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/ceph_features.h>
  14 #include <linux/ceph/messenger.h>
  15 #include <linux/ceph/decode.h>
  16 #include <linux/ceph/pagelist.h>
  17 #include <linux/ceph/auth.h>
  18 #include <linux/ceph/debugfs.h>
  19
  20 /*
  21  * A cluster of MDS (metadata server) daemons is responsible for
  22  * managing the file system namespace (the directory hierarchy and
  23  * inodes) and for coordinating shared access to storage.  Metadata is
  24  * partitioning hierarchically across a number of servers, and that
  25  * partition varies over time as the cluster adjusts the distribution
  26  * in order to balance load.
  27  *
  28  * The MDS client is primarily responsible to managing synchronous
  29  * metadata requests for operations like open, unlink, and so forth.
  30  * If there is a MDS failure, we find out about it when we (possibly
  31  * request and) receive a new MDS map, and can resubmit affected
  32  * requests.
  33  *
  34  * For the most part, though, we take advantage of a lossless
  35  * communications channel to the MDS, and do not need to worry about
  36  * timing out or resubmitting requests.
  37  *
  38  * We maintain a stateful "session" with each MDS we interact with.
  39  * Within each session, we sent periodic heartbeat messages to ensure
  40  * any capabilities or leases we have been issues remain valid.  If
  41  * the session times out and goes stale, our leases and capabilities
  42  * are no longer valid.
  43  */
  44
  45 struct ceph_reconnect_state {
  46         int nr_caps;
  47         struct ceph_pagelist *pagelist;
  48         bool flock;
  49 };
  50
  51 static void __wake_requests(struct ceph_mds_client *mdsc,
  52                             struct list_head *head);
  53
  54 static const struct ceph_connection_operations mds_con_ops;
  55
  56
  57 /*
  58  * mds reply parsing
  59  */
  60
  61 /*
  62  * parse individual inode info
  63  */
  64 static int parse_reply_info_in(void **p, void *end,
  65                                struct ceph_mds_reply_info_in *info,
  66                                int features)
  67 {
  68         int err = -EIO;
  69
  70         info->in = *p;
  71         *p += sizeof(struct ceph_mds_reply_inode) +
  72                 sizeof(*info->in->fragtree.splits) *
  73                 le32_to_cpu(info->in->fragtree.nsplits);
  74
  75         ceph_decode_32_safe(p, end, info->symlink_len, bad);
  76         ceph_decode_need(p, end, info->symlink_len, bad);
  77         info->symlink = *p;
  78         *p += info->symlink_len;
  79
  80         if (features & CEPH_FEATURE_DIRLAYOUTHASH)
  81                 ceph_decode_copy_safe(p, end, &info->dir_layout,
  82                                       sizeof(info->dir_layout), bad);
  83         else
  84                 memset(&info->dir_layout, 0, sizeof(info->dir_layout));
  85
  86         ceph_decode_32_safe(p, end, info->xattr_len, bad);
  87         ceph_decode_need(p, end, info->xattr_len, bad);
  88         info->xattr_data = *p;
  89         *p += info->xattr_len;
  90         return 0;
  91 bad:
  92         return err;
  93 }
  94
  95 /*
  96  * parse a normal reply, which may contain a (dir+)dentry and/or a
  97  * target inode.
  98  */
  99 static int parse_reply_info_trace(void **p, void *end,
 100                                   struct ceph_mds_reply_info_parsed *info,
 101                                   int features)
 102 {
 103         int err;
 104
 105         if (info->head->is_dentry) {
 106                 err = parse_reply_info_in(p, end, &info->diri, features);
 107                 if (err < 0)
 108                         goto out_bad;
 109
 110                 if (unlikely(*p + sizeof(*info->dirfrag) > end))
 111                         goto bad;
 112                 info->dirfrag = *p;
 113                 *p += sizeof(*info->dirfrag) +
 114                         sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
 115                 if (unlikely(*p > end))
 116                         goto bad;
 117
 118                 ceph_decode_32_safe(p, end, info->dname_len, bad);
 119                 ceph_decode_need(p, end, info->dname_len, bad);
 120                 info->dname = *p;
 121                 *p += info->dname_len;
 122                 info->dlease = *p;
 123                 *p += sizeof(*info->dlease);
 124         }
 125
 126         if (info->head->is_target) {
 127                 err = parse_reply_info_in(p, end, &info->targeti, features);
 128                 if (err < 0)
 129                         goto out_bad;
 130         }
 131
 132         if (unlikely(*p != end))
 133                 goto bad;
 134         return 0;
 135
 136 bad:
 137         err = -EIO;
 138 out_bad:
 139         pr_err("problem parsing mds trace %d\n", err);
 140         return err;
 141 }
 142
 143 /*
 144  * parse readdir results
 145  */
 146 static int parse_reply_info_dir(void **p, void *end,
 147                                 struct ceph_mds_reply_info_parsed *info,
 148                                 int features)
 149 {
 150         u32 num, i = 0;
 151         int err;
 152
 153         info->dir_dir = *p;
 154         if (*p + sizeof(*info->dir_dir) > end)
 155                 goto bad;
 156         *p += sizeof(*info->dir_dir) +
 157                 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
 158         if (*p > end)
 159                 goto bad;
 160
 161         ceph_decode_need(p, end, sizeof(num) + 2, bad);
 162         num = ceph_decode_32(p);
 163         info->dir_end = ceph_decode_8(p);
 164         info->dir_complete = ceph_decode_8(p);
 165         if (num == 0)
 166                 goto done;
 167
 168         /* alloc large array */
 169         info->dir_nr = num;
 170         info->dir_in = kcalloc(num, sizeof(*info->dir_in) +
 171                                sizeof(*info->dir_dname) +
 172                                sizeof(*info->dir_dname_len) +
 173                                sizeof(*info->dir_dlease),
 174                                GFP_NOFS);
 175         if (info->dir_in == NULL) {
 176                 err = -ENOMEM;
 177                 goto out_bad;
 178         }
 179         info->dir_dname = (void *)(info->dir_in + num);
 180         info->dir_dname_len = (void *)(info->dir_dname + num);
 181         info->dir_dlease = (void *)(info->dir_dname_len + num);
 182
 183         while (num) {
 184                 /* dentry */
 185                 ceph_decode_need(p, end, sizeof(u32)*2, bad);
 186                 info->dir_dname_len[i] = ceph_decode_32(p);
 187                 ceph_decode_need(p, end, info->dir_dname_len[i], bad);
 188                 info->dir_dname[i] = *p;
 189                 *p += info->dir_dname_len[i];
 190                 dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
 191                      info->dir_dname[i]);
 192                 info->dir_dlease[i] = *p;
 193                 *p += sizeof(struct ceph_mds_reply_lease);
 194
 195                 /* inode */
 196                 err = parse_reply_info_in(p, end, &info->dir_in[i], features);
 197                 if (err < 0)
 198                         goto out_bad;
 199                 i++;
 200                 num--;
 201         }
 202
 203 done:
 204         if (*p != end)
 205                 goto bad;
 206         return 0;
 207
 208 bad:
 209         err = -EIO;
 210 out_bad:
 211         pr_err("problem parsing dir contents %d\n", err);
 212         return err;
 213 }
 214
 215 /*
 216  * parse fcntl F_GETLK results
 217  */
 218 static int parse_reply_info_filelock(void **p, void *end,
 219                                      struct ceph_mds_reply_info_parsed *info,
 220                                      int features)
 221 {
 222         if (*p + sizeof(*info->filelock_reply) > end)
 223                 goto bad;
 224
 225         info->filelock_reply = *p;
 226         *p += sizeof(*info->filelock_reply);
 227
 228         if (unlikely(*p != end))
 229                 goto bad;
 230         return 0;
 231
 232 bad:
 233         return -EIO;
 234 }
 235
 236 /*
 237  * parse create results
 238  */
 239 static int parse_reply_info_create(void **p, void *end,
 240                                   struct ceph_mds_reply_info_parsed *info,
 241                                   int features)
 242 {
 243         if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
 244                 if (*p == end) {
 245                         info->has_create_ino = false;
 246                 } else {
 247                         info->has_create_ino = true;
 248                         info->ino = ceph_decode_64(p);
 249                 }
 250         }
 251
 252         if (unlikely(*p != end))
 253                 goto bad;
 254         return 0;
 255
 256 bad:
 257         return -EIO;
 258 }
 259
 260 /*
 261  * parse extra results
 262  */
 263 static int parse_reply_info_extra(void **p, void *end,
 264                                   struct ceph_mds_reply_info_parsed *info,
 265                                   int features)
 266 {
 267         if (info->head->op == CEPH_MDS_OP_GETFILELOCK)
 268                 return parse_reply_info_filelock(p, end, info, features);
 269         else if (info->head->op == CEPH_MDS_OP_READDIR ||
 270                  info->head->op == CEPH_MDS_OP_LSSNAP)
 271                 return parse_reply_info_dir(p, end, info, features);
 272         else if (info->head->op == CEPH_MDS_OP_CREATE)
 273                 return parse_reply_info_create(p, end, info, features);
 274         else
 275                 return -EIO;
 276 }
 277
 278 /*
 279  * parse entire mds reply
 280  */
 281 static int parse_reply_info(struct ceph_msg *msg,
 282                             struct ceph_mds_reply_info_parsed *info,
 283                             int features)
 284 {
 285         void *p, *end;
 286         u32 len;
 287         int err;
 288
 289         info->head = msg->front.iov_base;
 290         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 291         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 292
 293         /* trace */
 294         ceph_decode_32_safe(&p, end, len, bad);
 295         if (len > 0) {
 296                 ceph_decode_need(&p, end, len, bad);
 297                 err = parse_reply_info_trace(&p, p+len, info, features);
 298                 if (err < 0)
 299                         goto out_bad;
 300         }
 301
 302         /* extra */
 303         ceph_decode_32_safe(&p, end, len, bad);
 304         if (len > 0) {
 305                 ceph_decode_need(&p, end, len, bad);
 306                 err = parse_reply_info_extra(&p, p+len, info, features);
 307                 if (err < 0)
 308                         goto out_bad;
 309         }
 310
 311         /* snap blob */
 312         ceph_decode_32_safe(&p, end, len, bad);
 313         info->snapblob_len = len;
 314         info->snapblob = p;
 315         p += len;
 316
 317         if (p != end)
 318                 goto bad;
 319         return 0;
 320
 321 bad:
 322         err = -EIO;
 323 out_bad:
 324         pr_err("mds parse_reply err %d\n", err);
 325         return err;
 326 }
 327
 328 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
 329 {
 330         kfree(info->dir_in);
 331 }
 332
 333
 334 /*
 335  * sessions
 336  */
 337 static const char *session_state_name(int s)
 338 {
 339         switch (s) {
 340         case CEPH_MDS_SESSION_NEW: return "new";
 341         case CEPH_MDS_SESSION_OPENING: return "opening";
 342         case CEPH_MDS_SESSION_OPEN: return "open";
 343         case CEPH_MDS_SESSION_HUNG: return "hung";
 344         case CEPH_MDS_SESSION_CLOSING: return "closing";
 345         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
 346         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
 347         default: return "???";
 348         }
 349 }
 350
 351 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
 352 {
 353         if (atomic_inc_not_zero(&s->s_ref)) {
 354                 dout("mdsc get_session %p %d -> %d\n", s,
 355                      atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
 356                 return s;
 357         } else {
 358                 dout("mdsc get_session %p 0 -- FAIL", s);
 359                 return NULL;
 360         }
 361 }
 362
 363 void ceph_put_mds_session(struct ceph_mds_session *s)
 364 {
 365         dout("mdsc put_session %p %d -> %d\n", s,
 366              atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
 367         if (atomic_dec_and_test(&s->s_ref)) {
 368                 if (s->s_auth.authorizer)
 369                         ceph_auth_destroy_authorizer(
 370                                 s->s_mdsc->fsc->client->monc.auth,
 371                                 s->s_auth.authorizer);
 372                 kfree(s);
 373         }
 374 }
 375
 376 /*
 377  * called under mdsc->mutex
 378  */
 379 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
 380                                                    int mds)
 381 {
 382         struct ceph_mds_session *session;
 383
 384         if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
 385                 return NULL;
 386         session = mdsc->sessions[mds];
 387         dout("lookup_mds_session %p %d\n", session,
 388              atomic_read(&session->s_ref));
 389         get_session(session);
 390         return session;
 391 }
 392
 393 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
 394 {
 395         if (mds >= mdsc->max_sessions)
 396                 return false;
 397         return mdsc->sessions[mds];
 398 }
 399
 400 static int __verify_registered_session(struct ceph_mds_client *mdsc,
 401                                        struct ceph_mds_session *s)
 402 {
 403         if (s->s_mds >= mdsc->max_sessions ||
 404             mdsc->sessions[s->s_mds] != s)
 405                 return -ENOENT;
 406         return 0;
 407 }
 408
 409 /*
 410  * create+register a new session for given mds.
 411  * called under mdsc->mutex.
 412  */
 413 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 414                                                  int mds)
 415 {
 416         struct ceph_mds_session *s;
 417
 418         if (mds >= mdsc->mdsmap->m_max_mds)
 419                 return ERR_PTR(-EINVAL);
 420
 421         s = kzalloc(sizeof(*s), GFP_NOFS);
 422         if (!s)
 423                 return ERR_PTR(-ENOMEM);
 424         s->s_mdsc = mdsc;
 425         s->s_mds = mds;
 426         s->s_state = CEPH_MDS_SESSION_NEW;
 427         s->s_ttl = 0;
 428         s->s_seq = 0;
 429         mutex_init(&s->s_mutex);
 430
 431         ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
 432
 433         spin_lock_init(&s->s_gen_ttl_lock);
 434         s->s_cap_gen = 0;
 435         s->s_cap_ttl = jiffies - 1;
 436
 437         spin_lock_init(&s->s_cap_lock);
 438         s->s_renew_requested = 0;
 439         s->s_renew_seq = 0;
 440         INIT_LIST_HEAD(&s->s_caps);
 441         s->s_nr_caps = 0;
 442         s->s_trim_caps = 0;
 443         atomic_set(&s->s_ref, 1);
 444         INIT_LIST_HEAD(&s->s_waiting);
 445         INIT_LIST_HEAD(&s->s_unsafe);
 446         s->s_num_cap_releases = 0;
 447         s->s_cap_reconnect = 0;
 448         s->s_cap_iterator = NULL;
 449         INIT_LIST_HEAD(&s->s_cap_releases);
 450         INIT_LIST_HEAD(&s->s_cap_releases_done);
 451         INIT_LIST_HEAD(&s->s_cap_flushing);
 452         INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
 453
 454         dout("register_session mds%d\n", mds);
 455         if (mds >= mdsc->max_sessions) {
 456                 int newmax = 1 << get_count_order(mds+1);
 457                 struct ceph_mds_session **sa;
 458
 459                 dout("register_session realloc to %d\n", newmax);
 460                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 461                 if (sa == NULL)
 462                         goto fail_realloc;
 463                 if (mdsc->sessions) {
 464                         memcpy(sa, mdsc->sessions,
 465                                mdsc->max_sessions * sizeof(void *));
 466                         kfree(mdsc->sessions);
 467                 }
 468                 mdsc->sessions = sa;
 469                 mdsc->max_sessions = newmax;
 470         }
 471         mdsc->sessions[mds] = s;
 472         atomic_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
 473
 474         ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
 475                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
 476
 477         return s;
 478
 479 fail_realloc:
 480         kfree(s);
 481         return ERR_PTR(-ENOMEM);
 482 }
 483
 484 /*
 485  * called under mdsc->mutex
 486  */
 487 static void __unregister_session(struct ceph_mds_client *mdsc,
 488                                struct ceph_mds_session *s)
 489 {
 490         dout("__unregister_session mds%d %p\n", s->s_mds, s);
 491         BUG_ON(mdsc->sessions[s->s_mds] != s);
 492         mdsc->sessions[s->s_mds] = NULL;
 493         ceph_con_close(&s->s_con);
 494         ceph_put_mds_session(s);
 495 }
 496
 497 /*
 498  * drop session refs in request.
