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  45  * are no longer valid.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         u32 op = le32_to_cpu(info->head->op);
 292
 293         if (op == CEPH_MDS_OP_GETFILELOCK)
 294                 return parse_reply_info_filelock(p, end, info, features);
 295         else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
 296                 return parse_reply_info_dir(p, end, info, features);
 297         else if (op == CEPH_MDS_OP_CREATE)
 298                 return parse_reply_info_create(p, end, info, features);
 299         else
 300                 return -EIO;
 301 }
 302
 303 /*
 304  * parse entire mds reply
 305  */
 306 static int parse_reply_info(struct ceph_msg *msg,
 307                             struct ceph_mds_reply_info_parsed *info,
 308                             u64 features)
 309 {
 310         void *p, *end;
 311         u32 len;
 312         int err;
 313
 314         info->head = msg->front.iov_base;
 315         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 316         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 317
 318         /* trace */
 319         ceph_decode_32_safe(&p, end, len, bad);
 320         if (len > 0) {
 321                 ceph_decode_need(&p, end, len, bad);
 322                 err = parse_reply_info_trace(&p, p+len, info, features);
 323                 if (err < 0)
 324                         goto out_bad;
 325         }
 326
 327         /* extra */
 328         ceph_decode_32_safe(&p, end, len, bad);
 329         if (len > 0) {
 330                 ceph_decode_need(&p, end, len, bad);
 331                 err = parse_reply_info_extra(&p, p+len, info, features);
 332                 if (err < 0)
 333                         goto out_bad;
 334         }
 335
 336         /* snap blob */
 337         ceph_decode_32_safe(&p, end, len, bad);
 338         info->snapblob_len = len;
 339         info->snapblob = p;
 340         p += len;
 341
 342         if (p != end)
 343                 goto bad;
 344         return 0;
 345
 346 bad:
 347         err = -EIO;
 348 out_bad:
 349         pr_err("mds parse_reply err %d\n", err);
 350         return err;
 351 }
 352
 353 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
 354 {
 355         if (!info->dir_entries)
 356                 return;
 357         free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
 358 }
 359
 360
 361 /*
 362  * sessions
 363  */
 364 const char *ceph_session_state_name(int s)
 365 {
 366         switch (s) {
 367         case CEPH_MDS_SESSION_NEW: return "new";
 368         case CEPH_MDS_SESSION_OPENING: return "opening";
 369         case CEPH_MDS_SESSION_OPEN: return "open";
 370         case CEPH_MDS_SESSION_HUNG: return "hung";
 371         case CEPH_MDS_SESSION_CLOSING: return "closing";
 372         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
 373         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
 374         case CEPH_MDS_SESSION_REJECTED: return "rejected";
 375         default: return "???";
 376         }
 377 }
 378
 379 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
 380 {
 381         if (atomic_inc_not_zero(&s->s_ref)) {
 382                 dout("mdsc get_session %p %d -> %d\n", s,
 383                      atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
 384                 return s;
 385         } else {
 386                 dout("mdsc get_session %p 0 -- FAIL", s);
 387                 return NULL;
 388         }
 389 }
 390
 391 void ceph_put_mds_session(struct ceph_mds_session *s)
 392 {
 393         dout("mdsc put_session %p %d -> %d\n", s,
 394              atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
 395         if (atomic_dec_and_test(&s->s_ref)) {
 396                 if (s->s_auth.authorizer)
 397                         ceph_auth_destroy_authorizer(s->s_auth.authorizer);
 398                 kfree(s);
 399         }
 400 }
 401
 402 /*
 403  * called under mdsc->mutex
 404  */
 405 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
 406                                                    int mds)
 407 {
 408         struct ceph_mds_session *session;
 409
 410         if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
 411                 return NULL;
 412         session = mdsc->sessions[mds];
 413         dout("lookup_mds_session %p %d\n", session,
 414              atomic_read(&session->s_ref));
 415         get_session(session);
 416         return session;
 417 }
 418
 419 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
 420 {
 421         if (mds >= mdsc->max_sessions)
 422                 return false;
 423         return mdsc->sessions[mds];
 424 }
 425
 426 static int __verify_registered_session(struct ceph_mds_client *mdsc,
 427                                        struct ceph_mds_session *s)
 428 {
 429         if (s->s_mds >= mdsc->max_sessions ||
 430             mdsc->sessions[s->s_mds] != s)
 431                 return -ENOENT;
 432         return 0;
 433 }
 434
 435 /*
 436  * create+register a new session for given mds.
 437  * called under mdsc->mutex.
 438  */
 439 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 440                                                  int mds)
 441 {
 442         struct ceph_mds_session *s;
 443
 444         if (mds >= mdsc->mdsmap->m_max_mds)
 445                 return ERR_PTR(-EINVAL);
 446
 447         s = kzalloc(sizeof(*s), GFP_NOFS);
 448         if (!s)
 449                 return ERR_PTR(-ENOMEM);
 450         s->s_mdsc = mdsc;
 451         s->s_mds = mds;
 452         s->s_state = CEPH_MDS_SESSION_NEW;
 453         s->s_ttl = 0;
 454         s->s_seq = 0;
 455         mutex_init(&s->s_mutex);
 456
 457         ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
 458
 459         spin_lock_init(&s->s_gen_ttl_lock);
 460         s->s_cap_gen = 0;
 461         s->s_cap_ttl = jiffies - 1;
 462
 463         spin_lock_init(&s->s_cap_lock);
 464         s->s_renew_requested = 0;
 465         s->s_renew_seq = 0;
 466         INIT_LIST_HEAD(&s->s_caps);
 467         s->s_nr_caps = 0;
 468         s->s_trim_caps = 0;
 469         atomic_set(&s->s_ref, 1);
 470         INIT_LIST_HEAD(&s->s_waiting);
 471         INIT_LIST_HEAD(&s->s_unsafe);
 472         s->s_num_cap_releases = 0;
 473         s->s_cap_reconnect = 0;
 474         s->s_cap_iterator = NULL;
 475         INIT_LIST_HEAD(&s->s_cap_releases);
 476         INIT_LIST_HEAD(&s->s_cap_flushing);
 477
 478         dout("register_session mds%d\n", mds);
 479         if (mds >= mdsc->max_sessions) {
 480                 int newmax = 1 << get_count_order(mds+1);
 481                 struct ceph_mds_session **sa;
 482
 483                 dout("register_session realloc to %d\n", newmax);
 484                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 485                 if (sa == NULL)
 486                         goto fail_realloc;
 487                 if (mdsc->sessions) {
 488                         memcpy(sa, mdsc->sessions,
 489                                mdsc->max_sessions * sizeof(void *));
 490                         kfree(mdsc->sessions);
 491                 }
 492                 mdsc->sessions = sa;
 493                 mdsc->max_sessions = newmax;
 494         }
 495         mdsc->sessions[mds] = s;
 496         atomic_inc(&mdsc->num_sessions);
 497         atomic_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
 498
 499         ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
 500                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
 501
 502         return s;
 503
 504 fail_realloc:
 505         kfree(s);
 506         return ERR_PTR(-ENOMEM);
 507 }
 508
 509 /*
 510  * called under mdsc->mutex
 511  */
 512 static void __unregister_session(struct ceph_mds_client *mdsc,
 513                                struct ceph_mds_session *s)
 514 {
 515         dout("__unregister_session mds%d %p\n", s->s_mds, s);
 516         BUG_ON(mdsc->sessions[s->s_mds] != s);
 517         mdsc->sessions[s->s_mds] = NULL;
 518         ceph_con_close(&s->s_con);
 519         ceph_put_mds_session(s);
 520         atomic_dec(&mdsc->num_sessions);
 521 }
 522
 523 /*
 524  * drop session refs in request.
 525  *
 526  * should be last request ref, or hold mdsc->mutex
 527  */
 528 static void put_request_session(struct ceph_mds_request *req)
 529 {
 530         if (req->r_session) {
 531                 ceph_put_mds_session(req->r_session);
 532                 req->r_session = NULL;
 533         }
 534 }
 535
 536 void ceph_mdsc_release_request(struct kref *kref)
 537 {
 538         struct ceph_mds_request *req = container_of(kref,
 539                                                     struct ceph_mds_request,
 540                                                     r_kref);
 541         destroy_reply_info(&req->r_reply_info);
 542         if (req->r_request)
 543                 ceph_msg_put(req->r_request);
 544         if (req->r_reply)
 545                 ceph_msg_put(req->r_reply);
 546         if (req->r_inode) {
 547                 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
 548                 iput(req->r_inode);
 549         }
 550         if (req->r_locked_dir)
 551                 ceph_put_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
 552         iput(req->r_target_inode);
 553         if (req->r_dentry)
 554                 dput(req->r_dentry);
 555         if (req->r_old_dentry)
 556                 dput(req->r_old_dentry);
 557         if (req->r_old_dentry_dir) {
 558                 /*
 559                  * track (and drop pins for) r_old_dentry_dir
 560                  * separately, since r_old_dentry's d_parent may have
 561                  * changed between the dir mutex being dropped and
 562                  * this request being freed.
 563                  */
 564                 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
 565                                   CEPH_CAP_PIN);
 566                 iput(req->r_old_dentry_dir);
 567         }
 568         kfree(req->r_path1);
 569         kfree(req->r_path2);
 570         if (req->r_pagelist)
 571                 ceph_pagelist_release(req->r_pagelist);
 572         put_request_session(req);
 573         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
 574         kfree(req);
 575 }
 576
 577 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
 578
 579 /*
 580  * lookup session, bump ref if found.
 581  *
 582  * called under mdsc->mutex.
 583  */
 584 static struct ceph_mds_request *
 585 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
 586 {
 587         struct ceph_mds_request *req;
 588
 589         req = lookup_request(&mdsc->request_tree, tid);
 590         if (req)
 591                 ceph_mdsc_get_request(req);
 592
 593         return req;
 594 }
 595
 596 /*
 597  * Register an in-flight request, and assign a tid.  Link to directory
 598  * are modifying (if any).
 599  *
 600  * Called under mdsc->mutex.
 601  */
 602 static void __register_request(struct ceph_mds_client *mdsc,
 603                                struct ceph_mds_request *req,
 604                                struct inode *dir)
 605 {
 606         req->r_tid = ++mdsc->last_tid;
 607         if (req->r_num_caps)
 608                 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
 609                                   req->r_num_caps);
 610         dout("__register_request %p tid %lld\n", req, req->r_tid);
 611         ceph_mdsc_get_request(req);
 612         insert_request(&mdsc->request_tree, req);
 613
 614         req->r_uid = current_fsuid();
 615         req->r_gid = current_fsgid();
 616
 617         if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
 618                 mdsc->oldest_tid = req->r_tid;
 619
 620         if (dir) {
 621                 ihold(dir);
 622                 req->r_unsafe_dir = dir;
 623         }
 624 }
 625
 626 static void __unregister_request(struct ceph_mds_client *mdsc,
 627                                  struct ceph_mds_request *req)
 628 {
 629         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 630
 631         if (req->r_tid == mdsc->oldest_tid) {
 632                 struct rb_node *p = rb_next(&req->r_node);
 633                 mdsc->oldest_tid = 0;
 634                 while (p) {
 635                         struct ceph_mds_request *next_req =
 636                                 rb_entry(p, struct ceph_mds_request, r_node);
 637                         if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
 638                                 mdsc->oldest_tid = next_req->r_tid;
 639                                 break;
 640                         }
 641                         p = rb_next(p);
 642                 }
 643         }
 644
 645         erase_request(&mdsc->request_tree, req);
 646
 647         if (req->r_unsafe_dir && req->r_got_unsafe) {
 648                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 649                 spin_lock(&ci->i_unsafe_lock);
 650                 list_del_init(&req->r_unsafe_dir_item);
 651                 spin_unlock(&ci->i_unsafe_lock);
 652         }
 653         if (req->r_target_inode && req->r_got_unsafe) {
 654                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
 655                 spin_lock(&ci->i_unsafe_lock);
 656                 list_del_init(&req->r_unsafe_target_item);
 657                 spin_unlock(&ci->i_unsafe_lock);
 658         }
 659
 660         if (req->r_unsafe_dir) {
 661                 iput(req->r_unsafe_dir);
 662                 req->r_unsafe_dir = NULL;
 663         }
 664
 665         complete_all(&req->r_safe_completion);
 666
 667         ceph_mdsc_put_request(req);
 668 }
 669
 670 /*
 671  * Choose mds to send request to next.  If there is a hint set in the
 672  * request (e.g., due to a prior forward hint from the mds), use that.
 673  * Otherwise, consult frag tree and/or caps to identify the
 674  * appropriate mds.  If all else fails, choose randomly.
 675  *
 676  * Called under mdsc->mutex.
 677  */
 678 static struct dentry *get_nonsnap_parent(struct dentry *dentry)
 679 {
 680         /*
 681          * we don't need to worry about protecting the d_parent access
 682          * here because we never renaming inside the snapped namespace
 683          * except to resplice to another snapdir, and either the old or new
 684          * result is a valid result.
 685          */
 686         while (!IS_ROOT(dentry) && ceph_snap(d_inode(dentry)) != CEPH_NOSNAP)
 687                 dentry = dentry->d_parent;
 688         return dentry;
 689 }
 690
 691 static int __choose_mds(struct ceph_mds_client *mdsc,
 692                         struct ceph_mds_request *req)
 693 {
 694         struct inode *inode;
 695         struct ceph_inode_info *ci;
 696         struct ceph_cap *cap;
 697         int mode = req->r_direct_mode;
 698         int mds = -1;
 699         u32 hash = req->r_direct_hash;
 700         bool is_hash = req->r_direct_is_hash;
 701
 702         /*
 703          * is there a specific mds we should try?  ignore hint if we have
 704          * no session and the mds is not up (active or recovering).
