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  45  * are no longer valid.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 /* caller holds s_cap_lock, we drop it */
1043 static void cleanup_cap_releases(struct ceph_mds_client *mdsc,
1044                                  struct ceph_mds_session *session)
1045         __releases(session->s_cap_lock)
1046 {
1047         LIST_HEAD(tmp_list);
1048         list_splice_init(&session->s_cap_releases, &tmp_list);
1049         session->s_num_cap_releases = 0;
1050         spin_unlock(&session->s_cap_lock);
1051
1052         dout("cleanup_cap_releases mds%d\n", session->s_mds);
1053         while (!list_empty(&tmp_list)) {
1054                 struct ceph_cap *cap;
1055                 /* zero out the in-progress message */
1056                 cap = list_first_entry(&tmp_list,
1057                                         struct ceph_cap, session_caps);
1058                 list_del(&cap->session_caps);
1059                 ceph_put_cap(mdsc, cap);
1060         }
1061 }
1062
1063 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1064                                      struct ceph_mds_session *session)
1065 {
1066         struct ceph_mds_request *req;
1067         struct rb_node *p;
1068
1069         dout("cleanup_session_requests mds%d\n", session->s_mds);
1070         mutex_lock(&mdsc->mutex);
1071         while (!list_empty(&session->s_unsafe)) {
1072                 req = list_first_entry(&session->s_unsafe,
1073                                        struct ceph_mds_request, r_unsafe_item);
1074                 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1075                                     req->r_tid);
1076                 __unregister_request(mdsc, req);
1077         }
1078         /* zero r_attempts, so kick_requests() will re-send requests */
1079         p = rb_first(&mdsc->request_tree);
1080         while (p) {
1081                 req = rb_entry(p, struct ceph_mds_request, r_node);
1082                 p = rb_next(p);
1083                 if (req->r_session &&
1084                     req->r_session->s_mds == session->s_mds)
1085                         req->r_attempts = 0;
1086         }
1087         mutex_unlock(&mdsc->mutex);
1088 }
1089
1090 /*
1091  * Helper to safely iterate over all caps associated with a session, with
1092  * special care taken to handle a racing __ceph_remove_cap().
1093  *
1094  * Caller must hold session s_mutex.
1095  */
1096 static int iterate_session_caps(struct ceph_mds_session *session,
1097                                  int (*cb)(struct inode *, struct ceph_cap *,
1098                                             void *), void *arg)
1099 {
1100         struct list_head *p;
1101         struct ceph_cap *cap;
1102         struct inode *inode, *last_inode = NULL;
1103         struct ceph_cap *old_cap = NULL;
1104         int ret;
1105
1106         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1107         spin_lock(&session->s_cap_lock);
1108         p = session->s_caps.next;
1109         while (p != &session->s_caps) {
1110                 cap = list_entry(p, struct ceph_cap, session_caps);
1111                 inode = igrab(&cap->ci->vfs_inode);
1112                 if (!inode) {
1113                         p = p->next;
1114                         continue;
1115                 }
1116                 session->s_cap_iterator = cap;
1117                 spin_unlock(&session->s_cap_lock);
1118
1119                 if (last_inode) {
1120                         iput(last_inode);
1121                         last_inode = NULL;
1122                 }
1123                 if (old_cap) {
1124                         ceph_put_cap(session->s_mdsc, old_cap);
1125                         old_cap = NULL;
1126                 }
1127
1128                 ret = cb(inode, cap, arg);
1129                 last_inode = inode;
1130
1131                 spin_lock(&session->s_cap_lock);
1132                 p = p->next;
1133                 if (!cap->ci) {
1134                         dout("iterate_session_caps  finishing cap %p removal\n",
1135                              cap);
1136                         BUG_ON(cap->session != session);
1137                         cap->session = NULL;
1138                         list_del_init(&cap->session_caps);
1139                         session->s_nr_caps--;
1140                         if (cap->queue_release) {
1141                                 list_add_tail(&cap->session_caps,
1142                                               &session->s_cap_releases);
1143                                 session->s_num_cap_releases++;
1144                         } else {
1145                                 old_cap = cap;  /* put_cap it w/o locks held */
1146                         }
1147                 }
1148                 if (ret < 0)
1149                         goto out;
1150         }
1151         ret = 0;
1152 out:
1153         session->s_cap_iterator = NULL;
1154         spin_unlock(&session->s_cap_lock);
1155
1156         iput(last_inode);
1157         if (old_cap)
1158                 ceph_put_cap(session->s_mdsc, old_cap);
1159
1160         return ret;
1161 }
1162
1163 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1164                                   void *arg)
1165 {
1166         struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1167         struct ceph_inode_info *ci = ceph_inode(inode);
1168         LIST_HEAD(to_remove);
1169         bool drop = false;
1170         bool invalidate = false;
1171
1172         dout("removing cap %p, ci is %p, inode is %p\n",
1173              cap, ci, &ci->vfs_inode);
1174         spin_lock(&ci->i_ceph_lock);
1175         __ceph_remove_cap(cap, false);
1176         if (!ci->i_auth_cap) {
1177                 struct ceph_cap_flush *cf;
1178                 struct ceph_mds_client *mdsc = fsc->mdsc;
1179
1180                 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1181
1182                 if (ci->i_wrbuffer_ref > 0 &&
1183                     READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
1184                         invalidate = true;
1185
1186                 while (!list_empty(&ci->i_cap_flush_list)) {
1187                         cf = list_first_entry(&ci->i_cap_flush_list,
1188                                               struct ceph_cap_flush, i_list);
1189                         list_move(&cf->i_list, &to_remove);
1190                 }
1191
1192                 spin_lock(&mdsc->cap_dirty_lock);
1193
1194                 list_for_each_entry(cf, &to_remove, i_list)
1195                         list_del(&cf->g_list);
1196
1197                 if (!list_empty(&ci->i_dirty_item)) {
1198                         pr_warn_ratelimited(
1199                                 " dropping dirty %s state for %p %lld\n",
1200                                 ceph_cap_string(ci->i_dirty_caps),
1201                                 inode, ceph_ino(inode));
1202                         ci->i_dirty_caps = 0;
1203                         list_del_init(&ci->i_dirty_item);
1204                         drop = true;
1205                 }
1206                 if (!list_empty(&ci->i_flushing_item)) {
1207                         pr_warn_ratelimited(
1208                                 " dropping dirty+flushing %s state for %p %lld\n",
1209                                 ceph_cap_string(ci->i_flushing_caps),
1210                                 inode, ceph_ino(inode));
1211                         ci->i_flushing_caps = 0;
1212                         list_del_init(&ci->i_flushing_item);
1213                         mdsc->num_cap_flushing--;
1214                         drop = true;
1215                 }
1216                 spin_unlock(&mdsc->cap_dirty_lock);
1217
1218                 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1219                         list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1220                         ci->i_prealloc_cap_flush = NULL;
1221                 }
1222
1223                if (drop &&
1224                   ci->i_wrbuffer_ref_head == 0 &&
1225                   ci->i_wr_ref == 0 &&
1226                   ci->i_dirty_caps == 0 &&
1227                   ci->i_flushing_caps == 0) {
1228                       ceph_put_snap_context(ci->i_head_snapc);
1229                       ci->i_head_snapc = NULL;
1230                }
1231         }
1232         spin_unlock(&ci->i_ceph_lock);
1233         while (!list_empty(&to_remove)) {
1234                 struct ceph_cap_flush *cf;
1235                 cf = list_first_entry(&to_remove,
1236                                       struct ceph_cap_flush, i_list);
1237                 list_del(&cf->i_list);
1238                 ceph_free_cap_flush(cf);
1239         }
1240
1241         wake_up_all(&ci->i_cap_wq);
1242         if (invalidate)
1243                 ceph_queue_invalidate(inode);
1244         if (drop)
1245                 iput(inode);
1246         return 0;
1247 }
1248
1249 /*
1250  * caller must hold session s_mutex
1251  */
1252 static void remove_session_caps(struct ceph_mds_session *session)
1253 {
1254         struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1255         struct super_block *sb = fsc->sb;
1256         dout("remove_session_caps on %p\n", session);
1257         iterate_session_caps(session, remove_session_caps_cb, fsc);
1258
1259         wake_up_all(&fsc->mdsc->cap_flushing_wq);
1260
1261         spin_lock(&session->s_cap_lock);
1262         if (session->s_nr_caps > 0) {
1263                 struct inode *inode;
1264                 struct ceph_cap *cap, *prev = NULL;
1265                 struct ceph_vino vino;
1266                 /*
1267                  * iterate_session_caps() skips inodes that are being
1268                  * deleted, we need to wait until deletions are complete.
1269                  * __wait_on_freeing_inode() is designed for the job,
1270                  * but it is not exported, so use lookup inode function
1271                  * to access it.
1272                  */
1273                 while (!list_empty(&session->s_caps)) {
1274                         cap = list_entry(session->s_caps.next,
1275                                          struct ceph_cap, session_caps);
1276                         if (cap == prev)
1277                                 break;
1278                         prev = cap;
1279                         vino = cap->ci->i_vino;
1280                         spin_unlock(&session->s_cap_lock);
1281
1282                         inode = ceph_find_inode(sb, vino);
1283                         iput(inode);
1284
1285                         spin_lock(&session->s_cap_lock);
1286                 }
1287         }
1288
1289         // drop cap expires and unlock s_cap_lock
1290         cleanup_cap_releases(session->s_mdsc, session);
1291
1292         BUG_ON(session->s_nr_caps > 0);
1293         BUG_ON(!list_empty(&session->s_cap_flushing));
1294 }
1295
1296 /*
1297  * wake up any threads waiting on this session's caps.  if the cap is
1298  * old (didn't get renewed on the client reconnect), remove it now.
1299  *
1300  * caller must hold s_mutex.
1301  */
1302 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1303                               void *arg)
1304 {
1305         struct ceph_inode_info *ci = ceph_inode(inode);
1306
1307         if (arg) {
1308                 spin_lock(&ci->i_ceph_lock);
1309                 ci->i_wanted_max_size = 0;
1310                 ci->i_requested_max_size = 0;
1311                 spin_unlock(&ci->i_ceph_lock);
1312         }
1313         wake_up_all(&ci->i_cap_wq);
1314         return 0;
1315 }
1316
1317 static void wake_up_session_caps(struct ceph_mds_session *session,
1318                                  int reconnect)
1319 {
1320         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1321         iterate_session_caps(session, wake_up_session_cb,
1322                              (void *)(unsigned long)reconnect);
1323 }
1324
1325 /*
1326  * Send periodic message to MDS renewing all currently held caps.  The
1327  * ack will reset the expiration for all caps from this session.
1328  *
1329  * caller holds s_mutex
1330  */
1331 static int send_renew_caps(struct ceph_mds_client *mdsc,
1332                            struct ceph_mds_session *session)
1333 {
1334         struct ceph_msg *msg;
1335         int state;
1336
1337         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1338             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1339                 pr_info("mds%d caps stale\n", session->s_mds);
1340         session->s_renew_requested = jiffies;
1341
1342         /* do not try to renew caps until a recovering mds has reconnected
1343          * with its clients. */
1344         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1345         if (state < CEPH_MDS_STATE_RECONNECT) {
1346                 dout("send_renew_caps ignoring mds%d (%s)\n",
1347                      session->s_mds, ceph_mds_state_name(state));
1348                 return 0;
1349         }
1350
1351         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1352                 ceph_mds_state_name(state));
1353         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1354                                  ++session->s_renew_seq);
1355         if (!msg)
1356                 return -ENOMEM;
1357         ceph_con_send(&session->s_con, msg);
1358         return 0;
1359 }
1360
1361 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1362                              struct ceph_mds_session *session, u64 seq)
1363 {
1364         struct ceph_msg *msg;
1365
1366         dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1367              session->s_mds, ceph_session_state_name(session->s_state), seq);
1368         msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1369         if (!msg)
1370                 return -ENOMEM;
1371         ceph_con_send(&session->s_con, msg);
1372         return 0;
1373 }
1374
1375
1376 /*
1377  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1378  *
1379  * Called under session->s_mutex
1380  */
1381 static void renewed_caps(struct ceph_mds_client *mdsc,
1382                          struct ceph_mds_session *session, int is_renew)
1383 {
1384         int was_stale;
1385         int wake = 0;
1386
1387         spin_lock(&session->s_cap_lock);
1388         was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1389
1390         session->s_cap_ttl = session->s_renew_requested +
1391                 mdsc->mdsmap->m_session_timeout*HZ;
1392
1393         if (was_stale) {
1394                 if (time_before(jiffies, session->s_cap_ttl)) {
1395                         pr_info("mds%d caps renewed\n", session->s_mds);
1396                         wake = 1;
1397                 } else {
1398                         pr_info("mds%d caps still stale\n", session->s_mds);
1399                 }
1400         }
1401         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1402              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1403              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1404         spin_unlock(&session->s_cap_lock);
1405
1406         if (wake)
1407                 wake_up_session_caps(session, 0);
1408 }
1409
1410 /*
1411  * send a session close request
1412  */
1413 static int request_close_session(struct ceph_mds_client *mdsc,
1414                                  struct ceph_mds_session *session)
1415 {
1416         struct ceph_msg *msg;
1417
1418         dout("request_close_session mds%d state %s seq %lld\n",
1419              session->s_mds, ceph_session_state_name(session->s_state),
1420              session->s_seq);
1421         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1422         if (!msg)
1423                 return -ENOMEM;
1424         ceph_con_send(&session->s_con, msg);
1425         return 1;
1426 }
1427
1428 /*
1429  * Called with s_mutex held.
