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  45  * are no longer valid.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\n", 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         int ret = 0;
 608
 609         req->r_tid = ++mdsc->last_tid;
 610         if (req->r_num_caps) {
 611                 ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
 612                                         req->r_num_caps);
 613                 if (ret < 0) {
 614                         pr_err("__register_request %p "
 615                                "failed to reserve caps: %d\n", req, ret);
 616                         /* set req->r_err to fail early from __do_request */
 617                         req->r_err = ret;
 618                         return;
 619                 }
 620         }
 621         dout("__register_request %p tid %lld\n", req, req->r_tid);
 622         ceph_mdsc_get_request(req);
 623         insert_request(&mdsc->request_tree, req);
 624
 625         req->r_uid = current_fsuid();
 626         req->r_gid = current_fsgid();
 627
 628         if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
 629                 mdsc->oldest_tid = req->r_tid;
 630
 631         if (dir) {
 632                 ihold(dir);
 633                 req->r_unsafe_dir = dir;
 634         }
 635 }
 636
 637 static void __unregister_request(struct ceph_mds_client *mdsc,
 638                                  struct ceph_mds_request *req)
 639 {
 640         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 641
 642         /* Never leave an unregistered request on an unsafe list! */
 643         list_del_init(&req->r_unsafe_item);
 644
 645         if (req->r_tid == mdsc->oldest_tid) {
 646                 struct rb_node *p = rb_next(&req->r_node);
 647                 mdsc->oldest_tid = 0;
 648                 while (p) {
 649                         struct ceph_mds_request *next_req =
 650                                 rb_entry(p, struct ceph_mds_request, r_node);
 651                         if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
 652                                 mdsc->oldest_tid = next_req->r_tid;
 653                                 break;
 654                         }
 655                         p = rb_next(p);
 656                 }
 657         }
 658
 659         erase_request(&mdsc->request_tree, req);
 660
 661         if (req->r_unsafe_dir  &&
 662             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
 663                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 664                 spin_lock(&ci->i_unsafe_lock);
 665                 list_del_init(&req->r_unsafe_dir_item);
 666                 spin_unlock(&ci->i_unsafe_lock);
 667         }
 668         if (req->r_target_inode &&
 669             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
 670                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
 671                 spin_lock(&ci->i_unsafe_lock);
 672                 list_del_init(&req->r_unsafe_target_item);
 673                 spin_unlock(&ci->i_unsafe_lock);
 674         }
 675
 676         if (req->r_unsafe_dir) {
 677                 iput(req->r_unsafe_dir);
 678                 req->r_unsafe_dir = NULL;
 679         }
 680
 681         complete_all(&req->r_safe_completion);
 682
 683         ceph_mdsc_put_request(req);
 684 }
 685
 686 /*
 687  * Walk back up the dentry tree until we hit a dentry representing a
 688  * non-snapshot inode. We do this using the rcu_read_lock (which must be held
 689  * when calling this) to ensure that the objects won't disappear while we're
 690  * working with them. Once we hit a candidate dentry, we attempt to take a
 691  * reference to it, and return that as the result.
 692  */
 693 static struct inode *get_nonsnap_parent(struct dentry *dentry)
 694 {
 695         struct inode *inode = NULL;
 696
 697         while (dentry && !IS_ROOT(dentry)) {
 698                 inode = d_inode_rcu(dentry);
 699                 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
 700                         break;
 701                 dentry = dentry->d_parent;
 702         }
 703         if (inode)
 704                 inode = igrab(inode);
 705         return inode;
 706 }
 707
 708 /*
 709  * Choose mds to send request to next.  If there is a hint set in the
 710  * request (e.g., due to a prior forward hint from the mds), use that.
 711  * Otherwise, consult frag tree and/or caps to identify the
 712  * appropriate mds.  If all else fails, choose randomly.
 713  *
 714  * Called under mdsc->mutex.
 715  */
 716 static int __choose_mds(struct ceph_mds_client *mdsc,
 717                         struct ceph_mds_request *req)
 718 {
 719         struct inode *inode;
 720         struct ceph_inode_info *ci;
 721         struct ceph_cap *cap;
 722         int mode = req->r_direct_mode;
 723         int mds = -1;
 724         u32 hash = req->r_direct_hash;
 725         bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
 726
 727         /*
 728          * is there a specific mds we should try?  ignore hint if we have
 729          * no session and the mds is not up (active or recovering).
