);
2222         u32 next_mds;
2223         u32 fwd_seq;
2224         int err = -EINVAL;
2225         void *p = msg->front.iov_base;
2226         void *end = p + msg->front.iov_len;
2227
2228         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2229         next_mds = ceph_decode_32(&p);
2230         fwd_seq = ceph_decode_32(&p);
2231
2232         mutex_lock(&mdsc->mutex);
2233         req = __lookup_request(mdsc, tid);
2234         if (!req) {
2235                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2236                 goto out;  /* dup reply? */
2237         }
2238
2239         if (req->r_aborted) {
2240                 dout("forward tid %llu aborted, unregistering\n", tid);
2241                 __unregister_request(mdsc, req);
2242         } else if (fwd_seq <= req->r_num_fwd) {
2243                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2244                      tid, next_mds, req->r_num_fwd, fwd_seq);
2245         } else {
2246                 /* resend. forward race not possible; mds would drop */
2247                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2248                 BUG_ON(req->r_err);
2249                 BUG_ON(req->r_got_result);
2250                 req->r_num_fwd = fwd_seq;
2251                 req->r_resend_mds = next_mds;
2252                 put_request_session(req);
2253                 __do_request(mdsc, req);
2254         }
2255         ceph_mdsc_put_request(req);
2256 out:
2257         mutex_unlock(&mdsc->mutex);
2258         return;
2259
2260 bad:
2261         pr_err("mdsc_handle_forward decode error err=%d\n", err);
2262 }
2263
2264 /*
2265  * handle a mds session control message
2266  */
2267 static void handle_session(struct ceph_mds_session *session,
2268                            struct ceph_msg *msg)
2269 {
2270         struct ceph_mds_client *mdsc = session->s_mdsc;
2271         u32 op;
2272         u64 seq;
2273         int mds = session->s_mds;
2274         struct ceph_mds_session_head *h = msg->front.iov_base;
2275         int wake = 0;
2276
2277         /* decode */
2278         if (msg->front.iov_len != sizeof(*h))
2279                 goto bad;
2280         op = le32_to_cpu(h->op);
2281         seq = le64_to_cpu(h->seq);
2282
2283         mutex_lock(&mdsc->mutex);
2284         if (op == CEPH_SESSION_CLOSE)
2285                 __unregister_session(mdsc, session);
2286         /* FIXME: this ttl calculation is generous */
2287         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2288         mutex_unlock(&mdsc->mutex);
2289
2290         mutex_lock(&session->s_mutex);
2291
2292         dout("handle_session mds%d %s %p state %s seq %llu\n",
2293              mds, ceph_session_op_name(op), session,
2294              session_state_name(session->s_state), seq);
2295
2296         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2297                 session->s_state = CEPH_MDS_SESSION_OPEN;
2298                 pr_info("mds%d came back\n", session->s_mds);
2299         }
2300
2301         switch (op) {
2302         case CEPH_SESSION_OPEN:
2303                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2304                         pr_info("mds%d reconnect success\n", session->s_mds);
2305                 session->s_state = CEPH_MDS_SESSION_OPEN;
2306                 renewed_caps(mdsc, session, 0);
2307                 wake = 1;
2308                 if (mdsc->stopping)
2309                         __close_session(mdsc, session);
2310                 break;
2311
2312         case CEPH_SESSION_RENEWCAPS:
2313                 if (session->s_renew_seq == seq)
2314                         renewed_caps(mdsc, session, 1);
2315                 break;
2316
2317         case CEPH_SESSION_CLOSE:
2318                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2319                         pr_info("mds%d reconnect denied\n", session->s_mds);
2320                 remove_session_caps(session);
2321                 wake = 1; /* for good measure */
2322                 wake_up_all(&mdsc->session_close_wq);
2323                 kick_requests(mdsc, mds);
2324                 break;
2325
2326         case CEPH_SESSION_STALE:
2327                 pr_info("mds%d caps went stale, renewing\n",
2328                         session->s_mds);
2329                 spin_lock(&session->s_cap_lock);
2330                 session->s_cap_gen++;
2331                 session->s_cap_ttl = 0;
2332                 spin_unlock(&session->s_cap_lock);
2333                 send_renew_caps(mdsc, session);
2334                 break;
2335
2336         case CEPH_SESSION_RECALL_STATE:
2337                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2338                 break;
2339
2340         default:
2341                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2342                 WARN_ON(1);
2343         }
2344
2345         mutex_unlock(&session->s_mutex);
2346         if (wake) {
2347                 mutex_lock(&mdsc->mutex);
2348                 __wake_requests(mdsc, &session->s_waiting);
2349                 mutex_unlock(&mdsc->mutex);
2350         }
2351         return;
2352
2353 bad:
2354         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2355                (int)msg->front.iov_len);
2356         ceph_msg_dump(msg);
2357         return;
2358 }
2359
2360
2361 /*
2362  * called under session->mutex.
2363  */
2364 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2365                                    struct ceph_mds_session *session)
2366 {
2367         struct ceph_mds_request *req, *nreq;
2368         int err;
2369
2370         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2371
2372         mutex_lock(&mdsc->mutex);
2373         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2374                 err = __prepare_send_request(mdsc, req, session->s_mds);
2375                 if (!err) {
2376                         ceph_msg_get(req->r_request);
2377                         ceph_con_send(&session->s_con, req->r_request);
2378                 }
2379         }
2380         mutex_unlock(&mdsc->mutex);
2381 }
2382
2383 /*
2384  * Encode information about a cap for a reconnect with the MDS.