 705          */
 706         if (req->r_resend_mds >= 0 &&
 707             (__have_session(mdsc, req->r_resend_mds) ||
 708              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 709                 dout("choose_mds using resend_mds mds%d\n",
 710                      req->r_resend_mds);
 711                 return req->r_resend_mds;
 712         }
 713
 714         if (mode == USE_RANDOM_MDS)
 715                 goto random;
 716
 717         inode = NULL;
 718         if (req->r_inode) {
 719                 inode = req->r_inode;
 720         } else if (req->r_dentry) {
 721                 /* ignore race with rename; old or new d_parent is okay */
 722                 struct dentry *parent = req->r_dentry->d_parent;
 723                 struct inode *dir = d_inode(parent);
 724
 725                 if (dir->i_sb != mdsc->fsc->sb) {
 726                         /* not this fs! */
 727                         inode = d_inode(req->r_dentry);
 728                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
 729                         /* direct snapped/virtual snapdir requests
 730                          * based on parent dir inode */
 731                         struct dentry *dn = get_nonsnap_parent(parent);
 732                         inode = d_inode(dn);
 733                         dout("__choose_mds using nonsnap parent %p\n", inode);
 734                 } else {
 735                         /* dentry target */
 736                         inode = d_inode(req->r_dentry);
 737                         if (!inode || mode == USE_AUTH_MDS) {
 738                                 /* dir + name */
 739                                 inode = dir;
 740                                 hash = ceph_dentry_hash(dir, req->r_dentry);
 741                                 is_hash = true;
 742                         }
 743                 }
 744         }
 745
 746         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 747              (int)hash, mode);
 748         if (!inode)
 749                 goto random;
 750         ci = ceph_inode(inode);
 751
 752         if (is_hash && S_ISDIR(inode->i_mode)) {
 753                 struct ceph_inode_frag frag;
 754                 int found;
 755
 756                 ceph_choose_frag(ci, hash, &frag, &found);
 757                 if (found) {
 758                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 759                                 u8 r;
 760
 761                                 /* choose a random replica */
 762                                 get_random_bytes(&r, 1);
 763                                 r %= frag.ndist;
 764                                 mds = frag.dist[r];
 765                                 dout("choose_mds %p %llx.%llx "
 766                                      "frag %u mds%d (%d/%d)\n",
 767                                      inode, ceph_vinop(inode),
 768                                      frag.frag, mds,
 769                                      (int)r, frag.ndist);
 770                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 771                                     CEPH_MDS_STATE_ACTIVE)
 772                                         return mds;
 773                         }
 774
 775                         /* since this file/dir wasn't known to be
 776                          * replicated, then we want to look for the
 777                          * authoritative mds. */
 778                         mode = USE_AUTH_MDS;
 779                         if (frag.mds >= 0) {
 780                                 /* choose auth mds */
 781                                 mds = frag.mds;
 782                                 dout("choose_mds %p %llx.%llx "
 783                                      "frag %u mds%d (auth)\n",
 784                                      inode, ceph_vinop(inode), frag.frag, mds);
 785                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 786                                     CEPH_MDS_STATE_ACTIVE)
 787                                         return mds;
 788                         }
 789                 }
 790         }
 791
 792         spin_lock(&ci->i_ceph_lock);
 793         cap = NULL;
 794         if (mode == USE_AUTH_MDS)
 795                 cap = ci->i_auth_cap;
 796         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
 797                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
 798         if (!cap) {
 799                 spin_unlock(&ci->i_ceph_lock);
 800                 goto random;
 801         }
 802         mds = cap->session->s_mds;
 803         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
 804              inode, ceph_vinop(inode), mds,
 805              cap == ci->i_auth_cap ? "auth " : "", cap);
 806         spin_unlock(&ci->i_ceph_lock);
 807         return mds;
 808
 809 random:
 810         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
 811         dout("choose_mds chose random mds%d\n", mds);
 812         return mds;
 813 }
 814
 815
 816 /*
 817  * session messages
 818  */
 819 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
 820 {
 821         struct ceph_msg *msg;
 822         struct ceph_mds_session_head *h;
 823
 824         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
 825                            false);
 826         if (!msg) {
 827                 pr_err("create_session_msg ENOMEM creating msg\n");
 828                 return NULL;
 829         }
 830         h = msg->front.iov_base;
 831         h->op = cpu_to_le32(op);
 832         h->seq = cpu_to_le64(seq);
 833
 834         return msg;
 835 }
 836
 837 /*
 838  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
 839  * to include additional client metadata fields.
 840  */
 841 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
 842 {
 843         struct ceph_msg *msg;
 844         struct ceph_mds_session_head *h;
 845         int i = -1;
 846         int metadata_bytes = 0;
 847         int metadata_key_count = 0;
 848         struct ceph_options *opt = mdsc->fsc->client->options;
 849         struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
 850         void *p;
 851
 852         const char* metadata[][2] = {
 853                 {"hostname", utsname()->nodename},
 854                 {"kernel_version", utsname()->release},
 855                 {"entity_id", opt->name ? : ""},
 856                 {"root", fsopt->server_path ? : "/"},
 857                 {NULL, NULL}
 858         };
 859
 860         /* Calculate serialized length of metadata */
 861         metadata_bytes = 4;  /* map length */
 862         for (i = 0; metadata[i][0] != NULL; ++i) {
 863                 metadata_bytes += 8 + strlen(metadata[i][0]) +
 864                         strlen(metadata[i][1]);
 865                 metadata_key_count++;
 866         }
 867
 868         /* Allocate the message */
 869         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + metadata_bytes,
 870                            GFP_NOFS, false);
 871         if (!msg) {
 872                 pr_err("create_session_msg ENOMEM creating msg\n");
 873                 return NULL;
 874         }
 875         h = msg->front.iov_base;
 876         h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
 877         h->seq = cpu_to_le64(seq);
 878
 879         /*
 880          * Serialize client metadata into waiting buffer space, using
 881          * the format that userspace expects for map<string, string>
 882          *
 883          * ClientSession messages with metadata are v2
 884          */
 885         msg->hdr.version = cpu_to_le16(2);
 886         msg->hdr.compat_version = cpu_to_le16(1);
 887
 888         /* The write pointer, following the session_head structure */
 889         p = msg->front.iov_base + sizeof(*h);
 890
 891         /* Number of entries in the map */
 892         ceph_encode_32(&p, metadata_key_count);
 893
 894         /* Two length-prefixed strings for each entry in the map */
 895         for (i = 0; metadata[i][0] != NULL; ++i) {
 896                 size_t const key_len = strlen(metadata[i][0]);
 897                 size_t const val_len = strlen(metadata[i][1]);
 898
 899                 ceph_encode_32(&p, key_len);
 900                 memcpy(p, metadata[i][0], key_len);
 901                 p += key_len;
 902                 ceph_encode_32(&p, val_len);
 903                 memcpy(p, metadata[i][1], val_len);
 904                 p += val_len;
 905         }
 906
 907         return msg;
 908 }
 909
 910 /*
 911  * send session open request.
 912  *
 913  * called under mdsc->mutex
 914  */
 915 static int __open_session(struct ceph_mds_client *mdsc,
 916                           struct ceph_mds_session *session)
 917 {
 918         struct ceph_msg *msg;
 919         int mstate;
 920         int mds = session->s_mds;
 921
 922         /* wait for mds to go active? */
 923         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
 924         dout("open_session to mds%d (%s)\n", mds,
 925              ceph_mds_state_name(mstate));
 926         session->s_state = CEPH_MDS_SESSION_OPENING;
 927         session->s_renew_requested = jiffies;
 928
 929         /* send connect message */
 930         msg = create_session_open_msg(mdsc, session->s_seq);
 931         if (!msg)
 932                 return -ENOMEM;
 933         ceph_con_send(&session->s_con, msg);
 934         return 0;
 935 }
 936
 937 /*
 938  * open sessions for any export targets for the given mds
 939  *
 940  * called under mdsc->mutex
 941  */
 942 static struct ceph_mds_session *
 943 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
 944 {
 945         struct ceph_mds_session *session;
 946
 947         session = __ceph_lookup_mds_session(mdsc, target);
 948         if (!session) {
 949                 session = register_session(mdsc, target);
 950                 if (IS_ERR(session))
 951                         return session;
 952         }
 953         if (session->s_state == CEPH_MDS_SESSION_NEW ||
 954             session->s_state == CEPH_MDS_SESSION_CLOSING)
 955                 __open_session(mdsc, session);
 956
 957         return session;
 958 }
 959
 960 struct ceph_mds_session *
 961 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
 962 {
 963         struct ceph_mds_session *session;
 964
 965         dout("open_export_target_session to mds%d\n", target);
 966
 967         mutex_lock(&mdsc->mutex);
 968         session = __open_export_target_session(mdsc, target);
 969         mutex_unlock(&mdsc->mutex);
 970
 971         return session;
 972 }
 973
 974 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
 975                                           struct ceph_mds_session *session)
 976 {
 977         struct ceph_mds_info *mi;
 978         struct ceph_mds_session *ts;
 979         int i, mds = session->s_mds;
 980
 981         if (mds >= mdsc->mdsmap->m_max_mds)
 982                 return;
 983
 984         mi = &mdsc->mdsmap->m_info[mds];
 985         dout("open_export_target_sessions for mds%d (%d targets)\n",
 986              session->s_mds, mi->num_export_targets);
 987
 988         for (i = 0; i < mi->num_export_targets; i++) {
 989                 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
 990                 if (!IS_ERR(ts))
 991                         ceph_put_mds_session(ts);
 992         }
 993 }
 994
 995 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
 996                                            struct ceph_mds_session *session)
 997 {
 998         mutex_lock(&mdsc->mutex);
 999         __open_export_target_sessions(mdsc, session);
1000         mutex_unlock(&mdsc->mutex);
1001 }
1002
1003 /*
1004  * session caps
1005  */
1006
1007 /* caller holds s_cap_lock, we drop it */
1008 static void cleanup_cap_releases(struct ceph_mds_client *mdsc,
1009                                  struct ceph_mds_session *session)
1010         __releases(session->s_cap_lock)
1011 {
1012         LIST_HEAD(tmp_list);
1013         list_splice_init(&session->s_cap_releases, &tmp_list);
1014         session->s_num_cap_releases = 0;
1015         spin_unlock(&session->s_cap_lock);
1016
1017         dout("cleanup_cap_releases mds%d\n", session->s_mds);
1018         while (!list_empty(&tmp_list)) {
1019                 struct ceph_cap *cap;
1020                 /* zero out the in-progress message */
1021                 cap = list_first_entry(&tmp_list,
1022                                         struct ceph_cap, session_caps);
1023                 list_del(&cap->session_caps);
1024                 ceph_put_cap(mdsc, cap);
1025         }
1026 }
1027
1028 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1029                                      struct ceph_mds_session *session)
1030 {
1031         struct ceph_mds_request *req;
1032         struct rb_node *p;
1033
1034         dout("cleanup_session_requests mds%d\n", session->s_mds);
1035         mutex_lock(&mdsc->mutex);
1036         while (!list_empty(&session->s_unsafe)) {
1037                 req = list_first_entry(&session->s_unsafe,
1038                                        struct ceph_mds_request, r_unsafe_item);
1039                 list_del_init(&req->r_unsafe_item);
1040                 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1041                                     req->r_tid);
1042                 __unregister_request(mdsc, req);
1043         }
1044         /* zero r_attempts, so kick_requests() will re-send requests */
1045         p = rb_first(&mdsc->request_tree);
1046         while (p) {
1047                 req = rb_entry(p, struct ceph_mds_request, r_node);
1048                 p = rb_next(p);
1049                 if (req->r_session &&
1050                     req->r_session->s_mds == session->s_mds)
1051                         req->r_attempts = 0;
1052         }
1053         mutex_unlock(&mdsc->mutex);
1054 }
1055
1056 /*
1057  * Helper to safely iterate over all caps associated with a session, with
1058  * special care taken to handle a racing __ceph_remove_cap().
1059  *
1060  * Caller must hold session s_mutex.
1061  */
1062 static int iterate_session_caps(struct ceph_mds_session *session,
1063                                  int (*cb)(struct inode *, struct ceph_cap *,
1064                                             void *), void *arg)
1065 {
1066         struct list_head *p;
1067         struct ceph_cap *cap;
1068         struct inode *inode, *last_inode = NULL;
1069         struct ceph_cap *old_cap = NULL;
1070         int ret;
1071
1072         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1073         spin_lock(&session->s_cap_lock);
1074         p = session->s_caps.next;
1075         while (p != &session->s_caps) {
1076                 cap = list_entry(p, struct ceph_cap, session_caps);
1077                 inode = igrab(&cap->ci->vfs_inode);
1078                 if (!inode) {
1079                         p = p->next;
1080                         continue;
1081                 }
1082                 session->s_cap_iterator = cap;
1083                 spin_unlock(&session->s_cap_lock);
1084
1085                 if (last_inode) {
1086                         iput(last_inode);
1087                         last_inode = NULL;
1088                 }
1089                 if (old_cap) {
1090                         ceph_put_cap(session->s_mdsc, old_cap);
1091                         old_cap = NULL;
1092                 }
1093
1094                 ret = cb(inode, cap, arg);
1095                 last_inode = inode;
1096
1097                 spin_lock(&session->s_cap_lock);
1098                 p = p->next;
1099                 if (cap->ci == NULL) {
1100                         dout("iterate_session_caps  finishing cap %p removal\n",
1101                              cap);
1102                         BUG_ON(cap->session != session);
1103                         cap->session = NULL;
1104                         list_del_init(&cap->session_caps);
1105                         session->s_nr_caps--;
1106                         if (cap->queue_release) {
1107                                 list_add_tail(&cap->session_caps,
1108                                               &session->s_cap_releases);
1109                                 session->s_num_cap_releases++;
1110                         } else {
1111                                 old_cap = cap;  /* put_cap it w/o locks held */
1112                         }
1113                 }
1114                 if (ret < 0)
1115                         goto out;
1116         }
1117         ret = 0;
1118 out:
1119         session->s_cap_iterator = NULL;
1120         spin_unlock(&session->s_cap_lock);
1121
1122         iput(last_inode);
1123         if (old_cap)
1124                 ceph_put_cap(session->s_mdsc, old_cap);
1125
1126         return ret;
1127 }
1128
1129 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1130                                   void *arg)
1131 {
1132         struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1133         struct ceph_inode_info *ci = ceph_inode(inode);
1134         LIST_HEAD(to_remove);
1135         bool drop = false;
1136         bool invalidate = false;
1137
1138         dout("removing cap %p, ci is %p, inode is %p\n",
1139              cap, ci, &ci->vfs_inode);
1140         spin_lock(&ci->i_ceph_lock);
1141         __ceph_remove_cap(cap, false);
1142         if (!ci->i_auth_cap) {
1143                 struct ceph_cap_flush *cf;
1144                 struct ceph_mds_client *mdsc = fsc->mdsc;
1145
1146                 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1147
1148                 if (ci->i_wrbuffer_ref > 0 &&
1149                     ACCESS_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
1150                         invalidate = true;
1151
1152                 while (!list_empty(&ci->i_cap_flush_list)) {
1153                         cf = list_first_entry(&ci->i_cap_flush_list,
1154                                               struct ceph_cap_flush, i_list);
1155                         list_move(&cf->i_list, &to_remove);
1156                 }
1157
1158                 spin_lock(&mdsc->cap_dirty_lock);
1159
1160                 list_for_each_entry(cf, &to_remove, i_list)
1161                         list_del(&cf->g_list);
1162
1163                 if (!list_empty(&ci->i_dirty_item)) {
1164                         pr_warn_ratelimited(
1165                                 " dropping dirty %s state for %p %lld\n",
1166                                 ceph_cap_string(ci->i_dirty_caps),
1167                                 inode, ceph_ino(inode));
1168                         ci->i_dirty_caps = 0;
1169                         list_del_init(&ci->i_dirty_item);
1170                         drop = true;
1171                 }
1172                 if (!list_empty(&ci->i_flushing_item)) {
1173                         pr_warn_ratelimited(
1174                                 " dropping dirty+flushing %s state for %p %lld\n",
1175                                 ceph_cap_string(ci->i_flushing_caps),
1176                                 inode, ceph_ino(inode));
1177                         ci->i_flushing_caps = 0;
1178                         list_del_init(&ci->i_flushing_item);
1179                         mdsc->num_cap_flushing--;
1180                         drop = true;
1181                 }
1182                 spin_unlock(&mdsc->cap_dirty_lock);
1183
1184                 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1185                         list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1186                         ci->i_prealloc_cap_flush = NULL;
1187                 }
1188         }
1189         spin_unlock(&ci->i_ceph_lock);
1190         while (!list_empty(&to_remove)) {
1191                 struct ceph_cap_flush *cf;
1192                 cf = list_first_entry(&to_remove,
1193                                       struct ceph_cap_flush, i_list);
1194                 list_del(&cf->i_list);
1195                 ceph_free_cap_flush(cf);
1196         }
1197
1198         wake_up_all(&ci->i_cap_wq);
1199         if (invalidate)
1200                 ceph_queue_invalidate(inode);
1201         if (drop)
1202                 iput(inode);
1203         return 0;
1204 }
1205
1206 /*
1207  * caller must hold session s_mutex
1208  */
1209 static void remove_session_caps(struct ceph_mds_session *session)
1210 {
1211         struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1212         struct super_block *sb = fsc->sb;
1213         dout("remove_session_caps on %p\n", session);
1214         iterate_session_caps(session, remove_session_caps_cb, fsc);
1215
1216         wake_up_all(&fsc->mdsc->cap_flushing_wq);
1217
1218         spin_lock(&session->s_cap_lock);
1219         if (session->s_nr_caps > 0) {
1220                 struct inode *inode;
1221                 struct ceph_cap *cap, *prev = NULL;
1222                 struct ceph_vino vino;
1223                 /*
1224                  * iterate_session_caps() skips inodes that are being
1225                  * deleted, we need to wait until deletions are complete.
