fs/ceph/mds_client.c

   1 #include "ceph_debug.h"
   2
   3 #include <linux/wait.h>
   4 #include <linux/slab.h>
   5 #include <linux/sched.h>
   6
   7 #include "mds_client.h"
   8 #include "mon_client.h"
   9 #include "super.h"
  10 #include "messenger.h"
  11 #include "decode.h"
  12 #include "auth.h"
  13 #include "pagelist.h"
  14
  15 /*
  16  * A cluster of MDS (metadata server) daemons is responsible for
  17  * managing the file system namespace (the directory hierarchy and
  18  * inodes) and for coordinating shared access to storage.  Metadata is
  19  * partitioning hierarchically across a number of servers, and that
  20  * partition varies over time as the cluster adjusts the distribution
  21  * in order to balance load.
  22  *
  23  * The MDS client is primarily responsible to managing synchronous
  24  * metadata requests for operations like open, unlink, and so forth.
  25  * If there is a MDS failure, we find out about it when we (possibly
  26  * request and) receive a new MDS map, and can resubmit affected
  27  * requests.
  28  *
  29  * For the most part, though, we take advantage of a lossless
  30  * communications channel to the MDS, and do not need to worry about
  31  * timing out or resubmitting requests.
  32  *
  33  * We maintain a stateful "session" with each MDS we interact with.
  34  * Within each session, we sent periodic heartbeat messages to ensure
  35  * any capabilities or leases we have been issues remain valid.  If
  36  * the session times out and goes stale, our leases and capabilities
  37  * are no longer valid.
  38  */
  39
  40 static void __wake_requests(struct ceph_mds_client *mdsc,
  41                             struct list_head *head);
  42
  43 const static struct ceph_connection_operations mds_con_ops;
  44
  45
  46 /*
  47  * mds reply parsing
  48  */
  49
  50 /*
  51  * parse individual inode info
  52  */
  53 static int parse_reply_info_in(void **p, void *end,
  54                                struct ceph_mds_reply_info_in *info)
  55 {
  56         int err = -EIO;
  57
  58         info->in = *p;
  59         *p += sizeof(struct ceph_mds_reply_inode) +
  60                 sizeof(*info->in->fragtree.splits) *
  61                 le32_to_cpu(info->in->fragtree.nsplits);
  62
  63         ceph_decode_32_safe(p, end, info->symlink_len, bad);
  64         ceph_decode_need(p, end, info->symlink_len, bad);
  65         info->symlink = *p;
  66         *p += info->symlink_len;
  67
  68         ceph_decode_32_safe(p, end, info->xattr_len, bad);
  69         ceph_decode_need(p, end, info->xattr_len, bad);
  70         info->xattr_data = *p;
  71         *p += info->xattr_len;
  72         return 0;
  73 bad:
  74         return err;
  75 }
  76
  77 /*
  78  * parse a normal reply, which may contain a (dir+)dentry and/or a
  79  * target inode.
  80  */
  81 static int parse_reply_info_trace(void **p, void *end,
  82                                   struct ceph_mds_reply_info_parsed *info)
  83 {
  84         int err;
  85
  86         if (info->head->is_dentry) {
  87                 err = parse_reply_info_in(p, end, &info->diri);
  88                 if (err < 0)
  89                         goto out_bad;
  90
  91                 if (unlikely(*p + sizeof(*info->dirfrag) > end))
  92                         goto bad;
  93                 info->dirfrag = *p;
  94                 *p += sizeof(*info->dirfrag) +
  95                         sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
  96                 if (unlikely(*p > end))
  97                         goto bad;
  98
  99                 ceph_decode_32_safe(p, end, info->dname_len, bad);
 100                 ceph_decode_need(p, end, info->dname_len, bad);
 101                 info->dname = *p;
 102                 *p += info->dname_len;
 103                 info->dlease = *p;
 104                 *p += sizeof(*info->dlease);
 105         }
 106
 107         if (info->head->is_target) {
 108                 err = parse_reply_info_in(p, end, &info->targeti);
 109                 if (err < 0)
 110                         goto out_bad;
 111         }
 112
 113         if (unlikely(*p != end))
 114                 goto bad;
 115         return 0;
 116
 117 bad:
 118         err = -EIO;
 119 out_bad:
 120         pr_err("problem parsing mds trace %d\n", err);
 121         return err;
 122 }
 123
 124 /*
 125  * parse readdir results
 126  */
 127 static int parse_reply_info_dir(void **p, void *end,
 128                                 struct ceph_mds_reply_info_parsed *info)
 129 {
 130         u32 num, i = 0;
 131         int err;
 132
 133         info->dir_dir = *p;
 134         if (*p + sizeof(*info->dir_dir) > end)
 135                 goto bad;
 136         *p += sizeof(*info->dir_dir) +
 137                 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
 138         if (*p > end)
 139                 goto bad;
 140
 141         ceph_decode_need(p, end, sizeof(num) + 2, bad);
 142         num = ceph_decode_32(p);
 143         info->dir_end = ceph_decode_8(p);
 144         info->dir_complete = ceph_decode_8(p);
 145         if (num == 0)
 146                 goto done;
 147
 148         /* alloc large array */
 149         info->dir_nr = num;
 150         info->dir_in = kcalloc(num, sizeof(*info->dir_in) +
 151                                sizeof(*info->dir_dname) +
 152                                sizeof(*info->dir_dname_len) +
 153                                sizeof(*info->dir_dlease),
 154                                GFP_NOFS);
 155         if (info->dir_in == NULL) {
 156                 err = -ENOMEM;
 157                 goto out_bad;
 158         }
 159         info->dir_dname = (void *)(info->dir_in + num);
 160         info->dir_dname_len = (void *)(info->dir_dname + num);
 161         info->dir_dlease = (void *)(info->dir_dname_len + num);
 162
 163         while (num) {
 164                 /* dentry */
 165                 ceph_decode_need(p, end, sizeof(u32)*2, bad);
 166                 info->dir_dname_len[i] = ceph_decode_32(p);
 167                 ceph_decode_need(p, end, info->dir_dname_len[i], bad);
 168                 info->dir_dname[i] = *p;
 169                 *p += info->dir_dname_len[i];
 170                 dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
 171                      info->dir_dname[i]);
 172                 info->dir_dlease[i] = *p;
 173                 *p += sizeof(struct ceph_mds_reply_lease);
 174
 175                 /* inode */
 176                 err = parse_reply_info_in(p, end, &info->dir_in[i]);
 177                 if (err < 0)
 178                         goto out_bad;
 179                 i++;
 180                 num--;
 181         }
 182
 183 done:
 184         if (*p != end)
 185                 goto bad;
 186         return 0;
 187
 188 bad:
 189         err = -EIO;
 190 out_bad:
 191         pr_err("problem parsing dir contents %d\n", err);
 192         return err;
 193 }
 194
 195 /*
 196  * parse entire mds reply
 197  */
 198 static int parse_reply_info(struct ceph_msg *msg,
 199                             struct ceph_mds_reply_info_parsed *info)
 200 {
 201         void *p, *end;
 202         u32 len;
 203         int err;
 204
 205         info->head = msg->front.iov_base;
 206         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 207         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 208
 209         /* trace */
 210         ceph_decode_32_safe(&p, end, len, bad);
 211         if (len > 0) {
 212                 err = parse_reply_info_trace(&p, p+len, info);
 213                 if (err < 0)
 214                         goto out_bad;
 215         }
 216
 217         /* dir content */
 218         ceph_decode_32_safe(&p, end, len, bad);
 219         if (len > 0) {
 220                 err = parse_reply_info_dir(&p, p+len, info);
 221                 if (err < 0)
 222                         goto out_bad;
 223         }
 224
 225         /* snap blob */
 226         ceph_decode_32_safe(&p, end, len, bad);
 227         info->snapblob_len = len;
 228         info->snapblob = p;
 229         p += len;
 230
 231         if (p != end)
 232                 goto bad;
 233         return 0;
 234
 235 bad:
 236         err = -EIO;
 237 out_bad:
 238         pr_err("mds parse_reply err %d\n", err);
 239         return err;
 240 }
 241
 242 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
 243 {
 244         kfree(info->dir_in);
 245 }
 246
 247
 248 /*
 249  * sessions
 250  */
 251 static const char *session_state_name(int s)
 252 {
 253         switch (s) {
 254         case CEPH_MDS_SESSION_NEW: return "new";
 255         case CEPH_MDS_SESSION_OPENING: return "opening";
 256         case CEPH_MDS_SESSION_OPEN: return "open";
 257         case CEPH_MDS_SESSION_HUNG: return "hung";
 258         case CEPH_MDS_SESSION_CLOSING: return "closing";
 259         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
 260         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
 261         default: return "???";
 262         }
 263 }
 264
 265 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
 266 {
 267         if (atomic_inc_not_zero(&s->s_ref)) {
 268                 dout("mdsc get_session %p %d -> %d\n", s,
 269                      atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
 270                 return s;
 271         } else {
 272                 dout("mdsc get_session %p 0 -- FAIL", s);
 273                 return NULL;
 274         }
 275 }
 276
 277 void ceph_put_mds_session(struct ceph_mds_session *s)
 278 {
 279         dout("mdsc put_session %p %d -> %d\n", s,
 280              atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
 281         if (atomic_dec_and_test(&s->s_ref)) {
 282                 if (s->s_authorizer)
 283                         s->s_mdsc->client->monc.auth->ops->destroy_authorizer(
 284                                 s->s_mdsc->client->monc.auth, s->s_authorizer);
 285                 kfree(s);
 286         }
 287 }
 288
 289 /*
 290  * called under mdsc->mutex
 291  */
 292 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
 293                                                    int mds)
 294 {
 295         struct ceph_mds_session *session;
 296
 297         if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
 298                 return NULL;
 299         session = mdsc->sessions[mds];
 300         dout("lookup_mds_session %p %d\n", session,
 301              atomic_read(&session->s_ref));
 302         get_session(session);
 303         return session;
 304 }
 305
 306 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
 307 {
 308         if (mds >= mdsc->max_sessions)
 309                 return false;
 310         return mdsc->sessions[mds];
 311 }
 312
 313 static int __verify_registered_session(struct ceph_mds_client *mdsc,
 314                                        struct ceph_mds_session *s)
 315 {
 316         if (s->s_mds >= mdsc->max_sessions ||
 317             mdsc->sessions[s->s_mds] != s)
 318                 return -ENOENT;
 319         return 0;
 320 }
 321
 322 /*
 323  * create+register a new session for given mds.
 324  * called under mdsc->mutex.
 325  */
 326 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 327                                                  int mds)
 328 {
 329         struct ceph_mds_session *s;
 330
 331         s = kzalloc(sizeof(*s), GFP_NOFS);
 332         if (!s)
 333                 return ERR_PTR(-ENOMEM);
 334         s->s_mdsc = mdsc;
 335         s->s_mds = mds;
 336         s->s_state = CEPH_MDS_SESSION_NEW;
 337         s->s_ttl = 0;
 338         s->s_seq = 0;
 339         mutex_init(&s->s_mutex);
 340
 341         ceph_con_init(mdsc->client->msgr, &s->s_con);
 342         s->s_con.private = s;
 343         s->s_con.ops = &mds_con_ops;
 344         s->s_con.peer_name.type = CEPH_ENTITY_TYPE_MDS;
 345         s->s_con.peer_name.num = cpu_to_le64(mds);
 346
 347         spin_lock_init(&s->s_cap_lock);
 348         s->s_cap_gen = 0;
 349         s->s_cap_ttl = 0;
 350         s->s_renew_requested = 0;
 351         s->s_renew_seq = 0;
 352         INIT_LIST_HEAD(&s->s_caps);
 353         s->s_nr_caps = 0;
 354         s->s_trim_caps = 0;
 355         atomic_set(&s->s_ref, 1);
 356         INIT_LIST_HEAD(&s->s_waiting);
 357         INIT_LIST_HEAD(&s->s_unsafe);
 358         s->s_num_cap_releases = 0;
 359         s->s_cap_iterator = NULL;
 360         INIT_LIST_HEAD(&s->s_cap_releases);
 361         INIT_LIST_HEAD(&s->s_cap_releases_done);
 362         INIT_LIST_HEAD(&s->s_cap_flushing);
 363         INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
 364
 365         dout("register_session mds%d\n", mds);
 366         if (mds >= mdsc->max_sessions) {
 367                 int newmax = 1 << get_count_order(mds+1);
 368                 struct ceph_mds_session **sa;
 369
 370                 dout("register_session realloc to %d\n", newmax);
 371                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 372                 if (sa == NULL)
 373                         goto fail_realloc;
 374                 if (mdsc->sessions) {
 375                         memcpy(sa, mdsc->sessions,
 376                                mdsc->max_sessions * sizeof(void *));
 377                         kfree(mdsc->sessions);
 378                 }
 379                 mdsc->sessions = sa;
 380                 mdsc->max_sessions = newmax;
 381         }
 382         mdsc->sessions[mds] = s;
 383         atomic_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
 384
 385         ceph_con_open(&s->s_con, ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
 386
 387         return s;
 388
 389 fail_realloc:
 390         kfree(s);
 391         return ERR_PTR(-ENOMEM);
 392 }
 393
 394 /*
 395  * called under mdsc->mutex
 396  */
 397 static void __unregister_session(struct ceph_mds_client *mdsc,
 398                                struct ceph_mds_session *s)
 399 {
 400         dout("__unregister_session mds%d %p\n", s->s_mds, s);
 401         BUG_ON(mdsc->sessions[s->s_mds] != s);
 402         mdsc->sessions[s->s_mds] = NULL;
 403         ceph_con_close(&s->s_con);
 404         ceph_put_mds_session(s);
 405 }
 406
 407 /*
 408  * drop session refs in request.