2385  */
2386 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2387                           void *arg)
2388 {
2389         union {
2390                 struct ceph_mds_cap_reconnect v2;
2391                 struct ceph_mds_cap_reconnect_v1 v1;
2392         } rec;
2393         size_t reclen;
2394         struct ceph_inode_info *ci;
2395         struct ceph_reconnect_state *recon_state = arg;
2396         struct ceph_pagelist *pagelist = recon_state->pagelist;
2397         char *path;
2398         int pathlen, err;
2399         u64 pathbase;
2400         struct dentry *dentry;
2401
2402         ci = cap->ci;
2403
2404         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2405              inode, ceph_vinop(inode), cap, cap->cap_id,
2406              ceph_cap_string(cap->issued));
2407         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2408         if (err)
2409                 return err;
2410
2411         dentry = d_find_alias(inode);
2412         if (dentry) {
2413                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2414                 if (IS_ERR(path)) {
2415                         err = PTR_ERR(path);
2416                         goto out_dput;
2417                 }
2418         } else {
2419                 path = NULL;
2420                 pathlen = 0;
2421         }
2422         err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2423         if (err)
2424                 goto out_free;
2425
2426         spin_lock(&ci->i_ceph_lock);
2427         cap->seq = 0;        /* reset cap seq */
2428         cap->issue_seq = 0;  /* and issue_seq */
2429
2430         if (recon_state->flock) {
2431                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2432                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2433                 rec.v2.issued = cpu_to_le32(cap->issued);
2434                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2435                 rec.v2.pathbase = cpu_to_le64(pathbase);
2436                 rec.v2.flock_len = 0;
2437                 reclen = sizeof(rec.v2);
2438         } else {
2439                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2440                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2441                 rec.v1.issued = cpu_to_le32(cap->issued);
2442                 rec.v1.size = cpu_to_le64(inode->i_size);
2443                 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2444                 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2445                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2446                 rec.v1.pathbase = cpu_to_le64(pathbase);
2447                 reclen = sizeof(rec.v1);
2448         }
2449         spin_unlock(&ci->i_ceph_lock);
2450
2451         if (recon_state->flock) {
2452                 int num_fcntl_locks, num_flock_locks;
2453                 struct ceph_pagelist_cursor trunc_point;
2454
2455                 ceph_pagelist_set_cursor(pagelist, &trunc_point);
2456                 do {
2457                         lock_flocks();
2458                         ceph_count_locks(inode, &num_fcntl_locks,
2459                                          &num_flock_locks);
2460                         rec.v2.flock_len = (2*sizeof(u32) +
2461                                             (num_fcntl_locks+num_flock_locks) *
2462                                             sizeof(struct ceph_filelock));
2463                         unlock_flocks();
2464
2465                         /* pre-alloc pagelist */
2466                         ceph_pagelist_truncate(pagelist, &trunc_point);
2467                         err = ceph_pagelist_append(pagelist, &rec, reclen);
2468                         if (!err)
2469                                 err = ceph_pagelist_reserve(pagelist,
2470                                                             rec.v2.flock_len);
2471
2472                         /* encode locks */
2473                         if (!err) {
2474                                 lock_flocks();
2475                                 err = ceph_encode_locks(inode,
2476                                                         pagelist,
2477                                                         num_fcntl_locks,
2478                                                         num_flock_locks);
2479                                 unlock_flocks();
2480                         }
2481                 } while (err == -ENOSPC);
2482         } else {
2483                 err = ceph_pagelist_append(pagelist, &rec, reclen);
2484         }
2485
2486 out_free:
2487         kfree(path);
2488 out_dput:
2489         dput(dentry);
2490         return err;
2491 }
2492
2493
2494 /*
2495  * If an MDS fails and recovers, clients need to reconnect in order to
2496  * reestablish shared state.  This includes all caps issued through
2497  * this session _and_ the snap_realm hierarchy.  Because it's not
2498  * clear which snap realms the mds cares about, we send everything we
2499  * know about.. that ensures we'll then get any new info the
2500  * recovering MDS might have.
2501  *
2502  * This is a relatively heavyweight operation, but it's rare.
2503  *
2504  * called with mdsc->mutex held.
2505  */
2506 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2507                                struct ceph_mds_session *session)
2508 {
2509         struct ceph_msg *reply;
2510         struct rb_node *p;
2511         int mds = session->s_mds;
2512         int err = -ENOMEM;
2513         struct ceph_pagelist *pagelist;
2514         struct ceph_reconnect_state recon_state;
2515
2516         pr_info("mds%d reconnect start\n", mds);
2517
2518         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2519         if (!pagelist)
2520                 goto fail_nopagelist;
2521         ceph_pagelist_init(pagelist);
2522
2523         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2524         if (!reply)
2525                 goto fail_nomsg;
2526
2527         mutex_lock(&session->s_mutex);
2528         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2529         session->s_seq = 0;
2530
2531         ceph_con_open(&session->s_con,
2532                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2533
2534         /* replay unsafe requests */
2535         replay_unsafe_requests(mdsc, session);
2536
2537         down_read(&mdsc->snap_rwsem);
2538
2539         dout("session %p state %s\n", session,
2540              session_state_name(session->s_state));
2541
2542         /* drop old cap expires; we're about to reestablish that state */
2543         discard_cap_releases(mdsc, session);
2544
2545         /* traverse this session's caps */
2546         err = ceph_pagelist_encode_32(pagelist, session->s_nr_caps);
2547         if (err)
2548                 goto fail;
2549
2550         recon_state.pagelist = pagelist;
2551         recon_state.flock = session->s_con.peer_features & CEPH_FEATURE_FLOCK;
2552         err = iterate_session_caps(session, encode_caps_cb, &recon_state);
2553         if (err < 0)
2554                 goto fail;
2555
2556         /*
2557          * snaprealms.  we provide mds with the ino, seq (version), and
2558          * parent for all of our realms.  If the mds has any newer info,
2559          * it will tell us.