1226                  * __wait_on_freeing_inode() is designed for the job,
1227                  * but it is not exported, so use lookup inode function
1228                  * to access it.
1229                  */
1230                 while (!list_empty(&session->s_caps)) {
1231                         cap = list_entry(session->s_caps.next,
1232                                          struct ceph_cap, session_caps);
1233                         if (cap == prev)
1234                                 break;
1235                         prev = cap;
1236                         vino = cap->ci->i_vino;
1237                         spin_unlock(&session->s_cap_lock);
1238
1239                         inode = ceph_find_inode(sb, vino);
1240                         iput(inode);
1241
1242                         spin_lock(&session->s_cap_lock);
1243                 }
1244         }
1245
1246         // drop cap expires and unlock s_cap_lock
1247         cleanup_cap_releases(session->s_mdsc, session);
1248
1249         BUG_ON(session->s_nr_caps > 0);
1250         BUG_ON(!list_empty(&session->s_cap_flushing));
1251 }
1252
1253 /*
1254  * wake up any threads waiting on this session's caps.  if the cap is
1255  * old (didn't get renewed on the client reconnect), remove it now.
1256  *
1257  * caller must hold s_mutex.
1258  */
1259 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1260                               void *arg)
1261 {
1262         struct ceph_inode_info *ci = ceph_inode(inode);
1263
1264         if (arg) {
1265                 spin_lock(&ci->i_ceph_lock);
1266                 ci->i_wanted_max_size = 0;
1267                 ci->i_requested_max_size = 0;
1268                 spin_unlock(&ci->i_ceph_lock);
1269         }
1270         wake_up_all(&ci->i_cap_wq);
1271         return 0;
1272 }
1273
1274 static void wake_up_session_caps(struct ceph_mds_session *session,
1275                                  int reconnect)
1276 {
1277         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1278         iterate_session_caps(session, wake_up_session_cb,
1279                              (void *)(unsigned long)reconnect);
1280 }
1281
1282 /*
1283  * Send periodic message to MDS renewing all currently held caps.  The
1284  * ack will reset the expiration for all caps from this session.
1285  *
1286  * caller holds s_mutex
1287  */
1288 static int send_renew_caps(struct ceph_mds_client *mdsc,
1289                            struct ceph_mds_session *session)
1290 {
1291         struct ceph_msg *msg;
1292         int state;
1293
1294         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1295             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1296                 pr_info("mds%d caps stale\n", session->s_mds);
1297         session->s_renew_requested = jiffies;
1298
1299         /* do not try to renew caps until a recovering mds has reconnected
1300          * with its clients. */
1301         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1302         if (state < CEPH_MDS_STATE_RECONNECT) {
1303                 dout("send_renew_caps ignoring mds%d (%s)\n",
1304                      session->s_mds, ceph_mds_state_name(state));
1305                 return 0;
1306         }
1307
1308         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1309                 ceph_mds_state_name(state));
1310         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1311                                  ++session->s_renew_seq);
1312         if (!msg)
1313                 return -ENOMEM;
1314         ceph_con_send(&session->s_con, msg);
1315         return 0;
1316 }
1317
1318 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1319                              struct ceph_mds_session *session, u64 seq)
1320 {
1321         struct ceph_msg *msg;
1322
1323         dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1324              session->s_mds, ceph_session_state_name(session->s_state), seq);
1325         msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1326         if (!msg)
1327                 return -ENOMEM;
1328         ceph_con_send(&session->s_con, msg);
1329         return 0;
1330 }
1331
1332
1333 /*
1334  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1335  *
1336  * Called under session->s_mutex
1337  */
1338 static void renewed_caps(struct ceph_mds_client *mdsc,
1339                          struct ceph_mds_session *session, int is_renew)
1340 {
1341         int was_stale;
1342         int wake = 0;
1343
1344         spin_lock(&session->s_cap_lock);
1345         was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1346
1347         session->s_cap_ttl = session->s_renew_requested +
1348                 mdsc->mdsmap->m_session_timeout*HZ;
1349
1350         if (was_stale) {
1351                 if (time_before(jiffies, session->s_cap_ttl)) {
1352                         pr_info("mds%d caps renewed\n", session->s_mds);
1353                         wake = 1;
1354                 } else {
1355                         pr_info("mds%d caps still stale\n", session->s_mds);
1356                 }
1357         }
1358         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1359              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1360              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1361         spin_unlock(&session->s_cap_lock);
1362
1363         if (wake)
1364                 wake_up_session_caps(session, 0);
1365 }
1366
1367 /*
1368  * send a session close request
1369  */
1370 static int request_close_session(struct ceph_mds_client *mdsc,
1371                                  struct ceph_mds_session *session)
1372 {
1373         struct ceph_msg *msg;
1374
1375         dout("request_close_session mds%d state %s seq %lld\n",
1376              session->s_mds, ceph_session_state_name(session->s_state),
1377              session->s_seq);
1378         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1379         if (!msg)
1380                 return -ENOMEM;
1381         ceph_con_send(&session->s_con, msg);
1382         return 1;
1383 }
1384
1385 /*
1386  * Called with s_mutex held.
1387  */
1388 static int __close_session(struct ceph_mds_client *mdsc,
1389                          struct ceph_mds_session *session)
1390 {
1391         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1392                 return 0;
1393         session->s_state = CEPH_MDS_SESSION_CLOSING;
1394         return request_close_session(mdsc, session);
1395 }
1396
1397 /*
1398  * Trim old(er) caps.
1399  *
1400  * Because we can't cache an inode without one or more caps, we do
1401  * this indirectly: if a cap is unused, we prune its aliases, at which
1402  * point the inode will hopefully get dropped to.
1403  *
1404  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1405  * memory pressure from the MDS, though, so it needn't be perfect.
1406  */
1407 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1408 {
1409         struct ceph_mds_session *session = arg;
1410         struct ceph_inode_info *ci = ceph_inode(inode);
1411         int used, wanted, oissued, mine;
1412
1413         if (session->s_trim_caps <= 0)
1414                 return -1;
1415
1416         spin_lock(&ci->i_ceph_lock);
1417         mine = cap->issued | cap->implemented;
1418         used = __ceph_caps_used(ci);
1419         wanted = __ceph_caps_file_wanted(ci);
1420         oissued = __ceph_caps_issued_other(ci, cap);
1421
1422         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1423              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1424              ceph_cap_string(used), ceph_cap_string(wanted));
1425         if (cap == ci->i_auth_cap) {
1426                 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1427                     !list_empty(&ci->i_cap_snaps))
1428                         goto out;
1429                 if ((used | wanted) & CEPH_CAP_ANY_WR)
1430                         goto out;
1431         }
1432         /* The inode has cached pages, but it's no longer used.
1433          * we can safely drop it */
1434         if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1435             !(oissued & CEPH_CAP_FILE_CACHE)) {
1436           used = 0;
1437           oissued = 0;
1438         }
1439         if ((used | wanted) & ~oissued & mine)
1440                 goto out;   /* we need these caps */
1441
1442         session->s_trim_caps--;
1443         if (oissued) {
1444                 /* we aren't the only cap.. just remove us */
1445                 __ceph_remove_cap(cap, true);
1446         } else {
1447                 /* try dropping referring dentries */
1448                 spin_unlock(&ci->i_ceph_lock);
1449                 d_prune_aliases(inode);
1450                 dout("trim_caps_cb %p cap %p  pruned, count now %d\n",
1451                      inode, cap, atomic_read(&inode->i_count));
1452                 return 0;
1453         }
1454
1455 out:
1456         spin_unlock(&ci->i_ceph_lock);
1457         return 0;
1458 }
1459
1460 /*
1461  * Trim session cap count down to some max number.
1462  */
1463 static int trim_caps(struct ceph_mds_client *mdsc,
1464                      struct ceph_mds_session *session,
1465                      int max_caps)
1466 {
1467         int trim_caps = session->s_nr_caps - max_caps;
1468
1469         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1470              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1471         if (trim_caps > 0) {
1472                 session->s_trim_caps = trim_caps;
1473                 iterate_session_caps(session, trim_caps_cb, session);
1474                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1475                      session->s_mds, session->s_nr_caps, max_caps,
1476                         trim_caps - session->s_trim_caps);
1477                 session->s_trim_caps = 0;
1478         }
1479
1480         ceph_send_cap_releases(mdsc, session);
1481         return 0;
1482 }
1483
1484 static int check_caps_flush(struct ceph_mds_client *mdsc,
1485                             u64 want_flush_tid)
1486 {
1487         int ret = 1;
1488
1489         spin_lock(&mdsc->cap_dirty_lock);
1490         if (!list_empty(&mdsc->cap_flush_list)) {
1491                 struct ceph_cap_flush *cf =
1492                         list_first_entry(&mdsc->cap_flush_list,
1493                                          struct ceph_cap_flush, g_list);
1494                 if (cf->tid <= want_flush_tid) {
1495                         dout("check_caps_flush still flushing tid "
1496                              "%llu <= %llu\n", cf->tid, want_flush_tid);
1497                         ret = 0;
1498                 }
1499         }
1500         spin_unlock(&mdsc->cap_dirty_lock);
1501         return ret;
1502 }
1503
1504 /*
1505  * flush all dirty inode data to disk.
1506  *
1507  * returns true if we've flushed through want_flush_tid
1508  */
1509 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1510                             u64 want_flush_tid)
1511 {
1512         dout("check_caps_flush want %llu\n", want_flush_tid);
1513
1514         wait_event(mdsc->cap_flushing_wq,
1515                    check_caps_flush(mdsc, want_flush_tid));
1516
1517         dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1518 }
1519
1520 /*
1521  * called under s_mutex
1522  */
1523 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1524                             struct ceph_mds_session *session)
1525 {
1526         struct ceph_msg *msg = NULL;
1527         struct ceph_mds_cap_release *head;
1528         struct ceph_mds_cap_item *item;
1529         struct ceph_cap *cap;
1530         LIST_HEAD(tmp_list);
1531         int num_cap_releases;
1532
1533         spin_lock(&session->s_cap_lock);
1534 again:
1535         list_splice_init(&session->s_cap_releases, &tmp_list);
1536         num_cap_releases = session->s_num_cap_releases;
1537         session->s_num_cap_releases = 0;
1538         spin_unlock(&session->s_cap_lock);
1539
1540         while (!list_empty(&tmp_list)) {
1541                 if (!msg) {
1542                         msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1543                                         PAGE_SIZE, GFP_NOFS, false);
1544                         if (!msg)
1545                                 goto out_err;
1546                         head = msg->front.iov_base;
1547                         head->num = cpu_to_le32(0);
1548                         msg->front.iov_len = sizeof(*head);
1549                 }
1550                 cap = list_first_entry(&tmp_list, struct ceph_cap,
1551                                         session_caps);
1552                 list_del(&cap->session_caps);
1553                 num_cap_releases--;
1554
1555                 head = msg->front.iov_base;
1556                 le32_add_cpu(&head->num, 1);
1557                 item = msg->front.iov_base + msg->front.iov_len;
1558                 item->ino = cpu_to_le64(cap->cap_ino);
1559                 item->cap_id = cpu_to_le64(cap->cap_id);
1560                 item->migrate_seq = cpu_to_le32(cap->mseq);
1561                 item->seq = cpu_to_le32(cap->issue_seq);
1562                 msg->front.iov_len += sizeof(*item);
1563
1564                 ceph_put_cap(mdsc, cap);
1565
1566                 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1567                         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1568                         dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1569                         ceph_con_send(&session->s_con, msg);
1570                         msg = NULL;
1571                 }
1572         }
1573
1574         BUG_ON(num_cap_releases != 0);
1575
1576         spin_lock(&session->s_cap_lock);
1577         if (!list_empty(&session->s_cap_releases))
1578                 goto again;
1579         spin_unlock(&session->s_cap_lock);
1580
1581         if (msg) {
1582                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1583                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1584                 ceph_con_send(&session->s_con, msg);
1585         }
1586         return;
1587 out_err:
1588         pr_err("send_cap_releases mds%d, failed to allocate message\n",
1589                 session->s_mds);
1590         spin_lock(&session->s_cap_lock);
1591         list_splice(&tmp_list, &session->s_cap_releases);
1592         session->s_num_cap_releases += num_cap_releases;
1593         spin_unlock(&session->s_cap_lock);
1594 }
1595
1596 /*
1597  * requests
1598  */
1599
1600 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1601                                     struct inode *dir)
1602 {
1603         struct ceph_inode_info *ci = ceph_inode(dir);
1604         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1605         struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1606         size_t size = sizeof(struct ceph_mds_reply_dir_entry);
1607         int order, num_entries;
1608
1609         spin_lock(&ci->i_ceph_lock);
1610         num_entries = ci->i_files + ci->i_subdirs;
1611         spin_unlock(&ci->i_ceph_lock);
1612         num_entries = max(num_entries, 1);
1613         num_entries = min(num_entries, opt->max_readdir);
1614
1615         order = get_order(size * num_entries);
1616         while (order >= 0) {
1617                 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
1618                                                              __GFP_NOWARN,
1619                                                              order);
1620                 if (rinfo->dir_entries)
1621                         break;
1622                 order--;
1623         }
1624         if (!rinfo->dir_entries)
1625                 return -ENOMEM;
1626
1627         num_entries = (PAGE_SIZE << order) / size;
1628         num_entries = min(num_entries, opt->max_readdir);
1629
1630         rinfo->dir_buf_size = PAGE_SIZE << order;
1631         req->r_num_caps = num_entries + 1;
1632         req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1633         req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1634         return 0;
1635 }
1636
1637 /*
1638  * Create an mds request.
1639  */
1640 struct ceph_mds_request *
1641 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1642 {
1643         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1644
1645         if (!req)
1646                 return ERR_PTR(-ENOMEM);
1647
1648         mutex_init(&req->r_fill_mutex);
1649         req->r_mdsc = mdsc;
1650         req->r_started = jiffies;
1651         req->r_resend_mds = -1;
1652         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1653         INIT_LIST_HEAD(&req->r_unsafe_target_item);
1654         req->r_fmode = -1;
1655         kref_init(&req->r_kref);
1656         RB_CLEAR_NODE(&req->r_node);
1657         INIT_LIST_HEAD(&req->r_wait);
1658         init_completion(&req->r_completion);
1659         init_completion(&req->r_safe_completion);
1660         INIT_LIST_HEAD(&req->r_unsafe_item);
1661
1662         req->r_stamp = current_fs_time(mdsc->fsc->sb);
1663
1664         req->r_op = op;
1665         req->r_direct_mode = mode;
1666         return req;
1667 }
1668
1669 /*
1670  * return oldest (lowest) request, tid in request tree, 0 if none.
1671  *
1672  * called under mdsc->mutex.
1673  */
1674 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1675 {
1676         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1677                 return NULL;
1678         return rb_entry(rb_first(&mdsc->request_tree),
1679                         struct ceph_mds_request, r_node);
1680 }
1681
1682 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1683 {
1684         return mdsc->oldest_tid;
1685 }
1686
1687 /*
1688  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1689  * on build_path_from_dentry in fs/cifs/dir.c.
1690  *
1691  * If @stop_on_nosnap, generate path relative to the first non-snapped
1692  * inode.
1693  *
1694  * Encode hidden .snap dirs as a double /, i.e.
