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  37  * are no longer valid.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_all(&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 int ceph_add_cap_releases(struct ceph_mds_client *mdsc,
1070                           struct ceph_mds_session *session)
1071 {
1072         struct ceph_msg *msg, *partial = NULL;
1073         struct ceph_mds_cap_release *head;
1074         int err = -ENOMEM;
1075         int extra = mdsc->client->mount_args->cap_release_safety;
1076         int num;
1077
1078         dout("add_cap_releases %p mds%d extra %d\n", session, session->s_mds,
1079              extra);
1080
1081         spin_lock(&session->s_cap_lock);
1082
1083         if (!list_empty(&session->s_cap_releases)) {
1084                 msg = list_first_entry(&session->s_cap_releases,
1085                                        struct ceph_msg,
1086                                  list_head);
1087                 head = msg->front.iov_base;
1088                 num = le32_to_cpu(head->num);
1089                 if (num) {
1090                         dout(" partial %p with (%d/%d)\n", msg, num,
1091                              (int)CEPH_CAPS_PER_RELEASE);
1092                         extra += CEPH_CAPS_PER_RELEASE - num;
1093                         partial = msg;
1094                 }
1095         }
1096         while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1097                 spin_unlock(&session->s_cap_lock);
1098                 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
1099                                    GFP_NOFS);
1100                 if (!msg)
1101                         goto out_unlocked;
1102                 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1103                      (int)msg->front.iov_len);
1104                 head = msg->front.iov_base;
1105                 head->num = cpu_to_le32(0);
1106                 msg->front.iov_len = sizeof(*head);
1107                 spin_lock(&session->s_cap_lock);
1108                 list_add(&msg->list_head, &session->s_cap_releases);
1109                 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1110         }
1111
1112         if (partial) {
1113                 head = partial->front.iov_base;
1114                 num = le32_to_cpu(head->num);
1115                 dout(" queueing partial %p with %d/%d\n", partial, num,
1116                      (int)CEPH_CAPS_PER_RELEASE);
1117                 list_move_tail(&partial->list_head,
1118                                &session->s_cap_releases_done);
1119                 session->s_num_cap_releases -= CEPH_CAPS_PER_RELEASE - num;
1120         }
1121         err = 0;
1122         spin_unlock(&session->s_cap_lock);
1123 out_unlocked:
1124         return err;
1125 }
1126
1127 /*
1128  * flush all dirty inode data to disk.
1129  *
1130  * returns true if we've flushed through want_flush_seq
1131  */
1132 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1133 {
1134         int mds, ret = 1;
1135
1136         dout("check_cap_flush want %lld\n", want_flush_seq);
1137         mutex_lock(&mdsc->mutex);
1138         for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1139                 struct ceph_mds_session *session = mdsc->sessions[mds];
1140
1141                 if (!session)
1142                         continue;
1143                 get_session(session);
1144                 mutex_unlock(&mdsc->mutex);
1145
1146                 mutex_lock(&session->s_mutex);
1147                 if (!list_empty(&session->s_cap_flushing)) {
1148                         struct ceph_inode_info *ci =
1149                                 list_entry(session->s_cap_flushing.next,
1150                                            struct ceph_inode_info,
1151                                            i_flushing_item);
1152                         struct inode *inode = &ci->vfs_inode;
1153
1154                         spin_lock(&inode->i_lock);
1155                         if (ci->i_cap_flush_seq <= want_flush_seq) {
1156                                 dout("check_cap_flush still flushing %p "
1157                                      "seq %lld <= %lld to mds%d\n", inode,
1158                                      ci->i_cap_flush_seq, want_flush_seq,
1159                                      session->s_mds);
1160                                 ret = 0;
1161                         }
1162                         spin_unlock(&inode->i_lock);
1163                 }
1164                 mutex_unlock(&session->s_mutex);
1165                 ceph_put_mds_session(session);
1166
1167                 if (!ret)
1168                         return ret;
1169                 mutex_lock(&mdsc->mutex);
1170         }
1171
1172         mutex_unlock(&mdsc->mutex);
1173         dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1174         return ret;
1175 }
1176
1177 /*
1178  * called under s_mutex
1179  */
1180 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1181                             struct ceph_mds_session *session)
1182 {
1183         struct ceph_msg *msg;
1184
1185         dout("send_cap_releases mds%d\n", session->s_mds);
1186         spin_lock(&session->s_cap_lock);
1187         while (!list_empty(&session->s_cap_releases_done)) {
1188                 msg = list_first_entry(&session->s_cap_releases_done,
1189                                  struct ceph_msg, list_head);
1190                 list_del_init(&msg->list_head);
1191                 spin_unlock(&session->s_cap_lock);
1192                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1193                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1194                 ceph_con_send(&session->s_con, msg);
1195                 spin_lock(&session->s_cap_lock);
1196         }
1197         spin_unlock(&session->s_cap_lock);
1198 }
1199
1200 static void discard_cap_releases(struct ceph_mds_client *mdsc,
1201                                  struct ceph_mds_session *session)
1202 {
1203         struct ceph_msg *msg;
1204         struct ceph_mds_cap_release *head;
1205         unsigned num;
1206
1207         dout("discard_cap_releases mds%d\n", session->s_mds);
1208         spin_lock(&session->s_cap_lock);
1209
1210         /* zero out the in-progress message */
1211         msg = list_first_entry(&session->s_cap_releases,
1212                                struct ceph_msg, list_head);
1213         head = msg->front.iov_base;
1214         num = le32_to_cpu(head->num);
1215         dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg, num);
1216         head->num = cpu_to_le32(0);
1217         session->s_num_cap_releases += num;
1218
1219         /* requeue completed messages */
1220         while (!list_empty(&session->s_cap_releases_done)) {
1221                 msg = list_first_entry(&session->s_cap_releases_done,
1222                                  struct ceph_msg, list_head);
1223                 list_del_init(&msg->list_head);
1224
1225                 head = msg->front.iov_base;
1226                 num = le32_to_cpu(head->num);
1227                 dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg,
1228                      num);
1229                 session->s_num_cap_releases += num;
1230                 head->num = cpu_to_le32(0);
1231                 msg->front.iov_len = sizeof(*head);
1232                 list_add(&msg->list_head, &session->s_cap_releases);
1233         }
1234
1235         spin_unlock(&session->s_cap_lock);
1236 }
1237
1238 /*
1239  * requests
1240  */
1241
1242 /*
1243  * Create an mds request.
1244  */
1245 struct ceph_mds_request *
1246 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1247 {
1248         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1249
1250         if (!req)
1251                 return ERR_PTR(-ENOMEM);
1252
1253         mutex_init(&req->r_fill_mutex);
1254         req->r_started = jiffies;
1255         req->r_resend_mds = -1;
1256         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1257         req->r_fmode = -1;
1258         kref_init(&req->r_kref);
1259         INIT_LIST_HEAD(&req->r_wait);
1260         init_completion(&req->r_completion);
1261         init_completion(&req->r_safe_completion);
1262         INIT_LIST_HEAD(&req->r_unsafe_item);
1263
1264         req->r_op = op;
1265         req->r_direct_mode = mode;
1266         return req;
1267 }
1268
1269 /*
1270  * return oldest (lowest) request, tid in request tree, 0 if none.
1271  *
1272  * called under mdsc->mutex.
1273  */
1274 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1275 {
1276         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1277                 return NULL;
1278         return rb_entry(rb_first(&mdsc->request_tree),
1279                         struct ceph_mds_request, r_node);
1280 }
1281
1282 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1283 {
1284         struct ceph_mds_request *req = __get_oldest_req(mdsc);
1285
1286         if (req)
1287                 return req->r_tid;
1288         return 0;
1289 }
1290
1291 /*
1292  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1293  * on build_path_from_dentry in fs/cifs/dir.c.
1294  *
1295  * If @stop_on_nosnap, generate path relative to the first non-snapped
1296  * inode.
1297  *
1298  * Encode hidden .snap dirs as a double /, i.e.
