fs/ceph/mds_client.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/ceph/ceph_debug.h>
   3
   4 #include <linux/fs.h>
   5 #include <linux/wait.h>
   6 #include <linux/slab.h>
   7 #include <linux/gfp.h>
   8 #include <linux/sched.h>
   9 #include <linux/debugfs.h>
  10 #include <linux/seq_file.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  45  * are no longer valid.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                 info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
 193         }
 194         if (num == 0)
 195                 goto done;
 196
 197         BUG_ON(!info->dir_entries);
 198         if ((unsigned long)(info->dir_entries + num) >
 199             (unsigned long)info->dir_entries + info->dir_buf_size) {
 200                 pr_err("dir contents are larger than expected\n");
 201                 WARN_ON(1);
 202                 goto bad;
 203         }
 204
 205         info->dir_nr = num;
 206         while (num) {
 207                 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
 208                 /* dentry */
 209                 ceph_decode_need(p, end, sizeof(u32)*2, bad);
 210                 rde->name_len = ceph_decode_32(p);
 211                 ceph_decode_need(p, end, rde->name_len, bad);
 212                 rde->name = *p;
 213                 *p += rde->name_len;
 214                 dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
 215                 rde->lease = *p;
 216                 *p += sizeof(struct ceph_mds_reply_lease);
 217
 218                 /* inode */
 219                 err = parse_reply_info_in(p, end, &rde->inode, features);
 220                 if (err < 0)
 221                         goto out_bad;
 222                 /* ceph_readdir_prepopulate() will update it */
 223                 rde->offset = 0;
 224                 i++;
 225                 num--;
 226         }
 227
 228 done:
 229         if (*p != end)
 230                 goto bad;
 231         return 0;
 232
 233 bad:
 234         err = -EIO;
 235 out_bad:
 236         pr_err("problem parsing dir contents %d\n", err);
 237         return err;
 238 }
 239
 240 /*
 241  * parse fcntl F_GETLK results
 242  */
 243 static int parse_reply_info_filelock(void **p, void *end,
 244                                      struct ceph_mds_reply_info_parsed *info,
 245                                      u64 features)
 246 {
 247         if (*p + sizeof(*info->filelock_reply) > end)
 248                 goto bad;
 249
 250         info->filelock_reply = *p;
 251         *p += sizeof(*info->filelock_reply);
 252
 253         if (unlikely(*p != end))
 254                 goto bad;
 255         return 0;
 256
 257 bad:
 258         return -EIO;
 259 }
 260
 261 /*
 262  * parse create results
 263  */
 264 static int parse_reply_info_create(void **p, void *end,
 265                                   struct ceph_mds_reply_info_parsed *info,
 266                                   u64 features)
 267 {
 268         if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
 269                 if (*p == end) {
 270                         info->has_create_ino = false;
 271                 } else {
 272                         info->has_create_ino = true;
 273                         info->ino = ceph_decode_64(p);
 274                 }
 275         }
 276
 277         if (unlikely(*p != end))
 278                 goto bad;
 279         return 0;
 280
 281 bad:
 282         return -EIO;
 283 }
 284
 285 /*
 286  * parse extra results
 287  */
 288 static int parse_reply_info_extra(void **p, void *end,
 289                                   struct ceph_mds_reply_info_parsed *info,
 290                                   u64 features)
 291 {
 292         u32 op = le32_to_cpu(info->head->op);
 293
 294         if (op == CEPH_MDS_OP_GETFILELOCK)
 295                 return parse_reply_info_filelock(p, end, info, features);
 296         else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
 297                 return parse_reply_info_dir(p, end, info, features);
 298         else if (op == CEPH_MDS_OP_CREATE)
 299                 return parse_reply_info_create(p, end, info, features);
 300         else
 301                 return -EIO;
 302 }
 303
 304 /*
 305  * parse entire mds reply
 306  */
 307 static int parse_reply_info(struct ceph_msg *msg,
 308                             struct ceph_mds_reply_info_parsed *info,
 309                             u64 features)
 310 {
 311         void *p, *end;
 312         u32 len;
 313         int err;
 314
 315         info->head = msg->front.iov_base;
 316         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 317         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 318
 319         /* trace */
 320         ceph_decode_32_safe(&p, end, len, bad);
 321         if (len > 0) {
 322                 ceph_decode_need(&p, end, len, bad);
 323                 err = parse_reply_info_trace(&p, p+len, info, features);
 324                 if (err < 0)
 325                         goto out_bad;
 326         }
 327
 328         /* extra */
 329         ceph_decode_32_safe(&p, end, len, bad);
 330         if (len > 0) {
 331                 ceph_decode_need(&p, end, len, bad);
 332                 err = parse_reply_info_extra(&p, p+len, info, features);
 333                 if (err < 0)
 334                         goto out_bad;
 335         }
 336
 337         /* snap blob */
 338         ceph_decode_32_safe(&p, end, len, bad);
 339         info->snapblob_len = len;
 340         info->snapblob = p;
 341         p += len;
 342
 343         if (p != end)
 344                 goto bad;
 345         return 0;
 346
 347 bad:
 348         err = -EIO;
 349 out_bad:
 350         pr_err("mds parse_reply err %d\n", err);
 351         return err;
 352 }
 353
 354 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
 355 {
 356         if (!info->dir_entries)
 357                 return;
 358         free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
 359 }
 360
 361
 362 /*
 363  * sessions
 364  */
 365 const char *ceph_session_state_name(int s)
 366 {
 367         switch (s) {
 368         case CEPH_MDS_SESSION_NEW: return "new";
 369         case CEPH_MDS_SESSION_OPENING: return "opening";
 370         case CEPH_MDS_SESSION_OPEN: return "open";
 371         case CEPH_MDS_SESSION_HUNG: return "hung";
 372         case CEPH_MDS_SESSION_CLOSING: return "closing";
 373         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
 374         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
 375         case CEPH_MDS_SESSION_REJECTED: return "rejected";
 376         default: return "???";
 377         }
 378 }
 379
 380 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
 381 {
 382         if (refcount_inc_not_zero(&s->s_ref)) {
 383                 dout("mdsc get_session %p %d -> %d\n", s,
 384                      refcount_read(&s->s_ref)-1, refcount_read(&s->s_ref));
 385                 return s;
 386         } else {
 387                 dout("mdsc get_session %p 0 -- FAIL", s);
 388                 return NULL;
 389         }
 390 }
 391
 392 void ceph_put_mds_session(struct ceph_mds_session *s)
 393 {
 394         dout("mdsc put_session %p %d -> %d\n", s,
 395              refcount_read(&s->s_ref), refcount_read(&s->s_ref)-1);
 396         if (refcount_dec_and_test(&s->s_ref)) {
 397                 if (s->s_auth.authorizer)
 398                         ceph_auth_destroy_authorizer(s->s_auth.authorizer);
 399                 kfree(s);
 400         }
 401 }
 402
 403 /*
 404  * called under mdsc->mutex
 405  */
 406 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
 407                                                    int mds)
 408 {
 409         struct ceph_mds_session *session;
 410
 411         if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
 412                 return NULL;
 413         session = mdsc->sessions[mds];
 414         dout("lookup_mds_session %p %d\n", session,
 415              refcount_read(&session->s_ref));
 416         get_session(session);
 417         return session;
 418 }
 419
 420 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
 421 {
 422         if (mds >= mdsc->max_sessions)
 423                 return false;
 424         return mdsc->sessions[mds];
 425 }
 426
 427 static int __verify_registered_session(struct ceph_mds_client *mdsc,
 428                                        struct ceph_mds_session *s)
 429 {
 430         if (s->s_mds >= mdsc->max_sessions ||
 431             mdsc->sessions[s->s_mds] != s)
 432                 return -ENOENT;
 433         return 0;
 434 }
 435
 436 /*
 437  * create+register a new session for given mds.
 438  * called under mdsc->mutex.
 439  */
 440 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 441                                                  int mds)
 442 {
 443         struct ceph_mds_session *s;
 444
 445         if (mds >= mdsc->mdsmap->m_num_mds)
 446                 return ERR_PTR(-EINVAL);
 447
 448         s = kzalloc(sizeof(*s), GFP_NOFS);
 449         if (!s)
 450                 return ERR_PTR(-ENOMEM);
 451         s->s_mdsc = mdsc;
 452         s->s_mds = mds;
 453         s->s_state = CEPH_MDS_SESSION_NEW;
 454         s->s_ttl = 0;
 455         s->s_seq = 0;
 456         mutex_init(&s->s_mutex);
 457
 458         ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
 459
 460         spin_lock_init(&s->s_gen_ttl_lock);
 461         s->s_cap_gen = 0;
 462         s->s_cap_ttl = jiffies - 1;
 463
 464         spin_lock_init(&s->s_cap_lock);
 465         s->s_renew_requested = 0;
 466         s->s_renew_seq = 0;
 467         INIT_LIST_HEAD(&s->s_caps);
 468         s->s_nr_caps = 0;
 469         s->s_trim_caps = 0;
 470         refcount_set(&s->s_ref, 1);
 471         INIT_LIST_HEAD(&s->s_waiting);
 472         INIT_LIST_HEAD(&s->s_unsafe);
 473         s->s_num_cap_releases = 0;
 474         s->s_cap_reconnect = 0;
 475         s->s_cap_iterator = NULL;
 476         INIT_LIST_HEAD(&s->s_cap_releases);
 477         INIT_LIST_HEAD(&s->s_cap_flushing);
 478
 479         dout("register_session mds%d\n", mds);
 480         if (mds >= mdsc->max_sessions) {
 481                 int newmax = 1 << get_count_order(mds+1);
 482                 struct ceph_mds_session **sa;
 483
 484                 dout("register_session realloc to %d\n", newmax);
 485                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 486                 if (!sa)
 487                         goto fail_realloc;
 488                 if (mdsc->sessions) {
 489                         memcpy(sa, mdsc->sessions,
 490                                mdsc->max_sessions * sizeof(void *));
 491                         kfree(mdsc->sessions);
 492                 }
 493                 mdsc->sessions = sa;
 494                 mdsc->max_sessions = newmax;
 495         }
 496         mdsc->sessions[mds] = s;
 497         atomic_inc(&mdsc->num_sessions);
 498         refcount_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
 499
 500         ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
 501                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
 502
 503         return s;
 504
 505 fail_realloc:
 506         kfree(s);
 507         return ERR_PTR(-ENOMEM);
 508 }
 509
 510 /*
 511  * called under mdsc->mutex
 512  */
 513 static void __unregister_session(struct ceph_mds_client *mdsc,
 514                                struct ceph_mds_session *s)
 515 {
 516         dout("__unregister_session mds%d %p\n", s->s_mds, s);
 517         BUG_ON(mdsc->sessions[s->s_mds] != s);
 518         mdsc->sessions[s->s_mds] = NULL;
 519         ceph_con_close(&s->s_con);
 520         ceph_put_mds_session(s);
 521         atomic_dec(&mdsc->num_sessions);
 522 }
 523
 524 /*
 525  * drop session refs in request.
 526  *
 527  * should be last request ref, or hold mdsc->mutex
 528  */
 529 static void put_request_session(struct ceph_mds_request *req)
 530 {
 531         if (req->r_session) {
 532                 ceph_put_mds_session(req->r_session);
 533                 req->r_session = NULL;
 534         }
 535 }
 536
 537 void ceph_mdsc_release_request(struct kref *kref)
 538 {
 539         struct ceph_mds_request *req = container_of(kref,
 540                                                     struct ceph_mds_request,
 541                                                     r_kref);
 542         destroy_reply_info(&req->r_reply_info);
 543         if (req->r_request)
 544                 ceph_msg_put(req->r_request);
 545         if (req->r_reply)
 546                 ceph_msg_put(req->r_reply);
 547         if (req->r_inode) {
 548                 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
 549                 iput(req->r_inode);
 550         }
 551         if (req->r_parent)
 552                 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
 553         iput(req->r_target_inode);
 554         if (req->r_dentry)
 555                 dput(req->r_dentry);
 556         if (req->r_old_dentry)
 557                 dput(req->r_old_dentry);
 558         if (req->r_old_dentry_dir) {
 559                 /*
 560                  * track (and drop pins for) r_old_dentry_dir
 561                  * separately, since r_old_dentry's d_parent may have
 562                  * changed between the dir mutex being dropped and
 563                  * this request being freed.
 564                  */
 565                 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
 566                                   CEPH_CAP_PIN);
 567                 iput(req->r_old_dentry_dir);
 568         }
 569         kfree(req->r_path1);
 570         kfree(req->r_path2);
 571         if (req->r_pagelist)
 572                 ceph_pagelist_release(req->r_pagelist);
 573         put_request_session(req);
 574         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
 575         kfree(req);
 576 }
 577
 578 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
 579
 580 /*
 581  * lookup session, bump ref if found.
 582  *
 583  * called under mdsc->mutex.
 584  */
 585 static struct ceph_mds_request *
 586 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
 587 {
 588         struct ceph_mds_request *req;
 589
 590         req = lookup_request(&mdsc->request_tree, tid);
 591         if (req)
 592                 ceph_mdsc_get_request(req);
 593
 594         return req;
 595 }
 596
 597 /*
 598  * Register an in-flight request, and assign a tid.  Link to directory
 599  * are modifying (if any).
 600  *
 601  * Called under mdsc->mutex.
 602  */
 603 static void __register_request(struct ceph_mds_client *mdsc,
 604                                struct ceph_mds_request *req,
 605                                struct inode *dir)
 606 {
 607         req->r_tid = ++mdsc->last_tid;
 608         if (req->r_num_caps)
 609                 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
 610                                   req->r_num_caps);
 611         dout("__register_request %p tid %lld\n", req, req->r_tid);
 612         ceph_mdsc_get_request(req);
 613         insert_request(&mdsc->request_tree, req);
 614
 615         req->r_uid = current_fsuid();
 616         req->r_gid = current_fsgid();
 617
 618         if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
 619                 mdsc->oldest_tid = req->r_tid;
 620
 621         if (dir) {
 622                 ihold(dir);
 623                 req->r_unsafe_dir = dir;
 624         }
 625 }
 626
 627 static void __unregister_request(struct ceph_mds_client *mdsc,
 628                                  struct ceph_mds_request *req)
 629 {
 630         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 631
 632         /* Never leave an unregistered request on an unsafe list! */
 633         list_del_init(&req->r_unsafe_item);
 634
 635         if (req->r_tid == mdsc->oldest_tid) {
 636                 struct rb_node *p = rb_next(&req->r_node);
 637                 mdsc->oldest_tid = 0;
 638                 while (p) {
 639                         struct ceph_mds_request *next_req =
 640                                 rb_entry(p, struct ceph_mds_request, r_node);
 641                         if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
 642                                 mdsc->oldest_tid = next_req->r_tid;
 643                                 break;
 644                         }
 645                         p = rb_next(p);
 646                 }
 647         }
 648
 649         erase_request(&mdsc->request_tree, req);
 650
 651         if (req->r_unsafe_dir  &&
 652             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
 653                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 654                 spin_lock(&ci->i_unsafe_lock);
 655                 list_del_init(&req->r_unsafe_dir_item);
 656                 spin_unlock(&ci->i_unsafe_lock);
 657         }
 658         if (req->r_target_inode &&
 659             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
 660                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
 661                 spin_lock(&ci->i_unsafe_lock);
 662                 list_del_init(&req->r_unsafe_target_item);
 663                 spin_unlock(&ci->i_unsafe_lock);
 664         }
 665
 666         if (req->r_unsafe_dir) {
 667                 iput(req->r_unsafe_dir);
 668                 req->r_unsafe_dir = NULL;
 669         }
 670
 671         complete_all(&req->r_safe_completion);
 672
 673         ceph_mdsc_put_request(req);
 674 }
 675
 676 /*
 677  * Walk back up the dentry tree until we hit a dentry representing a
 678  * non-snapshot inode. We do this using the rcu_read_lock (which must be held
 679  * when calling this) to ensure that the objects won't disappear while we're
 680  * working with them. Once we hit a candidate dentry, we attempt to take a
 681  * reference to it, and return that as the result.
 682  */
 683 static struct inode *get_nonsnap_parent(struct dentry *dentry)
 684 {
 685         struct inode *inode = NULL;
 686
 687         while (dentry && !IS_ROOT(dentry)) {
 688                 inode = d_inode_rcu(dentry);
 689                 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
 690                         break;
 691                 dentry = dentry->d_parent;
 692         }
 693         if (inode)
 694                 inode = igrab(inode);
 695         return inode;
 696 }
 697
 698 /*
 699  * Choose mds to send request to next.  If there is a hint set in the
 700  * request (e.g., due to a prior forward hint from the mds), use that.
