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  45  * are no longer valid.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         /* Never leave an unregistered request on an unsafe list! */
 632         list_del_init(&req->r_unsafe_item);
 633
 634         if (req->r_tid == mdsc->oldest_tid) {
 635                 struct rb_node *p = rb_next(&req->r_node);
 636                 mdsc->oldest_tid = 0;
 637                 while (p) {
 638                         struct ceph_mds_request *next_req =
 639                                 rb_entry(p, struct ceph_mds_request, r_node);
 640                         if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
 641                                 mdsc->oldest_tid = next_req->r_tid;
 642                                 break;
 643                         }
 644                         p = rb_next(p);
 645                 }
 646         }
 647
 648         erase_request(&mdsc->request_tree, req);
 649
 650         if (req->r_unsafe_dir && req->r_got_unsafe) {
 651                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 652                 spin_lock(&ci->i_unsafe_lock);
 653                 list_del_init(&req->r_unsafe_dir_item);
 654                 spin_unlock(&ci->i_unsafe_lock);
 655         }
 656         if (req->r_target_inode && req->r_got_unsafe) {
 657                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
 658                 spin_lock(&ci->i_unsafe_lock);
 659                 list_del_init(&req->r_unsafe_target_item);
 660                 spin_unlock(&ci->i_unsafe_lock);
 661         }
 662
 663         if (req->r_unsafe_dir) {
 664                 iput(req->r_unsafe_dir);
 665                 req->r_unsafe_dir = NULL;
 666         }
 667
 668         complete_all(&req->r_safe_completion);
 669
 670         ceph_mdsc_put_request(req);
 671 }
 672
 673 /*
 674  * Choose mds to send request to next.  If there is a hint set in the
 675  * request (e.g., due to a prior forward hint from the mds), use that.
 676  * Otherwise, consult frag tree and/or caps to identify the
 677  * appropriate mds.  If all else fails, choose randomly.
 678  *
 679  * Called under mdsc->mutex.
 680  */
 681 static struct dentry *get_nonsnap_parent(struct dentry *dentry)
 682 {
 683         /*
 684          * we don't need to worry about protecting the d_parent access
 685          * here because we never renaming inside the snapped namespace
 686          * except to resplice to another snapdir, and either the old or new
 687          * result is a valid result.
 688          */
 689         while (!IS_ROOT(dentry) && ceph_snap(d_inode(dentry)) != CEPH_NOSNAP)
 690                 dentry = dentry->d_parent;
 691         return dentry;
 692 }
 693
 694 static int __choose_mds(struct ceph_mds_client *mdsc,
 695                         struct ceph_mds_request *req)
 696 {
 697         struct inode *inode;
 698         struct ceph_inode_info *ci;
 699         struct ceph_cap *cap;
 700         int mode = req->r_direct_mode;
 701         int mds = -1;
 702         u32 hash = req->r_direct_hash;
 703         bool is_hash = req->r_direct_is_hash;
 704
 705         /*
 706          * is there a specific mds we should try?  ignore hint if we have
 707          * no session and the mds is not up (active or recovering).
 708          */
 709         if (req->r_resend_mds >= 0 &&
 710             (__have_session(mdsc, req->r_resend_mds) ||
 711              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 712                 dout("choose_mds using resend_mds mds%d\n",
 713                      req->r_resend_mds);
 714                 return req->r_resend_mds;
 715         }
 716
 717         if (mode == USE_RANDOM_MDS)
 718                 goto random;
 719
 720         inode = NULL;
 721         if (req->r_inode) {
 722                 inode = req->r_inode;
 723         } else if (req->r_dentry) {
 724                 /* ignore race with rename; old or new d_parent is okay */
 725                 struct dentry *parent = req->r_dentry->d_parent;
 726                 struct inode *dir = d_inode(parent);
 727
 728                 if (dir->i_sb != mdsc->fsc->sb) {
 729                         /* not this fs! */
 730                         inode = d_inode(req->r_dentry);
 731                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
 732                         /* direct snapped/virtual snapdir requests
 733                          * based on parent dir inode */
 734                         struct dentry *dn = get_nonsnap_parent(parent);
 735                         inode = d_inode(dn);
 736                         dout("__choose_mds using nonsnap parent %p\n", inode);
 737                 } else {
 738                         /* dentry target */
 739                         inode = d_inode(req->r_dentry);
 740                         if (!inode || mode == USE_AUTH_MDS) {
 741                                 /* dir + name */
 742                                 inode = dir;
 743                                 hash = ceph_dentry_hash(dir, req->r_dentry);
 744                                 is_hash = true;
 745                         }
 746                 }
 747         }
 748
 749         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 750              (int)hash, mode);
 751         if (!inode)
 752                 goto random;
 753         ci = ceph_inode(inode);
 754
 755         if (is_hash && S_ISDIR(inode->i_mode)) {
 756                 struct ceph_inode_frag frag;
 757                 int found;
 758
 759                 ceph_choose_frag(ci, hash, &frag, &found);
 760                 if (found) {
 761                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 762                                 u8 r;
 763
 764                                 /* choose a random replica */
 765                                 get_random_bytes(&r, 1);
 766                                 r %= frag.ndist;
 767                                 mds = frag.dist[r];
 768                                 dout("choose_mds %p %llx.%llx "
 769                                      "frag %u mds%d (%d/%d)\n",
 770                                      inode, ceph_vinop(inode),
 771                                      frag.frag, mds,
 772                                      (int)r, frag.ndist);
 773                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 774                                     CEPH_MDS_STATE_ACTIVE)
 775                                         return mds;
 776                         }
 777
 778                         /* since this file/dir wasn't known to be
 779                          * replicated, then we want to look for the
 780                          * authoritative mds. */
 781                         mode = USE_AUTH_MDS;
 782                         if (frag.mds >= 0) {
 783                                 /* choose auth mds */
 784                                 mds = frag.mds;
 785                                 dout("choose_mds %p %llx.%llx "
 786                                      "frag %u mds%d (auth)\n",
 787                                      inode, ceph_vinop(inode), frag.frag, mds);
 788                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 789                                     CEPH_MDS_STATE_ACTIVE)
 790                                         return mds;
 791                         }
 792                 }
 793         }
 794
 795         spin_lock(&ci->i_ceph_lock);
 796         cap = NULL;
 797         if (mode == USE_AUTH_MDS)
 798                 cap = ci->i_auth_cap;
 799         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
 800                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
 801         if (!cap) {
 802                 spin_unlock(&ci->i_ceph_lock);
 803                 goto random;
 804         }
 805         mds = cap->session->s_mds;
 806         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
 807              inode, ceph_vinop(inode), mds,
 808              cap == ci->i_auth_cap ? "auth " : "", cap);
 809         spin_unlock(&ci->i_ceph_lock);
 810         return mds;
 811
 812 random:
 813         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
 814         dout("choose_mds chose random mds%d\n", mds);
 815         return mds;
 816 }
 817
 818
 819 /*
 820  * session messages
 821  */
 822 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
 823 {
 824         struct ceph_msg *msg;
 825         struct ceph_mds_session_head *h;
 826
 827         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
 828                            false);
 829         if (!msg) {
 830                 pr_err("create_session_msg ENOMEM creating msg\n");
 831                 return NULL;
 832         }
 833         h = msg->front.iov_base;
 834         h->op = cpu_to_le32(op);
 835         h->seq = cpu_to_le64(seq);
 836
 837         return msg;
 838 }
 839
 840 /*
 841  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
 842  * to include additional client metadata fields.
 843  */
 844 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
 845 {
 846         struct ceph_msg *msg;
 847         struct ceph_mds_session_head *h;
 848         int i = -1;
 849         int metadata_bytes = 0;
 850         int metadata_key_count = 0;
 851         struct ceph_options *opt = mdsc->fsc->client->options;
 852         struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
 853         void *p;
 854
 855         const char* metadata[][2] = {
 856                 {"hostname", utsname()->nodename},
 857                 {"kernel_version", utsname()->release},
 858                 {"entity_id", opt->name ? : ""},
 859                 {"root", fsopt->server_path ? : "/"},
 860                 {NULL, NULL}
 861         };
 862
 863         /* Calculate serialized length of metadata */
 864         metadata_bytes = 4;  /* map length */
 865         for (i = 0; metadata[i][0] != NULL; ++i) {
 866                 metadata_bytes += 8 + strlen(metadata[i][0]) +
 867                         strlen(metadata[i][1]);
 868                 metadata_key_count++;
 869         }
 870
 871         /* Allocate the message */
 872         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + metadata_bytes,
 873                            GFP_NOFS, false);
 874         if (!msg) {
 875                 pr_err("create_session_msg ENOMEM creating msg\n");
 876                 return NULL;
 877         }
 878         h = msg->front.iov_base;
 879         h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
 880         h->seq = cpu_to_le64(seq);
 881
 882         /*
 883          * Serialize client metadata into waiting buffer space, using
 884          * the format that userspace expects for map<string, string>
 885          *
 886          * ClientSession messages with metadata are v2
 887          */
 888         msg->hdr.version = cpu_to_le16(2);
 889         msg->hdr.compat_version = cpu_to_le16(1);
 890
 891         /* The write pointer, following the session_head structure */
 892         p = msg->front.iov_base + sizeof(*h);
 893
 894         /* Number of entries in the map */
 895         ceph_encode_32(&p, metadata_key_count);
 896
 897         /* Two length-prefixed strings for each entry in the map */
 898         for (i = 0; metadata[i][0] != NULL; ++i) {
 899                 size_t const key_len = strlen(metadata[i][0]);
 900                 size_t const val_len = strlen(metadata[i][1]);
 901
 902                 ceph_encode_32(&p, key_len);
 903                 memcpy(p, metadata[i][0], key_len);
 904                 p += key_len;
 905                 ceph_encode_32(&p, val_len);
 906                 memcpy(p, metadata[i][1], val_len);
 907                 p += val_len;
 908         }
 909
 910         return msg;
 911 }
 912
 913 /*
 914  * send session open request.
 915  *
 916  * called under mdsc->mutex
 917  */
 918 static int __open_session(struct ceph_mds_client *mdsc,
 919                           struct ceph_mds_session *session)
 920 {
 921         struct ceph_msg *msg;
 922         int mstate;
 923         int mds = session->s_mds;
 924
 925         /* wait for mds to go active? */
 926         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
 927         dout("open_session to mds%d (%s)\n", mds,
 928              ceph_mds_state_name(mstate));
 929         session->s_state = CEPH_MDS_SESSION_OPENING;
 930         session->s_renew_requested = jiffies;
 931
 932         /* send connect message */
 933         msg = create_session_open_msg(mdsc, session->s_seq);
 934         if (!msg)
 935                 return -ENOMEM;
 936         ceph_con_send(&session->s_con, msg);
 937         return 0;
 938 }
 939
 940 /*
 941  * open sessions for any export targets for the given mds
 942  *
 943  * called under mdsc->mutex
 944  */
 945 static struct ceph_mds_session *
 946 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
 947 {
 948         struct ceph_mds_session *session;
 949
 950         session = __ceph_lookup_mds_session(mdsc, target);
 951         if (!session) {
 952                 session = register_session(mdsc, target);
 953                 if (IS_ERR(session))
 954                         return session;
 955         }
 956         if (session->s_state == CEPH_MDS_SESSION_NEW ||
 957             session->s_state == CEPH_MDS_SESSION_CLOSING)
 958                 __open_session(mdsc, session);
 959
 960         return session;
 961 }
 962
 963 struct ceph_mds_session *
 964 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
 965 {
 966         struct ceph_mds_session *session;
 967
 968         dout("open_export_target_session to mds%d\n", target);
 969
 970         mutex_lock(&mdsc->mutex);
 971         session = __open_export_target_session(mdsc, target);
 972         mutex_unlock(&mdsc->mutex);
 973
 974         return session;
 975 }
 976
 977 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
 978                                           struct ceph_mds_session *session)
 979 {
 980         struct ceph_mds_info *mi;
 981         struct ceph_mds_session *ts;
 982         int i, mds = session->s_mds;
 983
 984         if (mds >= mdsc->mdsmap->m_max_mds)
 985                 return;
 986
 987         mi = &mdsc->mdsmap->m_info[mds];
 988         dout("open_export_target_sessions for mds%d (%d targets)\n",
 989              session->s_mds, mi->num_export_targets);
 990
 991         for (i = 0; i < mi->num_export_targets; i++) {
 992                 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
 993                 if (!IS_ERR(ts))
 994                         ceph_put_mds_session(ts);
 995         }
 996 }
 997
 998 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
 999                                            struct ceph_mds_session *session)
1000 {
1001         mutex_lock(&mdsc->mutex);
1002         __open_export_target_sessions(mdsc, session);
1003         mutex_unlock(&mdsc->mutex);
1004 }
1005
1006 /*
1007  * session caps
1008  */
1009
1010 /* caller holds s_cap_lock, we drop it */
1011 static void cleanup_cap_releases(struct ceph_mds_client *mdsc,
1012                                  struct ceph_mds_session *session)
1013         __releases(session->s_cap_lock)
1014 {
1015         LIST_HEAD(tmp_list);
1016         list_splice_init(&session->s_cap_releases, &tmp_list);
1017         session->s_num_cap_releases = 0;
1018         spin_unlock(&session->s_cap_lock);
1019
1020         dout("cleanup_cap_releases mds%d\n", session->s_mds);
1021         while (!list_empty(&tmp_list)) {
1022                 struct ceph_cap *cap;
1023                 /* zero out the in-progress message */
1024                 cap = list_first_entry(&tmp_list,
1025                                         struct ceph_cap, session_caps);
1026                 list_del(&cap->session_caps);
1027                 ceph_put_cap(mdsc, cap);
1028         }
1029 }
1030
1031 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1032                                      struct ceph_mds_session *session)
1033 {
1034         struct ceph_mds_request *req;
1035         struct rb_node *p;
1036
1037         dout("cleanup_session_requests mds%d\n", session->s_mds);
1038         mutex_lock(&mdsc->mutex);
1039         while (!list_empty(&session->s_unsafe)) {
1040                 req = list_first_entry(&session->s_unsafe,
1041                                        struct ceph_mds_request, r_unsafe_item);
1042                 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1043                                     req->r_tid);
1044                 __unregister_request(mdsc, req);
1045         }
1046         /* zero r_attempts, so kick_requests() will re-send requests */
1047         p = rb_first(&mdsc->request_tree);
1048         while (p) {
1049                 req = rb_entry(p, struct ceph_mds_request, r_node);
1050                 p = rb_next(p);
1051                 if (req->r_session &&
1052                     req->r_session->s_mds == session->s_mds)
1053                         req->r_attempts = 0;
1054         }
1055         mutex_unlock(&mdsc->mutex);
1056 }
1057
1058 /*
1059  * Helper to safely iterate over all caps associated with a session, with
1060  * special care taken to handle a racing __ceph_remove_cap().