1299  *   foo/.snap/bar -> foo//bar
1300  */
1301 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1302                            int stop_on_nosnap)
1303 {
1304         struct dentry *temp;
1305         char *path;
1306         int len, pos;
1307
1308         if (dentry == NULL)
1309                 return ERR_PTR(-EINVAL);
1310
1311 retry:
1312         len = 0;
1313         for (temp = dentry; !IS_ROOT(temp);) {
1314                 struct inode *inode = temp->d_inode;
1315                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1316                         len++;  /* slash only */
1317                 else if (stop_on_nosnap && inode &&
1318                          ceph_snap(inode) == CEPH_NOSNAP)
1319                         break;
1320                 else
1321                         len += 1 + temp->d_name.len;
1322                 temp = temp->d_parent;
1323                 if (temp == NULL) {
1324                         pr_err("build_path corrupt dentry %p\n", dentry);
1325                         return ERR_PTR(-EINVAL);
1326                 }
1327         }
1328         if (len)
1329                 len--;  /* no leading '/' */
1330
1331         path = kmalloc(len+1, GFP_NOFS);
1332         if (path == NULL)
1333                 return ERR_PTR(-ENOMEM);
1334         pos = len;
1335         path[pos] = 0;  /* trailing null */
1336         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1337                 struct inode *inode = temp->d_inode;
1338
1339                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1340                         dout("build_path path+%d: %p SNAPDIR\n",
1341                              pos, temp);
1342                 } else if (stop_on_nosnap && inode &&
1343                            ceph_snap(inode) == CEPH_NOSNAP) {
1344                         break;
1345                 } else {
1346                         pos -= temp->d_name.len;
1347                         if (pos < 0)
1348                                 break;
1349                         strncpy(path + pos, temp->d_name.name,
1350                                 temp->d_name.len);
1351                 }
1352                 if (pos)
1353                         path[--pos] = '/';
1354                 temp = temp->d_parent;
1355                 if (temp == NULL) {
1356                         pr_err("build_path corrupt dentry\n");
1357                         kfree(path);
1358                         return ERR_PTR(-EINVAL);
1359                 }
1360         }
1361         if (pos != 0) {
1362                 pr_err("build_path did not end path lookup where "
1363                        "expected, namelen is %d, pos is %d\n", len, pos);
1364                 /* presumably this is only possible if racing with a
1365                    rename of one of the parent directories (we can not
1366                    lock the dentries above us to prevent this, but
1367                    retrying should be harmless) */
1368                 kfree(path);
1369                 goto retry;
1370         }
1371
1372         *base = ceph_ino(temp->d_inode);
1373         *plen = len;
1374         dout("build_path on %p %d built %llx '%.*s'\n",
1375              dentry, atomic_read(&dentry->d_count), *base, len, path);
1376         return path;
1377 }
1378
1379 static int build_dentry_path(struct dentry *dentry,
1380                              const char **ppath, int *ppathlen, u64 *pino,
1381                              int *pfreepath)
1382 {
1383         char *path;
1384
1385         if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1386                 *pino = ceph_ino(dentry->d_parent->d_inode);
1387                 *ppath = dentry->d_name.name;
1388                 *ppathlen = dentry->d_name.len;
1389                 return 0;
1390         }
1391         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1392         if (IS_ERR(path))
1393                 return PTR_ERR(path);
1394         *ppath = path;
1395         *pfreepath = 1;
1396         return 0;
1397 }
1398
1399 static int build_inode_path(struct inode *inode,
1400                             const char **ppath, int *ppathlen, u64 *pino,
1401                             int *pfreepath)
1402 {
1403         struct dentry *dentry;
1404         char *path;
1405
1406         if (ceph_snap(inode) == CEPH_NOSNAP) {
1407                 *pino = ceph_ino(inode);
1408                 *ppathlen = 0;
1409                 return 0;
1410         }
1411         dentry = d_find_alias(inode);
1412         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1413         dput(dentry);
1414         if (IS_ERR(path))
1415                 return PTR_ERR(path);
1416         *ppath = path;
1417         *pfreepath = 1;
1418         return 0;
1419 }
1420
1421 /*
1422  * request arguments may be specified via an inode *, a dentry *, or
1423  * an explicit ino+path.
1424  */
1425 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1426                                   const char *rpath, u64 rino,
1427                                   const char **ppath, int *pathlen,
1428                                   u64 *ino, int *freepath)
1429 {
1430         int r = 0;
1431
1432         if (rinode) {
1433                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1434                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1435                      ceph_snap(rinode));
1436         } else if (rdentry) {
1437                 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1438                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1439                      *ppath);
1440         } else if (rpath) {
1441                 *ino = rino;
1442                 *ppath = rpath;
1443                 *pathlen = strlen(rpath);
1444                 dout(" path %.*s\n", *pathlen, rpath);
1445         }
1446
1447         return r;
1448 }
1449
1450 /*
1451  * called under mdsc->mutex
1452  */
1453 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1454                                                struct ceph_mds_request *req,
1455                                                int mds)
1456 {
1457         struct ceph_msg *msg;
1458         struct ceph_mds_request_head *head;
1459         const char *path1 = NULL;
1460         const char *path2 = NULL;
1461         u64 ino1 = 0, ino2 = 0;
1462         int pathlen1 = 0, pathlen2 = 0;
1463         int freepath1 = 0, freepath2 = 0;
1464         int len;
1465         u16 releases;
1466         void *p, *end;
1467         int ret;
1468
1469         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1470                               req->r_path1, req->r_ino1.ino,
1471                               &path1, &pathlen1, &ino1, &freepath1);
1472         if (ret < 0) {
1473                 msg = ERR_PTR(ret);
1474                 goto out;
1475         }
1476
1477         ret = set_request_path_attr(NULL, req->r_old_dentry,
1478                               req->r_path2, req->r_ino2.ino,
1479                               &path2, &pathlen2, &ino2, &freepath2);
1480         if (ret < 0) {
1481                 msg = ERR_PTR(ret);
1482                 goto out_free1;
1483         }
1484
1485         len = sizeof(*head) +
1486                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64));
1487
1488         /* calculate (max) length for cap releases */
1489         len += sizeof(struct ceph_mds_request_release) *
1490                 (!!req->r_inode_drop + !!req->r_dentry_drop +
1491                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1492         if (req->r_dentry_drop)
1493                 len += req->r_dentry->d_name.len;
1494         if (req->r_old_dentry_drop)
1495                 len += req->r_old_dentry->d_name.len;
1496
1497         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS);
1498         if (!msg) {
1499                 msg = ERR_PTR(-ENOMEM);
1500                 goto out_free2;
1501         }
1502
1503         msg->hdr.tid = cpu_to_le64(req->r_tid);
1504
1505         head = msg->front.iov_base;
1506         p = msg->front.iov_base + sizeof(*head);
1507         end = msg->front.iov_base + msg->front.iov_len;
1508
1509         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1510         head->op = cpu_to_le32(req->r_op);
1511         head->caller_uid = cpu_to_le32(current_fsuid());
1512         head->caller_gid = cpu_to_le32(current_fsgid());
1513         head->args = req->r_args;
1514
1515         ceph_encode_filepath(&p, end, ino1, path1);
1516         ceph_encode_filepath(&p, end, ino2, path2);
1517
1518         /* make note of release offset, in case we need to replay */
1519         req->r_request_release_offset = p - msg->front.iov_base;
1520
1521         /* cap releases */
1522         releases = 0;
1523         if (req->r_inode_drop)
1524                 releases += ceph_encode_inode_release(&p,
1525                       req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1526                       mds, req->r_inode_drop, req->r_inode_unless, 0);
1527         if (req->r_dentry_drop)
1528                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1529                        mds, req->r_dentry_drop, req->r_dentry_unless);
1530         if (req->r_old_dentry_drop)
1531                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1532                        mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1533         if (req->r_old_inode_drop)
1534                 releases += ceph_encode_inode_release(&p,
1535                       req->r_old_dentry->d_inode,
1536                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1537         head->num_releases = cpu_to_le16(releases);
1538
1539         BUG_ON(p > end);
1540         msg->front.iov_len = p - msg->front.iov_base;
1541         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1542
1543         msg->pages = req->r_pages;
1544         msg->nr_pages = req->r_num_pages;
1545         msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1546         msg->hdr.data_off = cpu_to_le16(0);
1547
1548 out_free2:
1549         if (freepath2)
1550                 kfree((char *)path2);
1551 out_free1:
1552         if (freepath1)
1553                 kfree((char *)path1);
1554 out:
1555         return msg;
1556 }
1557
1558 /*
1559  * called under mdsc->mutex if error, under no mutex if
1560  * success.
1561  */
1562 static void complete_request(struct ceph_mds_client *mdsc,
1563                              struct ceph_mds_request *req)
1564 {
1565         if (req->r_callback)
1566                 req->r_callback(mdsc, req);
1567         else
1568                 complete_all(&req->r_completion);
1569 }
1570
1571 /*
1572  * called under mdsc->mutex
1573  */
1574 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1575                                   struct ceph_mds_request *req,
1576                                   int mds)
1577 {
1578         struct ceph_mds_request_head *rhead;
1579         struct ceph_msg *msg;
1580         int flags = 0;
1581
1582         req->r_mds = mds;
1583         req->r_attempts++;
1584         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1585              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1586
1587         if (req->r_got_unsafe) {
1588                 /*
1589                  * Replay.  Do not regenerate message (and rebuild
1590                  * paths, etc.); just use the original message.
1591                  * Rebuilding paths will break for renames because
1592                  * d_move mangles the src name.
