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  37  * are no longer valid.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 const static 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_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(&req->r_caps_reservation, req->r_num_caps);
 516         dout("__register_request %p tid %lld\n", req, req->r_tid);
 517         ceph_mdsc_get_request(req);
 518         __insert_request(mdsc, req);
 519
 520         if (dir) {
 521                 struct ceph_inode_info *ci = ceph_inode(dir);
 522
 523                 spin_lock(&ci->i_unsafe_lock);
 524                 req->r_unsafe_dir = dir;
 525                 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
 526                 spin_unlock(&ci->i_unsafe_lock);
 527         }
 528 }
 529
 530 static void __unregister_request(struct ceph_mds_client *mdsc,
 531                                  struct ceph_mds_request *req)
 532 {
 533         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 534         rb_erase(&req->r_node, &mdsc->request_tree);
 535         RB_CLEAR_NODE(&req->r_node);
 536
 537         if (req->r_unsafe_dir) {
 538                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 539
 540                 spin_lock(&ci->i_unsafe_lock);
 541                 list_del_init(&req->r_unsafe_dir_item);
 542                 spin_unlock(&ci->i_unsafe_lock);
 543         }
 544
 545         ceph_mdsc_put_request(req);
 546 }
 547
 548 /*
 549  * Choose mds to send request to next.  If there is a hint set in the
 550  * request (e.g., due to a prior forward hint from the mds), use that.
 551  * Otherwise, consult frag tree and/or caps to identify the
 552  * appropriate mds.  If all else fails, choose randomly.
 553  *
 554  * Called under mdsc->mutex.
 555  */
 556 static int __choose_mds(struct ceph_mds_client *mdsc,
 557                         struct ceph_mds_request *req)
 558 {
 559         struct inode *inode;
 560         struct ceph_inode_info *ci;
 561         struct ceph_cap *cap;
 562         int mode = req->r_direct_mode;
 563         int mds = -1;
 564         u32 hash = req->r_direct_hash;
 565         bool is_hash = req->r_direct_is_hash;
 566
 567         /*
 568          * is there a specific mds we should try?  ignore hint if we have
 569          * no session and the mds is not up (active or recovering).
 570          */
 571         if (req->r_resend_mds >= 0 &&
 572             (__have_session(mdsc, req->r_resend_mds) ||
 573              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 574                 dout("choose_mds using resend_mds mds%d\n",
 575                      req->r_resend_mds);
 576                 return req->r_resend_mds;
 577         }
 578
 579         if (mode == USE_RANDOM_MDS)
 580                 goto random;
 581
 582         inode = NULL;
 583         if (req->r_inode) {
 584                 inode = req->r_inode;
 585         } else if (req->r_dentry) {
 586                 if (req->r_dentry->d_inode) {
 587                         inode = req->r_dentry->d_inode;
 588                 } else {
 589                         inode = req->r_dentry->d_parent->d_inode;
 590                         hash = req->r_dentry->d_name.hash;
 591                         is_hash = true;
 592                 }
 593         }
 594         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 595              (int)hash, mode);
 596         if (!inode)
 597                 goto random;
 598         ci = ceph_inode(inode);
 599
 600         if (is_hash && S_ISDIR(inode->i_mode)) {
 601                 struct ceph_inode_frag frag;
 602                 int found;
 603
 604                 ceph_choose_frag(ci, hash, &frag, &found);
 605                 if (found) {
 606                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 607                                 u8 r;
 608
 609                                 /* choose a random replica */
 610                                 get_random_bytes(&r, 1);
 611                                 r %= frag.ndist;
 612                                 mds = frag.dist[r];
 613                                 dout("choose_mds %p %llx.%llx "
 614                                      "frag %u mds%d (%d/%d)\n",
 615                                      inode, ceph_vinop(inode),
 616                                      frag.frag, frag.mds,
 617                                      (int)r, frag.ndist);
 618                                 return mds;
 619                         }
 620
 621                         /* since this file/dir wasn't known to be
 622                          * replicated, then we want to look for the
 623                          * authoritative mds. */
 624                         mode = USE_AUTH_MDS;
 625                         if (frag.mds >= 0) {
 626                                 /* choose auth mds */
 627                                 mds = frag.mds;
 628                                 dout("choose_mds %p %llx.%llx "
 629                                      "frag %u mds%d (auth)\n",
 630                                      inode, ceph_vinop(inode), frag.frag, mds);
 631                                 return mds;
 632                         }
 633                 }
 634         }
 635
 636         spin_lock(&inode->i_lock);
 637         cap = NULL;
 638         if (mode == USE_AUTH_MDS)
 639                 cap = ci->i_auth_cap;
 640         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
 641                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
 642         if (!cap) {
 643                 spin_unlock(&inode->i_lock);
 644                 goto random;
 645         }
 646         mds = cap->session->s_mds;
 647         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
 648              inode, ceph_vinop(inode), mds,
 649              cap == ci->i_auth_cap ? "auth " : "", cap);
 650         spin_unlock(&inode->i_lock);
 651         return mds;
 652
 653 random:
 654         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
 655         dout("choose_mds chose random mds%d\n", mds);
 656         return mds;
 657 }
 658
 659
 660 /*
 661  * session messages
 662  */
 663 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
 664 {
 665         struct ceph_msg *msg;
 666         struct ceph_mds_session_head *h;
 667
 668         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), 0, 0, NULL);
 669         if (IS_ERR(msg)) {
 670                 pr_err("create_session_msg ENOMEM creating msg\n");
 671                 return ERR_PTR(PTR_ERR(msg));
 672         }
 673         h = msg->front.iov_base;
 674         h->op = cpu_to_le32(op);
 675         h->seq = cpu_to_le64(seq);
 676         return msg;
 677 }
 678
 679 /*
 680  * send session open request.
 681  *
 682  * called under mdsc->mutex
 683  */
 684 static int __open_session(struct ceph_mds_client *mdsc,
 685                           struct ceph_mds_session *session)
 686 {
 687         struct ceph_msg *msg;
 688         int mstate;
 689         int mds = session->s_mds;
 690         int err = 0;
 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 (IS_ERR(msg)) {
 702                 err = PTR_ERR(msg);
 703                 goto out;
 704         }
 705         ceph_con_send(&session->s_con, msg);
 706
 707 out:
 708         return 0;
 709 }
 710
 711 /*
 712  * session caps
 713  */
 714
 715 /*
 716  * Free preallocated cap messages assigned to this session
 717  */
 718 static void cleanup_cap_releases(struct ceph_mds_session *session)
 719 {
 720         struct ceph_msg *msg;
 721
 722         spin_lock(&session->s_cap_lock);
 723         while (!list_empty(&session->s_cap_releases)) {
 724                 msg = list_first_entry(&session->s_cap_releases,
 725                                        struct ceph_msg, list_head);
 726                 list_del_init(&msg->list_head);
 727                 ceph_msg_put(msg);
 728         }
 729         while (!list_empty(&session->s_cap_releases_done)) {
 730                 msg = list_first_entry(&session->s_cap_releases_done,
 731                                        struct ceph_msg, list_head);
 732                 list_del_init(&msg->list_head);
 733                 ceph_msg_put(msg);
 734         }
 735         spin_unlock(&session->s_cap_lock);
 736 }
 737
 738 /*
 739  * Helper to safely iterate over all caps associated with a session, with
 740  * special care taken to handle a racing __ceph_remove_cap().
 741  *
 742  * Caller must hold session s_mutex.
 743  */
 744 static int iterate_session_caps(struct ceph_mds_session *session,
 745                                  int (*cb)(struct inode *, struct ceph_cap *,
 746                                             void *), void *arg)
 747 {
 748         struct list_head *p;
 749         struct ceph_cap *cap;
 750         struct inode *inode, *last_inode = NULL;
 751         struct ceph_cap *old_cap = NULL;
 752         int ret;
 753
 754         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
 755         spin_lock(&session->s_cap_lock);
 756         p = session->s_caps.next;
 757         while (p != &session->s_caps) {
 758                 cap = list_entry(p, struct ceph_cap, session_caps);
 759                 inode = igrab(&cap->ci->vfs_inode);
 760                 if (!inode) {
 761                         p = p->next;
 762                         continue;
 763                 }
 764                 session->s_cap_iterator = cap;
 765                 spin_unlock(&session->s_cap_lock);
 766
 767                 if (last_inode) {
 768                         iput(last_inode);
 769                         last_inode = NULL;
 770                 }
 771                 if (old_cap) {
 772                         ceph_put_cap(old_cap);
 773                         old_cap = NULL;
 774                 }
 775
 776                 ret = cb(inode, cap, arg);
 777                 last_inode = inode;
 778
 779                 spin_lock(&session->s_cap_lock);
 780                 p = p->next;
 781                 if (cap->ci == NULL) {
 782                         dout("iterate_session_caps  finishing cap %p removal\n",
 783                              cap);
 784                         BUG_ON(cap->session != session);
 785                         list_del_init(&cap->session_caps);
 786                         session->s_nr_caps--;
 787                         cap->session = NULL;
 788                         old_cap = cap;  /* put_cap it w/o locks held */
 789                 }
 790                 if (ret < 0)
 791                         goto out;
 792         }
 793         ret = 0;
 794 out:
 795         session->s_cap_iterator = NULL;
 796         spin_unlock(&session->s_cap_lock);
 797
 798         if (last_inode)
 799                 iput(last_inode);
 800         if (old_cap)
 801                 ceph_put_cap(old_cap);
 802
 803         return ret;
 804 }
 805
 806 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
 807                                    void *arg)
 808 {
 809         struct ceph_inode_info *ci = ceph_inode(inode);
 810         dout("removing cap %p, ci is %p, inode is %p\n",
 811              cap, ci, &ci->vfs_inode);
 812         ceph_remove_cap(cap);
 813         return 0;
 814 }
 815
 816 /*
 817  * caller must hold session s_mutex
 818  */
 819 static void remove_session_caps(struct ceph_mds_session *session)
 820 {
 821         dout("remove_session_caps on %p\n", session);
 822         iterate_session_caps(session, remove_session_caps_cb, NULL);
 823         BUG_ON(session->s_nr_caps > 0);
 824         cleanup_cap_releases(session);
 825 }
 826
 827 /*
 828  * wake up any threads waiting on this session's caps.  if the cap is
 829  * old (didn't get renewed on the client reconnect), remove it now.
 830  *
 831  * caller must hold s_mutex.