1061  *
1062  * Caller must hold session s_mutex.
1063  */
1064 static int iterate_session_caps(struct ceph_mds_session *session,
1065                                  int (*cb)(struct inode *, struct ceph_cap *,
1066                                             void *), void *arg)
1067 {
1068         struct list_head *p;
1069         struct ceph_cap *cap;
1070         struct inode *inode, *last_inode = NULL;
1071         struct ceph_cap *old_cap = NULL;
1072         int ret;
1073
1074         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1075         spin_lock(&session->s_cap_lock);
1076         p = session->s_caps.next;
1077         while (p != &session->s_caps) {
1078                 cap = list_entry(p, struct ceph_cap, session_caps);
1079                 inode = igrab(&cap->ci->vfs_inode);
1080                 if (!inode) {
1081                         p = p->next;
1082                         continue;
1083                 }
1084                 session->s_cap_iterator = cap;
1085                 spin_unlock(&session->s_cap_lock);
1086
1087                 if (last_inode) {
1088                         iput(last_inode);
1089                         last_inode = NULL;
1090                 }
1091                 if (old_cap) {
1092                         ceph_put_cap(session->s_mdsc, old_cap);
1093                         old_cap = NULL;
1094                 }
1095
1096                 ret = cb(inode, cap, arg);
1097                 last_inode = inode;
1098
1099                 spin_lock(&session->s_cap_lock);
1100                 p = p->next;
1101                 if (cap->ci == NULL) {
1102                         dout("iterate_session_caps  finishing cap %p removal\n",
1103                              cap);
1104                         BUG_ON(cap->session != session);
1105                         cap->session = NULL;
1106                         list_del_init(&cap->session_caps);
1107                         session->s_nr_caps--;
1108                         if (cap->queue_release) {
1109                                 list_add_tail(&cap->session_caps,
1110                                               &session->s_cap_releases);
1111                                 session->s_num_cap_releases++;
1112                         } else {
1113                                 old_cap = cap;  /* put_cap it w/o locks held */
1114                         }
1115                 }
1116                 if (ret < 0)
1117                         goto out;
1118         }
1119         ret = 0;
1120 out:
1121         session->s_cap_iterator = NULL;
1122         spin_unlock(&session->s_cap_lock);
1123
1124         iput(last_inode);
1125         if (old_cap)
1126                 ceph_put_cap(session->s_mdsc, old_cap);
1127
1128         return ret;
1129 }
1130
1131 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1132                                   void *arg)
1133 {
1134         struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1135         struct ceph_inode_info *ci = ceph_inode(inode);
1136         LIST_HEAD(to_remove);
1137         bool drop = false;
1138         bool invalidate = false;
1139
1140         dout("removing cap %p, ci is %p, inode is %p\n",
1141              cap, ci, &ci->vfs_inode);
1142         spin_lock(&ci->i_ceph_lock);
1143         __ceph_remove_cap(cap, false);
1144         if (!ci->i_auth_cap) {
1145                 struct ceph_cap_flush *cf;
1146                 struct ceph_mds_client *mdsc = fsc->mdsc;
1147
1148                 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1149
1150                 if (ci->i_wrbuffer_ref > 0 &&
1151                     ACCESS_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
1152                         invalidate = true;
1153
1154                 while (!list_empty(&ci->i_cap_flush_list)) {
1155                         cf = list_first_entry(&ci->i_cap_flush_list,
1156                                               struct ceph_cap_flush, i_list);
1157                         list_move(&cf->i_list, &to_remove);
1158                 }
1159
1160                 spin_lock(&mdsc->cap_dirty_lock);
1161
1162                 list_for_each_entry(cf, &to_remove, i_list)
1163                         list_del(&cf->g_list);
1164
1165                 if (!list_empty(&ci->i_dirty_item)) {
1166                         pr_warn_ratelimited(
1167                                 " dropping dirty %s state for %p %lld\n",
1168                                 ceph_cap_string(ci->i_dirty_caps),
1169                                 inode, ceph_ino(inode));
1170                         ci->i_dirty_caps = 0;
1171                         list_del_init(&ci->i_dirty_item);
1172                         drop = true;
1173                 }
1174                 if (!list_empty(&ci->i_flushing_item)) {
1175                         pr_warn_ratelimited(
1176                                 " dropping dirty+flushing %s state for %p %lld\n",
1177                                 ceph_cap_string(ci->i_flushing_caps),
1178                                 inode, ceph_ino(inode));
1179                         ci->i_flushing_caps = 0;
1180                         list_del_init(&ci->i_flushing_item);
1181                         mdsc->num_cap_flushing--;
1182                         drop = true;
1183                 }
1184                 spin_unlock(&mdsc->cap_dirty_lock);
1185
1186                 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1187                         list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1188                         ci->i_prealloc_cap_flush = NULL;
1189                 }
1190
1191                if (drop &&
1192                   ci->i_wrbuffer_ref_head == 0 &&
1193                   ci->i_wr_ref == 0 &&
1194                   ci->i_dirty_caps == 0 &&
1195                   ci->i_flushing_caps == 0) {
1196                       ceph_put_snap_context(ci->i_head_snapc);
1197                       ci->i_head_snapc = NULL;
1198                }
1199         }
1200         spin_unlock(&ci->i_ceph_lock);
1201         while (!list_empty(&to_remove)) {
1202                 struct ceph_cap_flush *cf;
1203                 cf = list_first_entry(&to_remove,
1204                                       struct ceph_cap_flush, i_list);
1205                 list_del(&cf->i_list);
1206                 ceph_free_cap_flush(cf);
1207         }
1208
1209         wake_up_all(&ci->i_cap_wq);
1210         if (invalidate)
1211                 ceph_queue_invalidate(inode);
1212         if (drop)
1213                 iput(inode);
1214         return 0;
1215 }
1216
1217 /*
1218  * caller must hold session s_mutex
1219  */
1220 static void remove_session_caps(struct ceph_mds_session *session)
1221 {
1222         struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1223         struct super_block *sb = fsc->sb;
1224         dout("remove_session_caps on %p\n", session);
1225         iterate_session_caps(session, remove_session_caps_cb, fsc);
1226
1227         wake_up_all(&fsc->mdsc->cap_flushing_wq);
1228
1229         spin_lock(&session->s_cap_lock);
1230         if (session->s_nr_caps > 0) {
1231                 struct inode *inode;
1232                 struct ceph_cap *cap, *prev = NULL;
1233                 struct ceph_vino vino;
1234                 /*
1235                  * iterate_session_caps() skips inodes that are being
1236                  * deleted, we need to wait until deletions are complete.
1237                  * __wait_on_freeing_inode() is designed for the job,
1238                  * but it is not exported, so use lookup inode function
1239                  * to access it.
1240                  */
1241                 while (!list_empty(&session->s_caps)) {
1242                         cap = list_entry(session->s_caps.next,
1243                                          struct ceph_cap, session_caps);
1244                         if (cap == prev)
1245                                 break;
1246                         prev = cap;
1247                         vino = cap->ci->i_vino;
1248                         spin_unlock(&session->s_cap_lock);
1249
1250                         inode = ceph_find_inode(sb, vino);
1251                         iput(inode);
1252
1253                         spin_lock(&session->s_cap_lock);
1254                 }
1255         }
1256
1257         // drop cap expires and unlock s_cap_lock
1258         cleanup_cap_releases(session->s_mdsc, session);
1259
1260         BUG_ON(session->s_nr_caps > 0);
1261         BUG_ON(!list_empty(&session->s_cap_flushing));
1262 }
1263
1264 /*
1265  * wake up any threads waiting on this session's caps.  if the cap is
1266  * old (didn't get renewed on the client reconnect), remove it now.
1267  *
1268  * caller must hold s_mutex.
1269  */
1270 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1271                               void *arg)
1272 {
1273         struct ceph_inode_info *ci = ceph_inode(inode);
1274
1275         if (arg) {
1276                 spin_lock(&ci->i_ceph_lock);
1277                 ci->i_wanted_max_size = 0;
1278                 ci->i_requested_max_size = 0;
1279                 spin_unlock(&ci->i_ceph_lock);
1280         }
1281         wake_up_all(&ci->i_cap_wq);
1282         return 0;
1283 }
1284
1285 static void wake_up_session_caps(struct ceph_mds_session *session,
1286                                  int reconnect)
1287 {
1288         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1289         iterate_session_caps(session, wake_up_session_cb,
1290                              (void *)(unsigned long)reconnect);
1291 }
1292
1293 /*
1294  * Send periodic message to MDS renewing all currently held caps.  The
1295  * ack will reset the expiration for all caps from this session.
1296  *
1297  * caller holds s_mutex
1298  */
1299 static int send_renew_caps(struct ceph_mds_client *mdsc,
1300                            struct ceph_mds_session *session)
1301 {
1302         struct ceph_msg *msg;
1303         int state;
1304
1305         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1306             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1307                 pr_info("mds%d caps stale\n", session->s_mds);
1308         session->s_renew_requested = jiffies;
1309
1310         /* do not try to renew caps until a recovering mds has reconnected
1311          * with its clients. */
1312         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1313         if (state < CEPH_MDS_STATE_RECONNECT) {
1314                 dout("send_renew_caps ignoring mds%d (%s)\n",
1315                      session->s_mds, ceph_mds_state_name(state));
1316                 return 0;
1317         }
1318
1319         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1320                 ceph_mds_state_name(state));
1321         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1322                                  ++session->s_renew_seq);
1323         if (!msg)
1324                 return -ENOMEM;
1325         ceph_con_send(&session->s_con, msg);
1326         return 0;
1327 }
1328
1329 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1330                              struct ceph_mds_session *session, u64 seq)
1331 {
1332         struct ceph_msg *msg;
1333
1334         dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1335              session->s_mds, ceph_session_state_name(session->s_state), seq);
1336         msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1337         if (!msg)
1338                 return -ENOMEM;
1339         ceph_con_send(&session->s_con, msg);
1340         return 0;
1341 }
1342
1343
1344 /*
1345  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1346  *
1347  * Called under session->s_mutex
1348  */
1349 static void renewed_caps(struct ceph_mds_client *mdsc,
1350                          struct ceph_mds_session *session, int is_renew)
1351 {
1352         int was_stale;
1353         int wake = 0;
1354
1355         spin_lock(&session->s_cap_lock);
1356         was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1357
1358         session->s_cap_ttl = session->s_renew_requested +
1359                 mdsc->mdsmap->m_session_timeout*HZ;
1360
1361         if (was_stale) {
1362                 if (time_before(jiffies, session->s_cap_ttl)) {
1363                         pr_info("mds%d caps renewed\n", session->s_mds);
1364                         wake = 1;
1365                 } else {
1366                         pr_info("mds%d caps still stale\n", session->s_mds);
1367                 }
1368         }
1369         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1370              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1371              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1372         spin_unlock(&session->s_cap_lock);
1373
1374         if (wake)
1375                 wake_up_session_caps(session, 0);
1376 }
1377
1378 /*
1379  * send a session close request
1380  */
1381 static int request_close_session(struct ceph_mds_client *mdsc,
1382                                  struct ceph_mds_session *session)
1383 {
1384         struct ceph_msg *msg;
1385
1386         dout("request_close_session mds%d state %s seq %lld\n",
1387              session->s_mds, ceph_session_state_name(session->s_state),
1388              session->s_seq);
1389         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1390         if (!msg)
1391                 return -ENOMEM;
1392         ceph_con_send(&session->s_con, msg);
1393         return 1;
1394 }
1395
1396 /*
1397  * Called with s_mutex held.
1398  */
1399 static int __close_session(struct ceph_mds_client *mdsc,
1400                          struct ceph_mds_session *session)
1401 {
1402         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1403                 return 0;
1404         session->s_state = CEPH_MDS_SESSION_CLOSING;
1405         return request_close_session(mdsc, session);
1406 }
1407
1408 static bool drop_negative_children(struct dentry *dentry)
1409 {
1410         struct dentry *child;
1411         bool all_negative = true;
1412
1413         if (!d_is_dir(dentry))
1414                 goto out;
1415
1416         spin_lock(&dentry->d_lock);
1417         list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1418                 if (d_really_is_positive(child)) {
1419                         all_negative = false;
1420                         break;
1421                 }
1422         }
1423         spin_unlock(&dentry->d_lock);
1424
1425         if (all_negative)
1426                 shrink_dcache_parent(dentry);
1427 out:
1428         return all_negative;
1429 }
1430
1431 /*
1432  * Trim old(er) caps.
1433  *
1434  * Because we can't cache an inode without one or more caps, we do
1435  * this indirectly: if a cap is unused, we prune its aliases, at which
1436  * point the inode will hopefully get dropped to.
1437  *
1438  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1439  * memory pressure from the MDS, though, so it needn't be perfect.
1440  */
1441 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1442 {
1443         struct ceph_mds_session *session = arg;
1444         struct ceph_inode_info *ci = ceph_inode(inode);
1445         int used, wanted, oissued, mine;
1446
1447         if (session->s_trim_caps <= 0)
1448                 return -1;
1449
1450         spin_lock(&ci->i_ceph_lock);
1451         mine = cap->issued | cap->implemented;
1452         used = __ceph_caps_used(ci);
1453         wanted = __ceph_caps_file_wanted(ci);
1454         oissued = __ceph_caps_issued_other(ci, cap);
1455
1456         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1457              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1458              ceph_cap_string(used), ceph_cap_string(wanted));
1459         if (cap == ci->i_auth_cap) {
1460                 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1461                     !list_empty(&ci->i_cap_snaps))
1462                         goto out;
1463                 if ((used | wanted) & CEPH_CAP_ANY_WR)
1464                         goto out;
1465         }
1466         /* The inode has cached pages, but it's no longer used.
1467          * we can safely drop it */
1468         if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1469             !(oissued & CEPH_CAP_FILE_CACHE)) {
1470           used = 0;
1471           oissued = 0;
1472         }
1473         if ((used | wanted) & ~oissued & mine)
1474                 goto out;   /* we need these caps */
1475
1476         if (oissued) {
1477                 /* we aren't the only cap.. just remove us */
1478                 __ceph_remove_cap(cap, true);
1479                 session->s_trim_caps--;
1480         } else {
1481                 struct dentry *dentry;
1482                 /* try dropping referring dentries */
1483                 spin_unlock(&ci->i_ceph_lock);
1484                 dentry = d_find_any_alias(inode);
1485                 if (dentry && drop_negative_children(dentry)) {
1486                         int count;
1487                         dput(dentry);
1488                         d_prune_aliases(inode);
1489                         count = atomic_read(&inode->i_count);
1490                         if (count == 1)
1491                                 session->s_trim_caps--;
1492                         dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1493                              inode, cap, count);
1494                 } else {
1495                         dput(dentry);
1496                 }
1497                 return 0;
1498         }
1499
1500 out:
1501         spin_unlock(&ci->i_ceph_lock);
1502         return 0;
1503 }
1504
1505 /*
1506  * Trim session cap count down to some max number.
