fs/ceph/mds_client.c

   1 #include "ceph_debug.h"
   2
   3 #include <linux/wait.h>
   4 #include <linux/slab.h>
   5 #include <linux/sched.h>
   6
   7 #include "mds_client.h"
   8 #include "mon_client.h"
   9 #include "super.h"
  10 #include "messenger.h"
  11 #include "decode.h"
  12 #include "auth.h"
  13 #include "pagelist.h"
  14
  15 /*
  16  * A cluster of MDS (metadata server) daemons is responsible for
  17  * managing the file system namespace (the directory hierarchy and
  18  * inodes) and for coordinating shared access to storage.  Metadata is
  19  * partitioning hierarchically across a number of servers, and that
  20  * partition varies over time as the cluster adjusts the distribution
  21  * in order to balance load.
  22  *
  23  * The MDS client is primarily responsible to managing synchronous
  24  * metadata requests for operations like open, unlink, and so forth.
  25  * If there is a MDS failure, we find out about it when we (possibly
  26  * request and) receive a new MDS map, and can resubmit affected
  27  * requests.
  28  *
  29  * For the most part, though, we take advantage of a lossless
  30  * communications channel to the MDS, and do not need to worry about
  31  * timing out or resubmitting requests.
  32  *
  33  * We maintain a stateful "session" with each MDS we interact with.
  34  * Within each session, we sent periodic heartbeat messages to ensure
  35  * any capabilities or leases we have been issues remain valid.  If
  36  * the session times out and goes stale, our leases and capabilities
  37  * are no longer valid.
  38  */
  39
  40 static void __wake_requests(struct ceph_mds_client *mdsc,
  41                             struct list_head *head);
  42
  43 const static struct ceph_connection_operations mds_con_ops;
  44
  45
  46 /*
  47  * mds reply parsing
  48  */
  49
  50 /*
  51  * parse individual inode info
  52  */
  53 static int parse_reply_info_in(void **p, void *end,
  54                                struct ceph_mds_reply_info_in *info)
  55 {
  56         int err = -EIO;
  57
  58         info->in = *p;
  59         *p += sizeof(struct ceph_mds_reply_inode) +
  60                 sizeof(*info->in->fragtree.splits) *
  61                 le32_to_cpu(info->in->fragtree.nsplits);
  62
  63         ceph_decode_32_safe(p, end, info->symlink_len, bad);
  64         ceph_decode_need(p, end, info->symlink_len, bad);
  65         info->symlink = *p;
  66         *p += info->symlink_len;
  67
  68         ceph_decode_32_safe(p, end, info->xattr_len, bad);
  69         ceph_decode_need(p, end, info->xattr_len, bad);
  70         info->xattr_data = *p;
  71         *p += info->xattr_len;
  72         return 0;
  73 bad:
  74         return err;
  75 }
  76
  77 /*
  78  * parse a normal reply, which may contain a (dir+)dentry and/or a
  79  * target inode.
  80  */
  81 static int parse_reply_info_trace(void **p, void *end,
  82                                   struct ceph_mds_reply_info_parsed *info)
  83 {
  84         int err;
  85
  86         if (info->head->is_dentry) {
  87                 err = parse_reply_info_in(p, end, &info->diri);
  88                 if (err < 0)
  89                         goto out_bad;
  90
  91                 if (unlikely(*p + sizeof(*info->dirfrag) > end))
  92                         goto bad;
  93                 info->dirfrag = *p;
  94                 *p += sizeof(*info->dirfrag) +
  95                         sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
  96                 if (unlikely(*p > end))
  97                         goto bad;
  98
  99                 ceph_decode_32_safe(p, end, info->dname_len, bad);
 100                 ceph_decode_need(p, end, info->dname_len, bad);
 101                 info->dname = *p;
 102                 *p += info->dname_len;
 103                 info->dlease = *p;
 104                 *p += sizeof(*info->dlease);
 105         }
 106
 107         if (info->head->is_target) {
 108                 err = parse_reply_info_in(p, end, &info->targeti);
 109                 if (err < 0)
 110                         goto out_bad;
 111         }
 112
 113         if (unlikely(*p != end))
 114                 goto bad;
 115         return 0;
 116
 117 bad:
 118         err = -EIO;
 119 out_bad:
 120         pr_err("problem parsing mds trace %d\n", err);
 121         return err;
 122 }
 123
 124 /*
 125  * parse readdir results
 126  */
 127 static int parse_reply_info_dir(void **p, void *end,
 128                                 struct ceph_mds_reply_info_parsed *info)
 129 {
 130         u32 num, i = 0;
 131         int err;
 132
 133         info->dir_dir = *p;
 134         if (*p + sizeof(*info->dir_dir) > end)
 135                 goto bad;
 136         *p += sizeof(*info->dir_dir) +
 137                 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
 138         if (*p > end)
 139                 goto bad;
 140
 141         ceph_decode_need(p, end, sizeof(num) + 2, bad);
 142         num = ceph_decode_32(p);
 143         info->dir_end = ceph_decode_8(p);
 144         info->dir_complete = ceph_decode_8(p);
 145         if (num == 0)
 146                 goto done;
 147
 148         /* alloc large array */
 149         info->dir_nr = num;
 150         info->dir_in = kcalloc(num, sizeof(*info->dir_in) +
 151                                sizeof(*info->dir_dname) +
 152                                sizeof(*info->dir_dname_len) +
 153                                sizeof(*info->dir_dlease),
 154                                GFP_NOFS);
 155         if (info->dir_in == NULL) {
 156                 err = -ENOMEM;
 157                 goto out_bad;
 158         }
 159         info->dir_dname = (void *)(info->dir_in + num);
 160         info->dir_dname_len = (void *)(info->dir_dname + num);
 161         info->dir_dlease = (void *)(info->dir_dname_len + num);
 162
 163         while (num) {
 164                 /* dentry */
 165                 ceph_decode_need(p, end, sizeof(u32)*2, bad);
 166                 info->dir_dname_len[i] = ceph_decode_32(p);
 167                 ceph_decode_need(p, end, info->dir_dname_len[i], bad);
 168                 info->dir_dname[i] = *p;
 169                 *p += info->dir_dname_len[i];
 170                 dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
 171                      info->dir_dname[i]);
 172                 info->dir_dlease[i] = *p;
 173                 *p += sizeof(struct ceph_mds_reply_lease);
 174
 175                 /* inode */
 176                 err = parse_reply_info_in(p, end, &info->dir_in[i]);
 177                 if (err < 0)
 178                         goto out_bad;
 179                 i++;
 180                 num--;
 181         }
 182
 183 done:
 184         if (*p != end)
 185                 goto bad;
 186         return 0;
 187
 188 bad:
 189         err = -EIO;
 190 out_bad:
 191         pr_err("problem parsing dir contents %d\n", err);
 192         return err;
 193 }
 194
 195 /*
 196  * parse entire mds reply
 197  */
 198 static int parse_reply_info(struct ceph_msg *msg,
 199                             struct ceph_mds_reply_info_parsed *info)
 200 {
 201         void *p, *end;
 202         u32 len;
 203         int err;
 204
 205         info->head = msg->front.iov_base;
 206         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 207         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 208
 209         /* trace */
 210         ceph_decode_32_safe(&p, end, len, bad);
 211         if (len > 0) {
 212                 err = parse_reply_info_trace(&p, p+len, info);
 213                 if (err < 0)
 214                         goto out_bad;
 215         }
 216
 217         /* dir content */
 218         ceph_decode_32_safe(&p, end, len, bad);
 219         if (len > 0) {
 220                 err = parse_reply_info_dir(&p, p+len, info);
 221                 if (err < 0)
 222                         goto out_bad;
 223         }
 224
 225         /* snap blob */
 226         ceph_decode_32_safe(&p, end, len, bad);
 227         info->snapblob_len = len;
 228         info->snapblob = p;
 229         p += len;
 230
 231         if (p != end)
 232                 goto bad;
 233         return 0;
 234
 235 bad:
 236         err = -EIO;
 237 out_bad:
 238         pr_err("mds parse_reply err %d\n", err);
 239         return err;
 240 }
 241
 242 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
 243 {
 244         kfree(info->dir_in);
 245 }
 246
 247
 248 /*
 249  * sessions
 250  */
 251 static const char *session_state_name(int s)
 252 {
 253         switch (s) {
 254         case CEPH_MDS_SESSION_NEW: return "new";
 255         case CEPH_MDS_SESSION_OPENING: return "opening";
 256         case CEPH_MDS_SESSION_OPEN: return "open";
 257         case CEPH_MDS_SESSION_HUNG: return "hung";
 258         case CEPH_MDS_SESSION_CLOSING: return "closing";
 259         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
 260         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
 261         default: return "???";
 262         }
 263 }
 264
 265 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
 266 {
 267         if (atomic_inc_not_zero(&s->s_ref)) {
 268                 dout("mdsc get_session %p %d -> %d\n", s,
 269                      atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
 270                 return s;
 271         } else {
 272                 dout("mdsc get_session %p 0 -- FAIL", s);
 273                 return NULL;
 274         }
 275 }
 276
 277 void ceph_put_mds_session(struct ceph_mds_session *s)
 278 {
 279         dout("mdsc put_session %p %d -> %d\n", s,
 280              atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
 281         if (atomic_dec_and_test(&s->s_ref)) {
 282                 if (s->s_authorizer)
 283                         s->s_mdsc->client->monc.auth->ops->destroy_authorizer(
 284                                 s->s_mdsc->client->monc.auth, s->s_authorizer);
 285                 kfree(s);
 286         }
 287 }
 288
 289 /*
 290  * called under mdsc->mutex
 291  */
 292 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
 293                                                    int mds)
 294 {
 295         struct ceph_mds_session *session;
 296
 297         if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
 298                 return NULL;
 299         session = mdsc->sessions[mds];
 300         dout("lookup_mds_session %p %d\n", session,
 301              atomic_read(&session->s_ref));
 302         get_session(session);
 303         return session;
 304 }
 305
 306 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
 307 {
 308         if (mds >= mdsc->max_sessions)
 309                 return false;
 310         return mdsc->sessions[mds];
 311 }
 312
 313 static int __verify_registered_session(struct ceph_mds_client *mdsc,
 314                                        struct ceph_mds_session *s)
 315 {
 316         if (s->s_mds >= mdsc->max_sessions ||
 317             mdsc->sessions[s->s_mds] != s)
 318                 return -ENOENT;
 319         return 0;
 320 }
 321
 322 /*
 323  * create+register a new session for given mds.
 324  * called under mdsc->mutex.
 325  */
 326 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 327                                                  int mds)
 328 {
 329         struct ceph_mds_session *s;
 330
 331         s = kzalloc(sizeof(*s), GFP_NOFS);
 332         if (!s)
 333                 return ERR_PTR(-ENOMEM);
 334         s->s_mdsc = mdsc;
 335         s->s_mds = mds;
 336         s->s_state = CEPH_MDS_SESSION_NEW;
 337         s->s_ttl = 0;
 338         s->s_seq = 0;
 339         mutex_init(&s->s_mutex);
 340
 341         ceph_con_init(mdsc->client->msgr, &s->s_con);
 342         s->s_con.private = s;
 343         s->s_con.ops = &mds_con_ops;
 344         s->s_con.peer_name.type = CEPH_ENTITY_TYPE_MDS;
 345         s->s_con.peer_name.num = cpu_to_le64(mds);
 346
 347         spin_lock_init(&s->s_cap_lock);
 348         s->s_cap_gen = 0;
 349         s->s_cap_ttl = 0;
 350         s->s_renew_requested = 0;
 351         s->s_renew_seq = 0;
 352         INIT_LIST_HEAD(&s->s_caps);
 353         s->s_nr_caps = 0;
 354         s->s_trim_caps = 0;
 355         atomic_set(&s->s_ref, 1);
 356         INIT_LIST_HEAD(&s->s_waiting);
 357         INIT_LIST_HEAD(&s->s_unsafe);
 358         s->s_num_cap_releases = 0;
 359         s->s_cap_iterator = NULL;
 360         INIT_LIST_HEAD(&s->s_cap_releases);
 361         INIT_LIST_HEAD(&s->s_cap_releases_done);
 362         INIT_LIST_HEAD(&s->s_cap_flushing);
 363         INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
 364
 365         dout("register_session mds%d\n", mds);
 366         if (mds >= mdsc->max_sessions) {
 367                 int newmax = 1 << get_count_order(mds+1);
 368                 struct ceph_mds_session **sa;
 369
 370                 dout("register_session realloc to %d\n", newmax);
 371                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 372                 if (sa == NULL)
 373                         goto fail_realloc;
 374                 if (mdsc->sessions) {
 375                         memcpy(sa, mdsc->sessions,
 376                                mdsc->max_sessions * sizeof(void *));
 377                         kfree(mdsc->sessions);
 378                 }
 379                 mdsc->sessions = sa;
 380                 mdsc->max_sessions = newmax;
 381         }
 382         mdsc->sessions[mds] = s;
 383         atomic_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
 384
 385         ceph_con_open(&s->s_con, ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
 386
 387         return s;
 388
 389 fail_realloc:
 390         kfree(s);
 391         return ERR_PTR(-ENOMEM);
 392 }
 393
 394 /*
 395  * called under mdsc->mutex
 396  */
 397 static void __unregister_session(struct ceph_mds_client *mdsc,
 398                                struct ceph_mds_session *s)
 399 {
 400         dout("__unregister_session mds%d %p\n", s->s_mds, s);
 401         BUG_ON(mdsc->sessions[s->s_mds] != s);
 402         mdsc->sessions[s->s_mds] = NULL;
 403         ceph_con_close(&s->s_con);
 404         ceph_put_mds_session(s);
 405 }
 406
 407 /*
 408  * drop session refs in request.