 701  * Otherwise, consult frag tree and/or caps to identify the
 702  * appropriate mds.  If all else fails, choose randomly.
 703  *
 704  * Called under mdsc->mutex.
 705  */
 706 static int __choose_mds(struct ceph_mds_client *mdsc,
 707                         struct ceph_mds_request *req)
 708 {
 709         struct inode *inode;
 710         struct ceph_inode_info *ci;
 711         struct ceph_cap *cap;
 712         int mode = req->r_direct_mode;
 713         int mds = -1;
 714         u32 hash = req->r_direct_hash;
 715         bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
 716
 717         /*
 718          * is there a specific mds we should try?  ignore hint if we have
 719          * no session and the mds is not up (active or recovering).
 720          */
 721         if (req->r_resend_mds >= 0 &&
 722             (__have_session(mdsc, req->r_resend_mds) ||
 723              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 724                 dout("choose_mds using resend_mds mds%d\n",
 725                      req->r_resend_mds);
 726                 return req->r_resend_mds;
 727         }
 728
 729         if (mode == USE_RANDOM_MDS)
 730                 goto random;
 731
 732         inode = NULL;
 733         if (req->r_inode) {
 734                 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
 735                         inode = req->r_inode;
 736                         ihold(inode);
 737                 } else {
 738                         /* req->r_dentry is non-null for LSSNAP request */
 739                         rcu_read_lock();
 740                         inode = get_nonsnap_parent(req->r_dentry);
 741                         rcu_read_unlock();
 742                         dout("__choose_mds using snapdir's parent %p\n", inode);
 743                 }
 744         } else if (req->r_dentry) {
 745                 /* ignore race with rename; old or new d_parent is okay */
 746                 struct dentry *parent;
 747                 struct inode *dir;
 748
 749                 rcu_read_lock();
 750                 parent = req->r_dentry->d_parent;
 751                 dir = req->r_parent ? : d_inode_rcu(parent);
 752
 753                 if (!dir || dir->i_sb != mdsc->fsc->sb) {
 754                         /*  not this fs or parent went negative */
 755                         inode = d_inode(req->r_dentry);
 756                         if (inode)
 757                                 ihold(inode);
 758                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
 759                         /* direct snapped/virtual snapdir requests
 760                          * based on parent dir inode */
 761                         inode = get_nonsnap_parent(parent);
 762                         dout("__choose_mds using nonsnap parent %p\n", inode);
 763                 } else {
 764                         /* dentry target */
 765                         inode = d_inode(req->r_dentry);
 766                         if (!inode || mode == USE_AUTH_MDS) {
 767                                 /* dir + name */
 768                                 inode = igrab(dir);
 769                                 hash = ceph_dentry_hash(dir, req->r_dentry);
 770                                 is_hash = true;
 771                         } else {
 772                                 ihold(inode);
 773                         }
 774                 }
 775                 rcu_read_unlock();
 776         }
 777
 778         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 779              (int)hash, mode);
 780         if (!inode)
 781                 goto random;
 782         ci = ceph_inode(inode);
 783
 784         if (is_hash && S_ISDIR(inode->i_mode)) {
 785                 struct ceph_inode_frag frag;
 786                 int found;
 787
 788                 ceph_choose_frag(ci, hash, &frag, &found);
 789                 if (found) {
 790                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 791                                 u8 r;
 792
 793                                 /* choose a random replica */
 794                                 get_random_bytes(&r, 1);
 795                                 r %= frag.ndist;
 796                                 mds = frag.dist[r];
 797                                 dout("choose_mds %p %llx.%llx "
 798                                      "frag %u mds%d (%d/%d)\n",
 799                                      inode, ceph_vinop(inode),
 800                                      frag.frag, mds,
 801                                      (int)r, frag.ndist);
 802                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 803                                     CEPH_MDS_STATE_ACTIVE)
 804                                         goto out;
 805                         }
 806
 807                         /* since this file/dir wasn't known to be
 808                          * replicated, then we want to look for the
 809                          * authoritative mds. */
 810                         mode = USE_AUTH_MDS;
 811                         if (frag.mds >= 0) {
 812                                 /* choose auth mds */
 813                                 mds = frag.mds;
 814                                 dout("choose_mds %p %llx.%llx "
 815                                      "frag %u mds%d (auth)\n",
 816                                      inode, ceph_vinop(inode), frag.frag, mds);
 817                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 818                                     CEPH_MDS_STATE_ACTIVE)
 819                                         goto out;
 820                         }
 821                 }
 822         }
 823
 824         spin_lock(&ci->i_ceph_lock);
 825         cap = NULL;
 826         if (mode == USE_AUTH_MDS)
 827                 cap = ci->i_auth_cap;
 828         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
 829                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
 830         if (!cap) {
 831                 spin_unlock(&ci->i_ceph_lock);
 832                 iput(inode);
 833                 goto random;
 834         }
 835         mds = cap->session->s_mds;
 836         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
 837              inode, ceph_vinop(inode), mds,
 838              cap == ci->i_auth_cap ? "auth " : "", cap);
 839         spin_unlock(&ci->i_ceph_lock);
 840 out:
 841         iput(inode);
 842         return mds;
 843
 844 random:
 845         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
 846         dout("choose_mds chose random mds%d\n", mds);
 847         return mds;
 848 }
 849
 850
 851 /*
 852  * session messages
 853  */
 854 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
 855 {
 856         struct ceph_msg *msg;
 857         struct ceph_mds_session_head *h;
 858
 859         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
 860                            false);
 861         if (!msg) {
 862                 pr_err("create_session_msg ENOMEM creating msg\n");
 863                 return NULL;
 864         }
 865         h = msg->front.iov_base;
 866         h->op = cpu_to_le32(op);
 867         h->seq = cpu_to_le64(seq);
 868
 869         return msg;
 870 }
 871
 872 /*
 873  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
 874  * to include additional client metadata fields.
 875  */
 876 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
 877 {
 878         struct ceph_msg *msg;
 879         struct ceph_mds_session_head *h;
 880         int i = -1;
 881         int metadata_bytes = 0;
 882         int metadata_key_count = 0;
 883         struct ceph_options *opt = mdsc->fsc->client->options;
 884         struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
 885         void *p;
 886
 887         const char* metadata[][2] = {
 888                 {"hostname", mdsc->nodename},
 889                 {"kernel_version", init_utsname()->release},
 890                 {"entity_id", opt->name ? : ""},
 891                 {"root", fsopt->server_path ? : "/"},
 892                 {NULL, NULL}
 893         };
 894
 895         /* Calculate serialized length of metadata */
 896         metadata_bytes = 4;  /* map length */
 897         for (i = 0; metadata[i][0]; ++i) {
 898                 metadata_bytes += 8 + strlen(metadata[i][0]) +
 899                         strlen(metadata[i][1]);
 900                 metadata_key_count++;
 901         }
 902
 903         /* Allocate the message */
 904         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + metadata_bytes,
 905                            GFP_NOFS, false);
 906         if (!msg) {
 907                 pr_err("create_session_msg ENOMEM creating msg\n");
 908                 return NULL;
 909         }
 910         h = msg->front.iov_base;
 911         h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
 912         h->seq = cpu_to_le64(seq);
 913
 914         /*
 915          * Serialize client metadata into waiting buffer space, using
 916          * the format that userspace expects for map<string, string>
 917          *
 918          * ClientSession messages with metadata are v2
 919          */
 920         msg->hdr.version = cpu_to_le16(2);
 921         msg->hdr.compat_version = cpu_to_le16(1);
 922
 923         /* The write pointer, following the session_head structure */
 924         p = msg->front.iov_base + sizeof(*h);
 925
 926         /* Number of entries in the map */
 927         ceph_encode_32(&p, metadata_key_count);
 928
 929         /* Two length-prefixed strings for each entry in the map */
 930         for (i = 0; metadata[i][0]; ++i) {
 931                 size_t const key_len = strlen(metadata[i][0]);
 932                 size_t const val_len = strlen(metadata[i][1]);
 933
 934                 ceph_encode_32(&p, key_len);
 935                 memcpy(p, metadata[i][0], key_len);
 936                 p += key_len;
 937                 ceph_encode_32(&p, val_len);
 938                 memcpy(p, metadata[i][1], val_len);
 939                 p += val_len;
 940         }
 941
 942         return msg;
 943 }
 944
 945 /*
 946  * send session open request.
 947  *
 948  * called under mdsc->mutex
 949  */
 950 static int __open_session(struct ceph_mds_client *mdsc,
 951                           struct ceph_mds_session *session)
 952 {
 953         struct ceph_msg *msg;
 954         int mstate;
 955         int mds = session->s_mds;
 956
 957         /* wait for mds to go active? */
 958         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
 959         dout("open_session to mds%d (%s)\n", mds,
 960              ceph_mds_state_name(mstate));
 961         session->s_state = CEPH_MDS_SESSION_OPENING;
 962         session->s_renew_requested = jiffies;
 963
 964         /* send connect message */
 965         msg = create_session_open_msg(mdsc, session->s_seq);
 966         if (!msg)
 967                 return -ENOMEM;
 968         ceph_con_send(&session->s_con, msg);
 969         return 0;
 970 }
 971
 972 /*
 973  * open sessions for any export targets for the given mds
 974  *
 975  * called under mdsc->mutex
 976  */
 977 static struct ceph_mds_session *
 978 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
 979 {
 980         struct ceph_mds_session *session;
 981
 982         session = __ceph_lookup_mds_session(mdsc, target);
 983         if (!session) {
 984                 session = register_session(mdsc, target);
 985                 if (IS_ERR(session))
 986                         return session;
 987         }
 988         if (session->s_state == CEPH_MDS_SESSION_NEW ||
 989             session->s_state == CEPH_MDS_SESSION_CLOSING)
 990                 __open_session(mdsc, session);
 991
 992         return session;
 993 }
 994
 995 struct ceph_mds_session *
 996 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
 997 {
 998         struct ceph_mds_session *session;
 999
1000         dout("open_export_target_session to mds%d\n", target);
1001
1002         mutex_lock(&mdsc->mutex);
1003         session = __open_export_target_session(mdsc, target);
1004         mutex_unlock(&mdsc->mutex);
1005
1006         return session;
1007 }
1008
1009 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1010                                           struct ceph_mds_session *session)
1011 {
1012         struct ceph_mds_info *mi;
1013         struct ceph_mds_session *ts;
1014         int i, mds = session->s_mds;
1015
1016         if (mds >= mdsc->mdsmap->m_num_mds)
1017                 return;
1018
1019         mi = &mdsc->mdsmap->m_info[mds];
1020         dout("open_export_target_sessions for mds%d (%d targets)\n",
1021              session->s_mds, mi->num_export_targets);
1022
1023         for (i = 0; i < mi->num_export_targets; i++) {
1024                 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1025                 if (!IS_ERR(ts))
1026                         ceph_put_mds_session(ts);
1027         }
1028 }
1029
1030 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1031                                            struct ceph_mds_session *session)
1032 {
1033         mutex_lock(&mdsc->mutex);
1034         __open_export_target_sessions(mdsc, session);
1035         mutex_unlock(&mdsc->mutex);
1036 }
1037
1038 /*
1039  * session caps
1040  */
1041
1042 static void detach_cap_releases(struct ceph_mds_session *session,
1043                                 struct list_head *target)
1044 {
1045         lockdep_assert_held(&session->s_cap_lock);
1046
1047         list_splice_init(&session->s_cap_releases, target);
1048         session->s_num_cap_releases = 0;
1049         dout("dispose_cap_releases mds%d\n", session->s_mds);
1050 }
1051
1052 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1053                                  struct list_head *dispose)
1054 {
1055         while (!list_empty(dispose)) {
1056                 struct ceph_cap *cap;
1057                 /* zero out the in-progress message */
1058                 cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1059                 list_del(&cap->session_caps);
1060                 ceph_put_cap(mdsc, cap);
1061         }
1062 }
1063
1064 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1065                                      struct ceph_mds_session *session)
1066 {
1067         struct ceph_mds_request *req;
1068         struct rb_node *p;
1069
1070         dout("cleanup_session_requests mds%d\n", session->s_mds);
1071         mutex_lock(&mdsc->mutex);
1072         while (!list_empty(&session->s_unsafe)) {
1073                 req = list_first_entry(&session->s_unsafe,
1074                                        struct ceph_mds_request, r_unsafe_item);
1075                 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1076                                     req->r_tid);
1077                 __unregister_request(mdsc, req);
1078         }
1079         /* zero r_attempts, so kick_requests() will re-send requests */
1080         p = rb_first(&mdsc->request_tree);
1081         while (p) {
1082                 req = rb_entry(p, struct ceph_mds_request, r_node);
1083                 p = rb_next(p);
1084                 if (req->r_session &&
1085                     req->r_session->s_mds == session->s_mds)
1086                         req->r_attempts = 0;
1087         }
1088         mutex_unlock(&mdsc->mutex);
1089 }
1090
1091 /*
1092  * Helper to safely iterate over all caps associated with a session, with
1093  * special care taken to handle a racing __ceph_remove_cap().
1094  *
1095  * Caller must hold session s_mutex.
1096  */
1097 static int iterate_session_caps(struct ceph_mds_session *session,
1098                                  int (*cb)(struct inode *, struct ceph_cap *,
1099                                             void *), void *arg)
1100 {
1101         struct list_head *p;
1102         struct ceph_cap *cap;
1103         struct inode *inode, *last_inode = NULL;
1104         struct ceph_cap *old_cap = NULL;
1105         int ret;
1106
1107         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1108         spin_lock(&session->s_cap_lock);
1109         p = session->s_caps.next;
1110         while (p != &session->s_caps) {
1111                 cap = list_entry(p, struct ceph_cap, session_caps);
1112                 inode = igrab(&cap->ci->vfs_inode);
1113                 if (!inode) {
1114                         p = p->next;
1115                         continue;
1116                 }
1117                 session->s_cap_iterator = cap;
1118                 spin_unlock(&session->s_cap_lock);
1119
1120                 if (last_inode) {
1121                         iput(last_inode);
1122                         last_inode = NULL;
1123                 }
1124                 if (old_cap) {
1125                         ceph_put_cap(session->s_mdsc, old_cap);
1126                         old_cap = NULL;
1127                 }
1128
1129                 ret = cb(inode, cap, arg);
1130                 last_inode = inode;
1131
1132                 spin_lock(&session->s_cap_lock);
1133                 p = p->next;
1134                 if (!cap->ci) {
1135                         dout("iterate_session_caps  finishing cap %p removal\n",
1136                              cap);
1137                         BUG_ON(cap->session != session);
1138                         cap->session = NULL;
1139                         list_del_init(&cap->session_caps);
1140                         session->s_nr_caps--;
1141                         if (cap->queue_release) {
1142                                 list_add_tail(&cap->session_caps,
1143                                               &session->s_cap_releases);
1144                                 session->s_num_cap_releases++;
1145                         } else {
1146                                 old_cap = cap;  /* put_cap it w/o locks held */
1147                         }
1148                 }
1149                 if (ret < 0)
1150                         goto out;
1151         }
1152         ret = 0;
1153 out:
1154         session->s_cap_iterator = NULL;
1155         spin_unlock(&session->s_cap_lock);
1156
1157         iput(last_inode);
1158         if (old_cap)
1159                 ceph_put_cap(session->s_mdsc, old_cap);
1160
1161         return ret;
1162 }
1163
1164 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1165                                   void *arg)
1166 {
1167         struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1168         struct ceph_inode_info *ci = ceph_inode(inode);
1169         LIST_HEAD(to_remove);
1170         bool drop = false;
1171         bool invalidate = false;
1172
1173         dout("removing cap %p, ci is %p, inode is %p\n",
1174              cap, ci, &ci->vfs_inode);
1175         spin_lock(&ci->i_ceph_lock);
1176         __ceph_remove_cap(cap, false);
1177         if (!ci->i_auth_cap) {
1178                 struct ceph_cap_flush *cf;
1179                 struct ceph_mds_client *mdsc = fsc->mdsc;
1180
1181                 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1182
1183                 if (ci->i_wrbuffer_ref > 0 &&
1184                     READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
1185                         invalidate = true;
1186
1187                 while (!list_empty(&ci->i_cap_flush_list)) {
1188                         cf = list_first_entry(&ci->i_cap_flush_list,
1189                                               struct ceph_cap_flush, i_list);
1190                         list_move(&cf->i_list, &to_remove);
1191                 }
1192
1193                 spin_lock(&mdsc->cap_dirty_lock);
1194
1195                 list_for_each_entry(cf, &to_remove, i_list)
1196                         list_del(&cf->g_list);
1197
1198                 if (!list_empty(&ci->i_dirty_item)) {
1199                         pr_warn_ratelimited(
1200                                 " dropping dirty %s state for %p %lld\n",
1201                                 ceph_cap_string(ci->i_dirty_caps),
1202                                 inode, ceph_ino(inode));
1203                         ci->i_dirty_caps = 0;
1204                         list_del_init(&ci->i_dirty_item);
1205                         drop = true;
1206                 }
1207                 if (!list_empty(&ci->i_flushing_item)) {
1208                         pr_warn_ratelimited(
1209                                 " dropping dirty+flushing %s state for %p %lld\n",
1210                                 ceph_cap_string(ci->i_flushing_caps),
1211                                 inode, ceph_ino(inode));
1212                         ci->i_flushing_caps = 0;
1213                         list_del_init(&ci->i_flushing_item);
1214                         mdsc->num_cap_flushing--;
1215                         drop = true;
1216                 }
1217                 spin_unlock(&mdsc->cap_dirty_lock);
1218
1219                 if (atomic_read(&ci->i_filelock_ref) > 0) {
1220                         /* make further file lock syscall return -EIO */
1221                         ci->i_ceph_flags |= CEPH_I_ERROR_FILELOCK;
1222                         pr_warn_ratelimited(" dropping file locks for %p %lld\n",
1223                                             inode, ceph_ino(inode));
1224                 }
1225
1226                 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1227                         list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1228                         ci->i_prealloc_cap_flush = NULL;
1229                 }
1230         }
1231         spin_unlock(&ci->i_ceph_lock);
1232         while (!list_empty(&to_remove)) {
1233                 struct ceph_cap_flush *cf;
1234                 cf = list_first_entry(&to_remove,
1235                                       struct ceph_cap_flush, i_list);
1236                 list_del(&cf->i_list);
1237                 ceph_free_cap_flush(cf);
1238         }
1239
1240         wake_up_all(&ci->i_cap_wq);
1241         if (invalidate)
1242                 ceph_queue_invalidate(inode);
1243         if (drop)
1244                 iput(inode);
1245         return 0;
1246 }
1247
1248 /*
1249  * caller must hold session s_mutex
1250  */
1251 static void remove_session_caps(struct ceph_mds_session *session)
1252 {
1253         struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1254         struct super_block *sb = fsc->sb;
1255         LIST_HEAD(dispose);
1256
1257         dout("remove_session_caps on %p\n", session);
1258         iterate_session_caps(session, remove_session_caps_cb, fsc);
1259
1260         wake_up_all(&fsc->mdsc->cap_flushing_wq);
1261
1262         spin_lock(&session->s_cap_lock);
1263         if (session->s_nr_caps > 0) {
1264                 struct inode *inode;
1265                 struct ceph_cap *cap, *prev = NULL;
1266                 struct ceph_vino vino;
1267                 /*
1268                  * iterate_session_caps() skips inodes that are being
1269                  * deleted, we need to wait until deletions are complete.