1507  */
1508 static int trim_caps(struct ceph_mds_client *mdsc,
1509                      struct ceph_mds_session *session,
1510                      int max_caps)
1511 {
1512         int trim_caps = session->s_nr_caps - max_caps;
1513
1514         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1515              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1516         if (trim_caps > 0) {
1517                 session->s_trim_caps = trim_caps;
1518                 iterate_session_caps(session, trim_caps_cb, session);
1519                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1520                      session->s_mds, session->s_nr_caps, max_caps,
1521                         trim_caps - session->s_trim_caps);
1522                 session->s_trim_caps = 0;
1523         }
1524
1525         ceph_send_cap_releases(mdsc, session);
1526         return 0;
1527 }
1528
1529 static int check_caps_flush(struct ceph_mds_client *mdsc,
1530                             u64 want_flush_tid)
1531 {
1532         int ret = 1;
1533
1534         spin_lock(&mdsc->cap_dirty_lock);
1535         if (!list_empty(&mdsc->cap_flush_list)) {
1536                 struct ceph_cap_flush *cf =
1537                         list_first_entry(&mdsc->cap_flush_list,
1538                                          struct ceph_cap_flush, g_list);
1539                 if (cf->tid <= want_flush_tid) {
1540                         dout("check_caps_flush still flushing tid "
1541                              "%llu <= %llu\n", cf->tid, want_flush_tid);
1542                         ret = 0;
1543                 }
1544         }
1545         spin_unlock(&mdsc->cap_dirty_lock);
1546         return ret;
1547 }
1548
1549 /*
1550  * flush all dirty inode data to disk.
1551  *
1552  * returns true if we've flushed through want_flush_tid
1553  */
1554 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1555                             u64 want_flush_tid)
1556 {
1557         dout("check_caps_flush want %llu\n", want_flush_tid);
1558
1559         wait_event(mdsc->cap_flushing_wq,
1560                    check_caps_flush(mdsc, want_flush_tid));
1561
1562         dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1563 }
1564
1565 /*
1566  * called under s_mutex
1567  */
1568 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1569                             struct ceph_mds_session *session)
1570 {
1571         struct ceph_msg *msg = NULL;
1572         struct ceph_mds_cap_release *head;
1573         struct ceph_mds_cap_item *item;
1574         struct ceph_cap *cap;
1575         LIST_HEAD(tmp_list);
1576         int num_cap_releases;
1577
1578         spin_lock(&session->s_cap_lock);
1579 again:
1580         list_splice_init(&session->s_cap_releases, &tmp_list);
1581         num_cap_releases = session->s_num_cap_releases;
1582         session->s_num_cap_releases = 0;
1583         spin_unlock(&session->s_cap_lock);
1584
1585         while (!list_empty(&tmp_list)) {
1586                 if (!msg) {
1587                         msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1588                                         PAGE_SIZE, GFP_NOFS, false);
1589                         if (!msg)
1590                                 goto out_err;
1591                         head = msg->front.iov_base;
1592                         head->num = cpu_to_le32(0);
1593                         msg->front.iov_len = sizeof(*head);
1594                 }
1595                 cap = list_first_entry(&tmp_list, struct ceph_cap,
1596                                         session_caps);
1597                 list_del(&cap->session_caps);
1598                 num_cap_releases--;
1599
1600                 head = msg->front.iov_base;
1601                 le32_add_cpu(&head->num, 1);
1602                 item = msg->front.iov_base + msg->front.iov_len;
1603                 item->ino = cpu_to_le64(cap->cap_ino);
1604                 item->cap_id = cpu_to_le64(cap->cap_id);
1605                 item->migrate_seq = cpu_to_le32(cap->mseq);
1606                 item->seq = cpu_to_le32(cap->issue_seq);
1607                 msg->front.iov_len += sizeof(*item);
1608
1609                 ceph_put_cap(mdsc, cap);
1610
1611                 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1612                         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1613                         dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1614                         ceph_con_send(&session->s_con, msg);
1615                         msg = NULL;
1616                 }
1617         }
1618
1619         BUG_ON(num_cap_releases != 0);
1620
1621         spin_lock(&session->s_cap_lock);
1622         if (!list_empty(&session->s_cap_releases))
1623                 goto again;
1624         spin_unlock(&session->s_cap_lock);
1625
1626         if (msg) {
1627                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1628                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1629                 ceph_con_send(&session->s_con, msg);
1630         }
1631         return;
1632 out_err:
1633         pr_err("send_cap_releases mds%d, failed to allocate message\n",
1634                 session->s_mds);
1635         spin_lock(&session->s_cap_lock);
1636         list_splice(&tmp_list, &session->s_cap_releases);
1637         session->s_num_cap_releases += num_cap_releases;
1638         spin_unlock(&session->s_cap_lock);
1639 }
1640
1641 /*
1642  * requests
1643  */
1644
1645 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1646                                     struct inode *dir)
1647 {
1648         struct ceph_inode_info *ci = ceph_inode(dir);
1649         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1650         struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1651         size_t size = sizeof(struct ceph_mds_reply_dir_entry);
1652         int order, num_entries;
1653
1654         spin_lock(&ci->i_ceph_lock);
1655         num_entries = ci->i_files + ci->i_subdirs;
1656         spin_unlock(&ci->i_ceph_lock);
1657         num_entries = max(num_entries, 1);
1658         num_entries = min(num_entries, opt->max_readdir);
1659
1660         order = get_order(size * num_entries);
1661         while (order >= 0) {
1662                 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
1663                                                              __GFP_NOWARN,
1664                                                              order);
1665                 if (rinfo->dir_entries)
1666                         break;
1667                 order--;
1668         }
1669         if (!rinfo->dir_entries)
1670                 return -ENOMEM;
1671
1672         num_entries = (PAGE_SIZE << order) / size;
1673         num_entries = min(num_entries, opt->max_readdir);
1674
1675         rinfo->dir_buf_size = PAGE_SIZE << order;
1676         req->r_num_caps = num_entries + 1;
1677         req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1678         req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1679         return 0;
1680 }
1681
1682 /*
1683  * Create an mds request.
1684  */
1685 struct ceph_mds_request *
1686 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1687 {
1688         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1689
1690         if (!req)
1691                 return ERR_PTR(-ENOMEM);
1692
1693         mutex_init(&req->r_fill_mutex);
1694         req->r_mdsc = mdsc;
1695         req->r_started = jiffies;
1696         req->r_resend_mds = -1;
1697         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1698         INIT_LIST_HEAD(&req->r_unsafe_target_item);
1699         req->r_fmode = -1;
1700         kref_init(&req->r_kref);
1701         RB_CLEAR_NODE(&req->r_node);
1702         INIT_LIST_HEAD(&req->r_wait);
1703         init_completion(&req->r_completion);
1704         init_completion(&req->r_safe_completion);
1705         INIT_LIST_HEAD(&req->r_unsafe_item);
1706
1707         req->r_stamp = current_fs_time(mdsc->fsc->sb);
1708
1709         req->r_op = op;
1710         req->r_direct_mode = mode;
1711         return req;
1712 }
1713
1714 /*
1715  * return oldest (lowest) request, tid in request tree, 0 if none.
1716  *
1717  * called under mdsc->mutex.
1718  */
1719 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1720 {
1721         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1722                 return NULL;
1723         return rb_entry(rb_first(&mdsc->request_tree),
1724                         struct ceph_mds_request, r_node);
1725 }
1726
1727 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1728 {
1729         return mdsc->oldest_tid;
1730 }
1731
1732 /*
1733  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1734  * on build_path_from_dentry in fs/cifs/dir.c.
1735  *
1736  * If @stop_on_nosnap, generate path relative to the first non-snapped
1737  * inode.
1738  *
1739  * Encode hidden .snap dirs as a double /, i.e.
1740  *   foo/.snap/bar -> foo//bar
1741  */
1742 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1743                            int stop_on_nosnap)
1744 {
1745         struct dentry *temp;
1746         char *path;
1747         int len, pos;
1748         unsigned seq;
1749
1750         if (dentry == NULL)
1751                 return ERR_PTR(-EINVAL);
1752
1753 retry:
1754         len = 0;
1755         seq = read_seqbegin(&rename_lock);
1756         rcu_read_lock();
1757         for (temp = dentry; !IS_ROOT(temp);) {
1758                 struct inode *inode = d_inode(temp);
1759                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1760                         len++;  /* slash only */
1761                 else if (stop_on_nosnap && inode &&
1762                          ceph_snap(inode) == CEPH_NOSNAP)
1763                         break;
1764                 else
1765                         len += 1 + temp->d_name.len;
1766                 temp = temp->d_parent;
1767         }
1768         rcu_read_unlock();
1769         if (len)
1770                 len--;  /* no leading '/' */
1771
1772         path = kmalloc(len+1, GFP_NOFS);
1773         if (path == NULL)
1774                 return ERR_PTR(-ENOMEM);
1775         pos = len;
1776         path[pos] = 0;  /* trailing null */
1777         rcu_read_lock();
1778         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1779                 struct inode *inode;
1780
1781                 spin_lock(&temp->d_lock);
1782                 inode = d_inode(temp);
1783                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1784                         dout("build_path path+%d: %p SNAPDIR\n",
1785                              pos, temp);
1786                 } else if (stop_on_nosnap && inode &&
1787                            ceph_snap(inode) == CEPH_NOSNAP) {
1788                         spin_unlock(&temp->d_lock);
1789                         break;
1790                 } else {
1791                         pos -= temp->d_name.len;
1792                         if (pos < 0) {
1793                                 spin_unlock(&temp->d_lock);
1794                                 break;
1795                         }
1796                         strncpy(path + pos, temp->d_name.name,
1797                                 temp->d_name.len);
1798                 }
1799                 spin_unlock(&temp->d_lock);
1800                 if (pos)
1801                         path[--pos] = '/';
1802                 temp = temp->d_parent;
1803         }
1804         rcu_read_unlock();
1805         if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1806                 pr_err("build_path did not end path lookup where "
1807                        "expected, namelen is %d, pos is %d\n", len, pos);
1808                 /* presumably this is only possible if racing with a
1809                    rename of one of the parent directories (we can not
1810                    lock the dentries above us to prevent this, but
1811                    retrying should be harmless) */
1812                 kfree(path);
1813                 goto retry;
1814         }
1815
1816         *base = ceph_ino(d_inode(temp));
1817         *plen = len;
1818         dout("build_path on %p %d built %llx '%.*s'\n",
1819              dentry, d_count(dentry), *base, len, path);
1820         return path;
1821 }
1822
1823 static int build_dentry_path(struct dentry *dentry,
1824                              const char **ppath, int *ppathlen, u64 *pino,
1825                              int *pfreepath)
1826 {
1827         char *path;
1828         struct inode *dir;
1829
1830         rcu_read_lock();
1831         dir = d_inode_rcu(dentry->d_parent);
1832         if (dir && ceph_snap(dir) == CEPH_NOSNAP) {
1833                 *pino = ceph_ino(dir);
1834                 rcu_read_unlock();
1835                 *ppath = dentry->d_name.name;
1836                 *ppathlen = dentry->d_name.len;
1837                 return 0;
1838         }
1839         rcu_read_unlock();
1840         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1841         if (IS_ERR(path))
1842                 return PTR_ERR(path);
1843         *ppath = path;
1844         *pfreepath = 1;
1845         return 0;
1846 }
1847
1848 static int build_inode_path(struct inode *inode,
1849                             const char **ppath, int *ppathlen, u64 *pino,
1850                             int *pfreepath)
1851 {
1852         struct dentry *dentry;
1853         char *path;
1854
1855         if (ceph_snap(inode) == CEPH_NOSNAP) {
1856                 *pino = ceph_ino(inode);
1857                 *ppathlen = 0;
1858                 return 0;
1859         }
1860         dentry = d_find_alias(inode);
1861         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1862         dput(dentry);
1863         if (IS_ERR(path))
1864                 return PTR_ERR(path);
1865         *ppath = path;
1866         *pfreepath = 1;
1867         return 0;
1868 }
1869
1870 /*
1871  * request arguments may be specified via an inode *, a dentry *, or
1872  * an explicit ino+path.
