fs/ceph/mds_client.c

   1 #include <linux/ceph/ceph_debug.h>
   2
   3 #include <linux/fs.h>
   4 #include <linux/wait.h>
   5 #include <linux/slab.h>
   6 #include <linux/sched.h>
   7 #include <linux/debugfs.h>
   8 #include <linux/seq_file.h>
   9
  10 #include "super.h"
  11 #include "mds_client.h"
  12
  13 #include <linux/ceph/messenger.h>
  14 #include <linux/ceph/decode.h>
  15 #include <linux/ceph/pagelist.h>
  16 #include <linux/ceph/auth.h>
  17 #include <linux/ceph/debugfs.h>
  18
  19 /*
  20  * A cluster of MDS (metadata server) daemons is responsible for
  21  * managing the file system namespace (the directory hierarchy and
  22  * inodes) and for coordinating shared access to storage.  Metadata is
  23  * partitioning hierarchically across a number of servers, and that
  24  * partition varies over time as the cluster adjusts the distribution
  25  * in order to balance load.
  26  *
  27  * The MDS client is primarily responsible to managing synchronous
  28  * metadata requests for operations like open, unlink, and so forth.
  29  * If there is a MDS failure, we find out about it when we (possibly
  30  * request and) receive a new MDS map, and can resubmit affected
  31  * requests.
  32  *
  33  * For the most part, though, we take advantage of a lossless
  34  * communications channel to the MDS, and do not need to worry about
  35  * timing out or resubmitting requests.
  36  *
  37  * We maintain a stateful "session" with each MDS we interact with.
  38  * Within each session, we sent periodic heartbeat messages to ensure
  39  * any capabilities or leases we have been issues remain valid.  If
  40  * the session times out and goes stale, our leases and capabilities
  41  * are no longer valid.
  42  */
  43
  44 struct ceph_reconnect_state {
  45         struct ceph_pagelist *pagelist;
  46         bool flock;
  47 };
  48
  49 static void __wake_requests(struct ceph_mds_client *mdsc,
  50                             struct list_head *head);
  51
  52 static const struct ceph_connection_operations mds_con_ops;
  53
  54
  55 /*
  56  * mds reply parsing
  57  */
  58
  59 /*
  60  * parse individual inode info
  61  */
  62 static int parse_reply_info_in(void **p, void *end,
  63                                struct ceph_mds_reply_info_in *info)
  64 {
  65         int err = -EIO;
  66
  67         info->in = *p;
  68         *p += sizeof(struct ceph_mds_reply_inode) +
  69                 sizeof(*info->in->fragtree.splits) *
  70                 le32_to_cpu(info->in->fragtree.nsplits);
  71
  72         ceph_decode_32_safe(p, end, info->symlink_len, bad);
  73         ceph_decode_need(p, end, info->symlink_len, bad);
  74         info->symlink = *p;
  75         *p += info->symlink_len;
  76
  77         ceph_decode_32_safe(p, end, info->xattr_len, bad);
  78         ceph_decode_need(p, end, info->xattr_len, bad);
  79         info->xattr_data = *p;
  80         *p += info->xattr_len;
  81         return 0;
  82 bad:
  83         return err;
  84 }
  85
  86 /*
  87  * parse a normal reply, which may contain a (dir+)dentry and/or a
  88  * target inode.
  89  */
  90 static int parse_reply_info_trace(void **p, void *end,
  91                                   struct ceph_mds_reply_info_parsed *info)
  92 {
  93         int err;
  94
  95         if (info->head->is_dentry) {
  96                 err = parse_reply_info_in(p, end, &info->diri);
  97                 if (err < 0)
  98                         goto out_bad;
  99
 100                 if (unlikely(*p + sizeof(*info->dirfrag) > end))
 101                         goto bad;
 102                 info->dirfrag = *p;
 103                 *p += sizeof(*info->dirfrag) +
 104                         sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
 105                 if (unlikely(*p > end))
 106                         goto bad;
 107
 108                 ceph_decode_32_safe(p, end, info->dname_len, bad);
 109                 ceph_decode_need(p, end, info->dname_len, bad);
 110                 info->dname = *p;
 111                 *p += info->dname_len;
 112                 info->dlease = *p;
 113                 *p += sizeof(*info->dlease);
 114         }
 115
 116         if (info->head->is_target) {
 117                 err = parse_reply_info_in(p, end, &info->targeti);
 118                 if (err < 0)
 119                         goto out_bad;
 120         }
 121
 122         if (unlikely(*p != end))
 123                 goto bad;
 124         return 0;
 125
 126 bad:
 127         err = -EIO;
 128 out_bad:
 129         pr_err("problem parsing mds trace %d\n", err);
 130         return err;
 131 }
 132
 133 /*
 134  * parse readdir results
 135  */
 136 static int parse_reply_info_dir(void **p, void *end,
 137                                 struct ceph_mds_reply_info_parsed *info)
 138 {
 139         u32 num, i = 0;
 140         int err;
 141
 142         info->dir_dir = *p;
 143         if (*p + sizeof(*info->dir_dir) > end)
 144                 goto bad;
 145         *p += sizeof(*info->dir_dir) +
 146                 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
 147         if (*p > end)
 148                 goto bad;
 149
 150         ceph_decode_need(p, end, sizeof(num) + 2, bad);
 151         num = ceph_decode_32(p);
 152         info->dir_end = ceph_decode_8(p);
 153         info->dir_complete = ceph_decode_8(p);
 154         if (num == 0)
 155                 goto done;
 156
 157         /* alloc large array */
 158         info->dir_nr = num;
 159         info->dir_in = kcalloc(num, sizeof(*info->dir_in) +
 160                                sizeof(*info->dir_dname) +
 161                                sizeof(*info->dir_dname_len) +
 162                                sizeof(*info->dir_dlease),
 163                                GFP_NOFS);
 164         if (info->dir_in == NULL) {
 165                 err = -ENOMEM;
 166                 goto out_bad;
 167         }
 168         info->dir_dname = (void *)(info->dir_in + num);
 169         info->dir_dname_len = (void *)(info->dir_dname + num);
 170         info->dir_dlease = (void *)(info->dir_dname_len + num);
 171
 172         while (num) {
 173                 /* dentry */
 174                 ceph_decode_need(p, end, sizeof(u32)*2, bad);
 175                 info->dir_dname_len[i] = ceph_decode_32(p);
 176                 ceph_decode_need(p, end, info->dir_dname_len[i], bad);
 177                 info->dir_dname[i] = *p;
 178                 *p += info->dir_dname_len[i];
 179                 dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
 180                      info->dir_dname[i]);
 181                 info->dir_dlease[i] = *p;
 182                 *p += sizeof(struct ceph_mds_reply_lease);
 183
 184                 /* inode */
 185                 err = parse_reply_info_in(p, end, &info->dir_in[i]);
 186                 if (err < 0)
 187                         goto out_bad;
 188                 i++;
 189                 num--;
 190         }
 191
 192 done:
 193         if (*p != end)
 194                 goto bad;
 195         return 0;
 196
 197 bad:
 198         err = -EIO;
 199 out_bad:
 200         pr_err("problem parsing dir contents %d\n", err);
 201         return err;
 202 }
 203
 204 /*
 205  * parse fcntl F_GETLK results
 206  */
 207 static int parse_reply_info_filelock(void **p, void *end,
 208                 struct ceph_mds_reply_info_parsed *info)
 209 {
 210         if (*p + sizeof(*info->filelock_reply) > end)
 211                 goto bad;
 212
 213         info->filelock_reply = *p;
 214         *p += sizeof(*info->filelock_reply);
 215
 216         if (unlikely(*p != end))
 217                 goto bad;
 218         return 0;
 219
 220 bad:
 221         return -EIO;
 222 }
 223
 224 /*
 225  * parse extra results
 226  */
 227 static int parse_reply_info_extra(void **p, void *end,
 228                 struct ceph_mds_reply_info_parsed *info)
 229 {
 230         if (info->head->op == CEPH_MDS_OP_GETFILELOCK)
 231                 return parse_reply_info_filelock(p, end, info);
 232         else
 233                 return parse_reply_info_dir(p, end, info);
 234 }
 235
 236 /*
 237  * parse entire mds reply
 238  */
 239 static int parse_reply_info(struct ceph_msg *msg,
 240                             struct ceph_mds_reply_info_parsed *info)
 241 {
 242         void *p, *end;
 243         u32 len;
 244         int err;
 245
 246         info->head = msg->front.iov_base;
 247         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 248         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 249
 250         /* trace */
 251         ceph_decode_32_safe(&p, end, len, bad);
 252         if (len > 0) {
 253                 err = parse_reply_info_trace(&p, p+len, info);
 254                 if (err < 0)
 255                         goto out_bad;
 256         }
 257
 258         /* extra */
 259         ceph_decode_32_safe(&p, end, len, bad);
 260         if (len > 0) {
 261                 err = parse_reply_info_extra(&p, p+len, info);
 262                 if (err < 0)
 263                         goto out_bad;
 264         }
 265
 266         /* snap blob */
 267         ceph_decode_32_safe(&p, end, len, bad);
 268         info->snapblob_len = len;
 269         info->snapblob = p;
 270         p += len;
 271
 272         if (p != end)
 273                 goto bad;
 274         return 0;
 275
 276 bad:
 277         err = -EIO;
 278 out_bad:
 279         pr_err("mds parse_reply err %d\n", err);
 280         return err;
 281 }
 282
 283 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
 284 {
 285         kfree(info->dir_in);
 286 }
 287
 288
 289 /*
 290  * sessions
 291  */
 292 static const char *session_state_name(int s)
 293 {
 294         switch (s) {
 295         case CEPH_MDS_SESSION_NEW: return "new";
 296         case CEPH_MDS_SESSION_OPENING: return "opening";
 297         case CEPH_MDS_SESSION_OPEN: return "open";
 298         case CEPH_MDS_SESSION_HUNG: return "hung";
 299         case CEPH_MDS_SESSION_CLOSING: return "closing";
 300         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
 301         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
 302         default: return "???";
 303         }
 304 }
 305
 306 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
 307 {
 308         if (atomic_inc_not_zero(&s->s_ref)) {
 309                 dout("mdsc get_session %p %d -> %d\n", s,
 310                      atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
 311                 return s;
 312         } else {
 313                 dout("mdsc get_session %p 0 -- FAIL", s);
 314                 return NULL;
 315         }
 316 }
 317
 318 void ceph_put_mds_session(struct ceph_mds_session *s)
 319 {
 320         dout("mdsc put_session %p %d -> %d\n", s,
 321              atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
 322         if (atomic_dec_and_test(&s->s_ref)) {
 323                 if (s->s_authorizer)
 324                      s->s_mdsc->fsc->client->monc.auth->ops->destroy_authorizer(
 325                              s->s_mdsc->fsc->client->monc.auth,
 326                              s->s_authorizer);
 327                 kfree(s);
 328         }
 329 }
 330
 331 /*
 332  * called under mdsc->mutex
 333  */
 334 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
 335                                                    int mds)
 336 {
 337         struct ceph_mds_session *session;
 338
 339         if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
 340                 return NULL;
 341         session = mdsc->sessions[mds];
 342         dout("lookup_mds_session %p %d\n", session,
 343              atomic_read(&session->s_ref));
 344         get_session(session);
 345         return session;
 346 }
 347
 348 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
 349 {
 350         if (mds >= mdsc->max_sessions)
 351                 return false;
 352         return mdsc->sessions[mds];
 353 }
 354
 355 static int __verify_registered_session(struct ceph_mds_client *mdsc,
 356                                        struct ceph_mds_session *s)
 357 {
 358         if (s->s_mds >= mdsc->max_sessions ||
 359             mdsc->sessions[s->s_mds] != s)
 360                 return -ENOENT;
 361         return 0;
 362 }
 363
 364 /*
 365  * create+register a new session for given mds.
 366  * called under mdsc->mutex.
 367  */
 368 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 369                                                  int mds)
 370 {
 371         struct ceph_mds_session *s;
 372
 373         s = kzalloc(sizeof(*s), GFP_NOFS);
 374         if (!s)
 375                 return ERR_PTR(-ENOMEM);
 376         s->s_mdsc = mdsc;
 377         s->s_mds = mds;
 378         s->s_state = CEPH_MDS_SESSION_NEW;
 379         s->s_ttl = 0;
 380         s->s_seq = 0;
 381         mutex_init(&s->s_mutex);
 382
 383         ceph_con_init(mdsc->fsc->client->msgr, &s->s_con);
 384         s->s_con.private = s;
 385         s->s_con.ops = &mds_con_ops;
 386         s->s_con.peer_name.type = CEPH_ENTITY_TYPE_MDS;
 387         s->s_con.peer_name.num = cpu_to_le64(mds);
 388
 389         spin_lock_init(&s->s_cap_lock);
 390         s->s_cap_gen = 0;
 391         s->s_cap_ttl = 0;
 392         s->s_renew_requested = 0;
 393         s->s_renew_seq = 0;
 394         INIT_LIST_HEAD(&s->s_caps);
 395         s->s_nr_caps = 0;
 396         s->s_trim_caps = 0;
 397         atomic_set(&s->s_ref, 1);
 398         INIT_LIST_HEAD(&s->s_waiting);
 399         INIT_LIST_HEAD(&s->s_unsafe);
 400         s->s_num_cap_releases = 0;
 401         s->s_cap_iterator = NULL;
 402         INIT_LIST_HEAD(&s->s_cap_releases);
 403         INIT_LIST_HEAD(&s->s_cap_releases_done);
 404         INIT_LIST_HEAD(&s->s_cap_flushing);
 405         INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
 406
 407         dout("register_session mds%d\n", mds);
 408         if (mds >= mdsc->max_sessions) {
 409                 int newmax = 1 << get_count_order(mds+1);
 410                 struct ceph_mds_session **sa;
 411
 412                 dout("register_session realloc to %d\n", newmax);
 413                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 414                 if (sa == NULL)
 415                         goto fail_realloc;
 416                 if (mdsc->sessions) {
 417                         memcpy(sa, mdsc->sessions,
 418                                mdsc->max_sessions * sizeof(void *));
 419                         kfree(mdsc->sessions);
 420                 }
 421                 mdsc->sessions = sa;
 422                 mdsc->max_sessions = newmax;
 423         }
 424         mdsc->sessions[mds] = s;
 425         atomic_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
 426
 427         ceph_con_open(&s->s_con, ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
 428
 429         return s;
 430
 431 fail_realloc:
 432         kfree(s);
 433         return ERR_PTR(-ENOMEM);
 434 }
 435
 436 /*
 437  * called under mdsc->mutex
 438  */
 439 static void __unregister_session(struct ceph_mds_client *mdsc,
 440                                struct ceph_mds_session *s)
 441 {
 442         dout("__unregister_session mds%d %p\n", s->s_mds, s);
 443         BUG_ON(mdsc->sessions[s->s_mds] != s);
 444         mdsc->sessions[s->s_mds] = NULL;
 445         ceph_con_close(&s->s_con);
 446         ceph_put_mds_session(s);
 447 }
 448
 449 /*
 450  * drop session refs in request.