1270                  * __wait_on_freeing_inode() is designed for the job,
1271                  * but it is not exported, so use lookup inode function
1272                  * to access it.
1273                  */
1274                 while (!list_empty(&session->s_caps)) {
1275                         cap = list_entry(session->s_caps.next,
1276                                          struct ceph_cap, session_caps);
1277                         if (cap == prev)
1278                                 break;
1279                         prev = cap;
1280                         vino = cap->ci->i_vino;
1281                         spin_unlock(&session->s_cap_lock);
1282
1283                         inode = ceph_find_inode(sb, vino);
1284                         iput(inode);
1285
1286                         spin_lock(&session->s_cap_lock);
1287                 }
1288         }
1289
1290         // drop cap expires and unlock s_cap_lock
1291         detach_cap_releases(session, &dispose);
1292
1293         BUG_ON(session->s_nr_caps > 0);
1294         BUG_ON(!list_empty(&session->s_cap_flushing));
1295         spin_unlock(&session->s_cap_lock);
1296         dispose_cap_releases(session->s_mdsc, &dispose);
1297 }
1298
1299 /*
1300  * wake up any threads waiting on this session's caps.  if the cap is
1301  * old (didn't get renewed on the client reconnect), remove it now.
1302  *
1303  * caller must hold s_mutex.
1304  */
1305 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1306                               void *arg)
1307 {
1308         struct ceph_inode_info *ci = ceph_inode(inode);
1309
1310         if (arg) {
1311                 spin_lock(&ci->i_ceph_lock);
1312                 ci->i_wanted_max_size = 0;
1313                 ci->i_requested_max_size = 0;
1314                 spin_unlock(&ci->i_ceph_lock);
1315         }
1316         wake_up_all(&ci->i_cap_wq);
1317         return 0;
1318 }
1319
1320 static void wake_up_session_caps(struct ceph_mds_session *session,
1321                                  int reconnect)
1322 {
1323         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1324         iterate_session_caps(session, wake_up_session_cb,
1325                              (void *)(unsigned long)reconnect);
1326 }
1327
1328 /*
1329  * Send periodic message to MDS renewing all currently held caps.  The
1330  * ack will reset the expiration for all caps from this session.
1331  *
1332  * caller holds s_mutex
1333  */
1334 static int send_renew_caps(struct ceph_mds_client *mdsc,
1335                            struct ceph_mds_session *session)
1336 {
1337         struct ceph_msg *msg;
1338         int state;
1339
1340         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1341             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1342                 pr_info("mds%d caps stale\n", session->s_mds);
1343         session->s_renew_requested = jiffies;
1344
1345         /* do not try to renew caps until a recovering mds has reconnected
1346          * with its clients. */
1347         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1348         if (state < CEPH_MDS_STATE_RECONNECT) {
1349                 dout("send_renew_caps ignoring mds%d (%s)\n",
1350                      session->s_mds, ceph_mds_state_name(state));
1351                 return 0;
1352         }
1353
1354         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1355                 ceph_mds_state_name(state));
1356         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1357                                  ++session->s_renew_seq);
1358         if (!msg)
1359                 return -ENOMEM;
1360         ceph_con_send(&session->s_con, msg);
1361         return 0;
1362 }
1363
1364 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1365                              struct ceph_mds_session *session, u64 seq)
1366 {
1367         struct ceph_msg *msg;
1368
1369         dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1370              session->s_mds, ceph_session_state_name(session->s_state), seq);
1371         msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1372         if (!msg)
1373                 return -ENOMEM;
1374         ceph_con_send(&session->s_con, msg);
1375         return 0;
1376 }
1377
1378
1379 /*
1380  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1381  *
1382  * Called under session->s_mutex
1383  */
1384 static void renewed_caps(struct ceph_mds_client *mdsc,
1385                          struct ceph_mds_session *session, int is_renew)
1386 {
1387         int was_stale;
1388         int wake = 0;
1389
1390         spin_lock(&session->s_cap_lock);
1391         was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1392
1393         session->s_cap_ttl = session->s_renew_requested +
1394                 mdsc->mdsmap->m_session_timeout*HZ;
1395
1396         if (was_stale) {
1397                 if (time_before(jiffies, session->s_cap_ttl)) {
1398                         pr_info("mds%d caps renewed\n", session->s_mds);
1399                         wake = 1;
1400                 } else {
1401                         pr_info("mds%d caps still stale\n", session->s_mds);
1402                 }
1403         }
1404         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1405              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1406              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1407         spin_unlock(&session->s_cap_lock);
1408
1409         if (wake)
1410                 wake_up_session_caps(session, 0);
1411 }
1412
1413 /*
1414  * send a session close request
1415  */
1416 static int request_close_session(struct ceph_mds_client *mdsc,
1417                                  struct ceph_mds_session *session)
1418 {
1419         struct ceph_msg *msg;
1420
1421         dout("request_close_session mds%d state %s seq %lld\n",
1422              session->s_mds, ceph_session_state_name(session->s_state),
1423              session->s_seq);
1424         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1425         if (!msg)
1426                 return -ENOMEM;
1427         ceph_con_send(&session->s_con, msg);
1428         return 1;
1429 }
1430
1431 /*
1432  * Called with s_mutex held.
1433  */
1434 static int __close_session(struct ceph_mds_client *mdsc,
1435                          struct ceph_mds_session *session)
1436 {
1437         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1438                 return 0;
1439         session->s_state = CEPH_MDS_SESSION_CLOSING;
1440         return request_close_session(mdsc, session);
1441 }
1442
1443 /*
1444  * Trim old(er) caps.
1445  *
1446  * Because we can't cache an inode without one or more caps, we do
1447  * this indirectly: if a cap is unused, we prune its aliases, at which
1448  * point the inode will hopefully get dropped to.
1449  *
1450  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1451  * memory pressure from the MDS, though, so it needn't be perfect.
1452  */
1453 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1454 {
1455         struct ceph_mds_session *session = arg;
1456         struct ceph_inode_info *ci = ceph_inode(inode);
1457         int used, wanted, oissued, mine;
1458
1459         if (session->s_trim_caps <= 0)
1460                 return -1;
1461
1462         spin_lock(&ci->i_ceph_lock);
1463         mine = cap->issued | cap->implemented;
1464         used = __ceph_caps_used(ci);
1465         wanted = __ceph_caps_file_wanted(ci);
1466         oissued = __ceph_caps_issued_other(ci, cap);
1467
1468         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1469              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1470              ceph_cap_string(used), ceph_cap_string(wanted));
1471         if (cap == ci->i_auth_cap) {
1472                 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1473                     !list_empty(&ci->i_cap_snaps))
1474                         goto out;
1475                 if ((used | wanted) & CEPH_CAP_ANY_WR)
1476                         goto out;
1477                 /* Note: it's possible that i_filelock_ref becomes non-zero
1478                  * after dropping auth caps. It doesn't hurt because reply
1479                  * of lock mds request will re-add auth caps. */
1480                 if (atomic_read(&ci->i_filelock_ref) > 0)
1481                         goto out;
1482         }
1483         /* The inode has cached pages, but it's no longer used.
1484          * we can safely drop it */
1485         if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1486             !(oissued & CEPH_CAP_FILE_CACHE)) {
1487           used = 0;
1488           oissued = 0;
1489         }
1490         if ((used | wanted) & ~oissued & mine)
1491                 goto out;   /* we need these caps */
1492
1493         session->s_trim_caps--;
1494         if (oissued) {
1495                 /* we aren't the only cap.. just remove us */
1496                 __ceph_remove_cap(cap, true);
1497         } else {
1498                 /* try dropping referring dentries */
1499                 spin_unlock(&ci->i_ceph_lock);
1500                 d_prune_aliases(inode);
1501                 dout("trim_caps_cb %p cap %p  pruned, count now %d\n",
1502                      inode, cap, atomic_read(&inode->i_count));
1503                 return 0;
1504         }
1505
1506 out:
1507         spin_unlock(&ci->i_ceph_lock);
1508         return 0;
1509 }
1510
1511 /*
1512  * Trim session cap count down to some max number.
1513  */
1514 static int trim_caps(struct ceph_mds_client *mdsc,
1515                      struct ceph_mds_session *session,
1516                      int max_caps)
1517 {
1518         int trim_caps = session->s_nr_caps - max_caps;
1519
1520         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1521              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1522         if (trim_caps > 0) {
1523                 session->s_trim_caps = trim_caps;
1524                 iterate_session_caps(session, trim_caps_cb, session);
1525                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1526                      session->s_mds, session->s_nr_caps, max_caps,
1527                         trim_caps - session->s_trim_caps);
1528                 session->s_trim_caps = 0;
1529         }
1530
1531         ceph_send_cap_releases(mdsc, session);
1532         return 0;
1533 }
1534
1535 static int check_caps_flush(struct ceph_mds_client *mdsc,
1536                             u64 want_flush_tid)
1537 {
1538         int ret = 1;
1539
1540         spin_lock(&mdsc->cap_dirty_lock);
1541         if (!list_empty(&mdsc->cap_flush_list)) {
1542                 struct ceph_cap_flush *cf =
1543                         list_first_entry(&mdsc->cap_flush_list,
1544                                          struct ceph_cap_flush, g_list);
1545                 if (cf->tid <= want_flush_tid) {
1546                         dout("check_caps_flush still flushing tid "
1547                              "%llu <= %llu\n", cf->tid, want_flush_tid);
1548                         ret = 0;
1549                 }
1550         }
1551         spin_unlock(&mdsc->cap_dirty_lock);
1552         return ret;
1553 }
1554
1555 /*
1556  * flush all dirty inode data to disk.
1557  *
1558  * returns true if we've flushed through want_flush_tid
1559  */
1560 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1561                             u64 want_flush_tid)
1562 {
1563         dout("check_caps_flush want %llu\n", want_flush_tid);
1564
1565         wait_event(mdsc->cap_flushing_wq,
1566                    check_caps_flush(mdsc, want_flush_tid));
1567
1568         dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1569 }
1570
1571 /*
1572  * called under s_mutex
1573  */
1574 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1575                             struct ceph_mds_session *session)
1576 {
1577         struct ceph_msg *msg = NULL;
1578         struct ceph_mds_cap_release *head;
1579         struct ceph_mds_cap_item *item;
1580         struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
1581         struct ceph_cap *cap;
1582         LIST_HEAD(tmp_list);
1583         int num_cap_releases;
1584         __le32  barrier, *cap_barrier;
1585
1586         down_read(&osdc->lock);
1587         barrier = cpu_to_le32(osdc->epoch_barrier);
1588         up_read(&osdc->lock);
1589
1590         spin_lock(&session->s_cap_lock);
1591 again:
1592         list_splice_init(&session->s_cap_releases, &tmp_list);
1593         num_cap_releases = session->s_num_cap_releases;
1594         session->s_num_cap_releases = 0;
1595         spin_unlock(&session->s_cap_lock);
1596
1597         while (!list_empty(&tmp_list)) {
1598                 if (!msg) {
1599                         msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1600                                         PAGE_SIZE, GFP_NOFS, false);
1601                         if (!msg)
1602                                 goto out_err;
1603                         head = msg->front.iov_base;
1604                         head->num = cpu_to_le32(0);
1605                         msg->front.iov_len = sizeof(*head);
1606
1607                         msg->hdr.version = cpu_to_le16(2);
1608                         msg->hdr.compat_version = cpu_to_le16(1);
1609                 }
1610
1611                 cap = list_first_entry(&tmp_list, struct ceph_cap,
1612                                         session_caps);
1613                 list_del(&cap->session_caps);
1614                 num_cap_releases--;
1615
1616                 head = msg->front.iov_base;
1617                 le32_add_cpu(&head->num, 1);
1618                 item = msg->front.iov_base + msg->front.iov_len;
1619                 item->ino = cpu_to_le64(cap->cap_ino);
1620                 item->cap_id = cpu_to_le64(cap->cap_id);
1621                 item->migrate_seq = cpu_to_le32(cap->mseq);
1622                 item->seq = cpu_to_le32(cap->issue_seq);
1623                 msg->front.iov_len += sizeof(*item);
1624
1625                 ceph_put_cap(mdsc, cap);
1626
1627                 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1628                         // Append cap_barrier field
1629                         cap_barrier = msg->front.iov_base + msg->front.iov_len;
1630                         *cap_barrier = barrier;
1631                         msg->front.iov_len += sizeof(*cap_barrier);
1632
1633                         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1634                         dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1635                         ceph_con_send(&session->s_con, msg);
1636                         msg = NULL;
1637                 }
1638         }
1639
1640         BUG_ON(num_cap_releases != 0);
1641
1642         spin_lock(&session->s_cap_lock);
1643         if (!list_empty(&session->s_cap_releases))
1644                 goto again;
1645         spin_unlock(&session->s_cap_lock);
1646
1647         if (msg) {
1648                 // Append cap_barrier field
1649                 cap_barrier = msg->front.iov_base + msg->front.iov_len;
1650                 *cap_barrier = barrier;
1651                 msg->front.iov_len += sizeof(*cap_barrier);
1652
1653                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1654                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1655                 ceph_con_send(&session->s_con, msg);
1656         }
1657         return;
1658 out_err:
1659         pr_err("send_cap_releases mds%d, failed to allocate message\n",
1660                 session->s_mds);
1661         spin_lock(&session->s_cap_lock);
1662         list_splice(&tmp_list, &session->s_cap_releases);
1663         session->s_num_cap_releases += num_cap_releases;
1664         spin_unlock(&session->s_cap_lock);
1665 }
1666
1667 /*
1668  * requests
1669  */
1670
1671 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1672                                     struct inode *dir)
1673 {
1674         struct ceph_inode_info *ci = ceph_inode(dir);
1675         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1676         struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1677         size_t size = sizeof(struct ceph_mds_reply_dir_entry);
1678         int order, num_entries;
1679
1680         spin_lock(&ci->i_ceph_lock);
1681         num_entries = ci->i_files + ci->i_subdirs;
1682         spin_unlock(&ci->i_ceph_lock);
1683         num_entries = max(num_entries, 1);
1684         num_entries = min(num_entries, opt->max_readdir);
1685
1686         order = get_order(size * num_entries);
1687         while (order >= 0) {
1688                 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
1689                                                              __GFP_NOWARN,
1690                                                              order);
1691                 if (rinfo->dir_entries)
1692                         break;
1693                 order--;
1694         }
1695         if (!rinfo->dir_entries)
1696                 return -ENOMEM;
1697
1698         num_entries = (PAGE_SIZE << order) / size;
1699         num_entries = min(num_entries, opt->max_readdir);
1700
1701         rinfo->dir_buf_size = PAGE_SIZE << order;
1702         req->r_num_caps = num_entries + 1;
1703         req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1704         req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1705         return 0;
1706 }
1707
1708 /*
1709  * Create an mds request.