1593                  */
1594                 msg = req->r_request;
1595                 rhead = msg->front.iov_base;
1596
1597                 flags = le32_to_cpu(rhead->flags);
1598                 flags |= CEPH_MDS_FLAG_REPLAY;
1599                 rhead->flags = cpu_to_le32(flags);
1600
1601                 if (req->r_target_inode)
1602                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1603
1604                 rhead->num_retry = req->r_attempts - 1;
1605
1606                 /* remove cap/dentry releases from message */
1607                 rhead->num_releases = 0;
1608                 msg->hdr.front_len = cpu_to_le32(req->r_request_release_offset);
1609                 msg->front.iov_len = req->r_request_release_offset;
1610                 return 0;
1611         }
1612
1613         if (req->r_request) {
1614                 ceph_msg_put(req->r_request);
1615                 req->r_request = NULL;
1616         }
1617         msg = create_request_message(mdsc, req, mds);
1618         if (IS_ERR(msg)) {
1619                 req->r_err = PTR_ERR(msg);
1620                 complete_request(mdsc, req);
1621                 return PTR_ERR(msg);
1622         }
1623         req->r_request = msg;
1624
1625         rhead = msg->front.iov_base;
1626         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1627         if (req->r_got_unsafe)
1628                 flags |= CEPH_MDS_FLAG_REPLAY;
1629         if (req->r_locked_dir)
1630                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1631         rhead->flags = cpu_to_le32(flags);
1632         rhead->num_fwd = req->r_num_fwd;
1633         rhead->num_retry = req->r_attempts - 1;
1634         rhead->ino = 0;
1635
1636         dout(" r_locked_dir = %p\n", req->r_locked_dir);
1637         return 0;
1638 }
1639
1640 /*
1641  * send request, or put it on the appropriate wait list.
1642  */
1643 static int __do_request(struct ceph_mds_client *mdsc,
1644                         struct ceph_mds_request *req)
1645 {
1646         struct ceph_mds_session *session = NULL;
1647         int mds = -1;
1648         int err = -EAGAIN;
1649
1650         if (req->r_err || req->r_got_result)
1651                 goto out;
1652
1653         if (req->r_timeout &&
1654             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1655                 dout("do_request timed out\n");
1656                 err = -EIO;
1657                 goto finish;
1658         }
1659
1660         mds = __choose_mds(mdsc, req);
1661         if (mds < 0 ||
1662             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1663                 dout("do_request no mds or not active, waiting for map\n");
1664                 list_add(&req->r_wait, &mdsc->waiting_for_map);
1665                 goto out;
1666         }
1667
1668         /* get, open session */
1669         session = __ceph_lookup_mds_session(mdsc, mds);
1670         if (!session) {
1671                 session = register_session(mdsc, mds);
1672                 if (IS_ERR(session)) {
1673                         err = PTR_ERR(session);
1674                         goto finish;
1675                 }
1676         }
1677         dout("do_request mds%d session %p state %s\n", mds, session,
1678              session_state_name(session->s_state));
1679         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1680             session->s_state != CEPH_MDS_SESSION_HUNG) {
1681                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1682                     session->s_state == CEPH_MDS_SESSION_CLOSING)
1683                         __open_session(mdsc, session);
1684                 list_add(&req->r_wait, &session->s_waiting);
1685                 goto out_session;
1686         }
1687
1688         /* send request */
1689         req->r_session = get_session(session);
1690         req->r_resend_mds = -1;   /* forget any previous mds hint */
1691
1692         if (req->r_request_started == 0)   /* note request start time */
1693                 req->r_request_started = jiffies;
1694
1695         err = __prepare_send_request(mdsc, req, mds);
1696         if (!err) {
1697                 ceph_msg_get(req->r_request);
1698                 ceph_con_send(&session->s_con, req->r_request);
1699         }
1700
1701 out_session:
1702         ceph_put_mds_session(session);
1703 out:
1704         return err;
1705
1706 finish:
1707         req->r_err = err;
1708         complete_request(mdsc, req);
1709         goto out;
1710 }
1711
1712 /*
1713  * called under mdsc->mutex
1714  */
1715 static void __wake_requests(struct ceph_mds_client *mdsc,
1716                             struct list_head *head)
1717 {
1718         struct ceph_mds_request *req, *nreq;
1719
1720         list_for_each_entry_safe(req, nreq, head, r_wait) {
1721                 list_del_init(&req->r_wait);
1722                 __do_request(mdsc, req);
1723         }
1724 }
1725
1726 /*
1727  * Wake up threads with requests pending for @mds, so that they can
1728  * resubmit their requests to a possibly different mds.
1729  */
1730 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
1731 {
1732         struct ceph_mds_request *req;
1733         struct rb_node *p;
1734
1735         dout("kick_requests mds%d\n", mds);
1736         for (p = rb_first(&mdsc->request_tree); p; p = rb_next(p)) {
1737                 req = rb_entry(p, struct ceph_mds_request, r_node);
1738                 if (req->r_got_unsafe)
1739                         continue;
1740                 if (req->r_session &&
1741                     req->r_session->s_mds == mds) {
1742                         dout(" kicking tid %llu\n", req->r_tid);
1743                         put_request_session(req);
1744                         __do_request(mdsc, req);
1745                 }
1746         }
1747 }
1748
1749 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
1750                               struct ceph_mds_request *req)
1751 {
1752         dout("submit_request on %p\n", req);
1753         mutex_lock(&mdsc->mutex);
1754         __register_request(mdsc, req, NULL);
1755         __do_request(mdsc, req);
1756         mutex_unlock(&mdsc->mutex);
1757 }
1758
1759 /*
1760  * Synchrously perform an mds request.  Take care of all of the
1761  * session setup, forwarding, retry details.
1762  */
1763 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
1764                          struct inode *dir,
1765                          struct ceph_mds_request *req)
1766 {
1767         int err;
1768
1769         dout("do_request on %p\n", req);
1770
1771         /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
1772         if (req->r_inode)
1773                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
1774         if (req->r_locked_dir)
1775                 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
1776         if (req->r_old_dentry)
1777                 ceph_get_cap_refs(
1778                         ceph_inode(req->r_old_dentry->d_parent->d_inode),
1779                         CEPH_CAP_PIN);
1780
1781         /* issue */
1782         mutex_lock(&mdsc->mutex);
1783         __register_request(mdsc, req, dir);
1784         __do_request(mdsc, req);
1785
1786         if (req->r_err) {
1787                 err = req->r_err;
1788                 __unregister_request(mdsc, req);
1789                 dout("do_request early error %d\n", err);
1790                 goto out;
1791         }
1792
1793         /* wait */
1794         mutex_unlock(&mdsc->mutex);
1795         dout("do_request waiting\n");
1796         if (req->r_timeout) {
1797                 err = (long)wait_for_completion_killable_timeout(
1798                         &req->r_completion, req->r_timeout);
1799                 if (err == 0)
1800                         err = -EIO;
1801         } else {
1802                 err = wait_for_completion_killable(&req->r_completion);
1803         }
1804         dout("do_request waited, got %d\n", err);
1805         mutex_lock(&mdsc->mutex);
1806
1807         /* only abort if we didn't race with a real reply */
1808         if (req->r_got_result) {
1809                 err = le32_to_cpu(req->r_reply_info.head->result);
1810         } else if (err < 0) {
1811                 dout("aborted request %lld with %d\n", req->r_tid, err);
1812
1813                 /*
1814                  * ensure we aren't running concurrently with
1815                  * ceph_fill_trace or ceph_readdir_prepopulate, which
1816                  * rely on locks (dir mutex) held by our caller.
1817                  */
1818                 mutex_lock(&req->r_fill_mutex);
1819                 req->r_err = err;
1820                 req->r_aborted = true;
1821                 mutex_unlock(&req->r_fill_mutex);
1822
1823                 if (req->r_locked_dir &&
1824                     (req->r_op & CEPH_MDS_OP_WRITE))
1825                         ceph_invalidate_dir_request(req);
1826         } else {
1827                 err = req->r_err;
1828         }
1829
1830 out:
1831         mutex_unlock(&mdsc->mutex);
1832         dout("do_request %p done, result %d\n", req, err);
1833         return err;
1834 }
1835
1836 /*
1837  * Invalidate dir I_COMPLETE, dentry lease state on an aborted MDS
1838  * namespace request.
1839  */
1840 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
1841 {
1842         struct inode *inode = req->r_locked_dir;
1843         struct ceph_inode_info *ci = ceph_inode(inode);
1844
1845         dout("invalidate_dir_request %p (I_COMPLETE, lease(s))\n", inode);
1846         spin_lock(&inode->i_lock);
1847         ci->i_ceph_flags &= ~CEPH_I_COMPLETE;
1848         ci->i_release_count++;
1849         spin_unlock(&inode->i_lock);
1850
1851         if (req->r_dentry)
1852                 ceph_invalidate_dentry_lease(req->r_dentry);
1853         if (req->r_old_dentry)
1854                 ceph_invalidate_dentry_lease(req->r_old_dentry);
1855 }
1856
1857 /*
1858  * Handle mds reply.
1859  *
1860  * We take the session mutex and parse and process the reply immediately.
1861  * This preserves the logical ordering of replies, capabilities, etc., sent
1862  * by the MDS as they are applied to our local cache.
