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/gfp.h>
   7 #include <linux/sched.h>
   8 #include <linux/debugfs.h>
   9 #include <linux/seq_file.h>
  10 #include <linux/utsname.h>
  11 #include <linux/ratelimit.h>
  12
  13 #include "super.h"
  14 #include "mds_client.h"
  15
  16 #include <linux/ceph/ceph_features.h>
  17 #include <linux/ceph/messenger.h>
  18 #include <linux/ceph/decode.h>
  19 #include <linux/ceph/pagelist.h>
  20 #include <linux/ceph/auth.h>
  21 #include <linux/ceph/debugfs.h>
  22
  23 /*
  24  * A cluster of MDS (metadata server) daemons is responsible for
  25  * managing the file system namespace (the directory hierarchy and
  26  * inodes) and for coordinating shared access to storage.  Metadata is
  27  * partitioning hierarchically across a number of servers, and that
  28  * partition varies over time as the cluster adjusts the distribution
  29  * in order to balance load.
  30  *
  31  * The MDS client is primarily responsible to managing synchronous
  32  * metadata requests for operations like open, unlink, and so forth.
  33  * If there is a MDS failure, we find out about it when we (possibly
  34  * request and) receive a new MDS map, and can resubmit affected
  35  * requests.
  36  *
  37  * For the most part, though, we take advantage of a lossless
  38  * communications channel to the MDS, and do not need to worry about
  39  * timing out or resubmitting requests.
  40  *
  41  * We maintain a stateful "session" with each MDS we interact with.
  42  * Within each session, we sent periodic heartbeat messages to ensure
  43  * any capabilities or leases we have been issues remain valid.  If
  44  * the session times out and goes stale, our leases and capabilities
  45  * are no longer valid.
  46  */
  47
  48 struct ceph_reconnect_state {
  49         int nr_caps;
  50         struct ceph_pagelist *pagelist;
  51         unsigned msg_version;
  52 };
  53
  54 static void __wake_requests(struct ceph_mds_client *mdsc,
  55                             struct list_head *head);
  56
  57 static const struct ceph_connection_operations mds_con_ops;
  58
  59
  60 /*
  61  * mds reply parsing
  62  */
  63
  64 /*
  65  * parse individual inode info
  66  */
  67 static int parse_reply_info_in(void **p, void *end,
  68                                struct ceph_mds_reply_info_in *info,
  69                                u64 features)
  70 {
  71         int err = -EIO;
  72
  73         info->in = *p;
  74         *p += sizeof(struct ceph_mds_reply_inode) +
  75                 sizeof(*info->in->fragtree.splits) *
  76                 le32_to_cpu(info->in->fragtree.nsplits);
  77
  78         ceph_decode_32_safe(p, end, info->symlink_len, bad);
  79         ceph_decode_need(p, end, info->symlink_len, bad);
  80         info->symlink = *p;
  81         *p += info->symlink_len;
  82
  83         if (features & CEPH_FEATURE_DIRLAYOUTHASH)
  84                 ceph_decode_copy_safe(p, end, &info->dir_layout,
  85                                       sizeof(info->dir_layout), bad);
  86         else
  87                 memset(&info->dir_layout, 0, sizeof(info->dir_layout));
  88
  89         ceph_decode_32_safe(p, end, info->xattr_len, bad);
  90         ceph_decode_need(p, end, info->xattr_len, bad);
  91         info->xattr_data = *p;
  92         *p += info->xattr_len;
  93
  94         if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
  95                 ceph_decode_64_safe(p, end, info->inline_version, bad);
  96                 ceph_decode_32_safe(p, end, info->inline_len, bad);
  97                 ceph_decode_need(p, end, info->inline_len, bad);
  98                 info->inline_data = *p;
  99                 *p += info->inline_len;
 100         } else
 101                 info->inline_version = CEPH_INLINE_NONE;
 102
 103         info->pool_ns_len = 0;
 104         info->pool_ns_data = NULL;
 105         if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
 106                 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
 107                 if (info->pool_ns_len > 0) {
 108                         ceph_decode_need(p, end, info->pool_ns_len, bad);
 109                         info->pool_ns_data = *p;
 110                         *p += info->pool_ns_len;
 111                 }
 112         }
 113
 114         return 0;
 115 bad:
 116         return err;
 117 }
 118
 119 /*
 120  * parse a normal reply, which may contain a (dir+)dentry and/or a
 121  * target inode.
 122  */
 123 static int parse_reply_info_trace(void **p, void *end,
 124                                   struct ceph_mds_reply_info_parsed *info,
 125                                   u64 features)
 126 {
 127         int err;
 128
 129         if (info->head->is_dentry) {
 130                 err = parse_reply_info_in(p, end, &info->diri, features);
 131                 if (err < 0)
 132                         goto out_bad;
 133
 134                 if (unlikely(*p + sizeof(*info->dirfrag) > end))
 135                         goto bad;
 136                 info->dirfrag = *p;
 137                 *p += sizeof(*info->dirfrag) +
 138                         sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
 139                 if (unlikely(*p > end))
 140                         goto bad;
 141
 142                 ceph_decode_32_safe(p, end, info->dname_len, bad);
 143                 ceph_decode_need(p, end, info->dname_len, bad);
 144                 info->dname = *p;
 145                 *p += info->dname_len;
 146                 info->dlease = *p;
 147                 *p += sizeof(*info->dlease);
 148         }
 149
 150         if (info->head->is_target) {
 151                 err = parse_reply_info_in(p, end, &info->targeti, features);
 152                 if (err < 0)
 153                         goto out_bad;
 154         }
 155
 156         if (unlikely(*p != end))
 157                 goto bad;
 158         return 0;
 159
 160 bad:
 161         err = -EIO;
 162 out_bad:
 163         pr_err("problem parsing mds trace %d\n", err);
 164         return err;
 165 }
 166
 167 /*
 168  * parse readdir results
 169  */
 170 static int parse_reply_info_dir(void **p, void *end,
 171                                 struct ceph_mds_reply_info_parsed *info,
 172                                 u64 features)
 173 {
 174         u32 num, i = 0;
 175         int err;
 176
 177         info->dir_dir = *p;
 178         if (*p + sizeof(*info->dir_dir) > end)
 179                 goto bad;
 180         *p += sizeof(*info->dir_dir) +
 181                 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
 182         if (*p > end)
 183                 goto bad;
 184
 185         ceph_decode_need(p, end, sizeof(num) + 2, bad);
 186         num = ceph_decode_32(p);
 187         {
 188                 u16 flags = ceph_decode_16(p);
 189                 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
 190                 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
 191                 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
 192         }
 193         if (num == 0)
 194                 goto done;
 195
 196         BUG_ON(!info->dir_entries);
 197         if ((unsigned long)(info->dir_entries + num) >
 198             (unsigned long)info->dir_entries + info->dir_buf_size) {
 199                 pr_err("dir contents are larger than expected\n");
 200                 WARN_ON(1);
 201                 goto bad;
 202         }
 203
 204         info->dir_nr = num;
 205         while (num) {
 206                 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
 207                 /* dentry */
 208                 ceph_decode_need(p, end, sizeof(u32)*2, bad);
 209                 rde->name_len = ceph_decode_32(p);
 210                 ceph_decode_need(p, end, rde->name_len, bad);
 211                 rde->name = *p;
 212                 *p += rde->name_len;
 213                 dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
 214                 rde->lease = *p;
 215                 *p += sizeof(struct ceph_mds_reply_lease);
 216
 217                 /* inode */
 218                 err = parse_reply_info_in(p, end, &rde->inode, features);
 219                 if (err < 0)
 220                         goto out_bad;
 221                 /* ceph_readdir_prepopulate() will update it */
 222                 rde->offset = 0;
 223                 i++;
 224                 num--;
 225         }
 226
 227 done:
 228         if (*p != end)
 229                 goto bad;
 230         return 0;
 231
 232 bad:
 233         err = -EIO;
 234 out_bad:
 235         pr_err("problem parsing dir contents %d\n", err);
 236         return err;
 237 }
 238
 239 /*
 240  * parse fcntl F_GETLK results
 241  */
 242 static int parse_reply_info_filelock(void **p, void *end,
 243                                      struct ceph_mds_reply_info_parsed *info,
 244                                      u64 features)
 245 {
 246         if (*p + sizeof(*info->filelock_reply) > end)
 247                 goto bad;
 248
 249         info->filelock_reply = *p;
 250         *p += sizeof(*info->filelock_reply);
 251
 252         if (unlikely(*p != end))
 253                 goto bad;
 254         return 0;
 255
 256 bad:
 257         return -EIO;
 258 }
 259
 260 /*
 261  * parse create results
 262  */
 263 static int parse_reply_info_create(void **p, void *end,
 264                                   struct ceph_mds_reply_info_parsed *info,
 265                                   u64 features)
 266 {
 267         if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
 268                 if (*p == end) {
 269                         info->has_create_ino = false;
 270                 } else {
 271                         info->has_create_ino = true;
 272                         info->ino = ceph_decode_64(p);
 273                 }
 274         }
 275
 276         if (unlikely(*p != end))
 277                 goto bad;
 278         return 0;
 279
 280 bad:
 281         return -EIO;
 282 }
 283
 284 /*
 285  * parse extra results
 286  */
 287 static int parse_reply_info_extra(void **p, void *end,
 288                                   struct ceph_mds_reply_info_parsed *info,
 289                                   u64 features)
 290 {
 291         if (info->head->op == CEPH_MDS_OP_GETFILELOCK)
 292                 return parse_reply_info_filelock(p, end, info, features);
 293         else if (info->head->op == CEPH_MDS_OP_READDIR ||
 294                  info->head->op == CEPH_MDS_OP_LSSNAP)
 295                 return parse_reply_info_dir(p, end, info, features);
 296         else if (info->head->op == CEPH_MDS_OP_CREATE)
 297                 return parse_reply_info_create(p, end, info, features);
 298         else
 299                 return -EIO;
 300 }
 301
 302 /*
 303  * parse entire mds reply
 304  */
 305 static int parse_reply_info(struct ceph_msg *msg,
 306                             struct ceph_mds_reply_info_parsed *info,
 307                             u64 features)
 308 {
 309         void *p, *end;
 310         u32 len;
 311         int err;
 312
 313         info->head = msg->front.iov_base;
 314         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 315         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 316
 317         /* trace */
 318         ceph_decode_32_safe(&p, end, len, bad);
 319         if (len > 0) {
 320                 ceph_decode_need(&p, end, len, bad);
 321                 err = parse_reply_info_trace(&p, p+len, info, features);
 322                 if (err < 0)
 323                         goto out_bad;
 324         }
 325
 326         /* extra */
 327         ceph_decode_32_safe(&p, end, len, bad);
 328         if (len > 0) {
 329                 ceph_decode_need(&p, end, len, bad);
 330                 err = parse_reply_info_extra(&p, p+len, info, features);
 331                 if (err < 0)
 332                         goto out_bad;
 333         }
 334
 335         /* snap blob */
 336         ceph_decode_32_safe(&p, end, len, bad);
 337         info->snapblob_len = len;
 338         info->snapblob = p;
 339         p += len;
 340
 341         if (p != end)
 342                 goto bad;
 343         return 0;
 344
 345 bad:
 346         err = -EIO;
 347 out_bad:
 348         pr_err("mds parse_reply err %d\n", err);
 349         return err;
 350 }
 351
 352 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
 353 {
 354         if (!info->dir_entries)
 355                 return;
 356         free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
 357 }
 358
 359
 360 /*
 361  * sessions
 362  */
 363 const char *ceph_session_state_name(int s)
 364 {
 365         switch (s) {
 366         case CEPH_MDS_SESSION_NEW: return "new";
 367         case CEPH_MDS_SESSION_OPENING: return "opening";
 368         case CEPH_MDS_SESSION_OPEN: return "open";
 369         case CEPH_MDS_SESSION_HUNG: return "hung";
 370         case CEPH_MDS_SESSION_CLOSING: return "closing";
 371         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
 372         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
 373         case CEPH_MDS_SESSION_REJECTED: return "rejected";
 374         default: return "???";
 375         }
 376 }
 377
 378 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
 379 {
 380         if (atomic_inc_not_zero(&s->s_ref)) {
 381                 dout("mdsc get_session %p %d -> %d\n", s,
 382                      atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
 383                 return s;
 384         } else {
 385                 dout("mdsc get_session %p 0 -- FAIL", s);
 386                 return NULL;
 387         }
 388 }
 389
 390 void ceph_put_mds_session(struct ceph_mds_session *s)
 391 {
 392         dout("mdsc put_session %p %d -> %d\n", s,
 393              atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
 394         if (atomic_dec_and_test(&s->s_ref)) {
 395                 if (s->s_auth.authorizer)
 396                         ceph_auth_destroy_authorizer(s->s_auth.authorizer);
 397                 kfree(s);
 398         }
 399 }
 400
 401 /*
 402  * called under mdsc->mutex
 403  */
 404 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
 405                                                    int mds)
 406 {
 407         struct ceph_mds_session *session;
 408
 409         if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
 410                 return NULL;
 411         session = mdsc->sessions[mds];
 412         dout("lookup_mds_session %p %d\n", session,
 413              atomic_read(&session->s_ref));
 414         get_session(session);
 415         return session;
 416 }
 417
 418 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
 419 {
 420         if (mds >= mdsc->max_sessions)
 421                 return false;
 422         return mdsc->sessions[mds];
 423 }
 424
 425 static int __verify_registered_session(struct ceph_mds_client *mdsc,
 426                                        struct ceph_mds_session *s)
 427 {
 428         if (s->s_mds >= mdsc->max_sessions ||
 429             mdsc->sessions[s->s_mds] != s)
 430                 return -ENOENT;
 431         return 0;
 432 }
 433
 434 /*
 435  * create+register a new session for given mds.
 436  * called under mdsc->mutex.
 437  */
 438 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 439                                                  int mds)
 440 {
 441         struct ceph_mds_session *s;
 442
 443         if (mds >= mdsc->mdsmap->m_max_mds)
 444                 return ERR_PTR(-EINVAL);
 445
 446         s = kzalloc(sizeof(*s), GFP_NOFS);
 447         if (!s)
 448                 return ERR_PTR(-ENOMEM);
 449         s->s_mdsc = mdsc;
 450         s->s_mds = mds;
 451         s->s_state = CEPH_MDS_SESSION_NEW;
 452         s->s_ttl = 0;
 453         s->s_seq = 0;
 454         mutex_init(&s->s_mutex);
 455
 456         ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
 457
 458         spin_lock_init(&s->s_gen_ttl_lock);
 459         s->s_cap_gen = 0;
 460         s->s_cap_ttl = jiffies - 1;
 461
 462         spin_lock_init(&s->s_cap_lock);
 463         s->s_renew_requested = 0;
 464         s->s_renew_seq = 0;
 465         INIT_LIST_HEAD(&s->s_caps);
 466         s->s_nr_caps = 0;
 467         s->s_trim_caps = 0;
 468         atomic_set(&s->s_ref, 1);
 469         INIT_LIST_HEAD(&s->s_waiting);
 470         INIT_LIST_HEAD(&s->s_unsafe);
 471         s->s_num_cap_releases = 0;
 472         s->s_cap_reconnect = 0;
 473         s->s_cap_iterator = NULL;
 474         INIT_LIST_HEAD(&s->s_cap_releases);
 475         INIT_LIST_HEAD(&s->s_cap_flushing);
 476
 477         dout("register_session mds%d\n", mds);
 478         if (mds >= mdsc->max_sessions) {
 479                 int newmax = 1 << get_count_order(mds+1);
 480                 struct ceph_mds_session **sa;
 481
 482                 dout("register_session realloc to %d\n", newmax);
 483                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 484                 if (sa == NULL)
 485                         goto fail_realloc;
 486                 if (mdsc->sessions) {
 487                         memcpy(sa, mdsc->sessions,
 488                                mdsc->max_sessions * sizeof(void *));
 489                         kfree(mdsc->sessions);
 490                 }
 491                 mdsc->sessions = sa;
 492                 mdsc->max_sessions = newmax;
 493         }
 494         mdsc->sessions[mds] = s;
 495         atomic_inc(&mdsc->num_sessions);
 496         atomic_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
 497
 498         ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
 499                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
 500
 501         return s;
 502
 503 fail_realloc:
 504         kfree(s);
 505         return ERR_PTR(-ENOMEM);
 506 }
 507
 508 /*
 509  * called under mdsc->mutex
 510  */
 511 static void __unregister_session(struct ceph_mds_client *mdsc,
 512                                struct ceph_mds_session *s)
 513 {
 514         dout("__unregister_session mds%d %p\n", s->s_mds, s);
 515         BUG_ON(mdsc->sessions[s->s_mds] != s);
 516         mdsc->sessions[s->s_mds] = NULL;
 517         ceph_con_close(&s->s_con);
 518         ceph_put_mds_session(s);
 519         atomic_dec(&mdsc->num_sessions);
 520 }
 521
 522 /*
 523  * drop session refs in request.