1873  */
1874 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1875                                   const char *rpath, u64 rino,
1876                                   const char **ppath, int *pathlen,
1877                                   u64 *ino, int *freepath)
1878 {
1879         int r = 0;
1880
1881         if (rinode) {
1882                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1883                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1884                      ceph_snap(rinode));
1885         } else if (rdentry) {
1886                 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1887                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1888                      *ppath);
1889         } else if (rpath || rino) {
1890                 *ino = rino;
1891                 *ppath = rpath;
1892                 *pathlen = rpath ? strlen(rpath) : 0;
1893                 dout(" path %.*s\n", *pathlen, rpath);
1894         }
1895
1896         return r;
1897 }
1898
1899 /*
1900  * called under mdsc->mutex
1901  */
1902 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1903                                                struct ceph_mds_request *req,
1904                                                int mds, bool drop_cap_releases)
1905 {
1906         struct ceph_msg *msg;
1907         struct ceph_mds_request_head *head;
1908         const char *path1 = NULL;
1909         const char *path2 = NULL;
1910         u64 ino1 = 0, ino2 = 0;
1911         int pathlen1 = 0, pathlen2 = 0;
1912         int freepath1 = 0, freepath2 = 0;
1913         int len;
1914         u16 releases;
1915         void *p, *end;
1916         int ret;
1917
1918         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1919                               req->r_path1, req->r_ino1.ino,
1920                               &path1, &pathlen1, &ino1, &freepath1);
1921         if (ret < 0) {
1922                 msg = ERR_PTR(ret);
1923                 goto out;
1924         }
1925
1926         ret = set_request_path_attr(NULL, req->r_old_dentry,
1927                               req->r_path2, req->r_ino2.ino,
1928                               &path2, &pathlen2, &ino2, &freepath2);
1929         if (ret < 0) {
1930                 msg = ERR_PTR(ret);
1931                 goto out_free1;
1932         }
1933
1934         len = sizeof(*head) +
1935                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
1936                 sizeof(struct ceph_timespec);
1937
1938         /* calculate (max) length for cap releases */
1939         len += sizeof(struct ceph_mds_request_release) *
1940                 (!!req->r_inode_drop + !!req->r_dentry_drop +
1941                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1942         if (req->r_dentry_drop)
1943                 len += req->r_dentry->d_name.len;
1944         if (req->r_old_dentry_drop)
1945                 len += req->r_old_dentry->d_name.len;
1946
1947         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1948         if (!msg) {
1949                 msg = ERR_PTR(-ENOMEM);
1950                 goto out_free2;
1951         }
1952
1953         msg->hdr.version = cpu_to_le16(2);
1954         msg->hdr.tid = cpu_to_le64(req->r_tid);
1955
1956         head = msg->front.iov_base;
1957         p = msg->front.iov_base + sizeof(*head);
1958         end = msg->front.iov_base + msg->front.iov_len;
1959
1960         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1961         head->op = cpu_to_le32(req->r_op);
1962         head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1963         head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1964         head->args = req->r_args;
1965
1966         ceph_encode_filepath(&p, end, ino1, path1);
1967         ceph_encode_filepath(&p, end, ino2, path2);
1968
1969         /* make note of release offset, in case we need to replay */
1970         req->r_request_release_offset = p - msg->front.iov_base;
1971
1972         /* cap releases */
1973         releases = 0;
1974         if (req->r_inode_drop)
1975                 releases += ceph_encode_inode_release(&p,
1976                       req->r_inode ? req->r_inode : d_inode(req->r_dentry),
1977                       mds, req->r_inode_drop, req->r_inode_unless, 0);
1978         if (req->r_dentry_drop)
1979                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1980                        mds, req->r_dentry_drop, req->r_dentry_unless);
1981         if (req->r_old_dentry_drop)
1982                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1983                        mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1984         if (req->r_old_inode_drop)
1985                 releases += ceph_encode_inode_release(&p,
1986                       d_inode(req->r_old_dentry),
1987                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1988
1989         if (drop_cap_releases) {
1990                 releases = 0;
1991                 p = msg->front.iov_base + req->r_request_release_offset;
1992         }
1993
1994         head->num_releases = cpu_to_le16(releases);
1995
1996         /* time stamp */
1997         {
1998                 struct ceph_timespec ts;
1999                 ceph_encode_timespec(&ts, &req->r_stamp);
2000                 ceph_encode_copy(&p, &ts, sizeof(ts));
2001         }
2002
2003         BUG_ON(p > end);
2004         msg->front.iov_len = p - msg->front.iov_base;
2005         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2006
2007         if (req->r_pagelist) {
2008                 struct ceph_pagelist *pagelist = req->r_pagelist;
2009                 atomic_inc(&pagelist->refcnt);
2010                 ceph_msg_data_add_pagelist(msg, pagelist);
2011                 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2012         } else {
2013                 msg->hdr.data_len = 0;
2014         }
2015
2016         msg->hdr.data_off = cpu_to_le16(0);
2017
2018 out_free2:
2019         if (freepath2)
2020                 kfree((char *)path2);
2021 out_free1:
2022         if (freepath1)
2023                 kfree((char *)path1);
2024 out:
2025         return msg;
2026 }
2027
2028 /*
2029  * called under mdsc->mutex if error, under no mutex if
2030  * success.
2031  */
2032 static void complete_request(struct ceph_mds_client *mdsc,
2033                              struct ceph_mds_request *req)
2034 {
2035         if (req->r_callback)
2036                 req->r_callback(mdsc, req);
2037         else
2038                 complete_all(&req->r_completion);
2039 }
2040
2041 /*
2042  * called under mdsc->mutex
2043  */
2044 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2045                                   struct ceph_mds_request *req,
2046                                   int mds, bool drop_cap_releases)
2047 {
2048         struct ceph_mds_request_head *rhead;
2049         struct ceph_msg *msg;
2050         int flags = 0;
2051
2052         req->r_attempts++;
2053         if (req->r_inode) {
2054                 struct ceph_cap *cap =
2055                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2056
2057                 if (cap)
2058                         req->r_sent_on_mseq = cap->mseq;
2059                 else
2060                         req->r_sent_on_mseq = -1;
2061         }
2062         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2063              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2064
2065         if (req->r_got_unsafe) {
2066                 void *p;
2067                 /*
2068                  * Replay.  Do not regenerate message (and rebuild
2069                  * paths, etc.); just use the original message.
2070                  * Rebuilding paths will break for renames because
2071                  * d_move mangles the src name.
2072                  */
2073                 msg = req->r_request;
2074                 rhead = msg->front.iov_base;
2075
2076                 flags = le32_to_cpu(rhead->flags);
2077                 flags |= CEPH_MDS_FLAG_REPLAY;
2078                 rhead->flags = cpu_to_le32(flags);
2079
2080                 if (req->r_target_inode)
2081                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2082
2083                 rhead->num_retry = req->r_attempts - 1;
2084
2085                 /* remove cap/dentry releases from message */
2086                 rhead->num_releases = 0;
2087
2088                 /* time stamp */
2089                 p = msg->front.iov_base + req->r_request_release_offset;
2090                 {
2091                         struct ceph_timespec ts;
2092                         ceph_encode_timespec(&ts, &req->r_stamp);
2093                         ceph_encode_copy(&p, &ts, sizeof(ts));
2094                 }
2095
2096                 msg->front.iov_len = p - msg->front.iov_base;
2097                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2098                 return 0;
2099         }
2100
2101         if (req->r_request) {
2102                 ceph_msg_put(req->r_request);
2103                 req->r_request = NULL;
2104         }
2105         msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2106         if (IS_ERR(msg)) {
2107                 req->r_err = PTR_ERR(msg);
2108                 return PTR_ERR(msg);
2109         }
2110         req->r_request = msg;
2111
2112         rhead = msg->front.iov_base;
2113         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2114         if (req->r_got_unsafe)
2115                 flags |= CEPH_MDS_FLAG_REPLAY;
2116         if (req->r_locked_dir)
2117                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2118         rhead->flags = cpu_to_le32(flags);
2119         rhead->num_fwd = req->r_num_fwd;
2120         rhead->num_retry = req->r_attempts - 1;
2121         rhead->ino = 0;
2122
2123         dout(" r_locked_dir = %p\n", req->r_locked_dir);
2124         return 0;
2125 }
2126
2127 /*
2128  * send request, or put it on the appropriate wait list.
2129  */
2130 static int __do_request(struct ceph_mds_client *mdsc,
2131                         struct ceph_mds_request *req)
2132 {
2133         struct ceph_mds_session *session = NULL;
2134         int mds = -1;
2135         int err = 0;
2136
2137         if (req->r_err || req->r_got_result) {
2138                 if (req->r_aborted)
2139                         __unregister_request(mdsc, req);
2140                 goto out;
2141         }
2142
2143         if (req->r_timeout &&
2144             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2145                 dout("do_request timed out\n");
2146                 err = -EIO;
2147                 goto finish;
2148         }
2149         if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2150                 dout("do_request forced umount\n");
2151                 err = -EIO;
2152                 goto finish;
2153         }
2154
2155         put_request_session(req);
2156
2157         mds = __choose_mds(mdsc, req);
2158         if (mds < 0 ||
2159             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2160                 if (mdsc->mdsmap_err) {
2161                         err = mdsc->mdsmap_err;
2162                         dout("do_request mdsmap err %d\n", err);
2163                         goto finish;
2164                 }
2165                 dout("do_request no mds or not active, waiting for map\n");
2166                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2167                 goto out;
2168         }
2169
2170         /* get, open session */
2171         session = __ceph_lookup_mds_session(mdsc, mds);
2172         if (!session) {
2173                 session = register_session(mdsc, mds);
2174                 if (IS_ERR(session)) {
2175                         err = PTR_ERR(session);
2176                         goto finish;
2177                 }
2178         }
2179         req->r_session = get_session(session);
2180
2181         dout("do_request mds%d session %p state %s\n", mds, session,
2182              ceph_session_state_name(session->s_state));
2183         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2184             session->s_state != CEPH_MDS_SESSION_HUNG) {
2185                 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2186                         err = -EACCES;
2187                         goto out_session;
2188                 }
2189                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2190                     session->s_state == CEPH_MDS_SESSION_CLOSING)
2191                         __open_session(mdsc, session);
2192                 list_add(&req->r_wait, &session->s_waiting);
2193                 goto out_session;
2194         }
2195
2196         /* send request */
2197         req->r_resend_mds = -1;   /* forget any previous mds hint */
2198
2199         if (req->r_request_started == 0)   /* note request start time */
2200                 req->r_request_started = jiffies;
2201
2202         err = __prepare_send_request(mdsc, req, mds, false);
2203         if (!err) {
2204                 ceph_msg_get(req->r_request);
2205                 ceph_con_send(&session->s_con, req->r_request);
2206         }
2207
2208 out_session:
2209         ceph_put_mds_session(session);
2210 finish:
2211         if (err) {
2212                 dout("__do_request early error %d\n", err);
2213                 req->r_err = err;
2214                 complete_request(mdsc, req);
2215                 __unregister_request(mdsc, req);
2216         }
2217 out:
2218         return err;
2219 }
2220
2221 /*
2222  * called under mdsc->mutex
2223  */
2224 static void __wake_requests(struct ceph_mds_client *mdsc,
2225                             struct list_head *head)
2226 {
2227         struct ceph_mds_request *req;
2228         LIST_HEAD(tmp_list);
2229
2230         list_splice_init(head, &tmp_list);
2231
2232         while (!list_empty(&tmp_list)) {
2233                 req = list_entry(tmp_list.next,
2234                                  struct ceph_mds_request, r_wait);
2235                 list_del_init(&req->r_wait);
2236                 dout(" wake request %p tid %llu\n", req, req->r_tid);
2237                 __do_request(mdsc, req);
2238         }
2239 }
2240
2241 /*
2242  * Wake up threads with requests pending for @mds, so that they can
2243  * resubmit their requests to a possibly different mds.
2244  */
2245 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2246 {
2247         struct ceph_mds_request *req;
2248         struct rb_node *p = rb_first(&mdsc->request_tree);
2249
2250         dout("kick_requests mds%d\n", mds);
2251         while (p) {
2252                 req = rb_entry(p, struct ceph_mds_request, r_node);
2253                 p = rb_next(p);
2254                 if (req->r_got_unsafe)
2255                         continue;
2256                 if (req->r_attempts > 0)
2257                         continue; /* only new requests */
2258                 if (req->r_session &&
2259                     req->r_session->s_mds == mds) {
2260                         dout(" kicking tid %llu\n", req->r_tid);
2261                         list_del_init(&req->r_wait);
2262                         __do_request(mdsc, req);
2263                 }
2264         }
2265 }
2266
2267 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2268                               struct ceph_mds_request *req)
2269 {
2270         dout("submit_request on %p\n", req);
2271         mutex_lock(&mdsc->mutex);
2272         __register_request(mdsc, req, NULL);
2273         __do_request(mdsc, req);
2274         mutex_unlock(&mdsc->mutex);
2275 }
2276
2277 /*
2278  * Synchrously perform an mds request.  Take care of all of the
2279  * session setup, forwarding, retry details.
2280  */
2281 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2282                          struct inode *dir,
2283                          struct ceph_mds_request *req)
2284 {
2285         int err;
2286
2287         dout("do_request on %p\n", req);
2288
2289         /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
2290         if (req->r_inode)
2291                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2292         if (req->r_locked_dir)
2293                 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
2294         if (req->r_old_dentry_dir)
2295                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2296                                   CEPH_CAP_PIN);
2297
2298         /* issue */
2299         mutex_lock(&mdsc->mutex);
2300         __register_request(mdsc, req, dir);
2301         __do_request(mdsc, req);
2302
2303         if (req->r_err) {
2304                 err = req->r_err;
2305                 goto out;
2306         }
2307
2308         /* wait */
2309         mutex_unlock(&mdsc->mutex);
2310         dout("do_request waiting\n");
2311         if (!req->r_timeout && req->r_wait_for_completion) {
2312                 err = req->r_wait_for_completion(mdsc, req);
2313         } else {
2314                 long timeleft = wait_for_completion_killable_timeout(
2315                                         &req->r_completion,
2316                                         ceph_timeout_jiffies(req->r_timeout));
2317                 if (timeleft > 0)
2318                         err = 0;
2319                 else if (!timeleft)
2320                         err = -EIO;  /* timed out */
2321                 else
2322                         err = timeleft;  /* killed */
2323         }
2324         dout("do_request waited, got %d\n", err);
2325         mutex_lock(&mdsc->mutex);
2326
2327         /* only abort if we didn't race with a real reply */
2328         if (req->r_got_result) {
2329                 err = le32_to_cpu(req->r_reply_info.head->result);
2330         } else if (err < 0) {
2331                 dout("aborted request %lld with %d\n", req->r_tid, err);
2332
2333                 /*
2334                  * ensure we aren't running concurrently with
2335                  * ceph_fill_trace or ceph_readdir_prepopulate, which
2336                  * rely on locks (dir mutex) held by our caller.
2337                  */
2338                 mutex_lock(&req->r_fill_mutex);
2339                 req->r_err = err;
2340                 req->r_aborted = true;
2341                 mutex_unlock(&req->r_fill_mutex);
2342
2343                 if (req->r_locked_dir &&
2344                     (req->r_op & CEPH_MDS_OP_WRITE))
2345                         ceph_invalidate_dir_request(req);
2346         } else {
2347                 err = req->r_err;
2348         }
2349
2350 out:
2351         mutex_unlock(&mdsc->mutex);
2352         dout("do_request %p done, result %d\n", req, err);
2353         return err;
2354 }
2355
2356 /*
2357  * Invalidate dir's completeness, dentry lease state on an aborted MDS
2358  * namespace request.
2359  */
2360 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2361 {
2362         struct inode *inode = req->r_locked_dir;
2363
2364         dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2365
2366         ceph_dir_clear_complete(inode);
2367         if (req->r_dentry)
2368                 ceph_invalidate_dentry_lease(req->r_dentry);
2369         if (req->r_old_dentry)
2370                 ceph_invalidate_dentry_lease(req->r_old_dentry);
2371 }
2372
2373 /*
2374  * Handle mds reply.
2375  *
2376  * We take the session mutex and parse and process the reply immediately.
2377  * This preserves the logical ordering of replies, capabilities, etc., sent
2378  * by the MDS as they are applied to our local cache.
