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  36  * are no longer valid.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