 409  *
 410  * should be last request ref, or hold mdsc->mutex
 411  */
 412 static void put_request_session(struct ceph_mds_request *req)
 413 {
 414         if (req->r_session) {
 415                 ceph_put_mds_session(req->r_session);
 416                 req->r_session = NULL;
 417         }
 418 }
 419
 420 void ceph_mdsc_release_request(struct kref *kref)
 421 {
 422         struct ceph_mds_request *req = container_of(kref,
 423                                                     struct ceph_mds_request,
 424                                                     r_kref);
 425         if (req->r_request)
 426                 ceph_msg_put(req->r_request);
 427         if (req->r_reply) {
 428                 ceph_msg_put(req->r_reply);
 429                 destroy_reply_info(&req->r_reply_info);
 430         }
 431         if (req->r_inode) {
 432                 ceph_put_cap_refs(ceph_inode(req->r_inode),
 433                                   CEPH_CAP_PIN);
 434                 iput(req->r_inode);
 435         }
 436         if (req->r_locked_dir)
 437                 ceph_put_cap_refs(ceph_inode(req->r_locked_dir),
 438                                   CEPH_CAP_PIN);
 439         if (req->r_target_inode)
 440                 iput(req->r_target_inode);
 441         if (req->r_dentry)
 442                 dput(req->r_dentry);
 443         if (req->r_old_dentry) {
 444                 ceph_put_cap_refs(
 445                         ceph_inode(req->r_old_dentry->d_parent->d_inode),
 446                         CEPH_CAP_PIN);
 447                 dput(req->r_old_dentry);
 448         }
 449         kfree(req->r_path1);
 450         kfree(req->r_path2);
 451         put_request_session(req);
 452         ceph_unreserve_caps(&req->r_caps_reservation);
 453         kfree(req);
 454 }
 455
 456 /*
 457  * lookup session, bump ref if found.
 458  *
 459  * called under mdsc->mutex.
 460  */
 461 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
 462                                              u64 tid)
 463 {
 464         struct ceph_mds_request *req;
 465         struct rb_node *n = mdsc->request_tree.rb_node;
 466
 467         while (n) {
 468                 req = rb_entry(n, struct ceph_mds_request, r_node);
 469                 if (tid < req->r_tid)
 470                         n = n->rb_left;
 471                 else if (tid > req->r_tid)
 472                         n = n->rb_right;
 473                 else {
 474                         ceph_mdsc_get_request(req);
 475                         return req;
 476                 }
 477         }
 478         return NULL;
 479 }
 480
 481 static void __insert_request(struct ceph_mds_client *mdsc,
 482                              struct ceph_mds_request *new)
 483 {
 484         struct rb_node **p = &mdsc->request_tree.rb_node;
 485         struct rb_node *parent = NULL;
 486         struct ceph_mds_request *req = NULL;
 487
 488         while (*p) {
 489                 parent = *p;
 490                 req = rb_entry(parent, struct ceph_mds_request, r_node);
 491                 if (new->r_tid < req->r_tid)
 492                         p = &(*p)->rb_left;
 493                 else if (new->r_tid > req->r_tid)
 494                         p = &(*p)->rb_right;
 495                 else
 496                         BUG();
 497         }
 498
 499         rb_link_node(&new->r_node, parent, p);
 500         rb_insert_color(&new->r_node, &mdsc->request_tree);
 501 }
 502
 503 /*
 504  * Register an in-flight request, and assign a tid.  Link to directory
 505  * are modifying (if any).
 506  *
 507  * Called under mdsc->mutex.
 508  */
 509 static void __register_request(struct ceph_mds_client *mdsc,
 510                                struct ceph_mds_request *req,
 511                                struct inode *dir)
 512 {
 513         req->r_tid = ++mdsc->last_tid;
 514         if (req->r_num_caps)
 515                 ceph_reserve_caps(&req->r_caps_reservation, req->r_num_caps);
 516         dout("__register_request %p tid %lld\n", req, req->r_tid);
 517         ceph_mdsc_get_request(req);
 518         __insert_request(mdsc, req);
 519
 520         if (dir) {
 521                 struct ceph_inode_info *ci = ceph_inode(dir);
 522
 523                 spin_lock(&ci->i_unsafe_lock);
 524                 req->r_unsafe_dir = dir;
 525                 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
 526                 spin_unlock(&ci->i_unsafe_lock);
 527         }
 528 }
 529
 530 static void __unregister_request(struct ceph_mds_client *mdsc,
 531                                  struct ceph_mds_request *req)
 532 {
 533         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 534         rb_erase(&req->r_node, &mdsc->request_tree);
 535         RB_CLEAR_NODE(&req->r_node);
 536
 537         if (req->r_unsafe_dir) {
 538                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 539
 540                 spin_lock(&ci->i_unsafe_lock);
 541                 list_del_init(&req->r_unsafe_dir_item);
 542                 spin_unlock(&ci->i_unsafe_lock);
 543         }
 544
 545         ceph_mdsc_put_request(req);
 546 }
 547
 548 /*
 549  * Choose mds to send request to next.  If there is a hint set in the
 550  * request (e.g., due to a prior forward hint from the mds), use that.
 551  * Otherwise, consult frag tree and/or caps to identify the
 552  * appropriate mds.  If all else fails, choose randomly.
 553  *
 554  * Called under mdsc->mutex.
 555  */
 556 static int __choose_mds(struct ceph_mds_client *mdsc,
 557                         struct ceph_mds_request *req)
 558 {
 559         struct inode *inode;
 560         struct ceph_inode_info *ci;
 561         struct ceph_cap *cap;
 562         int mode = req->r_direct_mode;
 563         int mds = -1;
 564         u32 hash = req->r_direct_hash;
 565         bool is_hash = req->r_direct_is_hash;
 566
 567         /*
 568          * is there a specific mds we should try?  ignore hint if we have
 569          * no session and the mds is not up (active or recovering).
 570          */
 571         if (req->r_resend_mds >= 0 &&
 572             (__have_session(mdsc, req->r_resend_mds) ||
 573              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 574                 dout("choose_mds using resend_mds mds%d\n",
 575                      req->r_resend_mds);
 576                 return req->r_resend_mds;
 577         }
 578
 579         if (mode == USE_RANDOM_MDS)
 580                 goto random;
 581
 582         inode = NULL;
 583         if (req->r_inode) {
 584                 inode = req->r_inode;
 585         } else if (req->r_dentry) {
 586                 if (req->r_dentry->d_inode) {
 587                         inode = req->r_dentry->d_inode;
 588                 } else {
 589                         inode = req->r_dentry->d_parent->d_inode;
 590                         hash = req->r_dentry->d_name.hash;
 591                         is_hash = true;
 592                 }
 593         }
 594         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 595              (int)hash, mode);
 596         if (!inode)
 597                 goto random;
 598         ci = ceph_inode(inode);
 599
 600         if (is_hash && S_ISDIR(inode->i_mode)) {
 601                 struct ceph_inode_frag frag;
 602                 int found;
 603
 604                 ceph_choose_frag(ci, hash, &frag, &found);
 605                 if (found) {
 606                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 607                                 u8 r;
 608
 609                                 /* choose a random replica */
 610                                 get_random_bytes(&r, 1);
 611                                 r %= frag.ndist;
 612                                 mds = frag.dist[r];
 613                                 dout("choose_mds %p %llx.%llx "
 614                                      "frag %u mds%d (%d/%d)\n",
 615                                      inode, ceph_vinop(inode),
 616                                      frag.frag, frag.mds,
 617                                      (int)r, frag.ndist);
 618                                 return mds;
 619                         }
 620
 621                         /* since this file/dir wasn't known to be
 622                          * replicated, then we want to look for the
 623                          * authoritative mds. */
 624                         mode = USE_AUTH_MDS;
 625                         if (frag.mds >= 0) {
 626                                 /* choose auth mds */
 627                                 mds = frag.mds;
 628                                 dout("choose_mds %p %llx.%llx "
 629                                      "frag %u mds%d (auth)\n",
 630                                      inode, ceph_vinop(inode), frag.frag, mds);
 631                                 return mds;
 632                         }
 633                 }
 634         }
 635
 636         spin_lock(&inode->i_lock);
 637         cap = NULL;
 638         if (mode == USE_AUTH_MDS)
 639                 cap = ci->i_auth_cap;
 640         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
 641                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
 642         if (!cap) {
 643                 spin_unlock(&inode->i_lock);
 644                 goto random;
 645         }
 646         mds = cap->session->s_mds;
 647         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
 648              inode, ceph_vinop(inode), mds,
 649              cap == ci->i_auth_cap ? "auth " : "", cap);
 650         spin_unlock(&inode->i_lock);
 651         return mds;
 652
 653 random:
 654         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
 655         dout("choose_mds chose random mds%d\n", mds);
 656         return mds;
 657 }
 658
 659
 660 /*
 661  * session messages
 662  */
 663 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
 664 {
 665         struct ceph_msg *msg;
 666         struct ceph_mds_session_head *h;
 667
 668         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h));
 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         dout("removing cap %p, ci is %p, inode is %p\n",
 806              cap, ci, &ci->vfs_inode);
 807         ceph_remove_cap(cap);
 808         return 0;
 809 }
 810
 811 /*
 812  * caller must hold session s_mutex
 813  */
 814 static void remove_session_caps(struct ceph_mds_session *session)
 815 {
 816         dout("remove_session_caps on %p\n", session);
 817         iterate_session_caps(session, remove_session_caps_cb, NULL);
 818         BUG_ON(session->s_nr_caps > 0);
 819         cleanup_cap_releases(session);
 820 }
 821
 822 /*
 823  * wake up any threads waiting on this session's caps.  if the cap is
 824  * old (didn't get renewed on the client reconnect), remove it now.
 825  *
 826  * caller must hold s_mutex.
 827  */
 828 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
 829                               void *arg)
 830 {
 831         struct ceph_inode_info *ci = ceph_inode(inode);
 832
 833         wake_up(&ci->i_cap_wq);
 834         if (arg) {
 835                 spin_lock(&inode->i_lock);
 836                 ci->i_wanted_max_size = 0;
 837                 ci->i_requested_max_size = 0;
 838                 spin_unlock(&inode->i_lock);
 839         }
 840         return 0;
 841 }
 842
 843 static void wake_up_session_caps(struct ceph_mds_session *session,
 844                                  int reconnect)
 845 {
 846         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
 847         iterate_session_caps(session, wake_up_session_cb,
 848                              (void *)(unsigned long)reconnect);
 849 }
 850
 851 /*
 852  * Send periodic message to MDS renewing all currently held caps.  The
 853  * ack will reset the expiration for all caps from this session.