 524  *
 525  * should be last request ref, or hold mdsc->mutex
 526  */
 527 static void put_request_session(struct ceph_mds_request *req)
 528 {
 529         if (req->r_session) {
 530                 ceph_put_mds_session(req->r_session);
 531                 req->r_session = NULL;
 532         }
 533 }
 534
 535 void ceph_mdsc_release_request(struct kref *kref)
 536 {
 537         struct ceph_mds_request *req = container_of(kref,
 538                                                     struct ceph_mds_request,
 539                                                     r_kref);
 540         destroy_reply_info(&req->r_reply_info);
 541         if (req->r_request)
 542                 ceph_msg_put(req->r_request);
 543         if (req->r_reply)
 544                 ceph_msg_put(req->r_reply);
 545         if (req->r_inode) {
 546                 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
 547                 iput(req->r_inode);
 548         }
 549         if (req->r_locked_dir)
 550                 ceph_put_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
 551         iput(req->r_target_inode);
 552         if (req->r_dentry)
 553                 dput(req->r_dentry);
 554         if (req->r_old_dentry)
 555                 dput(req->r_old_dentry);
 556         if (req->r_old_dentry_dir) {
 557                 /*
 558                  * track (and drop pins for) r_old_dentry_dir
 559                  * separately, since r_old_dentry's d_parent may have
 560                  * changed between the dir mutex being dropped and
 561                  * this request being freed.
 562                  */
 563                 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
 564                                   CEPH_CAP_PIN);
 565                 iput(req->r_old_dentry_dir);
 566         }
 567         kfree(req->r_path1);
 568         kfree(req->r_path2);
 569         if (req->r_pagelist)
 570                 ceph_pagelist_release(req->r_pagelist);
 571         put_request_session(req);
 572         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
 573         kfree(req);
 574 }
 575
 576 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
 577
 578 /*
 579  * lookup session, bump ref if found.
 580  *
 581  * called under mdsc->mutex.
 582  */
 583 static struct ceph_mds_request *
 584 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
 585 {
 586         struct ceph_mds_request *req;
 587
 588         req = lookup_request(&mdsc->request_tree, tid);
 589         if (req)
 590                 ceph_mdsc_get_request(req);
 591
 592         return req;
 593 }
 594
 595 /*
 596  * Register an in-flight request, and assign a tid.  Link to directory
 597  * are modifying (if any).
 598  *
 599  * Called under mdsc->mutex.
 600  */
 601 static void __register_request(struct ceph_mds_client *mdsc,
 602                                struct ceph_mds_request *req,
 603                                struct inode *dir)
 604 {
 605         req->r_tid = ++mdsc->last_tid;
 606         if (req->r_num_caps)
 607                 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
 608                                   req->r_num_caps);
 609         dout("__register_request %p tid %lld\n", req, req->r_tid);
 610         ceph_mdsc_get_request(req);
 611         insert_request(&mdsc->request_tree, req);
 612
 613         req->r_uid = current_fsuid();
 614         req->r_gid = current_fsgid();
 615
 616         if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
 617                 mdsc->oldest_tid = req->r_tid;
 618
 619         if (dir) {
 620                 ihold(dir);
 621                 req->r_unsafe_dir = dir;
 622         }
 623 }
 624
 625 static void __unregister_request(struct ceph_mds_client *mdsc,
 626                                  struct ceph_mds_request *req)
 627 {
 628         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 629
 630         if (req->r_tid == mdsc->oldest_tid) {
 631                 struct rb_node *p = rb_next(&req->r_node);
 632                 mdsc->oldest_tid = 0;
 633                 while (p) {
 634                         struct ceph_mds_request *next_req =
 635                                 rb_entry(p, struct ceph_mds_request, r_node);
 636                         if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
 637                                 mdsc->oldest_tid = next_req->r_tid;
 638                                 break;
 639                         }
 640                         p = rb_next(p);
 641                 }
 642         }
 643
 644         erase_request(&mdsc->request_tree, req);
 645
 646         if (req->r_unsafe_dir && req->r_got_unsafe) {
 647                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 648                 spin_lock(&ci->i_unsafe_lock);
 649                 list_del_init(&req->r_unsafe_dir_item);
 650                 spin_unlock(&ci->i_unsafe_lock);
 651         }
 652         if (req->r_target_inode && req->r_got_unsafe) {
 653                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
 654                 spin_lock(&ci->i_unsafe_lock);
 655                 list_del_init(&req->r_unsafe_target_item);
 656                 spin_unlock(&ci->i_unsafe_lock);
 657         }
 658
 659         if (req->r_unsafe_dir) {
 660                 iput(req->r_unsafe_dir);
 661                 req->r_unsafe_dir = NULL;
 662         }
 663
 664         complete_all(&req->r_safe_completion);
 665
 666         ceph_mdsc_put_request(req);
 667 }
 668
 669 /*
 670  * Choose mds to send request to next.  If there is a hint set in the
 671  * request (e.g., due to a prior forward hint from the mds), use that.
 672  * Otherwise, consult frag tree and/or caps to identify the
 673  * appropriate mds.  If all else fails, choose randomly.
 674  *
 675  * Called under mdsc->mutex.
 676  */
 677 static struct dentry *get_nonsnap_parent(struct dentry *dentry)
 678 {
 679         /*
 680          * we don't need to worry about protecting the d_parent access
 681          * here because we never renaming inside the snapped namespace
 682          * except to resplice to another snapdir, and either the old or new
 683          * result is a valid result.
 684          */
 685         while (!IS_ROOT(dentry) && ceph_snap(d_inode(dentry)) != CEPH_NOSNAP)
 686                 dentry = dentry->d_parent;
 687         return dentry;
 688 }
 689
 690 static int __choose_mds(struct ceph_mds_client *mdsc,
 691                         struct ceph_mds_request *req)
 692 {
 693         struct inode *inode;
 694         struct ceph_inode_info *ci;
 695         struct ceph_cap *cap;
 696         int mode = req->r_direct_mode;
 697         int mds = -1;
 698         u32 hash = req->r_direct_hash;
 699         bool is_hash = req->r_direct_is_hash;
 700
 701         /*
 702          * is there a specific mds we should try?  ignore hint if we have
 703          * no session and the mds is not up (active or recovering).
 704          */
 705         if (req->r_resend_mds >= 0 &&
 706             (__have_session(mdsc, req->r_resend_mds) ||
 707              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 708                 dout("choose_mds using resend_mds mds%d\n",
 709                      req->r_resend_mds);
 710                 return req->r_resend_mds;
 711         }
 712
 713         if (mode == USE_RANDOM_MDS)
 714                 goto random;
 715
 716         inode = NULL;
 717         if (req->r_inode) {
 718                 inode = req->r_inode;
 719         } else if (req->r_dentry) {
 720                 /* ignore race with rename; old or new d_parent is okay */
 721                 struct dentry *parent = req->r_dentry->d_parent;
 722                 struct inode *dir = d_inode(parent);
 723
 724                 if (dir->i_sb != mdsc->fsc->sb) {
 725                         /* not this fs! */
 726                         inode = d_inode(req->r_dentry);
 727                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
 728                         /* direct snapped/virtual snapdir requests
 729                          * based on parent dir inode */
 730                         struct dentry *dn = get_nonsnap_parent(parent);
 731                         inode = d_inode(dn);
 732                         dout("__choose_mds using nonsnap parent %p\n", inode);
 733                 } else {
 734                         /* dentry target */
 735                         inode = d_inode(req->r_dentry);
 736                         if (!inode || mode == USE_AUTH_MDS) {
 737                                 /* dir + name */
 738                                 inode = dir;
 739                                 hash = ceph_dentry_hash(dir, req->r_dentry);
 740                                 is_hash = true;
 741                         }
 742                 }
 743         }
 744
 745         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 746              (int)hash, mode);
 747         if (!inode)
 748                 goto random;
 749         ci = ceph_inode(inode);
 750
 751         if (is_hash && S_ISDIR(inode->i_mode)) {
 752                 struct ceph_inode_frag frag;
 753                 int found;
 754
 755                 ceph_choose_frag(ci, hash, &frag, &found);
 756                 if (found) {
 757                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 758                                 u8 r;
 759
 760                                 /* choose a random replica */
 761                                 get_random_bytes(&r, 1);
 762                                 r %= frag.ndist;
 763                                 mds = frag.dist[r];
 764                                 dout("choose_mds %p %llx.%llx "
 765                                      "frag %u mds%d (%d/%d)\n",
 766                                      inode, ceph_vinop(inode),
 767                                      frag.frag, mds,
 768                                      (int)r, frag.ndist);
 769                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 770                                     CEPH_MDS_STATE_ACTIVE)
 771                                         return mds;
 772                         }
 773
 774                         /* since this file/dir wasn't known to be
 775                          * replicated, then we want to look for the
 776                          * authoritative mds. */
 777                         mode = USE_AUTH_MDS;
 778                         if (frag.mds >= 0) {
 779                                 /* choose auth mds */
 780                                 mds = frag.mds;
 781                                 dout("choose_mds %p %llx.%llx "
 782                                      "frag %u mds%d (auth)\n",
 783                                      inode, ceph_vinop(inode), frag.frag, mds);
 784                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 785                                     CEPH_MDS_STATE_ACTIVE)
 786                                         return mds;
 787                         }
 788                 }
 789         }
 790
 791         spin_lock(&ci->i_ceph_lock);
 792         cap = NULL;
 793         if (mode == USE_AUTH_MDS)
 794                 cap = ci->i_auth_cap;
 795         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
 796                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
 797         if (!cap) {
 798                 spin_unlock(&ci->i_ceph_lock);
 799                 goto random;
 800         }
 801         mds = cap->session->s_mds;
 802         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
 803              inode, ceph_vinop(inode), mds,
 804              cap == ci->i_auth_cap ? "auth " : "", cap);
 805         spin_unlock(&ci->i_ceph_lock);
 806         return mds;
 807
 808 random:
 809         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
 810         dout("choose_mds chose random mds%d\n", mds);
 811         return mds;
 812 }
 813
 814
 815 /*
 816  * session messages
 817  */
 818 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
 819 {
 820         struct ceph_msg *msg;
 821         struct ceph_mds_session_head *h;
 822
 823         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
 824                            false);
 825         if (!msg) {
 826                 pr_err("create_session_msg ENOMEM creating msg\n");
 827                 return NULL;
 828         }
 829         h = msg->front.iov_base;
 830         h->op = cpu_to_le32(op);
 831         h->seq = cpu_to_le64(seq);
 832
 833         return msg;
 834 }
 835
 836 /*
 837  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
 838  * to include additional client metadata fields.
 839  */
 840 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
 841 {
 842         struct ceph_msg *msg;
 843         struct ceph_mds_session_head *h;
 844         int i = -1;
 845         int metadata_bytes = 0;
 846         int metadata_key_count = 0;
 847         struct ceph_options *opt = mdsc->fsc->client->options;
 848         struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
 849         void *p;
 850
 851         const char* metadata[][2] = {
 852                 {"hostname", utsname()->nodename},
 853                 {"kernel_version", utsname()->release},
 854                 {"entity_id", opt->name ? : ""},
 855                 {"root", fsopt->server_path ? : "/"},
 856                 {NULL, NULL}
 857         };
 858
 859         /* Calculate serialized length of metadata */
 860         metadata_bytes = 4;  /* map length */
 861         for (i = 0; metadata[i][0] != NULL; ++i) {
 862                 metadata_bytes += 8 + strlen(metadata[i][0]) +
 863                         strlen(metadata[i][1]);
 864                 metadata_key_count++;
 865         }
 866
 867         /* Allocate the message */
 868         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + metadata_bytes,
 869                            GFP_NOFS, false);
 870         if (!msg) {
 871                 pr_err("create_session_msg ENOMEM creating msg\n");
 872                 return NULL;
 873         }
 874         h = msg->front.iov_base;
 875         h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
 876         h->seq = cpu_to_le64(seq);
 877
 878         /*
 879          * Serialize client metadata into waiting buffer space, using
 880          * the format that userspace expects for map<string, string>
 881          *
 882          * ClientSession messages with metadata are v2
 883          */
 884         msg->hdr.version = cpu_to_le16(2);
 885         msg->hdr.compat_version = cpu_to_le16(1);
 886
 887         /* The write pointer, following the session_head structure */
 888         p = msg->front.iov_base + sizeof(*h);
 889
 890         /* Number of entries in the map */
 891         ceph_encode_32(&p, metadata_key_count);
 892
 893         /* Two length-prefixed strings for each entry in the map */
 894         for (i = 0; metadata[i][0] != NULL; ++i) {
 895                 size_t const key_len = strlen(metadata[i][0]);
 896                 size_t const val_len = strlen(metadata[i][1]);
 897
 898                 ceph_encode_32(&p, key_len);
 899                 memcpy(p, metadata[i][0], key_len);
 900                 p += key_len;
 901                 ceph_encode_32(&p, val_len);
 902                 memcpy(p, metadata[i][1], val_len);
 903                 p += val_len;
 904         }
 905
 906         return msg;
 907 }
 908
 909 /*
 910  * send session open request.
 911  *
 912  * called under mdsc->mutex
 913  */
 914 static int __open_session(struct ceph_mds_client *mdsc,
 915                           struct ceph_mds_session *session)
 916 {
 917         struct ceph_msg *msg;
 918         int mstate;
 919         int mds = session->s_mds;
 920
 921         /* wait for mds to go active? */
 922         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
 923         dout("open_session to mds%d (%s)\n", mds,
 924              ceph_mds_state_name(mstate));
 925         session->s_state = CEPH_MDS_SESSION_OPENING;
 926         session->s_renew_requested = jiffies;
 927
 928         /* send connect message */
 929         msg = create_session_open_msg(mdsc, session->s_seq);
 930         if (!msg)
 931                 return -ENOMEM;
 932         ceph_con_send(&session->s_con, msg);
 933         return 0;
 934 }
 935
 936 /*
 937  * open sessions for any export targets for the given mds
 938  *
 939  * called under mdsc->mutex
 940  */
 941 static struct ceph_mds_session *
 942 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
 943 {
 944         struct ceph_mds_session *session;
 945
 946         session = __ceph_lookup_mds_session(mdsc, target);
 947         if (!session) {
 948                 session = register_session(mdsc, target);
 949                 if (IS_ERR(session))
 950                         return session;
 951         }
 952         if (session->s_state == CEPH_MDS_SESSION_NEW ||
 953             session->s_state == CEPH_MDS_SESSION_CLOSING)
 954                 __open_session(mdsc, session);
 955
 956         return session;
 957 }
 958
 959 struct ceph_mds_session *
 960 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
 961 {
 962         struct ceph_mds_session *session;
 963
 964         dout("open_export_target_session to mds%d\n", target);
 965
 966         mutex_lock(&mdsc->mutex);
 967         session = __open_export_target_session(mdsc, target);
 968         mutex_unlock(&mdsc->mutex);
 969
 970         return session;
 971 }
 972
 973 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
 974                                           struct ceph_mds_session *session)
 975 {
 976         struct ceph_mds_info *mi;
 977         struct ceph_mds_session *ts;
 978         int i, mds = session->s_mds;
 979
 980         if (mds >= mdsc->mdsmap->m_max_mds)
 981                 return;
 982
 983         mi = &mdsc->mdsmap->m_info[mds];
 984         dout("open_export_target_sessions for mds%d (%d targets)\n",
 985              session->s_mds, mi->num_export_targets);
 986
 987         for (i = 0; i < mi->num_export_targets; i++) {
 988                 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
 989                 if (!IS_ERR(ts))
 990                         ceph_put_mds_session(ts);
 991         }
 992 }
 993
 994 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
 995                                            struct ceph_mds_session *session)
 996 {
 997         mutex_lock(&mdsc->mutex);
 998         __open_export_target_sessions(mdsc, session);
 999         mutex_unlock(&mdsc->mutex);
1000 }
1001
1002 /*
1003  * session caps
1004  */
1005
1006 /* caller holds s_cap_lock, we drop it */
1007 static void cleanup_cap_releases(struct ceph_mds_client *mdsc,
1008                                  struct ceph_mds_session *session)
1009         __releases(session->s_cap_lock)
1010 {
1011         LIST_HEAD(tmp_list);
1012         list_splice_init(&session->s_cap_releases, &tmp_list);
1013         session->s_num_cap_releases = 0;
1014         spin_unlock(&session->s_cap_lock);
1015
1016         dout("cleanup_cap_releases mds%d\n", session->s_mds);
1017         while (!list_empty(&tmp_list)) {
1018                 struct ceph_cap *cap;
1019                 /* zero out the in-progress message */
1020                 cap = list_first_entry(&tmp_list,
1021                                         struct ceph_cap, session_caps);
1022                 list_del(&cap->session_caps);
1023                 ceph_put_cap(mdsc, cap);
1024         }
1025 }
1026
1027 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1028                                      struct ceph_mds_session *session)
1029 {
1030         struct ceph_mds_request *req;
1031         struct rb_node *p;
1032
1033         dout("cleanup_session_requests mds%d\n", session->s_mds);
1034         mutex_lock(&mdsc->mutex);
1035         while (!list_empty(&session->s_unsafe)) {
1036                 req = list_first_entry(&session->s_unsafe,
1037                                        struct ceph_mds_request, r_unsafe_item);
1038                 list_del_init(&req->r_unsafe_item);
1039                 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1040                                     req->r_tid);
1041                 __unregister_request(mdsc, req);
1042         }
1043         /* zero r_attempts, so kick_requests() will re-send requests */
1044         p = rb_first(&mdsc->request_tree);
1045         while (p) {
1046                 req = rb_entry(p, struct ceph_mds_request, r_node);
1047                 p = rb_next(p);
1048                 if (req->r_session &&
1049                     req->r_session->s_mds == session->s_mds)
1050                         req->r_attempts = 0;
1051         }
1052         mutex_unlock(&mdsc->mutex);
1053 }
1054
1055 /*
1056  * Helper to safely iterate over all caps associated with a session, with
1057  * special care taken to handle a racing __ceph_remove_cap().