1430  */
1431 static int __close_session(struct ceph_mds_client *mdsc,
1432                          struct ceph_mds_session *session)
1433 {
1434         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1435                 return 0;
1436         session->s_state = CEPH_MDS_SESSION_CLOSING;
1437         return request_close_session(mdsc, session);
1438 }
1439
1440 static bool drop_negative_children(struct dentry *dentry)
1441 {
1442         struct dentry *child;
1443         bool all_negative = true;
1444
1445         if (!d_is_dir(dentry))
1446                 goto out;
1447
1448         spin_lock(&dentry->d_lock);
1449         list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1450                 if (d_really_is_positive(child)) {
1451                         all_negative = false;
1452                         break;
1453                 }
1454         }
1455         spin_unlock(&dentry->d_lock);
1456
1457         if (all_negative)
1458                 shrink_dcache_parent(dentry);
1459 out:
1460         return all_negative;
1461 }
1462
1463 /*
1464  * Trim old(er) caps.
1465  *
1466  * Because we can't cache an inode without one or more caps, we do
1467  * this indirectly: if a cap is unused, we prune its aliases, at which
1468  * point the inode will hopefully get dropped to.
1469  *
1470  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1471  * memory pressure from the MDS, though, so it needn't be perfect.
1472  */
1473 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1474 {
1475         struct ceph_mds_session *session = arg;
1476         struct ceph_inode_info *ci = ceph_inode(inode);
1477         int used, wanted, oissued, mine;
1478
1479         if (session->s_trim_caps <= 0)
1480                 return -1;
1481
1482         spin_lock(&ci->i_ceph_lock);
1483         mine = cap->issued | cap->implemented;
1484         used = __ceph_caps_used(ci);
1485         wanted = __ceph_caps_file_wanted(ci);
1486         oissued = __ceph_caps_issued_other(ci, cap);
1487
1488         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1489              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1490              ceph_cap_string(used), ceph_cap_string(wanted));
1491         if (cap == ci->i_auth_cap) {
1492                 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1493                     !list_empty(&ci->i_cap_snaps))
1494                         goto out;
1495                 if ((used | wanted) & CEPH_CAP_ANY_WR)
1496                         goto out;
1497         }
1498         /* The inode has cached pages, but it's no longer used.
1499          * we can safely drop it */
1500         if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1501             !(oissued & CEPH_CAP_FILE_CACHE)) {
1502           used = 0;
1503           oissued = 0;
1504         }
1505         if ((used | wanted) & ~oissued & mine)
1506                 goto out;   /* we need these caps */
1507
1508         if (oissued) {
1509                 /* we aren't the only cap.. just remove us */
1510                 __ceph_remove_cap(cap, true);
1511                 session->s_trim_caps--;
1512         } else {
1513                 struct dentry *dentry;
1514                 /* try dropping referring dentries */
1515                 spin_unlock(&ci->i_ceph_lock);
1516                 dentry = d_find_any_alias(inode);
1517                 if (dentry && drop_negative_children(dentry)) {
1518                         int count;
1519                         dput(dentry);
1520                         d_prune_aliases(inode);
1521                         count = atomic_read(&inode->i_count);
1522                         if (count == 1)
1523                                 session->s_trim_caps--;
1524                         dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1525                              inode, cap, count);
1526                 } else {
1527                         dput(dentry);
1528                 }
1529                 return 0;
1530         }
1531
1532 out:
1533         spin_unlock(&ci->i_ceph_lock);
1534         return 0;
1535 }
1536
1537 /*
1538  * Trim session cap count down to some max number.
1539  */
1540 static int trim_caps(struct ceph_mds_client *mdsc,
1541                      struct ceph_mds_session *session,
1542                      int max_caps)
1543 {
1544         int trim_caps = session->s_nr_caps - max_caps;
1545
1546         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1547              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1548         if (trim_caps > 0) {
1549                 session->s_trim_caps = trim_caps;
1550                 iterate_session_caps(session, trim_caps_cb, session);
1551                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1552                      session->s_mds, session->s_nr_caps, max_caps,
1553                         trim_caps - session->s_trim_caps);
1554                 session->s_trim_caps = 0;
1555         }
1556
1557         ceph_send_cap_releases(mdsc, session);
1558         return 0;
1559 }
1560
1561 static int check_caps_flush(struct ceph_mds_client *mdsc,
1562                             u64 want_flush_tid)
1563 {
1564         int ret = 1;
1565
1566         spin_lock(&mdsc->cap_dirty_lock);
1567         if (!list_empty(&mdsc->cap_flush_list)) {
1568                 struct ceph_cap_flush *cf =
1569                         list_first_entry(&mdsc->cap_flush_list,
1570                                          struct ceph_cap_flush, g_list);
1571                 if (cf->tid <= want_flush_tid) {
1572                         dout("check_caps_flush still flushing tid "
1573                              "%llu <= %llu\n", cf->tid, want_flush_tid);
1574                         ret = 0;
1575                 }
1576         }
1577         spin_unlock(&mdsc->cap_dirty_lock);
1578         return ret;
1579 }
1580
1581 /*
1582  * flush all dirty inode data to disk.
1583  *
1584  * returns true if we've flushed through want_flush_tid
1585  */
1586 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1587                             u64 want_flush_tid)
1588 {
1589         dout("check_caps_flush want %llu\n", want_flush_tid);
1590
1591         wait_event(mdsc->cap_flushing_wq,
1592                    check_caps_flush(mdsc, want_flush_tid));
1593
1594         dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1595 }
1596
1597 /*
1598  * called under s_mutex
1599  */
1600 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1601                             struct ceph_mds_session *session)
1602 {
1603         struct ceph_msg *msg = NULL;
1604         struct ceph_mds_cap_release *head;
1605         struct ceph_mds_cap_item *item;
1606         struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
1607         struct ceph_cap *cap;
1608         LIST_HEAD(tmp_list);
1609         int num_cap_releases;
1610         __le32  barrier, *cap_barrier;
1611
1612         down_read(&osdc->lock);
1613         barrier = cpu_to_le32(osdc->epoch_barrier);
1614         up_read(&osdc->lock);
1615
1616         spin_lock(&session->s_cap_lock);
1617 again:
1618         list_splice_init(&session->s_cap_releases, &tmp_list);
1619         num_cap_releases = session->s_num_cap_releases;
1620         session->s_num_cap_releases = 0;
1621         spin_unlock(&session->s_cap_lock);
1622
1623         while (!list_empty(&tmp_list)) {
1624                 if (!msg) {
1625                         msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1626                                         PAGE_SIZE, GFP_NOFS, false);
1627                         if (!msg)
1628                                 goto out_err;
1629                         head = msg->front.iov_base;
1630                         head->num = cpu_to_le32(0);
1631                         msg->front.iov_len = sizeof(*head);
1632
1633                         msg->hdr.version = cpu_to_le16(2);
1634                         msg->hdr.compat_version = cpu_to_le16(1);
1635                 }
1636
1637                 cap = list_first_entry(&tmp_list, struct ceph_cap,
1638                                         session_caps);
1639                 list_del(&cap->session_caps);
1640                 num_cap_releases--;
1641
1642                 head = msg->front.iov_base;
1643                 le32_add_cpu(&head->num, 1);
1644                 item = msg->front.iov_base + msg->front.iov_len;
1645                 item->ino = cpu_to_le64(cap->cap_ino);
1646                 item->cap_id = cpu_to_le64(cap->cap_id);
1647                 item->migrate_seq = cpu_to_le32(cap->mseq);
1648                 item->seq = cpu_to_le32(cap->issue_seq);
1649                 msg->front.iov_len += sizeof(*item);
1650
1651                 ceph_put_cap(mdsc, cap);
1652
1653                 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1654                         // Append cap_barrier field
1655                         cap_barrier = msg->front.iov_base + msg->front.iov_len;
1656                         *cap_barrier = barrier;
1657                         msg->front.iov_len += sizeof(*cap_barrier);
1658
1659                         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1660                         dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1661                         ceph_con_send(&session->s_con, msg);
1662                         msg = NULL;
1663                 }
1664         }
1665
1666         BUG_ON(num_cap_releases != 0);
1667
1668         spin_lock(&session->s_cap_lock);
1669         if (!list_empty(&session->s_cap_releases))
1670                 goto again;
1671         spin_unlock(&session->s_cap_lock);
1672
1673         if (msg) {
1674                 // Append cap_barrier field
1675                 cap_barrier = msg->front.iov_base + msg->front.iov_len;
1676                 *cap_barrier = barrier;
1677                 msg->front.iov_len += sizeof(*cap_barrier);
1678
1679                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1680                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1681                 ceph_con_send(&session->s_con, msg);
1682         }
1683         return;
1684 out_err:
1685         pr_err("send_cap_releases mds%d, failed to allocate message\n",
1686                 session->s_mds);
1687         spin_lock(&session->s_cap_lock);
1688         list_splice(&tmp_list, &session->s_cap_releases);
1689         session->s_num_cap_releases += num_cap_releases;
1690         spin_unlock(&session->s_cap_lock);
1691 }
1692
1693 /*
1694  * requests
1695  */
1696
1697 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1698                                     struct inode *dir)
1699 {
1700         struct ceph_inode_info *ci = ceph_inode(dir);
1701         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1702         struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1703         size_t size = sizeof(struct ceph_mds_reply_dir_entry);
1704         int order, num_entries;
1705
1706         spin_lock(&ci->i_ceph_lock);
1707         num_entries = ci->i_files + ci->i_subdirs;
1708         spin_unlock(&ci->i_ceph_lock);
1709         num_entries = max(num_entries, 1);
1710         num_entries = min(num_entries, opt->max_readdir);
1711
1712         order = get_order(size * num_entries);
1713         while (order >= 0) {
1714                 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
1715                                                              __GFP_NOWARN,
1716                                                              order);
1717                 if (rinfo->dir_entries)
1718                         break;
1719                 order--;
1720         }
1721         if (!rinfo->dir_entries)
1722                 return -ENOMEM;
1723
1724         num_entries = (PAGE_SIZE << order) / size;
1725         num_entries = min(num_entries, opt->max_readdir);
1726
1727         rinfo->dir_buf_size = PAGE_SIZE << order;
1728         req->r_num_caps = num_entries + 1;
1729         req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1730         req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1731         return 0;
1732 }
1733
1734 /*
1735  * Create an mds request.
1736  */
1737 struct ceph_mds_request *
1738 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1739 {
1740         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1741
1742         if (!req)
1743                 return ERR_PTR(-ENOMEM);
1744
1745         mutex_init(&req->r_fill_mutex);
1746         req->r_mdsc = mdsc;
1747         req->r_started = jiffies;
1748         req->r_resend_mds = -1;
1749         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1750         INIT_LIST_HEAD(&req->r_unsafe_target_item);
1751         req->r_fmode = -1;
1752         kref_init(&req->r_kref);
1753         RB_CLEAR_NODE(&req->r_node);
1754         INIT_LIST_HEAD(&req->r_wait);
1755         init_completion(&req->r_completion);
1756         init_completion(&req->r_safe_completion);
1757         INIT_LIST_HEAD(&req->r_unsafe_item);
1758
1759         req->r_stamp = timespec_trunc(current_kernel_time(), mdsc->fsc->sb->s_time_gran);
1760
1761         req->r_op = op;
1762         req->r_direct_mode = mode;
1763         return req;
1764 }
1765
1766 /*
1767  * return oldest (lowest) request, tid in request tree, 0 if none.
1768  *
1769  * called under mdsc->mutex.
1770  */
1771 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1772 {
1773         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1774                 return NULL;
1775         return rb_entry(rb_first(&mdsc->request_tree),
1776                         struct ceph_mds_request, r_node);
1777 }
1778
1779 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1780 {
1781         return mdsc->oldest_tid;
1782 }
1783
1784 /*
1785  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1786  * on build_path_from_dentry in fs/cifs/dir.c.
1787  *
1788  * If @stop_on_nosnap, generate path relative to the first non-snapped
1789  * inode.
1790  *
1791  * Encode hidden .snap dirs as a double /, i.e.