1710  */
1711 struct ceph_mds_request *
1712 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1713 {
1714         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1715
1716         if (!req)
1717                 return ERR_PTR(-ENOMEM);
1718
1719         mutex_init(&req->r_fill_mutex);
1720         req->r_mdsc = mdsc;
1721         req->r_started = jiffies;
1722         req->r_resend_mds = -1;
1723         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1724         INIT_LIST_HEAD(&req->r_unsafe_target_item);
1725         req->r_fmode = -1;
1726         kref_init(&req->r_kref);
1727         RB_CLEAR_NODE(&req->r_node);
1728         INIT_LIST_HEAD(&req->r_wait);
1729         init_completion(&req->r_completion);
1730         init_completion(&req->r_safe_completion);
1731         INIT_LIST_HEAD(&req->r_unsafe_item);
1732
1733         req->r_stamp = timespec_trunc(current_kernel_time(), mdsc->fsc->sb->s_time_gran);
1734
1735         req->r_op = op;
1736         req->r_direct_mode = mode;
1737         return req;
1738 }
1739
1740 /*
1741  * return oldest (lowest) request, tid in request tree, 0 if none.
1742  *
1743  * called under mdsc->mutex.
1744  */
1745 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1746 {
1747         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1748                 return NULL;
1749         return rb_entry(rb_first(&mdsc->request_tree),
1750                         struct ceph_mds_request, r_node);
1751 }
1752
1753 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1754 {
1755         return mdsc->oldest_tid;
1756 }
1757
1758 /*
1759  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1760  * on build_path_from_dentry in fs/cifs/dir.c.
1761  *
1762  * If @stop_on_nosnap, generate path relative to the first non-snapped
1763  * inode.
1764  *
1765  * Encode hidden .snap dirs as a double /, i.e.
1766  *   foo/.snap/bar -> foo//bar
1767  */
1768 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1769                            int stop_on_nosnap)
1770 {
1771         struct dentry *temp;
1772         char *path;
1773         int len, pos;
1774         unsigned seq;
1775
1776         if (!dentry)
1777                 return ERR_PTR(-EINVAL);
1778
1779 retry:
1780         len = 0;
1781         seq = read_seqbegin(&rename_lock);
1782         rcu_read_lock();
1783         for (temp = dentry; !IS_ROOT(temp);) {
1784                 struct inode *inode = d_inode(temp);
1785                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1786                         len++;  /* slash only */
1787                 else if (stop_on_nosnap && inode &&
1788                          ceph_snap(inode) == CEPH_NOSNAP)
1789                         break;
1790                 else
1791                         len += 1 + temp->d_name.len;
1792                 temp = temp->d_parent;
1793         }
1794         rcu_read_unlock();
1795         if (len)
1796                 len--;  /* no leading '/' */
1797
1798         path = kmalloc(len+1, GFP_NOFS);
1799         if (!path)
1800                 return ERR_PTR(-ENOMEM);
1801         pos = len;
1802         path[pos] = 0;  /* trailing null */
1803         rcu_read_lock();
1804         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1805                 struct inode *inode;
1806
1807                 spin_lock(&temp->d_lock);
1808                 inode = d_inode(temp);
1809                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1810                         dout("build_path path+%d: %p SNAPDIR\n",
1811                              pos, temp);
1812                 } else if (stop_on_nosnap && inode &&
1813                            ceph_snap(inode) == CEPH_NOSNAP) {
1814                         spin_unlock(&temp->d_lock);
1815                         break;
1816                 } else {
1817                         pos -= temp->d_name.len;
1818                         if (pos < 0) {
1819                                 spin_unlock(&temp->d_lock);
1820                                 break;
1821                         }
1822                         strncpy(path + pos, temp->d_name.name,
1823                                 temp->d_name.len);
1824                 }
1825                 spin_unlock(&temp->d_lock);
1826                 if (pos)
1827                         path[--pos] = '/';
1828                 temp = temp->d_parent;
1829         }
1830         rcu_read_unlock();
1831         if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1832                 pr_err("build_path did not end path lookup where "
1833                        "expected, namelen is %d, pos is %d\n", len, pos);
1834                 /* presumably this is only possible if racing with a
1835                    rename of one of the parent directories (we can not
1836                    lock the dentries above us to prevent this, but
1837                    retrying should be harmless) */
1838                 kfree(path);
1839                 goto retry;
1840         }
1841
1842         *base = ceph_ino(d_inode(temp));
1843         *plen = len;
1844         dout("build_path on %p %d built %llx '%.*s'\n",
1845              dentry, d_count(dentry), *base, len, path);
1846         return path;
1847 }
1848
1849 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
1850                              const char **ppath, int *ppathlen, u64 *pino,
1851                              int *pfreepath)
1852 {
1853         char *path;
1854
1855         rcu_read_lock();
1856         if (!dir)
1857                 dir = d_inode_rcu(dentry->d_parent);
1858         if (dir && ceph_snap(dir) == CEPH_NOSNAP) {
1859                 *pino = ceph_ino(dir);
1860                 rcu_read_unlock();
1861                 *ppath = dentry->d_name.name;
1862                 *ppathlen = dentry->d_name.len;
1863                 return 0;
1864         }
1865         rcu_read_unlock();
1866         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1867         if (IS_ERR(path))
1868                 return PTR_ERR(path);
1869         *ppath = path;
1870         *pfreepath = 1;
1871         return 0;
1872 }
1873
1874 static int build_inode_path(struct inode *inode,
1875                             const char **ppath, int *ppathlen, u64 *pino,
1876                             int *pfreepath)
1877 {
1878         struct dentry *dentry;
1879         char *path;
1880
1881         if (ceph_snap(inode) == CEPH_NOSNAP) {
1882                 *pino = ceph_ino(inode);
1883                 *ppathlen = 0;
1884                 return 0;
1885         }
1886         dentry = d_find_alias(inode);
1887         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1888         dput(dentry);
1889         if (IS_ERR(path))
1890                 return PTR_ERR(path);
1891         *ppath = path;
1892         *pfreepath = 1;
1893         return 0;
1894 }
1895
1896 /*
1897  * request arguments may be specified via an inode *, a dentry *, or
1898  * an explicit ino+path.
1899  */
1900 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1901                                   struct inode *rdiri, const char *rpath,
1902                                   u64 rino, const char **ppath, int *pathlen,
1903                                   u64 *ino, int *freepath)
1904 {
1905         int r = 0;
1906
1907         if (rinode) {
1908                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1909                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1910                      ceph_snap(rinode));
1911         } else if (rdentry) {
1912                 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
1913                                         freepath);
1914                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1915                      *ppath);
1916         } else if (rpath || rino) {
1917                 *ino = rino;
1918                 *ppath = rpath;
1919                 *pathlen = rpath ? strlen(rpath) : 0;
1920                 dout(" path %.*s\n", *pathlen, rpath);
1921         }
1922
1923         return r;
1924 }
1925
1926 /*
1927  * called under mdsc->mutex
1928  */
1929 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1930                                                struct ceph_mds_request *req,
1931                                                int mds, bool drop_cap_releases)
1932 {
1933         struct ceph_msg *msg;
1934         struct ceph_mds_request_head *head;
1935         const char *path1 = NULL;
1936         const char *path2 = NULL;
1937         u64 ino1 = 0, ino2 = 0;
1938         int pathlen1 = 0, pathlen2 = 0;
1939         int freepath1 = 0, freepath2 = 0;
1940         int len;
1941         u16 releases;
1942         void *p, *end;
1943         int ret;
1944
1945         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1946                               req->r_parent, req->r_path1, req->r_ino1.ino,
1947                               &path1, &pathlen1, &ino1, &freepath1);
1948         if (ret < 0) {
1949                 msg = ERR_PTR(ret);
1950                 goto out;
1951         }
1952
1953         ret = set_request_path_attr(NULL, req->r_old_dentry,
1954                               req->r_old_dentry_dir,
1955                               req->r_path2, req->r_ino2.ino,
1956                               &path2, &pathlen2, &ino2, &freepath2);
1957         if (ret < 0) {
1958                 msg = ERR_PTR(ret);
1959                 goto out_free1;
1960         }
1961
1962         len = sizeof(*head) +
1963                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
1964                 sizeof(struct ceph_timespec);
1965
1966         /* calculate (max) length for cap releases */
1967         len += sizeof(struct ceph_mds_request_release) *
1968                 (!!req->r_inode_drop + !!req->r_dentry_drop +
1969                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1970         if (req->r_dentry_drop)
1971                 len += req->r_dentry->d_name.len;
1972         if (req->r_old_dentry_drop)
1973                 len += req->r_old_dentry->d_name.len;
1974
1975         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1976         if (!msg) {
1977                 msg = ERR_PTR(-ENOMEM);
1978                 goto out_free2;
1979         }
1980
1981         msg->hdr.version = cpu_to_le16(2);
1982         msg->hdr.tid = cpu_to_le64(req->r_tid);
1983
1984         head = msg->front.iov_base;
1985         p = msg->front.iov_base + sizeof(*head);
1986         end = msg->front.iov_base + msg->front.iov_len;
1987
1988         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1989         head->op = cpu_to_le32(req->r_op);
1990         head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1991         head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1992         head->args = req->r_args;
1993
1994         ceph_encode_filepath(&p, end, ino1, path1);
1995         ceph_encode_filepath(&p, end, ino2, path2);
1996
1997         /* make note of release offset, in case we need to replay */
1998         req->r_request_release_offset = p - msg->front.iov_base;
1999
2000         /* cap releases */
2001         releases = 0;
2002         if (req->r_inode_drop)
2003                 releases += ceph_encode_inode_release(&p,
2004                       req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2005                       mds, req->r_inode_drop, req->r_inode_unless, 0);
2006         if (req->r_dentry_drop)
2007                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
2008                                 req->r_parent, mds, req->r_dentry_drop,
2009                                 req->r_dentry_unless);
2010         if (req->r_old_dentry_drop)
2011                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2012                                 req->r_old_dentry_dir, mds,
2013                                 req->r_old_dentry_drop,
2014                                 req->r_old_dentry_unless);
2015         if (req->r_old_inode_drop)
2016                 releases += ceph_encode_inode_release(&p,
2017                       d_inode(req->r_old_dentry),
2018                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2019
2020         if (drop_cap_releases) {
2021                 releases = 0;
2022                 p = msg->front.iov_base + req->r_request_release_offset;
2023         }
2024
2025         head->num_releases = cpu_to_le16(releases);
2026
2027         /* time stamp */
2028         {
2029                 struct ceph_timespec ts;
2030                 ceph_encode_timespec(&ts, &req->r_stamp);
2031                 ceph_encode_copy(&p, &ts, sizeof(ts));
2032         }
2033
2034         BUG_ON(p > end);
2035         msg->front.iov_len = p - msg->front.iov_base;
2036         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2037
2038         if (req->r_pagelist) {
2039                 struct ceph_pagelist *pagelist = req->r_pagelist;
2040                 refcount_inc(&pagelist->refcnt);
2041                 ceph_msg_data_add_pagelist(msg, pagelist);
2042                 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2043         } else {
2044                 msg->hdr.data_len = 0;
2045         }
2046
2047         msg->hdr.data_off = cpu_to_le16(0);
2048
2049 out_free2:
2050         if (freepath2)
2051                 kfree((char *)path2);
2052 out_free1:
2053         if (freepath1)
2054                 kfree((char *)path1);
2055 out:
2056         return msg;
2057 }
2058
2059 /*
2060  * called under mdsc->mutex if error, under no mutex if
2061  * success.
2062  */
2063 static void complete_request(struct ceph_mds_client *mdsc,
2064                              struct ceph_mds_request *req)
2065 {
2066         if (req->r_callback)
2067                 req->r_callback(mdsc, req);
2068         else
2069                 complete_all(&req->r_completion);
2070 }
2071
2072 /*
2073  * called under mdsc->mutex
2074  */
2075 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2076                                   struct ceph_mds_request *req,
2077                                   int mds, bool drop_cap_releases)
2078 {
2079         struct ceph_mds_request_head *rhead;
2080         struct ceph_msg *msg;
2081         int flags = 0;
2082
2083         req->r_attempts++;
2084         if (req->r_inode) {
2085                 struct ceph_cap *cap =
2086                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2087
2088                 if (cap)
2089                         req->r_sent_on_mseq = cap->mseq;
2090                 else
2091                         req->r_sent_on_mseq = -1;
2092         }
2093         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2094              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2095
2096         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2097                 void *p;
2098                 /*
2099                  * Replay.  Do not regenerate message (and rebuild
2100                  * paths, etc.); just use the original message.
2101                  * Rebuilding paths will break for renames because
2102                  * d_move mangles the src name.
2103                  */
2104                 msg = req->r_request;
2105                 rhead = msg->front.iov_base;
2106
2107                 flags = le32_to_cpu(rhead->flags);
2108                 flags |= CEPH_MDS_FLAG_REPLAY;
2109                 rhead->flags = cpu_to_le32(flags);
2110
2111                 if (req->r_target_inode)
2112                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2113
2114                 rhead->num_retry = req->r_attempts - 1;
2115
2116                 /* remove cap/dentry releases from message */
2117                 rhead->num_releases = 0;
2118
2119                 /* time stamp */
2120                 p = msg->front.iov_base + req->r_request_release_offset;
2121                 {
2122                         struct ceph_timespec ts;
2123                         ceph_encode_timespec(&ts, &req->r_stamp);
2124                         ceph_encode_copy(&p, &ts, sizeof(ts));
2125                 }
2126
2127                 msg->front.iov_len = p - msg->front.iov_base;
2128                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2129                 return 0;
2130         }
2131
2132         if (req->r_request) {
2133                 ceph_msg_put(req->r_request);
2134                 req->r_request = NULL;
2135         }
2136         msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2137         if (IS_ERR(msg)) {
2138                 req->r_err = PTR_ERR(msg);
2139                 return PTR_ERR(msg);
2140         }
2141         req->r_request = msg;
2142
2143         rhead = msg->front.iov_base;
2144         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2145         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2146                 flags |= CEPH_MDS_FLAG_REPLAY;
2147         if (req->r_parent)
2148                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2149         rhead->flags = cpu_to_le32(flags);
2150         rhead->num_fwd = req->r_num_fwd;
2151         rhead->num_retry = req->r_attempts - 1;
2152         rhead->ino = 0;
2153
2154         dout(" r_parent = %p\n", req->r_parent);
2155         return 0;
2156 }
2157
2158 /*
2159  * send request, or put it on the appropriate wait list.