 499  *
 500  * should be last request ref, or hold mdsc->mutex
 501  */
 502 static void put_request_session(struct ceph_mds_request *req)
 503 {
 504         if (req->r_session) {
 505                 ceph_put_mds_session(req->r_session);
 506                 req->r_session = NULL;
 507         }
 508 }
 509
 510 void ceph_mdsc_release_request(struct kref *kref)
 511 {
 512         struct ceph_mds_request *req = container_of(kref,
 513                                                     struct ceph_mds_request,
 514                                                     r_kref);
 515         if (req->r_request)
 516                 ceph_msg_put(req->r_request);
 517         if (req->r_reply) {
 518                 ceph_msg_put(req->r_reply);
 519                 destroy_reply_info(&req->r_reply_info);
 520         }
 521         if (req->r_inode) {
 522                 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
 523                 iput(req->r_inode);
 524         }
 525         if (req->r_locked_dir)
 526                 ceph_put_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
 527         if (req->r_target_inode)
 528                 iput(req->r_target_inode);
 529         if (req->r_dentry)
 530                 dput(req->r_dentry);
 531         if (req->r_old_dentry) {
 532                 /*
 533                  * track (and drop pins for) r_old_dentry_dir
 534                  * separately, since r_old_dentry's d_parent may have
 535                  * changed between the dir mutex being dropped and
 536                  * this request being freed.
 537                  */
 538                 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
 539                                   CEPH_CAP_PIN);
 540                 dput(req->r_old_dentry);
 541                 iput(req->r_old_dentry_dir);
 542         }
 543         kfree(req->r_path1);
 544         kfree(req->r_path2);
 545         put_request_session(req);
 546         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
 547         kfree(req);
 548 }
 549
 550 /*
 551  * lookup session, bump ref if found.
 552  *
 553  * called under mdsc->mutex.
 554  */
 555 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
 556                                              u64 tid)
 557 {
 558         struct ceph_mds_request *req;
 559         struct rb_node *n = mdsc->request_tree.rb_node;
 560
 561         while (n) {
 562                 req = rb_entry(n, struct ceph_mds_request, r_node);
 563                 if (tid < req->r_tid)
 564                         n = n->rb_left;
 565                 else if (tid > req->r_tid)
 566                         n = n->rb_right;
 567                 else {
 568                         ceph_mdsc_get_request(req);
 569                         return req;
 570                 }
 571         }
 572         return NULL;
 573 }
 574
 575 static void __insert_request(struct ceph_mds_client *mdsc,
 576                              struct ceph_mds_request *new)
 577 {
 578         struct rb_node **p = &mdsc->request_tree.rb_node;
 579         struct rb_node *parent = NULL;
 580         struct ceph_mds_request *req = NULL;
 581
 582         while (*p) {
 583                 parent = *p;
 584                 req = rb_entry(parent, struct ceph_mds_request, r_node);
 585                 if (new->r_tid < req->r_tid)
 586                         p = &(*p)->rb_left;
 587                 else if (new->r_tid > req->r_tid)
 588                         p = &(*p)->rb_right;
 589                 else
 590                         BUG();
 591         }
 592
 593         rb_link_node(&new->r_node, parent, p);
 594         rb_insert_color(&new->r_node, &mdsc->request_tree);
 595 }
 596
 597 /*
 598  * Register an in-flight request, and assign a tid.  Link to directory
 599  * are modifying (if any).
 600  *
 601  * Called under mdsc->mutex.
 602  */
 603 static void __register_request(struct ceph_mds_client *mdsc,
 604                                struct ceph_mds_request *req,
 605                                struct inode *dir)
 606 {
 607         req->r_tid = ++mdsc->last_tid;
 608         if (req->r_num_caps)
 609                 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
 610                                   req->r_num_caps);
 611         dout("__register_request %p tid %lld\n", req, req->r_tid);
 612         ceph_mdsc_get_request(req);
 613         __insert_request(mdsc, req);
 614
 615         req->r_uid = current_fsuid();
 616         req->r_gid = current_fsgid();
 617
 618         if (dir) {
 619                 struct ceph_inode_info *ci = ceph_inode(dir);
 620
 621                 ihold(dir);
 622                 spin_lock(&ci->i_unsafe_lock);
 623                 req->r_unsafe_dir = dir;
 624                 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
 625                 spin_unlock(&ci->i_unsafe_lock);
 626         }
 627 }
 628
 629 static void __unregister_request(struct ceph_mds_client *mdsc,
 630                                  struct ceph_mds_request *req)
 631 {
 632         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 633         rb_erase(&req->r_node, &mdsc->request_tree);
 634         RB_CLEAR_NODE(&req->r_node);
 635
 636         if (req->r_unsafe_dir) {
 637                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 638
 639                 spin_lock(&ci->i_unsafe_lock);
 640                 list_del_init(&req->r_unsafe_dir_item);
 641                 spin_unlock(&ci->i_unsafe_lock);
 642
 643                 iput(req->r_unsafe_dir);
 644                 req->r_unsafe_dir = NULL;
 645         }
 646
 647         ceph_mdsc_put_request(req);
 648 }
 649
 650 /*
 651  * Choose mds to send request to next.  If there is a hint set in the
 652  * request (e.g., due to a prior forward hint from the mds), use that.
 653  * Otherwise, consult frag tree and/or caps to identify the
 654  * appropriate mds.  If all else fails, choose randomly.
 655  *
 656  * Called under mdsc->mutex.
 657  */
 658 static struct dentry *get_nonsnap_parent(struct dentry *dentry)
 659 {
 660         /*
 661          * we don't need to worry about protecting the d_parent access
 662          * here because we never renaming inside the snapped namespace
 663          * except to resplice to another snapdir, and either the old or new
 664          * result is a valid result.
 665          */
 666         while (!IS_ROOT(dentry) && ceph_snap(dentry->d_inode) != CEPH_NOSNAP)
 667                 dentry = dentry->d_parent;
 668         return dentry;
 669 }
 670
 671 static int __choose_mds(struct ceph_mds_client *mdsc,
 672                         struct ceph_mds_request *req)
 673 {
 674         struct inode *inode;
 675         struct ceph_inode_info *ci;
 676         struct ceph_cap *cap;
 677         int mode = req->r_direct_mode;
 678         int mds = -1;
 679         u32 hash = req->r_direct_hash;
 680         bool is_hash = req->r_direct_is_hash;
 681
 682         /*
 683          * is there a specific mds we should try?  ignore hint if we have
 684          * no session and the mds is not up (active or recovering).
 685          */
 686         if (req->r_resend_mds >= 0 &&
 687             (__have_session(mdsc, req->r_resend_mds) ||
 688              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 689                 dout("choose_mds using resend_mds mds%d\n",
 690                      req->r_resend_mds);
 691                 return req->r_resend_mds;
 692         }
 693
 694         if (mode == USE_RANDOM_MDS)
 695                 goto random;
 696
 697         inode = NULL;
 698         if (req->r_inode) {
 699                 inode = req->r_inode;
 700         } else if (req->r_dentry) {
 701                 /* ignore race with rename; old or new d_parent is okay */
 702                 struct dentry *parent = req->r_dentry->d_parent;
 703                 struct inode *dir = parent->d_inode;
 704
 705                 if (dir->i_sb != mdsc->fsc->sb) {
 706                         /* not this fs! */
 707                         inode = req->r_dentry->d_inode;
 708                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
 709                         /* direct snapped/virtual snapdir requests
 710                          * based on parent dir inode */
 711                         struct dentry *dn = get_nonsnap_parent(parent);
 712                         inode = dn->d_inode;
 713                         dout("__choose_mds using nonsnap parent %p\n", inode);
 714                 } else if (req->r_dentry->d_inode) {
 715                         /* dentry target */
 716                         inode = req->r_dentry->d_inode;
 717                 } else {
 718                         /* dir + name */
 719                         inode = dir;
 720                         hash = ceph_dentry_hash(dir, req->r_dentry);
 721                         is_hash = true;
 722                 }
 723         }
 724
 725         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 726              (int)hash, mode);
 727         if (!inode)
 728                 goto random;
 729         ci = ceph_inode(inode);
 730
 731         if (is_hash && S_ISDIR(inode->i_mode)) {
 732                 struct ceph_inode_frag frag;
 733                 int found;
 734
 735                 ceph_choose_frag(ci, hash, &frag, &found);
 736                 if (found) {
 737                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 738                                 u8 r;
 739
 740                                 /* choose a random replica */
 741                                 get_random_bytes(&r, 1);
 742                                 r %= frag.ndist;
 743                                 mds = frag.dist[r];
 744                                 dout("choose_mds %p %llx.%llx "
 745                                      "frag %u mds%d (%d/%d)\n",
 746                                      inode, ceph_vinop(inode),
 747                                      frag.frag, mds,
 748                                      (int)r, frag.ndist);
 749                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 750                                     CEPH_MDS_STATE_ACTIVE)
 751                                         return mds;
 752                         }
 753
 754                         /* since this file/dir wasn't known to be
 755                          * replicated, then we want to look for the
 756                          * authoritative mds. */
 757                         mode = USE_AUTH_MDS;
 758                         if (frag.mds >= 0) {
 759                                 /* choose auth mds */
 760                                 mds = frag.mds;
 761                                 dout("choose_mds %p %llx.%llx "
 762                                      "frag %u mds%d (auth)\n",
 763                                      inode, ceph_vinop(inode), frag.frag, mds);
 764                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 765                                     CEPH_MDS_STATE_ACTIVE)
 766                                         return mds;
 767                         }
 768                 }
 769         }
 770
 771         spin_lock(&ci->i_ceph_lock);
 772         cap = NULL;
 773         if (mode == USE_AUTH_MDS)
 774                 cap = ci->i_auth_cap;
 775         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
 776                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
 777         if (!cap) {
 778                 spin_unlock(&ci->i_ceph_lock);
 779                 goto random;
 780         }
 781         mds = cap->session->s_mds;
 782         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
 783              inode, ceph_vinop(inode), mds,
 784              cap == ci->i_auth_cap ? "auth " : "", cap);
 785         spin_unlock(&ci->i_ceph_lock);
 786         return mds;
 787
 788 random:
 789         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
 790         dout("choose_mds chose random mds%d\n", mds);
 791         return mds;
 792 }
 793
 794
 795 /*
 796  * session messages
 797  */
 798 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
 799 {
 800         struct ceph_msg *msg;
 801         struct ceph_mds_session_head *h;
 802
 803         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
 804                            false);
 805         if (!msg) {
 806                 pr_err("create_session_msg ENOMEM creating msg\n");
 807                 return NULL;
 808         }
 809         h = msg->front.iov_base;
 810         h->op = cpu_to_le32(op);
 811         h->seq = cpu_to_le64(seq);
 812         return msg;
 813 }
 814
 815 /*
 816  * send session open request.
 817  *
 818  * called under mdsc->mutex
 819  */
 820 static int __open_session(struct ceph_mds_client *mdsc,
 821                           struct ceph_mds_session *session)
 822 {
 823         struct ceph_msg *msg;
 824         int mstate;
 825         int mds = session->s_mds;
 826
 827         /* wait for mds to go active? */
 828         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
 829         dout("open_session to mds%d (%s)\n", mds,
 830              ceph_mds_state_name(mstate));
 831         session->s_state = CEPH_MDS_SESSION_OPENING;
 832         session->s_renew_requested = jiffies;
 833
 834         /* send connect message */
 835         msg = create_session_msg(CEPH_SESSION_REQUEST_OPEN, session->s_seq);
 836         if (!msg)
 837                 return -ENOMEM;
 838         ceph_con_send(&session->s_con, msg);
 839         return 0;
 840 }
 841
 842 /*
 843  * open sessions for any export targets for the given mds
 844  *
 845  * called under mdsc->mutex
 846  */
 847 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
 848                                           struct ceph_mds_session *session)
 849 {
 850         struct ceph_mds_info *mi;
 851         struct ceph_mds_session *ts;
 852         int i, mds = session->s_mds;
 853         int target;
 854
 855         if (mds >= mdsc->mdsmap->m_max_mds)
 856                 return;
 857         mi = &mdsc->mdsmap->m_info[mds];
 858         dout("open_export_target_sessions for mds%d (%d targets)\n",
 859              session->s_mds, mi->num_export_targets);
 860
 861         for (i = 0; i < mi->num_export_targets; i++) {
 862                 target = mi->export_targets[i];
 863                 ts = __ceph_lookup_mds_session(mdsc, target);
 864                 if (!ts) {
 865                         ts = register_session(mdsc, target);
 866                         if (IS_ERR(ts))
 867                                 return;
 868                 }
 869                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
 870                     session->s_state == CEPH_MDS_SESSION_CLOSING)
 871                         __open_session(mdsc, session);
 872                 else
 873                         dout(" mds%d target mds%d %p is %s\n", session->s_mds,
 874                              i, ts, session_state_name(ts->s_state));
 875                 ceph_put_mds_session(ts);
 876         }
 877 }
 878
 879 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
 880                                            struct ceph_mds_session *session)
 881 {
 882         mutex_lock(&mdsc->mutex);
 883         __open_export_target_sessions(mdsc, session);
 884         mutex_unlock(&mdsc->mutex);
 885 }
 886
 887 /*
 888  * session caps
 889  */
 890
 891 /*
 892  * Free preallocated cap messages assigned to this session
 893  */
 894 static void cleanup_cap_releases(struct ceph_mds_session *session)
 895 {
 896         struct ceph_msg *msg;
 897
 898         spin_lock(&session->s_cap_lock);
 899         while (!list_empty(&session->s_cap_releases)) {
 900                 msg = list_first_entry(&session->s_cap_releases,
 901                                        struct ceph_msg, list_head);
 902                 list_del_init(&msg->list_head);
 903                 ceph_msg_put(msg);
 904         }
 905         while (!list_empty(&session->s_cap_releases_done)) {
 906                 msg = list_first_entry(&session->s_cap_releases_done,
 907                                        struct ceph_msg, list_head);
 908                 list_del_init(&msg->list_head);
 909                 ceph_msg_put(msg);
 910         }
 911         spin_unlock(&session->s_cap_lock);
 912 }
 913
 914 /*
 915  * Helper to safely iterate over all caps associated with a session, with
 916  * special care taken to handle a racing __ceph_remove_cap().