 451  *
 452  * should be last request ref, or hold mdsc->mutex
 453  */
 454 static void put_request_session(struct ceph_mds_request *req)
 455 {
 456         if (req->r_session) {
 457                 ceph_put_mds_session(req->r_session);
 458                 req->r_session = NULL;
 459         }
 460 }
 461
 462 void ceph_mdsc_release_request(struct kref *kref)
 463 {
 464         struct ceph_mds_request *req = container_of(kref,
 465                                                     struct ceph_mds_request,
 466                                                     r_kref);
 467         if (req->r_request)
 468                 ceph_msg_put(req->r_request);
 469         if (req->r_reply) {
 470                 ceph_msg_put(req->r_reply);
 471                 destroy_reply_info(&req->r_reply_info);
 472         }
 473         if (req->r_inode) {
 474                 ceph_put_cap_refs(ceph_inode(req->r_inode),
 475                                   CEPH_CAP_PIN);
 476                 iput(req->r_inode);
 477         }
 478         if (req->r_locked_dir)
 479                 ceph_put_cap_refs(ceph_inode(req->r_locked_dir),
 480                                   CEPH_CAP_PIN);
 481         if (req->r_target_inode)
 482                 iput(req->r_target_inode);
 483         if (req->r_dentry)
 484                 dput(req->r_dentry);
 485         if (req->r_old_dentry) {
 486                 ceph_put_cap_refs(
 487                         ceph_inode(req->r_old_dentry->d_parent->d_inode),
 488                         CEPH_CAP_PIN);
 489                 dput(req->r_old_dentry);
 490         }
 491         kfree(req->r_path1);
 492         kfree(req->r_path2);
 493         put_request_session(req);
 494         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
 495         kfree(req);
 496 }
 497
 498 /*
 499  * lookup session, bump ref if found.
 500  *
 501  * called under mdsc->mutex.
 502  */
 503 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
 504                                              u64 tid)
 505 {
 506         struct ceph_mds_request *req;
 507         struct rb_node *n = mdsc->request_tree.rb_node;
 508
 509         while (n) {
 510                 req = rb_entry(n, struct ceph_mds_request, r_node);
 511                 if (tid < req->r_tid)
 512                         n = n->rb_left;
 513                 else if (tid > req->r_tid)
 514                         n = n->rb_right;
 515                 else {
 516                         ceph_mdsc_get_request(req);
 517                         return req;
 518                 }
 519         }
 520         return NULL;
 521 }
 522
 523 static void __insert_request(struct ceph_mds_client *mdsc,
 524                              struct ceph_mds_request *new)
 525 {
 526         struct rb_node **p = &mdsc->request_tree.rb_node;
 527         struct rb_node *parent = NULL;
 528         struct ceph_mds_request *req = NULL;
 529
 530         while (*p) {
 531                 parent = *p;
 532                 req = rb_entry(parent, struct ceph_mds_request, r_node);
 533                 if (new->r_tid < req->r_tid)
 534                         p = &(*p)->rb_left;
 535                 else if (new->r_tid > req->r_tid)
 536                         p = &(*p)->rb_right;
 537                 else
 538                         BUG();
 539         }
 540
 541         rb_link_node(&new->r_node, parent, p);
 542         rb_insert_color(&new->r_node, &mdsc->request_tree);
 543 }
 544
 545 /*
 546  * Register an in-flight request, and assign a tid.  Link to directory
 547  * are modifying (if any).
 548  *
 549  * Called under mdsc->mutex.
 550  */
 551 static void __register_request(struct ceph_mds_client *mdsc,
 552                                struct ceph_mds_request *req,
 553                                struct inode *dir)
 554 {
 555         req->r_tid = ++mdsc->last_tid;
 556         if (req->r_num_caps)
 557                 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
 558                                   req->r_num_caps);
 559         dout("__register_request %p tid %lld\n", req, req->r_tid);
 560         ceph_mdsc_get_request(req);
 561         __insert_request(mdsc, req);
 562
 563         req->r_uid = current_fsuid();
 564         req->r_gid = current_fsgid();
 565
 566         if (dir) {
 567                 struct ceph_inode_info *ci = ceph_inode(dir);
 568
 569                 spin_lock(&ci->i_unsafe_lock);
 570                 req->r_unsafe_dir = dir;
 571                 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
 572                 spin_unlock(&ci->i_unsafe_lock);
 573         }
 574 }
 575
 576 static void __unregister_request(struct ceph_mds_client *mdsc,
 577                                  struct ceph_mds_request *req)
 578 {
 579         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 580         rb_erase(&req->r_node, &mdsc->request_tree);
 581         RB_CLEAR_NODE(&req->r_node);
 582
 583         if (req->r_unsafe_dir) {
 584                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 585
 586                 spin_lock(&ci->i_unsafe_lock);
 587                 list_del_init(&req->r_unsafe_dir_item);
 588                 spin_unlock(&ci->i_unsafe_lock);
 589         }
 590
 591         ceph_mdsc_put_request(req);
 592 }
 593
 594 /*
 595  * Choose mds to send request to next.  If there is a hint set in the
 596  * request (e.g., due to a prior forward hint from the mds), use that.
 597  * Otherwise, consult frag tree and/or caps to identify the
 598  * appropriate mds.  If all else fails, choose randomly.
 599  *
 600  * Called under mdsc->mutex.
 601  */
 602 struct dentry *get_nonsnap_parent(struct dentry *dentry)
 603 {
 604         while (!IS_ROOT(dentry) && ceph_snap(dentry->d_inode) != CEPH_NOSNAP)
 605                 dentry = dentry->d_parent;
 606         return dentry;
 607 }
 608
 609 static int __choose_mds(struct ceph_mds_client *mdsc,
 610                         struct ceph_mds_request *req)
 611 {
 612         struct inode *inode;
 613         struct ceph_inode_info *ci;
 614         struct ceph_cap *cap;
 615         int mode = req->r_direct_mode;
 616         int mds = -1;
 617         u32 hash = req->r_direct_hash;
 618         bool is_hash = req->r_direct_is_hash;
 619
 620         /*
 621          * is there a specific mds we should try?  ignore hint if we have
 622          * no session and the mds is not up (active or recovering).
 623          */
 624         if (req->r_resend_mds >= 0 &&
 625             (__have_session(mdsc, req->r_resend_mds) ||
 626              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 627                 dout("choose_mds using resend_mds mds%d\n",
 628                      req->r_resend_mds);
 629                 return req->r_resend_mds;
 630         }
 631
 632         if (mode == USE_RANDOM_MDS)
 633                 goto random;
 634
 635         inode = NULL;
 636         if (req->r_inode) {
 637                 inode = req->r_inode;
 638         } else if (req->r_dentry) {
 639                 struct inode *dir = req->r_dentry->d_parent->d_inode;
 640
 641                 if (dir->i_sb != mdsc->fsc->sb) {
 642                         /* not this fs! */
 643                         inode = req->r_dentry->d_inode;
 644                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
 645                         /* direct snapped/virtual snapdir requests
 646                          * based on parent dir inode */
 647                         struct dentry *dn =
 648                                 get_nonsnap_parent(req->r_dentry->d_parent);
 649                         inode = dn->d_inode;
 650                         dout("__choose_mds using nonsnap parent %p\n", inode);
 651                 } else if (req->r_dentry->d_inode) {
 652                         /* dentry target */
 653                         inode = req->r_dentry->d_inode;
 654                 } else {
 655                         /* dir + name */
 656                         inode = dir;
 657                         hash = req->r_dentry->d_name.hash;
 658                         is_hash = true;
 659                 }
 660         }
 661
 662         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 663              (int)hash, mode);
 664         if (!inode)
 665                 goto random;
 666         ci = ceph_inode(inode);
 667
 668         if (is_hash && S_ISDIR(inode->i_mode)) {
 669                 struct ceph_inode_frag frag;
 670                 int found;
 671
 672                 ceph_choose_frag(ci, hash, &frag, &found);
 673                 if (found) {
 674                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 675                                 u8 r;
 676
 677                                 /* choose a random replica */
 678                                 get_random_bytes(&r, 1);
 679                                 r %= frag.ndist;
 680                                 mds = frag.dist[r];
 681                                 dout("choose_mds %p %llx.%llx "
 682                                      "frag %u mds%d (%d/%d)\n",
 683                                      inode, ceph_vinop(inode),
 684                                      frag.frag, frag.mds,
 685                                      (int)r, frag.ndist);
 686                                 return mds;
 687                         }
 688
 689                         /* since this file/dir wasn't known to be
 690                          * replicated, then we want to look for the
 691                          * authoritative mds. */
 692                         mode = USE_AUTH_MDS;
 693                         if (frag.mds >= 0) {
 694                                 /* choose auth mds */
 695                                 mds = frag.mds;
 696                                 dout("choose_mds %p %llx.%llx "
 697                                      "frag %u mds%d (auth)\n",
 698                                      inode, ceph_vinop(inode), frag.frag, mds);
 699                                 return mds;
 700                         }
 701                 }
 702         }
 703
 704         spin_lock(&inode->i_lock);
 705         cap = NULL;
 706         if (mode == USE_AUTH_MDS)
 707                 cap = ci->i_auth_cap;
 708         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
 709                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
 710         if (!cap) {
 711                 spin_unlock(&inode->i_lock);
 712                 goto random;
 713         }
 714         mds = cap->session->s_mds;
 715         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
 716              inode, ceph_vinop(inode), mds,
 717              cap == ci->i_auth_cap ? "auth " : "", cap);
 718         spin_unlock(&inode->i_lock);
 719         return mds;
 720
 721 random:
 722         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
 723         dout("choose_mds chose random mds%d\n", mds);
 724         return mds;
 725 }
 726
 727
 728 /*
 729  * session messages
 730  */
 731 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
 732 {
 733         struct ceph_msg *msg;
 734         struct ceph_mds_session_head *h;
 735
 736         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS);
 737         if (!msg) {
 738                 pr_err("create_session_msg ENOMEM creating msg\n");
 739                 return NULL;
 740         }
 741         h = msg->front.iov_base;
 742         h->op = cpu_to_le32(op);
 743         h->seq = cpu_to_le64(seq);
 744         return msg;
 745 }
 746
 747 /*
 748  * send session open request.
 749  *
 750  * called under mdsc->mutex
 751  */
 752 static int __open_session(struct ceph_mds_client *mdsc,
 753                           struct ceph_mds_session *session)
 754 {
 755         struct ceph_msg *msg;
 756         int mstate;
 757         int mds = session->s_mds;
 758
 759         /* wait for mds to go active? */
 760         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
 761         dout("open_session to mds%d (%s)\n", mds,
 762              ceph_mds_state_name(mstate));
 763         session->s_state = CEPH_MDS_SESSION_OPENING;
 764         session->s_renew_requested = jiffies;
 765
 766         /* send connect message */
 767         msg = create_session_msg(CEPH_SESSION_REQUEST_OPEN, session->s_seq);
 768         if (!msg)
 769                 return -ENOMEM;
 770         ceph_con_send(&session->s_con, msg);
 771         return 0;
 772 }
 773
 774 /*
 775  * open sessions for any export targets for the given mds
 776  *
 777  * called under mdsc->mutex
 778  */
 779 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
 780                                           struct ceph_mds_session *session)
 781 {
 782         struct ceph_mds_info *mi;
 783         struct ceph_mds_session *ts;
 784         int i, mds = session->s_mds;
 785         int target;
 786
 787         if (mds >= mdsc->mdsmap->m_max_mds)
 788                 return;
 789         mi = &mdsc->mdsmap->m_info[mds];
 790         dout("open_export_target_sessions for mds%d (%d targets)\n",
 791              session->s_mds, mi->num_export_targets);
 792
 793         for (i = 0; i < mi->num_export_targets; i++) {
 794                 target = mi->export_targets[i];
 795                 ts = __ceph_lookup_mds_session(mdsc, target);
 796                 if (!ts) {
 797                         ts = register_session(mdsc, target);
 798                         if (IS_ERR(ts))
 799                                 return;
 800                 }
 801                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
 802                     session->s_state == CEPH_MDS_SESSION_CLOSING)
 803                         __open_session(mdsc, session);
 804                 else
 805                         dout(" mds%d target mds%d %p is %s\n", session->s_mds,
 806                              i, ts, session_state_name(ts->s_state));
 807                 ceph_put_mds_session(ts);
 808         }
 809 }
 810
 811 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
 812                                            struct ceph_mds_session *session)
 813 {
 814         mutex_lock(&mdsc->mutex);
 815         __open_export_target_sessions(mdsc, session);
 816         mutex_unlock(&mdsc->mutex);
 817 }
 818
 819 /*
 820  * session caps
 821  */
 822
 823 /*
 824  * Free preallocated cap messages assigned to this session
 825  */
 826 static void cleanup_cap_releases(struct ceph_mds_session *session)
 827 {
 828         struct ceph_msg *msg;
 829
 830         spin_lock(&session->s_cap_lock);
 831         while (!list_empty(&session->s_cap_releases)) {
 832                 msg = list_first_entry(&session->s_cap_releases,
 833                                        struct ceph_msg, list_head);
 834                 list_del_init(&msg->list_head);
 835                 ceph_msg_put(msg);
 836         }
 837         while (!list_empty(&session->s_cap_releases_done)) {
 838                 msg = list_first_entry(&session->s_cap_releases_done,
 839                                        struct ceph_msg, list_head);
 840                 list_del_init(&msg->list_head);
 841                 ceph_msg_put(msg);
 842         }
 843         spin_unlock(&session->s_cap_lock);
 844 }
 845
 846 /*
 847  * Helper to safely iterate over all caps associated with a session, with
 848  * special care taken to handle a racing __ceph_remove_cap().
 849  *
 850  * Caller must hold session s_mutex.