 854  *
 855  * caller holds s_mutex
 856  */
 857 static int send_renew_caps(struct ceph_mds_client *mdsc,
 858                            struct ceph_mds_session *session)
 859 {
 860         struct ceph_msg *msg;
 861         int state;
 862
 863         if (time_after_eq(jiffies, session->s_cap_ttl) &&
 864             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
 865                 pr_info("mds%d caps stale\n", session->s_mds);
 866         session->s_renew_requested = jiffies;
 867
 868         /* do not try to renew caps until a recovering mds has reconnected
 869          * with its clients. */
 870         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
 871         if (state < CEPH_MDS_STATE_RECONNECT) {
 872                 dout("send_renew_caps ignoring mds%d (%s)\n",
 873                      session->s_mds, ceph_mds_state_name(state));
 874                 return 0;
 875         }
 876
 877         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
 878                 ceph_mds_state_name(state));
 879         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
 880                                  ++session->s_renew_seq);
 881         if (!msg)
 882                 return -ENOMEM;
 883         ceph_con_send(&session->s_con, msg);
 884         return 0;
 885 }
 886
 887 /*
 888  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
 889  *
 890  * Called under session->s_mutex
 891  */
 892 static void renewed_caps(struct ceph_mds_client *mdsc,
 893                          struct ceph_mds_session *session, int is_renew)
 894 {
 895         int was_stale;
 896         int wake = 0;
 897
 898         spin_lock(&session->s_cap_lock);
 899         was_stale = is_renew && (session->s_cap_ttl == 0 ||
 900                                  time_after_eq(jiffies, session->s_cap_ttl));
 901
 902         session->s_cap_ttl = session->s_renew_requested +
 903                 mdsc->mdsmap->m_session_timeout*HZ;
 904
 905         if (was_stale) {
 906                 if (time_before(jiffies, session->s_cap_ttl)) {
 907                         pr_info("mds%d caps renewed\n", session->s_mds);
 908                         wake = 1;
 909                 } else {
 910                         pr_info("mds%d caps still stale\n", session->s_mds);
 911                 }
 912         }
 913         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
 914              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
 915              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
 916         spin_unlock(&session->s_cap_lock);
 917
 918         if (wake)
 919                 wake_up_session_caps(session, 0);
 920 }
 921
 922 /*
 923  * send a session close request
 924  */
 925 static int request_close_session(struct ceph_mds_client *mdsc,
 926                                  struct ceph_mds_session *session)
 927 {
 928         struct ceph_msg *msg;
 929
 930         dout("request_close_session mds%d state %s seq %lld\n",
 931              session->s_mds, session_state_name(session->s_state),
 932              session->s_seq);
 933         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
 934         if (!msg)
 935                 return -ENOMEM;
 936         ceph_con_send(&session->s_con, msg);
 937         return 0;
 938 }
 939
 940 /*
 941  * Called with s_mutex held.
 942  */
 943 static int __close_session(struct ceph_mds_client *mdsc,
 944                          struct ceph_mds_session *session)
 945 {
 946         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
 947                 return 0;
 948         session->s_state = CEPH_MDS_SESSION_CLOSING;
 949         return request_close_session(mdsc, session);
 950 }
 951
 952 /*
 953  * Trim old(er) caps.
 954  *
 955  * Because we can't cache an inode without one or more caps, we do
 956  * this indirectly: if a cap is unused, we prune its aliases, at which
 957  * point the inode will hopefully get dropped to.
 958  *
 959  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
 960  * memory pressure from the MDS, though, so it needn't be perfect.
 961  */
 962 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
 963 {
 964         struct ceph_mds_session *session = arg;
 965         struct ceph_inode_info *ci = ceph_inode(inode);
 966         int used, oissued, mine;
 967
 968         if (session->s_trim_caps <= 0)
 969                 return -1;
 970
 971         spin_lock(&inode->i_lock);
 972         mine = cap->issued | cap->implemented;
 973         used = __ceph_caps_used(ci);
 974         oissued = __ceph_caps_issued_other(ci, cap);
 975
 976         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
 977              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
 978              ceph_cap_string(used));
 979         if (ci->i_dirty_caps)
 980                 goto out;   /* dirty caps */
 981         if ((used & ~oissued) & mine)
 982                 goto out;   /* we need these caps */
 983
 984         session->s_trim_caps--;
 985         if (oissued) {
 986                 /* we aren't the only cap.. just remove us */
 987                 __ceph_remove_cap(cap);
 988         } else {
 989                 /* try to drop referring dentries */
 990                 spin_unlock(&inode->i_lock);
 991                 d_prune_aliases(inode);
 992                 dout("trim_caps_cb %p cap %p  pruned, count now %d\n",
 993                      inode, cap, atomic_read(&inode->i_count));
 994                 return 0;
 995         }
 996
 997 out:
 998         spin_unlock(&inode->i_lock);
 999         return 0;
1000 }
1001
1002 /*
1003  * Trim session cap count down to some max number.
1004  */
1005 static int trim_caps(struct ceph_mds_client *mdsc,
1006                      struct ceph_mds_session *session,
1007                      int max_caps)
1008 {
1009         int trim_caps = session->s_nr_caps - max_caps;
1010
1011         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1012              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1013         if (trim_caps > 0) {
1014                 session->s_trim_caps = trim_caps;
1015                 iterate_session_caps(session, trim_caps_cb, session);
1016                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1017                      session->s_mds, session->s_nr_caps, max_caps,
1018                         trim_caps - session->s_trim_caps);
1019                 session->s_trim_caps = 0;
1020         }
1021         return 0;
1022 }
1023
1024 /*
1025  * Allocate cap_release messages.  If there is a partially full message
1026  * in the queue, try to allocate enough to cover it's remainder, so that
1027  * we can send it immediately.
1028  *
1029  * Called under s_mutex.
1030  */
1031 static int add_cap_releases(struct ceph_mds_client *mdsc,
1032                             struct ceph_mds_session *session,
1033                             int extra)
1034 {
1035         struct ceph_msg *msg;
1036         struct ceph_mds_cap_release *head;
1037         int err = -ENOMEM;
1038
1039         if (extra < 0)
1040                 extra = mdsc->client->mount_args->cap_release_safety;
1041
1042         spin_lock(&session->s_cap_lock);
1043
1044         if (!list_empty(&session->s_cap_releases)) {
1045                 msg = list_first_entry(&session->s_cap_releases,
1046                                        struct ceph_msg,
1047                                  list_head);
1048                 head = msg->front.iov_base;
1049                 extra += CEPH_CAPS_PER_RELEASE - le32_to_cpu(head->num);
1050         }
1051
1052         while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1053                 spin_unlock(&session->s_cap_lock);
1054                 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE);
1055                 if (!msg)
1056                         goto out_unlocked;
1057                 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1058                      (int)msg->front.iov_len);
1059                 head = msg->front.iov_base;
1060                 head->num = cpu_to_le32(0);
1061                 msg->front.iov_len = sizeof(*head);
1062                 spin_lock(&session->s_cap_lock);
1063                 list_add(&msg->list_head, &session->s_cap_releases);
1064                 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1065         }
1066
1067         if (!list_empty(&session->s_cap_releases)) {
1068                 msg = list_first_entry(&session->s_cap_releases,
1069                                        struct ceph_msg,
1070                                        list_head);
1071                 head = msg->front.iov_base;
1072                 if (head->num) {
1073                         dout(" queueing non-full %p (%d)\n", msg,
1074                              le32_to_cpu(head->num));
1075                         list_move_tail(&msg->list_head,
1076                                       &session->s_cap_releases_done);
1077                         session->s_num_cap_releases -=
1078                                 CEPH_CAPS_PER_RELEASE - le32_to_cpu(head->num);
1079                 }
1080         }
1081         err = 0;
1082         spin_unlock(&session->s_cap_lock);
1083 out_unlocked:
1084         return err;
1085 }
1086
1087 /*
1088  * flush all dirty inode data to disk.
1089  *
1090  * returns true if we've flushed through want_flush_seq
1091  */
1092 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1093 {
1094         int mds, ret = 1;
1095
1096         dout("check_cap_flush want %lld\n", want_flush_seq);
1097         mutex_lock(&mdsc->mutex);
1098         for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1099                 struct ceph_mds_session *session = mdsc->sessions[mds];
1100
1101                 if (!session)
1102                         continue;
1103                 get_session(session);
1104                 mutex_unlock(&mdsc->mutex);
1105
1106                 mutex_lock(&session->s_mutex);
1107                 if (!list_empty(&session->s_cap_flushing)) {
1108                         struct ceph_inode_info *ci =
1109                                 list_entry(session->s_cap_flushing.next,
1110                                            struct ceph_inode_info,
1111                                            i_flushing_item);
1112                         struct inode *inode = &ci->vfs_inode;
1113
1114                         spin_lock(&inode->i_lock);
1115                         if (ci->i_cap_flush_seq <= want_flush_seq) {
1116                                 dout("check_cap_flush still flushing %p "
1117                                      "seq %lld <= %lld to mds%d\n", inode,
1118                                      ci->i_cap_flush_seq, want_flush_seq,
1119                                      session->s_mds);
1120                                 ret = 0;
1121                         }
1122                         spin_unlock(&inode->i_lock);
1123                 }
1124                 mutex_unlock(&session->s_mutex);
1125                 ceph_put_mds_session(session);
1126
1127                 if (!ret)
1128                         return ret;
1129                 mutex_lock(&mdsc->mutex);
1130         }
1131
1132         mutex_unlock(&mdsc->mutex);
1133         dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1134         return ret;
1135 }
1136
1137 /*
1138  * called under s_mutex
1139  */
1140 static void send_cap_releases(struct ceph_mds_client *mdsc,
1141                        struct ceph_mds_session *session)
1142 {
1143         struct ceph_msg *msg;
1144
1145         dout("send_cap_releases mds%d\n", session->s_mds);
1146         spin_lock(&session->s_cap_lock);
1147         while (!list_empty(&session->s_cap_releases_done)) {
1148                 msg = list_first_entry(&session->s_cap_releases_done,
1149                                  struct ceph_msg, list_head);
1150                 list_del_init(&msg->list_head);
1151                 spin_unlock(&session->s_cap_lock);
1152                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1153                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1154                 ceph_con_send(&session->s_con, msg);
1155                 spin_lock(&session->s_cap_lock);
1156         }
1157         spin_unlock(&session->s_cap_lock);
1158 }
1159
1160 /*
1161  * requests
1162  */
1163
1164 /*
1165  * Create an mds request.
1166  */
1167 struct ceph_mds_request *
1168 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1169 {
1170         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1171
1172         if (!req)
1173                 return ERR_PTR(-ENOMEM);
1174
1175         mutex_init(&req->r_fill_mutex);
1176         req->r_started = jiffies;
1177         req->r_resend_mds = -1;
1178         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1179         req->r_fmode = -1;
1180         kref_init(&req->r_kref);
1181         INIT_LIST_HEAD(&req->r_wait);
1182         init_completion(&req->r_completion);
1183         init_completion(&req->r_safe_completion);
1184         INIT_LIST_HEAD(&req->r_unsafe_item);
1185
1186         req->r_op = op;
1187         req->r_direct_mode = mode;
1188         return req;
1189 }
1190
1191 /*
1192  * return oldest (lowest) request, tid in request tree, 0 if none.
1193  *
1194  * called under mdsc->mutex.
1195  */
1196 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1197 {
1198         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1199                 return NULL;
1200         return rb_entry(rb_first(&mdsc->request_tree),
1201                         struct ceph_mds_request, r_node);
1202 }
1203
1204 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1205 {
1206         struct ceph_mds_request *req = __get_oldest_req(mdsc);
1207
1208         if (req)
1209                 return req->r_tid;
1210         return 0;
1211 }
1212
1213 /*
1214  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1215  * on build_path_from_dentry in fs/cifs/dir.c.
1216  *
1217  * If @stop_on_nosnap, generate path relative to the first non-snapped
1218  * inode.
1219  *
1220  * Encode hidden .snap dirs as a double /, i.e.