 832  */
 833 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
 834                               void *arg)
 835 {
 836         struct ceph_inode_info *ci = ceph_inode(inode);
 837
 838         wake_up(&ci->i_cap_wq);
 839         if (arg) {
 840                 spin_lock(&inode->i_lock);
 841                 ci->i_wanted_max_size = 0;
 842                 ci->i_requested_max_size = 0;
 843                 spin_unlock(&inode->i_lock);
 844         }
 845         return 0;
 846 }
 847
 848 static void wake_up_session_caps(struct ceph_mds_session *session,
 849                                  int reconnect)
 850 {
 851         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
 852         iterate_session_caps(session, wake_up_session_cb,
 853                              (void *)(unsigned long)reconnect);
 854 }
 855
 856 /*
 857  * Send periodic message to MDS renewing all currently held caps.  The
 858  * ack will reset the expiration for all caps from this session.
 859  *
 860  * caller holds s_mutex
 861  */
 862 static int send_renew_caps(struct ceph_mds_client *mdsc,
 863                            struct ceph_mds_session *session)
 864 {
 865         struct ceph_msg *msg;
 866         int state;
 867
 868         if (time_after_eq(jiffies, session->s_cap_ttl) &&
 869             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
 870                 pr_info("mds%d caps stale\n", session->s_mds);
 871         session->s_renew_requested = jiffies;
 872
 873         /* do not try to renew caps until a recovering mds has reconnected
 874          * with its clients. */
 875         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
 876         if (state < CEPH_MDS_STATE_RECONNECT) {
 877                 dout("send_renew_caps ignoring mds%d (%s)\n",
 878                      session->s_mds, ceph_mds_state_name(state));
 879                 return 0;
 880         }
 881
 882         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
 883                 ceph_mds_state_name(state));
 884         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
 885                                  ++session->s_renew_seq);
 886         if (IS_ERR(msg))
 887                 return PTR_ERR(msg);
 888         ceph_con_send(&session->s_con, msg);
 889         return 0;
 890 }
 891
 892 /*
 893  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
 894  *
 895  * Called under session->s_mutex
 896  */
 897 static void renewed_caps(struct ceph_mds_client *mdsc,
 898                          struct ceph_mds_session *session, int is_renew)
 899 {
 900         int was_stale;
 901         int wake = 0;
 902
 903         spin_lock(&session->s_cap_lock);
 904         was_stale = is_renew && (session->s_cap_ttl == 0 ||
 905                                  time_after_eq(jiffies, session->s_cap_ttl));
 906
 907         session->s_cap_ttl = session->s_renew_requested +
 908                 mdsc->mdsmap->m_session_timeout*HZ;
 909
 910         if (was_stale) {
 911                 if (time_before(jiffies, session->s_cap_ttl)) {
 912                         pr_info("mds%d caps renewed\n", session->s_mds);
 913                         wake = 1;
 914                 } else {
 915                         pr_info("mds%d caps still stale\n", session->s_mds);
 916                 }
 917         }
 918         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
 919              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
 920              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
 921         spin_unlock(&session->s_cap_lock);
 922
 923         if (wake)
 924                 wake_up_session_caps(session, 0);
 925 }
 926
 927 /*
 928  * send a session close request
 929  */
 930 static int request_close_session(struct ceph_mds_client *mdsc,
 931                                  struct ceph_mds_session *session)
 932 {
 933         struct ceph_msg *msg;
 934         int err = 0;
 935
 936         dout("request_close_session mds%d state %s seq %lld\n",
 937              session->s_mds, session_state_name(session->s_state),
 938              session->s_seq);
 939         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
 940         if (IS_ERR(msg))
 941                 err = PTR_ERR(msg);
 942         else
 943                 ceph_con_send(&session->s_con, msg);
 944         return err;
 945 }
 946
 947 /*
 948  * Called with s_mutex held.
 949  */
 950 static int __close_session(struct ceph_mds_client *mdsc,
 951                          struct ceph_mds_session *session)
 952 {
 953         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
 954                 return 0;
 955         session->s_state = CEPH_MDS_SESSION_CLOSING;
 956         return request_close_session(mdsc, session);
 957 }
 958
 959 /*
 960  * Trim old(er) caps.
 961  *
 962  * Because we can't cache an inode without one or more caps, we do
 963  * this indirectly: if a cap is unused, we prune its aliases, at which
 964  * point the inode will hopefully get dropped to.
 965  *
 966  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
 967  * memory pressure from the MDS, though, so it needn't be perfect.
 968  */
 969 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
 970 {
 971         struct ceph_mds_session *session = arg;
 972         struct ceph_inode_info *ci = ceph_inode(inode);
 973         int used, oissued, mine;
 974
 975         if (session->s_trim_caps <= 0)
 976                 return -1;
 977
 978         spin_lock(&inode->i_lock);
 979         mine = cap->issued | cap->implemented;
 980         used = __ceph_caps_used(ci);
 981         oissued = __ceph_caps_issued_other(ci, cap);
 982
 983         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
 984              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
 985              ceph_cap_string(used));
 986         if (ci->i_dirty_caps)
 987                 goto out;   /* dirty caps */
 988         if ((used & ~oissued) & mine)
 989                 goto out;   /* we need these caps */
 990
 991         session->s_trim_caps--;
 992         if (oissued) {
 993                 /* we aren't the only cap.. just remove us */
 994                 __ceph_remove_cap(cap);
 995         } else {
 996                 /* try to drop referring dentries */
 997                 spin_unlock(&inode->i_lock);
 998                 d_prune_aliases(inode);
 999                 dout("trim_caps_cb %p cap %p  pruned, count now %d\n",
1000                      inode, cap, atomic_read(&inode->i_count));
1001                 return 0;
1002         }
1003
1004 out:
1005         spin_unlock(&inode->i_lock);
1006         return 0;
1007 }
1008
1009 /*
1010  * Trim session cap count down to some max number.
1011  */
1012 static int trim_caps(struct ceph_mds_client *mdsc,
1013                      struct ceph_mds_session *session,
1014                      int max_caps)
1015 {
1016         int trim_caps = session->s_nr_caps - max_caps;
1017
1018         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1019              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1020         if (trim_caps > 0) {
1021                 session->s_trim_caps = trim_caps;
1022                 iterate_session_caps(session, trim_caps_cb, session);
1023                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1024                      session->s_mds, session->s_nr_caps, max_caps,
1025                         trim_caps - session->s_trim_caps);
1026                 session->s_trim_caps = 0;
1027         }
1028         return 0;
1029 }
1030
1031 /*
1032  * Allocate cap_release messages.  If there is a partially full message
1033  * in the queue, try to allocate enough to cover it's remainder, so that
1034  * we can send it immediately.
1035  *
1036  * Called under s_mutex.
1037  */
1038 static int add_cap_releases(struct ceph_mds_client *mdsc,
1039                             struct ceph_mds_session *session,
1040                             int extra)
1041 {
1042         struct ceph_msg *msg;
1043         struct ceph_mds_cap_release *head;
1044         int err = -ENOMEM;
1045
1046         if (extra < 0)
1047                 extra = mdsc->client->mount_args->cap_release_safety;
1048
1049         spin_lock(&session->s_cap_lock);
1050
1051         if (!list_empty(&session->s_cap_releases)) {
1052                 msg = list_first_entry(&session->s_cap_releases,
1053                                        struct ceph_msg,
1054                                  list_head);
1055                 head = msg->front.iov_base;
1056                 extra += CEPH_CAPS_PER_RELEASE - le32_to_cpu(head->num);
1057         }
1058
1059         while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1060                 spin_unlock(&session->s_cap_lock);
1061                 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
1062                                    0, 0, NULL);
1063                 if (!msg)
1064                         goto out_unlocked;
1065                 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1066                      (int)msg->front.iov_len);
1067                 head = msg->front.iov_base;
1068                 head->num = cpu_to_le32(0);
1069                 msg->front.iov_len = sizeof(*head);
1070                 spin_lock(&session->s_cap_lock);
1071                 list_add(&msg->list_head, &session->s_cap_releases);
1072                 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1073         }
1074
1075         if (!list_empty(&session->s_cap_releases)) {
1076                 msg = list_first_entry(&session->s_cap_releases,
1077                                        struct ceph_msg,
1078                                        list_head);
1079                 head = msg->front.iov_base;
1080                 if (head->num) {
1081                         dout(" queueing non-full %p (%d)\n", msg,
1082                              le32_to_cpu(head->num));
1083                         list_move_tail(&msg->list_head,
1084                                       &session->s_cap_releases_done);
1085                         session->s_num_cap_releases -=
1086                                 CEPH_CAPS_PER_RELEASE - le32_to_cpu(head->num);
1087                 }
1088         }
1089         err = 0;
1090         spin_unlock(&session->s_cap_lock);
1091 out_unlocked:
1092         return err;
1093 }
1094
1095 /*
1096  * flush all dirty inode data to disk.
1097  *
1098  * returns true if we've flushed through want_flush_seq
1099  */
1100 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1101 {
1102         int mds, ret = 1;
1103
1104         dout("check_cap_flush want %lld\n", want_flush_seq);
1105         mutex_lock(&mdsc->mutex);
1106         for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1107                 struct ceph_mds_session *session = mdsc->sessions[mds];
1108
1109                 if (!session)
1110                         continue;
1111                 get_session(session);
1112                 mutex_unlock(&mdsc->mutex);
1113
1114                 mutex_lock(&session->s_mutex);
1115                 if (!list_empty(&session->s_cap_flushing)) {
1116                         struct ceph_inode_info *ci =
1117                                 list_entry(session->s_cap_flushing.next,
1118                                            struct ceph_inode_info,
1119                                            i_flushing_item);
1120                         struct inode *inode = &ci->vfs_inode;
1121
1122                         spin_lock(&inode->i_lock);
1123                         if (ci->i_cap_flush_seq <= want_flush_seq) {
1124                                 dout("check_cap_flush still flushing %p "
1125                                      "seq %lld <= %lld to mds%d\n", inode,
1126                                      ci->i_cap_flush_seq, want_flush_seq,
1127                                      session->s_mds);
1128                                 ret = 0;
1129                         }
1130                         spin_unlock(&inode->i_lock);
1131                 }
1132                 mutex_unlock(&session->s_mutex);
1133                 ceph_put_mds_session(session);
1134
1135                 if (!ret)
1136                         return ret;
1137                 mutex_lock(&mdsc->mutex);
1138         }
1139
1140         mutex_unlock(&mdsc->mutex);
1141         dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1142         return ret;
1143 }
1144
1145 /*
1146  * called under s_mutex
1147  */
1148 static void send_cap_releases(struct ceph_mds_client *mdsc,
1149                        struct ceph_mds_session *session)
1150 {
1151         struct ceph_msg *msg;
1152
1153         dout("send_cap_releases mds%d\n", session->s_mds);
1154         while (1) {
1155                 spin_lock(&session->s_cap_lock);
1156                 if (list_empty(&session->s_cap_releases_done))
1157                         break;
1158                 msg = list_first_entry(&session->s_cap_releases_done,
1159                                  struct ceph_msg, list_head);
1160                 list_del_init(&msg->list_head);
1161                 spin_unlock(&session->s_cap_lock);
1162                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1163                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1164                 ceph_con_send(&session->s_con, msg);
1165         }
1166         spin_unlock(&session->s_cap_lock);
1167 }
1168
1169 /*
1170  * requests
1171  */
1172
1173 /*
1174  * Create an mds request.