 851  */
 852 static int iterate_session_caps(struct ceph_mds_session *session,
 853                                  int (*cb)(struct inode *, struct ceph_cap *,
 854                                             void *), void *arg)
 855 {
 856         struct list_head *p;
 857         struct ceph_cap *cap;
 858         struct inode *inode, *last_inode = NULL;
 859         struct ceph_cap *old_cap = NULL;
 860         int ret;
 861
 862         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
 863         spin_lock(&session->s_cap_lock);
 864         p = session->s_caps.next;
 865         while (p != &session->s_caps) {
 866                 cap = list_entry(p, struct ceph_cap, session_caps);
 867                 inode = igrab(&cap->ci->vfs_inode);
 868                 if (!inode) {
 869                         p = p->next;
 870                         continue;
 871                 }
 872                 session->s_cap_iterator = cap;
 873                 spin_unlock(&session->s_cap_lock);
 874
 875                 if (last_inode) {
 876                         iput(last_inode);
 877                         last_inode = NULL;
 878                 }
 879                 if (old_cap) {
 880                         ceph_put_cap(session->s_mdsc, old_cap);
 881                         old_cap = NULL;
 882                 }
 883
 884                 ret = cb(inode, cap, arg);
 885                 last_inode = inode;
 886
 887                 spin_lock(&session->s_cap_lock);
 888                 p = p->next;
 889                 if (cap->ci == NULL) {
 890                         dout("iterate_session_caps  finishing cap %p removal\n",
 891                              cap);
 892                         BUG_ON(cap->session != session);
 893                         list_del_init(&cap->session_caps);
 894                         session->s_nr_caps--;
 895                         cap->session = NULL;
 896                         old_cap = cap;  /* put_cap it w/o locks held */
 897                 }
 898                 if (ret < 0)
 899                         goto out;
 900         }
 901         ret = 0;
 902 out:
 903         session->s_cap_iterator = NULL;
 904         spin_unlock(&session->s_cap_lock);
 905
 906         if (last_inode)
 907                 iput(last_inode);
 908         if (old_cap)
 909                 ceph_put_cap(session->s_mdsc, old_cap);
 910
 911         return ret;
 912 }
 913
 914 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
 915                                   void *arg)
 916 {
 917         struct ceph_inode_info *ci = ceph_inode(inode);
 918         int drop = 0;
 919
 920         dout("removing cap %p, ci is %p, inode is %p\n",
 921              cap, ci, &ci->vfs_inode);
 922         spin_lock(&inode->i_lock);
 923         __ceph_remove_cap(cap);
 924         if (!__ceph_is_any_real_caps(ci)) {
 925                 struct ceph_mds_client *mdsc =
 926                         ceph_sb_to_client(inode->i_sb)->mdsc;
 927
 928                 spin_lock(&mdsc->cap_dirty_lock);
 929                 if (!list_empty(&ci->i_dirty_item)) {
 930                         pr_info(" dropping dirty %s state for %p %lld\n",
 931                                 ceph_cap_string(ci->i_dirty_caps),
 932                                 inode, ceph_ino(inode));
 933                         ci->i_dirty_caps = 0;
 934                         list_del_init(&ci->i_dirty_item);
 935                         drop = 1;
 936                 }
 937                 if (!list_empty(&ci->i_flushing_item)) {
 938                         pr_info(" dropping dirty+flushing %s state for %p %lld\n",
 939                                 ceph_cap_string(ci->i_flushing_caps),
 940                                 inode, ceph_ino(inode));
 941                         ci->i_flushing_caps = 0;
 942                         list_del_init(&ci->i_flushing_item);
 943                         mdsc->num_cap_flushing--;
 944                         drop = 1;
 945                 }
 946                 if (drop && ci->i_wrbuffer_ref) {
 947                         pr_info(" dropping dirty data for %p %lld\n",
 948                                 inode, ceph_ino(inode));
 949                         ci->i_wrbuffer_ref = 0;
 950                         ci->i_wrbuffer_ref_head = 0;
 951                         drop++;
 952                 }
 953                 spin_unlock(&mdsc->cap_dirty_lock);
 954         }
 955         spin_unlock(&inode->i_lock);
 956         while (drop--)
 957                 iput(inode);
 958         return 0;
 959 }
 960
 961 /*
 962  * caller must hold session s_mutex
 963  */
 964 static void remove_session_caps(struct ceph_mds_session *session)
 965 {
 966         dout("remove_session_caps on %p\n", session);
 967         iterate_session_caps(session, remove_session_caps_cb, NULL);
 968         BUG_ON(session->s_nr_caps > 0);
 969         BUG_ON(!list_empty(&session->s_cap_flushing));
 970         cleanup_cap_releases(session);
 971 }
 972
 973 /*
 974  * wake up any threads waiting on this session's caps.  if the cap is
 975  * old (didn't get renewed on the client reconnect), remove it now.
 976  *
 977  * caller must hold s_mutex.
 978  */
 979 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
 980                               void *arg)
 981 {
 982         struct ceph_inode_info *ci = ceph_inode(inode);
 983
 984         wake_up_all(&ci->i_cap_wq);
 985         if (arg) {
 986                 spin_lock(&inode->i_lock);
 987                 ci->i_wanted_max_size = 0;
 988                 ci->i_requested_max_size = 0;
 989                 spin_unlock(&inode->i_lock);
 990         }
 991         return 0;
 992 }
 993
 994 static void wake_up_session_caps(struct ceph_mds_session *session,
 995                                  int reconnect)
 996 {
 997         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
 998         iterate_session_caps(session, wake_up_session_cb,
 999                              (void *)(unsigned long)reconnect);
1000 }
1001
1002 /*
1003  * Send periodic message to MDS renewing all currently held caps.  The
1004  * ack will reset the expiration for all caps from this session.
1005  *
1006  * caller holds s_mutex
1007  */
1008 static int send_renew_caps(struct ceph_mds_client *mdsc,
1009                            struct ceph_mds_session *session)
1010 {
1011         struct ceph_msg *msg;
1012         int state;
1013
1014         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1015             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1016                 pr_info("mds%d caps stale\n", session->s_mds);
1017         session->s_renew_requested = jiffies;
1018
1019         /* do not try to renew caps until a recovering mds has reconnected
1020          * with its clients. */
1021         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1022         if (state < CEPH_MDS_STATE_RECONNECT) {
1023                 dout("send_renew_caps ignoring mds%d (%s)\n",
1024                      session->s_mds, ceph_mds_state_name(state));
1025                 return 0;
1026         }
1027
1028         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1029                 ceph_mds_state_name(state));
1030         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1031                                  ++session->s_renew_seq);
1032         if (!msg)
1033                 return -ENOMEM;
1034         ceph_con_send(&session->s_con, msg);
1035         return 0;
1036 }
1037
1038 /*
1039  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1040  *
1041  * Called under session->s_mutex
1042  */
1043 static void renewed_caps(struct ceph_mds_client *mdsc,
1044                          struct ceph_mds_session *session, int is_renew)
1045 {
1046         int was_stale;
1047         int wake = 0;
1048
1049         spin_lock(&session->s_cap_lock);
1050         was_stale = is_renew && (session->s_cap_ttl == 0 ||
1051                                  time_after_eq(jiffies, session->s_cap_ttl));
1052
1053         session->s_cap_ttl = session->s_renew_requested +
1054                 mdsc->mdsmap->m_session_timeout*HZ;
1055
1056         if (was_stale) {
1057                 if (time_before(jiffies, session->s_cap_ttl)) {
1058                         pr_info("mds%d caps renewed\n", session->s_mds);
1059                         wake = 1;
1060                 } else {
1061                         pr_info("mds%d caps still stale\n", session->s_mds);
1062                 }
1063         }
1064         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1065              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1066              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1067         spin_unlock(&session->s_cap_lock);
1068
1069         if (wake)
1070                 wake_up_session_caps(session, 0);
1071 }
1072
1073 /*
1074  * send a session close request
1075  */
1076 static int request_close_session(struct ceph_mds_client *mdsc,
1077                                  struct ceph_mds_session *session)
1078 {
1079         struct ceph_msg *msg;
1080
1081         dout("request_close_session mds%d state %s seq %lld\n",
1082              session->s_mds, session_state_name(session->s_state),
1083              session->s_seq);
1084         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1085         if (!msg)
1086                 return -ENOMEM;
1087         ceph_con_send(&session->s_con, msg);
1088         return 0;
1089 }
1090
1091 /*
1092  * Called with s_mutex held.
1093  */
1094 static int __close_session(struct ceph_mds_client *mdsc,
1095                          struct ceph_mds_session *session)
1096 {
1097         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1098                 return 0;
1099         session->s_state = CEPH_MDS_SESSION_CLOSING;
1100         return request_close_session(mdsc, session);
1101 }
1102
1103 /*
1104  * Trim old(er) caps.
1105  *
1106  * Because we can't cache an inode without one or more caps, we do
1107  * this indirectly: if a cap is unused, we prune its aliases, at which
1108  * point the inode will hopefully get dropped to.
1109  *
1110  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1111  * memory pressure from the MDS, though, so it needn't be perfect.
1112  */
1113 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1114 {
1115         struct ceph_mds_session *session = arg;
1116         struct ceph_inode_info *ci = ceph_inode(inode);
1117         int used, oissued, mine;
1118
1119         if (session->s_trim_caps <= 0)
1120                 return -1;
1121
1122         spin_lock(&inode->i_lock);
1123         mine = cap->issued | cap->implemented;
1124         used = __ceph_caps_used(ci);
1125         oissued = __ceph_caps_issued_other(ci, cap);
1126
1127         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
1128              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1129              ceph_cap_string(used));
1130         if (ci->i_dirty_caps)
1131                 goto out;   /* dirty caps */
1132         if ((used & ~oissued) & mine)
1133                 goto out;   /* we need these caps */
1134
1135         session->s_trim_caps--;
1136         if (oissued) {
1137                 /* we aren't the only cap.. just remove us */
1138                 __ceph_remove_cap(cap);
1139         } else {
1140                 /* try to drop referring dentries */
1141                 spin_unlock(&inode->i_lock);
1142                 d_prune_aliases(inode);
1143                 dout("trim_caps_cb %p cap %p  pruned, count now %d\n",
1144                      inode, cap, atomic_read(&inode->i_count));
1145                 return 0;
1146         }
1147
1148 out:
1149         spin_unlock(&inode->i_lock);
1150         return 0;
1151 }
1152
1153 /*
1154  * Trim session cap count down to some max number.
1155  */
1156 static int trim_caps(struct ceph_mds_client *mdsc,
1157                      struct ceph_mds_session *session,
1158                      int max_caps)
1159 {
1160         int trim_caps = session->s_nr_caps - max_caps;
1161
1162         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1163              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1164         if (trim_caps > 0) {
1165                 session->s_trim_caps = trim_caps;
1166                 iterate_session_caps(session, trim_caps_cb, session);
1167                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1168                      session->s_mds, session->s_nr_caps, max_caps,
1169                         trim_caps - session->s_trim_caps);
1170                 session->s_trim_caps = 0;
1171         }
1172         return 0;
1173 }
1174
1175 /*
1176  * Allocate cap_release messages.  If there is a partially full message
1177  * in the queue, try to allocate enough to cover it's remainder, so that
1178  * we can send it immediately.
1179  *
1180  * Called under s_mutex.
1181  */
1182 int ceph_add_cap_releases(struct ceph_mds_client *mdsc,
1183                           struct ceph_mds_session *session)
1184 {
1185         struct ceph_msg *msg, *partial = NULL;
1186         struct ceph_mds_cap_release *head;
1187         int err = -ENOMEM;
1188         int extra = mdsc->fsc->mount_options->cap_release_safety;
1189         int num;
1190
1191         dout("add_cap_releases %p mds%d extra %d\n", session, session->s_mds,
1192              extra);
1193
1194         spin_lock(&session->s_cap_lock);
1195
1196         if (!list_empty(&session->s_cap_releases)) {
1197                 msg = list_first_entry(&session->s_cap_releases,
1198                                        struct ceph_msg,
1199                                  list_head);
1200                 head = msg->front.iov_base;
1201                 num = le32_to_cpu(head->num);
1202                 if (num) {
1203                         dout(" partial %p with (%d/%d)\n", msg, num,
1204                              (int)CEPH_CAPS_PER_RELEASE);
1205                         extra += CEPH_CAPS_PER_RELEASE - num;
1206                         partial = msg;
1207                 }
1208         }
1209         while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1210                 spin_unlock(&session->s_cap_lock);
1211                 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
1212                                    GFP_NOFS);
1213                 if (!msg)
1214                         goto out_unlocked;
1215                 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1216                      (int)msg->front.iov_len);
1217                 head = msg->front.iov_base;
1218                 head->num = cpu_to_le32(0);
1219                 msg->front.iov_len = sizeof(*head);
1220                 spin_lock(&session->s_cap_lock);
1221                 list_add(&msg->list_head, &session->s_cap_releases);
1222                 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1223         }
1224
1225         if (partial) {
1226                 head = partial->front.iov_base;
1227                 num = le32_to_cpu(head->num);
1228                 dout(" queueing partial %p with %d/%d\n", partial, num,
1229                      (int)CEPH_CAPS_PER_RELEASE);
1230                 list_move_tail(&partial->list_head,
1231                                &session->s_cap_releases_done);
1232                 session->s_num_cap_releases -= CEPH_CAPS_PER_RELEASE - num;
1233         }
1234         err = 0;
1235         spin_unlock(&session->s_cap_lock);
1236 out_unlocked:
1237         return err;
1238 }
1239
1240 /*
1241  * flush all dirty inode data to disk.