1221  *   foo/.snap/bar -> foo//bar
1222  */
1223 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1224                            int stop_on_nosnap)
1225 {
1226         struct dentry *temp;
1227         char *path;
1228         int len, pos;
1229
1230         if (dentry == NULL)
1231                 return ERR_PTR(-EINVAL);
1232
1233 retry:
1234         len = 0;
1235         for (temp = dentry; !IS_ROOT(temp);) {
1236                 struct inode *inode = temp->d_inode;
1237                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1238                         len++;  /* slash only */
1239                 else if (stop_on_nosnap && inode &&
1240                          ceph_snap(inode) == CEPH_NOSNAP)
1241                         break;
1242                 else
1243                         len += 1 + temp->d_name.len;
1244                 temp = temp->d_parent;
1245                 if (temp == NULL) {
1246                         pr_err("build_path_dentry corrupt dentry %p\n", dentry);
1247                         return ERR_PTR(-EINVAL);
1248                 }
1249         }
1250         if (len)
1251                 len--;  /* no leading '/' */
1252
1253         path = kmalloc(len+1, GFP_NOFS);
1254         if (path == NULL)
1255                 return ERR_PTR(-ENOMEM);
1256         pos = len;
1257         path[pos] = 0;  /* trailing null */
1258         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1259                 struct inode *inode = temp->d_inode;
1260
1261                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1262                         dout("build_path path+%d: %p SNAPDIR\n",
1263                              pos, temp);
1264                 } else if (stop_on_nosnap && inode &&
1265                            ceph_snap(inode) == CEPH_NOSNAP) {
1266                         break;
1267                 } else {
1268                         pos -= temp->d_name.len;
1269                         if (pos < 0)
1270                                 break;
1271                         strncpy(path + pos, temp->d_name.name,
1272                                 temp->d_name.len);
1273                 }
1274                 if (pos)
1275                         path[--pos] = '/';
1276                 temp = temp->d_parent;
1277                 if (temp == NULL) {
1278                         pr_err("build_path corrupt dentry\n");
1279                         kfree(path);
1280                         return ERR_PTR(-EINVAL);
1281                 }
1282         }
1283         if (pos != 0) {
1284                 pr_err("build_path did not end path lookup where "
1285                        "expected, namelen is %d, pos is %d\n", len, pos);
1286                 /* presumably this is only possible if racing with a
1287                    rename of one of the parent directories (we can not
1288                    lock the dentries above us to prevent this, but
1289                    retrying should be harmless) */
1290                 kfree(path);
1291                 goto retry;
1292         }
1293
1294         *base = ceph_ino(temp->d_inode);
1295         *plen = len;
1296         dout("build_path on %p %d built %llx '%.*s'\n",
1297              dentry, atomic_read(&dentry->d_count), *base, len, path);
1298         return path;
1299 }
1300
1301 static int build_dentry_path(struct dentry *dentry,
1302                              const char **ppath, int *ppathlen, u64 *pino,
1303                              int *pfreepath)
1304 {
1305         char *path;
1306
1307         if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1308                 *pino = ceph_ino(dentry->d_parent->d_inode);
1309                 *ppath = dentry->d_name.name;
1310                 *ppathlen = dentry->d_name.len;
1311                 return 0;
1312         }
1313         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1314         if (IS_ERR(path))
1315                 return PTR_ERR(path);
1316         *ppath = path;
1317         *pfreepath = 1;
1318         return 0;
1319 }
1320
1321 static int build_inode_path(struct inode *inode,
1322                             const char **ppath, int *ppathlen, u64 *pino,
1323                             int *pfreepath)
1324 {
1325         struct dentry *dentry;
1326         char *path;
1327
1328         if (ceph_snap(inode) == CEPH_NOSNAP) {
1329                 *pino = ceph_ino(inode);
1330                 *ppathlen = 0;
1331                 return 0;
1332         }
1333         dentry = d_find_alias(inode);
1334         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1335         dput(dentry);
1336         if (IS_ERR(path))
1337                 return PTR_ERR(path);
1338         *ppath = path;
1339         *pfreepath = 1;
1340         return 0;
1341 }
1342
1343 /*
1344  * request arguments may be specified via an inode *, a dentry *, or
1345  * an explicit ino+path.
1346  */
1347 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1348                                   const char *rpath, u64 rino,
1349                                   const char **ppath, int *pathlen,
1350                                   u64 *ino, int *freepath)
1351 {
1352         int r = 0;
1353
1354         if (rinode) {
1355                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1356                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1357                      ceph_snap(rinode));
1358         } else if (rdentry) {
1359                 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1360                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1361                      *ppath);
1362         } else if (rpath) {
1363                 *ino = rino;
1364                 *ppath = rpath;
1365                 *pathlen = strlen(rpath);
1366                 dout(" path %.*s\n", *pathlen, rpath);
1367         }
1368
1369         return r;
1370 }
1371
1372 /*
1373  * called under mdsc->mutex
1374  */
1375 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1376                                                struct ceph_mds_request *req,
1377                                                int mds)
1378 {
1379         struct ceph_msg *msg;
1380         struct ceph_mds_request_head *head;
1381         const char *path1 = NULL;
1382         const char *path2 = NULL;
1383         u64 ino1 = 0, ino2 = 0;
1384         int pathlen1 = 0, pathlen2 = 0;
1385         int freepath1 = 0, freepath2 = 0;
1386         int len;
1387         u16 releases;
1388         void *p, *end;
1389         int ret;
1390
1391         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1392                               req->r_path1, req->r_ino1.ino,
1393                               &path1, &pathlen1, &ino1, &freepath1);
1394         if (ret < 0) {
1395                 msg = ERR_PTR(ret);
1396                 goto out;
1397         }
1398
1399         ret = set_request_path_attr(NULL, req->r_old_dentry,
1400                               req->r_path2, req->r_ino2.ino,
1401                               &path2, &pathlen2, &ino2, &freepath2);
1402         if (ret < 0) {
1403                 msg = ERR_PTR(ret);
1404                 goto out_free1;
1405         }
1406
1407         len = sizeof(*head) +
1408                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64));
1409
1410         /* calculate (max) length for cap releases */
1411         len += sizeof(struct ceph_mds_request_release) *
1412                 (!!req->r_inode_drop + !!req->r_dentry_drop +
1413                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1414         if (req->r_dentry_drop)
1415                 len += req->r_dentry->d_name.len;
1416         if (req->r_old_dentry_drop)
1417                 len += req->r_old_dentry->d_name.len;
1418
1419         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len);
1420         if (!msg) {
1421                 msg = ERR_PTR(-ENOMEM);
1422                 goto out_free2;
1423         }
1424
1425         msg->hdr.tid = cpu_to_le64(req->r_tid);
1426
1427         head = msg->front.iov_base;
1428         p = msg->front.iov_base + sizeof(*head);
1429         end = msg->front.iov_base + msg->front.iov_len;
1430
1431         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1432         head->op = cpu_to_le32(req->r_op);
1433         head->caller_uid = cpu_to_le32(current_fsuid());
1434         head->caller_gid = cpu_to_le32(current_fsgid());
1435         head->args = req->r_args;
1436
1437         ceph_encode_filepath(&p, end, ino1, path1);
1438         ceph_encode_filepath(&p, end, ino2, path2);
1439
1440         /* cap releases */
1441         releases = 0;
1442         if (req->r_inode_drop)
1443                 releases += ceph_encode_inode_release(&p,
1444                       req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1445                       mds, req->r_inode_drop, req->r_inode_unless, 0);
1446         if (req->r_dentry_drop)
1447                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1448                        mds, req->r_dentry_drop, req->r_dentry_unless);
1449         if (req->r_old_dentry_drop)
1450                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1451                        mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1452         if (req->r_old_inode_drop)
1453                 releases += ceph_encode_inode_release(&p,
1454                       req->r_old_dentry->d_inode,
1455                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1456         head->num_releases = cpu_to_le16(releases);
1457
1458         BUG_ON(p > end);
1459         msg->front.iov_len = p - msg->front.iov_base;
1460         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1461
1462         msg->pages = req->r_pages;
1463         msg->nr_pages = req->r_num_pages;
1464         msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1465         msg->hdr.data_off = cpu_to_le16(0);
1466
1467 out_free2:
1468         if (freepath2)
1469                 kfree((char *)path2);
1470 out_free1:
1471         if (freepath1)
1472                 kfree((char *)path1);
1473 out:
1474         return msg;
1475 }
1476
1477 /*
1478  * called under mdsc->mutex if error, under no mutex if
1479  * success.
1480  */
1481 static void complete_request(struct ceph_mds_client *mdsc,
1482                              struct ceph_mds_request *req)
1483 {
1484         if (req->r_callback)
1485                 req->r_callback(mdsc, req);
1486         else
1487                 complete(&req->r_completion);
1488 }
1489
1490 /*
1491  * called under mdsc->mutex
1492  */
1493 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1494                                   struct ceph_mds_request *req,
1495                                   int mds)
1496 {
1497         struct ceph_mds_request_head *rhead;
1498         struct ceph_msg *msg;
1499         int flags = 0;
1500
1501         req->r_mds = mds;
1502         req->r_attempts++;
1503         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1504              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1505
1506         if (req->r_request) {
1507                 ceph_msg_put(req->r_request);
1508                 req->r_request = NULL;
1509         }
1510         msg = create_request_message(mdsc, req, mds);
1511         if (IS_ERR(msg)) {
1512                 req->r_err = PTR_ERR(msg);
1513                 complete_request(mdsc, req);
1514                 return PTR_ERR(msg);
1515         }
1516         req->r_request = msg;
1517
1518         rhead = msg->front.iov_base;
1519         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1520         if (req->r_got_unsafe)
1521                 flags |= CEPH_MDS_FLAG_REPLAY;
1522         if (req->r_locked_dir)
1523                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1524         rhead->flags = cpu_to_le32(flags);
1525         rhead->num_fwd = req->r_num_fwd;
1526         rhead->num_retry = req->r_attempts - 1;
1527
1528         dout(" r_locked_dir = %p\n", req->r_locked_dir);
1529
1530         if (req->r_target_inode && req->r_got_unsafe)
1531                 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1532         else
1533                 rhead->ino = 0;
1534         return 0;
1535 }
1536
1537 /*
1538  * send request, or put it on the appropriate wait list.
1539  */
1540 static int __do_request(struct ceph_mds_client *mdsc,
1541                         struct ceph_mds_request *req)
1542 {
1543         struct ceph_mds_session *session = NULL;
1544         int mds = -1;
1545         int err = -EAGAIN;
1546
1547         if (req->r_err || req->r_got_result)
1548                 goto out;
1549
1550         if (req->r_timeout &&
1551             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1552                 dout("do_request timed out\n");
1553                 err = -EIO;
1554                 goto finish;
1555         }
1556
1557         mds = __choose_mds(mdsc, req);
1558         if (mds < 0 ||
1559             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1560                 dout("do_request no mds or not active, waiting for map\n");
1561                 list_add(&req->r_wait, &mdsc->waiting_for_map);
1562                 goto out;
1563         }
1564
1565         /* get, open session */
1566         session = __ceph_lookup_mds_session(mdsc, mds);
1567         if (!session) {
1568                 session = register_session(mdsc, mds);
1569                 if (IS_ERR(session)) {
1570                         err = PTR_ERR(session);
1571                         goto finish;
1572                 }
1573         }
1574         dout("do_request mds%d session %p state %s\n", mds, session,
1575              session_state_name(session->s_state));
1576         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1577             session->s_state != CEPH_MDS_SESSION_HUNG) {
1578                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1579                     session->s_state == CEPH_MDS_SESSION_CLOSING)
1580                         __open_session(mdsc, session);
1581                 list_add(&req->r_wait, &session->s_waiting);
1582                 goto out_session;
1583         }
1584
1585         /* send request */
1586         req->r_session = get_session(session);
1587         req->r_resend_mds = -1;   /* forget any previous mds hint */
1588
1589         if (req->r_request_started == 0)   /* note request start time */
1590                 req->r_request_started = jiffies;
1591
1592         err = __prepare_send_request(mdsc, req, mds);
1593         if (!err) {
1594                 ceph_msg_get(req->r_request);
1595                 ceph_con_send(&session->s_con, req->r_request);
1596         }
1597
1598 out_session:
1599         ceph_put_mds_session(session);
1600 out:
1601         return err;
1602
1603 finish:
1604         req->r_err = err;
1605         complete_request(mdsc, req);
1606         goto out;
1607 }
1608
1609 /*
1610  * called under mdsc->mutex
1611  */
1612 static void __wake_requests(struct ceph_mds_client *mdsc,
1613                             struct list_head *head)
1614 {
1615         struct ceph_mds_request *req, *nreq;
1616
1617         list_for_each_entry_safe(req, nreq, head, r_wait) {
1618                 list_del_init(&req->r_wait);
1619                 __do_request(mdsc, req);
1620         }
1621 }
1622
1623 /*
1624  * Wake up threads with requests pending for @mds, so that they can
1625  * resubmit their requests to a possibly different mds.  If @all is set,
1626  * wake up if their requests has been forwarded to @mds, too.