2560          */
2561         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2562                 struct ceph_snap_realm *realm =
2563                         rb_entry(p, struct ceph_snap_realm, node);
2564                 struct ceph_mds_snaprealm_reconnect sr_rec;
2565
2566                 dout(" adding snap realm %llx seq %lld parent %llx\n",
2567                      realm->ino, realm->seq, realm->parent_ino);
2568                 sr_rec.ino = cpu_to_le64(realm->ino);
2569                 sr_rec.seq = cpu_to_le64(realm->seq);
2570                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2571                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2572                 if (err)
2573                         goto fail;
2574         }
2575
2576         reply->pagelist = pagelist;
2577         if (recon_state.flock)
2578                 reply->hdr.version = cpu_to_le16(2);
2579         reply->hdr.data_len = cpu_to_le32(pagelist->length);
2580         reply->nr_pages = calc_pages_for(0, pagelist->length);
2581         ceph_con_send(&session->s_con, reply);
2582
2583         mutex_unlock(&session->s_mutex);
2584
2585         mutex_lock(&mdsc->mutex);
2586         __wake_requests(mdsc, &session->s_waiting);
2587         mutex_unlock(&mdsc->mutex);
2588
2589         up_read(&mdsc->snap_rwsem);
2590         return;
2591
2592 fail:
2593         ceph_msg_put(reply);
2594         up_read(&mdsc->snap_rwsem);
2595         mutex_unlock(&session->s_mutex);
2596 fail_nomsg:
2597         ceph_pagelist_release(pagelist);
2598         kfree(pagelist);
2599 fail_nopagelist:
2600         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
2601         return;
2602 }
2603
2604
2605 /*
2606  * compare old and new mdsmaps, kicking requests
2607  * and closing out old connections as necessary
2608  *
2609  * called under mdsc->mutex.
2610  */
2611 static void check_new_map(struct ceph_mds_client *mdsc,
2612                           struct ceph_mdsmap *newmap,
2613                           struct ceph_mdsmap *oldmap)
2614 {
2615         int i;
2616         int oldstate, newstate;
2617         struct ceph_mds_session *s;
2618
2619         dout("check_new_map new %u old %u\n",
2620              newmap->m_epoch, oldmap->m_epoch);
2621
2622         for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2623                 if (mdsc->sessions[i] == NULL)
2624                         continue;
2625                 s = mdsc->sessions[i];
2626                 oldstate = ceph_mdsmap_get_state(oldmap, i);
2627                 newstate = ceph_mdsmap_get_state(newmap, i);
2628
2629                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
2630                      i, ceph_mds_state_name(oldstate),
2631                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
2632                      ceph_mds_state_name(newstate),
2633                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
2634                      session_state_name(s->s_state));
2635
2636                 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
2637                            ceph_mdsmap_get_addr(newmap, i),
2638                            sizeof(struct ceph_entity_addr))) {
2639                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2640                                 /* the session never opened, just close it
2641                                  * out now */
2642                                 __wake_requests(mdsc, &s->s_waiting);
2643                                 __unregister_session(mdsc, s);
2644                         } else {
2645                                 /* just close it */
2646                                 mutex_unlock(&mdsc->mutex);
2647                                 mutex_lock(&s->s_mutex);
2648                                 mutex_lock(&mdsc->mutex);
2649                                 ceph_con_close(&s->s_con);
2650                                 mutex_unlock(&s->s_mutex);
2651                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2652                         }
2653
2654                         /* kick any requests waiting on the recovering mds */
2655                         kick_requests(mdsc, i);
2656                 } else if (oldstate == newstate) {
2657                         continue;  /* nothing new with this mds */
2658                 }
2659
2660                 /*
2661                  * send reconnect?
2662                  */
2663                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2664                     newstate >= CEPH_MDS_STATE_RECONNECT) {
2665                         mutex_unlock(&mdsc->mutex);
2666                         send_mds_reconnect(mdsc, s);
2667                         mutex_lock(&mdsc->mutex);
2668                 }
2669
2670                 /*
2671                  * kick request on any mds that has gone active.