1695  *   foo/.snap/bar -> foo//bar
1696  */
1697 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1698                            int stop_on_nosnap)
1699 {
1700         struct dentry *temp;
1701         char *path;
1702         int len, pos;
1703         unsigned seq;
1704
1705         if (dentry == NULL)
1706                 return ERR_PTR(-EINVAL);
1707
1708 retry:
1709         len = 0;
1710         seq = read_seqbegin(&rename_lock);
1711         rcu_read_lock();
1712         for (temp = dentry; !IS_ROOT(temp);) {
1713                 struct inode *inode = d_inode(temp);
1714                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1715                         len++;  /* slash only */
1716                 else if (stop_on_nosnap && inode &&
1717                          ceph_snap(inode) == CEPH_NOSNAP)
1718                         break;
1719                 else
1720                         len += 1 + temp->d_name.len;
1721                 temp = temp->d_parent;
1722         }
1723         rcu_read_unlock();
1724         if (len)
1725                 len--;  /* no leading '/' */
1726
1727         path = kmalloc(len+1, GFP_NOFS);
1728         if (path == NULL)
1729                 return ERR_PTR(-ENOMEM);
1730         pos = len;
1731         path[pos] = 0;  /* trailing null */
1732         rcu_read_lock();
1733         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1734                 struct inode *inode;
1735
1736                 spin_lock(&temp->d_lock);
1737                 inode = d_inode(temp);
1738                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1739                         dout("build_path path+%d: %p SNAPDIR\n",
1740                              pos, temp);
1741                 } else if (stop_on_nosnap && inode &&
1742                            ceph_snap(inode) == CEPH_NOSNAP) {
1743                         spin_unlock(&temp->d_lock);
1744                         break;
1745                 } else {
1746                         pos -= temp->d_name.len;
1747                         if (pos < 0) {
1748                                 spin_unlock(&temp->d_lock);
1749                                 break;
1750                         }
1751                         strncpy(path + pos, temp->d_name.name,
1752                                 temp->d_name.len);
1753                 }
1754                 spin_unlock(&temp->d_lock);
1755                 if (pos)
1756                         path[--pos] = '/';
1757                 temp = temp->d_parent;
1758         }
1759         rcu_read_unlock();
1760         if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1761                 pr_err("build_path did not end path lookup where "
1762                        "expected, namelen is %d, pos is %d\n", len, pos);
1763                 /* presumably this is only possible if racing with a
1764                    rename of one of the parent directories (we can not
1765                    lock the dentries above us to prevent this, but
1766                    retrying should be harmless) */
1767                 kfree(path);
1768                 goto retry;
1769         }
1770
1771         *base = ceph_ino(d_inode(temp));
1772         *plen = len;
1773         dout("build_path on %p %d built %llx '%.*s'\n",
1774              dentry, d_count(dentry), *base, len, path);
1775         return path;
1776 }
1777
1778 static int build_dentry_path(struct dentry *dentry,
1779                              const char **ppath, int *ppathlen, u64 *pino,
1780                              int *pfreepath)
1781 {
1782         char *path;
1783
1784         if (ceph_snap(d_inode(dentry->d_parent)) == CEPH_NOSNAP) {
1785                 *pino = ceph_ino(d_inode(dentry->d_parent));
1786                 *ppath = dentry->d_name.name;
1787                 *ppathlen = dentry->d_name.len;
1788                 return 0;
1789         }
1790         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1791         if (IS_ERR(path))
1792                 return PTR_ERR(path);
1793         *ppath = path;
1794         *pfreepath = 1;
1795         return 0;
1796 }
1797
1798 static int build_inode_path(struct inode *inode,
1799                             const char **ppath, int *ppathlen, u64 *pino,
1800                             int *pfreepath)
1801 {
1802         struct dentry *dentry;
1803         char *path;
1804
1805         if (ceph_snap(inode) == CEPH_NOSNAP) {
1806                 *pino = ceph_ino(inode);
1807                 *ppathlen = 0;
1808                 return 0;
1809         }
1810         dentry = d_find_alias(inode);
1811         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1812         dput(dentry);
1813         if (IS_ERR(path))
1814                 return PTR_ERR(path);
1815         *ppath = path;
1816         *pfreepath = 1;
1817         return 0;
1818 }
1819
1820 /*
1821  * request arguments may be specified via an inode *, a dentry *, or
1822  * an explicit ino+path.
1823  */
1824 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1825                                   const char *rpath, u64 rino,
1826                                   const char **ppath, int *pathlen,
1827                                   u64 *ino, int *freepath)
1828 {
1829         int r = 0;
1830
1831         if (rinode) {
1832                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1833                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1834                      ceph_snap(rinode));
1835         } else if (rdentry) {
1836                 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1837                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1838                      *ppath);
1839         } else if (rpath || rino) {
1840                 *ino = rino;
1841                 *ppath = rpath;
1842                 *pathlen = rpath ? strlen(rpath) : 0;
1843                 dout(" path %.*s\n", *pathlen, rpath);
1844         }
1845
1846         return r;
1847 }
1848
1849 /*
1850  * called under mdsc->mutex
1851  */
1852 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1853                                                struct ceph_mds_request *req,
1854                                                int mds, bool drop_cap_releases)
1855 {
1856         struct ceph_msg *msg;
1857         struct ceph_mds_request_head *head;
1858         const char *path1 = NULL;
1859         const char *path2 = NULL;
1860         u64 ino1 = 0, ino2 = 0;
1861         int pathlen1 = 0, pathlen2 = 0;
1862         int freepath1 = 0, freepath2 = 0;
1863         int len;
1864         u16 releases;
1865         void *p, *end;
1866         int ret;
1867
1868         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1869                               req->r_path1, req->r_ino1.ino,
1870                               &path1, &pathlen1, &ino1, &freepath1);
1871         if (ret < 0) {
1872                 msg = ERR_PTR(ret);
1873                 goto out;
1874         }
1875
1876         ret = set_request_path_attr(NULL, req->r_old_dentry,
1877                               req->r_path2, req->r_ino2.ino,
1878                               &path2, &pathlen2, &ino2, &freepath2);
1879         if (ret < 0) {
1880                 msg = ERR_PTR(ret);
1881                 goto out_free1;
1882         }
1883
1884         len = sizeof(*head) +
1885                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
1886                 sizeof(struct ceph_timespec);
1887
1888         /* calculate (max) length for cap releases */
1889         len += sizeof(struct ceph_mds_request_release) *
1890                 (!!req->r_inode_drop + !!req->r_dentry_drop +
1891                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1892         if (req->r_dentry_drop)
1893                 len += req->r_dentry->d_name.len;
1894         if (req->r_old_dentry_drop)
1895                 len += req->r_old_dentry->d_name.len;
1896
1897         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1898         if (!msg) {
1899                 msg = ERR_PTR(-ENOMEM);
1900                 goto out_free2;
1901         }
1902
1903         msg->hdr.version = cpu_to_le16(2);
1904         msg->hdr.tid = cpu_to_le64(req->r_tid);
1905
1906         head = msg->front.iov_base;
1907         p = msg->front.iov_base + sizeof(*head);
1908         end = msg->front.iov_base + msg->front.iov_len;
1909
1910         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1911         head->op = cpu_to_le32(req->r_op);
1912         head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1913         head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1914         head->args = req->r_args;
1915
1916         ceph_encode_filepath(&p, end, ino1, path1);
1917         ceph_encode_filepath(&p, end, ino2, path2);
1918
1919         /* make note of release offset, in case we need to replay */
1920         req->r_request_release_offset = p - msg->front.iov_base;
1921
1922         /* cap releases */
1923         releases = 0;
1924         if (req->r_inode_drop)
1925                 releases += ceph_encode_inode_release(&p,
1926                       req->r_inode ? req->r_inode : d_inode(req->r_dentry),
1927                       mds, req->r_inode_drop, req->r_inode_unless, 0);
1928         if (req->r_dentry_drop)
1929                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1930                        mds, req->r_dentry_drop, req->r_dentry_unless);
1931         if (req->r_old_dentry_drop)
1932                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1933                        mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1934         if (req->r_old_inode_drop)
1935                 releases += ceph_encode_inode_release(&p,
1936                       d_inode(req->r_old_dentry),
1937                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1938
1939         if (drop_cap_releases) {
1940                 releases = 0;
1941                 p = msg->front.iov_base + req->r_request_release_offset;
1942         }
1943
1944         head->num_releases = cpu_to_le16(releases);
1945
1946         /* time stamp */
1947         {
1948                 struct ceph_timespec ts;
1949                 ceph_encode_timespec(&ts, &req->r_stamp);
1950                 ceph_encode_copy(&p, &ts, sizeof(ts));
1951         }
1952
1953         BUG_ON(p > end);
1954         msg->front.iov_len = p - msg->front.iov_base;
1955         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1956
1957         if (req->r_pagelist) {
1958                 struct ceph_pagelist *pagelist = req->r_pagelist;
1959                 atomic_inc(&pagelist->refcnt);
1960                 ceph_msg_data_add_pagelist(msg, pagelist);
1961                 msg->hdr.data_len = cpu_to_le32(pagelist->length);
1962         } else {
1963                 msg->hdr.data_len = 0;
1964         }
1965
1966         msg->hdr.data_off = cpu_to_le16(0);
1967
1968 out_free2:
1969         if (freepath2)
1970                 kfree((char *)path2);
1971 out_free1:
1972         if (freepath1)
1973                 kfree((char *)path1);
1974 out:
1975         return msg;
1976 }
1977
1978 /*
1979  * called under mdsc->mutex if error, under no mutex if
1980  * success.
1981  */
1982 static void complete_request(struct ceph_mds_client *mdsc,
1983                              struct ceph_mds_request *req)
1984 {
1985         if (req->r_callback)
1986                 req->r_callback(mdsc, req);
1987         else
1988                 complete_all(&req->r_completion);
1989 }
1990
1991 /*
1992  * called under mdsc->mutex
1993  */
1994 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1995                                   struct ceph_mds_request *req,
1996                                   int mds, bool drop_cap_releases)
1997 {
1998         struct ceph_mds_request_head *rhead;
1999         struct ceph_msg *msg;
2000         int flags = 0;
2001
2002         req->r_attempts++;
2003         if (req->r_inode) {
2004                 struct ceph_cap *cap =
2005                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2006
2007                 if (cap)
2008                         req->r_sent_on_mseq = cap->mseq;
2009                 else
2010                         req->r_sent_on_mseq = -1;
2011         }
2012         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2013              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2014
2015         if (req->r_got_unsafe) {
2016                 void *p;
2017                 /*
2018                  * Replay.  Do not regenerate message (and rebuild
2019                  * paths, etc.); just use the original message.
2020                  * Rebuilding paths will break for renames because
2021                  * d_move mangles the src name.
2022                  */
2023                 msg = req->r_request;
2024                 rhead = msg->front.iov_base;
2025
2026                 flags = le32_to_cpu(rhead->flags);
2027                 flags |= CEPH_MDS_FLAG_REPLAY;
2028                 rhead->flags = cpu_to_le32(flags);
2029
2030                 if (req->r_target_inode)
2031                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2032
2033                 rhead->num_retry = req->r_attempts - 1;
2034
2035                 /* remove cap/dentry releases from message */
2036                 rhead->num_releases = 0;
2037
2038                 /* time stamp */
2039                 p = msg->front.iov_base + req->r_request_release_offset;
2040                 {
2041                         struct ceph_timespec ts;
2042                         ceph_encode_timespec(&ts, &req->r_stamp);
2043                         ceph_encode_copy(&p, &ts, sizeof(ts));
2044                 }
2045
2046                 msg->front.iov_len = p - msg->front.iov_base;
2047                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2048                 return 0;
2049         }
2050
2051         if (req->r_request) {
2052                 ceph_msg_put(req->r_request);
2053                 req->r_request = NULL;
2054         }
2055         msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2056         if (IS_ERR(msg)) {
2057                 req->r_err = PTR_ERR(msg);
2058                 return PTR_ERR(msg);
2059         }
2060         req->r_request = msg;
2061
2062         rhead = msg->front.iov_base;
2063         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2064         if (req->r_got_unsafe)
2065                 flags |= CEPH_MDS_FLAG_REPLAY;
2066         if (req->r_locked_dir)
2067                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2068         rhead->flags = cpu_to_le32(flags);
2069         rhead->num_fwd = req->r_num_fwd;
2070         rhead->num_retry = req->r_attempts - 1;
2071         rhead->ino = 0;
2072
2073         dout(" r_locked_dir = %p\n", req->r_locked_dir);
2074         return 0;
2075 }
2076
2077 /*
2078  * send request, or put it on the appropriate wait list.
2079  */
2080 static int __do_request(struct ceph_mds_client *mdsc,
2081                         struct ceph_mds_request *req)
2082 {
2083         struct ceph_mds_session *session = NULL;
2084         int mds = -1;
2085         int err = 0;
2086
2087         if (req->r_err || req->r_got_result) {
2088                 if (req->r_aborted)
2089                         __unregister_request(mdsc, req);
2090                 goto out;
2091         }
2092
2093         if (req->r_timeout &&
2094             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2095                 dout("do_request timed out\n");
2096                 err = -EIO;
2097                 goto finish;
2098         }
2099         if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2100                 dout("do_request forced umount\n");
2101                 err = -EIO;
2102                 goto finish;
2103         }
2104         if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2105                 if (mdsc->mdsmap_err) {
2106                         err = mdsc->mdsmap_err;
2107                         dout("do_request mdsmap err %d\n", err);
2108                         goto finish;
2109                 }
2110                 if (mdsc->mdsmap->m_epoch == 0) {
2111                         dout("do_request no mdsmap, waiting for map\n");
2112                         list_add(&req->r_wait, &mdsc->waiting_for_map);
2113                         goto finish;
2114                 }
2115                 if (!(mdsc->fsc->mount_options->flags &
2116                       CEPH_MOUNT_OPT_MOUNTWAIT) &&
2117                     !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2118                         err = -ENOENT;
2119                         pr_info("probably no mds server is up\n");
2120                         goto finish;
2121                 }
2122         }
2123
2124         put_request_session(req);
2125
2126         mds = __choose_mds(mdsc, req);
2127         if (mds < 0 ||
2128             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2129                 dout("do_request no mds or not active, waiting for map\n");
2130                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2131                 goto out;
2132         }
2133
2134         /* get, open session */
2135         session = __ceph_lookup_mds_session(mdsc, mds);
2136         if (!session) {
2137                 session = register_session(mdsc, mds);
2138                 if (IS_ERR(session)) {
2139                         err = PTR_ERR(session);
2140                         goto finish;
2141                 }
2142         }
2143         req->r_session = get_session(session);
2144
2145         dout("do_request mds%d session %p state %s\n", mds, session,
2146              ceph_session_state_name(session->s_state));
2147         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2148             session->s_state != CEPH_MDS_SESSION_HUNG) {
2149                 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2150                         err = -EACCES;
2151                         goto out_session;
2152                 }
2153                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2154                     session->s_state == CEPH_MDS_SESSION_CLOSING)
2155                         __open_session(mdsc, session);
2156                 list_add(&req->r_wait, &session->s_waiting);
2157                 goto out_session;
2158         }
2159
2160         /* send request */
2161         req->r_resend_mds = -1;   /* forget any previous mds hint */
2162
2163         if (req->r_request_started == 0)   /* note request start time */
2164                 req->r_request_started = jiffies;
2165
2166         err = __prepare_send_request(mdsc, req, mds, false);
2167         if (!err) {
2168                 ceph_msg_get(req->r_request);
2169                 ceph_con_send(&session->s_con, req->r_request);
2170         }
2171
2172 out_session:
2173         ceph_put_mds_session(session);
2174 finish:
2175         if (err) {
2176                 dout("__do_request early error %d\n", err);
2177                 req->r_err = err;
2178                 complete_request(mdsc, req);
2179                 __unregister_request(mdsc, req);
2180         }
2181 out:
2182         return err;
2183 }
2184
2185 /*
2186  * called under mdsc->mutex
2187  */
2188 static void __wake_requests(struct ceph_mds_client *mdsc,
2189                             struct list_head *head)
2190 {
2191         struct ceph_mds_request *req;
2192         LIST_HEAD(tmp_list);
2193
2194         list_splice_init(head, &tmp_list);
2195
2196         while (!list_empty(&tmp_list)) {
2197                 req = list_entry(tmp_list.next,
2198                                  struct ceph_mds_request, r_wait);
2199                 list_del_init(&req->r_wait);
2200                 dout(" wake request %p tid %llu\n", req, req->r_tid);
2201                 __do_request(mdsc, req);
2202         }
2203 }
2204
2205 /*
2206  * Wake up threads with requests pending for @mds, so that they can
2207  * resubmit their requests to a possibly different mds.