1175  */
1176 struct ceph_mds_request *
1177 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1178 {
1179         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1180
1181         if (!req)
1182                 return ERR_PTR(-ENOMEM);
1183
1184         mutex_init(&req->r_fill_mutex);
1185         req->r_started = jiffies;
1186         req->r_resend_mds = -1;
1187         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1188         req->r_fmode = -1;
1189         kref_init(&req->r_kref);
1190         INIT_LIST_HEAD(&req->r_wait);
1191         init_completion(&req->r_completion);
1192         init_completion(&req->r_safe_completion);
1193         INIT_LIST_HEAD(&req->r_unsafe_item);
1194
1195         req->r_op = op;
1196         req->r_direct_mode = mode;
1197         return req;
1198 }
1199
1200 /*
1201  * return oldest (lowest) request, tid in request tree, 0 if none.
1202  *
1203  * called under mdsc->mutex.
1204  */
1205 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1206 {
1207         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1208                 return NULL;
1209         return rb_entry(rb_first(&mdsc->request_tree),
1210                         struct ceph_mds_request, r_node);
1211 }
1212
1213 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1214 {
1215         struct ceph_mds_request *req = __get_oldest_req(mdsc);
1216
1217         if (req)
1218                 return req->r_tid;
1219         return 0;
1220 }
1221
1222 /*
1223  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1224  * on build_path_from_dentry in fs/cifs/dir.c.
1225  *
1226  * If @stop_on_nosnap, generate path relative to the first non-snapped
1227  * inode.
1228  *
1229  * Encode hidden .snap dirs as a double /, i.e.
1230  *   foo/.snap/bar -> foo//bar
1231  */
1232 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1233                            int stop_on_nosnap)
1234 {
1235         struct dentry *temp;
1236         char *path;
1237         int len, pos;
1238
1239         if (dentry == NULL)
1240                 return ERR_PTR(-EINVAL);
1241
1242 retry:
1243         len = 0;
1244         for (temp = dentry; !IS_ROOT(temp);) {
1245                 struct inode *inode = temp->d_inode;
1246                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1247                         len++;  /* slash only */
1248                 else if (stop_on_nosnap && inode &&
1249                          ceph_snap(inode) == CEPH_NOSNAP)
1250                         break;
1251                 else
1252                         len += 1 + temp->d_name.len;
1253                 temp = temp->d_parent;
1254                 if (temp == NULL) {
1255                         pr_err("build_path_dentry corrupt dentry %p\n", dentry);
1256                         return ERR_PTR(-EINVAL);
1257                 }
1258         }
1259         if (len)
1260                 len--;  /* no leading '/' */
1261
1262         path = kmalloc(len+1, GFP_NOFS);
1263         if (path == NULL)
1264                 return ERR_PTR(-ENOMEM);
1265         pos = len;
1266         path[pos] = 0;  /* trailing null */
1267         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1268                 struct inode *inode = temp->d_inode;
1269
1270                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1271                         dout("build_path_dentry path+%d: %p SNAPDIR\n",
1272                              pos, temp);
1273                 } else if (stop_on_nosnap && inode &&
1274                            ceph_snap(inode) == CEPH_NOSNAP) {
1275                         break;
1276                 } else {
1277                         pos -= temp->d_name.len;
1278                         if (pos < 0)
1279                                 break;
1280                         strncpy(path + pos, temp->d_name.name,
1281                                 temp->d_name.len);
1282                         dout("build_path_dentry path+%d: %p '%.*s'\n",
1283                              pos, temp, temp->d_name.len, path + pos);
1284                 }
1285                 if (pos)
1286                         path[--pos] = '/';
1287                 temp = temp->d_parent;
1288                 if (temp == NULL) {
1289                         pr_err("build_path_dentry corrupt dentry\n");
1290                         kfree(path);
1291                         return ERR_PTR(-EINVAL);
1292                 }
1293         }
1294         if (pos != 0) {
1295                 pr_err("build_path_dentry did not end path lookup where "
1296                        "expected, namelen is %d, pos is %d\n", len, pos);
1297                 /* presumably this is only possible if racing with a
1298                    rename of one of the parent directories (we can not
1299                    lock the dentries above us to prevent this, but
1300                    retrying should be harmless) */
1301                 kfree(path);
1302                 goto retry;
1303         }
1304
1305         *base = ceph_ino(temp->d_inode);
1306         *plen = len;
1307         dout("build_path_dentry on %p %d built %llx '%.*s'\n",
1308              dentry, atomic_read(&dentry->d_count), *base, len, path);
1309         return path;
1310 }
1311
1312 static int build_dentry_path(struct dentry *dentry,
1313                              const char **ppath, int *ppathlen, u64 *pino,
1314                              int *pfreepath)
1315 {
1316         char *path;
1317
1318         if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1319                 *pino = ceph_ino(dentry->d_parent->d_inode);
1320                 *ppath = dentry->d_name.name;
1321                 *ppathlen = dentry->d_name.len;
1322                 return 0;
1323         }
1324         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1325         if (IS_ERR(path))
1326                 return PTR_ERR(path);
1327         *ppath = path;
1328         *pfreepath = 1;
1329         return 0;
1330 }
1331
1332 static int build_inode_path(struct inode *inode,
1333                             const char **ppath, int *ppathlen, u64 *pino,
1334                             int *pfreepath)
1335 {
1336         struct dentry *dentry;
1337         char *path;
1338
1339         if (ceph_snap(inode) == CEPH_NOSNAP) {
1340                 *pino = ceph_ino(inode);
1341                 *ppathlen = 0;
1342                 return 0;
1343         }
1344         dentry = d_find_alias(inode);
1345         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1346         dput(dentry);
1347         if (IS_ERR(path))
1348                 return PTR_ERR(path);
1349         *ppath = path;
1350         *pfreepath = 1;
1351         return 0;
1352 }
1353
1354 /*
1355  * request arguments may be specified via an inode *, a dentry *, or
1356  * an explicit ino+path.
1357  */
1358 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1359                                   const char *rpath, u64 rino,
1360                                   const char **ppath, int *pathlen,
1361                                   u64 *ino, int *freepath)
1362 {
1363         int r = 0;
1364
1365         if (rinode) {
1366                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1367                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1368                      ceph_snap(rinode));
1369         } else if (rdentry) {
1370                 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1371                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1372                      *ppath);
1373         } else if (rpath) {
1374                 *ino = rino;
1375                 *ppath = rpath;
1376                 *pathlen = strlen(rpath);
1377                 dout(" path %.*s\n", *pathlen, rpath);
1378         }
1379
1380         return r;
1381 }
1382
1383 /*
1384  * called under mdsc->mutex
1385  */
1386 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1387                                                struct ceph_mds_request *req,
1388                                                int mds)
1389 {
1390         struct ceph_msg *msg;
1391         struct ceph_mds_request_head *head;
1392         const char *path1 = NULL;
1393         const char *path2 = NULL;
1394         u64 ino1 = 0, ino2 = 0;
1395         int pathlen1 = 0, pathlen2 = 0;
1396         int freepath1 = 0, freepath2 = 0;
1397         int len;
1398         u16 releases;
1399         void *p, *end;
1400         int ret;
1401
1402         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1403                               req->r_path1, req->r_ino1.ino,
1404                               &path1, &pathlen1, &ino1, &freepath1);
1405         if (ret < 0) {
1406                 msg = ERR_PTR(ret);
1407                 goto out;
1408         }
1409
1410         ret = set_request_path_attr(NULL, req->r_old_dentry,
1411                               req->r_path2, req->r_ino2.ino,
1412                               &path2, &pathlen2, &ino2, &freepath2);
1413         if (ret < 0) {
1414                 msg = ERR_PTR(ret);
1415                 goto out_free1;
1416         }
1417
1418         len = sizeof(*head) +
1419                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64));
1420
1421         /* calculate (max) length for cap releases */
1422         len += sizeof(struct ceph_mds_request_release) *
1423                 (!!req->r_inode_drop + !!req->r_dentry_drop +
1424                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1425         if (req->r_dentry_drop)
1426                 len += req->r_dentry->d_name.len;
1427         if (req->r_old_dentry_drop)
1428                 len += req->r_old_dentry->d_name.len;
1429
1430         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, 0, 0, NULL);
1431         if (IS_ERR(msg))
1432                 goto out_free2;
1433
1434         msg->hdr.tid = cpu_to_le64(req->r_tid);
1435
1436         head = msg->front.iov_base;
1437         p = msg->front.iov_base + sizeof(*head);
1438         end = msg->front.iov_base + msg->front.iov_len;
1439
1440         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1441         head->op = cpu_to_le32(req->r_op);
1442         head->caller_uid = cpu_to_le32(current_fsuid());
1443         head->caller_gid = cpu_to_le32(current_fsgid());
1444         head->args = req->r_args;
1445
1446         ceph_encode_filepath(&p, end, ino1, path1);
1447         ceph_encode_filepath(&p, end, ino2, path2);
1448
1449         /* cap releases */
1450         releases = 0;
1451         if (req->r_inode_drop)
1452                 releases += ceph_encode_inode_release(&p,
1453                       req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1454                       mds, req->r_inode_drop, req->r_inode_unless, 0);
1455         if (req->r_dentry_drop)
1456                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1457                        mds, req->r_dentry_drop, req->r_dentry_unless);
1458         if (req->r_old_dentry_drop)
1459                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1460                        mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1461         if (req->r_old_inode_drop)
1462                 releases += ceph_encode_inode_release(&p,
1463                       req->r_old_dentry->d_inode,
1464                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1465         head->num_releases = cpu_to_le16(releases);
1466
1467         BUG_ON(p > end);
1468         msg->front.iov_len = p - msg->front.iov_base;
1469         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1470
1471         msg->pages = req->r_pages;
1472         msg->nr_pages = req->r_num_pages;
1473         msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1474         msg->hdr.data_off = cpu_to_le16(0);
1475
1476 out_free2:
1477         if (freepath2)
1478                 kfree((char *)path2);
1479 out_free1:
1480         if (freepath1)
1481                 kfree((char *)path1);
1482 out:
1483         return msg;
1484 }
1485
1486 /*
1487  * called under mdsc->mutex if error, under no mutex if
1488  * success.
