fs/ceph/mds_client.c

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  37  * are no longer valid.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