2208  */
2209 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2210 {
2211         struct ceph_mds_request *req;
2212         struct rb_node *p = rb_first(&mdsc->request_tree);
2213
2214         dout("kick_requests mds%d\n", mds);
2215         while (p) {
2216                 req = rb_entry(p, struct ceph_mds_request, r_node);
2217                 p = rb_next(p);
2218                 if (req->r_got_unsafe)
2219                         continue;
2220                 if (req->r_attempts > 0)
2221                         continue; /* only new requests */
2222                 if (req->r_session &&
2223                     req->r_session->s_mds == mds) {
2224                         dout(" kicking tid %llu\n", req->r_tid);
2225                         list_del_init(&req->r_wait);
2226                         __do_request(mdsc, req);
2227                 }
2228         }
2229 }
2230
2231 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2232                               struct ceph_mds_request *req)
2233 {
2234         dout("submit_request on %p\n", req);
2235         mutex_lock(&mdsc->mutex);
2236         __register_request(mdsc, req, NULL);
2237         __do_request(mdsc, req);
2238         mutex_unlock(&mdsc->mutex);
2239 }
2240
2241 /*
2242  * Synchrously perform an mds request.  Take care of all of the
2243  * session setup, forwarding, retry details.
2244  */
2245 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2246                          struct inode *dir,
2247                          struct ceph_mds_request *req)
2248 {
2249         int err;
2250
2251         dout("do_request on %p\n", req);
2252
2253         /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
2254         if (req->r_inode)
2255                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2256         if (req->r_locked_dir)
2257                 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
2258         if (req->r_old_dentry_dir)
2259                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2260                                   CEPH_CAP_PIN);
2261
2262         /* issue */
2263         mutex_lock(&mdsc->mutex);
2264         __register_request(mdsc, req, dir);
2265         __do_request(mdsc, req);
2266
2267         if (req->r_err) {
2268                 err = req->r_err;
2269                 goto out;
2270         }
2271
2272         /* wait */
2273         mutex_unlock(&mdsc->mutex);
2274         dout("do_request waiting\n");
2275         if (!req->r_timeout && req->r_wait_for_completion) {
2276                 err = req->r_wait_for_completion(mdsc, req);
2277         } else {
2278                 long timeleft = wait_for_completion_killable_timeout(
2279                                         &req->r_completion,
2280                                         ceph_timeout_jiffies(req->r_timeout));
2281                 if (timeleft > 0)
2282                         err = 0;
2283                 else if (!timeleft)
2284                         err = -EIO;  /* timed out */
2285                 else
2286                         err = timeleft;  /* killed */
2287         }
2288         dout("do_request waited, got %d\n", err);
2289         mutex_lock(&mdsc->mutex);
2290
2291         /* only abort if we didn't race with a real reply */
2292         if (req->r_got_result) {
2293                 err = le32_to_cpu(req->r_reply_info.head->result);
2294         } else if (err < 0) {
2295                 dout("aborted request %lld with %d\n", req->r_tid, err);
2296
2297                 /*
2298                  * ensure we aren't running concurrently with
2299                  * ceph_fill_trace or ceph_readdir_prepopulate, which
2300                  * rely on locks (dir mutex) held by our caller.
2301                  */
2302                 mutex_lock(&req->r_fill_mutex);
2303                 req->r_err = err;
2304                 req->r_aborted = true;
2305                 mutex_unlock(&req->r_fill_mutex);
2306
2307                 if (req->r_locked_dir &&
2308                     (req->r_op & CEPH_MDS_OP_WRITE))
2309                         ceph_invalidate_dir_request(req);
2310         } else {
2311                 err = req->r_err;
2312         }
2313
2314 out:
2315         mutex_unlock(&mdsc->mutex);
2316         dout("do_request %p done, result %d\n", req, err);
2317         return err;
2318 }
2319
2320 /*
2321  * Invalidate dir's completeness, dentry lease state on an aborted MDS
2322  * namespace request.
2323  */
2324 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2325 {
2326         struct inode *inode = req->r_locked_dir;
2327
2328         dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2329
2330         ceph_dir_clear_complete(inode);
2331         if (req->r_dentry)
2332                 ceph_invalidate_dentry_lease(req->r_dentry);
2333         if (req->r_old_dentry)
2334                 ceph_invalidate_dentry_lease(req->r_old_dentry);
2335 }
2336
2337 /*
2338  * Handle mds reply.
2339  *
2340  * We take the session mutex and parse and process the reply immediately.
2341  * This preserves the logical ordering of replies, capabilities, etc., sent
2342  * by the MDS as they are applied to our local cache.
2343  */
2344 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2345 {
2346         struct ceph_mds_client *mdsc = session->s_mdsc;
2347         struct ceph_mds_request *req;
2348         struct ceph_mds_reply_head *head = msg->front.iov_base;
2349         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2350         struct ceph_snap_realm *realm;
2351         u64 tid;
2352         int err, result;
2353         int mds = session->s_mds;
2354
2355         if (msg->front.iov_len < sizeof(*head)) {
2356                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2357                 ceph_msg_dump(msg);
2358                 return;
2359         }
2360
2361         /* get request, session */
2362         tid = le64_to_cpu(msg->hdr.tid);
2363         mutex_lock(&mdsc->mutex);
2364         req = lookup_get_request(mdsc, tid);
2365         if (!req) {
2366                 dout("handle_reply on unknown tid %llu\n", tid);
2367                 mutex_unlock(&mdsc->mutex);
2368                 return;
2369         }
2370         dout("handle_reply %p\n", req);
2371
2372         /* correct session? */
2373         if (req->r_session != session) {
2374                 pr_err("mdsc_handle_reply got %llu on session mds%d"
2375                        " not mds%d\n", tid, session->s_mds,
2376                        req->r_session ? req->r_session->s_mds : -1);
2377                 mutex_unlock(&mdsc->mutex);
2378                 goto out;
2379         }
2380
2381         /* dup? */
2382         if ((req->r_got_unsafe && !head->safe) ||
2383             (req->r_got_safe && head->safe)) {
2384                 pr_warn("got a dup %s reply on %llu from mds%d\n",
2385                            head->safe ? "safe" : "unsafe", tid, mds);
2386                 mutex_unlock(&mdsc->mutex);
2387                 goto out;
2388         }
2389         if (req->r_got_safe) {
2390                 pr_warn("got unsafe after safe on %llu from mds%d\n",
2391                            tid, mds);
2392                 mutex_unlock(&mdsc->mutex);
2393                 goto out;
2394         }
2395
2396         result = le32_to_cpu(head->result);
2397
2398         /*
2399          * Handle an ESTALE
2400          * if we're not talking to the authority, send to them
2401          * if the authority has changed while we weren't looking,
2402          * send to new authority
2403          * Otherwise we just have to return an ESTALE
2404          */
2405         if (result == -ESTALE) {
2406                 dout("got ESTALE on request %llu", req->r_tid);
2407                 req->r_resend_mds = -1;
2408                 if (req->r_direct_mode != USE_AUTH_MDS) {
2409                         dout("not using auth, setting for that now");
2410                         req->r_direct_mode = USE_AUTH_MDS;
2411                         __do_request(mdsc, req);
2412                         mutex_unlock(&mdsc->mutex);
2413                         goto out;
2414                 } else  {
2415                         int mds = __choose_mds(mdsc, req);
2416                         if (mds >= 0 && mds != req->r_session->s_mds) {
2417                                 dout("but auth changed, so resending");
2418                                 __do_request(mdsc, req);
2419                                 mutex_unlock(&mdsc->mutex);
2420                                 goto out;
2421                         }
2422                 }
2423                 dout("have to return ESTALE on request %llu", req->r_tid);
2424         }
2425
2426
2427         if (head->safe) {
2428                 req->r_got_safe = true;
2429                 __unregister_request(mdsc, req);
2430
2431                 if (req->r_got_unsafe) {
2432                         /*
2433                          * We already handled the unsafe response, now do the
2434                          * cleanup.  No need to examine the response; the MDS
2435                          * doesn't include any result info in the safe
2436                          * response.  And even if it did, there is nothing
2437                          * useful we could do with a revised return value.
2438                          */
2439                         dout("got safe reply %llu, mds%d\n", tid, mds);
2440                         list_del_init(&req->r_unsafe_item);
2441
2442                         /* last unsafe request during umount? */
2443                         if (mdsc->stopping && !__get_oldest_req(mdsc))
2444                                 complete_all(&mdsc->safe_umount_waiters);
2445                         mutex_unlock(&mdsc->mutex);
2446                         goto out;
2447                 }
2448         } else {
2449                 req->r_got_unsafe = true;
2450                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2451                 if (req->r_unsafe_dir) {
2452                         struct ceph_inode_info *ci =
2453                                         ceph_inode(req->r_unsafe_dir);
2454                         spin_lock(&ci->i_unsafe_lock);
2455                         list_add_tail(&req->r_unsafe_dir_item,
2456                                       &ci->i_unsafe_dirops);
2457                         spin_unlock(&ci->i_unsafe_lock);
2458                 }
2459         }
2460
2461         dout("handle_reply tid %lld result %d\n", tid, result);
2462         rinfo = &req->r_reply_info;
2463         err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2464         mutex_unlock(&mdsc->mutex);
2465
2466         mutex_lock(&session->s_mutex);
2467         if (err < 0) {
2468                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2469                 ceph_msg_dump(msg);
2470                 goto out_err;
2471         }
2472
2473         /* snap trace */
2474         realm = NULL;
2475         if (rinfo->snapblob_len) {
2476                 down_write(&mdsc->snap_rwsem);
2477                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2478                                 rinfo->snapblob + rinfo->snapblob_len,
2479                                 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2480                                 &realm);
2481                 downgrade_write(&mdsc->snap_rwsem);
2482         } else {
2483                 down_read(&mdsc->snap_rwsem);
2484         }
2485
2486         /* insert trace into our cache */
2487         mutex_lock(&req->r_fill_mutex);
2488         current->journal_info = req;
2489         err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2490         if (err == 0) {
2491                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2492                                     req->r_op == CEPH_MDS_OP_LSSNAP))
2493                         ceph_readdir_prepopulate(req, req->r_session);
2494                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2495         }
2496         current->journal_info = NULL;
2497         mutex_unlock(&req->r_fill_mutex);
2498
2499         up_read(&mdsc->snap_rwsem);
2500         if (realm)
2501                 ceph_put_snap_realm(mdsc, realm);
2502
2503         if (err == 0 && req->r_got_unsafe && req->r_target_inode) {
2504                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
2505                 spin_lock(&ci->i_unsafe_lock);
2506                 list_add_tail(&req->r_unsafe_target_item, &ci->i_unsafe_iops);
2507                 spin_unlock(&ci->i_unsafe_lock);
2508         }
2509 out_err:
2510         mutex_lock(&mdsc->mutex);
2511         if (!req->r_aborted) {
2512                 if (err) {
2513                         req->r_err = err;
2514                 } else {
2515                         req->r_reply =  ceph_msg_get(msg);
2516                         req->r_got_result = true;
2517                 }
2518         } else {
2519                 dout("reply arrived after request %lld was aborted\n", tid);
2520         }
2521         mutex_unlock(&mdsc->mutex);
2522
2523         mutex_unlock(&session->s_mutex);
2524
2525         /* kick calling process */
2526         complete_request(mdsc, req);
2527 out:
2528         ceph_mdsc_put_request(req);
2529         return;
2530 }
2531
2532
2533
2534 /*
2535  * handle mds notification that our request has been forwarded.
2536  */
2537 static void handle_forward(struct ceph_mds_client *mdsc,
2538                            struct ceph_mds_session *session,
2539                            struct ceph_msg *msg)
2540 {
2541         struct ceph_mds_request *req;
2542         u64 tid = le64_to_cpu(msg->hdr.tid);
2543         u32 next_mds;
2544         u32 fwd_seq;
2545         int err = -EINVAL;
2546         void *p = msg->front.iov_base;
2547         void *end = p + msg->front.iov_len;
2548
2549         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2550         next_mds = ceph_decode_32(&p);
2551         fwd_seq = ceph_decode_32(&p);
2552
2553         mutex_lock(&mdsc->mutex);
2554         req = lookup_get_request(mdsc, tid);
2555         if (!req) {
2556                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2557                 goto out;  /* dup reply? */
2558         }
2559
2560         if (req->r_aborted) {
2561                 dout("forward tid %llu aborted, unregistering\n", tid);
2562                 __unregister_request(mdsc, req);
2563         } else if (fwd_seq <= req->r_num_fwd) {
2564                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2565                      tid, next_mds, req->r_num_fwd, fwd_seq);
2566         } else {
2567                 /* resend. forward race not possible; mds would drop */
2568                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2569                 BUG_ON(req->r_err);
2570                 BUG_ON(req->r_got_result);
2571                 req->r_attempts = 0;
2572                 req->r_num_fwd = fwd_seq;
2573                 req->r_resend_mds = next_mds;
2574                 put_request_session(req);
2575                 __do_request(mdsc, req);
2576         }
2577         ceph_mdsc_put_request(req);
2578 out:
2579         mutex_unlock(&mdsc->mutex);
2580         return;
2581
2582 bad:
2583         pr_err("mdsc_handle_forward decode error err=%d\n", err);
2584 }
2585
2586 /*
2587  * handle a mds session control message
2588  */
2589 static void handle_session(struct ceph_mds_session *session,
2590                            struct ceph_msg *msg)
2591 {
2592         struct ceph_mds_client *mdsc = session->s_mdsc;
2593         u32 op;
2594         u64 seq;
2595         int mds = session->s_mds;
2596         struct ceph_mds_session_head *h = msg->front.iov_base;
2597         int wake = 0;
2598
2599         /* decode */
2600         if (msg->front.iov_len != sizeof(*h))
2601                 goto bad;
2602         op = le32_to_cpu(h->op);
2603         seq = le64_to_cpu(h->seq);
2604
2605         mutex_lock(&mdsc->mutex);
2606         if (op == CEPH_SESSION_CLOSE)
2607                 __unregister_session(mdsc, session);
2608         /* FIXME: this ttl calculation is generous */
2609         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2610         mutex_unlock(&mdsc->mutex);
2611
2612         mutex_lock(&session->s_mutex);
2613
2614         dout("handle_session mds%d %s %p state %s seq %llu\n",
2615              mds, ceph_session_op_name(op), session,
2616              ceph_session_state_name(session->s_state), seq);
2617
2618         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2619                 session->s_state = CEPH_MDS_SESSION_OPEN;
2620                 pr_info("mds%d came back\n", session->s_mds);
2621         }
2622
2623         switch (op) {
2624         case CEPH_SESSION_OPEN:
2625                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2626                         pr_info("mds%d reconnect success\n", session->s_mds);
2627                 session->s_state = CEPH_MDS_SESSION_OPEN;
2628                 renewed_caps(mdsc, session, 0);
2629                 wake = 1;
2630                 if (mdsc->stopping)
2631                         __close_session(mdsc, session);
2632                 break;
2633
2634         case CEPH_SESSION_RENEWCAPS:
2635                 if (session->s_renew_seq == seq)
2636                         renewed_caps(mdsc, session, 1);
2637                 break;
2638
2639         case CEPH_SESSION_CLOSE:
2640                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2641                         pr_info("mds%d reconnect denied\n", session->s_mds);
2642                 cleanup_session_requests(mdsc, session);
2643                 remove_session_caps(session);
2644                 wake = 2; /* for good measure */
2645                 wake_up_all(&mdsc->session_close_wq);
2646                 break;
2647
2648         case CEPH_SESSION_STALE:
2649                 pr_info("mds%d caps went stale, renewing\n",
2650                         session->s_mds);
2651                 spin_lock(&session->s_gen_ttl_lock);
2652                 session->s_cap_gen++;
2653                 session->s_cap_ttl = jiffies - 1;
2654                 spin_unlock(&session->s_gen_ttl_lock);
2655                 send_renew_caps(mdsc, session);
2656                 break;
2657
2658         case CEPH_SESSION_RECALL_STATE:
2659                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2660                 break;
2661
2662         case CEPH_SESSION_FLUSHMSG:
2663                 send_flushmsg_ack(mdsc, session, seq);
2664                 break;
2665
2666         case CEPH_SESSION_FORCE_RO:
2667                 dout("force_session_readonly %p\n", session);
2668                 spin_lock(&session->s_cap_lock);
2669                 session->s_readonly = true;
2670                 spin_unlock(&session->s_cap_lock);
2671                 wake_up_session_caps(session, 0);
2672                 break;
2673
2674         case CEPH_SESSION_REJECT:
2675                 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
2676                 pr_info("mds%d rejected session\n", session->s_mds);
2677                 session->s_state = CEPH_MDS_SESSION_REJECTED;
2678                 cleanup_session_requests(mdsc, session);
2679                 remove_session_caps(session);
2680                 wake = 2; /* for good measure */
2681                 break;
2682
2683         default:
2684                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2685                 WARN_ON(1);
2686         }
2687
2688         mutex_unlock(&session->s_mutex);
2689         if (wake) {
2690                 mutex_lock(&mdsc->mutex);
2691                 __wake_requests(mdsc, &session->s_waiting);
2692                 if (wake == 2)
2693                         kick_requests(mdsc, mds);
2694                 mutex_unlock(&mdsc->mutex);
2695         }
2696         return;
2697
2698 bad:
2699         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2700                (int)msg->front.iov_len);
2701         ceph_msg_dump(msg);
2702         return;
2703 }
2704
2705
2706 /*
2707  * called under session->mutex.