1863  */
1864 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
1865 {
1866         struct ceph_mds_client *mdsc = session->s_mdsc;
1867         struct ceph_mds_request *req;
1868         struct ceph_mds_reply_head *head = msg->front.iov_base;
1869         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
1870         u64 tid;
1871         int err, result;
1872         int mds = session->s_mds;
1873
1874         if (msg->front.iov_len < sizeof(*head)) {
1875                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
1876                 ceph_msg_dump(msg);
1877                 return;
1878         }
1879
1880         /* get request, session */
1881         tid = le64_to_cpu(msg->hdr.tid);
1882         mutex_lock(&mdsc->mutex);
1883         req = __lookup_request(mdsc, tid);
1884         if (!req) {
1885                 dout("handle_reply on unknown tid %llu\n", tid);
1886                 mutex_unlock(&mdsc->mutex);
1887                 return;
1888         }
1889         dout("handle_reply %p\n", req);
1890
1891         /* correct session? */
1892         if (req->r_session != session) {
1893                 pr_err("mdsc_handle_reply got %llu on session mds%d"
1894                        " not mds%d\n", tid, session->s_mds,
1895                        req->r_session ? req->r_session->s_mds : -1);
1896                 mutex_unlock(&mdsc->mutex);
1897                 goto out;
1898         }
1899
1900         /* dup? */
1901         if ((req->r_got_unsafe && !head->safe) ||
1902             (req->r_got_safe && head->safe)) {
1903                 pr_warning("got a dup %s reply on %llu from mds%d\n",
1904                            head->safe ? "safe" : "unsafe", tid, mds);
1905                 mutex_unlock(&mdsc->mutex);
1906                 goto out;
1907         }
1908         if (req->r_got_safe && !head->safe) {
1909                 pr_warning("got unsafe after safe on %llu from mds%d\n",
1910                            tid, mds);
1911                 mutex_unlock(&mdsc->mutex);
1912                 goto out;
1913         }
1914
1915         result = le32_to_cpu(head->result);
1916
1917         /*
1918          * Tolerate 2 consecutive ESTALEs from the same mds.
1919          * FIXME: we should be looking at the cap migrate_seq.
1920          */
1921         if (result == -ESTALE) {
1922                 req->r_direct_mode = USE_AUTH_MDS;
1923                 req->r_num_stale++;
1924                 if (req->r_num_stale <= 2) {
1925                         __do_request(mdsc, req);
1926                         mutex_unlock(&mdsc->mutex);
1927                         goto out;
1928                 }
1929         } else {
1930                 req->r_num_stale = 0;
1931         }
1932
1933         if (head->safe) {
1934                 req->r_got_safe = true;
1935                 __unregister_request(mdsc, req);
1936                 complete_all(&req->r_safe_completion);
1937
1938                 if (req->r_got_unsafe) {
1939                         /*
1940                          * We already handled the unsafe response, now do the
1941                          * cleanup.  No need to examine the response; the MDS
1942                          * doesn't include any result info in the safe
1943                          * response.  And even if it did, there is nothing
1944                          * useful we could do with a revised return value.
1945                          */
1946                         dout("got safe reply %llu, mds%d\n", tid, mds);
1947                         list_del_init(&req->r_unsafe_item);
1948
1949                         /* last unsafe request during umount? */
1950                         if (mdsc->stopping && !__get_oldest_req(mdsc))
1951                                 complete_all(&mdsc->safe_umount_waiters);
1952                         mutex_unlock(&mdsc->mutex);
1953                         goto out;
1954                 }
1955         } else {
1956                 req->r_got_unsafe = true;
1957                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
1958         }
1959
1960         dout("handle_reply tid %lld result %d\n", tid, result);
1961         rinfo = &req->r_reply_info;
1962         err = parse_reply_info(msg, rinfo);
1963         mutex_unlock(&mdsc->mutex);
1964
1965         mutex_lock(&session->s_mutex);
1966         if (err < 0) {
1967                 pr_err("mdsc_handle_reply got corrupt reply mds%d\n", mds);
1968                 ceph_msg_dump(msg);
1969                 goto out_err;
1970         }
1971
1972         /* snap trace */
1973         if (rinfo->snapblob_len) {
1974                 down_write(&mdsc->snap_rwsem);
1975                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
1976                                rinfo->snapblob + rinfo->snapblob_len,
1977                                le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
1978                 downgrade_write(&mdsc->snap_rwsem);
1979         } else {
1980                 down_read(&mdsc->snap_rwsem);
1981         }
1982
1983         /* insert trace into our cache */
1984         mutex_lock(&req->r_fill_mutex);
1985         err = ceph_fill_trace(mdsc->client->sb, req, req->r_session);
1986         if (err == 0) {
1987                 if (result == 0 && rinfo->dir_nr)
1988                         ceph_readdir_prepopulate(req, req->r_session);
1989                 ceph_unreserve_caps(&req->r_caps_reservation);
1990         }
1991         mutex_unlock(&req->r_fill_mutex);
1992
1993         up_read(&mdsc->snap_rwsem);
1994 out_err:
1995         mutex_lock(&mdsc->mutex);
1996         if (!req->r_aborted) {
1997                 if (err) {
1998                         req->r_err = err;
1999                 } else {
2000                         req->r_reply = msg;
2001                         ceph_msg_get(msg);
2002                         req->r_got_result = true;
2003                 }
2004         } else {
2005                 dout("reply arrived after request %lld was aborted\n", tid);
2006         }
2007         mutex_unlock(&mdsc->mutex);
2008
2009         ceph_add_cap_releases(mdsc, req->r_session);
2010         mutex_unlock(&session->s_mutex);
2011
2012         /* kick calling process */
2013         complete_request(mdsc, req);
2014 out:
2015         ceph_mdsc_put_request(req);
2016         return;
2017 }
2018
2019
2020
2021 /*
2022  * handle mds notification that our request has been forwarded.
2023  */
2024 static void handle_forward(struct ceph_mds_client *mdsc,
2025                            struct ceph_mds_session *session,
2026                            struct ceph_msg *msg)
2027 {
2028         struct ceph_mds_request *req;
2029         u64 tid = le64_to_cpu(msg->hdr.tid);
2030         u32 next_mds;
2031         u32 fwd_seq;
2032         int err = -EINVAL;
2033         void *p = msg->front.iov_base;
2034         void *end = p + msg->front.iov_len;
2035
2036         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2037         next_mds = ceph_decode_32(&p);
2038         fwd_seq = ceph_decode_32(&p);
2039
2040         mutex_lock(&mdsc->mutex);
2041         req = __lookup_request(mdsc, tid);
2042         if (!req) {
2043                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2044                 goto out;  /* dup reply? */
2045         }
2046
2047         if (req->r_aborted) {
2048                 dout("forward tid %llu aborted, unregistering\n", tid);
2049                 __unregister_request(mdsc, req);
2050         } else if (fwd_seq <= req->r_num_fwd) {
2051                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2052                      tid, next_mds, req->r_num_fwd, fwd_seq);
2053         } else {
2054                 /* resend. forward race not possible; mds would drop */
2055                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2056                 BUG_ON(req->r_err);
2057                 BUG_ON(req->r_got_result);
2058                 req->r_num_fwd = fwd_seq;
2059                 req->r_resend_mds = next_mds;
2060                 put_request_session(req);
2061                 __do_request(mdsc, req);
2062         }
2063         ceph_mdsc_put_request(req);
2064 out:
2065         mutex_unlock(&mdsc->mutex);
2066         return;
2067
2068 bad:
2069         pr_err("mdsc_handle_forward decode error err=%d\n", err);
2070 }
2071
2072 /*
2073  * handle a mds session control message
2074  */
2075 static void handle_session(struct ceph_mds_session *session,
2076                            struct ceph_msg *msg)
2077 {
2078         struct ceph_mds_client *mdsc = session->s_mdsc;
2079         u32 op;
2080         u64 seq;
2081         int mds = session->s_mds;
2082         struct ceph_mds_session_head *h = msg->front.iov_base;
2083         int wake = 0;
2084
2085         /* decode */
2086         if (msg->front.iov_len != sizeof(*h))
2087                 goto bad;
2088         op = le32_to_cpu(h->op);
2089         seq = le64_to_cpu(h->seq);
2090
2091         mutex_lock(&mdsc->mutex);
2092         if (op == CEPH_SESSION_CLOSE)
2093                 __unregister_session(mdsc, session);
2094         /* FIXME: this ttl calculation is generous */
2095         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2096         mutex_unlock(&mdsc->mutex);
2097
2098         mutex_lock(&session->s_mutex);
2099
2100         dout("handle_session mds%d %s %p state %s seq %llu\n",
2101              mds, ceph_session_op_name(op), session,
2102              session_state_name(session->s_state), seq);
2103
2104         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2105                 session->s_state = CEPH_MDS_SESSION_OPEN;
2106                 pr_info("mds%d came back\n", session->s_mds);
2107         }
2108
2109         switch (op) {
2110         case CEPH_SESSION_OPEN:
2111                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2112                         pr_info("mds%d reconnect success\n", session->s_mds);
2113                 session->s_state = CEPH_MDS_SESSION_OPEN;
2114                 renewed_caps(mdsc, session, 0);
2115                 wake = 1;
2116                 if (mdsc->stopping)
2117                         __close_session(mdsc, session);
2118                 break;
2119
2120         case CEPH_SESSION_RENEWCAPS:
2121                 if (session->s_renew_seq == seq)
2122                         renewed_caps(mdsc, session, 1);
2123                 break;
2124
2125         case CEPH_SESSION_CLOSE:
2126                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2127                         pr_info("mds%d reconnect denied\n", session->s_mds);
2128                 remove_session_caps(session);
2129                 wake = 1; /* for good measure */
2130                 complete_all(&mdsc->session_close_waiters);
2131                 kick_requests(mdsc, mds);
2132                 break;
2133
2134         case CEPH_SESSION_STALE:
2135                 pr_info("mds%d caps went stale, renewing\n",
2136                         session->s_mds);
2137                 spin_lock(&session->s_cap_lock);
2138                 session->s_cap_gen++;
2139                 session->s_cap_ttl = 0;
2140                 spin_unlock(&session->s_cap_lock);
2141                 send_renew_caps(mdsc, session);
2142                 break;
2143
2144         case CEPH_SESSION_RECALL_STATE:
2145                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2146                 break;
2147
2148         default:
2149                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2150                 WARN_ON(1);
2151         }
2152
2153         mutex_unlock(&session->s_mutex);
2154         if (wake) {
2155                 mutex_lock(&mdsc->mutex);
2156                 __wake_requests(mdsc, &session->s_waiting);
2157                 mutex_unlock(&mdsc->mutex);
2158         }
2159         return;
2160
2161 bad:
2162         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2163                (int)msg->front.iov_len);
2164         ceph_msg_dump(msg);
2165         return;
2166 }
2167
2168
2169 /*
2170  * called under session->mutex.