1489  */
1490 static void complete_request(struct ceph_mds_client *mdsc,
1491                              struct ceph_mds_request *req)
1492 {
1493         if (req->r_callback)
1494                 req->r_callback(mdsc, req);
1495         else
1496                 complete(&req->r_completion);
1497 }
1498
1499 /*
1500  * called under mdsc->mutex
1501  */
1502 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1503                                   struct ceph_mds_request *req,
1504                                   int mds)
1505 {
1506         struct ceph_mds_request_head *rhead;
1507         struct ceph_msg *msg;
1508         int flags = 0;
1509
1510         req->r_mds = mds;
1511         req->r_attempts++;
1512         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1513              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1514
1515         if (req->r_request) {
1516                 ceph_msg_put(req->r_request);
1517                 req->r_request = NULL;
1518         }
1519         msg = create_request_message(mdsc, req, mds);
1520         if (IS_ERR(msg)) {
1521                 req->r_err = PTR_ERR(msg);
1522                 complete_request(mdsc, req);
1523                 return -PTR_ERR(msg);
1524         }
1525         req->r_request = msg;
1526
1527         rhead = msg->front.iov_base;
1528         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1529         if (req->r_got_unsafe)
1530                 flags |= CEPH_MDS_FLAG_REPLAY;
1531         if (req->r_locked_dir)
1532                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1533         rhead->flags = cpu_to_le32(flags);
1534         rhead->num_fwd = req->r_num_fwd;
1535         rhead->num_retry = req->r_attempts - 1;
1536
1537         dout(" r_locked_dir = %p\n", req->r_locked_dir);
1538
1539         if (req->r_target_inode && req->r_got_unsafe)
1540                 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1541         else
1542                 rhead->ino = 0;
1543         return 0;
1544 }
1545
1546 /*
1547  * send request, or put it on the appropriate wait list.
1548  */
1549 static int __do_request(struct ceph_mds_client *mdsc,
1550                         struct ceph_mds_request *req)
1551 {
1552         struct ceph_mds_session *session = NULL;
1553         int mds = -1;
1554         int err = -EAGAIN;
1555
1556         if (req->r_err || req->r_got_result)
1557                 goto out;
1558
1559         if (req->r_timeout &&
1560             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1561                 dout("do_request timed out\n");
1562                 err = -EIO;
1563                 goto finish;
1564         }
1565
1566         mds = __choose_mds(mdsc, req);
1567         if (mds < 0 ||
1568             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1569                 dout("do_request no mds or not active, waiting for map\n");
1570                 list_add(&req->r_wait, &mdsc->waiting_for_map);
1571                 goto out;
1572         }
1573
1574         /* get, open session */
1575         session = __ceph_lookup_mds_session(mdsc, mds);
1576         if (!session) {
1577                 session = register_session(mdsc, mds);
1578                 if (IS_ERR(session)) {
1579                         err = PTR_ERR(session);
1580                         goto finish;
1581                 }
1582         }
1583         dout("do_request mds%d session %p state %s\n", mds, session,
1584              session_state_name(session->s_state));
1585         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1586             session->s_state != CEPH_MDS_SESSION_HUNG) {
1587                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1588                     session->s_state == CEPH_MDS_SESSION_CLOSING)
1589                         __open_session(mdsc, session);
1590                 list_add(&req->r_wait, &session->s_waiting);
1591                 goto out_session;
1592         }
1593
1594         /* send request */
1595         req->r_session = get_session(session);
1596         req->r_resend_mds = -1;   /* forget any previous mds hint */
1597
1598         if (req->r_request_started == 0)   /* note request start time */
1599                 req->r_request_started = jiffies;
1600
1601         err = __prepare_send_request(mdsc, req, mds);
1602         if (!err) {
1603                 ceph_msg_get(req->r_request);
1604                 ceph_con_send(&session->s_con, req->r_request);
1605         }
1606
1607 out_session:
1608         ceph_put_mds_session(session);
1609 out:
1610         return err;
1611
1612 finish:
1613         req->r_err = err;
1614         complete_request(mdsc, req);
1615         goto out;
1616 }
1617
1618 /*
1619  * called under mdsc->mutex
1620  */
1621 static void __wake_requests(struct ceph_mds_client *mdsc,
1622                             struct list_head *head)
1623 {
1624         struct ceph_mds_request *req, *nreq;
1625
1626         list_for_each_entry_safe(req, nreq, head, r_wait) {
1627                 list_del_init(&req->r_wait);
1628                 __do_request(mdsc, req);
1629         }
1630 }
1631
1632 /*
1633  * Wake up threads with requests pending for @mds, so that they can
1634  * resubmit their requests to a possibly different mds.  If @all is set,
1635  * wake up if their requests has been forwarded to @mds, too.
1636  */
1637 static void kick_requests(struct ceph_mds_client *mdsc, int mds, int all)
1638 {
1639         struct ceph_mds_request *req;
1640         struct rb_node *p;
1641
1642         dout("kick_requests mds%d\n", mds);
1643         for (p = rb_first(&mdsc->request_tree); p; p = rb_next(p)) {
1644                 req = rb_entry(p, struct ceph_mds_request, r_node);
1645                 if (req->r_got_unsafe)
1646                         continue;
1647                 if (req->r_session &&
1648                     req->r_session->s_mds == mds) {
1649                         dout(" kicking tid %llu\n", req->r_tid);
1650                         put_request_session(req);
1651                         __do_request(mdsc, req);
1652                 }
1653         }
1654 }
1655
1656 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
1657                               struct ceph_mds_request *req)
1658 {
1659         dout("submit_request on %p\n", req);
1660         mutex_lock(&mdsc->mutex);
1661         __register_request(mdsc, req, NULL);
1662         __do_request(mdsc, req);
1663         mutex_unlock(&mdsc->mutex);
1664 }
1665
1666 /*
1667  * Synchrously perform an mds request.  Take care of all of the
1668  * session setup, forwarding, retry details.
1669  */
1670 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
1671                          struct inode *dir,
1672                          struct ceph_mds_request *req)
1673 {
1674         int err;
1675
1676         dout("do_request on %p\n", req);
1677
1678         /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
1679         if (req->r_inode)
1680                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
1681         if (req->r_locked_dir)
1682                 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
1683         if (req->r_old_dentry)
1684                 ceph_get_cap_refs(
1685                         ceph_inode(req->r_old_dentry->d_parent->d_inode),
1686                         CEPH_CAP_PIN);
1687
1688         /* issue */
1689         mutex_lock(&mdsc->mutex);
1690         __register_request(mdsc, req, dir);
1691         __do_request(mdsc, req);
1692
1693         if (req->r_err) {
1694                 err = req->r_err;
1695                 __unregister_request(mdsc, req);
1696                 dout("do_request early error %d\n", err);
1697                 goto out;
1698         }
1699
1700         /* wait */
1701         mutex_unlock(&mdsc->mutex);
1702         dout("do_request waiting\n");
1703         if (req->r_timeout) {
1704                 err = (long)wait_for_completion_interruptible_timeout(
1705                         &req->r_completion, req->r_timeout);
1706                 if (err == 0)
1707                         err = -EIO;
1708         } else {
1709                 err = wait_for_completion_interruptible(&req->r_completion);
1710         }
1711         dout("do_request waited, got %d\n", err);
1712         mutex_lock(&mdsc->mutex);
1713
1714         /* only abort if we didn't race with a real reply */
1715         if (req->r_got_result) {
1716                 err = le32_to_cpu(req->r_reply_info.head->result);
1717         } else if (err < 0) {
1718                 dout("aborted request %lld with %d\n", req->r_tid, err);
1719
1720                 /*
1721                  * ensure we aren't running concurrently with
1722                  * ceph_fill_trace or ceph_readdir_prepopulate, which
1723                  * rely on locks (dir mutex) held by our caller.
1724                  */
1725                 mutex_lock(&req->r_fill_mutex);
1726                 req->r_err = err;
1727                 req->r_aborted = true;
1728                 mutex_unlock(&req->r_fill_mutex);
1729
1730                 if (req->r_locked_dir &&
1731                     (req->r_op & CEPH_MDS_OP_WRITE)) {
1732                         struct ceph_inode_info *ci =
1733                                 ceph_inode(req->r_locked_dir);
1734
1735                         dout("aborted, clearing I_COMPLETE on %p\n",
1736                              req->r_locked_dir);
1737                         spin_lock(&req->r_locked_dir->i_lock);
1738                         ci->i_ceph_flags &= ~CEPH_I_COMPLETE;
1739                         ci->i_release_count++;
1740                         spin_unlock(&req->r_locked_dir->i_lock);
1741                 }
1742         } else {
1743                 err = req->r_err;
1744         }
1745
1746 out:
1747         mutex_unlock(&mdsc->mutex);
1748         dout("do_request %p done, result %d\n", req, err);
1749         return err;
1750 }
1751
1752 /*
1753  * Handle mds reply.
1754  *
1755  * We take the session mutex and parse and process the reply immediately.
1756  * This preserves the logical ordering of replies, capabilities, etc., sent
1757  * by the MDS as they are applied to our local cache.
1758  */
1759 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
1760 {
1761         struct ceph_mds_client *mdsc = session->s_mdsc;
1762         struct ceph_mds_request *req;
1763         struct ceph_mds_reply_head *head = msg->front.iov_base;
1764         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
1765         u64 tid;
1766         int err, result;
1767         int mds = session->s_mds;
1768
1769         if (msg->front.iov_len < sizeof(*head)) {
1770                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
1771                 ceph_msg_dump(msg);
1772                 return;
1773         }
1774
1775         /* get request, session */
1776         tid = le64_to_cpu(msg->hdr.tid);
1777         mutex_lock(&mdsc->mutex);
1778         req = __lookup_request(mdsc, tid);
1779         if (!req) {
1780                 dout("handle_reply on unknown tid %llu\n", tid);
1781                 mutex_unlock(&mdsc->mutex);
1782                 return;
1783         }
1784         dout("handle_reply %p\n", req);
1785
1786         /* correct session? */
1787         if (req->r_session != session) {
1788                 pr_err("mdsc_handle_reply got %llu on session mds%d"
1789                        " not mds%d\n", tid, session->s_mds,
1790                        req->r_session ? req->r_session->s_mds : -1);
1791                 mutex_unlock(&mdsc->mutex);
1792                 goto out;
1793         }
1794
1795         /* dup? */
1796         if ((req->r_got_unsafe && !head->safe) ||
1797             (req->r_got_safe && head->safe)) {
1798                 pr_warning("got a dup %s reply on %llu from mds%d\n",
1799                            head->safe ? "safe" : "unsafe", tid, mds);
1800                 mutex_unlock(&mdsc->mutex);
1801                 goto out;
1802         }
1803
1804         result = le32_to_cpu(head->result);
1805
1806         /*
1807          * Tolerate 2 consecutive ESTALEs from the same mds.
1808          * FIXME: we should be looking at the cap migrate_seq.