2708  */
2709 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2710                                    struct ceph_mds_session *session)
2711 {
2712         struct ceph_mds_request *req, *nreq;
2713         struct rb_node *p;
2714         int err;
2715
2716         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2717
2718         mutex_lock(&mdsc->mutex);
2719         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2720                 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2721                 if (!err) {
2722                         ceph_msg_get(req->r_request);
2723                         ceph_con_send(&session->s_con, req->r_request);
2724                 }
2725         }
2726
2727         /*
2728          * also re-send old requests when MDS enters reconnect stage. So that MDS
2729          * can process completed request in clientreplay stage.
2730          */
2731         p = rb_first(&mdsc->request_tree);
2732         while (p) {
2733                 req = rb_entry(p, struct ceph_mds_request, r_node);
2734                 p = rb_next(p);
2735                 if (req->r_got_unsafe)
2736                         continue;
2737                 if (req->r_attempts == 0)
2738                         continue; /* only old requests */
2739                 if (req->r_session &&
2740                     req->r_session->s_mds == session->s_mds) {
2741                         err = __prepare_send_request(mdsc, req,
2742                                                      session->s_mds, true);
2743                         if (!err) {
2744                                 ceph_msg_get(req->r_request);
2745                                 ceph_con_send(&session->s_con, req->r_request);
2746                         }
2747                 }
2748         }
2749         mutex_unlock(&mdsc->mutex);
2750 }
2751
2752 /*
2753  * Encode information about a cap for a reconnect with the MDS.
2754  */
2755 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2756                           void *arg)
2757 {
2758         union {
2759                 struct ceph_mds_cap_reconnect v2;
2760                 struct ceph_mds_cap_reconnect_v1 v1;
2761         } rec;
2762         struct ceph_inode_info *ci;
2763         struct ceph_reconnect_state *recon_state = arg;
2764         struct ceph_pagelist *pagelist = recon_state->pagelist;
2765         char *path;
2766         int pathlen, err;
2767         u64 pathbase;
2768         u64 snap_follows;
2769         struct dentry *dentry;
2770
2771         ci = cap->ci;
2772
2773         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2774              inode, ceph_vinop(inode), cap, cap->cap_id,
2775              ceph_cap_string(cap->issued));
2776         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2777         if (err)
2778                 return err;
2779
2780         dentry = d_find_alias(inode);
2781         if (dentry) {
2782                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2783                 if (IS_ERR(path)) {
2784                         err = PTR_ERR(path);
2785                         goto out_dput;
2786                 }
2787         } else {
2788                 path = NULL;
2789                 pathlen = 0;
2790                 pathbase = 0;
2791         }
2792
2793         spin_lock(&ci->i_ceph_lock);
2794         cap->seq = 0;        /* reset cap seq */
2795         cap->issue_seq = 0;  /* and issue_seq */
2796         cap->mseq = 0;       /* and migrate_seq */
2797         cap->cap_gen = cap->session->s_cap_gen;
2798
2799         if (recon_state->msg_version >= 2) {
2800                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2801                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2802                 rec.v2.issued = cpu_to_le32(cap->issued);
2803                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2804                 rec.v2.pathbase = cpu_to_le64(pathbase);
2805                 rec.v2.flock_len = 0;
2806         } else {
2807                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2808                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2809                 rec.v1.issued = cpu_to_le32(cap->issued);
2810                 rec.v1.size = cpu_to_le64(inode->i_size);
2811                 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2812                 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2813                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2814                 rec.v1.pathbase = cpu_to_le64(pathbase);
2815         }
2816
2817         if (list_empty(&ci->i_cap_snaps)) {
2818                 snap_follows = 0;
2819         } else {
2820                 struct ceph_cap_snap *capsnap =
2821                         list_first_entry(&ci->i_cap_snaps,
2822                                          struct ceph_cap_snap, ci_item);
2823                 snap_follows = capsnap->follows;
2824         }
2825         spin_unlock(&ci->i_ceph_lock);
2826
2827         if (recon_state->msg_version >= 2) {
2828                 int num_fcntl_locks, num_flock_locks;
2829                 struct ceph_filelock *flocks;
2830                 size_t struct_len, total_len = 0;
2831                 u8 struct_v = 0;
2832
2833 encode_again:
2834                 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2835                 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2836                                  sizeof(struct ceph_filelock), GFP_NOFS);
2837                 if (!flocks) {
2838                         err = -ENOMEM;
2839                         goto out_free;
2840                 }
2841                 err = ceph_encode_locks_to_buffer(inode, flocks,
2842                                                   num_fcntl_locks,
2843                                                   num_flock_locks);
2844                 if (err) {
2845                         kfree(flocks);
2846                         if (err == -ENOSPC)
2847                                 goto encode_again;
2848                         goto out_free;
2849                 }
2850
2851                 if (recon_state->msg_version >= 3) {
2852                         /* version, compat_version and struct_len */
2853                         total_len = 2 * sizeof(u8) + sizeof(u32);
2854                         struct_v = 2;
2855                 }
2856                 /*
2857                  * number of encoded locks is stable, so copy to pagelist
2858                  */
2859                 struct_len = 2 * sizeof(u32) +
2860                             (num_fcntl_locks + num_flock_locks) *
2861                             sizeof(struct ceph_filelock);
2862                 rec.v2.flock_len = cpu_to_le32(struct_len);
2863
2864                 struct_len += sizeof(rec.v2);
2865                 struct_len += sizeof(u32) + pathlen;
2866
2867                 if (struct_v >= 2)
2868                         struct_len += sizeof(u64); /* snap_follows */
2869
2870                 total_len += struct_len;
2871                 err = ceph_pagelist_reserve(pagelist, total_len);
2872
2873                 if (!err) {
2874                         if (recon_state->msg_version >= 3) {
2875                                 ceph_pagelist_encode_8(pagelist, struct_v);
2876                                 ceph_pagelist_encode_8(pagelist, 1);
2877                                 ceph_pagelist_encode_32(pagelist, struct_len);
2878                         }
2879                         ceph_pagelist_encode_string(pagelist, path, pathlen);
2880                         ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
2881                         ceph_locks_to_pagelist(flocks, pagelist,
2882                                                num_fcntl_locks,
2883                                                num_flock_locks);
2884                         if (struct_v >= 2)
2885                                 ceph_pagelist_encode_64(pagelist, snap_follows);
2886                 }
2887                 kfree(flocks);
2888         } else {
2889                 size_t size = sizeof(u32) + pathlen + sizeof(rec.v1);
2890                 err = ceph_pagelist_reserve(pagelist, size);
2891                 if (!err) {
2892                         ceph_pagelist_encode_string(pagelist, path, pathlen);
2893                         ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
2894                 }
2895         }
2896
2897         recon_state->nr_caps++;
2898 out_free:
2899         kfree(path);
2900 out_dput:
2901         dput(dentry);
2902         return err;
2903 }
2904
2905
2906 /*
2907  * If an MDS fails and recovers, clients need to reconnect in order to
2908  * reestablish shared state.  This includes all caps issued through
2909  * this session _and_ the snap_realm hierarchy.  Because it's not
2910  * clear which snap realms the mds cares about, we send everything we
2911  * know about.. that ensures we'll then get any new info the
2912  * recovering MDS might have.
2913  *
2914  * This is a relatively heavyweight operation, but it's rare.
2915  *
2916  * called with mdsc->mutex held.
2917  */
2918 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2919                                struct ceph_mds_session *session)
2920 {
2921         struct ceph_msg *reply;
2922         struct rb_node *p;
2923         int mds = session->s_mds;
2924         int err = -ENOMEM;
2925         int s_nr_caps;
2926         struct ceph_pagelist *pagelist;
2927         struct ceph_reconnect_state recon_state;
2928
2929         pr_info("mds%d reconnect start\n", mds);
2930
2931         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2932         if (!pagelist)
2933                 goto fail_nopagelist;
2934         ceph_pagelist_init(pagelist);
2935
2936         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2937         if (!reply)
2938                 goto fail_nomsg;
2939
2940         mutex_lock(&session->s_mutex);
2941         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2942         session->s_seq = 0;
2943
2944         dout("session %p state %s\n", session,
2945              ceph_session_state_name(session->s_state));
2946
2947         spin_lock(&session->s_gen_ttl_lock);
2948         session->s_cap_gen++;
2949         spin_unlock(&session->s_gen_ttl_lock);
2950
2951         spin_lock(&session->s_cap_lock);
2952         /* don't know if session is readonly */
2953         session->s_readonly = 0;
2954         /*
2955          * notify __ceph_remove_cap() that we are composing cap reconnect.
2956          * If a cap get released before being added to the cap reconnect,
2957          * __ceph_remove_cap() should skip queuing cap release.
2958          */
2959         session->s_cap_reconnect = 1;
2960         /* drop old cap expires; we're about to reestablish that state */
2961         cleanup_cap_releases(mdsc, session);
2962
2963         /* trim unused caps to reduce MDS's cache rejoin time */
2964         if (mdsc->fsc->sb->s_root)
2965                 shrink_dcache_parent(mdsc->fsc->sb->s_root);
2966
2967         ceph_con_close(&session->s_con);
2968         ceph_con_open(&session->s_con,
2969                       CEPH_ENTITY_TYPE_MDS, mds,
2970                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2971
2972         /* replay unsafe requests */
2973         replay_unsafe_requests(mdsc, session);
2974
2975         down_read(&mdsc->snap_rwsem);
2976
2977         /* traverse this session's caps */
2978         s_nr_caps = session->s_nr_caps;
2979         err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
2980         if (err)
2981                 goto fail;
2982
2983         recon_state.nr_caps = 0;
2984         recon_state.pagelist = pagelist;
2985         if (session->s_con.peer_features & CEPH_FEATURE_MDSENC)
2986                 recon_state.msg_version = 3;
2987         else if (session->s_con.peer_features & CEPH_FEATURE_FLOCK)
2988                 recon_state.msg_version = 2;
2989         else
2990                 recon_state.msg_version = 1;
2991         err = iterate_session_caps(session, encode_caps_cb, &recon_state);
2992         if (err < 0)
2993                 goto fail;
2994
2995         spin_lock(&session->s_cap_lock);
2996         session->s_cap_reconnect = 0;
2997         spin_unlock(&session->s_cap_lock);
2998
2999         /*
3000          * snaprealms.  we provide mds with the ino, seq (version), and
3001          * parent for all of our realms.  If the mds has any newer info,
3002          * it will tell us.
3003          */
3004         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3005                 struct ceph_snap_realm *realm =
3006                         rb_entry(p, struct ceph_snap_realm, node);
3007                 struct ceph_mds_snaprealm_reconnect sr_rec;
3008
3009                 dout(" adding snap realm %llx seq %lld parent %llx\n",
3010                      realm->ino, realm->seq, realm->parent_ino);
3011                 sr_rec.ino = cpu_to_le64(realm->ino);
3012                 sr_rec.seq = cpu_to_le64(realm->seq);
3013                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3014                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3015                 if (err)
3016                         goto fail;
3017         }
3018
3019         reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3020
3021         /* raced with cap release? */
3022         if (s_nr_caps != recon_state.nr_caps) {
3023                 struct page *page = list_first_entry(&pagelist->head,
3024                                                      struct page, lru);
3025                 __le32 *addr = kmap_atomic(page);
3026                 *addr = cpu_to_le32(recon_state.nr_caps);
3027                 kunmap_atomic(addr);
3028         }
3029
3030         reply->hdr.data_len = cpu_to_le32(pagelist->length);
3031         ceph_msg_data_add_pagelist(reply, pagelist);
3032
3033         ceph_early_kick_flushing_caps(mdsc, session);
3034
3035         ceph_con_send(&session->s_con, reply);
3036
3037         mutex_unlock(&session->s_mutex);
3038
3039         mutex_lock(&mdsc->mutex);
3040         __wake_requests(mdsc, &session->s_waiting);
3041         mutex_unlock(&mdsc->mutex);
3042
3043         up_read(&mdsc->snap_rwsem);
3044         return;
3045
3046 fail:
3047         ceph_msg_put(reply);
3048         up_read(&mdsc->snap_rwsem);
3049         mutex_unlock(&session->s_mutex);
3050 fail_nomsg:
3051         ceph_pagelist_release(pagelist);
3052 fail_nopagelist:
3053         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3054         return;
3055 }
3056
3057
3058 /*
3059  * compare old and new mdsmaps, kicking requests
3060  * and closing out old connections as necessary
3061  *
3062  * called under mdsc->mutex.
3063  */
3064 static void check_new_map(struct ceph_mds_client *mdsc,
3065                           struct ceph_mdsmap *newmap,
3066                           struct ceph_mdsmap *oldmap)
3067 {
3068         int i;
3069         int oldstate, newstate;
3070         struct ceph_mds_session *s;
3071
3072         dout("check_new_map new %u old %u\n",
3073              newmap->m_epoch, oldmap->m_epoch);
3074
3075         for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
3076                 if (mdsc->sessions[i] == NULL)
3077                         continue;
3078                 s = mdsc->sessions[i];
3079                 oldstate = ceph_mdsmap_get_state(oldmap, i);
3080                 newstate = ceph_mdsmap_get_state(newmap, i);
3081
3082                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3083                      i, ceph_mds_state_name(oldstate),
3084                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3085                      ceph_mds_state_name(newstate),
3086                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3087                      ceph_session_state_name(s->s_state));
3088
3089                 if (i >= newmap->m_max_mds ||
3090                     memcmp(ceph_mdsmap_get_addr(oldmap, i),
3091                            ceph_mdsmap_get_addr(newmap, i),
3092                            sizeof(struct ceph_entity_addr))) {
3093                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3094                                 /* the session never opened, just close it
3095                                  * out now */
3096                                 __wake_requests(mdsc, &s->s_waiting);
3097                                 __unregister_session(mdsc, s);
3098                         } else {
3099                                 /* just close it */
3100                                 mutex_unlock(&mdsc->mutex);
3101                                 mutex_lock(&s->s_mutex);
3102                                 mutex_lock(&mdsc->mutex);
3103                                 ceph_con_close(&s->s_con);
3104                                 mutex_unlock(&s->s_mutex);
3105                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3106                         }
3107                 } else if (oldstate == newstate) {
3108                         continue;  /* nothing new with this mds */
3109                 }
3110
3111                 /*
3112                  * send reconnect?