1058  *
1059  * Caller must hold session s_mutex.
1060  */
1061 static int iterate_session_caps(struct ceph_mds_session *session,
1062                                  int (*cb)(struct inode *, struct ceph_cap *,
1063                                             void *), void *arg)
1064 {
1065         struct list_head *p;
1066         struct ceph_cap *cap;
1067         struct inode *inode, *last_inode = NULL;
1068         struct ceph_cap *old_cap = NULL;
1069         int ret;
1070
1071         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1072         spin_lock(&session->s_cap_lock);
1073         p = session->s_caps.next;
1074         while (p != &session->s_caps) {
1075                 cap = list_entry(p, struct ceph_cap, session_caps);
1076                 inode = igrab(&cap->ci->vfs_inode);
1077                 if (!inode) {
1078                         p = p->next;
1079                         continue;
1080                 }
1081                 session->s_cap_iterator = cap;
1082                 spin_unlock(&session->s_cap_lock);
1083
1084                 if (last_inode) {
1085                         iput(last_inode);
1086                         last_inode = NULL;
1087                 }
1088                 if (old_cap) {
1089                         ceph_put_cap(session->s_mdsc, old_cap);
1090                         old_cap = NULL;
1091                 }
1092
1093                 ret = cb(inode, cap, arg);
1094                 last_inode = inode;
1095
1096                 spin_lock(&session->s_cap_lock);
1097                 p = p->next;
1098                 if (cap->ci == NULL) {
1099                         dout("iterate_session_caps  finishing cap %p removal\n",
1100                              cap);
1101                         BUG_ON(cap->session != session);
1102                         cap->session = NULL;
1103                         list_del_init(&cap->session_caps);
1104                         session->s_nr_caps--;
1105                         if (cap->queue_release) {
1106                                 list_add_tail(&cap->session_caps,
1107                                               &session->s_cap_releases);
1108                                 session->s_num_cap_releases++;
1109                         } else {
1110                                 old_cap = cap;  /* put_cap it w/o locks held */
1111                         }
1112                 }
1113                 if (ret < 0)
1114                         goto out;
1115         }
1116         ret = 0;
1117 out:
1118         session->s_cap_iterator = NULL;
1119         spin_unlock(&session->s_cap_lock);
1120
1121         iput(last_inode);
1122         if (old_cap)
1123                 ceph_put_cap(session->s_mdsc, old_cap);
1124
1125         return ret;
1126 }
1127
1128 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1129                                   void *arg)
1130 {
1131         struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1132         struct ceph_inode_info *ci = ceph_inode(inode);
1133         LIST_HEAD(to_remove);
1134         bool drop = false;
1135         bool invalidate = false;
1136
1137         dout("removing cap %p, ci is %p, inode is %p\n",
1138              cap, ci, &ci->vfs_inode);
1139         spin_lock(&ci->i_ceph_lock);
1140         __ceph_remove_cap(cap, false);
1141         if (!ci->i_auth_cap) {
1142                 struct ceph_cap_flush *cf;
1143                 struct ceph_mds_client *mdsc = fsc->mdsc;
1144
1145                 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1146
1147                 if (ci->i_wrbuffer_ref > 0 &&
1148                     ACCESS_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
1149                         invalidate = true;
1150
1151                 while (!list_empty(&ci->i_cap_flush_list)) {
1152                         cf = list_first_entry(&ci->i_cap_flush_list,
1153                                               struct ceph_cap_flush, i_list);
1154                         list_move(&cf->i_list, &to_remove);
1155                 }
1156
1157                 spin_lock(&mdsc->cap_dirty_lock);
1158
1159                 list_for_each_entry(cf, &to_remove, i_list)
1160                         list_del(&cf->g_list);
1161
1162                 if (!list_empty(&ci->i_dirty_item)) {
1163                         pr_warn_ratelimited(
1164                                 " dropping dirty %s state for %p %lld\n",
1165                                 ceph_cap_string(ci->i_dirty_caps),
1166                                 inode, ceph_ino(inode));
1167                         ci->i_dirty_caps = 0;
1168                         list_del_init(&ci->i_dirty_item);
1169                         drop = true;
1170                 }
1171                 if (!list_empty(&ci->i_flushing_item)) {
1172                         pr_warn_ratelimited(
1173                                 " dropping dirty+flushing %s state for %p %lld\n",
1174                                 ceph_cap_string(ci->i_flushing_caps),
1175                                 inode, ceph_ino(inode));
1176                         ci->i_flushing_caps = 0;
1177                         list_del_init(&ci->i_flushing_item);
1178                         mdsc->num_cap_flushing--;
1179                         drop = true;
1180                 }
1181                 spin_unlock(&mdsc->cap_dirty_lock);
1182
1183                 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1184                         list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1185                         ci->i_prealloc_cap_flush = NULL;
1186                 }
1187         }
1188         spin_unlock(&ci->i_ceph_lock);
1189         while (!list_empty(&to_remove)) {
1190                 struct ceph_cap_flush *cf;
1191                 cf = list_first_entry(&to_remove,
1192                                       struct ceph_cap_flush, i_list);
1193                 list_del(&cf->i_list);
1194                 ceph_free_cap_flush(cf);
1195         }
1196
1197         wake_up_all(&ci->i_cap_wq);
1198         if (invalidate)
1199                 ceph_queue_invalidate(inode);
1200         if (drop)
1201                 iput(inode);
1202         return 0;
1203 }
1204
1205 /*
1206  * caller must hold session s_mutex
1207  */
1208 static void remove_session_caps(struct ceph_mds_session *session)
1209 {
1210         struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1211         struct super_block *sb = fsc->sb;
1212         dout("remove_session_caps on %p\n", session);
1213         iterate_session_caps(session, remove_session_caps_cb, fsc);
1214
1215         wake_up_all(&fsc->mdsc->cap_flushing_wq);
1216
1217         spin_lock(&session->s_cap_lock);
1218         if (session->s_nr_caps > 0) {
1219                 struct inode *inode;
1220                 struct ceph_cap *cap, *prev = NULL;
1221                 struct ceph_vino vino;
1222                 /*
1223                  * iterate_session_caps() skips inodes that are being
1224                  * deleted, we need to wait until deletions are complete.
1225                  * __wait_on_freeing_inode() is designed for the job,
1226                  * but it is not exported, so use lookup inode function
1227                  * to access it.
1228                  */
1229                 while (!list_empty(&session->s_caps)) {
1230                         cap = list_entry(session->s_caps.next,
1231                                          struct ceph_cap, session_caps);
1232                         if (cap == prev)
1233                                 break;
1234                         prev = cap;
1235                         vino = cap->ci->i_vino;
1236                         spin_unlock(&session->s_cap_lock);
1237
1238                         inode = ceph_find_inode(sb, vino);
1239                         iput(inode);
1240
1241                         spin_lock(&session->s_cap_lock);
1242                 }
1243         }
1244
1245         // drop cap expires and unlock s_cap_lock
1246         cleanup_cap_releases(session->s_mdsc, session);
1247
1248         BUG_ON(session->s_nr_caps > 0);
1249         BUG_ON(!list_empty(&session->s_cap_flushing));
1250 }
1251
1252 /*
1253  * wake up any threads waiting on this session's caps.  if the cap is
1254  * old (didn't get renewed on the client reconnect), remove it now.
1255  *
1256  * caller must hold s_mutex.
1257  */
1258 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1259                               void *arg)
1260 {
1261         struct ceph_inode_info *ci = ceph_inode(inode);
1262
1263         if (arg) {
1264                 spin_lock(&ci->i_ceph_lock);
1265                 ci->i_wanted_max_size = 0;
1266                 ci->i_requested_max_size = 0;
1267                 spin_unlock(&ci->i_ceph_lock);
1268         }
1269         wake_up_all(&ci->i_cap_wq);
1270         return 0;
1271 }
1272
1273 static void wake_up_session_caps(struct ceph_mds_session *session,
1274                                  int reconnect)
1275 {
1276         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1277         iterate_session_caps(session, wake_up_session_cb,
1278                              (void *)(unsigned long)reconnect);
1279 }
1280
1281 /*
1282  * Send periodic message to MDS renewing all currently held caps.  The
1283  * ack will reset the expiration for all caps from this session.
1284  *
1285  * caller holds s_mutex
1286  */
1287 static int send_renew_caps(struct ceph_mds_client *mdsc,
1288                            struct ceph_mds_session *session)
1289 {
1290         struct ceph_msg *msg;
1291         int state;
1292
1293         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1294             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1295                 pr_info("mds%d caps stale\n", session->s_mds);
1296         session->s_renew_requested = jiffies;
1297
1298         /* do not try to renew caps until a recovering mds has reconnected
1299          * with its clients. */
1300         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1301         if (state < CEPH_MDS_STATE_RECONNECT) {
1302                 dout("send_renew_caps ignoring mds%d (%s)\n",
1303                      session->s_mds, ceph_mds_state_name(state));
1304                 return 0;
1305         }
1306
1307         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1308                 ceph_mds_state_name(state));
1309         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1310                                  ++session->s_renew_seq);
1311         if (!msg)
1312                 return -ENOMEM;
1313         ceph_con_send(&session->s_con, msg);
1314         return 0;
1315 }
1316
1317 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1318                              struct ceph_mds_session *session, u64 seq)
1319 {
1320         struct ceph_msg *msg;
1321
1322         dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1323              session->s_mds, ceph_session_state_name(session->s_state), seq);
1324         msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1325         if (!msg)
1326                 return -ENOMEM;
1327         ceph_con_send(&session->s_con, msg);
1328         return 0;
1329 }
1330
1331
1332 /*
1333  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1334  *
1335  * Called under session->s_mutex
1336  */
1337 static void renewed_caps(struct ceph_mds_client *mdsc,
1338                          struct ceph_mds_session *session, int is_renew)
1339 {
1340         int was_stale;
1341         int wake = 0;
1342
1343         spin_lock(&session->s_cap_lock);
1344         was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1345
1346         session->s_cap_ttl = session->s_renew_requested +
1347                 mdsc->mdsmap->m_session_timeout*HZ;
1348
1349         if (was_stale) {
1350                 if (time_before(jiffies, session->s_cap_ttl)) {
1351                         pr_info("mds%d caps renewed\n", session->s_mds);
1352                         wake = 1;
1353                 } else {
1354                         pr_info("mds%d caps still stale\n", session->s_mds);
1355                 }
1356         }
1357         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1358              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1359              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1360         spin_unlock(&session->s_cap_lock);
1361
1362         if (wake)
1363                 wake_up_session_caps(session, 0);
1364 }
1365
1366 /*
1367  * send a session close request
1368  */
1369 static int request_close_session(struct ceph_mds_client *mdsc,
1370                                  struct ceph_mds_session *session)
1371 {
1372         struct ceph_msg *msg;
1373
1374         dout("request_close_session mds%d state %s seq %lld\n",
1375              session->s_mds, ceph_session_state_name(session->s_state),
1376              session->s_seq);
1377         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1378         if (!msg)
1379                 return -ENOMEM;
1380         ceph_con_send(&session->s_con, msg);
1381         return 1;
1382 }
1383
1384 /*
1385  * Called with s_mutex held.
1386  */
1387 static int __close_session(struct ceph_mds_client *mdsc,
1388                          struct ceph_mds_session *session)
1389 {
1390         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1391                 return 0;
1392         session->s_state = CEPH_MDS_SESSION_CLOSING;
1393         return request_close_session(mdsc, session);
1394 }
1395
1396 /*
1397  * Trim old(er) caps.
1398  *
1399  * Because we can't cache an inode without one or more caps, we do
1400  * this indirectly: if a cap is unused, we prune its aliases, at which
1401  * point the inode will hopefully get dropped to.
1402  *
1403  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1404  * memory pressure from the MDS, though, so it needn't be perfect.
1405  */
1406 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1407 {
1408         struct ceph_mds_session *session = arg;
1409         struct ceph_inode_info *ci = ceph_inode(inode);
1410         int used, wanted, oissued, mine;
1411
1412         if (session->s_trim_caps <= 0)
1413                 return -1;
1414
1415         spin_lock(&ci->i_ceph_lock);
1416         mine = cap->issued | cap->implemented;
1417         used = __ceph_caps_used(ci);
1418         wanted = __ceph_caps_file_wanted(ci);
1419         oissued = __ceph_caps_issued_other(ci, cap);
1420
1421         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1422              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1423              ceph_cap_string(used), ceph_cap_string(wanted));
1424         if (cap == ci->i_auth_cap) {
1425                 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1426                     !list_empty(&ci->i_cap_snaps))
1427                         goto out;
1428                 if ((used | wanted) & CEPH_CAP_ANY_WR)
1429                         goto out;
1430         }
1431         /* The inode has cached pages, but it's no longer used.
1432          * we can safely drop it */
1433         if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1434             !(oissued & CEPH_CAP_FILE_CACHE)) {
1435           used = 0;
1436           oissued = 0;
1437         }
1438         if ((used | wanted) & ~oissued & mine)
1439                 goto out;   /* we need these caps */
1440
1441         session->s_trim_caps--;
1442         if (oissued) {
1443                 /* we aren't the only cap.. just remove us */
1444                 __ceph_remove_cap(cap, true);
1445         } else {
1446                 /* try dropping referring dentries */
1447                 spin_unlock(&ci->i_ceph_lock);
1448                 d_prune_aliases(inode);
1449                 dout("trim_caps_cb %p cap %p  pruned, count now %d\n",
1450                      inode, cap, atomic_read(&inode->i_count));
1451                 return 0;
1452         }
1453
1454 out:
1455         spin_unlock(&ci->i_ceph_lock);
1456         return 0;
1457 }
1458
1459 /*
1460  * Trim session cap count down to some max number.
1461  */
1462 static int trim_caps(struct ceph_mds_client *mdsc,
1463                      struct ceph_mds_session *session,
1464                      int max_caps)
1465 {
1466         int trim_caps = session->s_nr_caps - max_caps;
1467
1468         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1469              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1470         if (trim_caps > 0) {
1471                 session->s_trim_caps = trim_caps;
1472                 iterate_session_caps(session, trim_caps_cb, session);
1473                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1474                      session->s_mds, session->s_nr_caps, max_caps,
1475                         trim_caps - session->s_trim_caps);
1476                 session->s_trim_caps = 0;
1477         }
1478
1479         ceph_send_cap_releases(mdsc, session);
1480         return 0;
1481 }
1482
1483 static int check_caps_flush(struct ceph_mds_client *mdsc,
1484                             u64 want_flush_tid)
1485 {
1486         int ret = 1;
1487
1488         spin_lock(&mdsc->cap_dirty_lock);
1489         if (!list_empty(&mdsc->cap_flush_list)) {
1490                 struct ceph_cap_flush *cf =
1491                         list_first_entry(&mdsc->cap_flush_list,
1492                                          struct ceph_cap_flush, g_list);
1493                 if (cf->tid <= want_flush_tid) {
1494                         dout("check_caps_flush still flushing tid "
1495                              "%llu <= %llu\n", cf->tid, want_flush_tid);
1496                         ret = 0;
1497                 }
1498         }
1499         spin_unlock(&mdsc->cap_dirty_lock);
1500         return ret;
1501 }
1502
1503 /*
1504  * flush all dirty inode data to disk.
