fs/ceph/mds_client.c

   1 #include "ceph_debug.h"
   2
   3 #include <linux/wait.h>
   4 #include <linux/sched.h>
   5
   6 #include "mds_client.h"
   7 #include "mon_client.h"
   8 #include "super.h"
   9 #include "messenger.h"
  10 #include "decode.h"
  11 #include "auth.h"
  12 #include "pagelist.h"
  13
  14 /*
  15  * A cluster of MDS (metadata server) daemons is responsible for
  16  * managing the file system namespace (the directory hierarchy and
  17  * inodes) and for coordinating shared access to storage.  Metadata is
  18  * partitioning hierarchically across a number of servers, and that
  19  * partition varies over time as the cluster adjusts the distribution
  20  * in order to balance load.
  21  *
  22  * The MDS client is primarily responsible to managing synchronous
  23  * metadata requests for operations like open, unlink, and so forth.
  24  * If there is a MDS failure, we find out about it when we (possibly
  25  * request and) receive a new MDS map, and can resubmit affected
  26  * requests.
  27  *
  28  * For the most part, though, we take advantage of a lossless
  29  * communications channel to the MDS, and do not need to worry about
  30  * timing out or resubmitting requests.
  31  *
  32  * We maintain a stateful "session" with each MDS we interact with.
  33  * Within each session, we sent periodic heartbeat messages to ensure
  34  * any capabilities or leases we have been issues remain valid.  If
  35  * the session times out and goes stale, our leases and capabilities
  36  * are no longer valid.
  37  */
  38
  39 static void __wake_requests(struct ceph_mds_client *mdsc,
  40                             struct list_head *head);
  41
  42 const static struct ceph_connection_operations mds_con_ops;
  43
  44
  45 /*
  46  * mds reply parsing
  47  */
  48
  49 /*
  50  * parse individual inode info
  51  */
  52 static int parse_reply_info_in(void **p, void *end,
  53                                struct ceph_mds_reply_info_in *info)
  54 {
  55         int err = -EIO;
  56
  57         info->in = *p;
  58         *p += sizeof(struct ceph_mds_reply_inode) +
  59                 sizeof(*info->in->fragtree.splits) *
  60                 le32_to_cpu(info->in->fragtree.nsplits);
  61
  62         ceph_decode_32_safe(p, end, info->symlink_len, bad);
  63         ceph_decode_need(p, end, info->symlink_len, bad);
  64         info->symlink = *p;
  65         *p += info->symlink_len;
  66
  67         ceph_decode_32_safe(p, end, info->xattr_len, bad);
  68         ceph_decode_need(p, end, info->xattr_len, bad);
  69         info->xattr_data = *p;
  70         *p += info->xattr_len;
  71         return 0;
  72 bad:
  73         return err;
  74 }
  75
  76 /*
  77  * parse a normal reply, which may contain a (dir+)dentry and/or a
  78  * target inode.
  79  */
  80 static int parse_reply_info_trace(void **p, void *end,
  81                                   struct ceph_mds_reply_info_parsed *info)
  82 {
  83         int err;
  84
  85         if (info->head->is_dentry) {
  86                 err = parse_reply_info_in(p, end, &info->diri);
  87                 if (err < 0)
  88                         goto out_bad;
  89
  90                 if (unlikely(*p + sizeof(*info->dirfrag) > end))
  91                         goto bad;
  92                 info->dirfrag = *p;
  93                 *p += sizeof(*info->dirfrag) +
  94                         sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
  95                 if (unlikely(*p > end))
  96                         goto bad;
  97
  98                 ceph_decode_32_safe(p, end, info->dname_len, bad);
  99                 ceph_decode_need(p, end, info->dname_len, bad);
 100                 info->dname = *p;
 101                 *p += info->dname_len;
 102                 info->dlease = *p;
 103                 *p += sizeof(*info->dlease);
 104         }
 105
 106         if (info->head->is_target) {
 107                 err = parse_reply_info_in(p, end, &info->targeti);
 108                 if (err < 0)
 109                         goto out_bad;
 110         }
 111
 112         if (unlikely(*p != end))
 113                 goto bad;
 114         return 0;
 115
 116 bad:
 117         err = -EIO;
 118 out_bad:
 119         pr_err("problem parsing mds trace %d\n", err);
 120         return err;
 121 }
 122
 123 /*
 124  * parse readdir results
 125  */
 126 static int parse_reply_info_dir(void **p, void *end,
 127                                 struct ceph_mds_reply_info_parsed *info)
 128 {
 129         u32 num, i = 0;
 130         int err;
 131
 132         info->dir_dir = *p;
 133         if (*p + sizeof(*info->dir_dir) > end)
 134                 goto bad;
 135         *p += sizeof(*info->dir_dir) +
 136                 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
 137         if (*p > end)
 138                 goto bad;
 139
 140         ceph_decode_need(p, end, sizeof(num) + 2, bad);
 141         num = ceph_decode_32(p);
 142         info->dir_end = ceph_decode_8(p);
 143         info->dir_complete = ceph_decode_8(p);
 144         if (num == 0)
 145                 goto done;
 146
 147         /* alloc large array */
 148         info->dir_nr = num;
 149         info->dir_in = kcalloc(num, sizeof(*info->dir_in) +
 150                                sizeof(*info->dir_dname) +
 151                                sizeof(*info->dir_dname_len) +
 152                                sizeof(*info->dir_dlease),
 153                                GFP_NOFS);
 154         if (info->dir_in == NULL) {
 155                 err = -ENOMEM;
 156                 goto out_bad;
 157         }
 158         info->dir_dname = (void *)(info->dir_in + num);
 159         info->dir_dname_len = (void *)(info->dir_dname + num);
 160         info->dir_dlease = (void *)(info->dir_dname_len + num);
 161
 162         while (num) {
 163                 /* dentry */
 164                 ceph_decode_need(p, end, sizeof(u32)*2, bad);
 165                 info->dir_dname_len[i] = ceph_decode_32(p);
 166                 ceph_decode_need(p, end, info->dir_dname_len[i], bad);
 167                 info->dir_dname[i] = *p;
 168                 *p += info->dir_dname_len[i];
 169                 dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
 170                      info->dir_dname[i]);
 171                 info->dir_dlease[i] = *p;
 172                 *p += sizeof(struct ceph_mds_reply_lease);
 173
 174                 /* inode */
 175                 err = parse_reply_info_in(p, end, &info->dir_in[i]);
 176                 if (err < 0)
 177                         goto out_bad;
 178                 i++;
 179                 num--;
 180         }
 181
 182 done:
 183         if (*p != end)
 184                 goto bad;
 185         return 0;
 186
 187 bad:
 188         err = -EIO;
 189 out_bad:
 190         pr_err("problem parsing dir contents %d\n", err);
 191         return err;
 192 }
 193
 194 /*
 195  * parse entire mds reply
 196  */
 197 static int parse_reply_info(struct ceph_msg *msg,
 198                             struct ceph_mds_reply_info_parsed *info)
 199 {
 200         void *p, *end;
 201         u32 len;
 202         int err;
 203
 204         info->head = msg->front.iov_base;
 205         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 206         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 207
 208         /* trace */
 209         ceph_decode_32_safe(&p, end, len, bad);
 210         if (len > 0) {
 211                 err = parse_reply_info_trace(&p, p+len, info);
 212                 if (err < 0)
 213                         goto out_bad;
 214         }
 215
 216         /* dir content */
 217         ceph_decode_32_safe(&p, end, len, bad);
 218         if (len > 0) {
 219                 err = parse_reply_info_dir(&p, p+len, info);
 220                 if (err < 0)
 221                         goto out_bad;
 222         }
 223
 224         /* snap blob */
 225         ceph_decode_32_safe(&p, end, len, bad);
 226         info->snapblob_len = len;
 227         info->snapblob = p;
 228         p += len;
 229
 230         if (p != end)
 231                 goto bad;
 232         return 0;
 233
 234 bad:
 235         err = -EIO;
 236 out_bad:
 237         pr_err("mds parse_reply err %d\n", err);
 238         return err;
 239 }
 240
 241 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
 242 {
 243         kfree(info->dir_in);
 244 }
 245
 246
 247 /*
 248  * sessions
 249  */
 250 static const char *session_state_name(int s)
 251 {
 252         switch (s) {
 253         case CEPH_MDS_SESSION_NEW: return "new";
 254         case CEPH_MDS_SESSION_OPENING: return "opening";
 255         case CEPH_MDS_SESSION_OPEN: return "open";
 256         case CEPH_MDS_SESSION_HUNG: return "hung";
 257         case CEPH_MDS_SESSION_CLOSING: return "closing";
 258         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
 259         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
 260         default: return "???";
 261         }
 262 }
 263
 264 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
 265 {
 266         if (atomic_inc_not_zero(&s->s_ref)) {
 267                 dout("mdsc get_session %p %d -> %d\n", s,
 268                      atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
 269                 return s;
 270         } else {
 271                 dout("mdsc get_session %p 0 -- FAIL", s);
 272                 return NULL;
 273         }
 274 }
 275
 276 void ceph_put_mds_session(struct ceph_mds_session *s)
 277 {
 278         dout("mdsc put_session %p %d -> %d\n", s,
 279              atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
 280         if (atomic_dec_and_test(&s->s_ref)) {
 281                 if (s->s_authorizer)
 282                         s->s_mdsc->client->monc.auth->ops->destroy_authorizer(
 283                                 s->s_mdsc->client->monc.auth, s->s_authorizer);
 284                 kfree(s);
 285         }
 286 }
 287
 288 /*
 289  * called under mdsc->mutex
 290  */
 291 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
 292                                                    int mds)
 293 {
 294         struct ceph_mds_session *session;
 295
 296         if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
 297                 return NULL;
 298         session = mdsc->sessions[mds];
 299         dout("lookup_mds_session %p %d\n", session,
 300              atomic_read(&session->s_ref));
 301         get_session(session);
 302         return session;
 303 }
 304
 305 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
 306 {
 307         if (mds >= mdsc->max_sessions)
 308                 return false;
 309         return mdsc->sessions[mds];
 310 }
 311
 312 static int __verify_registered_session(struct ceph_mds_client *mdsc,
 313                                        struct ceph_mds_session *s)
 314 {
 315         if (s->s_mds >= mdsc->max_sessions ||
 316             mdsc->sessions[s->s_mds] != s)
 317                 return -ENOENT;
 318         return 0;
 319 }
 320
 321 /*
 322  * create+register a new session for given mds.
 323  * called under mdsc->mutex.
 324  */
 325 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 326                                                  int mds)
 327 {
 328         struct ceph_mds_session *s;
 329
 330         s = kzalloc(sizeof(*s), GFP_NOFS);
 331         s->s_mdsc = mdsc;
 332         s->s_mds = mds;
 333         s->s_state = CEPH_MDS_SESSION_NEW;
 334         s->s_ttl = 0;
 335         s->s_seq = 0;
 336         mutex_init(&s->s_mutex);
 337
 338         ceph_con_init(mdsc->client->msgr, &s->s_con);
 339         s->s_con.private = s;
 340         s->s_con.ops = &mds_con_ops;
 341         s->s_con.peer_name.type = CEPH_ENTITY_TYPE_MDS;
 342         s->s_con.peer_name.num = cpu_to_le64(mds);
 343
 344         spin_lock_init(&s->s_cap_lock);
 345         s->s_cap_gen = 0;
 346         s->s_cap_ttl = 0;
 347         s->s_renew_requested = 0;
 348         s->s_renew_seq = 0;
 349         INIT_LIST_HEAD(&s->s_caps);
 350         s->s_nr_caps = 0;
 351         s->s_trim_caps = 0;
 352         atomic_set(&s->s_ref, 1);
 353         INIT_LIST_HEAD(&s->s_waiting);
 354         INIT_LIST_HEAD(&s->s_unsafe);
 355         s->s_num_cap_releases = 0;
 356         s->s_cap_iterator = NULL;
 357         INIT_LIST_HEAD(&s->s_cap_releases);
 358         INIT_LIST_HEAD(&s->s_cap_releases_done);
 359         INIT_LIST_HEAD(&s->s_cap_flushing);
 360         INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
 361
 362         dout("register_session mds%d\n", mds);
 363         if (mds >= mdsc->max_sessions) {
 364                 int newmax = 1 << get_count_order(mds+1);
 365                 struct ceph_mds_session **sa;
 366
 367                 dout("register_session realloc to %d\n", newmax);
 368                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 369                 if (sa == NULL)
 370                         goto fail_realloc;
 371                 if (mdsc->sessions) {
 372                         memcpy(sa, mdsc->sessions,
 373                                mdsc->max_sessions * sizeof(void *));
 374                         kfree(mdsc->sessions);
 375                 }
 376                 mdsc->sessions = sa;
 377                 mdsc->max_sessions = newmax;
 378         }
 379         mdsc->sessions[mds] = s;
 380         atomic_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
 381
 382         ceph_con_open(&s->s_con, ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
 383
 384         return s;
 385
 386 fail_realloc:
 387         kfree(s);
 388         return ERR_PTR(-ENOMEM);
 389 }
 390
 391 /*
 392  * called under mdsc->mutex
 393  */
 394 static void __unregister_session(struct ceph_mds_client *mdsc,
 395                                struct ceph_mds_session *s)
 396 {
 397         dout("__unregister_session mds%d %p\n", s->s_mds, s);
 398         BUG_ON(mdsc->sessions[s->s_mds] != s);
 399         mdsc->sessions[s->s_mds] = NULL;
 400         ceph_con_close(&s->s_con);
 401         ceph_put_mds_session(s);
 402 }
 403
 404 /*
 405  * drop session refs in request.