2160  */
2161 static int __do_request(struct ceph_mds_client *mdsc,
2162                         struct ceph_mds_request *req)
2163 {
2164         struct ceph_mds_session *session = NULL;
2165         int mds = -1;
2166         int err = 0;
2167
2168         if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2169                 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2170                         __unregister_request(mdsc, req);
2171                 goto out;
2172         }
2173
2174         if (req->r_timeout &&
2175             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2176                 dout("do_request timed out\n");
2177                 err = -EIO;
2178                 goto finish;
2179         }
2180         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2181                 dout("do_request forced umount\n");
2182                 err = -EIO;
2183                 goto finish;
2184         }
2185         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2186                 if (mdsc->mdsmap_err) {
2187                         err = mdsc->mdsmap_err;
2188                         dout("do_request mdsmap err %d\n", err);
2189                         goto finish;
2190                 }
2191                 if (mdsc->mdsmap->m_epoch == 0) {
2192                         dout("do_request no mdsmap, waiting for map\n");
2193                         list_add(&req->r_wait, &mdsc->waiting_for_map);
2194                         goto finish;
2195                 }
2196                 if (!(mdsc->fsc->mount_options->flags &
2197                       CEPH_MOUNT_OPT_MOUNTWAIT) &&
2198                     !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2199                         err = -ENOENT;
2200                         pr_info("probably no mds server is up\n");
2201                         goto finish;
2202                 }
2203         }
2204
2205         put_request_session(req);
2206
2207         mds = __choose_mds(mdsc, req);
2208         if (mds < 0 ||
2209             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2210                 dout("do_request no mds or not active, waiting for map\n");
2211                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2212                 goto out;
2213         }
2214
2215         /* get, open session */
2216         session = __ceph_lookup_mds_session(mdsc, mds);
2217         if (!session) {
2218                 session = register_session(mdsc, mds);
2219                 if (IS_ERR(session)) {
2220                         err = PTR_ERR(session);
2221                         goto finish;
2222                 }
2223         }
2224         req->r_session = get_session(session);
2225
2226         dout("do_request mds%d session %p state %s\n", mds, session,
2227              ceph_session_state_name(session->s_state));
2228         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2229             session->s_state != CEPH_MDS_SESSION_HUNG) {
2230                 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2231                         err = -EACCES;
2232                         goto out_session;
2233                 }
2234                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2235                     session->s_state == CEPH_MDS_SESSION_CLOSING)
2236                         __open_session(mdsc, session);
2237                 list_add(&req->r_wait, &session->s_waiting);
2238                 goto out_session;
2239         }
2240
2241         /* send request */
2242         req->r_resend_mds = -1;   /* forget any previous mds hint */
2243
2244         if (req->r_request_started == 0)   /* note request start time */
2245                 req->r_request_started = jiffies;
2246
2247         err = __prepare_send_request(mdsc, req, mds, false);
2248         if (!err) {
2249                 ceph_msg_get(req->r_request);
2250                 ceph_con_send(&session->s_con, req->r_request);
2251         }
2252
2253 out_session:
2254         ceph_put_mds_session(session);
2255 finish:
2256         if (err) {
2257                 dout("__do_request early error %d\n", err);
2258                 req->r_err = err;
2259                 complete_request(mdsc, req);
2260                 __unregister_request(mdsc, req);
2261         }
2262 out:
2263         return err;
2264 }
2265
2266 /*
2267  * called under mdsc->mutex
2268  */
2269 static void __wake_requests(struct ceph_mds_client *mdsc,
2270                             struct list_head *head)
2271 {
2272         struct ceph_mds_request *req;
2273         LIST_HEAD(tmp_list);
2274
2275         list_splice_init(head, &tmp_list);
2276
2277         while (!list_empty(&tmp_list)) {
2278                 req = list_entry(tmp_list.next,
2279                                  struct ceph_mds_request, r_wait);
2280                 list_del_init(&req->r_wait);
2281                 dout(" wake request %p tid %llu\n", req, req->r_tid);
2282                 __do_request(mdsc, req);
2283         }
2284 }
2285
2286 /*
2287  * Wake up threads with requests pending for @mds, so that they can
2288  * resubmit their requests to a possibly different mds.
2289  */
2290 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2291 {
2292         struct ceph_mds_request *req;
2293         struct rb_node *p = rb_first(&mdsc->request_tree);
2294
2295         dout("kick_requests mds%d\n", mds);
2296         while (p) {
2297                 req = rb_entry(p, struct ceph_mds_request, r_node);
2298                 p = rb_next(p);
2299                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2300                         continue;
2301                 if (req->r_attempts > 0)
2302                         continue; /* only new requests */
2303                 if (req->r_session &&
2304                     req->r_session->s_mds == mds) {
2305                         dout(" kicking tid %llu\n", req->r_tid);
2306                         list_del_init(&req->r_wait);
2307                         __do_request(mdsc, req);
2308                 }
2309         }
2310 }
2311
2312 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2313                               struct ceph_mds_request *req)
2314 {
2315         dout("submit_request on %p\n", req);
2316         mutex_lock(&mdsc->mutex);
2317         __register_request(mdsc, req, NULL);
2318         __do_request(mdsc, req);
2319         mutex_unlock(&mdsc->mutex);
2320 }
2321
2322 /*
2323  * Synchrously perform an mds request.  Take care of all of the
2324  * session setup, forwarding, retry details.
2325  */
2326 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2327                          struct inode *dir,
2328                          struct ceph_mds_request *req)
2329 {
2330         int err;
2331
2332         dout("do_request on %p\n", req);
2333
2334         /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2335         if (req->r_inode)
2336                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2337         if (req->r_parent)
2338                 ceph_get_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
2339         if (req->r_old_dentry_dir)
2340                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2341                                   CEPH_CAP_PIN);
2342
2343         /* issue */
2344         mutex_lock(&mdsc->mutex);
2345         __register_request(mdsc, req, dir);
2346         __do_request(mdsc, req);
2347
2348         if (req->r_err) {
2349                 err = req->r_err;
2350                 goto out;
2351         }
2352
2353         /* wait */
2354         mutex_unlock(&mdsc->mutex);
2355         dout("do_request waiting\n");
2356         if (!req->r_timeout && req->r_wait_for_completion) {
2357                 err = req->r_wait_for_completion(mdsc, req);
2358         } else {
2359                 long timeleft = wait_for_completion_killable_timeout(
2360                                         &req->r_completion,
2361                                         ceph_timeout_jiffies(req->r_timeout));
2362                 if (timeleft > 0)
2363                         err = 0;
2364                 else if (!timeleft)
2365                         err = -EIO;  /* timed out */
2366                 else
2367                         err = timeleft;  /* killed */
2368         }
2369         dout("do_request waited, got %d\n", err);
2370         mutex_lock(&mdsc->mutex);
2371
2372         /* only abort if we didn't race with a real reply */
2373         if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2374                 err = le32_to_cpu(req->r_reply_info.head->result);
2375         } else if (err < 0) {
2376                 dout("aborted request %lld with %d\n", req->r_tid, err);
2377
2378                 /*
2379                  * ensure we aren't running concurrently with
2380                  * ceph_fill_trace or ceph_readdir_prepopulate, which
2381                  * rely on locks (dir mutex) held by our caller.
2382                  */
2383                 mutex_lock(&req->r_fill_mutex);
2384                 req->r_err = err;
2385                 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2386                 mutex_unlock(&req->r_fill_mutex);
2387
2388                 if (req->r_parent &&
2389                     (req->r_op & CEPH_MDS_OP_WRITE))
2390                         ceph_invalidate_dir_request(req);
2391         } else {
2392                 err = req->r_err;
2393         }
2394
2395 out:
2396         mutex_unlock(&mdsc->mutex);
2397         dout("do_request %p done, result %d\n", req, err);
2398         return err;
2399 }
2400
2401 /*
2402  * Invalidate dir's completeness, dentry lease state on an aborted MDS
2403  * namespace request.
2404  */
2405 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2406 {
2407         struct inode *inode = req->r_parent;
2408
2409         dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2410
2411         ceph_dir_clear_complete(inode);
2412         if (req->r_dentry)
2413                 ceph_invalidate_dentry_lease(req->r_dentry);
2414         if (req->r_old_dentry)
2415                 ceph_invalidate_dentry_lease(req->r_old_dentry);
2416 }
2417
2418 /*
2419  * Handle mds reply.
2420  *
2421  * We take the session mutex and parse and process the reply immediately.
2422  * This preserves the logical ordering of replies, capabilities, etc., sent
2423  * by the MDS as they are applied to our local cache.
2424  */
2425 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2426 {
2427         struct ceph_mds_client *mdsc = session->s_mdsc;
2428         struct ceph_mds_request *req;
2429         struct ceph_mds_reply_head *head = msg->front.iov_base;
2430         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2431         struct ceph_snap_realm *realm;
2432         u64 tid;
2433         int err, result;
2434         int mds = session->s_mds;
2435
2436         if (msg->front.iov_len < sizeof(*head)) {
2437                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2438                 ceph_msg_dump(msg);
2439                 return;
2440         }
2441
2442         /* get request, session */
2443         tid = le64_to_cpu(msg->hdr.tid);
2444         mutex_lock(&mdsc->mutex);
2445         req = lookup_get_request(mdsc, tid);
2446         if (!req) {
2447                 dout("handle_reply on unknown tid %llu\n", tid);
2448                 mutex_unlock(&mdsc->mutex);
2449                 return;
2450         }
2451         dout("handle_reply %p\n", req);
2452
2453         /* correct session? */
2454         if (req->r_session != session) {
2455                 pr_err("mdsc_handle_reply got %llu on session mds%d"
2456                        " not mds%d\n", tid, session->s_mds,
2457                        req->r_session ? req->r_session->s_mds : -1);
2458                 mutex_unlock(&mdsc->mutex);
2459                 goto out;
2460         }
2461
2462         /* dup? */
2463         if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
2464             (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
2465                 pr_warn("got a dup %s reply on %llu from mds%d\n",
2466                            head->safe ? "safe" : "unsafe", tid, mds);
2467                 mutex_unlock(&mdsc->mutex);
2468                 goto out;
2469         }
2470         if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
2471                 pr_warn("got unsafe after safe on %llu from mds%d\n",
2472                            tid, mds);
2473                 mutex_unlock(&mdsc->mutex);
2474                 goto out;
2475         }
2476
2477         result = le32_to_cpu(head->result);
2478
2479         /*
2480          * Handle an ESTALE
2481          * if we're not talking to the authority, send to them
2482          * if the authority has changed while we weren't looking,
2483          * send to new authority
2484          * Otherwise we just have to return an ESTALE
2485          */
2486         if (result == -ESTALE) {
2487                 dout("got ESTALE on request %llu", req->r_tid);
2488                 req->r_resend_mds = -1;
2489                 if (req->r_direct_mode != USE_AUTH_MDS) {
2490                         dout("not using auth, setting for that now");
2491                         req->r_direct_mode = USE_AUTH_MDS;
2492                         __do_request(mdsc, req);
2493                         mutex_unlock(&mdsc->mutex);
2494                         goto out;
2495                 } else  {
2496                         int mds = __choose_mds(mdsc, req);
2497                         if (mds >= 0 && mds != req->r_session->s_mds) {
2498                                 dout("but auth changed, so resending");
2499                                 __do_request(mdsc, req);
2500                                 mutex_unlock(&mdsc->mutex);
2501                                 goto out;
2502                         }
2503                 }
2504                 dout("have to return ESTALE on request %llu", req->r_tid);
2505         }
2506
2507
2508         if (head->safe) {
2509                 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
2510                 __unregister_request(mdsc, req);
2511
2512                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2513                         /*
2514                          * We already handled the unsafe response, now do the
2515                          * cleanup.  No need to examine the response; the MDS
2516                          * doesn't include any result info in the safe
2517                          * response.  And even if it did, there is nothing
2518                          * useful we could do with a revised return value.
2519                          */
2520                         dout("got safe reply %llu, mds%d\n", tid, mds);
2521
2522                         /* last unsafe request during umount? */
2523                         if (mdsc->stopping && !__get_oldest_req(mdsc))
2524                                 complete_all(&mdsc->safe_umount_waiters);
2525                         mutex_unlock(&mdsc->mutex);
2526                         goto out;
2527                 }
2528         } else {
2529                 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
2530                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2531                 if (req->r_unsafe_dir) {
2532                         struct ceph_inode_info *ci =
2533                                         ceph_inode(req->r_unsafe_dir);
2534                         spin_lock(&ci->i_unsafe_lock);
2535                         list_add_tail(&req->r_unsafe_dir_item,
2536                                       &ci->i_unsafe_dirops);
2537                         spin_unlock(&ci->i_unsafe_lock);
2538                 }
2539         }
2540
2541         dout("handle_reply tid %lld result %d\n", tid, result);
2542         rinfo = &req->r_reply_info;
2543         err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2544         mutex_unlock(&mdsc->mutex);
2545
2546         mutex_lock(&session->s_mutex);
2547         if (err < 0) {
2548                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2549                 ceph_msg_dump(msg);
2550                 goto out_err;
2551         }
2552
2553         /* snap trace */
2554         realm = NULL;
2555         if (rinfo->snapblob_len) {
2556                 down_write(&mdsc->snap_rwsem);
2557                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2558                                 rinfo->snapblob + rinfo->snapblob_len,
2559                                 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2560                                 &realm);
2561                 downgrade_write(&mdsc->snap_rwsem);
2562         } else {
2563                 down_read(&mdsc->snap_rwsem);
2564         }
2565
2566         /* insert trace into our cache */
2567         mutex_lock(&req->r_fill_mutex);
2568         current->journal_info = req;
2569         err = ceph_fill_trace(mdsc->fsc->sb, req);
2570         if (err == 0) {
2571                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2572                                     req->r_op == CEPH_MDS_OP_LSSNAP))
2573                         ceph_readdir_prepopulate(req, req->r_session);
2574                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2575         }
2576         current->journal_info = NULL;
2577         mutex_unlock(&req->r_fill_mutex);
2578
2579         up_read(&mdsc->snap_rwsem);
2580         if (realm)
2581                 ceph_put_snap_realm(mdsc, realm);
2582
2583         if (err == 0 && req->r_target_inode &&
2584             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2585                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
2586                 spin_lock(&ci->i_unsafe_lock);
2587                 list_add_tail(&req->r_unsafe_target_item, &ci->i_unsafe_iops);
2588                 spin_unlock(&ci->i_unsafe_lock);
2589         }
2590 out_err:
2591         mutex_lock(&mdsc->mutex);
2592         if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2593                 if (err) {
2594                         req->r_err = err;
2595                 } else {
2596                         req->r_reply =  ceph_msg_get(msg);
2597                         set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
2598                 }
2599         } else {
2600                 dout("reply arrived after request %lld was aborted\n", tid);
2601         }
2602         mutex_unlock(&mdsc->mutex);
2603
2604         mutex_unlock(&session->s_mutex);
2605
2606         /* kick calling process */
2607         complete_request(mdsc, req);
2608 out:
2609         ceph_mdsc_put_request(req);
2610         return;
2611 }
2612
2613
2614
2615 /*
2616  * handle mds notification that our request has been forwarded.
