fs/ceph/mds_client.c

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  37  * are no longer valid.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