 406  *
 407  * should be last request ref, or hold mdsc->mutex
 408  */
 409 static void put_request_session(struct ceph_mds_request *req)
 410 {
 411         if (req->r_session) {
 412                 ceph_put_mds_session(req->r_session);
 413                 req->r_session = NULL;
 414         }
 415 }
 416
 417 void ceph_mdsc_release_request(struct kref *kref)
 418 {
 419         struct ceph_mds_request *req = container_of(kref,
 420                                                     struct ceph_mds_request,
 421                                                     r_kref);
 422         if (req->r_request)
 423                 ceph_msg_put(req->r_request);
 424         if (req->r_reply) {
 425                 ceph_msg_put(req->r_reply);
 426                 destroy_reply_info(&req->r_reply_info);
 427         }
 428         if (req->r_inode) {
 429                 ceph_put_cap_refs(ceph_inode(req->r_inode),
 430                                   CEPH_CAP_PIN);
 431                 iput(req->r_inode);
 432         }
 433         if (req->r_locked_dir)
 434                 ceph_put_cap_refs(ceph_inode(req->r_locked_dir),
 435                                   CEPH_CAP_PIN);
 436         if (req->r_target_inode)
 437                 iput(req->r_target_inode);
 438         if (req->r_dentry)
 439                 dput(req->r_dentry);
 440         if (req->r_old_dentry) {
 441                 ceph_put_cap_refs(
 442                         ceph_inode(req->r_old_dentry->d_parent->d_inode),
 443                         CEPH_CAP_PIN);
 444                 dput(req->r_old_dentry);
 445         }
 446         kfree(req->r_path1);
 447         kfree(req->r_path2);
 448         put_request_session(req);
 449         ceph_unreserve_caps(&req->r_caps_reservation);
 450         kfree(req);
 451 }
 452
 453 /*
 454  * lookup session, bump ref if found.
 455  *
 456  * called under mdsc->mutex.
 457  */
 458 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
 459                                              u64 tid)
 460 {
 461         struct ceph_mds_request *req;
 462         struct rb_node *n = mdsc->request_tree.rb_node;
 463
 464         while (n) {
 465                 req = rb_entry(n, struct ceph_mds_request, r_node);
 466                 if (tid < req->r_tid)
 467                         n = n->rb_left;
 468                 else if (tid > req->r_tid)
 469                         n = n->rb_right;
 470                 else {
 471                         ceph_mdsc_get_request(req);
 472                         return req;
 473                 }
 474         }
 475         return NULL;
 476 }
 477
 478 static void __insert_request(struct ceph_mds_client *mdsc,
 479                              struct ceph_mds_request *new)
 480 {
 481         struct rb_node **p = &mdsc->request_tree.rb_node;
 482         struct rb_node *parent = NULL;
 483         struct ceph_mds_request *req = NULL;
 484
 485         while (*p) {
 486                 parent = *p;
 487                 req = rb_entry(parent, struct ceph_mds_request, r_node);
 488                 if (new->r_tid < req->r_tid)
 489                         p = &(*p)->rb_left;
 490                 else if (new->r_tid > req->r_tid)
 491                         p = &(*p)->rb_right;
 492                 else
 493                         BUG();
 494         }
 495
 496         rb_link_node(&new->r_node, parent, p);
 497         rb_insert_color(&new->r_node, &mdsc->request_tree);
 498 }
 499
 500 /*
 501  * Register an in-flight request, and assign a tid.  Link to directory
 502  * are modifying (if any).
 503  *
 504  * Called under mdsc->mutex.
 505  */
 506 static void __register_request(struct ceph_mds_client *mdsc,
 507                                struct ceph_mds_request *req,
 508                                struct inode *dir)
 509 {
 510         req->r_tid = ++mdsc->last_tid;
 511         if (req->r_num_caps)
 512                 ceph_reserve_caps(&req->r_caps_reservation, req->r_num_caps);
 513         dout("__register_request %p tid %lld\n", req, req->r_tid);
 514         ceph_mdsc_get_request(req);
 515         __insert_request(mdsc, req);
 516
 517         if (dir) {
 518                 struct ceph_inode_info *ci = ceph_inode(dir);
 519
 520                 spin_lock(&ci->i_unsafe_lock);
 521                 req->r_unsafe_dir = dir;
 522                 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
 523                 spin_unlock(&ci->i_unsafe_lock);
 524         }
 525 }
 526
 527 static void __unregister_request(struct ceph_mds_client *mdsc,
 528                                  struct ceph_mds_request *req)
 529 {
 530         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 531         rb_erase(&req->r_node, &mdsc->request_tree);
 532         RB_CLEAR_NODE(&req->r_node);
 533         ceph_mdsc_put_request(req);
 534
 535         if (req->r_unsafe_dir) {
 536                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 537
 538                 spin_lock(&ci->i_unsafe_lock);
 539                 list_del_init(&req->r_unsafe_dir_item);
 540                 spin_unlock(&ci->i_unsafe_lock);
 541         }
 542 }
 543
 544 /*
 545  * Choose mds to send request to next.  If there is a hint set in the
 546  * request (e.g., due to a prior forward hint from the mds), use that.
 547  * Otherwise, consult frag tree and/or caps to identify the
 548  * appropriate mds.  If all else fails, choose randomly.
 549  *
 550  * Called under mdsc->mutex.
 551  */
 552 static int __choose_mds(struct ceph_mds_client *mdsc,
 553                         struct ceph_mds_request *req)
 554 {
 555         struct inode *inode;
 556         struct ceph_inode_info *ci;
 557         struct ceph_cap *cap;
 558         int mode = req->r_direct_mode;
 559         int mds = -1;
 560         u32 hash = req->r_direct_hash;
 561         bool is_hash = req->r_direct_is_hash;
 562
 563         /*
 564          * is there a specific mds we should try?  ignore hint if we have
 565          * no session and the mds is not up (active or recovering).
 566          */
 567         if (req->r_resend_mds >= 0 &&
 568             (__have_session(mdsc, req->r_resend_mds) ||
 569              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 570                 dout("choose_mds using resend_mds mds%d\n",
 571                      req->r_resend_mds);
 572                 return req->r_resend_mds;
 573         }
 574
 575         if (mode == USE_RANDOM_MDS)
 576                 goto random;
 577
 578         inode = NULL;
 579         if (req->r_inode) {
 580                 inode = req->r_inode;
 581         } else if (req->r_dentry) {
 582                 if (req->r_dentry->d_inode) {
 583                         inode = req->r_dentry->d_inode;
 584                 } else {
 585                         inode = req->r_dentry->d_parent->d_inode;
 586                         hash = req->r_dentry->d_name.hash;
 587                         is_hash = true;
 588                 }
 589         }
 590         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 591              (int)hash, mode);
 592         if (!inode)
 593                 goto random;
 594         ci = ceph_inode(inode);
 595
 596         if (is_hash && S_ISDIR(inode->i_mode)) {
 597                 struct ceph_inode_frag frag;
 598                 int found;
 599
 600                 ceph_choose_frag(ci, hash, &frag, &found);
 601                 if (found) {
 602                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 603                                 u8 r;
 604
 605                                 /* choose a random replica */
 606                                 get_random_bytes(&r, 1);
 607                                 r %= frag.ndist;
 608                                 mds = frag.dist[r];
 609                                 dout("choose_mds %p %llx.%llx "
 610                                      "frag %u mds%d (%d/%d)\n",
 611                                      inode, ceph_vinop(inode),
 612                                      frag.frag, frag.mds,
 613                                      (int)r, frag.ndist);
 614                                 return mds;
 615                         }
 616
 617                         /* since this file/dir wasn't known to be
 618                          * replicated, then we want to look for the
 619                          * authoritative mds. */
 620                         mode = USE_AUTH_MDS;
 621                         if (frag.mds >= 0) {
 622                                 /* choose auth mds */
 623                                 mds = frag.mds;
 624                                 dout("choose_mds %p %llx.%llx "
 625                                      "frag %u mds%d (auth)\n",
 626                                      inode, ceph_vinop(inode), frag.frag, mds);
 627                                 return mds;
 628                         }
 629                 }
 630         }
 631
 632         spin_lock(&inode->i_lock);
 633         cap = NULL;
 634         if (mode == USE_AUTH_MDS)
 635                 cap = ci->i_auth_cap;
 636         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
 637                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
 638         if (!cap) {
 639                 spin_unlock(&inode->i_lock);
 640                 goto random;
 641         }
 642         mds = cap->session->s_mds;
 643         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
 644              inode, ceph_vinop(inode), mds,
 645              cap == ci->i_auth_cap ? "auth " : "", cap);
 646         spin_unlock(&inode->i_lock);
 647         return mds;
 648
 649 random:
 650         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
 651         dout("choose_mds chose random mds%d\n", mds);
 652         return mds;
 653 }
 654
 655
 656 /*
 657  * session messages
 658  */
 659 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
 660 {
 661         struct ceph_msg *msg;
 662         struct ceph_mds_session_head *h;
 663
 664         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), 0, 0, NULL);
 665         if (IS_ERR(msg)) {
 666                 pr_err("create_session_msg ENOMEM creating msg\n");
 667                 return ERR_PTR(PTR_ERR(msg));
 668         }
 669         h = msg->front.iov_base;
 670         h->op = cpu_to_le32(op);
 671         h->seq = cpu_to_le64(seq);
 672         return msg;
 673 }
 674
 675 /*
 676  * send session open request.
 677  *
 678  * called under mdsc->mutex
 679  */
 680 static int __open_session(struct ceph_mds_client *mdsc,
 681                           struct ceph_mds_session *session)
 682 {
 683         struct ceph_msg *msg;
 684         int mstate;
 685         int mds = session->s_mds;
 686         int err = 0;
 687
 688         /* wait for mds to go active? */
 689         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
 690         dout("open_session to mds%d (%s)\n", mds,
 691              ceph_mds_state_name(mstate));
 692         session->s_state = CEPH_MDS_SESSION_OPENING;
 693         session->s_renew_requested = jiffies;
 694
 695         /* send connect message */
 696         msg = create_session_msg(CEPH_SESSION_REQUEST_OPEN, session->s_seq);
 697         if (IS_ERR(msg)) {
 698                 err = PTR_ERR(msg);
 699                 goto out;
 700         }
 701         ceph_con_send(&session->s_con, msg);
 702
 703 out:
 704         return 0;
 705 }
 706
 707 /*
 708  * session caps
 709  */
 710
 711 /*
 712  * Free preallocated cap messages assigned to this session
 713  */
 714 static void cleanup_cap_releases(struct ceph_mds_session *session)
 715 {
 716         struct ceph_msg *msg;
 717
 718         spin_lock(&session->s_cap_lock);
 719         while (!list_empty(&session->s_cap_releases)) {
 720                 msg = list_first_entry(&session->s_cap_releases,
 721                                        struct ceph_msg, list_head);
 722                 list_del_init(&msg->list_head);
 723                 ceph_msg_put(msg);
 724         }
 725         while (!list_empty(&session->s_cap_releases_done)) {
 726                 msg = list_first_entry(&session->s_cap_releases_done,
 727                                        struct ceph_msg, list_head);
 728                 list_del_init(&msg->list_head);
 729                 ceph_msg_put(msg);
 730         }
 731         spin_unlock(&session->s_cap_lock);
 732 }
 733
 734 /*
 735  * Helper to safely iterate over all caps associated with a session.