 409  *
 410  * should be last request ref, or hold mdsc->mutex
 411  */
 412 static void put_request_session(struct ceph_mds_request *req)
 413 {
 414         if (req->r_session) {
 415                 ceph_put_mds_session(req->r_session);
 416                 req->r_session = NULL;
 417         }
 418 }
 419
 420 void ceph_mdsc_release_request(struct kref *kref)
 421 {
 422         struct ceph_mds_request *req = container_of(kref,
 423                                                     struct ceph_mds_request,
 424                                                     r_kref);
 425         if (req->r_request)
 426                 ceph_msg_put(req->r_request);
 427         if (req->r_reply) {
 428                 ceph_msg_put(req->r_reply);
 429                 destroy_reply_info(&req->r_reply_info);
 430         }
 431         if (req->r_inode) {
 432                 ceph_put_cap_refs(ceph_inode(req->r_inode),
 433                                   CEPH_CAP_PIN);
 434                 iput(req->r_inode);
 435         }
 436         if (req->r_locked_dir)
 437                 ceph_put_cap_refs(ceph_inode(req->r_locked_dir),
 438                                   CEPH_CAP_PIN);
 439         if (req->r_target_inode)
 440                 iput(req->r_target_inode);
 441         if (req->r_dentry)
 442                 dput(req->r_dentry);
 443         if (req->r_old_dentry) {
 444                 ceph_put_cap_refs(
 445                         ceph_inode(req->r_old_dentry->d_parent->d_inode),
 446                         CEPH_CAP_PIN);
 447                 dput(req->r_old_dentry);
 448         }
 449         kfree(req->r_path1);
 450         kfree(req->r_path2);
 451         put_request_session(req);
 452         ceph_unreserve_caps(&req->r_caps_reservation);
 453         kfree(req);
 454 }
 455
 456 /*
 457  * lookup session, bump ref if found.
 458  *
 459  * called under mdsc->mutex.
 460  */
 461 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
 462                                              u64 tid)
 463 {
 464         struct ceph_mds_request *req;
 465         struct rb_node *n = mdsc->request_tree.rb_node;
 466
 467         while (n) {
 468                 req = rb_entry(n, struct ceph_mds_request, r_node);
 469                 if (tid < req->r_tid)
 470                         n = n->rb_left;
 471                 else if (tid > req->r_tid)
 472                         n = n->rb_right;
 473                 else {
 474                         ceph_mdsc_get_request(req);
 475                         return req;
 476                 }
 477         }
 478         return NULL;
 479 }
 480
 481 static void __insert_request(struct ceph_mds_client *mdsc,
 482                              struct ceph_mds_request *new)
 483 {
 484         struct rb_node **p = &mdsc->request_tree.rb_node;
 485         struct rb_node *parent = NULL;
 486         struct ceph_mds_request *req = NULL;
 487
 488         while (*p) {
 489                 parent = *p;
 490                 req = rb_entry(parent, struct ceph_mds_request, r_node);
 491                 if (new->r_tid < req->r_tid)
 492                         p = &(*p)->rb_left;
 493                 else if (new->r_tid > req->r_tid)
 494                         p = &(*p)->rb_right;
 495                 else
 496                         BUG();
 497         }
 498
 499         rb_link_node(&new->r_node, parent, p);
 500         rb_insert_color(&new->r_node, &mdsc->request_tree);
 501 }
 502
 503 /*
 504  * Register an in-flight request, and assign a tid.  Link to directory
 505  * are modifying (if any).
 506  *
 507  * Called under mdsc->mutex.
 508  */
 509 static void __register_request(struct ceph_mds_client *mdsc,
 510                                struct ceph_mds_request *req,
 511                                struct inode *dir)
 512 {
 513         req->r_tid = ++mdsc->last_tid;
 514         if (req->r_num_caps)
 515                 ceph_reserve_caps(&req->r_caps_reservation, req->r_num_caps);
 516         dout("__register_request %p tid %lld\n", req, req->r_tid);
 517         ceph_mdsc_get_request(req);
 518         __insert_request(mdsc, req);
 519
 520         if (dir) {
 521                 struct ceph_inode_info *ci = ceph_inode(dir);
 522
 523                 spin_lock(&ci->i_unsafe_lock);
 524                 req->r_unsafe_dir = dir;
 525                 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
 526                 spin_unlock(&ci->i_unsafe_lock);
 527         }
 528 }
 529
 530 static void __unregister_request(struct ceph_mds_client *mdsc,
 531                                  struct ceph_mds_request *req)
 532 {
 533         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 534         rb_erase(&req->r_node, &mdsc->request_tree);
 535         RB_CLEAR_NODE(&req->r_node);
 536
 537         if (req->r_unsafe_dir) {
 538                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 539
 540                 spin_lock(&ci->i_unsafe_lock);
 541                 list_del_init(&req->r_unsafe_dir_item);
 542                 spin_unlock(&ci->i_unsafe_lock);
 543         }
 544
 545         ceph_mdsc_put_request(req);
 546 }
 547
 548 /*
 549  * Choose mds to send request to next.  If there is a hint set in the
 550  * request (e.g., due to a prior forward hint from the mds), use that.
 551  * Otherwise, consult frag tree and/or caps to identify the
 552  * appropriate mds.  If all else fails, choose randomly.
 553  *
 554  * Called under mdsc->mutex.
 555  */
 556 static int __choose_mds(struct ceph_mds_client *mdsc,
 557                         struct ceph_mds_request *req)
 558 {
 559         struct inode *inode;
 560         struct ceph_inode_info *ci;
 561         struct ceph_cap *cap;
 562         int mode = req->r_direct_mode;
 563         int mds = -1;
 564         u32 hash = req->r_direct_hash;
 565         bool is_hash = req->r_direct_is_hash;
 566
 567         /*
 568          * is there a specific mds we should try?  ignore hint if we have
 569          * no session and the mds is not up (active or recovering).
 570          */
 571         if (req->r_resend_mds >= 0 &&
 572             (__have_session(mdsc, req->r_resend_mds) ||
 573              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 574                 dout("choose_mds using resend_mds mds%d\n",
 575                      req->r_resend_mds);
 576                 return req->r_resend_mds;
 577         }
 578
 579         if (mode == USE_RANDOM_MDS)
 580                 goto random;
 581
 582         inode = NULL;
 583         if (req->r_inode) {
 584                 inode = req->r_inode;
 585         } else if (req->r_dentry) {
 586                 if (req->r_dentry->d_inode) {
 587                         inode = req->r_dentry->d_inode;
 588                 } else {
 589                         inode = req->r_dentry->d_parent->d_inode;
 590                         hash = req->r_dentry->d_name.hash;
 591                         is_hash = true;
 592                 }
 593         }
 594         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 595              (int)hash, mode);
 596         if (!inode)
 597                 goto random;
 598         ci = ceph_inode(inode);
 599
 600         if (is_hash && S_ISDIR(inode->i_mode)) {
 601                 struct ceph_inode_frag frag;
 602                 int found;
 603
 604                 ceph_choose_frag(ci, hash, &frag, &found);
 605                 if (found) {
 606                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 607                                 u8 r;
 608
 609                                 /* choose a random replica */
 610                                 get_random_bytes(&r, 1);
 611                                 r %= frag.ndist;
 612                                 mds = frag.dist[r];
 613                                 dout("choose_mds %p %llx.%llx "
 614                                      "frag %u mds%d (%d/%d)\n",
 615                                      inode, ceph_vinop(inode),
 616                                      frag.frag, frag.mds,
 617                                      (int)r, frag.ndist);
 618                                 return mds;
 619                         }
 620
 621                         /* since this file/dir wasn't known to be
 622                          * replicated, then we want to look for the
 623                          * authoritative mds. */
 624                         mode = USE_AUTH_MDS;
 625                         if (frag.mds >= 0) {
 626                                 /* choose auth mds */
 627                                 mds = frag.mds;
 628                                 dout("choose_mds %p %llx.%llx "
 629                                      "frag %u mds%d (auth)\n",
 630                                      inode, ceph_vinop(inode), frag.frag, mds);
 631                                 return mds;
 632                         }
 633                 }
 634         }
 635
 636         spin_lock(&inode->i_lock);
 637         cap = NULL;
 638         if (mode == USE_AUTH_MDS)
 639                 cap = ci->i_auth_cap;
 640         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
 641                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
 642         if (!cap) {
 643                 spin_unlock(&inode->i_lock);
 644                 goto random;
 645         }
 646         mds = cap->session->s_mds;
 647         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
 648              inode, ceph_vinop(inode), mds,
 649              cap == ci->i_auth_cap ? "auth " : "", cap);
 650         spin_unlock(&inode->i_lock);
 651         return mds;
 652
 653 random:
 654         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
 655         dout("choose_mds chose random mds%d\n", mds);
 656         return mds;
 657 }
 658
 659
 660 /*
 661  * session messages
 662  */
 663 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
 664 {
 665         struct ceph_msg *msg;
 666         struct ceph_mds_session_head *h;
 667
 668         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), 0, 0, NULL);
 669         if (IS_ERR(msg)) {
 670                 pr_err("create_session_msg ENOMEM creating msg\n");
 671                 return ERR_PTR(PTR_ERR(msg));
 672         }
 673         h = msg->front.iov_base;
 674         h->op = cpu_to_le32(op);
 675         h->seq = cpu_to_le64(seq);
 676         return msg;
 677 }
 678
 679 /*
 680  * send session open request.
 681  *
 682  * called under mdsc->mutex
 683  */
 684 static int __open_session(struct ceph_mds_client *mdsc,
 685                           struct ceph_mds_session *session)
 686 {
 687         struct ceph_msg *msg;
 688         int mstate;
 689         int mds = session->s_mds;
 690         int err = 0;
 691
 692         /* wait for mds to go active? */
 693         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
 694         dout("open_session to mds%d (%s)\n", mds,
 695              ceph_mds_state_name(mstate));
 696         session->s_state = CEPH_MDS_SESSION_OPENING;
 697         session->s_renew_requested = jiffies;
 698
 699         /* send connect message */
 700         msg = create_session_msg(CEPH_SESSION_REQUEST_OPEN, session->s_seq);
 701         if (IS_ERR(msg)) {
 702                 err = PTR_ERR(msg);
 703                 goto out;
 704         }
 705         ceph_con_send(&session->s_con, msg);
 706
 707 out:
 708         return 0;
 709 }
 710
 711 /*
 712  * session caps
 713  */
 714
 715 /*
 716  * Free preallocated cap messages assigned to this session
 717  */
 718 static void cleanup_cap_releases(struct ceph_mds_session *session)
 719 {
 720         struct ceph_msg *msg;
 721
 722         spin_lock(&session->s_cap_lock);
 723         while (!list_empty(&session->s_cap_releases)) {
 724                 msg = list_first_entry(&session->s_cap_releases,
 725                                        struct ceph_msg, list_head);
 726                 list_del_init(&msg->list_head);
 727                 ceph_msg_put(msg);
 728         }
 729         while (!list_empty(&session->s_cap_releases_done)) {
 730                 msg = list_first_entry(&session->s_cap_releases_done,
 731                                        struct ceph_msg, list_head);
 732                 list_del_init(&msg->list_head);
 733                 ceph_msg_put(msg);
 734         }
 735         spin_unlock(&session->s_cap_lock);
 736 }
 737
 738 /*
 739  * Helper to safely iterate over all caps associated with a session, with
 740  * special care taken to handle a racing __ceph_remove_cap().
 741  *
 742  * Caller must hold session s_mutex.
