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  41  * are no longer valid.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                                int features)
  65 {
  66         int err = -EIO;
  67
  68         info->in = *p;
  69         *p += sizeof(struct ceph_mds_reply_inode) +
  70                 sizeof(*info->in->fragtree.splits) *
  71                 le32_to_cpu(info->in->fragtree.nsplits);
  72
  73         ceph_decode_32_safe(p, end, info->symlink_len, bad);
  74         ceph_decode_need(p, end, info->symlink_len, bad);
  75         info->symlink = *p;
  76         *p += info->symlink_len;
  77
  78         if (features & CEPH_FEATURE_DIRLAYOUTHASH)
  79                 ceph_decode_copy_safe(p, end, &info->dir_layout,
  80                                       sizeof(info->dir_layout), bad);
  81         else
  82                 memset(&info->dir_layout, 0, sizeof(info->dir_layout));
  83
  84         ceph_decode_32_safe(p, end, info->xattr_len, bad);
  85         ceph_decode_need(p, end, info->xattr_len, bad);
  86         info->xattr_data = *p;
  87         *p += info->xattr_len;
  88         return 0;
  89 bad:
  90         return err;
  91 }
  92
  93 /*
  94  * parse a normal reply, which may contain a (dir+)dentry and/or a
  95  * target inode.
  96  */
  97 static int parse_reply_info_trace(void **p, void *end,
  98                                   struct ceph_mds_reply_info_parsed *info,
  99                                   int features)
 100 {
 101         int err;
 102
 103         if (info->head->is_dentry) {
 104                 err = parse_reply_info_in(p, end, &info->diri, features);
 105                 if (err < 0)
 106                         goto out_bad;
 107
 108                 if (unlikely(*p + sizeof(*info->dirfrag) > end))
 109                         goto bad;
 110                 info->dirfrag = *p;
 111                 *p += sizeof(*info->dirfrag) +
 112                         sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
 113                 if (unlikely(*p > end))
 114                         goto bad;
 115
 116                 ceph_decode_32_safe(p, end, info->dname_len, bad);
 117                 ceph_decode_need(p, end, info->dname_len, bad);
 118                 info->dname = *p;
 119                 *p += info->dname_len;
 120                 info->dlease = *p;
 121                 *p += sizeof(*info->dlease);
 122         }
 123
 124         if (info->head->is_target) {
 125                 err = parse_reply_info_in(p, end, &info->targeti, features);
 126                 if (err < 0)
 127                         goto out_bad;
 128         }
 129
 130         if (unlikely(*p != end))
 131                 goto bad;
 132         return 0;
 133
 134 bad:
 135         err = -EIO;
 136 out_bad:
 137         pr_err("problem parsing mds trace %d\n", err);
 138         return err;
 139 }
 140
 141 /*
 142  * parse readdir results
 143  */
 144 static int parse_reply_info_dir(void **p, void *end,
 145                                 struct ceph_mds_reply_info_parsed *info,
 146                                 int features)
 147 {
 148         u32 num, i = 0;
 149         int err;
 150
 151         info->dir_dir = *p;
 152         if (*p + sizeof(*info->dir_dir) > end)
 153                 goto bad;
 154         *p += sizeof(*info->dir_dir) +
 155                 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
 156         if (*p > end)
 157                 goto bad;
 158
 159         ceph_decode_need(p, end, sizeof(num) + 2, bad);
 160         num = ceph_decode_32(p);
 161         info->dir_end = ceph_decode_8(p);
 162         info->dir_complete = ceph_decode_8(p);
 163         if (num == 0)
 164                 goto done;
 165
 166         /* alloc large array */
 167         info->dir_nr = num;
 168         info->dir_in = kcalloc(num, sizeof(*info->dir_in) +
 169                                sizeof(*info->dir_dname) +
 170                                sizeof(*info->dir_dname_len) +
 171                                sizeof(*info->dir_dlease),
 172                                GFP_NOFS);
 173         if (info->dir_in == NULL) {
 174                 err = -ENOMEM;
 175                 goto out_bad;
 176         }
 177         info->dir_dname = (void *)(info->dir_in + num);
 178         info->dir_dname_len = (void *)(info->dir_dname + num);
 179         info->dir_dlease = (void *)(info->dir_dname_len + num);
 180
 181         while (num) {
 182                 /* dentry */
 183                 ceph_decode_need(p, end, sizeof(u32)*2, bad);
 184                 info->dir_dname_len[i] = ceph_decode_32(p);
 185                 ceph_decode_need(p, end, info->dir_dname_len[i], bad);
 186                 info->dir_dname[i] = *p;
 187                 *p += info->dir_dname_len[i];
 188                 dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
 189                      info->dir_dname[i]);
 190                 info->dir_dlease[i] = *p;
 191                 *p += sizeof(struct ceph_mds_reply_lease);
 192
 193                 /* inode */
 194                 err = parse_reply_info_in(p, end, &info->dir_in[i], features);
 195                 if (err < 0)
 196                         goto out_bad;
 197                 i++;
 198                 num--;
 199         }
 200
 201 done:
 202         if (*p != end)
 203                 goto bad;
 204         return 0;
 205
 206 bad:
 207         err = -EIO;
 208 out_bad:
 209         pr_err("problem parsing dir contents %d\n", err);
 210         return err;
 211 }
 212
 213 /*
 214  * parse fcntl F_GETLK results
 215  */
 216 static int parse_reply_info_filelock(void **p, void *end,
 217                                      struct ceph_mds_reply_info_parsed *info,
 218                                      int features)
 219 {
 220         if (*p + sizeof(*info->filelock_reply) > end)
 221                 goto bad;
 222
 223         info->filelock_reply = *p;
 224         *p += sizeof(*info->filelock_reply);
 225
 226         if (unlikely(*p != end))
 227                 goto bad;
 228         return 0;
 229
 230 bad:
 231         return -EIO;
 232 }
 233
 234 /*
 235  * parse extra results
 236  */
 237 static int parse_reply_info_extra(void **p, void *end,
 238                                   struct ceph_mds_reply_info_parsed *info,
 239                                   int features)
 240 {
 241         if (info->head->op == CEPH_MDS_OP_GETFILELOCK)
 242                 return parse_reply_info_filelock(p, end, info, features);
 243         else
 244                 return parse_reply_info_dir(p, end, info, features);
 245 }
 246
 247 /*
 248  * parse entire mds reply
 249  */
 250 static int parse_reply_info(struct ceph_msg *msg,
 251                             struct ceph_mds_reply_info_parsed *info,
 252                             int features)
 253 {
 254         void *p, *end;
 255         u32 len;
 256         int err;
 257
 258         info->head = msg->front.iov_base;
 259         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 260         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 261
 262         /* trace */
 263         ceph_decode_32_safe(&p, end, len, bad);
 264         if (len > 0) {
 265                 err = parse_reply_info_trace(&p, p+len, info, features);
 266                 if (err < 0)
 267                         goto out_bad;
 268         }
 269
 270         /* extra */
 271         ceph_decode_32_safe(&p, end, len, bad);
 272         if (len > 0) {
 273                 err = parse_reply_info_extra(&p, p+len, info, features);
 274                 if (err < 0)
 275                         goto out_bad;
 276         }
 277
 278         /* snap blob */
 279         ceph_decode_32_safe(&p, end, len, bad);
 280         info->snapblob_len = len;
 281         info->snapblob = p;
 282         p += len;
 283
 284         if (p != end)
 285                 goto bad;
 286         return 0;
 287
 288 bad:
 289         err = -EIO;
 290 out_bad:
 291         pr_err("mds parse_reply err %d\n", err);
 292         return err;
 293 }
 294
 295 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
 296 {
 297         kfree(info->dir_in);
 298 }
 299
 300
 301 /*
 302  * sessions
 303  */
 304 static const char *session_state_name(int s)
 305 {
 306         switch (s) {
 307         case CEPH_MDS_SESSION_NEW: return "new";
 308         case CEPH_MDS_SESSION_OPENING: return "opening";
 309         case CEPH_MDS_SESSION_OPEN: return "open";
 310         case CEPH_MDS_SESSION_HUNG: return "hung";
 311         case CEPH_MDS_SESSION_CLOSING: return "closing";
 312         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
 313         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
 314         default: return "???";
 315         }
 316 }
 317
 318 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
 319 {
 320         if (atomic_inc_not_zero(&s->s_ref)) {
 321                 dout("mdsc get_session %p %d -> %d\n", s,
 322                      atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
 323                 return s;
 324         } else {
 325                 dout("mdsc get_session %p 0 -- FAIL", s);
 326                 return NULL;
 327         }
 328 }
 329
 330 void ceph_put_mds_session(struct ceph_mds_session *s)
 331 {
 332         dout("mdsc put_session %p %d -> %d\n", s,
 333              atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
 334         if (atomic_dec_and_test(&s->s_ref)) {
 335                 if (s->s_authorizer)
 336                      s->s_mdsc->fsc->client->monc.auth->ops->destroy_authorizer(
 337                              s->s_mdsc->fsc->client->monc.auth,
 338                              s->s_authorizer);
 339                 kfree(s);
 340         }
 341 }
 342
 343 /*
 344  * called under mdsc->mutex
 345  */
 346 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
 347                                                    int mds)
 348 {
 349         struct ceph_mds_session *session;
 350
 351         if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
 352                 return NULL;
 353         session = mdsc->sessions[mds];
 354         dout("lookup_mds_session %p %d\n", session,
 355              atomic_read(&session->s_ref));
 356         get_session(session);
 357         return session;
 358 }
 359
 360 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
 361 {
 362         if (mds >= mdsc->max_sessions)
 363                 return false;
 364         return mdsc->sessions[mds];
 365 }
 366
 367 static int __verify_registered_session(struct ceph_mds_client *mdsc,
 368                                        struct ceph_mds_session *s)
 369 {
 370         if (s->s_mds >= mdsc->max_sessions ||
 371             mdsc->sessions[s->s_mds] != s)
 372                 return -ENOENT;
 373         return 0;
 374 }
 375
 376 /*
 377  * create+register a new session for given mds.
 378  * called under mdsc->mutex.
 379  */
 380 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 381                                                  int mds)
 382 {
 383         struct ceph_mds_session *s;
 384
 385         s = kzalloc(sizeof(*s), GFP_NOFS);
 386         if (!s)
 387                 return ERR_PTR(-ENOMEM);
 388         s->s_mdsc = mdsc;
 389         s->s_mds = mds;
 390         s->s_state = CEPH_MDS_SESSION_NEW;
 391         s->s_ttl = 0;
 392         s->s_seq = 0;
 393         mutex_init(&s->s_mutex);
 394
 395         ceph_con_init(mdsc->fsc->client->msgr, &s->s_con);
 396         s->s_con.private = s;
 397         s->s_con.ops = &mds_con_ops;
 398         s->s_con.peer_name.type = CEPH_ENTITY_TYPE_MDS;
 399         s->s_con.peer_name.num = cpu_to_le64(mds);
 400
 401         spin_lock_init(&s->s_cap_lock);
 402         s->s_cap_gen = 0;
 403         s->s_cap_ttl = 0;
 404         s->s_renew_requested = 0;
 405         s->s_renew_seq = 0;
 406         INIT_LIST_HEAD(&s->s_caps);
 407         s->s_nr_caps = 0;
 408         s->s_trim_caps = 0;
 409         atomic_set(&s->s_ref, 1);
 410         INIT_LIST_HEAD(&s->s_waiting);
 411         INIT_LIST_HEAD(&s->s_unsafe);
 412         s->s_num_cap_releases = 0;
 413         s->s_cap_iterator = NULL;
 414         INIT_LIST_HEAD(&s->s_cap_releases);
 415         INIT_LIST_HEAD(&s->s_cap_releases_done);
 416         INIT_LIST_HEAD(&s->s_cap_flushing);
 417         INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
 418
 419         dout("register_session mds%d\n", mds);
 420         if (mds >= mdsc->max_sessions) {
 421                 int newmax = 1 << get_count_order(mds+1);
 422                 struct ceph_mds_session **sa;
 423
 424                 dout("register_session realloc to %d\n", newmax);
 425                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 426                 if (sa == NULL)
 427                         goto fail_realloc;
 428                 if (mdsc->sessions) {
 429                         memcpy(sa, mdsc->sessions,
 430                                mdsc->max_sessions * sizeof(void *));
 431                         kfree(mdsc->sessions);
 432                 }
 433                 mdsc->sessions = sa;
 434                 mdsc->max_sessions = newmax;
 435         }
 436         mdsc->sessions[mds] = s;
 437         atomic_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
 438
 439         ceph_con_open(&s->s_con, ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
 440
 441         return s;
 442
 443 fail_realloc:
 444         kfree(s);
 445         return ERR_PTR(-ENOMEM);
 446 }
 447
 448 /*
 449  * called under mdsc->mutex
 450  */
 451 static void __unregister_session(struct ceph_mds_client *mdsc,
 452                                struct ceph_mds_session *s)
 453 {
 454         dout("__unregister_session mds%d %p\n", s->s_mds, s);
 455         BUG_ON(mdsc->sessions[s->s_mds] != s);
 456         mdsc->sessions[s->s_mds] = NULL;
 457         ceph_con_close(&s->s_con);
 458         ceph_put_mds_session(s);
 459 }
 460
 461 /*
 462  * drop session refs in request.
 463  *
 464  * should be last request ref, or hold mdsc->mutex
 465  */
 466 static void put_request_session(struct ceph_mds_request *req)
 467 {
 468         if (req->r_session) {
 469                 ceph_put_mds_session(req->r_session);
 470                 req->r_session = NULL;
 471         }
 472 }
 473
 474 void ceph_mdsc_release_request(struct kref *kref)
 475 {
 476         struct ceph_mds_request *req = container_of(kref,
 477                                                     struct ceph_mds_request,
 478                                                     r_kref);
 479         if (req->r_request)
 480                 ceph_msg_put(req->r_request);
 481         if (req->r_reply) {
 482                 ceph_msg_put(req->r_reply);
 483                 destroy_reply_info(&req->r_reply_info);
 484         }
 485         if (req->r_inode) {
 486                 ceph_put_cap_refs(ceph_inode(req->r_inode),
 487                                   CEPH_CAP_PIN);
 488                 iput(req->r_inode);
 489         }
 490         if (req->r_locked_dir)
 491                 ceph_put_cap_refs(ceph_inode(req->r_locked_dir),
 492                                   CEPH_CAP_PIN);
 493         if (req->r_target_inode)
 494                 iput(req->r_target_inode);
 495         if (req->r_dentry)
 496                 dput(req->r_dentry);
 497         if (req->r_old_dentry) {
 498                 ceph_put_cap_refs(
 499                         ceph_inode(req->r_old_dentry->d_parent->d_inode),
 500                         CEPH_CAP_PIN);
 501                 dput(req->r_old_dentry);
 502         }
 503         kfree(req->r_path1);
 504         kfree(req->r_path2);
 505         put_request_session(req);
 506         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
 507         kfree(req);
 508 }
 509
 510 /*
 511  * lookup session, bump ref if found.