 736  *
 737  * caller must hold session s_mutex
 738  */
 739 static int iterate_session_caps(struct ceph_mds_session *session,
 740                                  int (*cb)(struct inode *, struct ceph_cap *,
 741                                             void *), void *arg)
 742 {
 743         struct list_head *p;
 744         struct ceph_cap *cap;
 745         struct inode *inode, *last_inode = NULL;
 746         struct ceph_cap *old_cap = NULL;
 747         int ret;
 748
 749         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
 750         spin_lock(&session->s_cap_lock);
 751         p = session->s_caps.next;
 752         while (p != &session->s_caps) {
 753                 cap = list_entry(p, struct ceph_cap, session_caps);
 754                 inode = igrab(&cap->ci->vfs_inode);
 755                 if (!inode) {
 756                         p = p->next;
 757                         continue;
 758                 }
 759                 session->s_cap_iterator = cap;
 760                 spin_unlock(&session->s_cap_lock);
 761
 762                 if (last_inode) {
 763                         iput(last_inode);
 764                         last_inode = NULL;
 765                 }
 766                 if (old_cap) {
 767                         ceph_put_cap(old_cap);
 768                         old_cap = NULL;
 769                 }
 770
 771                 ret = cb(inode, cap, arg);
 772                 last_inode = inode;
 773
 774                 spin_lock(&session->s_cap_lock);
 775                 p = p->next;
 776                 if (cap->ci == NULL) {
 777                         dout("iterate_session_caps  finishing cap %p removal\n",
 778                              cap);
 779                         BUG_ON(cap->session != session);
 780                         list_del_init(&cap->session_caps);
 781                         session->s_nr_caps--;
 782                         cap->session = NULL;
 783                         old_cap = cap;  /* put_cap it w/o locks held */
 784                 }
 785                 if (ret < 0)
 786                         goto out;
 787         }
 788         ret = 0;
 789 out:
 790         session->s_cap_iterator = NULL;
 791         spin_unlock(&session->s_cap_lock);
 792
 793         if (last_inode)
 794                 iput(last_inode);
 795         if (old_cap)
 796                 ceph_put_cap(old_cap);
 797
 798         return ret;
 799 }
 800
 801 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
 802                                    void *arg)
 803 {
 804         struct ceph_inode_info *ci = ceph_inode(inode);
 805         dout("removing cap %p, ci is %p, inode is %p\n",
 806              cap, ci, &ci->vfs_inode);
 807         ceph_remove_cap(cap);
 808         return 0;
 809 }
 810
 811 /*
 812  * caller must hold session s_mutex
 813  */
 814 static void remove_session_caps(struct ceph_mds_session *session)
 815 {
 816         dout("remove_session_caps on %p\n", session);
 817         iterate_session_caps(session, remove_session_caps_cb, NULL);
 818         BUG_ON(session->s_nr_caps > 0);
 819         cleanup_cap_releases(session);
 820 }
 821
 822 /*
 823  * wake up any threads waiting on this session's caps.  if the cap is
 824  * old (didn't get renewed on the client reconnect), remove it now.
 825  *
 826  * caller must hold s_mutex.
 827  */
 828 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
 829                               void *arg)
 830 {
 831         struct ceph_inode_info *ci = ceph_inode(inode);
 832
 833         wake_up(&ci->i_cap_wq);
 834         if (arg) {
 835                 spin_lock(&inode->i_lock);
 836                 ci->i_wanted_max_size = 0;
 837                 ci->i_requested_max_size = 0;
 838                 spin_unlock(&inode->i_lock);
 839         }
 840         return 0;
 841 }
 842
 843 static void wake_up_session_caps(struct ceph_mds_session *session,
 844                                  int reconnect)
 845 {
 846         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
 847         iterate_session_caps(session, wake_up_session_cb,
 848                              (void *)(unsigned long)reconnect);
 849 }
 850
 851 /*
 852  * Send periodic message to MDS renewing all currently held caps.  The
 853  * ack will reset the expiration for all caps from this session.
 854  *
 855  * caller holds s_mutex
 856  */
 857 static int send_renew_caps(struct ceph_mds_client *mdsc,
 858                            struct ceph_mds_session *session)
 859 {
 860         struct ceph_msg *msg;
 861         int state;
 862
 863         if (time_after_eq(jiffies, session->s_cap_ttl) &&
 864             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
 865                 pr_info("mds%d caps stale\n", session->s_mds);
 866         session->s_renew_requested = jiffies;
 867
 868         /* do not try to renew caps until a recovering mds has reconnected
 869          * with its clients. */
 870         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
 871         if (state < CEPH_MDS_STATE_RECONNECT) {
 872                 dout("send_renew_caps ignoring mds%d (%s)\n",
 873                      session->s_mds, ceph_mds_state_name(state));
 874                 return 0;
 875         }
 876
 877         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
 878                 ceph_mds_state_name(state));
 879         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
 880                                  ++session->s_renew_seq);
 881         if (IS_ERR(msg))
 882                 return PTR_ERR(msg);
 883         ceph_con_send(&session->s_con, msg);
 884         return 0;
 885 }
 886
 887 /*
 888  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
 889  *
 890  * Called under session->s_mutex
 891  */
 892 static void renewed_caps(struct ceph_mds_client *mdsc,
 893                          struct ceph_mds_session *session, int is_renew)
 894 {
 895         int was_stale;
 896         int wake = 0;
 897
 898         spin_lock(&session->s_cap_lock);
 899         was_stale = is_renew && (session->s_cap_ttl == 0 ||
 900                                  time_after_eq(jiffies, session->s_cap_ttl));
 901
 902         session->s_cap_ttl = session->s_renew_requested +
 903                 mdsc->mdsmap->m_session_timeout*HZ;
 904
 905         if (was_stale) {
 906                 if (time_before(jiffies, session->s_cap_ttl)) {
 907                         pr_info("mds%d caps renewed\n", session->s_mds);
 908                         wake = 1;
 909                 } else {
 910                         pr_info("mds%d caps still stale\n", session->s_mds);
 911                 }
 912         }
 913         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
 914              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
 915              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
 916         spin_unlock(&session->s_cap_lock);
 917
 918         if (wake)
 919                 wake_up_session_caps(session, 0);
 920 }
 921
 922 /*
 923  * send a session close request
 924  */
 925 static int request_close_session(struct ceph_mds_client *mdsc,
 926                                  struct ceph_mds_session *session)
 927 {
 928         struct ceph_msg *msg;
 929         int err = 0;
 930
 931         dout("request_close_session mds%d state %s seq %lld\n",
 932              session->s_mds, session_state_name(session->s_state),
 933              session->s_seq);
 934         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
 935         if (IS_ERR(msg))
 936                 err = PTR_ERR(msg);
 937         else
 938                 ceph_con_send(&session->s_con, msg);
 939         return err;
 940 }
 941
 942 /*
 943  * Called with s_mutex held.
 944  */
 945 static int __close_session(struct ceph_mds_client *mdsc,
 946                          struct ceph_mds_session *session)
 947 {
 948         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
 949                 return 0;
 950         session->s_state = CEPH_MDS_SESSION_CLOSING;
 951         return request_close_session(mdsc, session);
 952 }
 953
 954 /*
 955  * Trim old(er) caps.
 956  *
 957  * Because we can't cache an inode without one or more caps, we do
 958  * this indirectly: if a cap is unused, we prune its aliases, at which
 959  * point the inode will hopefully get dropped to.
 960  *
 961  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
 962  * memory pressure from the MDS, though, so it needn't be perfect.
 963  */
 964 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
 965 {
 966         struct ceph_mds_session *session = arg;
 967         struct ceph_inode_info *ci = ceph_inode(inode);
 968         int used, oissued, mine;
 969
 970         if (session->s_trim_caps <= 0)
 971                 return -1;
 972
 973         spin_lock(&inode->i_lock);
 974         mine = cap->issued | cap->implemented;
 975         used = __ceph_caps_used(ci);
 976         oissued = __ceph_caps_issued_other(ci, cap);
 977
 978         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
 979              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
 980              ceph_cap_string(used));
 981         if (ci->i_dirty_caps)
 982                 goto out;   /* dirty caps */
 983         if ((used & ~oissued) & mine)
 984                 goto out;   /* we need these caps */
 985
 986         session->s_trim_caps--;
 987         if (oissued) {
 988                 /* we aren't the only cap.. just remove us */
 989                 __ceph_remove_cap(cap);
 990         } else {
 991                 /* try to drop referring dentries */
 992                 spin_unlock(&inode->i_lock);
 993                 d_prune_aliases(inode);
 994                 dout("trim_caps_cb %p cap %p  pruned, count now %d\n",
 995                      inode, cap, atomic_read(&inode->i_count));
 996                 return 0;
 997         }
 998
 999 out:
1000         spin_unlock(&inode->i_lock);
1001         return 0;
1002 }
1003
1004 /*
1005  * Trim session cap count down to some max number.
1006  */
1007 static int trim_caps(struct ceph_mds_client *mdsc,
1008                      struct ceph_mds_session *session,
1009                      int max_caps)
1010 {
1011         int trim_caps = session->s_nr_caps - max_caps;
1012
1013         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1014              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1015         if (trim_caps > 0) {
1016                 session->s_trim_caps = trim_caps;
1017                 iterate_session_caps(session, trim_caps_cb, session);
1018                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1019                      session->s_mds, session->s_nr_caps, max_caps,
1020                         trim_caps - session->s_trim_caps);
1021                 session->s_trim_caps = 0;
1022         }
1023         return 0;
1024 }
1025
1026 /*
1027  * Allocate cap_release messages.  If there is a partially full message
1028  * in the queue, try to allocate enough to cover it's remainder, so that
1029  * we can send it immediately.
1030  *
1031  * Called under s_mutex.
1032  */
1033 static int add_cap_releases(struct ceph_mds_client *mdsc,
1034                             struct ceph_mds_session *session,
1035                             int extra)
1036 {
1037         struct ceph_msg *msg;
1038         struct ceph_mds_cap_release *head;
1039         int err = -ENOMEM;
1040
1041         if (extra < 0)
1042                 extra = mdsc->client->mount_args->cap_release_safety;
1043
1044         spin_lock(&session->s_cap_lock);
1045
1046         if (!list_empty(&session->s_cap_releases)) {
1047                 msg = list_first_entry(&session->s_cap_releases,
1048                                        struct ceph_msg,
1049                                  list_head);
1050                 head = msg->front.iov_base;
1051                 extra += CEPH_CAPS_PER_RELEASE - le32_to_cpu(head->num);
1052         }
1053
1054         while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1055                 spin_unlock(&session->s_cap_lock);
1056                 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
1057                                    0, 0, NULL);
1058                 if (!msg)
1059                         goto out_unlocked;
1060                 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1061                      (int)msg->front.iov_len);
1062                 head = msg->front.iov_base;
1063                 head->num = cpu_to_le32(0);
1064                 msg->front.iov_len = sizeof(*head);
1065                 spin_lock(&session->s_cap_lock);
1066                 list_add(&msg->list_head, &session->s_cap_releases);
1067                 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1068         }
1069
1070         if (!list_empty(&session->s_cap_releases)) {
1071                 msg = list_first_entry(&session->s_cap_releases,
1072                                        struct ceph_msg,
1073                                        list_head);
1074                 head = msg->front.iov_base;
1075                 if (head->num) {
1076                         dout(" queueing non-full %p (%d)\n", msg,
1077                              le32_to_cpu(head->num));
1078                         list_move_tail(&msg->list_head,
1079                                       &session->s_cap_releases_done);
1080                         session->s_num_cap_releases -=
1081                                 CEPH_CAPS_PER_RELEASE - le32_to_cpu(head->num);
1082                 }
1083         }
1084         err = 0;
1085         spin_unlock(&session->s_cap_lock);
1086 out_unlocked:
1087         return err;
1088 }
1089
1090 /*
1091  * flush all dirty inode data to disk.
1092  *
1093  * returns true if we've flushed through want_flush_seq
1094  */
1095 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1096 {
1097         int mds, ret = 1;
1098
1099         dout("check_cap_flush want %lld\n", want_flush_seq);
1100         mutex_lock(&mdsc->mutex);
1101         for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1102                 struct ceph_mds_session *session = mdsc->sessions[mds];
1103
1104                 if (!session)
1105                         continue;
1106                 get_session(session);
1107                 mutex_unlock(&mdsc->mutex);
1108
1109                 mutex_lock(&session->s_mutex);
1110                 if (!list_empty(&session->s_cap_flushing)) {
1111                         struct ceph_inode_info *ci =
1112                                 list_entry(session->s_cap_flushing.next,
1113                                            struct ceph_inode_info,
1114                                            i_flushing_item);
1115                         struct inode *inode = &ci->vfs_inode;
1116
1117                         spin_lock(&inode->i_lock);
1118                         if (ci->i_cap_flush_seq <= want_flush_seq) {
1119                                 dout("check_cap_flush still flushing %p "
1120                                      "seq %lld <= %lld to mds%d\n", inode,
1121                                      ci->i_cap_flush_seq, want_flush_seq,
1122                                      session->s_mds);
1123                                 ret = 0;
1124                         }
1125                         spin_unlock(&inode->i_lock);
1126                 }
1127                 mutex_unlock(&session->s_mutex);
1128                 ceph_put_mds_session(session);
1129
1130                 if (!ret)
1131                         return ret;
1132                 mutex_lock(&mdsc->mutex);
1133         }
1134
1135         mutex_unlock(&mdsc->mutex);
1136         dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1137         return ret;
1138 }
1139
1140 /*
1141  * called under s_mutex
1142  */
1143 static void send_cap_releases(struct ceph_mds_client *mdsc,
1144                        struct ceph_mds_session *session)
1145 {
1146         struct ceph_msg *msg;
1147
1148         dout("send_cap_releases mds%d\n", session->s_mds);
1149         while (1) {
1150                 spin_lock(&session->s_cap_lock);
1151                 if (list_empty(&session->s_cap_releases_done))
1152                         break;
1153                 msg = list_first_entry(&session->s_cap_releases_done,
1154                                  struct ceph_msg, list_head);
1155                 list_del_init(&msg->list_head);
1156                 spin_unlock(&session->s_cap_lock);
1157                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1158                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1159                 ceph_con_send(&session->s_con, msg);
1160         }
1161         spin_unlock(&session->s_cap_lock);
1162 }
1163
1164 /*
1165  * requests
1166  */
1167
1168 /*
1169  * Create an mds request.