2672                  */
2673                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2674                     newstate >= CEPH_MDS_STATE_ACTIVE) {
2675                         if (oldstate != CEPH_MDS_STATE_CREATING &&
2676                             oldstate != CEPH_MDS_STATE_STARTING)
2677                                 pr_info("mds%d recovery completed\n", s->s_mds);
2678                         kick_requests(mdsc, i);
2679                         ceph_kick_flushing_caps(mdsc, s);
2680                         wake_up_session_caps(s, 1);
2681                 }
2682         }
2683
2684         for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
2685                 s = mdsc->sessions[i];
2686                 if (!s)
2687                         continue;
2688                 if (!ceph_mdsmap_is_laggy(newmap, i))
2689                         continue;
2690                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2691                     s->s_state == CEPH_MDS_SESSION_HUNG ||
2692                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
2693                         dout(" connecting to export targets of laggy mds%d\n",
2694                              i);
2695                         __open_export_target_sessions(mdsc, s);
2696                 }
2697         }
2698 }
2699
2700
2701
2702 /*
2703  * leases
2704  */
2705
2706 /*
2707  * caller must hold session s_mutex, dentry->d_lock
2708  */
2709 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2710 {
2711         struct ceph_dentry_info *di = ceph_dentry(dentry);
2712
2713         ceph_put_mds_session(di->lease_session);
2714         di->lease_session = NULL;
2715 }
2716
2717 static void handle_lease(struct ceph_mds_client *mdsc,
2718                          struct ceph_mds_session *session,
2719                          struct ceph_msg *msg)
2720 {
2721         struct super_block *sb = mdsc->fsc->sb;
2722         struct inode *inode;
2723         struct dentry *parent, *dentry;
2724         struct ceph_dentry_info *di;
2725         int mds = session->s_mds;
2726         struct ceph_mds_lease *h = msg->front.iov_base;
2727         u32 seq;
2728         struct ceph_vino vino;
2729         struct qstr dname;
2730         int release = 0;
2731
2732         dout("handle_lease from mds%d\n", mds);
2733
2734         /* decode */
2735         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2736                 goto bad;
2737         vino.ino = le64_to_cpu(h->ino);
2738         vino.snap = CEPH_NOSNAP;
2739         seq = le32_to_cpu(h->seq);
2740         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2741         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2742         if (dname.len != get_unaligned_le32(h+1))
2743                 goto bad;
2744
2745         mutex_lock(&session->s_mutex);
2746         session->s_seq++;
2747
2748         /* lookup inode */
2749         inode = ceph_find_inode(sb, vino);
2750         dout("handle_lease %s, ino %llx %p %.*s\n",
2751              ceph_lease_op_name(h->action), vino.ino, inode,
2752              dname.len, dname.name);
2753         if (inode == NULL) {
2754                 dout("handle_lease no inode %llx\n", vino.ino);
2755                 goto release;
2756         }
2757
2758         /* dentry */
2759         parent = d_find_alias(inode);
2760         if (!parent) {
2761                 dout("no parent dentry on inode %p\n", inode);
2762                 WARN_ON(1);
2763                 goto release;  /* hrm... */
2764         }
2765         dname.hash = full_name_hash(dname.name, dname.len);
2766         dentry = d_lookup(parent, &dname);
2767         dput(parent);
2768         if (!dentry)
2769                 goto release;
2770
2771         spin_lock(&dentry->d_lock);
2772         di = ceph_dentry(dentry);
2773         switch (h->action) {
2774         case CEPH_MDS_LEASE_REVOKE:
2775                 if (di && di->lease_session == session) {
2776                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
2777                                 h->seq = cpu_to_le32(di->lease_seq);
2778                         __ceph_mdsc_drop_dentry_lease(dentry);
2779                 }
2780                 release = 1;
2781                 break;
2782
2783         case CEPH_MDS_LEASE_RENEW:
2784                 if (di && di->lease_session == session &&
2785                     di->lease_gen == session->s_cap_gen &&
2786                     di->lease_renew_from &&
2787                     di->lease_renew_after == 0) {
2788                         unsigned long duration =
2789                                 le32_to_cpu(h->duration_ms) * HZ / 1000;
2790
2791                         di->lease_seq = seq;
2792                         dentry->d_time = di->lease_renew_from + duration;
2793                         di->lease_renew_after = di->lease_renew_from +
2794                                 (duration >> 1);
2795                         di->lease_renew_from = 0;
2796                 }
2797                 break;
2798         }
2799         spin_unlock(&dentry->d_lock);
2800         dput(dentry);
2801
2802         if (!release)
2803                 goto out;
2804
2805 release:
2806         /* let's just reuse the same message */
2807         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2808         ceph_msg_get(msg);
2809         ceph_con_send(&session->s_con, msg);
2810
2811 out:
2812         iput(inode);
2813         mutex_unlock(&session->s_mutex);
2814         return;
2815
2816 bad:
2817         pr_err("corrupt lease message\n");
2818         ceph_msg_dump(msg);
2819 }
2820
2821 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
2822                               struct inode *inode,
2823                               struct dentry *dentry, char action,
2824                               u32 seq)
2825 {
2826         struct ceph_msg *msg;
2827         struct ceph_mds_lease *lease;
2828         int len = sizeof(*lease) + sizeof(u32);
2829         int dnamelen = 0;
2830
2831         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2832              inode, dentry, ceph_lease_op_name(action), session->s_mds);
2833         dnamelen = dentry->d_name.len;
2834         len += dnamelen;
2835
2836         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
2837         if (!msg)
2838                 return;
2839         lease = msg->front.iov_base;
2840         lease->action = action;
2841         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
2842         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
2843         lease->seq = cpu_to_le32(seq);
2844         put_unaligned_le32(dnamelen, lease + 1);
2845         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
2846
2847         /*
2848          * if this is a preemptive lease RELEASE, no need to
2849          * flush request stream, since the actual request will
2850          * soon follow.