2171  */
2172 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2173                                    struct ceph_mds_session *session)
2174 {
2175         struct ceph_mds_request *req, *nreq;
2176         int err;
2177
2178         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2179
2180         mutex_lock(&mdsc->mutex);
2181         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2182                 err = __prepare_send_request(mdsc, req, session->s_mds);
2183                 if (!err) {
2184                         ceph_msg_get(req->r_request);
2185                         ceph_con_send(&session->s_con, req->r_request);
2186                 }
2187         }
2188         mutex_unlock(&mdsc->mutex);
2189 }
2190
2191 /*
2192  * Encode information about a cap for a reconnect with the MDS.
2193  */
2194 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2195                           void *arg)
2196 {
2197         struct ceph_mds_cap_reconnect rec;
2198         struct ceph_inode_info *ci;
2199         struct ceph_pagelist *pagelist = arg;
2200         char *path;
2201         int pathlen, err;
2202         u64 pathbase;
2203         struct dentry *dentry;
2204
2205         ci = cap->ci;
2206
2207         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2208              inode, ceph_vinop(inode), cap, cap->cap_id,
2209              ceph_cap_string(cap->issued));
2210         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2211         if (err)
2212                 return err;
2213
2214         dentry = d_find_alias(inode);
2215         if (dentry) {
2216                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2217                 if (IS_ERR(path)) {
2218                         err = PTR_ERR(path);
2219                         BUG_ON(err);
2220                 }
2221         } else {
2222                 path = NULL;
2223                 pathlen = 0;
2224         }
2225         err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2226         if (err)
2227                 goto out;
2228
2229         spin_lock(&inode->i_lock);
2230         cap->seq = 0;        /* reset cap seq */
2231         cap->issue_seq = 0;  /* and issue_seq */
2232         rec.cap_id = cpu_to_le64(cap->cap_id);
2233         rec.pathbase = cpu_to_le64(pathbase);
2234         rec.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2235         rec.issued = cpu_to_le32(cap->issued);
2236         rec.size = cpu_to_le64(inode->i_size);
2237         ceph_encode_timespec(&rec.mtime, &inode->i_mtime);
2238         ceph_encode_timespec(&rec.atime, &inode->i_atime);
2239         rec.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2240         spin_unlock(&inode->i_lock);
2241
2242         err = ceph_pagelist_append(pagelist, &rec, sizeof(rec));
2243
2244 out:
2245         kfree(path);
2246         dput(dentry);
2247         return err;
2248 }
2249
2250
2251 /*
2252  * If an MDS fails and recovers, clients need to reconnect in order to
2253  * reestablish shared state.  This includes all caps issued through
2254  * this session _and_ the snap_realm hierarchy.  Because it's not
2255  * clear which snap realms the mds cares about, we send everything we
2256  * know about.. that ensures we'll then get any new info the
2257  * recovering MDS might have.
2258  *
2259  * This is a relatively heavyweight operation, but it's rare.
2260  *
2261  * called with mdsc->mutex held.
2262  */
2263 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2264                                struct ceph_mds_session *session)
2265 {
2266         struct ceph_msg *reply;
2267         struct rb_node *p;
2268         int mds = session->s_mds;
2269         int err = -ENOMEM;
2270         struct ceph_pagelist *pagelist;
2271
2272         pr_info("mds%d reconnect start\n", mds);
2273
2274         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2275         if (!pagelist)
2276                 goto fail_nopagelist;
2277         ceph_pagelist_init(pagelist);
2278
2279         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS);
2280         if (!reply)
2281                 goto fail_nomsg;
2282
2283         mutex_lock(&session->s_mutex);
2284         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2285         session->s_seq = 0;
2286
2287         ceph_con_open(&session->s_con,
2288                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2289
2290         /* replay unsafe requests */
2291         replay_unsafe_requests(mdsc, session);
2292
2293         down_read(&mdsc->snap_rwsem);
2294
2295         dout("session %p state %s\n", session,
2296              session_state_name(session->s_state));
2297
2298         /* drop old cap expires; we're about to reestablish that state */
2299         discard_cap_releases(mdsc, session);
2300
2301         /* traverse this session's caps */
2302         err = ceph_pagelist_encode_32(pagelist, session->s_nr_caps);
2303         if (err)
2304                 goto fail;
2305         err = iterate_session_caps(session, encode_caps_cb, pagelist);
2306         if (err < 0)
2307                 goto fail;
2308
2309         /*
2310          * snaprealms.  we provide mds with the ino, seq (version), and
2311          * parent for all of our realms.  If the mds has any newer info,
2312          * it will tell us.
2313          */
2314         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2315                 struct ceph_snap_realm *realm =
2316                         rb_entry(p, struct ceph_snap_realm, node);
2317                 struct ceph_mds_snaprealm_reconnect sr_rec;
2318
2319                 dout(" adding snap realm %llx seq %lld parent %llx\n",
2320                      realm->ino, realm->seq, realm->parent_ino);
2321                 sr_rec.ino = cpu_to_le64(realm->ino);
2322                 sr_rec.seq = cpu_to_le64(realm->seq);
2323                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2324                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2325                 if (err)
2326                         goto fail;
2327         }
2328
2329         reply->pagelist = pagelist;
2330         reply->hdr.data_len = cpu_to_le32(pagelist->length);
2331         reply->nr_pages = calc_pages_for(0, pagelist->length);
2332         ceph_con_send(&session->s_con, reply);
2333
2334         mutex_unlock(&session->s_mutex);
2335
2336         mutex_lock(&mdsc->mutex);
2337         __wake_requests(mdsc, &session->s_waiting);
2338         mutex_unlock(&mdsc->mutex);
2339
2340         up_read(&mdsc->snap_rwsem);
2341         return;
2342
2343 fail:
2344         ceph_msg_put(reply);
2345         up_read(&mdsc->snap_rwsem);
2346         mutex_unlock(&session->s_mutex);
2347 fail_nomsg:
2348         ceph_pagelist_release(pagelist);
2349         kfree(pagelist);
2350 fail_nopagelist:
2351         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
2352         return;
2353 }
2354
2355
2356 /*
2357  * compare old and new mdsmaps, kicking requests
2358  * and closing out old connections as necessary
2359  *
2360  * called under mdsc->mutex.
2361  */
2362 static void check_new_map(struct ceph_mds_client *mdsc,
2363                           struct ceph_mdsmap *newmap,
2364                           struct ceph_mdsmap *oldmap)
2365 {
2366         int i;
2367         int oldstate, newstate;
2368         struct ceph_mds_session *s;
2369
2370         dout("check_new_map new %u old %u\n",
2371              newmap->m_epoch, oldmap->m_epoch);
2372
2373         for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2374                 if (mdsc->sessions[i] == NULL)
2375                         continue;
2376                 s = mdsc->sessions[i];
2377                 oldstate = ceph_mdsmap_get_state(oldmap, i);
2378                 newstate = ceph_mdsmap_get_state(newmap, i);
2379
2380                 dout("check_new_map mds%d state %s -> %s (session %s)\n",
2381                      i, ceph_mds_state_name(oldstate),
2382                      ceph_mds_state_name(newstate),
2383                      session_state_name(s->s_state));
2384
2385                 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
2386                            ceph_mdsmap_get_addr(newmap, i),
2387                            sizeof(struct ceph_entity_addr))) {
2388                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2389                                 /* the session never opened, just close it
2390                                  * out now */
2391                                 __wake_requests(mdsc, &s->s_waiting);
2392                                 __unregister_session(mdsc, s);
2393                         } else {
2394                                 /* just close it */
2395                                 mutex_unlock(&mdsc->mutex);
2396                                 mutex_lock(&s->s_mutex);
2397                                 mutex_lock(&mdsc->mutex);
2398                                 ceph_con_close(&s->s_con);
2399                                 mutex_unlock(&s->s_mutex);
2400                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2401                         }
2402
2403                         /* kick any requests waiting on the recovering mds */
2404                         kick_requests(mdsc, i);
2405                 } else if (oldstate == newstate) {
2406                         continue;  /* nothing new with this mds */
2407                 }
2408
2409                 /*
2410                  * send reconnect?