2617  */
2618 static void handle_forward(struct ceph_mds_client *mdsc,
2619                            struct ceph_mds_session *session,
2620                            struct ceph_msg *msg)
2621 {
2622         struct ceph_mds_request *req;
2623         u64 tid = le64_to_cpu(msg->hdr.tid);
2624         u32 next_mds;
2625         u32 fwd_seq;
2626         int err = -EINVAL;
2627         void *p = msg->front.iov_base;
2628         void *end = p + msg->front.iov_len;
2629
2630         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2631         next_mds = ceph_decode_32(&p);
2632         fwd_seq = ceph_decode_32(&p);
2633
2634         mutex_lock(&mdsc->mutex);
2635         req = lookup_get_request(mdsc, tid);
2636         if (!req) {
2637                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2638                 goto out;  /* dup reply? */
2639         }
2640
2641         if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2642                 dout("forward tid %llu aborted, unregistering\n", tid);
2643                 __unregister_request(mdsc, req);
2644         } else if (fwd_seq <= req->r_num_fwd) {
2645                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2646                      tid, next_mds, req->r_num_fwd, fwd_seq);
2647         } else {
2648                 /* resend. forward race not possible; mds would drop */
2649                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2650                 BUG_ON(req->r_err);
2651                 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
2652                 req->r_attempts = 0;
2653                 req->r_num_fwd = fwd_seq;
2654                 req->r_resend_mds = next_mds;
2655                 put_request_session(req);
2656                 __do_request(mdsc, req);
2657         }
2658         ceph_mdsc_put_request(req);
2659 out:
2660         mutex_unlock(&mdsc->mutex);
2661         return;
2662
2663 bad:
2664         pr_err("mdsc_handle_forward decode error err=%d\n", err);
2665 }
2666
2667 /*
2668  * handle a mds session control message
2669  */
2670 static void handle_session(struct ceph_mds_session *session,
2671                            struct ceph_msg *msg)
2672 {
2673         struct ceph_mds_client *mdsc = session->s_mdsc;
2674         u32 op;
2675         u64 seq;
2676         int mds = session->s_mds;
2677         struct ceph_mds_session_head *h = msg->front.iov_base;
2678         int wake = 0;
2679
2680         /* decode */
2681         if (msg->front.iov_len != sizeof(*h))
2682                 goto bad;
2683         op = le32_to_cpu(h->op);
2684         seq = le64_to_cpu(h->seq);
2685
2686         mutex_lock(&mdsc->mutex);
2687         if (op == CEPH_SESSION_CLOSE) {
2688                 get_session(session);
2689                 __unregister_session(mdsc, session);
2690         }
2691         /* FIXME: this ttl calculation is generous */
2692         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2693         mutex_unlock(&mdsc->mutex);
2694
2695         mutex_lock(&session->s_mutex);
2696
2697         dout("handle_session mds%d %s %p state %s seq %llu\n",
2698              mds, ceph_session_op_name(op), session,
2699              ceph_session_state_name(session->s_state), seq);
2700
2701         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2702                 session->s_state = CEPH_MDS_SESSION_OPEN;
2703                 pr_info("mds%d came back\n", session->s_mds);
2704         }
2705
2706         switch (op) {
2707         case CEPH_SESSION_OPEN:
2708                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2709                         pr_info("mds%d reconnect success\n", session->s_mds);
2710                 session->s_state = CEPH_MDS_SESSION_OPEN;
2711                 renewed_caps(mdsc, session, 0);
2712                 wake = 1;
2713                 if (mdsc->stopping)
2714                         __close_session(mdsc, session);
2715                 break;
2716
2717         case CEPH_SESSION_RENEWCAPS:
2718                 if (session->s_renew_seq == seq)
2719                         renewed_caps(mdsc, session, 1);
2720                 break;
2721
2722         case CEPH_SESSION_CLOSE:
2723                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2724                         pr_info("mds%d reconnect denied\n", session->s_mds);
2725                 cleanup_session_requests(mdsc, session);
2726                 remove_session_caps(session);
2727                 wake = 2; /* for good measure */
2728                 wake_up_all(&mdsc->session_close_wq);
2729                 break;
2730
2731         case CEPH_SESSION_STALE:
2732                 pr_info("mds%d caps went stale, renewing\n",
2733                         session->s_mds);
2734                 spin_lock(&session->s_gen_ttl_lock);
2735                 session->s_cap_gen++;
2736                 session->s_cap_ttl = jiffies - 1;
2737                 spin_unlock(&session->s_gen_ttl_lock);
2738                 send_renew_caps(mdsc, session);
2739                 break;
2740
2741         case CEPH_SESSION_RECALL_STATE:
2742                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2743                 break;
2744
2745         case CEPH_SESSION_FLUSHMSG:
2746                 send_flushmsg_ack(mdsc, session, seq);
2747                 break;
2748
2749         case CEPH_SESSION_FORCE_RO:
2750                 dout("force_session_readonly %p\n", session);
2751                 spin_lock(&session->s_cap_lock);
2752                 session->s_readonly = true;
2753                 spin_unlock(&session->s_cap_lock);
2754                 wake_up_session_caps(session, 0);
2755                 break;
2756
2757         case CEPH_SESSION_REJECT:
2758                 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
2759                 pr_info("mds%d rejected session\n", session->s_mds);
2760                 session->s_state = CEPH_MDS_SESSION_REJECTED;
2761                 cleanup_session_requests(mdsc, session);
2762                 remove_session_caps(session);
2763                 wake = 2; /* for good measure */
2764                 break;
2765
2766         default:
2767                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2768                 WARN_ON(1);
2769         }
2770
2771         mutex_unlock(&session->s_mutex);
2772         if (wake) {
2773                 mutex_lock(&mdsc->mutex);
2774                 __wake_requests(mdsc, &session->s_waiting);
2775                 if (wake == 2)
2776                         kick_requests(mdsc, mds);
2777                 mutex_unlock(&mdsc->mutex);
2778         }
2779         if (op == CEPH_SESSION_CLOSE)
2780                 ceph_put_mds_session(session);
2781         return;
2782
2783 bad:
2784         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2785                (int)msg->front.iov_len);
2786         ceph_msg_dump(msg);
2787         return;
2788 }
2789
2790
2791 /*
2792  * called under session->mutex.
2793  */
2794 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2795                                    struct ceph_mds_session *session)
2796 {
2797         struct ceph_mds_request *req, *nreq;
2798         struct rb_node *p;
2799         int err;
2800
2801         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2802
2803         mutex_lock(&mdsc->mutex);
2804         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2805                 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2806                 if (!err) {
2807                         ceph_msg_get(req->r_request);
2808                         ceph_con_send(&session->s_con, req->r_request);
2809                 }
2810         }
2811
2812         /*
2813          * also re-send old requests when MDS enters reconnect stage. So that MDS
2814          * can process completed request in clientreplay stage.
2815          */
2816         p = rb_first(&mdsc->request_tree);
2817         while (p) {
2818                 req = rb_entry(p, struct ceph_mds_request, r_node);
2819                 p = rb_next(p);
2820                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2821                         continue;
2822                 if (req->r_attempts == 0)
2823                         continue; /* only old requests */
2824                 if (req->r_session &&
2825                     req->r_session->s_mds == session->s_mds) {
2826                         err = __prepare_send_request(mdsc, req,
2827                                                      session->s_mds, true);
2828                         if (!err) {
2829                                 ceph_msg_get(req->r_request);
2830                                 ceph_con_send(&session->s_con, req->r_request);
2831                         }
2832                 }
2833         }
2834         mutex_unlock(&mdsc->mutex);
2835 }
2836
2837 /*
2838  * Encode information about a cap for a reconnect with the MDS.
2839  */
2840 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2841                           void *arg)
2842 {
2843         union {
2844                 struct ceph_mds_cap_reconnect v2;
2845                 struct ceph_mds_cap_reconnect_v1 v1;
2846         } rec;
2847         struct ceph_inode_info *ci = cap->ci;
2848         struct ceph_reconnect_state *recon_state = arg;
2849         struct ceph_pagelist *pagelist = recon_state->pagelist;
2850         char *path;
2851         int pathlen, err;
2852         u64 pathbase;
2853         u64 snap_follows;
2854         struct dentry *dentry;
2855
2856         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2857              inode, ceph_vinop(inode), cap, cap->cap_id,
2858              ceph_cap_string(cap->issued));
2859         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2860         if (err)
2861                 return err;
2862
2863         dentry = d_find_alias(inode);
2864         if (dentry) {
2865                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2866                 if (IS_ERR(path)) {
2867                         err = PTR_ERR(path);
2868                         goto out_dput;
2869                 }
2870         } else {
2871                 path = NULL;
2872                 pathlen = 0;
2873                 pathbase = 0;
2874         }
2875
2876         spin_lock(&ci->i_ceph_lock);
2877         cap->seq = 0;        /* reset cap seq */
2878         cap->issue_seq = 0;  /* and issue_seq */
2879         cap->mseq = 0;       /* and migrate_seq */
2880         cap->cap_gen = cap->session->s_cap_gen;
2881
2882         if (recon_state->msg_version >= 2) {
2883                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2884                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2885                 rec.v2.issued = cpu_to_le32(cap->issued);
2886                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2887                 rec.v2.pathbase = cpu_to_le64(pathbase);
2888                 rec.v2.flock_len = (__force __le32)
2889                         ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
2890         } else {
2891                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2892                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2893                 rec.v1.issued = cpu_to_le32(cap->issued);
2894                 rec.v1.size = cpu_to_le64(inode->i_size);
2895                 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2896                 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2897                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2898                 rec.v1.pathbase = cpu_to_le64(pathbase);
2899         }
2900
2901         if (list_empty(&ci->i_cap_snaps)) {
2902                 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
2903         } else {
2904                 struct ceph_cap_snap *capsnap =
2905                         list_first_entry(&ci->i_cap_snaps,
2906                                          struct ceph_cap_snap, ci_item);
2907                 snap_follows = capsnap->follows;
2908         }
2909         spin_unlock(&ci->i_ceph_lock);
2910
2911         if (recon_state->msg_version >= 2) {
2912                 int num_fcntl_locks, num_flock_locks;
2913                 struct ceph_filelock *flocks = NULL;
2914                 size_t struct_len, total_len = 0;
2915                 u8 struct_v = 0;
2916
2917 encode_again:
2918                 if (rec.v2.flock_len) {
2919                         ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2920                 } else {
2921                         num_fcntl_locks = 0;
2922                         num_flock_locks = 0;
2923                 }
2924                 if (num_fcntl_locks + num_flock_locks > 0) {
2925                         flocks = kmalloc((num_fcntl_locks + num_flock_locks) *
2926                                          sizeof(struct ceph_filelock), GFP_NOFS);
2927                         if (!flocks) {
2928                                 err = -ENOMEM;
2929                                 goto out_free;
2930                         }
2931                         err = ceph_encode_locks_to_buffer(inode, flocks,
2932                                                           num_fcntl_locks,
2933                                                           num_flock_locks);
2934                         if (err) {
2935                                 kfree(flocks);
2936                                 flocks = NULL;
2937                                 if (err == -ENOSPC)
2938                                         goto encode_again;
2939                                 goto out_free;
2940                         }
2941                 } else {
2942                         kfree(flocks);
2943                         flocks = NULL;
2944                 }
2945
2946                 if (recon_state->msg_version >= 3) {
2947                         /* version, compat_version and struct_len */
2948                         total_len = 2 * sizeof(u8) + sizeof(u32);
2949                         struct_v = 2;
2950                 }
2951                 /*
2952                  * number of encoded locks is stable, so copy to pagelist
2953                  */
2954                 struct_len = 2 * sizeof(u32) +
2955                             (num_fcntl_locks + num_flock_locks) *
2956                             sizeof(struct ceph_filelock);
2957                 rec.v2.flock_len = cpu_to_le32(struct_len);
2958
2959                 struct_len += sizeof(rec.v2);
2960                 struct_len += sizeof(u32) + pathlen;
2961
2962                 if (struct_v >= 2)
2963                         struct_len += sizeof(u64); /* snap_follows */
2964
2965                 total_len += struct_len;
2966                 err = ceph_pagelist_reserve(pagelist, total_len);
2967
2968                 if (!err) {
2969                         if (recon_state->msg_version >= 3) {
2970                                 ceph_pagelist_encode_8(pagelist, struct_v);
2971                                 ceph_pagelist_encode_8(pagelist, 1);
2972                                 ceph_pagelist_encode_32(pagelist, struct_len);
2973                         }
2974                         ceph_pagelist_encode_string(pagelist, path, pathlen);
2975                         ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
2976                         ceph_locks_to_pagelist(flocks, pagelist,
2977                                                num_fcntl_locks,
2978                                                num_flock_locks);
2979                         if (struct_v >= 2)
2980                                 ceph_pagelist_encode_64(pagelist, snap_follows);
2981                 }
2982                 kfree(flocks);
2983         } else {
2984                 size_t size = sizeof(u32) + pathlen + sizeof(rec.v1);
2985                 err = ceph_pagelist_reserve(pagelist, size);
2986                 if (!err) {
2987                         ceph_pagelist_encode_string(pagelist, path, pathlen);
2988                         ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
2989                 }
2990         }
2991
2992         recon_state->nr_caps++;
2993 out_free:
2994         kfree(path);
2995 out_dput:
2996         dput(dentry);
2997         return err;
2998 }
2999
3000
3001 /*
3002  * If an MDS fails and recovers, clients need to reconnect in order to
3003  * reestablish shared state.  This includes all caps issued through
3004  * this session _and_ the snap_realm hierarchy.  Because it's not
3005  * clear which snap realms the mds cares about, we send everything we
3006  * know about.. that ensures we'll then get any new info the
3007  * recovering MDS might have.
3008  *
3009  * This is a relatively heavyweight operation, but it's rare.
3010  *
3011  * called with mdsc->mutex held.
3012  */
3013 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3014                                struct ceph_mds_session *session)
3015 {
3016         struct ceph_msg *reply;
3017         struct rb_node *p;
3018         int mds = session->s_mds;
3019         int err = -ENOMEM;
3020         int s_nr_caps;
3021         struct ceph_pagelist *pagelist;
3022         struct ceph_reconnect_state recon_state;
3023         LIST_HEAD(dispose);
3024
3025         pr_info("mds%d reconnect start\n", mds);
3026
3027         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
3028         if (!pagelist)
3029                 goto fail_nopagelist;
3030         ceph_pagelist_init(pagelist);
3031
3032         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
3033         if (!reply)
3034                 goto fail_nomsg;
3035
3036         mutex_lock(&session->s_mutex);
3037         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3038         session->s_seq = 0;
3039
3040         dout("session %p state %s\n", session,
3041              ceph_session_state_name(session->s_state));
3042
3043         spin_lock(&session->s_gen_ttl_lock);
3044         session->s_cap_gen++;
3045         spin_unlock(&session->s_gen_ttl_lock);
3046
3047         spin_lock(&session->s_cap_lock);
3048         /* don't know if session is readonly */
3049         session->s_readonly = 0;
3050         /*
3051          * notify __ceph_remove_cap() that we are composing cap reconnect.
3052          * If a cap get released before being added to the cap reconnect,
3053          * __ceph_remove_cap() should skip queuing cap release.
3054          */
3055         session->s_cap_reconnect = 1;
3056         /* drop old cap expires; we're about to reestablish that state */
3057         detach_cap_releases(session, &dispose);
3058         spin_unlock(&session->s_cap_lock);
3059         dispose_cap_releases(mdsc, &dispose);
3060
3061         /* trim unused caps to reduce MDS's cache rejoin time */
3062         if (mdsc->fsc->sb->s_root)
3063                 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3064
3065         ceph_con_close(&session->s_con);
3066         ceph_con_open(&session->s_con,
3067                       CEPH_ENTITY_TYPE_MDS, mds,
3068                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3069
3070         /* replay unsafe requests */
3071         replay_unsafe_requests(mdsc, session);
3072
3073         down_read(&mdsc->snap_rwsem);
3074
3075         /* traverse this session's caps */
3076         s_nr_caps = session->s_nr_caps;
3077         err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
3078         if (err)
3079                 goto fail;
3080
3081         recon_state.nr_caps = 0;
3082         recon_state.pagelist = pagelist;
3083         if (session->s_con.peer_features & CEPH_FEATURE_MDSENC)
3084                 recon_state.msg_version = 3;
3085         else if (session->s_con.peer_features & CEPH_FEATURE_FLOCK)
3086                 recon_state.msg_version = 2;
3087         else
3088                 recon_state.msg_version = 1;
3089         err = iterate_session_caps(session, encode_caps_cb, &recon_state);
3090         if (err < 0)
3091                 goto fail;
3092
3093         spin_lock(&session->s_cap_lock);
3094         session->s_cap_reconnect = 0;
3095         spin_unlock(&session->s_cap_lock);
3096
3097         /*
3098          * snaprealms.  we provide mds with the ino, seq (version), and
3099          * parent for all of our realms.  If the mds has any newer info,
3100          * it will tell us.
3101          */
3102         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3103                 struct ceph_snap_realm *realm =
3104                         rb_entry(p, struct ceph_snap_realm, node);
3105                 struct ceph_mds_snaprealm_reconnect sr_rec;
3106
3107                 dout(" adding snap realm %llx seq %lld parent %llx\n",
3108                      realm->ino, realm->seq, realm->parent_ino);
3109                 sr_rec.ino = cpu_to_le64(realm->ino);
3110                 sr_rec.seq = cpu_to_le64(realm->seq);
3111                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3112                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3113                 if (err)
3114                         goto fail;
3115         }
3116
3117         reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3118
3119         /* raced with cap release? */
3120         if (s_nr_caps != recon_state.nr_caps) {
3121                 struct page *page = list_first_entry(&pagelist->head,
3122                                                      struct page, lru);
3123                 __le32 *addr = kmap_atomic(page);
3124                 *addr = cpu_to_le32(recon_state.nr_caps);
3125                 kunmap_atomic(addr);
3126         }
3127
3128         reply->hdr.data_len = cpu_to_le32(pagelist->length);
3129         ceph_msg_data_add_pagelist(reply, pagelist);
3130
3131         ceph_early_kick_flushing_caps(mdsc, session);
3132
3133         ceph_con_send(&session->s_con, reply);
3134
3135         mutex_unlock(&session->s_mutex);
3136
3137         mutex_lock(&mdsc->mutex);
3138         __wake_requests(mdsc, &session->s_waiting);
3139         mutex_unlock(&mdsc->mutex);
3140
3141         up_read(&mdsc->snap_rwsem);
3142         return;
3143
3144 fail:
3145         ceph_msg_put(reply);
3146         up_read(&mdsc->snap_rwsem);
3147         mutex_unlock(&session->s_mutex);
3148 fail_nomsg:
3149         ceph_pagelist_release(pagelist);
3150 fail_nopagelist:
3151         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3152         return;
3153 }
3154
3155
3156 /*
3157  * compare old and new mdsmaps, kicking requests
3158  * and closing out old connections as necessary
3159  *
3160  * called under mdsc->mutex.