1627  */
1628 static void kick_requests(struct ceph_mds_client *mdsc, int mds, int all)
1629 {
1630         struct ceph_mds_request *req;
1631         struct rb_node *p;
1632
1633         dout("kick_requests mds%d\n", mds);
1634         for (p = rb_first(&mdsc->request_tree); p; p = rb_next(p)) {
1635                 req = rb_entry(p, struct ceph_mds_request, r_node);
1636                 if (req->r_got_unsafe)
1637                         continue;
1638                 if (req->r_session &&
1639                     req->r_session->s_mds == mds) {
1640                         dout(" kicking tid %llu\n", req->r_tid);
1641                         put_request_session(req);
1642                         __do_request(mdsc, req);
1643                 }
1644         }
1645 }
1646
1647 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
1648                               struct ceph_mds_request *req)
1649 {
1650         dout("submit_request on %p\n", req);
1651         mutex_lock(&mdsc->mutex);
1652         __register_request(mdsc, req, NULL);
1653         __do_request(mdsc, req);
1654         mutex_unlock(&mdsc->mutex);
1655 }
1656
1657 /*
1658  * Synchrously perform an mds request.  Take care of all of the
1659  * session setup, forwarding, retry details.
1660  */
1661 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
1662                          struct inode *dir,
1663                          struct ceph_mds_request *req)
1664 {
1665         int err;
1666
1667         dout("do_request on %p\n", req);
1668
1669         /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
1670         if (req->r_inode)
1671                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
1672         if (req->r_locked_dir)
1673                 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
1674         if (req->r_old_dentry)
1675                 ceph_get_cap_refs(
1676                         ceph_inode(req->r_old_dentry->d_parent->d_inode),
1677                         CEPH_CAP_PIN);
1678
1679         /* issue */
1680         mutex_lock(&mdsc->mutex);
1681         __register_request(mdsc, req, dir);
1682         __do_request(mdsc, req);
1683
1684         if (req->r_err) {
1685                 err = req->r_err;
1686                 __unregister_request(mdsc, req);
1687                 dout("do_request early error %d\n", err);
1688                 goto out;
1689         }
1690
1691         /* wait */
1692         mutex_unlock(&mdsc->mutex);
1693         dout("do_request waiting\n");
1694         if (req->r_timeout) {
1695                 err = (long)wait_for_completion_interruptible_timeout(
1696                         &req->r_completion, req->r_timeout);
1697                 if (err == 0)
1698                         err = -EIO;
1699         } else {
1700                 err = wait_for_completion_interruptible(&req->r_completion);
1701         }
1702         dout("do_request waited, got %d\n", err);
1703         mutex_lock(&mdsc->mutex);
1704
1705         /* only abort if we didn't race with a real reply */
1706         if (req->r_got_result) {
1707                 err = le32_to_cpu(req->r_reply_info.head->result);
1708         } else if (err < 0) {
1709                 dout("aborted request %lld with %d\n", req->r_tid, err);
1710
1711                 /*
1712                  * ensure we aren't running concurrently with
1713                  * ceph_fill_trace or ceph_readdir_prepopulate, which
1714                  * rely on locks (dir mutex) held by our caller.
1715                  */
1716                 mutex_lock(&req->r_fill_mutex);
1717                 req->r_err = err;
1718                 req->r_aborted = true;
1719                 mutex_unlock(&req->r_fill_mutex);
1720
1721                 if (req->r_locked_dir &&
1722                     (req->r_op & CEPH_MDS_OP_WRITE)) {
1723                         struct ceph_inode_info *ci =
1724                                 ceph_inode(req->r_locked_dir);
1725
1726                         dout("aborted, clearing I_COMPLETE on %p, leases\n",
1727                              req->r_locked_dir);
1728                         spin_lock(&req->r_locked_dir->i_lock);
1729                         ci->i_ceph_flags &= ~CEPH_I_COMPLETE;
1730                         ci->i_release_count++;
1731                         spin_unlock(&req->r_locked_dir->i_lock);
1732
1733                         if (req->r_dentry)
1734                                 ceph_invalidate_dentry_lease(req->r_dentry);
1735                         if (req->r_old_dentry)
1736                                 ceph_invalidate_dentry_lease(req->r_old_dentry);
1737                 }
1738         } else {
1739                 err = req->r_err;
1740         }
1741
1742 out:
1743         mutex_unlock(&mdsc->mutex);
1744         dout("do_request %p done, result %d\n", req, err);
1745         return err;
1746 }
1747
1748 /*
1749  * Handle mds reply.
1750  *
1751  * We take the session mutex and parse and process the reply immediately.
1752  * This preserves the logical ordering of replies, capabilities, etc., sent
1753  * by the MDS as they are applied to our local cache.
1754  */
1755 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
1756 {
1757         struct ceph_mds_client *mdsc = session->s_mdsc;
1758         struct ceph_mds_request *req;
1759         struct ceph_mds_reply_head *head = msg->front.iov_base;
1760         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
1761         u64 tid;
1762         int err, result;
1763         int mds = session->s_mds;
1764
1765         if (msg->front.iov_len < sizeof(*head)) {
1766                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
1767                 ceph_msg_dump(msg);
1768                 return;
1769         }
1770
1771         /* get request, session */
1772         tid = le64_to_cpu(msg->hdr.tid);
1773         mutex_lock(&mdsc->mutex);
1774         req = __lookup_request(mdsc, tid);
1775         if (!req) {
1776                 dout("handle_reply on unknown tid %llu\n", tid);
1777                 mutex_unlock(&mdsc->mutex);
1778                 return;
1779         }
1780         dout("handle_reply %p\n", req);
1781
1782         /* correct session? */
1783         if (req->r_session != session) {
1784                 pr_err("mdsc_handle_reply got %llu on session mds%d"
1785                        " not mds%d\n", tid, session->s_mds,
1786                        req->r_session ? req->r_session->s_mds : -1);
1787                 mutex_unlock(&mdsc->mutex);
1788                 goto out;
1789         }
1790
1791         /* dup? */
1792         if ((req->r_got_unsafe && !head->safe) ||
1793             (req->r_got_safe && head->safe)) {
1794                 pr_warning("got a dup %s reply on %llu from mds%d\n",
1795                            head->safe ? "safe" : "unsafe", tid, mds);
1796                 mutex_unlock(&mdsc->mutex);
1797                 goto out;
1798         }
1799
1800         result = le32_to_cpu(head->result);
1801
1802         /*
1803          * Tolerate 2 consecutive ESTALEs from the same mds.
1804          * FIXME: we should be looking at the cap migrate_seq.
1805          */
1806         if (result == -ESTALE) {
1807                 req->r_direct_mode = USE_AUTH_MDS;
1808                 req->r_num_stale++;
1809                 if (req->r_num_stale <= 2) {
1810                         __do_request(mdsc, req);
1811                         mutex_unlock(&mdsc->mutex);
1812                         goto out;
1813                 }
1814         } else {
1815                 req->r_num_stale = 0;
1816         }
1817
1818         if (head->safe) {
1819                 req->r_got_safe = true;
1820                 __unregister_request(mdsc, req);
1821                 complete(&req->r_safe_completion);
1822
1823                 if (req->r_got_unsafe) {
1824                         /*
1825                          * We already handled the unsafe response, now do the
1826                          * cleanup.  No need to examine the response; the MDS
1827                          * doesn't include any result info in the safe
1828                          * response.  And even if it did, there is nothing
1829                          * useful we could do with a revised return value.
1830                          */
1831                         dout("got safe reply %llu, mds%d\n", tid, mds);
1832                         list_del_init(&req->r_unsafe_item);
1833
1834                         /* last unsafe request during umount? */
1835                         if (mdsc->stopping && !__get_oldest_req(mdsc))
1836                                 complete(&mdsc->safe_umount_waiters);
1837                         mutex_unlock(&mdsc->mutex);
1838                         goto out;
1839                 }
1840         } else {
1841                 req->r_got_unsafe = true;
1842                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
1843         }
1844
1845         dout("handle_reply tid %lld result %d\n", tid, result);
1846         rinfo = &req->r_reply_info;
1847         err = parse_reply_info(msg, rinfo);
1848         mutex_unlock(&mdsc->mutex);
1849
1850         mutex_lock(&session->s_mutex);
1851         if (err < 0) {
1852                 pr_err("mdsc_handle_reply got corrupt reply mds%d\n", mds);
1853                 ceph_msg_dump(msg);
1854                 goto out_err;
1855         }
1856
1857         /* snap trace */
1858         if (rinfo->snapblob_len) {
1859                 down_write(&mdsc->snap_rwsem);
1860                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
1861                                rinfo->snapblob + rinfo->snapblob_len,
1862                                le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
1863                 downgrade_write(&mdsc->snap_rwsem);
1864         } else {
1865                 down_read(&mdsc->snap_rwsem);
1866         }
1867
1868         /* insert trace into our cache */
1869         mutex_lock(&req->r_fill_mutex);
1870         err = ceph_fill_trace(mdsc->client->sb, req, req->r_session);
1871         if (err == 0) {
1872                 if (result == 0 && rinfo->dir_nr)
1873                         ceph_readdir_prepopulate(req, req->r_session);
1874                 ceph_unreserve_caps(&req->r_caps_reservation);
1875         }
1876         mutex_unlock(&req->r_fill_mutex);
1877
1878         up_read(&mdsc->snap_rwsem);
1879 out_err:
1880         mutex_lock(&mdsc->mutex);
1881         if (!req->r_aborted) {
1882                 if (err) {
1883                         req->r_err = err;
1884                 } else {
1885                         req->r_reply = msg;
1886                         ceph_msg_get(msg);
1887                         req->r_got_result = true;
1888                 }
1889         } else {
1890                 dout("reply arrived after request %lld was aborted\n", tid);
1891         }
1892         mutex_unlock(&mdsc->mutex);
1893
1894         add_cap_releases(mdsc, req->r_session, -1);
1895         mutex_unlock(&session->s_mutex);
1896
1897         /* kick calling process */
1898         complete_request(mdsc, req);
1899 out:
1900         ceph_mdsc_put_request(req);
1901         return;
1902 }
1903
1904
1905
1906 /*
1907  * handle mds notification that our request has been forwarded.
1908  */
1909 static void handle_forward(struct ceph_mds_client *mdsc,
1910                            struct ceph_mds_session *session,
1911                            struct ceph_msg *msg)
1912 {
1913         struct ceph_mds_request *req;
1914         u64 tid = le64_to_cpu(msg->hdr.tid);
1915         u32 next_mds;
1916         u32 fwd_seq;
1917         int err = -EINVAL;
1918         void *p = msg->front.iov_base;
1919         void *end = p + msg->front.iov_len;
1920
1921         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1922         next_mds = ceph_decode_32(&p);
1923         fwd_seq = ceph_decode_32(&p);
1924
1925         mutex_lock(&mdsc->mutex);
1926         req = __lookup_request(mdsc, tid);
1927         if (!req) {
1928                 dout("forward %llu to mds%d - req dne\n", tid, next_mds);
1929                 goto out;  /* dup reply? */
1930         }
1931
1932         if (fwd_seq <= req->r_num_fwd) {
1933                 dout("forward %llu to mds%d - old seq %d <= %d\n",
1934                      tid, next_mds, req->r_num_fwd, fwd_seq);
1935         } else {
1936                 /* resend. forward race not possible; mds would drop */
1937                 dout("forward %llu to mds%d (we resend)\n", tid, next_mds);
1938                 req->r_num_fwd = fwd_seq;
1939                 req->r_resend_mds = next_mds;
1940                 put_request_session(req);
1941                 __do_request(mdsc, req);
1942         }
1943         ceph_mdsc_put_request(req);
1944 out:
1945         mutex_unlock(&mdsc->mutex);
1946         return;
1947
1948 bad:
1949         pr_err("mdsc_handle_forward decode error err=%d\n", err);
1950 }
1951
1952 /*
1953  * handle a mds session control message
1954  */
1955 static void handle_session(struct ceph_mds_session *session,
1956                            struct ceph_msg *msg)
1957 {
1958         struct ceph_mds_client *mdsc = session->s_mdsc;
1959         u32 op;
1960         u64 seq;
1961         int mds = session->s_mds;
1962         struct ceph_mds_session_head *h = msg->front.iov_base;
1963         int wake = 0;
1964
1965         /* decode */
1966         if (msg->front.iov_len != sizeof(*h))
1967                 goto bad;
1968         op = le32_to_cpu(h->op);
1969         seq = le64_to_cpu(h->seq);
1970
1971         mutex_lock(&mdsc->mutex);
1972         if (op == CEPH_SESSION_CLOSE)
1973                 __unregister_session(mdsc, session);
1974         /* FIXME: this ttl calculation is generous */
1975         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
1976         mutex_unlock(&mdsc->mutex);
1977
1978         mutex_lock(&session->s_mutex);
1979
1980         dout("handle_session mds%d %s %p state %s seq %llu\n",
1981              mds, ceph_session_op_name(op), session,
1982              session_state_name(session->s_state), seq);
1983
1984         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
1985                 session->s_state = CEPH_MDS_SESSION_OPEN;
1986                 pr_info("mds%d came back\n", session->s_mds);
1987         }
1988
1989         switch (op) {
1990         case CEPH_SESSION_OPEN:
1991                 session->s_state = CEPH_MDS_SESSION_OPEN;
1992                 renewed_caps(mdsc, session, 0);
1993                 wake = 1;
1994                 if (mdsc->stopping)
1995                         __close_session(mdsc, session);
1996                 break;
1997
1998         case CEPH_SESSION_RENEWCAPS:
1999                 if (session->s_renew_seq == seq)
2000                         renewed_caps(mdsc, session, 1);
2001                 break;
2002
2003         case CEPH_SESSION_CLOSE:
2004                 remove_session_caps(session);
2005                 wake = 1; /* for good measure */
2006                 complete(&mdsc->session_close_waiters);
2007                 kick_requests(mdsc, mds, 0);      /* cur only */
2008                 break;
2009
2010         case CEPH_SESSION_STALE:
2011                 pr_info("mds%d caps went stale, renewing\n",
2012                         session->s_mds);
2013                 spin_lock(&session->s_cap_lock);
2014                 session->s_cap_gen++;
2015                 session->s_cap_ttl = 0;
2016                 spin_unlock(&session->s_cap_lock);
2017                 send_renew_caps(mdsc, session);
2018                 break;
2019
2020         case CEPH_SESSION_RECALL_STATE:
2021                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2022                 break;
2023
2024         default:
2025                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2026                 WARN_ON(1);
2027         }
2028
2029         mutex_unlock(&session->s_mutex);
2030         if (wake) {
2031                 mutex_lock(&mdsc->mutex);
2032                 __wake_requests(mdsc, &session->s_waiting);
2033                 mutex_unlock(&mdsc->mutex);
2034         }
2035         return;
2036
2037 bad:
2038         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2039                (int)msg->front.iov_len);
2040         ceph_msg_dump(msg);
2041         return;
2042 }
2043
2044
2045 /*
2046  * called under session->mutex.