 512  *
 513  * called under mdsc->mutex.
 514  */
 515 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
 516                                              u64 tid)
 517 {
 518         struct ceph_mds_request *req;
 519         struct rb_node *n = mdsc->request_tree.rb_node;
 520
 521         while (n) {
 522                 req = rb_entry(n, struct ceph_mds_request, r_node);
 523                 if (tid < req->r_tid)
 524                         n = n->rb_left;
 525                 else if (tid > req->r_tid)
 526                         n = n->rb_right;
 527                 else {
 528                         ceph_mdsc_get_request(req);
 529                         return req;
 530                 }
 531         }
 532         return NULL;
 533 }
 534
 535 static void __insert_request(struct ceph_mds_client *mdsc,
 536                              struct ceph_mds_request *new)
 537 {
 538         struct rb_node **p = &mdsc->request_tree.rb_node;
 539         struct rb_node *parent = NULL;
 540         struct ceph_mds_request *req = NULL;
 541
 542         while (*p) {
 543                 parent = *p;
 544                 req = rb_entry(parent, struct ceph_mds_request, r_node);
 545                 if (new->r_tid < req->r_tid)
 546                         p = &(*p)->rb_left;
 547                 else if (new->r_tid > req->r_tid)
 548                         p = &(*p)->rb_right;
 549                 else
 550                         BUG();
 551         }
 552
 553         rb_link_node(&new->r_node, parent, p);
 554         rb_insert_color(&new->r_node, &mdsc->request_tree);
 555 }
 556
 557 /*
 558  * Register an in-flight request, and assign a tid.  Link to directory
 559  * are modifying (if any).
 560  *
 561  * Called under mdsc->mutex.
 562  */
 563 static void __register_request(struct ceph_mds_client *mdsc,
 564                                struct ceph_mds_request *req,
 565                                struct inode *dir)
 566 {
 567         req->r_tid = ++mdsc->last_tid;
 568         if (req->r_num_caps)
 569                 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
 570                                   req->r_num_caps);
 571         dout("__register_request %p tid %lld\n", req, req->r_tid);
 572         ceph_mdsc_get_request(req);
 573         __insert_request(mdsc, req);
 574
 575         req->r_uid = current_fsuid();
 576         req->r_gid = current_fsgid();
 577
 578         if (dir) {
 579                 struct ceph_inode_info *ci = ceph_inode(dir);
 580
 581                 ihold(dir);
 582                 spin_lock(&ci->i_unsafe_lock);
 583                 req->r_unsafe_dir = dir;
 584                 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
 585                 spin_unlock(&ci->i_unsafe_lock);
 586         }
 587 }
 588
 589 static void __unregister_request(struct ceph_mds_client *mdsc,
 590                                  struct ceph_mds_request *req)
 591 {
 592         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 593         rb_erase(&req->r_node, &mdsc->request_tree);
 594         RB_CLEAR_NODE(&req->r_node);
 595
 596         if (req->r_unsafe_dir) {
 597                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 598
 599                 spin_lock(&ci->i_unsafe_lock);
 600                 list_del_init(&req->r_unsafe_dir_item);
 601                 spin_unlock(&ci->i_unsafe_lock);
 602
 603                 iput(req->r_unsafe_dir);
 604                 req->r_unsafe_dir = NULL;
 605         }
 606
 607         ceph_mdsc_put_request(req);
 608 }
 609
 610 /*
 611  * Choose mds to send request to next.  If there is a hint set in the
 612  * request (e.g., due to a prior forward hint from the mds), use that.
 613  * Otherwise, consult frag tree and/or caps to identify the
 614  * appropriate mds.  If all else fails, choose randomly.
 615  *
 616  * Called under mdsc->mutex.
 617  */
 618 struct dentry *get_nonsnap_parent(struct dentry *dentry)
 619 {
 620         while (!IS_ROOT(dentry) && ceph_snap(dentry->d_inode) != CEPH_NOSNAP)
 621                 dentry = dentry->d_parent;
 622         return dentry;
 623 }
 624
 625 static int __choose_mds(struct ceph_mds_client *mdsc,
 626                         struct ceph_mds_request *req)
 627 {
 628         struct inode *inode;
 629         struct ceph_inode_info *ci;
 630         struct ceph_cap *cap;
 631         int mode = req->r_direct_mode;
 632         int mds = -1;
 633         u32 hash = req->r_direct_hash;
 634         bool is_hash = req->r_direct_is_hash;
 635
 636         /*
 637          * is there a specific mds we should try?  ignore hint if we have
 638          * no session and the mds is not up (active or recovering).
 639          */
 640         if (req->r_resend_mds >= 0 &&
 641             (__have_session(mdsc, req->r_resend_mds) ||
 642              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 643                 dout("choose_mds using resend_mds mds%d\n",
 644                      req->r_resend_mds);
 645                 return req->r_resend_mds;
 646         }
 647
 648         if (mode == USE_RANDOM_MDS)
 649                 goto random;
 650
 651         inode = NULL;
 652         if (req->r_inode) {
 653                 inode = req->r_inode;
 654         } else if (req->r_dentry) {
 655                 struct inode *dir = req->r_dentry->d_parent->d_inode;
 656
 657                 if (dir->i_sb != mdsc->fsc->sb) {
 658                         /* not this fs! */
 659                         inode = req->r_dentry->d_inode;
 660                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
 661                         /* direct snapped/virtual snapdir requests
 662                          * based on parent dir inode */
 663                         struct dentry *dn =
 664                                 get_nonsnap_parent(req->r_dentry->d_parent);
 665                         inode = dn->d_inode;
 666                         dout("__choose_mds using nonsnap parent %p\n", inode);
 667                 } else if (req->r_dentry->d_inode) {
 668                         /* dentry target */
 669                         inode = req->r_dentry->d_inode;
 670                 } else {
 671                         /* dir + name */
 672                         inode = dir;
 673                         hash = ceph_dentry_hash(req->r_dentry);
 674                         is_hash = true;
 675                 }
 676         }
 677
 678         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 679              (int)hash, mode);
 680         if (!inode)
 681                 goto random;
 682         ci = ceph_inode(inode);
 683
 684         if (is_hash && S_ISDIR(inode->i_mode)) {
 685                 struct ceph_inode_frag frag;
 686                 int found;
 687
 688                 ceph_choose_frag(ci, hash, &frag, &found);
 689                 if (found) {
 690                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 691                                 u8 r;
 692
 693                                 /* choose a random replica */
 694                                 get_random_bytes(&r, 1);
 695                                 r %= frag.ndist;
 696                                 mds = frag.dist[r];
 697                                 dout("choose_mds %p %llx.%llx "
 698                                      "frag %u mds%d (%d/%d)\n",
 699                                      inode, ceph_vinop(inode),
 700                                      frag.frag, mds,
 701                                      (int)r, frag.ndist);
 702                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 703                                     CEPH_MDS_STATE_ACTIVE)
 704                                         return mds;
 705                         }
 706
 707                         /* since this file/dir wasn't known to be
 708                          * replicated, then we want to look for the
 709                          * authoritative mds. */
 710                         mode = USE_AUTH_MDS;
 711                         if (frag.mds >= 0) {
 712                                 /* choose auth mds */
 713                                 mds = frag.mds;
 714                                 dout("choose_mds %p %llx.%llx "
 715                                      "frag %u mds%d (auth)\n",
 716                                      inode, ceph_vinop(inode), frag.frag, mds);
 717                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 718                                     CEPH_MDS_STATE_ACTIVE)
 719                                         return mds;
 720                         }
 721                 }
 722         }
 723
 724         spin_lock(&inode->i_lock);
 725         cap = NULL;
 726         if (mode == USE_AUTH_MDS)
 727                 cap = ci->i_auth_cap;
 728         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
 729                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
 730         if (!cap) {
 731                 spin_unlock(&inode->i_lock);
 732                 goto random;
 733         }
 734         mds = cap->session->s_mds;
 735         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
 736              inode, ceph_vinop(inode), mds,
 737              cap == ci->i_auth_cap ? "auth " : "", cap);
 738         spin_unlock(&inode->i_lock);
 739         return mds;
 740
 741 random:
 742         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
 743         dout("choose_mds chose random mds%d\n", mds);
 744         return mds;
 745 }
 746
 747
 748 /*
 749  * session messages
 750  */
 751 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
 752 {
 753         struct ceph_msg *msg;
 754         struct ceph_mds_session_head *h;
 755
 756         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS);
 757         if (!msg) {
 758                 pr_err("create_session_msg ENOMEM creating msg\n");
 759                 return NULL;
 760         }
 761         h = msg->front.iov_base;
 762         h->op = cpu_to_le32(op);
 763         h->seq = cpu_to_le64(seq);
 764         return msg;
 765 }
 766
 767 /*
 768  * send session open request.
 769  *
 770  * called under mdsc->mutex
 771  */
 772 static int __open_session(struct ceph_mds_client *mdsc,
 773                           struct ceph_mds_session *session)
 774 {
 775         struct ceph_msg *msg;
 776         int mstate;
 777         int mds = session->s_mds;
 778
 779         /* wait for mds to go active? */
 780         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
 781         dout("open_session to mds%d (%s)\n", mds,
 782              ceph_mds_state_name(mstate));
 783         session->s_state = CEPH_MDS_SESSION_OPENING;
 784         session->s_renew_requested = jiffies;
 785
 786         /* send connect message */
 787         msg = create_session_msg(CEPH_SESSION_REQUEST_OPEN, session->s_seq);
 788         if (!msg)
 789                 return -ENOMEM;
 790         ceph_con_send(&session->s_con, msg);
 791         return 0;
 792 }
 793
 794 /*
 795  * open sessions for any export targets for the given mds
 796  *
 797  * called under mdsc->mutex
 798  */
 799 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
 800                                           struct ceph_mds_session *session)
 801 {
 802         struct ceph_mds_info *mi;
 803         struct ceph_mds_session *ts;
 804         int i, mds = session->s_mds;
 805         int target;
 806
 807         if (mds >= mdsc->mdsmap->m_max_mds)
 808                 return;
 809         mi = &mdsc->mdsmap->m_info[mds];
 810         dout("open_export_target_sessions for mds%d (%d targets)\n",
 811              session->s_mds, mi->num_export_targets);
 812
 813         for (i = 0; i < mi->num_export_targets; i++) {
 814                 target = mi->export_targets[i];
 815                 ts = __ceph_lookup_mds_session(mdsc, target);
 816                 if (!ts) {
 817                         ts = register_session(mdsc, target);
 818                         if (IS_ERR(ts))
 819                                 return;
 820                 }
 821                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
 822                     session->s_state == CEPH_MDS_SESSION_CLOSING)
 823                         __open_session(mdsc, session);
 824                 else
 825                         dout(" mds%d target mds%d %p is %s\n", session->s_mds,
 826                              i, ts, session_state_name(ts->s_state));
 827                 ceph_put_mds_session(ts);
 828         }
 829 }
 830
 831 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
 832                                            struct ceph_mds_session *session)
 833 {
 834         mutex_lock(&mdsc->mutex);
 835         __open_export_target_sessions(mdsc, session);
 836         mutex_unlock(&mdsc->mutex);
 837 }
 838
 839 /*
 840  * session caps
 841  */
 842
 843 /*
 844  * Free preallocated cap messages assigned to this session
 845  */
 846 static void cleanup_cap_releases(struct ceph_mds_session *session)
 847 {
 848         struct ceph_msg *msg;
 849
 850         spin_lock(&session->s_cap_lock);
 851         while (!list_empty(&session->s_cap_releases)) {
 852                 msg = list_first_entry(&session->s_cap_releases,
 853                                        struct ceph_msg, list_head);
 854                 list_del_init(&msg->list_head);
 855                 ceph_msg_put(msg);
 856         }
 857         while (!list_empty(&session->s_cap_releases_done)) {
 858                 msg = list_first_entry(&session->s_cap_releases_done,
 859                                        struct ceph_msg, list_head);
 860                 list_del_init(&msg->list_head);
 861                 ceph_msg_put(msg);
 862         }
 863         spin_unlock(&session->s_cap_lock);
 864 }
 865
 866 /*
 867  * Helper to safely iterate over all caps associated with a session, with
 868  * special care taken to handle a racing __ceph_remove_cap().
 869  *
 870  * Caller must hold session s_mutex.