2851          */
2852         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
2853
2854         ceph_con_send(&session->s_con, msg);
2855 }
2856
2857 /*
2858  * Preemptively release a lease we expect to invalidate anyway.
2859  * Pass @inode always, @dentry is optional.
2860  */
2861 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2862                              struct dentry *dentry)
2863 {
2864         struct ceph_dentry_info *di;
2865         struct ceph_mds_session *session;
2866         u32 seq;
2867
2868         BUG_ON(inode == NULL);
2869         BUG_ON(dentry == NULL);
2870
2871         /* is dentry lease valid? */
2872         spin_lock(&dentry->d_lock);
2873         di = ceph_dentry(dentry);
2874         if (!di || !di->lease_session ||
2875             di->lease_session->s_mds < 0 ||
2876             di->lease_gen != di->lease_session->s_cap_gen ||
2877             !time_before(jiffies, dentry->d_time)) {
2878                 dout("lease_release inode %p dentry %p -- "
2879                      "no lease\n",
2880                      inode, dentry);
2881                 spin_unlock(&dentry->d_lock);
2882                 return;
2883         }
2884
2885         /* we do have a lease on this dentry; note mds and seq */
2886         session = ceph_get_mds_session(di->lease_session);
2887         seq = di->lease_seq;
2888         __ceph_mdsc_drop_dentry_lease(dentry);
2889         spin_unlock(&dentry->d_lock);
2890
2891         dout("lease_release inode %p dentry %p to mds%d\n",
2892              inode, dentry, session->s_mds);
2893         ceph_mdsc_lease_send_msg(session, inode, dentry,
2894                                  CEPH_MDS_LEASE_RELEASE, seq);
2895         ceph_put_mds_session(session);
2896 }
2897
2898 /*
2899  * drop all leases (and dentry refs) in preparation for umount
2900  */
2901 static void drop_leases(struct ceph_mds_client *mdsc)
2902 {
2903         int i;
2904
2905         dout("drop_leases\n");
2906         mutex_lock(&mdsc->mutex);
2907         for (i = 0; i < mdsc->max_sessions; i++) {
2908                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2909                 if (!s)
2910                         continue;
2911                 mutex_unlock(&mdsc->mutex);
2912                 mutex_lock(&s->s_mutex);
2913                 mutex_unlock(&s->s_mutex);
2914                 ceph_put_mds_session(s);
2915                 mutex_lock(&mdsc->mutex);
2916         }
2917         mutex_unlock(&mdsc->mutex);
2918 }
2919
2920
2921
2922 /*
2923  * delayed work -- periodically trim expired leases, renew caps with mds
2924  */
2925 static void schedule_delayed(struct ceph_mds_client *mdsc)
2926 {
2927         int delay = 5;
2928         unsigned hz = round_jiffies_relative(HZ * delay);
2929         schedule_delayed_work(&mdsc->delayed_work, hz);
2930 }
2931
2932 static void delayed_work(struct work_struct *work)
2933 {
2934         int i;
2935         struct ceph_mds_client *mdsc =
2936                 container_of(work, struct ceph_mds_client, delayed_work.work);
2937         int renew_interval;
2938         int renew_caps;
2939
2940         dout("mdsc delayed_work\n");
2941         ceph_check_delayed_caps(mdsc);
2942
2943         mutex_lock(&mdsc->mutex);
2944         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
2945         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
2946                                    mdsc->last_renew_caps);
2947         if (renew_caps)
2948                 mdsc->last_renew_caps = jiffies;
2949
2950         for (i = 0; i < mdsc->max_sessions; i++) {
2951                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2952                 if (s == NULL)
2953                         continue;
2954                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
2955                         dout("resending session close request for mds%d\n",
2956                              s->s_mds);
2957                         request_close_session(mdsc, s);
2958                         ceph_put_mds_session(s);
2959                         continue;
2960                 }
2961                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
2962                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
2963                                 s->s_state = CEPH_MDS_SESSION_HUNG;
2964                                 pr_info("mds%d hung\n", s->s_mds);
2965                         }
2966                 }
2967                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
2968                         /* this mds is failed or recovering, just wait */
2969                         ceph_put_mds_session(s);
2970                         continue;
2971                 }
2972                 mutex_unlock(&mdsc->mutex);
2973
2974                 mutex_lock(&s->s_mutex);
2975                 if (renew_caps)
2976                         send_renew_caps(mdsc, s);
2977                 else
2978                         ceph_con_keepalive(&s->s_con);
2979                 ceph_add_cap_releases(mdsc, s);
2980                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2981                     s->s_state == CEPH_MDS_SESSION_HUNG)
2982                         ceph_send_cap_releases(mdsc, s);
2983                 mutex_unlock(&s->s_mutex);
2984                 ceph_put_mds_session(s);
2985
2986                 mutex_lock(&mdsc->mutex);
2987         }
2988         mutex_unlock(&mdsc->mutex);
2989
2990         schedule_delayed(mdsc);
2991 }
2992
2993 int ceph_mdsc_init(struct ceph_fs_client *fsc)
2994
2995 {
2996         struct ceph_mds_client *mdsc;
2997
2998         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
2999         if (!mdsc)
3000                 return -ENOMEM;
3001         mdsc->fsc = fsc;
3002         fsc->mdsc = mdsc;
3003         mutex_init(&mdsc->mutex);
3004         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3005         if (mdsc->mdsmap == NULL)
3006                 return -ENOMEM;
3007
3008         init_completion(&mdsc->safe_umount_waiters);
3009         init_waitqueue_head(&mdsc->session_close_wq);
3010         INIT_LIST_HEAD(&mdsc->waiting_for_map);
3011         mdsc->sessions = NULL;