1792  *   foo/.snap/bar -> foo//bar
1793  */
1794 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1795                            int stop_on_nosnap)
1796 {
1797         struct dentry *temp;
1798         char *path;
1799         int len, pos;
1800         unsigned seq;
1801
1802         if (!dentry)
1803                 return ERR_PTR(-EINVAL);
1804
1805 retry:
1806         len = 0;
1807         seq = read_seqbegin(&rename_lock);
1808         rcu_read_lock();
1809         for (temp = dentry; !IS_ROOT(temp);) {
1810                 struct inode *inode = d_inode(temp);
1811                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1812                         len++;  /* slash only */
1813                 else if (stop_on_nosnap && inode &&
1814                          ceph_snap(inode) == CEPH_NOSNAP)
1815                         break;
1816                 else
1817                         len += 1 + temp->d_name.len;
1818                 temp = temp->d_parent;
1819         }
1820         rcu_read_unlock();
1821         if (len)
1822                 len--;  /* no leading '/' */
1823
1824         path = kmalloc(len+1, GFP_NOFS);
1825         if (!path)
1826                 return ERR_PTR(-ENOMEM);
1827         pos = len;
1828         path[pos] = 0;  /* trailing null */
1829         rcu_read_lock();
1830         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1831                 struct inode *inode;
1832
1833                 spin_lock(&temp->d_lock);
1834                 inode = d_inode(temp);
1835                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1836                         dout("build_path path+%d: %p SNAPDIR\n",
1837                              pos, temp);
1838                 } else if (stop_on_nosnap && inode &&
1839                            ceph_snap(inode) == CEPH_NOSNAP) {
1840                         spin_unlock(&temp->d_lock);
1841                         break;
1842                 } else {
1843                         pos -= temp->d_name.len;
1844                         if (pos < 0) {
1845                                 spin_unlock(&temp->d_lock);
1846                                 break;
1847                         }
1848                         strncpy(path + pos, temp->d_name.name,
1849                                 temp->d_name.len);
1850                 }
1851                 spin_unlock(&temp->d_lock);
1852                 if (pos)
1853                         path[--pos] = '/';
1854                 temp = temp->d_parent;
1855         }
1856         rcu_read_unlock();
1857         if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1858                 pr_err("build_path did not end path lookup where "
1859                        "expected, namelen is %d, pos is %d\n", len, pos);
1860                 /* presumably this is only possible if racing with a
1861                    rename of one of the parent directories (we can not
1862                    lock the dentries above us to prevent this, but
1863                    retrying should be harmless) */
1864                 kfree(path);
1865                 goto retry;
1866         }
1867
1868         *base = ceph_ino(d_inode(temp));
1869         *plen = len;
1870         dout("build_path on %p %d built %llx '%.*s'\n",
1871              dentry, d_count(dentry), *base, len, path);
1872         return path;
1873 }
1874
1875 /* Duplicate the dentry->d_name.name safely */
1876 static int clone_dentry_name(struct dentry *dentry, const char **ppath,
1877                              int *ppathlen)
1878 {
1879         u32 len;
1880         char *name;
1881
1882 retry:
1883         len = READ_ONCE(dentry->d_name.len);
1884         name = kmalloc(len + 1, GFP_NOFS);
1885         if (!name)
1886                 return -ENOMEM;
1887
1888         spin_lock(&dentry->d_lock);
1889         if (dentry->d_name.len != len) {
1890                 spin_unlock(&dentry->d_lock);
1891                 kfree(name);
1892                 goto retry;
1893         }
1894         memcpy(name, dentry->d_name.name, len);
1895         spin_unlock(&dentry->d_lock);
1896
1897         name[len] = '\0';
1898         *ppath = name;
1899         *ppathlen = len;
1900         return 0;
1901 }
1902
1903 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
1904                              const char **ppath, int *ppathlen, u64 *pino,
1905                              bool *pfreepath, bool parent_locked)
1906 {
1907         int ret;
1908         char *path;
1909
1910         rcu_read_lock();
1911         if (!dir)
1912                 dir = d_inode_rcu(dentry->d_parent);
1913         if (dir && ceph_snap(dir) == CEPH_NOSNAP) {
1914                 *pino = ceph_ino(dir);
1915                 rcu_read_unlock();
1916                 if (parent_locked) {
1917                         *ppath = dentry->d_name.name;
1918                         *ppathlen = dentry->d_name.len;
1919                 } else {
1920                         ret = clone_dentry_name(dentry, ppath, ppathlen);
1921                         if (ret)
1922                                 return ret;
1923                         *pfreepath = true;
1924                 }
1925                 return 0;
1926         }
1927         rcu_read_unlock();
1928         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1929         if (IS_ERR(path))
1930                 return PTR_ERR(path);
1931         *ppath = path;
1932         *pfreepath = true;
1933         return 0;
1934 }
1935
1936 static int build_inode_path(struct inode *inode,
1937                             const char **ppath, int *ppathlen, u64 *pino,
1938                             bool *pfreepath)
1939 {
1940         struct dentry *dentry;
1941         char *path;
1942
1943         if (ceph_snap(inode) == CEPH_NOSNAP) {
1944                 *pino = ceph_ino(inode);
1945                 *ppathlen = 0;
1946                 return 0;
1947         }
1948         dentry = d_find_alias(inode);
1949         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1950         dput(dentry);
1951         if (IS_ERR(path))
1952                 return PTR_ERR(path);
1953         *ppath = path;
1954         *pfreepath = true;
1955         return 0;
1956 }
1957
1958 /*
1959  * request arguments may be specified via an inode *, a dentry *, or
1960  * an explicit ino+path.
1961  */
1962 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1963                                   struct inode *rdiri, const char *rpath,
1964                                   u64 rino, const char **ppath, int *pathlen,
1965                                   u64 *ino, bool *freepath, bool parent_locked)
1966 {
1967         int r = 0;
1968
1969         if (rinode) {
1970                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1971                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1972                      ceph_snap(rinode));
1973         } else if (rdentry) {
1974                 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
1975                                         freepath, parent_locked);
1976                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1977                      *ppath);
1978         } else if (rpath || rino) {
1979                 *ino = rino;
1980                 *ppath = rpath;
1981                 *pathlen = rpath ? strlen(rpath) : 0;
1982                 dout(" path %.*s\n", *pathlen, rpath);
1983         }
1984
1985         return r;
1986 }
1987
1988 /*
1989  * called under mdsc->mutex
1990  */
1991 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1992                                                struct ceph_mds_request *req,
1993                                                int mds, bool drop_cap_releases)
1994 {
1995         struct ceph_msg *msg;
1996         struct ceph_mds_request_head *head;
1997         const char *path1 = NULL;
1998         const char *path2 = NULL;
1999         u64 ino1 = 0, ino2 = 0;
2000         int pathlen1 = 0, pathlen2 = 0;
2001         bool freepath1 = false, freepath2 = false;
2002         int len;
2003         u16 releases;
2004         void *p, *end;
2005         int ret;
2006
2007         ret = set_request_path_attr(req->r_inode, req->r_dentry,
2008                               req->r_parent, req->r_path1, req->r_ino1.ino,
2009                               &path1, &pathlen1, &ino1, &freepath1,
2010                               test_bit(CEPH_MDS_R_PARENT_LOCKED,
2011                                         &req->r_req_flags));
2012         if (ret < 0) {
2013                 msg = ERR_PTR(ret);
2014                 goto out;
2015         }
2016
2017         /* If r_old_dentry is set, then assume that its parent is locked */
2018         ret = set_request_path_attr(NULL, req->r_old_dentry,
2019                               req->r_old_dentry_dir,
2020                               req->r_path2, req->r_ino2.ino,
2021                               &path2, &pathlen2, &ino2, &freepath2, true);
2022         if (ret < 0) {
2023                 msg = ERR_PTR(ret);
2024                 goto out_free1;
2025         }
2026
2027         len = sizeof(*head) +
2028                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2029                 sizeof(struct ceph_timespec);
2030
2031         /* calculate (max) length for cap releases */
2032         len += sizeof(struct ceph_mds_request_release) *
2033                 (!!req->r_inode_drop + !!req->r_dentry_drop +
2034                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2035         if (req->r_dentry_drop)
2036                 len += req->r_dentry->d_name.len;
2037         if (req->r_old_dentry_drop)
2038                 len += req->r_old_dentry->d_name.len;
2039
2040         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
2041         if (!msg) {
2042                 msg = ERR_PTR(-ENOMEM);
2043                 goto out_free2;
2044         }
2045
2046         msg->hdr.version = cpu_to_le16(2);
2047         msg->hdr.tid = cpu_to_le64(req->r_tid);
2048
2049         head = msg->front.iov_base;
2050         p = msg->front.iov_base + sizeof(*head);
2051         end = msg->front.iov_base + msg->front.iov_len;
2052
2053         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2054         head->op = cpu_to_le32(req->r_op);
2055         head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
2056         head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
2057         head->args = req->r_args;
2058
2059         ceph_encode_filepath(&p, end, ino1, path1);
2060         ceph_encode_filepath(&p, end, ino2, path2);
2061
2062         /* make note of release offset, in case we need to replay */
2063         req->r_request_release_offset = p - msg->front.iov_base;
2064
2065         /* cap releases */
2066         releases = 0;
2067         if (req->r_inode_drop)
2068                 releases += ceph_encode_inode_release(&p,
2069                       req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2070                       mds, req->r_inode_drop, req->r_inode_unless, 0);
2071         if (req->r_dentry_drop)
2072                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
2073                                 req->r_parent, mds, req->r_dentry_drop,
2074                                 req->r_dentry_unless);
2075         if (req->r_old_dentry_drop)
2076                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2077                                 req->r_old_dentry_dir, mds,
2078                                 req->r_old_dentry_drop,
2079                                 req->r_old_dentry_unless);
2080         if (req->r_old_inode_drop)
2081                 releases += ceph_encode_inode_release(&p,
2082                       d_inode(req->r_old_dentry),
2083                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2084
2085         if (drop_cap_releases) {
2086                 releases = 0;
2087                 p = msg->front.iov_base + req->r_request_release_offset;
2088         }
2089
2090         head->num_releases = cpu_to_le16(releases);
2091
2092         /* time stamp */
2093         {
2094                 struct ceph_timespec ts;
2095                 ceph_encode_timespec(&ts, &req->r_stamp);
2096                 ceph_encode_copy(&p, &ts, sizeof(ts));
2097         }
2098
2099         BUG_ON(p > end);
2100         msg->front.iov_len = p - msg->front.iov_base;
2101         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2102
2103         if (req->r_pagelist) {
2104                 struct ceph_pagelist *pagelist = req->r_pagelist;
2105                 refcount_inc(&pagelist->refcnt);
2106                 ceph_msg_data_add_pagelist(msg, pagelist);
2107                 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2108         } else {
2109                 msg->hdr.data_len = 0;
2110         }
2111
2112         msg->hdr.data_off = cpu_to_le16(0);
2113
2114 out_free2:
2115         if (freepath2)
2116                 kfree((char *)path2);
2117 out_free1:
2118         if (freepath1)
2119                 kfree((char *)path1);
2120 out:
2121         return msg;
2122 }
2123
2124 /*
2125  * called under mdsc->mutex if error, under no mutex if
2126  * success.
2127  */
2128 static void complete_request(struct ceph_mds_client *mdsc,
2129                              struct ceph_mds_request *req)
2130 {
2131         if (req->r_callback)
2132                 req->r_callback(mdsc, req);
2133         else
2134                 complete_all(&req->r_completion);
2135 }
2136
2137 /*
2138  * called under mdsc->mutex
2139  */
2140 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2141                                   struct ceph_mds_request *req,
2142                                   int mds, bool drop_cap_releases)
2143 {
2144         struct ceph_mds_request_head *rhead;
2145         struct ceph_msg *msg;
2146         int flags = 0;
2147
2148         req->r_attempts++;
2149         if (req->r_inode) {
2150                 struct ceph_cap *cap =
2151                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2152
2153                 if (cap)
2154                         req->r_sent_on_mseq = cap->mseq;
2155                 else
2156                         req->r_sent_on_mseq = -1;
2157         }
2158         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2159              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2160
2161         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2162                 void *p;
2163                 /*
2164                  * Replay.  Do not regenerate message (and rebuild
2165                  * paths, etc.); just use the original message.
2166                  * Rebuilding paths will break for renames because
2167                  * d_move mangles the src name.
2168                  */
2169                 msg = req->r_request;
2170                 rhead = msg->front.iov_base;
2171
2172                 flags = le32_to_cpu(rhead->flags);
2173                 flags |= CEPH_MDS_FLAG_REPLAY;
2174                 rhead->flags = cpu_to_le32(flags);
2175
2176                 if (req->r_target_inode)
2177                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2178
2179                 rhead->num_retry = req->r_attempts - 1;
2180
2181                 /* remove cap/dentry releases from message */
2182                 rhead->num_releases = 0;
2183
2184                 /* time stamp */
2185                 p = msg->front.iov_base + req->r_request_release_offset;
2186                 {
2187                         struct ceph_timespec ts;
2188                         ceph_encode_timespec(&ts, &req->r_stamp);
2189                         ceph_encode_copy(&p, &ts, sizeof(ts));
2190                 }
2191
2192                 msg->front.iov_len = p - msg->front.iov_base;
2193                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2194                 return 0;
2195         }
2196
2197         if (req->r_request) {
2198                 ceph_msg_put(req->r_request);
2199                 req->r_request = NULL;
2200         }
2201         msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2202         if (IS_ERR(msg)) {
2203                 req->r_err = PTR_ERR(msg);
2204                 return PTR_ERR(msg);
2205         }
2206         req->r_request = msg;
2207
2208         rhead = msg->front.iov_base;
2209         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2210         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2211                 flags |= CEPH_MDS_FLAG_REPLAY;
2212         if (req->r_parent)
2213                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2214         rhead->flags = cpu_to_le32(flags);
2215         rhead->num_fwd = req->r_num_fwd;
2216         rhead->num_retry = req->r_attempts - 1;
2217         rhead->ino = 0;
2218
2219         dout(" r_parent = %p\n", req->r_parent);
2220         return 0;
2221 }
2222
2223 /*
2224  * send request, or put it on the appropriate wait list.