1242  *
1243  * returns true if we've flushed through want_flush_seq
1244  */
1245 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1246 {
1247         int mds, ret = 1;
1248
1249         dout("check_cap_flush want %lld\n", want_flush_seq);
1250         mutex_lock(&mdsc->mutex);
1251         for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1252                 struct ceph_mds_session *session = mdsc->sessions[mds];
1253
1254                 if (!session)
1255                         continue;
1256                 get_session(session);
1257                 mutex_unlock(&mdsc->mutex);
1258
1259                 mutex_lock(&session->s_mutex);
1260                 if (!list_empty(&session->s_cap_flushing)) {
1261                         struct ceph_inode_info *ci =
1262                                 list_entry(session->s_cap_flushing.next,
1263                                            struct ceph_inode_info,
1264                                            i_flushing_item);
1265                         struct inode *inode = &ci->vfs_inode;
1266
1267                         spin_lock(&inode->i_lock);
1268                         if (ci->i_cap_flush_seq <= want_flush_seq) {
1269                                 dout("check_cap_flush still flushing %p "
1270                                      "seq %lld <= %lld to mds%d\n", inode,
1271                                      ci->i_cap_flush_seq, want_flush_seq,
1272                                      session->s_mds);
1273                                 ret = 0;
1274                         }
1275                         spin_unlock(&inode->i_lock);
1276                 }
1277                 mutex_unlock(&session->s_mutex);
1278                 ceph_put_mds_session(session);
1279
1280                 if (!ret)
1281                         return ret;
1282                 mutex_lock(&mdsc->mutex);
1283         }
1284
1285         mutex_unlock(&mdsc->mutex);
1286         dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1287         return ret;
1288 }
1289
1290 /*
1291  * called under s_mutex
1292  */
1293 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1294                             struct ceph_mds_session *session)
1295 {
1296         struct ceph_msg *msg;
1297
1298         dout("send_cap_releases mds%d\n", session->s_mds);
1299         spin_lock(&session->s_cap_lock);
1300         while (!list_empty(&session->s_cap_releases_done)) {
1301                 msg = list_first_entry(&session->s_cap_releases_done,
1302                                  struct ceph_msg, list_head);
1303                 list_del_init(&msg->list_head);
1304                 spin_unlock(&session->s_cap_lock);
1305                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1306                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1307                 ceph_con_send(&session->s_con, msg);
1308                 spin_lock(&session->s_cap_lock);
1309         }
1310         spin_unlock(&session->s_cap_lock);
1311 }
1312
1313 static void discard_cap_releases(struct ceph_mds_client *mdsc,
1314                                  struct ceph_mds_session *session)
1315 {
1316         struct ceph_msg *msg;
1317         struct ceph_mds_cap_release *head;
1318         unsigned num;
1319
1320         dout("discard_cap_releases mds%d\n", session->s_mds);
1321         spin_lock(&session->s_cap_lock);
1322
1323         /* zero out the in-progress message */
1324         msg = list_first_entry(&session->s_cap_releases,
1325                                struct ceph_msg, list_head);
1326         head = msg->front.iov_base;
1327         num = le32_to_cpu(head->num);
1328         dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg, num);
1329         head->num = cpu_to_le32(0);
1330         session->s_num_cap_releases += num;
1331
1332         /* requeue completed messages */
1333         while (!list_empty(&session->s_cap_releases_done)) {
1334                 msg = list_first_entry(&session->s_cap_releases_done,
1335                                  struct ceph_msg, list_head);
1336                 list_del_init(&msg->list_head);
1337
1338                 head = msg->front.iov_base;
1339                 num = le32_to_cpu(head->num);
1340                 dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg,
1341                      num);
1342                 session->s_num_cap_releases += num;
1343                 head->num = cpu_to_le32(0);
1344                 msg->front.iov_len = sizeof(*head);
1345                 list_add(&msg->list_head, &session->s_cap_releases);
1346         }
1347
1348         spin_unlock(&session->s_cap_lock);
1349 }
1350
1351 /*
1352  * requests
1353  */
1354
1355 /*
1356  * Create an mds request.
1357  */
1358 struct ceph_mds_request *
1359 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1360 {
1361         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1362
1363         if (!req)
1364                 return ERR_PTR(-ENOMEM);
1365
1366         mutex_init(&req->r_fill_mutex);
1367         req->r_mdsc = mdsc;
1368         req->r_started = jiffies;
1369         req->r_resend_mds = -1;
1370         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1371         req->r_fmode = -1;
1372         kref_init(&req->r_kref);
1373         INIT_LIST_HEAD(&req->r_wait);
1374         init_completion(&req->r_completion);
1375         init_completion(&req->r_safe_completion);
1376         INIT_LIST_HEAD(&req->r_unsafe_item);
1377
1378         req->r_op = op;
1379         req->r_direct_mode = mode;
1380         return req;
1381 }
1382
1383 /*
1384  * return oldest (lowest) request, tid in request tree, 0 if none.
1385  *
1386  * called under mdsc->mutex.
1387  */
1388 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1389 {
1390         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1391                 return NULL;
1392         return rb_entry(rb_first(&mdsc->request_tree),
1393                         struct ceph_mds_request, r_node);
1394 }
1395
1396 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1397 {
1398         struct ceph_mds_request *req = __get_oldest_req(mdsc);
1399
1400         if (req)
1401                 return req->r_tid;
1402         return 0;
1403 }
1404
1405 /*
1406  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1407  * on build_path_from_dentry in fs/cifs/dir.c.
1408  *
1409  * If @stop_on_nosnap, generate path relative to the first non-snapped
1410  * inode.
1411  *
1412  * Encode hidden .snap dirs as a double /, i.e.
1413  *   foo/.snap/bar -> foo//bar
1414  */
1415 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1416                            int stop_on_nosnap)
1417 {
1418         struct dentry *temp;
1419         char *path;
1420         int len, pos;
1421
1422         if (dentry == NULL)
1423                 return ERR_PTR(-EINVAL);
1424
1425 retry:
1426         len = 0;
1427         for (temp = dentry; !IS_ROOT(temp);) {
1428                 struct inode *inode = temp->d_inode;
1429                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1430                         len++;  /* slash only */
1431                 else if (stop_on_nosnap && inode &&
1432                          ceph_snap(inode) == CEPH_NOSNAP)
1433                         break;
1434                 else
1435                         len += 1 + temp->d_name.len;
1436                 temp = temp->d_parent;
1437                 if (temp == NULL) {
1438                         pr_err("build_path corrupt dentry %p\n", dentry);
1439                         return ERR_PTR(-EINVAL);
1440                 }
1441         }
1442         if (len)
1443                 len--;  /* no leading '/' */
1444
1445         path = kmalloc(len+1, GFP_NOFS);
1446         if (path == NULL)
1447                 return ERR_PTR(-ENOMEM);
1448         pos = len;
1449         path[pos] = 0;  /* trailing null */
1450         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1451                 struct inode *inode = temp->d_inode;
1452
1453                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1454                         dout("build_path path+%d: %p SNAPDIR\n",
1455                              pos, temp);
1456                 } else if (stop_on_nosnap && inode &&
1457                            ceph_snap(inode) == CEPH_NOSNAP) {
1458                         break;
1459                 } else {
1460                         pos -= temp->d_name.len;
1461                         if (pos < 0)
1462                                 break;
1463                         strncpy(path + pos, temp->d_name.name,
1464                                 temp->d_name.len);
1465                 }
1466                 if (pos)
1467                         path[--pos] = '/';
1468                 temp = temp->d_parent;
1469                 if (temp == NULL) {
1470                         pr_err("build_path corrupt dentry\n");
1471                         kfree(path);
1472                         return ERR_PTR(-EINVAL);
1473                 }
1474         }
1475         if (pos != 0) {
1476                 pr_err("build_path did not end path lookup where "
1477                        "expected, namelen is %d, pos is %d\n", len, pos);
1478                 /* presumably this is only possible if racing with a
1479                    rename of one of the parent directories (we can not
1480                    lock the dentries above us to prevent this, but
1481                    retrying should be harmless) */
1482                 kfree(path);
1483                 goto retry;
1484         }
1485
1486         *base = ceph_ino(temp->d_inode);
1487         *plen = len;
1488         dout("build_path on %p %d built %llx '%.*s'\n",
1489              dentry, dentry->d_count, *base, len, path);
1490         return path;
1491 }
1492
1493 static int build_dentry_path(struct dentry *dentry,
1494                              const char **ppath, int *ppathlen, u64 *pino,
1495                              int *pfreepath)
1496 {
1497         char *path;
1498
1499         if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1500                 *pino = ceph_ino(dentry->d_parent->d_inode);
1501                 *ppath = dentry->d_name.name;
1502                 *ppathlen = dentry->d_name.len;
1503                 return 0;
1504         }
1505         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1506         if (IS_ERR(path))
1507                 return PTR_ERR(path);
1508         *ppath = path;
1509         *pfreepath = 1;
1510         return 0;
1511 }
1512
1513 static int build_inode_path(struct inode *inode,
1514                             const char **ppath, int *ppathlen, u64 *pino,
1515                             int *pfreepath)
1516 {
1517         struct dentry *dentry;
1518         char *path;
1519
1520         if (ceph_snap(inode) == CEPH_NOSNAP) {
1521                 *pino = ceph_ino(inode);
1522                 *ppathlen = 0;
1523                 return 0;
1524         }
1525         dentry = d_find_alias(inode);
1526         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1527         dput(dentry);
1528         if (IS_ERR(path))
1529                 return PTR_ERR(path);
1530         *ppath = path;
1531         *pfreepath = 1;
1532         return 0;
1533 }
1534
1535 /*
1536  * request arguments may be specified via an inode *, a dentry *, or
1537  * an explicit ino+path.
1538  */
1539 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1540                                   const char *rpath, u64 rino,
1541                                   const char **ppath, int *pathlen,
1542                                   u64 *ino, int *freepath)
1543 {
1544         int r = 0;
1545
1546         if (rinode) {
1547                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1548                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1549                      ceph_snap(rinode));
1550         } else if (rdentry) {
1551                 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1552                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1553                      *ppath);
1554         } else if (rpath) {
1555                 *ino = rino;
1556                 *ppath = rpath;
1557                 *pathlen = strlen(rpath);
1558                 dout(" path %.*s\n", *pathlen, rpath);
1559         }
1560
1561         return r;
1562 }
1563
1564 /*
1565  * called under mdsc->mutex
1566  */
1567 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1568                                                struct ceph_mds_request *req,
1569                                                int mds)
1570 {
1571         struct ceph_msg *msg;
1572         struct ceph_mds_request_head *head;
1573         const char *path1 = NULL;
1574         const char *path2 = NULL;
1575         u64 ino1 = 0, ino2 = 0;
1576         int pathlen1 = 0, pathlen2 = 0;
1577         int freepath1 = 0, freepath2 = 0;
1578         int len;
1579         u16 releases;
1580         void *p, *end;
1581         int ret;
1582
1583         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1584                               req->r_path1, req->r_ino1.ino,
1585                               &path1, &pathlen1, &ino1, &freepath1);
1586         if (ret < 0) {
1587                 msg = ERR_PTR(ret);
1588                 goto out;
1589         }
1590
1591         ret = set_request_path_attr(NULL, req->r_old_dentry,
1592                               req->r_path2, req->r_ino2.ino,
1593                               &path2, &pathlen2, &ino2, &freepath2);
1594         if (ret < 0) {
1595                 msg = ERR_PTR(ret);
1596                 goto out_free1;
1597         }
1598
1599         len = sizeof(*head) +
1600                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64));
1601
1602         /* calculate (max) length for cap releases */
1603         len += sizeof(struct ceph_mds_request_release) *
1604                 (!!req->r_inode_drop + !!req->r_dentry_drop +
1605                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1606         if (req->r_dentry_drop)
1607                 len += req->r_dentry->d_name.len;
1608         if (req->r_old_dentry_drop)
1609                 len += req->r_old_dentry->d_name.len;
1610
1611         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS);
1612         if (!msg) {
1613                 msg = ERR_PTR(-ENOMEM);
1614                 goto out_free2;
1615         }
1616
1617         msg->hdr.tid = cpu_to_le64(req->r_tid);
1618
1619         head = msg->front.iov_base;
1620         p = msg->front.iov_base + sizeof(*head);
1621         end = msg->front.iov_base + msg->front.iov_len;
1622
1623         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1624         head->op = cpu_to_le32(req->r_op);
1625         head->caller_uid = cpu_to_le32(req->r_uid);
1626         head->caller_gid = cpu_to_le32(req->r_gid);
1627         head->args = req->r_args;
1628
1629         ceph_encode_filepath(&p, end, ino1, path1);
1630         ceph_encode_filepath(&p, end, ino2, path2);
1631
1632         /* make note of release offset, in case we need to replay */
1633         req->r_request_release_offset = p - msg->front.iov_base;
1634
1635         /* cap releases */
1636         releases = 0;
1637         if (req->r_inode_drop)
1638                 releases += ceph_encode_inode_release(&p,
1639                       req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1640                       mds, req->r_inode_drop, req->r_inode_unless, 0);
1641         if (req->r_dentry_drop)
1642                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1643                        mds, req->r_dentry_drop, req->r_dentry_unless);
1644         if (req->r_old_dentry_drop)
1645                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1646                        mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1647         if (req->r_old_inode_drop)
1648                 releases += ceph_encode_inode_release(&p,
1649                       req->r_old_dentry->d_inode,
1650                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1651         head->num_releases = cpu_to_le16(releases);
1652
1653         BUG_ON(p > end);
1654         msg->front.iov_len = p - msg->front.iov_base;
1655         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1656
1657         msg->pages = req->r_pages;
1658         msg->nr_pages = req->r_num_pages;
1659         msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1660         msg->hdr.data_off = cpu_to_le16(0);
1661
1662 out_free2:
1663         if (freepath2)
1664                 kfree((char *)path2);
1665 out_free1:
1666         if (freepath1)
1667                 kfree((char *)path1);
1668 out:
1669         return msg;
1670 }
1671
1672 /*
1673  * called under mdsc->mutex if error, under no mutex if
1674  * success.
1675  */
1676 static void complete_request(struct ceph_mds_client *mdsc,
1677                              struct ceph_mds_request *req)
1678 {
1679         if (req->r_callback)
1680                 req->r_callback(mdsc, req);
1681         else
1682                 complete_all(&req->r_completion);
1683 }
1684
1685 /*
1686  * called under mdsc->mutex
1687  */
1688 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1689                                   struct ceph_mds_request *req,
1690                                   int mds)
1691 {
1692         struct ceph_mds_request_head *rhead;
1693         struct ceph_msg *msg;
1694         int flags = 0;
1695
1696         req->r_mds = mds;
1697         req->r_attempts++;
1698         if (req->r_inode) {
1699                 struct ceph_cap *cap =
1700                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
1701
1702                 if (cap)
1703                         req->r_sent_on_mseq = cap->mseq;
1704                 else
1705                         req->r_sent_on_mseq = -1;
1706         }
1707         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1708              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1709
1710         if (req->r_got_unsafe) {
1711                 /*
1712                  * Replay.  Do not regenerate message (and rebuild
1713                  * paths, etc.); just use the original message.
1714                  * Rebuilding paths will break for renames because
1715                  * d_move mangles the src name.