3012         mdsc->max_sessions = 0;
3013         mdsc->stopping = 0;
3014         init_rwsem(&mdsc->snap_rwsem);
3015         mdsc->snap_realms = RB_ROOT;
3016         INIT_LIST_HEAD(&mdsc->snap_empty);
3017         spin_lock_init(&mdsc->snap_empty_lock);
3018         mdsc->last_tid = 0;
3019         mdsc->request_tree = RB_ROOT;
3020         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3021         mdsc->last_renew_caps = jiffies;
3022         INIT_LIST_HEAD(&mdsc->cap_delay_list);
3023         spin_lock_init(&mdsc->cap_delay_lock);
3024         INIT_LIST_HEAD(&mdsc->snap_flush_list);
3025         spin_lock_init(&mdsc->snap_flush_lock);
3026         mdsc->cap_flush_seq = 0;
3027         INIT_LIST_HEAD(&mdsc->cap_dirty);
3028         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3029         mdsc->num_cap_flushing = 0;
3030         spin_lock_init(&mdsc->cap_dirty_lock);
3031         init_waitqueue_head(&mdsc->cap_flushing_wq);
3032         spin_lock_init(&mdsc->dentry_lru_lock);
3033         INIT_LIST_HEAD(&mdsc->dentry_lru);
3034
3035         ceph_caps_init(mdsc);
3036         ceph_adjust_min_caps(mdsc, fsc->min_caps);
3037
3038         return 0;
3039 }
3040
3041 /*
3042  * Wait for safe replies on open mds requests.  If we time out, drop
3043  * all requests from the tree to avoid dangling dentry refs.
3044  */
3045 static void wait_requests(struct ceph_mds_client *mdsc)
3046 {
3047         struct ceph_mds_request *req;
3048         struct ceph_fs_client *fsc = mdsc->fsc;
3049
3050         mutex_lock(&mdsc->mutex);
3051         if (__get_oldest_req(mdsc)) {
3052                 mutex_unlock(&mdsc->mutex);
3053
3054                 dout("wait_requests waiting for requests\n");
3055                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3056                                     fsc->client->options->mount_timeout * HZ);
3057
3058                 /* tear down remaining requests */
3059                 mutex_lock(&mdsc->mutex);
3060                 while ((req = __get_oldest_req(mdsc))) {
3061                         dout("wait_requests timed out on tid %llu\n",
3062                              req->r_tid);
3063                         __unregister_request(mdsc, req);
3064                 }
3065         }
3066         mutex_unlock(&mdsc->mutex);
3067         dout("wait_requests done\n");
3068 }
3069
3070 /*
3071  * called before mount is ro, and before dentries are torn down.
3072  * (hmm, does this still race with new lookups?)
3073  */
3074 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3075 {
3076         dout("pre_umount\n");
3077         mdsc->stopping = 1;
3078
3079         drop_leases(mdsc);
3080         ceph_flush_dirty_caps(mdsc);
3081         wait_requests(mdsc);
3082
3083         /*
3084          * wait for reply handlers to drop their request refs and
3085          * their inode/dcache refs
3086          */
3087         ceph_msgr_flush();
3088 }
3089
3090 /*
3091  * wait for all write mds requests to flush.
3092  */
3093 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3094 {
3095         struct ceph_mds_request *req = NULL, *nextreq;
3096         struct rb_node *n;
3097
3098         mutex_lock(&mdsc->mutex);
3099         dout("wait_unsafe_requests want %lld\n", want_tid);
3100 restart:
3101         req = __get_oldest_req(mdsc);
3102         while (req && req->r_tid <= want_tid) {
3103                 /* find next request */
3104                 n = rb_next(&req->r_node);
3105                 if (n)
3106                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3107                 else
3108                         nextreq = NULL;
3109                 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
3110                         /* write op */
3111                         ceph_mdsc_get_request(req);
3112                         if (nextreq)
3113                                 ceph_mdsc_get_request(nextreq);
3114                         mutex_unlock(&mdsc->mutex);
3115                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
3116                              req->r_tid, want_tid);
3117                         wait_for_completion(&req->r_safe_completion);
3118                         mutex_lock(&mdsc->mutex);
3119                         ceph_mdsc_put_request(req);
3120                         if (!nextreq)
3121                                 break;  /* next dne before, so we're done! */
3122                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
3123                                 /* next request was removed from tree */
3124                                 ceph_mdsc_put_request(nextreq);
3125                                 goto restart;
3126                         }
3127                         ceph_mdsc_put_request(nextreq);  /* won't go away */
3128                 }
3129                 req = nextreq;
3130         }
3131         mutex_unlock(&mdsc->mutex);
3132         dout("wait_unsafe_requests done\n");
3133 }
3134
3135 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3136 {
3137         u64 want_tid, want_flush;
3138
3139         if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3140                 return;
3141
3142         dout("sync\n");
3143         mutex_lock(&mdsc->mutex);
3144         want_tid = mdsc->last_tid;
3145         want_flush = mdsc->cap_flush_seq;
3146         mutex_unlock(&mdsc->mutex);
3147         dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
3148
3149         ceph_flush_dirty_caps(mdsc);
3150
3151         wait_unsafe_requests(mdsc, want_tid);
3152         wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
3153 }
3154
3155 /*
3156  * true if all sessions are closed, or we force unmount
3157  */
3158 static bool done_closing_sessions(struct ceph_mds_client *mdsc)
3159 {
3160         int i, n = 0;
3161
3162         if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3163                 return true;
3164
3165         mutex_lock(&mdsc->mutex);
3166         for (i = 0; i < mdsc->max_sessions; i++)
3167                 if (mdsc->sessions[i])
3168                         n++;
3169         mutex_unlock(&mdsc->mutex);
3170         return n == 0;
3171 }
3172
3173 /*
3174  * called after sb is ro.