2225  */
2226 static int __do_request(struct ceph_mds_client *mdsc,
2227                         struct ceph_mds_request *req)
2228 {
2229         struct ceph_mds_session *session = NULL;
2230         int mds = -1;
2231         int err = 0;
2232
2233         if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2234                 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2235                         __unregister_request(mdsc, req);
2236                 goto out;
2237         }
2238
2239         if (req->r_timeout &&
2240             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2241                 dout("do_request timed out\n");
2242                 err = -EIO;
2243                 goto finish;
2244         }
2245         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2246                 dout("do_request forced umount\n");
2247                 err = -EIO;
2248                 goto finish;
2249         }
2250         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2251                 if (mdsc->mdsmap_err) {
2252                         err = mdsc->mdsmap_err;
2253                         dout("do_request mdsmap err %d\n", err);
2254                         goto finish;
2255                 }
2256                 if (mdsc->mdsmap->m_epoch == 0) {
2257                         dout("do_request no mdsmap, waiting for map\n");
2258                         list_add(&req->r_wait, &mdsc->waiting_for_map);
2259                         goto finish;
2260                 }
2261                 if (!(mdsc->fsc->mount_options->flags &
2262                       CEPH_MOUNT_OPT_MOUNTWAIT) &&
2263                     !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2264                         err = -EHOSTUNREACH;
2265                         goto finish;
2266                 }
2267         }
2268
2269         put_request_session(req);
2270
2271         mds = __choose_mds(mdsc, req);
2272         if (mds < 0 ||
2273             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2274                 dout("do_request no mds or not active, waiting for map\n");
2275                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2276                 goto out;
2277         }
2278
2279         /* get, open session */
2280         session = __ceph_lookup_mds_session(mdsc, mds);
2281         if (!session) {
2282                 session = register_session(mdsc, mds);
2283                 if (IS_ERR(session)) {
2284                         err = PTR_ERR(session);
2285                         goto finish;
2286                 }
2287         }
2288         req->r_session = get_session(session);
2289
2290         dout("do_request mds%d session %p state %s\n", mds, session,
2291              ceph_session_state_name(session->s_state));
2292         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2293             session->s_state != CEPH_MDS_SESSION_HUNG) {
2294                 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2295                         err = -EACCES;
2296                         goto out_session;
2297                 }
2298                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2299                     session->s_state == CEPH_MDS_SESSION_CLOSING)
2300                         __open_session(mdsc, session);
2301                 list_add(&req->r_wait, &session->s_waiting);
2302                 goto out_session;
2303         }
2304
2305         /* send request */
2306         req->r_resend_mds = -1;   /* forget any previous mds hint */
2307
2308         if (req->r_request_started == 0)   /* note request start time */
2309                 req->r_request_started = jiffies;
2310
2311         err = __prepare_send_request(mdsc, req, mds, false);
2312         if (!err) {
2313                 ceph_msg_get(req->r_request);
2314                 ceph_con_send(&session->s_con, req->r_request);
2315         }
2316
2317 out_session:
2318         ceph_put_mds_session(session);
2319 finish:
2320         if (err) {
2321                 dout("__do_request early error %d\n", err);
2322                 req->r_err = err;
2323                 complete_request(mdsc, req);
2324                 __unregister_request(mdsc, req);
2325         }
2326 out:
2327         return err;
2328 }
2329
2330 /*
2331  * called under mdsc->mutex
2332  */
2333 static void __wake_requests(struct ceph_mds_client *mdsc,
2334                             struct list_head *head)
2335 {
2336         struct ceph_mds_request *req;
2337         LIST_HEAD(tmp_list);
2338
2339         list_splice_init(head, &tmp_list);
2340
2341         while (!list_empty(&tmp_list)) {
2342                 req = list_entry(tmp_list.next,
2343                                  struct ceph_mds_request, r_wait);
2344                 list_del_init(&req->r_wait);
2345                 dout(" wake request %p tid %llu\n", req, req->r_tid);
2346                 __do_request(mdsc, req);
2347         }
2348 }
2349
2350 /*
2351  * Wake up threads with requests pending for @mds, so that they can
2352  * resubmit their requests to a possibly different mds.
2353  */
2354 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2355 {
2356         struct ceph_mds_request *req;
2357         struct rb_node *p = rb_first(&mdsc->request_tree);
2358
2359         dout("kick_requests mds%d\n", mds);
2360         while (p) {
2361                 req = rb_entry(p, struct ceph_mds_request, r_node);
2362                 p = rb_next(p);
2363                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2364                         continue;
2365                 if (req->r_attempts > 0)
2366                         continue; /* only new requests */
2367                 if (req->r_session &&
2368                     req->r_session->s_mds == mds) {
2369                         dout(" kicking tid %llu\n", req->r_tid);
2370                         list_del_init(&req->r_wait);
2371                         __do_request(mdsc, req);
2372                 }
2373         }
2374 }
2375
2376 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2377                               struct ceph_mds_request *req)
2378 {
2379         dout("submit_request on %p\n", req);
2380         mutex_lock(&mdsc->mutex);
2381         __register_request(mdsc, req, NULL);
2382         __do_request(mdsc, req);
2383         mutex_unlock(&mdsc->mutex);
2384 }
2385
2386 /*
2387  * Synchrously perform an mds request.  Take care of all of the
2388  * session setup, forwarding, retry details.
2389  */
2390 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2391                          struct inode *dir,
2392                          struct ceph_mds_request *req)
2393 {
2394         int err;
2395
2396         dout("do_request on %p\n", req);
2397
2398         /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2399         if (req->r_inode)
2400                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2401         if (req->r_parent)
2402                 ceph_get_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
2403         if (req->r_old_dentry_dir)
2404                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2405                                   CEPH_CAP_PIN);
2406
2407         /* issue */
2408         mutex_lock(&mdsc->mutex);
2409         __register_request(mdsc, req, dir);
2410         __do_request(mdsc, req);
2411
2412         if (req->r_err) {
2413                 err = req->r_err;
2414                 goto out;
2415         }
2416
2417         /* wait */
2418         mutex_unlock(&mdsc->mutex);
2419         dout("do_request waiting\n");
2420         if (!req->r_timeout && req->r_wait_for_completion) {
2421                 err = req->r_wait_for_completion(mdsc, req);
2422         } else {
2423                 long timeleft = wait_for_completion_killable_timeout(
2424                                         &req->r_completion,
2425                                         ceph_timeout_jiffies(req->r_timeout));
2426                 if (timeleft > 0)
2427                         err = 0;
2428                 else if (!timeleft)
2429                         err = -EIO;  /* timed out */
2430                 else
2431                         err = timeleft;  /* killed */
2432         }
2433         dout("do_request waited, got %d\n", err);
2434         mutex_lock(&mdsc->mutex);
2435
2436         /* only abort if we didn't race with a real reply */
2437         if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2438                 err = le32_to_cpu(req->r_reply_info.head->result);
2439         } else if (err < 0) {
2440                 dout("aborted request %lld with %d\n", req->r_tid, err);
2441
2442                 /*
2443                  * ensure we aren't running concurrently with
2444                  * ceph_fill_trace or ceph_readdir_prepopulate, which
2445                  * rely on locks (dir mutex) held by our caller.
2446                  */
2447                 mutex_lock(&req->r_fill_mutex);
2448                 req->r_err = err;
2449                 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2450                 mutex_unlock(&req->r_fill_mutex);
2451
2452                 if (req->r_parent &&
2453                     (req->r_op & CEPH_MDS_OP_WRITE))
2454                         ceph_invalidate_dir_request(req);
2455         } else {
2456                 err = req->r_err;
2457         }
2458
2459 out:
2460         mutex_unlock(&mdsc->mutex);
2461         dout("do_request %p done, result %d\n", req, err);
2462         return err;
2463 }
2464
2465 /*
2466  * Invalidate dir's completeness, dentry lease state on an aborted MDS
2467  * namespace request.
2468  */
2469 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2470 {
2471         struct inode *inode = req->r_parent;
2472
2473         dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2474
2475         ceph_dir_clear_complete(inode);
2476         if (req->r_dentry)
2477                 ceph_invalidate_dentry_lease(req->r_dentry);
2478         if (req->r_old_dentry)
2479                 ceph_invalidate_dentry_lease(req->r_old_dentry);
2480 }
2481
2482 /*
2483  * Handle mds reply.
2484  *
2485  * We take the session mutex and parse and process the reply immediately.
2486  * This preserves the logical ordering of replies, capabilities, etc., sent
2487  * by the MDS as they are applied to our local cache.
2488  */
2489 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2490 {
2491         struct ceph_mds_client *mdsc = session->s_mdsc;
2492         struct ceph_mds_request *req;
2493         struct ceph_mds_reply_head *head = msg->front.iov_base;
2494         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2495         struct ceph_snap_realm *realm;
2496         u64 tid;
2497         int err, result;
2498         int mds = session->s_mds;
2499
2500         if (msg->front.iov_len < sizeof(*head)) {
2501                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2502                 ceph_msg_dump(msg);
2503                 return;
2504         }
2505
2506         /* get request, session */
2507         tid = le64_to_cpu(msg->hdr.tid);
2508         mutex_lock(&mdsc->mutex);
2509         req = lookup_get_request(mdsc, tid);
2510         if (!req) {
2511                 dout("handle_reply on unknown tid %llu\n", tid);
2512                 mutex_unlock(&mdsc->mutex);
2513                 return;
2514         }
2515         dout("handle_reply %p\n", req);
2516
2517         /* correct session? */
2518         if (req->r_session != session) {
2519                 pr_err("mdsc_handle_reply got %llu on session mds%d"
2520                        " not mds%d\n", tid, session->s_mds,
2521                        req->r_session ? req->r_session->s_mds : -1);
2522                 mutex_unlock(&mdsc->mutex);
2523                 goto out;
2524         }
2525
2526         /* dup? */
2527         if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
2528             (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
2529                 pr_warn("got a dup %s reply on %llu from mds%d\n",
2530                            head->safe ? "safe" : "unsafe", tid, mds);
2531                 mutex_unlock(&mdsc->mutex);
2532                 goto out;
2533         }
2534         if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
2535                 pr_warn("got unsafe after safe on %llu from mds%d\n",
2536                            tid, mds);
2537                 mutex_unlock(&mdsc->mutex);
2538                 goto out;
2539         }
2540
2541         result = le32_to_cpu(head->result);
2542
2543         /*
2544          * Handle an ESTALE
2545          * if we're not talking to the authority, send to them
2546          * if the authority has changed while we weren't looking,
2547          * send to new authority
2548          * Otherwise we just have to return an ESTALE
2549          */
2550         if (result == -ESTALE) {
2551                 dout("got ESTALE on request %llu", req->r_tid);
2552                 req->r_resend_mds = -1;
2553                 if (req->r_direct_mode != USE_AUTH_MDS) {
2554                         dout("not using auth, setting for that now");
2555                         req->r_direct_mode = USE_AUTH_MDS;
2556                         __do_request(mdsc, req);
2557                         mutex_unlock(&mdsc->mutex);
2558                         goto out;
2559                 } else  {
2560                         int mds = __choose_mds(mdsc, req);
2561                         if (mds >= 0 && mds != req->r_session->s_mds) {
2562                                 dout("but auth changed, so resending");
2563                                 __do_request(mdsc, req);
2564                                 mutex_unlock(&mdsc->mutex);
2565                                 goto out;
2566                         }
2567                 }
2568                 dout("have to return ESTALE on request %llu", req->r_tid);
2569         }
2570
2571
2572         if (head->safe) {
2573                 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
2574                 __unregister_request(mdsc, req);
2575
2576                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2577                         /*
2578                          * We already handled the unsafe response, now do the
2579                          * cleanup.  No need to examine the response; the MDS
2580                          * doesn't include any result info in the safe
2581                          * response.  And even if it did, there is nothing
2582                          * useful we could do with a revised return value.
2583                          */
2584                         dout("got safe reply %llu, mds%d\n", tid, mds);
2585
2586                         /* last unsafe request during umount? */
2587                         if (mdsc->stopping && !__get_oldest_req(mdsc))
2588                                 complete_all(&mdsc->safe_umount_waiters);
2589                         mutex_unlock(&mdsc->mutex);
2590                         goto out;
2591                 }
2592         } else {
2593                 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
2594                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2595                 if (req->r_unsafe_dir) {
2596                         struct ceph_inode_info *ci =
2597                                         ceph_inode(req->r_unsafe_dir);
2598                         spin_lock(&ci->i_unsafe_lock);
2599                         list_add_tail(&req->r_unsafe_dir_item,
2600                                       &ci->i_unsafe_dirops);
2601                         spin_unlock(&ci->i_unsafe_lock);
2602                 }
2603         }
2604
2605         dout("handle_reply tid %lld result %d\n", tid, result);
2606         rinfo = &req->r_reply_info;
2607         err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2608         mutex_unlock(&mdsc->mutex);
2609
2610         mutex_lock(&session->s_mutex);
2611         if (err < 0) {
2612                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2613                 ceph_msg_dump(msg);
2614                 goto out_err;
2615         }
2616
2617         /* snap trace */
2618         realm = NULL;
2619         if (rinfo->snapblob_len) {
2620                 down_write(&mdsc->snap_rwsem);
2621                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2622                                 rinfo->snapblob + rinfo->snapblob_len,
2623                                 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2624                                 &realm);
2625                 downgrade_write(&mdsc->snap_rwsem);
2626         } else {
2627                 down_read(&mdsc->snap_rwsem);
2628         }
2629
2630         /* insert trace into our cache */
2631         mutex_lock(&req->r_fill_mutex);
2632         current->journal_info = req;
2633         err = ceph_fill_trace(mdsc->fsc->sb, req);
2634         if (err == 0) {
2635                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2636                                     req->r_op == CEPH_MDS_OP_LSSNAP))
2637                         ceph_readdir_prepopulate(req, req->r_session);
2638                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2639         }
2640         current->journal_info = NULL;
2641         mutex_unlock(&req->r_fill_mutex);
2642
2643         up_read(&mdsc->snap_rwsem);
2644         if (realm)
2645                 ceph_put_snap_realm(mdsc, realm);
2646
2647         if (err == 0 && req->r_target_inode &&
2648             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2649                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
2650                 spin_lock(&ci->i_unsafe_lock);
2651                 list_add_tail(&req->r_unsafe_target_item, &ci->i_unsafe_iops);
2652                 spin_unlock(&ci->i_unsafe_lock);
2653         }
2654 out_err:
2655         mutex_lock(&mdsc->mutex);
2656         if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2657                 if (err) {
2658                         req->r_err = err;
2659                 } else {
2660                         req->r_reply =  ceph_msg_get(msg);
2661                         set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
2662                 }
2663         } else {
2664                 dout("reply arrived after request %lld was aborted\n", tid);
2665         }
2666         mutex_unlock(&mdsc->mutex);
2667
2668         mutex_unlock(&session->s_mutex);
2669
2670         /* kick calling process */
2671         complete_request(mdsc, req);
2672 out:
2673         ceph_mdsc_put_request(req);
2674         return;
2675 }
2676
2677
2678
2679 /*
2680  * handle mds notification that our request has been forwarded.