 743  */
 744 static int iterate_session_caps(struct ceph_mds_session *session,
 745                                  int (*cb)(struct inode *, struct ceph_cap *,
 746                                             void *), void *arg)
 747 {
 748         struct list_head *p;
 749         struct ceph_cap *cap;
 750         struct inode *inode, *last_inode = NULL;
 751         struct ceph_cap *old_cap = NULL;
 752         int ret;
 753
 754         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
 755         spin_lock(&session->s_cap_lock);
 756         p = session->s_caps.next;
 757         while (p != &session->s_caps) {
 758                 cap = list_entry(p, struct ceph_cap, session_caps);
 759                 inode = igrab(&cap->ci->vfs_inode);
 760                 if (!inode) {
 761                         p = p->next;
 762                         continue;
 763                 }
 764                 session->s_cap_iterator = cap;
 765                 spin_unlock(&session->s_cap_lock);
 766
 767                 if (last_inode) {
 768                         iput(last_inode);
 769                         last_inode = NULL;
 770                 }
 771                 if (old_cap) {
 772                         ceph_put_cap(old_cap);
 773                         old_cap = NULL;
 774                 }
 775
 776                 ret = cb(inode, cap, arg);
 777                 last_inode = inode;
 778
 779                 spin_lock(&session->s_cap_lock);
 780                 p = p->next;
 781                 if (cap->ci == NULL) {
 782                         dout("iterate_session_caps  finishing cap %p removal\n",
 783                              cap);
 784                         BUG_ON(cap->session != session);
 785                         list_del_init(&cap->session_caps);
 786                         session->s_nr_caps--;
 787                         cap->session = NULL;
 788                         old_cap = cap;  /* put_cap it w/o locks held */
 789                 }
 790                 if (ret < 0)
 791                         goto out;
 792         }
 793         ret = 0;
 794 out:
 795         session->s_cap_iterator = NULL;
 796         spin_unlock(&session->s_cap_lock);
 797
 798         if (last_inode)
 799                 iput(last_inode);
 800         if (old_cap)
 801                 ceph_put_cap(old_cap);
 802
 803         return ret;
 804 }
 805
 806 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
 807                                    void *arg)
 808 {
 809         struct ceph_inode_info *ci = ceph_inode(inode);
 810         dout("removing cap %p, ci is %p, inode is %p\n",
 811              cap, ci, &ci->vfs_inode);
 812         ceph_remove_cap(cap);
 813         return 0;
 814 }
 815
 816 /*
 817  * caller must hold session s_mutex
 818  */
 819 static void remove_session_caps(struct ceph_mds_session *session)
 820 {
 821         dout("remove_session_caps on %p\n", session);
 822         iterate_session_caps(session, remove_session_caps_cb, NULL);
 823         BUG_ON(session->s_nr_caps > 0);
 824         cleanup_cap_releases(session);
 825 }
 826
 827 /*
 828  * wake up any threads waiting on this session's caps.  if the cap is
 829  * old (didn't get renewed on the client reconnect), remove it now.
 830  *
 831  * caller must hold s_mutex.
 832  */
 833 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
 834                               void *arg)
 835 {
 836         struct ceph_inode_info *ci = ceph_inode(inode);
 837
 838         wake_up(&ci->i_cap_wq);
 839         if (arg) {
 840                 spin_lock(&inode->i_lock);
 841                 ci->i_wanted_max_size = 0;
 842                 ci->i_requested_max_size = 0;
 843                 spin_unlock(&inode->i_lock);
 844         }
 845         return 0;
 846 }
 847
 848 static void wake_up_session_caps(struct ceph_mds_session *session,
 849                                  int reconnect)
 850 {
 851         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
 852         iterate_session_caps(session, wake_up_session_cb,
 853                              (void *)(unsigned long)reconnect);
 854 }
 855
 856 /*
 857  * Send periodic message to MDS renewing all currently held caps.  The
 858  * ack will reset the expiration for all caps from this session.
 859  *
 860  * caller holds s_mutex
 861  */
 862 static int send_renew_caps(struct ceph_mds_client *mdsc,
 863                            struct ceph_mds_session *session)
 864 {
 865         struct ceph_msg *msg;
 866         int state;
 867
 868         if (time_after_eq(jiffies, session->s_cap_ttl) &&
 869             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
 870                 pr_info("mds%d caps stale\n", session->s_mds);
 871         session->s_renew_requested = jiffies;
 872
 873         /* do not try to renew caps until a recovering mds has reconnected
 874          * with its clients. */
 875         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
 876         if (state < CEPH_MDS_STATE_RECONNECT) {
 877                 dout("send_renew_caps ignoring mds%d (%s)\n",
 878                      session->s_mds, ceph_mds_state_name(state));
 879                 return 0;
 880         }
 881
 882         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
 883                 ceph_mds_state_name(state));
 884         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
 885                                  ++session->s_renew_seq);
 886         if (IS_ERR(msg))
 887                 return PTR_ERR(msg);
 888         ceph_con_send(&session->s_con, msg);
 889         return 0;
 890 }
 891
 892 /*
 893  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
 894  *
 895  * Called under session->s_mutex
 896  */
 897 static void renewed_caps(struct ceph_mds_client *mdsc,
 898                          struct ceph_mds_session *session, int is_renew)
 899 {
 900         int was_stale;
 901         int wake = 0;
 902
 903         spin_lock(&session->s_cap_lock);
 904         was_stale = is_renew && (session->s_cap_ttl == 0 ||
 905                                  time_after_eq(jiffies, session->s_cap_ttl));
 906
 907         session->s_cap_ttl = session->s_renew_requested +
 908                 mdsc->mdsmap->m_session_timeout*HZ;
 909
 910         if (was_stale) {
 911                 if (time_before(jiffies, session->s_cap_ttl)) {
 912                         pr_info("mds%d caps renewed\n", session->s_mds);
 913                         wake = 1;
 914                 } else {
 915                         pr_info("mds%d caps still stale\n", session->s_mds);
 916                 }
 917         }
 918         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
 919              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
 920              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
 921         spin_unlock(&session->s_cap_lock);
 922
 923         if (wake)
 924                 wake_up_session_caps(session, 0);
 925 }
 926
 927 /*
 928  * send a session close request
 929  */
 930 static int request_close_session(struct ceph_mds_client *mdsc,
 931                                  struct ceph_mds_session *session)
 932 {
 933         struct ceph_msg *msg;
 934         int err = 0;
 935
 936         dout("request_close_session mds%d state %s seq %lld\n",
 937              session->s_mds, session_state_name(session->s_state),
 938              session->s_seq);
 939         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
 940         if (IS_ERR(msg))
 941                 err = PTR_ERR(msg);
 942         else
 943                 ceph_con_send(&session->s_con, msg);
 944         return err;
 945 }
 946
 947 /*
 948  * Called with s_mutex held.
 949  */
 950 static int __close_session(struct ceph_mds_client *mdsc,
 951                          struct ceph_mds_session *session)
 952 {
 953         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
 954                 return 0;
 955         session->s_state = CEPH_MDS_SESSION_CLOSING;
 956         return request_close_session(mdsc, session);
 957 }
 958
 959 /*
 960  * Trim old(er) caps.
 961  *
 962  * Because we can't cache an inode without one or more caps, we do
 963  * this indirectly: if a cap is unused, we prune its aliases, at which
 964  * point the inode will hopefully get dropped to.
 965  *
 966  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
 967  * memory pressure from the MDS, though, so it needn't be perfect.
 968  */
 969 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
 970 {
 971         struct ceph_mds_session *session = arg;
 972         struct ceph_inode_info *ci = ceph_inode(inode);
 973         int used, oissued, mine;
 974
 975         if (session->s_trim_caps <= 0)
 976                 return -1;
 977
 978         spin_lock(&inode->i_lock);
 979         mine = cap->issued | cap->implemented;
 980         used = __ceph_caps_used(ci);
 981         oissued = __ceph_caps_issued_other(ci, cap);
 982
 983         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
 984              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
 985              ceph_cap_string(used));
 986         if (ci->i_dirty_caps)
 987                 goto out;   /* dirty caps */
 988         if ((used & ~oissued) & mine)
 989                 goto out;   /* we need these caps */
 990
 991         session->s_trim_caps--;
 992         if (oissued) {
 993                 /* we aren't the only cap.. just remove us */
 994                 __ceph_remove_cap(cap);
 995         } else {
 996                 /* try to drop referring dentries */
 997                 spin_unlock(&inode->i_lock);
 998                 d_prune_aliases(inode);
 999                 dout("trim_caps_cb %p cap %p  pruned, count now %d\n",
1000                      inode, cap, atomic_read(&inode->i_count));
1001                 return 0;
1002         }
1003
1004 out:
1005         spin_unlock(&inode->i_lock);
1006         return 0;
1007 }
1008
1009 /*
1010  * Trim session cap count down to some max number.
1011  */
1012 static int trim_caps(struct ceph_mds_client *mdsc,
1013                      struct ceph_mds_session *session,
1014                      int max_caps)
1015 {
1016         int trim_caps = session->s_nr_caps - max_caps;
1017
1018         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1019              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1020         if (trim_caps > 0) {
1021                 session->s_trim_caps = trim_caps;
1022                 iterate_session_caps(session, trim_caps_cb, session);
1023                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1024                      session->s_mds, session->s_nr_caps, max_caps,
1025                         trim_caps - session->s_trim_caps);
1026                 session->s_trim_caps = 0;
1027         }
1028         return 0;
1029 }
1030
1031 /*
1032  * Allocate cap_release messages.  If there is a partially full message
1033  * in the queue, try to allocate enough to cover it's remainder, so that
1034  * we can send it immediately.
1035  *
1036  * Called under s_mutex.
1037  */
1038 static int add_cap_releases(struct ceph_mds_client *mdsc,
1039                             struct ceph_mds_session *session,
1040                             int extra)
1041 {
1042         struct ceph_msg *msg;
1043         struct ceph_mds_cap_release *head;
1044         int err = -ENOMEM;
1045
1046         if (extra < 0)
1047                 extra = mdsc->client->mount_args->cap_release_safety;
1048
1049         spin_lock(&session->s_cap_lock);
1050
1051         if (!list_empty(&session->s_cap_releases)) {
1052                 msg = list_first_entry(&session->s_cap_releases,
1053                                        struct ceph_msg,
1054                                  list_head);
1055                 head = msg->front.iov_base;
1056                 extra += CEPH_CAPS_PER_RELEASE - le32_to_cpu(head->num);
1057         }
1058
1059         while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1060                 spin_unlock(&session->s_cap_lock);
1061                 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
1062                                    0, 0, NULL);
1063                 if (!msg)
1064                         goto out_unlocked;
1065                 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1066                      (int)msg->front.iov_len);
1067                 head = msg->front.iov_base;
1068                 head->num = cpu_to_le32(0);
1069                 msg->front.iov_len = sizeof(*head);
1070                 spin_lock(&session->s_cap_lock);
1071                 list_add(&msg->list_head, &session->s_cap_releases);
1072                 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1073         }
1074
1075         if (!list_empty(&session->s_cap_releases)) {
1076                 msg = list_first_entry(&session->s_cap_releases,
1077                                        struct ceph_msg,
1078                                        list_head);
1079                 head = msg->front.iov_base;
1080                 if (head->num) {
1081                         dout(" queueing non-full %p (%d)\n", msg,
1082                              le32_to_cpu(head->num));
1083                         list_move_tail(&msg->list_head,
1084                                       &session->s_cap_releases_done);
1085                         session->s_num_cap_releases -=
1086                                 CEPH_CAPS_PER_RELEASE - le32_to_cpu(head->num);
1087                 }
1088         }
1089         err = 0;
1090         spin_unlock(&session->s_cap_lock);
1091 out_unlocked:
1092         return err;
1093 }
1094
1095 /*
1096  * flush all dirty inode data to disk.
1097  *
1098  * returns true if we've flushed through want_flush_seq
1099  */
1100 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1101 {
1102         int mds, ret = 1;
1103
1104         dout("check_cap_flush want %lld\n", want_flush_seq);
1105         mutex_lock(&mdsc->mutex);
1106         for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1107                 struct ceph_mds_session *session = mdsc->sessions[mds];
1108
1109                 if (!session)
1110                         continue;
1111                 get_session(session);
1112                 mutex_unlock(&mdsc->mutex);
1113
1114                 mutex_lock(&session->s_mutex);
1115                 if (!list_empty(&session->s_cap_flushing)) {
1116                         struct ceph_inode_info *ci =
1117                                 list_entry(session->s_cap_flushing.next,
1118                                            struct ceph_inode_info,
1119                                            i_flushing_item);
1120                         struct inode *inode = &ci->vfs_inode;
1121
1122                         spin_lock(&inode->i_lock);
1123                         if (ci->i_cap_flush_seq <= want_flush_seq) {
1124                                 dout("check_cap_flush still flushing %p "
1125                                      "seq %lld <= %lld to mds%d\n", inode,
1126                                      ci->i_cap_flush_seq, want_flush_seq,
1127                                      session->s_mds);
1128                                 ret = 0;
1129                         }
1130                         spin_unlock(&inode->i_lock);
1131                 }
1132                 mutex_unlock(&session->s_mutex);
1133                 ceph_put_mds_session(session);
1134
1135                 if (!ret)
1136                         return ret;
1137                 mutex_lock(&mdsc->mutex);
1138         }
1139
1140         mutex_unlock(&mdsc->mutex);
1141         dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1142         return ret;
1143 }
1144
1145 /*
1146  * called under s_mutex
1147  */
1148 static void send_cap_releases(struct ceph_mds_client *mdsc,
1149                        struct ceph_mds_session *session)
1150 {
1151         struct ceph_msg *msg;
1152
1153         dout("send_cap_releases mds%d\n", session->s_mds);
1154         while (1) {
1155                 spin_lock(&session->s_cap_lock);
1156                 if (list_empty(&session->s_cap_releases_done))
1157                         break;
1158                 msg = list_first_entry(&session->s_cap_releases_done,
1159                                  struct ceph_msg, list_head);
1160                 list_del_init(&msg->list_head);
1161                 spin_unlock(&session->s_cap_lock);
1162                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1163                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1164                 ceph_con_send(&session->s_con, msg);
1165         }
1166         spin_unlock(&session->s_cap_lock);
1167 }
1168
1169 /*
1170  * requests
1171  */
1172
1173 /*
1174  * Create an mds request.