3175  */
3176 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3177 {
3178         struct ceph_mds_session *session;
3179         int i;
3180         struct ceph_fs_client *fsc = mdsc->fsc;
3181         unsigned long timeout = fsc->client->options->mount_timeout * HZ;
3182
3183         dout("close_sessions\n");
3184
3185         /* close sessions */
3186         mutex_lock(&mdsc->mutex);
3187         for (i = 0; i < mdsc->max_sessions; i++) {
3188                 session = __ceph_lookup_mds_session(mdsc, i);
3189                 if (!session)
3190                         continue;
3191                 mutex_unlock(&mdsc->mutex);
3192                 mutex_lock(&session->s_mutex);
3193                 __close_session(mdsc, session);
3194                 mutex_unlock(&session->s_mutex);
3195                 ceph_put_mds_session(session);
3196                 mutex_lock(&mdsc->mutex);
3197         }
3198         mutex_unlock(&mdsc->mutex);
3199
3200         dout("waiting for sessions to close\n");
3201         wait_event_timeout(mdsc->session_close_wq, done_closing_sessions(mdsc),
3202                            timeout);
3203
3204         /* tear down remaining sessions */
3205         mutex_lock(&mdsc->mutex);
3206         for (i = 0; i < mdsc->max_sessions; i++) {
3207                 if (mdsc->sessions[i]) {
3208                         session = get_session(mdsc->sessions[i]);
3209                         __unregister_session(mdsc, session);
3210                         mutex_unlock(&mdsc->mutex);
3211                         mutex_lock(&session->s_mutex);
3212                         remove_session_caps(session);
3213                         mutex_unlock(&session->s_mutex);
3214                         ceph_put_mds_session(session);
3215                         mutex_lock(&mdsc->mutex);
3216                 }
3217         }
3218         WARN_ON(!list_empty(&mdsc->cap_delay_list));
3219         mutex_unlock(&mdsc->mutex);
3220
3221         ceph_cleanup_empty_realms(mdsc);
3222
3223         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3224
3225         dout("stopped\n");
3226 }
3227
3228 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3229 {
3230         dout("stop\n");
3231         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3232         if (mdsc->mdsmap)
3233                 ceph_mdsmap_destroy(mdsc->mdsmap);
3234         kfree(mdsc->sessions);
3235         ceph_caps_finalize(mdsc);
3236 }
3237
3238 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3239 {
3240         struct ceph_mds_client *mdsc = fsc->mdsc;
3241
3242         dout("mdsc_destroy %p\n", mdsc);
3243         ceph_mdsc_stop(mdsc);
3244
3245         /* flush out any connection work with references to us */
3246         ceph_msgr_flush();
3247
3248         fsc->mdsc = NULL;
3249         kfree(mdsc);
3250         dout("mdsc_destroy %p done\n", mdsc);
3251 }
3252
3253
3254 /*
3255  * handle mds map update.
3256  */
3257 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3258 {
3259         u32 epoch;
3260         u32 maplen;
3261         void *p = msg->front.iov_base;
3262         void *end = p + msg->front.iov_len;
3263         struct ceph_mdsmap *newmap, *oldmap;
3264         struct ceph_fsid fsid;
3265         int err = -EINVAL;
3266
3267         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3268         ceph_decode_copy(&p, &fsid, sizeof(fsid));
3269         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3270                 return;
3271         epoch = ceph_decode_32(&p);
3272         maplen = ceph_decode_32(&p);
3273         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3274
3275         /* do we need it? */
3276         ceph_monc_got_mdsmap(&mdsc->fsc->client->monc, epoch);
3277         mutex_lock(&mdsc->mutex);
3278         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3279                 dout("handle_map epoch %u <= our %u\n",
3280                      epoch, mdsc->mdsmap->m_epoch);
3281                 mutex_unlock(&mdsc->mutex);
3282                 return;
3283         }
3284
3285         newmap = ceph_mdsmap_decode(&p, end);
3286         if (IS_ERR(newmap)) {
3287                 err = PTR_ERR(newmap);
3288                 goto bad_unlock;
3289         }
3290
3291         /* swap into place */
3292         if (mdsc->mdsmap) {
3293                 oldmap = mdsc->mdsmap;
3294                 mdsc->mdsmap = newmap;
3295                 check_new_map(mdsc, newmap, oldmap);
3296                 ceph_mdsmap_destroy(oldmap);
3297         } else {
3298                 mdsc->mdsmap = newmap;  /* first mds map */
3299         }
3300         mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3301
3302         __wake_requests(mdsc, &mdsc->waiting_for_map);
3303
3304         mutex_unlock(&mdsc->mutex);
3305         schedule_delayed(mdsc);
3306         return;
3307
3308 bad_unlock:
3309         mutex_unlock(&mdsc->mutex);
3310 bad:
3311         pr_err("error decoding mdsmap %d\n", err);
3312         return;
3313 }
3314
3315 static struct ceph_connection *con_get(struct ceph_connection *con)
3316 {
3317         struct ceph_mds_session *s = con->private;
3318
3319         if (get_session(s)) {
3320                 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3321                 return con;
3322         }
3323         dout("mdsc con_get %p FAIL\n", s);
3324         return NULL;
3325 }
3326
3327 static void con_put(struct ceph_connection *con)
3328 {
3329         struct ceph_mds_session *s = con->private;
3330
3331         dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3332         ceph_put_mds_session(s);
3333 }
3334
3335 /*
3336  * if the client is unresponsive for long enough, the mds will kill
3337  * the session entirely.