2379  */
2380 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2381 {
2382         struct ceph_mds_client *mdsc = session->s_mdsc;
2383         struct ceph_mds_request *req;
2384         struct ceph_mds_reply_head *head = msg->front.iov_base;
2385         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2386         struct ceph_snap_realm *realm;
2387         u64 tid;
2388         int err, result;
2389         int mds = session->s_mds;
2390
2391         if (msg->front.iov_len < sizeof(*head)) {
2392                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2393                 ceph_msg_dump(msg);
2394                 return;
2395         }
2396
2397         /* get request, session */
2398         tid = le64_to_cpu(msg->hdr.tid);
2399         mutex_lock(&mdsc->mutex);
2400         req = lookup_get_request(mdsc, tid);
2401         if (!req) {
2402                 dout("handle_reply on unknown tid %llu\n", tid);
2403                 mutex_unlock(&mdsc->mutex);
2404                 return;
2405         }
2406         dout("handle_reply %p\n", req);
2407
2408         /* correct session? */
2409         if (req->r_session != session) {
2410                 pr_err("mdsc_handle_reply got %llu on session mds%d"
2411                        " not mds%d\n", tid, session->s_mds,
2412                        req->r_session ? req->r_session->s_mds : -1);
2413                 mutex_unlock(&mdsc->mutex);
2414                 goto out;
2415         }
2416
2417         /* dup? */
2418         if ((req->r_got_unsafe && !head->safe) ||
2419             (req->r_got_safe && head->safe)) {
2420                 pr_warn("got a dup %s reply on %llu from mds%d\n",
2421                            head->safe ? "safe" : "unsafe", tid, mds);
2422                 mutex_unlock(&mdsc->mutex);
2423                 goto out;
2424         }
2425         if (req->r_got_safe) {
2426                 pr_warn("got unsafe after safe on %llu from mds%d\n",
2427                            tid, mds);
2428                 mutex_unlock(&mdsc->mutex);
2429                 goto out;
2430         }
2431
2432         result = le32_to_cpu(head->result);
2433
2434         /*
2435          * Handle an ESTALE
2436          * if we're not talking to the authority, send to them
2437          * if the authority has changed while we weren't looking,
2438          * send to new authority
2439          * Otherwise we just have to return an ESTALE
2440          */
2441         if (result == -ESTALE) {
2442                 dout("got ESTALE on request %llu", req->r_tid);
2443                 req->r_resend_mds = -1;
2444                 if (req->r_direct_mode != USE_AUTH_MDS) {
2445                         dout("not using auth, setting for that now");
2446                         req->r_direct_mode = USE_AUTH_MDS;
2447                         __do_request(mdsc, req);
2448                         mutex_unlock(&mdsc->mutex);
2449                         goto out;
2450                 } else  {
2451                         int mds = __choose_mds(mdsc, req);
2452                         if (mds >= 0 && mds != req->r_session->s_mds) {
2453                                 dout("but auth changed, so resending");
2454                                 __do_request(mdsc, req);
2455                                 mutex_unlock(&mdsc->mutex);
2456                                 goto out;
2457                         }
2458                 }
2459                 dout("have to return ESTALE on request %llu", req->r_tid);
2460         }
2461
2462
2463         if (head->safe) {
2464                 req->r_got_safe = true;
2465                 __unregister_request(mdsc, req);
2466
2467                 if (req->r_got_unsafe) {
2468                         /*
2469                          * We already handled the unsafe response, now do the
2470                          * cleanup.  No need to examine the response; the MDS
2471                          * doesn't include any result info in the safe
2472                          * response.  And even if it did, there is nothing
2473                          * useful we could do with a revised return value.
2474                          */
2475                         dout("got safe reply %llu, mds%d\n", tid, mds);
2476
2477                         /* last unsafe request during umount? */
2478                         if (mdsc->stopping && !__get_oldest_req(mdsc))
2479                                 complete_all(&mdsc->safe_umount_waiters);
2480                         mutex_unlock(&mdsc->mutex);
2481                         goto out;
2482                 }
2483         } else {
2484                 req->r_got_unsafe = true;
2485                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2486                 if (req->r_unsafe_dir) {
2487                         struct ceph_inode_info *ci =
2488                                         ceph_inode(req->r_unsafe_dir);
2489                         spin_lock(&ci->i_unsafe_lock);
2490                         list_add_tail(&req->r_unsafe_dir_item,
2491                                       &ci->i_unsafe_dirops);
2492                         spin_unlock(&ci->i_unsafe_lock);
2493                 }
2494         }
2495
2496         dout("handle_reply tid %lld result %d\n", tid, result);
2497         rinfo = &req->r_reply_info;
2498         err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2499         mutex_unlock(&mdsc->mutex);
2500
2501         mutex_lock(&session->s_mutex);
2502         if (err < 0) {
2503                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2504                 ceph_msg_dump(msg);
2505                 goto out_err;
2506         }
2507
2508         /* snap trace */
2509         realm = NULL;
2510         if (rinfo->snapblob_len) {
2511                 down_write(&mdsc->snap_rwsem);
2512                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2513                                 rinfo->snapblob + rinfo->snapblob_len,
2514                                 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2515                                 &realm);
2516                 downgrade_write(&mdsc->snap_rwsem);
2517         } else {
2518                 down_read(&mdsc->snap_rwsem);
2519         }
2520
2521         /* insert trace into our cache */
2522         mutex_lock(&req->r_fill_mutex);
2523         current->journal_info = req;
2524         err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2525         if (err == 0) {
2526                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2527                                     req->r_op == CEPH_MDS_OP_LSSNAP))
2528                         ceph_readdir_prepopulate(req, req->r_session);
2529                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2530         }
2531         current->journal_info = NULL;
2532         mutex_unlock(&req->r_fill_mutex);
2533
2534         up_read(&mdsc->snap_rwsem);
2535         if (realm)
2536                 ceph_put_snap_realm(mdsc, realm);
2537
2538         if (err == 0 && req->r_got_unsafe && req->r_target_inode) {
2539                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
2540                 spin_lock(&ci->i_unsafe_lock);
2541                 list_add_tail(&req->r_unsafe_target_item, &ci->i_unsafe_iops);
2542                 spin_unlock(&ci->i_unsafe_lock);
2543         }
2544 out_err:
2545         mutex_lock(&mdsc->mutex);
2546         if (!req->r_aborted) {
2547                 if (err) {
2548                         req->r_err = err;
2549                 } else {
2550                         req->r_reply =  ceph_msg_get(msg);
2551                         req->r_got_result = true;
2552                 }
2553         } else {
2554                 dout("reply arrived after request %lld was aborted\n", tid);
2555         }
2556         mutex_unlock(&mdsc->mutex);
2557
2558         mutex_unlock(&session->s_mutex);
2559
2560         /* kick calling process */
2561         complete_request(mdsc, req);
2562 out:
2563         ceph_mdsc_put_request(req);
2564         return;
2565 }
2566
2567
2568
2569 /*
2570  * handle mds notification that our request has been forwarded.
2571  */
2572 static void handle_forward(struct ceph_mds_client *mdsc,
2573                            struct ceph_mds_session *session,
2574                            struct ceph_msg *msg)
2575 {
2576         struct ceph_mds_request *req;
2577         u64 tid = le64_to_cpu(msg->hdr.tid);
2578         u32 next_mds;
2579         u32 fwd_seq;
2580         int err = -EINVAL;
2581         void *p = msg->front.iov_base;
2582         void *end = p + msg->front.iov_len;
2583
2584         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2585         next_mds = ceph_decode_32(&p);
2586         fwd_seq = ceph_decode_32(&p);
2587
2588         mutex_lock(&mdsc->mutex);
2589         req = lookup_get_request(mdsc, tid);
2590         if (!req) {
2591                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2592                 goto out;  /* dup reply? */
2593         }
2594
2595         if (req->r_aborted) {
2596                 dout("forward tid %llu aborted, unregistering\n", tid);
2597                 __unregister_request(mdsc, req);
2598         } else if (fwd_seq <= req->r_num_fwd) {
2599                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2600                      tid, next_mds, req->r_num_fwd, fwd_seq);
2601         } else {
2602                 /* resend. forward race not possible; mds would drop */
2603                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2604                 BUG_ON(req->r_err);
2605                 BUG_ON(req->r_got_result);
2606                 req->r_attempts = 0;
2607                 req->r_num_fwd = fwd_seq;
2608                 req->r_resend_mds = next_mds;
2609                 put_request_session(req);
2610                 __do_request(mdsc, req);
2611         }
2612         ceph_mdsc_put_request(req);
2613 out:
2614         mutex_unlock(&mdsc->mutex);
2615         return;
2616
2617 bad:
2618         pr_err("mdsc_handle_forward decode error err=%d\n", err);
2619 }
2620
2621 /*
2622  * handle a mds session control message
2623  */
2624 static void handle_session(struct ceph_mds_session *session,
2625                            struct ceph_msg *msg)
2626 {
2627         struct ceph_mds_client *mdsc = session->s_mdsc;
2628         u32 op;
2629         u64 seq;
2630         int mds = session->s_mds;
2631         struct ceph_mds_session_head *h = msg->front.iov_base;
2632         int wake = 0;
2633
2634         /* decode */
2635         if (msg->front.iov_len != sizeof(*h))
2636                 goto bad;
2637         op = le32_to_cpu(h->op);
2638         seq = le64_to_cpu(h->seq);
2639
2640         mutex_lock(&mdsc->mutex);
2641         if (op == CEPH_SESSION_CLOSE)
2642                 __unregister_session(mdsc, session);
2643         /* FIXME: this ttl calculation is generous */
2644         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2645         mutex_unlock(&mdsc->mutex);
2646
2647         mutex_lock(&session->s_mutex);
2648
2649         dout("handle_session mds%d %s %p state %s seq %llu\n",
2650              mds, ceph_session_op_name(op), session,
2651              ceph_session_state_name(session->s_state), seq);
2652
2653         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2654                 session->s_state = CEPH_MDS_SESSION_OPEN;
2655                 pr_info("mds%d came back\n", session->s_mds);
2656         }
2657
2658         switch (op) {
2659         case CEPH_SESSION_OPEN:
2660                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2661                         pr_info("mds%d reconnect success\n", session->s_mds);
2662                 session->s_state = CEPH_MDS_SESSION_OPEN;
2663                 renewed_caps(mdsc, session, 0);
2664                 wake = 1;
2665                 if (mdsc->stopping)
2666                         __close_session(mdsc, session);
2667                 break;
2668
2669         case CEPH_SESSION_RENEWCAPS:
2670                 if (session->s_renew_seq == seq)
2671                         renewed_caps(mdsc, session, 1);
2672                 break;
2673
2674         case CEPH_SESSION_CLOSE:
2675                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2676                         pr_info("mds%d reconnect denied\n", session->s_mds);
2677                 cleanup_session_requests(mdsc, session);
2678                 remove_session_caps(session);
2679                 wake = 2; /* for good measure */
2680                 wake_up_all(&mdsc->session_close_wq);
2681                 break;
2682
2683         case CEPH_SESSION_STALE:
2684                 pr_info("mds%d caps went stale, renewing\n",
2685                         session->s_mds);
2686                 spin_lock(&session->s_gen_ttl_lock);
2687                 session->s_cap_gen++;
2688                 session->s_cap_ttl = jiffies - 1;
2689                 spin_unlock(&session->s_gen_ttl_lock);
2690                 send_renew_caps(mdsc, session);
2691                 break;
2692
2693         case CEPH_SESSION_RECALL_STATE:
2694                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2695                 break;
2696
2697         case CEPH_SESSION_FLUSHMSG:
2698                 send_flushmsg_ack(mdsc, session, seq);
2699                 break;
2700
2701         case CEPH_SESSION_FORCE_RO:
2702                 dout("force_session_readonly %p\n", session);
2703                 spin_lock(&session->s_cap_lock);
2704                 session->s_readonly = true;
2705                 spin_unlock(&session->s_cap_lock);
2706                 wake_up_session_caps(session, 0);
2707                 break;
2708
2709         case CEPH_SESSION_REJECT:
2710                 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
2711                 pr_info("mds%d rejected session\n", session->s_mds);
2712                 session->s_state = CEPH_MDS_SESSION_REJECTED;
2713                 cleanup_session_requests(mdsc, session);
2714                 remove_session_caps(session);
2715                 wake = 2; /* for good measure */
2716                 break;
2717
2718         default:
2719                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2720                 WARN_ON(1);
2721         }
2722
2723         mutex_unlock(&session->s_mutex);
2724         if (wake) {
2725                 mutex_lock(&mdsc->mutex);
2726                 __wake_requests(mdsc, &session->s_waiting);
2727                 if (wake == 2)
2728                         kick_requests(mdsc, mds);
2729                 mutex_unlock(&mdsc->mutex);
2730         }
2731         return;
2732
2733 bad:
2734         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2735                (int)msg->front.iov_len);
2736         ceph_msg_dump(msg);
2737         return;
2738 }
2739
2740
2741 /*
2742  * called under session->mutex.
2743  */
2744 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2745                                    struct ceph_mds_session *session)
2746 {
2747         struct ceph_mds_request *req, *nreq;
2748         struct rb_node *p;
2749         int err;
2750
2751         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2752
2753         mutex_lock(&mdsc->mutex);
2754         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2755                 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2756                 if (!err) {
2757                         ceph_msg_get(req->r_request);
2758                         ceph_con_send(&session->s_con, req->r_request);
2759                 }
2760         }
2761
2762         /*
2763          * also re-send old requests when MDS enters reconnect stage. So that MDS
2764          * can process completed request in clientreplay stage.
2765          */
2766         p = rb_first(&mdsc->request_tree);
2767         while (p) {
2768                 req = rb_entry(p, struct ceph_mds_request, r_node);
2769                 p = rb_next(p);
2770                 if (req->r_got_unsafe)
2771                         continue;
2772                 if (req->r_attempts == 0)
2773                         continue; /* only old requests */
2774                 if (req->r_session &&
2775                     req->r_session->s_mds == session->s_mds) {
2776                         err = __prepare_send_request(mdsc, req,
2777                                                      session->s_mds, true);
2778                         if (!err) {
2779                                 ceph_msg_get(req->r_request);
2780                                 ceph_con_send(&session->s_con, req->r_request);
2781                         }
2782                 }
2783         }
2784         mutex_unlock(&mdsc->mutex);
2785 }
2786
2787 /*
2788  * Encode information about a cap for a reconnect with the MDS.