 871  */
 872 static int iterate_session_caps(struct ceph_mds_session *session,
 873                                  int (*cb)(struct inode *, struct ceph_cap *,
 874                                             void *), void *arg)
 875 {
 876         struct list_head *p;
 877         struct ceph_cap *cap;
 878         struct inode *inode, *last_inode = NULL;
 879         struct ceph_cap *old_cap = NULL;
 880         int ret;
 881
 882         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
 883         spin_lock(&session->s_cap_lock);
 884         p = session->s_caps.next;
 885         while (p != &session->s_caps) {
 886                 cap = list_entry(p, struct ceph_cap, session_caps);
 887                 inode = igrab(&cap->ci->vfs_inode);
 888                 if (!inode) {
 889                         p = p->next;
 890                         continue;
 891                 }
 892                 session->s_cap_iterator = cap;
 893                 spin_unlock(&session->s_cap_lock);
 894
 895                 if (last_inode) {
 896                         iput(last_inode);
 897                         last_inode = NULL;
 898                 }
 899                 if (old_cap) {
 900                         ceph_put_cap(session->s_mdsc, old_cap);
 901                         old_cap = NULL;
 902                 }
 903
 904                 ret = cb(inode, cap, arg);
 905                 last_inode = inode;
 906
 907                 spin_lock(&session->s_cap_lock);
 908                 p = p->next;
 909                 if (cap->ci == NULL) {
 910                         dout("iterate_session_caps  finishing cap %p removal\n",
 911                              cap);
 912                         BUG_ON(cap->session != session);
 913                         list_del_init(&cap->session_caps);
 914                         session->s_nr_caps--;
 915                         cap->session = NULL;
 916                         old_cap = cap;  /* put_cap it w/o locks held */
 917                 }
 918                 if (ret < 0)
 919                         goto out;
 920         }
 921         ret = 0;
 922 out:
 923         session->s_cap_iterator = NULL;
 924         spin_unlock(&session->s_cap_lock);
 925
 926         if (last_inode)
 927                 iput(last_inode);
 928         if (old_cap)
 929                 ceph_put_cap(session->s_mdsc, old_cap);
 930
 931         return ret;
 932 }
 933
 934 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
 935                                   void *arg)
 936 {
 937         struct ceph_inode_info *ci = ceph_inode(inode);
 938         int drop = 0;
 939
 940         dout("removing cap %p, ci is %p, inode is %p\n",
 941              cap, ci, &ci->vfs_inode);
 942         spin_lock(&inode->i_lock);
 943         __ceph_remove_cap(cap);
 944         if (!__ceph_is_any_real_caps(ci)) {
 945                 struct ceph_mds_client *mdsc =
 946                         ceph_sb_to_client(inode->i_sb)->mdsc;
 947
 948                 spin_lock(&mdsc->cap_dirty_lock);
 949                 if (!list_empty(&ci->i_dirty_item)) {
 950                         pr_info(" dropping dirty %s state for %p %lld\n",
 951                                 ceph_cap_string(ci->i_dirty_caps),
 952                                 inode, ceph_ino(inode));
 953                         ci->i_dirty_caps = 0;
 954                         list_del_init(&ci->i_dirty_item);
 955                         drop = 1;
 956                 }
 957                 if (!list_empty(&ci->i_flushing_item)) {
 958                         pr_info(" dropping dirty+flushing %s state for %p %lld\n",
 959                                 ceph_cap_string(ci->i_flushing_caps),
 960                                 inode, ceph_ino(inode));
 961                         ci->i_flushing_caps = 0;
 962                         list_del_init(&ci->i_flushing_item);
 963                         mdsc->num_cap_flushing--;
 964                         drop = 1;
 965                 }
 966                 if (drop && ci->i_wrbuffer_ref) {
 967                         pr_info(" dropping dirty data for %p %lld\n",
 968                                 inode, ceph_ino(inode));
 969                         ci->i_wrbuffer_ref = 0;
 970                         ci->i_wrbuffer_ref_head = 0;
 971                         drop++;
 972                 }
 973                 spin_unlock(&mdsc->cap_dirty_lock);
 974         }
 975         spin_unlock(&inode->i_lock);
 976         while (drop--)
 977                 iput(inode);
 978         return 0;
 979 }
 980
 981 /*
 982  * caller must hold session s_mutex
 983  */
 984 static void remove_session_caps(struct ceph_mds_session *session)
 985 {
 986         dout("remove_session_caps on %p\n", session);
 987         iterate_session_caps(session, remove_session_caps_cb, NULL);
 988         BUG_ON(session->s_nr_caps > 0);
 989         BUG_ON(!list_empty(&session->s_cap_flushing));
 990         cleanup_cap_releases(session);
 991 }
 992
 993 /*
 994  * wake up any threads waiting on this session's caps.  if the cap is
 995  * old (didn't get renewed on the client reconnect), remove it now.
 996  *
 997  * caller must hold s_mutex.
 998  */
 999 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1000                               void *arg)
1001 {
1002         struct ceph_inode_info *ci = ceph_inode(inode);
1003
1004         wake_up_all(&ci->i_cap_wq);
1005         if (arg) {
1006                 spin_lock(&inode->i_lock);
1007                 ci->i_wanted_max_size = 0;
1008                 ci->i_requested_max_size = 0;
1009                 spin_unlock(&inode->i_lock);
1010         }
1011         return 0;
1012 }
1013
1014 static void wake_up_session_caps(struct ceph_mds_session *session,
1015                                  int reconnect)
1016 {
1017         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1018         iterate_session_caps(session, wake_up_session_cb,
1019                              (void *)(unsigned long)reconnect);
1020 }
1021
1022 /*
1023  * Send periodic message to MDS renewing all currently held caps.  The
1024  * ack will reset the expiration for all caps from this session.
1025  *
1026  * caller holds s_mutex
1027  */
1028 static int send_renew_caps(struct ceph_mds_client *mdsc,
1029                            struct ceph_mds_session *session)
1030 {
1031         struct ceph_msg *msg;
1032         int state;
1033
1034         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1035             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1036                 pr_info("mds%d caps stale\n", session->s_mds);
1037         session->s_renew_requested = jiffies;
1038
1039         /* do not try to renew caps until a recovering mds has reconnected
1040          * with its clients. */
1041         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1042         if (state < CEPH_MDS_STATE_RECONNECT) {
1043                 dout("send_renew_caps ignoring mds%d (%s)\n",
1044                      session->s_mds, ceph_mds_state_name(state));
1045                 return 0;
1046         }
1047
1048         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1049                 ceph_mds_state_name(state));
1050         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1051                                  ++session->s_renew_seq);
1052         if (!msg)
1053                 return -ENOMEM;
1054         ceph_con_send(&session->s_con, msg);
1055         return 0;
1056 }
1057
1058 /*
1059  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1060  *
1061  * Called under session->s_mutex
1062  */
1063 static void renewed_caps(struct ceph_mds_client *mdsc,
1064                          struct ceph_mds_session *session, int is_renew)
1065 {
1066         int was_stale;
1067         int wake = 0;
1068
1069         spin_lock(&session->s_cap_lock);
1070         was_stale = is_renew && (session->s_cap_ttl == 0 ||
1071                                  time_after_eq(jiffies, session->s_cap_ttl));
1072
1073         session->s_cap_ttl = session->s_renew_requested +
1074                 mdsc->mdsmap->m_session_timeout*HZ;
1075
1076         if (was_stale) {
1077                 if (time_before(jiffies, session->s_cap_ttl)) {
1078                         pr_info("mds%d caps renewed\n", session->s_mds);
1079                         wake = 1;
1080                 } else {
1081                         pr_info("mds%d caps still stale\n", session->s_mds);
1082                 }
1083         }
1084         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1085              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1086              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1087         spin_unlock(&session->s_cap_lock);
1088
1089         if (wake)
1090                 wake_up_session_caps(session, 0);
1091 }
1092
1093 /*
1094  * send a session close request
1095  */
1096 static int request_close_session(struct ceph_mds_client *mdsc,
1097                                  struct ceph_mds_session *session)
1098 {
1099         struct ceph_msg *msg;
1100
1101         dout("request_close_session mds%d state %s seq %lld\n",
1102              session->s_mds, session_state_name(session->s_state),
1103              session->s_seq);
1104         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1105         if (!msg)
1106                 return -ENOMEM;
1107         ceph_con_send(&session->s_con, msg);
1108         return 0;
1109 }
1110
1111 /*
1112  * Called with s_mutex held.
1113  */
1114 static int __close_session(struct ceph_mds_client *mdsc,
1115                          struct ceph_mds_session *session)
1116 {
1117         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1118                 return 0;
1119         session->s_state = CEPH_MDS_SESSION_CLOSING;
1120         return request_close_session(mdsc, session);
1121 }
1122
1123 /*
1124  * Trim old(er) caps.
1125  *
1126  * Because we can't cache an inode without one or more caps, we do
1127  * this indirectly: if a cap is unused, we prune its aliases, at which
1128  * point the inode will hopefully get dropped to.
1129  *
1130  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1131  * memory pressure from the MDS, though, so it needn't be perfect.
1132  */
1133 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1134 {
1135         struct ceph_mds_session *session = arg;
1136         struct ceph_inode_info *ci = ceph_inode(inode);
1137         int used, oissued, mine;
1138
1139         if (session->s_trim_caps <= 0)
1140                 return -1;
1141
1142         spin_lock(&inode->i_lock);
1143         mine = cap->issued | cap->implemented;
1144         used = __ceph_caps_used(ci);
1145         oissued = __ceph_caps_issued_other(ci, cap);
1146
1147         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
1148              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1149              ceph_cap_string(used));
1150         if (ci->i_dirty_caps)
1151                 goto out;   /* dirty caps */
1152         if ((used & ~oissued) & mine)
1153                 goto out;   /* we need these caps */
1154
1155         session->s_trim_caps--;
1156         if (oissued) {
1157                 /* we aren't the only cap.. just remove us */
1158                 __ceph_remove_cap(cap);
1159         } else {
1160                 /* try to drop referring dentries */
1161                 spin_unlock(&inode->i_lock);
1162                 d_prune_aliases(inode);
1163                 dout("trim_caps_cb %p cap %p  pruned, count now %d\n",
1164                      inode, cap, atomic_read(&inode->i_count));
1165                 return 0;
1166         }
1167
1168 out:
1169         spin_unlock(&inode->i_lock);
1170         return 0;
1171 }
1172
1173 /*
1174  * Trim session cap count down to some max number.
1175  */
1176 static int trim_caps(struct ceph_mds_client *mdsc,
1177                      struct ceph_mds_session *session,
1178                      int max_caps)
1179 {
1180         int trim_caps = session->s_nr_caps - max_caps;
1181
1182         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1183              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1184         if (trim_caps > 0) {
1185                 session->s_trim_caps = trim_caps;
1186                 iterate_session_caps(session, trim_caps_cb, session);
1187                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1188                      session->s_mds, session->s_nr_caps, max_caps,
1189                         trim_caps - session->s_trim_caps);
1190                 session->s_trim_caps = 0;
1191         }
1192         return 0;
1193 }
1194
1195 /*
1196  * Allocate cap_release messages.  If there is a partially full message
1197  * in the queue, try to allocate enough to cover it's remainder, so that
1198  * we can send it immediately.
1199  *
1200  * Called under s_mutex.
1201  */
1202 int ceph_add_cap_releases(struct ceph_mds_client *mdsc,
1203                           struct ceph_mds_session *session)
1204 {
1205         struct ceph_msg *msg, *partial = NULL;
1206         struct ceph_mds_cap_release *head;
1207         int err = -ENOMEM;
1208         int extra = mdsc->fsc->mount_options->cap_release_safety;
1209         int num;
1210
1211         dout("add_cap_releases %p mds%d extra %d\n", session, session->s_mds,
1212              extra);
1213
1214         spin_lock(&session->s_cap_lock);
1215
1216         if (!list_empty(&session->s_cap_releases)) {
1217                 msg = list_first_entry(&session->s_cap_releases,
1218                                        struct ceph_msg,
1219                                  list_head);
1220                 head = msg->front.iov_base;
1221                 num = le32_to_cpu(head->num);
1222                 if (num) {
1223                         dout(" partial %p with (%d/%d)\n", msg, num,
1224                              (int)CEPH_CAPS_PER_RELEASE);
1225                         extra += CEPH_CAPS_PER_RELEASE - num;
1226                         partial = msg;
1227                 }
1228         }
1229         while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1230                 spin_unlock(&session->s_cap_lock);
1231                 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
1232                                    GFP_NOFS);
1233                 if (!msg)
1234                         goto out_unlocked;
1235                 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1236                      (int)msg->front.iov_len);
1237                 head = msg->front.iov_base;
1238                 head->num = cpu_to_le32(0);
1239                 msg->front.iov_len = sizeof(*head);
1240                 spin_lock(&session->s_cap_lock);
1241                 list_add(&msg->list_head, &session->s_cap_releases);
1242                 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1243         }
1244
1245         if (partial) {
1246                 head = partial->front.iov_base;
1247                 num = le32_to_cpu(head->num);
1248                 dout(" queueing partial %p with %d/%d\n", partial, num,
1249                      (int)CEPH_CAPS_PER_RELEASE);
1250                 list_move_tail(&partial->list_head,
1251                                &session->s_cap_releases_done);
1252                 session->s_num_cap_releases -= CEPH_CAPS_PER_RELEASE - num;
1253         }
1254         err = 0;
1255         spin_unlock(&session->s_cap_lock);
1256 out_unlocked:
1257         return err;
1258 }
1259
1260 /*
1261  * flush all dirty inode data to disk.
1262  *
1263  * returns true if we've flushed through want_flush_seq
1264  */
1265 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1266 {
1267         int mds, ret = 1;
1268
1269         dout("check_cap_flush want %lld\n", want_flush_seq);
1270         mutex_lock(&mdsc->mutex);
1271         for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1272                 struct ceph_mds_session *session = mdsc->sessions[mds];
1273
1274                 if (!session)
1275                         continue;
1276                 get_session(session);
1277                 mutex_unlock(&mdsc->mutex);
1278
1279                 mutex_lock(&session->s_mutex);
1280                 if (!list_empty(&session->s_cap_flushing)) {
1281                         struct ceph_inode_info *ci =
1282                                 list_entry(session->s_cap_flushing.next,
1283                                            struct ceph_inode_info,
1284                                            i_flushing_item);
1285                         struct inode *inode = &ci->vfs_inode;
1286
1287                         spin_lock(&inode->i_lock);
1288                         if (ci->i_cap_flush_seq <= want_flush_seq) {
1289                                 dout("check_cap_flush still flushing %p "
1290                                      "seq %lld <= %lld to mds%d\n", inode,
1291                                      ci->i_cap_flush_seq, want_flush_seq,
1292                                      session->s_mds);
1293                                 ret = 0;
1294                         }
1295                         spin_unlock(&inode->i_lock);
1296                 }
1297                 mutex_unlock(&session->s_mutex);
1298                 ceph_put_mds_session(session);
1299
1300                 if (!ret)
1301                         return ret;
1302                 mutex_lock(&mdsc->mutex);
1303         }
1304
1305         mutex_unlock(&mdsc->mutex);
1306         dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1307         return ret;
1308 }
1309
1310 /*
1311  * called under s_mutex
1312  */
1313 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1314                             struct ceph_mds_session *session)
1315 {
1316         struct ceph_msg *msg;
1317
1318         dout("send_cap_releases mds%d\n", session->s_mds);
1319         spin_lock(&session->s_cap_lock);
1320         while (!list_empty(&session->s_cap_releases_done)) {
1321                 msg = list_first_entry(&session->s_cap_releases_done,
1322                                  struct ceph_msg, list_head);
1323                 list_del_init(&msg->list_head);
1324                 spin_unlock(&session->s_cap_lock);
1325                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1326                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1327                 ceph_con_send(&session->s_con, msg);
1328                 spin_lock(&session->s_cap_lock);
1329         }
1330         spin_unlock(&session->s_cap_lock);
1331 }
1332
1333 static void discard_cap_releases(struct ceph_mds_client *mdsc,
1334                                  struct ceph_mds_session *session)
1335 {
1336         struct ceph_msg *msg;
1337         struct ceph_mds_cap_release *head;
1338         unsigned num;
1339
1340         dout("discard_cap_releases mds%d\n", session->s_mds);
1341         spin_lock(&session->s_cap_lock);
1342
1343         /* zero out the in-progress message */
1344         msg = list_first_entry(&session->s_cap_releases,
1345                                struct ceph_msg, list_head);
1346         head = msg->front.iov_base;
1347         num = le32_to_cpu(head->num);
1348         dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg, num);
1349         head->num = cpu_to_le32(0);
1350         session->s_num_cap_releases += num;
1351
1352         /* requeue completed messages */
1353         while (!list_empty(&session->s_cap_releases_done)) {
1354                 msg = list_first_entry(&session->s_cap_releases_done,
1355                                  struct ceph_msg, list_head);
1356                 list_del_init(&msg->list_head);
1357
1358                 head = msg->front.iov_base;
1359                 num = le32_to_cpu(head->num);
1360                 dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg,
1361                      num);
1362                 session->s_num_cap_releases += num;
1363                 head->num = cpu_to_le32(0);
1364                 msg->front.iov_len = sizeof(*head);
1365                 list_add(&msg->list_head, &session->s_cap_releases);
1366         }
1367
1368         spin_unlock(&session->s_cap_lock);
1369 }
1370
1371 /*
1372  * requests
1373  */
1374
1375 /*
1376  * Create an mds request.