3113                  */
3114                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3115                     newstate >= CEPH_MDS_STATE_RECONNECT) {
3116                         mutex_unlock(&mdsc->mutex);
3117                         send_mds_reconnect(mdsc, s);
3118                         mutex_lock(&mdsc->mutex);
3119                 }
3120
3121                 /*
3122                  * kick request on any mds that has gone active.
3123                  */
3124                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3125                     newstate >= CEPH_MDS_STATE_ACTIVE) {
3126                         if (oldstate != CEPH_MDS_STATE_CREATING &&
3127                             oldstate != CEPH_MDS_STATE_STARTING)
3128                                 pr_info("mds%d recovery completed\n", s->s_mds);
3129                         kick_requests(mdsc, i);
3130                         ceph_kick_flushing_caps(mdsc, s);
3131                         wake_up_session_caps(s, 1);
3132                 }
3133         }
3134
3135         for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
3136                 s = mdsc->sessions[i];
3137                 if (!s)
3138                         continue;
3139                 if (!ceph_mdsmap_is_laggy(newmap, i))
3140                         continue;
3141                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3142                     s->s_state == CEPH_MDS_SESSION_HUNG ||
3143                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
3144                         dout(" connecting to export targets of laggy mds%d\n",
3145                              i);
3146                         __open_export_target_sessions(mdsc, s);
3147                 }
3148         }
3149 }
3150
3151
3152
3153 /*
3154  * leases
3155  */
3156
3157 /*
3158  * caller must hold session s_mutex, dentry->d_lock
3159  */
3160 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3161 {
3162         struct ceph_dentry_info *di = ceph_dentry(dentry);
3163
3164         ceph_put_mds_session(di->lease_session);
3165         di->lease_session = NULL;
3166 }
3167
3168 static void handle_lease(struct ceph_mds_client *mdsc,
3169                          struct ceph_mds_session *session,
3170                          struct ceph_msg *msg)
3171 {
3172         struct super_block *sb = mdsc->fsc->sb;
3173         struct inode *inode;
3174         struct dentry *parent, *dentry;
3175         struct ceph_dentry_info *di;
3176         int mds = session->s_mds;
3177         struct ceph_mds_lease *h = msg->front.iov_base;
3178         u32 seq;
3179         struct ceph_vino vino;
3180         struct qstr dname;
3181         int release = 0;
3182
3183         dout("handle_lease from mds%d\n", mds);
3184
3185         /* decode */
3186         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3187                 goto bad;
3188         vino.ino = le64_to_cpu(h->ino);
3189         vino.snap = CEPH_NOSNAP;
3190         seq = le32_to_cpu(h->seq);
3191         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3192         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3193         if (dname.len != get_unaligned_le32(h+1))
3194                 goto bad;
3195
3196         /* lookup inode */
3197         inode = ceph_find_inode(sb, vino);
3198         dout("handle_lease %s, ino %llx %p %.*s\n",
3199              ceph_lease_op_name(h->action), vino.ino, inode,
3200              dname.len, dname.name);
3201
3202         mutex_lock(&session->s_mutex);
3203         session->s_seq++;
3204
3205         if (inode == NULL) {
3206                 dout("handle_lease no inode %llx\n", vino.ino);
3207                 goto release;
3208         }
3209
3210         /* dentry */
3211         parent = d_find_alias(inode);
3212         if (!parent) {
3213                 dout("no parent dentry on inode %p\n", inode);
3214                 WARN_ON(1);
3215                 goto release;  /* hrm... */
3216         }
3217         dname.hash = full_name_hash(parent, dname.name, dname.len);
3218         dentry = d_lookup(parent, &dname);
3219         dput(parent);
3220         if (!dentry)
3221                 goto release;
3222
3223         spin_lock(&dentry->d_lock);
3224         di = ceph_dentry(dentry);
3225         switch (h->action) {
3226         case CEPH_MDS_LEASE_REVOKE:
3227                 if (di->lease_session == session) {
3228                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3229                                 h->seq = cpu_to_le32(di->lease_seq);
3230                         __ceph_mdsc_drop_dentry_lease(dentry);
3231                 }
3232                 release = 1;
3233                 break;
3234
3235         case CEPH_MDS_LEASE_RENEW:
3236                 if (di->lease_session == session &&
3237                     di->lease_gen == session->s_cap_gen &&
3238                     di->lease_renew_from &&
3239                     di->lease_renew_after == 0) {
3240                         unsigned long duration =
3241                                 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3242
3243                         di->lease_seq = seq;
3244                         di->time = di->lease_renew_from + duration;
3245                         di->lease_renew_after = di->lease_renew_from +
3246                                 (duration >> 1);
3247                         di->lease_renew_from = 0;
3248                 }
3249                 break;
3250         }
3251         spin_unlock(&dentry->d_lock);
3252         dput(dentry);
3253
3254         if (!release)
3255                 goto out;
3256
3257 release:
3258         /* let's just reuse the same message */
3259         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3260         ceph_msg_get(msg);
3261         ceph_con_send(&session->s_con, msg);
3262
3263 out:
3264         iput(inode);
3265         mutex_unlock(&session->s_mutex);
3266         return;
3267
3268 bad:
3269         pr_err("corrupt lease message\n");
3270         ceph_msg_dump(msg);
3271 }
3272
3273 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3274                               struct inode *inode,
3275                               struct dentry *dentry, char action,
3276                               u32 seq)
3277 {
3278         struct ceph_msg *msg;
3279         struct ceph_mds_lease *lease;
3280         int len = sizeof(*lease) + sizeof(u32);
3281         int dnamelen = 0;
3282
3283         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3284              inode, dentry, ceph_lease_op_name(action), session->s_mds);
3285         dnamelen = dentry->d_name.len;
3286         len += dnamelen;
3287
3288         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3289         if (!msg)
3290                 return;
3291         lease = msg->front.iov_base;
3292         lease->action = action;
3293         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3294         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3295         lease->seq = cpu_to_le32(seq);
3296         put_unaligned_le32(dnamelen, lease + 1);
3297         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3298
3299         /*
3300          * if this is a preemptive lease RELEASE, no need to
3301          * flush request stream, since the actual request will
3302          * soon follow.
3303          */
3304         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3305
3306         ceph_con_send(&session->s_con, msg);
3307 }
3308
3309 /*
3310  * drop all leases (and dentry refs) in preparation for umount
3311  */
3312 static void drop_leases(struct ceph_mds_client *mdsc)
3313 {
3314         int i;
3315
3316         dout("drop_leases\n");
3317         mutex_lock(&mdsc->mutex);
3318         for (i = 0; i < mdsc->max_sessions; i++) {
3319                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3320                 if (!s)
3321                         continue;
3322                 mutex_unlock(&mdsc->mutex);
3323                 mutex_lock(&s->s_mutex);
3324                 mutex_unlock(&s->s_mutex);
3325                 ceph_put_mds_session(s);
3326                 mutex_lock(&mdsc->mutex);
3327         }
3328         mutex_unlock(&mdsc->mutex);
3329 }
3330
3331
3332
3333 /*
3334  * delayed work -- periodically trim expired leases, renew caps with mds
3335  */
3336 static void schedule_delayed(struct ceph_mds_client *mdsc)
3337 {
3338         int delay = 5;
3339         unsigned hz = round_jiffies_relative(HZ * delay);
3340         schedule_delayed_work(&mdsc->delayed_work, hz);
3341 }
3342
3343 static void delayed_work(struct work_struct *work)
3344 {
3345         int i;
3346         struct ceph_mds_client *mdsc =
3347                 container_of(work, struct ceph_mds_client, delayed_work.work);
3348         int renew_interval;
3349         int renew_caps;
3350
3351         dout("mdsc delayed_work\n");
3352         ceph_check_delayed_caps(mdsc);
3353
3354         mutex_lock(&mdsc->mutex);
3355         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3356         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3357                                    mdsc->last_renew_caps);
3358         if (renew_caps)
3359                 mdsc->last_renew_caps = jiffies;
3360
3361         for (i = 0; i < mdsc->max_sessions; i++) {
3362                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3363                 if (s == NULL)
3364                         continue;
3365                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3366                         dout("resending session close request for mds%d\n",
3367                              s->s_mds);
3368                         request_close_session(mdsc, s);
3369                         ceph_put_mds_session(s);
3370                         continue;
3371                 }
3372                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3373                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3374                                 s->s_state = CEPH_MDS_SESSION_HUNG;
3375                                 pr_info("mds%d hung\n", s->s_mds);
3376                         }
3377                 }
3378                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3379                         /* this mds is failed or recovering, just wait */
3380                         ceph_put_mds_session(s);
3381                         continue;
3382                 }
3383                 mutex_unlock(&mdsc->mutex);
3384
3385                 mutex_lock(&s->s_mutex);
3386                 if (renew_caps)
3387                         send_renew_caps(mdsc, s);
3388                 else
3389                         ceph_con_keepalive(&s->s_con);
3390                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3391                     s->s_state == CEPH_MDS_SESSION_HUNG)
3392                         ceph_send_cap_releases(mdsc, s);
3393                 mutex_unlock(&s->s_mutex);
3394                 ceph_put_mds_session(s);
3395
3396                 mutex_lock(&mdsc->mutex);
3397         }
3398         mutex_unlock(&mdsc->mutex);
3399
3400         schedule_delayed(mdsc);
3401 }
3402
3403 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3404
3405 {
3406         struct ceph_mds_client *mdsc;
3407
3408         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3409         if (!mdsc)
3410                 return -ENOMEM;
3411         mdsc->fsc = fsc;
3412         fsc->mdsc = mdsc;
3413         mutex_init(&mdsc->mutex);
3414         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3415         if (mdsc->mdsmap == NULL) {
3416                 kfree(mdsc);
3417                 return -ENOMEM;
3418         }
3419
3420         init_completion(&mdsc->safe_umount_waiters);
3421         init_waitqueue_head(&mdsc->session_close_wq);
3422         INIT_LIST_HEAD(&mdsc->waiting_for_map);
3423         mdsc->sessions = NULL;
3424         atomic_set(&mdsc->num_sessions, 0);
3425         mdsc->max_sessions = 0;
3426         mdsc->stopping = 0;
3427         mdsc->last_snap_seq = 0;
3428         init_rwsem(&mdsc->snap_rwsem);
3429         mdsc->snap_realms = RB_ROOT;
3430         INIT_LIST_HEAD(&mdsc->snap_empty);
3431         spin_lock_init(&mdsc->snap_empty_lock);
3432         mdsc->last_tid = 0;
3433         mdsc->oldest_tid = 0;
3434         mdsc->request_tree = RB_ROOT;
3435         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3436         mdsc->last_renew_caps = jiffies;
3437         INIT_LIST_HEAD(&mdsc->cap_delay_list);
3438         spin_lock_init(&mdsc->cap_delay_lock);
3439         INIT_LIST_HEAD(&mdsc->snap_flush_list);
3440         spin_lock_init(&mdsc->snap_flush_lock);
3441         mdsc->last_cap_flush_tid = 1;
3442         INIT_LIST_HEAD(&mdsc->cap_flush_list);
3443         INIT_LIST_HEAD(&mdsc->cap_dirty);
3444         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3445         mdsc->num_cap_flushing = 0;
3446         spin_lock_init(&mdsc->cap_dirty_lock);
3447         init_waitqueue_head(&mdsc->cap_flushing_wq);
3448         spin_lock_init(&mdsc->dentry_lru_lock);
3449         INIT_LIST_HEAD(&mdsc->dentry_lru);
3450
3451         ceph_caps_init(mdsc);
3452         ceph_adjust_min_caps(mdsc, fsc->min_caps);
3453
3454         init_rwsem(&mdsc->pool_perm_rwsem);
3455         mdsc->pool_perm_tree = RB_ROOT;
3456
3457         return 0;
3458 }
3459
3460 /*
3461  * Wait for safe replies on open mds requests.  If we time out, drop
3462  * all requests from the tree to avoid dangling dentry refs.
3463  */
3464 static void wait_requests(struct ceph_mds_client *mdsc)
3465 {
3466         struct ceph_options *opts = mdsc->fsc->client->options;
3467         struct ceph_mds_request *req;
3468
3469         mutex_lock(&mdsc->mutex);
3470         if (__get_oldest_req(mdsc)) {
3471                 mutex_unlock(&mdsc->mutex);
3472
3473                 dout("wait_requests waiting for requests\n");
3474                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3475                                     ceph_timeout_jiffies(opts->mount_timeout));
3476
3477                 /* tear down remaining requests */
3478                 mutex_lock(&mdsc->mutex);
3479                 while ((req = __get_oldest_req(mdsc))) {
3480                         dout("wait_requests timed out on tid %llu\n",
3481                              req->r_tid);
3482                         __unregister_request(mdsc, req);
3483                 }
3484         }
3485         mutex_unlock(&mdsc->mutex);
3486         dout("wait_requests done\n");
3487 }
3488
3489 /*
3490  * called before mount is ro, and before dentries are torn down.
3491  * (hmm, does this still race with new lookups?)
3492  */
3493 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3494 {
3495         dout("pre_umount\n");
3496         mdsc->stopping = 1;
3497
3498         drop_leases(mdsc);
3499         ceph_flush_dirty_caps(mdsc);
3500         wait_requests(mdsc);
3501
3502         /*
3503          * wait for reply handlers to drop their request refs and
3504          * their inode/dcache refs
3505          */
3506         ceph_msgr_flush();
3507 }
3508
3509 /*
3510  * wait for all write mds requests to flush.
3511  */
3512 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3513 {
3514         struct ceph_mds_request *req = NULL, *nextreq;
3515         struct rb_node *n;
3516
3517         mutex_lock(&mdsc->mutex);
3518         dout("wait_unsafe_requests want %lld\n", want_tid);
3519 restart:
3520         req = __get_oldest_req(mdsc);
3521         while (req && req->r_tid <= want_tid) {
3522                 /* find next request */
3523                 n = rb_next(&req->r_node);
3524                 if (n)
3525                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3526                 else
3527                         nextreq = NULL;
3528                 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3529                     (req->r_op & CEPH_MDS_OP_WRITE)) {
3530                         /* write op */
3531                         ceph_mdsc_get_request(req);
3532                         if (nextreq)
3533                                 ceph_mdsc_get_request(nextreq);
3534                         mutex_unlock(&mdsc->mutex);
3535                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
3536                              req->r_tid, want_tid);
3537                         wait_for_completion(&req->r_safe_completion);
3538                         mutex_lock(&mdsc->mutex);
3539                         ceph_mdsc_put_request(req);
3540                         if (!nextreq)
3541                                 break;  /* next dne before, so we're done! */
3542                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
3543                                 /* next request was removed from tree */
3544                                 ceph_mdsc_put_request(nextreq);
3545                                 goto restart;
3546                         }
3547                         ceph_mdsc_put_request(nextreq);  /* won't go away */
3548                 }
3549                 req = nextreq;
3550         }
3551         mutex_unlock(&mdsc->mutex);
3552         dout("wait_unsafe_requests done\n");
3553 }
3554
3555 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3556 {
3557         u64 want_tid, want_flush;
3558
3559         if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3560                 return;
3561
3562         dout("sync\n");
3563         mutex_lock(&mdsc->mutex);
3564         want_tid = mdsc->last_tid;
3565         mutex_unlock(&mdsc->mutex);
3566
3567         ceph_flush_dirty_caps(mdsc);
3568         spin_lock(&mdsc->cap_dirty_lock);
3569         want_flush = mdsc->last_cap_flush_tid;
3570         if (!list_empty(&mdsc->cap_flush_list)) {
3571                 struct ceph_cap_flush *cf =
3572                         list_last_entry(&mdsc->cap_flush_list,
3573                                         struct ceph_cap_flush, g_list);
3574                 cf->wake = true;
3575         }
3576         spin_unlock(&mdsc->cap_dirty_lock);
3577
3578         dout("sync want tid %lld flush_seq %lld\n",
3579              want_tid, want_flush);
3580
3581         wait_unsafe_requests(mdsc, want_tid);
3582         wait_caps_flush(mdsc, want_flush);
3583 }
3584
3585 /*
3586  * true if all sessions are closed, or we force unmount
3587  */
3588 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
3589 {
3590         if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3591                 return true;
3592         return atomic_read(&mdsc->num_sessions) <= skipped;
3593 }
3594
3595 /*
3596  * called after sb is ro.