1505  *
1506  * returns true if we've flushed through want_flush_tid
1507  */
1508 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1509                             u64 want_flush_tid)
1510 {
1511         dout("check_caps_flush want %llu\n", want_flush_tid);
1512
1513         wait_event(mdsc->cap_flushing_wq,
1514                    check_caps_flush(mdsc, want_flush_tid));
1515
1516         dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1517 }
1518
1519 /*
1520  * called under s_mutex
1521  */
1522 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1523                             struct ceph_mds_session *session)
1524 {
1525         struct ceph_msg *msg = NULL;
1526         struct ceph_mds_cap_release *head;
1527         struct ceph_mds_cap_item *item;
1528         struct ceph_cap *cap;
1529         LIST_HEAD(tmp_list);
1530         int num_cap_releases;
1531
1532         spin_lock(&session->s_cap_lock);
1533 again:
1534         list_splice_init(&session->s_cap_releases, &tmp_list);
1535         num_cap_releases = session->s_num_cap_releases;
1536         session->s_num_cap_releases = 0;
1537         spin_unlock(&session->s_cap_lock);
1538
1539         while (!list_empty(&tmp_list)) {
1540                 if (!msg) {
1541                         msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1542                                         PAGE_SIZE, GFP_NOFS, false);
1543                         if (!msg)
1544                                 goto out_err;
1545                         head = msg->front.iov_base;
1546                         head->num = cpu_to_le32(0);
1547                         msg->front.iov_len = sizeof(*head);
1548                 }
1549                 cap = list_first_entry(&tmp_list, struct ceph_cap,
1550                                         session_caps);
1551                 list_del(&cap->session_caps);
1552                 num_cap_releases--;
1553
1554                 head = msg->front.iov_base;
1555                 le32_add_cpu(&head->num, 1);
1556                 item = msg->front.iov_base + msg->front.iov_len;
1557                 item->ino = cpu_to_le64(cap->cap_ino);
1558                 item->cap_id = cpu_to_le64(cap->cap_id);
1559                 item->migrate_seq = cpu_to_le32(cap->mseq);
1560                 item->seq = cpu_to_le32(cap->issue_seq);
1561                 msg->front.iov_len += sizeof(*item);
1562
1563                 ceph_put_cap(mdsc, cap);
1564
1565                 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1566                         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1567                         dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1568                         ceph_con_send(&session->s_con, msg);
1569                         msg = NULL;
1570                 }
1571         }
1572
1573         BUG_ON(num_cap_releases != 0);
1574
1575         spin_lock(&session->s_cap_lock);
1576         if (!list_empty(&session->s_cap_releases))
1577                 goto again;
1578         spin_unlock(&session->s_cap_lock);
1579
1580         if (msg) {
1581                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1582                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1583                 ceph_con_send(&session->s_con, msg);
1584         }
1585         return;
1586 out_err:
1587         pr_err("send_cap_releases mds%d, failed to allocate message\n",
1588                 session->s_mds);
1589         spin_lock(&session->s_cap_lock);
1590         list_splice(&tmp_list, &session->s_cap_releases);
1591         session->s_num_cap_releases += num_cap_releases;
1592         spin_unlock(&session->s_cap_lock);
1593 }
1594
1595 /*
1596  * requests
1597  */
1598
1599 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1600                                     struct inode *dir)
1601 {
1602         struct ceph_inode_info *ci = ceph_inode(dir);
1603         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1604         struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1605         size_t size = sizeof(struct ceph_mds_reply_dir_entry);
1606         int order, num_entries;
1607
1608         spin_lock(&ci->i_ceph_lock);
1609         num_entries = ci->i_files + ci->i_subdirs;
1610         spin_unlock(&ci->i_ceph_lock);
1611         num_entries = max(num_entries, 1);
1612         num_entries = min(num_entries, opt->max_readdir);
1613
1614         order = get_order(size * num_entries);
1615         while (order >= 0) {
1616                 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
1617                                                              __GFP_NOWARN,
1618                                                              order);
1619                 if (rinfo->dir_entries)
1620                         break;
1621                 order--;
1622         }
1623         if (!rinfo->dir_entries)
1624                 return -ENOMEM;
1625
1626         num_entries = (PAGE_SIZE << order) / size;
1627         num_entries = min(num_entries, opt->max_readdir);
1628
1629         rinfo->dir_buf_size = PAGE_SIZE << order;
1630         req->r_num_caps = num_entries + 1;
1631         req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1632         req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1633         return 0;
1634 }
1635
1636 /*
1637  * Create an mds request.
1638  */
1639 struct ceph_mds_request *
1640 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1641 {
1642         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1643
1644         if (!req)
1645                 return ERR_PTR(-ENOMEM);
1646
1647         mutex_init(&req->r_fill_mutex);
1648         req->r_mdsc = mdsc;
1649         req->r_started = jiffies;
1650         req->r_resend_mds = -1;
1651         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1652         INIT_LIST_HEAD(&req->r_unsafe_target_item);
1653         req->r_fmode = -1;
1654         kref_init(&req->r_kref);
1655         RB_CLEAR_NODE(&req->r_node);
1656         INIT_LIST_HEAD(&req->r_wait);
1657         init_completion(&req->r_completion);
1658         init_completion(&req->r_safe_completion);
1659         INIT_LIST_HEAD(&req->r_unsafe_item);
1660
1661         req->r_stamp = current_fs_time(mdsc->fsc->sb);
1662
1663         req->r_op = op;
1664         req->r_direct_mode = mode;
1665         return req;
1666 }
1667
1668 /*
1669  * return oldest (lowest) request, tid in request tree, 0 if none.
1670  *
1671  * called under mdsc->mutex.
1672  */
1673 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1674 {
1675         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1676                 return NULL;
1677         return rb_entry(rb_first(&mdsc->request_tree),
1678                         struct ceph_mds_request, r_node);
1679 }
1680
1681 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1682 {
1683         return mdsc->oldest_tid;
1684 }
1685
1686 /*
1687  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1688  * on build_path_from_dentry in fs/cifs/dir.c.
1689  *
1690  * If @stop_on_nosnap, generate path relative to the first non-snapped
1691  * inode.
1692  *
1693  * Encode hidden .snap dirs as a double /, i.e.
1694  *   foo/.snap/bar -> foo//bar
1695  */
1696 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1697                            int stop_on_nosnap)
1698 {
1699         struct dentry *temp;
1700         char *path;
1701         int len, pos;
1702         unsigned seq;
1703
1704         if (dentry == NULL)
1705                 return ERR_PTR(-EINVAL);
1706
1707 retry:
1708         len = 0;
1709         seq = read_seqbegin(&rename_lock);
1710         rcu_read_lock();
1711         for (temp = dentry; !IS_ROOT(temp);) {
1712                 struct inode *inode = d_inode(temp);
1713                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1714                         len++;  /* slash only */
1715                 else if (stop_on_nosnap && inode &&
1716                          ceph_snap(inode) == CEPH_NOSNAP)
1717                         break;
1718                 else
1719                         len += 1 + temp->d_name.len;
1720                 temp = temp->d_parent;
1721         }
1722         rcu_read_unlock();
1723         if (len)
1724                 len--;  /* no leading '/' */
1725
1726         path = kmalloc(len+1, GFP_NOFS);
1727         if (path == NULL)
1728                 return ERR_PTR(-ENOMEM);
1729         pos = len;
1730         path[pos] = 0;  /* trailing null */
1731         rcu_read_lock();
1732         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1733                 struct inode *inode;
1734
1735                 spin_lock(&temp->d_lock);
1736                 inode = d_inode(temp);
1737                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1738                         dout("build_path path+%d: %p SNAPDIR\n",
1739                              pos, temp);
1740                 } else if (stop_on_nosnap && inode &&
1741                            ceph_snap(inode) == CEPH_NOSNAP) {
1742                         spin_unlock(&temp->d_lock);
1743                         break;
1744                 } else {
1745                         pos -= temp->d_name.len;
1746                         if (pos < 0) {
1747                                 spin_unlock(&temp->d_lock);
1748                                 break;
1749                         }
1750                         strncpy(path + pos, temp->d_name.name,
1751                                 temp->d_name.len);
1752                 }
1753                 spin_unlock(&temp->d_lock);
1754                 if (pos)
1755                         path[--pos] = '/';
1756                 temp = temp->d_parent;
1757         }
1758         rcu_read_unlock();
1759         if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1760                 pr_err("build_path did not end path lookup where "
1761                        "expected, namelen is %d, pos is %d\n", len, pos);
1762                 /* presumably this is only possible if racing with a
1763                    rename of one of the parent directories (we can not
1764                    lock the dentries above us to prevent this, but
1765                    retrying should be harmless) */
1766                 kfree(path);
1767                 goto retry;
1768         }
1769
1770         *base = ceph_ino(d_inode(temp));
1771         *plen = len;
1772         dout("build_path on %p %d built %llx '%.*s'\n",
1773              dentry, d_count(dentry), *base, len, path);
1774         return path;
1775 }
1776
1777 static int build_dentry_path(struct dentry *dentry,
1778                              const char **ppath, int *ppathlen, u64 *pino,
1779                              int *pfreepath)
1780 {
1781         char *path;
1782
1783         if (ceph_snap(d_inode(dentry->d_parent)) == CEPH_NOSNAP) {
1784                 *pino = ceph_ino(d_inode(dentry->d_parent));
1785                 *ppath = dentry->d_name.name;
1786                 *ppathlen = dentry->d_name.len;
1787                 return 0;
1788         }
1789         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1790         if (IS_ERR(path))
1791                 return PTR_ERR(path);
1792         *ppath = path;
1793         *pfreepath = 1;
1794         return 0;
1795 }
1796
1797 static int build_inode_path(struct inode *inode,
1798                             const char **ppath, int *ppathlen, u64 *pino,
1799                             int *pfreepath)
1800 {
1801         struct dentry *dentry;
1802         char *path;
1803
1804         if (ceph_snap(inode) == CEPH_NOSNAP) {
1805                 *pino = ceph_ino(inode);
1806                 *ppathlen = 0;
1807                 return 0;
1808         }
1809         dentry = d_find_alias(inode);
1810         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1811         dput(dentry);
1812         if (IS_ERR(path))
1813                 return PTR_ERR(path);
1814         *ppath = path;
1815         *pfreepath = 1;
1816         return 0;
1817 }
1818
1819 /*
1820  * request arguments may be specified via an inode *, a dentry *, or
1821  * an explicit ino+path.
1822  */
1823 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1824                                   const char *rpath, u64 rino,
1825                                   const char **ppath, int *pathlen,
1826                                   u64 *ino, int *freepath)
1827 {
1828         int r = 0;
1829
1830         if (rinode) {
1831                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1832                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1833                      ceph_snap(rinode));
1834         } else if (rdentry) {
1835                 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1836                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1837                      *ppath);
1838         } else if (rpath || rino) {
1839                 *ino = rino;
1840                 *ppath = rpath;
1841                 *pathlen = rpath ? strlen(rpath) : 0;
1842                 dout(" path %.*s\n", *pathlen, rpath);
1843         }
1844
1845         return r;
1846 }
1847
1848 /*
1849  * called under mdsc->mutex
1850  */
1851 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1852                                                struct ceph_mds_request *req,
1853                                                int mds, bool drop_cap_releases)
1854 {
1855         struct ceph_msg *msg;
1856         struct ceph_mds_request_head *head;
1857         const char *path1 = NULL;
1858         const char *path2 = NULL;
1859         u64 ino1 = 0, ino2 = 0;
1860         int pathlen1 = 0, pathlen2 = 0;
1861         int freepath1 = 0, freepath2 = 0;
1862         int len;
1863         u16 releases;
1864         void *p, *end;
1865         int ret;
1866
1867         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1868                               req->r_path1, req->r_ino1.ino,
1869                               &path1, &pathlen1, &ino1, &freepath1);
1870         if (ret < 0) {
1871                 msg = ERR_PTR(ret);
1872                 goto out;
1873         }
1874
1875         ret = set_request_path_attr(NULL, req->r_old_dentry,
1876                               req->r_path2, req->r_ino2.ino,
1877                               &path2, &pathlen2, &ino2, &freepath2);
1878         if (ret < 0) {
1879                 msg = ERR_PTR(ret);
1880                 goto out_free1;
1881         }
1882
1883         len = sizeof(*head) +
1884                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
1885                 sizeof(struct ceph_timespec);
1886
1887         /* calculate (max) length for cap releases */
1888         len += sizeof(struct ceph_mds_request_release) *
1889                 (!!req->r_inode_drop + !!req->r_dentry_drop +
1890                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1891         if (req->r_dentry_drop)
1892                 len += req->r_dentry->d_name.len;
1893         if (req->r_old_dentry_drop)
1894                 len += req->r_old_dentry->d_name.len;
1895
1896         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1897         if (!msg) {
1898                 msg = ERR_PTR(-ENOMEM);
1899                 goto out_free2;
1900         }
1901
1902         msg->hdr.version = cpu_to_le16(2);
1903         msg->hdr.tid = cpu_to_le64(req->r_tid);
1904
1905         head = msg->front.iov_base;
1906         p = msg->front.iov_base + sizeof(*head);
1907         end = msg->front.iov_base + msg->front.iov_len;
1908
1909         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1910         head->op = cpu_to_le32(req->r_op);
1911         head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1912         head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1913         head->args = req->r_args;
1914
1915         ceph_encode_filepath(&p, end, ino1, path1);
1916         ceph_encode_filepath(&p, end, ino2, path2);
1917
1918         /* make note of release offset, in case we need to replay */
1919         req->r_request_release_offset = p - msg->front.iov_base;
1920
1921         /* cap releases */
1922         releases = 0;
1923         if (req->r_inode_drop)
1924                 releases += ceph_encode_inode_release(&p,
1925                       req->r_inode ? req->r_inode : d_inode(req->r_dentry),
1926                       mds, req->r_inode_drop, req->r_inode_unless, 0);
1927         if (req->r_dentry_drop)
1928                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1929                        mds, req->r_dentry_drop, req->r_dentry_unless);
1930         if (req->r_old_dentry_drop)
1931                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1932                        mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1933         if (req->r_old_inode_drop)
1934                 releases += ceph_encode_inode_release(&p,
1935                       d_inode(req->r_old_dentry),
1936                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1937
1938         if (drop_cap_releases) {
1939                 releases = 0;
1940                 p = msg->front.iov_base + req->r_request_release_offset;
1941         }
1942
1943         head->num_releases = cpu_to_le16(releases);
1944
1945         /* time stamp */
1946         {
1947                 struct ceph_timespec ts;
1948                 ceph_encode_timespec(&ts, &req->r_stamp);
1949                 ceph_encode_copy(&p, &ts, sizeof(ts));
1950         }
1951
1952         BUG_ON(p > end);
1953         msg->front.iov_len = p - msg->front.iov_base;
1954         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1955
1956         if (req->r_pagelist) {
1957                 struct ceph_pagelist *pagelist = req->r_pagelist;
1958                 atomic_inc(&pagelist->refcnt);
1959                 ceph_msg_data_add_pagelist(msg, pagelist);
1960                 msg->hdr.data_len = cpu_to_le32(pagelist->length);
1961         } else {
1962                 msg->hdr.data_len = 0;
1963         }
1964
1965         msg->hdr.data_off = cpu_to_le16(0);
1966
1967 out_free2:
1968         if (freepath2)
1969                 kfree((char *)path2);
1970 out_free1:
1971         if (freepath1)
1972                 kfree((char *)path1);
1973 out:
1974         return msg;
1975 }
1976
1977 /*
1978  * called under mdsc->mutex if error, under no mutex if
1979  * success.
1980  */
1981 static void complete_request(struct ceph_mds_client *mdsc,
1982                              struct ceph_mds_request *req)
1983 {
1984         if (req->r_callback)
1985                 req->r_callback(mdsc, req);
1986         else
1987                 complete_all(&req->r_completion);
1988 }
1989
1990 /*
1991  * called under mdsc->mutex
1992  */
1993 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1994                                   struct ceph_mds_request *req,
1995                                   int mds, bool drop_cap_releases)
1996 {
1997         struct ceph_mds_request_head *rhead;
1998         struct ceph_msg *msg;
1999         int flags = 0;
2000
2001         req->r_attempts++;
2002         if (req->r_inode) {
2003                 struct ceph_cap *cap =
2004                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2005
2006                 if (cap)
2007                         req->r_sent_on_mseq = cap->mseq;
2008                 else
2009                         req->r_sent_on_mseq = -1;
2010         }
2011         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2012              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2013
2014         if (req->r_got_unsafe) {
2015                 void *p;
2016                 /*
2017                  * Replay.  Do not regenerate message (and rebuild
2018                  * paths, etc.); just use the original message.
2019                  * Rebuilding paths will break for renames because
2020                  * d_move mangles the src name.