1377  */
1378 struct ceph_mds_request *
1379 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1380 {
1381         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1382
1383         if (!req)
1384                 return ERR_PTR(-ENOMEM);
1385
1386         mutex_init(&req->r_fill_mutex);
1387         req->r_mdsc = mdsc;
1388         req->r_started = jiffies;
1389         req->r_resend_mds = -1;
1390         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1391         req->r_fmode = -1;
1392         kref_init(&req->r_kref);
1393         INIT_LIST_HEAD(&req->r_wait);
1394         init_completion(&req->r_completion);
1395         init_completion(&req->r_safe_completion);
1396         INIT_LIST_HEAD(&req->r_unsafe_item);
1397
1398         req->r_op = op;
1399         req->r_direct_mode = mode;
1400         return req;
1401 }
1402
1403 /*
1404  * return oldest (lowest) request, tid in request tree, 0 if none.
1405  *
1406  * called under mdsc->mutex.
1407  */
1408 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1409 {
1410         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1411                 return NULL;
1412         return rb_entry(rb_first(&mdsc->request_tree),
1413                         struct ceph_mds_request, r_node);
1414 }
1415
1416 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1417 {
1418         struct ceph_mds_request *req = __get_oldest_req(mdsc);
1419
1420         if (req)
1421                 return req->r_tid;
1422         return 0;
1423 }
1424
1425 /*
1426  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1427  * on build_path_from_dentry in fs/cifs/dir.c.
1428  *
1429  * If @stop_on_nosnap, generate path relative to the first non-snapped
1430  * inode.
1431  *
1432  * Encode hidden .snap dirs as a double /, i.e.
1433  *   foo/.snap/bar -> foo//bar
1434  */
1435 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1436                            int stop_on_nosnap)
1437 {
1438         struct dentry *temp;
1439         char *path;
1440         int len, pos;
1441
1442         if (dentry == NULL)
1443                 return ERR_PTR(-EINVAL);
1444
1445 retry:
1446         len = 0;
1447         for (temp = dentry; !IS_ROOT(temp);) {
1448                 struct inode *inode = temp->d_inode;
1449                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1450                         len++;  /* slash only */
1451                 else if (stop_on_nosnap && inode &&
1452                          ceph_snap(inode) == CEPH_NOSNAP)
1453                         break;
1454                 else
1455                         len += 1 + temp->d_name.len;
1456                 temp = temp->d_parent;
1457                 if (temp == NULL) {
1458                         pr_err("build_path corrupt dentry %p\n", dentry);
1459                         return ERR_PTR(-EINVAL);
1460                 }
1461         }
1462         if (len)
1463                 len--;  /* no leading '/' */
1464
1465         path = kmalloc(len+1, GFP_NOFS);
1466         if (path == NULL)
1467                 return ERR_PTR(-ENOMEM);
1468         pos = len;
1469         path[pos] = 0;  /* trailing null */
1470         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1471                 struct inode *inode = temp->d_inode;
1472
1473                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1474                         dout("build_path path+%d: %p SNAPDIR\n",
1475                              pos, temp);
1476                 } else if (stop_on_nosnap && inode &&
1477                            ceph_snap(inode) == CEPH_NOSNAP) {
1478                         break;
1479                 } else {
1480                         pos -= temp->d_name.len;
1481                         if (pos < 0)
1482                                 break;
1483                         strncpy(path + pos, temp->d_name.name,
1484                                 temp->d_name.len);
1485                 }
1486                 if (pos)
1487                         path[--pos] = '/';
1488                 temp = temp->d_parent;
1489                 if (temp == NULL) {
1490                         pr_err("build_path corrupt dentry\n");
1491                         kfree(path);
1492                         return ERR_PTR(-EINVAL);
1493                 }
1494         }
1495         if (pos != 0) {
1496                 pr_err("build_path did not end path lookup where "
1497                        "expected, namelen is %d, pos is %d\n", len, pos);
1498                 /* presumably this is only possible if racing with a
1499                    rename of one of the parent directories (we can not
1500                    lock the dentries above us to prevent this, but
1501                    retrying should be harmless) */
1502                 kfree(path);
1503                 goto retry;
1504         }
1505
1506         *base = ceph_ino(temp->d_inode);
1507         *plen = len;
1508         dout("build_path on %p %d built %llx '%.*s'\n",
1509              dentry, dentry->d_count, *base, len, path);
1510         return path;
1511 }
1512
1513 static int build_dentry_path(struct dentry *dentry,
1514                              const char **ppath, int *ppathlen, u64 *pino,
1515                              int *pfreepath)
1516 {
1517         char *path;
1518
1519         if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1520                 *pino = ceph_ino(dentry->d_parent->d_inode);
1521                 *ppath = dentry->d_name.name;
1522                 *ppathlen = dentry->d_name.len;
1523                 return 0;
1524         }
1525         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1526         if (IS_ERR(path))
1527                 return PTR_ERR(path);
1528         *ppath = path;
1529         *pfreepath = 1;
1530         return 0;
1531 }
1532
1533 static int build_inode_path(struct inode *inode,
1534                             const char **ppath, int *ppathlen, u64 *pino,
1535                             int *pfreepath)
1536 {
1537         struct dentry *dentry;
1538         char *path;
1539
1540         if (ceph_snap(inode) == CEPH_NOSNAP) {
1541                 *pino = ceph_ino(inode);
1542                 *ppathlen = 0;
1543                 return 0;
1544         }
1545         dentry = d_find_alias(inode);
1546         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1547         dput(dentry);
1548         if (IS_ERR(path))
1549                 return PTR_ERR(path);
1550         *ppath = path;
1551         *pfreepath = 1;
1552         return 0;
1553 }
1554
1555 /*
1556  * request arguments may be specified via an inode *, a dentry *, or
1557  * an explicit ino+path.
1558  */
1559 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1560                                   const char *rpath, u64 rino,
1561                                   const char **ppath, int *pathlen,
1562                                   u64 *ino, int *freepath)
1563 {
1564         int r = 0;
1565
1566         if (rinode) {
1567                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1568                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1569                      ceph_snap(rinode));
1570         } else if (rdentry) {
1571                 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1572                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1573                      *ppath);
1574         } else if (rpath) {
1575                 *ino = rino;
1576                 *ppath = rpath;
1577                 *pathlen = strlen(rpath);
1578                 dout(" path %.*s\n", *pathlen, rpath);
1579         }
1580
1581         return r;
1582 }
1583
1584 /*
1585  * called under mdsc->mutex
1586  */
1587 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1588                                                struct ceph_mds_request *req,
1589                                                int mds)
1590 {
1591         struct ceph_msg *msg;
1592         struct ceph_mds_request_head *head;
1593         const char *path1 = NULL;
1594         const char *path2 = NULL;
1595         u64 ino1 = 0, ino2 = 0;
1596         int pathlen1 = 0, pathlen2 = 0;
1597         int freepath1 = 0, freepath2 = 0;
1598         int len;
1599         u16 releases;
1600         void *p, *end;
1601         int ret;
1602
1603         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1604                               req->r_path1, req->r_ino1.ino,
1605                               &path1, &pathlen1, &ino1, &freepath1);
1606         if (ret < 0) {
1607                 msg = ERR_PTR(ret);
1608                 goto out;
1609         }
1610
1611         ret = set_request_path_attr(NULL, req->r_old_dentry,
1612                               req->r_path2, req->r_ino2.ino,
1613                               &path2, &pathlen2, &ino2, &freepath2);
1614         if (ret < 0) {
1615                 msg = ERR_PTR(ret);
1616                 goto out_free1;
1617         }
1618
1619         len = sizeof(*head) +
1620                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64));
1621
1622         /* calculate (max) length for cap releases */
1623         len += sizeof(struct ceph_mds_request_release) *
1624                 (!!req->r_inode_drop + !!req->r_dentry_drop +
1625                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1626         if (req->r_dentry_drop)
1627                 len += req->r_dentry->d_name.len;
1628         if (req->r_old_dentry_drop)
1629                 len += req->r_old_dentry->d_name.len;
1630
1631         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS);
1632         if (!msg) {
1633                 msg = ERR_PTR(-ENOMEM);
1634                 goto out_free2;
1635         }
1636
1637         msg->hdr.tid = cpu_to_le64(req->r_tid);
1638
1639         head = msg->front.iov_base;
1640         p = msg->front.iov_base + sizeof(*head);
1641         end = msg->front.iov_base + msg->front.iov_len;
1642
1643         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1644         head->op = cpu_to_le32(req->r_op);
1645         head->caller_uid = cpu_to_le32(req->r_uid);
1646         head->caller_gid = cpu_to_le32(req->r_gid);
1647         head->args = req->r_args;
1648
1649         ceph_encode_filepath(&p, end, ino1, path1);
1650         ceph_encode_filepath(&p, end, ino2, path2);
1651
1652         /* make note of release offset, in case we need to replay */
1653         req->r_request_release_offset = p - msg->front.iov_base;
1654
1655         /* cap releases */
1656         releases = 0;
1657         if (req->r_inode_drop)
1658                 releases += ceph_encode_inode_release(&p,
1659                       req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1660                       mds, req->r_inode_drop, req->r_inode_unless, 0);
1661         if (req->r_dentry_drop)
1662                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1663                        mds, req->r_dentry_drop, req->r_dentry_unless);
1664         if (req->r_old_dentry_drop)
1665                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1666                        mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1667         if (req->r_old_inode_drop)
1668                 releases += ceph_encode_inode_release(&p,
1669                       req->r_old_dentry->d_inode,
1670                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1671         head->num_releases = cpu_to_le16(releases);
1672
1673         BUG_ON(p > end);
1674         msg->front.iov_len = p - msg->front.iov_base;
1675         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1676
1677         msg->pages = req->r_pages;
1678         msg->nr_pages = req->r_num_pages;
1679         msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1680         msg->hdr.data_off = cpu_to_le16(0);
1681
1682 out_free2:
1683         if (freepath2)
1684                 kfree((char *)path2);
1685 out_free1:
1686         if (freepath1)
1687                 kfree((char *)path1);
1688 out:
1689         return msg;
1690 }
1691
1692 /*
1693  * called under mdsc->mutex if error, under no mutex if
1694  * success.
1695  */
1696 static void complete_request(struct ceph_mds_client *mdsc,
1697                              struct ceph_mds_request *req)
1698 {
1699         if (req->r_callback)
1700                 req->r_callback(mdsc, req);
1701         else
1702                 complete_all(&req->r_completion);
1703 }
1704
1705 /*
1706  * called under mdsc->mutex
1707  */
1708 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1709                                   struct ceph_mds_request *req,
1710                                   int mds)
1711 {
1712         struct ceph_mds_request_head *rhead;
1713         struct ceph_msg *msg;
1714         int flags = 0;
1715
1716         req->r_attempts++;
1717         if (req->r_inode) {
1718                 struct ceph_cap *cap =
1719                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
1720
1721                 if (cap)
1722                         req->r_sent_on_mseq = cap->mseq;
1723                 else
1724                         req->r_sent_on_mseq = -1;
1725         }
1726         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1727              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1728
1729         if (req->r_got_unsafe) {
1730                 /*
1731                  * Replay.  Do not regenerate message (and rebuild
1732                  * paths, etc.); just use the original message.
1733                  * Rebuilding paths will break for renames because
1734                  * d_move mangles the src name.
1735                  */
1736                 msg = req->r_request;
1737                 rhead = msg->front.iov_base;
1738
1739                 flags = le32_to_cpu(rhead->flags);
1740                 flags |= CEPH_MDS_FLAG_REPLAY;
1741                 rhead->flags = cpu_to_le32(flags);
1742
1743                 if (req->r_target_inode)
1744                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1745
1746                 rhead->num_retry = req->r_attempts - 1;
1747
1748                 /* remove cap/dentry releases from message */
1749                 rhead->num_releases = 0;
1750                 msg->hdr.front_len = cpu_to_le32(req->r_request_release_offset);
1751                 msg->front.iov_len = req->r_request_release_offset;
1752                 return 0;
1753         }
1754
1755         if (req->r_request) {
1756                 ceph_msg_put(req->r_request);
1757                 req->r_request = NULL;
1758         }
1759         msg = create_request_message(mdsc, req, mds);
1760         if (IS_ERR(msg)) {
1761                 req->r_err = PTR_ERR(msg);
1762                 complete_request(mdsc, req);
1763                 return PTR_ERR(msg);
1764         }
1765         req->r_request = msg;
1766
1767         rhead = msg->front.iov_base;
1768         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1769         if (req->r_got_unsafe)
1770                 flags |= CEPH_MDS_FLAG_REPLAY;
1771         if (req->r_locked_dir)
1772                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1773         rhead->flags = cpu_to_le32(flags);
1774         rhead->num_fwd = req->r_num_fwd;
1775         rhead->num_retry = req->r_attempts - 1;
1776         rhead->ino = 0;
1777
1778         dout(" r_locked_dir = %p\n", req->r_locked_dir);
1779         return 0;
1780 }
1781
1782 /*
1783  * send request, or put it on the appropriate wait list.