1716                  */
1717                 msg = req->r_request;
1718                 rhead = msg->front.iov_base;
1719
1720                 flags = le32_to_cpu(rhead->flags);
1721                 flags |= CEPH_MDS_FLAG_REPLAY;
1722                 rhead->flags = cpu_to_le32(flags);
1723
1724                 if (req->r_target_inode)
1725                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1726
1727                 rhead->num_retry = req->r_attempts - 1;
1728
1729                 /* remove cap/dentry releases from message */
1730                 rhead->num_releases = 0;
1731                 msg->hdr.front_len = cpu_to_le32(req->r_request_release_offset);
1732                 msg->front.iov_len = req->r_request_release_offset;
1733                 return 0;
1734         }
1735
1736         if (req->r_request) {
1737                 ceph_msg_put(req->r_request);
1738                 req->r_request = NULL;
1739         }
1740         msg = create_request_message(mdsc, req, mds);
1741         if (IS_ERR(msg)) {
1742                 req->r_err = PTR_ERR(msg);
1743                 complete_request(mdsc, req);
1744                 return PTR_ERR(msg);
1745         }
1746         req->r_request = msg;
1747
1748         rhead = msg->front.iov_base;
1749         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1750         if (req->r_got_unsafe)
1751                 flags |= CEPH_MDS_FLAG_REPLAY;
1752         if (req->r_locked_dir)
1753                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1754         rhead->flags = cpu_to_le32(flags);
1755         rhead->num_fwd = req->r_num_fwd;
1756         rhead->num_retry = req->r_attempts - 1;
1757         rhead->ino = 0;
1758
1759         dout(" r_locked_dir = %p\n", req->r_locked_dir);
1760         return 0;
1761 }
1762
1763 /*
1764  * send request, or put it on the appropriate wait list.
1765  */
1766 static int __do_request(struct ceph_mds_client *mdsc,
1767                         struct ceph_mds_request *req)
1768 {
1769         struct ceph_mds_session *session = NULL;
1770         int mds = -1;
1771         int err = -EAGAIN;
1772
1773         if (req->r_err || req->r_got_result)
1774                 goto out;
1775
1776         if (req->r_timeout &&
1777             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1778                 dout("do_request timed out\n");
1779                 err = -EIO;
1780                 goto finish;
1781         }
1782
1783         mds = __choose_mds(mdsc, req);
1784         if (mds < 0 ||
1785             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1786                 dout("do_request no mds or not active, waiting for map\n");
1787                 list_add(&req->r_wait, &mdsc->waiting_for_map);
1788                 goto out;
1789         }
1790
1791         /* get, open session */
1792         session = __ceph_lookup_mds_session(mdsc, mds);
1793         if (!session) {
1794                 session = register_session(mdsc, mds);
1795                 if (IS_ERR(session)) {
1796                         err = PTR_ERR(session);
1797                         goto finish;
1798                 }
1799         }
1800         dout("do_request mds%d session %p state %s\n", mds, session,
1801              session_state_name(session->s_state));
1802         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1803             session->s_state != CEPH_MDS_SESSION_HUNG) {
1804                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1805                     session->s_state == CEPH_MDS_SESSION_CLOSING)
1806                         __open_session(mdsc, session);
1807                 list_add(&req->r_wait, &session->s_waiting);
1808                 goto out_session;
1809         }
1810
1811         /* send request */
1812         req->r_session = get_session(session);
1813         req->r_resend_mds = -1;   /* forget any previous mds hint */
1814
1815         if (req->r_request_started == 0)   /* note request start time */
1816                 req->r_request_started = jiffies;
1817
1818         err = __prepare_send_request(mdsc, req, mds);
1819         if (!err) {
1820                 ceph_msg_get(req->r_request);
1821                 ceph_con_send(&session->s_con, req->r_request);
1822         }
1823
1824 out_session:
1825         ceph_put_mds_session(session);
1826 out:
1827         return err;
1828
1829 finish:
1830         req->r_err = err;
1831         complete_request(mdsc, req);
1832         goto out;
1833 }
1834
1835 /*
1836  * called under mdsc->mutex
1837  */
1838 static void __wake_requests(struct ceph_mds_client *mdsc,
1839                             struct list_head *head)
1840 {
1841         struct ceph_mds_request *req, *nreq;
1842
1843         list_for_each_entry_safe(req, nreq, head, r_wait) {
1844                 list_del_init(&req->r_wait);
1845                 __do_request(mdsc, req);
1846         }
1847 }
1848
1849 /*
1850  * Wake up threads with requests pending for @mds, so that they can
1851  * resubmit their requests to a possibly different mds.
1852  */
1853 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
1854 {
1855         struct ceph_mds_request *req;
1856         struct rb_node *p;
1857
1858         dout("kick_requests mds%d\n", mds);
1859         for (p = rb_first(&mdsc->request_tree); p; p = rb_next(p)) {
1860                 req = rb_entry(p, struct ceph_mds_request, r_node);
1861                 if (req->r_got_unsafe)
1862                         continue;
1863                 if (req->r_session &&
1864                     req->r_session->s_mds == mds) {
1865                         dout(" kicking tid %llu\n", req->r_tid);
1866                         put_request_session(req);
1867                         __do_request(mdsc, req);
1868                 }
1869         }
1870 }
1871
1872 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
1873                               struct ceph_mds_request *req)
1874 {
1875         dout("submit_request on %p\n", req);
1876         mutex_lock(&mdsc->mutex);
1877         __register_request(mdsc, req, NULL);
1878         __do_request(mdsc, req);
1879         mutex_unlock(&mdsc->mutex);
1880 }
1881
1882 /*
1883  * Synchrously perform an mds request.  Take care of all of the
1884  * session setup, forwarding, retry details.
1885  */
1886 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
1887                          struct inode *dir,
1888                          struct ceph_mds_request *req)
1889 {
1890         int err;
1891
1892         dout("do_request on %p\n", req);
1893
1894         /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
1895         if (req->r_inode)
1896                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
1897         if (req->r_locked_dir)
1898                 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
1899         if (req->r_old_dentry)
1900                 ceph_get_cap_refs(
1901                         ceph_inode(req->r_old_dentry->d_parent->d_inode),
1902                         CEPH_CAP_PIN);
1903
1904         /* issue */
1905         mutex_lock(&mdsc->mutex);
1906         __register_request(mdsc, req, dir);
1907         __do_request(mdsc, req);
1908
1909         if (req->r_err) {
1910                 err = req->r_err;
1911                 __unregister_request(mdsc, req);
1912                 dout("do_request early error %d\n", err);
1913                 goto out;
1914         }
1915
1916         /* wait */
1917         mutex_unlock(&mdsc->mutex);
1918         dout("do_request waiting\n");
1919         if (req->r_timeout) {
1920                 err = (long)wait_for_completion_killable_timeout(
1921                         &req->r_completion, req->r_timeout);
1922                 if (err == 0)
1923                         err = -EIO;
1924         } else {
1925                 err = wait_for_completion_killable(&req->r_completion);
1926         }
1927         dout("do_request waited, got %d\n", err);
1928         mutex_lock(&mdsc->mutex);
1929
1930         /* only abort if we didn't race with a real reply */
1931         if (req->r_got_result) {
1932                 err = le32_to_cpu(req->r_reply_info.head->result);
1933         } else if (err < 0) {
1934                 dout("aborted request %lld with %d\n", req->r_tid, err);
1935
1936                 /*
1937                  * ensure we aren't running concurrently with
1938                  * ceph_fill_trace or ceph_readdir_prepopulate, which
1939                  * rely on locks (dir mutex) held by our caller.
1940                  */
1941                 mutex_lock(&req->r_fill_mutex);
1942                 req->r_err = err;
1943                 req->r_aborted = true;
1944                 mutex_unlock(&req->r_fill_mutex);
1945
1946                 if (req->r_locked_dir &&
1947                     (req->r_op & CEPH_MDS_OP_WRITE))
1948                         ceph_invalidate_dir_request(req);
1949         } else {
1950                 err = req->r_err;
1951         }
1952
1953 out:
1954         mutex_unlock(&mdsc->mutex);
1955         dout("do_request %p done, result %d\n", req, err);
1956         return err;
1957 }
1958
1959 /*
1960  * Invalidate dir I_COMPLETE, dentry lease state on an aborted MDS
1961  * namespace request.
1962  */
1963 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
1964 {
1965         struct inode *inode = req->r_locked_dir;
1966         struct ceph_inode_info *ci = ceph_inode(inode);
1967
1968         dout("invalidate_dir_request %p (I_COMPLETE, lease(s))\n", inode);
1969         spin_lock(&inode->i_lock);
1970         ci->i_ceph_flags &= ~CEPH_I_COMPLETE;
1971         ci->i_release_count++;
1972         spin_unlock(&inode->i_lock);
1973
1974         if (req->r_dentry)
1975                 ceph_invalidate_dentry_lease(req->r_dentry);
1976         if (req->r_old_dentry)
1977                 ceph_invalidate_dentry_lease(req->r_old_dentry);
1978 }
1979
1980 /*
1981  * Handle mds reply.
1982  *
1983  * We take the session mutex and parse and process the reply immediately.
1984  * This preserves the logical ordering of replies, capabilities, etc., sent
1985  * by the MDS as they are applied to our local cache.
1986  */
1987 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
1988 {
1989         struct ceph_mds_client *mdsc = session->s_mdsc;
1990         struct ceph_mds_request *req;
1991         struct ceph_mds_reply_head *head = msg->front.iov_base;
1992         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
1993         u64 tid;
1994         int err, result;
1995         int mds = session->s_mds;
1996
1997         if (msg->front.iov_len < sizeof(*head)) {
1998                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
1999                 ceph_msg_dump(msg);
2000                 return;
2001         }
2002
2003         /* get request, session */
2004         tid = le64_to_cpu(msg->hdr.tid);
2005         mutex_lock(&mdsc->mutex);
2006         req = __lookup_request(mdsc, tid);
2007         if (!req) {
2008                 dout("handle_reply on unknown tid %llu\n", tid);
2009                 mutex_unlock(&mdsc->mutex);
2010                 return;
2011         }
2012         dout("handle_reply %p\n", req);
2013
2014         /* correct session? */
2015         if (req->r_session != session) {
2016                 pr_err("mdsc_handle_reply got %llu on session mds%d"
2017                        " not mds%d\n", tid, session->s_mds,
2018                        req->r_session ? req->r_session->s_mds : -1);
2019                 mutex_unlock(&mdsc->mutex);
2020                 goto out;
2021         }
2022
2023         /* dup? */
2024         if ((req->r_got_unsafe && !head->safe) ||
2025             (req->r_got_safe && head->safe)) {
2026                 pr_warning("got a dup %s reply on %llu from mds%d\n",
2027                            head->safe ? "safe" : "unsafe", tid, mds);
2028                 mutex_unlock(&mdsc->mutex);
2029                 goto out;
2030         }
2031         if (req->r_got_safe && !head->safe) {
2032                 pr_warning("got unsafe after safe on %llu from mds%d\n",
2033                            tid, mds);
2034                 mutex_unlock(&mdsc->mutex);
2035                 goto out;
2036         }
2037
2038         result = le32_to_cpu(head->result);
2039
2040         /*
2041          * Handle an ESTALE
2042          * if we're not talking to the authority, send to them
2043          * if the authority has changed while we weren't looking,
2044          * send to new authority
2045          * Otherwise we just have to return an ESTALE
2046          */
2047         if (result == -ESTALE) {
2048                 dout("got ESTALE on request %llu", req->r_tid);
2049                 if (!req->r_inode) {
2050                         /* do nothing; not an authority problem */
2051                 } else if (req->r_direct_mode != USE_AUTH_MDS) {
2052                         dout("not using auth, setting for that now");
2053                         req->r_direct_mode = USE_AUTH_MDS;
2054                         __do_request(mdsc, req);
2055                         mutex_unlock(&mdsc->mutex);
2056                         goto out;
2057                 } else  {
2058                         struct ceph_inode_info *ci = ceph_inode(req->r_inode);
2059                         struct ceph_cap *cap =
2060                                 ceph_get_cap_for_mds(ci, req->r_mds);;
2061
2062                         dout("already using auth");
2063                         if ((!cap || cap != ci->i_auth_cap) ||
2064                             (cap->mseq != req->r_sent_on_mseq)) {
2065                                 dout("but cap changed, so resending");
2066                                 __do_request(mdsc, req);
2067                                 mutex_unlock(&mdsc->mutex);
2068                                 goto out;
2069                         }
2070                 }
2071                 dout("have to return ESTALE on request %llu", req->r_tid);
2072         }
2073
2074
2075         if (head->safe) {
2076                 req->r_got_safe = true;
2077                 __unregister_request(mdsc, req);
2078                 complete_all(&req->r_safe_completion);
2079
2080                 if (req->r_got_unsafe) {
2081                         /*
2082                          * We already handled the unsafe response, now do the
2083                          * cleanup.  No need to examine the response; the MDS
2084                          * doesn't include any result info in the safe
2085                          * response.  And even if it did, there is nothing
2086                          * useful we could do with a revised return value.
2087                          */
2088                         dout("got safe reply %llu, mds%d\n", tid, mds);
2089                         list_del_init(&req->r_unsafe_item);
2090
2091                         /* last unsafe request during umount? */
2092                         if (mdsc->stopping && !__get_oldest_req(mdsc))
2093                                 complete_all(&mdsc->safe_umount_waiters);
2094                         mutex_unlock(&mdsc->mutex);
2095                         goto out;
2096                 }
2097         } else {
2098                 req->r_got_unsafe = true;
2099                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2100         }
2101
2102         dout("handle_reply tid %lld result %d\n", tid, result);
2103         rinfo = &req->r_reply_info;
2104         err = parse_reply_info(msg, rinfo);
2105         mutex_unlock(&mdsc->mutex);
2106
2107         mutex_lock(&session->s_mutex);
2108         if (err < 0) {
2109                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2110                 ceph_msg_dump(msg);
2111                 goto out_err;
2112         }
2113
2114         /* snap trace */
2115         if (rinfo->snapblob_len) {
2116                 down_write(&mdsc->snap_rwsem);
2117                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2118                                rinfo->snapblob + rinfo->snapblob_len,
2119                                le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
2120                 downgrade_write(&mdsc->snap_rwsem);
2121         } else {
2122                 down_read(&mdsc->snap_rwsem);
2123         }
2124
2125         /* insert trace into our cache */
2126         mutex_lock(&req->r_fill_mutex);
2127         err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2128         if (err == 0) {
2129                 if (result == 0 && req->r_op != CEPH_MDS_OP_GETFILELOCK &&
2130                     rinfo->dir_nr)
2131                         ceph_readdir_prepopulate(req, req->r_session);
2132                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2133         }
2134         mutex_unlock(&req->r_fill_mutex);
2135
2136         up_read(&mdsc->snap_rwsem);
2137 out_err:
2138         mutex_lock(&mdsc->mutex);
2139         if (!req->r_aborted) {
2140                 if (err) {
2141                         req->r_err = err;
2142                 } else {
2143                         req->r_reply = msg;
2144                         ceph_msg_get(msg);
2145                         req->r_got_result = true;
2146                 }
2147         } else {
2148                 dout("reply arrived after request %lld was aborted\n", tid);
2149         }
2150         mutex_unlock(&mdsc->mutex);
2151
2152         ceph_add_cap_releases(mdsc, req->r_session);
2153         mutex_unlock(&session->s_mutex);
2154
2155         /* kick calling process */
2156         complete_request(mdsc, req);
2157 out:
2158         ceph_mdsc_put_request(req);
2159         return;
2160 }
2161
2162
2163
2164 /*
2165  * handle mds notification that our request has been forwarded.