1170  */
1171 struct ceph_mds_request *
1172 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1173 {
1174         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1175
1176         if (!req)
1177                 return ERR_PTR(-ENOMEM);
1178
1179         req->r_started = jiffies;
1180         req->r_resend_mds = -1;
1181         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1182         req->r_fmode = -1;
1183         kref_init(&req->r_kref);
1184         INIT_LIST_HEAD(&req->r_wait);
1185         init_completion(&req->r_completion);
1186         init_completion(&req->r_safe_completion);
1187         INIT_LIST_HEAD(&req->r_unsafe_item);
1188
1189         req->r_op = op;
1190         req->r_direct_mode = mode;
1191         return req;
1192 }
1193
1194 /*
1195  * return oldest (lowest) request, tid in request tree, 0 if none.
1196  *
1197  * called under mdsc->mutex.
1198  */
1199 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1200 {
1201         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1202                 return NULL;
1203         return rb_entry(rb_first(&mdsc->request_tree),
1204                         struct ceph_mds_request, r_node);
1205 }
1206
1207 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1208 {
1209         struct ceph_mds_request *req = __get_oldest_req(mdsc);
1210
1211         if (req)
1212                 return req->r_tid;
1213         return 0;
1214 }
1215
1216 /*
1217  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1218  * on build_path_from_dentry in fs/cifs/dir.c.
1219  *
1220  * If @stop_on_nosnap, generate path relative to the first non-snapped
1221  * inode.
1222  *
1223  * Encode hidden .snap dirs as a double /, i.e.
1224  *   foo/.snap/bar -> foo//bar
1225  */
1226 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1227                            int stop_on_nosnap)
1228 {
1229         struct dentry *temp;
1230         char *path;
1231         int len, pos;
1232
1233         if (dentry == NULL)
1234                 return ERR_PTR(-EINVAL);
1235
1236 retry:
1237         len = 0;
1238         for (temp = dentry; !IS_ROOT(temp);) {
1239                 struct inode *inode = temp->d_inode;
1240                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1241                         len++;  /* slash only */
1242                 else if (stop_on_nosnap && inode &&
1243                          ceph_snap(inode) == CEPH_NOSNAP)
1244                         break;
1245                 else
1246                         len += 1 + temp->d_name.len;
1247                 temp = temp->d_parent;
1248                 if (temp == NULL) {
1249                         pr_err("build_path_dentry corrupt dentry %p\n", dentry);
1250                         return ERR_PTR(-EINVAL);
1251                 }
1252         }
1253         if (len)
1254                 len--;  /* no leading '/' */
1255
1256         path = kmalloc(len+1, GFP_NOFS);
1257         if (path == NULL)
1258                 return ERR_PTR(-ENOMEM);
1259         pos = len;
1260         path[pos] = 0;  /* trailing null */
1261         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1262                 struct inode *inode = temp->d_inode;
1263
1264                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1265                         dout("build_path_dentry path+%d: %p SNAPDIR\n",
1266                              pos, temp);
1267                 } else if (stop_on_nosnap && inode &&
1268                            ceph_snap(inode) == CEPH_NOSNAP) {
1269                         break;
1270                 } else {
1271                         pos -= temp->d_name.len;
1272                         if (pos < 0)
1273                                 break;
1274                         strncpy(path + pos, temp->d_name.name,
1275                                 temp->d_name.len);
1276                         dout("build_path_dentry path+%d: %p '%.*s'\n",
1277                              pos, temp, temp->d_name.len, path + pos);
1278                 }
1279                 if (pos)
1280                         path[--pos] = '/';
1281                 temp = temp->d_parent;
1282                 if (temp == NULL) {
1283                         pr_err("build_path_dentry corrupt dentry\n");
1284                         kfree(path);
1285                         return ERR_PTR(-EINVAL);
1286                 }
1287         }
1288         if (pos != 0) {
1289                 pr_err("build_path_dentry did not end path lookup where "
1290                        "expected, namelen is %d, pos is %d\n", len, pos);
1291                 /* presumably this is only possible if racing with a
1292                    rename of one of the parent directories (we can not
1293                    lock the dentries above us to prevent this, but
1294                    retrying should be harmless) */
1295                 kfree(path);
1296                 goto retry;
1297         }
1298
1299         *base = ceph_ino(temp->d_inode);
1300         *plen = len;
1301         dout("build_path_dentry on %p %d built %llx '%.*s'\n",
1302              dentry, atomic_read(&dentry->d_count), *base, len, path);
1303         return path;
1304 }
1305
1306 static int build_dentry_path(struct dentry *dentry,
1307                              const char **ppath, int *ppathlen, u64 *pino,
1308                              int *pfreepath)
1309 {
1310         char *path;
1311
1312         if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1313                 *pino = ceph_ino(dentry->d_parent->d_inode);
1314                 *ppath = dentry->d_name.name;
1315                 *ppathlen = dentry->d_name.len;
1316                 return 0;
1317         }
1318         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1319         if (IS_ERR(path))
1320                 return PTR_ERR(path);
1321         *ppath = path;
1322         *pfreepath = 1;
1323         return 0;
1324 }
1325
1326 static int build_inode_path(struct inode *inode,
1327                             const char **ppath, int *ppathlen, u64 *pino,
1328                             int *pfreepath)
1329 {
1330         struct dentry *dentry;
1331         char *path;
1332
1333         if (ceph_snap(inode) == CEPH_NOSNAP) {
1334                 *pino = ceph_ino(inode);
1335                 *ppathlen = 0;
1336                 return 0;
1337         }
1338         dentry = d_find_alias(inode);
1339         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1340         dput(dentry);
1341         if (IS_ERR(path))
1342                 return PTR_ERR(path);
1343         *ppath = path;
1344         *pfreepath = 1;
1345         return 0;
1346 }
1347
1348 /*
1349  * request arguments may be specified via an inode *, a dentry *, or
1350  * an explicit ino+path.
1351  */
1352 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1353                                   const char *rpath, u64 rino,
1354                                   const char **ppath, int *pathlen,
1355                                   u64 *ino, int *freepath)
1356 {
1357         int r = 0;
1358
1359         if (rinode) {
1360                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1361                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1362                      ceph_snap(rinode));
1363         } else if (rdentry) {
1364                 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1365                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1366                      *ppath);
1367         } else if (rpath) {
1368                 *ino = rino;
1369                 *ppath = rpath;
1370                 *pathlen = strlen(rpath);
1371                 dout(" path %.*s\n", *pathlen, rpath);
1372         }
1373
1374         return r;
1375 }
1376
1377 /*
1378  * called under mdsc->mutex
1379  */
1380 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1381                                                struct ceph_mds_request *req,
1382                                                int mds)
1383 {
1384         struct ceph_msg *msg;
1385         struct ceph_mds_request_head *head;
1386         const char *path1 = NULL;
1387         const char *path2 = NULL;
1388         u64 ino1 = 0, ino2 = 0;
1389         int pathlen1 = 0, pathlen2 = 0;
1390         int freepath1 = 0, freepath2 = 0;
1391         int len;
1392         u16 releases;
1393         void *p, *end;
1394         int ret;
1395
1396         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1397                               req->r_path1, req->r_ino1.ino,
1398                               &path1, &pathlen1, &ino1, &freepath1);
1399         if (ret < 0) {
1400                 msg = ERR_PTR(ret);
1401                 goto out;
1402         }
1403
1404         ret = set_request_path_attr(NULL, req->r_old_dentry,
1405                               req->r_path2, req->r_ino2.ino,
1406                               &path2, &pathlen2, &ino2, &freepath2);
1407         if (ret < 0) {
1408                 msg = ERR_PTR(ret);
1409                 goto out_free1;
1410         }
1411
1412         len = sizeof(*head) +
1413                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64));
1414
1415         /* calculate (max) length for cap releases */
1416         len += sizeof(struct ceph_mds_request_release) *
1417                 (!!req->r_inode_drop + !!req->r_dentry_drop +
1418                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1419         if (req->r_dentry_drop)
1420                 len += req->r_dentry->d_name.len;
1421         if (req->r_old_dentry_drop)
1422                 len += req->r_old_dentry->d_name.len;
1423
1424         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, 0, 0, NULL);
1425         if (IS_ERR(msg))
1426                 goto out_free2;
1427
1428         msg->hdr.tid = cpu_to_le64(req->r_tid);
1429
1430         head = msg->front.iov_base;
1431         p = msg->front.iov_base + sizeof(*head);
1432         end = msg->front.iov_base + msg->front.iov_len;
1433
1434         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1435         head->op = cpu_to_le32(req->r_op);
1436         head->caller_uid = cpu_to_le32(current_fsuid());
1437         head->caller_gid = cpu_to_le32(current_fsgid());
1438         head->args = req->r_args;
1439
1440         ceph_encode_filepath(&p, end, ino1, path1);
1441         ceph_encode_filepath(&p, end, ino2, path2);
1442
1443         /* cap releases */
1444         releases = 0;
1445         if (req->r_inode_drop)
1446                 releases += ceph_encode_inode_release(&p,
1447                       req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1448                       mds, req->r_inode_drop, req->r_inode_unless, 0);
1449         if (req->r_dentry_drop)
1450                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1451                        mds, req->r_dentry_drop, req->r_dentry_unless);
1452         if (req->r_old_dentry_drop)
1453                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1454                        mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1455         if (req->r_old_inode_drop)
1456                 releases += ceph_encode_inode_release(&p,
1457                       req->r_old_dentry->d_inode,
1458                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1459         head->num_releases = cpu_to_le16(releases);
1460
1461         BUG_ON(p > end);
1462         msg->front.iov_len = p - msg->front.iov_base;
1463         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1464
1465         msg->pages = req->r_pages;
1466         msg->nr_pages = req->r_num_pages;
1467         msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1468         msg->hdr.data_off = cpu_to_le16(0);
1469
1470 out_free2:
1471         if (freepath2)
1472                 kfree((char *)path2);
1473 out_free1:
1474         if (freepath1)
1475                 kfree((char *)path1);
1476 out:
1477         return msg;
1478 }
1479
1480 /*
1481  * called under mdsc->mutex if error, under no mutex if
1482  * success.
1483  */
1484 static void complete_request(struct ceph_mds_client *mdsc,
1485                              struct ceph_mds_request *req)
1486 {
1487         if (req->r_callback)
1488                 req->r_callback(mdsc, req);
1489         else
1490                 complete(&req->r_completion);
1491 }
1492
1493 /*
1494  * called under mdsc->mutex
1495  */
1496 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1497                                   struct ceph_mds_request *req,
1498                                   int mds)
1499 {
1500         struct ceph_mds_request_head *rhead;
1501         struct ceph_msg *msg;
1502         int flags = 0;
1503
1504         req->r_mds = mds;
1505         req->r_attempts++;
1506         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1507              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1508
1509         if (req->r_request) {
1510                 ceph_msg_put(req->r_request);
1511                 req->r_request = NULL;
1512         }
1513         msg = create_request_message(mdsc, req, mds);
1514         if (IS_ERR(msg)) {
1515                 req->r_reply = ERR_PTR(PTR_ERR(msg));
1516                 complete_request(mdsc, req);
1517                 return -PTR_ERR(msg);
1518         }
1519         req->r_request = msg;
1520
1521         rhead = msg->front.iov_base;
1522         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1523         if (req->r_got_unsafe)
1524                 flags |= CEPH_MDS_FLAG_REPLAY;
1525         if (req->r_locked_dir)
1526                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1527         rhead->flags = cpu_to_le32(flags);
1528         rhead->num_fwd = req->r_num_fwd;
1529         rhead->num_retry = req->r_attempts - 1;
1530
1531         dout(" r_locked_dir = %p\n", req->r_locked_dir);
1532
1533         if (req->r_target_inode && req->r_got_unsafe)
1534                 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1535         else
1536                 rhead->ino = 0;
1537         return 0;
1538 }
1539
1540 /*
1541  * send request, or put it on the appropriate wait list.