2789  */
2790 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2791                           void *arg)
2792 {
2793         union {
2794                 struct ceph_mds_cap_reconnect v2;
2795                 struct ceph_mds_cap_reconnect_v1 v1;
2796         } rec;
2797         struct ceph_inode_info *ci;
2798         struct ceph_reconnect_state *recon_state = arg;
2799         struct ceph_pagelist *pagelist = recon_state->pagelist;
2800         char *path;
2801         int pathlen, err;
2802         u64 pathbase;
2803         u64 snap_follows;
2804         struct dentry *dentry;
2805
2806         ci = cap->ci;
2807
2808         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2809              inode, ceph_vinop(inode), cap, cap->cap_id,
2810              ceph_cap_string(cap->issued));
2811         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2812         if (err)
2813                 return err;
2814
2815         dentry = d_find_alias(inode);
2816         if (dentry) {
2817                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2818                 if (IS_ERR(path)) {
2819                         err = PTR_ERR(path);
2820                         goto out_dput;
2821                 }
2822         } else {
2823                 path = NULL;
2824                 pathlen = 0;
2825                 pathbase = 0;
2826         }
2827
2828         spin_lock(&ci->i_ceph_lock);
2829         cap->seq = 0;        /* reset cap seq */
2830         cap->issue_seq = 0;  /* and issue_seq */
2831         cap->mseq = 0;       /* and migrate_seq */
2832         cap->cap_gen = cap->session->s_cap_gen;
2833
2834         if (recon_state->msg_version >= 2) {
2835                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2836                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2837                 rec.v2.issued = cpu_to_le32(cap->issued);
2838                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2839                 rec.v2.pathbase = cpu_to_le64(pathbase);
2840                 rec.v2.flock_len = 0;
2841         } else {
2842                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2843                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2844                 rec.v1.issued = cpu_to_le32(cap->issued);
2845                 rec.v1.size = cpu_to_le64(inode->i_size);
2846                 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2847                 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2848                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2849                 rec.v1.pathbase = cpu_to_le64(pathbase);
2850         }
2851
2852         if (list_empty(&ci->i_cap_snaps)) {
2853                 snap_follows = 0;
2854         } else {
2855                 struct ceph_cap_snap *capsnap =
2856                         list_first_entry(&ci->i_cap_snaps,
2857                                          struct ceph_cap_snap, ci_item);
2858                 snap_follows = capsnap->follows;
2859         }
2860         spin_unlock(&ci->i_ceph_lock);
2861
2862         if (recon_state->msg_version >= 2) {
2863                 int num_fcntl_locks, num_flock_locks;
2864                 struct ceph_filelock *flocks;
2865                 size_t struct_len, total_len = 0;
2866                 u8 struct_v = 0;
2867
2868 encode_again:
2869                 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2870                 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2871                                  sizeof(struct ceph_filelock), GFP_NOFS);
2872                 if (!flocks) {
2873                         err = -ENOMEM;
2874                         goto out_free;
2875                 }
2876                 err = ceph_encode_locks_to_buffer(inode, flocks,
2877                                                   num_fcntl_locks,
2878                                                   num_flock_locks);
2879                 if (err) {
2880                         kfree(flocks);
2881                         if (err == -ENOSPC)
2882                                 goto encode_again;
2883                         goto out_free;
2884                 }
2885
2886                 if (recon_state->msg_version >= 3) {
2887                         /* version, compat_version and struct_len */
2888                         total_len = 2 * sizeof(u8) + sizeof(u32);
2889                         struct_v = 2;
2890                 }
2891                 /*
2892                  * number of encoded locks is stable, so copy to pagelist
2893                  */
2894                 struct_len = 2 * sizeof(u32) +
2895                             (num_fcntl_locks + num_flock_locks) *
2896                             sizeof(struct ceph_filelock);
2897                 rec.v2.flock_len = cpu_to_le32(struct_len);
2898
2899                 struct_len += sizeof(rec.v2);
2900                 struct_len += sizeof(u32) + pathlen;
2901
2902                 if (struct_v >= 2)
2903                         struct_len += sizeof(u64); /* snap_follows */
2904
2905                 total_len += struct_len;
2906                 err = ceph_pagelist_reserve(pagelist, total_len);
2907
2908                 if (!err) {
2909                         if (recon_state->msg_version >= 3) {
2910                                 ceph_pagelist_encode_8(pagelist, struct_v);
2911                                 ceph_pagelist_encode_8(pagelist, 1);
2912                                 ceph_pagelist_encode_32(pagelist, struct_len);
2913                         }
2914                         ceph_pagelist_encode_string(pagelist, path, pathlen);
2915                         ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
2916                         ceph_locks_to_pagelist(flocks, pagelist,
2917                                                num_fcntl_locks,
2918                                                num_flock_locks);
2919                         if (struct_v >= 2)
2920                                 ceph_pagelist_encode_64(pagelist, snap_follows);
2921                 }
2922                 kfree(flocks);
2923         } else {
2924                 size_t size = sizeof(u32) + pathlen + sizeof(rec.v1);
2925                 err = ceph_pagelist_reserve(pagelist, size);
2926                 if (!err) {
2927                         ceph_pagelist_encode_string(pagelist, path, pathlen);
2928                         ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
2929                 }
2930         }
2931
2932         recon_state->nr_caps++;
2933 out_free:
2934         kfree(path);
2935 out_dput:
2936         dput(dentry);
2937         return err;
2938 }
2939
2940
2941 /*
2942  * If an MDS fails and recovers, clients need to reconnect in order to
2943  * reestablish shared state.  This includes all caps issued through
2944  * this session _and_ the snap_realm hierarchy.  Because it's not
2945  * clear which snap realms the mds cares about, we send everything we
2946  * know about.. that ensures we'll then get any new info the
2947  * recovering MDS might have.
2948  *
2949  * This is a relatively heavyweight operation, but it's rare.
2950  *
2951  * called with mdsc->mutex held.
2952  */
2953 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2954                                struct ceph_mds_session *session)
2955 {
2956         struct ceph_msg *reply;
2957         struct rb_node *p;
2958         int mds = session->s_mds;
2959         int err = -ENOMEM;
2960         int s_nr_caps;
2961         struct ceph_pagelist *pagelist;
2962         struct ceph_reconnect_state recon_state;
2963
2964         pr_info("mds%d reconnect start\n", mds);
2965
2966         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2967         if (!pagelist)
2968                 goto fail_nopagelist;
2969         ceph_pagelist_init(pagelist);
2970
2971         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2972         if (!reply)
2973                 goto fail_nomsg;
2974
2975         mutex_lock(&session->s_mutex);
2976         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2977         session->s_seq = 0;
2978
2979         dout("session %p state %s\n", session,
2980              ceph_session_state_name(session->s_state));
2981
2982         spin_lock(&session->s_gen_ttl_lock);
2983         session->s_cap_gen++;
2984         spin_unlock(&session->s_gen_ttl_lock);
2985
2986         spin_lock(&session->s_cap_lock);
2987         /* don't know if session is readonly */
2988         session->s_readonly = 0;
2989         /*
2990          * notify __ceph_remove_cap() that we are composing cap reconnect.
2991          * If a cap get released before being added to the cap reconnect,
2992          * __ceph_remove_cap() should skip queuing cap release.
2993          */
2994         session->s_cap_reconnect = 1;
2995         /* drop old cap expires; we're about to reestablish that state */
2996         cleanup_cap_releases(mdsc, session);
2997
2998         /* trim unused caps to reduce MDS's cache rejoin time */
2999         if (mdsc->fsc->sb->s_root)
3000                 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3001
3002         ceph_con_close(&session->s_con);
3003         ceph_con_open(&session->s_con,
3004                       CEPH_ENTITY_TYPE_MDS, mds,
3005                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3006
3007         /* replay unsafe requests */
3008         replay_unsafe_requests(mdsc, session);
3009
3010         down_read(&mdsc->snap_rwsem);
3011
3012         /* traverse this session's caps */
3013         s_nr_caps = session->s_nr_caps;
3014         err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
3015         if (err)
3016                 goto fail;
3017
3018         recon_state.nr_caps = 0;
3019         recon_state.pagelist = pagelist;
3020         if (session->s_con.peer_features & CEPH_FEATURE_MDSENC)
3021                 recon_state.msg_version = 3;
3022         else if (session->s_con.peer_features & CEPH_FEATURE_FLOCK)
3023                 recon_state.msg_version = 2;
3024         else
3025                 recon_state.msg_version = 1;
3026         err = iterate_session_caps(session, encode_caps_cb, &recon_state);
3027         if (err < 0)
3028                 goto fail;
3029
3030         spin_lock(&session->s_cap_lock);
3031         session->s_cap_reconnect = 0;
3032         spin_unlock(&session->s_cap_lock);
3033
3034         /*
3035          * snaprealms.  we provide mds with the ino, seq (version), and
3036          * parent for all of our realms.  If the mds has any newer info,
3037          * it will tell us.
3038          */
3039         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3040                 struct ceph_snap_realm *realm =
3041                         rb_entry(p, struct ceph_snap_realm, node);
3042                 struct ceph_mds_snaprealm_reconnect sr_rec;
3043
3044                 dout(" adding snap realm %llx seq %lld parent %llx\n",
3045                      realm->ino, realm->seq, realm->parent_ino);
3046                 sr_rec.ino = cpu_to_le64(realm->ino);
3047                 sr_rec.seq = cpu_to_le64(realm->seq);
3048                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3049                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3050                 if (err)
3051                         goto fail;
3052         }
3053
3054         reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3055
3056         /* raced with cap release? */
3057         if (s_nr_caps != recon_state.nr_caps) {
3058                 struct page *page = list_first_entry(&pagelist->head,
3059                                                      struct page, lru);
3060                 __le32 *addr = kmap_atomic(page);
3061                 *addr = cpu_to_le32(recon_state.nr_caps);
3062                 kunmap_atomic(addr);
3063         }
3064
3065         reply->hdr.data_len = cpu_to_le32(pagelist->length);
3066         ceph_msg_data_add_pagelist(reply, pagelist);
3067
3068         ceph_early_kick_flushing_caps(mdsc, session);
3069
3070         ceph_con_send(&session->s_con, reply);
3071
3072         mutex_unlock(&session->s_mutex);
3073
3074         mutex_lock(&mdsc->mutex);
3075         __wake_requests(mdsc, &session->s_waiting);
3076         mutex_unlock(&mdsc->mutex);
3077
3078         up_read(&mdsc->snap_rwsem);
3079         return;
3080
3081 fail:
3082         ceph_msg_put(reply);
3083         up_read(&mdsc->snap_rwsem);
3084         mutex_unlock(&session->s_mutex);
3085 fail_nomsg:
3086         ceph_pagelist_release(pagelist);
3087 fail_nopagelist:
3088         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3089         return;
3090 }
3091
3092
3093 /*
3094  * compare old and new mdsmaps, kicking requests
3095  * and closing out old connections as necessary
3096  *
3097  * called under mdsc->mutex.
3098  */
3099 static void check_new_map(struct ceph_mds_client *mdsc,
3100                           struct ceph_mdsmap *newmap,
3101                           struct ceph_mdsmap *oldmap)
3102 {
3103         int i;
3104         int oldstate, newstate;
3105         struct ceph_mds_session *s;
3106
3107         dout("check_new_map new %u old %u\n",
3108              newmap->m_epoch, oldmap->m_epoch);
3109
3110         for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
3111                 if (mdsc->sessions[i] == NULL)
3112                         continue;
3113                 s = mdsc->sessions[i];
3114                 oldstate = ceph_mdsmap_get_state(oldmap, i);
3115                 newstate = ceph_mdsmap_get_state(newmap, i);
3116
3117                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3118                      i, ceph_mds_state_name(oldstate),
3119                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3120                      ceph_mds_state_name(newstate),
3121                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3122                      ceph_session_state_name(s->s_state));
3123
3124                 if (i >= newmap->m_max_mds ||
3125                     memcmp(ceph_mdsmap_get_addr(oldmap, i),
3126                            ceph_mdsmap_get_addr(newmap, i),
3127                            sizeof(struct ceph_entity_addr))) {
3128                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3129                                 /* the session never opened, just close it
3130                                  * out now */
3131                                 __wake_requests(mdsc, &s->s_waiting);
3132                                 __unregister_session(mdsc, s);
3133                         } else {
3134                                 /* just close it */
3135                                 mutex_unlock(&mdsc->mutex);
3136                                 mutex_lock(&s->s_mutex);
3137                                 mutex_lock(&mdsc->mutex);
3138                                 ceph_con_close(&s->s_con);
3139                                 mutex_unlock(&s->s_mutex);
3140                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3141                         }
3142                 } else if (oldstate == newstate) {
3143                         continue;  /* nothing new with this mds */
3144                 }
3145
3146                 /*
3147                  * send reconnect?
3148                  */
3149                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3150                     newstate >= CEPH_MDS_STATE_RECONNECT) {
3151                         mutex_unlock(&mdsc->mutex);
3152                         send_mds_reconnect(mdsc, s);
3153                         mutex_lock(&mdsc->mutex);
3154                 }
3155
3156                 /*
3157                  * kick request on any mds that has gone active.