2411                  */
2412                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2413                     newstate >= CEPH_MDS_STATE_RECONNECT) {
2414                         mutex_unlock(&mdsc->mutex);
2415                         send_mds_reconnect(mdsc, s);
2416                         mutex_lock(&mdsc->mutex);
2417                 }
2418
2419                 /*
2420                  * kick request on any mds that has gone active.
2421                  */
2422                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2423                     newstate >= CEPH_MDS_STATE_ACTIVE) {
2424                         if (oldstate != CEPH_MDS_STATE_CREATING &&
2425                             oldstate != CEPH_MDS_STATE_STARTING)
2426                                 pr_info("mds%d recovery completed\n", s->s_mds);
2427                         kick_requests(mdsc, i);
2428                         ceph_kick_flushing_caps(mdsc, s);
2429                         wake_up_session_caps(s, 1);
2430                 }
2431         }
2432 }
2433
2434
2435
2436 /*
2437  * leases
2438  */
2439
2440 /*
2441  * caller must hold session s_mutex, dentry->d_lock
2442  */
2443 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2444 {
2445         struct ceph_dentry_info *di = ceph_dentry(dentry);
2446
2447         ceph_put_mds_session(di->lease_session);
2448         di->lease_session = NULL;
2449 }
2450
2451 static void handle_lease(struct ceph_mds_client *mdsc,
2452                          struct ceph_mds_session *session,
2453                          struct ceph_msg *msg)
2454 {
2455         struct super_block *sb = mdsc->client->sb;
2456         struct inode *inode;
2457         struct ceph_inode_info *ci;
2458         struct dentry *parent, *dentry;
2459         struct ceph_dentry_info *di;
2460         int mds = session->s_mds;
2461         struct ceph_mds_lease *h = msg->front.iov_base;
2462         u32 seq;
2463         struct ceph_vino vino;
2464         int mask;
2465         struct qstr dname;
2466         int release = 0;
2467
2468         dout("handle_lease from mds%d\n", mds);
2469
2470         /* decode */
2471         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2472                 goto bad;
2473         vino.ino = le64_to_cpu(h->ino);
2474         vino.snap = CEPH_NOSNAP;
2475         mask = le16_to_cpu(h->mask);
2476         seq = le32_to_cpu(h->seq);
2477         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2478         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2479         if (dname.len != get_unaligned_le32(h+1))
2480                 goto bad;
2481
2482         mutex_lock(&session->s_mutex);
2483         session->s_seq++;
2484
2485         /* lookup inode */
2486         inode = ceph_find_inode(sb, vino);
2487         dout("handle_lease %s, mask %d, ino %llx %p %.*s\n",
2488              ceph_lease_op_name(h->action), mask, vino.ino, inode,
2489              dname.len, dname.name);
2490         if (inode == NULL) {
2491                 dout("handle_lease no inode %llx\n", vino.ino);
2492                 goto release;
2493         }
2494         ci = ceph_inode(inode);
2495
2496         /* dentry */
2497         parent = d_find_alias(inode);
2498         if (!parent) {
2499                 dout("no parent dentry on inode %p\n", inode);
2500                 WARN_ON(1);
2501                 goto release;  /* hrm... */
2502         }
2503         dname.hash = full_name_hash(dname.name, dname.len);
2504         dentry = d_lookup(parent, &dname);
2505         dput(parent);
2506         if (!dentry)
2507                 goto release;
2508
2509         spin_lock(&dentry->d_lock);
2510         di = ceph_dentry(dentry);
2511         switch (h->action) {
2512         case CEPH_MDS_LEASE_REVOKE:
2513                 if (di && di->lease_session == session) {
2514                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
2515                                 h->seq = cpu_to_le32(di->lease_seq);
2516                         __ceph_mdsc_drop_dentry_lease(dentry);
2517                 }
2518                 release = 1;
2519                 break;
2520
2521         case CEPH_MDS_LEASE_RENEW:
2522                 if (di && di->lease_session == session &&
2523                     di->lease_gen == session->s_cap_gen &&
2524                     di->lease_renew_from &&
2525                     di->lease_renew_after == 0) {
2526                         unsigned long duration =
2527                                 le32_to_cpu(h->duration_ms) * HZ / 1000;
2528
2529                         di->lease_seq = seq;
2530                         dentry->d_time = di->lease_renew_from + duration;
2531                         di->lease_renew_after = di->lease_renew_from +
2532                                 (duration >> 1);
2533                         di->lease_renew_from = 0;
2534                 }
2535                 break;
2536         }
2537         spin_unlock(&dentry->d_lock);
2538         dput(dentry);
2539
2540         if (!release)
2541                 goto out;
2542
2543 release:
2544         /* let's just reuse the same message */
2545         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2546         ceph_msg_get(msg);
2547         ceph_con_send(&session->s_con, msg);
2548
2549 out:
2550         iput(inode);
2551         mutex_unlock(&session->s_mutex);
2552         return;
2553
2554 bad:
2555         pr_err("corrupt lease message\n");
2556         ceph_msg_dump(msg);
2557 }
2558
2559 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
2560                               struct inode *inode,
2561                               struct dentry *dentry, char action,
2562                               u32 seq)
2563 {
2564         struct ceph_msg *msg;
2565         struct ceph_mds_lease *lease;
2566         int len = sizeof(*lease) + sizeof(u32);
2567         int dnamelen = 0;
2568
2569         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2570              inode, dentry, ceph_lease_op_name(action), session->s_mds);
2571         dnamelen = dentry->d_name.len;
2572         len += dnamelen;
2573
2574         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS);
2575         if (!msg)
2576                 return;
2577         lease = msg->front.iov_base;
2578         lease->action = action;
2579         lease->mask = cpu_to_le16(1);
2580         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
2581         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
2582         lease->seq = cpu_to_le32(seq);
2583         put_unaligned_le32(dnamelen, lease + 1);
2584         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
2585
2586         /*
2587          * if this is a preemptive lease RELEASE, no need to
2588          * flush request stream, since the actual request will
2589          * soon follow.
2590          */
2591         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
2592
2593         ceph_con_send(&session->s_con, msg);
2594 }
2595
2596 /*
2597  * Preemptively release a lease we expect to invalidate anyway.
2598  * Pass @inode always, @dentry is optional.