1784  */
1785 static int __do_request(struct ceph_mds_client *mdsc,
1786                         struct ceph_mds_request *req)
1787 {
1788         struct ceph_mds_session *session = NULL;
1789         int mds = -1;
1790         int err = -EAGAIN;
1791
1792         if (req->r_err || req->r_got_result)
1793                 goto out;
1794
1795         if (req->r_timeout &&
1796             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1797                 dout("do_request timed out\n");
1798                 err = -EIO;
1799                 goto finish;
1800         }
1801
1802         put_request_session(req);
1803
1804         mds = __choose_mds(mdsc, req);
1805         if (mds < 0 ||
1806             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1807                 dout("do_request no mds or not active, waiting for map\n");
1808                 list_add(&req->r_wait, &mdsc->waiting_for_map);
1809                 goto out;
1810         }
1811
1812         /* get, open session */
1813         session = __ceph_lookup_mds_session(mdsc, mds);
1814         if (!session) {
1815                 session = register_session(mdsc, mds);
1816                 if (IS_ERR(session)) {
1817                         err = PTR_ERR(session);
1818                         goto finish;
1819                 }
1820         }
1821         req->r_session = get_session(session);
1822
1823         dout("do_request mds%d session %p state %s\n", mds, session,
1824              session_state_name(session->s_state));
1825         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1826             session->s_state != CEPH_MDS_SESSION_HUNG) {
1827                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1828                     session->s_state == CEPH_MDS_SESSION_CLOSING)
1829                         __open_session(mdsc, session);
1830                 list_add(&req->r_wait, &session->s_waiting);
1831                 goto out_session;
1832         }
1833
1834         /* send request */
1835         req->r_resend_mds = -1;   /* forget any previous mds hint */
1836
1837         if (req->r_request_started == 0)   /* note request start time */
1838                 req->r_request_started = jiffies;
1839
1840         err = __prepare_send_request(mdsc, req, mds);
1841         if (!err) {
1842                 ceph_msg_get(req->r_request);
1843                 ceph_con_send(&session->s_con, req->r_request);
1844         }
1845
1846 out_session:
1847         ceph_put_mds_session(session);
1848 out:
1849         return err;
1850
1851 finish:
1852         req->r_err = err;
1853         complete_request(mdsc, req);
1854         goto out;
1855 }
1856
1857 /*
1858  * called under mdsc->mutex
1859  */
1860 static void __wake_requests(struct ceph_mds_client *mdsc,
1861                             struct list_head *head)
1862 {
1863         struct ceph_mds_request *req, *nreq;
1864
1865         list_for_each_entry_safe(req, nreq, head, r_wait) {
1866                 list_del_init(&req->r_wait);
1867                 __do_request(mdsc, req);
1868         }
1869 }
1870
1871 /*
1872  * Wake up threads with requests pending for @mds, so that they can
1873  * resubmit their requests to a possibly different mds.
1874  */
1875 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
1876 {
1877         struct ceph_mds_request *req;
1878         struct rb_node *p;
1879
1880         dout("kick_requests mds%d\n", mds);
1881         for (p = rb_first(&mdsc->request_tree); p; p = rb_next(p)) {
1882                 req = rb_entry(p, struct ceph_mds_request, r_node);
1883                 if (req->r_got_unsafe)
1884                         continue;
1885                 if (req->r_session &&
1886                     req->r_session->s_mds == mds) {
1887                         dout(" kicking tid %llu\n", req->r_tid);
1888                         __do_request(mdsc, req);
1889                 }
1890         }
1891 }
1892
1893 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
1894                               struct ceph_mds_request *req)
1895 {
1896         dout("submit_request on %p\n", req);
1897         mutex_lock(&mdsc->mutex);
1898         __register_request(mdsc, req, NULL);
1899         __do_request(mdsc, req);
1900         mutex_unlock(&mdsc->mutex);
1901 }
1902
1903 /*
1904  * Synchrously perform an mds request.  Take care of all of the
1905  * session setup, forwarding, retry details.
1906  */
1907 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
1908                          struct inode *dir,
1909                          struct ceph_mds_request *req)
1910 {
1911         int err;
1912
1913         dout("do_request on %p\n", req);
1914
1915         /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
1916         if (req->r_inode)
1917                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
1918         if (req->r_locked_dir)
1919                 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
1920         if (req->r_old_dentry)
1921                 ceph_get_cap_refs(
1922                         ceph_inode(req->r_old_dentry->d_parent->d_inode),
1923                         CEPH_CAP_PIN);
1924
1925         /* issue */
1926         mutex_lock(&mdsc->mutex);
1927         __register_request(mdsc, req, dir);
1928         __do_request(mdsc, req);
1929
1930         if (req->r_err) {
1931                 err = req->r_err;
1932                 __unregister_request(mdsc, req);
1933                 dout("do_request early error %d\n", err);
1934                 goto out;
1935         }
1936
1937         /* wait */
1938         mutex_unlock(&mdsc->mutex);
1939         dout("do_request waiting\n");
1940         if (req->r_timeout) {
1941                 err = (long)wait_for_completion_killable_timeout(
1942                         &req->r_completion, req->r_timeout);
1943                 if (err == 0)
1944                         err = -EIO;
1945         } else {
1946                 err = wait_for_completion_killable(&req->r_completion);
1947         }
1948         dout("do_request waited, got %d\n", err);
1949         mutex_lock(&mdsc->mutex);
1950
1951         /* only abort if we didn't race with a real reply */
1952         if (req->r_got_result) {
1953                 err = le32_to_cpu(req->r_reply_info.head->result);
1954         } else if (err < 0) {
1955                 dout("aborted request %lld with %d\n", req->r_tid, err);
1956
1957                 /*
1958                  * ensure we aren't running concurrently with
1959                  * ceph_fill_trace or ceph_readdir_prepopulate, which
1960                  * rely on locks (dir mutex) held by our caller.
1961                  */
1962                 mutex_lock(&req->r_fill_mutex);
1963                 req->r_err = err;
1964                 req->r_aborted = true;
1965                 mutex_unlock(&req->r_fill_mutex);
1966
1967                 if (req->r_locked_dir &&
1968                     (req->r_op & CEPH_MDS_OP_WRITE))
1969                         ceph_invalidate_dir_request(req);
1970         } else {
1971                 err = req->r_err;
1972         }
1973
1974 out:
1975         mutex_unlock(&mdsc->mutex);
1976         dout("do_request %p done, result %d\n", req, err);
1977         return err;
1978 }
1979
1980 /*
1981  * Invalidate dir I_COMPLETE, dentry lease state on an aborted MDS
1982  * namespace request.
1983  */
1984 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
1985 {
1986         struct inode *inode = req->r_locked_dir;
1987         struct ceph_inode_info *ci = ceph_inode(inode);
1988
1989         dout("invalidate_dir_request %p (I_COMPLETE, lease(s))\n", inode);
1990         spin_lock(&inode->i_lock);
1991         ci->i_ceph_flags &= ~CEPH_I_COMPLETE;
1992         ci->i_release_count++;
1993         spin_unlock(&inode->i_lock);
1994
1995         if (req->r_dentry)
1996                 ceph_invalidate_dentry_lease(req->r_dentry);
1997         if (req->r_old_dentry)
1998                 ceph_invalidate_dentry_lease(req->r_old_dentry);
1999 }
2000
2001 /*
2002  * Handle mds reply.
2003  *
2004  * We take the session mutex and parse and process the reply immediately.
2005  * This preserves the logical ordering of replies, capabilities, etc., sent
2006  * by the MDS as they are applied to our local cache.
2007  */
2008 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2009 {
2010         struct ceph_mds_client *mdsc = session->s_mdsc;
2011         struct ceph_mds_request *req;
2012         struct ceph_mds_reply_head *head = msg->front.iov_base;
2013         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2014         u64 tid;
2015         int err, result;
2016         int mds = session->s_mds;
2017
2018         if (msg->front.iov_len < sizeof(*head)) {
2019                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2020                 ceph_msg_dump(msg);
2021                 return;
2022         }
2023
2024         /* get request, session */
2025         tid = le64_to_cpu(msg->hdr.tid);
2026         mutex_lock(&mdsc->mutex);
2027         req = __lookup_request(mdsc, tid);
2028         if (!req) {
2029                 dout("handle_reply on unknown tid %llu\n", tid);
2030                 mutex_unlock(&mdsc->mutex);
2031                 return;
2032         }
2033         dout("handle_reply %p\n", req);
2034
2035         /* correct session? */
2036         if (req->r_session != session) {
2037                 pr_err("mdsc_handle_reply got %llu on session mds%d"
2038                        " not mds%d\n", tid, session->s_mds,
2039                        req->r_session ? req->r_session->s_mds : -1);
2040                 mutex_unlock(&mdsc->mutex);
2041                 goto out;
2042         }
2043
2044         /* dup? */
2045         if ((req->r_got_unsafe && !head->safe) ||
2046             (req->r_got_safe && head->safe)) {
2047                 pr_warning("got a dup %s reply on %llu from mds%d\n",
2048                            head->safe ? "safe" : "unsafe", tid, mds);
2049                 mutex_unlock(&mdsc->mutex);
2050                 goto out;
2051         }
2052         if (req->r_got_safe && !head->safe) {
2053                 pr_warning("got unsafe after safe on %llu from mds%d\n",
2054                            tid, mds);
2055                 mutex_unlock(&mdsc->mutex);
2056                 goto out;
2057         }
2058
2059         result = le32_to_cpu(head->result);
2060
2061         /*
2062          * Handle an ESTALE
2063          * if we're not talking to the authority, send to them
2064          * if the authority has changed while we weren't looking,
2065          * send to new authority
2066          * Otherwise we just have to return an ESTALE
2067          */
2068         if (result == -ESTALE) {
2069                 dout("got ESTALE on request %llu", req->r_tid);
2070                 if (!req->r_inode) {
2071                         /* do nothing; not an authority problem */
2072                 } else if (req->r_direct_mode != USE_AUTH_MDS) {
2073                         dout("not using auth, setting for that now");
2074                         req->r_direct_mode = USE_AUTH_MDS;
2075                         __do_request(mdsc, req);
2076                         mutex_unlock(&mdsc->mutex);
2077                         goto out;
2078                 } else  {
2079                         struct ceph_inode_info *ci = ceph_inode(req->r_inode);
2080                         struct ceph_cap *cap = NULL;
2081
2082                         if (req->r_session)
2083                                 cap = ceph_get_cap_for_mds(ci,
2084                                                    req->r_session->s_mds);
2085
2086                         dout("already using auth");
2087                         if ((!cap || cap != ci->i_auth_cap) ||
2088                             (cap->mseq != req->r_sent_on_mseq)) {
2089                                 dout("but cap changed, so resending");
2090                                 __do_request(mdsc, req);
2091                                 mutex_unlock(&mdsc->mutex);
2092                                 goto out;
2093                         }
2094                 }
2095                 dout("have to return ESTALE on request %llu", req->r_tid);
2096         }
2097
2098
2099         if (head->safe) {
2100                 req->r_got_safe = true;
2101                 __unregister_request(mdsc, req);
2102                 complete_all(&req->r_safe_completion);
2103
2104                 if (req->r_got_unsafe) {
2105                         /*
2106                          * We already handled the unsafe response, now do the
2107                          * cleanup.  No need to examine the response; the MDS
2108                          * doesn't include any result info in the safe
2109                          * response.  And even if it did, there is nothing
2110                          * useful we could do with a revised return value.
2111                          */
2112                         dout("got safe reply %llu, mds%d\n", tid, mds);
2113                         list_del_init(&req->r_unsafe_item);
2114
2115                         /* last unsafe request during umount? */
2116                         if (mdsc->stopping && !__get_oldest_req(mdsc))
2117                                 complete_all(&mdsc->safe_umount_waiters);
2118                         mutex_unlock(&mdsc->mutex);
2119                         goto out;
2120                 }
2121         } else {
2122                 req->r_got_unsafe = true;
2123                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2124         }
2125
2126         dout("handle_reply tid %lld result %d\n", tid, result);
2127         rinfo = &req->r_reply_info;
2128         err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2129         mutex_unlock(&mdsc->mutex);
2130
2131         mutex_lock(&session->s_mutex);
2132         if (err < 0) {
2133                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2134                 ceph_msg_dump(msg);
2135                 goto out_err;
2136         }
2137
2138         /* snap trace */
2139         if (rinfo->snapblob_len) {
2140                 down_write(&mdsc->snap_rwsem);
2141                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2142                                rinfo->snapblob + rinfo->snapblob_len,
2143                                le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
2144                 downgrade_write(&mdsc->snap_rwsem);
2145         } else {
2146                 down_read(&mdsc->snap_rwsem);
2147         }
2148
2149         /* insert trace into our cache */
2150         mutex_lock(&req->r_fill_mutex);
2151         err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2152         if (err == 0) {
2153                 if (result == 0 && req->r_op != CEPH_MDS_OP_GETFILELOCK &&
2154                     rinfo->dir_nr)
2155                         ceph_readdir_prepopulate(req, req->r_session);
2156                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2157         }
2158         mutex_unlock(&req->r_fill_mutex);
2159
2160         up_read(&mdsc->snap_rwsem);
2161 out_err:
2162         mutex_lock(&mdsc->mutex);
2163         if (!req->r_aborted) {
2164                 if (err) {
2165                         req->r_err = err;
2166                 } else {
2167                         req->r_reply = msg;
2168                         ceph_msg_get(msg);
2169                         req->r_got_result = true;
2170                 }
2171         } else {
2172                 dout("reply arrived after request %lld was aborted\n", tid);
2173         }
2174         mutex_unlock(&mdsc->mutex);
2175
2176         ceph_add_cap_releases(mdsc, req->r_session);
2177         mutex_unlock(&session->s_mutex);
2178
2179         /* kick calling process */
2180         complete_request(mdsc, req);
2181 out:
2182         ceph_mdsc_put_request(req);
2183         return;
2184 }
2185
2186
2187
2188 /*
2189  * handle mds notification that our request has been forwarded.