1542  */
1543 static int __do_request(struct ceph_mds_client *mdsc,
1544                         struct ceph_mds_request *req)
1545 {
1546         struct ceph_mds_session *session = NULL;
1547         int mds = -1;
1548         int err = -EAGAIN;
1549
1550         if (req->r_reply)
1551                 goto out;
1552
1553         if (req->r_timeout &&
1554             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1555                 dout("do_request timed out\n");
1556                 err = -EIO;
1557                 goto finish;
1558         }
1559
1560         mds = __choose_mds(mdsc, req);
1561         if (mds < 0 ||
1562             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1563                 dout("do_request no mds or not active, waiting for map\n");
1564                 list_add(&req->r_wait, &mdsc->waiting_for_map);
1565                 goto out;
1566         }
1567
1568         /* get, open session */
1569         session = __ceph_lookup_mds_session(mdsc, mds);
1570         if (!session) {
1571                 session = register_session(mdsc, mds);
1572                 if (IS_ERR(session)) {
1573                         err = PTR_ERR(session);
1574                         goto finish;
1575                 }
1576         }
1577         dout("do_request mds%d session %p state %s\n", mds, session,
1578              session_state_name(session->s_state));
1579         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1580             session->s_state != CEPH_MDS_SESSION_HUNG) {
1581                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1582                     session->s_state == CEPH_MDS_SESSION_CLOSING)
1583                         __open_session(mdsc, session);
1584                 list_add(&req->r_wait, &session->s_waiting);
1585                 goto out_session;
1586         }
1587
1588         /* send request */
1589         req->r_session = get_session(session);
1590         req->r_resend_mds = -1;   /* forget any previous mds hint */
1591
1592         if (req->r_request_started == 0)   /* note request start time */
1593                 req->r_request_started = jiffies;
1594
1595         err = __prepare_send_request(mdsc, req, mds);
1596         if (!err) {
1597                 ceph_msg_get(req->r_request);
1598                 ceph_con_send(&session->s_con, req->r_request);
1599         }
1600
1601 out_session:
1602         ceph_put_mds_session(session);
1603 out:
1604         return err;
1605
1606 finish:
1607         req->r_reply = ERR_PTR(err);
1608         complete_request(mdsc, req);
1609         goto out;
1610 }
1611
1612 /*
1613  * called under mdsc->mutex
1614  */
1615 static void __wake_requests(struct ceph_mds_client *mdsc,
1616                             struct list_head *head)
1617 {
1618         struct ceph_mds_request *req, *nreq;
1619
1620         list_for_each_entry_safe(req, nreq, head, r_wait) {
1621                 list_del_init(&req->r_wait);
1622                 __do_request(mdsc, req);
1623         }
1624 }
1625
1626 /*
1627  * Wake up threads with requests pending for @mds, so that they can
1628  * resubmit their requests to a possibly different mds.  If @all is set,
1629  * wake up if their requests has been forwarded to @mds, too.
1630  */
1631 static void kick_requests(struct ceph_mds_client *mdsc, int mds, int all)
1632 {
1633         struct ceph_mds_request *req;
1634         struct rb_node *p;
1635
1636         dout("kick_requests mds%d\n", mds);
1637         for (p = rb_first(&mdsc->request_tree); p; p = rb_next(p)) {
1638                 req = rb_entry(p, struct ceph_mds_request, r_node);
1639                 if (req->r_got_unsafe)
1640                         continue;
1641                 if (req->r_session &&
1642                     req->r_session->s_mds == mds) {
1643                         dout(" kicking tid %llu\n", req->r_tid);
1644                         put_request_session(req);
1645                         __do_request(mdsc, req);
1646                 }
1647         }
1648 }
1649
1650 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
1651                               struct ceph_mds_request *req)
1652 {
1653         dout("submit_request on %p\n", req);
1654         mutex_lock(&mdsc->mutex);
1655         __register_request(mdsc, req, NULL);
1656         __do_request(mdsc, req);
1657         mutex_unlock(&mdsc->mutex);
1658 }
1659
1660 /*
1661  * Synchrously perform an mds request.  Take care of all of the
1662  * session setup, forwarding, retry details.
1663  */
1664 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
1665                          struct inode *dir,
1666                          struct ceph_mds_request *req)
1667 {
1668         int err;
1669
1670         dout("do_request on %p\n", req);
1671
1672         /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
1673         if (req->r_inode)
1674                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
1675         if (req->r_locked_dir)
1676                 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
1677         if (req->r_old_dentry)
1678                 ceph_get_cap_refs(
1679                         ceph_inode(req->r_old_dentry->d_parent->d_inode),
1680                         CEPH_CAP_PIN);
1681
1682         /* issue */
1683         mutex_lock(&mdsc->mutex);
1684         __register_request(mdsc, req, dir);
1685         __do_request(mdsc, req);
1686
1687         /* wait */
1688         if (!req->r_reply) {
1689                 mutex_unlock(&mdsc->mutex);
1690                 if (req->r_timeout) {
1691                         err = (long)wait_for_completion_interruptible_timeout(
1692                                 &req->r_completion, req->r_timeout);
1693                         if (err == 0)
1694                                 req->r_reply = ERR_PTR(-EIO);
1695                         else if (err < 0)
1696                                 req->r_reply = ERR_PTR(err);
1697                 } else {
1698                         err = wait_for_completion_interruptible(
1699                                 &req->r_completion);
1700                         if (err)
1701                                 req->r_reply = ERR_PTR(err);
1702                 }
1703                 mutex_lock(&mdsc->mutex);
1704         }
1705
1706         if (IS_ERR(req->r_reply)) {
1707                 err = PTR_ERR(req->r_reply);
1708                 req->r_reply = NULL;
1709
1710                 if (err == -ERESTARTSYS) {
1711                         /* aborted */
1712                         req->r_aborted = true;
1713
1714                         if (req->r_locked_dir &&
1715                             (req->r_op & CEPH_MDS_OP_WRITE)) {
1716                                 struct ceph_inode_info *ci =
1717                                         ceph_inode(req->r_locked_dir);
1718
1719                                 dout("aborted, clearing I_COMPLETE on %p\n",
1720                                      req->r_locked_dir);
1721                                 spin_lock(&req->r_locked_dir->i_lock);
1722                                 ci->i_ceph_flags &= ~CEPH_I_COMPLETE;
1723                                 ci->i_release_count++;
1724                                 spin_unlock(&req->r_locked_dir->i_lock);
1725                         }
1726                 } else {
1727                         /* clean up this request */
1728                         __unregister_request(mdsc, req);
1729                         if (!list_empty(&req->r_unsafe_item))
1730                                 list_del_init(&req->r_unsafe_item);
1731                         complete(&req->r_safe_completion);
1732                 }
1733         } else if (req->r_err) {
1734                 err = req->r_err;
1735         } else {
1736                 err = le32_to_cpu(req->r_reply_info.head->result);
1737         }
1738         mutex_unlock(&mdsc->mutex);
1739
1740         dout("do_request %p done, result %d\n", req, err);
1741         return err;
1742 }
1743
1744 /*
1745  * Handle mds reply.
1746  *
1747  * We take the session mutex and parse and process the reply immediately.
1748  * This preserves the logical ordering of replies, capabilities, etc., sent
1749  * by the MDS as they are applied to our local cache.
1750  */
1751 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
1752 {
1753         struct ceph_mds_client *mdsc = session->s_mdsc;
1754         struct ceph_mds_request *req;
1755         struct ceph_mds_reply_head *head = msg->front.iov_base;
1756         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
1757         u64 tid;
1758         int err, result;
1759         int mds = session->s_mds;
1760
1761         if (msg->front.iov_len < sizeof(*head)) {
1762                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
1763                 ceph_msg_dump(msg);
1764                 return;
1765         }
1766
1767         /* get request, session */
1768         tid = le64_to_cpu(msg->hdr.tid);
1769         mutex_lock(&mdsc->mutex);
1770         req = __lookup_request(mdsc, tid);
1771         if (!req) {
1772                 dout("handle_reply on unknown tid %llu\n", tid);
1773                 mutex_unlock(&mdsc->mutex);
1774                 return;
1775         }
1776         dout("handle_reply %p\n", req);
1777
1778         /* correct session? */
1779         if (!req->r_session && req->r_session != session) {
1780                 pr_err("mdsc_handle_reply got %llu on session mds%d"
1781                        " not mds%d\n", tid, session->s_mds,
1782                        req->r_session ? req->r_session->s_mds : -1);
1783                 mutex_unlock(&mdsc->mutex);
1784                 goto out;
1785         }
1786
1787         /* dup? */
1788         if ((req->r_got_unsafe && !head->safe) ||
1789             (req->r_got_safe && head->safe)) {
1790                 pr_warning("got a dup %s reply on %llu from mds%d\n",
1791                            head->safe ? "safe" : "unsafe", tid, mds);
1792                 mutex_unlock(&mdsc->mutex);
1793                 goto out;
1794         }
1795
1796         result = le32_to_cpu(head->result);
1797
1798         /*
1799          * Tolerate 2 consecutive ESTALEs from the same mds.
1800          * FIXME: we should be looking at the cap migrate_seq.
1801          */
1802         if (result == -ESTALE) {
1803                 req->r_direct_mode = USE_AUTH_MDS;
1804                 req->r_num_stale++;
1805                 if (req->r_num_stale <= 2) {
1806                         __do_request(mdsc, req);
1807                         mutex_unlock(&mdsc->mutex);
1808                         goto out;
1809                 }
1810         } else {
1811                 req->r_num_stale = 0;
1812         }
1813
1814         if (head->safe) {
1815                 req->r_got_safe = true;
1816                 __unregister_request(mdsc, req);
1817                 complete(&req->r_safe_completion);
1818
1819                 if (req->r_got_unsafe) {
1820                         /*
1821                          * We already handled the unsafe response, now do the
1822                          * cleanup.  No need to examine the response; the MDS
1823                          * doesn't include any result info in the safe
1824                          * response.  And even if it did, there is nothing
1825                          * useful we could do with a revised return value.
1826                          */
1827                         dout("got safe reply %llu, mds%d\n", tid, mds);
1828                         list_del_init(&req->r_unsafe_item);
1829
1830                         /* last unsafe request during umount? */
1831                         if (mdsc->stopping && !__get_oldest_req(mdsc))
1832                                 complete(&mdsc->safe_umount_waiters);
1833                         mutex_unlock(&mdsc->mutex);
1834                         goto out;
1835                 }
1836         }
1837
1838         BUG_ON(req->r_reply);
1839
1840         if (!head->safe) {
1841                 req->r_got_unsafe = true;
1842                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
1843         }
1844
1845         dout("handle_reply tid %lld result %d\n", tid, result);
1846         rinfo = &req->r_reply_info;
1847         err = parse_reply_info(msg, rinfo);
1848         mutex_unlock(&mdsc->mutex);
1849
1850         mutex_lock(&session->s_mutex);
1851         if (err < 0) {
1852                 pr_err("mdsc_handle_reply got corrupt reply mds%d\n", mds);
1853                 ceph_msg_dump(msg);
1854                 goto out_err;
1855         }
1856
1857         /* snap trace */
1858         if (rinfo->snapblob_len) {
1859                 down_write(&mdsc->snap_rwsem);
1860                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
1861                                rinfo->snapblob + rinfo->snapblob_len,
1862                                le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
1863                 downgrade_write(&mdsc->snap_rwsem);
1864         } else {
1865                 down_read(&mdsc->snap_rwsem);
1866         }
1867
1868         /* insert trace into our cache */
1869         err = ceph_fill_trace(mdsc->client->sb, req, req->r_session);
1870         if (err == 0) {
1871                 if (result == 0 && rinfo->dir_nr)
1872                         ceph_readdir_prepopulate(req, req->r_session);
1873                 ceph_unreserve_caps(&req->r_caps_reservation);
1874         }
1875
1876         up_read(&mdsc->snap_rwsem);
1877 out_err:
1878         if (err) {
1879                 req->r_err = err;
1880         } else {
1881                 req->r_reply = msg;
1882                 ceph_msg_get(msg);
1883         }
1884
1885         add_cap_releases(mdsc, req->r_session, -1);
1886         mutex_unlock(&session->s_mutex);
1887
1888         /* kick calling process */
1889         complete_request(mdsc, req);
1890 out:
1891         ceph_mdsc_put_request(req);
1892         return;
1893 }
1894
1895
1896
1897 /*
1898  * handle mds notification that our request has been forwarded.