2021                  */
2022                 msg = req->r_request;
2023                 rhead = msg->front.iov_base;
2024
2025                 flags = le32_to_cpu(rhead->flags);
2026                 flags |= CEPH_MDS_FLAG_REPLAY;
2027                 rhead->flags = cpu_to_le32(flags);
2028
2029                 if (req->r_target_inode)
2030                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2031
2032                 rhead->num_retry = req->r_attempts - 1;
2033
2034                 /* remove cap/dentry releases from message */
2035                 rhead->num_releases = 0;
2036
2037                 /* time stamp */
2038                 p = msg->front.iov_base + req->r_request_release_offset;
2039                 {
2040                         struct ceph_timespec ts;
2041                         ceph_encode_timespec(&ts, &req->r_stamp);
2042                         ceph_encode_copy(&p, &ts, sizeof(ts));
2043                 }
2044
2045                 msg->front.iov_len = p - msg->front.iov_base;
2046                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2047                 return 0;
2048         }
2049
2050         if (req->r_request) {
2051                 ceph_msg_put(req->r_request);
2052                 req->r_request = NULL;
2053         }
2054         msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2055         if (IS_ERR(msg)) {
2056                 req->r_err = PTR_ERR(msg);
2057                 return PTR_ERR(msg);
2058         }
2059         req->r_request = msg;
2060
2061         rhead = msg->front.iov_base;
2062         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2063         if (req->r_got_unsafe)
2064                 flags |= CEPH_MDS_FLAG_REPLAY;
2065         if (req->r_locked_dir)
2066                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2067         rhead->flags = cpu_to_le32(flags);
2068         rhead->num_fwd = req->r_num_fwd;
2069         rhead->num_retry = req->r_attempts - 1;
2070         rhead->ino = 0;
2071
2072         dout(" r_locked_dir = %p\n", req->r_locked_dir);
2073         return 0;
2074 }
2075
2076 /*
2077  * send request, or put it on the appropriate wait list.
2078  */
2079 static int __do_request(struct ceph_mds_client *mdsc,
2080                         struct ceph_mds_request *req)
2081 {
2082         struct ceph_mds_session *session = NULL;
2083         int mds = -1;
2084         int err = 0;
2085
2086         if (req->r_err || req->r_got_result) {
2087                 if (req->r_aborted)
2088                         __unregister_request(mdsc, req);
2089                 goto out;
2090         }
2091
2092         if (req->r_timeout &&
2093             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2094                 dout("do_request timed out\n");
2095                 err = -EIO;
2096                 goto finish;
2097         }
2098         if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2099                 dout("do_request forced umount\n");
2100                 err = -EIO;
2101                 goto finish;
2102         }
2103
2104         put_request_session(req);
2105
2106         mds = __choose_mds(mdsc, req);
2107         if (mds < 0 ||
2108             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2109                 if (mdsc->mdsmap_err) {
2110                         err = mdsc->mdsmap_err;
2111                         dout("do_request mdsmap err %d\n", err);
2112                         goto finish;
2113                 }
2114                 dout("do_request no mds or not active, waiting for map\n");
2115                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2116                 goto out;
2117         }
2118
2119         /* get, open session */
2120         session = __ceph_lookup_mds_session(mdsc, mds);
2121         if (!session) {
2122                 session = register_session(mdsc, mds);
2123                 if (IS_ERR(session)) {
2124                         err = PTR_ERR(session);
2125                         goto finish;
2126                 }
2127         }
2128         req->r_session = get_session(session);
2129
2130         dout("do_request mds%d session %p state %s\n", mds, session,
2131              ceph_session_state_name(session->s_state));
2132         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2133             session->s_state != CEPH_MDS_SESSION_HUNG) {
2134                 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2135                         err = -EACCES;
2136                         goto out_session;
2137                 }
2138                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2139                     session->s_state == CEPH_MDS_SESSION_CLOSING)
2140                         __open_session(mdsc, session);
2141                 list_add(&req->r_wait, &session->s_waiting);
2142                 goto out_session;
2143         }
2144
2145         /* send request */
2146         req->r_resend_mds = -1;   /* forget any previous mds hint */
2147
2148         if (req->r_request_started == 0)   /* note request start time */
2149                 req->r_request_started = jiffies;
2150
2151         err = __prepare_send_request(mdsc, req, mds, false);
2152         if (!err) {
2153                 ceph_msg_get(req->r_request);
2154                 ceph_con_send(&session->s_con, req->r_request);
2155         }
2156
2157 out_session:
2158         ceph_put_mds_session(session);
2159 finish:
2160         if (err) {
2161                 dout("__do_request early error %d\n", err);
2162                 req->r_err = err;
2163                 complete_request(mdsc, req);
2164                 __unregister_request(mdsc, req);
2165         }
2166 out:
2167         return err;
2168 }
2169
2170 /*
2171  * called under mdsc->mutex
2172  */
2173 static void __wake_requests(struct ceph_mds_client *mdsc,
2174                             struct list_head *head)
2175 {
2176         struct ceph_mds_request *req;
2177         LIST_HEAD(tmp_list);
2178
2179         list_splice_init(head, &tmp_list);
2180
2181         while (!list_empty(&tmp_list)) {
2182                 req = list_entry(tmp_list.next,
2183                                  struct ceph_mds_request, r_wait);
2184                 list_del_init(&req->r_wait);
2185                 dout(" wake request %p tid %llu\n", req, req->r_tid);
2186                 __do_request(mdsc, req);
2187         }
2188 }
2189
2190 /*
2191  * Wake up threads with requests pending for @mds, so that they can
2192  * resubmit their requests to a possibly different mds.
2193  */
2194 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2195 {
2196         struct ceph_mds_request *req;
2197         struct rb_node *p = rb_first(&mdsc->request_tree);
2198
2199         dout("kick_requests mds%d\n", mds);
2200         while (p) {
2201                 req = rb_entry(p, struct ceph_mds_request, r_node);
2202                 p = rb_next(p);
2203                 if (req->r_got_unsafe)
2204                         continue;
2205                 if (req->r_attempts > 0)
2206                         continue; /* only new requests */
2207                 if (req->r_session &&
2208                     req->r_session->s_mds == mds) {
2209                         dout(" kicking tid %llu\n", req->r_tid);
2210                         list_del_init(&req->r_wait);
2211                         __do_request(mdsc, req);
2212                 }
2213         }
2214 }
2215
2216 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2217                               struct ceph_mds_request *req)
2218 {
2219         dout("submit_request on %p\n", req);
2220         mutex_lock(&mdsc->mutex);
2221         __register_request(mdsc, req, NULL);
2222         __do_request(mdsc, req);
2223         mutex_unlock(&mdsc->mutex);
2224 }
2225
2226 /*
2227  * Synchrously perform an mds request.  Take care of all of the
2228  * session setup, forwarding, retry details.
2229  */
2230 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2231                          struct inode *dir,
2232                          struct ceph_mds_request *req)
2233 {
2234         int err;
2235
2236         dout("do_request on %p\n", req);
2237
2238         /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
2239         if (req->r_inode)
2240                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2241         if (req->r_locked_dir)
2242                 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
2243         if (req->r_old_dentry_dir)
2244                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2245                                   CEPH_CAP_PIN);
2246
2247         /* issue */
2248         mutex_lock(&mdsc->mutex);
2249         __register_request(mdsc, req, dir);
2250         __do_request(mdsc, req);
2251
2252         if (req->r_err) {
2253                 err = req->r_err;
2254                 goto out;
2255         }
2256
2257         /* wait */
2258         mutex_unlock(&mdsc->mutex);
2259         dout("do_request waiting\n");
2260         if (!req->r_timeout && req->r_wait_for_completion) {
2261                 err = req->r_wait_for_completion(mdsc, req);
2262         } else {
2263                 long timeleft = wait_for_completion_killable_timeout(
2264                                         &req->r_completion,
2265                                         ceph_timeout_jiffies(req->r_timeout));
2266                 if (timeleft > 0)
2267                         err = 0;
2268                 else if (!timeleft)
2269                         err = -EIO;  /* timed out */
2270                 else
2271                         err = timeleft;  /* killed */
2272         }
2273         dout("do_request waited, got %d\n", err);
2274         mutex_lock(&mdsc->mutex);
2275
2276         /* only abort if we didn't race with a real reply */
2277         if (req->r_got_result) {
2278                 err = le32_to_cpu(req->r_reply_info.head->result);
2279         } else if (err < 0) {
2280                 dout("aborted request %lld with %d\n", req->r_tid, err);
2281
2282                 /*
2283                  * ensure we aren't running concurrently with
2284                  * ceph_fill_trace or ceph_readdir_prepopulate, which
2285                  * rely on locks (dir mutex) held by our caller.
2286                  */
2287                 mutex_lock(&req->r_fill_mutex);
2288                 req->r_err = err;
2289                 req->r_aborted = true;
2290                 mutex_unlock(&req->r_fill_mutex);
2291
2292                 if (req->r_locked_dir &&
2293                     (req->r_op & CEPH_MDS_OP_WRITE))
2294                         ceph_invalidate_dir_request(req);
2295         } else {
2296                 err = req->r_err;
2297         }
2298
2299 out:
2300         mutex_unlock(&mdsc->mutex);
2301         dout("do_request %p done, result %d\n", req, err);
2302         return err;
2303 }
2304
2305 /*
2306  * Invalidate dir's completeness, dentry lease state on an aborted MDS
2307  * namespace request.
2308  */
2309 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2310 {
2311         struct inode *inode = req->r_locked_dir;
2312
2313         dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2314
2315         ceph_dir_clear_complete(inode);
2316         if (req->r_dentry)
2317                 ceph_invalidate_dentry_lease(req->r_dentry);
2318         if (req->r_old_dentry)
2319                 ceph_invalidate_dentry_lease(req->r_old_dentry);
2320 }
2321
2322 /*
2323  * Handle mds reply.
2324  *
2325  * We take the session mutex and parse and process the reply immediately.
2326  * This preserves the logical ordering of replies, capabilities, etc., sent
2327  * by the MDS as they are applied to our local cache.
2328  */
2329 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2330 {
2331         struct ceph_mds_client *mdsc = session->s_mdsc;
2332         struct ceph_mds_request *req;
2333         struct ceph_mds_reply_head *head = msg->front.iov_base;
2334         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2335         struct ceph_snap_realm *realm;
2336         u64 tid;
2337         int err, result;
2338         int mds = session->s_mds;
2339
2340         if (msg->front.iov_len < sizeof(*head)) {
2341                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2342                 ceph_msg_dump(msg);
2343                 return;
2344         }
2345
2346         /* get request, session */
2347         tid = le64_to_cpu(msg->hdr.tid);
2348         mutex_lock(&mdsc->mutex);
2349         req = lookup_get_request(mdsc, tid);
2350         if (!req) {
2351                 dout("handle_reply on unknown tid %llu\n", tid);
2352                 mutex_unlock(&mdsc->mutex);
2353                 return;
2354         }
2355         dout("handle_reply %p\n", req);
2356
2357         /* correct session? */
2358         if (req->r_session != session) {
2359                 pr_err("mdsc_handle_reply got %llu on session mds%d"
2360                        " not mds%d\n", tid, session->s_mds,
2361                        req->r_session ? req->r_session->s_mds : -1);
2362                 mutex_unlock(&mdsc->mutex);
2363                 goto out;
2364         }
2365
2366         /* dup? */
2367         if ((req->r_got_unsafe && !head->safe) ||
2368             (req->r_got_safe && head->safe)) {
2369                 pr_warn("got a dup %s reply on %llu from mds%d\n",
2370                            head->safe ? "safe" : "unsafe", tid, mds);
2371                 mutex_unlock(&mdsc->mutex);
2372                 goto out;
2373         }
2374         if (req->r_got_safe) {
2375                 pr_warn("got unsafe after safe on %llu from mds%d\n",
2376                            tid, mds);
2377                 mutex_unlock(&mdsc->mutex);
2378                 goto out;
2379         }
2380
2381         result = le32_to_cpu(head->result);
2382
2383         /*
2384          * Handle an ESTALE
2385          * if we're not talking to the authority, send to them
2386          * if the authority has changed while we weren't looking,
2387          * send to new authority
2388          * Otherwise we just have to return an ESTALE
2389          */
2390         if (result == -ESTALE) {
2391                 dout("got ESTALE on request %llu", req->r_tid);
2392                 req->r_resend_mds = -1;
2393                 if (req->r_direct_mode != USE_AUTH_MDS) {
2394                         dout("not using auth, setting for that now");
2395                         req->r_direct_mode = USE_AUTH_MDS;
2396                         __do_request(mdsc, req);
2397                         mutex_unlock(&mdsc->mutex);
2398                         goto out;
2399                 } else  {
2400                         int mds = __choose_mds(mdsc, req);
2401                         if (mds >= 0 && mds != req->r_session->s_mds) {
2402                                 dout("but auth changed, so resending");
2403                                 __do_request(mdsc, req);
2404                                 mutex_unlock(&mdsc->mutex);
2405                                 goto out;
2406                         }
2407                 }
2408                 dout("have to return ESTALE on request %llu", req->r_tid);
2409         }
2410
2411
2412         if (head->safe) {
2413                 req->r_got_safe = true;
2414                 __unregister_request(mdsc, req);
2415
2416                 if (req->r_got_unsafe) {
2417                         /*
2418                          * We already handled the unsafe response, now do the
2419                          * cleanup.  No need to examine the response; the MDS
2420                          * doesn't include any result info in the safe
2421                          * response.  And even if it did, there is nothing
2422                          * useful we could do with a revised return value.
2423                          */
2424                         dout("got safe reply %llu, mds%d\n", tid, mds);
2425                         list_del_init(&req->r_unsafe_item);
2426
2427                         /* last unsafe request during umount? */
2428                         if (mdsc->stopping && !__get_oldest_req(mdsc))
2429                                 complete_all(&mdsc->safe_umount_waiters);
2430                         mutex_unlock(&mdsc->mutex);
2431                         goto out;
2432                 }
2433         } else {
2434                 req->r_got_unsafe = true;
2435                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2436                 if (req->r_unsafe_dir) {
2437                         struct ceph_inode_info *ci =
2438                                         ceph_inode(req->r_unsafe_dir);
2439                         spin_lock(&ci->i_unsafe_lock);
2440                         list_add_tail(&req->r_unsafe_dir_item,
2441                                       &ci->i_unsafe_dirops);
2442                         spin_unlock(&ci->i_unsafe_lock);
2443                 }
2444         }
2445
2446         dout("handle_reply tid %lld result %d\n", tid, result);
2447         rinfo = &req->r_reply_info;
2448         err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2449         mutex_unlock(&mdsc->mutex);
2450
2451         mutex_lock(&session->s_mutex);
2452         if (err < 0) {
2453                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2454                 ceph_msg_dump(msg);
2455                 goto out_err;
2456         }
2457
2458         /* snap trace */
2459         realm = NULL;
2460         if (rinfo->snapblob_len) {
2461                 down_write(&mdsc->snap_rwsem);
2462                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2463                                 rinfo->snapblob + rinfo->snapblob_len,
2464                                 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2465                                 &realm);
2466                 downgrade_write(&mdsc->snap_rwsem);
2467         } else {
2468                 down_read(&mdsc->snap_rwsem);
2469         }
2470
2471         /* insert trace into our cache */
2472         mutex_lock(&req->r_fill_mutex);
2473         current->journal_info = req;
2474         err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2475         if (err == 0) {
2476                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2477                                     req->r_op == CEPH_MDS_OP_LSSNAP))
2478                         ceph_readdir_prepopulate(req, req->r_session);
2479                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2480         }
2481         current->journal_info = NULL;
2482         mutex_unlock(&req->r_fill_mutex);
2483
2484         up_read(&mdsc->snap_rwsem);
2485         if (realm)
2486                 ceph_put_snap_realm(mdsc, realm);
2487
2488         if (err == 0 && req->r_got_unsafe && req->r_target_inode) {
2489                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
2490                 spin_lock(&ci->i_unsafe_lock);
2491                 list_add_tail(&req->r_unsafe_target_item, &ci->i_unsafe_iops);
2492                 spin_unlock(&ci->i_unsafe_lock);
2493         }
2494 out_err:
2495         mutex_lock(&mdsc->mutex);
2496         if (!req->r_aborted) {
2497                 if (err) {
2498                         req->r_err = err;
2499                 } else {
2500                         req->r_reply =  ceph_msg_get(msg);
2501                         req->r_got_result = true;
2502                 }
2503         } else {
2504                 dout("reply arrived after request %lld was aborted\n", tid);
2505         }
2506         mutex_unlock(&mdsc->mutex);
2507
2508         mutex_unlock(&session->s_mutex);
2509
2510         /* kick calling process */
2511         complete_request(mdsc, req);
2512 out:
2513         ceph_mdsc_put_request(req);
2514         return;
2515 }
2516
2517
2518
2519 /*
2520  * handle mds notification that our request has been forwarded.