1809          */
1810         if (result == -ESTALE) {
1811                 req->r_direct_mode = USE_AUTH_MDS;
1812                 req->r_num_stale++;
1813                 if (req->r_num_stale <= 2) {
1814                         __do_request(mdsc, req);
1815                         mutex_unlock(&mdsc->mutex);
1816                         goto out;
1817                 }
1818         } else {
1819                 req->r_num_stale = 0;
1820         }
1821
1822         if (head->safe) {
1823                 req->r_got_safe = true;
1824                 __unregister_request(mdsc, req);
1825                 complete(&req->r_safe_completion);
1826
1827                 if (req->r_got_unsafe) {
1828                         /*
1829                          * We already handled the unsafe response, now do the
1830                          * cleanup.  No need to examine the response; the MDS
1831                          * doesn't include any result info in the safe
1832                          * response.  And even if it did, there is nothing
1833                          * useful we could do with a revised return value.
1834                          */
1835                         dout("got safe reply %llu, mds%d\n", tid, mds);
1836                         list_del_init(&req->r_unsafe_item);
1837
1838                         /* last unsafe request during umount? */
1839                         if (mdsc->stopping && !__get_oldest_req(mdsc))
1840                                 complete(&mdsc->safe_umount_waiters);
1841                         mutex_unlock(&mdsc->mutex);
1842                         goto out;
1843                 }
1844         } else {
1845                 req->r_got_unsafe = true;
1846                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
1847         }
1848
1849         dout("handle_reply tid %lld result %d\n", tid, result);
1850         rinfo = &req->r_reply_info;
1851         err = parse_reply_info(msg, rinfo);
1852         mutex_unlock(&mdsc->mutex);
1853
1854         mutex_lock(&session->s_mutex);
1855         if (err < 0) {
1856                 pr_err("mdsc_handle_reply got corrupt reply mds%d\n", mds);
1857                 ceph_msg_dump(msg);
1858                 goto out_err;
1859         }
1860
1861         /* snap trace */
1862         if (rinfo->snapblob_len) {
1863                 down_write(&mdsc->snap_rwsem);
1864                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
1865                                rinfo->snapblob + rinfo->snapblob_len,
1866                                le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
1867                 downgrade_write(&mdsc->snap_rwsem);
1868         } else {
1869                 down_read(&mdsc->snap_rwsem);
1870         }
1871
1872         /* insert trace into our cache */
1873         mutex_lock(&req->r_fill_mutex);
1874         err = ceph_fill_trace(mdsc->client->sb, req, req->r_session);
1875         if (err == 0) {
1876                 if (result == 0 && rinfo->dir_nr)
1877                         ceph_readdir_prepopulate(req, req->r_session);
1878                 ceph_unreserve_caps(&req->r_caps_reservation);
1879         }
1880         mutex_unlock(&req->r_fill_mutex);
1881
1882         up_read(&mdsc->snap_rwsem);
1883 out_err:
1884         mutex_lock(&mdsc->mutex);
1885         if (!req->r_aborted) {
1886                 if (err) {
1887                         req->r_err = err;
1888                 } else {
1889                         req->r_reply = msg;
1890                         ceph_msg_get(msg);
1891                         req->r_got_result = true;
1892                 }
1893         } else {
1894                 dout("reply arrived after request %lld was aborted\n", tid);
1895         }
1896         mutex_unlock(&mdsc->mutex);
1897
1898         add_cap_releases(mdsc, req->r_session, -1);
1899         mutex_unlock(&session->s_mutex);
1900
1901         /* kick calling process */
1902         complete_request(mdsc, req);
1903 out:
1904         ceph_mdsc_put_request(req);
1905         return;
1906 }
1907
1908
1909
1910 /*
1911  * handle mds notification that our request has been forwarded.
1912  */
1913 static void handle_forward(struct ceph_mds_client *mdsc,
1914                            struct ceph_mds_session *session,
1915                            struct ceph_msg *msg)
1916 {
1917         struct ceph_mds_request *req;
1918         u64 tid = le64_to_cpu(msg->hdr.tid);
1919         u32 next_mds;
1920         u32 fwd_seq;
1921         int err = -EINVAL;
1922         void *p = msg->front.iov_base;
1923         void *end = p + msg->front.iov_len;
1924
1925         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1926         next_mds = ceph_decode_32(&p);
1927         fwd_seq = ceph_decode_32(&p);
1928
1929         mutex_lock(&mdsc->mutex);
1930         req = __lookup_request(mdsc, tid);
1931         if (!req) {
1932                 dout("forward %llu to mds%d - req dne\n", tid, next_mds);
1933                 goto out;  /* dup reply? */
1934         }
1935
1936         if (fwd_seq <= req->r_num_fwd) {
1937                 dout("forward %llu to mds%d - old seq %d <= %d\n",
1938                      tid, next_mds, req->r_num_fwd, fwd_seq);
1939         } else {
1940                 /* resend. forward race not possible; mds would drop */
1941                 dout("forward %llu to mds%d (we resend)\n", tid, next_mds);
1942                 req->r_num_fwd = fwd_seq;
1943                 req->r_resend_mds = next_mds;
1944                 put_request_session(req);
1945                 __do_request(mdsc, req);
1946         }
1947         ceph_mdsc_put_request(req);
1948 out:
1949         mutex_unlock(&mdsc->mutex);
1950         return;
1951
1952 bad:
1953         pr_err("mdsc_handle_forward decode error err=%d\n", err);
1954 }
1955
1956 /*
1957  * handle a mds session control message
1958  */
1959 static void handle_session(struct ceph_mds_session *session,
1960                            struct ceph_msg *msg)
1961 {
1962         struct ceph_mds_client *mdsc = session->s_mdsc;
1963         u32 op;
1964         u64 seq;
1965         int mds = session->s_mds;
1966         struct ceph_mds_session_head *h = msg->front.iov_base;
1967         int wake = 0;
1968
1969         /* decode */
1970         if (msg->front.iov_len != sizeof(*h))
1971                 goto bad;
1972         op = le32_to_cpu(h->op);
1973         seq = le64_to_cpu(h->seq);
1974
1975         mutex_lock(&mdsc->mutex);
1976         if (op == CEPH_SESSION_CLOSE)
1977                 __unregister_session(mdsc, session);
1978         /* FIXME: this ttl calculation is generous */
1979         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
1980         mutex_unlock(&mdsc->mutex);
1981
1982         mutex_lock(&session->s_mutex);
1983
1984         dout("handle_session mds%d %s %p state %s seq %llu\n",
1985              mds, ceph_session_op_name(op), session,
1986              session_state_name(session->s_state), seq);
1987
1988         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
1989                 session->s_state = CEPH_MDS_SESSION_OPEN;
1990                 pr_info("mds%d came back\n", session->s_mds);
1991         }
1992
1993         switch (op) {
1994         case CEPH_SESSION_OPEN:
1995                 session->s_state = CEPH_MDS_SESSION_OPEN;
1996                 renewed_caps(mdsc, session, 0);
1997                 wake = 1;
1998                 if (mdsc->stopping)
1999                         __close_session(mdsc, session);
2000                 break;
2001
2002         case CEPH_SESSION_RENEWCAPS:
2003                 if (session->s_renew_seq == seq)
2004                         renewed_caps(mdsc, session, 1);
2005                 break;
2006
2007         case CEPH_SESSION_CLOSE:
2008                 remove_session_caps(session);
2009                 wake = 1; /* for good measure */
2010                 complete(&mdsc->session_close_waiters);
2011                 kick_requests(mdsc, mds, 0);      /* cur only */
2012                 break;
2013
2014         case CEPH_SESSION_STALE:
2015                 pr_info("mds%d caps went stale, renewing\n",
2016                         session->s_mds);
2017                 spin_lock(&session->s_cap_lock);
2018                 session->s_cap_gen++;
2019                 session->s_cap_ttl = 0;
2020                 spin_unlock(&session->s_cap_lock);
2021                 send_renew_caps(mdsc, session);
2022                 break;
2023
2024         case CEPH_SESSION_RECALL_STATE:
2025                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2026                 break;
2027
2028         default:
2029                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2030                 WARN_ON(1);
2031         }
2032
2033         mutex_unlock(&session->s_mutex);
2034         if (wake) {
2035                 mutex_lock(&mdsc->mutex);
2036                 __wake_requests(mdsc, &session->s_waiting);
2037                 mutex_unlock(&mdsc->mutex);
2038         }
2039         return;
2040
2041 bad:
2042         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2043                (int)msg->front.iov_len);
2044         ceph_msg_dump(msg);
2045         return;
2046 }
2047
2048
2049 /*
2050  * called under session->mutex.
2051  */
2052 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2053                                    struct ceph_mds_session *session)
2054 {
2055         struct ceph_mds_request *req, *nreq;
2056         int err;
2057
2058         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2059
2060         mutex_lock(&mdsc->mutex);
2061         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2062                 err = __prepare_send_request(mdsc, req, session->s_mds);
2063                 if (!err) {
2064                         ceph_msg_get(req->r_request);
2065                         ceph_con_send(&session->s_con, req->r_request);
2066                 }
2067         }
2068         mutex_unlock(&mdsc->mutex);
2069 }
2070
2071 /*
2072  * Encode information about a cap for a reconnect with the MDS.
2073  */
2074 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2075                           void *arg)
2076 {
2077         struct ceph_mds_cap_reconnect rec;
2078         struct ceph_inode_info *ci;
2079         struct ceph_pagelist *pagelist = arg;
2080         char *path;
2081         int pathlen, err;
2082         u64 pathbase;
2083         struct dentry *dentry;
2084
2085         ci = cap->ci;
2086
2087         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2088              inode, ceph_vinop(inode), cap, cap->cap_id,
2089              ceph_cap_string(cap->issued));
2090         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2091         if (err)
2092                 return err;
2093
2094         dentry = d_find_alias(inode);
2095         if (dentry) {
2096                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2097                 if (IS_ERR(path)) {
2098                         err = PTR_ERR(path);
2099                         BUG_ON(err);
2100                 }
2101         } else {
2102                 path = NULL;
2103                 pathlen = 0;
2104         }
2105         err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2106         if (err)
2107                 goto out;
2108
2109         spin_lock(&inode->i_lock);
2110         cap->seq = 0;        /* reset cap seq */
2111         cap->issue_seq = 0;  /* and issue_seq */
2112         rec.cap_id = cpu_to_le64(cap->cap_id);
2113         rec.pathbase = cpu_to_le64(pathbase);
2114         rec.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2115         rec.issued = cpu_to_le32(cap->issued);
2116         rec.size = cpu_to_le64(inode->i_size);
2117         ceph_encode_timespec(&rec.mtime, &inode->i_mtime);
2118         ceph_encode_timespec(&rec.atime, &inode->i_atime);
2119         rec.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2120         spin_unlock(&inode->i_lock);
2121
2122         err = ceph_pagelist_append(pagelist, &rec, sizeof(rec));
2123
2124 out:
2125         kfree(path);
2126         dput(dentry);
2127         return err;
2128 }
2129
2130
2131 /*
2132  * If an MDS fails and recovers, clients need to reconnect in order to
2133  * reestablish shared state.  This includes all caps issued through
2134  * this session _and_ the snap_realm hierarchy.  Because it's not
2135  * clear which snap realms the mds cares about, we send everything we
2136  * know about.. that ensures we'll then get any new info the
2137  * recovering MDS might have.