1899  */
1900 static void handle_forward(struct ceph_mds_client *mdsc,
1901                            struct ceph_mds_session *session,
1902                            struct ceph_msg *msg)
1903 {
1904         struct ceph_mds_request *req;
1905         u64 tid = le64_to_cpu(msg->hdr.tid);
1906         u32 next_mds;
1907         u32 fwd_seq;
1908         int err = -EINVAL;
1909         void *p = msg->front.iov_base;
1910         void *end = p + msg->front.iov_len;
1911
1912         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1913         next_mds = ceph_decode_32(&p);
1914         fwd_seq = ceph_decode_32(&p);
1915
1916         mutex_lock(&mdsc->mutex);
1917         req = __lookup_request(mdsc, tid);
1918         if (!req) {
1919                 dout("forward %llu to mds%d - req dne\n", tid, next_mds);
1920                 goto out;  /* dup reply? */
1921         }
1922
1923         if (fwd_seq <= req->r_num_fwd) {
1924                 dout("forward %llu to mds%d - old seq %d <= %d\n",
1925                      tid, next_mds, req->r_num_fwd, fwd_seq);
1926         } else {
1927                 /* resend. forward race not possible; mds would drop */
1928                 dout("forward %llu to mds%d (we resend)\n", tid, next_mds);
1929                 req->r_num_fwd = fwd_seq;
1930                 req->r_resend_mds = next_mds;
1931                 put_request_session(req);
1932                 __do_request(mdsc, req);
1933         }
1934         ceph_mdsc_put_request(req);
1935 out:
1936         mutex_unlock(&mdsc->mutex);
1937         return;
1938
1939 bad:
1940         pr_err("mdsc_handle_forward decode error err=%d\n", err);
1941 }
1942
1943 /*
1944  * handle a mds session control message
1945  */
1946 static void handle_session(struct ceph_mds_session *session,
1947                            struct ceph_msg *msg)
1948 {
1949         struct ceph_mds_client *mdsc = session->s_mdsc;
1950         u32 op;
1951         u64 seq;
1952         int mds = session->s_mds;
1953         struct ceph_mds_session_head *h = msg->front.iov_base;
1954         int wake = 0;
1955
1956         /* decode */
1957         if (msg->front.iov_len != sizeof(*h))
1958                 goto bad;
1959         op = le32_to_cpu(h->op);
1960         seq = le64_to_cpu(h->seq);
1961
1962         mutex_lock(&mdsc->mutex);
1963         if (op == CEPH_SESSION_CLOSE)
1964                 __unregister_session(mdsc, session);
1965         /* FIXME: this ttl calculation is generous */
1966         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
1967         mutex_unlock(&mdsc->mutex);
1968
1969         mutex_lock(&session->s_mutex);
1970
1971         dout("handle_session mds%d %s %p state %s seq %llu\n",
1972              mds, ceph_session_op_name(op), session,
1973              session_state_name(session->s_state), seq);
1974
1975         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
1976                 session->s_state = CEPH_MDS_SESSION_OPEN;
1977                 pr_info("mds%d came back\n", session->s_mds);
1978         }
1979
1980         switch (op) {
1981         case CEPH_SESSION_OPEN:
1982                 session->s_state = CEPH_MDS_SESSION_OPEN;
1983                 renewed_caps(mdsc, session, 0);
1984                 wake = 1;
1985                 if (mdsc->stopping)
1986                         __close_session(mdsc, session);
1987                 break;
1988
1989         case CEPH_SESSION_RENEWCAPS:
1990                 if (session->s_renew_seq == seq)
1991                         renewed_caps(mdsc, session, 1);
1992                 break;
1993
1994         case CEPH_SESSION_CLOSE:
1995                 remove_session_caps(session);
1996                 wake = 1; /* for good measure */
1997                 complete(&mdsc->session_close_waiters);
1998                 kick_requests(mdsc, mds, 0);      /* cur only */
1999                 break;
2000
2001         case CEPH_SESSION_STALE:
2002                 pr_info("mds%d caps went stale, renewing\n",
2003                         session->s_mds);
2004                 spin_lock(&session->s_cap_lock);
2005                 session->s_cap_gen++;
2006                 session->s_cap_ttl = 0;
2007                 spin_unlock(&session->s_cap_lock);
2008                 send_renew_caps(mdsc, session);
2009                 break;
2010
2011         case CEPH_SESSION_RECALL_STATE:
2012                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2013                 break;
2014
2015         default:
2016                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2017                 WARN_ON(1);
2018         }
2019
2020         mutex_unlock(&session->s_mutex);
2021         if (wake) {
2022                 mutex_lock(&mdsc->mutex);
2023                 __wake_requests(mdsc, &session->s_waiting);
2024                 mutex_unlock(&mdsc->mutex);
2025         }
2026         return;
2027
2028 bad:
2029         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2030                (int)msg->front.iov_len);
2031         ceph_msg_dump(msg);
2032         return;
2033 }
2034
2035
2036 /*
2037  * called under session->mutex.
2038  */
2039 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2040                                    struct ceph_mds_session *session)
2041 {
2042         struct ceph_mds_request *req, *nreq;
2043         int err;
2044
2045         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2046
2047         mutex_lock(&mdsc->mutex);
2048         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2049                 err = __prepare_send_request(mdsc, req, session->s_mds);
2050                 if (!err) {
2051                         ceph_msg_get(req->r_request);
2052                         ceph_con_send(&session->s_con, req->r_request);
2053                 }
2054         }
2055         mutex_unlock(&mdsc->mutex);
2056 }
2057
2058 /*
2059  * Encode information about a cap for a reconnect with the MDS.
2060  */
2061 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2062                           void *arg)
2063 {
2064         struct ceph_mds_cap_reconnect rec;
2065         struct ceph_inode_info *ci;
2066         struct ceph_pagelist *pagelist = arg;
2067         char *path;
2068         int pathlen, err;
2069         u64 pathbase;
2070         struct dentry *dentry;
2071
2072         ci = cap->ci;
2073
2074         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2075              inode, ceph_vinop(inode), cap, cap->cap_id,
2076              ceph_cap_string(cap->issued));
2077         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2078         if (err)
2079                 return err;
2080
2081         dentry = d_find_alias(inode);
2082         if (dentry) {
2083                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2084                 if (IS_ERR(path)) {
2085                         err = PTR_ERR(path);
2086                         BUG_ON(err);
2087                 }
2088         } else {
2089                 path = NULL;
2090                 pathlen = 0;
2091         }
2092         err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2093         if (err)
2094                 goto out;
2095
2096         spin_lock(&inode->i_lock);
2097         cap->seq = 0;        /* reset cap seq */
2098         cap->issue_seq = 0;  /* and issue_seq */
2099         rec.cap_id = cpu_to_le64(cap->cap_id);
2100         rec.pathbase = cpu_to_le64(pathbase);
2101         rec.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2102         rec.issued = cpu_to_le32(cap->issued);
2103         rec.size = cpu_to_le64(inode->i_size);
2104         ceph_encode_timespec(&rec.mtime, &inode->i_mtime);
2105         ceph_encode_timespec(&rec.atime, &inode->i_atime);
2106         rec.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2107         spin_unlock(&inode->i_lock);
2108
2109         err = ceph_pagelist_append(pagelist, &rec, sizeof(rec));
2110
2111 out:
2112         kfree(path);
2113         dput(dentry);
2114         return err;
2115 }
2116
2117
2118 /*
2119  * If an MDS fails and recovers, clients need to reconnect in order to
2120  * reestablish shared state.  This includes all caps issued through
2121  * this session _and_ the snap_realm hierarchy.  Because it's not
2122  * clear which snap realms the mds cares about, we send everything we
2123  * know about.. that ensures we'll then get any new info the
2124  * recovering MDS might have.
2125  *
2126  * This is a relatively heavyweight operation, but it's rare.
2127  *
2128  * called with mdsc->mutex held.
2129  */
2130 static void send_mds_reconnect(struct ceph_mds_client *mdsc, int mds)
2131 {
2132         struct ceph_mds_session *session = NULL;
2133         struct ceph_msg *reply;
2134         struct rb_node *p;
2135         int err;
2136         struct ceph_pagelist *pagelist;
2137
2138         pr_info("reconnect to recovering mds%d\n", mds);
2139
2140         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2141         if (!pagelist)
2142                 goto fail_nopagelist;
2143         ceph_pagelist_init(pagelist);
2144
2145         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, 0, 0, NULL);
2146         if (IS_ERR(reply)) {
2147                 err = PTR_ERR(reply);
2148                 goto fail_nomsg;
2149         }
2150
2151         /* find session */
2152         session = __ceph_lookup_mds_session(mdsc, mds);
2153         mutex_unlock(&mdsc->mutex);    /* drop lock for duration */
2154
2155         if (session) {
2156                 mutex_lock(&session->s_mutex);
2157
2158                 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2159                 session->s_seq = 0;
2160
2161                 ceph_con_open(&session->s_con,
2162                               ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2163
2164                 /* replay unsafe requests */
2165                 replay_unsafe_requests(mdsc, session);
2166         } else {
2167                 dout("no session for mds%d, will send short reconnect\n",
2168                      mds);
2169         }
2170
2171         down_read(&mdsc->snap_rwsem);
2172
2173         if (!session)
2174                 goto send;
2175         dout("session %p state %s\n", session,
2176              session_state_name(session->s_state));
2177
2178         /* traverse this session's caps */
2179         err = ceph_pagelist_encode_32(pagelist, session->s_nr_caps);
2180         if (err)
2181                 goto fail;
2182         err = iterate_session_caps(session, encode_caps_cb, pagelist);
2183         if (err < 0)
2184                 goto out;
2185
2186         /*
2187          * snaprealms.  we provide mds with the ino, seq (version), and
2188          * parent for all of our realms.  If the mds has any newer info,
2189          * it will tell us.
2190          */
2191         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2192                 struct ceph_snap_realm *realm =
2193                         rb_entry(p, struct ceph_snap_realm, node);
2194                 struct ceph_mds_snaprealm_reconnect sr_rec;
2195
2196                 dout(" adding snap realm %llx seq %lld parent %llx\n",
2197                      realm->ino, realm->seq, realm->parent_ino);
2198                 sr_rec.ino = cpu_to_le64(realm->ino);
2199                 sr_rec.seq = cpu_to_le64(realm->seq);
2200                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2201                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2202                 if (err)
2203                         goto fail;
2204         }
2205
2206 send:
2207         reply->pagelist = pagelist;
2208         reply->hdr.data_len = cpu_to_le32(pagelist->length);
2209         reply->nr_pages = calc_pages_for(0, pagelist->length);
2210         ceph_con_send(&session->s_con, reply);
2211
2212         if (session) {
2213                 session->s_state = CEPH_MDS_SESSION_OPEN;
2214                 __wake_requests(mdsc, &session->s_waiting);
2215         }
2216
2217 out:
2218         up_read(&mdsc->snap_rwsem);
2219         if (session) {
2220                 mutex_unlock(&session->s_mutex);
2221                 ceph_put_mds_session(session);
2222         }
2223         mutex_lock(&mdsc->mutex);
2224         return;
2225
2226 fail:
2227         ceph_msg_put(reply);
2228 fail_nomsg:
2229         ceph_pagelist_release(pagelist);
2230         kfree(pagelist);
2231 fail_nopagelist:
2232         pr_err("ENOMEM preparing reconnect for mds%d\n", mds);
2233         goto out;
2234 }
2235
2236
2237 /*
2238  * compare old and new mdsmaps, kicking requests
2239  * and closing out old connections as necessary
2240  *
2241  * called under mdsc->mutex.
2242  */
2243 static void check_new_map(struct ceph_mds_client *mdsc,
2244                           struct ceph_mdsmap *newmap,
2245                           struct ceph_mdsmap *oldmap)
2246 {
2247         int i;
2248         int oldstate, newstate;
2249         struct ceph_mds_session *s;
2250
2251         dout("check_new_map new %u old %u\n",
2252              newmap->m_epoch, oldmap->m_epoch);
2253
2254         for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2255                 if (mdsc->sessions[i] == NULL)
2256                         continue;
2257                 s = mdsc->sessions[i];
2258                 oldstate = ceph_mdsmap_get_state(oldmap, i);
2259                 newstate = ceph_mdsmap_get_state(newmap, i);
2260
2261                 dout("check_new_map mds%d state %s -> %s (session %s)\n",
2262                      i, ceph_mds_state_name(oldstate),
2263                      ceph_mds_state_name(newstate),
2264                      session_state_name(s->s_state));
2265
2266                 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
2267                            ceph_mdsmap_get_addr(newmap, i),
2268                            sizeof(struct ceph_entity_addr))) {
2269                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2270                                 /* the session never opened, just close it
2271                                  * out now */
2272                                 __wake_requests(mdsc, &s->s_waiting);
2273                                 __unregister_session(mdsc, s);
2274                         } else {
2275                                 /* just close it */
2276                                 mutex_unlock(&mdsc->mutex);
2277                                 mutex_lock(&s->s_mutex);
2278                                 mutex_lock(&mdsc->mutex);
2279                                 ceph_con_close(&s->s_con);
2280                                 mutex_unlock(&s->s_mutex);
2281                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2282                         }
2283
2284                         /* kick any requests waiting on the recovering mds */
2285                         kick_requests(mdsc, i, 1);
2286                 } else if (oldstate == newstate) {
2287                         continue;  /* nothing new with this mds */
2288                 }
2289
2290                 /*
2291                  * send reconnect?
2292                  */
2293                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2294                     newstate >= CEPH_MDS_STATE_RECONNECT)
2295                         send_mds_reconnect(mdsc, i);
2296
2297                 /*
2298                  * kick requests on any mds that has gone active.