1175  */
1176 struct ceph_mds_request *
1177 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1178 {
1179         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1180
1181         if (!req)
1182                 return ERR_PTR(-ENOMEM);
1183
1184         req->r_started = jiffies;
1185         req->r_resend_mds = -1;
1186         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1187         req->r_fmode = -1;
1188         kref_init(&req->r_kref);
1189         INIT_LIST_HEAD(&req->r_wait);
1190         init_completion(&req->r_completion);
1191         init_completion(&req->r_safe_completion);
1192         INIT_LIST_HEAD(&req->r_unsafe_item);
1193
1194         req->r_op = op;
1195         req->r_direct_mode = mode;
1196         return req;
1197 }
1198
1199 /*
1200  * return oldest (lowest) request, tid in request tree, 0 if none.
1201  *
1202  * called under mdsc->mutex.
1203  */
1204 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1205 {
1206         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1207                 return NULL;
1208         return rb_entry(rb_first(&mdsc->request_tree),
1209                         struct ceph_mds_request, r_node);
1210 }
1211
1212 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1213 {
1214         struct ceph_mds_request *req = __get_oldest_req(mdsc);
1215
1216         if (req)
1217                 return req->r_tid;
1218         return 0;
1219 }
1220
1221 /*
1222  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1223  * on build_path_from_dentry in fs/cifs/dir.c.
1224  *
1225  * If @stop_on_nosnap, generate path relative to the first non-snapped
1226  * inode.
1227  *
1228  * Encode hidden .snap dirs as a double /, i.e.
1229  *   foo/.snap/bar -> foo//bar
1230  */
1231 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1232                            int stop_on_nosnap)
1233 {
1234         struct dentry *temp;
1235         char *path;
1236         int len, pos;
1237
1238         if (dentry == NULL)
1239                 return ERR_PTR(-EINVAL);
1240
1241 retry:
1242         len = 0;
1243         for (temp = dentry; !IS_ROOT(temp);) {
1244                 struct inode *inode = temp->d_inode;
1245                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1246                         len++;  /* slash only */
1247                 else if (stop_on_nosnap && inode &&
1248                          ceph_snap(inode) == CEPH_NOSNAP)
1249                         break;
1250                 else
1251                         len += 1 + temp->d_name.len;
1252                 temp = temp->d_parent;
1253                 if (temp == NULL) {
1254                         pr_err("build_path_dentry corrupt dentry %p\n", dentry);
1255                         return ERR_PTR(-EINVAL);
1256                 }
1257         }
1258         if (len)
1259                 len--;  /* no leading '/' */
1260
1261         path = kmalloc(len+1, GFP_NOFS);
1262         if (path == NULL)
1263                 return ERR_PTR(-ENOMEM);
1264         pos = len;
1265         path[pos] = 0;  /* trailing null */
1266         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1267                 struct inode *inode = temp->d_inode;
1268
1269                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1270                         dout("build_path_dentry path+%d: %p SNAPDIR\n",
1271                              pos, temp);
1272                 } else if (stop_on_nosnap && inode &&
1273                            ceph_snap(inode) == CEPH_NOSNAP) {
1274                         break;
1275                 } else {
1276                         pos -= temp->d_name.len;
1277                         if (pos < 0)
1278                                 break;
1279                         strncpy(path + pos, temp->d_name.name,
1280                                 temp->d_name.len);
1281                         dout("build_path_dentry path+%d: %p '%.*s'\n",
1282                              pos, temp, temp->d_name.len, path + pos);
1283                 }
1284                 if (pos)
1285                         path[--pos] = '/';
1286                 temp = temp->d_parent;
1287                 if (temp == NULL) {
1288                         pr_err("build_path_dentry corrupt dentry\n");
1289                         kfree(path);
1290                         return ERR_PTR(-EINVAL);
1291                 }
1292         }
1293         if (pos != 0) {
1294                 pr_err("build_path_dentry did not end path lookup where "
1295                        "expected, namelen is %d, pos is %d\n", len, pos);
1296                 /* presumably this is only possible if racing with a
1297                    rename of one of the parent directories (we can not
1298                    lock the dentries above us to prevent this, but
1299                    retrying should be harmless) */
1300                 kfree(path);
1301                 goto retry;
1302         }
1303
1304         *base = ceph_ino(temp->d_inode);
1305         *plen = len;
1306         dout("build_path_dentry on %p %d built %llx '%.*s'\n",
1307              dentry, atomic_read(&dentry->d_count), *base, len, path);
1308         return path;
1309 }
1310
1311 static int build_dentry_path(struct dentry *dentry,
1312                              const char **ppath, int *ppathlen, u64 *pino,
1313                              int *pfreepath)
1314 {
1315         char *path;
1316
1317         if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1318                 *pino = ceph_ino(dentry->d_parent->d_inode);
1319                 *ppath = dentry->d_name.name;
1320                 *ppathlen = dentry->d_name.len;
1321                 return 0;
1322         }
1323         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1324         if (IS_ERR(path))
1325                 return PTR_ERR(path);
1326         *ppath = path;
1327         *pfreepath = 1;
1328         return 0;
1329 }
1330
1331 static int build_inode_path(struct inode *inode,
1332                             const char **ppath, int *ppathlen, u64 *pino,
1333                             int *pfreepath)
1334 {
1335         struct dentry *dentry;
1336         char *path;
1337
1338         if (ceph_snap(inode) == CEPH_NOSNAP) {
1339                 *pino = ceph_ino(inode);
1340                 *ppathlen = 0;
1341                 return 0;
1342         }
1343         dentry = d_find_alias(inode);
1344         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1345         dput(dentry);
1346         if (IS_ERR(path))
1347                 return PTR_ERR(path);
1348         *ppath = path;
1349         *pfreepath = 1;
1350         return 0;
1351 }
1352
1353 /*
1354  * request arguments may be specified via an inode *, a dentry *, or
1355  * an explicit ino+path.
1356  */
1357 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1358                                   const char *rpath, u64 rino,
1359                                   const char **ppath, int *pathlen,
1360                                   u64 *ino, int *freepath)
1361 {
1362         int r = 0;
1363
1364         if (rinode) {
1365                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1366                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1367                      ceph_snap(rinode));
1368         } else if (rdentry) {
1369                 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1370                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1371                      *ppath);
1372         } else if (rpath) {
1373                 *ino = rino;
1374                 *ppath = rpath;
1375                 *pathlen = strlen(rpath);
1376                 dout(" path %.*s\n", *pathlen, rpath);
1377         }
1378
1379         return r;
1380 }
1381
1382 /*
1383  * called under mdsc->mutex
1384  */
1385 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1386                                                struct ceph_mds_request *req,
1387                                                int mds)
1388 {
1389         struct ceph_msg *msg;
1390         struct ceph_mds_request_head *head;
1391         const char *path1 = NULL;
1392         const char *path2 = NULL;
1393         u64 ino1 = 0, ino2 = 0;
1394         int pathlen1 = 0, pathlen2 = 0;
1395         int freepath1 = 0, freepath2 = 0;
1396         int len;
1397         u16 releases;
1398         void *p, *end;
1399         int ret;
1400
1401         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1402                               req->r_path1, req->r_ino1.ino,
1403                               &path1, &pathlen1, &ino1, &freepath1);
1404         if (ret < 0) {
1405                 msg = ERR_PTR(ret);
1406                 goto out;
1407         }
1408
1409         ret = set_request_path_attr(NULL, req->r_old_dentry,
1410                               req->r_path2, req->r_ino2.ino,
1411                               &path2, &pathlen2, &ino2, &freepath2);
1412         if (ret < 0) {
1413                 msg = ERR_PTR(ret);
1414                 goto out_free1;
1415         }
1416
1417         len = sizeof(*head) +
1418                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64));
1419
1420         /* calculate (max) length for cap releases */
1421         len += sizeof(struct ceph_mds_request_release) *
1422                 (!!req->r_inode_drop + !!req->r_dentry_drop +
1423                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1424         if (req->r_dentry_drop)
1425                 len += req->r_dentry->d_name.len;
1426         if (req->r_old_dentry_drop)
1427                 len += req->r_old_dentry->d_name.len;
1428
1429         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, 0, 0, NULL);
1430         if (IS_ERR(msg))
1431                 goto out_free2;
1432
1433         msg->hdr.tid = cpu_to_le64(req->r_tid);
1434
1435         head = msg->front.iov_base;
1436         p = msg->front.iov_base + sizeof(*head);
1437         end = msg->front.iov_base + msg->front.iov_len;
1438
1439         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1440         head->op = cpu_to_le32(req->r_op);
1441         head->caller_uid = cpu_to_le32(current_fsuid());
1442         head->caller_gid = cpu_to_le32(current_fsgid());
1443         head->args = req->r_args;
1444
1445         ceph_encode_filepath(&p, end, ino1, path1);
1446         ceph_encode_filepath(&p, end, ino2, path2);
1447
1448         /* cap releases */
1449         releases = 0;
1450         if (req->r_inode_drop)
1451                 releases += ceph_encode_inode_release(&p,
1452                       req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1453                       mds, req->r_inode_drop, req->r_inode_unless, 0);
1454         if (req->r_dentry_drop)
1455                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1456                        mds, req->r_dentry_drop, req->r_dentry_unless);
1457         if (req->r_old_dentry_drop)
1458                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1459                        mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1460         if (req->r_old_inode_drop)
1461                 releases += ceph_encode_inode_release(&p,
1462                       req->r_old_dentry->d_inode,
1463                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1464         head->num_releases = cpu_to_le16(releases);
1465
1466         BUG_ON(p > end);
1467         msg->front.iov_len = p - msg->front.iov_base;
1468         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1469
1470         msg->pages = req->r_pages;
1471         msg->nr_pages = req->r_num_pages;
1472         msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1473         msg->hdr.data_off = cpu_to_le16(0);
1474
1475 out_free2:
1476         if (freepath2)
1477                 kfree((char *)path2);
1478 out_free1:
1479         if (freepath1)
1480                 kfree((char *)path1);
1481 out:
1482         return msg;
1483 }
1484
1485 /*
1486  * called under mdsc->mutex if error, under no mutex if
1487  * success.
1488  */
1489 static void complete_request(struct ceph_mds_client *mdsc,
1490                              struct ceph_mds_request *req)
1491 {
1492         if (req->r_callback)
1493                 req->r_callback(mdsc, req);
1494         else
1495                 complete(&req->r_completion);
1496 }
1497
1498 /*
1499  * called under mdsc->mutex
1500  */
1501 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1502                                   struct ceph_mds_request *req,
1503                                   int mds)
1504 {
1505         struct ceph_mds_request_head *rhead;
1506         struct ceph_msg *msg;
1507         int flags = 0;
1508
1509         req->r_mds = mds;
1510         req->r_attempts++;
1511         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1512              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1513
1514         if (req->r_request) {
1515                 ceph_msg_put(req->r_request);
1516                 req->r_request = NULL;
1517         }
1518         msg = create_request_message(mdsc, req, mds);
1519         if (IS_ERR(msg)) {
1520                 req->r_err = PTR_ERR(msg);
1521                 complete_request(mdsc, req);
1522                 return -PTR_ERR(msg);
1523         }
1524         req->r_request = msg;
1525
1526         rhead = msg->front.iov_base;
1527         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1528         if (req->r_got_unsafe)
1529                 flags |= CEPH_MDS_FLAG_REPLAY;
1530         if (req->r_locked_dir)
1531                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1532         rhead->flags = cpu_to_le32(flags);
1533         rhead->num_fwd = req->r_num_fwd;
1534         rhead->num_retry = req->r_attempts - 1;
1535
1536         dout(" r_locked_dir = %p\n", req->r_locked_dir);
1537
1538         if (req->r_target_inode && req->r_got_unsafe)
1539                 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1540         else
1541                 rhead->ino = 0;
1542         return 0;
1543 }
1544
1545 /*
1546  * send request, or put it on the appropriate wait list.