2047  */
2048 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2049                                    struct ceph_mds_session *session)
2050 {
2051         struct ceph_mds_request *req, *nreq;
2052         int err;
2053
2054         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2055
2056         mutex_lock(&mdsc->mutex);
2057         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2058                 err = __prepare_send_request(mdsc, req, session->s_mds);
2059                 if (!err) {
2060                         ceph_msg_get(req->r_request);
2061                         ceph_con_send(&session->s_con, req->r_request);
2062                 }
2063         }
2064         mutex_unlock(&mdsc->mutex);
2065 }
2066
2067 /*
2068  * Encode information about a cap for a reconnect with the MDS.
2069  */
2070 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2071                           void *arg)
2072 {
2073         struct ceph_mds_cap_reconnect rec;
2074         struct ceph_inode_info *ci;
2075         struct ceph_pagelist *pagelist = arg;
2076         char *path;
2077         int pathlen, err;
2078         u64 pathbase;
2079         struct dentry *dentry;
2080
2081         ci = cap->ci;
2082
2083         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2084              inode, ceph_vinop(inode), cap, cap->cap_id,
2085              ceph_cap_string(cap->issued));
2086         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2087         if (err)
2088                 return err;
2089
2090         dentry = d_find_alias(inode);
2091         if (dentry) {
2092                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2093                 if (IS_ERR(path)) {
2094                         err = PTR_ERR(path);
2095                         BUG_ON(err);
2096                 }
2097         } else {
2098                 path = NULL;
2099                 pathlen = 0;
2100         }
2101         err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2102         if (err)
2103                 goto out;
2104
2105         spin_lock(&inode->i_lock);
2106         cap->seq = 0;        /* reset cap seq */
2107         cap->issue_seq = 0;  /* and issue_seq */
2108         rec.cap_id = cpu_to_le64(cap->cap_id);
2109         rec.pathbase = cpu_to_le64(pathbase);
2110         rec.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2111         rec.issued = cpu_to_le32(cap->issued);
2112         rec.size = cpu_to_le64(inode->i_size);
2113         ceph_encode_timespec(&rec.mtime, &inode->i_mtime);
2114         ceph_encode_timespec(&rec.atime, &inode->i_atime);
2115         rec.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2116         spin_unlock(&inode->i_lock);
2117
2118         err = ceph_pagelist_append(pagelist, &rec, sizeof(rec));
2119
2120 out:
2121         kfree(path);
2122         dput(dentry);
2123         return err;
2124 }
2125
2126
2127 /*
2128  * If an MDS fails and recovers, clients need to reconnect in order to
2129  * reestablish shared state.  This includes all caps issued through
2130  * this session _and_ the snap_realm hierarchy.  Because it's not
2131  * clear which snap realms the mds cares about, we send everything we
2132  * know about.. that ensures we'll then get any new info the
2133  * recovering MDS might have.
2134  *
2135  * This is a relatively heavyweight operation, but it's rare.
2136  *
2137  * called with mdsc->mutex held.
2138  */
2139 static void send_mds_reconnect(struct ceph_mds_client *mdsc, int mds)
2140 {
2141         struct ceph_mds_session *session = NULL;
2142         struct ceph_msg *reply;
2143         struct rb_node *p;
2144         int err = -ENOMEM;
2145         struct ceph_pagelist *pagelist;
2146
2147         pr_info("reconnect to recovering mds%d\n", mds);
2148
2149         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2150         if (!pagelist)
2151                 goto fail_nopagelist;
2152         ceph_pagelist_init(pagelist);
2153
2154         err = -ENOMEM;
2155         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0);
2156         if (!reply)
2157                 goto fail_nomsg;
2158
2159         /* find session */
2160         session = __ceph_lookup_mds_session(mdsc, mds);
2161         mutex_unlock(&mdsc->mutex);    /* drop lock for duration */
2162
2163         if (session) {
2164                 mutex_lock(&session->s_mutex);
2165
2166                 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2167                 session->s_seq = 0;
2168
2169                 ceph_con_open(&session->s_con,
2170                               ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2171
2172                 /* replay unsafe requests */
2173                 replay_unsafe_requests(mdsc, session);
2174         } else {
2175                 dout("no session for mds%d, will send short reconnect\n",
2176                      mds);
2177         }
2178
2179         down_read(&mdsc->snap_rwsem);
2180
2181         if (!session)
2182                 goto send;
2183         dout("session %p state %s\n", session,
2184              session_state_name(session->s_state));
2185
2186         /* traverse this session's caps */
2187         err = ceph_pagelist_encode_32(pagelist, session->s_nr_caps);
2188         if (err)
2189                 goto fail;
2190         err = iterate_session_caps(session, encode_caps_cb, pagelist);
2191         if (err < 0)
2192                 goto fail;
2193
2194         /*
2195          * snaprealms.  we provide mds with the ino, seq (version), and
2196          * parent for all of our realms.  If the mds has any newer info,
2197          * it will tell us.
2198          */
2199         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2200                 struct ceph_snap_realm *realm =
2201                         rb_entry(p, struct ceph_snap_realm, node);
2202                 struct ceph_mds_snaprealm_reconnect sr_rec;
2203
2204                 dout(" adding snap realm %llx seq %lld parent %llx\n",
2205                      realm->ino, realm->seq, realm->parent_ino);
2206                 sr_rec.ino = cpu_to_le64(realm->ino);
2207                 sr_rec.seq = cpu_to_le64(realm->seq);
2208                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2209                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2210                 if (err)
2211                         goto fail;
2212         }
2213
2214 send:
2215         reply->pagelist = pagelist;
2216         reply->hdr.data_len = cpu_to_le32(pagelist->length);
2217         reply->nr_pages = calc_pages_for(0, pagelist->length);
2218         ceph_con_send(&session->s_con, reply);
2219
2220         session->s_state = CEPH_MDS_SESSION_OPEN;
2221         mutex_unlock(&session->s_mutex);
2222
2223         mutex_lock(&mdsc->mutex);
2224         __wake_requests(mdsc, &session->s_waiting);
2225         mutex_unlock(&mdsc->mutex);
2226
2227         ceph_put_mds_session(session);
2228
2229         up_read(&mdsc->snap_rwsem);
2230         mutex_lock(&mdsc->mutex);
2231         return;
2232
2233 fail:
2234         ceph_msg_put(reply);
2235         up_read(&mdsc->snap_rwsem);
2236         mutex_unlock(&session->s_mutex);
2237         ceph_put_mds_session(session);
2238 fail_nomsg:
2239         ceph_pagelist_release(pagelist);
2240         kfree(pagelist);
2241 fail_nopagelist:
2242         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
2243         mutex_lock(&mdsc->mutex);
2244         return;
2245 }
2246
2247
2248 /*
2249  * compare old and new mdsmaps, kicking requests
2250  * and closing out old connections as necessary
2251  *
2252  * called under mdsc->mutex.
2253  */
2254 static void check_new_map(struct ceph_mds_client *mdsc,
2255                           struct ceph_mdsmap *newmap,
2256                           struct ceph_mdsmap *oldmap)
2257 {
2258         int i;
2259         int oldstate, newstate;
2260         struct ceph_mds_session *s;
2261
2262         dout("check_new_map new %u old %u\n",
2263              newmap->m_epoch, oldmap->m_epoch);
2264
2265         for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2266                 if (mdsc->sessions[i] == NULL)
2267                         continue;
2268                 s = mdsc->sessions[i];
2269                 oldstate = ceph_mdsmap_get_state(oldmap, i);
2270                 newstate = ceph_mdsmap_get_state(newmap, i);
2271
2272                 dout("check_new_map mds%d state %s -> %s (session %s)\n",
2273                      i, ceph_mds_state_name(oldstate),
2274                      ceph_mds_state_name(newstate),
2275                      session_state_name(s->s_state));
2276
2277                 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
2278                            ceph_mdsmap_get_addr(newmap, i),
2279                            sizeof(struct ceph_entity_addr))) {
2280                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2281                                 /* the session never opened, just close it
2282                                  * out now */
2283                                 __wake_requests(mdsc, &s->s_waiting);
2284                                 __unregister_session(mdsc, s);
2285                         } else {
2286                                 /* just close it */
2287                                 mutex_unlock(&mdsc->mutex);
2288                                 mutex_lock(&s->s_mutex);
2289                                 mutex_lock(&mdsc->mutex);
2290                                 ceph_con_close(&s->s_con);
2291                                 mutex_unlock(&s->s_mutex);
2292                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2293                         }
2294
2295                         /* kick any requests waiting on the recovering mds */
2296                         kick_requests(mdsc, i, 1);
2297                 } else if (oldstate == newstate) {
2298                         continue;  /* nothing new with this mds */
2299                 }
2300
2301                 /*
2302                  * send reconnect?
2303                  */
2304                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2305                     newstate >= CEPH_MDS_STATE_RECONNECT)
2306                         send_mds_reconnect(mdsc, i);
2307
2308                 /*
2309                  * kick requests on any mds that has gone active.
2310                  *
2311                  * kick requests on cur or forwarder: we may have sent
2312                  * the request to mds1, mds1 told us it forwarded it
2313                  * to mds2, but then we learn mds1 failed and can't be
2314                  * sure it successfully forwarded our request before
2315                  * it died.