3161  */
3162 static void check_new_map(struct ceph_mds_client *mdsc,
3163                           struct ceph_mdsmap *newmap,
3164                           struct ceph_mdsmap *oldmap)
3165 {
3166         int i;
3167         int oldstate, newstate;
3168         struct ceph_mds_session *s;
3169
3170         dout("check_new_map new %u old %u\n",
3171              newmap->m_epoch, oldmap->m_epoch);
3172
3173         for (i = 0; i < oldmap->m_num_mds && i < mdsc->max_sessions; i++) {
3174                 if (!mdsc->sessions[i])
3175                         continue;
3176                 s = mdsc->sessions[i];
3177                 oldstate = ceph_mdsmap_get_state(oldmap, i);
3178                 newstate = ceph_mdsmap_get_state(newmap, i);
3179
3180                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3181                      i, ceph_mds_state_name(oldstate),
3182                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3183                      ceph_mds_state_name(newstate),
3184                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3185                      ceph_session_state_name(s->s_state));
3186
3187                 if (i >= newmap->m_num_mds ||
3188                     memcmp(ceph_mdsmap_get_addr(oldmap, i),
3189                            ceph_mdsmap_get_addr(newmap, i),
3190                            sizeof(struct ceph_entity_addr))) {
3191                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3192                                 /* the session never opened, just close it
3193                                  * out now */
3194                                 get_session(s);
3195                                 __unregister_session(mdsc, s);
3196                                 __wake_requests(mdsc, &s->s_waiting);
3197                                 ceph_put_mds_session(s);
3198                         } else if (i >= newmap->m_num_mds) {
3199                                 /* force close session for stopped mds */
3200                                 get_session(s);
3201                                 __unregister_session(mdsc, s);
3202                                 __wake_requests(mdsc, &s->s_waiting);
3203                                 kick_requests(mdsc, i);
3204                                 mutex_unlock(&mdsc->mutex);
3205
3206                                 mutex_lock(&s->s_mutex);
3207                                 cleanup_session_requests(mdsc, s);
3208                                 remove_session_caps(s);
3209                                 mutex_unlock(&s->s_mutex);
3210
3211                                 ceph_put_mds_session(s);
3212
3213                                 mutex_lock(&mdsc->mutex);
3214                         } else {
3215                                 /* just close it */
3216                                 mutex_unlock(&mdsc->mutex);
3217                                 mutex_lock(&s->s_mutex);
3218                                 mutex_lock(&mdsc->mutex);
3219                                 ceph_con_close(&s->s_con);
3220                                 mutex_unlock(&s->s_mutex);
3221                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3222                         }
3223                 } else if (oldstate == newstate) {
3224                         continue;  /* nothing new with this mds */
3225                 }
3226
3227                 /*
3228                  * send reconnect?
3229                  */
3230                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3231                     newstate >= CEPH_MDS_STATE_RECONNECT) {
3232                         mutex_unlock(&mdsc->mutex);
3233                         send_mds_reconnect(mdsc, s);
3234                         mutex_lock(&mdsc->mutex);
3235                 }
3236
3237                 /*
3238                  * kick request on any mds that has gone active.
3239                  */
3240                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3241                     newstate >= CEPH_MDS_STATE_ACTIVE) {
3242                         if (oldstate != CEPH_MDS_STATE_CREATING &&
3243                             oldstate != CEPH_MDS_STATE_STARTING)
3244                                 pr_info("mds%d recovery completed\n", s->s_mds);
3245                         kick_requests(mdsc, i);
3246                         ceph_kick_flushing_caps(mdsc, s);
3247                         wake_up_session_caps(s, 1);
3248                 }
3249         }
3250
3251         for (i = 0; i < newmap->m_num_mds && i < mdsc->max_sessions; i++) {
3252                 s = mdsc->sessions[i];
3253                 if (!s)
3254                         continue;
3255                 if (!ceph_mdsmap_is_laggy(newmap, i))
3256                         continue;
3257                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3258                     s->s_state == CEPH_MDS_SESSION_HUNG ||
3259                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
3260                         dout(" connecting to export targets of laggy mds%d\n",
3261                              i);
3262                         __open_export_target_sessions(mdsc, s);
3263                 }
3264         }
3265 }
3266
3267
3268
3269 /*
3270  * leases
3271  */
3272
3273 /*
3274  * caller must hold session s_mutex, dentry->d_lock
3275  */
3276 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3277 {
3278         struct ceph_dentry_info *di = ceph_dentry(dentry);
3279
3280         ceph_put_mds_session(di->lease_session);
3281         di->lease_session = NULL;
3282 }
3283
3284 static void handle_lease(struct ceph_mds_client *mdsc,
3285                          struct ceph_mds_session *session,
3286                          struct ceph_msg *msg)
3287 {
3288         struct super_block *sb = mdsc->fsc->sb;
3289         struct inode *inode;
3290         struct dentry *parent, *dentry;
3291         struct ceph_dentry_info *di;
3292         int mds = session->s_mds;
3293         struct ceph_mds_lease *h = msg->front.iov_base;
3294         u32 seq;
3295         struct ceph_vino vino;
3296         struct qstr dname;
3297         int release = 0;
3298
3299         dout("handle_lease from mds%d\n", mds);
3300
3301         /* decode */
3302         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3303                 goto bad;
3304         vino.ino = le64_to_cpu(h->ino);
3305         vino.snap = CEPH_NOSNAP;
3306         seq = le32_to_cpu(h->seq);
3307         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3308         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3309         if (dname.len != get_unaligned_le32(h+1))
3310                 goto bad;
3311
3312         /* lookup inode */
3313         inode = ceph_find_inode(sb, vino);
3314         dout("handle_lease %s, ino %llx %p %.*s\n",
3315              ceph_lease_op_name(h->action), vino.ino, inode,
3316              dname.len, dname.name);
3317
3318         mutex_lock(&session->s_mutex);
3319         session->s_seq++;
3320
3321         if (!inode) {
3322                 dout("handle_lease no inode %llx\n", vino.ino);
3323                 goto release;
3324         }
3325
3326         /* dentry */
3327         parent = d_find_alias(inode);
3328         if (!parent) {
3329                 dout("no parent dentry on inode %p\n", inode);
3330                 WARN_ON(1);
3331                 goto release;  /* hrm... */
3332         }
3333         dname.hash = full_name_hash(parent, dname.name, dname.len);
3334         dentry = d_lookup(parent, &dname);
3335         dput(parent);
3336         if (!dentry)
3337                 goto release;
3338
3339         spin_lock(&dentry->d_lock);
3340         di = ceph_dentry(dentry);
3341         switch (h->action) {
3342         case CEPH_MDS_LEASE_REVOKE:
3343                 if (di->lease_session == session) {
3344                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3345                                 h->seq = cpu_to_le32(di->lease_seq);
3346                         __ceph_mdsc_drop_dentry_lease(dentry);
3347                 }
3348                 release = 1;
3349                 break;
3350
3351         case CEPH_MDS_LEASE_RENEW:
3352                 if (di->lease_session == session &&
3353                     di->lease_gen == session->s_cap_gen &&
3354                     di->lease_renew_from &&
3355                     di->lease_renew_after == 0) {
3356                         unsigned long duration =
3357                                 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3358
3359                         di->lease_seq = seq;
3360                         di->time = di->lease_renew_from + duration;
3361                         di->lease_renew_after = di->lease_renew_from +
3362                                 (duration >> 1);
3363                         di->lease_renew_from = 0;
3364                 }
3365                 break;
3366         }
3367         spin_unlock(&dentry->d_lock);
3368         dput(dentry);
3369
3370         if (!release)
3371                 goto out;
3372
3373 release:
3374         /* let's just reuse the same message */
3375         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3376         ceph_msg_get(msg);
3377         ceph_con_send(&session->s_con, msg);
3378
3379 out:
3380         iput(inode);
3381         mutex_unlock(&session->s_mutex);
3382         return;
3383
3384 bad:
3385         pr_err("corrupt lease message\n");
3386         ceph_msg_dump(msg);
3387 }
3388
3389 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3390                               struct inode *inode,
3391                               struct dentry *dentry, char action,
3392                               u32 seq)
3393 {
3394         struct ceph_msg *msg;
3395         struct ceph_mds_lease *lease;
3396         int len = sizeof(*lease) + sizeof(u32);
3397         int dnamelen = 0;
3398
3399         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3400              inode, dentry, ceph_lease_op_name(action), session->s_mds);
3401         dnamelen = dentry->d_name.len;
3402         len += dnamelen;
3403
3404         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3405         if (!msg)
3406                 return;
3407         lease = msg->front.iov_base;
3408         lease->action = action;
3409         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3410         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3411         lease->seq = cpu_to_le32(seq);
3412         put_unaligned_le32(dnamelen, lease + 1);
3413         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3414
3415         /*
3416          * if this is a preemptive lease RELEASE, no need to
3417          * flush request stream, since the actual request will
3418          * soon follow.
3419          */
3420         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3421
3422         ceph_con_send(&session->s_con, msg);
3423 }
3424
3425 /*
3426  * drop all leases (and dentry refs) in preparation for umount
3427  */
3428 static void drop_leases(struct ceph_mds_client *mdsc)
3429 {
3430         int i;
3431
3432         dout("drop_leases\n");
3433         mutex_lock(&mdsc->mutex);
3434         for (i = 0; i < mdsc->max_sessions; i++) {
3435                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3436                 if (!s)
3437                         continue;
3438                 mutex_unlock(&mdsc->mutex);
3439                 mutex_lock(&s->s_mutex);
3440                 mutex_unlock(&s->s_mutex);
3441                 ceph_put_mds_session(s);
3442                 mutex_lock(&mdsc->mutex);
3443         }
3444         mutex_unlock(&mdsc->mutex);
3445 }
3446
3447
3448
3449 /*
3450  * delayed work -- periodically trim expired leases, renew caps with mds
3451  */
3452 static void schedule_delayed(struct ceph_mds_client *mdsc)
3453 {
3454         int delay = 5;
3455         unsigned hz = round_jiffies_relative(HZ * delay);
3456         schedule_delayed_work(&mdsc->delayed_work, hz);
3457 }
3458
3459 static void delayed_work(struct work_struct *work)
3460 {
3461         int i;
3462         struct ceph_mds_client *mdsc =
3463                 container_of(work, struct ceph_mds_client, delayed_work.work);
3464         int renew_interval;
3465         int renew_caps;
3466
3467         dout("mdsc delayed_work\n");
3468         ceph_check_delayed_caps(mdsc);
3469
3470         mutex_lock(&mdsc->mutex);
3471         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3472         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3473                                    mdsc->last_renew_caps);
3474         if (renew_caps)
3475                 mdsc->last_renew_caps = jiffies;
3476
3477         for (i = 0; i < mdsc->max_sessions; i++) {
3478                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3479                 if (!s)
3480                         continue;
3481                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3482                         dout("resending session close request for mds%d\n",
3483                              s->s_mds);
3484                         request_close_session(mdsc, s);
3485                         ceph_put_mds_session(s);
3486                         continue;
3487                 }
3488                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3489                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3490                                 s->s_state = CEPH_MDS_SESSION_HUNG;
3491                                 pr_info("mds%d hung\n", s->s_mds);
3492                         }
3493                 }
3494                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3495                         /* this mds is failed or recovering, just wait */
3496                         ceph_put_mds_session(s);
3497                         continue;
3498                 }
3499                 mutex_unlock(&mdsc->mutex);
3500
3501                 mutex_lock(&s->s_mutex);
3502                 if (renew_caps)
3503                         send_renew_caps(mdsc, s);
3504                 else
3505                         ceph_con_keepalive(&s->s_con);
3506                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3507                     s->s_state == CEPH_MDS_SESSION_HUNG)
3508                         ceph_send_cap_releases(mdsc, s);
3509                 mutex_unlock(&s->s_mutex);
3510                 ceph_put_mds_session(s);
3511
3512                 mutex_lock(&mdsc->mutex);
3513         }
3514         mutex_unlock(&mdsc->mutex);
3515
3516         schedule_delayed(mdsc);
3517 }
3518
3519 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3520
3521 {
3522         struct ceph_mds_client *mdsc;
3523
3524         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3525         if (!mdsc)
3526                 return -ENOMEM;
3527         mdsc->fsc = fsc;
3528         fsc->mdsc = mdsc;
3529         mutex_init(&mdsc->mutex);
3530         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3531         if (!mdsc->mdsmap) {
3532                 kfree(mdsc);
3533                 return -ENOMEM;
3534         }
3535
3536         init_completion(&mdsc->safe_umount_waiters);
3537         init_waitqueue_head(&mdsc->session_close_wq);
3538         INIT_LIST_HEAD(&mdsc->waiting_for_map);
3539         mdsc->sessions = NULL;
3540         atomic_set(&mdsc->num_sessions, 0);
3541         mdsc->max_sessions = 0;
3542         mdsc->stopping = 0;
3543         mdsc->last_snap_seq = 0;
3544         init_rwsem(&mdsc->snap_rwsem);
3545         mdsc->snap_realms = RB_ROOT;
3546         INIT_LIST_HEAD(&mdsc->snap_empty);
3547         spin_lock_init(&mdsc->snap_empty_lock);
3548         mdsc->last_tid = 0;
3549         mdsc->oldest_tid = 0;
3550         mdsc->request_tree = RB_ROOT;
3551         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3552         mdsc->last_renew_caps = jiffies;
3553         INIT_LIST_HEAD(&mdsc->cap_delay_list);
3554         spin_lock_init(&mdsc->cap_delay_lock);
3555         INIT_LIST_HEAD(&mdsc->snap_flush_list);
3556         spin_lock_init(&mdsc->snap_flush_lock);
3557         mdsc->last_cap_flush_tid = 1;
3558         INIT_LIST_HEAD(&mdsc->cap_flush_list);
3559         INIT_LIST_HEAD(&mdsc->cap_dirty);
3560         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3561         mdsc->num_cap_flushing = 0;
3562         spin_lock_init(&mdsc->cap_dirty_lock);
3563         init_waitqueue_head(&mdsc->cap_flushing_wq);
3564         spin_lock_init(&mdsc->dentry_lru_lock);
3565         INIT_LIST_HEAD(&mdsc->dentry_lru);
3566
3567         ceph_caps_init(mdsc);
3568         ceph_adjust_min_caps(mdsc, fsc->min_caps);
3569
3570         init_rwsem(&mdsc->pool_perm_rwsem);
3571         mdsc->pool_perm_tree = RB_ROOT;
3572
3573         strncpy(mdsc->nodename, utsname()->nodename,
3574                 sizeof(mdsc->nodename) - 1);
3575         return 0;
3576 }
3577
3578 /*
3579  * Wait for safe replies on open mds requests.  If we time out, drop
3580  * all requests from the tree to avoid dangling dentry refs.
3581  */
3582 static void wait_requests(struct ceph_mds_client *mdsc)
3583 {
3584         struct ceph_options *opts = mdsc->fsc->client->options;
3585         struct ceph_mds_request *req;
3586
3587         mutex_lock(&mdsc->mutex);
3588         if (__get_oldest_req(mdsc)) {
3589                 mutex_unlock(&mdsc->mutex);
3590
3591                 dout("wait_requests waiting for requests\n");
3592                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3593                                     ceph_timeout_jiffies(opts->mount_timeout));
3594
3595                 /* tear down remaining requests */
3596                 mutex_lock(&mdsc->mutex);
3597                 while ((req = __get_oldest_req(mdsc))) {
3598                         dout("wait_requests timed out on tid %llu\n",
3599                              req->r_tid);
3600                         __unregister_request(mdsc, req);
3601                 }
3602         }
3603         mutex_unlock(&mdsc->mutex);
3604         dout("wait_requests done\n");
3605 }
3606
3607 /*
3608  * called before mount is ro, and before dentries are torn down.
3609  * (hmm, does this still race with new lookups?)
3610  */
3611 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3612 {
3613         dout("pre_umount\n");
3614         mdsc->stopping = 1;
3615
3616         drop_leases(mdsc);
3617         ceph_flush_dirty_caps(mdsc);
3618         wait_requests(mdsc);
3619
3620         /*
3621          * wait for reply handlers to drop their request refs and
3622          * their inode/dcache refs
3623          */
3624         ceph_msgr_flush();
3625 }
3626
3627 /*
3628  * wait for all write mds requests to flush.