2299                  *
2300                  * kick requests on cur or forwarder: we may have sent
2301                  * the request to mds1, mds1 told us it forwarded it
2302                  * to mds2, but then we learn mds1 failed and can't be
2303                  * sure it successfully forwarded our request before
2304                  * it died.
2305                  */
2306                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2307                     newstate >= CEPH_MDS_STATE_ACTIVE) {
2308                         pr_info("mds%d reconnect completed\n", s->s_mds);
2309                         kick_requests(mdsc, i, 1);
2310                         ceph_kick_flushing_caps(mdsc, s);
2311                         wake_up_session_caps(s, 1);
2312                 }
2313         }
2314 }
2315
2316
2317
2318 /*
2319  * leases
2320  */
2321
2322 /*
2323  * caller must hold session s_mutex, dentry->d_lock
2324  */
2325 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2326 {
2327         struct ceph_dentry_info *di = ceph_dentry(dentry);
2328
2329         ceph_put_mds_session(di->lease_session);
2330         di->lease_session = NULL;
2331 }
2332
2333 static void handle_lease(struct ceph_mds_client *mdsc,
2334                          struct ceph_mds_session *session,
2335                          struct ceph_msg *msg)
2336 {
2337         struct super_block *sb = mdsc->client->sb;
2338         struct inode *inode;
2339         struct ceph_inode_info *ci;
2340         struct dentry *parent, *dentry;
2341         struct ceph_dentry_info *di;
2342         int mds = session->s_mds;
2343         struct ceph_mds_lease *h = msg->front.iov_base;
2344         struct ceph_vino vino;
2345         int mask;
2346         struct qstr dname;
2347         int release = 0;
2348
2349         dout("handle_lease from mds%d\n", mds);
2350
2351         /* decode */
2352         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2353                 goto bad;
2354         vino.ino = le64_to_cpu(h->ino);
2355         vino.snap = CEPH_NOSNAP;
2356         mask = le16_to_cpu(h->mask);
2357         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2358         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2359         if (dname.len != get_unaligned_le32(h+1))
2360                 goto bad;
2361
2362         mutex_lock(&session->s_mutex);
2363         session->s_seq++;
2364
2365         /* lookup inode */
2366         inode = ceph_find_inode(sb, vino);
2367         dout("handle_lease '%s', mask %d, ino %llx %p\n",
2368              ceph_lease_op_name(h->action), mask, vino.ino, inode);
2369         if (inode == NULL) {
2370                 dout("handle_lease no inode %llx\n", vino.ino);
2371                 goto release;
2372         }
2373         ci = ceph_inode(inode);
2374
2375         /* dentry */
2376         parent = d_find_alias(inode);
2377         if (!parent) {
2378                 dout("no parent dentry on inode %p\n", inode);
2379                 WARN_ON(1);
2380                 goto release;  /* hrm... */
2381         }
2382         dname.hash = full_name_hash(dname.name, dname.len);
2383         dentry = d_lookup(parent, &dname);
2384         dput(parent);
2385         if (!dentry)
2386                 goto release;
2387
2388         spin_lock(&dentry->d_lock);
2389         di = ceph_dentry(dentry);
2390         switch (h->action) {
2391         case CEPH_MDS_LEASE_REVOKE:
2392                 if (di && di->lease_session == session) {
2393                         h->seq = cpu_to_le32(di->lease_seq);
2394                         __ceph_mdsc_drop_dentry_lease(dentry);
2395                 }
2396                 release = 1;
2397                 break;
2398
2399         case CEPH_MDS_LEASE_RENEW:
2400                 if (di && di->lease_session == session &&
2401                     di->lease_gen == session->s_cap_gen &&
2402                     di->lease_renew_from &&
2403                     di->lease_renew_after == 0) {
2404                         unsigned long duration =
2405                                 le32_to_cpu(h->duration_ms) * HZ / 1000;
2406
2407                         di->lease_seq = le32_to_cpu(h->seq);
2408                         dentry->d_time = di->lease_renew_from + duration;
2409                         di->lease_renew_after = di->lease_renew_from +
2410                                 (duration >> 1);
2411                         di->lease_renew_from = 0;
2412                 }
2413                 break;
2414         }
2415         spin_unlock(&dentry->d_lock);
2416         dput(dentry);
2417
2418         if (!release)
2419                 goto out;
2420
2421 release:
2422         /* let's just reuse the same message */
2423         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2424         ceph_msg_get(msg);
2425         ceph_con_send(&session->s_con, msg);
2426
2427 out:
2428         iput(inode);
2429         mutex_unlock(&session->s_mutex);
2430         return;
2431
2432 bad:
2433         pr_err("corrupt lease message\n");
2434         ceph_msg_dump(msg);
2435 }
2436
2437 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
2438                               struct inode *inode,
2439                               struct dentry *dentry, char action,
2440                               u32 seq)
2441 {
2442         struct ceph_msg *msg;
2443         struct ceph_mds_lease *lease;
2444         int len = sizeof(*lease) + sizeof(u32);
2445         int dnamelen = 0;
2446
2447         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2448              inode, dentry, ceph_lease_op_name(action), session->s_mds);
2449         dnamelen = dentry->d_name.len;
2450         len += dnamelen;
2451
2452         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, 0, 0, NULL);
2453         if (IS_ERR(msg))
2454                 return;
2455         lease = msg->front.iov_base;
2456         lease->action = action;
2457         lease->mask = cpu_to_le16(CEPH_LOCK_DN);
2458         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
2459         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
2460         lease->seq = cpu_to_le32(seq);
2461         put_unaligned_le32(dnamelen, lease + 1);
2462         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
2463
2464         /*
2465          * if this is a preemptive lease RELEASE, no need to
2466          * flush request stream, since the actual request will
2467          * soon follow.
2468          */
2469         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
2470
2471         ceph_con_send(&session->s_con, msg);
2472 }
2473
2474 /*
2475  * Preemptively release a lease we expect to invalidate anyway.
2476  * Pass @inode always, @dentry is optional.
2477  */
2478 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2479                              struct dentry *dentry, int mask)
2480 {
2481         struct ceph_dentry_info *di;
2482         struct ceph_mds_session *session;
2483         u32 seq;
2484
2485         BUG_ON(inode == NULL);
2486         BUG_ON(dentry == NULL);
2487         BUG_ON(mask != CEPH_LOCK_DN);
2488
2489         /* is dentry lease valid? */
2490         spin_lock(&dentry->d_lock);
2491         di = ceph_dentry(dentry);
2492         if (!di || !di->lease_session ||
2493             di->lease_session->s_mds < 0 ||
2494             di->lease_gen != di->lease_session->s_cap_gen ||
2495             !time_before(jiffies, dentry->d_time)) {
2496                 dout("lease_release inode %p dentry %p -- "
2497                      "no lease on %d\n",
2498                      inode, dentry, mask);
2499                 spin_unlock(&dentry->d_lock);
2500                 return;
2501         }
2502
2503         /* we do have a lease on this dentry; note mds and seq */
2504         session = ceph_get_mds_session(di->lease_session);
2505         seq = di->lease_seq;
2506         __ceph_mdsc_drop_dentry_lease(dentry);
2507         spin_unlock(&dentry->d_lock);
2508
2509         dout("lease_release inode %p dentry %p mask %d to mds%d\n",
2510              inode, dentry, mask, session->s_mds);
2511         ceph_mdsc_lease_send_msg(session, inode, dentry,
2512                                  CEPH_MDS_LEASE_RELEASE, seq);
2513         ceph_put_mds_session(session);
2514 }
2515
2516 /*
2517  * drop all leases (and dentry refs) in preparation for umount
2518  */
2519 static void drop_leases(struct ceph_mds_client *mdsc)
2520 {
2521         int i;
2522
2523         dout("drop_leases\n");
2524         mutex_lock(&mdsc->mutex);
2525         for (i = 0; i < mdsc->max_sessions; i++) {
2526                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2527                 if (!s)
2528                         continue;
2529                 mutex_unlock(&mdsc->mutex);
2530                 mutex_lock(&s->s_mutex);
2531                 mutex_unlock(&s->s_mutex);
2532                 ceph_put_mds_session(s);
2533                 mutex_lock(&mdsc->mutex);
2534         }
2535         mutex_unlock(&mdsc->mutex);
2536 }
2537
2538
2539
2540 /*
2541  * delayed work -- periodically trim expired leases, renew caps with mds
2542  */
2543 static void schedule_delayed(struct ceph_mds_client *mdsc)
2544 {
2545         int delay = 5;
2546         unsigned hz = round_jiffies_relative(HZ * delay);
2547         schedule_delayed_work(&mdsc->delayed_work, hz);
2548 }
2549
2550 static void delayed_work(struct work_struct *work)
2551 {
2552         int i;
2553         struct ceph_mds_client *mdsc =
2554                 container_of(work, struct ceph_mds_client, delayed_work.work);
2555         int renew_interval;
2556         int renew_caps;
2557
2558         dout("mdsc delayed_work\n");
2559         ceph_check_delayed_caps(mdsc);
2560
2561         mutex_lock(&mdsc->mutex);
2562         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
2563         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
2564                                    mdsc->last_renew_caps);
2565         if (renew_caps)
2566                 mdsc->last_renew_caps = jiffies;
2567
2568         for (i = 0; i < mdsc->max_sessions; i++) {
2569                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2570                 if (s == NULL)
2571                         continue;
2572                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
2573                         dout("resending session close request for mds%d\n",
2574                              s->s_mds);
2575                         request_close_session(mdsc, s);
2576                         ceph_put_mds_session(s);
2577                         continue;
2578                 }
2579                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
2580                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
2581                                 s->s_state = CEPH_MDS_SESSION_HUNG;
2582                                 pr_info("mds%d hung\n", s->s_mds);
2583                         }
2584                 }
2585                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
2586                         /* this mds is failed or recovering, just wait */
2587                         ceph_put_mds_session(s);
2588                         continue;
2589                 }
2590                 mutex_unlock(&mdsc->mutex);
2591
2592                 mutex_lock(&s->s_mutex);
2593                 if (renew_caps)
2594                         send_renew_caps(mdsc, s);
2595                 else
2596                         ceph_con_keepalive(&s->s_con);
2597                 add_cap_releases(mdsc, s, -1);
2598                 send_cap_releases(mdsc, s);
2599                 mutex_unlock(&s->s_mutex);
2600                 ceph_put_mds_session(s);
2601
2602                 mutex_lock(&mdsc->mutex);
2603         }
2604         mutex_unlock(&mdsc->mutex);
2605
2606         schedule_delayed(mdsc);
2607 }
2608
2609
2610 int ceph_mdsc_init(struct ceph_mds_client *mdsc, struct ceph_client *client)
2611 {
2612         mdsc->client = client;
2613         mutex_init(&mdsc->mutex);
2614         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
2615         init_completion(&mdsc->safe_umount_waiters);
2616         init_completion(&mdsc->session_close_waiters);
2617         INIT_LIST_HEAD(&mdsc->waiting_for_map);
2618         mdsc->sessions = NULL;
2619         mdsc->max_sessions = 0;
2620         mdsc->stopping = 0;
2621         init_rwsem(&mdsc->snap_rwsem);
2622         mdsc->snap_realms = RB_ROOT;
2623         INIT_LIST_HEAD(&mdsc->snap_empty);
2624         spin_lock_init(&mdsc->snap_empty_lock);
2625         mdsc->last_tid = 0;
2626         mdsc->request_tree = RB_ROOT;
2627         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
2628         mdsc->last_renew_caps = jiffies;
2629         INIT_LIST_HEAD(&mdsc->cap_delay_list);
2630         spin_lock_init(&mdsc->cap_delay_lock);
2631         INIT_LIST_HEAD(&mdsc->snap_flush_list);
2632         spin_lock_init(&mdsc->snap_flush_lock);
2633         mdsc->cap_flush_seq = 0;
2634         INIT_LIST_HEAD(&mdsc->cap_dirty);
2635         mdsc->num_cap_flushing = 0;
2636         spin_lock_init(&mdsc->cap_dirty_lock);
2637         init_waitqueue_head(&mdsc->cap_flushing_wq);
2638         spin_lock_init(&mdsc->dentry_lru_lock);
2639         INIT_LIST_HEAD(&mdsc->dentry_lru);
2640         return 0;
2641 }
2642
2643 /*
2644  * Wait for safe replies on open mds requests.  If we time out, drop
2645  * all requests from the tree to avoid dangling dentry refs.