1547  */
1548 static int __do_request(struct ceph_mds_client *mdsc,
1549                         struct ceph_mds_request *req)
1550 {
1551         struct ceph_mds_session *session = NULL;
1552         int mds = -1;
1553         int err = -EAGAIN;
1554
1555         if (req->r_err || req->r_got_result)
1556                 goto out;
1557
1558         if (req->r_timeout &&
1559             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1560                 dout("do_request timed out\n");
1561                 err = -EIO;
1562                 goto finish;
1563         }
1564
1565         mds = __choose_mds(mdsc, req);
1566         if (mds < 0 ||
1567             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1568                 dout("do_request no mds or not active, waiting for map\n");
1569                 list_add(&req->r_wait, &mdsc->waiting_for_map);
1570                 goto out;
1571         }
1572
1573         /* get, open session */
1574         session = __ceph_lookup_mds_session(mdsc, mds);
1575         if (!session) {
1576                 session = register_session(mdsc, mds);
1577                 if (IS_ERR(session)) {
1578                         err = PTR_ERR(session);
1579                         goto finish;
1580                 }
1581         }
1582         dout("do_request mds%d session %p state %s\n", mds, session,
1583              session_state_name(session->s_state));
1584         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1585             session->s_state != CEPH_MDS_SESSION_HUNG) {
1586                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1587                     session->s_state == CEPH_MDS_SESSION_CLOSING)
1588                         __open_session(mdsc, session);
1589                 list_add(&req->r_wait, &session->s_waiting);
1590                 goto out_session;
1591         }
1592
1593         /* send request */
1594         req->r_session = get_session(session);
1595         req->r_resend_mds = -1;   /* forget any previous mds hint */
1596
1597         if (req->r_request_started == 0)   /* note request start time */
1598                 req->r_request_started = jiffies;
1599
1600         err = __prepare_send_request(mdsc, req, mds);
1601         if (!err) {
1602                 ceph_msg_get(req->r_request);
1603                 ceph_con_send(&session->s_con, req->r_request);
1604         }
1605
1606 out_session:
1607         ceph_put_mds_session(session);
1608 out:
1609         return err;
1610
1611 finish:
1612         req->r_err = err;
1613         complete_request(mdsc, req);
1614         goto out;
1615 }
1616
1617 /*
1618  * called under mdsc->mutex
1619  */
1620 static void __wake_requests(struct ceph_mds_client *mdsc,
1621                             struct list_head *head)
1622 {
1623         struct ceph_mds_request *req, *nreq;
1624
1625         list_for_each_entry_safe(req, nreq, head, r_wait) {
1626                 list_del_init(&req->r_wait);
1627                 __do_request(mdsc, req);
1628         }
1629 }
1630
1631 /*
1632  * Wake up threads with requests pending for @mds, so that they can
1633  * resubmit their requests to a possibly different mds.  If @all is set,
1634  * wake up if their requests has been forwarded to @mds, too.
1635  */
1636 static void kick_requests(struct ceph_mds_client *mdsc, int mds, int all)
1637 {
1638         struct ceph_mds_request *req;
1639         struct rb_node *p;
1640
1641         dout("kick_requests mds%d\n", mds);
1642         for (p = rb_first(&mdsc->request_tree); p; p = rb_next(p)) {
1643                 req = rb_entry(p, struct ceph_mds_request, r_node);
1644                 if (req->r_got_unsafe)
1645                         continue;
1646                 if (req->r_session &&
1647                     req->r_session->s_mds == mds) {
1648                         dout(" kicking tid %llu\n", req->r_tid);
1649                         put_request_session(req);
1650                         __do_request(mdsc, req);
1651                 }
1652         }
1653 }
1654
1655 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
1656                               struct ceph_mds_request *req)
1657 {
1658         dout("submit_request on %p\n", req);
1659         mutex_lock(&mdsc->mutex);
1660         __register_request(mdsc, req, NULL);
1661         __do_request(mdsc, req);
1662         mutex_unlock(&mdsc->mutex);
1663 }
1664
1665 /*
1666  * Synchrously perform an mds request.  Take care of all of the
1667  * session setup, forwarding, retry details.
1668  */
1669 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
1670                          struct inode *dir,
1671                          struct ceph_mds_request *req)
1672 {
1673         int err;
1674
1675         dout("do_request on %p\n", req);
1676
1677         /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
1678         if (req->r_inode)
1679                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
1680         if (req->r_locked_dir)
1681                 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
1682         if (req->r_old_dentry)
1683                 ceph_get_cap_refs(
1684                         ceph_inode(req->r_old_dentry->d_parent->d_inode),
1685                         CEPH_CAP_PIN);
1686
1687         /* issue */
1688         mutex_lock(&mdsc->mutex);
1689         __register_request(mdsc, req, dir);
1690         __do_request(mdsc, req);
1691
1692         if (req->r_err) {
1693                 err = req->r_err;
1694                 __unregister_request(mdsc, req);
1695                 dout("do_request early error %d\n", err);
1696                 goto out;
1697         }
1698
1699         /* wait */
1700         mutex_unlock(&mdsc->mutex);
1701         dout("do_request waiting\n");
1702         if (req->r_timeout) {
1703                 err = (long)wait_for_completion_interruptible_timeout(
1704                         &req->r_completion, req->r_timeout);
1705                 if (err == 0)
1706                         err = -EIO;
1707         } else {
1708                 err = wait_for_completion_interruptible(&req->r_completion);
1709         }
1710         dout("do_request waited, got %d\n", err);
1711         mutex_lock(&mdsc->mutex);
1712
1713         /* only abort if we didn't race with a real reply */
1714         if (req->r_got_result) {
1715                 err = le32_to_cpu(req->r_reply_info.head->result);
1716         } else if (err < 0) {
1717                 dout("aborted request %lld with %d\n", req->r_tid, err);
1718                 req->r_err = err;
1719                 req->r_aborted = true;
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\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         } else {
1734                 err = req->r_err;
1735         }
1736
1737 out:
1738         mutex_unlock(&mdsc->mutex);
1739         dout("do_request %p done, result %d\n", req, err);
1740         return err;
1741 }
1742
1743 /*
1744  * Handle mds reply.
1745  *
1746  * We take the session mutex and parse and process the reply immediately.
1747  * This preserves the logical ordering of replies, capabilities, etc., sent
1748  * by the MDS as they are applied to our local cache.
1749  */
1750 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
1751 {
1752         struct ceph_mds_client *mdsc = session->s_mdsc;
1753         struct ceph_mds_request *req;
1754         struct ceph_mds_reply_head *head = msg->front.iov_base;
1755         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
1756         u64 tid;
1757         int err, result;
1758         int mds = session->s_mds;
1759
1760         if (msg->front.iov_len < sizeof(*head)) {
1761                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
1762                 ceph_msg_dump(msg);
1763                 return;
1764         }
1765
1766         /* get request, session */
1767         tid = le64_to_cpu(msg->hdr.tid);
1768         mutex_lock(&mdsc->mutex);
1769         req = __lookup_request(mdsc, tid);
1770         if (!req) {
1771                 dout("handle_reply on unknown tid %llu\n", tid);
1772                 mutex_unlock(&mdsc->mutex);
1773                 return;
1774         }
1775         dout("handle_reply %p\n", req);
1776
1777         /* correct session? */
1778         if (req->r_session != session) {
1779                 pr_err("mdsc_handle_reply got %llu on session mds%d"
1780                        " not mds%d\n", tid, session->s_mds,
1781                        req->r_session ? req->r_session->s_mds : -1);
1782                 mutex_unlock(&mdsc->mutex);
1783                 goto out;
1784         }
1785
1786         /* dup? */
1787         if ((req->r_got_unsafe && !head->safe) ||
1788             (req->r_got_safe && head->safe)) {
1789                 pr_warning("got a dup %s reply on %llu from mds%d\n",
1790                            head->safe ? "safe" : "unsafe", tid, mds);
1791                 mutex_unlock(&mdsc->mutex);
1792                 goto out;
1793         }
1794
1795         result = le32_to_cpu(head->result);
1796
1797         /*
1798          * Tolerate 2 consecutive ESTALEs from the same mds.
1799          * FIXME: we should be looking at the cap migrate_seq.
1800          */
1801         if (result == -ESTALE) {
1802                 req->r_direct_mode = USE_AUTH_MDS;
1803                 req->r_num_stale++;
1804                 if (req->r_num_stale <= 2) {
1805                         __do_request(mdsc, req);
1806                         mutex_unlock(&mdsc->mutex);
1807                         goto out;
1808                 }
1809         } else {
1810                 req->r_num_stale = 0;
1811         }
1812
1813         if (head->safe) {
1814                 req->r_got_safe = true;
1815                 __unregister_request(mdsc, req);
1816                 complete(&req->r_safe_completion);
1817
1818                 if (req->r_got_unsafe) {
1819                         /*
1820                          * We already handled the unsafe response, now do the
1821                          * cleanup.  No need to examine the response; the MDS
1822                          * doesn't include any result info in the safe
1823                          * response.  And even if it did, there is nothing
1824                          * useful we could do with a revised return value.
1825                          */
1826                         dout("got safe reply %llu, mds%d\n", tid, mds);
1827                         list_del_init(&req->r_unsafe_item);
1828
1829                         /* last unsafe request during umount? */
1830                         if (mdsc->stopping && !__get_oldest_req(mdsc))
1831                                 complete(&mdsc->safe_umount_waiters);
1832                         mutex_unlock(&mdsc->mutex);
1833                         goto out;
1834                 }
1835         } else {
1836                 req->r_got_unsafe = true;
1837                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
1838         }
1839
1840         dout("handle_reply tid %lld result %d\n", tid, result);
1841         rinfo = &req->r_reply_info;
1842         err = parse_reply_info(msg, rinfo);
1843         mutex_unlock(&mdsc->mutex);
1844
1845         mutex_lock(&session->s_mutex);
1846         if (err < 0) {
1847                 pr_err("mdsc_handle_reply got corrupt reply mds%d\n", mds);
1848                 ceph_msg_dump(msg);
1849                 goto out_err;
1850         }
1851
1852         /* snap trace */
1853         if (rinfo->snapblob_len) {
1854                 down_write(&mdsc->snap_rwsem);
1855                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
1856                                rinfo->snapblob + rinfo->snapblob_len,
1857                                le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
1858                 downgrade_write(&mdsc->snap_rwsem);
1859         } else {
1860                 down_read(&mdsc->snap_rwsem);
1861         }
1862
1863         /* insert trace into our cache */
1864         err = ceph_fill_trace(mdsc->client->sb, req, req->r_session);
1865         if (err == 0) {
1866                 if (result == 0 && rinfo->dir_nr)
1867                         ceph_readdir_prepopulate(req, req->r_session);
1868                 ceph_unreserve_caps(&req->r_caps_reservation);
1869         }
1870
1871         up_read(&mdsc->snap_rwsem);
1872 out_err:
1873         mutex_lock(&mdsc->mutex);
1874         if (!req->r_aborted) {
1875                 if (err) {
1876                         req->r_err = err;
1877                 } else {
1878                         req->r_reply = msg;
1879                         ceph_msg_get(msg);
1880                         req->r_got_result = true;
1881                 }
1882         } else {
1883                 dout("reply arrived after request %lld was aborted\n", tid);
1884         }
1885         mutex_unlock(&mdsc->mutex);
1886
1887         add_cap_releases(mdsc, req->r_session, -1);
1888         mutex_unlock(&session->s_mutex);
1889
1890         /* kick calling process */
1891         complete_request(mdsc, req);
1892 out:
1893         ceph_mdsc_put_request(req);
1894         return;
1895 }
1896
1897
1898
1899 /*
1900  * handle mds notification that our request has been forwarded.