2138  *
2139  * This is a relatively heavyweight operation, but it's rare.
2140  *
2141  * called with mdsc->mutex held.
2142  */
2143 static void send_mds_reconnect(struct ceph_mds_client *mdsc, int mds)
2144 {
2145         struct ceph_mds_session *session = NULL;
2146         struct ceph_msg *reply;
2147         struct rb_node *p;
2148         int err = -ENOMEM;
2149         struct ceph_pagelist *pagelist;
2150
2151         pr_info("reconnect to recovering mds%d\n", mds);
2152
2153         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2154         if (!pagelist)
2155                 goto fail_nopagelist;
2156         ceph_pagelist_init(pagelist);
2157
2158         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, 0, 0, NULL);
2159         if (IS_ERR(reply)) {
2160                 err = PTR_ERR(reply);
2161                 goto fail_nomsg;
2162         }
2163
2164         /* find session */
2165         session = __ceph_lookup_mds_session(mdsc, mds);
2166         mutex_unlock(&mdsc->mutex);    /* drop lock for duration */
2167
2168         if (session) {
2169                 mutex_lock(&session->s_mutex);
2170
2171                 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2172                 session->s_seq = 0;
2173
2174                 ceph_con_open(&session->s_con,
2175                               ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2176
2177                 /* replay unsafe requests */
2178                 replay_unsafe_requests(mdsc, session);
2179         } else {
2180                 dout("no session for mds%d, will send short reconnect\n",
2181                      mds);
2182         }
2183
2184         down_read(&mdsc->snap_rwsem);
2185
2186         if (!session)
2187                 goto send;
2188         dout("session %p state %s\n", session,
2189              session_state_name(session->s_state));
2190
2191         /* traverse this session's caps */
2192         err = ceph_pagelist_encode_32(pagelist, session->s_nr_caps);
2193         if (err)
2194                 goto fail;
2195         err = iterate_session_caps(session, encode_caps_cb, pagelist);
2196         if (err < 0)
2197                 goto fail;
2198
2199         /*
2200          * snaprealms.  we provide mds with the ino, seq (version), and
2201          * parent for all of our realms.  If the mds has any newer info,
2202          * it will tell us.
2203          */
2204         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2205                 struct ceph_snap_realm *realm =
2206                         rb_entry(p, struct ceph_snap_realm, node);
2207                 struct ceph_mds_snaprealm_reconnect sr_rec;
2208
2209                 dout(" adding snap realm %llx seq %lld parent %llx\n",
2210                      realm->ino, realm->seq, realm->parent_ino);
2211                 sr_rec.ino = cpu_to_le64(realm->ino);
2212                 sr_rec.seq = cpu_to_le64(realm->seq);
2213                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2214                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2215                 if (err)
2216                         goto fail;
2217         }
2218
2219 send:
2220         reply->pagelist = pagelist;
2221         reply->hdr.data_len = cpu_to_le32(pagelist->length);
2222         reply->nr_pages = calc_pages_for(0, pagelist->length);
2223         ceph_con_send(&session->s_con, reply);
2224
2225         session->s_state = CEPH_MDS_SESSION_OPEN;
2226         mutex_unlock(&session->s_mutex);
2227
2228         mutex_lock(&mdsc->mutex);
2229         __wake_requests(mdsc, &session->s_waiting);
2230         mutex_unlock(&mdsc->mutex);
2231
2232         ceph_put_mds_session(session);
2233
2234         up_read(&mdsc->snap_rwsem);
2235         mutex_lock(&mdsc->mutex);
2236         return;
2237
2238 fail:
2239         ceph_msg_put(reply);
2240         up_read(&mdsc->snap_rwsem);
2241         mutex_unlock(&session->s_mutex);
2242         ceph_put_mds_session(session);
2243 fail_nomsg:
2244         ceph_pagelist_release(pagelist);
2245         kfree(pagelist);
2246 fail_nopagelist:
2247         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
2248         mutex_lock(&mdsc->mutex);
2249         return;
2250 }
2251
2252
2253 /*
2254  * compare old and new mdsmaps, kicking requests
2255  * and closing out old connections as necessary
2256  *
2257  * called under mdsc->mutex.
2258  */
2259 static void check_new_map(struct ceph_mds_client *mdsc,
2260                           struct ceph_mdsmap *newmap,
2261                           struct ceph_mdsmap *oldmap)
2262 {
2263         int i;
2264         int oldstate, newstate;
2265         struct ceph_mds_session *s;
2266
2267         dout("check_new_map new %u old %u\n",
2268              newmap->m_epoch, oldmap->m_epoch);
2269
2270         for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2271                 if (mdsc->sessions[i] == NULL)
2272                         continue;
2273                 s = mdsc->sessions[i];
2274                 oldstate = ceph_mdsmap_get_state(oldmap, i);
2275                 newstate = ceph_mdsmap_get_state(newmap, i);
2276
2277                 dout("check_new_map mds%d state %s -> %s (session %s)\n",
2278                      i, ceph_mds_state_name(oldstate),
2279                      ceph_mds_state_name(newstate),
2280                      session_state_name(s->s_state));
2281
2282                 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
2283                            ceph_mdsmap_get_addr(newmap, i),
2284                            sizeof(struct ceph_entity_addr))) {
2285                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2286                                 /* the session never opened, just close it
2287                                  * out now */
2288                                 __wake_requests(mdsc, &s->s_waiting);
2289                                 __unregister_session(mdsc, s);
2290                         } else {
2291                                 /* just close it */
2292                                 mutex_unlock(&mdsc->mutex);
2293                                 mutex_lock(&s->s_mutex);
2294                                 mutex_lock(&mdsc->mutex);
2295                                 ceph_con_close(&s->s_con);
2296                                 mutex_unlock(&s->s_mutex);
2297                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2298                         }
2299
2300                         /* kick any requests waiting on the recovering mds */
2301                         kick_requests(mdsc, i, 1);
2302                 } else if (oldstate == newstate) {
2303                         continue;  /* nothing new with this mds */
2304                 }
2305
2306                 /*
2307                  * send reconnect?
2308                  */
2309                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2310                     newstate >= CEPH_MDS_STATE_RECONNECT)
2311                         send_mds_reconnect(mdsc, i);
2312
2313                 /*
2314                  * kick requests on any mds that has gone active.
2315                  *
2316                  * kick requests on cur or forwarder: we may have sent
2317                  * the request to mds1, mds1 told us it forwarded it
2318                  * to mds2, but then we learn mds1 failed and can't be
2319                  * sure it successfully forwarded our request before
2320                  * it died.
2321                  */
2322                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2323                     newstate >= CEPH_MDS_STATE_ACTIVE) {
2324                         pr_info("mds%d reconnect completed\n", s->s_mds);
2325                         kick_requests(mdsc, i, 1);
2326                         ceph_kick_flushing_caps(mdsc, s);
2327                         wake_up_session_caps(s, 1);
2328                 }
2329         }
2330 }
2331
2332
2333
2334 /*
2335  * leases
2336  */
2337
2338 /*
2339  * caller must hold session s_mutex, dentry->d_lock
2340  */
2341 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2342 {
2343         struct ceph_dentry_info *di = ceph_dentry(dentry);
2344
2345         ceph_put_mds_session(di->lease_session);
2346         di->lease_session = NULL;
2347 }
2348
2349 static void handle_lease(struct ceph_mds_client *mdsc,
2350                          struct ceph_mds_session *session,
2351                          struct ceph_msg *msg)
2352 {
2353         struct super_block *sb = mdsc->client->sb;
2354         struct inode *inode;
2355         struct ceph_inode_info *ci;
2356         struct dentry *parent, *dentry;
2357         struct ceph_dentry_info *di;
2358         int mds = session->s_mds;
2359         struct ceph_mds_lease *h = msg->front.iov_base;
2360         struct ceph_vino vino;
2361         int mask;
2362         struct qstr dname;
2363         int release = 0;
2364
2365         dout("handle_lease from mds%d\n", mds);
2366
2367         /* decode */
2368         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2369                 goto bad;
2370         vino.ino = le64_to_cpu(h->ino);
2371         vino.snap = CEPH_NOSNAP;
2372         mask = le16_to_cpu(h->mask);
2373         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2374         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2375         if (dname.len != get_unaligned_le32(h+1))
2376                 goto bad;
2377
2378         mutex_lock(&session->s_mutex);
2379         session->s_seq++;
2380
2381         /* lookup inode */
2382         inode = ceph_find_inode(sb, vino);
2383         dout("handle_lease '%s', mask %d, ino %llx %p\n",
2384              ceph_lease_op_name(h->action), mask, vino.ino, inode);
2385         if (inode == NULL) {
2386                 dout("handle_lease no inode %llx\n", vino.ino);
2387                 goto release;
2388         }
2389         ci = ceph_inode(inode);
2390
2391         /* dentry */
2392         parent = d_find_alias(inode);
2393         if (!parent) {
2394                 dout("no parent dentry on inode %p\n", inode);
2395                 WARN_ON(1);
2396                 goto release;  /* hrm... */
2397         }
2398         dname.hash = full_name_hash(dname.name, dname.len);
2399         dentry = d_lookup(parent, &dname);
2400         dput(parent);
2401         if (!dentry)
2402                 goto release;
2403
2404         spin_lock(&dentry->d_lock);
2405         di = ceph_dentry(dentry);
2406         switch (h->action) {
2407         case CEPH_MDS_LEASE_REVOKE:
2408                 if (di && di->lease_session == session) {
2409                         h->seq = cpu_to_le32(di->lease_seq);
2410                         __ceph_mdsc_drop_dentry_lease(dentry);
2411                 }
2412                 release = 1;
2413                 break;
2414
2415         case CEPH_MDS_LEASE_RENEW:
2416                 if (di && di->lease_session == session &&
2417                     di->lease_gen == session->s_cap_gen &&
2418                     di->lease_renew_from &&
2419                     di->lease_renew_after == 0) {
2420                         unsigned long duration =
2421                                 le32_to_cpu(h->duration_ms) * HZ / 1000;
2422
2423                         di->lease_seq = le32_to_cpu(h->seq);
2424                         dentry->d_time = di->lease_renew_from + duration;
2425                         di->lease_renew_after = di->lease_renew_from +
2426                                 (duration >> 1);
2427                         di->lease_renew_from = 0;
2428                 }
2429                 break;
2430         }
2431         spin_unlock(&dentry->d_lock);
2432         dput(dentry);
2433
2434         if (!release)
2435                 goto out;
2436
2437 release:
2438         /* let's just reuse the same message */
2439         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2440         ceph_msg_get(msg);
2441         ceph_con_send(&session->s_con, msg);
2442
2443 out:
2444         iput(inode);
2445         mutex_unlock(&session->s_mutex);
2446         return;
2447
2448 bad:
2449         pr_err("corrupt lease message\n");
2450         ceph_msg_dump(msg);
2451 }
2452
2453 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
2454                               struct inode *inode,
2455                               struct dentry *dentry, char action,
2456                               u32 seq)
2457 {
2458         struct ceph_msg *msg;
2459         struct ceph_mds_lease *lease;
2460         int len = sizeof(*lease) + sizeof(u32);
2461         int dnamelen = 0;
2462
2463         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2464              inode, dentry, ceph_lease_op_name(action), session->s_mds);
2465         dnamelen = dentry->d_name.len;
2466         len += dnamelen;
2467
2468         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, 0, 0, NULL);
2469         if (IS_ERR(msg))
2470                 return;
2471         lease = msg->front.iov_base;
2472         lease->action = action;
2473         lease->mask = cpu_to_le16(CEPH_LOCK_DN);
2474         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
2475         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
2476         lease->seq = cpu_to_le32(seq);
2477         put_unaligned_le32(dnamelen, lease + 1);
2478         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
2479
2480         /*
2481          * if this is a preemptive lease RELEASE, no need to
2482          * flush request stream, since the actual request will
2483          * soon follow.