2166  */
2167 static void handle_forward(struct ceph_mds_client *mdsc,
2168                            struct ceph_mds_session *session,
2169                            struct ceph_msg *msg)
2170 {
2171         struct ceph_mds_request *req;
2172         u64 tid = le64_to_cpu(msg->hdr.tid);
2173         u32 next_mds;
2174         u32 fwd_seq;
2175         int err = -EINVAL;
2176         void *p = msg->front.iov_base;
2177         void *end = p + msg->front.iov_len;
2178
2179         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2180         next_mds = ceph_decode_32(&p);
2181         fwd_seq = ceph_decode_32(&p);
2182
2183         mutex_lock(&mdsc->mutex);
2184         req = __lookup_request(mdsc, tid);
2185         if (!req) {
2186                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2187                 goto out;  /* dup reply? */
2188         }
2189
2190         if (req->r_aborted) {
2191                 dout("forward tid %llu aborted, unregistering\n", tid);
2192                 __unregister_request(mdsc, req);
2193         } else if (fwd_seq <= req->r_num_fwd) {
2194                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2195                      tid, next_mds, req->r_num_fwd, fwd_seq);
2196         } else {
2197                 /* resend. forward race not possible; mds would drop */
2198                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2199                 BUG_ON(req->r_err);
2200                 BUG_ON(req->r_got_result);
2201                 req->r_num_fwd = fwd_seq;
2202                 req->r_resend_mds = next_mds;
2203                 put_request_session(req);
2204                 __do_request(mdsc, req);
2205         }
2206         ceph_mdsc_put_request(req);
2207 out:
2208         mutex_unlock(&mdsc->mutex);
2209         return;
2210
2211 bad:
2212         pr_err("mdsc_handle_forward decode error err=%d\n", err);
2213 }
2214
2215 /*
2216  * handle a mds session control message
2217  */
2218 static void handle_session(struct ceph_mds_session *session,
2219                            struct ceph_msg *msg)
2220 {
2221         struct ceph_mds_client *mdsc = session->s_mdsc;
2222         u32 op;
2223         u64 seq;
2224         int mds = session->s_mds;
2225         struct ceph_mds_session_head *h = msg->front.iov_base;
2226         int wake = 0;
2227
2228         /* decode */
2229         if (msg->front.iov_len != sizeof(*h))
2230                 goto bad;
2231         op = le32_to_cpu(h->op);
2232         seq = le64_to_cpu(h->seq);
2233
2234         mutex_lock(&mdsc->mutex);
2235         if (op == CEPH_SESSION_CLOSE)
2236                 __unregister_session(mdsc, session);
2237         /* FIXME: this ttl calculation is generous */
2238         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2239         mutex_unlock(&mdsc->mutex);
2240
2241         mutex_lock(&session->s_mutex);
2242
2243         dout("handle_session mds%d %s %p state %s seq %llu\n",
2244              mds, ceph_session_op_name(op), session,
2245              session_state_name(session->s_state), seq);
2246
2247         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2248                 session->s_state = CEPH_MDS_SESSION_OPEN;
2249                 pr_info("mds%d came back\n", session->s_mds);
2250         }
2251
2252         switch (op) {
2253         case CEPH_SESSION_OPEN:
2254                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2255                         pr_info("mds%d reconnect success\n", session->s_mds);
2256                 session->s_state = CEPH_MDS_SESSION_OPEN;
2257                 renewed_caps(mdsc, session, 0);
2258                 wake = 1;
2259                 if (mdsc->stopping)
2260                         __close_session(mdsc, session);
2261                 break;
2262
2263         case CEPH_SESSION_RENEWCAPS:
2264                 if (session->s_renew_seq == seq)
2265                         renewed_caps(mdsc, session, 1);
2266                 break;
2267
2268         case CEPH_SESSION_CLOSE:
2269                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2270                         pr_info("mds%d reconnect denied\n", session->s_mds);
2271                 remove_session_caps(session);
2272                 wake = 1; /* for good measure */
2273                 wake_up_all(&mdsc->session_close_wq);
2274                 kick_requests(mdsc, mds);
2275                 break;
2276
2277         case CEPH_SESSION_STALE:
2278                 pr_info("mds%d caps went stale, renewing\n",
2279                         session->s_mds);
2280                 spin_lock(&session->s_cap_lock);
2281                 session->s_cap_gen++;
2282                 session->s_cap_ttl = 0;
2283                 spin_unlock(&session->s_cap_lock);
2284                 send_renew_caps(mdsc, session);
2285                 break;
2286
2287         case CEPH_SESSION_RECALL_STATE:
2288                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2289                 break;
2290
2291         default:
2292                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2293                 WARN_ON(1);
2294         }
2295
2296         mutex_unlock(&session->s_mutex);
2297         if (wake) {
2298                 mutex_lock(&mdsc->mutex);
2299                 __wake_requests(mdsc, &session->s_waiting);
2300                 mutex_unlock(&mdsc->mutex);
2301         }
2302         return;
2303
2304 bad:
2305         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2306                (int)msg->front.iov_len);
2307         ceph_msg_dump(msg);
2308         return;
2309 }
2310
2311
2312 /*
2313  * called under session->mutex.
2314  */
2315 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2316                                    struct ceph_mds_session *session)
2317 {
2318         struct ceph_mds_request *req, *nreq;
2319         int err;
2320
2321         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2322
2323         mutex_lock(&mdsc->mutex);
2324         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2325                 err = __prepare_send_request(mdsc, req, session->s_mds);
2326                 if (!err) {
2327                         ceph_msg_get(req->r_request);
2328                         ceph_con_send(&session->s_con, req->r_request);
2329                 }
2330         }
2331         mutex_unlock(&mdsc->mutex);
2332 }
2333
2334 /*
2335  * Encode information about a cap for a reconnect with the MDS.
2336  */
2337 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2338                           void *arg)
2339 {
2340         union {
2341                 struct ceph_mds_cap_reconnect v2;
2342                 struct ceph_mds_cap_reconnect_v1 v1;
2343         } rec;
2344         size_t reclen;
2345         struct ceph_inode_info *ci;
2346         struct ceph_reconnect_state *recon_state = arg;
2347         struct ceph_pagelist *pagelist = recon_state->pagelist;
2348         char *path;
2349         int pathlen, err;
2350         u64 pathbase;
2351         struct dentry *dentry;
2352
2353         ci = cap->ci;
2354
2355         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2356              inode, ceph_vinop(inode), cap, cap->cap_id,
2357              ceph_cap_string(cap->issued));
2358         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2359         if (err)
2360                 return err;
2361
2362         dentry = d_find_alias(inode);
2363         if (dentry) {
2364                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2365                 if (IS_ERR(path)) {
2366                         err = PTR_ERR(path);
2367                         goto out_dput;
2368                 }
2369         } else {
2370                 path = NULL;
2371                 pathlen = 0;
2372         }
2373         err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2374         if (err)
2375                 goto out_free;
2376
2377         spin_lock(&inode->i_lock);
2378         cap->seq = 0;        /* reset cap seq */
2379         cap->issue_seq = 0;  /* and issue_seq */
2380
2381         if (recon_state->flock) {
2382                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2383                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2384                 rec.v2.issued = cpu_to_le32(cap->issued);
2385                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2386                 rec.v2.pathbase = cpu_to_le64(pathbase);
2387                 rec.v2.flock_len = 0;
2388                 reclen = sizeof(rec.v2);
2389         } else {
2390                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2391                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2392                 rec.v1.issued = cpu_to_le32(cap->issued);
2393                 rec.v1.size = cpu_to_le64(inode->i_size);
2394                 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2395                 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2396                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2397                 rec.v1.pathbase = cpu_to_le64(pathbase);
2398                 reclen = sizeof(rec.v1);
2399         }
2400         spin_unlock(&inode->i_lock);
2401
2402         if (recon_state->flock) {
2403                 int num_fcntl_locks, num_flock_locks;
2404                 struct ceph_pagelist_cursor trunc_point;
2405
2406                 ceph_pagelist_set_cursor(pagelist, &trunc_point);
2407                 do {
2408                         lock_flocks();
2409                         ceph_count_locks(inode, &num_fcntl_locks,
2410                                          &num_flock_locks);
2411                         rec.v2.flock_len = (2*sizeof(u32) +
2412                                             (num_fcntl_locks+num_flock_locks) *
2413                                             sizeof(struct ceph_filelock));
2414                         unlock_flocks();
2415
2416                         /* pre-alloc pagelist */
2417                         ceph_pagelist_truncate(pagelist, &trunc_point);
2418                         err = ceph_pagelist_append(pagelist, &rec, reclen);
2419                         if (!err)
2420                                 err = ceph_pagelist_reserve(pagelist,
2421                                                             rec.v2.flock_len);
2422
2423                         /* encode locks */
2424                         if (!err) {
2425                                 lock_flocks();
2426                                 err = ceph_encode_locks(inode,
2427                                                         pagelist,
2428                                                         num_fcntl_locks,
2429                                                         num_flock_locks);
2430                                 unlock_flocks();
2431                         }
2432                 } while (err == -ENOSPC);
2433         } else {
2434                 err = ceph_pagelist_append(pagelist, &rec, reclen);
2435         }
2436
2437 out_free:
2438         kfree(path);
2439 out_dput:
2440         dput(dentry);
2441         return err;
2442 }
2443
2444
2445 /*
2446  * If an MDS fails and recovers, clients need to reconnect in order to
2447  * reestablish shared state.  This includes all caps issued through
2448  * this session _and_ the snap_realm hierarchy.  Because it's not
2449  * clear which snap realms the mds cares about, we send everything we
2450  * know about.. that ensures we'll then get any new info the
2451  * recovering MDS might have.
2452  *
2453  * This is a relatively heavyweight operation, but it's rare.
2454  *
2455  * called with mdsc->mutex held.
2456  */
2457 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2458                                struct ceph_mds_session *session)
2459 {
2460         struct ceph_msg *reply;
2461         struct rb_node *p;
2462         int mds = session->s_mds;
2463         int err = -ENOMEM;
2464         struct ceph_pagelist *pagelist;
2465         struct ceph_reconnect_state recon_state;
2466
2467         pr_info("mds%d reconnect start\n", mds);
2468
2469         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2470         if (!pagelist)
2471                 goto fail_nopagelist;
2472         ceph_pagelist_init(pagelist);
2473
2474         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS);
2475         if (!reply)
2476                 goto fail_nomsg;
2477
2478         mutex_lock(&session->s_mutex);
2479         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2480         session->s_seq = 0;
2481
2482         ceph_con_open(&session->s_con,
2483                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2484
2485         /* replay unsafe requests */
2486         replay_unsafe_requests(mdsc, session);
2487
2488         down_read(&mdsc->snap_rwsem);
2489
2490         dout("session %p state %s\n", session,
2491              session_state_name(session->s_state));
2492
2493         /* drop old cap expires; we're about to reestablish that state */
2494         discard_cap_releases(mdsc, session);
2495
2496         /* traverse this session's caps */
2497         err = ceph_pagelist_encode_32(pagelist, session->s_nr_caps);
2498         if (err)
2499                 goto fail;
2500
2501         recon_state.pagelist = pagelist;
2502         recon_state.flock = session->s_con.peer_features & CEPH_FEATURE_FLOCK;
2503         err = iterate_session_caps(session, encode_caps_cb, &recon_state);
2504         if (err < 0)
2505                 goto fail;
2506
2507         /*
2508          * snaprealms.  we provide mds with the ino, seq (version), and
2509          * parent for all of our realms.  If the mds has any newer info,
2510          * it will tell us.
2511          */
2512         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2513                 struct ceph_snap_realm *realm =
2514                         rb_entry(p, struct ceph_snap_realm, node);
2515                 struct ceph_mds_snaprealm_reconnect sr_rec;
2516
2517                 dout(" adding snap realm %llx seq %lld parent %llx\n",
2518                      realm->ino, realm->seq, realm->parent_ino);
2519                 sr_rec.ino = cpu_to_le64(realm->ino);
2520                 sr_rec.seq = cpu_to_le64(realm->seq);
2521                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2522                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2523                 if (err)
2524                         goto fail;
2525         }
2526
2527         reply->pagelist = pagelist;
2528         if (recon_state.flock)
2529                 reply->hdr.version = cpu_to_le16(2);
2530         reply->hdr.data_len = cpu_to_le32(pagelist->length);
2531         reply->nr_pages = calc_pages_for(0, pagelist->length);
2532         ceph_con_send(&session->s_con, reply);
2533
2534         mutex_unlock(&session->s_mutex);
2535
2536         mutex_lock(&mdsc->mutex);
2537         __wake_requests(mdsc, &session->s_waiting);
2538         mutex_unlock(&mdsc->mutex);
2539
2540         up_read(&mdsc->snap_rwsem);
2541         return;
2542
2543 fail:
2544         ceph_msg_put(reply);
2545         up_read(&mdsc->snap_rwsem);
2546         mutex_unlock(&session->s_mutex);
2547 fail_nomsg:
2548         ceph_pagelist_release(pagelist);
2549         kfree(pagelist);
2550 fail_nopagelist:
2551         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
2552         return;
2553 }
2554
2555
2556 /*
2557  * compare old and new mdsmaps, kicking requests
2558  * and closing out old connections as necessary
2559  *
2560  * called under mdsc->mutex.