2190  */
2191 static void handle_forward(struct ceph_mds_client *mdsc,
2192                            struct ceph_mds_session *session,
2193                            struct ceph_msg *msg)
2194 {
2195         struct ceph_mds_request *req;
2196         u64 tid = le64_to_cpu(msg->hdr.tid);
2197         u32 next_mds;
2198         u32 fwd_seq;
2199         int err = -EINVAL;
2200         void *p = msg->front.iov_base;
2201         void *end = p + msg->front.iov_len;
2202
2203         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2204         next_mds = ceph_decode_32(&p);
2205         fwd_seq = ceph_decode_32(&p);
2206
2207         mutex_lock(&mdsc->mutex);
2208         req = __lookup_request(mdsc, tid);
2209         if (!req) {
2210                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2211                 goto out;  /* dup reply? */
2212         }
2213
2214         if (req->r_aborted) {
2215                 dout("forward tid %llu aborted, unregistering\n", tid);
2216                 __unregister_request(mdsc, req);
2217         } else if (fwd_seq <= req->r_num_fwd) {
2218                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2219                      tid, next_mds, req->r_num_fwd, fwd_seq);
2220         } else {
2221                 /* resend. forward race not possible; mds would drop */
2222                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2223                 BUG_ON(req->r_err);
2224                 BUG_ON(req->r_got_result);
2225                 req->r_num_fwd = fwd_seq;
2226                 req->r_resend_mds = next_mds;
2227                 put_request_session(req);
2228                 __do_request(mdsc, req);
2229         }
2230         ceph_mdsc_put_request(req);
2231 out:
2232         mutex_unlock(&mdsc->mutex);
2233         return;
2234
2235 bad:
2236         pr_err("mdsc_handle_forward decode error err=%d\n", err);
2237 }
2238
2239 /*
2240  * handle a mds session control message
2241  */
2242 static void handle_session(struct ceph_mds_session *session,
2243                            struct ceph_msg *msg)
2244 {
2245         struct ceph_mds_client *mdsc = session->s_mdsc;
2246         u32 op;
2247         u64 seq;
2248         int mds = session->s_mds;
2249         struct ceph_mds_session_head *h = msg->front.iov_base;
2250         int wake = 0;
2251
2252         /* decode */
2253         if (msg->front.iov_len != sizeof(*h))
2254                 goto bad;
2255         op = le32_to_cpu(h->op);
2256         seq = le64_to_cpu(h->seq);
2257
2258         mutex_lock(&mdsc->mutex);
2259         if (op == CEPH_SESSION_CLOSE)
2260                 __unregister_session(mdsc, session);
2261         /* FIXME: this ttl calculation is generous */
2262         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2263         mutex_unlock(&mdsc->mutex);
2264
2265         mutex_lock(&session->s_mutex);
2266
2267         dout("handle_session mds%d %s %p state %s seq %llu\n",
2268              mds, ceph_session_op_name(op), session,
2269              session_state_name(session->s_state), seq);
2270
2271         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2272                 session->s_state = CEPH_MDS_SESSION_OPEN;
2273                 pr_info("mds%d came back\n", session->s_mds);
2274         }
2275
2276         switch (op) {
2277         case CEPH_SESSION_OPEN:
2278                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2279                         pr_info("mds%d reconnect success\n", session->s_mds);
2280                 session->s_state = CEPH_MDS_SESSION_OPEN;
2281                 renewed_caps(mdsc, session, 0);
2282                 wake = 1;
2283                 if (mdsc->stopping)
2284                         __close_session(mdsc, session);
2285                 break;
2286
2287         case CEPH_SESSION_RENEWCAPS:
2288                 if (session->s_renew_seq == seq)
2289                         renewed_caps(mdsc, session, 1);
2290                 break;
2291
2292         case CEPH_SESSION_CLOSE:
2293                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2294                         pr_info("mds%d reconnect denied\n", session->s_mds);
2295                 remove_session_caps(session);
2296                 wake = 1; /* for good measure */
2297                 wake_up_all(&mdsc->session_close_wq);
2298                 kick_requests(mdsc, mds);
2299                 break;
2300
2301         case CEPH_SESSION_STALE:
2302                 pr_info("mds%d caps went stale, renewing\n",
2303                         session->s_mds);
2304                 spin_lock(&session->s_cap_lock);
2305                 session->s_cap_gen++;
2306                 session->s_cap_ttl = 0;
2307                 spin_unlock(&session->s_cap_lock);
2308                 send_renew_caps(mdsc, session);
2309                 break;
2310
2311         case CEPH_SESSION_RECALL_STATE:
2312                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2313                 break;
2314
2315         default:
2316                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2317                 WARN_ON(1);
2318         }
2319
2320         mutex_unlock(&session->s_mutex);
2321         if (wake) {
2322                 mutex_lock(&mdsc->mutex);
2323                 __wake_requests(mdsc, &session->s_waiting);
2324                 mutex_unlock(&mdsc->mutex);
2325         }
2326         return;
2327
2328 bad:
2329         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2330                (int)msg->front.iov_len);
2331         ceph_msg_dump(msg);
2332         return;
2333 }
2334
2335
2336 /*
2337  * called under session->mutex.
2338  */
2339 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2340                                    struct ceph_mds_session *session)
2341 {
2342         struct ceph_mds_request *req, *nreq;
2343         int err;
2344
2345         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2346
2347         mutex_lock(&mdsc->mutex);
2348         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2349                 err = __prepare_send_request(mdsc, req, session->s_mds);
2350                 if (!err) {
2351                         ceph_msg_get(req->r_request);
2352                         ceph_con_send(&session->s_con, req->r_request);
2353                 }
2354         }
2355         mutex_unlock(&mdsc->mutex);
2356 }
2357
2358 /*
2359  * Encode information about a cap for a reconnect with the MDS.
2360  */
2361 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2362                           void *arg)
2363 {
2364         union {
2365                 struct ceph_mds_cap_reconnect v2;
2366                 struct ceph_mds_cap_reconnect_v1 v1;
2367         } rec;
2368         size_t reclen;
2369         struct ceph_inode_info *ci;
2370         struct ceph_reconnect_state *recon_state = arg;
2371         struct ceph_pagelist *pagelist = recon_state->pagelist;
2372         char *path;
2373         int pathlen, err;
2374         u64 pathbase;
2375         struct dentry *dentry;
2376
2377         ci = cap->ci;
2378
2379         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2380              inode, ceph_vinop(inode), cap, cap->cap_id,
2381              ceph_cap_string(cap->issued));
2382         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2383         if (err)
2384                 return err;
2385
2386         dentry = d_find_alias(inode);
2387         if (dentry) {
2388                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2389                 if (IS_ERR(path)) {
2390                         err = PTR_ERR(path);
2391                         goto out_dput;
2392                 }
2393         } else {
2394                 path = NULL;
2395                 pathlen = 0;
2396         }
2397         err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2398         if (err)
2399                 goto out_free;
2400
2401         spin_lock(&inode->i_lock);
2402         cap->seq = 0;        /* reset cap seq */
2403         cap->issue_seq = 0;  /* and issue_seq */
2404
2405         if (recon_state->flock) {
2406                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2407                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2408                 rec.v2.issued = cpu_to_le32(cap->issued);
2409                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2410                 rec.v2.pathbase = cpu_to_le64(pathbase);
2411                 rec.v2.flock_len = 0;
2412                 reclen = sizeof(rec.v2);
2413         } else {
2414                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2415                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2416                 rec.v1.issued = cpu_to_le32(cap->issued);
2417                 rec.v1.size = cpu_to_le64(inode->i_size);
2418                 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2419                 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2420                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2421                 rec.v1.pathbase = cpu_to_le64(pathbase);
2422                 reclen = sizeof(rec.v1);
2423         }
2424         spin_unlock(&inode->i_lock);
2425
2426         if (recon_state->flock) {
2427                 int num_fcntl_locks, num_flock_locks;
2428                 struct ceph_pagelist_cursor trunc_point;
2429
2430                 ceph_pagelist_set_cursor(pagelist, &trunc_point);
2431                 do {
2432                         lock_flocks();
2433                         ceph_count_locks(inode, &num_fcntl_locks,
2434                                          &num_flock_locks);
2435                         rec.v2.flock_len = (2*sizeof(u32) +
2436                                             (num_fcntl_locks+num_flock_locks) *
2437                                             sizeof(struct ceph_filelock));
2438                         unlock_flocks();
2439
2440                         /* pre-alloc pagelist */
2441                         ceph_pagelist_truncate(pagelist, &trunc_point);
2442                         err = ceph_pagelist_append(pagelist, &rec, reclen);
2443                         if (!err)
2444                                 err = ceph_pagelist_reserve(pagelist,
2445                                                             rec.v2.flock_len);
2446
2447                         /* encode locks */
2448                         if (!err) {
2449                                 lock_flocks();
2450                                 err = ceph_encode_locks(inode,
2451                                                         pagelist,
2452                                                         num_fcntl_locks,
2453                                                         num_flock_locks);
2454                                 unlock_flocks();
2455                         }
2456                 } while (err == -ENOSPC);
2457         } else {
2458                 err = ceph_pagelist_append(pagelist, &rec, reclen);
2459         }
2460
2461 out_free:
2462         kfree(path);
2463 out_dput:
2464         dput(dentry);
2465         return err;
2466 }
2467
2468
2469 /*
2470  * If an MDS fails and recovers, clients need to reconnect in order to
2471  * reestablish shared state.  This includes all caps issued through
2472  * this session _and_ the snap_realm hierarchy.  Because it's not
2473  * clear which snap realms the mds cares about, we send everything we
2474  * know about.. that ensures we'll then get any new info the
2475  * recovering MDS might have.
2476  *
2477  * This is a relatively heavyweight operation, but it's rare.
2478  *
2479  * called with mdsc->mutex held.
2480  */
2481 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2482                                struct ceph_mds_session *session)
2483 {
2484         struct ceph_msg *reply;
2485         struct rb_node *p;
2486         int mds = session->s_mds;
2487         int err = -ENOMEM;
2488         struct ceph_pagelist *pagelist;
2489         struct ceph_reconnect_state recon_state;
2490
2491         pr_info("mds%d reconnect start\n", mds);
2492
2493         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2494         if (!pagelist)
2495                 goto fail_nopagelist;
2496         ceph_pagelist_init(pagelist);
2497
2498         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS);
2499         if (!reply)
2500                 goto fail_nomsg;
2501
2502         mutex_lock(&session->s_mutex);
2503         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2504         session->s_seq = 0;
2505
2506         ceph_con_open(&session->s_con,
2507                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2508
2509         /* replay unsafe requests */
2510         replay_unsafe_requests(mdsc, session);
2511
2512         down_read(&mdsc->snap_rwsem);
2513
2514         dout("session %p state %s\n", session,
2515              session_state_name(session->s_state));
2516
2517         /* drop old cap expires; we're about to reestablish that state */
2518         discard_cap_releases(mdsc, session);
2519
2520         /* traverse this session's caps */
2521         err = ceph_pagelist_encode_32(pagelist, session->s_nr_caps);
2522         if (err)
2523                 goto fail;
2524
2525         recon_state.pagelist = pagelist;
2526         recon_state.flock = session->s_con.peer_features & CEPH_FEATURE_FLOCK;
2527         err = iterate_session_caps(session, encode_caps_cb, &recon_state);
2528         if (err < 0)
2529                 goto fail;
2530
2531         /*
2532          * snaprealms.  we provide mds with the ino, seq (version), and
2533          * parent for all of our realms.  If the mds has any newer info,
2534          * it will tell us.
2535          */
2536         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2537                 struct ceph_snap_realm *realm =
2538                         rb_entry(p, struct ceph_snap_realm, node);
2539                 struct ceph_mds_snaprealm_reconnect sr_rec;
2540
2541                 dout(" adding snap realm %llx seq %lld parent %llx\n",
2542                      realm->ino, realm->seq, realm->parent_ino);
2543                 sr_rec.ino = cpu_to_le64(realm->ino);
2544                 sr_rec.seq = cpu_to_le64(realm->seq);
2545                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2546                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2547                 if (err)
2548                         goto fail;
2549         }
2550
2551         reply->pagelist = pagelist;
2552         if (recon_state.flock)
2553                 reply->hdr.version = cpu_to_le16(2);
2554         reply->hdr.data_len = cpu_to_le32(pagelist->length);
2555         reply->nr_pages = calc_pages_for(0, pagelist->length);
2556         ceph_con_send(&session->s_con, reply);
2557
2558         mutex_unlock(&session->s_mutex);
2559
2560         mutex_lock(&mdsc->mutex);
2561         __wake_requests(mdsc, &session->s_waiting);
2562         mutex_unlock(&mdsc->mutex);
2563
2564         up_read(&mdsc->snap_rwsem);
2565         return;
2566
2567 fail:
2568         ceph_msg_put(reply);
2569         up_read(&mdsc->snap_rwsem);
2570         mutex_unlock(&session->s_mutex);
2571 fail_nomsg:
2572         ceph_pagelist_release(pagelist);
2573         kfree(pagelist);
2574 fail_nopagelist:
2575         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
2576         return;
2577 }
2578
2579
2580 /*
2581  * compare old and new mdsmaps, kicking requests
2582  * and closing out old connections as necessary
2583  *
2584  * called under mdsc->mutex.