1901  */
1902 static void handle_forward(struct ceph_mds_client *mdsc,
1903                            struct ceph_mds_session *session,
1904                            struct ceph_msg *msg)
1905 {
1906         struct ceph_mds_request *req;
1907         u64 tid = le64_to_cpu(msg->hdr.tid);
1908         u32 next_mds;
1909         u32 fwd_seq;
1910         int err = -EINVAL;
1911         void *p = msg->front.iov_base;
1912         void *end = p + msg->front.iov_len;
1913
1914         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1915         next_mds = ceph_decode_32(&p);
1916         fwd_seq = ceph_decode_32(&p);
1917
1918         mutex_lock(&mdsc->mutex);
1919         req = __lookup_request(mdsc, tid);
1920         if (!req) {
1921                 dout("forward %llu to mds%d - req dne\n", tid, next_mds);
1922                 goto out;  /* dup reply? */
1923         }
1924
1925         if (fwd_seq <= req->r_num_fwd) {
1926                 dout("forward %llu to mds%d - old seq %d <= %d\n",
1927                      tid, next_mds, req->r_num_fwd, fwd_seq);
1928         } else {
1929                 /* resend. forward race not possible; mds would drop */
1930                 dout("forward %llu to mds%d (we resend)\n", tid, next_mds);
1931                 req->r_num_fwd = fwd_seq;
1932                 req->r_resend_mds = next_mds;
1933                 put_request_session(req);
1934                 __do_request(mdsc, req);
1935         }
1936         ceph_mdsc_put_request(req);
1937 out:
1938         mutex_unlock(&mdsc->mutex);
1939         return;
1940
1941 bad:
1942         pr_err("mdsc_handle_forward decode error err=%d\n", err);
1943 }
1944
1945 /*
1946  * handle a mds session control message
1947  */
1948 static void handle_session(struct ceph_mds_session *session,
1949                            struct ceph_msg *msg)
1950 {
1951         struct ceph_mds_client *mdsc = session->s_mdsc;
1952         u32 op;
1953         u64 seq;
1954         int mds = session->s_mds;
1955         struct ceph_mds_session_head *h = msg->front.iov_base;
1956         int wake = 0;
1957
1958         /* decode */
1959         if (msg->front.iov_len != sizeof(*h))
1960                 goto bad;
1961         op = le32_to_cpu(h->op);
1962         seq = le64_to_cpu(h->seq);
1963
1964         mutex_lock(&mdsc->mutex);
1965         if (op == CEPH_SESSION_CLOSE)
1966                 __unregister_session(mdsc, session);
1967         /* FIXME: this ttl calculation is generous */
1968         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
1969         mutex_unlock(&mdsc->mutex);
1970
1971         mutex_lock(&session->s_mutex);
1972
1973         dout("handle_session mds%d %s %p state %s seq %llu\n",
1974              mds, ceph_session_op_name(op), session,
1975              session_state_name(session->s_state), seq);
1976
1977         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
1978                 session->s_state = CEPH_MDS_SESSION_OPEN;
1979                 pr_info("mds%d came back\n", session->s_mds);
1980         }
1981
1982         switch (op) {
1983         case CEPH_SESSION_OPEN:
1984                 session->s_state = CEPH_MDS_SESSION_OPEN;
1985                 renewed_caps(mdsc, session, 0);
1986                 wake = 1;
1987                 if (mdsc->stopping)
1988                         __close_session(mdsc, session);
1989                 break;
1990
1991         case CEPH_SESSION_RENEWCAPS:
1992                 if (session->s_renew_seq == seq)
1993                         renewed_caps(mdsc, session, 1);
1994                 break;
1995
1996         case CEPH_SESSION_CLOSE:
1997                 remove_session_caps(session);
1998                 wake = 1; /* for good measure */
1999                 complete(&mdsc->session_close_waiters);
2000                 kick_requests(mdsc, mds, 0);      /* cur only */
2001                 break;
2002
2003         case CEPH_SESSION_STALE:
2004                 pr_info("mds%d caps went stale, renewing\n",
2005                         session->s_mds);
2006                 spin_lock(&session->s_cap_lock);
2007                 session->s_cap_gen++;
2008                 session->s_cap_ttl = 0;
2009                 spin_unlock(&session->s_cap_lock);
2010                 send_renew_caps(mdsc, session);
2011                 break;
2012
2013         case CEPH_SESSION_RECALL_STATE:
2014                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2015                 break;
2016
2017         default:
2018                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2019                 WARN_ON(1);
2020         }
2021
2022         mutex_unlock(&session->s_mutex);
2023         if (wake) {
2024                 mutex_lock(&mdsc->mutex);
2025                 __wake_requests(mdsc, &session->s_waiting);
2026                 mutex_unlock(&mdsc->mutex);
2027         }
2028         return;
2029
2030 bad:
2031         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2032                (int)msg->front.iov_len);
2033         ceph_msg_dump(msg);
2034         return;
2035 }
2036
2037
2038 /*
2039  * called under session->mutex.
2040  */
2041 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2042                                    struct ceph_mds_session *session)
2043 {
2044         struct ceph_mds_request *req, *nreq;
2045         int err;
2046
2047         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2048
2049         mutex_lock(&mdsc->mutex);
2050         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2051                 err = __prepare_send_request(mdsc, req, session->s_mds);
2052                 if (!err) {
2053                         ceph_msg_get(req->r_request);
2054                         ceph_con_send(&session->s_con, req->r_request);
2055                 }
2056         }
2057         mutex_unlock(&mdsc->mutex);
2058 }
2059
2060 /*
2061  * Encode information about a cap for a reconnect with the MDS.
2062  */
2063 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2064                           void *arg)
2065 {
2066         struct ceph_mds_cap_reconnect rec;
2067         struct ceph_inode_info *ci;
2068         struct ceph_pagelist *pagelist = arg;
2069         char *path;
2070         int pathlen, err;
2071         u64 pathbase;
2072         struct dentry *dentry;
2073
2074         ci = cap->ci;
2075
2076         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2077              inode, ceph_vinop(inode), cap, cap->cap_id,
2078              ceph_cap_string(cap->issued));
2079         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2080         if (err)
2081                 return err;
2082
2083         dentry = d_find_alias(inode);
2084         if (dentry) {
2085                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2086                 if (IS_ERR(path)) {
2087                         err = PTR_ERR(path);
2088                         BUG_ON(err);
2089                 }
2090         } else {
2091                 path = NULL;
2092                 pathlen = 0;
2093         }
2094         err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2095         if (err)
2096                 goto out;
2097
2098         spin_lock(&inode->i_lock);
2099         cap->seq = 0;        /* reset cap seq */
2100         cap->issue_seq = 0;  /* and issue_seq */
2101         rec.cap_id = cpu_to_le64(cap->cap_id);
2102         rec.pathbase = cpu_to_le64(pathbase);
2103         rec.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2104         rec.issued = cpu_to_le32(cap->issued);
2105         rec.size = cpu_to_le64(inode->i_size);
2106         ceph_encode_timespec(&rec.mtime, &inode->i_mtime);
2107         ceph_encode_timespec(&rec.atime, &inode->i_atime);
2108         rec.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2109         spin_unlock(&inode->i_lock);
2110
2111         err = ceph_pagelist_append(pagelist, &rec, sizeof(rec));
2112
2113 out:
2114         kfree(path);
2115         dput(dentry);
2116         return err;
2117 }
2118
2119
2120 /*
2121  * If an MDS fails and recovers, clients need to reconnect in order to
2122  * reestablish shared state.  This includes all caps issued through
2123  * this session _and_ the snap_realm hierarchy.  Because it's not
2124  * clear which snap realms the mds cares about, we send everything we
2125  * know about.. that ensures we'll then get any new info the
2126  * recovering MDS might have.
2127  *
2128  * This is a relatively heavyweight operation, but it's rare.
2129  *
2130  * called with mdsc->mutex held.
2131  */
2132 static void send_mds_reconnect(struct ceph_mds_client *mdsc, int mds)
2133 {
2134         struct ceph_mds_session *session = NULL;
2135         struct ceph_msg *reply;
2136         struct rb_node *p;
2137         int err = -ENOMEM;
2138         struct ceph_pagelist *pagelist;
2139
2140         pr_info("reconnect to recovering mds%d\n", mds);
2141
2142         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2143         if (!pagelist)
2144                 goto fail_nopagelist;
2145         ceph_pagelist_init(pagelist);
2146
2147         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, 0, 0, NULL);
2148         if (IS_ERR(reply)) {
2149                 err = PTR_ERR(reply);
2150                 goto fail_nomsg;
2151         }
2152
2153         /* find session */
2154         session = __ceph_lookup_mds_session(mdsc, mds);
2155         mutex_unlock(&mdsc->mutex);    /* drop lock for duration */
2156
2157         if (session) {
2158                 mutex_lock(&session->s_mutex);
2159
2160                 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2161                 session->s_seq = 0;
2162
2163                 ceph_con_open(&session->s_con,
2164                               ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2165
2166                 /* replay unsafe requests */
2167                 replay_unsafe_requests(mdsc, session);
2168         } else {
2169                 dout("no session for mds%d, will send short reconnect\n",
2170                      mds);
2171         }
2172
2173         down_read(&mdsc->snap_rwsem);
2174
2175         if (!session)
2176                 goto send;
2177         dout("session %p state %s\n", session,
2178              session_state_name(session->s_state));
2179
2180         /* traverse this session's caps */
2181         err = ceph_pagelist_encode_32(pagelist, session->s_nr_caps);
2182         if (err)
2183                 goto fail;
2184         err = iterate_session_caps(session, encode_caps_cb, pagelist);
2185         if (err < 0)
2186                 goto fail;
2187
2188         /*
2189          * snaprealms.  we provide mds with the ino, seq (version), and
2190          * parent for all of our realms.  If the mds has any newer info,
2191          * it will tell us.
2192          */
2193         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2194                 struct ceph_snap_realm *realm =
2195                         rb_entry(p, struct ceph_snap_realm, node);
2196                 struct ceph_mds_snaprealm_reconnect sr_rec;
2197
2198                 dout(" adding snap realm %llx seq %lld parent %llx\n",
2199                      realm->ino, realm->seq, realm->parent_ino);
2200                 sr_rec.ino = cpu_to_le64(realm->ino);
2201                 sr_rec.seq = cpu_to_le64(realm->seq);
2202                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2203                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2204                 if (err)
2205                         goto fail;
2206         }
2207
2208 send:
2209         reply->pagelist = pagelist;
2210         reply->hdr.data_len = cpu_to_le32(pagelist->length);
2211         reply->nr_pages = calc_pages_for(0, pagelist->length);
2212         ceph_con_send(&session->s_con, reply);
2213
2214         session->s_state = CEPH_MDS_SESSION_OPEN;
2215         mutex_unlock(&session->s_mutex);
2216
2217         mutex_lock(&mdsc->mutex);
2218         __wake_requests(mdsc, &session->s_waiting);
2219         mutex_unlock(&mdsc->mutex);
2220
2221         ceph_put_mds_session(session);
2222
2223         up_read(&mdsc->snap_rwsem);
2224         mutex_lock(&mdsc->mutex);
2225         return;
2226
2227 fail:
2228         ceph_msg_put(reply);
2229         up_read(&mdsc->snap_rwsem);
2230         mutex_unlock(&session->s_mutex);
2231         ceph_put_mds_session(session);
2232 fail_nomsg:
2233         ceph_pagelist_release(pagelist);
2234         kfree(pagelist);
2235 fail_nopagelist:
2236         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
2237         mutex_lock(&mdsc->mutex);
2238         return;
2239 }
2240
2241
2242 /*
2243  * compare old and new mdsmaps, kicking requests
2244  * and closing out old connections as necessary
2245  *
2246  * called under mdsc->mutex.
2247  */
2248 static void check_new_map(struct ceph_mds_client *mdsc,
2249                           struct ceph_mdsmap *newmap,
2250                           struct ceph_mdsmap *oldmap)
2251 {
2252         int i;
2253         int oldstate, newstate;
2254         struct ceph_mds_session *s;
2255
2256         dout("check_new_map new %u old %u\n",
2257              newmap->m_epoch, oldmap->m_epoch);
2258
2259         for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2260                 if (mdsc->sessions[i] == NULL)
2261                         continue;
2262                 s = mdsc->sessions[i];
2263                 oldstate = ceph_mdsmap_get_state(oldmap, i);
2264                 newstate = ceph_mdsmap_get_state(newmap, i);
2265
2266                 dout("check_new_map mds%d state %s -> %s (session %s)\n",
2267                      i, ceph_mds_state_name(oldstate),
2268                      ceph_mds_state_name(newstate),
2269                      session_state_name(s->s_state));
2270
2271                 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
2272                            ceph_mdsmap_get_addr(newmap, i),
2273                            sizeof(struct ceph_entity_addr))) {
2274                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2275                                 /* the session never opened, just close it
2276                                  * out now */
2277                                 __wake_requests(mdsc, &s->s_waiting);
2278                                 __unregister_session(mdsc, s);
2279                         } else {
2280                                 /* just close it */
2281                                 mutex_unlock(&mdsc->mutex);
2282                                 mutex_lock(&s->s_mutex);
2283                                 mutex_lock(&mdsc->mutex);
2284                                 ceph_con_close(&s->s_con);
2285                                 mutex_unlock(&s->s_mutex);
2286                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2287                         }
2288
2289                         /* kick any requests waiting on the recovering mds */
2290                         kick_requests(mdsc, i, 1);
2291                 } else if (oldstate == newstate) {
2292                         continue;  /* nothing new with this mds */
2293                 }
2294
2295                 /*
2296                  * send reconnect?
2297                  */
2298                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2299                     newstate >= CEPH_MDS_STATE_RECONNECT)
2300                         send_mds_reconnect(mdsc, i);
2301
2302                 /*
2303                  * kick requests on any mds that has gone active.
