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  41  * are no longer valid.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