3629  */
3630 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3631 {
3632         struct ceph_mds_request *req = NULL, *nextreq;
3633         struct rb_node *n;
3634
3635         mutex_lock(&mdsc->mutex);
3636         dout("wait_unsafe_requests want %lld\n", want_tid);
3637 restart:
3638         req = __get_oldest_req(mdsc);
3639         while (req && req->r_tid <= want_tid) {
3640                 /* find next request */
3641                 n = rb_next(&req->r_node);
3642                 if (n)
3643                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3644                 else
3645                         nextreq = NULL;
3646                 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3647                     (req->r_op & CEPH_MDS_OP_WRITE)) {
3648                         /* write op */
3649                         ceph_mdsc_get_request(req);
3650                         if (nextreq)
3651                                 ceph_mdsc_get_request(nextreq);
3652                         mutex_unlock(&mdsc->mutex);
3653                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
3654                              req->r_tid, want_tid);
3655                         wait_for_completion(&req->r_safe_completion);
3656                         mutex_lock(&mdsc->mutex);
3657                         ceph_mdsc_put_request(req);
3658                         if (!nextreq)
3659                                 break;  /* next dne before, so we're done! */
3660                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
3661                                 /* next request was removed from tree */
3662                                 ceph_mdsc_put_request(nextreq);
3663                                 goto restart;
3664                         }
3665                         ceph_mdsc_put_request(nextreq);  /* won't go away */
3666                 }
3667                 req = nextreq;
3668         }
3669         mutex_unlock(&mdsc->mutex);
3670         dout("wait_unsafe_requests done\n");
3671 }
3672
3673 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3674 {
3675         u64 want_tid, want_flush;
3676
3677         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3678                 return;
3679
3680         dout("sync\n");
3681         mutex_lock(&mdsc->mutex);
3682         want_tid = mdsc->last_tid;
3683         mutex_unlock(&mdsc->mutex);
3684
3685         ceph_flush_dirty_caps(mdsc);
3686         spin_lock(&mdsc->cap_dirty_lock);
3687         want_flush = mdsc->last_cap_flush_tid;
3688         if (!list_empty(&mdsc->cap_flush_list)) {
3689                 struct ceph_cap_flush *cf =
3690                         list_last_entry(&mdsc->cap_flush_list,
3691                                         struct ceph_cap_flush, g_list);
3692                 cf->wake = true;
3693         }
3694         spin_unlock(&mdsc->cap_dirty_lock);
3695
3696         dout("sync want tid %lld flush_seq %lld\n",
3697              want_tid, want_flush);
3698
3699         wait_unsafe_requests(mdsc, want_tid);
3700         wait_caps_flush(mdsc, want_flush);
3701 }
3702
3703 /*
3704  * true if all sessions are closed, or we force unmount
3705  */
3706 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
3707 {
3708         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3709                 return true;
3710         return atomic_read(&mdsc->num_sessions) <= skipped;
3711 }
3712
3713 /*
3714  * called after sb is ro.
3715  */
3716 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3717 {
3718         struct ceph_options *opts = mdsc->fsc->client->options;
3719         struct ceph_mds_session *session;
3720         int i;
3721         int skipped = 0;
3722
3723         dout("close_sessions\n");
3724
3725         /* close sessions */
3726         mutex_lock(&mdsc->mutex);
3727         for (i = 0; i < mdsc->max_sessions; i++) {
3728                 session = __ceph_lookup_mds_session(mdsc, i);
3729                 if (!session)
3730                         continue;
3731                 mutex_unlock(&mdsc->mutex);
3732                 mutex_lock(&session->s_mutex);
3733                 if (__close_session(mdsc, session) <= 0)
3734                         skipped++;
3735                 mutex_unlock(&session->s_mutex);
3736                 ceph_put_mds_session(session);
3737                 mutex_lock(&mdsc->mutex);
3738         }
3739         mutex_unlock(&mdsc->mutex);
3740
3741         dout("waiting for sessions to close\n");
3742         wait_event_timeout(mdsc->session_close_wq,
3743                            done_closing_sessions(mdsc, skipped),
3744                            ceph_timeout_jiffies(opts->mount_timeout));
3745
3746         /* tear down remaining sessions */
3747         mutex_lock(&mdsc->mutex);
3748         for (i = 0; i < mdsc->max_sessions; i++) {
3749                 if (mdsc->sessions[i]) {
3750                         session = get_session(mdsc->sessions[i]);
3751                         __unregister_session(mdsc, session);
3752                         mutex_unlock(&mdsc->mutex);
3753                         mutex_lock(&session->s_mutex);
3754                         remove_session_caps(session);
3755                         mutex_unlock(&session->s_mutex);
3756                         ceph_put_mds_session(session);
3757                         mutex_lock(&mdsc->mutex);
3758                 }
3759         }
3760         WARN_ON(!list_empty(&mdsc->cap_delay_list));
3761         mutex_unlock(&mdsc->mutex);
3762
3763         ceph_cleanup_empty_realms(mdsc);
3764
3765         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3766
3767         dout("stopped\n");
3768 }
3769
3770 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
3771 {
3772         struct ceph_mds_session *session;
3773         int mds;
3774
3775         dout("force umount\n");
3776
3777         mutex_lock(&mdsc->mutex);
3778         for (mds = 0; mds < mdsc->max_sessions; mds++) {
3779                 session = __ceph_lookup_mds_session(mdsc, mds);
3780                 if (!session)
3781                         continue;
3782                 mutex_unlock(&mdsc->mutex);
3783                 mutex_lock(&session->s_mutex);
3784                 __close_session(mdsc, session);
3785                 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
3786                         cleanup_session_requests(mdsc, session);
3787                         remove_session_caps(session);
3788                 }
3789                 mutex_unlock(&session->s_mutex);
3790                 ceph_put_mds_session(session);
3791                 mutex_lock(&mdsc->mutex);
3792                 kick_requests(mdsc, mds);
3793         }
3794         __wake_requests(mdsc, &mdsc->waiting_for_map);
3795         mutex_unlock(&mdsc->mutex);
3796 }
3797
3798 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3799 {
3800         dout("stop\n");
3801         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3802         if (mdsc->mdsmap)
3803                 ceph_mdsmap_destroy(mdsc->mdsmap);
3804         kfree(mdsc->sessions);
3805         ceph_caps_finalize(mdsc);
3806         ceph_pool_perm_destroy(mdsc);
3807 }
3808
3809 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3810 {
3811         struct ceph_mds_client *mdsc = fsc->mdsc;
3812         dout("mdsc_destroy %p\n", mdsc);
3813
3814         /* flush out any connection work with references to us */
3815         ceph_msgr_flush();
3816
3817         ceph_mdsc_stop(mdsc);
3818
3819         fsc->mdsc = NULL;
3820         kfree(mdsc);
3821         dout("mdsc_destroy %p done\n", mdsc);
3822 }
3823
3824 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3825 {
3826         struct ceph_fs_client *fsc = mdsc->fsc;
3827         const char *mds_namespace = fsc->mount_options->mds_namespace;
3828         void *p = msg->front.iov_base;
3829         void *end = p + msg->front.iov_len;
3830         u32 epoch;
3831         u32 map_len;
3832         u32 num_fs;
3833         u32 mount_fscid = (u32)-1;
3834         u8 struct_v, struct_cv;
3835         int err = -EINVAL;
3836
3837         ceph_decode_need(&p, end, sizeof(u32), bad);
3838         epoch = ceph_decode_32(&p);
3839
3840         dout("handle_fsmap epoch %u\n", epoch);
3841
3842         ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3843         struct_v = ceph_decode_8(&p);
3844         struct_cv = ceph_decode_8(&p);
3845         map_len = ceph_decode_32(&p);
3846
3847         ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
3848         p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
3849
3850         num_fs = ceph_decode_32(&p);
3851         while (num_fs-- > 0) {
3852                 void *info_p, *info_end;
3853                 u32 info_len;
3854                 u8 info_v, info_cv;
3855                 u32 fscid, namelen;
3856
3857                 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3858                 info_v = ceph_decode_8(&p);
3859                 info_cv = ceph_decode_8(&p);
3860                 info_len = ceph_decode_32(&p);
3861                 ceph_decode_need(&p, end, info_len, bad);
3862                 info_p = p;
3863                 info_end = p + info_len;
3864                 p = info_end;
3865
3866                 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
3867                 fscid = ceph_decode_32(&info_p);
3868                 namelen = ceph_decode_32(&info_p);
3869                 ceph_decode_need(&info_p, info_end, namelen, bad);
3870
3871                 if (mds_namespace &&
3872                     strlen(mds_namespace) == namelen &&
3873                     !strncmp(mds_namespace, (char *)info_p, namelen)) {
3874                         mount_fscid = fscid;
3875                         break;
3876                 }
3877         }
3878
3879         ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
3880         if (mount_fscid != (u32)-1) {
3881                 fsc->client->monc.fs_cluster_id = mount_fscid;
3882                 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
3883                                    0, true);
3884                 ceph_monc_renew_subs(&fsc->client->monc);
3885         } else {
3886                 err = -ENOENT;
3887                 goto err_out;
3888         }
3889         return;
3890
3891 bad:
3892         pr_err("error decoding fsmap\n");
3893 err_out:
3894         mutex_lock(&mdsc->mutex);
3895         mdsc->mdsmap_err = err;
3896         __wake_requests(mdsc, &mdsc->waiting_for_map);
3897         mutex_unlock(&mdsc->mutex);
3898 }
3899
3900 /*
3901  * handle mds map update.
3902  */
3903 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3904 {
3905         u32 epoch;
3906         u32 maplen;
3907         void *p = msg->front.iov_base;
3908         void *end = p + msg->front.iov_len;
3909         struct ceph_mdsmap *newmap, *oldmap;
3910         struct ceph_fsid fsid;
3911         int err = -EINVAL;
3912
3913         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3914         ceph_decode_copy(&p, &fsid, sizeof(fsid));
3915         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3916                 return;
3917         epoch = ceph_decode_32(&p);
3918         maplen = ceph_decode_32(&p);
3919         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3920
3921         /* do we need it? */
3922         mutex_lock(&mdsc->mutex);
3923         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3924                 dout("handle_map epoch %u <= our %u\n",
3925                      epoch, mdsc->mdsmap->m_epoch);
3926                 mutex_unlock(&mdsc->mutex);
3927                 return;
3928         }
3929
3930         newmap = ceph_mdsmap_decode(&p, end);
3931         if (IS_ERR(newmap)) {
3932                 err = PTR_ERR(newmap);
3933                 goto bad_unlock;
3934         }
3935
3936         /* swap into place */
3937         if (mdsc->mdsmap) {
3938                 oldmap = mdsc->mdsmap;
3939                 mdsc->mdsmap = newmap;
3940                 check_new_map(mdsc, newmap, oldmap);
3941                 ceph_mdsmap_destroy(oldmap);
3942         } else {
3943                 mdsc->mdsmap = newmap;  /* first mds map */
3944         }
3945         mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3946
3947         __wake_requests(mdsc, &mdsc->waiting_for_map);
3948         ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
3949                           mdsc->mdsmap->m_epoch);
3950
3951         mutex_unlock(&mdsc->mutex);
3952         schedule_delayed(mdsc);
3953         return;
3954
3955 bad_unlock:
3956         mutex_unlock(&mdsc->mutex);
3957 bad:
3958         pr_err("error decoding mdsmap %d\n", err);
3959         return;
3960 }
3961
3962 static struct ceph_connection *con_get(struct ceph_connection *con)
3963 {
3964         struct ceph_mds_session *s = con->private;
3965
3966         if (get_session(s)) {
3967                 dout("mdsc con_get %p ok (%d)\n", s, refcount_read(&s->s_ref));
3968                 return con;
3969         }
3970         dout("mdsc con_get %p FAIL\n", s);
3971         return NULL;
3972 }
3973
3974 static void con_put(struct ceph_connection *con)
3975 {
3976         struct ceph_mds_session *s = con->private;
3977
3978         dout("mdsc con_put %p (%d)\n", s, refcount_read(&s->s_ref) - 1);
3979         ceph_put_mds_session(s);
3980 }
3981
3982 /*
3983  * if the client is unresponsive for long enough, the mds will kill
3984  * the session entirely.
3985  */
3986 static void peer_reset(struct ceph_connection *con)
3987 {
3988         struct ceph_mds_session *s = con->private;
3989         struct ceph_mds_client *mdsc = s->s_mdsc;
3990
3991         pr_warn("mds%d closed our session\n", s->s_mds);
3992         send_mds_reconnect(mdsc, s);
3993 }
3994
3995 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3996 {
3997         struct ceph_mds_session *s = con->private;
3998         struct ceph_mds_client *mdsc = s->s_mdsc;
3999         int type = le16_to_cpu(msg->hdr.type);
4000
4001         mutex_lock(&mdsc->mutex);
4002         if (__verify_registered_session(mdsc, s) < 0) {
4003                 mutex_unlock(&mdsc->mutex);
4004                 goto out;
4005         }
4006         mutex_unlock(&mdsc->mutex);
4007
4008         switch (type) {
4009         case CEPH_MSG_MDS_MAP:
4010                 ceph_mdsc_handle_mdsmap(mdsc, msg);
4011                 break;
4012         case CEPH_MSG_FS_MAP_USER:
4013                 ceph_mdsc_handle_fsmap(mdsc, msg);
4014                 break;
4015         case CEPH_MSG_CLIENT_SESSION:
4016                 handle_session(s, msg);
4017                 break;
4018         case CEPH_MSG_CLIENT_REPLY:
4019                 handle_reply(s, msg);
4020                 break;
4021         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
4022                 handle_forward(mdsc, s, msg);
4023                 break;
4024         case CEPH_MSG_CLIENT_CAPS:
4025                 ceph_handle_caps(s, msg);
4026                 break;
4027         case CEPH_MSG_CLIENT_SNAP:
4028                 ceph_handle_snap(mdsc, s, msg);
4029                 break;
4030         case CEPH_MSG_CLIENT_LEASE:
4031                 handle_lease(mdsc, s, msg);
4032                 break;
4033
4034         default:
4035                 pr_err("received unknown message type %d %s\n", type,
4036                        ceph_msg_type_name(type));
4037         }
4038 out:
4039         ceph_msg_put(msg);
4040 }
4041
4042 /*
4043  * authentication
4044  */
4045
4046 /*
4047  * Note: returned pointer is the address of a structure that's
4048  * managed separately.  Caller must *not* attempt to free it.
4049  */
4050 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
4051                                         int *proto, int force_new)
4052 {
4053         struct ceph_mds_session *s = con->private;
4054         struct ceph_mds_client *mdsc = s->s_mdsc;
4055         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4056         struct ceph_auth_handshake *auth = &s->s_auth;
4057
4058         if (force_new && auth->authorizer) {
4059                 ceph_auth_destroy_authorizer(auth->authorizer);
4060                 auth->authorizer = NULL;
4061         }
4062         if (!auth->authorizer) {
4063                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4064                                                       auth);
4065                 if (ret)
4066                         return ERR_PTR(ret);
4067         } else {
4068                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4069                                                       auth);
4070                 if (ret)
4071                         return ERR_PTR(ret);
4072         }
4073         *proto = ac->protocol;
4074
4075         return auth;
4076 }
4077
4078
4079 static int verify_authorizer_reply(struct ceph_connection *con)
4080 {
4081         struct ceph_mds_session *s = con->private;
4082         struct ceph_mds_client *mdsc = s->s_mdsc;
4083         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4084
4085         return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
4086 }
4087
4088 static int invalidate_authorizer(struct ceph_connection *con)
4089 {
4090         struct ceph_mds_session *s = con->private;
4091         struct ceph_mds_client *mdsc = s->s_mdsc;
4092         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4093
4094         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
4095
4096         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
4097 }
4098
4099 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
4100                                 struct ceph_msg_header *hdr, int *skip)
4101 {
4102         struct ceph_msg *msg;
4103         int type = (int) le16_to_cpu(hdr->type);
4104         int front_len = (int) le32_to_cpu(hdr->front_len);
4105
4106         if (con->in_msg)
4107                 return con->in_msg;
4108
4109         *skip = 0;
4110         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
4111         if (!msg) {
4112                 pr_err("unable to allocate msg type %d len %d\n",
4113                        type, front_len);
4114                 return NULL;
4115         }
4116
4117         return msg;
4118 }
4119
4120 static int mds_sign_message(struct ceph_msg *msg)
4121 {
4122        struct ceph_mds_session *s = msg->con->private;
4123        struct ceph_auth_handshake *auth = &s->s_auth;
4124
4125        return ceph_auth_sign_message(auth, msg);
4126 }
4127
4128 static int mds_check_message_signature(struct ceph_msg *msg)
4129 {
4130        struct ceph_mds_session *s = msg->con->private;
4131        struct ceph_auth_handshake *auth = &s->s_auth;
4132
4133        return ceph_auth_check_message_signature(auth, msg);
4134 }
4135
4136 static const struct ceph_connection_operations mds_con_ops = {
4137         .get = con_get,
4138         .put = con_put,
4139         .dispatch = dispatch,
4140         .get_authorizer = get_authorizer,
4141         .verify_authorizer_reply = verify_authorizer_reply,
4142         .invalidate_authorizer = invalidate_authorizer,
4143         .peer_reset = peer_reset,
4144         .alloc_msg = mds_alloc_msg,
4145         .sign_message = mds_sign_message,
4146         .check_message_signature = mds_check_message_signature,
4147 };
4148
4149 /* eof */