2521  */
2522 static void handle_forward(struct ceph_mds_client *mdsc,
2523                            struct ceph_mds_session *session,
2524                            struct ceph_msg *msg)
2525 {
2526         struct ceph_mds_request *req;
2527         u64 tid = le64_to_cpu(msg->hdr.tid);
2528         u32 next_mds;
2529         u32 fwd_seq;
2530         int err = -EINVAL;
2531         void *p = msg->front.iov_base;
2532         void *end = p + msg->front.iov_len;
2533
2534         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2535         next_mds = ceph_decode_32(&p);
2536         fwd_seq = ceph_decode_32(&p);
2537
2538         mutex_lock(&mdsc->mutex);
2539         req = lookup_get_request(mdsc, tid);
2540         if (!req) {
2541                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2542                 goto out;  /* dup reply? */
2543         }
2544
2545         if (req->r_aborted) {
2546                 dout("forward tid %llu aborted, unregistering\n", tid);
2547                 __unregister_request(mdsc, req);
2548         } else if (fwd_seq <= req->r_num_fwd) {
2549                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2550                      tid, next_mds, req->r_num_fwd, fwd_seq);
2551         } else {
2552                 /* resend. forward race not possible; mds would drop */
2553                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2554                 BUG_ON(req->r_err);
2555                 BUG_ON(req->r_got_result);
2556                 req->r_attempts = 0;
2557                 req->r_num_fwd = fwd_seq;
2558                 req->r_resend_mds = next_mds;
2559                 put_request_session(req);
2560                 __do_request(mdsc, req);
2561         }
2562         ceph_mdsc_put_request(req);
2563 out:
2564         mutex_unlock(&mdsc->mutex);
2565         return;
2566
2567 bad:
2568         pr_err("mdsc_handle_forward decode error err=%d\n", err);
2569 }
2570
2571 /*
2572  * handle a mds session control message
2573  */
2574 static void handle_session(struct ceph_mds_session *session,
2575                            struct ceph_msg *msg)
2576 {
2577         struct ceph_mds_client *mdsc = session->s_mdsc;
2578         u32 op;
2579         u64 seq;
2580         int mds = session->s_mds;
2581         struct ceph_mds_session_head *h = msg->front.iov_base;
2582         int wake = 0;
2583
2584         /* decode */
2585         if (msg->front.iov_len != sizeof(*h))
2586                 goto bad;
2587         op = le32_to_cpu(h->op);
2588         seq = le64_to_cpu(h->seq);
2589
2590         mutex_lock(&mdsc->mutex);
2591         if (op == CEPH_SESSION_CLOSE)
2592                 __unregister_session(mdsc, session);
2593         /* FIXME: this ttl calculation is generous */
2594         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2595         mutex_unlock(&mdsc->mutex);
2596
2597         mutex_lock(&session->s_mutex);
2598
2599         dout("handle_session mds%d %s %p state %s seq %llu\n",
2600              mds, ceph_session_op_name(op), session,
2601              ceph_session_state_name(session->s_state), seq);
2602
2603         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2604                 session->s_state = CEPH_MDS_SESSION_OPEN;
2605                 pr_info("mds%d came back\n", session->s_mds);
2606         }
2607
2608         switch (op) {
2609         case CEPH_SESSION_OPEN:
2610                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2611                         pr_info("mds%d reconnect success\n", session->s_mds);
2612                 session->s_state = CEPH_MDS_SESSION_OPEN;
2613                 renewed_caps(mdsc, session, 0);
2614                 wake = 1;
2615                 if (mdsc->stopping)
2616                         __close_session(mdsc, session);
2617                 break;
2618
2619         case CEPH_SESSION_RENEWCAPS:
2620                 if (session->s_renew_seq == seq)
2621                         renewed_caps(mdsc, session, 1);
2622                 break;
2623
2624         case CEPH_SESSION_CLOSE:
2625                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2626                         pr_info("mds%d reconnect denied\n", session->s_mds);
2627                 cleanup_session_requests(mdsc, session);
2628                 remove_session_caps(session);
2629                 wake = 2; /* for good measure */
2630                 wake_up_all(&mdsc->session_close_wq);
2631                 break;
2632
2633         case CEPH_SESSION_STALE:
2634                 pr_info("mds%d caps went stale, renewing\n",
2635                         session->s_mds);
2636                 spin_lock(&session->s_gen_ttl_lock);
2637                 session->s_cap_gen++;
2638                 session->s_cap_ttl = jiffies - 1;
2639                 spin_unlock(&session->s_gen_ttl_lock);
2640                 send_renew_caps(mdsc, session);
2641                 break;
2642
2643         case CEPH_SESSION_RECALL_STATE:
2644                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2645                 break;
2646
2647         case CEPH_SESSION_FLUSHMSG:
2648                 send_flushmsg_ack(mdsc, session, seq);
2649                 break;
2650
2651         case CEPH_SESSION_FORCE_RO:
2652                 dout("force_session_readonly %p\n", session);
2653                 spin_lock(&session->s_cap_lock);
2654                 session->s_readonly = true;
2655                 spin_unlock(&session->s_cap_lock);
2656                 wake_up_session_caps(session, 0);
2657                 break;
2658
2659         case CEPH_SESSION_REJECT:
2660                 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
2661                 pr_info("mds%d rejected session\n", session->s_mds);
2662                 session->s_state = CEPH_MDS_SESSION_REJECTED;
2663                 cleanup_session_requests(mdsc, session);
2664                 remove_session_caps(session);
2665                 wake = 2; /* for good measure */
2666                 break;
2667
2668         default:
2669                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2670                 WARN_ON(1);
2671         }
2672
2673         mutex_unlock(&session->s_mutex);
2674         if (wake) {
2675                 mutex_lock(&mdsc->mutex);
2676                 __wake_requests(mdsc, &session->s_waiting);
2677                 if (wake == 2)
2678                         kick_requests(mdsc, mds);
2679                 mutex_unlock(&mdsc->mutex);
2680         }
2681         return;
2682
2683 bad:
2684         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2685                (int)msg->front.iov_len);
2686         ceph_msg_dump(msg);
2687         return;
2688 }
2689
2690
2691 /*
2692  * called under session->mutex.
2693  */
2694 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2695                                    struct ceph_mds_session *session)
2696 {
2697         struct ceph_mds_request *req, *nreq;
2698         struct rb_node *p;
2699         int err;
2700
2701         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2702
2703         mutex_lock(&mdsc->mutex);
2704         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2705                 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2706                 if (!err) {
2707                         ceph_msg_get(req->r_request);
2708                         ceph_con_send(&session->s_con, req->r_request);
2709                 }
2710         }
2711
2712         /*
2713          * also re-send old requests when MDS enters reconnect stage. So that MDS
2714          * can process completed request in clientreplay stage.
2715          */
2716         p = rb_first(&mdsc->request_tree);
2717         while (p) {
2718                 req = rb_entry(p, struct ceph_mds_request, r_node);
2719                 p = rb_next(p);
2720                 if (req->r_got_unsafe)
2721                         continue;
2722                 if (req->r_attempts == 0)
2723                         continue; /* only old requests */
2724                 if (req->r_session &&
2725                     req->r_session->s_mds == session->s_mds) {
2726                         err = __prepare_send_request(mdsc, req,
2727                                                      session->s_mds, true);
2728                         if (!err) {
2729                                 ceph_msg_get(req->r_request);
2730                                 ceph_con_send(&session->s_con, req->r_request);
2731                         }
2732                 }
2733         }
2734         mutex_unlock(&mdsc->mutex);
2735 }
2736
2737 /*
2738  * Encode information about a cap for a reconnect with the MDS.
2739  */
2740 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2741                           void *arg)
2742 {
2743         union {
2744                 struct ceph_mds_cap_reconnect v2;
2745                 struct ceph_mds_cap_reconnect_v1 v1;
2746         } rec;
2747         struct ceph_inode_info *ci;
2748         struct ceph_reconnect_state *recon_state = arg;
2749         struct ceph_pagelist *pagelist = recon_state->pagelist;
2750         char *path;
2751         int pathlen, err;
2752         u64 pathbase;
2753         u64 snap_follows;
2754         struct dentry *dentry;
2755
2756         ci = cap->ci;
2757
2758         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2759              inode, ceph_vinop(inode), cap, cap->cap_id,
2760              ceph_cap_string(cap->issued));
2761         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2762         if (err)
2763                 return err;
2764
2765         dentry = d_find_alias(inode);
2766         if (dentry) {
2767                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2768                 if (IS_ERR(path)) {
2769                         err = PTR_ERR(path);
2770                         goto out_dput;
2771                 }
2772         } else {
2773                 path = NULL;
2774                 pathlen = 0;
2775                 pathbase = 0;
2776         }
2777
2778         spin_lock(&ci->i_ceph_lock);
2779         cap->seq = 0;        /* reset cap seq */
2780         cap->issue_seq = 0;  /* and issue_seq */
2781         cap->mseq = 0;       /* and migrate_seq */
2782         cap->cap_gen = cap->session->s_cap_gen;
2783
2784         if (recon_state->msg_version >= 2) {
2785                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2786                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2787                 rec.v2.issued = cpu_to_le32(cap->issued);
2788                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2789                 rec.v2.pathbase = cpu_to_le64(pathbase);
2790                 rec.v2.flock_len = 0;
2791         } else {
2792                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2793                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2794                 rec.v1.issued = cpu_to_le32(cap->issued);
2795                 rec.v1.size = cpu_to_le64(inode->i_size);
2796                 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2797                 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2798                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2799                 rec.v1.pathbase = cpu_to_le64(pathbase);
2800         }
2801
2802         if (list_empty(&ci->i_cap_snaps)) {
2803                 snap_follows = 0;
2804         } else {
2805                 struct ceph_cap_snap *capsnap =
2806                         list_first_entry(&ci->i_cap_snaps,
2807                                          struct ceph_cap_snap, ci_item);
2808                 snap_follows = capsnap->follows;
2809         }
2810         spin_unlock(&ci->i_ceph_lock);
2811
2812         if (recon_state->msg_version >= 2) {
2813                 int num_fcntl_locks, num_flock_locks;
2814                 struct ceph_filelock *flocks;
2815                 size_t struct_len, total_len = 0;
2816                 u8 struct_v = 0;
2817
2818 encode_again:
2819                 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2820                 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2821                                  sizeof(struct ceph_filelock), GFP_NOFS);
2822                 if (!flocks) {
2823                         err = -ENOMEM;
2824                         goto out_free;
2825                 }
2826                 err = ceph_encode_locks_to_buffer(inode, flocks,
2827                                                   num_fcntl_locks,
2828                                                   num_flock_locks);
2829                 if (err) {
2830                         kfree(flocks);
2831                         if (err == -ENOSPC)
2832                                 goto encode_again;
2833                         goto out_free;
2834                 }
2835
2836                 if (recon_state->msg_version >= 3) {
2837                         /* version, compat_version and struct_len */
2838                         total_len = 2 * sizeof(u8) + sizeof(u32);
2839                         struct_v = 2;
2840                 }
2841                 /*
2842                  * number of encoded locks is stable, so copy to pagelist
2843                  */
2844                 struct_len = 2 * sizeof(u32) +
2845                             (num_fcntl_locks + num_flock_locks) *
2846                             sizeof(struct ceph_filelock);
2847                 rec.v2.flock_len = cpu_to_le32(struct_len);
2848
2849                 struct_len += sizeof(rec.v2);
2850                 struct_len += sizeof(u32) + pathlen;
2851
2852                 if (struct_v >= 2)
2853                         struct_len += sizeof(u64); /* snap_follows */
2854
2855                 total_len += struct_len;
2856                 err = ceph_pagelist_reserve(pagelist, total_len);
2857
2858                 if (!err) {
2859                         if (recon_state->msg_version >= 3) {
2860                                 ceph_pagelist_encode_8(pagelist, struct_v);
2861                                 ceph_pagelist_encode_8(pagelist, 1);
2862                                 ceph_pagelist_encode_32(pagelist, struct_len);
2863                         }
2864                         ceph_pagelist_encode_string(pagelist, path, pathlen);
2865                         ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
2866                         ceph_locks_to_pagelist(flocks, pagelist,
2867                                                num_fcntl_locks,
2868                                                num_flock_locks);
2869                         if (struct_v >= 2)
2870                                 ceph_pagelist_encode_64(pagelist, snap_follows);
2871                 }
2872                 kfree(flocks);
2873         } else {
2874                 size_t size = sizeof(u32) + pathlen + sizeof(rec.v1);
2875                 err = ceph_pagelist_reserve(pagelist, size);
2876                 if (!err) {
2877                         ceph_pagelist_encode_string(pagelist, path, pathlen);
2878                         ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
2879                 }
2880         }
2881
2882         recon_state->nr_caps++;
2883 out_free:
2884         kfree(path);
2885 out_dput:
2886         dput(dentry);
2887         return err;
2888 }
2889
2890
2891 /*
2892  * If an MDS fails and recovers, clients need to reconnect in order to
2893  * reestablish shared state.  This includes all caps issued through
2894  * this session _and_ the snap_realm hierarchy.  Because it's not
2895  * clear which snap realms the mds cares about, we send everything we
2896  * know about.. that ensures we'll then get any new info the
2897  * recovering MDS might have.
2898  *
2899  * This is a relatively heavyweight operation, but it's rare.
2900  *
2901  * called with mdsc->mutex held.
2902  */
2903 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2904                                struct ceph_mds_session *session)
2905 {
2906         struct ceph_msg *reply;
2907         struct rb_node *p;
2908         int mds = session->s_mds;
2909         int err = -ENOMEM;
2910         int s_nr_caps;
2911         struct ceph_pagelist *pagelist;
2912         struct ceph_reconnect_state recon_state;
2913
2914         pr_info("mds%d reconnect start\n", mds);
2915
2916         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2917         if (!pagelist)
2918                 goto fail_nopagelist;
2919         ceph_pagelist_init(pagelist);
2920
2921         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2922         if (!reply)
2923                 goto fail_nomsg;
2924
2925         mutex_lock(&session->s_mutex);
2926         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2927         session->s_seq = 0;
2928
2929         dout("session %p state %s\n", session,
2930              ceph_session_state_name(session->s_state));
2931
2932         spin_lock(&session->s_gen_ttl_lock);
2933         session->s_cap_gen++;
2934         spin_unlock(&session->s_gen_ttl_lock);
2935
2936         spin_lock(&session->s_cap_lock);
2937         /* don't know if session is readonly */
2938         session->s_readonly = 0;
2939         /*
2940          * notify __ceph_remove_cap() that we are composing cap reconnect.
2941          * If a cap get released before being added to the cap reconnect,
2942          * __ceph_remove_cap() should skip queuing cap release.
2943          */
2944         session->s_cap_reconnect = 1;
2945         /* drop old cap expires; we're about to reestablish that state */
2946         cleanup_cap_releases(mdsc, session);
2947
2948         /* trim unused caps to reduce MDS's cache rejoin time */
2949         if (mdsc->fsc->sb->s_root)
2950                 shrink_dcache_parent(mdsc->fsc->sb->s_root);
2951
2952         ceph_con_close(&session->s_con);
2953         ceph_con_open(&session->s_con,
2954                       CEPH_ENTITY_TYPE_MDS, mds,
2955                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2956
2957         /* replay unsafe requests */
2958         replay_unsafe_requests(mdsc, session);
2959
2960         down_read(&mdsc->snap_rwsem);
2961
2962         /* traverse this session's caps */
2963         s_nr_caps = session->s_nr_caps;
2964         err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
2965         if (err)
2966                 goto fail;
2967
2968         recon_state.nr_caps = 0;
2969         recon_state.pagelist = pagelist;
2970         if (session->s_con.peer_features & CEPH_FEATURE_MDSENC)
2971                 recon_state.msg_version = 3;
2972         else if (session->s_con.peer_features & CEPH_FEATURE_FLOCK)
2973                 recon_state.msg_version = 2;
2974         else
2975                 recon_state.msg_version = 1;
2976         err = iterate_session_caps(session, encode_caps_cb, &recon_state);
2977         if (err < 0)
2978                 goto fail;
2979
2980         spin_lock(&session->s_cap_lock);
2981         session->s_cap_reconnect = 0;
2982         spin_unlock(&session->s_cap_lock);
2983
2984         /*
2985          * snaprealms.  we provide mds with the ino, seq (version), and
2986          * parent for all of our realms.  If the mds has any newer info,
2987          * it will tell us.