3158                  */
3159                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3160                     newstate >= CEPH_MDS_STATE_ACTIVE) {
3161                         if (oldstate != CEPH_MDS_STATE_CREATING &&
3162                             oldstate != CEPH_MDS_STATE_STARTING)
3163                                 pr_info("mds%d recovery completed\n", s->s_mds);
3164                         kick_requests(mdsc, i);
3165                         ceph_kick_flushing_caps(mdsc, s);
3166                         wake_up_session_caps(s, 1);
3167                 }
3168         }
3169
3170         for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
3171                 s = mdsc->sessions[i];
3172                 if (!s)
3173                         continue;
3174                 if (!ceph_mdsmap_is_laggy(newmap, i))
3175                         continue;
3176                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3177                     s->s_state == CEPH_MDS_SESSION_HUNG ||
3178                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
3179                         dout(" connecting to export targets of laggy mds%d\n",
3180                              i);
3181                         __open_export_target_sessions(mdsc, s);
3182                 }
3183         }
3184 }
3185
3186
3187
3188 /*
3189  * leases
3190  */
3191
3192 /*
3193  * caller must hold session s_mutex, dentry->d_lock
3194  */
3195 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3196 {
3197         struct ceph_dentry_info *di = ceph_dentry(dentry);
3198
3199         ceph_put_mds_session(di->lease_session);
3200         di->lease_session = NULL;
3201 }
3202
3203 static void handle_lease(struct ceph_mds_client *mdsc,
3204                          struct ceph_mds_session *session,
3205                          struct ceph_msg *msg)
3206 {
3207         struct super_block *sb = mdsc->fsc->sb;
3208         struct inode *inode;
3209         struct dentry *parent, *dentry;
3210         struct ceph_dentry_info *di;
3211         int mds = session->s_mds;
3212         struct ceph_mds_lease *h = msg->front.iov_base;
3213         u32 seq;
3214         struct ceph_vino vino;
3215         struct qstr dname;
3216         int release = 0;
3217
3218         dout("handle_lease from mds%d\n", mds);
3219
3220         /* decode */
3221         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3222                 goto bad;
3223         vino.ino = le64_to_cpu(h->ino);
3224         vino.snap = CEPH_NOSNAP;
3225         seq = le32_to_cpu(h->seq);
3226         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3227         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3228         if (dname.len != get_unaligned_le32(h+1))
3229                 goto bad;
3230
3231         /* lookup inode */
3232         inode = ceph_find_inode(sb, vino);
3233         dout("handle_lease %s, ino %llx %p %.*s\n",
3234              ceph_lease_op_name(h->action), vino.ino, inode,
3235              dname.len, dname.name);
3236
3237         mutex_lock(&session->s_mutex);
3238         session->s_seq++;
3239
3240         if (inode == NULL) {
3241                 dout("handle_lease no inode %llx\n", vino.ino);
3242                 goto release;
3243         }
3244
3245         /* dentry */
3246         parent = d_find_alias(inode);
3247         if (!parent) {
3248                 dout("no parent dentry on inode %p\n", inode);
3249                 WARN_ON(1);
3250                 goto release;  /* hrm... */
3251         }
3252         dname.hash = full_name_hash(parent, dname.name, dname.len);
3253         dentry = d_lookup(parent, &dname);
3254         dput(parent);
3255         if (!dentry)
3256                 goto release;
3257
3258         spin_lock(&dentry->d_lock);
3259         di = ceph_dentry(dentry);
3260         switch (h->action) {
3261         case CEPH_MDS_LEASE_REVOKE:
3262                 if (di->lease_session == session) {
3263                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3264                                 h->seq = cpu_to_le32(di->lease_seq);
3265                         __ceph_mdsc_drop_dentry_lease(dentry);
3266                 }
3267                 release = 1;
3268                 break;
3269
3270         case CEPH_MDS_LEASE_RENEW:
3271                 if (di->lease_session == session &&
3272                     di->lease_gen == session->s_cap_gen &&
3273                     di->lease_renew_from &&
3274                     di->lease_renew_after == 0) {
3275                         unsigned long duration =
3276                                 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3277
3278                         di->lease_seq = seq;
3279                         di->time = di->lease_renew_from + duration;
3280                         di->lease_renew_after = di->lease_renew_from +
3281                                 (duration >> 1);
3282                         di->lease_renew_from = 0;
3283                 }
3284                 break;
3285         }
3286         spin_unlock(&dentry->d_lock);
3287         dput(dentry);
3288
3289         if (!release)
3290                 goto out;
3291
3292 release:
3293         /* let's just reuse the same message */
3294         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3295         ceph_msg_get(msg);
3296         ceph_con_send(&session->s_con, msg);
3297
3298 out:
3299         iput(inode);
3300         mutex_unlock(&session->s_mutex);
3301         return;
3302
3303 bad:
3304         pr_err("corrupt lease message\n");
3305         ceph_msg_dump(msg);
3306 }
3307
3308 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3309                               struct inode *inode,
3310                               struct dentry *dentry, char action,
3311                               u32 seq)
3312 {
3313         struct ceph_msg *msg;
3314         struct ceph_mds_lease *lease;
3315         int len = sizeof(*lease) + sizeof(u32);
3316         int dnamelen = 0;
3317
3318         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3319              inode, dentry, ceph_lease_op_name(action), session->s_mds);
3320         dnamelen = dentry->d_name.len;
3321         len += dnamelen;
3322
3323         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3324         if (!msg)
3325                 return;
3326         lease = msg->front.iov_base;
3327         lease->action = action;
3328         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3329         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3330         lease->seq = cpu_to_le32(seq);
3331         put_unaligned_le32(dnamelen, lease + 1);
3332         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3333
3334         /*
3335          * if this is a preemptive lease RELEASE, no need to
3336          * flush request stream, since the actual request will
3337          * soon follow.
3338          */
3339         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3340
3341         ceph_con_send(&session->s_con, msg);
3342 }
3343
3344 /*
3345  * drop all leases (and dentry refs) in preparation for umount
3346  */
3347 static void drop_leases(struct ceph_mds_client *mdsc)
3348 {
3349         int i;
3350
3351         dout("drop_leases\n");
3352         mutex_lock(&mdsc->mutex);
3353         for (i = 0; i < mdsc->max_sessions; i++) {
3354                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3355                 if (!s)
3356                         continue;
3357                 mutex_unlock(&mdsc->mutex);
3358                 mutex_lock(&s->s_mutex);
3359                 mutex_unlock(&s->s_mutex);
3360                 ceph_put_mds_session(s);
3361                 mutex_lock(&mdsc->mutex);
3362         }
3363         mutex_unlock(&mdsc->mutex);
3364 }
3365
3366
3367
3368 /*
3369  * delayed work -- periodically trim expired leases, renew caps with mds
3370  */
3371 static void schedule_delayed(struct ceph_mds_client *mdsc)
3372 {
3373         int delay = 5;
3374         unsigned hz = round_jiffies_relative(HZ * delay);
3375         schedule_delayed_work(&mdsc->delayed_work, hz);
3376 }
3377
3378 static void delayed_work(struct work_struct *work)
3379 {
3380         int i;
3381         struct ceph_mds_client *mdsc =
3382                 container_of(work, struct ceph_mds_client, delayed_work.work);
3383         int renew_interval;
3384         int renew_caps;
3385
3386         dout("mdsc delayed_work\n");
3387         ceph_check_delayed_caps(mdsc);
3388
3389         if (mdsc->stopping)
3390                 return;
3391
3392         mutex_lock(&mdsc->mutex);
3393         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3394         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3395                                    mdsc->last_renew_caps);
3396         if (renew_caps)
3397                 mdsc->last_renew_caps = jiffies;
3398
3399         for (i = 0; i < mdsc->max_sessions; i++) {
3400                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3401                 if (s == NULL)
3402                         continue;
3403                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3404                         dout("resending session close request for mds%d\n",
3405                              s->s_mds);
3406                         request_close_session(mdsc, s);
3407                         ceph_put_mds_session(s);
3408                         continue;
3409                 }
3410                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3411                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3412                                 s->s_state = CEPH_MDS_SESSION_HUNG;
3413                                 pr_info("mds%d hung\n", s->s_mds);
3414                         }
3415                 }
3416                 if (s->s_state == CEPH_MDS_SESSION_NEW ||
3417                     s->s_state == CEPH_MDS_SESSION_RESTARTING ||
3418                     s->s_state == CEPH_MDS_SESSION_REJECTED) {
3419                         /* this mds is failed or recovering, just wait */
3420                         ceph_put_mds_session(s);
3421                         continue;
3422                 }
3423                 mutex_unlock(&mdsc->mutex);
3424
3425                 mutex_lock(&s->s_mutex);
3426                 if (renew_caps)
3427                         send_renew_caps(mdsc, s);
3428                 else
3429                         ceph_con_keepalive(&s->s_con);
3430                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3431                     s->s_state == CEPH_MDS_SESSION_HUNG)
3432                         ceph_send_cap_releases(mdsc, s);
3433                 mutex_unlock(&s->s_mutex);
3434                 ceph_put_mds_session(s);
3435
3436                 mutex_lock(&mdsc->mutex);
3437         }
3438         mutex_unlock(&mdsc->mutex);
3439
3440         schedule_delayed(mdsc);
3441 }
3442
3443 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3444
3445 {
3446         struct ceph_mds_client *mdsc;
3447
3448         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3449         if (!mdsc)
3450                 return -ENOMEM;
3451         mdsc->fsc = fsc;
3452         fsc->mdsc = mdsc;
3453         mutex_init(&mdsc->mutex);
3454         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3455         if (mdsc->mdsmap == NULL) {
3456                 kfree(mdsc);
3457                 return -ENOMEM;
3458         }
3459
3460         init_completion(&mdsc->safe_umount_waiters);
3461         init_waitqueue_head(&mdsc->session_close_wq);
3462         INIT_LIST_HEAD(&mdsc->waiting_for_map);
3463         mdsc->sessions = NULL;
3464         atomic_set(&mdsc->num_sessions, 0);
3465         mdsc->max_sessions = 0;
3466         mdsc->stopping = 0;
3467         mdsc->last_snap_seq = 0;
3468         init_rwsem(&mdsc->snap_rwsem);
3469         mdsc->snap_realms = RB_ROOT;
3470         INIT_LIST_HEAD(&mdsc->snap_empty);
3471         spin_lock_init(&mdsc->snap_empty_lock);
3472         mdsc->last_tid = 0;
3473         mdsc->oldest_tid = 0;
3474         mdsc->request_tree = RB_ROOT;
3475         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3476         mdsc->last_renew_caps = jiffies;
3477         INIT_LIST_HEAD(&mdsc->cap_delay_list);
3478         spin_lock_init(&mdsc->cap_delay_lock);
3479         INIT_LIST_HEAD(&mdsc->snap_flush_list);
3480         spin_lock_init(&mdsc->snap_flush_lock);
3481         mdsc->last_cap_flush_tid = 1;
3482         INIT_LIST_HEAD(&mdsc->cap_flush_list);
3483         INIT_LIST_HEAD(&mdsc->cap_dirty);
3484         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3485         mdsc->num_cap_flushing = 0;
3486         spin_lock_init(&mdsc->cap_dirty_lock);
3487         init_waitqueue_head(&mdsc->cap_flushing_wq);
3488         spin_lock_init(&mdsc->dentry_lru_lock);
3489         INIT_LIST_HEAD(&mdsc->dentry_lru);
3490
3491         ceph_caps_init(mdsc);
3492         ceph_adjust_min_caps(mdsc, fsc->min_caps);
3493
3494         init_rwsem(&mdsc->pool_perm_rwsem);
3495         mdsc->pool_perm_tree = RB_ROOT;
3496
3497         return 0;
3498 }
3499
3500 /*
3501  * Wait for safe replies on open mds requests.  If we time out, drop
3502  * all requests from the tree to avoid dangling dentry refs.
3503  */
3504 static void wait_requests(struct ceph_mds_client *mdsc)
3505 {
3506         struct ceph_options *opts = mdsc->fsc->client->options;
3507         struct ceph_mds_request *req;
3508
3509         mutex_lock(&mdsc->mutex);
3510         if (__get_oldest_req(mdsc)) {
3511                 mutex_unlock(&mdsc->mutex);
3512
3513                 dout("wait_requests waiting for requests\n");
3514                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3515                                     ceph_timeout_jiffies(opts->mount_timeout));
3516
3517                 /* tear down remaining requests */
3518                 mutex_lock(&mdsc->mutex);
3519                 while ((req = __get_oldest_req(mdsc))) {
3520                         dout("wait_requests timed out on tid %llu\n",
3521                              req->r_tid);
3522                         __unregister_request(mdsc, req);
3523                 }
3524         }
3525         mutex_unlock(&mdsc->mutex);
3526         dout("wait_requests done\n");
3527 }
3528
3529 /*
3530  * called before mount is ro, and before dentries are torn down.
3531  * (hmm, does this still race with new lookups?)
3532  */
3533 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3534 {
3535         dout("pre_umount\n");
3536         mdsc->stopping = 1;
3537
3538         drop_leases(mdsc);
3539         ceph_flush_dirty_caps(mdsc);
3540         wait_requests(mdsc);
3541
3542         /*
3543          * wait for reply handlers to drop their request refs and
3544          * their inode/dcache refs
3545          */
3546         ceph_msgr_flush();
3547 }
3548
3549 /*
3550  * wait for all write mds requests to flush.
3551  */
3552 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3553 {
3554         struct ceph_mds_request *req = NULL, *nextreq;
3555         struct rb_node *n;
3556
3557         mutex_lock(&mdsc->mutex);
3558         dout("wait_unsafe_requests want %lld\n", want_tid);
3559 restart:
3560         req = __get_oldest_req(mdsc);
3561         while (req && req->r_tid <= want_tid) {
3562                 /* find next request */
3563                 n = rb_next(&req->r_node);
3564                 if (n)
3565                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3566                 else
3567                         nextreq = NULL;
3568                 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3569                     (req->r_op & CEPH_MDS_OP_WRITE)) {
3570                         /* write op */
3571                         ceph_mdsc_get_request(req);
3572                         if (nextreq)
3573                                 ceph_mdsc_get_request(nextreq);
3574                         mutex_unlock(&mdsc->mutex);
3575                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
3576                              req->r_tid, want_tid);
3577                         wait_for_completion(&req->r_safe_completion);
3578                         mutex_lock(&mdsc->mutex);
3579                         ceph_mdsc_put_request(req);
3580                         if (!nextreq)
3581                                 break;  /* next dne before, so we're done! */
3582                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
3583                                 /* next request was removed from tree */
3584                                 ceph_mdsc_put_request(nextreq);
3585                                 goto restart;
3586                         }
3587                         ceph_mdsc_put_request(nextreq);  /* won't go away */
3588                 }
3589                 req = nextreq;
3590         }
3591         mutex_unlock(&mdsc->mutex);
3592         dout("wait_unsafe_requests done\n");
3593 }
3594
3595 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3596 {
3597         u64 want_tid, want_flush;
3598
3599         if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3600                 return;
3601
3602         dout("sync\n");
3603         mutex_lock(&mdsc->mutex);
3604         want_tid = mdsc->last_tid;
3605         mutex_unlock(&mdsc->mutex);
3606
3607         ceph_flush_dirty_caps(mdsc);
3608         spin_lock(&mdsc->cap_dirty_lock);
3609         want_flush = mdsc->last_cap_flush_tid;
3610         if (!list_empty(&mdsc->cap_flush_list)) {
3611                 struct ceph_cap_flush *cf =
3612                         list_last_entry(&mdsc->cap_flush_list,
3613                                         struct ceph_cap_flush, g_list);
3614                 cf->wake = true;
3615         }
3616         spin_unlock(&mdsc->cap_dirty_lock);
3617
3618         dout("sync want tid %lld flush_seq %lld\n",
3619              want_tid, want_flush);
3620
3621         wait_unsafe_requests(mdsc, want_tid);
3622         wait_caps_flush(mdsc, want_flush);
3623 }
3624
3625 /*
3626  * true if all sessions are closed, or we force unmount
3627  */
3628 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
3629 {
3630         if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3631                 return true;
3632         return atomic_read(&mdsc->num_sessions) <= skipped;
3633 }
3634
3635 /*
3636  * called after sb is ro.