2681  */
2682 static void handle_forward(struct ceph_mds_client *mdsc,
2683                            struct ceph_mds_session *session,
2684                            struct ceph_msg *msg)
2685 {
2686         struct ceph_mds_request *req;
2687         u64 tid = le64_to_cpu(msg->hdr.tid);
2688         u32 next_mds;
2689         u32 fwd_seq;
2690         int err = -EINVAL;
2691         void *p = msg->front.iov_base;
2692         void *end = p + msg->front.iov_len;
2693
2694         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2695         next_mds = ceph_decode_32(&p);
2696         fwd_seq = ceph_decode_32(&p);
2697
2698         mutex_lock(&mdsc->mutex);
2699         req = lookup_get_request(mdsc, tid);
2700         if (!req) {
2701                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2702                 goto out;  /* dup reply? */
2703         }
2704
2705         if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2706                 dout("forward tid %llu aborted, unregistering\n", tid);
2707                 __unregister_request(mdsc, req);
2708         } else if (fwd_seq <= req->r_num_fwd) {
2709                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2710                      tid, next_mds, req->r_num_fwd, fwd_seq);
2711         } else {
2712                 /* resend. forward race not possible; mds would drop */
2713                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2714                 BUG_ON(req->r_err);
2715                 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
2716                 req->r_attempts = 0;
2717                 req->r_num_fwd = fwd_seq;
2718                 req->r_resend_mds = next_mds;
2719                 put_request_session(req);
2720                 __do_request(mdsc, req);
2721         }
2722         ceph_mdsc_put_request(req);
2723 out:
2724         mutex_unlock(&mdsc->mutex);
2725         return;
2726
2727 bad:
2728         pr_err("mdsc_handle_forward decode error err=%d\n", err);
2729 }
2730
2731 /*
2732  * handle a mds session control message
2733  */
2734 static void handle_session(struct ceph_mds_session *session,
2735                            struct ceph_msg *msg)
2736 {
2737         struct ceph_mds_client *mdsc = session->s_mdsc;
2738         u32 op;
2739         u64 seq;
2740         int mds = session->s_mds;
2741         struct ceph_mds_session_head *h = msg->front.iov_base;
2742         int wake = 0;
2743
2744         /* decode */
2745         if (msg->front.iov_len != sizeof(*h))
2746                 goto bad;
2747         op = le32_to_cpu(h->op);
2748         seq = le64_to_cpu(h->seq);
2749
2750         mutex_lock(&mdsc->mutex);
2751         if (op == CEPH_SESSION_CLOSE) {
2752                 get_session(session);
2753                 __unregister_session(mdsc, session);
2754         }
2755         /* FIXME: this ttl calculation is generous */
2756         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2757         mutex_unlock(&mdsc->mutex);
2758
2759         mutex_lock(&session->s_mutex);
2760
2761         dout("handle_session mds%d %s %p state %s seq %llu\n",
2762              mds, ceph_session_op_name(op), session,
2763              ceph_session_state_name(session->s_state), seq);
2764
2765         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2766                 session->s_state = CEPH_MDS_SESSION_OPEN;
2767                 pr_info("mds%d came back\n", session->s_mds);
2768         }
2769
2770         switch (op) {
2771         case CEPH_SESSION_OPEN:
2772                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2773                         pr_info("mds%d reconnect success\n", session->s_mds);
2774                 session->s_state = CEPH_MDS_SESSION_OPEN;
2775                 renewed_caps(mdsc, session, 0);
2776                 wake = 1;
2777                 if (mdsc->stopping)
2778                         __close_session(mdsc, session);
2779                 break;
2780
2781         case CEPH_SESSION_RENEWCAPS:
2782                 if (session->s_renew_seq == seq)
2783                         renewed_caps(mdsc, session, 1);
2784                 break;
2785
2786         case CEPH_SESSION_CLOSE:
2787                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2788                         pr_info("mds%d reconnect denied\n", session->s_mds);
2789                 cleanup_session_requests(mdsc, session);
2790                 remove_session_caps(session);
2791                 wake = 2; /* for good measure */
2792                 wake_up_all(&mdsc->session_close_wq);
2793                 break;
2794
2795         case CEPH_SESSION_STALE:
2796                 pr_info("mds%d caps went stale, renewing\n",
2797                         session->s_mds);
2798                 spin_lock(&session->s_gen_ttl_lock);
2799                 session->s_cap_gen++;
2800                 session->s_cap_ttl = jiffies - 1;
2801                 spin_unlock(&session->s_gen_ttl_lock);
2802                 send_renew_caps(mdsc, session);
2803                 break;
2804
2805         case CEPH_SESSION_RECALL_STATE:
2806                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2807                 break;
2808
2809         case CEPH_SESSION_FLUSHMSG:
2810                 send_flushmsg_ack(mdsc, session, seq);
2811                 break;
2812
2813         case CEPH_SESSION_FORCE_RO:
2814                 dout("force_session_readonly %p\n", session);
2815                 spin_lock(&session->s_cap_lock);
2816                 session->s_readonly = true;
2817                 spin_unlock(&session->s_cap_lock);
2818                 wake_up_session_caps(session, 0);
2819                 break;
2820
2821         case CEPH_SESSION_REJECT:
2822                 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
2823                 pr_info("mds%d rejected session\n", session->s_mds);
2824                 session->s_state = CEPH_MDS_SESSION_REJECTED;
2825                 cleanup_session_requests(mdsc, session);
2826                 remove_session_caps(session);
2827                 wake = 2; /* for good measure */
2828                 break;
2829
2830         default:
2831                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2832                 WARN_ON(1);
2833         }
2834
2835         mutex_unlock(&session->s_mutex);
2836         if (wake) {
2837                 mutex_lock(&mdsc->mutex);
2838                 __wake_requests(mdsc, &session->s_waiting);
2839                 if (wake == 2)
2840                         kick_requests(mdsc, mds);
2841                 mutex_unlock(&mdsc->mutex);
2842         }
2843         if (op == CEPH_SESSION_CLOSE)
2844                 ceph_put_mds_session(session);
2845         return;
2846
2847 bad:
2848         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2849                (int)msg->front.iov_len);
2850         ceph_msg_dump(msg);
2851         return;
2852 }
2853
2854
2855 /*
2856  * called under session->mutex.
2857  */
2858 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2859                                    struct ceph_mds_session *session)
2860 {
2861         struct ceph_mds_request *req, *nreq;
2862         struct rb_node *p;
2863         int err;
2864
2865         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2866
2867         mutex_lock(&mdsc->mutex);
2868         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2869                 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2870                 if (!err) {
2871                         ceph_msg_get(req->r_request);
2872                         ceph_con_send(&session->s_con, req->r_request);
2873                 }
2874         }
2875
2876         /*
2877          * also re-send old requests when MDS enters reconnect stage. So that MDS
2878          * can process completed request in clientreplay stage.
2879          */
2880         p = rb_first(&mdsc->request_tree);
2881         while (p) {
2882                 req = rb_entry(p, struct ceph_mds_request, r_node);
2883                 p = rb_next(p);
2884                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2885                         continue;
2886                 if (req->r_attempts == 0)
2887                         continue; /* only old requests */
2888                 if (req->r_session &&
2889                     req->r_session->s_mds == session->s_mds) {
2890                         err = __prepare_send_request(mdsc, req,
2891                                                      session->s_mds, true);
2892                         if (!err) {
2893                                 ceph_msg_get(req->r_request);
2894                                 ceph_con_send(&session->s_con, req->r_request);
2895                         }
2896                 }
2897         }
2898         mutex_unlock(&mdsc->mutex);
2899 }
2900
2901 /*
2902  * Encode information about a cap for a reconnect with the MDS.
2903  */
2904 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2905                           void *arg)
2906 {
2907         union {
2908                 struct ceph_mds_cap_reconnect v2;
2909                 struct ceph_mds_cap_reconnect_v1 v1;
2910         } rec;
2911         struct ceph_inode_info *ci;
2912         struct ceph_reconnect_state *recon_state = arg;
2913         struct ceph_pagelist *pagelist = recon_state->pagelist;
2914         char *path;
2915         int pathlen, err;
2916         u64 pathbase;
2917         u64 snap_follows;
2918         struct dentry *dentry;
2919
2920         ci = cap->ci;
2921
2922         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2923              inode, ceph_vinop(inode), cap, cap->cap_id,
2924              ceph_cap_string(cap->issued));
2925         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2926         if (err)
2927                 return err;
2928
2929         dentry = d_find_alias(inode);
2930         if (dentry) {
2931                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2932                 if (IS_ERR(path)) {
2933                         err = PTR_ERR(path);
2934                         goto out_dput;
2935                 }
2936         } else {
2937                 path = NULL;
2938                 pathlen = 0;
2939                 pathbase = 0;
2940         }
2941
2942         spin_lock(&ci->i_ceph_lock);
2943         cap->seq = 0;        /* reset cap seq */
2944         cap->issue_seq = 0;  /* and issue_seq */
2945         cap->mseq = 0;       /* and migrate_seq */
2946         cap->cap_gen = cap->session->s_cap_gen;
2947
2948         if (recon_state->msg_version >= 2) {
2949                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2950                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2951                 rec.v2.issued = cpu_to_le32(cap->issued);
2952                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2953                 rec.v2.pathbase = cpu_to_le64(pathbase);
2954                 rec.v2.flock_len = 0;
2955         } else {
2956                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2957                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2958                 rec.v1.issued = cpu_to_le32(cap->issued);
2959                 rec.v1.size = cpu_to_le64(inode->i_size);
2960                 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2961                 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2962                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2963                 rec.v1.pathbase = cpu_to_le64(pathbase);
2964         }
2965
2966         if (list_empty(&ci->i_cap_snaps)) {
2967                 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
2968         } else {
2969                 struct ceph_cap_snap *capsnap =
2970                         list_first_entry(&ci->i_cap_snaps,
2971                                          struct ceph_cap_snap, ci_item);
2972                 snap_follows = capsnap->follows;
2973         }
2974         spin_unlock(&ci->i_ceph_lock);
2975
2976         if (recon_state->msg_version >= 2) {
2977                 int num_fcntl_locks, num_flock_locks;
2978                 struct ceph_filelock *flocks;
2979                 size_t struct_len, total_len = 0;
2980                 u8 struct_v = 0;
2981
2982 encode_again:
2983                 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2984                 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2985                                  sizeof(struct ceph_filelock), GFP_NOFS);
2986                 if (!flocks) {
2987                         err = -ENOMEM;
2988                         goto out_free;
2989                 }
2990                 err = ceph_encode_locks_to_buffer(inode, flocks,
2991                                                   num_fcntl_locks,
2992                                                   num_flock_locks);
2993                 if (err) {
2994                         kfree(flocks);
2995                         if (err == -ENOSPC)
2996                                 goto encode_again;
2997                         goto out_free;
2998                 }
2999
3000                 if (recon_state->msg_version >= 3) {
3001                         /* version, compat_version and struct_len */
3002                         total_len = 2 * sizeof(u8) + sizeof(u32);
3003                         struct_v = 2;
3004                 }
3005                 /*
3006                  * number of encoded locks is stable, so copy to pagelist
3007                  */
3008                 struct_len = 2 * sizeof(u32) +
3009                             (num_fcntl_locks + num_flock_locks) *
3010                             sizeof(struct ceph_filelock);
3011                 rec.v2.flock_len = cpu_to_le32(struct_len);
3012
3013                 struct_len += sizeof(rec.v2);
3014                 struct_len += sizeof(u32) + pathlen;
3015
3016                 if (struct_v >= 2)
3017                         struct_len += sizeof(u64); /* snap_follows */
3018
3019                 total_len += struct_len;
3020                 err = ceph_pagelist_reserve(pagelist, total_len);
3021
3022                 if (!err) {
3023                         if (recon_state->msg_version >= 3) {
3024                                 ceph_pagelist_encode_8(pagelist, struct_v);
3025                                 ceph_pagelist_encode_8(pagelist, 1);
3026                                 ceph_pagelist_encode_32(pagelist, struct_len);
3027                         }
3028                         ceph_pagelist_encode_string(pagelist, path, pathlen);
3029                         ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
3030                         ceph_locks_to_pagelist(flocks, pagelist,
3031                                                num_fcntl_locks,
3032                                                num_flock_locks);
3033                         if (struct_v >= 2)
3034                                 ceph_pagelist_encode_64(pagelist, snap_follows);
3035                 }
3036                 kfree(flocks);
3037         } else {
3038                 size_t size = sizeof(u32) + pathlen + sizeof(rec.v1);
3039                 err = ceph_pagelist_reserve(pagelist, size);
3040                 if (!err) {
3041                         ceph_pagelist_encode_string(pagelist, path, pathlen);
3042                         ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
3043                 }
3044         }
3045
3046         recon_state->nr_caps++;
3047 out_free:
3048         kfree(path);
3049 out_dput:
3050         dput(dentry);
3051         return err;
3052 }
3053
3054
3055 /*
3056  * If an MDS fails and recovers, clients need to reconnect in order to
3057  * reestablish shared state.  This includes all caps issued through
3058  * this session _and_ the snap_realm hierarchy.  Because it's not
3059  * clear which snap realms the mds cares about, we send everything we
3060  * know about.. that ensures we'll then get any new info the
3061  * recovering MDS might have.