 730          */
 731         if (req->r_resend_mds >= 0 &&
 732             (__have_session(mdsc, req->r_resend_mds) ||
 733              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 734                 dout("choose_mds using resend_mds mds%d\n",
 735                      req->r_resend_mds);
 736                 return req->r_resend_mds;
 737         }
 738
 739         if (mode == USE_RANDOM_MDS)
 740                 goto random;
 741
 742         inode = NULL;
 743         if (req->r_inode) {
 744                 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
 745                         inode = req->r_inode;
 746                         ihold(inode);
 747                 } else {
 748                         /* req->r_dentry is non-null for LSSNAP request */
 749                         rcu_read_lock();
 750                         inode = get_nonsnap_parent(req->r_dentry);
 751                         rcu_read_unlock();
 752                         dout("__choose_mds using snapdir's parent %p\n", inode);
 753                 }
 754         } else if (req->r_dentry) {
 755                 /* ignore race with rename; old or new d_parent is okay */
 756                 struct dentry *parent;
 757                 struct inode *dir;
 758
 759                 rcu_read_lock();
 760                 parent = req->r_dentry->d_parent;
 761                 dir = req->r_parent ? : d_inode_rcu(parent);
 762
 763                 if (!dir || dir->i_sb != mdsc->fsc->sb) {
 764                         /*  not this fs or parent went negative */
 765                         inode = d_inode(req->r_dentry);
 766                         if (inode)
 767                                 ihold(inode);
 768                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
 769                         /* direct snapped/virtual snapdir requests
 770                          * based on parent dir inode */
 771                         inode = get_nonsnap_parent(parent);
 772                         dout("__choose_mds using nonsnap parent %p\n", inode);
 773                 } else {
 774                         /* dentry target */
 775                         inode = d_inode(req->r_dentry);
 776                         if (!inode || mode == USE_AUTH_MDS) {
 777                                 /* dir + name */
 778                                 inode = igrab(dir);
 779                                 hash = ceph_dentry_hash(dir, req->r_dentry);
 780                                 is_hash = true;
 781                         } else {
 782                                 ihold(inode);
 783                         }
 784                 }
 785                 rcu_read_unlock();
 786         }
 787
 788         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 789              (int)hash, mode);
 790         if (!inode)
 791                 goto random;
 792         ci = ceph_inode(inode);
 793
 794         if (is_hash && S_ISDIR(inode->i_mode)) {
 795                 struct ceph_inode_frag frag;
 796                 int found;
 797
 798                 ceph_choose_frag(ci, hash, &frag, &found);
 799                 if (found) {
 800                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 801                                 u8 r;
 802
 803                                 /* choose a random replica */
 804                                 get_random_bytes(&r, 1);
 805                                 r %= frag.ndist;
 806                                 mds = frag.dist[r];
 807                                 dout("choose_mds %p %llx.%llx "
 808                                      "frag %u mds%d (%d/%d)\n",
 809                                      inode, ceph_vinop(inode),
 810                                      frag.frag, mds,
 811                                      (int)r, frag.ndist);
 812                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 813                                     CEPH_MDS_STATE_ACTIVE)
 814                                         goto out;
 815                         }
 816
 817                         /* since this file/dir wasn't known to be
 818                          * replicated, then we want to look for the
 819                          * authoritative mds. */
 820                         mode = USE_AUTH_MDS;
 821                         if (frag.mds >= 0) {
 822                                 /* choose auth mds */
 823                                 mds = frag.mds;
 824                                 dout("choose_mds %p %llx.%llx "
 825                                      "frag %u mds%d (auth)\n",
 826                                      inode, ceph_vinop(inode), frag.frag, mds);
 827                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 828                                     CEPH_MDS_STATE_ACTIVE)
 829                                         goto out;
 830                         }
 831                 }
 832         }
 833
 834         spin_lock(&ci->i_ceph_lock);
 835         cap = NULL;
 836         if (mode == USE_AUTH_MDS)
 837                 cap = ci->i_auth_cap;
 838         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
 839                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
 840         if (!cap) {
 841                 spin_unlock(&ci->i_ceph_lock);
 842                 iput(inode);
 843                 goto random;
 844         }
 845         mds = cap->session->s_mds;
 846         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
 847              inode, ceph_vinop(inode), mds,
 848              cap == ci->i_auth_cap ? "auth " : "", cap);
 849         spin_unlock(&ci->i_ceph_lock);
 850 out:
 851         iput(inode);
 852         return mds;
 853
 854 random:
 855         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
 856         dout("choose_mds chose random mds%d\n", mds);
 857         return mds;
 858 }
 859
 860
 861 /*
 862  * session messages
 863  */
 864 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
 865 {
 866         struct ceph_msg *msg;
 867         struct ceph_mds_session_head *h;
 868
 869         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
 870                            false);
 871         if (!msg) {
 872                 pr_err("create_session_msg ENOMEM creating msg\n");
 873                 return NULL;
 874         }
 875         h = msg->front.iov_base;
 876         h->op = cpu_to_le32(op);
 877         h->seq = cpu_to_le64(seq);
 878
 879         return msg;
 880 }
 881
 882 /*
 883  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
 884  * to include additional client metadata fields.