 917  *
 918  * Caller must hold session s_mutex.
 919  */
 920 static int iterate_session_caps(struct ceph_mds_session *session,
 921                                  int (*cb)(struct inode *, struct ceph_cap *,
 922                                             void *), void *arg)
 923 {
 924         struct list_head *p;
 925         struct ceph_cap *cap;
 926         struct inode *inode, *last_inode = NULL;
 927         struct ceph_cap *old_cap = NULL;
 928         int ret;
 929
 930         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
 931         spin_lock(&session->s_cap_lock);
 932         p = session->s_caps.next;
 933         while (p != &session->s_caps) {
 934                 cap = list_entry(p, struct ceph_cap, session_caps);
 935                 inode = igrab(&cap->ci->vfs_inode);
 936                 if (!inode) {
 937                         p = p->next;
 938                         continue;
 939                 }
 940                 session->s_cap_iterator = cap;
 941                 spin_unlock(&session->s_cap_lock);
 942
 943                 if (last_inode) {
 944                         iput(last_inode);
 945                         last_inode = NULL;
 946                 }
 947                 if (old_cap) {
 948                         ceph_put_cap(session->s_mdsc, old_cap);
 949                         old_cap = NULL;
 950                 }
 951
 952                 ret = cb(inode, cap, arg);
 953                 last_inode = inode;
 954
 955                 spin_lock(&session->s_cap_lock);
 956                 p = p->next;
 957                 if (cap->ci == NULL) {
 958                         dout("iterate_session_caps  finishing cap %p removal\n",
 959                              cap);
 960                         BUG_ON(cap->session != session);
 961                         list_del_init(&cap->session_caps);
 962                         session->s_nr_caps--;
 963                         cap->session = NULL;
 964                         old_cap = cap;  /* put_cap it w/o locks held */
 965                 }
 966                 if (ret < 0)
 967                         goto out;
 968         }
 969         ret = 0;
 970 out:
 971         session->s_cap_iterator = NULL;
 972         spin_unlock(&session->s_cap_lock);
 973
 974         if (last_inode)
 975                 iput(last_inode);
 976         if (old_cap)
 977                 ceph_put_cap(session->s_mdsc, old_cap);
 978
 979         return ret;
 980 }
 981
 982 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
 983                                   void *arg)
 984 {
 985         struct ceph_inode_info *ci = ceph_inode(inode);
 986         int drop = 0;
 987
 988         dout("removing cap %p, ci is %p, inode is %p\n",
 989              cap, ci, &ci->vfs_inode);
 990         spin_lock(&ci->i_ceph_lock);
 991         __ceph_remove_cap(cap, false);
 992         if (!__ceph_is_any_real_caps(ci)) {
 993                 struct ceph_mds_client *mdsc =
 994                         ceph_sb_to_client(inode->i_sb)->mdsc;
 995
 996                 spin_lock(&mdsc->cap_dirty_lock);
 997                 if (!list_empty(&ci->i_dirty_item)) {
 998                         pr_info(" dropping dirty %s state for %p %lld\n",
 999                                 ceph_cap_string(ci->i_dirty_caps),
1000                                 inode, ceph_ino(inode));
1001                         ci->i_dirty_caps = 0;
1002                         list_del_init(&ci->i_dirty_item);
1003                         drop = 1;
1004                 }
1005                 if (!list_empty(&ci->i_flushing_item)) {
1006                         pr_info(" dropping dirty+flushing %s state for %p %lld\n",
1007                                 ceph_cap_string(ci->i_flushing_caps),
1008                                 inode, ceph_ino(inode));
1009                         ci->i_flushing_caps = 0;
1010                         list_del_init(&ci->i_flushing_item);
1011                         mdsc->num_cap_flushing--;
1012                         drop = 1;
1013                 }
1014                 if (drop && ci->i_wrbuffer_ref) {
1015                         pr_info(" dropping dirty data for %p %lld\n",
1016                                 inode, ceph_ino(inode));
1017                         ci->i_wrbuffer_ref = 0;
1018                         ci->i_wrbuffer_ref_head = 0;
1019                         drop++;
1020                 }
1021                 spin_unlock(&mdsc->cap_dirty_lock);
1022         }
1023         spin_unlock(&ci->i_ceph_lock);
1024         while (drop--)
1025                 iput(inode);
1026         return 0;
1027 }
1028
1029 /*
1030  * caller must hold session s_mutex
1031  */
1032 static void remove_session_caps(struct ceph_mds_session *session)
1033 {
1034         dout("remove_session_caps on %p\n", session);
1035         iterate_session_caps(session, remove_session_caps_cb, NULL);
1036
1037         spin_lock(&session->s_cap_lock);
1038         if (session->s_nr_caps > 0) {
1039                 struct super_block *sb = session->s_mdsc->fsc->sb;
1040                 struct inode *inode;
1041                 struct ceph_cap *cap, *prev = NULL;
1042                 struct ceph_vino vino;
1043                 /*
1044                  * iterate_session_caps() skips inodes that are being
1045                  * deleted, we need to wait until deletions are complete.
1046                  * __wait_on_freeing_inode() is designed for the job,
1047                  * but it is not exported, so use lookup inode function
1048                  * to access it.
1049                  */
1050                 while (!list_empty(&session->s_caps)) {
1051                         cap = list_entry(session->s_caps.next,
1052                                          struct ceph_cap, session_caps);
1053                         if (cap == prev)
1054                                 break;
1055                         prev = cap;
1056                         vino = cap->ci->i_vino;
1057                         spin_unlock(&session->s_cap_lock);
1058
1059                         inode = ceph_find_inode(sb, vino);
1060                         iput(inode);
1061
1062                         spin_lock(&session->s_cap_lock);
1063                 }
1064         }
1065         spin_unlock(&session->s_cap_lock);
1066
1067         BUG_ON(session->s_nr_caps > 0);
1068         BUG_ON(!list_empty(&session->s_cap_flushing));
1069         cleanup_cap_releases(session);
1070 }
1071
1072 /*
1073  * wake up any threads waiting on this session's caps.  if the cap is
1074  * old (didn't get renewed on the client reconnect), remove it now.
1075  *
1076  * caller must hold s_mutex.
1077  */
1078 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1079                               void *arg)
1080 {
1081         struct ceph_inode_info *ci = ceph_inode(inode);
1082
1083         wake_up_all(&ci->i_cap_wq);
1084         if (arg) {
1085                 spin_lock(&ci->i_ceph_lock);
1086                 ci->i_wanted_max_size = 0;
1087                 ci->i_requested_max_size = 0;
1088                 spin_unlock(&ci->i_ceph_lock);
1089         }
1090         return 0;
1091 }
1092
1093 static void wake_up_session_caps(struct ceph_mds_session *session,
1094                                  int reconnect)
1095 {
1096         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1097         iterate_session_caps(session, wake_up_session_cb,
1098                              (void *)(unsigned long)reconnect);
1099 }
1100
1101 /*
1102  * Send periodic message to MDS renewing all currently held caps.  The
1103  * ack will reset the expiration for all caps from this session.
1104  *
1105  * caller holds s_mutex
1106  */
1107 static int send_renew_caps(struct ceph_mds_client *mdsc,
1108                            struct ceph_mds_session *session)
1109 {
1110         struct ceph_msg *msg;
1111         int state;
1112
1113         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1114             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1115                 pr_info("mds%d caps stale\n", session->s_mds);
1116         session->s_renew_requested = jiffies;
1117
1118         /* do not try to renew caps until a recovering mds has reconnected
1119          * with its clients. */
1120         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1121         if (state < CEPH_MDS_STATE_RECONNECT) {
1122                 dout("send_renew_caps ignoring mds%d (%s)\n",
1123                      session->s_mds, ceph_mds_state_name(state));
1124                 return 0;
1125         }
1126
1127         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1128                 ceph_mds_state_name(state));
1129         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1130                                  ++session->s_renew_seq);
1131         if (!msg)
1132                 return -ENOMEM;
1133         ceph_con_send(&session->s_con, msg);
1134         return 0;
1135 }
1136
1137 /*
1138  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1139  *
1140  * Called under session->s_mutex
1141  */
1142 static void renewed_caps(struct ceph_mds_client *mdsc,
1143                          struct ceph_mds_session *session, int is_renew)
1144 {
1145         int was_stale;
1146         int wake = 0;
1147
1148         spin_lock(&session->s_cap_lock);
1149         was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1150
1151         session->s_cap_ttl = session->s_renew_requested +
1152                 mdsc->mdsmap->m_session_timeout*HZ;
1153
1154         if (was_stale) {
1155                 if (time_before(jiffies, session->s_cap_ttl)) {
1156                         pr_info("mds%d caps renewed\n", session->s_mds);
1157                         wake = 1;
1158                 } else {
1159                         pr_info("mds%d caps still stale\n", session->s_mds);
1160                 }
1161         }
1162         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1163              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1164              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1165         spin_unlock(&session->s_cap_lock);
1166
1167         if (wake)
1168                 wake_up_session_caps(session, 0);
1169 }
1170
1171 /*
1172  * send a session close request
1173  */
1174 static int request_close_session(struct ceph_mds_client *mdsc,
1175                                  struct ceph_mds_session *session)
1176 {
1177         struct ceph_msg *msg;
1178
1179         dout("request_close_session mds%d state %s seq %lld\n",
1180              session->s_mds, session_state_name(session->s_state),
1181              session->s_seq);
1182         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1183         if (!msg)
1184                 return -ENOMEM;
1185         ceph_con_send(&session->s_con, msg);
1186         return 0;
1187 }
1188
1189 /*
1190  * Called with s_mutex held.
1191  */
1192 static int __close_session(struct ceph_mds_client *mdsc,
1193                          struct ceph_mds_session *session)
1194 {
1195         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1196                 return 0;
1197         session->s_state = CEPH_MDS_SESSION_CLOSING;
1198         return request_close_session(mdsc, session);
1199 }
1200
1201 /*
1202  * Trim old(er) caps.
1203  *
1204  * Because we can't cache an inode without one or more caps, we do
1205  * this indirectly: if a cap is unused, we prune its aliases, at which
1206  * point the inode will hopefully get dropped to.
1207  *
1208  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1209  * memory pressure from the MDS, though, so it needn't be perfect.
1210  */
1211 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1212 {
1213         struct ceph_mds_session *session = arg;
1214         struct ceph_inode_info *ci = ceph_inode(inode);
1215         int used, oissued, mine;
1216
1217         if (session->s_trim_caps <= 0)
1218                 return -1;
1219
1220         spin_lock(&ci->i_ceph_lock);
1221         mine = cap->issued | cap->implemented;
1222         used = __ceph_caps_used(ci);
1223         oissued = __ceph_caps_issued_other(ci, cap);
1224
1225         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
1226              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1227              ceph_cap_string(used));
1228         if (ci->i_dirty_caps)
1229                 goto out;   /* dirty caps */
1230         if ((used & ~oissued) & mine)
1231                 goto out;   /* we need these caps */
1232
1233         session->s_trim_caps--;
1234         if (oissued) {
1235                 /* we aren't the only cap.. just remove us */
1236                 __ceph_remove_cap(cap, true);
1237         } else {
1238                 /* try to drop referring dentries */
1239                 spin_unlock(&ci->i_ceph_lock);
1240                 d_prune_aliases(inode);
1241                 dout("trim_caps_cb %p cap %p  pruned, count now %d\n",
1242                      inode, cap, atomic_read(&inode->i_count));
1243                 return 0;
1244         }
1245
1246 out:
1247         spin_unlock(&ci->i_ceph_lock);
1248         return 0;
1249 }
1250
1251 /*
1252  * Trim session cap count down to some max number.
1253  */
1254 static int trim_caps(struct ceph_mds_client *mdsc,
1255                      struct ceph_mds_session *session,
1256                      int max_caps)
1257 {
1258         int trim_caps = session->s_nr_caps - max_caps;
1259
1260         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1261              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1262         if (trim_caps > 0) {
1263                 session->s_trim_caps = trim_caps;
1264                 iterate_session_caps(session, trim_caps_cb, session);
1265                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1266                      session->s_mds, session->s_nr_caps, max_caps,
1267                         trim_caps - session->s_trim_caps);
1268                 session->s_trim_caps = 0;
1269         }
1270         return 0;
1271 }
1272
1273 /*
1274  * Allocate cap_release messages.  If there is a partially full message
1275  * in the queue, try to allocate enough to cover it's remainder, so that
1276  * we can send it immediately.
1277  *
1278  * Called under s_mutex.
1279  */
1280 int ceph_add_cap_releases(struct ceph_mds_client *mdsc,
1281                           struct ceph_mds_session *session)
1282 {
1283         struct ceph_msg *msg, *partial = NULL;
1284         struct ceph_mds_cap_release *head;
1285         int err = -ENOMEM;
1286         int extra = mdsc->fsc->mount_options->cap_release_safety;
1287         int num;
1288
1289         dout("add_cap_releases %p mds%d extra %d\n", session, session->s_mds,
1290              extra);
1291
1292         spin_lock(&session->s_cap_lock);
1293
1294         if (!list_empty(&session->s_cap_releases)) {
1295                 msg = list_first_entry(&session->s_cap_releases,
1296                                        struct ceph_msg,
1297                                  list_head);
1298                 head = msg->front.iov_base;
1299                 num = le32_to_cpu(head->num);
1300                 if (num) {
1301                         dout(" partial %p with (%d/%d)\n", msg, num,
1302                              (int)CEPH_CAPS_PER_RELEASE);
1303                         extra += CEPH_CAPS_PER_RELEASE - num;
1304                         partial = msg;
1305                 }
1306         }
1307         while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1308                 spin_unlock(&session->s_cap_lock);
1309                 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
1310                                    GFP_NOFS, false);
1311                 if (!msg)
1312                         goto out_unlocked;
1313                 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1314                      (int)msg->front.iov_len);
1315                 head = msg->front.iov_base;
1316                 head->num = cpu_to_le32(0);
1317                 msg->front.iov_len = sizeof(*head);
1318                 spin_lock(&session->s_cap_lock);
1319                 list_add(&msg->list_head, &session->s_cap_releases);
1320                 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1321         }
1322
1323         if (partial) {
1324                 head = partial->front.iov_base;
1325                 num = le32_to_cpu(head->num);
1326                 dout(" queueing partial %p with %d/%d\n", partial, num,
1327                      (int)CEPH_CAPS_PER_RELEASE);
1328                 list_move_tail(&partial->list_head,
1329                                &session->s_cap_releases_done);
1330                 session->s_num_cap_releases -= CEPH_CAPS_PER_RELEASE - num;
1331         }
1332         err = 0;
1333         spin_unlock(&session->s_cap_lock);
1334 out_unlocked:
1335         return err;
1336 }
1337
1338 /*
1339  * flush all dirty inode data to disk.