2988          */
2989         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2990                 struct ceph_snap_realm *realm =
2991                         rb_entry(p, struct ceph_snap_realm, node);
2992                 struct ceph_mds_snaprealm_reconnect sr_rec;
2993
2994                 dout(" adding snap realm %llx seq %lld parent %llx\n",
2995                      realm->ino, realm->seq, realm->parent_ino);
2996                 sr_rec.ino = cpu_to_le64(realm->ino);
2997                 sr_rec.seq = cpu_to_le64(realm->seq);
2998                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2999                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3000                 if (err)
3001                         goto fail;
3002         }
3003
3004         reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3005
3006         /* raced with cap release? */
3007         if (s_nr_caps != recon_state.nr_caps) {
3008                 struct page *page = list_first_entry(&pagelist->head,
3009                                                      struct page, lru);
3010                 __le32 *addr = kmap_atomic(page);
3011                 *addr = cpu_to_le32(recon_state.nr_caps);
3012                 kunmap_atomic(addr);
3013         }
3014
3015         reply->hdr.data_len = cpu_to_le32(pagelist->length);
3016         ceph_msg_data_add_pagelist(reply, pagelist);
3017
3018         ceph_early_kick_flushing_caps(mdsc, session);
3019
3020         ceph_con_send(&session->s_con, reply);
3021
3022         mutex_unlock(&session->s_mutex);
3023
3024         mutex_lock(&mdsc->mutex);
3025         __wake_requests(mdsc, &session->s_waiting);
3026         mutex_unlock(&mdsc->mutex);
3027
3028         up_read(&mdsc->snap_rwsem);
3029         return;
3030
3031 fail:
3032         ceph_msg_put(reply);
3033         up_read(&mdsc->snap_rwsem);
3034         mutex_unlock(&session->s_mutex);
3035 fail_nomsg:
3036         ceph_pagelist_release(pagelist);
3037 fail_nopagelist:
3038         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3039         return;
3040 }
3041
3042
3043 /*
3044  * compare old and new mdsmaps, kicking requests
3045  * and closing out old connections as necessary
3046  *
3047  * called under mdsc->mutex.
3048  */
3049 static void check_new_map(struct ceph_mds_client *mdsc,
3050                           struct ceph_mdsmap *newmap,
3051                           struct ceph_mdsmap *oldmap)
3052 {
3053         int i;
3054         int oldstate, newstate;
3055         struct ceph_mds_session *s;
3056
3057         dout("check_new_map new %u old %u\n",
3058              newmap->m_epoch, oldmap->m_epoch);
3059
3060         for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
3061                 if (mdsc->sessions[i] == NULL)
3062                         continue;
3063                 s = mdsc->sessions[i];
3064                 oldstate = ceph_mdsmap_get_state(oldmap, i);
3065                 newstate = ceph_mdsmap_get_state(newmap, i);
3066
3067                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3068                      i, ceph_mds_state_name(oldstate),
3069                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3070                      ceph_mds_state_name(newstate),
3071                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3072                      ceph_session_state_name(s->s_state));
3073
3074                 if (i >= newmap->m_max_mds ||
3075                     memcmp(ceph_mdsmap_get_addr(oldmap, i),
3076                            ceph_mdsmap_get_addr(newmap, i),
3077                            sizeof(struct ceph_entity_addr))) {
3078                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3079                                 /* the session never opened, just close it
3080                                  * out now */
3081                                 __wake_requests(mdsc, &s->s_waiting);
3082                                 __unregister_session(mdsc, s);
3083                         } else {
3084                                 /* just close it */
3085                                 mutex_unlock(&mdsc->mutex);
3086                                 mutex_lock(&s->s_mutex);
3087                                 mutex_lock(&mdsc->mutex);
3088                                 ceph_con_close(&s->s_con);
3089                                 mutex_unlock(&s->s_mutex);
3090                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3091                         }
3092                 } else if (oldstate == newstate) {
3093                         continue;  /* nothing new with this mds */
3094                 }
3095
3096                 /*
3097                  * send reconnect?
3098                  */
3099                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3100                     newstate >= CEPH_MDS_STATE_RECONNECT) {
3101                         mutex_unlock(&mdsc->mutex);
3102                         send_mds_reconnect(mdsc, s);
3103                         mutex_lock(&mdsc->mutex);
3104                 }
3105
3106                 /*
3107                  * kick request on any mds that has gone active.
3108                  */
3109                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3110                     newstate >= CEPH_MDS_STATE_ACTIVE) {
3111                         if (oldstate != CEPH_MDS_STATE_CREATING &&
3112                             oldstate != CEPH_MDS_STATE_STARTING)
3113                                 pr_info("mds%d recovery completed\n", s->s_mds);
3114                         kick_requests(mdsc, i);
3115                         ceph_kick_flushing_caps(mdsc, s);
3116                         wake_up_session_caps(s, 1);
3117                 }
3118         }
3119
3120         for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
3121                 s = mdsc->sessions[i];
3122                 if (!s)
3123                         continue;
3124                 if (!ceph_mdsmap_is_laggy(newmap, i))
3125                         continue;
3126                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3127                     s->s_state == CEPH_MDS_SESSION_HUNG ||
3128                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
3129                         dout(" connecting to export targets of laggy mds%d\n",
3130                              i);
3131                         __open_export_target_sessions(mdsc, s);
3132                 }
3133         }
3134 }
3135
3136
3137
3138 /*
3139  * leases
3140  */
3141
3142 /*
3143  * caller must hold session s_mutex, dentry->d_lock
3144  */
3145 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3146 {
3147         struct ceph_dentry_info *di = ceph_dentry(dentry);
3148
3149         ceph_put_mds_session(di->lease_session);
3150         di->lease_session = NULL;
3151 }
3152
3153 static void handle_lease(struct ceph_mds_client *mdsc,
3154                          struct ceph_mds_session *session,
3155                          struct ceph_msg *msg)
3156 {
3157         struct super_block *sb = mdsc->fsc->sb;
3158         struct inode *inode;
3159         struct dentry *parent, *dentry;
3160         struct ceph_dentry_info *di;
3161         int mds = session->s_mds;
3162         struct ceph_mds_lease *h = msg->front.iov_base;
3163         u32 seq;
3164         struct ceph_vino vino;
3165         struct qstr dname;
3166         int release = 0;
3167
3168         dout("handle_lease from mds%d\n", mds);
3169
3170         /* decode */
3171         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3172                 goto bad;
3173         vino.ino = le64_to_cpu(h->ino);
3174         vino.snap = CEPH_NOSNAP;
3175         seq = le32_to_cpu(h->seq);
3176         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3177         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3178         if (dname.len != get_unaligned_le32(h+1))
3179                 goto bad;
3180
3181         /* lookup inode */
3182         inode = ceph_find_inode(sb, vino);
3183         dout("handle_lease %s, ino %llx %p %.*s\n",
3184              ceph_lease_op_name(h->action), vino.ino, inode,
3185              dname.len, dname.name);
3186
3187         mutex_lock(&session->s_mutex);
3188         session->s_seq++;
3189
3190         if (inode == NULL) {
3191                 dout("handle_lease no inode %llx\n", vino.ino);
3192                 goto release;
3193         }
3194
3195         /* dentry */
3196         parent = d_find_alias(inode);
3197         if (!parent) {
3198                 dout("no parent dentry on inode %p\n", inode);
3199                 WARN_ON(1);
3200                 goto release;  /* hrm... */
3201         }
3202         dname.hash = full_name_hash(parent, dname.name, dname.len);
3203         dentry = d_lookup(parent, &dname);
3204         dput(parent);
3205         if (!dentry)
3206                 goto release;
3207
3208         spin_lock(&dentry->d_lock);
3209         di = ceph_dentry(dentry);
3210         switch (h->action) {
3211         case CEPH_MDS_LEASE_REVOKE:
3212                 if (di->lease_session == session) {
3213                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3214                                 h->seq = cpu_to_le32(di->lease_seq);
3215                         __ceph_mdsc_drop_dentry_lease(dentry);
3216                 }
3217                 release = 1;
3218                 break;
3219
3220         case CEPH_MDS_LEASE_RENEW:
3221                 if (di->lease_session == session &&
3222                     di->lease_gen == session->s_cap_gen &&
3223                     di->lease_renew_from &&
3224                     di->lease_renew_after == 0) {
3225                         unsigned long duration =
3226                                 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3227
3228                         di->lease_seq = seq;
3229                         di->time = di->lease_renew_from + duration;
3230                         di->lease_renew_after = di->lease_renew_from +
3231                                 (duration >> 1);
3232                         di->lease_renew_from = 0;
3233                 }
3234                 break;
3235         }
3236         spin_unlock(&dentry->d_lock);
3237         dput(dentry);
3238
3239         if (!release)
3240                 goto out;
3241
3242 release:
3243         /* let's just reuse the same message */
3244         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3245         ceph_msg_get(msg);
3246         ceph_con_send(&session->s_con, msg);
3247
3248 out:
3249         iput(inode);
3250         mutex_unlock(&session->s_mutex);
3251         return;
3252
3253 bad:
3254         pr_err("corrupt lease message\n");
3255         ceph_msg_dump(msg);
3256 }
3257
3258 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3259                               struct inode *inode,
3260                               struct dentry *dentry, char action,
3261                               u32 seq)
3262 {
3263         struct ceph_msg *msg;
3264         struct ceph_mds_lease *lease;
3265         int len = sizeof(*lease) + sizeof(u32);
3266         int dnamelen = 0;
3267
3268         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3269              inode, dentry, ceph_lease_op_name(action), session->s_mds);
3270         dnamelen = dentry->d_name.len;
3271         len += dnamelen;
3272
3273         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3274         if (!msg)
3275                 return;
3276         lease = msg->front.iov_base;
3277         lease->action = action;
3278         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3279         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3280         lease->seq = cpu_to_le32(seq);
3281         put_unaligned_le32(dnamelen, lease + 1);
3282         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3283
3284         /*
3285          * if this is a preemptive lease RELEASE, no need to
3286          * flush request stream, since the actual request will
3287          * soon follow.
3288          */
3289         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3290
3291         ceph_con_send(&session->s_con, msg);
3292 }
3293
3294 /*
3295  * drop all leases (and dentry refs) in preparation for umount
3296  */
3297 static void drop_leases(struct ceph_mds_client *mdsc)
3298 {
3299         int i;
3300
3301         dout("drop_leases\n");
3302         mutex_lock(&mdsc->mutex);
3303         for (i = 0; i < mdsc->max_sessions; i++) {
3304                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3305                 if (!s)
3306                         continue;
3307                 mutex_unlock(&mdsc->mutex);
3308                 mutex_lock(&s->s_mutex);
3309                 mutex_unlock(&s->s_mutex);
3310                 ceph_put_mds_session(s);
3311                 mutex_lock(&mdsc->mutex);
3312         }
3313         mutex_unlock(&mdsc->mutex);
3314 }
3315
3316
3317
3318 /*
3319  * delayed work -- periodically trim expired leases, renew caps with mds
3320  */
3321 static void schedule_delayed(struct ceph_mds_client *mdsc)
3322 {
3323         int delay = 5;
3324         unsigned hz = round_jiffies_relative(HZ * delay);
3325         schedule_delayed_work(&mdsc->delayed_work, hz);
3326 }
3327
3328 static void delayed_work(struct work_struct *work)
3329 {
3330         int i;
3331         struct ceph_mds_client *mdsc =
3332                 container_of(work, struct ceph_mds_client, delayed_work.work);
3333         int renew_interval;
3334         int renew_caps;
3335
3336         dout("mdsc delayed_work\n");
3337         ceph_check_delayed_caps(mdsc);
3338
3339         mutex_lock(&mdsc->mutex);
3340         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3341         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3342                                    mdsc->last_renew_caps);
3343         if (renew_caps)
3344                 mdsc->last_renew_caps = jiffies;
3345
3346         for (i = 0; i < mdsc->max_sessions; i++) {
3347                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3348                 if (s == NULL)
3349                         continue;
3350                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3351                         dout("resending session close request for mds%d\n",
3352                              s->s_mds);
3353                         request_close_session(mdsc, s);
3354                         ceph_put_mds_session(s);
3355                         continue;
3356                 }
3357                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3358                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3359                                 s->s_state = CEPH_MDS_SESSION_HUNG;
3360                                 pr_info("mds%d hung\n", s->s_mds);
3361                         }
3362                 }
3363                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3364                         /* this mds is failed or recovering, just wait */
3365                         ceph_put_mds_session(s);
3366                         continue;
3367                 }
3368                 mutex_unlock(&mdsc->mutex);
3369
3370                 mutex_lock(&s->s_mutex);
3371                 if (renew_caps)
3372                         send_renew_caps(mdsc, s);
3373                 else
3374                         ceph_con_keepalive(&s->s_con);
3375                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3376                     s->s_state == CEPH_MDS_SESSION_HUNG)
3377                         ceph_send_cap_releases(mdsc, s);
3378                 mutex_unlock(&s->s_mutex);
3379                 ceph_put_mds_session(s);
3380
3381                 mutex_lock(&mdsc->mutex);
3382         }
3383         mutex_unlock(&mdsc->mutex);
3384
3385         schedule_delayed(mdsc);
3386 }
3387
3388 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3389
3390 {
3391         struct ceph_mds_client *mdsc;
3392
3393         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3394         if (!mdsc)
3395                 return -ENOMEM;
3396         mdsc->fsc = fsc;
3397         fsc->mdsc = mdsc;
3398         mutex_init(&mdsc->mutex);
3399         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3400         if (mdsc->mdsmap == NULL) {
3401                 kfree(mdsc);
3402                 return -ENOMEM;
3403         }
3404
3405         init_completion(&mdsc->safe_umount_waiters);
3406         init_waitqueue_head(&mdsc->session_close_wq);
3407         INIT_LIST_HEAD(&mdsc->waiting_for_map);
3408         mdsc->sessions = NULL;
3409         atomic_set(&mdsc->num_sessions, 0);
3410         mdsc->max_sessions = 0;
3411         mdsc->stopping = 0;
3412         mdsc->last_snap_seq = 0;
3413         init_rwsem(&mdsc->snap_rwsem);
3414         mdsc->snap_realms = RB_ROOT;
3415         INIT_LIST_HEAD(&mdsc->snap_empty);
3416         spin_lock_init(&mdsc->snap_empty_lock);
3417         mdsc->last_tid = 0;
3418         mdsc->oldest_tid = 0;
3419         mdsc->request_tree = RB_ROOT;
3420         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3421         mdsc->last_renew_caps = jiffies;
3422         INIT_LIST_HEAD(&mdsc->cap_delay_list);
3423         spin_lock_init(&mdsc->cap_delay_lock);
3424         INIT_LIST_HEAD(&mdsc->snap_flush_list);
3425         spin_lock_init(&mdsc->snap_flush_lock);
3426         mdsc->last_cap_flush_tid = 1;
3427         INIT_LIST_HEAD(&mdsc->cap_flush_list);
3428         INIT_LIST_HEAD(&mdsc->cap_dirty);
3429         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3430         mdsc->num_cap_flushing = 0;
3431         spin_lock_init(&mdsc->cap_dirty_lock);
3432         init_waitqueue_head(&mdsc->cap_flushing_wq);
3433         spin_lock_init(&mdsc->dentry_lru_lock);
3434         INIT_LIST_HEAD(&mdsc->dentry_lru);
3435
3436         ceph_caps_init(mdsc);
3437         ceph_adjust_min_caps(mdsc, fsc->min_caps);
3438
3439         init_rwsem(&mdsc->pool_perm_rwsem);
3440         mdsc->pool_perm_tree = RB_ROOT;
3441
3442         return 0;
3443 }
3444
3445 /*
3446  * Wait for safe replies on open mds requests.  If we time out, drop
3447  * all requests from the tree to avoid dangling dentry refs.
3448  */
3449 static void wait_requests(struct ceph_mds_client *mdsc)
3450 {
3451         struct ceph_options *opts = mdsc->fsc->client->options;
3452         struct ceph_mds_request *req;
3453
3454         mutex_lock(&mdsc->mutex);
3455         if (__get_oldest_req(mdsc)) {
3456                 mutex_unlock(&mdsc->mutex);
3457
3458                 dout("wait_requests waiting for requests\n");
3459                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3460                                     ceph_timeout_jiffies(opts->mount_timeout));
3461
3462                 /* tear down remaining requests */
3463                 mutex_lock(&mdsc->mutex);
3464                 while ((req = __get_oldest_req(mdsc))) {
3465                         dout("wait_requests timed out on tid %llu\n",
3466                              req->r_tid);
3467                         __unregister_request(mdsc, req);
3468                 }
3469         }
3470         mutex_unlock(&mdsc->mutex);
3471         dout("wait_requests done\n");
3472 }
3473
3474 /*
3475  * called before mount is ro, and before dentries are torn down.
3476  * (hmm, does this still race with new lookups?)
3477  */
3478 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3479 {
3480         dout("pre_umount\n");
3481         mdsc->stopping = 1;
3482
3483         drop_leases(mdsc);
3484         ceph_flush_dirty_caps(mdsc);
3485         wait_requests(mdsc);
3486
3487         /*
3488          * wait for reply handlers to drop their request refs and
3489          * their inode/dcache refs
3490          */
3491         ceph_msgr_flush();
3492 }
3493
3494 /*
3495  * wait for all write mds requests to flush.