2585  */
2586 static void check_new_map(struct ceph_mds_client *mdsc,
2587                           struct ceph_mdsmap *newmap,
2588                           struct ceph_mdsmap *oldmap)
2589 {
2590         int i;
2591         int oldstate, newstate;
2592         struct ceph_mds_session *s;
2593
2594         dout("check_new_map new %u old %u\n",
2595              newmap->m_epoch, oldmap->m_epoch);
2596
2597         for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2598                 if (mdsc->sessions[i] == NULL)
2599                         continue;
2600                 s = mdsc->sessions[i];
2601                 oldstate = ceph_mdsmap_get_state(oldmap, i);
2602                 newstate = ceph_mdsmap_get_state(newmap, i);
2603
2604                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
2605                      i, ceph_mds_state_name(oldstate),
2606                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
2607                      ceph_mds_state_name(newstate),
2608                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
2609                      session_state_name(s->s_state));
2610
2611                 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
2612                            ceph_mdsmap_get_addr(newmap, i),
2613                            sizeof(struct ceph_entity_addr))) {
2614                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2615                                 /* the session never opened, just close it
2616                                  * out now */
2617                                 __wake_requests(mdsc, &s->s_waiting);
2618                                 __unregister_session(mdsc, s);
2619                         } else {
2620                                 /* just close it */
2621                                 mutex_unlock(&mdsc->mutex);
2622                                 mutex_lock(&s->s_mutex);
2623                                 mutex_lock(&mdsc->mutex);
2624                                 ceph_con_close(&s->s_con);
2625                                 mutex_unlock(&s->s_mutex);
2626                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2627                         }
2628
2629                         /* kick any requests waiting on the recovering mds */
2630                         kick_requests(mdsc, i);
2631                 } else if (oldstate == newstate) {
2632                         continue;  /* nothing new with this mds */
2633                 }
2634
2635                 /*
2636                  * send reconnect?
2637                  */
2638                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2639                     newstate >= CEPH_MDS_STATE_RECONNECT) {
2640                         mutex_unlock(&mdsc->mutex);
2641                         send_mds_reconnect(mdsc, s);
2642                         mutex_lock(&mdsc->mutex);
2643                 }
2644
2645                 /*
2646                  * kick request on any mds that has gone active.
2647                  */
2648                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2649                     newstate >= CEPH_MDS_STATE_ACTIVE) {
2650                         if (oldstate != CEPH_MDS_STATE_CREATING &&
2651                             oldstate != CEPH_MDS_STATE_STARTING)
2652                                 pr_info("mds%d recovery completed\n", s->s_mds);
2653                         kick_requests(mdsc, i);
2654                         ceph_kick_flushing_caps(mdsc, s);
2655                         wake_up_session_caps(s, 1);
2656                 }
2657         }
2658
2659         for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
2660                 s = mdsc->sessions[i];
2661                 if (!s)
2662                         continue;
2663                 if (!ceph_mdsmap_is_laggy(newmap, i))
2664                         continue;
2665                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2666                     s->s_state == CEPH_MDS_SESSION_HUNG ||
2667                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
2668                         dout(" connecting to export targets of laggy mds%d\n",
2669                              i);
2670                         __open_export_target_sessions(mdsc, s);
2671                 }
2672         }
2673 }
2674
2675
2676
2677 /*
2678  * leases
2679  */
2680
2681 /*
2682  * caller must hold session s_mutex, dentry->d_lock
2683  */
2684 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2685 {
2686         struct ceph_dentry_info *di = ceph_dentry(dentry);
2687
2688         ceph_put_mds_session(di->lease_session);
2689         di->lease_session = NULL;
2690 }
2691
2692 static void handle_lease(struct ceph_mds_client *mdsc,
2693                          struct ceph_mds_session *session,
2694                          struct ceph_msg *msg)
2695 {
2696         struct super_block *sb = mdsc->fsc->sb;
2697         struct inode *inode;
2698         struct dentry *parent, *dentry;
2699         struct ceph_dentry_info *di;
2700         int mds = session->s_mds;
2701         struct ceph_mds_lease *h = msg->front.iov_base;
2702         u32 seq;
2703         struct ceph_vino vino;
2704         int mask;
2705         struct qstr dname;
2706         int release = 0;
2707
2708         dout("handle_lease from mds%d\n", mds);
2709
2710         /* decode */
2711         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2712                 goto bad;
2713         vino.ino = le64_to_cpu(h->ino);
2714         vino.snap = CEPH_NOSNAP;
2715         mask = le16_to_cpu(h->mask);
2716         seq = le32_to_cpu(h->seq);
2717         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2718         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2719         if (dname.len != get_unaligned_le32(h+1))
2720                 goto bad;
2721
2722         mutex_lock(&session->s_mutex);
2723         session->s_seq++;
2724
2725         /* lookup inode */
2726         inode = ceph_find_inode(sb, vino);
2727         dout("handle_lease %s, mask %d, ino %llx %p %.*s\n",
2728              ceph_lease_op_name(h->action), mask, vino.ino, inode,
2729              dname.len, dname.name);
2730         if (inode == NULL) {
2731                 dout("handle_lease no inode %llx\n", vino.ino);
2732                 goto release;
2733         }
2734
2735         /* dentry */
2736         parent = d_find_alias(inode);
2737         if (!parent) {
2738                 dout("no parent dentry on inode %p\n", inode);
2739                 WARN_ON(1);
2740                 goto release;  /* hrm... */
2741         }
2742         dname.hash = full_name_hash(dname.name, dname.len);
2743         dentry = d_lookup(parent, &dname);
2744         dput(parent);
2745         if (!dentry)
2746                 goto release;
2747
2748         spin_lock(&dentry->d_lock);
2749         di = ceph_dentry(dentry);
2750         switch (h->action) {
2751         case CEPH_MDS_LEASE_REVOKE:
2752                 if (di && di->lease_session == session) {
2753                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
2754                                 h->seq = cpu_to_le32(di->lease_seq);
2755                         __ceph_mdsc_drop_dentry_lease(dentry);
2756                 }
2757                 release = 1;
2758                 break;
2759
2760         case CEPH_MDS_LEASE_RENEW:
2761                 if (di && di->lease_session == session &&
2762                     di->lease_gen == session->s_cap_gen &&
2763                     di->lease_renew_from &&
2764                     di->lease_renew_after == 0) {
2765                         unsigned long duration =
2766                                 le32_to_cpu(h->duration_ms) * HZ / 1000;
2767
2768                         di->lease_seq = seq;
2769                         dentry->d_time = di->lease_renew_from + duration;
2770                         di->lease_renew_after = di->lease_renew_from +
2771                                 (duration >> 1);
2772                         di->lease_renew_from = 0;
2773                 }
2774                 break;
2775         }
2776         spin_unlock(&dentry->d_lock);
2777         dput(dentry);
2778
2779         if (!release)
2780                 goto out;
2781
2782 release:
2783         /* let's just reuse the same message */
2784         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2785         ceph_msg_get(msg);
2786         ceph_con_send(&session->s_con, msg);
2787
2788 out:
2789         iput(inode);
2790         mutex_unlock(&session->s_mutex);
2791         return;
2792
2793 bad:
2794         pr_err("corrupt lease message\n");
2795         ceph_msg_dump(msg);
2796 }
2797
2798 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
2799                               struct inode *inode,
2800                               struct dentry *dentry, char action,
2801                               u32 seq)
2802 {
2803         struct ceph_msg *msg;
2804         struct ceph_mds_lease *lease;
2805         int len = sizeof(*lease) + sizeof(u32);
2806         int dnamelen = 0;
2807
2808         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2809              inode, dentry, ceph_lease_op_name(action), session->s_mds);
2810         dnamelen = dentry->d_name.len;
2811         len += dnamelen;
2812
2813         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS);
2814         if (!msg)
2815                 return;
2816         lease = msg->front.iov_base;
2817         lease->action = action;
2818         lease->mask = cpu_to_le16(1);
2819         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
2820         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
2821         lease->seq = cpu_to_le32(seq);
2822         put_unaligned_le32(dnamelen, lease + 1);
2823         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
2824
2825         /*
2826          * if this is a preemptive lease RELEASE, no need to
2827          * flush request stream, since the actual request will
2828          * soon follow.
2829          */
2830         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
2831
2832         ceph_con_send(&session->s_con, msg);
2833 }
2834
2835 /*
2836  * Preemptively release a lease we expect to invalidate anyway.
2837  * Pass @inode always, @dentry is optional.
2838  */
2839 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2840                              struct dentry *dentry, int mask)
2841 {
2842         struct ceph_dentry_info *di;
2843         struct ceph_mds_session *session;
2844         u32 seq;
2845
2846         BUG_ON(inode == NULL);
2847         BUG_ON(dentry == NULL);
2848         BUG_ON(mask == 0);
2849
2850         /* is dentry lease valid? */
2851         spin_lock(&dentry->d_lock);
2852         di = ceph_dentry(dentry);
2853         if (!di || !di->lease_session ||
2854             di->lease_session->s_mds < 0 ||
2855             di->lease_gen != di->lease_session->s_cap_gen ||
2856             !time_before(jiffies, dentry->d_time)) {
2857                 dout("lease_release inode %p dentry %p -- "
2858                      "no lease on %d\n",
2859                      inode, dentry, mask);
2860                 spin_unlock(&dentry->d_lock);
2861                 return;
2862         }
2863
2864         /* we do have a lease on this dentry; note mds and seq */
2865         session = ceph_get_mds_session(di->lease_session);
2866         seq = di->lease_seq;
2867         __ceph_mdsc_drop_dentry_lease(dentry);
2868         spin_unlock(&dentry->d_lock);
2869
2870         dout("lease_release inode %p dentry %p mask %d to mds%d\n",
2871              inode, dentry, mask, session->s_mds);
2872         ceph_mdsc_lease_send_msg(session, inode, dentry,
2873                                  CEPH_MDS_LEASE_RELEASE, seq);
2874         ceph_put_mds_session(session);
2875 }
2876
2877 /*
2878  * drop all leases (and dentry refs) in preparation for umount
2879  */
2880 static void drop_leases(struct ceph_mds_client *mdsc)
2881 {
2882         int i;
2883
2884         dout("drop_leases\n");
2885         mutex_lock(&mdsc->mutex);
2886         for (i = 0; i < mdsc->max_sessions; i++) {
2887                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2888                 if (!s)
2889                         continue;
2890                 mutex_unlock(&mdsc->mutex);
2891                 mutex_lock(&s->s_mutex);
2892                 mutex_unlock(&s->s_mutex);
2893                 ceph_put_mds_session(s);
2894                 mutex_lock(&mdsc->mutex);
2895         }
2896         mutex_unlock(&mdsc->mutex);
2897 }
2898
2899
2900
2901 /*
2902  * delayed work -- periodically trim expired leases, renew caps with mds
2903  */
2904 static void schedule_delayed(struct ceph_mds_client *mdsc)
2905 {
2906         int delay = 5;
2907         unsigned hz = round_jiffies_relative(HZ * delay);
2908         schedule_delayed_work(&mdsc->delayed_work, hz);
2909 }
2910
2911 static void delayed_work(struct work_struct *work)
2912 {
2913         int i;
2914         struct ceph_mds_client *mdsc =
2915                 container_of(work, struct ceph_mds_client, delayed_work.work);
2916         int renew_interval;
2917         int renew_caps;
2918
2919         dout("mdsc delayed_work\n");
2920         ceph_check_delayed_caps(mdsc);
2921
2922         mutex_lock(&mdsc->mutex);
2923         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
2924         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
2925                                    mdsc->last_renew_caps);
2926         if (renew_caps)
2927                 mdsc->last_renew_caps = jiffies;
2928
2929         for (i = 0; i < mdsc->max_sessions; i++) {
2930                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2931                 if (s == NULL)
2932                         continue;
2933                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
2934                         dout("resending session close request for mds%d\n",
2935                              s->s_mds);
2936                         request_close_session(mdsc, s);
2937                         ceph_put_mds_session(s);
2938                         continue;
2939                 }
2940                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
2941                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
2942                                 s->s_state = CEPH_MDS_SESSION_HUNG;
2943                                 pr_info("mds%d hung\n", s->s_mds);
2944                         }
2945                 }
2946                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
2947                         /* this mds is failed or recovering, just wait */
2948                         ceph_put_mds_session(s);
2949                         continue;
2950                 }
2951                 mutex_unlock(&mdsc->mutex);
2952
2953                 mutex_lock(&s->s_mutex);
2954                 if (renew_caps)
2955                         send_renew_caps(mdsc, s);
2956                 else
2957                         ceph_con_keepalive(&s->s_con);
2958                 ceph_add_cap_releases(mdsc, s);
2959                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2960                     s->s_state == CEPH_MDS_SESSION_HUNG)
2961                         ceph_send_cap_releases(mdsc, s);
2962                 mutex_unlock(&s->s_mutex);
2963                 ceph_put_mds_session(s);
2964
2965                 mutex_lock(&mdsc->mutex);
2966         }
2967         mutex_unlock(&mdsc->mutex);
2968
2969         schedule_delayed(mdsc);
2970 }
2971
2972 int ceph_mdsc_init(struct ceph_fs_client *fsc)
2973
2974 {
2975         struct ceph_mds_client *mdsc;
2976
2977         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
2978         if (!mdsc)
2979                 return -ENOMEM;
2980         mdsc->fsc = fsc;
2981         fsc->mdsc = mdsc;
2982         mutex_init(&mdsc->mutex);
2983         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
2984         if (mdsc->mdsmap == NULL)
2985                 return -ENOMEM;
2986
2987         init_completion(&mdsc->safe_umount_waiters);
2988         init_waitqueue_head(&mdsc->session_close_wq);
2989         INIT_LIST_HEAD(&mdsc->waiting_for_map);
2990         mdsc->sessions = NULL;
2991         mdsc->max_sessions = 0;
2992         mdsc->stopping = 0;
2993         init_rwsem(&mdsc->snap_rwsem);
2994         mdsc->snap_realms = RB_ROOT;
2995         INIT_LIST_HEAD(&mdsc->snap_empty);
2996         spin_lock_init(&mdsc->snap_empty_lock);
2997         mdsc->last_tid = 0;
2998         mdsc->request_tree = RB_ROOT;
2999         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3000         mdsc->last_renew_caps = jiffies;
3001         INIT_LIST_HEAD(&mdsc->cap_delay_list);
3002         spin_lock_init(&mdsc->cap_delay_lock);
3003         INIT_LIST_HEAD(&mdsc->snap_flush_list);
3004         spin_lock_init(&mdsc->snap_flush_lock);
3005         mdsc->cap_flush_seq = 0;
3006         INIT_LIST_HEAD(&mdsc->cap_dirty);
3007         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3008         mdsc->num_cap_flushing = 0;
3009         spin_lock_init(&mdsc->cap_dirty_lock);
3010         init_waitqueue_head(&mdsc->cap_flushing_wq);
3011         spin_lock_init(&mdsc->dentry_lru_lock);
3012         INIT_LIST_HEAD(&mdsc->dentry_lru);
3013
3014         ceph_caps_init(mdsc);
3015         ceph_adjust_min_caps(mdsc, fsc->min_caps);
3016
3017         return 0;
3018 }
3019
3020 /*
3021  * Wait for safe replies on open mds requests.  If we time out, drop
3022  * all requests from the tree to avoid dangling dentry refs.