1340  *
1341  * returns true if we've flushed through want_flush_seq
1342  */
1343 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1344 {
1345         int mds, ret = 1;
1346
1347         dout("check_cap_flush want %lld\n", want_flush_seq);
1348         mutex_lock(&mdsc->mutex);
1349         for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1350                 struct ceph_mds_session *session = mdsc->sessions[mds];
1351
1352                 if (!session)
1353                         continue;
1354                 get_session(session);
1355                 mutex_unlock(&mdsc->mutex);
1356
1357                 mutex_lock(&session->s_mutex);
1358                 if (!list_empty(&session->s_cap_flushing)) {
1359                         struct ceph_inode_info *ci =
1360                                 list_entry(session->s_cap_flushing.next,
1361                                            struct ceph_inode_info,
1362                                            i_flushing_item);
1363                         struct inode *inode = &ci->vfs_inode;
1364
1365                         spin_lock(&ci->i_ceph_lock);
1366                         if (ci->i_cap_flush_seq <= want_flush_seq) {
1367                                 dout("check_cap_flush still flushing %p "
1368                                      "seq %lld <= %lld to mds%d\n", inode,
1369                                      ci->i_cap_flush_seq, want_flush_seq,
1370                                      session->s_mds);
1371                                 ret = 0;
1372                         }
1373                         spin_unlock(&ci->i_ceph_lock);
1374                 }
1375                 mutex_unlock(&session->s_mutex);
1376                 ceph_put_mds_session(session);
1377
1378                 if (!ret)
1379                         return ret;
1380                 mutex_lock(&mdsc->mutex);
1381         }
1382
1383         mutex_unlock(&mdsc->mutex);
1384         dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1385         return ret;
1386 }
1387
1388 /*
1389  * called under s_mutex
1390  */
1391 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1392                             struct ceph_mds_session *session)
1393 {
1394         struct ceph_msg *msg;
1395
1396         dout("send_cap_releases mds%d\n", session->s_mds);
1397         spin_lock(&session->s_cap_lock);
1398         while (!list_empty(&session->s_cap_releases_done)) {
1399                 msg = list_first_entry(&session->s_cap_releases_done,
1400                                  struct ceph_msg, list_head);
1401                 list_del_init(&msg->list_head);
1402                 spin_unlock(&session->s_cap_lock);
1403                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1404                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1405                 ceph_con_send(&session->s_con, msg);
1406                 spin_lock(&session->s_cap_lock);
1407         }
1408         spin_unlock(&session->s_cap_lock);
1409 }
1410
1411 static void discard_cap_releases(struct ceph_mds_client *mdsc,
1412                                  struct ceph_mds_session *session)
1413 {
1414         struct ceph_msg *msg;
1415         struct ceph_mds_cap_release *head;
1416         unsigned num;
1417
1418         dout("discard_cap_releases mds%d\n", session->s_mds);
1419
1420         /* zero out the in-progress message */
1421         msg = list_first_entry(&session->s_cap_releases,
1422                                struct ceph_msg, list_head);
1423         head = msg->front.iov_base;
1424         num = le32_to_cpu(head->num);
1425         dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg, num);
1426         head->num = cpu_to_le32(0);
1427         msg->front.iov_len = sizeof(*head);
1428         session->s_num_cap_releases += num;
1429
1430         /* requeue completed messages */
1431         while (!list_empty(&session->s_cap_releases_done)) {
1432                 msg = list_first_entry(&session->s_cap_releases_done,
1433                                  struct ceph_msg, list_head);
1434                 list_del_init(&msg->list_head);
1435
1436                 head = msg->front.iov_base;
1437                 num = le32_to_cpu(head->num);
1438                 dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg,
1439                      num);
1440                 session->s_num_cap_releases += num;
1441                 head->num = cpu_to_le32(0);
1442                 msg->front.iov_len = sizeof(*head);
1443                 list_add(&msg->list_head, &session->s_cap_releases);
1444         }
1445 }
1446
1447 /*
1448  * requests
1449  */
1450
1451 /*
1452  * Create an mds request.
1453  */
1454 struct ceph_mds_request *
1455 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1456 {
1457         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1458
1459         if (!req)
1460                 return ERR_PTR(-ENOMEM);
1461
1462         mutex_init(&req->r_fill_mutex);
1463         req->r_mdsc = mdsc;
1464         req->r_started = jiffies;
1465         req->r_resend_mds = -1;
1466         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1467         req->r_fmode = -1;
1468         kref_init(&req->r_kref);
1469         INIT_LIST_HEAD(&req->r_wait);
1470         init_completion(&req->r_completion);
1471         init_completion(&req->r_safe_completion);
1472         INIT_LIST_HEAD(&req->r_unsafe_item);
1473
1474         req->r_op = op;
1475         req->r_direct_mode = mode;
1476         return req;
1477 }
1478
1479 /*
1480  * return oldest (lowest) request, tid in request tree, 0 if none.
1481  *
1482  * called under mdsc->mutex.
1483  */
1484 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1485 {
1486         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1487                 return NULL;
1488         return rb_entry(rb_first(&mdsc->request_tree),
1489                         struct ceph_mds_request, r_node);
1490 }
1491
1492 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1493 {
1494         struct ceph_mds_request *req = __get_oldest_req(mdsc);
1495
1496         if (req)
1497                 return req->r_tid;
1498         return 0;
1499 }
1500
1501 /*
1502  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1503  * on build_path_from_dentry in fs/cifs/dir.c.
1504  *
1505  * If @stop_on_nosnap, generate path relative to the first non-snapped
1506  * inode.
1507  *
1508  * Encode hidden .snap dirs as a double /, i.e.
1509  *   foo/.snap/bar -> foo//bar
1510  */
1511 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1512                            int stop_on_nosnap)
1513 {
1514         struct dentry *temp;
1515         char *path;
1516         int len, pos;
1517         unsigned seq;
1518
1519         if (dentry == NULL)
1520                 return ERR_PTR(-EINVAL);
1521
1522 retry:
1523         len = 0;
1524         seq = read_seqbegin(&rename_lock);
1525         rcu_read_lock();
1526         for (temp = dentry; !IS_ROOT(temp);) {
1527                 struct inode *inode = temp->d_inode;
1528                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1529                         len++;  /* slash only */
1530                 else if (stop_on_nosnap && inode &&
1531                          ceph_snap(inode) == CEPH_NOSNAP)
1532                         break;
1533                 else
1534                         len += 1 + temp->d_name.len;
1535                 temp = temp->d_parent;
1536         }
1537         rcu_read_unlock();
1538         if (len)
1539                 len--;  /* no leading '/' */
1540
1541         path = kmalloc(len+1, GFP_NOFS);
1542         if (path == NULL)
1543                 return ERR_PTR(-ENOMEM);
1544         pos = len;
1545         path[pos] = 0;  /* trailing null */
1546         rcu_read_lock();
1547         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1548                 struct inode *inode;
1549
1550                 spin_lock(&temp->d_lock);
1551                 inode = temp->d_inode;
1552                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1553                         dout("build_path path+%d: %p SNAPDIR\n",
1554                              pos, temp);
1555                 } else if (stop_on_nosnap && inode &&
1556                            ceph_snap(inode) == CEPH_NOSNAP) {
1557                         spin_unlock(&temp->d_lock);
1558                         break;
1559                 } else {
1560                         pos -= temp->d_name.len;
1561                         if (pos < 0) {
1562                                 spin_unlock(&temp->d_lock);
1563                                 break;
1564                         }
1565                         strncpy(path + pos, temp->d_name.name,
1566                                 temp->d_name.len);
1567                 }
1568                 spin_unlock(&temp->d_lock);
1569                 if (pos)
1570                         path[--pos] = '/';
1571                 temp = temp->d_parent;
1572         }
1573         rcu_read_unlock();
1574         if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1575                 pr_err("build_path did not end path lookup where "
1576                        "expected, namelen is %d, pos is %d\n", len, pos);
1577                 /* presumably this is only possible if racing with a
1578                    rename of one of the parent directories (we can not
1579                    lock the dentries above us to prevent this, but
1580                    retrying should be harmless) */
1581                 kfree(path);
1582                 goto retry;
1583         }
1584
1585         *base = ceph_ino(temp->d_inode);
1586         *plen = len;
1587         dout("build_path on %p %d built %llx '%.*s'\n",
1588              dentry, d_count(dentry), *base, len, path);
1589         return path;
1590 }
1591
1592 static int build_dentry_path(struct dentry *dentry,
1593                              const char **ppath, int *ppathlen, u64 *pino,
1594                              int *pfreepath)
1595 {
1596         char *path;
1597
1598         if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1599                 *pino = ceph_ino(dentry->d_parent->d_inode);
1600                 *ppath = dentry->d_name.name;
1601                 *ppathlen = dentry->d_name.len;
1602                 return 0;
1603         }
1604         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1605         if (IS_ERR(path))
1606                 return PTR_ERR(path);
1607         *ppath = path;
1608         *pfreepath = 1;
1609         return 0;
1610 }
1611
1612 static int build_inode_path(struct inode *inode,
1613                             const char **ppath, int *ppathlen, u64 *pino,
1614                             int *pfreepath)
1615 {
1616         struct dentry *dentry;
1617         char *path;
1618
1619         if (ceph_snap(inode) == CEPH_NOSNAP) {
1620                 *pino = ceph_ino(inode);
1621                 *ppathlen = 0;
1622                 return 0;
1623         }
1624         dentry = d_find_alias(inode);
1625         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1626         dput(dentry);
1627         if (IS_ERR(path))
1628                 return PTR_ERR(path);
1629         *ppath = path;
1630         *pfreepath = 1;
1631         return 0;
1632 }
1633
1634 /*
1635  * request arguments may be specified via an inode *, a dentry *, or
1636  * an explicit ino+path.
1637  */
1638 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1639                                   const char *rpath, u64 rino,
1640                                   const char **ppath, int *pathlen,
1641                                   u64 *ino, int *freepath)
1642 {
1643         int r = 0;
1644
1645         if (rinode) {
1646                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1647                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1648                      ceph_snap(rinode));
1649         } else if (rdentry) {
1650                 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1651                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1652                      *ppath);
1653         } else if (rpath || rino) {
1654                 *ino = rino;
1655                 *ppath = rpath;
1656                 *pathlen = rpath ? strlen(rpath) : 0;
1657                 dout(" path %.*s\n", *pathlen, rpath);
1658         }
1659
1660         return r;
1661 }
1662
1663 /*
1664  * called under mdsc->mutex
1665  */
1666 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1667                                                struct ceph_mds_request *req,
1668                                                int mds)
1669 {
1670         struct ceph_msg *msg;
1671         struct ceph_mds_request_head *head;
1672         const char *path1 = NULL;
1673         const char *path2 = NULL;
1674         u64 ino1 = 0, ino2 = 0;
1675         int pathlen1 = 0, pathlen2 = 0;
1676         int freepath1 = 0, freepath2 = 0;
1677         int len;
1678         u16 releases;
1679         void *p, *end;
1680         int ret;
1681
1682         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1683                               req->r_path1, req->r_ino1.ino,
1684                               &path1, &pathlen1, &ino1, &freepath1);
1685         if (ret < 0) {
1686                 msg = ERR_PTR(ret);
1687                 goto out;
1688         }
1689
1690         ret = set_request_path_attr(NULL, req->r_old_dentry,
1691                               req->r_path2, req->r_ino2.ino,
1692                               &path2, &pathlen2, &ino2, &freepath2);
1693         if (ret < 0) {
1694                 msg = ERR_PTR(ret);
1695                 goto out_free1;
1696         }
1697
1698         len = sizeof(*head) +
1699                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64));
1700
1701         /* calculate (max) length for cap releases */
1702         len += sizeof(struct ceph_mds_request_release) *
1703                 (!!req->r_inode_drop + !!req->r_dentry_drop +
1704                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1705         if (req->r_dentry_drop)
1706                 len += req->r_dentry->d_name.len;
1707         if (req->r_old_dentry_drop)
1708                 len += req->r_old_dentry->d_name.len;
1709
1710         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1711         if (!msg) {
1712                 msg = ERR_PTR(-ENOMEM);
1713                 goto out_free2;
1714         }
1715
1716         msg->hdr.tid = cpu_to_le64(req->r_tid);
1717
1718         head = msg->front.iov_base;
1719         p = msg->front.iov_base + sizeof(*head);
1720         end = msg->front.iov_base + msg->front.iov_len;
1721
1722         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1723         head->op = cpu_to_le32(req->r_op);
1724         head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1725         head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1726         head->args = req->r_args;
1727
1728         ceph_encode_filepath(&p, end, ino1, path1);
1729         ceph_encode_filepath(&p, end, ino2, path2);
1730
1731         /* make note of release offset, in case we need to replay */
1732         req->r_request_release_offset = p - msg->front.iov_base;
1733
1734         /* cap releases */
1735         releases = 0;
1736         if (req->r_inode_drop)
1737                 releases += ceph_encode_inode_release(&p,
1738                       req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1739                       mds, req->r_inode_drop, req->r_inode_unless, 0);
1740         if (req->r_dentry_drop)
1741                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1742                        mds, req->r_dentry_drop, req->r_dentry_unless);
1743         if (req->r_old_dentry_drop)
1744                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1745                        mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1746         if (req->r_old_inode_drop)
1747                 releases += ceph_encode_inode_release(&p,
1748                       req->r_old_dentry->d_inode,
1749                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1750         head->num_releases = cpu_to_le16(releases);
1751
1752         BUG_ON(p > end);
1753         msg->front.iov_len = p - msg->front.iov_base;
1754         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1755
1756         if (req->r_data_len) {
1757                 /* outbound data set only by ceph_sync_setxattr() */
1758                 BUG_ON(!req->r_pages);
1759                 ceph_msg_data_add_pages(msg, req->r_pages, req->r_data_len, 0);
1760         }
1761
1762         msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1763         msg->hdr.data_off = cpu_to_le16(0);
1764
1765 out_free2:
1766         if (freepath2)
1767                 kfree((char *)path2);
1768 out_free1:
1769         if (freepath1)
1770                 kfree((char *)path1);
1771 out:
1772         return msg;
1773 }
1774
1775 /*
1776  * called under mdsc->mutex if error, under no mutex if
1777  * success.
1778  */
1779 static void complete_request(struct ceph_mds_client *mdsc,
1780                              struct ceph_mds_request *req)
1781 {
1782         if (req->r_callback)
1783                 req->r_callback(mdsc, req);
1784         else
1785                 complete_all(&req->r_completion);
1786 }
1787
1788 /*
1789  * called under mdsc->mutex
1790  */
1791 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1792                                   struct ceph_mds_request *req,
1793                                   int mds)
1794 {
1795         struct ceph_mds_request_head *rhead;
1796         struct ceph_msg *msg;
1797         int flags = 0;
1798
1799         req->r_attempts++;
1800         if (req->r_inode) {
1801                 struct ceph_cap *cap =
1802                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
1803
1804                 if (cap)
1805                         req->r_sent_on_mseq = cap->mseq;
1806                 else
1807                         req->r_sent_on_mseq = -1;
1808         }
1809         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1810              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1811
1812         if (req->r_got_unsafe) {
1813                 /*
1814                  * Replay.  Do not regenerate message (and rebuild
1815                  * paths, etc.); just use the original message.