2484          */
2485         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
2486
2487         ceph_con_send(&session->s_con, msg);
2488 }
2489
2490 /*
2491  * Preemptively release a lease we expect to invalidate anyway.
2492  * Pass @inode always, @dentry is optional.
2493  */
2494 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2495                              struct dentry *dentry, int mask)
2496 {
2497         struct ceph_dentry_info *di;
2498         struct ceph_mds_session *session;
2499         u32 seq;
2500
2501         BUG_ON(inode == NULL);
2502         BUG_ON(dentry == NULL);
2503         BUG_ON(mask != CEPH_LOCK_DN);
2504
2505         /* is dentry lease valid? */
2506         spin_lock(&dentry->d_lock);
2507         di = ceph_dentry(dentry);
2508         if (!di || !di->lease_session ||
2509             di->lease_session->s_mds < 0 ||
2510             di->lease_gen != di->lease_session->s_cap_gen ||
2511             !time_before(jiffies, dentry->d_time)) {
2512                 dout("lease_release inode %p dentry %p -- "
2513                      "no lease on %d\n",
2514                      inode, dentry, mask);
2515                 spin_unlock(&dentry->d_lock);
2516                 return;
2517         }
2518
2519         /* we do have a lease on this dentry; note mds and seq */
2520         session = ceph_get_mds_session(di->lease_session);
2521         seq = di->lease_seq;
2522         __ceph_mdsc_drop_dentry_lease(dentry);
2523         spin_unlock(&dentry->d_lock);
2524
2525         dout("lease_release inode %p dentry %p mask %d to mds%d\n",
2526              inode, dentry, mask, session->s_mds);
2527         ceph_mdsc_lease_send_msg(session, inode, dentry,
2528                                  CEPH_MDS_LEASE_RELEASE, seq);
2529         ceph_put_mds_session(session);
2530 }
2531
2532 /*
2533  * drop all leases (and dentry refs) in preparation for umount
2534  */
2535 static void drop_leases(struct ceph_mds_client *mdsc)
2536 {
2537         int i;
2538
2539         dout("drop_leases\n");
2540         mutex_lock(&mdsc->mutex);
2541         for (i = 0; i < mdsc->max_sessions; i++) {
2542                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2543                 if (!s)
2544                         continue;
2545                 mutex_unlock(&mdsc->mutex);
2546                 mutex_lock(&s->s_mutex);
2547                 mutex_unlock(&s->s_mutex);
2548                 ceph_put_mds_session(s);
2549                 mutex_lock(&mdsc->mutex);
2550         }
2551         mutex_unlock(&mdsc->mutex);
2552 }
2553
2554
2555
2556 /*
2557  * delayed work -- periodically trim expired leases, renew caps with mds
2558  */
2559 static void schedule_delayed(struct ceph_mds_client *mdsc)
2560 {
2561         int delay = 5;
2562         unsigned hz = round_jiffies_relative(HZ * delay);
2563         schedule_delayed_work(&mdsc->delayed_work, hz);
2564 }
2565
2566 static void delayed_work(struct work_struct *work)
2567 {
2568         int i;
2569         struct ceph_mds_client *mdsc =
2570                 container_of(work, struct ceph_mds_client, delayed_work.work);
2571         int renew_interval;
2572         int renew_caps;
2573
2574         dout("mdsc delayed_work\n");
2575         ceph_check_delayed_caps(mdsc);
2576
2577         mutex_lock(&mdsc->mutex);
2578         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
2579         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
2580                                    mdsc->last_renew_caps);
2581         if (renew_caps)
2582                 mdsc->last_renew_caps = jiffies;
2583
2584         for (i = 0; i < mdsc->max_sessions; i++) {
2585                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2586                 if (s == NULL)
2587                         continue;
2588                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
2589                         dout("resending session close request for mds%d\n",
2590                              s->s_mds);
2591                         request_close_session(mdsc, s);
2592                         ceph_put_mds_session(s);
2593                         continue;
2594                 }
2595                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
2596                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
2597                                 s->s_state = CEPH_MDS_SESSION_HUNG;
2598                                 pr_info("mds%d hung\n", s->s_mds);
2599                         }
2600                 }
2601                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
2602                         /* this mds is failed or recovering, just wait */
2603                         ceph_put_mds_session(s);
2604                         continue;
2605                 }
2606                 mutex_unlock(&mdsc->mutex);
2607
2608                 mutex_lock(&s->s_mutex);
2609                 if (renew_caps)
2610                         send_renew_caps(mdsc, s);
2611                 else
2612                         ceph_con_keepalive(&s->s_con);
2613                 add_cap_releases(mdsc, s, -1);
2614                 send_cap_releases(mdsc, s);
2615                 mutex_unlock(&s->s_mutex);
2616                 ceph_put_mds_session(s);
2617
2618                 mutex_lock(&mdsc->mutex);
2619         }
2620         mutex_unlock(&mdsc->mutex);
2621
2622         schedule_delayed(mdsc);
2623 }
2624
2625
2626 int ceph_mdsc_init(struct ceph_mds_client *mdsc, struct ceph_client *client)
2627 {
2628         mdsc->client = client;
2629         mutex_init(&mdsc->mutex);
2630         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
2631         init_completion(&mdsc->safe_umount_waiters);
2632         init_completion(&mdsc->session_close_waiters);
2633         INIT_LIST_HEAD(&mdsc->waiting_for_map);
2634         mdsc->sessions = NULL;
2635         mdsc->max_sessions = 0;
2636         mdsc->stopping = 0;
2637         init_rwsem(&mdsc->snap_rwsem);
2638         mdsc->snap_realms = RB_ROOT;
2639         INIT_LIST_HEAD(&mdsc->snap_empty);
2640         spin_lock_init(&mdsc->snap_empty_lock);
2641         mdsc->last_tid = 0;
2642         mdsc->request_tree = RB_ROOT;
2643         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
2644         mdsc->last_renew_caps = jiffies;
2645         INIT_LIST_HEAD(&mdsc->cap_delay_list);
2646         spin_lock_init(&mdsc->cap_delay_lock);
2647         INIT_LIST_HEAD(&mdsc->snap_flush_list);
2648         spin_lock_init(&mdsc->snap_flush_lock);
2649         mdsc->cap_flush_seq = 0;
2650         INIT_LIST_HEAD(&mdsc->cap_dirty);
2651         mdsc->num_cap_flushing = 0;
2652         spin_lock_init(&mdsc->cap_dirty_lock);
2653         init_waitqueue_head(&mdsc->cap_flushing_wq);
2654         spin_lock_init(&mdsc->dentry_lru_lock);
2655         INIT_LIST_HEAD(&mdsc->dentry_lru);
2656         return 0;
2657 }
2658
2659 /*
2660  * Wait for safe replies on open mds requests.  If we time out, drop
2661  * all requests from the tree to avoid dangling dentry refs.
2662  */
2663 static void wait_requests(struct ceph_mds_client *mdsc)
2664 {
2665         struct ceph_mds_request *req;
2666         struct ceph_client *client = mdsc->client;
2667
2668         mutex_lock(&mdsc->mutex);
2669         if (__get_oldest_req(mdsc)) {
2670                 mutex_unlock(&mdsc->mutex);
2671
2672                 dout("wait_requests waiting for requests\n");
2673                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
2674                                     client->mount_args->mount_timeout * HZ);
2675
2676                 /* tear down remaining requests */
2677                 mutex_lock(&mdsc->mutex);
2678                 while ((req = __get_oldest_req(mdsc))) {
2679                         dout("wait_requests timed out on tid %llu\n",
2680                              req->r_tid);
2681                         __unregister_request(mdsc, req);
2682                 }
2683         }
2684         mutex_unlock(&mdsc->mutex);
2685         dout("wait_requests done\n");
2686 }
2687
2688 /*
2689  * called before mount is ro, and before dentries are torn down.
2690  * (hmm, does this still race with new lookups?)
2691  */
2692 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
2693 {
2694         dout("pre_umount\n");
2695         mdsc->stopping = 1;
2696
2697         drop_leases(mdsc);
2698         ceph_flush_dirty_caps(mdsc);
2699         wait_requests(mdsc);
2700 }
2701
2702 /*
2703  * wait for all write mds requests to flush.
2704  */
2705 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
2706 {
2707         struct ceph_mds_request *req = NULL, *nextreq;
2708         struct rb_node *n;
2709
2710         mutex_lock(&mdsc->mutex);
2711         dout("wait_unsafe_requests want %lld\n", want_tid);
2712 restart:
2713         req = __get_oldest_req(mdsc);
2714         while (req && req->r_tid <= want_tid) {
2715                 /* find next request */
2716                 n = rb_next(&req->r_node);
2717                 if (n)
2718                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
2719                 else
2720                         nextreq = NULL;
2721                 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
2722                         /* write op */
2723                         ceph_mdsc_get_request(req);
2724                         if (nextreq)
2725                                 ceph_mdsc_get_request(nextreq);
2726                         mutex_unlock(&mdsc->mutex);
2727                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
2728                              req->r_tid, want_tid);
2729                         wait_for_completion(&req->r_safe_completion);
2730                         mutex_lock(&mdsc->mutex);
2731                         ceph_mdsc_put_request(req);
2732                         if (!nextreq)
2733                                 break;  /* next dne before, so we're done! */
2734                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
2735                                 /* next request was removed from tree */
2736                                 ceph_mdsc_put_request(nextreq);
2737                                 goto restart;
2738                         }
2739                         ceph_mdsc_put_request(nextreq);  /* won't go away */
2740                 }
2741                 req = nextreq;
2742         }
2743         mutex_unlock(&mdsc->mutex);
2744         dout("wait_unsafe_requests done\n");
2745 }
2746
2747 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
2748 {
2749         u64 want_tid, want_flush;
2750
2751         dout("sync\n");
2752         mutex_lock(&mdsc->mutex);
2753         want_tid = mdsc->last_tid;
2754         want_flush = mdsc->cap_flush_seq;
2755         mutex_unlock(&mdsc->mutex);
2756         dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
2757
2758         ceph_flush_dirty_caps(mdsc);
2759
2760         wait_unsafe_requests(mdsc, want_tid);
2761         wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
2762 }
2763
2764
2765 /*
2766  * called after sb is ro.