 885  */
 886 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
 887 {
 888         struct ceph_msg *msg;
 889         struct ceph_mds_session_head *h;
 890         int i = -1;
 891         int metadata_bytes = 0;
 892         int metadata_key_count = 0;
 893         struct ceph_options *opt = mdsc->fsc->client->options;
 894         struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
 895         void *p;
 896
 897         const char* metadata[][2] = {
 898                 {"hostname", mdsc->nodename},
 899                 {"kernel_version", init_utsname()->release},
 900                 {"entity_id", opt->name ? : ""},
 901                 {"root", fsopt->server_path ? : "/"},
 902                 {NULL, NULL}
 903         };
 904
 905         /* Calculate serialized length of metadata */
 906         metadata_bytes = 4;  /* map length */
 907         for (i = 0; metadata[i][0]; ++i) {
 908                 metadata_bytes += 8 + strlen(metadata[i][0]) +
 909                         strlen(metadata[i][1]);
 910                 metadata_key_count++;
 911         }
 912
 913         /* Allocate the message */
 914         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + metadata_bytes,
 915                            GFP_NOFS, false);
 916         if (!msg) {
 917                 pr_err("create_session_msg ENOMEM creating msg\n");
 918                 return NULL;
 919         }
 920         h = msg->front.iov_base;
 921         h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
 922         h->seq = cpu_to_le64(seq);
 923
 924         /*
 925          * Serialize client metadata into waiting buffer space, using
 926          * the format that userspace expects for map<string, string>
 927          *
 928          * ClientSession messages with metadata are v2
 929          */
 930         msg->hdr.version = cpu_to_le16(2);
 931         msg->hdr.compat_version = cpu_to_le16(1);
 932
 933         /* The write pointer, following the session_head structure */
 934         p = msg->front.iov_base + sizeof(*h);
 935
 936         /* Number of entries in the map */
 937         ceph_encode_32(&p, metadata_key_count);
 938
 939         /* Two length-prefixed strings for each entry in the map */
 940         for (i = 0; metadata[i][0]; ++i) {
 941                 size_t const key_len = strlen(metadata[i][0]);
 942                 size_t const val_len = strlen(metadata[i][1]);
 943
 944                 ceph_encode_32(&p, key_len);
 945                 memcpy(p, metadata[i][0], key_len);
 946                 p += key_len;
 947                 ceph_encode_32(&p, val_len);
 948                 memcpy(p, metadata[i][1], val_len);
 949                 p += val_len;
 950         }
 951
 952         return msg;
 953 }
 954
 955 /*
 956  * send session open request.