1816                  * Rebuilding paths will break for renames because
1817                  * d_move mangles the src name.
1818                  */
1819                 msg = req->r_request;
1820                 rhead = msg->front.iov_base;
1821
1822                 flags = le32_to_cpu(rhead->flags);
1823                 flags |= CEPH_MDS_FLAG_REPLAY;
1824                 rhead->flags = cpu_to_le32(flags);
1825
1826                 if (req->r_target_inode)
1827                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1828
1829                 rhead->num_retry = req->r_attempts - 1;
1830
1831                 /* remove cap/dentry releases from message */
1832                 rhead->num_releases = 0;
1833                 msg->hdr.front_len = cpu_to_le32(req->r_request_release_offset);
1834                 msg->front.iov_len = req->r_request_release_offset;
1835                 return 0;
1836         }
1837
1838         if (req->r_request) {
1839                 ceph_msg_put(req->r_request);
1840                 req->r_request = NULL;
1841         }
1842         msg = create_request_message(mdsc, req, mds);
1843         if (IS_ERR(msg)) {
1844                 req->r_err = PTR_ERR(msg);
1845                 complete_request(mdsc, req);
1846                 return PTR_ERR(msg);
1847         }
1848         req->r_request = msg;
1849
1850         rhead = msg->front.iov_base;
1851         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1852         if (req->r_got_unsafe)
1853                 flags |= CEPH_MDS_FLAG_REPLAY;
1854         if (req->r_locked_dir)
1855                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1856         rhead->flags = cpu_to_le32(flags);
1857         rhead->num_fwd = req->r_num_fwd;
1858         rhead->num_retry = req->r_attempts - 1;
1859         rhead->ino = 0;
1860
1861         dout(" r_locked_dir = %p\n", req->r_locked_dir);
1862         return 0;
1863 }
1864
1865 /*
1866  * send request, or put it on the appropriate wait list.
1867  */
1868 static int __do_request(struct ceph_mds_client *mdsc,
1869                         struct ceph_mds_request *req)
1870 {
1871         struct ceph_mds_session *session = NULL;
1872         int mds = -1;
1873         int err = -EAGAIN;
1874
1875         if (req->r_err || req->r_got_result)
1876                 goto out;
1877
1878         if (req->r_timeout &&
1879             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1880                 dout("do_request timed out\n");
1881                 err = -EIO;
1882                 goto finish;
1883         }
1884
1885         put_request_session(req);
1886
1887         mds = __choose_mds(mdsc, req);
1888         if (mds < 0 ||
1889             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1890                 dout("do_request no mds or not active, waiting for map\n");
1891                 list_add(&req->r_wait, &mdsc->waiting_for_map);
1892                 goto out;
1893         }
1894
1895         /* get, open session */
1896         session = __ceph_lookup_mds_session(mdsc, mds);
1897         if (!session) {
1898                 session = register_session(mdsc, mds);
1899                 if (IS_ERR(session)) {
1900                         err = PTR_ERR(session);
1901                         goto finish;
1902                 }
1903         }
1904         req->r_session = get_session(session);
1905
1906         dout("do_request mds%d session %p state %s\n", mds, session,
1907              session_state_name(session->s_state));
1908         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1909             session->s_state != CEPH_MDS_SESSION_HUNG) {
1910                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1911                     session->s_state == CEPH_MDS_SESSION_CLOSING)
1912                         __open_session(mdsc, session);
1913                 list_add(&req->r_wait, &session->s_waiting);
1914                 goto out_session;
1915         }
1916
1917         /* send request */
1918         req->r_resend_mds = -1;   /* forget any previous mds hint */
1919
1920         if (req->r_request_started == 0)   /* note request start time */
1921                 req->r_request_started = jiffies;
1922
1923         err = __prepare_send_request(mdsc, req, mds);
1924         if (!err) {
1925                 ceph_msg_get(req->r_request);
1926                 ceph_con_send(&session->s_con, req->r_request);
1927         }
1928
1929 out_session:
1930         ceph_put_mds_session(session);
1931 out:
1932         return err;
1933
1934 finish:
1935         req->r_err = err;
1936         complete_request(mdsc, req);
1937         goto out;
1938 }
1939
1940 /*
1941  * called under mdsc->mutex
1942  */
1943 static void __wake_requests(struct ceph_mds_client *mdsc,
1944                             struct list_head *head)
1945 {
1946         struct ceph_mds_request *req;
1947         LIST_HEAD(tmp_list);
1948
1949         list_splice_init(head, &tmp_list);
1950
1951         while (!list_empty(&tmp_list)) {
1952                 req = list_entry(tmp_list.next,
1953                                  struct ceph_mds_request, r_wait);
1954                 list_del_init(&req->r_wait);
1955                 dout(" wake request %p tid %llu\n", req, req->r_tid);
1956                 __do_request(mdsc, req);
1957         }
1958 }
1959
1960 /*
1961  * Wake up threads with requests pending for @mds, so that they can
1962  * resubmit their requests to a possibly different mds.
1963  */
1964 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
1965 {
1966         struct ceph_mds_request *req;
1967         struct rb_node *p;
1968
1969         dout("kick_requests mds%d\n", mds);
1970         for (p = rb_first(&mdsc->request_tree); p; p = rb_next(p)) {
1971                 req = rb_entry(p, struct ceph_mds_request, r_node);
1972                 if (req->r_got_unsafe)
1973                         continue;
1974                 if (req->r_session &&
1975                     req->r_session->s_mds == mds) {
1976                         dout(" kicking tid %llu\n", req->r_tid);
1977                         __do_request(mdsc, req);
1978                 }
1979         }
1980 }
1981
1982 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
1983                               struct ceph_mds_request *req)
1984 {
1985         dout("submit_request on %p\n", req);
1986         mutex_lock(&mdsc->mutex);
1987         __register_request(mdsc, req, NULL);
1988         __do_request(mdsc, req);
1989         mutex_unlock(&mdsc->mutex);
1990 }
1991
1992 /*
1993  * Synchrously perform an mds request.  Take care of all of the
1994  * session setup, forwarding, retry details.
1995  */
1996 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
1997                          struct inode *dir,
1998                          struct ceph_mds_request *req)
1999 {
2000         int err;
2001
2002         dout("do_request on %p\n", req);
2003
2004         /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
2005         if (req->r_inode)
2006                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2007         if (req->r_locked_dir)
2008                 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
2009         if (req->r_old_dentry)
2010                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2011                                   CEPH_CAP_PIN);
2012
2013         /* issue */
2014         mutex_lock(&mdsc->mutex);
2015         __register_request(mdsc, req, dir);
2016         __do_request(mdsc, req);
2017
2018         if (req->r_err) {
2019                 err = req->r_err;
2020                 __unregister_request(mdsc, req);
2021                 dout("do_request early error %d\n", err);
2022                 goto out;
2023         }
2024
2025         /* wait */
2026         mutex_unlock(&mdsc->mutex);
2027         dout("do_request waiting\n");
2028         if (req->r_timeout) {
2029                 err = (long)wait_for_completion_killable_timeout(
2030                         &req->r_completion, req->r_timeout);
2031                 if (err == 0)
2032                         err = -EIO;
2033         } else {
2034                 err = wait_for_completion_killable(&req->r_completion);
2035         }
2036         dout("do_request waited, got %d\n", err);
2037         mutex_lock(&mdsc->mutex);
2038
2039         /* only abort if we didn't race with a real reply */
2040         if (req->r_got_result) {
2041                 err = le32_to_cpu(req->r_reply_info.head->result);
2042         } else if (err < 0) {
2043                 dout("aborted request %lld with %d\n", req->r_tid, err);
2044
2045                 /*
2046                  * ensure we aren't running concurrently with
2047                  * ceph_fill_trace or ceph_readdir_prepopulate, which
2048                  * rely on locks (dir mutex) held by our caller.
2049                  */
2050                 mutex_lock(&req->r_fill_mutex);
2051                 req->r_err = err;
2052                 req->r_aborted = true;
2053                 mutex_unlock(&req->r_fill_mutex);
2054
2055                 if (req->r_locked_dir &&
2056                     (req->r_op & CEPH_MDS_OP_WRITE))
2057                         ceph_invalidate_dir_request(req);
2058         } else {
2059                 err = req->r_err;
2060         }
2061
2062 out:
2063         mutex_unlock(&mdsc->mutex);
2064         dout("do_request %p done, result %d\n", req, err);
2065         return err;
2066 }
2067
2068 /*
2069  * Invalidate dir's completeness, dentry lease state on an aborted MDS
2070  * namespace request.
2071  */
2072 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2073 {
2074         struct inode *inode = req->r_locked_dir;
2075
2076         dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2077
2078         ceph_dir_clear_complete(inode);
2079         if (req->r_dentry)
2080                 ceph_invalidate_dentry_lease(req->r_dentry);
2081         if (req->r_old_dentry)
2082                 ceph_invalidate_dentry_lease(req->r_old_dentry);
2083 }
2084
2085 /*
2086  * Handle mds reply.
2087  *
2088  * We take the session mutex and parse and process the reply immediately.
2089  * This preserves the logical ordering of replies, capabilities, etc., sent
2090  * by the MDS as they are applied to our local cache.
2091  */
2092 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2093 {
2094         struct ceph_mds_client *mdsc = session->s_mdsc;
2095         struct ceph_mds_request *req;
2096         struct ceph_mds_reply_head *head = msg->front.iov_base;
2097         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2098         u64 tid;
2099         int err, result;
2100         int mds = session->s_mds;
2101
2102         if (msg->front.iov_len < sizeof(*head)) {
2103                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2104                 ceph_msg_dump(msg);
2105                 return;
2106         }
2107
2108         /* get request, session */
2109         tid = le64_to_cpu(msg->hdr.tid);
2110         mutex_lock(&mdsc->mutex);
2111         req = __lookup_request(mdsc, tid);
2112         if (!req) {
2113                 dout("handle_reply on unknown tid %llu\n", tid);
2114                 mutex_unlock(&mdsc->mutex);
2115                 return;
2116         }
2117         dout("handle_reply %p\n", req);
2118
2119         /* correct session? */
2120         if (req->r_session != session) {
2121                 pr_err("mdsc_handle_reply got %llu on session mds%d"
2122                        " not mds%d\n", tid, session->s_mds,
2123                        req->r_session ? req->r_session->s_mds : -1);
2124                 mutex_unlock(&mdsc->mutex);
2125                 goto out;
2126         }
2127
2128         /* dup? */
2129         if ((req->r_got_unsafe && !head->safe) ||
2130             (req->r_got_safe && head->safe)) {
2131                 pr_warning("got a dup %s reply on %llu from mds%d\n",
2132                            head->safe ? "safe" : "unsafe", tid, mds);
2133                 mutex_unlock(&mdsc->mutex);
2134                 goto out;
2135         }
2136         if (req->r_got_safe && !head->safe) {
2137                 pr_warning("got unsafe after safe on %llu from mds%d\n",
2138                            tid, mds);
2139                 mutex_unlock(&mdsc->mutex);
2140                 goto out;
2141         }
2142
2143         result = le32_to_cpu(head->result);
2144
2145         /*
2146          * Handle an ESTALE
2147          * if we're not talking to the authority, send to them
2148          * if the authority has changed while we weren't looking,
2149          * send to new authority
2150          * Otherwise we just have to return an ESTALE
2151          */
2152         if (result == -ESTALE) {
2153                 dout("got ESTALE on request %llu", req->r_tid);
2154                 if (!req->r_inode) {
2155                         /* do nothing; not an authority problem */
2156                 } else if (req->r_direct_mode != USE_AUTH_MDS) {
2157                         dout("not using auth, setting for that now");
2158                         req->r_direct_mode = USE_AUTH_MDS;
2159                         __do_request(mdsc, req);
2160                         mutex_unlock(&mdsc->mutex);
2161                         goto out;
2162                 } else  {
2163                         struct ceph_inode_info *ci = ceph_inode(req->r_inode);
2164                         struct ceph_cap *cap = NULL;
2165
2166                         if (req->r_session)
2167                                 cap = ceph_get_cap_for_mds(ci,
2168                                                    req->r_session->s_mds);
2169
2170                         dout("already using auth");
2171                         if ((!cap || cap != ci->i_auth_cap) ||
2172                             (cap->mseq != req->r_sent_on_mseq)) {
2173                                 dout("but cap changed, so resending");
2174                                 __do_request(mdsc, req);
2175                                 mutex_unlock(&mdsc->mutex);
2176                                 goto out;
2177                         }
2178                 }
2179                 dout("have to return ESTALE on request %llu", req->r_tid);
2180         }
2181
2182
2183         if (head->safe) {
2184                 req->r_got_safe = true;
2185                 __unregister_request(mdsc, req);
2186                 complete_all(&req->r_safe_completion);
2187
2188                 if (req->r_got_unsafe) {
2189                         /*
2190                          * We already handled the unsafe response, now do the
2191                          * cleanup.  No need to examine the response; the MDS
2192                          * doesn't include any result info in the safe
2193                          * response.  And even if it did, there is nothing
2194                          * useful we could do with a revised return value.
2195                          */
2196                         dout("got safe reply %llu, mds%d\n", tid, mds);
2197                         list_del_init(&req->r_unsafe_item);
2198
2199                         /* last unsafe request during umount? */
2200                         if (mdsc->stopping && !__get_oldest_req(mdsc))
2201                                 complete_all(&mdsc->safe_umount_waiters);
2202                         mutex_unlock(&mdsc->mutex);
2203                         goto out;
2204                 }
2205         } else {
2206                 req->r_got_unsafe = true;
2207                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2208         }
2209
2210         dout("handle_reply tid %lld result %d\n", tid, result);
2211         rinfo = &req->r_reply_info;
2212         err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2213         mutex_unlock(&mdsc->mutex);
2214
2215         mutex_lock(&session->s_mutex);
2216         if (err < 0) {
2217                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2218                 ceph_msg_dump(msg);
2219                 goto out_err;
2220         }
2221
2222         /* snap trace */
2223         if (rinfo->snapblob_len) {
2224                 down_write(&mdsc->snap_rwsem);
2225                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2226                                rinfo->snapblob + rinfo->snapblob_len,
2227                                le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
2228                 downgrade_write(&mdsc->snap_rwsem);
2229         } else {
2230                 down_read(&mdsc->snap_rwsem);
2231         }
2232
2233         /* insert trace into our cache */
2234         mutex_lock(&req->r_fill_mutex);
2235         err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2236         if (err == 0) {
2237                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2238                                     req->r_op == CEPH_MDS_OP_LSSNAP))
2239                         ceph_readdir_prepopulate(req, req->r_session);
2240                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2241         }
2242         mutex_unlock(&req->r_fill_mutex);
2243
2244         up_read(&mdsc->snap_rwsem);
2245 out_err:
2246         mutex_lock(&mdsc->mutex);
2247         if (!req->r_aborted) {
2248                 if (err) {
2249                         req->r_err = err;
2250                 } else {
2251                         req->r_reply = msg;
2252                         ceph_msg_get(msg);
2253                         req->r_got_result = true;
2254                 }
2255         } else {
2256                 dout("reply arrived after request %lld was aborted\n", tid);
2257         }
2258         mutex_unlock(&mdsc->mutex);
2259
2260         ceph_add_cap_releases(mdsc, req->r_session);
2261         mutex_unlock(&session->s_mutex);
2262
2263         /* kick calling process */
2264         complete_request(mdsc, req);
2265 out:
2266         ceph_mdsc_put_request(req);
2267         return;
2268 }
2269
2270
2271
2272 /*
2273  * handle mds notification that our request has been forwarded.