2316                  */
2317                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2318                     newstate >= CEPH_MDS_STATE_ACTIVE) {
2319                         pr_info("mds%d reconnect completed\n", s->s_mds);
2320                         kick_requests(mdsc, i, 1);
2321                         ceph_kick_flushing_caps(mdsc, s);
2322                         wake_up_session_caps(s, 1);
2323                 }
2324         }
2325 }
2326
2327
2328
2329 /*
2330  * leases
2331  */
2332
2333 /*
2334  * caller must hold session s_mutex, dentry->d_lock
2335  */
2336 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2337 {
2338         struct ceph_dentry_info *di = ceph_dentry(dentry);
2339
2340         ceph_put_mds_session(di->lease_session);
2341         di->lease_session = NULL;
2342 }
2343
2344 static void handle_lease(struct ceph_mds_client *mdsc,
2345                          struct ceph_mds_session *session,
2346                          struct ceph_msg *msg)
2347 {
2348         struct super_block *sb = mdsc->client->sb;
2349         struct inode *inode;
2350         struct ceph_inode_info *ci;
2351         struct dentry *parent, *dentry;
2352         struct ceph_dentry_info *di;
2353         int mds = session->s_mds;
2354         struct ceph_mds_lease *h = msg->front.iov_base;
2355         struct ceph_vino vino;
2356         int mask;
2357         struct qstr dname;
2358         int release = 0;
2359
2360         dout("handle_lease from mds%d\n", mds);
2361
2362         /* decode */
2363         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2364                 goto bad;
2365         vino.ino = le64_to_cpu(h->ino);
2366         vino.snap = CEPH_NOSNAP;
2367         mask = le16_to_cpu(h->mask);
2368         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2369         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2370         if (dname.len != get_unaligned_le32(h+1))
2371                 goto bad;
2372
2373         mutex_lock(&session->s_mutex);
2374         session->s_seq++;
2375
2376         /* lookup inode */
2377         inode = ceph_find_inode(sb, vino);
2378         dout("handle_lease '%s', mask %d, ino %llx %p\n",
2379              ceph_lease_op_name(h->action), mask, vino.ino, inode);
2380         if (inode == NULL) {
2381                 dout("handle_lease no inode %llx\n", vino.ino);
2382                 goto release;
2383         }
2384         ci = ceph_inode(inode);
2385
2386         /* dentry */
2387         parent = d_find_alias(inode);
2388         if (!parent) {
2389                 dout("no parent dentry on inode %p\n", inode);
2390                 WARN_ON(1);
2391                 goto release;  /* hrm... */
2392         }
2393         dname.hash = full_name_hash(dname.name, dname.len);
2394         dentry = d_lookup(parent, &dname);
2395         dput(parent);
2396         if (!dentry)
2397                 goto release;
2398
2399         spin_lock(&dentry->d_lock);
2400         di = ceph_dentry(dentry);
2401         switch (h->action) {
2402         case CEPH_MDS_LEASE_REVOKE:
2403                 if (di && di->lease_session == session) {
2404                         h->seq = cpu_to_le32(di->lease_seq);
2405                         __ceph_mdsc_drop_dentry_lease(dentry);
2406                 }
2407                 release = 1;
2408                 break;
2409
2410         case CEPH_MDS_LEASE_RENEW:
2411                 if (di && di->lease_session == session &&
2412                     di->lease_gen == session->s_cap_gen &&
2413                     di->lease_renew_from &&
2414                     di->lease_renew_after == 0) {
2415                         unsigned long duration =
2416                                 le32_to_cpu(h->duration_ms) * HZ / 1000;
2417
2418                         di->lease_seq = le32_to_cpu(h->seq);
2419                         dentry->d_time = di->lease_renew_from + duration;
2420                         di->lease_renew_after = di->lease_renew_from +
2421                                 (duration >> 1);
2422                         di->lease_renew_from = 0;
2423                 }
2424                 break;
2425         }
2426         spin_unlock(&dentry->d_lock);
2427         dput(dentry);
2428
2429         if (!release)
2430                 goto out;
2431
2432 release:
2433         /* let's just reuse the same message */
2434         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2435         ceph_msg_get(msg);
2436         ceph_con_send(&session->s_con, msg);
2437
2438 out:
2439         iput(inode);
2440         mutex_unlock(&session->s_mutex);
2441         return;
2442
2443 bad:
2444         pr_err("corrupt lease message\n");
2445         ceph_msg_dump(msg);
2446 }
2447
2448 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
2449                               struct inode *inode,
2450                               struct dentry *dentry, char action,
2451                               u32 seq)
2452 {
2453         struct ceph_msg *msg;
2454         struct ceph_mds_lease *lease;
2455         int len = sizeof(*lease) + sizeof(u32);
2456         int dnamelen = 0;
2457
2458         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2459              inode, dentry, ceph_lease_op_name(action), session->s_mds);
2460         dnamelen = dentry->d_name.len;
2461         len += dnamelen;
2462
2463         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len);
2464         if (!msg)
2465                 return;
2466         lease = msg->front.iov_base;
2467         lease->action = action;
2468         lease->mask = cpu_to_le16(CEPH_LOCK_DN);
2469         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
2470         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
2471         lease->seq = cpu_to_le32(seq);
2472         put_unaligned_le32(dnamelen, lease + 1);
2473         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
2474
2475         /*
2476          * if this is a preemptive lease RELEASE, no need to
2477          * flush request stream, since the actual request will
2478          * soon follow.
2479          */
2480         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
2481
2482         ceph_con_send(&session->s_con, msg);
2483 }
2484
2485 /*
2486  * Preemptively release a lease we expect to invalidate anyway.
2487  * Pass @inode always, @dentry is optional.
2488  */
2489 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2490                              struct dentry *dentry, int mask)
2491 {
2492         struct ceph_dentry_info *di;
2493         struct ceph_mds_session *session;
2494         u32 seq;
2495
2496         BUG_ON(inode == NULL);
2497         BUG_ON(dentry == NULL);
2498         BUG_ON(mask != CEPH_LOCK_DN);
2499
2500         /* is dentry lease valid? */
2501         spin_lock(&dentry->d_lock);
2502         di = ceph_dentry(dentry);
2503         if (!di || !di->lease_session ||
2504             di->lease_session->s_mds < 0 ||
2505             di->lease_gen != di->lease_session->s_cap_gen ||
2506             !time_before(jiffies, dentry->d_time)) {
2507                 dout("lease_release inode %p dentry %p -- "
2508                      "no lease on %d\n",
2509                      inode, dentry, mask);
2510                 spin_unlock(&dentry->d_lock);
2511                 return;
2512         }
2513
2514         /* we do have a lease on this dentry; note mds and seq */
2515         session = ceph_get_mds_session(di->lease_session);
2516         seq = di->lease_seq;
2517         __ceph_mdsc_drop_dentry_lease(dentry);
2518         spin_unlock(&dentry->d_lock);
2519
2520         dout("lease_release inode %p dentry %p mask %d to mds%d\n",
2521              inode, dentry, mask, session->s_mds);
2522         ceph_mdsc_lease_send_msg(session, inode, dentry,
2523                                  CEPH_MDS_LEASE_RELEASE, seq);
2524         ceph_put_mds_session(session);
2525 }
2526
2527 /*
2528  * drop all leases (and dentry refs) in preparation for umount
2529  */
2530 static void drop_leases(struct ceph_mds_client *mdsc)
2531 {
2532         int i;
2533
2534         dout("drop_leases\n");
2535         mutex_lock(&mdsc->mutex);
2536         for (i = 0; i < mdsc->max_sessions; i++) {
2537                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2538                 if (!s)
2539                         continue;
2540                 mutex_unlock(&mdsc->mutex);
2541                 mutex_lock(&s->s_mutex);
2542                 mutex_unlock(&s->s_mutex);
2543                 ceph_put_mds_session(s);
2544                 mutex_lock(&mdsc->mutex);
2545         }
2546         mutex_unlock(&mdsc->mutex);
2547 }
2548
2549
2550
2551 /*
2552  * delayed work -- periodically trim expired leases, renew caps with mds
2553  */
2554 static void schedule_delayed(struct ceph_mds_client *mdsc)
2555 {
2556         int delay = 5;
2557         unsigned hz = round_jiffies_relative(HZ * delay);
2558         schedule_delayed_work(&mdsc->delayed_work, hz);
2559 }
2560
2561 static void delayed_work(struct work_struct *work)
2562 {
2563         int i;
2564         struct ceph_mds_client *mdsc =
2565                 container_of(work, struct ceph_mds_client, delayed_work.work);
2566         int renew_interval;
2567         int renew_caps;
2568
2569         dout("mdsc delayed_work\n");
2570         ceph_check_delayed_caps(mdsc);
2571
2572         mutex_lock(&mdsc->mutex);
2573         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
2574         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
2575                                    mdsc->last_renew_caps);
2576         if (renew_caps)
2577                 mdsc->last_renew_caps = jiffies;
2578
2579         for (i = 0; i < mdsc->max_sessions; i++) {
2580                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2581                 if (s == NULL)
2582                         continue;
2583                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
2584                         dout("resending session close request for mds%d\n",
2585                              s->s_mds);
2586                         request_close_session(mdsc, s);
2587                         ceph_put_mds_session(s);
2588                         continue;
2589                 }
2590                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
2591                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
2592                                 s->s_state = CEPH_MDS_SESSION_HUNG;
2593                                 pr_info("mds%d hung\n", s->s_mds);
2594                         }
2595                 }
2596                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
2597                         /* this mds is failed or recovering, just wait */
2598                         ceph_put_mds_session(s);
2599                         continue;
2600                 }
2601                 mutex_unlock(&mdsc->mutex);
2602
2603                 mutex_lock(&s->s_mutex);
2604                 if (renew_caps)
2605                         send_renew_caps(mdsc, s);
2606                 else
2607                         ceph_con_keepalive(&s->s_con);
2608                 add_cap_releases(mdsc, s, -1);
2609                 send_cap_releases(mdsc, s);
2610                 mutex_unlock(&s->s_mutex);
2611                 ceph_put_mds_session(s);
2612
2613                 mutex_lock(&mdsc->mutex);
2614         }
2615         mutex_unlock(&mdsc->mutex);
2616
2617         schedule_delayed(mdsc);
2618 }
2619
2620
2621 int ceph_mdsc_init(struct ceph_mds_client *mdsc, struct ceph_client *client)
2622 {
2623         mdsc->client = client;
2624         mutex_init(&mdsc->mutex);
2625         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
2626         if (mdsc->mdsmap == NULL)
2627                 return -ENOMEM;
2628
2629         init_completion(&mdsc->safe_umount_waiters);
2630         init_completion(&mdsc->session_close_waiters);
2631         INIT_LIST_HEAD(&mdsc->waiting_for_map);
2632         mdsc->sessions = NULL;
2633         mdsc->max_sessions = 0;
2634         mdsc->stopping = 0;
2635         init_rwsem(&mdsc->snap_rwsem);
2636         mdsc->snap_realms = RB_ROOT;
2637         INIT_LIST_HEAD(&mdsc->snap_empty);
2638         spin_lock_init(&mdsc->snap_empty_lock);
2639         mdsc->last_tid = 0;
2640         mdsc->request_tree = RB_ROOT;
2641         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
2642         mdsc->last_renew_caps = jiffies;
2643         INIT_LIST_HEAD(&mdsc->cap_delay_list);
2644         spin_lock_init(&mdsc->cap_delay_lock);
2645         INIT_LIST_HEAD(&mdsc->snap_flush_list);
2646         spin_lock_init(&mdsc->snap_flush_lock);
2647         mdsc->cap_flush_seq = 0;
2648         INIT_LIST_HEAD(&mdsc->cap_dirty);
2649         mdsc->num_cap_flushing = 0;
2650         spin_lock_init(&mdsc->cap_dirty_lock);
2651         init_waitqueue_head(&mdsc->cap_flushing_wq);
2652         spin_lock_init(&mdsc->dentry_lru_lock);
2653         INIT_LIST_HEAD(&mdsc->dentry_lru);
2654
2655         return 0;
2656 }
2657
2658 /*
2659  * Wait for safe replies on open mds requests.  If we time out, drop
2660  * all requests from the tree to avoid dangling dentry refs.
2661  */
2662 static void wait_requests(struct ceph_mds_client *mdsc)
2663 {
2664         struct ceph_mds_request *req;
2665         struct ceph_client *client = mdsc->client;
2666
2667         mutex_lock(&mdsc->mutex);
2668         if (__get_oldest_req(mdsc)) {
2669                 mutex_unlock(&mdsc->mutex);
2670
2671                 dout("wait_requests waiting for requests\n");
2672                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
2673                                     client->mount_args->mount_timeout * HZ);
2674
2675                 /* tear down remaining requests */
2676                 mutex_lock(&mdsc->mutex);
2677                 while ((req = __get_oldest_req(mdsc))) {
2678                         dout("wait_requests timed out on tid %llu\n",
2679                              req->r_tid);
2680                         __unregister_request(mdsc, req);
2681                 }
2682         }
2683         mutex_unlock(&mdsc->mutex);
2684         dout("wait_requests done\n");
2685 }
2686
2687 /*
2688  * called before mount is ro, and before dentries are torn down.
2689  * (hmm, does this still race with new lookups?)