3062  *
3063  * This is a relatively heavyweight operation, but it's rare.
3064  *
3065  * called with mdsc->mutex held.
3066  */
3067 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3068                                struct ceph_mds_session *session)
3069 {
3070         struct ceph_msg *reply;
3071         struct rb_node *p;
3072         int mds = session->s_mds;
3073         int err = -ENOMEM;
3074         int s_nr_caps;
3075         struct ceph_pagelist *pagelist;
3076         struct ceph_reconnect_state recon_state;
3077
3078         pr_info("mds%d reconnect start\n", mds);
3079
3080         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
3081         if (!pagelist)
3082                 goto fail_nopagelist;
3083         ceph_pagelist_init(pagelist);
3084
3085         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
3086         if (!reply)
3087                 goto fail_nomsg;
3088
3089         mutex_lock(&session->s_mutex);
3090         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3091         session->s_seq = 0;
3092
3093         dout("session %p state %s\n", session,
3094              ceph_session_state_name(session->s_state));
3095
3096         spin_lock(&session->s_gen_ttl_lock);
3097         session->s_cap_gen++;
3098         spin_unlock(&session->s_gen_ttl_lock);
3099
3100         spin_lock(&session->s_cap_lock);
3101         /* don't know if session is readonly */
3102         session->s_readonly = 0;
3103         /*
3104          * notify __ceph_remove_cap() that we are composing cap reconnect.
3105          * If a cap get released before being added to the cap reconnect,
3106          * __ceph_remove_cap() should skip queuing cap release.
3107          */
3108         session->s_cap_reconnect = 1;
3109         /* drop old cap expires; we're about to reestablish that state */
3110         cleanup_cap_releases(mdsc, session);
3111
3112         /* trim unused caps to reduce MDS's cache rejoin time */
3113         if (mdsc->fsc->sb->s_root)
3114                 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3115
3116         ceph_con_close(&session->s_con);
3117         ceph_con_open(&session->s_con,
3118                       CEPH_ENTITY_TYPE_MDS, mds,
3119                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3120
3121         /* replay unsafe requests */
3122         replay_unsafe_requests(mdsc, session);
3123
3124         down_read(&mdsc->snap_rwsem);
3125
3126         /* traverse this session's caps */
3127         s_nr_caps = session->s_nr_caps;
3128         err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
3129         if (err)
3130                 goto fail;
3131
3132         recon_state.nr_caps = 0;
3133         recon_state.pagelist = pagelist;
3134         if (session->s_con.peer_features & CEPH_FEATURE_MDSENC)
3135                 recon_state.msg_version = 3;
3136         else if (session->s_con.peer_features & CEPH_FEATURE_FLOCK)
3137                 recon_state.msg_version = 2;
3138         else
3139                 recon_state.msg_version = 1;
3140         err = iterate_session_caps(session, encode_caps_cb, &recon_state);
3141         if (err < 0)
3142                 goto fail;
3143
3144         spin_lock(&session->s_cap_lock);
3145         session->s_cap_reconnect = 0;
3146         spin_unlock(&session->s_cap_lock);
3147
3148         /*
3149          * snaprealms.  we provide mds with the ino, seq (version), and
3150          * parent for all of our realms.  If the mds has any newer info,
3151          * it will tell us.
3152          */
3153         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3154                 struct ceph_snap_realm *realm =
3155                         rb_entry(p, struct ceph_snap_realm, node);
3156                 struct ceph_mds_snaprealm_reconnect sr_rec;
3157
3158                 dout(" adding snap realm %llx seq %lld parent %llx\n",
3159                      realm->ino, realm->seq, realm->parent_ino);
3160                 sr_rec.ino = cpu_to_le64(realm->ino);
3161                 sr_rec.seq = cpu_to_le64(realm->seq);
3162                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3163                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3164                 if (err)
3165                         goto fail;
3166         }
3167
3168         reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3169
3170         /* raced with cap release? */
3171         if (s_nr_caps != recon_state.nr_caps) {
3172                 struct page *page = list_first_entry(&pagelist->head,
3173                                                      struct page, lru);
3174                 __le32 *addr = kmap_atomic(page);
3175                 *addr = cpu_to_le32(recon_state.nr_caps);
3176                 kunmap_atomic(addr);
3177         }
3178
3179         reply->hdr.data_len = cpu_to_le32(pagelist->length);
3180         ceph_msg_data_add_pagelist(reply, pagelist);
3181
3182         ceph_early_kick_flushing_caps(mdsc, session);
3183
3184         ceph_con_send(&session->s_con, reply);
3185
3186         mutex_unlock(&session->s_mutex);
3187
3188         mutex_lock(&mdsc->mutex);
3189         __wake_requests(mdsc, &session->s_waiting);
3190         mutex_unlock(&mdsc->mutex);
3191
3192         up_read(&mdsc->snap_rwsem);
3193         return;
3194
3195 fail:
3196         ceph_msg_put(reply);
3197         up_read(&mdsc->snap_rwsem);
3198         mutex_unlock(&session->s_mutex);
3199 fail_nomsg:
3200         ceph_pagelist_release(pagelist);
3201 fail_nopagelist:
3202         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3203         return;
3204 }
3205
3206
3207 /*
3208  * compare old and new mdsmaps, kicking requests
3209  * and closing out old connections as necessary
3210  *
3211  * called under mdsc->mutex.
3212  */
3213 static void check_new_map(struct ceph_mds_client *mdsc,
3214                           struct ceph_mdsmap *newmap,
3215                           struct ceph_mdsmap *oldmap)
3216 {
3217         int i;
3218         int oldstate, newstate;
3219         struct ceph_mds_session *s;
3220
3221         dout("check_new_map new %u old %u\n",
3222              newmap->m_epoch, oldmap->m_epoch);
3223
3224         for (i = 0; i < oldmap->m_num_mds && i < mdsc->max_sessions; i++) {
3225                 if (!mdsc->sessions[i])
3226                         continue;
3227                 s = mdsc->sessions[i];
3228                 oldstate = ceph_mdsmap_get_state(oldmap, i);
3229                 newstate = ceph_mdsmap_get_state(newmap, i);
3230
3231                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3232                      i, ceph_mds_state_name(oldstate),
3233                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3234                      ceph_mds_state_name(newstate),
3235                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3236                      ceph_session_state_name(s->s_state));
3237
3238                 if (i >= newmap->m_num_mds ||
3239                     memcmp(ceph_mdsmap_get_addr(oldmap, i),
3240                            ceph_mdsmap_get_addr(newmap, i),
3241                            sizeof(struct ceph_entity_addr))) {
3242                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3243                                 /* the session never opened, just close it
3244                                  * out now */
3245                                 get_session(s);
3246                                 __unregister_session(mdsc, s);
3247                                 __wake_requests(mdsc, &s->s_waiting);
3248                                 ceph_put_mds_session(s);
3249                         } else if (i >= newmap->m_num_mds) {
3250                                 /* force close session for stopped mds */
3251                                 get_session(s);
3252                                 __unregister_session(mdsc, s);
3253                                 __wake_requests(mdsc, &s->s_waiting);
3254                                 kick_requests(mdsc, i);
3255                                 mutex_unlock(&mdsc->mutex);
3256
3257                                 mutex_lock(&s->s_mutex);
3258                                 cleanup_session_requests(mdsc, s);
3259                                 remove_session_caps(s);
3260                                 mutex_unlock(&s->s_mutex);
3261
3262                                 ceph_put_mds_session(s);
3263
3264                                 mutex_lock(&mdsc->mutex);
3265                         } else {
3266                                 /* just close it */
3267                                 mutex_unlock(&mdsc->mutex);
3268                                 mutex_lock(&s->s_mutex);
3269                                 mutex_lock(&mdsc->mutex);
3270                                 ceph_con_close(&s->s_con);
3271                                 mutex_unlock(&s->s_mutex);
3272                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3273                         }
3274                 } else if (oldstate == newstate) {
3275                         continue;  /* nothing new with this mds */
3276                 }
3277
3278                 /*
3279                  * send reconnect?
3280                  */
3281                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3282                     newstate >= CEPH_MDS_STATE_RECONNECT) {
3283                         mutex_unlock(&mdsc->mutex);
3284                         send_mds_reconnect(mdsc, s);
3285                         mutex_lock(&mdsc->mutex);
3286                 }
3287
3288                 /*
3289                  * kick request on any mds that has gone active.
3290                  */
3291                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3292                     newstate >= CEPH_MDS_STATE_ACTIVE) {
3293                         if (oldstate != CEPH_MDS_STATE_CREATING &&
3294                             oldstate != CEPH_MDS_STATE_STARTING)
3295                                 pr_info("mds%d recovery completed\n", s->s_mds);
3296                         kick_requests(mdsc, i);
3297                         ceph_kick_flushing_caps(mdsc, s);
3298                         wake_up_session_caps(s, 1);
3299                 }
3300         }
3301
3302         for (i = 0; i < newmap->m_num_mds && i < mdsc->max_sessions; i++) {
3303                 s = mdsc->sessions[i];
3304                 if (!s)
3305                         continue;
3306                 if (!ceph_mdsmap_is_laggy(newmap, i))
3307                         continue;
3308                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3309                     s->s_state == CEPH_MDS_SESSION_HUNG ||
3310                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
3311                         dout(" connecting to export targets of laggy mds%d\n",
3312                              i);
3313                         __open_export_target_sessions(mdsc, s);
3314                 }
3315         }
3316 }
3317
3318
3319
3320 /*
3321  * leases
3322  */
3323
3324 /*
3325  * caller must hold session s_mutex, dentry->d_lock
3326  */
3327 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3328 {
3329         struct ceph_dentry_info *di = ceph_dentry(dentry);
3330
3331         ceph_put_mds_session(di->lease_session);
3332         di->lease_session = NULL;
3333 }
3334
3335 static void handle_lease(struct ceph_mds_client *mdsc,
3336                          struct ceph_mds_session *session,
3337                          struct ceph_msg *msg)
3338 {
3339         struct super_block *sb = mdsc->fsc->sb;
3340         struct inode *inode;
3341         struct dentry *parent, *dentry;
3342         struct ceph_dentry_info *di;
3343         int mds = session->s_mds;
3344         struct ceph_mds_lease *h = msg->front.iov_base;
3345         u32 seq;
3346         struct ceph_vino vino;
3347         struct qstr dname;
3348         int release = 0;
3349
3350         dout("handle_lease from mds%d\n", mds);
3351
3352         /* decode */
3353         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3354                 goto bad;
3355         vino.ino = le64_to_cpu(h->ino);
3356         vino.snap = CEPH_NOSNAP;
3357         seq = le32_to_cpu(h->seq);
3358         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3359         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3360         if (dname.len != get_unaligned_le32(h+1))
3361                 goto bad;
3362
3363         /* lookup inode */
3364         inode = ceph_find_inode(sb, vino);
3365         dout("handle_lease %s, ino %llx %p %.*s\n",
3366              ceph_lease_op_name(h->action), vino.ino, inode,
3367              dname.len, dname.name);
3368
3369         mutex_lock(&session->s_mutex);
3370         session->s_seq++;
3371
3372         if (!inode) {
3373                 dout("handle_lease no inode %llx\n", vino.ino);
3374                 goto release;
3375         }
3376
3377         /* dentry */
3378         parent = d_find_alias(inode);
3379         if (!parent) {
3380                 dout("no parent dentry on inode %p\n", inode);
3381                 WARN_ON(1);
3382                 goto release;  /* hrm... */
3383         }
3384         dname.hash = full_name_hash(parent, dname.name, dname.len);
3385         dentry = d_lookup(parent, &dname);
3386         dput(parent);
3387         if (!dentry)
3388                 goto release;
3389
3390         spin_lock(&dentry->d_lock);
3391         di = ceph_dentry(dentry);
3392         switch (h->action) {
3393         case CEPH_MDS_LEASE_REVOKE:
3394                 if (di->lease_session == session) {
3395                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3396                                 h->seq = cpu_to_le32(di->lease_seq);
3397                         __ceph_mdsc_drop_dentry_lease(dentry);
3398                 }
3399                 release = 1;
3400                 break;
3401
3402         case CEPH_MDS_LEASE_RENEW:
3403                 if (di->lease_session == session &&
3404                     di->lease_gen == session->s_cap_gen &&
3405                     di->lease_renew_from &&
3406                     di->lease_renew_after == 0) {
3407                         unsigned long duration =
3408                                 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3409
3410                         di->lease_seq = seq;
3411                         di->time = di->lease_renew_from + duration;
3412                         di->lease_renew_after = di->lease_renew_from +
3413                                 (duration >> 1);
3414                         di->lease_renew_from = 0;
3415                 }
3416                 break;
3417         }
3418         spin_unlock(&dentry->d_lock);
3419         dput(dentry);
3420
3421         if (!release)
3422                 goto out;
3423
3424 release:
3425         /* let's just reuse the same message */
3426         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3427         ceph_msg_get(msg);
3428         ceph_con_send(&session->s_con, msg);
3429
3430 out:
3431         iput(inode);
3432         mutex_unlock(&session->s_mutex);
3433         return;
3434
3435 bad:
3436         pr_err("corrupt lease message\n");
3437         ceph_msg_dump(msg);
3438 }
3439
3440 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3441                               struct inode *inode,
3442                               struct dentry *dentry, char action,
3443                               u32 seq)
3444 {
3445         struct ceph_msg *msg;
3446         struct ceph_mds_lease *lease;
3447         int len = sizeof(*lease) + sizeof(u32);
3448         int dnamelen = 0;
3449
3450         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3451              inode, dentry, ceph_lease_op_name(action), session->s_mds);
3452         dnamelen = dentry->d_name.len;
3453         len += dnamelen;
3454
3455         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3456         if (!msg)
3457                 return;
3458         lease = msg->front.iov_base;
3459         lease->action = action;
3460         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3461         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3462         lease->seq = cpu_to_le32(seq);
3463         put_unaligned_le32(dnamelen, lease + 1);
3464         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3465
3466         /*
3467          * if this is a preemptive lease RELEASE, no need to
3468          * flush request stream, since the actual request will
3469          * soon follow.