2304                  *
2305                  * kick requests on cur or forwarder: we may have sent
2306                  * the request to mds1, mds1 told us it forwarded it
2307                  * to mds2, but then we learn mds1 failed and can't be
2308                  * sure it successfully forwarded our request before
2309                  * it died.
2310                  */
2311                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2312                     newstate >= CEPH_MDS_STATE_ACTIVE) {
2313                         pr_info("mds%d reconnect completed\n", s->s_mds);
2314                         kick_requests(mdsc, i, 1);
2315                         ceph_kick_flushing_caps(mdsc, s);
2316                         wake_up_session_caps(s, 1);
2317                 }
2318         }
2319 }
2320
2321
2322
2323 /*
2324  * leases
2325  */
2326
2327 /*
2328  * caller must hold session s_mutex, dentry->d_lock
2329  */
2330 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2331 {
2332         struct ceph_dentry_info *di = ceph_dentry(dentry);
2333
2334         ceph_put_mds_session(di->lease_session);
2335         di->lease_session = NULL;
2336 }
2337
2338 static void handle_lease(struct ceph_mds_client *mdsc,
2339                          struct ceph_mds_session *session,
2340                          struct ceph_msg *msg)
2341 {
2342         struct super_block *sb = mdsc->client->sb;
2343         struct inode *inode;
2344         struct ceph_inode_info *ci;
2345         struct dentry *parent, *dentry;
2346         struct ceph_dentry_info *di;
2347         int mds = session->s_mds;
2348         struct ceph_mds_lease *h = msg->front.iov_base;
2349         struct ceph_vino vino;
2350         int mask;
2351         struct qstr dname;
2352         int release = 0;
2353
2354         dout("handle_lease from mds%d\n", mds);
2355
2356         /* decode */
2357         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2358                 goto bad;
2359         vino.ino = le64_to_cpu(h->ino);
2360         vino.snap = CEPH_NOSNAP;
2361         mask = le16_to_cpu(h->mask);
2362         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2363         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2364         if (dname.len != get_unaligned_le32(h+1))
2365                 goto bad;
2366
2367         mutex_lock(&session->s_mutex);
2368         session->s_seq++;
2369
2370         /* lookup inode */
2371         inode = ceph_find_inode(sb, vino);
2372         dout("handle_lease '%s', mask %d, ino %llx %p\n",
2373              ceph_lease_op_name(h->action), mask, vino.ino, inode);
2374         if (inode == NULL) {
2375                 dout("handle_lease no inode %llx\n", vino.ino);
2376                 goto release;
2377         }
2378         ci = ceph_inode(inode);
2379
2380         /* dentry */
2381         parent = d_find_alias(inode);
2382         if (!parent) {
2383                 dout("no parent dentry on inode %p\n", inode);
2384                 WARN_ON(1);
2385                 goto release;  /* hrm... */
2386         }
2387         dname.hash = full_name_hash(dname.name, dname.len);
2388         dentry = d_lookup(parent, &dname);
2389         dput(parent);
2390         if (!dentry)
2391                 goto release;
2392
2393         spin_lock(&dentry->d_lock);
2394         di = ceph_dentry(dentry);
2395         switch (h->action) {
2396         case CEPH_MDS_LEASE_REVOKE:
2397                 if (di && di->lease_session == session) {
2398                         h->seq = cpu_to_le32(di->lease_seq);
2399                         __ceph_mdsc_drop_dentry_lease(dentry);
2400                 }
2401                 release = 1;
2402                 break;
2403
2404         case CEPH_MDS_LEASE_RENEW:
2405                 if (di && di->lease_session == session &&
2406                     di->lease_gen == session->s_cap_gen &&
2407                     di->lease_renew_from &&
2408                     di->lease_renew_after == 0) {
2409                         unsigned long duration =
2410                                 le32_to_cpu(h->duration_ms) * HZ / 1000;
2411
2412                         di->lease_seq = le32_to_cpu(h->seq);
2413                         dentry->d_time = di->lease_renew_from + duration;
2414                         di->lease_renew_after = di->lease_renew_from +
2415                                 (duration >> 1);
2416                         di->lease_renew_from = 0;
2417                 }
2418                 break;
2419         }
2420         spin_unlock(&dentry->d_lock);
2421         dput(dentry);
2422
2423         if (!release)
2424                 goto out;
2425
2426 release:
2427         /* let's just reuse the same message */
2428         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2429         ceph_msg_get(msg);
2430         ceph_con_send(&session->s_con, msg);
2431
2432 out:
2433         iput(inode);
2434         mutex_unlock(&session->s_mutex);
2435         return;
2436
2437 bad:
2438         pr_err("corrupt lease message\n");
2439         ceph_msg_dump(msg);
2440 }
2441
2442 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
2443                               struct inode *inode,
2444                               struct dentry *dentry, char action,
2445                               u32 seq)
2446 {
2447         struct ceph_msg *msg;
2448         struct ceph_mds_lease *lease;
2449         int len = sizeof(*lease) + sizeof(u32);
2450         int dnamelen = 0;
2451
2452         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2453              inode, dentry, ceph_lease_op_name(action), session->s_mds);
2454         dnamelen = dentry->d_name.len;
2455         len += dnamelen;
2456
2457         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, 0, 0, NULL);
2458         if (IS_ERR(msg))
2459                 return;
2460         lease = msg->front.iov_base;
2461         lease->action = action;
2462         lease->mask = cpu_to_le16(CEPH_LOCK_DN);
2463         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
2464         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
2465         lease->seq = cpu_to_le32(seq);
2466         put_unaligned_le32(dnamelen, lease + 1);
2467         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
2468
2469         /*
2470          * if this is a preemptive lease RELEASE, no need to
2471          * flush request stream, since the actual request will
2472          * soon follow.
2473          */
2474         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
2475
2476         ceph_con_send(&session->s_con, msg);
2477 }
2478
2479 /*
2480  * Preemptively release a lease we expect to invalidate anyway.
2481  * Pass @inode always, @dentry is optional.
2482  */
2483 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2484                              struct dentry *dentry, int mask)
2485 {
2486         struct ceph_dentry_info *di;
2487         struct ceph_mds_session *session;
2488         u32 seq;
2489
2490         BUG_ON(inode == NULL);
2491         BUG_ON(dentry == NULL);
2492         BUG_ON(mask != CEPH_LOCK_DN);
2493
2494         /* is dentry lease valid? */
2495         spin_lock(&dentry->d_lock);
2496         di = ceph_dentry(dentry);
2497         if (!di || !di->lease_session ||
2498             di->lease_session->s_mds < 0 ||
2499             di->lease_gen != di->lease_session->s_cap_gen ||
2500             !time_before(jiffies, dentry->d_time)) {
2501                 dout("lease_release inode %p dentry %p -- "
2502                      "no lease on %d\n",
2503                      inode, dentry, mask);
2504                 spin_unlock(&dentry->d_lock);
2505                 return;
2506         }
2507
2508         /* we do have a lease on this dentry; note mds and seq */
2509         session = ceph_get_mds_session(di->lease_session);
2510         seq = di->lease_seq;
2511         __ceph_mdsc_drop_dentry_lease(dentry);
2512         spin_unlock(&dentry->d_lock);
2513
2514         dout("lease_release inode %p dentry %p mask %d to mds%d\n",
2515              inode, dentry, mask, session->s_mds);
2516         ceph_mdsc_lease_send_msg(session, inode, dentry,
2517                                  CEPH_MDS_LEASE_RELEASE, seq);
2518         ceph_put_mds_session(session);
2519 }
2520
2521 /*
2522  * drop all leases (and dentry refs) in preparation for umount
2523  */
2524 static void drop_leases(struct ceph_mds_client *mdsc)
2525 {
2526         int i;
2527
2528         dout("drop_leases\n");
2529         mutex_lock(&mdsc->mutex);
2530         for (i = 0; i < mdsc->max_sessions; i++) {
2531                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2532                 if (!s)
2533                         continue;
2534                 mutex_unlock(&mdsc->mutex);
2535                 mutex_lock(&s->s_mutex);
2536                 mutex_unlock(&s->s_mutex);
2537                 ceph_put_mds_session(s);
2538                 mutex_lock(&mdsc->mutex);
2539         }
2540         mutex_unlock(&mdsc->mutex);
2541 }
2542
2543
2544
2545 /*
2546  * delayed work -- periodically trim expired leases, renew caps with mds
2547  */
2548 static void schedule_delayed(struct ceph_mds_client *mdsc)
2549 {
2550         int delay = 5;
2551         unsigned hz = round_jiffies_relative(HZ * delay);
2552         schedule_delayed_work(&mdsc->delayed_work, hz);
2553 }
2554
2555 static void delayed_work(struct work_struct *work)
2556 {
2557         int i;
2558         struct ceph_mds_client *mdsc =
2559                 container_of(work, struct ceph_mds_client, delayed_work.work);
2560         int renew_interval;
2561         int renew_caps;
2562
2563         dout("mdsc delayed_work\n");
2564         ceph_check_delayed_caps(mdsc);
2565
2566         mutex_lock(&mdsc->mutex);
2567         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
2568         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
2569                                    mdsc->last_renew_caps);
2570         if (renew_caps)
2571                 mdsc->last_renew_caps = jiffies;
2572
2573         for (i = 0; i < mdsc->max_sessions; i++) {
2574                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2575                 if (s == NULL)
2576                         continue;
2577                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
2578                         dout("resending session close request for mds%d\n",
2579                              s->s_mds);
2580                         request_close_session(mdsc, s);
2581                         ceph_put_mds_session(s);
2582                         continue;
2583                 }
2584                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
2585                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
2586                                 s->s_state = CEPH_MDS_SESSION_HUNG;
2587                                 pr_info("mds%d hung\n", s->s_mds);
2588                         }
2589                 }
2590                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
2591                         /* this mds is failed or recovering, just wait */
2592                         ceph_put_mds_session(s);
2593                         continue;
2594                 }
2595                 mutex_unlock(&mdsc->mutex);
2596
2597                 mutex_lock(&s->s_mutex);
2598                 if (renew_caps)
2599                         send_renew_caps(mdsc, s);
2600                 else
2601                         ceph_con_keepalive(&s->s_con);
2602                 add_cap_releases(mdsc, s, -1);
2603                 send_cap_releases(mdsc, s);
2604                 mutex_unlock(&s->s_mutex);
2605                 ceph_put_mds_session(s);
2606
2607                 mutex_lock(&mdsc->mutex);
2608         }
2609         mutex_unlock(&mdsc->mutex);
2610
2611         schedule_delayed(mdsc);
2612 }
2613
2614
2615 int ceph_mdsc_init(struct ceph_mds_client *mdsc, struct ceph_client *client)
2616 {
2617         mdsc->client = client;
2618         mutex_init(&mdsc->mutex);
2619         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
2620         init_completion(&mdsc->safe_umount_waiters);
2621         init_completion(&mdsc->session_close_waiters);
2622         INIT_LIST_HEAD(&mdsc->waiting_for_map);
2623         mdsc->sessions = NULL;
2624         mdsc->max_sessions = 0;
2625         mdsc->stopping = 0;
2626         init_rwsem(&mdsc->snap_rwsem);
2627         mdsc->snap_realms = RB_ROOT;
2628         INIT_LIST_HEAD(&mdsc->snap_empty);
2629         spin_lock_init(&mdsc->snap_empty_lock);
2630         mdsc->last_tid = 0;
2631         mdsc->request_tree = RB_ROOT;
2632         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
2633         mdsc->last_renew_caps = jiffies;
2634         INIT_LIST_HEAD(&mdsc->cap_delay_list);
2635         spin_lock_init(&mdsc->cap_delay_lock);
2636         INIT_LIST_HEAD(&mdsc->snap_flush_list);
2637         spin_lock_init(&mdsc->snap_flush_lock);
2638         mdsc->cap_flush_seq = 0;
2639         INIT_LIST_HEAD(&mdsc->cap_dirty);
2640         mdsc->num_cap_flushing = 0;
2641         spin_lock_init(&mdsc->cap_dirty_lock);
2642         init_waitqueue_head(&mdsc->cap_flushing_wq);
2643         spin_lock_init(&mdsc->dentry_lru_lock);
2644         INIT_LIST_HEAD(&mdsc->dentry_lru);
2645         return 0;
2646 }
2647
2648 /*
2649  * Wait for safe replies on open mds requests.  If we time out, drop
2650  * all requests from the tree to avoid dangling dentry refs.