2690  */
2691 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
2692 {
2693         dout("pre_umount\n");
2694         mdsc->stopping = 1;
2695
2696         drop_leases(mdsc);
2697         ceph_flush_dirty_caps(mdsc);
2698         wait_requests(mdsc);
2699 }
2700
2701 /*
2702  * wait for all write mds requests to flush.
2703  */
2704 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
2705 {
2706         struct ceph_mds_request *req = NULL, *nextreq;
2707         struct rb_node *n;
2708
2709         mutex_lock(&mdsc->mutex);
2710         dout("wait_unsafe_requests want %lld\n", want_tid);
2711 restart:
2712         req = __get_oldest_req(mdsc);
2713         while (req && req->r_tid <= want_tid) {
2714                 /* find next request */
2715                 n = rb_next(&req->r_node);
2716                 if (n)
2717                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
2718                 else
2719                         nextreq = NULL;
2720                 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
2721                         /* write op */
2722                         ceph_mdsc_get_request(req);
2723                         if (nextreq)
2724                                 ceph_mdsc_get_request(nextreq);
2725                         mutex_unlock(&mdsc->mutex);
2726                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
2727                              req->r_tid, want_tid);
2728                         wait_for_completion(&req->r_safe_completion);
2729                         mutex_lock(&mdsc->mutex);
2730                         ceph_mdsc_put_request(req);
2731                         if (!nextreq)
2732                                 break;  /* next dne before, so we're done! */
2733                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
2734                                 /* next request was removed from tree */
2735                                 ceph_mdsc_put_request(nextreq);
2736                                 goto restart;
2737                         }
2738                         ceph_mdsc_put_request(nextreq);  /* won't go away */
2739                 }
2740                 req = nextreq;
2741         }
2742         mutex_unlock(&mdsc->mutex);
2743         dout("wait_unsafe_requests done\n");
2744 }
2745
2746 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
2747 {
2748         u64 want_tid, want_flush;
2749
2750         if (mdsc->client->mount_state == CEPH_MOUNT_SHUTDOWN)
2751                 return;
2752
2753         dout("sync\n");
2754         mutex_lock(&mdsc->mutex);
2755         want_tid = mdsc->last_tid;
2756         want_flush = mdsc->cap_flush_seq;
2757         mutex_unlock(&mdsc->mutex);
2758         dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
2759
2760         ceph_flush_dirty_caps(mdsc);
2761
2762         wait_unsafe_requests(mdsc, want_tid);
2763         wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
2764 }
2765
2766
2767 /*
2768  * called after sb is ro.
2769  */
2770 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
2771 {
2772         struct ceph_mds_session *session;
2773         int i;
2774         int n;
2775         struct ceph_client *client = mdsc->client;
2776         unsigned long started, timeout = client->mount_args->mount_timeout * HZ;
2777
2778         dout("close_sessions\n");
2779
2780         mutex_lock(&mdsc->mutex);
2781
2782         /* close sessions */
2783         started = jiffies;
2784         while (time_before(jiffies, started + timeout)) {
2785                 dout("closing sessions\n");
2786                 n = 0;
2787                 for (i = 0; i < mdsc->max_sessions; i++) {
2788                         session = __ceph_lookup_mds_session(mdsc, i);
2789                         if (!session)
2790                                 continue;
2791                         mutex_unlock(&mdsc->mutex);
2792                         mutex_lock(&session->s_mutex);
2793                         __close_session(mdsc, session);
2794                         mutex_unlock(&session->s_mutex);
2795                         ceph_put_mds_session(session);
2796                         mutex_lock(&mdsc->mutex);
2797                         n++;
2798                 }
2799                 if (n == 0)
2800                         break;
2801
2802                 if (client->mount_state == CEPH_MOUNT_SHUTDOWN)
2803                         break;
2804
2805                 dout("waiting for sessions to close\n");
2806                 mutex_unlock(&mdsc->mutex);
2807                 wait_for_completion_timeout(&mdsc->session_close_waiters,
2808                                             timeout);
2809                 mutex_lock(&mdsc->mutex);
2810         }
2811
2812         /* tear down remaining sessions */
2813         for (i = 0; i < mdsc->max_sessions; i++) {
2814                 if (mdsc->sessions[i]) {
2815                         session = get_session(mdsc->sessions[i]);
2816                         __unregister_session(mdsc, session);
2817                         mutex_unlock(&mdsc->mutex);
2818                         mutex_lock(&session->s_mutex);
2819                         remove_session_caps(session);
2820                         mutex_unlock(&session->s_mutex);
2821                         ceph_put_mds_session(session);
2822                         mutex_lock(&mdsc->mutex);
2823                 }
2824         }
2825
2826         WARN_ON(!list_empty(&mdsc->cap_delay_list));
2827
2828         mutex_unlock(&mdsc->mutex);
2829
2830         ceph_cleanup_empty_realms(mdsc);
2831
2832         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2833
2834         dout("stopped\n");
2835 }
2836
2837 void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
2838 {
2839         dout("stop\n");
2840         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2841         if (mdsc->mdsmap)
2842                 ceph_mdsmap_destroy(mdsc->mdsmap);
2843         kfree(mdsc->sessions);
2844 }
2845
2846
2847 /*
2848  * handle mds map update.
2849  */
2850 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
2851 {
2852         u32 epoch;
2853         u32 maplen;
2854         void *p = msg->front.iov_base;
2855         void *end = p + msg->front.iov_len;
2856         struct ceph_mdsmap *newmap, *oldmap;
2857         struct ceph_fsid fsid;
2858         int err = -EINVAL;
2859
2860         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
2861         ceph_decode_copy(&p, &fsid, sizeof(fsid));
2862         if (ceph_check_fsid(mdsc->client, &fsid) < 0)
2863                 return;
2864         epoch = ceph_decode_32(&p);
2865         maplen = ceph_decode_32(&p);
2866         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
2867
2868         /* do we need it? */
2869         ceph_monc_got_mdsmap(&mdsc->client->monc, epoch);
2870         mutex_lock(&mdsc->mutex);
2871         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
2872                 dout("handle_map epoch %u <= our %u\n",
2873                      epoch, mdsc->mdsmap->m_epoch);
2874                 mutex_unlock(&mdsc->mutex);
2875                 return;
2876         }
2877
2878         newmap = ceph_mdsmap_decode(&p, end);
2879         if (IS_ERR(newmap)) {
2880                 err = PTR_ERR(newmap);
2881                 goto bad_unlock;
2882         }
2883
2884         /* swap into place */
2885         if (mdsc->mdsmap) {
2886                 oldmap = mdsc->mdsmap;
2887                 mdsc->mdsmap = newmap;
2888                 check_new_map(mdsc, newmap, oldmap);
2889                 ceph_mdsmap_destroy(oldmap);
2890         } else {
2891                 mdsc->mdsmap = newmap;  /* first mds map */
2892         }
2893         mdsc->client->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
2894
2895         __wake_requests(mdsc, &mdsc->waiting_for_map);
2896
2897         mutex_unlock(&mdsc->mutex);
2898         schedule_delayed(mdsc);
2899         return;
2900
2901 bad_unlock:
2902         mutex_unlock(&mdsc->mutex);
2903 bad:
2904         pr_err("error decoding mdsmap %d\n", err);
2905         return;
2906 }
2907
2908 static struct ceph_connection *con_get(struct ceph_connection *con)
2909 {
2910         struct ceph_mds_session *s = con->private;
2911
2912         if (get_session(s)) {
2913                 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
2914                 return con;
2915         }
2916         dout("mdsc con_get %p FAIL\n", s);
2917         return NULL;
2918 }
2919
2920 static void con_put(struct ceph_connection *con)
2921 {
2922         struct ceph_mds_session *s = con->private;
2923
2924         ceph_put_mds_session(s);
2925         dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref));
2926 }
2927
2928 /*
2929  * if the client is unresponsive for long enough, the mds will kill
2930  * the session entirely.
2931  */
2932 static void peer_reset(struct ceph_connection *con)
2933 {
2934         struct ceph_mds_session *s = con->private;
2935
2936         pr_err("mds%d gave us the boot.  IMPLEMENT RECONNECT.\n",
2937                s->s_mds);
2938 }
2939
2940 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2941 {
2942         struct ceph_mds_session *s = con->private;
2943         struct ceph_mds_client *mdsc = s->s_mdsc;
2944         int type = le16_to_cpu(msg->hdr.type);
2945
2946         mutex_lock(&mdsc->mutex);
2947         if (__verify_registered_session(mdsc, s) < 0) {
2948                 mutex_unlock(&mdsc->mutex);
2949                 goto out;
2950         }
2951         mutex_unlock(&mdsc->mutex);
2952
2953         switch (type) {
2954         case CEPH_MSG_MDS_MAP:
2955                 ceph_mdsc_handle_map(mdsc, msg);
2956                 break;
2957         case CEPH_MSG_CLIENT_SESSION:
2958                 handle_session(s, msg);
2959                 break;
2960         case CEPH_MSG_CLIENT_REPLY:
2961                 handle_reply(s, msg);
2962                 break;
2963         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
2964                 handle_forward(mdsc, s, msg);
2965                 break;
2966         case CEPH_MSG_CLIENT_CAPS:
2967                 ceph_handle_caps(s, msg);
2968                 break;
2969         case CEPH_MSG_CLIENT_SNAP:
2970                 ceph_handle_snap(mdsc, s, msg);
2971                 break;
2972         case CEPH_MSG_CLIENT_LEASE:
2973                 handle_lease(mdsc, s, msg);
2974                 break;
2975
2976         default:
2977                 pr_err("received unknown message type %d %s\n", type,
2978                        ceph_msg_type_name(type));
2979         }
2980 out:
2981         ceph_msg_put(msg);
2982 }
2983
2984 /*
2985  * authentication
2986  */
2987 static int get_authorizer(struct ceph_connection *con,
2988                           void **buf, int *len, int *proto,
2989                           void **reply_buf, int *reply_len, int force_new)
2990 {
2991         struct ceph_mds_session *s = con->private;
2992         struct ceph_mds_client *mdsc = s->s_mdsc;
2993         struct ceph_auth_client *ac = mdsc->client->monc.auth;
2994         int ret = 0;
2995
2996         if (force_new && s->s_authorizer) {
2997                 ac->ops->destroy_authorizer(ac, s->s_authorizer);
2998                 s->s_authorizer = NULL;
2999         }
3000         if (s->s_authorizer == NULL) {
3001                 if (ac->ops->create_authorizer) {
3002                         ret = ac->ops->create_authorizer(
3003                                 ac, CEPH_ENTITY_TYPE_MDS,
3004                                 &s->s_authorizer,
3005                                 &s->s_authorizer_buf,
3006                                 &s->s_authorizer_buf_len,
3007                                 &s->s_authorizer_reply_buf,
3008                                 &s->s_authorizer_reply_buf_len);
3009                         if (ret)
3010                                 return ret;
3011                 }
3012         }
3013
3014         *proto = ac->protocol;
3015         *buf = s->s_authorizer_buf;
3016         *len = s->s_authorizer_buf_len;
3017         *reply_buf = s->s_authorizer_reply_buf;
3018         *reply_len = s->s_authorizer_reply_buf_len;
3019         return 0;
3020 }
3021
3022
3023 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3024 {
3025         struct ceph_mds_session *s = con->private;
3026         struct ceph_mds_client *mdsc = s->s_mdsc;
3027         struct ceph_auth_client *ac = mdsc->client->monc.auth;
3028
3029         return ac->ops->verify_authorizer_reply(ac, s->s_authorizer, len);
3030 }
3031
3032 static int invalidate_authorizer(struct ceph_connection *con)
3033 {
3034         struct ceph_mds_session *s = con->private;
3035         struct ceph_mds_client *mdsc = s->s_mdsc;
3036         struct ceph_auth_client *ac = mdsc->client->monc.auth;
3037
3038         if (ac->ops->invalidate_authorizer)
3039                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3040
3041         return ceph_monc_validate_auth(&mdsc->client->monc);
3042 }
3043
3044 const static struct ceph_connection_operations mds_con_ops = {
3045         .get = con_get,
3046         .put = con_put,
3047         .dispatch = dispatch,
3048         .get_authorizer = get_authorizer,
3049         .verify_authorizer_reply = verify_authorizer_reply,
3050         .invalidate_authorizer = invalidate_authorizer,
3051         .peer_reset = peer_reset,
3052 };
3053
3054
3055
3056
3057 /* eof */