2599  */
2600 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2601                              struct dentry *dentry, int mask)
2602 {
2603         struct ceph_dentry_info *di;
2604         struct ceph_mds_session *session;
2605         u32 seq;
2606
2607         BUG_ON(inode == NULL);
2608         BUG_ON(dentry == NULL);
2609         BUG_ON(mask == 0);
2610
2611         /* is dentry lease valid? */
2612         spin_lock(&dentry->d_lock);
2613         di = ceph_dentry(dentry);
2614         if (!di || !di->lease_session ||
2615             di->lease_session->s_mds < 0 ||
2616             di->lease_gen != di->lease_session->s_cap_gen ||
2617             !time_before(jiffies, dentry->d_time)) {
2618                 dout("lease_release inode %p dentry %p -- "
2619                      "no lease on %d\n",
2620                      inode, dentry, mask);
2621                 spin_unlock(&dentry->d_lock);
2622                 return;
2623         }
2624
2625         /* we do have a lease on this dentry; note mds and seq */
2626         session = ceph_get_mds_session(di->lease_session);
2627         seq = di->lease_seq;
2628         __ceph_mdsc_drop_dentry_lease(dentry);
2629         spin_unlock(&dentry->d_lock);
2630
2631         dout("lease_release inode %p dentry %p mask %d to mds%d\n",
2632              inode, dentry, mask, session->s_mds);
2633         ceph_mdsc_lease_send_msg(session, inode, dentry,
2634                                  CEPH_MDS_LEASE_RELEASE, seq);
2635         ceph_put_mds_session(session);
2636 }
2637
2638 /*
2639  * drop all leases (and dentry refs) in preparation for umount
2640  */
2641 static void drop_leases(struct ceph_mds_client *mdsc)
2642 {
2643         int i;
2644
2645         dout("drop_leases\n");
2646         mutex_lock(&mdsc->mutex);
2647         for (i = 0; i < mdsc->max_sessions; i++) {
2648                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2649                 if (!s)
2650                         continue;
2651                 mutex_unlock(&mdsc->mutex);
2652                 mutex_lock(&s->s_mutex);
2653                 mutex_unlock(&s->s_mutex);
2654                 ceph_put_mds_session(s);
2655                 mutex_lock(&mdsc->mutex);
2656         }
2657         mutex_unlock(&mdsc->mutex);
2658 }
2659
2660
2661
2662 /*
2663  * delayed work -- periodically trim expired leases, renew caps with mds
2664  */
2665 static void schedule_delayed(struct ceph_mds_client *mdsc)
2666 {
2667         int delay = 5;
2668         unsigned hz = round_jiffies_relative(HZ * delay);
2669         schedule_delayed_work(&mdsc->delayed_work, hz);
2670 }
2671
2672 static void delayed_work(struct work_struct *work)
2673 {
2674         int i;
2675         struct ceph_mds_client *mdsc =
2676                 container_of(work, struct ceph_mds_client, delayed_work.work);
2677         int renew_interval;
2678         int renew_caps;
2679
2680         dout("mdsc delayed_work\n");
2681         ceph_check_delayed_caps(mdsc);
2682
2683         mutex_lock(&mdsc->mutex);
2684         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
2685         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
2686                                    mdsc->last_renew_caps);
2687         if (renew_caps)
2688                 mdsc->last_renew_caps = jiffies;
2689
2690         for (i = 0; i < mdsc->max_sessions; i++) {
2691                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2692                 if (s == NULL)
2693                         continue;
2694                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
2695                         dout("resending session close request for mds%d\n",
2696                              s->s_mds);
2697                         request_close_session(mdsc, s);
2698                         ceph_put_mds_session(s);
2699                         continue;
2700                 }
2701                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
2702                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
2703                                 s->s_state = CEPH_MDS_SESSION_HUNG;
2704                                 pr_info("mds%d hung\n", s->s_mds);
2705                         }
2706                 }
2707                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
2708                         /* this mds is failed or recovering, just wait */
2709                         ceph_put_mds_session(s);
2710                         continue;
2711                 }
2712                 mutex_unlock(&mdsc->mutex);
2713
2714                 mutex_lock(&s->s_mutex);
2715                 if (renew_caps)
2716                         send_renew_caps(mdsc, s);
2717                 else
2718                         ceph_con_keepalive(&s->s_con);
2719                 ceph_add_cap_releases(mdsc, s);
2720                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2721                     s->s_state == CEPH_MDS_SESSION_HUNG)
2722                         ceph_send_cap_releases(mdsc, s);
2723                 mutex_unlock(&s->s_mutex);
2724                 ceph_put_mds_session(s);
2725
2726                 mutex_lock(&mdsc->mutex);
2727         }
2728         mutex_unlock(&mdsc->mutex);
2729
2730         schedule_delayed(mdsc);
2731 }
2732
2733
2734 int ceph_mdsc_init(struct ceph_mds_client *mdsc, struct ceph_client *client)
2735 {
2736         mdsc->client = client;
2737         mutex_init(&mdsc->mutex);
2738         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
2739         if (mdsc->mdsmap == NULL)
2740                 return -ENOMEM;
2741
2742         init_completion(&mdsc->safe_umount_waiters);
2743         init_completion(&mdsc->session_close_waiters);
2744         INIT_LIST_HEAD(&mdsc->waiting_for_map);
2745         mdsc->sessions = NULL;
2746         mdsc->max_sessions = 0;
2747         mdsc->stopping = 0;
2748         init_rwsem(&mdsc->snap_rwsem);
2749         mdsc->snap_realms = RB_ROOT;
2750         INIT_LIST_HEAD(&mdsc->snap_empty);
2751         spin_lock_init(&mdsc->snap_empty_lock);
2752         mdsc->last_tid = 0;
2753         mdsc->request_tree = RB_ROOT;
2754         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
2755         mdsc->last_renew_caps = jiffies;
2756         INIT_LIST_HEAD(&mdsc->cap_delay_list);
2757         spin_lock_init(&mdsc->cap_delay_lock);
2758         INIT_LIST_HEAD(&mdsc->snap_flush_list);
2759         spin_lock_init(&mdsc->snap_flush_lock);
2760         mdsc->cap_flush_seq = 0;
2761         INIT_LIST_HEAD(&mdsc->cap_dirty);
2762         mdsc->num_cap_flushing = 0;
2763         spin_lock_init(&mdsc->cap_dirty_lock);
2764         init_waitqueue_head(&mdsc->cap_flushing_wq);
2765         spin_lock_init(&mdsc->dentry_lru_lock);
2766         INIT_LIST_HEAD(&mdsc->dentry_lru);
2767
2768         return 0;
2769 }
2770
2771 /*
2772  * Wait for safe replies on open mds requests.  If we time out, drop
2773  * all requests from the tree to avoid dangling dentry refs.
2774  */
2775 static void wait_requests(struct ceph_mds_client *mdsc)
2776 {
2777         struct ceph_mds_request *req;
2778         struct ceph_client *client = mdsc->client;
2779
2780         mutex_lock(&mdsc->mutex);
2781         if (__get_oldest_req(mdsc)) {
2782                 mutex_unlock(&mdsc->mutex);
2783
2784                 dout("wait_requests waiting for requests\n");
2785                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
2786                                     client->mount_args->mount_timeout * HZ);
2787
2788                 /* tear down remaining requests */
2789                 mutex_lock(&mdsc->mutex);
2790                 while ((req = __get_oldest_req(mdsc))) {
2791                         dout("wait_requests timed out on tid %llu\n",
2792                              req->r_tid);
2793                         __unregister_request(mdsc, req);
2794                 }
2795         }
2796         mutex_unlock(&mdsc->mutex);
2797         dout("wait_requests done\n");
2798 }
2799
2800 /*
2801  * called before mount is ro, and before dentries are torn down.
2802  * (hmm, does this still race with new lookups?)
2803  */
2804 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
2805 {
2806         dout("pre_umount\n");
2807         mdsc->stopping = 1;
2808
2809         drop_leases(mdsc);
2810         ceph_flush_dirty_caps(mdsc);
2811         wait_requests(mdsc);
2812
2813         /*
2814          * wait for reply handlers to drop their request refs and
2815          * their inode/dcache refs
2816          */
2817         ceph_msgr_flush();
2818 }
2819
2820 /*
2821  * wait for all write mds requests to flush.
2822  */
2823 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
2824 {
2825         struct ceph_mds_request *req = NULL, *nextreq;
2826         struct rb_node *n;
2827
2828         mutex_lock(&mdsc->mutex);
2829         dout("wait_unsafe_requests want %lld\n", want_tid);
2830 restart:
2831         req = __get_oldest_req(mdsc);
2832         while (req && req->r_tid <= want_tid) {
2833                 /* find next request */
2834                 n = rb_next(&req->r_node);
2835                 if (n)
2836                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
2837                 else
2838                         nextreq = NULL;
2839                 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
2840                         /* write op */
2841                         ceph_mdsc_get_request(req);
2842                         if (nextreq)
2843                                 ceph_mdsc_get_request(nextreq);
2844                         mutex_unlock(&mdsc->mutex);
2845                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
2846                              req->r_tid, want_tid);
2847                         wait_for_completion(&req->r_safe_completion);
2848                         mutex_lock(&mdsc->mutex);
2849                         ceph_mdsc_put_request(req);
2850                         if (!nextreq)
2851                                 break;  /* next dne before, so we're done! */
2852                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
2853                                 /* next request was removed from tree */
2854                                 ceph_mdsc_put_request(nextreq);
2855                                 goto restart;
2856                         }
2857                         ceph_mdsc_put_request(nextreq);  /* won't go away */
2858                 }
2859                 req = nextreq;
2860         }
2861         mutex_unlock(&mdsc->mutex);
2862         dout("wait_unsafe_requests done\n");
2863 }
2864
2865 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
2866 {
2867         u64 want_tid, want_flush;
2868
2869         if (mdsc->client->mount_state == CEPH_MOUNT_SHUTDOWN)
2870                 return;
2871
2872         dout("sync\n");
2873         mutex_lock(&mdsc->mutex);
2874         want_tid = mdsc->last_tid;
2875         want_flush = mdsc->cap_flush_seq;
2876         mutex_unlock(&mdsc->mutex);
2877         dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
2878
2879         ceph_flush_dirty_caps(mdsc);
2880
2881         wait_unsafe_requests(mdsc, want_tid);
2882         wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
2883 }
2884
2885
2886 /*
2887  * called after sb is ro.