3597  */
3598 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3599 {
3600         struct ceph_options *opts = mdsc->fsc->client->options;
3601         struct ceph_mds_session *session;
3602         int i;
3603         int skipped = 0;
3604
3605         dout("close_sessions\n");
3606
3607         /* close sessions */
3608         mutex_lock(&mdsc->mutex);
3609         for (i = 0; i < mdsc->max_sessions; i++) {
3610                 session = __ceph_lookup_mds_session(mdsc, i);
3611                 if (!session)
3612                         continue;
3613                 mutex_unlock(&mdsc->mutex);
3614                 mutex_lock(&session->s_mutex);
3615                 if (__close_session(mdsc, session) <= 0)
3616                         skipped++;
3617                 mutex_unlock(&session->s_mutex);
3618                 ceph_put_mds_session(session);
3619                 mutex_lock(&mdsc->mutex);
3620         }
3621         mutex_unlock(&mdsc->mutex);
3622
3623         dout("waiting for sessions to close\n");
3624         wait_event_timeout(mdsc->session_close_wq,
3625                            done_closing_sessions(mdsc, skipped),
3626                            ceph_timeout_jiffies(opts->mount_timeout));
3627
3628         /* tear down remaining sessions */
3629         mutex_lock(&mdsc->mutex);
3630         for (i = 0; i < mdsc->max_sessions; i++) {
3631                 if (mdsc->sessions[i]) {
3632                         session = get_session(mdsc->sessions[i]);
3633                         __unregister_session(mdsc, session);
3634                         mutex_unlock(&mdsc->mutex);
3635                         mutex_lock(&session->s_mutex);
3636                         remove_session_caps(session);
3637                         mutex_unlock(&session->s_mutex);
3638                         ceph_put_mds_session(session);
3639                         mutex_lock(&mdsc->mutex);
3640                 }
3641         }
3642         WARN_ON(!list_empty(&mdsc->cap_delay_list));
3643         mutex_unlock(&mdsc->mutex);
3644
3645         ceph_cleanup_empty_realms(mdsc);
3646
3647         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3648
3649         dout("stopped\n");
3650 }
3651
3652 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
3653 {
3654         struct ceph_mds_session *session;
3655         int mds;
3656
3657         dout("force umount\n");
3658
3659         mutex_lock(&mdsc->mutex);
3660         for (mds = 0; mds < mdsc->max_sessions; mds++) {
3661                 session = __ceph_lookup_mds_session(mdsc, mds);
3662                 if (!session)
3663                         continue;
3664                 mutex_unlock(&mdsc->mutex);
3665                 mutex_lock(&session->s_mutex);
3666                 __close_session(mdsc, session);
3667                 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
3668                         cleanup_session_requests(mdsc, session);
3669                         remove_session_caps(session);
3670                 }
3671                 mutex_unlock(&session->s_mutex);
3672                 ceph_put_mds_session(session);
3673                 mutex_lock(&mdsc->mutex);
3674                 kick_requests(mdsc, mds);
3675         }
3676         __wake_requests(mdsc, &mdsc->waiting_for_map);
3677         mutex_unlock(&mdsc->mutex);
3678 }
3679
3680 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3681 {
3682         dout("stop\n");
3683         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3684         if (mdsc->mdsmap)
3685                 ceph_mdsmap_destroy(mdsc->mdsmap);
3686         kfree(mdsc->sessions);
3687         ceph_caps_finalize(mdsc);
3688         ceph_pool_perm_destroy(mdsc);
3689 }
3690
3691 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3692 {
3693         struct ceph_mds_client *mdsc = fsc->mdsc;
3694
3695         dout("mdsc_destroy %p\n", mdsc);
3696         ceph_mdsc_stop(mdsc);
3697
3698         /* flush out any connection work with references to us */
3699         ceph_msgr_flush();
3700
3701         fsc->mdsc = NULL;
3702         kfree(mdsc);
3703         dout("mdsc_destroy %p done\n", mdsc);
3704 }
3705
3706 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3707 {
3708         struct ceph_fs_client *fsc = mdsc->fsc;
3709         const char *mds_namespace = fsc->mount_options->mds_namespace;
3710         void *p = msg->front.iov_base;
3711         void *end = p + msg->front.iov_len;
3712         u32 epoch;
3713         u32 map_len;
3714         u32 num_fs;
3715         u32 mount_fscid = (u32)-1;
3716         u8 struct_v, struct_cv;
3717         int err = -EINVAL;
3718
3719         ceph_decode_need(&p, end, sizeof(u32), bad);
3720         epoch = ceph_decode_32(&p);
3721
3722         dout("handle_fsmap epoch %u\n", epoch);
3723
3724         ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3725         struct_v = ceph_decode_8(&p);
3726         struct_cv = ceph_decode_8(&p);
3727         map_len = ceph_decode_32(&p);
3728
3729         ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
3730         p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
3731
3732         num_fs = ceph_decode_32(&p);
3733         while (num_fs-- > 0) {
3734                 void *info_p, *info_end;
3735                 u32 info_len;
3736                 u8 info_v, info_cv;
3737                 u32 fscid, namelen;
3738
3739                 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3740                 info_v = ceph_decode_8(&p);
3741                 info_cv = ceph_decode_8(&p);
3742                 info_len = ceph_decode_32(&p);
3743                 ceph_decode_need(&p, end, info_len, bad);
3744                 info_p = p;
3745                 info_end = p + info_len;
3746                 p = info_end;
3747
3748                 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
3749                 fscid = ceph_decode_32(&info_p);
3750                 namelen = ceph_decode_32(&info_p);
3751                 ceph_decode_need(&info_p, info_end, namelen, bad);
3752
3753                 if (mds_namespace &&
3754                     strlen(mds_namespace) == namelen &&
3755                     !strncmp(mds_namespace, (char *)info_p, namelen)) {
3756                         mount_fscid = fscid;
3757                         break;
3758                 }
3759         }
3760
3761         ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
3762         if (mount_fscid != (u32)-1) {
3763                 fsc->client->monc.fs_cluster_id = mount_fscid;
3764                 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
3765                                    0, true);
3766                 ceph_monc_renew_subs(&fsc->client->monc);
3767         } else {
3768                 err = -ENOENT;
3769                 goto err_out;
3770         }
3771         return;
3772 bad:
3773         pr_err("error decoding fsmap\n");
3774 err_out:
3775         mutex_lock(&mdsc->mutex);
3776         mdsc->mdsmap_err = -ENOENT;
3777         __wake_requests(mdsc, &mdsc->waiting_for_map);
3778         mutex_unlock(&mdsc->mutex);
3779         return;
3780 }
3781
3782 /*
3783  * handle mds map update.
3784  */
3785 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3786 {
3787         u32 epoch;
3788         u32 maplen;
3789         void *p = msg->front.iov_base;
3790         void *end = p + msg->front.iov_len;
3791         struct ceph_mdsmap *newmap, *oldmap;
3792         struct ceph_fsid fsid;
3793         int err = -EINVAL;
3794
3795         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3796         ceph_decode_copy(&p, &fsid, sizeof(fsid));
3797         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3798                 return;
3799         epoch = ceph_decode_32(&p);
3800         maplen = ceph_decode_32(&p);
3801         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3802
3803         /* do we need it? */
3804         mutex_lock(&mdsc->mutex);
3805         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3806                 dout("handle_map epoch %u <= our %u\n",
3807                      epoch, mdsc->mdsmap->m_epoch);
3808                 mutex_unlock(&mdsc->mutex);
3809                 return;
3810         }
3811
3812         newmap = ceph_mdsmap_decode(&p, end);
3813         if (IS_ERR(newmap)) {
3814                 err = PTR_ERR(newmap);
3815                 goto bad_unlock;
3816         }
3817
3818         /* swap into place */
3819         if (mdsc->mdsmap) {
3820                 oldmap = mdsc->mdsmap;
3821                 mdsc->mdsmap = newmap;
3822                 check_new_map(mdsc, newmap, oldmap);
3823                 ceph_mdsmap_destroy(oldmap);
3824         } else {
3825                 mdsc->mdsmap = newmap;  /* first mds map */
3826         }
3827         mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3828
3829         __wake_requests(mdsc, &mdsc->waiting_for_map);
3830         ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
3831                           mdsc->mdsmap->m_epoch);
3832
3833         mutex_unlock(&mdsc->mutex);
3834         schedule_delayed(mdsc);
3835         return;
3836
3837 bad_unlock:
3838         mutex_unlock(&mdsc->mutex);
3839 bad:
3840         pr_err("error decoding mdsmap %d\n", err);
3841         return;
3842 }
3843
3844 static struct ceph_connection *con_get(struct ceph_connection *con)
3845 {
3846         struct ceph_mds_session *s = con->private;
3847
3848         if (get_session(s)) {
3849                 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3850                 return con;
3851         }
3852         dout("mdsc con_get %p FAIL\n", s);
3853         return NULL;
3854 }
3855
3856 static void con_put(struct ceph_connection *con)
3857 {
3858         struct ceph_mds_session *s = con->private;
3859
3860         dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3861         ceph_put_mds_session(s);
3862 }
3863
3864 /*
3865  * if the client is unresponsive for long enough, the mds will kill
3866  * the session entirely.
3867  */
3868 static void peer_reset(struct ceph_connection *con)
3869 {
3870         struct ceph_mds_session *s = con->private;
3871         struct ceph_mds_client *mdsc = s->s_mdsc;
3872
3873         pr_warn("mds%d closed our session\n", s->s_mds);
3874         send_mds_reconnect(mdsc, s);
3875 }
3876
3877 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3878 {
3879         struct ceph_mds_session *s = con->private;
3880         struct ceph_mds_client *mdsc = s->s_mdsc;
3881         int type = le16_to_cpu(msg->hdr.type);
3882
3883         mutex_lock(&mdsc->mutex);
3884         if (__verify_registered_session(mdsc, s) < 0) {
3885                 mutex_unlock(&mdsc->mutex);
3886                 goto out;
3887         }
3888         mutex_unlock(&mdsc->mutex);
3889
3890         switch (type) {
3891         case CEPH_MSG_MDS_MAP:
3892                 ceph_mdsc_handle_mdsmap(mdsc, msg);
3893                 break;
3894         case CEPH_MSG_FS_MAP_USER:
3895                 ceph_mdsc_handle_fsmap(mdsc, msg);
3896                 break;
3897         case CEPH_MSG_CLIENT_SESSION:
3898                 handle_session(s, msg);
3899                 break;
3900         case CEPH_MSG_CLIENT_REPLY:
3901                 handle_reply(s, msg);
3902                 break;
3903         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3904                 handle_forward(mdsc, s, msg);
3905                 break;
3906         case CEPH_MSG_CLIENT_CAPS:
3907                 ceph_handle_caps(s, msg);
3908                 break;
3909         case CEPH_MSG_CLIENT_SNAP:
3910                 ceph_handle_snap(mdsc, s, msg);
3911                 break;
3912         case CEPH_MSG_CLIENT_LEASE:
3913                 handle_lease(mdsc, s, msg);
3914                 break;
3915
3916         default:
3917                 pr_err("received unknown message type %d %s\n", type,
3918                        ceph_msg_type_name(type));
3919         }
3920 out:
3921         ceph_msg_put(msg);
3922 }
3923
3924 /*
3925  * authentication
3926  */
3927
3928 /*
3929  * Note: returned pointer is the address of a structure that's
3930  * managed separately.  Caller must *not* attempt to free it.
3931  */
3932 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
3933                                         int *proto, int force_new)
3934 {
3935         struct ceph_mds_session *s = con->private;
3936         struct ceph_mds_client *mdsc = s->s_mdsc;
3937         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3938         struct ceph_auth_handshake *auth = &s->s_auth;
3939
3940         if (force_new && auth->authorizer) {
3941                 ceph_auth_destroy_authorizer(auth->authorizer);
3942                 auth->authorizer = NULL;
3943         }
3944         if (!auth->authorizer) {
3945                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3946                                                       auth);
3947                 if (ret)
3948                         return ERR_PTR(ret);
3949         } else {
3950                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3951                                                       auth);
3952                 if (ret)
3953                         return ERR_PTR(ret);
3954         }
3955         *proto = ac->protocol;
3956
3957         return auth;
3958 }
3959
3960
3961 static int verify_authorizer_reply(struct ceph_connection *con)
3962 {
3963         struct ceph_mds_session *s = con->private;
3964         struct ceph_mds_client *mdsc = s->s_mdsc;
3965         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3966
3967         return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
3968 }
3969
3970 static int invalidate_authorizer(struct ceph_connection *con)
3971 {
3972         struct ceph_mds_session *s = con->private;
3973         struct ceph_mds_client *mdsc = s->s_mdsc;
3974         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3975
3976         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3977
3978         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
3979 }
3980
3981 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
3982                                 struct ceph_msg_header *hdr, int *skip)
3983 {
3984         struct ceph_msg *msg;
3985         int type = (int) le16_to_cpu(hdr->type);
3986         int front_len = (int) le32_to_cpu(hdr->front_len);
3987
3988         if (con->in_msg)
3989                 return con->in_msg;
3990
3991         *skip = 0;
3992         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
3993         if (!msg) {
3994                 pr_err("unable to allocate msg type %d len %d\n",
3995                        type, front_len);
3996                 return NULL;
3997         }
3998
3999         return msg;
4000 }
4001
4002 static int mds_sign_message(struct ceph_msg *msg)
4003 {
4004        struct ceph_mds_session *s = msg->con->private;
4005        struct ceph_auth_handshake *auth = &s->s_auth;
4006
4007        return ceph_auth_sign_message(auth, msg);
4008 }
4009
4010 static int mds_check_message_signature(struct ceph_msg *msg)
4011 {
4012        struct ceph_mds_session *s = msg->con->private;
4013        struct ceph_auth_handshake *auth = &s->s_auth;
4014
4015        return ceph_auth_check_message_signature(auth, msg);
4016 }
4017
4018 static const struct ceph_connection_operations mds_con_ops = {
4019         .get = con_get,
4020         .put = con_put,
4021         .dispatch = dispatch,
4022         .get_authorizer = get_authorizer,
4023         .verify_authorizer_reply = verify_authorizer_reply,
4024         .invalidate_authorizer = invalidate_authorizer,
4025         .peer_reset = peer_reset,
4026         .alloc_msg = mds_alloc_msg,
4027         .sign_message = mds_sign_message,
4028         .check_message_signature = mds_check_message_signature,
4029 };
4030
4031 /* eof */