2274  */
2275 static void handle_forward(struct ceph_mds_client *mdsc,
2276                            struct ceph_mds_session *session,
2277                            struct ceph_msg *msg)
2278 {
2279         struct ceph_mds_request *req;
2280         u64 tid = le64_to_cpu(msg->hdr.tid);
2281         u32 next_mds;
2282         u32 fwd_seq;
2283         int err = -EINVAL;
2284         void *p = msg->front.iov_base;
2285         void *end = p + msg->front.iov_len;
2286
2287         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2288         next_mds = ceph_decode_32(&p);
2289         fwd_seq = ceph_decode_32(&p);
2290
2291         mutex_lock(&mdsc->mutex);
2292         req = __lookup_request(mdsc, tid);
2293         if (!req) {
2294                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2295                 goto out;  /* dup reply? */
2296         }
2297
2298         if (req->r_aborted) {
2299                 dout("forward tid %llu aborted, unregistering\n", tid);
2300                 __unregister_request(mdsc, req);
2301         } else if (fwd_seq <= req->r_num_fwd) {
2302                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2303                      tid, next_mds, req->r_num_fwd, fwd_seq);
2304         } else {
2305                 /* resend. forward race not possible; mds would drop */
2306                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2307                 BUG_ON(req->r_err);
2308                 BUG_ON(req->r_got_result);
2309                 req->r_num_fwd = fwd_seq;
2310                 req->r_resend_mds = next_mds;
2311                 put_request_session(req);
2312                 __do_request(mdsc, req);
2313         }
2314         ceph_mdsc_put_request(req);
2315 out:
2316         mutex_unlock(&mdsc->mutex);
2317         return;
2318
2319 bad:
2320         pr_err("mdsc_handle_forward decode error err=%d\n", err);
2321 }
2322
2323 /*
2324  * handle a mds session control message
2325  */
2326 static void handle_session(struct ceph_mds_session *session,
2327                            struct ceph_msg *msg)
2328 {
2329         struct ceph_mds_client *mdsc = session->s_mdsc;
2330         u32 op;
2331         u64 seq;
2332         int mds = session->s_mds;
2333         struct ceph_mds_session_head *h = msg->front.iov_base;
2334         int wake = 0;
2335
2336         /* decode */
2337         if (msg->front.iov_len != sizeof(*h))
2338                 goto bad;
2339         op = le32_to_cpu(h->op);
2340         seq = le64_to_cpu(h->seq);
2341
2342         mutex_lock(&mdsc->mutex);
2343         if (op == CEPH_SESSION_CLOSE)
2344                 __unregister_session(mdsc, session);
2345         /* FIXME: this ttl calculation is generous */
2346         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2347         mutex_unlock(&mdsc->mutex);
2348
2349         mutex_lock(&session->s_mutex);
2350
2351         dout("handle_session mds%d %s %p state %s seq %llu\n",
2352              mds, ceph_session_op_name(op), session,
2353              session_state_name(session->s_state), seq);
2354
2355         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2356                 session->s_state = CEPH_MDS_SESSION_OPEN;
2357                 pr_info("mds%d came back\n", session->s_mds);
2358         }
2359
2360         switch (op) {
2361         case CEPH_SESSION_OPEN:
2362                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2363                         pr_info("mds%d reconnect success\n", session->s_mds);
2364                 session->s_state = CEPH_MDS_SESSION_OPEN;
2365                 renewed_caps(mdsc, session, 0);
2366                 wake = 1;
2367                 if (mdsc->stopping)
2368                         __close_session(mdsc, session);
2369                 break;
2370
2371         case CEPH_SESSION_RENEWCAPS:
2372                 if (session->s_renew_seq == seq)
2373                         renewed_caps(mdsc, session, 1);
2374                 break;
2375
2376         case CEPH_SESSION_CLOSE:
2377                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2378                         pr_info("mds%d reconnect denied\n", session->s_mds);
2379                 remove_session_caps(session);
2380                 wake = 1; /* for good measure */
2381                 wake_up_all(&mdsc->session_close_wq);
2382                 kick_requests(mdsc, mds);
2383                 break;
2384
2385         case CEPH_SESSION_STALE:
2386                 pr_info("mds%d caps went stale, renewing\n",
2387                         session->s_mds);
2388                 spin_lock(&session->s_gen_ttl_lock);
2389                 session->s_cap_gen++;
2390                 session->s_cap_ttl = jiffies - 1;
2391                 spin_unlock(&session->s_gen_ttl_lock);
2392                 send_renew_caps(mdsc, session);
2393                 break;
2394
2395         case CEPH_SESSION_RECALL_STATE:
2396                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2397                 break;
2398
2399         default:
2400                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2401                 WARN_ON(1);
2402         }
2403
2404         mutex_unlock(&session->s_mutex);
2405         if (wake) {
2406                 mutex_lock(&mdsc->mutex);
2407                 __wake_requests(mdsc, &session->s_waiting);
2408                 mutex_unlock(&mdsc->mutex);
2409         }
2410         return;
2411
2412 bad:
2413         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2414                (int)msg->front.iov_len);
2415         ceph_msg_dump(msg);
2416         return;
2417 }
2418
2419
2420 /*
2421  * called under session->mutex.
2422  */
2423 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2424                                    struct ceph_mds_session *session)
2425 {
2426         struct ceph_mds_request *req, *nreq;
2427         int err;
2428
2429         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2430
2431         mutex_lock(&mdsc->mutex);
2432         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2433                 err = __prepare_send_request(mdsc, req, session->s_mds);
2434                 if (!err) {
2435                         ceph_msg_get(req->r_request);
2436                         ceph_con_send(&session->s_con, req->r_request);
2437                 }
2438         }
2439         mutex_unlock(&mdsc->mutex);
2440 }
2441
2442 /*
2443  * Encode information about a cap for a reconnect with the MDS.
2444  */
2445 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2446                           void *arg)
2447 {
2448         union {
2449                 struct ceph_mds_cap_reconnect v2;
2450                 struct ceph_mds_cap_reconnect_v1 v1;
2451         } rec;
2452         size_t reclen;
2453         struct ceph_inode_info *ci;
2454         struct ceph_reconnect_state *recon_state = arg;
2455         struct ceph_pagelist *pagelist = recon_state->pagelist;
2456         char *path;
2457         int pathlen, err;
2458         u64 pathbase;
2459         struct dentry *dentry;
2460
2461         ci = cap->ci;
2462
2463         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2464              inode, ceph_vinop(inode), cap, cap->cap_id,
2465              ceph_cap_string(cap->issued));
2466         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2467         if (err)
2468                 return err;
2469
2470         dentry = d_find_alias(inode);
2471         if (dentry) {
2472                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2473                 if (IS_ERR(path)) {
2474                         err = PTR_ERR(path);
2475                         goto out_dput;
2476                 }
2477         } else {
2478                 path = NULL;
2479                 pathlen = 0;
2480         }
2481         err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2482         if (err)
2483                 goto out_free;
2484
2485         spin_lock(&ci->i_ceph_lock);
2486         cap->seq = 0;        /* reset cap seq */
2487         cap->issue_seq = 0;  /* and issue_seq */
2488         cap->mseq = 0;       /* and migrate_seq */
2489         cap->cap_gen = cap->session->s_cap_gen;
2490
2491         if (recon_state->flock) {
2492                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2493                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2494                 rec.v2.issued = cpu_to_le32(cap->issued);
2495                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2496                 rec.v2.pathbase = cpu_to_le64(pathbase);
2497                 rec.v2.flock_len = 0;
2498                 reclen = sizeof(rec.v2);
2499         } else {
2500                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2501                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2502                 rec.v1.issued = cpu_to_le32(cap->issued);
2503                 rec.v1.size = cpu_to_le64(inode->i_size);
2504                 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2505                 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2506                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2507                 rec.v1.pathbase = cpu_to_le64(pathbase);
2508                 reclen = sizeof(rec.v1);
2509         }
2510         spin_unlock(&ci->i_ceph_lock);
2511
2512         if (recon_state->flock) {
2513                 int num_fcntl_locks, num_flock_locks;
2514                 struct ceph_filelock *flocks;
2515
2516 encode_again:
2517                 spin_lock(&inode->i_lock);
2518                 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2519                 spin_unlock(&inode->i_lock);
2520                 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2521                                  sizeof(struct ceph_filelock), GFP_NOFS);
2522                 if (!flocks) {
2523                         err = -ENOMEM;
2524                         goto out_free;
2525                 }
2526                 spin_lock(&inode->i_lock);
2527                 err = ceph_encode_locks_to_buffer(inode, flocks,
2528                                                   num_fcntl_locks,
2529                                                   num_flock_locks);
2530                 spin_unlock(&inode->i_lock);
2531                 if (err) {
2532                         kfree(flocks);
2533                         if (err == -ENOSPC)
2534                                 goto encode_again;
2535                         goto out_free;
2536                 }
2537                 /*
2538                  * number of encoded locks is stable, so copy to pagelist
2539                  */
2540                 rec.v2.flock_len = cpu_to_le32(2*sizeof(u32) +
2541                                     (num_fcntl_locks+num_flock_locks) *
2542                                     sizeof(struct ceph_filelock));
2543                 err = ceph_pagelist_append(pagelist, &rec, reclen);
2544                 if (!err)
2545                         err = ceph_locks_to_pagelist(flocks, pagelist,
2546                                                      num_fcntl_locks,
2547                                                      num_flock_locks);
2548                 kfree(flocks);
2549         } else {
2550                 err = ceph_pagelist_append(pagelist, &rec, reclen);
2551         }
2552
2553         recon_state->nr_caps++;
2554 out_free:
2555         kfree(path);
2556 out_dput:
2557         dput(dentry);
2558         return err;
2559 }
2560
2561
2562 /*
2563  * If an MDS fails and recovers, clients need to reconnect in order to
2564  * reestablish shared state.  This includes all caps issued through
2565  * this session _and_ the snap_realm hierarchy.  Because it's not
2566  * clear which snap realms the mds cares about, we send everything we
2567  * know about.. that ensures we'll then get any new info the
2568  * recovering MDS might have.
2569  *
2570  * This is a relatively heavyweight operation, but it's rare.
2571  *
2572  * called with mdsc->mutex held.
2573  */
2574 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2575                                struct ceph_mds_session *session)
2576 {
2577         struct ceph_msg *reply;
2578         struct rb_node *p;
2579         int mds = session->s_mds;
2580         int err = -ENOMEM;
2581         int s_nr_caps;
2582         struct ceph_pagelist *pagelist;
2583         struct ceph_reconnect_state recon_state;
2584
2585         pr_info("mds%d reconnect start\n", mds);
2586
2587         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2588         if (!pagelist)
2589                 goto fail_nopagelist;
2590         ceph_pagelist_init(pagelist);
2591
2592         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2593         if (!reply)
2594                 goto fail_nomsg;
2595
2596         mutex_lock(&session->s_mutex);
2597         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2598         session->s_seq = 0;
2599
2600         ceph_con_close(&session->s_con);
2601         ceph_con_open(&session->s_con,
2602                       CEPH_ENTITY_TYPE_MDS, mds,
2603                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2604
2605         /* replay unsafe requests */
2606         replay_unsafe_requests(mdsc, session);
2607
2608         down_read(&mdsc->snap_rwsem);
2609
2610         dout("session %p state %s\n", session,
2611              session_state_name(session->s_state));
2612
2613         spin_lock(&session->s_gen_ttl_lock);
2614         session->s_cap_gen++;
2615         spin_unlock(&session->s_gen_ttl_lock);
2616
2617         spin_lock(&session->s_cap_lock);
2618         /*
2619          * notify __ceph_remove_cap() that we are composing cap reconnect.
2620          * If a cap get released before being added to the cap reconnect,
2621          * __ceph_remove_cap() should skip queuing cap release.
2622          */
2623         session->s_cap_reconnect = 1;
2624         /* drop old cap expires; we're about to reestablish that state */
2625         discard_cap_releases(mdsc, session);
2626         spin_unlock(&session->s_cap_lock);
2627
2628         /* traverse this session's caps */
2629         s_nr_caps = session->s_nr_caps;
2630         err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
2631         if (err)
2632                 goto fail;
2633
2634         recon_state.nr_caps = 0;
2635         recon_state.pagelist = pagelist;
2636         recon_state.flock = session->s_con.peer_features & CEPH_FEATURE_FLOCK;
2637         err = iterate_session_caps(session, encode_caps_cb, &recon_state);
2638         if (err < 0)
2639                 goto fail;
2640
2641         spin_lock(&session->s_cap_lock);
2642         session->s_cap_reconnect = 0;
2643         spin_unlock(&session->s_cap_lock);
2644
2645         /*
2646          * snaprealms.  we provide mds with the ino, seq (version), and
2647          * parent for all of our realms.  If the mds has any newer info,
2648          * it will tell us.
2649          */
2650         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2651                 struct ceph_snap_realm *realm =
2652                         rb_entry(p, struct ceph_snap_realm, node);
2653                 struct ceph_mds_snaprealm_reconnect sr_rec;
2654
2655                 dout(" adding snap realm %llx seq %lld parent %llx\n",
2656                      realm->ino, realm->seq, realm->parent_ino);
2657                 sr_rec.ino = cpu_to_le64(realm->ino);
2658                 sr_rec.seq = cpu_to_le64(realm->seq);
2659                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2660                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2661                 if (err)
2662                         goto fail;
2663         }
2664
2665         if (recon_state.flock)
2666                 reply->hdr.version = cpu_to_le16(2);
2667
2668         /* raced with cap release? */
2669         if (s_nr_caps != recon_state.nr_caps) {
2670                 struct page *page = list_first_entry(&pagelist->head,
2671                                                      struct page, lru);
2672                 __le32 *addr = kmap_atomic(page);
2673                 *addr = cpu_to_le32(recon_state.nr_caps);
2674                 kunmap_atomic(addr);
2675         }
2676
2677         reply->hdr.data_len = cpu_to_le32(pagelist->length);
2678         ceph_msg_data_add_pagelist(reply, pagelist);
2679         ceph_con_send(&session->s_con, reply);
2680
2681         mutex_unlock(&session->s_mutex);
2682
2683         mutex_lock(&mdsc->mutex);
2684         __wake_requests(mdsc, &session->s_waiting);
2685         mutex_unlock(&mdsc->mutex);
2686
2687         up_read(&mdsc->snap_rwsem);
2688         return;
2689
2690 fail:
2691         ceph_msg_put(reply);
2692         up_read(&mdsc->snap_rwsem);
2693         mutex_unlock(&session->s_mutex);
2694 fail_nomsg:
2695         ceph_pagelist_release(pagelist);
2696         kfree(pagelist);
2697 fail_nopagelist:
2698         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
2699         return;
2700 }
2701
2702
2703 /*
2704  * compare old and new mdsmaps, kicking requests
2705  * and closing out old connections as necessary
2706  *
2707  * called under mdsc->mutex.