2888  */
2889 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
2890 {
2891         struct ceph_mds_session *session;
2892         int i;
2893         int n;
2894         struct ceph_client *client = mdsc->client;
2895         unsigned long started, timeout = client->mount_args->mount_timeout * HZ;
2896
2897         dout("close_sessions\n");
2898
2899         mutex_lock(&mdsc->mutex);
2900
2901         /* close sessions */
2902         started = jiffies;
2903         while (time_before(jiffies, started + timeout)) {
2904                 dout("closing sessions\n");
2905                 n = 0;
2906                 for (i = 0; i < mdsc->max_sessions; i++) {
2907                         session = __ceph_lookup_mds_session(mdsc, i);
2908                         if (!session)
2909                                 continue;
2910                         mutex_unlock(&mdsc->mutex);
2911                         mutex_lock(&session->s_mutex);
2912                         __close_session(mdsc, session);
2913                         mutex_unlock(&session->s_mutex);
2914                         ceph_put_mds_session(session);
2915                         mutex_lock(&mdsc->mutex);
2916                         n++;
2917                 }
2918                 if (n == 0)
2919                         break;
2920
2921                 if (client->mount_state == CEPH_MOUNT_SHUTDOWN)
2922                         break;
2923
2924                 dout("waiting for sessions to close\n");
2925                 mutex_unlock(&mdsc->mutex);
2926                 wait_for_completion_timeout(&mdsc->session_close_waiters,
2927                                             timeout);
2928                 mutex_lock(&mdsc->mutex);
2929         }
2930
2931         /* tear down remaining sessions */
2932         for (i = 0; i < mdsc->max_sessions; i++) {
2933                 if (mdsc->sessions[i]) {
2934                         session = get_session(mdsc->sessions[i]);
2935                         __unregister_session(mdsc, session);
2936                         mutex_unlock(&mdsc->mutex);
2937                         mutex_lock(&session->s_mutex);
2938                         remove_session_caps(session);
2939                         mutex_unlock(&session->s_mutex);
2940                         ceph_put_mds_session(session);
2941                         mutex_lock(&mdsc->mutex);
2942                 }
2943         }
2944
2945         WARN_ON(!list_empty(&mdsc->cap_delay_list));
2946
2947         mutex_unlock(&mdsc->mutex);
2948
2949         ceph_cleanup_empty_realms(mdsc);
2950
2951         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2952
2953         dout("stopped\n");
2954 }
2955
2956 void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
2957 {
2958         dout("stop\n");
2959         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2960         if (mdsc->mdsmap)
2961                 ceph_mdsmap_destroy(mdsc->mdsmap);
2962         kfree(mdsc->sessions);
2963 }
2964
2965
2966 /*
2967  * handle mds map update.
2968  */
2969 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
2970 {
2971         u32 epoch;
2972         u32 maplen;
2973         void *p = msg->front.iov_base;
2974         void *end = p + msg->front.iov_len;
2975         struct ceph_mdsmap *newmap, *oldmap;
2976         struct ceph_fsid fsid;
2977         int err = -EINVAL;
2978
2979         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
2980         ceph_decode_copy(&p, &fsid, sizeof(fsid));
2981         if (ceph_check_fsid(mdsc->client, &fsid) < 0)
2982                 return;
2983         epoch = ceph_decode_32(&p);
2984         maplen = ceph_decode_32(&p);
2985         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
2986
2987         /* do we need it? */
2988         ceph_monc_got_mdsmap(&mdsc->client->monc, epoch);
2989         mutex_lock(&mdsc->mutex);
2990         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
2991                 dout("handle_map epoch %u <= our %u\n",
2992                      epoch, mdsc->mdsmap->m_epoch);
2993                 mutex_unlock(&mdsc->mutex);
2994                 return;
2995         }
2996
2997         newmap = ceph_mdsmap_decode(&p, end);
2998         if (IS_ERR(newmap)) {
2999                 err = PTR_ERR(newmap);
3000                 goto bad_unlock;
3001         }
3002
3003         /* swap into place */
3004         if (mdsc->mdsmap) {
3005                 oldmap = mdsc->mdsmap;
3006                 mdsc->mdsmap = newmap;
3007                 check_new_map(mdsc, newmap, oldmap);
3008                 ceph_mdsmap_destroy(oldmap);
3009         } else {
3010                 mdsc->mdsmap = newmap;  /* first mds map */
3011         }
3012         mdsc->client->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3013
3014         __wake_requests(mdsc, &mdsc->waiting_for_map);
3015
3016         mutex_unlock(&mdsc->mutex);
3017         schedule_delayed(mdsc);
3018         return;
3019
3020 bad_unlock:
3021         mutex_unlock(&mdsc->mutex);
3022 bad:
3023         pr_err("error decoding mdsmap %d\n", err);
3024         return;
3025 }
3026
3027 static struct ceph_connection *con_get(struct ceph_connection *con)
3028 {
3029         struct ceph_mds_session *s = con->private;
3030
3031         if (get_session(s)) {
3032                 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3033                 return con;
3034         }
3035         dout("mdsc con_get %p FAIL\n", s);
3036         return NULL;
3037 }
3038
3039 static void con_put(struct ceph_connection *con)
3040 {
3041         struct ceph_mds_session *s = con->private;
3042
3043         ceph_put_mds_session(s);
3044         dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref));
3045 }
3046
3047 /*
3048  * if the client is unresponsive for long enough, the mds will kill
3049  * the session entirely.
3050  */
3051 static void peer_reset(struct ceph_connection *con)
3052 {
3053         struct ceph_mds_session *s = con->private;
3054         struct ceph_mds_client *mdsc = s->s_mdsc;
3055
3056         pr_warning("mds%d closed our session\n", s->s_mds);
3057         send_mds_reconnect(mdsc, s);
3058 }
3059
3060 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3061 {
3062         struct ceph_mds_session *s = con->private;
3063         struct ceph_mds_client *mdsc = s->s_mdsc;
3064         int type = le16_to_cpu(msg->hdr.type);
3065
3066         mutex_lock(&mdsc->mutex);
3067         if (__verify_registered_session(mdsc, s) < 0) {
3068                 mutex_unlock(&mdsc->mutex);
3069                 goto out;
3070         }
3071         mutex_unlock(&mdsc->mutex);
3072
3073         switch (type) {
3074         case CEPH_MSG_MDS_MAP:
3075                 ceph_mdsc_handle_map(mdsc, msg);
3076                 break;
3077         case CEPH_MSG_CLIENT_SESSION:
3078                 handle_session(s, msg);
3079                 break;
3080         case CEPH_MSG_CLIENT_REPLY:
3081                 handle_reply(s, msg);
3082                 break;
3083         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3084                 handle_forward(mdsc, s, msg);
3085                 break;
3086         case CEPH_MSG_CLIENT_CAPS:
3087                 ceph_handle_caps(s, msg);
3088                 break;
3089         case CEPH_MSG_CLIENT_SNAP:
3090                 ceph_handle_snap(mdsc, s, msg);
3091                 break;
3092         case CEPH_MSG_CLIENT_LEASE:
3093                 handle_lease(mdsc, s, msg);
3094                 break;
3095
3096         default:
3097                 pr_err("received unknown message type %d %s\n", type,
3098                        ceph_msg_type_name(type));
3099         }
3100 out:
3101         ceph_msg_put(msg);
3102 }
3103
3104 /*
3105  * authentication
3106  */
3107 static int get_authorizer(struct ceph_connection *con,
3108                           void **buf, int *len, int *proto,
3109                           void **reply_buf, int *reply_len, int force_new)
3110 {
3111         struct ceph_mds_session *s = con->private;
3112         struct ceph_mds_client *mdsc = s->s_mdsc;
3113         struct ceph_auth_client *ac = mdsc->client->monc.auth;
3114         int ret = 0;
3115
3116         if (force_new && s->s_authorizer) {
3117                 ac->ops->destroy_authorizer(ac, s->s_authorizer);
3118                 s->s_authorizer = NULL;
3119         }
3120         if (s->s_authorizer == NULL) {
3121                 if (ac->ops->create_authorizer) {
3122                         ret = ac->ops->create_authorizer(
3123                                 ac, CEPH_ENTITY_TYPE_MDS,
3124                                 &s->s_authorizer,
3125                                 &s->s_authorizer_buf,
3126                                 &s->s_authorizer_buf_len,
3127                                 &s->s_authorizer_reply_buf,
3128                                 &s->s_authorizer_reply_buf_len);
3129                         if (ret)
3130                                 return ret;
3131                 }
3132         }
3133
3134         *proto = ac->protocol;
3135         *buf = s->s_authorizer_buf;
3136         *len = s->s_authorizer_buf_len;
3137         *reply_buf = s->s_authorizer_reply_buf;
3138         *reply_len = s->s_authorizer_reply_buf_len;
3139         return 0;
3140 }
3141
3142
3143 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3144 {
3145         struct ceph_mds_session *s = con->private;
3146         struct ceph_mds_client *mdsc = s->s_mdsc;
3147         struct ceph_auth_client *ac = mdsc->client->monc.auth;
3148
3149         return ac->ops->verify_authorizer_reply(ac, s->s_authorizer, len);
3150 }
3151
3152 static int invalidate_authorizer(struct ceph_connection *con)
3153 {
3154         struct ceph_mds_session *s = con->private;
3155         struct ceph_mds_client *mdsc = s->s_mdsc;
3156         struct ceph_auth_client *ac = mdsc->client->monc.auth;
3157
3158         if (ac->ops->invalidate_authorizer)
3159                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3160
3161         return ceph_monc_validate_auth(&mdsc->client->monc);
3162 }
3163
3164 static const struct ceph_connection_operations mds_con_ops = {
3165         .get = con_get,
3166         .put = con_put,
3167         .dispatch = dispatch,
3168         .get_authorizer = get_authorizer,
3169         .verify_authorizer_reply = verify_authorizer_reply,
3170         .invalidate_authorizer = invalidate_authorizer,
3171         .peer_reset = peer_reset,
3172 };
3173
3174
3175
3176
3177 /* eof */