3637  */
3638 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3639 {
3640         struct ceph_options *opts = mdsc->fsc->client->options;
3641         struct ceph_mds_session *session;
3642         int i;
3643         int skipped = 0;
3644
3645         dout("close_sessions\n");
3646
3647         /* close sessions */
3648         mutex_lock(&mdsc->mutex);
3649         for (i = 0; i < mdsc->max_sessions; i++) {
3650                 session = __ceph_lookup_mds_session(mdsc, i);
3651                 if (!session)
3652                         continue;
3653                 mutex_unlock(&mdsc->mutex);
3654                 mutex_lock(&session->s_mutex);
3655                 if (__close_session(mdsc, session) <= 0)
3656                         skipped++;
3657                 mutex_unlock(&session->s_mutex);
3658                 ceph_put_mds_session(session);
3659                 mutex_lock(&mdsc->mutex);
3660         }
3661         mutex_unlock(&mdsc->mutex);
3662
3663         dout("waiting for sessions to close\n");
3664         wait_event_timeout(mdsc->session_close_wq,
3665                            done_closing_sessions(mdsc, skipped),
3666                            ceph_timeout_jiffies(opts->mount_timeout));
3667
3668         /* tear down remaining sessions */
3669         mutex_lock(&mdsc->mutex);
3670         for (i = 0; i < mdsc->max_sessions; i++) {
3671                 if (mdsc->sessions[i]) {
3672                         session = get_session(mdsc->sessions[i]);
3673                         __unregister_session(mdsc, session);
3674                         mutex_unlock(&mdsc->mutex);
3675                         mutex_lock(&session->s_mutex);
3676                         remove_session_caps(session);
3677                         mutex_unlock(&session->s_mutex);
3678                         ceph_put_mds_session(session);
3679                         mutex_lock(&mdsc->mutex);
3680                 }
3681         }
3682         WARN_ON(!list_empty(&mdsc->cap_delay_list));
3683         mutex_unlock(&mdsc->mutex);
3684
3685         ceph_cleanup_empty_realms(mdsc);
3686
3687         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3688
3689         dout("stopped\n");
3690 }
3691
3692 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
3693 {
3694         struct ceph_mds_session *session;
3695         int mds;
3696
3697         dout("force umount\n");
3698
3699         mutex_lock(&mdsc->mutex);
3700         for (mds = 0; mds < mdsc->max_sessions; mds++) {
3701                 session = __ceph_lookup_mds_session(mdsc, mds);
3702                 if (!session)
3703                         continue;
3704                 mutex_unlock(&mdsc->mutex);
3705                 mutex_lock(&session->s_mutex);
3706                 __close_session(mdsc, session);
3707                 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
3708                         cleanup_session_requests(mdsc, session);
3709                         remove_session_caps(session);
3710                 }
3711                 mutex_unlock(&session->s_mutex);
3712                 ceph_put_mds_session(session);
3713                 mutex_lock(&mdsc->mutex);
3714                 kick_requests(mdsc, mds);
3715         }
3716         __wake_requests(mdsc, &mdsc->waiting_for_map);
3717         mutex_unlock(&mdsc->mutex);
3718 }
3719
3720 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3721 {
3722         dout("stop\n");
3723         /*
3724          * Make sure the delayed work stopped before releasing
3725          * the resources.
3726          *
3727          * Because the cancel_delayed_work_sync() will only
3728          * guarantee that the work finishes executing. But the
3729          * delayed work will re-arm itself again after that.
3730          */
3731         flush_delayed_work(&mdsc->delayed_work);
3732
3733         if (mdsc->mdsmap)
3734                 ceph_mdsmap_destroy(mdsc->mdsmap);
3735         kfree(mdsc->sessions);
3736         ceph_caps_finalize(mdsc);
3737         ceph_pool_perm_destroy(mdsc);
3738 }
3739
3740 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3741 {
3742         struct ceph_mds_client *mdsc = fsc->mdsc;
3743
3744         dout("mdsc_destroy %p\n", mdsc);
3745         ceph_mdsc_stop(mdsc);
3746
3747         /* flush out any connection work with references to us */
3748         ceph_msgr_flush();
3749
3750         fsc->mdsc = NULL;
3751         kfree(mdsc);
3752         dout("mdsc_destroy %p done\n", mdsc);
3753 }
3754
3755 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3756 {
3757         struct ceph_fs_client *fsc = mdsc->fsc;
3758         const char *mds_namespace = fsc->mount_options->mds_namespace;
3759         void *p = msg->front.iov_base;
3760         void *end = p + msg->front.iov_len;
3761         u32 epoch;
3762         u32 map_len;
3763         u32 num_fs;
3764         u32 mount_fscid = (u32)-1;
3765         u8 struct_v, struct_cv;
3766         int err = -EINVAL;
3767
3768         ceph_decode_need(&p, end, sizeof(u32), bad);
3769         epoch = ceph_decode_32(&p);
3770
3771         dout("handle_fsmap epoch %u\n", epoch);
3772
3773         ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3774         struct_v = ceph_decode_8(&p);
3775         struct_cv = ceph_decode_8(&p);
3776         map_len = ceph_decode_32(&p);
3777
3778         ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
3779         p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
3780
3781         num_fs = ceph_decode_32(&p);
3782         while (num_fs-- > 0) {
3783                 void *info_p, *info_end;
3784                 u32 info_len;
3785                 u8 info_v, info_cv;
3786                 u32 fscid, namelen;
3787
3788                 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3789                 info_v = ceph_decode_8(&p);
3790                 info_cv = ceph_decode_8(&p);
3791                 info_len = ceph_decode_32(&p);
3792                 ceph_decode_need(&p, end, info_len, bad);
3793                 info_p = p;
3794                 info_end = p + info_len;
3795                 p = info_end;
3796
3797                 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
3798                 fscid = ceph_decode_32(&info_p);
3799                 namelen = ceph_decode_32(&info_p);
3800                 ceph_decode_need(&info_p, info_end, namelen, bad);
3801
3802                 if (mds_namespace &&
3803                     strlen(mds_namespace) == namelen &&
3804                     !strncmp(mds_namespace, (char *)info_p, namelen)) {
3805                         mount_fscid = fscid;
3806                         break;
3807                 }
3808         }
3809
3810         ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
3811         if (mount_fscid != (u32)-1) {
3812                 fsc->client->monc.fs_cluster_id = mount_fscid;
3813                 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
3814                                    0, true);
3815                 ceph_monc_renew_subs(&fsc->client->monc);
3816         } else {
3817                 err = -ENOENT;
3818                 goto err_out;
3819         }
3820         return;
3821 bad:
3822         pr_err("error decoding fsmap\n");
3823 err_out:
3824         mutex_lock(&mdsc->mutex);
3825         mdsc->mdsmap_err = -ENOENT;
3826         __wake_requests(mdsc, &mdsc->waiting_for_map);
3827         mutex_unlock(&mdsc->mutex);
3828         return;
3829 }
3830
3831 /*
3832  * handle mds map update.
3833  */
3834 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3835 {
3836         u32 epoch;
3837         u32 maplen;
3838         void *p = msg->front.iov_base;
3839         void *end = p + msg->front.iov_len;
3840         struct ceph_mdsmap *newmap, *oldmap;
3841         struct ceph_fsid fsid;
3842         int err = -EINVAL;
3843
3844         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3845         ceph_decode_copy(&p, &fsid, sizeof(fsid));
3846         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3847                 return;
3848         epoch = ceph_decode_32(&p);
3849         maplen = ceph_decode_32(&p);
3850         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3851
3852         /* do we need it? */
3853         mutex_lock(&mdsc->mutex);
3854         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3855                 dout("handle_map epoch %u <= our %u\n",
3856                      epoch, mdsc->mdsmap->m_epoch);
3857                 mutex_unlock(&mdsc->mutex);
3858                 return;
3859         }
3860
3861         newmap = ceph_mdsmap_decode(&p, end);
3862         if (IS_ERR(newmap)) {
3863                 err = PTR_ERR(newmap);
3864                 goto bad_unlock;
3865         }
3866
3867         /* swap into place */
3868         if (mdsc->mdsmap) {
3869                 oldmap = mdsc->mdsmap;
3870                 mdsc->mdsmap = newmap;
3871                 check_new_map(mdsc, newmap, oldmap);
3872                 ceph_mdsmap_destroy(oldmap);
3873         } else {
3874                 mdsc->mdsmap = newmap;  /* first mds map */
3875         }
3876         mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3877
3878         __wake_requests(mdsc, &mdsc->waiting_for_map);
3879         ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
3880                           mdsc->mdsmap->m_epoch);
3881
3882         mutex_unlock(&mdsc->mutex);
3883         schedule_delayed(mdsc);
3884         return;
3885
3886 bad_unlock:
3887         mutex_unlock(&mdsc->mutex);
3888 bad:
3889         pr_err("error decoding mdsmap %d\n", err);
3890         return;
3891 }
3892
3893 static struct ceph_connection *con_get(struct ceph_connection *con)
3894 {
3895         struct ceph_mds_session *s = con->private;
3896
3897         if (get_session(s)) {
3898                 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3899                 return con;
3900         }
3901         dout("mdsc con_get %p FAIL\n", s);
3902         return NULL;
3903 }
3904
3905 static void con_put(struct ceph_connection *con)
3906 {
3907         struct ceph_mds_session *s = con->private;
3908
3909         dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3910         ceph_put_mds_session(s);
3911 }
3912
3913 /*
3914  * if the client is unresponsive for long enough, the mds will kill
3915  * the session entirely.
3916  */
3917 static void peer_reset(struct ceph_connection *con)
3918 {
3919         struct ceph_mds_session *s = con->private;
3920         struct ceph_mds_client *mdsc = s->s_mdsc;
3921
3922         pr_warn("mds%d closed our session\n", s->s_mds);
3923         send_mds_reconnect(mdsc, s);
3924 }
3925
3926 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3927 {
3928         struct ceph_mds_session *s = con->private;
3929         struct ceph_mds_client *mdsc = s->s_mdsc;
3930         int type = le16_to_cpu(msg->hdr.type);
3931
3932         mutex_lock(&mdsc->mutex);
3933         if (__verify_registered_session(mdsc, s) < 0) {
3934                 mutex_unlock(&mdsc->mutex);
3935                 goto out;
3936         }
3937         mutex_unlock(&mdsc->mutex);
3938
3939         switch (type) {
3940         case CEPH_MSG_MDS_MAP:
3941                 ceph_mdsc_handle_mdsmap(mdsc, msg);
3942                 break;
3943         case CEPH_MSG_FS_MAP_USER:
3944                 ceph_mdsc_handle_fsmap(mdsc, msg);
3945                 break;
3946         case CEPH_MSG_CLIENT_SESSION:
3947                 handle_session(s, msg);
3948                 break;
3949         case CEPH_MSG_CLIENT_REPLY:
3950                 handle_reply(s, msg);
3951                 break;
3952         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3953                 handle_forward(mdsc, s, msg);
3954                 break;
3955         case CEPH_MSG_CLIENT_CAPS:
3956                 ceph_handle_caps(s, msg);
3957                 break;
3958         case CEPH_MSG_CLIENT_SNAP:
3959                 ceph_handle_snap(mdsc, s, msg);
3960                 break;
3961         case CEPH_MSG_CLIENT_LEASE:
3962                 handle_lease(mdsc, s, msg);
3963                 break;
3964
3965         default:
3966                 pr_err("received unknown message type %d %s\n", type,
3967                        ceph_msg_type_name(type));
3968         }
3969 out:
3970         ceph_msg_put(msg);
3971 }
3972
3973 /*
3974  * authentication
3975  */
3976
3977 /*
3978  * Note: returned pointer is the address of a structure that's
3979  * managed separately.  Caller must *not* attempt to free it.
3980  */
3981 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
3982                                         int *proto, int force_new)
3983 {
3984         struct ceph_mds_session *s = con->private;
3985         struct ceph_mds_client *mdsc = s->s_mdsc;
3986         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3987         struct ceph_auth_handshake *auth = &s->s_auth;
3988
3989         if (force_new && auth->authorizer) {
3990                 ceph_auth_destroy_authorizer(auth->authorizer);
3991                 auth->authorizer = NULL;
3992         }
3993         if (!auth->authorizer) {
3994                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3995                                                       auth);
3996                 if (ret)
3997                         return ERR_PTR(ret);
3998         } else {
3999                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4000                                                       auth);
4001                 if (ret)
4002                         return ERR_PTR(ret);
4003         }
4004         *proto = ac->protocol;
4005
4006         return auth;
4007 }
4008
4009 static int add_authorizer_challenge(struct ceph_connection *con,
4010                                     void *challenge_buf, int challenge_buf_len)
4011 {
4012         struct ceph_mds_session *s = con->private;
4013         struct ceph_mds_client *mdsc = s->s_mdsc;
4014         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4015
4016         return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
4017                                             challenge_buf, challenge_buf_len);
4018 }
4019
4020 static int verify_authorizer_reply(struct ceph_connection *con)
4021 {
4022         struct ceph_mds_session *s = con->private;
4023         struct ceph_mds_client *mdsc = s->s_mdsc;
4024         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4025
4026         return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
4027 }
4028
4029 static int invalidate_authorizer(struct ceph_connection *con)
4030 {
4031         struct ceph_mds_session *s = con->private;
4032         struct ceph_mds_client *mdsc = s->s_mdsc;
4033         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4034
4035         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
4036
4037         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
4038 }
4039
4040 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
4041                                 struct ceph_msg_header *hdr, int *skip)
4042 {
4043         struct ceph_msg *msg;
4044         int type = (int) le16_to_cpu(hdr->type);
4045         int front_len = (int) le32_to_cpu(hdr->front_len);
4046
4047         if (con->in_msg)
4048                 return con->in_msg;
4049
4050         *skip = 0;
4051         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
4052         if (!msg) {
4053                 pr_err("unable to allocate msg type %d len %d\n",
4054                        type, front_len);
4055                 return NULL;
4056         }
4057
4058         return msg;
4059 }
4060
4061 static int mds_sign_message(struct ceph_msg *msg)
4062 {
4063        struct ceph_mds_session *s = msg->con->private;
4064        struct ceph_auth_handshake *auth = &s->s_auth;
4065
4066        return ceph_auth_sign_message(auth, msg);
4067 }
4068
4069 static int mds_check_message_signature(struct ceph_msg *msg)
4070 {
4071        struct ceph_mds_session *s = msg->con->private;
4072        struct ceph_auth_handshake *auth = &s->s_auth;
4073
4074        return ceph_auth_check_message_signature(auth, msg);
4075 }
4076
4077 static const struct ceph_connection_operations mds_con_ops = {
4078         .get = con_get,
4079         .put = con_put,
4080         .dispatch = dispatch,
4081         .get_authorizer = get_authorizer,
4082         .add_authorizer_challenge = add_authorizer_challenge,
4083         .verify_authorizer_reply = verify_authorizer_reply,
4084         .invalidate_authorizer = invalidate_authorizer,
4085         .peer_reset = peer_reset,
4086         .alloc_msg = mds_alloc_msg,
4087         .sign_message = mds_sign_message,
4088         .check_message_signature = mds_check_message_signature,
4089 };
4090
4091 /* eof */