2561  */
2562 static void check_new_map(struct ceph_mds_client *mdsc,
2563                           struct ceph_mdsmap *newmap,
2564                           struct ceph_mdsmap *oldmap)
2565 {
2566         int i;
2567         int oldstate, newstate;
2568         struct ceph_mds_session *s;
2569
2570         dout("check_new_map new %u old %u\n",
2571              newmap->m_epoch, oldmap->m_epoch);
2572
2573         for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2574                 if (mdsc->sessions[i] == NULL)
2575                         continue;
2576                 s = mdsc->sessions[i];
2577                 oldstate = ceph_mdsmap_get_state(oldmap, i);
2578                 newstate = ceph_mdsmap_get_state(newmap, i);
2579
2580                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
2581                      i, ceph_mds_state_name(oldstate),
2582                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
2583                      ceph_mds_state_name(newstate),
2584                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
2585                      session_state_name(s->s_state));
2586
2587                 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
2588                            ceph_mdsmap_get_addr(newmap, i),
2589                            sizeof(struct ceph_entity_addr))) {
2590                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2591                                 /* the session never opened, just close it
2592                                  * out now */
2593                                 __wake_requests(mdsc, &s->s_waiting);
2594                                 __unregister_session(mdsc, s);
2595                         } else {
2596                                 /* just close it */
2597                                 mutex_unlock(&mdsc->mutex);
2598                                 mutex_lock(&s->s_mutex);
2599                                 mutex_lock(&mdsc->mutex);
2600                                 ceph_con_close(&s->s_con);
2601                                 mutex_unlock(&s->s_mutex);
2602                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2603                         }
2604
2605                         /* kick any requests waiting on the recovering mds */
2606                         kick_requests(mdsc, i);
2607                 } else if (oldstate == newstate) {
2608                         continue;  /* nothing new with this mds */
2609                 }
2610
2611                 /*
2612                  * send reconnect?
2613                  */
2614                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2615                     newstate >= CEPH_MDS_STATE_RECONNECT) {
2616                         mutex_unlock(&mdsc->mutex);
2617                         send_mds_reconnect(mdsc, s);
2618                         mutex_lock(&mdsc->mutex);
2619                 }
2620
2621                 /*
2622                  * kick request on any mds that has gone active.
2623                  */
2624                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2625                     newstate >= CEPH_MDS_STATE_ACTIVE) {
2626                         if (oldstate != CEPH_MDS_STATE_CREATING &&
2627                             oldstate != CEPH_MDS_STATE_STARTING)
2628                                 pr_info("mds%d recovery completed\n", s->s_mds);
2629                         kick_requests(mdsc, i);
2630                         ceph_kick_flushing_caps(mdsc, s);
2631                         wake_up_session_caps(s, 1);
2632                 }
2633         }
2634
2635         for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
2636                 s = mdsc->sessions[i];
2637                 if (!s)
2638                         continue;
2639                 if (!ceph_mdsmap_is_laggy(newmap, i))
2640                         continue;
2641                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2642                     s->s_state == CEPH_MDS_SESSION_HUNG ||
2643                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
2644                         dout(" connecting to export targets of laggy mds%d\n",
2645                              i);
2646                         __open_export_target_sessions(mdsc, s);
2647                 }
2648         }
2649 }
2650
2651
2652
2653 /*
2654  * leases
2655  */
2656
2657 /*
2658  * caller must hold session s_mutex, dentry->d_lock
2659  */
2660 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2661 {
2662         struct ceph_dentry_info *di = ceph_dentry(dentry);
2663
2664         ceph_put_mds_session(di->lease_session);
2665         di->lease_session = NULL;
2666 }
2667
2668 static void handle_lease(struct ceph_mds_client *mdsc,
2669                          struct ceph_mds_session *session,
2670                          struct ceph_msg *msg)
2671 {
2672         struct super_block *sb = mdsc->fsc->sb;
2673         struct inode *inode;
2674         struct ceph_inode_info *ci;
2675         struct dentry *parent, *dentry;
2676         struct ceph_dentry_info *di;
2677         int mds = session->s_mds;
2678         struct ceph_mds_lease *h = msg->front.iov_base;
2679         u32 seq;
2680         struct ceph_vino vino;
2681         int mask;
2682         struct qstr dname;
2683         int release = 0;
2684
2685         dout("handle_lease from mds%d\n", mds);
2686
2687         /* decode */
2688         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2689                 goto bad;
2690         vino.ino = le64_to_cpu(h->ino);
2691         vino.snap = CEPH_NOSNAP;
2692         mask = le16_to_cpu(h->mask);
2693         seq = le32_to_cpu(h->seq);
2694         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2695         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2696         if (dname.len != get_unaligned_le32(h+1))
2697                 goto bad;
2698
2699         mutex_lock(&session->s_mutex);
2700         session->s_seq++;
2701
2702         /* lookup inode */
2703         inode = ceph_find_inode(sb, vino);
2704         dout("handle_lease %s, mask %d, ino %llx %p %.*s\n",
2705              ceph_lease_op_name(h->action), mask, vino.ino, inode,
2706              dname.len, dname.name);
2707         if (inode == NULL) {
2708                 dout("handle_lease no inode %llx\n", vino.ino);
2709                 goto release;
2710         }
2711         ci = ceph_inode(inode);
2712
2713         /* dentry */
2714         parent = d_find_alias(inode);
2715         if (!parent) {
2716                 dout("no parent dentry on inode %p\n", inode);
2717                 WARN_ON(1);
2718                 goto release;  /* hrm... */
2719         }
2720         dname.hash = full_name_hash(dname.name, dname.len);
2721         dentry = d_lookup(parent, &dname);
2722         dput(parent);
2723         if (!dentry)
2724                 goto release;
2725
2726         spin_lock(&dentry->d_lock);
2727         di = ceph_dentry(dentry);
2728         switch (h->action) {
2729         case CEPH_MDS_LEASE_REVOKE:
2730                 if (di && di->lease_session == session) {
2731                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
2732                                 h->seq = cpu_to_le32(di->lease_seq);
2733                         __ceph_mdsc_drop_dentry_lease(dentry);
2734                 }
2735                 release = 1;
2736                 break;
2737
2738         case CEPH_MDS_LEASE_RENEW:
2739                 if (di && di->lease_session == session &&
2740                     di->lease_gen == session->s_cap_gen &&
2741                     di->lease_renew_from &&
2742                     di->lease_renew_after == 0) {
2743                         unsigned long duration =
2744                                 le32_to_cpu(h->duration_ms) * HZ / 1000;
2745
2746                         di->lease_seq = seq;
2747                         dentry->d_time = di->lease_renew_from + duration;
2748                         di->lease_renew_after = di->lease_renew_from +
2749                                 (duration >> 1);
2750                         di->lease_renew_from = 0;
2751                 }
2752                 break;
2753         }
2754         spin_unlock(&dentry->d_lock);
2755         dput(dentry);
2756
2757         if (!release)
2758                 goto out;
2759
2760 release:
2761         /* let's just reuse the same message */
2762         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2763         ceph_msg_get(msg);
2764         ceph_con_send(&session->s_con, msg);
2765
2766 out:
2767         iput(inode);
2768         mutex_unlock(&session->s_mutex);
2769         return;
2770
2771 bad:
2772         pr_err("corrupt lease message\n");
2773         ceph_msg_dump(msg);
2774 }
2775
2776 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
2777                               struct inode *inode,
2778                               struct dentry *dentry, char action,
2779                               u32 seq)
2780 {
2781         struct ceph_msg *msg;
2782         struct ceph_mds_lease *lease;
2783         int len = sizeof(*lease) + sizeof(u32);
2784         int dnamelen = 0;
2785
2786         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2787              inode, dentry, ceph_lease_op_name(action), session->s_mds);
2788         dnamelen = dentry->d_name.len;
2789         len += dnamelen;
2790
2791         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS);
2792         if (!msg)
2793                 return;
2794         lease = msg->front.iov_base;
2795         lease->action = action;
2796         lease->mask = cpu_to_le16(1);
2797         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
2798         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
2799         lease->seq = cpu_to_le32(seq);
2800         put_unaligned_le32(dnamelen, lease + 1);
2801         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
2802
2803         /*
2804          * if this is a preemptive lease RELEASE, no need to
2805          * flush request stream, since the actual request will
2806          * soon follow.
2807          */
2808         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
2809
2810         ceph_con_send(&session->s_con, msg);
2811 }
2812
2813 /*
2814  * Preemptively release a lease we expect to invalidate anyway.
2815  * Pass @inode always, @dentry is optional.
2816  */
2817 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2818                              struct dentry *dentry, int mask)
2819 {
2820         struct ceph_dentry_info *di;
2821         struct ceph_mds_session *session;
2822         u32 seq;
2823
2824         BUG_ON(inode == NULL);
2825         BUG_ON(dentry == NULL);
2826         BUG_ON(mask == 0);
2827
2828         /* is dentry lease valid? */
2829         spin_lock(&dentry->d_lock);
2830         di = ceph_dentry(dentry);
2831         if (!di || !di->lease_session ||
2832             di->lease_session->s_mds < 0 ||
2833             di->lease_gen != di->lease_session->s_cap_gen ||
2834             !time_before(jiffies, dentry->d_time)) {
2835                 dout("lease_release inode %p dentry %p -- "
2836                      "no lease on %d\n",
2837                      inode, dentry, mask);
2838                 spin_unlock(&dentry->d_lock);
2839                 return;
2840         }
2841
2842         /* we do have a lease on this dentry; note mds and seq */
2843         session = ceph_get_mds_session(di->lease_session);
2844         seq = di->lease_seq;
2845         __ceph_mdsc_drop_dentry_lease(dentry);
2846         spin_unlock(&dentry->d_lock);
2847
2848         dout("lease_release inode %p dentry %p mask %d to mds%d\n",
2849              inode, dentry, mask, session->s_mds);
2850         ceph_mdsc_lease_send_msg(session, inode, dentry,
2851                                  CEPH_MDS_LEASE_RELEASE, seq);
2852         ceph_put_mds_session(session);
2853 }
2854
2855 /*
2856  * drop all leases (and dentry refs) in preparation for umount
2857  */
2858 static void drop_leases(struct ceph_mds_client *mdsc)
2859 {
2860         int i;
2861
2862         dout("drop_leases\n");
2863         mutex_lock(&mdsc->mutex);
2864         for (i = 0; i < mdsc->max_sessions; i++) {
2865                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2866                 if (!s)
2867                         continue;
2868                 mutex_unlock(&mdsc->mutex);
2869                 mutex_lock(&s->s_mutex);
2870                 mutex_unlock(&s->s_mutex);
2871                 ceph_put_mds_session(s);
2872                 mutex_lock(&mdsc->mutex);
2873         }
2874         mutex_unlock(&mdsc->mutex);
2875 }
2876
2877
2878
2879 /*
2880  * delayed work -- periodically trim expired leases, renew caps with mds
2881  */
2882 static void schedule_delayed(struct ceph_mds_client *mdsc)
2883 {
2884         int delay = 5;
2885         unsigned hz = round_jiffies_relative(HZ * delay);
2886         schedule_delayed_work(&mdsc->delayed_work, hz);
2887 }
2888
2889 static void delayed_work(struct work_struct *work)
2890 {
2891         int i;
2892         struct ceph_mds_client *mdsc =
2893                 container_of(work, struct ceph_mds_client, delayed_work.work);
2894         int renew_interval;
2895         int renew_caps;
2896
2897         dout("mdsc delayed_work\n");
2898         ceph_check_delayed_caps(mdsc);
2899
2900         mutex_lock(&mdsc->mutex);
2901         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
2902         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
2903                                    mdsc->last_renew_caps);
2904         if (renew_caps)
2905                 mdsc->last_renew_caps = jiffies;
2906
2907         for (i = 0; i < mdsc->max_sessions; i++) {
2908                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2909                 if (s == NULL)
2910                         continue;
2911                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
2912                         dout("resending session close request for mds%d\n",
2913                              s->s_mds);
2914                         request_close_session(mdsc, s);
2915                         ceph_put_mds_session(s);
2916                         continue;
2917                 }
2918                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
2919                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
2920                                 s->s_state = CEPH_MDS_SESSION_HUNG;
2921                                 pr_info("mds%d hung\n", s->s_mds);
2922                         }
2923                 }
2924                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
2925                         /* this mds is failed or recovering, just wait */
2926                         ceph_put_mds_session(s);
2927                         continue;
2928                 }
2929                 mutex_unlock(&mdsc->mutex);
2930
2931                 mutex_lock(&s->s_mutex);
2932                 if (renew_caps)
2933                         send_renew_caps(mdsc, s);
2934                 else
2935                         ceph_con_keepalive(&s->s_con);
2936                 ceph_add_cap_releases(mdsc, s);
2937                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2938                     s->s_state == CEPH_MDS_SESSION_HUNG)
2939                         ceph_send_cap_releases(mdsc, s);
2940                 mutex_unlock(&s->s_mutex);
2941                 ceph_put_mds_session(s);
2942
2943                 mutex_lock(&mdsc->mutex);
2944         }
2945         mutex_unlock(&mdsc->mutex);
2946
2947         schedule_delayed(mdsc);
2948 }
2949
2950 int ceph_mdsc_init(struct ceph_fs_client *fsc)
2951
2952 {
2953         struct ceph_mds_client *mdsc;
2954
2955         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
2956         if (!mdsc)
2957                 return -ENOMEM;
2958         mdsc->fsc = fsc;
2959         fsc->mdsc = mdsc;
2960         mutex_init(&mdsc->mutex);
2961         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
2962         if (mdsc->mdsmap == NULL)
2963                 return -ENOMEM;
2964
2965         init_completion(&mdsc->safe_umount_waiters);
2966         init_waitqueue_head(&mdsc->session_close_wq);
2967         INIT_LIST_HEAD(&mdsc->waiting_for_map);
2968         mdsc->sessions = NULL;
2969         mdsc->max_sessions = 0;
2970         mdsc->stopping = 0;
2971         init_rwsem(&mdsc->snap_rwsem);
2972         mdsc->snap_realms = RB_ROOT;
2973         INIT_LIST_HEAD(&mdsc->snap_empty);
2974         spin_lock_init(&mdsc->snap_empty_lock);
2975         mdsc->last_tid = 0;
2976         mdsc->request_tree = RB_ROOT;
2977         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
2978         mdsc->last_renew_caps = jiffies;
2979         INIT_LIST_HEAD(&mdsc->cap_delay_list);
2980         spin_lock_init(&mdsc->cap_delay_lock);
2981         INIT_LIST_HEAD(&mdsc->snap_flush_list);
2982         spin_lock_init(&mdsc->snap_flush_lock);
2983         mdsc->cap_flush_seq = 0;
2984         INIT_LIST_HEAD(&mdsc->cap_dirty);
2985         mdsc->num_cap_flushing = 0;
2986         spin_lock_init(&mdsc->cap_dirty_lock);
2987         init_waitqueue_head(&mdsc->cap_flushing_wq);
2988         spin_lock_init(&mdsc->dentry_lru_lock);
2989         INIT_LIST_HEAD(&mdsc->dentry_lru);
2990
2991         ceph_caps_init(mdsc);
2992         ceph_adjust_min_caps(mdsc, fsc->min_caps);
2993
2994         return 0;
2995 }
2996
2997 /*
2998  * Wait for safe replies on open mds requests.  If we time out, drop
2999  * all requests from the tree to avoid dangling dentry refs.