2651  */
2652 static void wait_requests(struct ceph_mds_client *mdsc)
2653 {
2654         struct ceph_mds_request *req;
2655         struct ceph_client *client = mdsc->client;
2656
2657         mutex_lock(&mdsc->mutex);
2658         if (__get_oldest_req(mdsc)) {
2659                 mutex_unlock(&mdsc->mutex);
2660
2661                 dout("wait_requests waiting for requests\n");
2662                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
2663                                     client->mount_args->mount_timeout * HZ);
2664
2665                 /* tear down remaining requests */
2666                 mutex_lock(&mdsc->mutex);
2667                 while ((req = __get_oldest_req(mdsc))) {
2668                         dout("wait_requests timed out on tid %llu\n",
2669                              req->r_tid);
2670                         __unregister_request(mdsc, req);
2671                 }
2672         }
2673         mutex_unlock(&mdsc->mutex);
2674         dout("wait_requests done\n");
2675 }
2676
2677 /*
2678  * called before mount is ro, and before dentries are torn down.
2679  * (hmm, does this still race with new lookups?)
2680  */
2681 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
2682 {
2683         dout("pre_umount\n");
2684         mdsc->stopping = 1;
2685
2686         drop_leases(mdsc);
2687         ceph_flush_dirty_caps(mdsc);
2688         wait_requests(mdsc);
2689 }
2690
2691 /*
2692  * wait for all write mds requests to flush.
2693  */
2694 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
2695 {
2696         struct ceph_mds_request *req = NULL, *nextreq;
2697         struct rb_node *n;
2698
2699         mutex_lock(&mdsc->mutex);
2700         dout("wait_unsafe_requests want %lld\n", want_tid);
2701 restart:
2702         req = __get_oldest_req(mdsc);
2703         while (req && req->r_tid <= want_tid) {
2704                 /* find next request */
2705                 n = rb_next(&req->r_node);
2706                 if (n)
2707                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
2708                 else
2709                         nextreq = NULL;
2710                 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
2711                         /* write op */
2712                         ceph_mdsc_get_request(req);
2713                         if (nextreq)
2714                                 ceph_mdsc_get_request(nextreq);
2715                         mutex_unlock(&mdsc->mutex);
2716                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
2717                              req->r_tid, want_tid);
2718                         wait_for_completion(&req->r_safe_completion);
2719                         mutex_lock(&mdsc->mutex);
2720                         ceph_mdsc_put_request(req);
2721                         if (!nextreq)
2722                                 break;  /* next dne before, so we're done! */
2723                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
2724                                 /* next request was removed from tree */
2725                                 ceph_mdsc_put_request(nextreq);
2726                                 goto restart;
2727                         }
2728                         ceph_mdsc_put_request(nextreq);  /* won't go away */
2729                 }
2730                 req = nextreq;
2731         }
2732         mutex_unlock(&mdsc->mutex);
2733         dout("wait_unsafe_requests done\n");
2734 }
2735
2736 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
2737 {
2738         u64 want_tid, want_flush;
2739
2740         dout("sync\n");
2741         mutex_lock(&mdsc->mutex);
2742         want_tid = mdsc->last_tid;
2743         want_flush = mdsc->cap_flush_seq;
2744         mutex_unlock(&mdsc->mutex);
2745         dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
2746
2747         ceph_flush_dirty_caps(mdsc);
2748
2749         wait_unsafe_requests(mdsc, want_tid);
2750         wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
2751 }
2752
2753
2754 /*
2755  * called after sb is ro.
2756  */
2757 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
2758 {
2759         struct ceph_mds_session *session;
2760         int i;
2761         int n;
2762         struct ceph_client *client = mdsc->client;
2763         unsigned long started, timeout = client->mount_args->mount_timeout * HZ;
2764
2765         dout("close_sessions\n");
2766
2767         mutex_lock(&mdsc->mutex);
2768
2769         /* close sessions */
2770         started = jiffies;
2771         while (time_before(jiffies, started + timeout)) {
2772                 dout("closing sessions\n");
2773                 n = 0;
2774                 for (i = 0; i < mdsc->max_sessions; i++) {
2775                         session = __ceph_lookup_mds_session(mdsc, i);
2776                         if (!session)
2777                                 continue;
2778                         mutex_unlock(&mdsc->mutex);
2779                         mutex_lock(&session->s_mutex);
2780                         __close_session(mdsc, session);
2781                         mutex_unlock(&session->s_mutex);
2782                         ceph_put_mds_session(session);
2783                         mutex_lock(&mdsc->mutex);
2784                         n++;
2785                 }
2786                 if (n == 0)
2787                         break;
2788
2789                 if (client->mount_state == CEPH_MOUNT_SHUTDOWN)
2790                         break;
2791
2792                 dout("waiting for sessions to close\n");
2793                 mutex_unlock(&mdsc->mutex);
2794                 wait_for_completion_timeout(&mdsc->session_close_waiters,
2795                                             timeout);
2796                 mutex_lock(&mdsc->mutex);
2797         }
2798
2799         /* tear down remaining sessions */
2800         for (i = 0; i < mdsc->max_sessions; i++) {
2801                 if (mdsc->sessions[i]) {
2802                         session = get_session(mdsc->sessions[i]);
2803                         __unregister_session(mdsc, session);
2804                         mutex_unlock(&mdsc->mutex);
2805                         mutex_lock(&session->s_mutex);
2806                         remove_session_caps(session);
2807                         mutex_unlock(&session->s_mutex);
2808                         ceph_put_mds_session(session);
2809                         mutex_lock(&mdsc->mutex);
2810                 }
2811         }
2812
2813         WARN_ON(!list_empty(&mdsc->cap_delay_list));
2814
2815         mutex_unlock(&mdsc->mutex);
2816
2817         ceph_cleanup_empty_realms(mdsc);
2818
2819         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2820
2821         dout("stopped\n");
2822 }
2823
2824 void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
2825 {
2826         dout("stop\n");
2827         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2828         if (mdsc->mdsmap)
2829                 ceph_mdsmap_destroy(mdsc->mdsmap);
2830         kfree(mdsc->sessions);
2831 }
2832
2833
2834 /*
2835  * handle mds map update.
2836  */
2837 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
2838 {
2839         u32 epoch;
2840         u32 maplen;
2841         void *p = msg->front.iov_base;
2842         void *end = p + msg->front.iov_len;
2843         struct ceph_mdsmap *newmap, *oldmap;
2844         struct ceph_fsid fsid;
2845         int err = -EINVAL;
2846
2847         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
2848         ceph_decode_copy(&p, &fsid, sizeof(fsid));
2849         if (ceph_check_fsid(mdsc->client, &fsid) < 0)
2850                 return;
2851         epoch = ceph_decode_32(&p);
2852         maplen = ceph_decode_32(&p);
2853         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
2854
2855         /* do we need it? */
2856         ceph_monc_got_mdsmap(&mdsc->client->monc, epoch);
2857         mutex_lock(&mdsc->mutex);
2858         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
2859                 dout("handle_map epoch %u <= our %u\n",
2860                      epoch, mdsc->mdsmap->m_epoch);
2861                 mutex_unlock(&mdsc->mutex);
2862                 return;
2863         }
2864
2865         newmap = ceph_mdsmap_decode(&p, end);
2866         if (IS_ERR(newmap)) {
2867                 err = PTR_ERR(newmap);
2868                 goto bad_unlock;
2869         }
2870
2871         /* swap into place */
2872         if (mdsc->mdsmap) {
2873                 oldmap = mdsc->mdsmap;
2874                 mdsc->mdsmap = newmap;
2875                 check_new_map(mdsc, newmap, oldmap);
2876                 ceph_mdsmap_destroy(oldmap);
2877         } else {
2878                 mdsc->mdsmap = newmap;  /* first mds map */
2879         }
2880         mdsc->client->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
2881
2882         __wake_requests(mdsc, &mdsc->waiting_for_map);
2883
2884         mutex_unlock(&mdsc->mutex);
2885         schedule_delayed(mdsc);
2886         return;
2887
2888 bad_unlock:
2889         mutex_unlock(&mdsc->mutex);
2890 bad:
2891         pr_err("error decoding mdsmap %d\n", err);
2892         return;
2893 }
2894
2895 static struct ceph_connection *con_get(struct ceph_connection *con)
2896 {
2897         struct ceph_mds_session *s = con->private;
2898
2899         if (get_session(s)) {
2900                 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
2901                 return con;
2902         }
2903         dout("mdsc con_get %p FAIL\n", s);
2904         return NULL;
2905 }
2906
2907 static void con_put(struct ceph_connection *con)
2908 {
2909         struct ceph_mds_session *s = con->private;
2910
2911         ceph_put_mds_session(s);
2912         dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref));
2913 }
2914
2915 /*
2916  * if the client is unresponsive for long enough, the mds will kill
2917  * the session entirely.
2918  */
2919 static void peer_reset(struct ceph_connection *con)
2920 {
2921         struct ceph_mds_session *s = con->private;
2922
2923         pr_err("mds%d gave us the boot.  IMPLEMENT RECONNECT.\n",
2924                s->s_mds);
2925 }
2926
2927 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2928 {
2929         struct ceph_mds_session *s = con->private;
2930         struct ceph_mds_client *mdsc = s->s_mdsc;
2931         int type = le16_to_cpu(msg->hdr.type);
2932
2933         mutex_lock(&mdsc->mutex);
2934         if (__verify_registered_session(mdsc, s) < 0) {
2935                 mutex_unlock(&mdsc->mutex);
2936                 goto out;
2937         }
2938         mutex_unlock(&mdsc->mutex);
2939
2940         switch (type) {
2941         case CEPH_MSG_MDS_MAP:
2942                 ceph_mdsc_handle_map(mdsc, msg);
2943                 break;
2944         case CEPH_MSG_CLIENT_SESSION:
2945                 handle_session(s, msg);
2946                 break;
2947         case CEPH_MSG_CLIENT_REPLY:
2948                 handle_reply(s, msg);
2949                 break;
2950         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
2951                 handle_forward(mdsc, s, msg);
2952                 break;
2953         case CEPH_MSG_CLIENT_CAPS:
2954                 ceph_handle_caps(s, msg);
2955                 break;
2956         case CEPH_MSG_CLIENT_SNAP:
2957                 ceph_handle_snap(mdsc, s, msg);
2958                 break;
2959         case CEPH_MSG_CLIENT_LEASE:
2960                 handle_lease(mdsc, s, msg);
2961                 break;
2962
2963         default:
2964                 pr_err("received unknown message type %d %s\n", type,
2965                        ceph_msg_type_name(type));
2966         }
2967 out:
2968         ceph_msg_put(msg);
2969 }
2970
2971 /*
2972  * authentication
2973  */
2974 static int get_authorizer(struct ceph_connection *con,
2975                           void **buf, int *len, int *proto,
2976                           void **reply_buf, int *reply_len, int force_new)
2977 {
2978         struct ceph_mds_session *s = con->private;
2979         struct ceph_mds_client *mdsc = s->s_mdsc;
2980         struct ceph_auth_client *ac = mdsc->client->monc.auth;
2981         int ret = 0;
2982
2983         if (force_new && s->s_authorizer) {
2984                 ac->ops->destroy_authorizer(ac, s->s_authorizer);
2985                 s->s_authorizer = NULL;
2986         }
2987         if (s->s_authorizer == NULL) {
2988                 if (ac->ops->create_authorizer) {
2989                         ret = ac->ops->create_authorizer(
2990                                 ac, CEPH_ENTITY_TYPE_MDS,
2991                                 &s->s_authorizer,
2992                                 &s->s_authorizer_buf,
2993                                 &s->s_authorizer_buf_len,
2994                                 &s->s_authorizer_reply_buf,
2995                                 &s->s_authorizer_reply_buf_len);
2996                         if (ret)
2997                                 return ret;
2998                 }
2999         }
3000
3001         *proto = ac->protocol;
3002         *buf = s->s_authorizer_buf;
3003         *len = s->s_authorizer_buf_len;
3004         *reply_buf = s->s_authorizer_reply_buf;
3005         *reply_len = s->s_authorizer_reply_buf_len;
3006         return 0;
3007 }
3008
3009
3010 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3011 {
3012         struct ceph_mds_session *s = con->private;
3013         struct ceph_mds_client *mdsc = s->s_mdsc;
3014         struct ceph_auth_client *ac = mdsc->client->monc.auth;
3015
3016         return ac->ops->verify_authorizer_reply(ac, s->s_authorizer, len);
3017 }
3018
3019 static int invalidate_authorizer(struct ceph_connection *con)
3020 {
3021         struct ceph_mds_session *s = con->private;
3022         struct ceph_mds_client *mdsc = s->s_mdsc;
3023         struct ceph_auth_client *ac = mdsc->client->monc.auth;
3024
3025         if (ac->ops->invalidate_authorizer)
3026                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3027
3028         return ceph_monc_validate_auth(&mdsc->client->monc);
3029 }
3030
3031 const static struct ceph_connection_operations mds_con_ops = {
3032         .get = con_get,
3033         .put = con_put,
3034         .dispatch = dispatch,
3035         .get_authorizer = get_authorizer,
3036         .verify_authorizer_reply = verify_authorizer_reply,
3037         .invalidate_authorizer = invalidate_authorizer,
3038         .peer_reset = peer_reset,
3039 };
3040
3041
3042
3043
3044 /* eof */