2646  */
2647 static void wait_requests(struct ceph_mds_client *mdsc)
2648 {
2649         struct ceph_mds_request *req;
2650         struct ceph_client *client = mdsc->client;
2651
2652         mutex_lock(&mdsc->mutex);
2653         if (__get_oldest_req(mdsc)) {
2654                 mutex_unlock(&mdsc->mutex);
2655
2656                 dout("wait_requests waiting for requests\n");
2657                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
2658                                     client->mount_args->mount_timeout * HZ);
2659
2660                 /* tear down remaining requests */
2661                 mutex_lock(&mdsc->mutex);
2662                 while ((req = __get_oldest_req(mdsc))) {
2663                         dout("wait_requests timed out on tid %llu\n",
2664                              req->r_tid);
2665                         __unregister_request(mdsc, req);
2666                 }
2667         }
2668         mutex_unlock(&mdsc->mutex);
2669         dout("wait_requests done\n");
2670 }
2671
2672 /*
2673  * called before mount is ro, and before dentries are torn down.
2674  * (hmm, does this still race with new lookups?)
2675  */
2676 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
2677 {
2678         dout("pre_umount\n");
2679         mdsc->stopping = 1;
2680
2681         drop_leases(mdsc);
2682         ceph_flush_dirty_caps(mdsc);
2683         wait_requests(mdsc);
2684 }
2685
2686 /*
2687  * wait for all write mds requests to flush.
2688  */
2689 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
2690 {
2691         struct ceph_mds_request *req = NULL, *nextreq;
2692         struct rb_node *n;
2693
2694         mutex_lock(&mdsc->mutex);
2695         dout("wait_unsafe_requests want %lld\n", want_tid);
2696 restart:
2697         req = __get_oldest_req(mdsc);
2698         while (req && req->r_tid <= want_tid) {
2699                 /* find next request */
2700                 n = rb_next(&req->r_node);
2701                 if (n)
2702                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
2703                 else
2704                         nextreq = NULL;
2705                 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
2706                         /* write op */
2707                         ceph_mdsc_get_request(req);
2708                         if (nextreq)
2709                                 ceph_mdsc_get_request(nextreq);
2710                         mutex_unlock(&mdsc->mutex);
2711                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
2712                              req->r_tid, want_tid);
2713                         wait_for_completion(&req->r_safe_completion);
2714                         mutex_lock(&mdsc->mutex);
2715                         ceph_mdsc_put_request(req);
2716                         if (!nextreq)
2717                                 break;  /* next dne before, so we're done! */
2718                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
2719                                 /* next request was removed from tree */
2720                                 ceph_mdsc_put_request(nextreq);
2721                                 goto restart;
2722                         }
2723                         ceph_mdsc_put_request(nextreq);  /* won't go away */
2724                 }
2725                 req = nextreq;
2726         }
2727         mutex_unlock(&mdsc->mutex);
2728         dout("wait_unsafe_requests done\n");
2729 }
2730
2731 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
2732 {
2733         u64 want_tid, want_flush;
2734
2735         dout("sync\n");
2736         mutex_lock(&mdsc->mutex);
2737         want_tid = mdsc->last_tid;
2738         want_flush = mdsc->cap_flush_seq;
2739         mutex_unlock(&mdsc->mutex);
2740         dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
2741
2742         ceph_flush_dirty_caps(mdsc);
2743
2744         wait_unsafe_requests(mdsc, want_tid);
2745         wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
2746 }
2747
2748
2749 /*
2750  * called after sb is ro.
2751  */
2752 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
2753 {
2754         struct ceph_mds_session *session;
2755         int i;
2756         int n;
2757         struct ceph_client *client = mdsc->client;
2758         unsigned long started, timeout = client->mount_args->mount_timeout * HZ;
2759
2760         dout("close_sessions\n");
2761
2762         mutex_lock(&mdsc->mutex);
2763
2764         /* close sessions */
2765         started = jiffies;
2766         while (time_before(jiffies, started + timeout)) {
2767                 dout("closing sessions\n");
2768                 n = 0;
2769                 for (i = 0; i < mdsc->max_sessions; i++) {
2770                         session = __ceph_lookup_mds_session(mdsc, i);
2771                         if (!session)
2772                                 continue;
2773                         mutex_unlock(&mdsc->mutex);
2774                         mutex_lock(&session->s_mutex);
2775                         __close_session(mdsc, session);
2776                         mutex_unlock(&session->s_mutex);
2777                         ceph_put_mds_session(session);
2778                         mutex_lock(&mdsc->mutex);
2779                         n++;
2780                 }
2781                 if (n == 0)
2782                         break;
2783
2784                 if (client->mount_state == CEPH_MOUNT_SHUTDOWN)
2785                         break;
2786
2787                 dout("waiting for sessions to close\n");
2788                 mutex_unlock(&mdsc->mutex);
2789                 wait_for_completion_timeout(&mdsc->session_close_waiters,
2790                                             timeout);
2791                 mutex_lock(&mdsc->mutex);
2792         }
2793
2794         /* tear down remaining sessions */
2795         for (i = 0; i < mdsc->max_sessions; i++) {
2796                 if (mdsc->sessions[i]) {
2797                         session = get_session(mdsc->sessions[i]);
2798                         __unregister_session(mdsc, session);
2799                         mutex_unlock(&mdsc->mutex);
2800                         mutex_lock(&session->s_mutex);
2801                         remove_session_caps(session);
2802                         mutex_unlock(&session->s_mutex);
2803                         ceph_put_mds_session(session);
2804                         mutex_lock(&mdsc->mutex);
2805                 }
2806         }
2807
2808         WARN_ON(!list_empty(&mdsc->cap_delay_list));
2809
2810         mutex_unlock(&mdsc->mutex);
2811
2812         ceph_cleanup_empty_realms(mdsc);
2813
2814         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2815
2816         dout("stopped\n");
2817 }
2818
2819 void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
2820 {
2821         dout("stop\n");
2822         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2823         if (mdsc->mdsmap)
2824                 ceph_mdsmap_destroy(mdsc->mdsmap);
2825         kfree(mdsc->sessions);
2826 }
2827
2828
2829 /*
2830  * handle mds map update.
2831  */
2832 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
2833 {
2834         u32 epoch;
2835         u32 maplen;
2836         void *p = msg->front.iov_base;
2837         void *end = p + msg->front.iov_len;
2838         struct ceph_mdsmap *newmap, *oldmap;
2839         struct ceph_fsid fsid;
2840         int err = -EINVAL;
2841
2842         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
2843         ceph_decode_copy(&p, &fsid, sizeof(fsid));
2844         if (ceph_check_fsid(mdsc->client, &fsid) < 0)
2845                 return;
2846         epoch = ceph_decode_32(&p);
2847         maplen = ceph_decode_32(&p);
2848         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
2849
2850         /* do we need it? */
2851         ceph_monc_got_mdsmap(&mdsc->client->monc, epoch);
2852         mutex_lock(&mdsc->mutex);
2853         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
2854                 dout("handle_map epoch %u <= our %u\n",
2855                      epoch, mdsc->mdsmap->m_epoch);
2856                 mutex_unlock(&mdsc->mutex);
2857                 return;
2858         }
2859
2860         newmap = ceph_mdsmap_decode(&p, end);
2861         if (IS_ERR(newmap)) {
2862                 err = PTR_ERR(newmap);
2863                 goto bad_unlock;
2864         }
2865
2866         /* swap into place */
2867         if (mdsc->mdsmap) {
2868                 oldmap = mdsc->mdsmap;
2869                 mdsc->mdsmap = newmap;
2870                 check_new_map(mdsc, newmap, oldmap);
2871                 ceph_mdsmap_destroy(oldmap);
2872         } else {
2873                 mdsc->mdsmap = newmap;  /* first mds map */
2874         }
2875         mdsc->client->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
2876
2877         __wake_requests(mdsc, &mdsc->waiting_for_map);
2878
2879         mutex_unlock(&mdsc->mutex);
2880         schedule_delayed(mdsc);
2881         return;
2882
2883 bad_unlock:
2884         mutex_unlock(&mdsc->mutex);
2885 bad:
2886         pr_err("error decoding mdsmap %d\n", err);
2887         return;
2888 }
2889
2890 static struct ceph_connection *con_get(struct ceph_connection *con)
2891 {
2892         struct ceph_mds_session *s = con->private;
2893
2894         if (get_session(s)) {
2895                 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
2896                 return con;
2897         }
2898         dout("mdsc con_get %p FAIL\n", s);
2899         return NULL;
2900 }
2901
2902 static void con_put(struct ceph_connection *con)
2903 {
2904         struct ceph_mds_session *s = con->private;
2905
2906         ceph_put_mds_session(s);
2907         dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref));
2908 }
2909
2910 /*
2911  * if the client is unresponsive for long enough, the mds will kill
2912  * the session entirely.
2913  */
2914 static void peer_reset(struct ceph_connection *con)
2915 {
2916         struct ceph_mds_session *s = con->private;
2917
2918         pr_err("mds%d gave us the boot.  IMPLEMENT RECONNECT.\n",
2919                s->s_mds);
2920 }
2921
2922 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2923 {
2924         struct ceph_mds_session *s = con->private;
2925         struct ceph_mds_client *mdsc = s->s_mdsc;
2926         int type = le16_to_cpu(msg->hdr.type);
2927
2928         mutex_lock(&mdsc->mutex);
2929         if (__verify_registered_session(mdsc, s) < 0) {
2930                 mutex_unlock(&mdsc->mutex);
2931                 goto out;
2932         }
2933         mutex_unlock(&mdsc->mutex);
2934
2935         switch (type) {
2936         case CEPH_MSG_MDS_MAP:
2937                 ceph_mdsc_handle_map(mdsc, msg);
2938                 break;
2939         case CEPH_MSG_CLIENT_SESSION:
2940                 handle_session(s, msg);
2941                 break;
2942         case CEPH_MSG_CLIENT_REPLY:
2943                 handle_reply(s, msg);
2944                 break;
2945         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
2946                 handle_forward(mdsc, s, msg);
2947                 break;
2948         case CEPH_MSG_CLIENT_CAPS:
2949                 ceph_handle_caps(s, msg);
2950                 break;
2951         case CEPH_MSG_CLIENT_SNAP:
2952                 ceph_handle_snap(mdsc, s, msg);
2953                 break;
2954         case CEPH_MSG_CLIENT_LEASE:
2955                 handle_lease(mdsc, s, msg);
2956                 break;
2957
2958         default:
2959                 pr_err("received unknown message type %d %s\n", type,
2960                        ceph_msg_type_name(type));
2961         }
2962 out:
2963         ceph_msg_put(msg);
2964 }
2965
2966 /*
2967  * authentication
2968  */
2969 static int get_authorizer(struct ceph_connection *con,
2970                           void **buf, int *len, int *proto,
2971                           void **reply_buf, int *reply_len, int force_new)
2972 {
2973         struct ceph_mds_session *s = con->private;
2974         struct ceph_mds_client *mdsc = s->s_mdsc;
2975         struct ceph_auth_client *ac = mdsc->client->monc.auth;
2976         int ret = 0;
2977
2978         if (force_new && s->s_authorizer) {
2979                 ac->ops->destroy_authorizer(ac, s->s_authorizer);
2980                 s->s_authorizer = NULL;
2981         }
2982         if (s->s_authorizer == NULL) {
2983                 if (ac->ops->create_authorizer) {
2984                         ret = ac->ops->create_authorizer(
2985                                 ac, CEPH_ENTITY_TYPE_MDS,
2986                                 &s->s_authorizer,
2987                                 &s->s_authorizer_buf,
2988                                 &s->s_authorizer_buf_len,
2989                                 &s->s_authorizer_reply_buf,
2990                                 &s->s_authorizer_reply_buf_len);
2991                         if (ret)
2992                                 return ret;
2993                 }
2994         }
2995
2996         *proto = ac->protocol;
2997         *buf = s->s_authorizer_buf;
2998         *len = s->s_authorizer_buf_len;
2999         *reply_buf = s->s_authorizer_reply_buf;
3000         *reply_len = s->s_authorizer_reply_buf_len;
3001         return 0;
3002 }
3003
3004
3005 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3006 {
3007         struct ceph_mds_session *s = con->private;
3008         struct ceph_mds_client *mdsc = s->s_mdsc;
3009         struct ceph_auth_client *ac = mdsc->client->monc.auth;
3010
3011         return ac->ops->verify_authorizer_reply(ac, s->s_authorizer, len);
3012 }
3013
3014 static int invalidate_authorizer(struct ceph_connection *con)
3015 {
3016         struct ceph_mds_session *s = con->private;
3017         struct ceph_mds_client *mdsc = s->s_mdsc;
3018         struct ceph_auth_client *ac = mdsc->client->monc.auth;
3019
3020         if (ac->ops->invalidate_authorizer)
3021                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3022
3023         return ceph_monc_validate_auth(&mdsc->client->monc);
3024 }
3025
3026 const static struct ceph_connection_operations mds_con_ops = {
3027         .get = con_get,
3028         .put = con_put,
3029         .dispatch = dispatch,
3030         .get_authorizer = get_authorizer,
3031         .verify_authorizer_reply = verify_authorizer_reply,
3032         .invalidate_authorizer = invalidate_authorizer,
3033         .peer_reset = peer_reset,
3034 };
3035
3036
3037
3038
3039 /* eof */