3000  */
3001 static void wait_requests(struct ceph_mds_client *mdsc)
3002 {
3003         struct ceph_mds_request *req;
3004         struct ceph_fs_client *fsc = mdsc->fsc;
3005
3006         mutex_lock(&mdsc->mutex);
3007         if (__get_oldest_req(mdsc)) {
3008                 mutex_unlock(&mdsc->mutex);
3009
3010                 dout("wait_requests waiting for requests\n");
3011                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3012                                     fsc->client->options->mount_timeout * HZ);
3013
3014                 /* tear down remaining requests */
3015                 mutex_lock(&mdsc->mutex);
3016                 while ((req = __get_oldest_req(mdsc))) {
3017                         dout("wait_requests timed out on tid %llu\n",
3018                              req->r_tid);
3019                         __unregister_request(mdsc, req);
3020                 }
3021         }
3022         mutex_unlock(&mdsc->mutex);
3023         dout("wait_requests done\n");
3024 }
3025
3026 /*
3027  * called before mount is ro, and before dentries are torn down.
3028  * (hmm, does this still race with new lookups?)
3029  */
3030 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3031 {
3032         dout("pre_umount\n");
3033         mdsc->stopping = 1;
3034
3035         drop_leases(mdsc);
3036         ceph_flush_dirty_caps(mdsc);
3037         wait_requests(mdsc);
3038
3039         /*
3040          * wait for reply handlers to drop their request refs and
3041          * their inode/dcache refs
3042          */
3043         ceph_msgr_flush();
3044 }
3045
3046 /*
3047  * wait for all write mds requests to flush.
3048  */
3049 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3050 {
3051         struct ceph_mds_request *req = NULL, *nextreq;
3052         struct rb_node *n;
3053
3054         mutex_lock(&mdsc->mutex);
3055         dout("wait_unsafe_requests want %lld\n", want_tid);
3056 restart:
3057         req = __get_oldest_req(mdsc);
3058         while (req && req->r_tid <= want_tid) {
3059                 /* find next request */
3060                 n = rb_next(&req->r_node);
3061                 if (n)
3062                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3063                 else
3064                         nextreq = NULL;
3065                 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
3066                         /* write op */
3067                         ceph_mdsc_get_request(req);
3068                         if (nextreq)
3069                                 ceph_mdsc_get_request(nextreq);
3070                         mutex_unlock(&mdsc->mutex);
3071                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
3072                              req->r_tid, want_tid);
3073                         wait_for_completion(&req->r_safe_completion);
3074                         mutex_lock(&mdsc->mutex);
3075                         ceph_mdsc_put_request(req);
3076                         if (!nextreq)
3077                                 break;  /* next dne before, so we're done! */
3078                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
3079                                 /* next request was removed from tree */
3080                                 ceph_mdsc_put_request(nextreq);
3081                                 goto restart;
3082                         }
3083                         ceph_mdsc_put_request(nextreq);  /* won't go away */
3084                 }
3085                 req = nextreq;
3086         }
3087         mutex_unlock(&mdsc->mutex);
3088         dout("wait_unsafe_requests done\n");
3089 }
3090
3091 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3092 {
3093         u64 want_tid, want_flush;
3094
3095         if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3096                 return;
3097
3098         dout("sync\n");
3099         mutex_lock(&mdsc->mutex);
3100         want_tid = mdsc->last_tid;
3101         want_flush = mdsc->cap_flush_seq;
3102         mutex_unlock(&mdsc->mutex);
3103         dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
3104
3105         ceph_flush_dirty_caps(mdsc);
3106
3107         wait_unsafe_requests(mdsc, want_tid);
3108         wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
3109 }
3110
3111 /*
3112  * true if all sessions are closed, or we force unmount
3113  */
3114 bool done_closing_sessions(struct ceph_mds_client *mdsc)
3115 {
3116         int i, n = 0;
3117
3118         if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3119                 return true;
3120
3121         mutex_lock(&mdsc->mutex);
3122         for (i = 0; i < mdsc->max_sessions; i++)
3123                 if (mdsc->sessions[i])
3124                         n++;
3125         mutex_unlock(&mdsc->mutex);
3126         return n == 0;
3127 }
3128
3129 /*
3130  * called after sb is ro.
3131  */
3132 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3133 {
3134         struct ceph_mds_session *session;
3135         int i;
3136         struct ceph_fs_client *fsc = mdsc->fsc;
3137         unsigned long timeout = fsc->client->options->mount_timeout * HZ;
3138
3139         dout("close_sessions\n");
3140
3141         /* close sessions */
3142         mutex_lock(&mdsc->mutex);
3143         for (i = 0; i < mdsc->max_sessions; i++) {
3144                 session = __ceph_lookup_mds_session(mdsc, i);
3145                 if (!session)
3146                         continue;
3147                 mutex_unlock(&mdsc->mutex);
3148                 mutex_lock(&session->s_mutex);
3149                 __close_session(mdsc, session);
3150                 mutex_unlock(&session->s_mutex);
3151                 ceph_put_mds_session(session);
3152                 mutex_lock(&mdsc->mutex);
3153         }
3154         mutex_unlock(&mdsc->mutex);
3155
3156         dout("waiting for sessions to close\n");
3157         wait_event_timeout(mdsc->session_close_wq, done_closing_sessions(mdsc),
3158                            timeout);
3159
3160         /* tear down remaining sessions */
3161         mutex_lock(&mdsc->mutex);
3162         for (i = 0; i < mdsc->max_sessions; i++) {
3163                 if (mdsc->sessions[i]) {
3164                         session = get_session(mdsc->sessions[i]);
3165                         __unregister_session(mdsc, session);
3166                         mutex_unlock(&mdsc->mutex);
3167                         mutex_lock(&session->s_mutex);
3168                         remove_session_caps(session);
3169                         mutex_unlock(&session->s_mutex);
3170                         ceph_put_mds_session(session);
3171                         mutex_lock(&mdsc->mutex);
3172                 }
3173         }
3174         WARN_ON(!list_empty(&mdsc->cap_delay_list));
3175         mutex_unlock(&mdsc->mutex);
3176
3177         ceph_cleanup_empty_realms(mdsc);
3178
3179         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3180
3181         dout("stopped\n");
3182 }
3183
3184 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3185 {
3186         dout("stop\n");
3187         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3188         if (mdsc->mdsmap)
3189                 ceph_mdsmap_destroy(mdsc->mdsmap);
3190         kfree(mdsc->sessions);
3191         ceph_caps_finalize(mdsc);
3192 }
3193
3194 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3195 {
3196         struct ceph_mds_client *mdsc = fsc->mdsc;
3197
3198         ceph_mdsc_stop(mdsc);
3199         fsc->mdsc = NULL;
3200         kfree(mdsc);
3201 }
3202
3203
3204 /*
3205  * handle mds map update.
3206  */
3207 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3208 {
3209         u32 epoch;
3210         u32 maplen;
3211         void *p = msg->front.iov_base;
3212         void *end = p + msg->front.iov_len;
3213         struct ceph_mdsmap *newmap, *oldmap;
3214         struct ceph_fsid fsid;
3215         int err = -EINVAL;
3216
3217         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3218         ceph_decode_copy(&p, &fsid, sizeof(fsid));
3219         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3220                 return;
3221         epoch = ceph_decode_32(&p);
3222         maplen = ceph_decode_32(&p);
3223         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3224
3225         /* do we need it? */
3226         ceph_monc_got_mdsmap(&mdsc->fsc->client->monc, epoch);
3227         mutex_lock(&mdsc->mutex);
3228         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3229                 dout("handle_map epoch %u <= our %u\n",
3230                      epoch, mdsc->mdsmap->m_epoch);
3231                 mutex_unlock(&mdsc->mutex);
3232                 return;
3233         }
3234
3235         newmap = ceph_mdsmap_decode(&p, end);
3236         if (IS_ERR(newmap)) {
3237                 err = PTR_ERR(newmap);
3238                 goto bad_unlock;
3239         }
3240
3241         /* swap into place */
3242         if (mdsc->mdsmap) {
3243                 oldmap = mdsc->mdsmap;
3244                 mdsc->mdsmap = newmap;
3245                 check_new_map(mdsc, newmap, oldmap);
3246                 ceph_mdsmap_destroy(oldmap);
3247         } else {
3248                 mdsc->mdsmap = newmap;  /* first mds map */
3249         }
3250         mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3251
3252         __wake_requests(mdsc, &mdsc->waiting_for_map);
3253
3254         mutex_unlock(&mdsc->mutex);
3255         schedule_delayed(mdsc);
3256         return;
3257
3258 bad_unlock:
3259         mutex_unlock(&mdsc->mutex);
3260 bad:
3261         pr_err("error decoding mdsmap %d\n", err);
3262         return;
3263 }
3264
3265 static struct ceph_connection *con_get(struct ceph_connection *con)
3266 {
3267         struct ceph_mds_session *s = con->private;
3268
3269         if (get_session(s)) {
3270                 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3271                 return con;
3272         }
3273         dout("mdsc con_get %p FAIL\n", s);
3274         return NULL;
3275 }
3276
3277 static void con_put(struct ceph_connection *con)
3278 {
3279         struct ceph_mds_session *s = con->private;
3280
3281         ceph_put_mds_session(s);
3282         dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref));
3283 }
3284
3285 /*
3286  * if the client is unresponsive for long enough, the mds will kill
3287  * the session entirely.
3288  */
3289 static void peer_reset(struct ceph_connection *con)
3290 {
3291         struct ceph_mds_session *s = con->private;
3292         struct ceph_mds_client *mdsc = s->s_mdsc;
3293
3294         pr_warning("mds%d closed our session\n", s->s_mds);
3295         send_mds_reconnect(mdsc, s);
3296 }
3297
3298 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3299 {
3300         struct ceph_mds_session *s = con->private;
3301         struct ceph_mds_client *mdsc = s->s_mdsc;
3302         int type = le16_to_cpu(msg->hdr.type);
3303
3304         mutex_lock(&mdsc->mutex);
3305         if (__verify_registered_session(mdsc, s) < 0) {
3306                 mutex_unlock(&mdsc->mutex);
3307                 goto out;
3308         }
3309         mutex_unlock(&mdsc->mutex);
3310
3311         switch (type) {
3312         case CEPH_MSG_MDS_MAP:
3313                 ceph_mdsc_handle_map(mdsc, msg);
3314                 break;
3315         case CEPH_MSG_CLIENT_SESSION:
3316                 handle_session(s, msg);
3317                 break;
3318         case CEPH_MSG_CLIENT_REPLY:
3319                 handle_reply(s, msg);
3320                 break;
3321         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3322                 handle_forward(mdsc, s, msg);
3323                 break;
3324         case CEPH_MSG_CLIENT_CAPS:
3325                 ceph_handle_caps(s, msg);
3326                 break;
3327         case CEPH_MSG_CLIENT_SNAP:
3328                 ceph_handle_snap(mdsc, s, msg);
3329                 break;
3330         case CEPH_MSG_CLIENT_LEASE:
3331                 handle_lease(mdsc, s, msg);
3332                 break;
3333
3334         default:
3335                 pr_err("received unknown message type %d %s\n", type,
3336                        ceph_msg_type_name(type));
3337         }
3338 out:
3339         ceph_msg_put(msg);
3340 }
3341
3342 /*
3343  * authentication
3344  */
3345 static int get_authorizer(struct ceph_connection *con,
3346                           void **buf, int *len, int *proto,
3347                           void **reply_buf, int *reply_len, int force_new)
3348 {
3349         struct ceph_mds_session *s = con->private;
3350         struct ceph_mds_client *mdsc = s->s_mdsc;
3351         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3352         int ret = 0;
3353
3354         if (force_new && s->s_authorizer) {
3355                 ac->ops->destroy_authorizer(ac, s->s_authorizer);
3356                 s->s_authorizer = NULL;
3357         }
3358         if (s->s_authorizer == NULL) {
3359                 if (ac->ops->create_authorizer) {
3360                         ret = ac->ops->create_authorizer(
3361                                 ac, CEPH_ENTITY_TYPE_MDS,
3362                                 &s->s_authorizer,
3363                                 &s->s_authorizer_buf,
3364                                 &s->s_authorizer_buf_len,
3365                                 &s->s_authorizer_reply_buf,
3366                                 &s->s_authorizer_reply_buf_len);
3367                         if (ret)
3368                                 return ret;
3369                 }
3370         }
3371
3372         *proto = ac->protocol;
3373         *buf = s->s_authorizer_buf;
3374         *len = s->s_authorizer_buf_len;
3375         *reply_buf = s->s_authorizer_reply_buf;
3376         *reply_len = s->s_authorizer_reply_buf_len;
3377         return 0;
3378 }
3379
3380
3381 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3382 {
3383         struct ceph_mds_session *s = con->private;
3384         struct ceph_mds_client *mdsc = s->s_mdsc;
3385         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3386
3387         return ac->ops->verify_authorizer_reply(ac, s->s_authorizer, len);
3388 }
3389
3390 static int invalidate_authorizer(struct ceph_connection *con)
3391 {
3392         struct ceph_mds_session *s = con->private;
3393         struct ceph_mds_client *mdsc = s->s_mdsc;
3394         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3395
3396         if (ac->ops->invalidate_authorizer)
3397                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3398
3399         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
3400 }
3401
3402 static const struct ceph_connection_operations mds_con_ops = {
3403         .get = con_get,
3404         .put = con_put,
3405         .dispatch = dispatch,
3406         .get_authorizer = get_authorizer,
3407         .verify_authorizer_reply = verify_authorizer_reply,
3408         .invalidate_authorizer = invalidate_authorizer,
3409         .peer_reset = peer_reset,
3410 };
3411
3412 /* eof */