2708  */
2709 static void check_new_map(struct ceph_mds_client *mdsc,
2710                           struct ceph_mdsmap *newmap,
2711                           struct ceph_mdsmap *oldmap)
2712 {
2713         int i;
2714         int oldstate, newstate;
2715         struct ceph_mds_session *s;
2716
2717         dout("check_new_map new %u old %u\n",
2718              newmap->m_epoch, oldmap->m_epoch);
2719
2720         for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2721                 if (mdsc->sessions[i] == NULL)
2722                         continue;
2723                 s = mdsc->sessions[i];
2724                 oldstate = ceph_mdsmap_get_state(oldmap, i);
2725                 newstate = ceph_mdsmap_get_state(newmap, i);
2726
2727                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
2728                      i, ceph_mds_state_name(oldstate),
2729                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
2730                      ceph_mds_state_name(newstate),
2731                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
2732                      session_state_name(s->s_state));
2733
2734                 if (i >= newmap->m_max_mds ||
2735                     memcmp(ceph_mdsmap_get_addr(oldmap, i),
2736                            ceph_mdsmap_get_addr(newmap, i),
2737                            sizeof(struct ceph_entity_addr))) {
2738                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2739                                 /* the session never opened, just close it
2740                                  * out now */
2741                                 __wake_requests(mdsc, &s->s_waiting);
2742                                 __unregister_session(mdsc, s);
2743                         } else {
2744                                 /* just close it */
2745                                 mutex_unlock(&mdsc->mutex);
2746                                 mutex_lock(&s->s_mutex);
2747                                 mutex_lock(&mdsc->mutex);
2748                                 ceph_con_close(&s->s_con);
2749                                 mutex_unlock(&s->s_mutex);
2750                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2751                         }
2752
2753                         /* kick any requests waiting on the recovering mds */
2754                         kick_requests(mdsc, i);
2755                 } else if (oldstate == newstate) {
2756                         continue;  /* nothing new with this mds */
2757                 }
2758
2759                 /*
2760                  * send reconnect?
2761                  */
2762                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2763                     newstate >= CEPH_MDS_STATE_RECONNECT) {
2764                         mutex_unlock(&mdsc->mutex);
2765                         send_mds_reconnect(mdsc, s);
2766                         mutex_lock(&mdsc->mutex);
2767                 }
2768
2769                 /*
2770                  * kick request on any mds that has gone active.
2771                  */
2772                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2773                     newstate >= CEPH_MDS_STATE_ACTIVE) {
2774                         if (oldstate != CEPH_MDS_STATE_CREATING &&
2775                             oldstate != CEPH_MDS_STATE_STARTING)
2776                                 pr_info("mds%d recovery completed\n", s->s_mds);
2777                         kick_requests(mdsc, i);
2778                         ceph_kick_flushing_caps(mdsc, s);
2779                         wake_up_session_caps(s, 1);
2780                 }
2781         }
2782
2783         for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
2784                 s = mdsc->sessions[i];
2785                 if (!s)
2786                         continue;
2787                 if (!ceph_mdsmap_is_laggy(newmap, i))
2788                         continue;
2789                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2790                     s->s_state == CEPH_MDS_SESSION_HUNG ||
2791                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
2792                         dout(" connecting to export targets of laggy mds%d\n",
2793                              i);
2794                         __open_export_target_sessions(mdsc, s);
2795                 }
2796         }
2797 }
2798
2799
2800
2801 /*
2802  * leases
2803  */
2804
2805 /*
2806  * caller must hold session s_mutex, dentry->d_lock
2807  */
2808 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2809 {
2810         struct ceph_dentry_info *di = ceph_dentry(dentry);
2811
2812         ceph_put_mds_session(di->lease_session);
2813         di->lease_session = NULL;
2814 }
2815
2816 static void handle_lease(struct ceph_mds_client *mdsc,
2817                          struct ceph_mds_session *session,
2818                          struct ceph_msg *msg)
2819 {
2820         struct super_block *sb = mdsc->fsc->sb;
2821         struct inode *inode;
2822         struct dentry *parent, *dentry;
2823         struct ceph_dentry_info *di;
2824         int mds = session->s_mds;
2825         struct ceph_mds_lease *h = msg->front.iov_base;
2826         u32 seq;
2827         struct ceph_vino vino;
2828         struct qstr dname;
2829         int release = 0;
2830
2831         dout("handle_lease from mds%d\n", mds);
2832
2833         /* decode */
2834         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2835                 goto bad;
2836         vino.ino = le64_to_cpu(h->ino);
2837         vino.snap = CEPH_NOSNAP;
2838         seq = le32_to_cpu(h->seq);
2839         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2840         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2841         if (dname.len != get_unaligned_le32(h+1))
2842                 goto bad;
2843
2844         mutex_lock(&session->s_mutex);
2845         session->s_seq++;
2846
2847         /* lookup inode */
2848         inode = ceph_find_inode(sb, vino);
2849         dout("handle_lease %s, ino %llx %p %.*s\n",
2850              ceph_lease_op_name(h->action), vino.ino, inode,
2851              dname.len, dname.name);
2852         if (inode == NULL) {
2853                 dout("handle_lease no inode %llx\n", vino.ino);
2854                 goto release;
2855         }
2856
2857         /* dentry */
2858         parent = d_find_alias(inode);
2859         if (!parent) {
2860                 dout("no parent dentry on inode %p\n", inode);
2861                 WARN_ON(1);
2862                 goto release;  /* hrm... */
2863         }
2864         dname.hash = full_name_hash(dname.name, dname.len);
2865         dentry = d_lookup(parent, &dname);
2866         dput(parent);
2867         if (!dentry)
2868                 goto release;
2869
2870         spin_lock(&dentry->d_lock);
2871         di = ceph_dentry(dentry);
2872         switch (h->action) {
2873         case CEPH_MDS_LEASE_REVOKE:
2874                 if (di->lease_session == session) {
2875                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
2876                                 h->seq = cpu_to_le32(di->lease_seq);
2877                         __ceph_mdsc_drop_dentry_lease(dentry);
2878                 }
2879                 release = 1;
2880                 break;
2881
2882         case CEPH_MDS_LEASE_RENEW:
2883                 if (di->lease_session == session &&
2884                     di->lease_gen == session->s_cap_gen &&
2885                     di->lease_renew_from &&
2886                     di->lease_renew_after == 0) {
2887                         unsigned long duration =
2888                                 le32_to_cpu(h->duration_ms) * HZ / 1000;
2889
2890                         di->lease_seq = seq;
2891                         dentry->d_time = di->lease_renew_from + duration;
2892                         di->lease_renew_after = di->lease_renew_from +
2893                                 (duration >> 1);
2894                         di->lease_renew_from = 0;
2895                 }
2896                 break;
2897         }
2898         spin_unlock(&dentry->d_lock);
2899         dput(dentry);
2900
2901         if (!release)
2902                 goto out;
2903
2904 release:
2905         /* let's just reuse the same message */
2906         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2907         ceph_msg_get(msg);
2908         ceph_con_send(&session->s_con, msg);
2909
2910 out:
2911         iput(inode);
2912         mutex_unlock(&session->s_mutex);
2913         return;
2914
2915 bad:
2916         pr_err("corrupt lease message\n");
2917         ceph_msg_dump(msg);
2918 }
2919
2920 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
2921                               struct inode *inode,
2922                               struct dentry *dentry, char action,
2923                               u32 seq)
2924 {
2925         struct ceph_msg *msg;
2926         struct ceph_mds_lease *lease;
2927         int len = sizeof(*lease) + sizeof(u32);
2928         int dnamelen = 0;
2929
2930         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2931              inode, dentry, ceph_lease_op_name(action), session->s_mds);
2932         dnamelen = dentry->d_name.len;
2933         len += dnamelen;
2934
2935         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
2936         if (!msg)
2937                 return;
2938         lease = msg->front.iov_base;
2939         lease->action = action;
2940         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
2941         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
2942         lease->seq = cpu_to_le32(seq);
2943         put_unaligned_le32(dnamelen, lease + 1);
2944         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
2945
2946         /*
2947          * if this is a preemptive lease RELEASE, no need to
2948          * flush request stream, since the actual request will
2949          * soon follow.
2950          */
2951         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
2952
2953         ceph_con_send(&session->s_con, msg);
2954 }
2955
2956 /*
2957  * Preemptively release a lease we expect to invalidate anyway.
2958  * Pass @inode always, @dentry is optional.
2959  */
2960 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2961                              struct dentry *dentry)
2962 {
2963         struct ceph_dentry_info *di;
2964         struct ceph_mds_session *session;
2965         u32 seq;
2966
2967         BUG_ON(inode == NULL);
2968         BUG_ON(dentry == NULL);
2969
2970         /* is dentry lease valid? */
2971         spin_lock(&dentry->d_lock);
2972         di = ceph_dentry(dentry);
2973         if (!di || !di->lease_session ||
2974             di->lease_session->s_mds < 0 ||
2975             di->lease_gen != di->lease_session->s_cap_gen ||
2976             !time_before(jiffies, dentry->d_time)) {
2977                 dout("lease_release inode %p dentry %p -- "
2978                      "no lease\n",
2979                      inode, dentry);
2980                 spin_unlock(&dentry->d_lock);
2981                 return;
2982         }
2983
2984         /* we do have a lease on this dentry; note mds and seq */
2985         session = ceph_get_mds_session(di->lease_session);
2986         seq = di->lease_seq;
2987         __ceph_mdsc_drop_dentry_lease(dentry);
2988         spin_unlock(&dentry->d_lock);
2989
2990         dout("lease_release inode %p dentry %p to mds%d\n",
2991              inode, dentry, session->s_mds);
2992         ceph_mdsc_lease_send_msg(session, inode, dentry,
2993                                  CEPH_MDS_LEASE_RELEASE, seq);
2994         ceph_put_mds_session(session);
2995 }
2996
2997 /*
2998  * drop all leases (and dentry refs) in preparation for umount
2999  */
3000 static void drop_leases(struct ceph_mds_client *mdsc)
3001 {
3002         int i;
3003
3004         dout("drop_leases\n");
3005         mutex_lock(&mdsc->mutex);
3006         for (i = 0; i < mdsc->max_sessions; i++) {
3007                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3008                 if (!s)
3009                         continue;
3010                 mutex_unlock(&mdsc->mutex);
3011                 mutex_lock(&s->s_mutex);
3012                 mutex_unlock(&s->s_mutex);
3013                 ceph_put_mds_session(s);
3014                 mutex_lock(&mdsc->mutex);
3015         }
3016         mutex_unlock(&mdsc->mutex);
3017 }
3018
3019
3020
3021 /*
3022  * delayed work -- periodically trim expired leases, renew caps with mds
3023  */
3024 static void schedule_delayed(struct ceph_mds_client *mdsc)
3025 {
3026         int delay = 5;
3027         unsigned hz = round_jiffies_relative(HZ * delay);
3028         schedule_delayed_work(&mdsc->delayed_work, hz);
3029 }
3030
3031 static void delayed_work(struct work_struct *work)
3032 {
3033         int i;
3034         struct ceph_mds_client *mdsc =
3035                 container_of(work, struct ceph_mds_client, delayed_work.work);
3036         int renew_interval;
3037         int renew_caps;
3038
3039         dout("mdsc delayed_work\n");
3040         ceph_check_delayed_caps(mdsc);
3041
3042         mutex_lock(&mdsc->mutex);
3043         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3044         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3045                                    mdsc->last_renew_caps);
3046         if (renew_caps)
3047                 mdsc->last_renew_caps = jiffies;
3048
3049         for (i = 0; i < mdsc->max_sessions; i++) {
3050                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3051                 if (s == NULL)
3052                         continue;
3053                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3054                         dout("resending session close request for mds%d\n",
3055                              s->s_mds);
3056                         request_close_session(mdsc, s);
3057                         ceph_put_mds_session(s);
3058                         continue;
3059                 }
3060                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3061                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3062                                 s->s_state = CEPH_MDS_SESSION_HUNG;
3063                                 pr_info("mds%d hung\n", s->s_mds);
3064                         }
3065                 }
3066                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3067                         /* this mds is failed or recovering, just wait */
3068                         ceph_put_mds_session(s);
3069                         continue;
3070                 }
3071                 mutex_unlock(&mdsc->mutex);
3072
3073                 mutex_lock(&s->s_mutex);
3074                 if (renew_caps)
3075                         send_renew_caps(mdsc, s);
3076                 else
3077                         ceph_con_keepalive(&s->s_con);
3078                 ceph_add_cap_releases(mdsc, s);
3079                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3080                     s->s_state == CEPH_MDS_SESSION_HUNG)
3081                         ceph_send_cap_releases(mdsc, s);
3082                 mutex_unlock(&s->s_mutex);
3083                 ceph_put_mds_session(s);
3084
3085                 mutex_lock(&mdsc->mutex);
3086         }
3087         mutex_unlock(&mdsc->mutex);
3088
3089         schedule_delayed(mdsc);
3090 }
3091
3092 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3093
3094 {
3095         struct ceph_mds_client *mdsc;
3096
3097         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3098         if (!mdsc)
3099                 return -ENOMEM;
3100         mdsc->fsc = fsc;
3101         fsc->mdsc = mdsc;
3102         mutex_init(&mdsc->mutex);
3103         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3104         if (mdsc->mdsmap == NULL) {
3105                 kfree(mdsc);
3106                 return -ENOMEM;
3107         }
3108
3109         init_completion(&mdsc->safe_umount_waiters);
3110         init_waitqueue_head(&mdsc->session_close_wq);
3111         INIT_LIST_HEAD(&mdsc->waiting_for_map);
3112         mdsc->sessions = NULL;
3113         mdsc->max_sessions = 0;
3114         mdsc->stopping = 0;
3115         init_rwsem(&mdsc->snap_rwsem);
3116         mdsc->snap_realms = RB_ROOT;
3117         INIT_LIST_HEAD(&mdsc->snap_empty);
3118         spin_lock_init(&mdsc->snap_empty_lock);
3119         mdsc->last_tid = 0;
3120         mdsc->request_tree = RB_ROOT;
3121         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3122         mdsc->last_renew_caps = jiffies;
3123         INIT_LIST_HEAD(&mdsc->cap_delay_list);
3124         spin_lock_init(&mdsc->cap_delay_lock);
3125         INIT_LIST_HEAD(&mdsc->snap_flush_list);
3126         spin_lock_init(&mdsc->snap_flush_lock);
3127         mdsc->cap_flush_seq = 0;
3128         INIT_LIST_HEAD(&mdsc->cap_dirty);
3129         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3130         mdsc->num_cap_flushing = 0;
3131         spin_lock_init(&mdsc->cap_dirty_lock);
3132         init_waitqueue_head(&mdsc->cap_flushing_wq);
3133         spin_lock_init(&mdsc->dentry_lru_lock);
3134         INIT_LIST_HEAD(&mdsc->dentry_lru);
3135
3136         ceph_caps_init(mdsc);
3137         ceph_adjust_min_caps(mdsc, fsc->min_caps);
3138
3139         return 0;
3140 }
3141
3142 /*
3143  * Wait for safe replies on open mds requests.  If we time out, drop
3144  * all requests from the tree to avoid dangling dentry refs.