3023  */
3024 static void wait_requests(struct ceph_mds_client *mdsc)
3025 {
3026         struct ceph_mds_request *req;
3027         struct ceph_fs_client *fsc = mdsc->fsc;
3028
3029         mutex_lock(&mdsc->mutex);
3030         if (__get_oldest_req(mdsc)) {
3031                 mutex_unlock(&mdsc->mutex);
3032
3033                 dout("wait_requests waiting for requests\n");
3034                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3035                                     fsc->client->options->mount_timeout * HZ);
3036
3037                 /* tear down remaining requests */
3038                 mutex_lock(&mdsc->mutex);
3039                 while ((req = __get_oldest_req(mdsc))) {
3040                         dout("wait_requests timed out on tid %llu\n",
3041                              req->r_tid);
3042                         __unregister_request(mdsc, req);
3043                 }
3044         }
3045         mutex_unlock(&mdsc->mutex);
3046         dout("wait_requests done\n");
3047 }
3048
3049 /*
3050  * called before mount is ro, and before dentries are torn down.
3051  * (hmm, does this still race with new lookups?)
3052  */
3053 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3054 {
3055         dout("pre_umount\n");
3056         mdsc->stopping = 1;
3057
3058         drop_leases(mdsc);
3059         ceph_flush_dirty_caps(mdsc);
3060         wait_requests(mdsc);
3061
3062         /*
3063          * wait for reply handlers to drop their request refs and
3064          * their inode/dcache refs
3065          */
3066         ceph_msgr_flush();
3067 }
3068
3069 /*
3070  * wait for all write mds requests to flush.
3071  */
3072 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3073 {
3074         struct ceph_mds_request *req = NULL, *nextreq;
3075         struct rb_node *n;
3076
3077         mutex_lock(&mdsc->mutex);
3078         dout("wait_unsafe_requests want %lld\n", want_tid);
3079 restart:
3080         req = __get_oldest_req(mdsc);
3081         while (req && req->r_tid <= want_tid) {
3082                 /* find next request */
3083                 n = rb_next(&req->r_node);
3084                 if (n)
3085                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3086                 else
3087                         nextreq = NULL;
3088                 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
3089                         /* write op */
3090                         ceph_mdsc_get_request(req);
3091                         if (nextreq)
3092                                 ceph_mdsc_get_request(nextreq);
3093                         mutex_unlock(&mdsc->mutex);
3094                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
3095                              req->r_tid, want_tid);
3096                         wait_for_completion(&req->r_safe_completion);
3097                         mutex_lock(&mdsc->mutex);
3098                         ceph_mdsc_put_request(req);
3099                         if (!nextreq)
3100                                 break;  /* next dne before, so we're done! */
3101                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
3102                                 /* next request was removed from tree */
3103                                 ceph_mdsc_put_request(nextreq);
3104                                 goto restart;
3105                         }
3106                         ceph_mdsc_put_request(nextreq);  /* won't go away */
3107                 }
3108                 req = nextreq;
3109         }
3110         mutex_unlock(&mdsc->mutex);
3111         dout("wait_unsafe_requests done\n");
3112 }
3113
3114 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3115 {
3116         u64 want_tid, want_flush;
3117
3118         if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3119                 return;
3120
3121         dout("sync\n");
3122         mutex_lock(&mdsc->mutex);
3123         want_tid = mdsc->last_tid;
3124         want_flush = mdsc->cap_flush_seq;
3125         mutex_unlock(&mdsc->mutex);
3126         dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
3127
3128         ceph_flush_dirty_caps(mdsc);
3129
3130         wait_unsafe_requests(mdsc, want_tid);
3131         wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
3132 }
3133
3134 /*
3135  * true if all sessions are closed, or we force unmount
3136  */
3137 bool done_closing_sessions(struct ceph_mds_client *mdsc)
3138 {
3139         int i, n = 0;
3140
3141         if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3142                 return true;
3143
3144         mutex_lock(&mdsc->mutex);
3145         for (i = 0; i < mdsc->max_sessions; i++)
3146                 if (mdsc->sessions[i])
3147                         n++;
3148         mutex_unlock(&mdsc->mutex);
3149         return n == 0;
3150 }
3151
3152 /*
3153  * called after sb is ro.
3154  */
3155 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3156 {
3157         struct ceph_mds_session *session;
3158         int i;
3159         struct ceph_fs_client *fsc = mdsc->fsc;
3160         unsigned long timeout = fsc->client->options->mount_timeout * HZ;
3161
3162         dout("close_sessions\n");
3163
3164         /* close sessions */
3165         mutex_lock(&mdsc->mutex);
3166         for (i = 0; i < mdsc->max_sessions; i++) {
3167                 session = __ceph_lookup_mds_session(mdsc, i);
3168                 if (!session)
3169                         continue;
3170                 mutex_unlock(&mdsc->mutex);
3171                 mutex_lock(&session->s_mutex);
3172                 __close_session(mdsc, session);
3173                 mutex_unlock(&session->s_mutex);
3174                 ceph_put_mds_session(session);
3175                 mutex_lock(&mdsc->mutex);
3176         }
3177         mutex_unlock(&mdsc->mutex);
3178
3179         dout("waiting for sessions to close\n");
3180         wait_event_timeout(mdsc->session_close_wq, done_closing_sessions(mdsc),
3181                            timeout);
3182
3183         /* tear down remaining sessions */
3184         mutex_lock(&mdsc->mutex);
3185         for (i = 0; i < mdsc->max_sessions; i++) {
3186                 if (mdsc->sessions[i]) {
3187                         session = get_session(mdsc->sessions[i]);
3188                         __unregister_session(mdsc, session);
3189                         mutex_unlock(&mdsc->mutex);
3190                         mutex_lock(&session->s_mutex);
3191                         remove_session_caps(session);
3192                         mutex_unlock(&session->s_mutex);
3193                         ceph_put_mds_session(session);
3194                         mutex_lock(&mdsc->mutex);
3195                 }
3196         }
3197         WARN_ON(!list_empty(&mdsc->cap_delay_list));
3198         mutex_unlock(&mdsc->mutex);
3199
3200         ceph_cleanup_empty_realms(mdsc);
3201
3202         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3203
3204         dout("stopped\n");
3205 }
3206
3207 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3208 {
3209         dout("stop\n");
3210         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3211         if (mdsc->mdsmap)
3212                 ceph_mdsmap_destroy(mdsc->mdsmap);
3213         kfree(mdsc->sessions);
3214         ceph_caps_finalize(mdsc);
3215 }
3216
3217 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3218 {
3219         struct ceph_mds_client *mdsc = fsc->mdsc;
3220
3221         dout("mdsc_destroy %p\n", mdsc);
3222         ceph_mdsc_stop(mdsc);
3223
3224         /* flush out any connection work with references to us */
3225         ceph_msgr_flush();
3226
3227         fsc->mdsc = NULL;
3228         kfree(mdsc);
3229         dout("mdsc_destroy %p done\n", mdsc);
3230 }
3231
3232
3233 /*
3234  * handle mds map update.
3235  */
3236 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3237 {
3238         u32 epoch;
3239         u32 maplen;
3240         void *p = msg->front.iov_base;
3241         void *end = p + msg->front.iov_len;
3242         struct ceph_mdsmap *newmap, *oldmap;
3243         struct ceph_fsid fsid;
3244         int err = -EINVAL;
3245
3246         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3247         ceph_decode_copy(&p, &fsid, sizeof(fsid));
3248         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3249                 return;
3250         epoch = ceph_decode_32(&p);
3251         maplen = ceph_decode_32(&p);
3252         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3253
3254         /* do we need it? */
3255         ceph_monc_got_mdsmap(&mdsc->fsc->client->monc, epoch);
3256         mutex_lock(&mdsc->mutex);
3257         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3258                 dout("handle_map epoch %u <= our %u\n",
3259                      epoch, mdsc->mdsmap->m_epoch);
3260                 mutex_unlock(&mdsc->mutex);
3261                 return;
3262         }
3263
3264         newmap = ceph_mdsmap_decode(&p, end);
3265         if (IS_ERR(newmap)) {
3266                 err = PTR_ERR(newmap);
3267                 goto bad_unlock;
3268         }
3269
3270         /* swap into place */
3271         if (mdsc->mdsmap) {
3272                 oldmap = mdsc->mdsmap;
3273                 mdsc->mdsmap = newmap;
3274                 check_new_map(mdsc, newmap, oldmap);
3275                 ceph_mdsmap_destroy(oldmap);
3276         } else {
3277                 mdsc->mdsmap = newmap;  /* first mds map */
3278         }
3279         mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3280
3281         __wake_requests(mdsc, &mdsc->waiting_for_map);
3282
3283         mutex_unlock(&mdsc->mutex);
3284         schedule_delayed(mdsc);
3285         return;
3286
3287 bad_unlock:
3288         mutex_unlock(&mdsc->mutex);
3289 bad:
3290         pr_err("error decoding mdsmap %d\n", err);
3291         return;
3292 }
3293
3294 static struct ceph_connection *con_get(struct ceph_connection *con)
3295 {
3296         struct ceph_mds_session *s = con->private;
3297
3298         if (get_session(s)) {
3299                 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3300                 return con;
3301         }
3302         dout("mdsc con_get %p FAIL\n", s);
3303         return NULL;
3304 }
3305
3306 static void con_put(struct ceph_connection *con)
3307 {
3308         struct ceph_mds_session *s = con->private;
3309
3310         dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3311         ceph_put_mds_session(s);
3312 }
3313
3314 /*
3315  * if the client is unresponsive for long enough, the mds will kill
3316  * the session entirely.
3317  */
3318 static void peer_reset(struct ceph_connection *con)
3319 {
3320         struct ceph_mds_session *s = con->private;
3321         struct ceph_mds_client *mdsc = s->s_mdsc;
3322
3323         pr_warning("mds%d closed our session\n", s->s_mds);
3324         send_mds_reconnect(mdsc, s);
3325 }
3326
3327 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3328 {
3329         struct ceph_mds_session *s = con->private;
3330         struct ceph_mds_client *mdsc = s->s_mdsc;
3331         int type = le16_to_cpu(msg->hdr.type);
3332
3333         mutex_lock(&mdsc->mutex);
3334         if (__verify_registered_session(mdsc, s) < 0) {
3335                 mutex_unlock(&mdsc->mutex);
3336                 goto out;
3337         }
3338         mutex_unlock(&mdsc->mutex);
3339
3340         switch (type) {
3341         case CEPH_MSG_MDS_MAP:
3342                 ceph_mdsc_handle_map(mdsc, msg);
3343                 break;
3344         case CEPH_MSG_CLIENT_SESSION:
3345                 handle_session(s, msg);
3346                 break;
3347         case CEPH_MSG_CLIENT_REPLY:
3348                 handle_reply(s, msg);
3349                 break;
3350         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3351                 handle_forward(mdsc, s, msg);
3352                 break;
3353         case CEPH_MSG_CLIENT_CAPS:
3354                 ceph_handle_caps(s, msg);
3355                 break;
3356         case CEPH_MSG_CLIENT_SNAP:
3357                 ceph_handle_snap(mdsc, s, msg);
3358                 break;
3359         case CEPH_MSG_CLIENT_LEASE:
3360                 handle_lease(mdsc, s, msg);
3361                 break;
3362
3363         default:
3364                 pr_err("received unknown message type %d %s\n", type,
3365                        ceph_msg_type_name(type));
3366         }
3367 out:
3368         ceph_msg_put(msg);
3369 }
3370
3371 /*
3372  * authentication
3373  */
3374 static int get_authorizer(struct ceph_connection *con,
3375                           void **buf, int *len, int *proto,
3376                           void **reply_buf, int *reply_len, int force_new)
3377 {
3378         struct ceph_mds_session *s = con->private;
3379         struct ceph_mds_client *mdsc = s->s_mdsc;
3380         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3381         int ret = 0;
3382
3383         if (force_new && s->s_authorizer) {
3384                 ac->ops->destroy_authorizer(ac, s->s_authorizer);
3385                 s->s_authorizer = NULL;
3386         }
3387         if (s->s_authorizer == NULL) {
3388                 if (ac->ops->create_authorizer) {
3389                         ret = ac->ops->create_authorizer(
3390                                 ac, CEPH_ENTITY_TYPE_MDS,
3391                                 &s->s_authorizer,
3392                                 &s->s_authorizer_buf,
3393                                 &s->s_authorizer_buf_len,
3394                                 &s->s_authorizer_reply_buf,
3395                                 &s->s_authorizer_reply_buf_len);
3396                         if (ret)
3397                                 return ret;
3398                 }
3399         }
3400
3401         *proto = ac->protocol;
3402         *buf = s->s_authorizer_buf;
3403         *len = s->s_authorizer_buf_len;
3404         *reply_buf = s->s_authorizer_reply_buf;
3405         *reply_len = s->s_authorizer_reply_buf_len;
3406         return 0;
3407 }
3408
3409
3410 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3411 {
3412         struct ceph_mds_session *s = con->private;
3413         struct ceph_mds_client *mdsc = s->s_mdsc;
3414         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3415
3416         return ac->ops->verify_authorizer_reply(ac, s->s_authorizer, len);
3417 }
3418
3419 static int invalidate_authorizer(struct ceph_connection *con)
3420 {
3421         struct ceph_mds_session *s = con->private;
3422         struct ceph_mds_client *mdsc = s->s_mdsc;
3423         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3424
3425         if (ac->ops->invalidate_authorizer)
3426                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3427
3428         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
3429 }
3430
3431 static const struct ceph_connection_operations mds_con_ops = {
3432         .get = con_get,
3433         .put = con_put,
3434         .dispatch = dispatch,
3435         .get_authorizer = get_authorizer,
3436         .verify_authorizer_reply = verify_authorizer_reply,
3437         .invalidate_authorizer = invalidate_authorizer,
3438         .peer_reset = peer_reset,
3439 };
3440
3441 /* eof */