3470          */
3471         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3472
3473         ceph_con_send(&session->s_con, msg);
3474 }
3475
3476 /*
3477  * drop all leases (and dentry refs) in preparation for umount
3478  */
3479 static void drop_leases(struct ceph_mds_client *mdsc)
3480 {
3481         int i;
3482
3483         dout("drop_leases\n");
3484         mutex_lock(&mdsc->mutex);
3485         for (i = 0; i < mdsc->max_sessions; i++) {
3486                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3487                 if (!s)
3488                         continue;
3489                 mutex_unlock(&mdsc->mutex);
3490                 mutex_lock(&s->s_mutex);
3491                 mutex_unlock(&s->s_mutex);
3492                 ceph_put_mds_session(s);
3493                 mutex_lock(&mdsc->mutex);
3494         }
3495         mutex_unlock(&mdsc->mutex);
3496 }
3497
3498
3499
3500 /*
3501  * delayed work -- periodically trim expired leases, renew caps with mds
3502  */
3503 static void schedule_delayed(struct ceph_mds_client *mdsc)
3504 {
3505         int delay = 5;
3506         unsigned hz = round_jiffies_relative(HZ * delay);
3507         schedule_delayed_work(&mdsc->delayed_work, hz);
3508 }
3509
3510 static void delayed_work(struct work_struct *work)
3511 {
3512         int i;
3513         struct ceph_mds_client *mdsc =
3514                 container_of(work, struct ceph_mds_client, delayed_work.work);
3515         int renew_interval;
3516         int renew_caps;
3517
3518         dout("mdsc delayed_work\n");
3519         ceph_check_delayed_caps(mdsc);
3520
3521         if (mdsc->stopping)
3522                 return;
3523
3524         mutex_lock(&mdsc->mutex);
3525         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3526         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3527                                    mdsc->last_renew_caps);
3528         if (renew_caps)
3529                 mdsc->last_renew_caps = jiffies;
3530
3531         for (i = 0; i < mdsc->max_sessions; i++) {
3532                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3533                 if (!s)
3534                         continue;
3535                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3536                         dout("resending session close request for mds%d\n",
3537                              s->s_mds);
3538                         request_close_session(mdsc, s);
3539                         ceph_put_mds_session(s);
3540                         continue;
3541                 }
3542                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3543                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3544                                 s->s_state = CEPH_MDS_SESSION_HUNG;
3545                                 pr_info("mds%d hung\n", s->s_mds);
3546                         }
3547                 }
3548                 if (s->s_state == CEPH_MDS_SESSION_NEW ||
3549                     s->s_state == CEPH_MDS_SESSION_RESTARTING ||
3550                     s->s_state == CEPH_MDS_SESSION_REJECTED) {
3551                         /* this mds is failed or recovering, just wait */
3552                         ceph_put_mds_session(s);
3553                         continue;
3554                 }
3555                 mutex_unlock(&mdsc->mutex);
3556
3557                 mutex_lock(&s->s_mutex);
3558                 if (renew_caps)
3559                         send_renew_caps(mdsc, s);
3560                 else
3561                         ceph_con_keepalive(&s->s_con);
3562                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3563                     s->s_state == CEPH_MDS_SESSION_HUNG)
3564                         ceph_send_cap_releases(mdsc, s);
3565                 mutex_unlock(&s->s_mutex);
3566                 ceph_put_mds_session(s);
3567
3568                 mutex_lock(&mdsc->mutex);
3569         }
3570         mutex_unlock(&mdsc->mutex);
3571
3572         schedule_delayed(mdsc);
3573 }
3574
3575 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3576
3577 {
3578         struct ceph_mds_client *mdsc;
3579
3580         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3581         if (!mdsc)
3582                 return -ENOMEM;
3583         mdsc->fsc = fsc;
3584         fsc->mdsc = mdsc;
3585         mutex_init(&mdsc->mutex);
3586         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3587         if (!mdsc->mdsmap) {
3588                 kfree(mdsc);
3589                 return -ENOMEM;
3590         }
3591
3592         init_completion(&mdsc->safe_umount_waiters);
3593         init_waitqueue_head(&mdsc->session_close_wq);
3594         INIT_LIST_HEAD(&mdsc->waiting_for_map);
3595         mdsc->sessions = NULL;
3596         atomic_set(&mdsc->num_sessions, 0);
3597         mdsc->max_sessions = 0;
3598         mdsc->stopping = 0;
3599         mdsc->last_snap_seq = 0;
3600         init_rwsem(&mdsc->snap_rwsem);
3601         mdsc->snap_realms = RB_ROOT;
3602         INIT_LIST_HEAD(&mdsc->snap_empty);
3603         spin_lock_init(&mdsc->snap_empty_lock);
3604         mdsc->last_tid = 0;
3605         mdsc->oldest_tid = 0;
3606         mdsc->request_tree = RB_ROOT;
3607         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3608         mdsc->last_renew_caps = jiffies;
3609         INIT_LIST_HEAD(&mdsc->cap_delay_list);
3610         spin_lock_init(&mdsc->cap_delay_lock);
3611         INIT_LIST_HEAD(&mdsc->snap_flush_list);
3612         spin_lock_init(&mdsc->snap_flush_lock);
3613         mdsc->last_cap_flush_tid = 1;
3614         INIT_LIST_HEAD(&mdsc->cap_flush_list);
3615         INIT_LIST_HEAD(&mdsc->cap_dirty);
3616         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3617         mdsc->num_cap_flushing = 0;
3618         spin_lock_init(&mdsc->cap_dirty_lock);
3619         init_waitqueue_head(&mdsc->cap_flushing_wq);
3620         spin_lock_init(&mdsc->dentry_lru_lock);
3621         INIT_LIST_HEAD(&mdsc->dentry_lru);
3622
3623         ceph_caps_init(mdsc);
3624         ceph_adjust_min_caps(mdsc, fsc->min_caps);
3625
3626         init_rwsem(&mdsc->pool_perm_rwsem);
3627         mdsc->pool_perm_tree = RB_ROOT;
3628
3629         strncpy(mdsc->nodename, utsname()->nodename,
3630                 sizeof(mdsc->nodename) - 1);
3631         return 0;
3632 }
3633
3634 /*
3635  * Wait for safe replies on open mds requests.  If we time out, drop
3636  * all requests from the tree to avoid dangling dentry refs.
3637  */
3638 static void wait_requests(struct ceph_mds_client *mdsc)
3639 {
3640         struct ceph_options *opts = mdsc->fsc->client->options;
3641         struct ceph_mds_request *req;
3642
3643         mutex_lock(&mdsc->mutex);
3644         if (__get_oldest_req(mdsc)) {
3645                 mutex_unlock(&mdsc->mutex);
3646
3647                 dout("wait_requests waiting for requests\n");
3648                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3649                                     ceph_timeout_jiffies(opts->mount_timeout));
3650
3651                 /* tear down remaining requests */
3652                 mutex_lock(&mdsc->mutex);
3653                 while ((req = __get_oldest_req(mdsc))) {
3654                         dout("wait_requests timed out on tid %llu\n",
3655                              req->r_tid);
3656                         __unregister_request(mdsc, req);
3657                 }
3658         }
3659         mutex_unlock(&mdsc->mutex);
3660         dout("wait_requests done\n");
3661 }
3662
3663 /*
3664  * called before mount is ro, and before dentries are torn down.
3665  * (hmm, does this still race with new lookups?)
3666  */
3667 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3668 {
3669         dout("pre_umount\n");
3670         mdsc->stopping = 1;
3671
3672         drop_leases(mdsc);
3673         ceph_flush_dirty_caps(mdsc);
3674         wait_requests(mdsc);
3675
3676         /*
3677          * wait for reply handlers to drop their request refs and
3678          * their inode/dcache refs
3679          */
3680         ceph_msgr_flush();
3681 }
3682
3683 /*
3684  * wait for all write mds requests to flush.
3685  */
3686 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3687 {
3688         struct ceph_mds_request *req = NULL, *nextreq;
3689         struct rb_node *n;
3690
3691         mutex_lock(&mdsc->mutex);
3692         dout("wait_unsafe_requests want %lld\n", want_tid);
3693 restart:
3694         req = __get_oldest_req(mdsc);
3695         while (req && req->r_tid <= want_tid) {
3696                 /* find next request */
3697                 n = rb_next(&req->r_node);
3698                 if (n)
3699                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3700                 else
3701                         nextreq = NULL;
3702                 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3703                     (req->r_op & CEPH_MDS_OP_WRITE)) {
3704                         /* write op */
3705                         ceph_mdsc_get_request(req);
3706                         if (nextreq)
3707                                 ceph_mdsc_get_request(nextreq);
3708                         mutex_unlock(&mdsc->mutex);
3709                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
3710                              req->r_tid, want_tid);
3711                         wait_for_completion(&req->r_safe_completion);
3712                         mutex_lock(&mdsc->mutex);
3713                         ceph_mdsc_put_request(req);
3714                         if (!nextreq)
3715                                 break;  /* next dne before, so we're done! */
3716                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
3717                                 /* next request was removed from tree */
3718                                 ceph_mdsc_put_request(nextreq);
3719                                 goto restart;
3720                         }
3721                         ceph_mdsc_put_request(nextreq);  /* won't go away */
3722                 }
3723                 req = nextreq;
3724         }
3725         mutex_unlock(&mdsc->mutex);
3726         dout("wait_unsafe_requests done\n");
3727 }
3728
3729 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3730 {
3731         u64 want_tid, want_flush;
3732
3733         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3734                 return;
3735
3736         dout("sync\n");
3737         mutex_lock(&mdsc->mutex);
3738         want_tid = mdsc->last_tid;
3739         mutex_unlock(&mdsc->mutex);
3740
3741         ceph_flush_dirty_caps(mdsc);
3742         spin_lock(&mdsc->cap_dirty_lock);
3743         want_flush = mdsc->last_cap_flush_tid;
3744         if (!list_empty(&mdsc->cap_flush_list)) {
3745                 struct ceph_cap_flush *cf =
3746                         list_last_entry(&mdsc->cap_flush_list,
3747                                         struct ceph_cap_flush, g_list);
3748                 cf->wake = true;
3749         }
3750         spin_unlock(&mdsc->cap_dirty_lock);
3751
3752         dout("sync want tid %lld flush_seq %lld\n",
3753              want_tid, want_flush);
3754
3755         wait_unsafe_requests(mdsc, want_tid);
3756         wait_caps_flush(mdsc, want_flush);
3757 }
3758
3759 /*
3760  * true if all sessions are closed, or we force unmount
3761  */
3762 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
3763 {
3764         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3765                 return true;
3766         return atomic_read(&mdsc->num_sessions) <= skipped;
3767 }
3768
3769 /*
3770  * called after sb is ro.
3771  */
3772 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3773 {
3774         struct ceph_options *opts = mdsc->fsc->client->options;
3775         struct ceph_mds_session *session;
3776         int i;
3777         int skipped = 0;
3778
3779         dout("close_sessions\n");
3780
3781         /* close sessions */
3782         mutex_lock(&mdsc->mutex);
3783         for (i = 0; i < mdsc->max_sessions; i++) {
3784                 session = __ceph_lookup_mds_session(mdsc, i);
3785                 if (!session)
3786                         continue;
3787                 mutex_unlock(&mdsc->mutex);
3788                 mutex_lock(&session->s_mutex);
3789                 if (__close_session(mdsc, session) <= 0)
3790                         skipped++;
3791                 mutex_unlock(&session->s_mutex);
3792                 ceph_put_mds_session(session);
3793                 mutex_lock(&mdsc->mutex);
3794         }
3795         mutex_unlock(&mdsc->mutex);
3796
3797         dout("waiting for sessions to close\n");
3798         wait_event_timeout(mdsc->session_close_wq,
3799                            done_closing_sessions(mdsc, skipped),
3800                            ceph_timeout_jiffies(opts->mount_timeout));
3801
3802         /* tear down remaining sessions */
3803         mutex_lock(&mdsc->mutex);
3804         for (i = 0; i < mdsc->max_sessions; i++) {
3805                 if (mdsc->sessions[i]) {
3806                         session = get_session(mdsc->sessions[i]);
3807                         __unregister_session(mdsc, session);
3808                         mutex_unlock(&mdsc->mutex);
3809                         mutex_lock(&session->s_mutex);
3810                         remove_session_caps(session);
3811                         mutex_unlock(&session->s_mutex);
3812                         ceph_put_mds_session(session);
3813                         mutex_lock(&mdsc->mutex);
3814                 }
3815         }
3816         WARN_ON(!list_empty(&mdsc->cap_delay_list));
3817         mutex_unlock(&mdsc->mutex);
3818
3819         ceph_cleanup_empty_realms(mdsc);
3820
3821         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3822
3823         dout("stopped\n");
3824 }
3825
3826 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
3827 {
3828         struct ceph_mds_session *session;
3829         int mds;
3830
3831         dout("force umount\n");
3832
3833         mutex_lock(&mdsc->mutex);
3834         for (mds = 0; mds < mdsc->max_sessions; mds++) {
3835                 session = __ceph_lookup_mds_session(mdsc, mds);
3836                 if (!session)
3837                         continue;
3838                 mutex_unlock(&mdsc->mutex);
3839                 mutex_lock(&session->s_mutex);
3840                 __close_session(mdsc, session);
3841                 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
3842                         cleanup_session_requests(mdsc, session);
3843                         remove_session_caps(session);
3844                 }
3845                 mutex_unlock(&session->s_mutex);
3846                 ceph_put_mds_session(session);
3847                 mutex_lock(&mdsc->mutex);
3848                 kick_requests(mdsc, mds);
3849         }
3850         __wake_requests(mdsc, &mdsc->waiting_for_map);
3851         mutex_unlock(&mdsc->mutex);
3852 }
3853
3854 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3855 {
3856         dout("stop\n");
3857         /*
3858          * Make sure the delayed work stopped before releasing
3859          * the resources.