3496  */
3497 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3498 {
3499         struct ceph_mds_request *req = NULL, *nextreq;
3500         struct rb_node *n;
3501
3502         mutex_lock(&mdsc->mutex);
3503         dout("wait_unsafe_requests want %lld\n", want_tid);
3504 restart:
3505         req = __get_oldest_req(mdsc);
3506         while (req && req->r_tid <= want_tid) {
3507                 /* find next request */
3508                 n = rb_next(&req->r_node);
3509                 if (n)
3510                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3511                 else
3512                         nextreq = NULL;
3513                 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3514                     (req->r_op & CEPH_MDS_OP_WRITE)) {
3515                         /* write op */
3516                         ceph_mdsc_get_request(req);
3517                         if (nextreq)
3518                                 ceph_mdsc_get_request(nextreq);
3519                         mutex_unlock(&mdsc->mutex);
3520                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
3521                              req->r_tid, want_tid);
3522                         wait_for_completion(&req->r_safe_completion);
3523                         mutex_lock(&mdsc->mutex);
3524                         ceph_mdsc_put_request(req);
3525                         if (!nextreq)
3526                                 break;  /* next dne before, so we're done! */
3527                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
3528                                 /* next request was removed from tree */
3529                                 ceph_mdsc_put_request(nextreq);
3530                                 goto restart;
3531                         }
3532                         ceph_mdsc_put_request(nextreq);  /* won't go away */
3533                 }
3534                 req = nextreq;
3535         }
3536         mutex_unlock(&mdsc->mutex);
3537         dout("wait_unsafe_requests done\n");
3538 }
3539
3540 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3541 {
3542         u64 want_tid, want_flush;
3543
3544         if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3545                 return;
3546
3547         dout("sync\n");
3548         mutex_lock(&mdsc->mutex);
3549         want_tid = mdsc->last_tid;
3550         mutex_unlock(&mdsc->mutex);
3551
3552         ceph_flush_dirty_caps(mdsc);
3553         spin_lock(&mdsc->cap_dirty_lock);
3554         want_flush = mdsc->last_cap_flush_tid;
3555         if (!list_empty(&mdsc->cap_flush_list)) {
3556                 struct ceph_cap_flush *cf =
3557                         list_last_entry(&mdsc->cap_flush_list,
3558                                         struct ceph_cap_flush, g_list);
3559                 cf->wake = true;
3560         }
3561         spin_unlock(&mdsc->cap_dirty_lock);
3562
3563         dout("sync want tid %lld flush_seq %lld\n",
3564              want_tid, want_flush);
3565
3566         wait_unsafe_requests(mdsc, want_tid);
3567         wait_caps_flush(mdsc, want_flush);
3568 }
3569
3570 /*
3571  * true if all sessions are closed, or we force unmount
3572  */
3573 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
3574 {
3575         if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3576                 return true;
3577         return atomic_read(&mdsc->num_sessions) <= skipped;
3578 }
3579
3580 /*
3581  * called after sb is ro.
3582  */
3583 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3584 {
3585         struct ceph_options *opts = mdsc->fsc->client->options;
3586         struct ceph_mds_session *session;
3587         int i;
3588         int skipped = 0;
3589
3590         dout("close_sessions\n");
3591
3592         /* close sessions */
3593         mutex_lock(&mdsc->mutex);
3594         for (i = 0; i < mdsc->max_sessions; i++) {
3595                 session = __ceph_lookup_mds_session(mdsc, i);
3596                 if (!session)
3597                         continue;
3598                 mutex_unlock(&mdsc->mutex);
3599                 mutex_lock(&session->s_mutex);
3600                 if (__close_session(mdsc, session) <= 0)
3601                         skipped++;
3602                 mutex_unlock(&session->s_mutex);
3603                 ceph_put_mds_session(session);
3604                 mutex_lock(&mdsc->mutex);
3605         }
3606         mutex_unlock(&mdsc->mutex);
3607
3608         dout("waiting for sessions to close\n");
3609         wait_event_timeout(mdsc->session_close_wq,
3610                            done_closing_sessions(mdsc, skipped),
3611                            ceph_timeout_jiffies(opts->mount_timeout));
3612
3613         /* tear down remaining sessions */
3614         mutex_lock(&mdsc->mutex);
3615         for (i = 0; i < mdsc->max_sessions; i++) {
3616                 if (mdsc->sessions[i]) {
3617                         session = get_session(mdsc->sessions[i]);
3618                         __unregister_session(mdsc, session);
3619                         mutex_unlock(&mdsc->mutex);
3620                         mutex_lock(&session->s_mutex);
3621                         remove_session_caps(session);
3622                         mutex_unlock(&session->s_mutex);
3623                         ceph_put_mds_session(session);
3624                         mutex_lock(&mdsc->mutex);
3625                 }
3626         }
3627         WARN_ON(!list_empty(&mdsc->cap_delay_list));
3628         mutex_unlock(&mdsc->mutex);
3629
3630         ceph_cleanup_empty_realms(mdsc);
3631
3632         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3633
3634         dout("stopped\n");
3635 }
3636
3637 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
3638 {
3639         struct ceph_mds_session *session;
3640         int mds;
3641
3642         dout("force umount\n");
3643
3644         mutex_lock(&mdsc->mutex);
3645         for (mds = 0; mds < mdsc->max_sessions; mds++) {
3646                 session = __ceph_lookup_mds_session(mdsc, mds);
3647                 if (!session)
3648                         continue;
3649                 mutex_unlock(&mdsc->mutex);
3650                 mutex_lock(&session->s_mutex);
3651                 __close_session(mdsc, session);
3652                 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
3653                         cleanup_session_requests(mdsc, session);
3654                         remove_session_caps(session);
3655                 }
3656                 mutex_unlock(&session->s_mutex);
3657                 ceph_put_mds_session(session);
3658                 mutex_lock(&mdsc->mutex);
3659                 kick_requests(mdsc, mds);
3660         }
3661         __wake_requests(mdsc, &mdsc->waiting_for_map);
3662         mutex_unlock(&mdsc->mutex);
3663 }
3664
3665 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3666 {
3667         dout("stop\n");
3668         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3669         if (mdsc->mdsmap)
3670                 ceph_mdsmap_destroy(mdsc->mdsmap);
3671         kfree(mdsc->sessions);
3672         ceph_caps_finalize(mdsc);
3673         ceph_pool_perm_destroy(mdsc);
3674 }
3675
3676 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3677 {
3678         struct ceph_mds_client *mdsc = fsc->mdsc;
3679
3680         dout("mdsc_destroy %p\n", mdsc);
3681         ceph_mdsc_stop(mdsc);
3682
3683         /* flush out any connection work with references to us */
3684         ceph_msgr_flush();
3685
3686         fsc->mdsc = NULL;
3687         kfree(mdsc);
3688         dout("mdsc_destroy %p done\n", mdsc);
3689 }
3690
3691 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3692 {
3693         struct ceph_fs_client *fsc = mdsc->fsc;
3694         const char *mds_namespace = fsc->mount_options->mds_namespace;
3695         void *p = msg->front.iov_base;
3696         void *end = p + msg->front.iov_len;
3697         u32 epoch;
3698         u32 map_len;
3699         u32 num_fs;
3700         u32 mount_fscid = (u32)-1;
3701         u8 struct_v, struct_cv;
3702         int err = -EINVAL;
3703
3704         ceph_decode_need(&p, end, sizeof(u32), bad);
3705         epoch = ceph_decode_32(&p);
3706
3707         dout("handle_fsmap epoch %u\n", epoch);
3708
3709         ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3710         struct_v = ceph_decode_8(&p);
3711         struct_cv = ceph_decode_8(&p);
3712         map_len = ceph_decode_32(&p);
3713
3714         ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
3715         p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
3716
3717         num_fs = ceph_decode_32(&p);
3718         while (num_fs-- > 0) {
3719                 void *info_p, *info_end;
3720                 u32 info_len;
3721                 u8 info_v, info_cv;
3722                 u32 fscid, namelen;
3723
3724                 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3725                 info_v = ceph_decode_8(&p);
3726                 info_cv = ceph_decode_8(&p);
3727                 info_len = ceph_decode_32(&p);
3728                 ceph_decode_need(&p, end, info_len, bad);
3729                 info_p = p;
3730                 info_end = p + info_len;
3731                 p = info_end;
3732
3733                 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
3734                 fscid = ceph_decode_32(&info_p);
3735                 namelen = ceph_decode_32(&info_p);
3736                 ceph_decode_need(&info_p, info_end, namelen, bad);
3737
3738                 if (mds_namespace &&
3739                     strlen(mds_namespace) == namelen &&
3740                     !strncmp(mds_namespace, (char *)info_p, namelen)) {
3741                         mount_fscid = fscid;
3742                         break;
3743                 }
3744         }
3745
3746         ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
3747         if (mount_fscid != (u32)-1) {
3748                 fsc->client->monc.fs_cluster_id = mount_fscid;
3749                 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
3750                                    0, true);
3751                 ceph_monc_renew_subs(&fsc->client->monc);
3752         } else {
3753                 err = -ENOENT;
3754                 goto err_out;
3755         }
3756         return;
3757 bad:
3758         pr_err("error decoding fsmap\n");
3759 err_out:
3760         mutex_lock(&mdsc->mutex);
3761         mdsc->mdsmap_err = -ENOENT;
3762         __wake_requests(mdsc, &mdsc->waiting_for_map);
3763         mutex_unlock(&mdsc->mutex);
3764         return;
3765 }
3766
3767 /*
3768  * handle mds map update.
3769  */
3770 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3771 {
3772         u32 epoch;
3773         u32 maplen;
3774         void *p = msg->front.iov_base;
3775         void *end = p + msg->front.iov_len;
3776         struct ceph_mdsmap *newmap, *oldmap;
3777         struct ceph_fsid fsid;
3778         int err = -EINVAL;
3779
3780         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3781         ceph_decode_copy(&p, &fsid, sizeof(fsid));
3782         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3783                 return;
3784         epoch = ceph_decode_32(&p);
3785         maplen = ceph_decode_32(&p);
3786         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3787
3788         /* do we need it? */
3789         mutex_lock(&mdsc->mutex);
3790         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3791                 dout("handle_map epoch %u <= our %u\n",
3792                      epoch, mdsc->mdsmap->m_epoch);
3793                 mutex_unlock(&mdsc->mutex);
3794                 return;
3795         }
3796
3797         newmap = ceph_mdsmap_decode(&p, end);
3798         if (IS_ERR(newmap)) {
3799                 err = PTR_ERR(newmap);
3800                 goto bad_unlock;
3801         }
3802
3803         /* swap into place */
3804         if (mdsc->mdsmap) {
3805                 oldmap = mdsc->mdsmap;
3806                 mdsc->mdsmap = newmap;
3807                 check_new_map(mdsc, newmap, oldmap);
3808                 ceph_mdsmap_destroy(oldmap);
3809         } else {
3810                 mdsc->mdsmap = newmap;  /* first mds map */
3811         }
3812         mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3813
3814         __wake_requests(mdsc, &mdsc->waiting_for_map);
3815         ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
3816                           mdsc->mdsmap->m_epoch);
3817
3818         mutex_unlock(&mdsc->mutex);
3819         schedule_delayed(mdsc);
3820         return;
3821
3822 bad_unlock:
3823         mutex_unlock(&mdsc->mutex);
3824 bad:
3825         pr_err("error decoding mdsmap %d\n", err);
3826         return;
3827 }
3828
3829 static struct ceph_connection *con_get(struct ceph_connection *con)
3830 {
3831         struct ceph_mds_session *s = con->private;
3832
3833         if (get_session(s)) {
3834                 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3835                 return con;
3836         }
3837         dout("mdsc con_get %p FAIL\n", s);
3838         return NULL;
3839 }
3840
3841 static void con_put(struct ceph_connection *con)
3842 {
3843         struct ceph_mds_session *s = con->private;
3844
3845         dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3846         ceph_put_mds_session(s);
3847 }
3848
3849 /*
3850  * if the client is unresponsive for long enough, the mds will kill
3851  * the session entirely.
3852  */
3853 static void peer_reset(struct ceph_connection *con)
3854 {
3855         struct ceph_mds_session *s = con->private;
3856         struct ceph_mds_client *mdsc = s->s_mdsc;
3857
3858         pr_warn("mds%d closed our session\n", s->s_mds);
3859         send_mds_reconnect(mdsc, s);
3860 }
3861
3862 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3863 {
3864         struct ceph_mds_session *s = con->private;
3865         struct ceph_mds_client *mdsc = s->s_mdsc;
3866         int type = le16_to_cpu(msg->hdr.type);
3867
3868         mutex_lock(&mdsc->mutex);
3869         if (__verify_registered_session(mdsc, s) < 0) {
3870                 mutex_unlock(&mdsc->mutex);
3871                 goto out;
3872         }
3873         mutex_unlock(&mdsc->mutex);
3874
3875         switch (type) {
3876         case CEPH_MSG_MDS_MAP:
3877                 ceph_mdsc_handle_mdsmap(mdsc, msg);
3878                 break;
3879         case CEPH_MSG_FS_MAP_USER:
3880                 ceph_mdsc_handle_fsmap(mdsc, msg);
3881                 break;
3882         case CEPH_MSG_CLIENT_SESSION:
3883                 handle_session(s, msg);
3884                 break;
3885         case CEPH_MSG_CLIENT_REPLY:
3886                 handle_reply(s, msg);
3887                 break;
3888         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3889                 handle_forward(mdsc, s, msg);
3890                 break;
3891         case CEPH_MSG_CLIENT_CAPS:
3892                 ceph_handle_caps(s, msg);
3893                 break;
3894         case CEPH_MSG_CLIENT_SNAP:
3895                 ceph_handle_snap(mdsc, s, msg);
3896                 break;
3897         case CEPH_MSG_CLIENT_LEASE:
3898                 handle_lease(mdsc, s, msg);
3899                 break;
3900
3901         default:
3902                 pr_err("received unknown message type %d %s\n", type,
3903                        ceph_msg_type_name(type));
3904         }
3905 out:
3906         ceph_msg_put(msg);
3907 }
3908
3909 /*
3910  * authentication
3911  */
3912
3913 /*
3914  * Note: returned pointer is the address of a structure that's
3915  * managed separately.  Caller must *not* attempt to free it.
3916  */
3917 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
3918                                         int *proto, int force_new)
3919 {
3920         struct ceph_mds_session *s = con->private;
3921         struct ceph_mds_client *mdsc = s->s_mdsc;
3922         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3923         struct ceph_auth_handshake *auth = &s->s_auth;
3924
3925         if (force_new && auth->authorizer) {
3926                 ceph_auth_destroy_authorizer(auth->authorizer);
3927                 auth->authorizer = NULL;
3928         }
3929         if (!auth->authorizer) {
3930                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3931                                                       auth);
3932                 if (ret)
3933                         return ERR_PTR(ret);
3934         } else {
3935                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3936                                                       auth);
3937                 if (ret)
3938                         return ERR_PTR(ret);
3939         }
3940         *proto = ac->protocol;
3941
3942         return auth;
3943 }
3944
3945
3946 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3947 {
3948         struct ceph_mds_session *s = con->private;
3949         struct ceph_mds_client *mdsc = s->s_mdsc;
3950         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3951
3952         return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer, len);
3953 }
3954
3955 static int invalidate_authorizer(struct ceph_connection *con)
3956 {
3957         struct ceph_mds_session *s = con->private;
3958         struct ceph_mds_client *mdsc = s->s_mdsc;
3959         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3960
3961         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3962
3963         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
3964 }
3965
3966 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
3967                                 struct ceph_msg_header *hdr, int *skip)
3968 {
3969         struct ceph_msg *msg;
3970         int type = (int) le16_to_cpu(hdr->type);
3971         int front_len = (int) le32_to_cpu(hdr->front_len);
3972
3973         if (con->in_msg)
3974                 return con->in_msg;
3975
3976         *skip = 0;
3977         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
3978         if (!msg) {
3979                 pr_err("unable to allocate msg type %d len %d\n",
3980                        type, front_len);
3981                 return NULL;
3982         }
3983
3984         return msg;
3985 }
3986
3987 static int mds_sign_message(struct ceph_msg *msg)
3988 {
3989        struct ceph_mds_session *s = msg->con->private;
3990        struct ceph_auth_handshake *auth = &s->s_auth;
3991
3992        return ceph_auth_sign_message(auth, msg);
3993 }
3994
3995 static int mds_check_message_signature(struct ceph_msg *msg)
3996 {
3997        struct ceph_mds_session *s = msg->con->private;
3998        struct ceph_auth_handshake *auth = &s->s_auth;
3999
4000        return ceph_auth_check_message_signature(auth, msg);
4001 }
4002
4003 static const struct ceph_connection_operations mds_con_ops = {
4004         .get = con_get,
4005         .put = con_put,
4006         .dispatch = dispatch,
4007         .get_authorizer = get_authorizer,
4008         .verify_authorizer_reply = verify_authorizer_reply,
4009         .invalidate_authorizer = invalidate_authorizer,
4010         .peer_reset = peer_reset,
4011         .alloc_msg = mds_alloc_msg,
4012         .sign_message = mds_sign_message,
4013         .check_message_signature = mds_check_message_signature,
4014 };
4015
4016 /* eof */