3338  */
3339 static void peer_reset(struct ceph_connection *con)
3340 {
3341         struct ceph_mds_session *s = con->private;
3342         struct ceph_mds_client *mdsc = s->s_mdsc;
3343
3344         pr_warning("mds%d closed our session\n", s->s_mds);
3345         send_mds_reconnect(mdsc, s);
3346 }
3347
3348 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3349 {
3350         struct ceph_mds_session *s = con->private;
3351         struct ceph_mds_client *mdsc = s->s_mdsc;
3352         int type = le16_to_cpu(msg->hdr.type);
3353
3354         mutex_lock(&mdsc->mutex);
3355         if (__verify_registered_session(mdsc, s) < 0) {
3356                 mutex_unlock(&mdsc->mutex);
3357                 goto out;
3358         }
3359         mutex_unlock(&mdsc->mutex);
3360
3361         switch (type) {
3362         case CEPH_MSG_MDS_MAP:
3363                 ceph_mdsc_handle_map(mdsc, msg);
3364                 break;
3365         case CEPH_MSG_CLIENT_SESSION:
3366                 handle_session(s, msg);
3367                 break;
3368         case CEPH_MSG_CLIENT_REPLY:
3369                 handle_reply(s, msg);
3370                 break;
3371         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3372                 handle_forward(mdsc, s, msg);
3373                 break;
3374         case CEPH_MSG_CLIENT_CAPS:
3375                 ceph_handle_caps(s, msg);
3376                 break;
3377         case CEPH_MSG_CLIENT_SNAP:
3378                 ceph_handle_snap(mdsc, s, msg);
3379                 break;
3380         case CEPH_MSG_CLIENT_LEASE:
3381                 handle_lease(mdsc, s, msg);
3382                 break;
3383
3384         default:
3385                 pr_err("received unknown message type %d %s\n", type,
3386                        ceph_msg_type_name(type));
3387         }
3388 out:
3389         ceph_msg_put(msg);
3390 }
3391
3392 /*
3393  * authentication
3394  */
3395 static int get_authorizer(struct ceph_connection *con,
3396                           void **buf, int *len, int *proto,
3397                           void **reply_buf, int *reply_len, int force_new)
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         int ret = 0;
3403
3404         if (force_new && s->s_authorizer) {
3405                 ac->ops->destroy_authorizer(ac, s->s_authorizer);
3406                 s->s_authorizer = NULL;
3407         }
3408         if (s->s_authorizer == NULL) {
3409                 if (ac->ops->create_authorizer) {
3410                         ret = ac->ops->create_authorizer(
3411                                 ac, CEPH_ENTITY_TYPE_MDS,
3412                                 &s->s_authorizer,
3413                                 &s->s_authorizer_buf,
3414                                 &s->s_authorizer_buf_len,
3415                                 &s->s_authorizer_reply_buf,
3416                                 &s->s_authorizer_reply_buf_len);
3417                         if (ret)
3418                                 return ret;
3419                 }
3420         }
3421
3422         *proto = ac->protocol;
3423         *buf = s->s_authorizer_buf;
3424         *len = s->s_authorizer_buf_len;
3425         *reply_buf = s->s_authorizer_reply_buf;
3426         *reply_len = s->s_authorizer_reply_buf_len;
3427         return 0;
3428 }
3429
3430
3431 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3432 {
3433         struct ceph_mds_session *s = con->private;
3434         struct ceph_mds_client *mdsc = s->s_mdsc;
3435         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3436
3437         return ac->ops->verify_authorizer_reply(ac, s->s_authorizer, len);
3438 }
3439
3440 static int invalidate_authorizer(struct ceph_connection *con)
3441 {
3442         struct ceph_mds_session *s = con->private;
3443         struct ceph_mds_client *mdsc = s->s_mdsc;
3444         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3445
3446         if (ac->ops->invalidate_authorizer)
3447                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3448
3449         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
3450 }
3451
3452 static const struct ceph_connection_operations mds_con_ops = {
3453         .get = con_get,
3454         .put = con_put,
3455         .dispatch = dispatch,
3456         .get_authorizer = get_authorizer,
3457         .verify_authorizer_reply = verify_authorizer_reply,
3458         .invalidate_authorizer = invalidate_authorizer,
3459         .peer_reset = peer_reset,
3460 };
3461
3462 /* eof */