3860          *
3861          * Because the cancel_delayed_work_sync() will only
3862          * guarantee that the work finishes executing. But the
3863          * delayed work will re-arm itself again after that.
3864          */
3865         flush_delayed_work(&mdsc->delayed_work);
3866
3867         if (mdsc->mdsmap)
3868                 ceph_mdsmap_destroy(mdsc->mdsmap);
3869         kfree(mdsc->sessions);
3870         ceph_caps_finalize(mdsc);
3871         ceph_pool_perm_destroy(mdsc);
3872 }
3873
3874 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3875 {
3876         struct ceph_mds_client *mdsc = fsc->mdsc;
3877         dout("mdsc_destroy %p\n", mdsc);
3878
3879         /* flush out any connection work with references to us */
3880         ceph_msgr_flush();
3881
3882         ceph_mdsc_stop(mdsc);
3883
3884         fsc->mdsc = NULL;
3885         kfree(mdsc);
3886         dout("mdsc_destroy %p done\n", mdsc);
3887 }
3888
3889 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3890 {
3891         struct ceph_fs_client *fsc = mdsc->fsc;
3892         const char *mds_namespace = fsc->mount_options->mds_namespace;
3893         void *p = msg->front.iov_base;
3894         void *end = p + msg->front.iov_len;
3895         u32 epoch;
3896         u32 map_len;
3897         u32 num_fs;
3898         u32 mount_fscid = (u32)-1;
3899         u8 struct_v, struct_cv;
3900         int err = -EINVAL;
3901
3902         ceph_decode_need(&p, end, sizeof(u32), bad);
3903         epoch = ceph_decode_32(&p);
3904
3905         dout("handle_fsmap epoch %u\n", epoch);
3906
3907         ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3908         struct_v = ceph_decode_8(&p);
3909         struct_cv = ceph_decode_8(&p);
3910         map_len = ceph_decode_32(&p);
3911
3912         ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
3913         p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
3914
3915         num_fs = ceph_decode_32(&p);
3916         while (num_fs-- > 0) {
3917                 void *info_p, *info_end;
3918                 u32 info_len;
3919                 u8 info_v, info_cv;
3920                 u32 fscid, namelen;
3921
3922                 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3923                 info_v = ceph_decode_8(&p);
3924                 info_cv = ceph_decode_8(&p);
3925                 info_len = ceph_decode_32(&p);
3926                 ceph_decode_need(&p, end, info_len, bad);
3927                 info_p = p;
3928                 info_end = p + info_len;
3929                 p = info_end;
3930
3931                 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
3932                 fscid = ceph_decode_32(&info_p);
3933                 namelen = ceph_decode_32(&info_p);
3934                 ceph_decode_need(&info_p, info_end, namelen, bad);
3935
3936                 if (mds_namespace &&
3937                     strlen(mds_namespace) == namelen &&
3938                     !strncmp(mds_namespace, (char *)info_p, namelen)) {
3939                         mount_fscid = fscid;
3940                         break;
3941                 }
3942         }
3943
3944         ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
3945         if (mount_fscid != (u32)-1) {
3946                 fsc->client->monc.fs_cluster_id = mount_fscid;
3947                 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
3948                                    0, true);
3949                 ceph_monc_renew_subs(&fsc->client->monc);
3950         } else {
3951                 err = -ENOENT;
3952                 goto err_out;
3953         }
3954         return;
3955 bad:
3956         pr_err("error decoding fsmap\n");
3957 err_out:
3958         mutex_lock(&mdsc->mutex);
3959         mdsc->mdsmap_err = -ENOENT;
3960         __wake_requests(mdsc, &mdsc->waiting_for_map);
3961         mutex_unlock(&mdsc->mutex);
3962         return;
3963 }
3964
3965 /*
3966  * handle mds map update.
3967  */
3968 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3969 {
3970         u32 epoch;
3971         u32 maplen;
3972         void *p = msg->front.iov_base;
3973         void *end = p + msg->front.iov_len;
3974         struct ceph_mdsmap *newmap, *oldmap;
3975         struct ceph_fsid fsid;
3976         int err = -EINVAL;
3977
3978         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3979         ceph_decode_copy(&p, &fsid, sizeof(fsid));
3980         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3981                 return;
3982         epoch = ceph_decode_32(&p);
3983         maplen = ceph_decode_32(&p);
3984         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3985
3986         /* do we need it? */
3987         mutex_lock(&mdsc->mutex);
3988         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3989                 dout("handle_map epoch %u <= our %u\n",
3990                      epoch, mdsc->mdsmap->m_epoch);
3991                 mutex_unlock(&mdsc->mutex);
3992                 return;
3993         }
3994
3995         newmap = ceph_mdsmap_decode(&p, end);
3996         if (IS_ERR(newmap)) {
3997                 err = PTR_ERR(newmap);
3998                 goto bad_unlock;
3999         }
4000
4001         /* swap into place */
4002         if (mdsc->mdsmap) {
4003                 oldmap = mdsc->mdsmap;
4004                 mdsc->mdsmap = newmap;
4005                 check_new_map(mdsc, newmap, oldmap);
4006                 ceph_mdsmap_destroy(oldmap);
4007         } else {
4008                 mdsc->mdsmap = newmap;  /* first mds map */
4009         }
4010         mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
4011
4012         __wake_requests(mdsc, &mdsc->waiting_for_map);
4013         ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
4014                           mdsc->mdsmap->m_epoch);
4015
4016         mutex_unlock(&mdsc->mutex);
4017         schedule_delayed(mdsc);
4018         return;
4019
4020 bad_unlock:
4021         mutex_unlock(&mdsc->mutex);
4022 bad:
4023         pr_err("error decoding mdsmap %d\n", err);
4024         return;
4025 }
4026
4027 static struct ceph_connection *con_get(struct ceph_connection *con)
4028 {
4029         struct ceph_mds_session *s = con->private;
4030
4031         if (get_session(s)) {
4032                 dout("mdsc con_get %p ok (%d)\n", s, refcount_read(&s->s_ref));
4033                 return con;
4034         }
4035         dout("mdsc con_get %p FAIL\n", s);
4036         return NULL;
4037 }
4038
4039 static void con_put(struct ceph_connection *con)
4040 {
4041         struct ceph_mds_session *s = con->private;
4042
4043         dout("mdsc con_put %p (%d)\n", s, refcount_read(&s->s_ref) - 1);
4044         ceph_put_mds_session(s);
4045 }
4046
4047 /*
4048  * if the client is unresponsive for long enough, the mds will kill
4049  * the session entirely.
4050  */
4051 static void peer_reset(struct ceph_connection *con)
4052 {
4053         struct ceph_mds_session *s = con->private;
4054         struct ceph_mds_client *mdsc = s->s_mdsc;
4055
4056         pr_warn("mds%d closed our session\n", s->s_mds);
4057         send_mds_reconnect(mdsc, s);
4058 }
4059
4060 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
4061 {
4062         struct ceph_mds_session *s = con->private;
4063         struct ceph_mds_client *mdsc = s->s_mdsc;
4064         int type = le16_to_cpu(msg->hdr.type);
4065
4066         mutex_lock(&mdsc->mutex);
4067         if (__verify_registered_session(mdsc, s) < 0) {
4068                 mutex_unlock(&mdsc->mutex);
4069                 goto out;
4070         }
4071         mutex_unlock(&mdsc->mutex);
4072
4073         switch (type) {
4074         case CEPH_MSG_MDS_MAP:
4075                 ceph_mdsc_handle_mdsmap(mdsc, msg);
4076                 break;
4077         case CEPH_MSG_FS_MAP_USER:
4078                 ceph_mdsc_handle_fsmap(mdsc, msg);
4079                 break;
4080         case CEPH_MSG_CLIENT_SESSION:
4081                 handle_session(s, msg);
4082                 break;
4083         case CEPH_MSG_CLIENT_REPLY:
4084                 handle_reply(s, msg);
4085                 break;
4086         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
4087                 handle_forward(mdsc, s, msg);
4088                 break;
4089         case CEPH_MSG_CLIENT_CAPS:
4090                 ceph_handle_caps(s, msg);
4091                 break;
4092         case CEPH_MSG_CLIENT_SNAP:
4093                 ceph_handle_snap(mdsc, s, msg);
4094                 break;
4095         case CEPH_MSG_CLIENT_LEASE:
4096                 handle_lease(mdsc, s, msg);
4097                 break;
4098
4099         default:
4100                 pr_err("received unknown message type %d %s\n", type,
4101                        ceph_msg_type_name(type));
4102         }
4103 out:
4104         ceph_msg_put(msg);
4105 }
4106
4107 /*
4108  * authentication
4109  */
4110
4111 /*
4112  * Note: returned pointer is the address of a structure that's
4113  * managed separately.  Caller must *not* attempt to free it.
4114  */
4115 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
4116                                         int *proto, int force_new)
4117 {
4118         struct ceph_mds_session *s = con->private;
4119         struct ceph_mds_client *mdsc = s->s_mdsc;
4120         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4121         struct ceph_auth_handshake *auth = &s->s_auth;
4122
4123         if (force_new && auth->authorizer) {
4124                 ceph_auth_destroy_authorizer(auth->authorizer);
4125                 auth->authorizer = NULL;
4126         }
4127         if (!auth->authorizer) {
4128                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4129                                                       auth);
4130                 if (ret)
4131                         return ERR_PTR(ret);
4132         } else {
4133                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4134                                                       auth);
4135                 if (ret)
4136                         return ERR_PTR(ret);
4137         }
4138         *proto = ac->protocol;
4139
4140         return auth;
4141 }
4142
4143 static int add_authorizer_challenge(struct ceph_connection *con,
4144                                     void *challenge_buf, int challenge_buf_len)
4145 {
4146         struct ceph_mds_session *s = con->private;
4147         struct ceph_mds_client *mdsc = s->s_mdsc;
4148         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4149
4150         return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
4151                                             challenge_buf, challenge_buf_len);
4152 }
4153
4154 static int verify_authorizer_reply(struct ceph_connection *con)
4155 {
4156         struct ceph_mds_session *s = con->private;
4157         struct ceph_mds_client *mdsc = s->s_mdsc;
4158         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4159
4160         return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
4161 }
4162
4163 static int invalidate_authorizer(struct ceph_connection *con)
4164 {
4165         struct ceph_mds_session *s = con->private;
4166         struct ceph_mds_client *mdsc = s->s_mdsc;
4167         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4168
4169         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
4170
4171         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
4172 }
4173
4174 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
4175                                 struct ceph_msg_header *hdr, int *skip)
4176 {
4177         struct ceph_msg *msg;
4178         int type = (int) le16_to_cpu(hdr->type);
4179         int front_len = (int) le32_to_cpu(hdr->front_len);
4180
4181         if (con->in_msg)
4182                 return con->in_msg;
4183
4184         *skip = 0;
4185         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
4186         if (!msg) {
4187                 pr_err("unable to allocate msg type %d len %d\n",
4188                        type, front_len);
4189                 return NULL;
4190         }
4191
4192         return msg;
4193 }
4194
4195 static int mds_sign_message(struct ceph_msg *msg)
4196 {
4197        struct ceph_mds_session *s = msg->con->private;
4198        struct ceph_auth_handshake *auth = &s->s_auth;
4199
4200        return ceph_auth_sign_message(auth, msg);
4201 }
4202
4203 static int mds_check_message_signature(struct ceph_msg *msg)
4204 {
4205        struct ceph_mds_session *s = msg->con->private;
4206        struct ceph_auth_handshake *auth = &s->s_auth;
4207
4208        return ceph_auth_check_message_signature(auth, msg);
4209 }
4210
4211 static const struct ceph_connection_operations mds_con_ops = {
4212         .get = con_get,
4213         .put = con_put,
4214         .dispatch = dispatch,
4215         .get_authorizer = get_authorizer,
4216         .add_authorizer_challenge = add_authorizer_challenge,
4217         .verify_authorizer_reply = verify_authorizer_reply,
4218         .invalidate_authorizer = invalidate_authorizer,
4219         .peer_reset = peer_reset,
4220         .alloc_msg = mds_alloc_msg,
4221         .sign_message = mds_sign_message,
4222         .check_message_signature = mds_check_message_signature,
4223 };
4224
4225 /* eof */