3145  */
3146 static void wait_requests(struct ceph_mds_client *mdsc)
3147 {
3148         struct ceph_mds_request *req;
3149         struct ceph_fs_client *fsc = mdsc->fsc;
3150
3151         mutex_lock(&mdsc->mutex);
3152         if (__get_oldest_req(mdsc)) {
3153                 mutex_unlock(&mdsc->mutex);
3154
3155                 dout("wait_requests waiting for requests\n");
3156                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3157                                     fsc->client->options->mount_timeout * HZ);
3158
3159                 /* tear down remaining requests */
3160                 mutex_lock(&mdsc->mutex);
3161                 while ((req = __get_oldest_req(mdsc))) {
3162                         dout("wait_requests timed out on tid %llu\n",
3163                              req->r_tid);
3164                         __unregister_request(mdsc, req);
3165                 }
3166         }
3167         mutex_unlock(&mdsc->mutex);
3168         dout("wait_requests done\n");
3169 }
3170
3171 /*
3172  * called before mount is ro, and before dentries are torn down.
3173  * (hmm, does this still race with new lookups?)
3174  */
3175 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3176 {
3177         dout("pre_umount\n");
3178         mdsc->stopping = 1;
3179
3180         drop_leases(mdsc);
3181         ceph_flush_dirty_caps(mdsc);
3182         wait_requests(mdsc);
3183
3184         /*
3185          * wait for reply handlers to drop their request refs and
3186          * their inode/dcache refs
3187          */
3188         ceph_msgr_flush();
3189 }
3190
3191 /*
3192  * wait for all write mds requests to flush.
3193  */
3194 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3195 {
3196         struct ceph_mds_request *req = NULL, *nextreq;
3197         struct rb_node *n;
3198
3199         mutex_lock(&mdsc->mutex);
3200         dout("wait_unsafe_requests want %lld\n", want_tid);
3201 restart:
3202         req = __get_oldest_req(mdsc);
3203         while (req && req->r_tid <= want_tid) {
3204                 /* find next request */
3205                 n = rb_next(&req->r_node);
3206                 if (n)
3207                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3208                 else
3209                         nextreq = NULL;
3210                 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
3211                         /* write op */
3212                         ceph_mdsc_get_request(req);
3213                         if (nextreq)
3214                                 ceph_mdsc_get_request(nextreq);
3215                         mutex_unlock(&mdsc->mutex);
3216                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
3217                              req->r_tid, want_tid);
3218                         wait_for_completion(&req->r_safe_completion);
3219                         mutex_lock(&mdsc->mutex);
3220                         ceph_mdsc_put_request(req);
3221                         if (!nextreq)
3222                                 break;  /* next dne before, so we're done! */
3223                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
3224                                 /* next request was removed from tree */
3225                                 ceph_mdsc_put_request(nextreq);
3226                                 goto restart;
3227                         }
3228                         ceph_mdsc_put_request(nextreq);  /* won't go away */
3229                 }
3230                 req = nextreq;
3231         }
3232         mutex_unlock(&mdsc->mutex);
3233         dout("wait_unsafe_requests done\n");
3234 }
3235
3236 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3237 {
3238         u64 want_tid, want_flush;
3239
3240         if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3241                 return;
3242
3243         dout("sync\n");
3244         mutex_lock(&mdsc->mutex);
3245         want_tid = mdsc->last_tid;
3246         want_flush = mdsc->cap_flush_seq;
3247         mutex_unlock(&mdsc->mutex);
3248         dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
3249
3250         ceph_flush_dirty_caps(mdsc);
3251
3252         wait_unsafe_requests(mdsc, want_tid);
3253         wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
3254 }
3255
3256 /*
3257  * true if all sessions are closed, or we force unmount
3258  */
3259 static bool done_closing_sessions(struct ceph_mds_client *mdsc)
3260 {
3261         int i, n = 0;
3262
3263         if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3264                 return true;
3265
3266         mutex_lock(&mdsc->mutex);
3267         for (i = 0; i < mdsc->max_sessions; i++)
3268                 if (mdsc->sessions[i])
3269                         n++;
3270         mutex_unlock(&mdsc->mutex);
3271         return n == 0;
3272 }
3273
3274 /*
3275  * called after sb is ro.
3276  */
3277 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3278 {
3279         struct ceph_mds_session *session;
3280         int i;
3281         struct ceph_fs_client *fsc = mdsc->fsc;
3282         unsigned long timeout = fsc->client->options->mount_timeout * HZ;
3283
3284         dout("close_sessions\n");
3285
3286         /* close sessions */
3287         mutex_lock(&mdsc->mutex);
3288         for (i = 0; i < mdsc->max_sessions; i++) {
3289                 session = __ceph_lookup_mds_session(mdsc, i);
3290                 if (!session)
3291                         continue;
3292                 mutex_unlock(&mdsc->mutex);
3293                 mutex_lock(&session->s_mutex);
3294                 __close_session(mdsc, session);
3295                 mutex_unlock(&session->s_mutex);
3296                 ceph_put_mds_session(session);
3297                 mutex_lock(&mdsc->mutex);
3298         }
3299         mutex_unlock(&mdsc->mutex);
3300
3301         dout("waiting for sessions to close\n");
3302         wait_event_timeout(mdsc->session_close_wq, done_closing_sessions(mdsc),
3303                            timeout);
3304
3305         /* tear down remaining sessions */
3306         mutex_lock(&mdsc->mutex);
3307         for (i = 0; i < mdsc->max_sessions; i++) {
3308                 if (mdsc->sessions[i]) {
3309                         session = get_session(mdsc->sessions[i]);
3310                         __unregister_session(mdsc, session);
3311                         mutex_unlock(&mdsc->mutex);
3312                         mutex_lock(&session->s_mutex);
3313                         remove_session_caps(session);
3314                         mutex_unlock(&session->s_mutex);
3315                         ceph_put_mds_session(session);
3316                         mutex_lock(&mdsc->mutex);
3317                 }
3318         }
3319         WARN_ON(!list_empty(&mdsc->cap_delay_list));
3320         mutex_unlock(&mdsc->mutex);
3321
3322         ceph_cleanup_empty_realms(mdsc);
3323
3324         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3325
3326         dout("stopped\n");
3327 }
3328
3329 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3330 {
3331         dout("stop\n");
3332         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3333         if (mdsc->mdsmap)
3334                 ceph_mdsmap_destroy(mdsc->mdsmap);
3335         kfree(mdsc->sessions);
3336         ceph_caps_finalize(mdsc);
3337 }
3338
3339 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3340 {
3341         struct ceph_mds_client *mdsc = fsc->mdsc;
3342
3343         dout("mdsc_destroy %p\n", mdsc);
3344         ceph_mdsc_stop(mdsc);
3345
3346         /* flush out any connection work with references to us */
3347         ceph_msgr_flush();
3348
3349         fsc->mdsc = NULL;
3350         kfree(mdsc);
3351         dout("mdsc_destroy %p done\n", mdsc);
3352 }
3353
3354
3355 /*
3356  * handle mds map update.
3357  */
3358 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3359 {
3360         u32 epoch;
3361         u32 maplen;
3362         void *p = msg->front.iov_base;
3363         void *end = p + msg->front.iov_len;
3364         struct ceph_mdsmap *newmap, *oldmap;
3365         struct ceph_fsid fsid;
3366         int err = -EINVAL;
3367
3368         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3369         ceph_decode_copy(&p, &fsid, sizeof(fsid));
3370         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3371                 return;
3372         epoch = ceph_decode_32(&p);
3373         maplen = ceph_decode_32(&p);
3374         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3375
3376         /* do we need it? */
3377         ceph_monc_got_mdsmap(&mdsc->fsc->client->monc, epoch);
3378         mutex_lock(&mdsc->mutex);
3379         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3380                 dout("handle_map epoch %u <= our %u\n",
3381                      epoch, mdsc->mdsmap->m_epoch);
3382                 mutex_unlock(&mdsc->mutex);
3383                 return;
3384         }
3385
3386         newmap = ceph_mdsmap_decode(&p, end);
3387         if (IS_ERR(newmap)) {
3388                 err = PTR_ERR(newmap);
3389                 goto bad_unlock;
3390         }
3391
3392         /* swap into place */
3393         if (mdsc->mdsmap) {
3394                 oldmap = mdsc->mdsmap;
3395                 mdsc->mdsmap = newmap;
3396                 check_new_map(mdsc, newmap, oldmap);
3397                 ceph_mdsmap_destroy(oldmap);
3398         } else {
3399                 mdsc->mdsmap = newmap;  /* first mds map */
3400         }
3401         mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3402
3403         __wake_requests(mdsc, &mdsc->waiting_for_map);
3404
3405         mutex_unlock(&mdsc->mutex);
3406         schedule_delayed(mdsc);
3407         return;
3408
3409 bad_unlock:
3410         mutex_unlock(&mdsc->mutex);
3411 bad:
3412         pr_err("error decoding mdsmap %d\n", err);
3413         return;
3414 }
3415
3416 static struct ceph_connection *con_get(struct ceph_connection *con)
3417 {
3418         struct ceph_mds_session *s = con->private;
3419
3420         if (get_session(s)) {
3421                 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3422                 return con;
3423         }
3424         dout("mdsc con_get %p FAIL\n", s);
3425         return NULL;
3426 }
3427
3428 static void con_put(struct ceph_connection *con)
3429 {
3430         struct ceph_mds_session *s = con->private;
3431
3432         dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3433         ceph_put_mds_session(s);
3434 }
3435
3436 /*
3437  * if the client is unresponsive for long enough, the mds will kill
3438  * the session entirely.
3439  */
3440 static void peer_reset(struct ceph_connection *con)
3441 {
3442         struct ceph_mds_session *s = con->private;
3443         struct ceph_mds_client *mdsc = s->s_mdsc;
3444
3445         pr_warning("mds%d closed our session\n", s->s_mds);
3446         send_mds_reconnect(mdsc, s);
3447 }
3448
3449 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3450 {
3451         struct ceph_mds_session *s = con->private;
3452         struct ceph_mds_client *mdsc = s->s_mdsc;
3453         int type = le16_to_cpu(msg->hdr.type);
3454
3455         mutex_lock(&mdsc->mutex);
3456         if (__verify_registered_session(mdsc, s) < 0) {
3457                 mutex_unlock(&mdsc->mutex);
3458                 goto out;
3459         }
3460         mutex_unlock(&mdsc->mutex);
3461
3462         switch (type) {
3463         case CEPH_MSG_MDS_MAP:
3464                 ceph_mdsc_handle_map(mdsc, msg);
3465                 break;
3466         case CEPH_MSG_CLIENT_SESSION:
3467                 handle_session(s, msg);
3468                 break;
3469         case CEPH_MSG_CLIENT_REPLY:
3470                 handle_reply(s, msg);
3471                 break;
3472         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3473                 handle_forward(mdsc, s, msg);
3474                 break;
3475         case CEPH_MSG_CLIENT_CAPS:
3476                 ceph_handle_caps(s, msg);
3477                 break;
3478         case CEPH_MSG_CLIENT_SNAP:
3479                 ceph_handle_snap(mdsc, s, msg);
3480                 break;
3481         case CEPH_MSG_CLIENT_LEASE:
3482                 handle_lease(mdsc, s, msg);
3483                 break;
3484
3485         default:
3486                 pr_err("received unknown message type %d %s\n", type,
3487                        ceph_msg_type_name(type));
3488         }
3489 out:
3490         ceph_msg_put(msg);
3491 }
3492
3493 /*
3494  * authentication
3495  */
3496
3497 /*
3498  * Note: returned pointer is the address of a structure that's
3499  * managed separately.  Caller must *not* attempt to free it.
3500  */
3501 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
3502                                         int *proto, int force_new)
3503 {
3504         struct ceph_mds_session *s = con->private;
3505         struct ceph_mds_client *mdsc = s->s_mdsc;
3506         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3507         struct ceph_auth_handshake *auth = &s->s_auth;
3508
3509         if (force_new && auth->authorizer) {
3510                 ceph_auth_destroy_authorizer(ac, auth->authorizer);
3511                 auth->authorizer = NULL;
3512         }
3513         if (!auth->authorizer) {
3514                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3515                                                       auth);
3516                 if (ret)
3517                         return ERR_PTR(ret);
3518         } else {
3519                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3520                                                       auth);
3521                 if (ret)
3522                         return ERR_PTR(ret);
3523         }
3524         *proto = ac->protocol;
3525
3526         return auth;
3527 }
3528
3529
3530 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3531 {
3532         struct ceph_mds_session *s = con->private;
3533         struct ceph_mds_client *mdsc = s->s_mdsc;
3534         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3535
3536         return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer, len);
3537 }
3538
3539 static int invalidate_authorizer(struct ceph_connection *con)
3540 {
3541         struct ceph_mds_session *s = con->private;
3542         struct ceph_mds_client *mdsc = s->s_mdsc;
3543         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3544
3545         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3546
3547         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
3548 }
3549
3550 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
3551                                 struct ceph_msg_header *hdr, int *skip)
3552 {
3553         struct ceph_msg *msg;
3554         int type = (int) le16_to_cpu(hdr->type);
3555         int front_len = (int) le32_to_cpu(hdr->front_len);
3556
3557         if (con->in_msg)
3558                 return con->in_msg;
3559
3560         *skip = 0;
3561         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
3562         if (!msg) {
3563                 pr_err("unable to allocate msg type %d len %d\n",
3564                        type, front_len);
3565                 return NULL;
3566         }
3567
3568         return msg;
3569 }
3570
3571 static const struct ceph_connection_operations mds_con_ops = {
3572         .get = con_get,
3573         .put = con_put,
3574         .dispatch = dispatch,
3575         .get_authorizer = get_authorizer,
3576         .verify_authorizer_reply = verify_authorizer_reply,
3577         .invalidate_authorizer = invalidate_authorizer,
3578         .peer_reset = peer_reset,
3579         .alloc_msg = mds_alloc_msg,
3580 };
3581
3582 /* eof */