2767  */
2768 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
2769 {
2770         struct ceph_mds_session *session;
2771         int i;
2772         int n;
2773         struct ceph_client *client = mdsc->client;
2774         unsigned long started, timeout = client->mount_args->mount_timeout * HZ;
2775
2776         dout("close_sessions\n");
2777
2778         mutex_lock(&mdsc->mutex);
2779
2780         /* close sessions */
2781         started = jiffies;
2782         while (time_before(jiffies, started + timeout)) {
2783                 dout("closing sessions\n");
2784                 n = 0;
2785                 for (i = 0; i < mdsc->max_sessions; i++) {
2786                         session = __ceph_lookup_mds_session(mdsc, i);
2787                         if (!session)
2788                                 continue;
2789                         mutex_unlock(&mdsc->mutex);
2790                         mutex_lock(&session->s_mutex);
2791                         __close_session(mdsc, session);
2792                         mutex_unlock(&session->s_mutex);
2793                         ceph_put_mds_session(session);
2794                         mutex_lock(&mdsc->mutex);
2795                         n++;
2796                 }
2797                 if (n == 0)
2798                         break;
2799
2800                 if (client->mount_state == CEPH_MOUNT_SHUTDOWN)
2801                         break;
2802
2803                 dout("waiting for sessions to close\n");
2804                 mutex_unlock(&mdsc->mutex);
2805                 wait_for_completion_timeout(&mdsc->session_close_waiters,
2806                                             timeout);
2807                 mutex_lock(&mdsc->mutex);
2808         }
2809
2810         /* tear down remaining sessions */
2811         for (i = 0; i < mdsc->max_sessions; i++) {
2812                 if (mdsc->sessions[i]) {
2813                         session = get_session(mdsc->sessions[i]);
2814                         __unregister_session(mdsc, session);
2815                         mutex_unlock(&mdsc->mutex);
2816                         mutex_lock(&session->s_mutex);
2817                         remove_session_caps(session);
2818                         mutex_unlock(&session->s_mutex);
2819                         ceph_put_mds_session(session);
2820                         mutex_lock(&mdsc->mutex);
2821                 }
2822         }
2823
2824         WARN_ON(!list_empty(&mdsc->cap_delay_list));
2825
2826         mutex_unlock(&mdsc->mutex);
2827
2828         ceph_cleanup_empty_realms(mdsc);
2829
2830         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2831
2832         dout("stopped\n");
2833 }
2834
2835 void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
2836 {
2837         dout("stop\n");
2838         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2839         if (mdsc->mdsmap)
2840                 ceph_mdsmap_destroy(mdsc->mdsmap);
2841         kfree(mdsc->sessions);
2842 }
2843
2844
2845 /*
2846  * handle mds map update.
2847  */
2848 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
2849 {
2850         u32 epoch;
2851         u32 maplen;
2852         void *p = msg->front.iov_base;
2853         void *end = p + msg->front.iov_len;
2854         struct ceph_mdsmap *newmap, *oldmap;
2855         struct ceph_fsid fsid;
2856         int err = -EINVAL;
2857
2858         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
2859         ceph_decode_copy(&p, &fsid, sizeof(fsid));
2860         if (ceph_check_fsid(mdsc->client, &fsid) < 0)
2861                 return;
2862         epoch = ceph_decode_32(&p);
2863         maplen = ceph_decode_32(&p);
2864         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
2865
2866         /* do we need it? */
2867         ceph_monc_got_mdsmap(&mdsc->client->monc, epoch);
2868         mutex_lock(&mdsc->mutex);
2869         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
2870                 dout("handle_map epoch %u <= our %u\n",
2871                      epoch, mdsc->mdsmap->m_epoch);
2872                 mutex_unlock(&mdsc->mutex);
2873                 return;
2874         }
2875
2876         newmap = ceph_mdsmap_decode(&p, end);
2877         if (IS_ERR(newmap)) {
2878                 err = PTR_ERR(newmap);
2879                 goto bad_unlock;
2880         }
2881
2882         /* swap into place */
2883         if (mdsc->mdsmap) {
2884                 oldmap = mdsc->mdsmap;
2885                 mdsc->mdsmap = newmap;
2886                 check_new_map(mdsc, newmap, oldmap);
2887                 ceph_mdsmap_destroy(oldmap);
2888         } else {
2889                 mdsc->mdsmap = newmap;  /* first mds map */
2890         }
2891         mdsc->client->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
2892
2893         __wake_requests(mdsc, &mdsc->waiting_for_map);
2894
2895         mutex_unlock(&mdsc->mutex);
2896         schedule_delayed(mdsc);
2897         return;
2898
2899 bad_unlock:
2900         mutex_unlock(&mdsc->mutex);
2901 bad:
2902         pr_err("error decoding mdsmap %d\n", err);
2903         return;
2904 }
2905
2906 static struct ceph_connection *con_get(struct ceph_connection *con)
2907 {
2908         struct ceph_mds_session *s = con->private;
2909
2910         if (get_session(s)) {
2911                 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
2912                 return con;
2913         }
2914         dout("mdsc con_get %p FAIL\n", s);
2915         return NULL;
2916 }
2917
2918 static void con_put(struct ceph_connection *con)
2919 {
2920         struct ceph_mds_session *s = con->private;
2921
2922         ceph_put_mds_session(s);
2923         dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref));
2924 }
2925
2926 /*
2927  * if the client is unresponsive for long enough, the mds will kill
2928  * the session entirely.
2929  */
2930 static void peer_reset(struct ceph_connection *con)
2931 {
2932         struct ceph_mds_session *s = con->private;
2933
2934         pr_err("mds%d gave us the boot.  IMPLEMENT RECONNECT.\n",
2935                s->s_mds);
2936 }
2937
2938 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2939 {
2940         struct ceph_mds_session *s = con->private;
2941         struct ceph_mds_client *mdsc = s->s_mdsc;
2942         int type = le16_to_cpu(msg->hdr.type);
2943
2944         mutex_lock(&mdsc->mutex);
2945         if (__verify_registered_session(mdsc, s) < 0) {
2946                 mutex_unlock(&mdsc->mutex);
2947                 goto out;
2948         }
2949         mutex_unlock(&mdsc->mutex);
2950
2951         switch (type) {
2952         case CEPH_MSG_MDS_MAP:
2953                 ceph_mdsc_handle_map(mdsc, msg);
2954                 break;
2955         case CEPH_MSG_CLIENT_SESSION:
2956                 handle_session(s, msg);
2957                 break;
2958         case CEPH_MSG_CLIENT_REPLY:
2959                 handle_reply(s, msg);
2960                 break;
2961         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
2962                 handle_forward(mdsc, s, msg);
2963                 break;
2964         case CEPH_MSG_CLIENT_CAPS:
2965                 ceph_handle_caps(s, msg);
2966                 break;
2967         case CEPH_MSG_CLIENT_SNAP:
2968                 ceph_handle_snap(mdsc, s, msg);
2969                 break;
2970         case CEPH_MSG_CLIENT_LEASE:
2971                 handle_lease(mdsc, s, msg);
2972                 break;
2973
2974         default:
2975                 pr_err("received unknown message type %d %s\n", type,
2976                        ceph_msg_type_name(type));
2977         }
2978 out:
2979         ceph_msg_put(msg);
2980 }
2981
2982 /*
2983  * authentication
2984  */
2985 static int get_authorizer(struct ceph_connection *con,
2986                           void **buf, int *len, int *proto,
2987                           void **reply_buf, int *reply_len, int force_new)
2988 {
2989         struct ceph_mds_session *s = con->private;
2990         struct ceph_mds_client *mdsc = s->s_mdsc;
2991         struct ceph_auth_client *ac = mdsc->client->monc.auth;
2992         int ret = 0;
2993
2994         if (force_new && s->s_authorizer) {
2995                 ac->ops->destroy_authorizer(ac, s->s_authorizer);
2996                 s->s_authorizer = NULL;
2997         }
2998         if (s->s_authorizer == NULL) {
2999                 if (ac->ops->create_authorizer) {
3000                         ret = ac->ops->create_authorizer(
3001                                 ac, CEPH_ENTITY_TYPE_MDS,
3002                                 &s->s_authorizer,
3003                                 &s->s_authorizer_buf,
3004                                 &s->s_authorizer_buf_len,
3005                                 &s->s_authorizer_reply_buf,
3006                                 &s->s_authorizer_reply_buf_len);
3007                         if (ret)
3008                                 return ret;
3009                 }
3010         }
3011
3012         *proto = ac->protocol;
3013         *buf = s->s_authorizer_buf;
3014         *len = s->s_authorizer_buf_len;
3015         *reply_buf = s->s_authorizer_reply_buf;
3016         *reply_len = s->s_authorizer_reply_buf_len;
3017         return 0;
3018 }
3019
3020
3021 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3022 {
3023         struct ceph_mds_session *s = con->private;
3024         struct ceph_mds_client *mdsc = s->s_mdsc;
3025         struct ceph_auth_client *ac = mdsc->client->monc.auth;
3026
3027         return ac->ops->verify_authorizer_reply(ac, s->s_authorizer, len);
3028 }
3029
3030 static int invalidate_authorizer(struct ceph_connection *con)
3031 {
3032         struct ceph_mds_session *s = con->private;
3033         struct ceph_mds_client *mdsc = s->s_mdsc;
3034         struct ceph_auth_client *ac = mdsc->client->monc.auth;
3035
3036         if (ac->ops->invalidate_authorizer)
3037                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3038
3039         return ceph_monc_validate_auth(&mdsc->client->monc);
3040 }
3041
3042 const static struct ceph_connection_operations mds_con_ops = {
3043         .get = con_get,
3044         .put = con_put,
3045         .dispatch = dispatch,
3046         .get_authorizer = get_authorizer,
3047         .verify_authorizer_reply = verify_authorizer_reply,
3048         .invalidate_authorizer = invalidate_authorizer,
3049         .peer_reset = peer_reset,
3050 };
3051
3052
3053
3054
3055 /* eof */