 957  *
 958  * called under mdsc->mutex
 959  */
 960 static int __open_session(struct ceph_mds_client *mdsc,
 961                           struct ceph_mds_session *session)
 962 {
 963         struct ceph_msg *msg;
 964         int mstate;
 965         int mds = session->s_mds;
 966
 967         /* wait for mds to go active? */
 968         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
 969         dout("open_session to mds%d (%s)\n", mds,
 970              ceph_mds_state_name(mstate));
 971         session->s_state = CEPH_MDS_SESSION_OPENING;
 972         session->s_renew_requested = jiffies;
 973
 974         /* send connect message */
 975         msg = create_session_open_msg(mdsc, session->s_seq);
 976         if (!msg)
 977                 return -ENOMEM;
 978         ceph_con_send(&session->s_con, msg);
 979         return 0;
 980 }
 981
 982 /*
 983  * open sessions for any export targets for the given mds
 984  *
 985  * called under mdsc->mutex
 986  */
 987 static struct ceph_mds_session *
 988 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
 989 {
 990         struct ceph_mds_session *session;
 991
 992         session = __ceph_lookup_mds_session(mdsc, target);
 993         if (!session) {
 994                 session = register_session(mdsc, target);
 995                 if (IS_ERR(session))
 996                         return session;
 997         }
 998         if (session->s_state == CEPH_MDS_SESSION_NEW ||
 999             session->s_state == CEPH_MDS_SESSION_CLOSING)
1000                 __open_session(mdsc, session);
1001
1002         return session;
1003 }
1004
1005 struct ceph_mds_session *
1006 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1007 {
1008         struct ceph_mds_session *session;
1009
1010         dout("open_export_target_session to mds%d\n", target);
1011
1012         mutex_lock(&mdsc->mutex);
1013         session = __open_export_target_session(mdsc, target);
1014         mutex_unlock(&mdsc->mutex);
1015
1016         return session;
1017 }
1018
1019 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1020                                           struct ceph_mds_session *session)
1021 {
1022         struct ceph_mds_info *mi;
1023         struct ceph_mds_session *ts;
1024         int i, mds = session->s_mds;
1025
1026         if (mds >= mdsc->mdsmap->m_num_mds)
1027                 return;
1028
1029         mi = &mdsc->mdsmap->m_info[mds];
1030         dout("open_export_target_sessions for mds%d (%d targets)\n",
1031              session->s_mds, mi->num_export_targets);
1032
1033         for (i = 0; i < mi->num_export_targets; i++) {
1034                 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1035                 if (!IS_ERR(ts))
1036                         ceph_put_mds_session(ts);
1037         }
1038 }
1039
1040 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1041                                            struct ceph_mds_session *session)
1042 {
1043         mutex_lock(&mdsc->mutex);
1044         __open_export_target_sessions(mdsc, session);
1045         mutex_unlock(&mdsc->mutex);
1046 }
1047
1048 /*
1049  * session caps
1050  */
1051
1052 static void detach_cap_releases(struct ceph_mds_session *session,
1053                                 struct list_head *target)
1054 {
1055         lockdep_assert_held(&session->s_cap_lock);
1056
1057         list_splice_init(&session->s_cap_releases, target);
1058         session->s_num_cap_releases = 0;
1059         dout("dispose_cap_releases mds%d\n", session->s_mds);
1060 }
1061
1062 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1063                                  struct list_head *dispose)
1064 {
1065         while (!list_empty(dispose)) {
1066                 struct ceph_cap *cap;
1067                 /* zero out the in-progress message */
1068                 cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1069                 list_del(&cap->session_caps);
1070                 ceph_put_cap(mdsc, cap);
1071         }
1072 }
1073
1074 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1075                                      struct ceph_mds_session *session)
1076 {
1077         struct ceph_mds_request *req;
1078         struct rb_node *p;
1079
1080         dout("cleanup_session_requests mds%d\n", session->s_mds);
1081         mutex_lock(&mdsc->mutex);
1082         while (!list_empty(&session->s_unsafe)) {
1083                 req = list_first_entry(&session->s_unsafe,
1084                                        struct ceph_mds_request, r_unsafe_item);
1085                 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1086                                     req->r_tid);
1087                 __unregister_request(mdsc, req);
1088         }
1089         /* zero r_attempts, so kick_requests() will re-send requests */
1090         p = rb_first(&mdsc->request_tree);
1091         while (p) {
1092                 req = rb_entry(p, struct ceph_mds_request, r_node);
1093                 p = rb_next(p);
1094                 if (req->r_session &&
1095                     req->r_session->s_mds == session->s_mds)
1096                         req->r_attempts = 0;
1097         }
1098         mutex_unlock(&mdsc->mutex);
1099 }
1100
1101 /*
1102  * Helper to safely iterate over all caps associated with a session, with
1103  * special care taken to handle a racing __ceph_remove_cap().
1104  *
1105  * Caller must hold session s_mutex.
1106  */
1107 static int iterate_session_caps(struct ceph_mds_session *session,
1108                                  int (*cb)(struct inode *, struct ceph_cap *,
1109                                             void *), void *arg)
1110 {
1111         struct list_head *p;
1112         struct ceph_cap *cap;
1113         struct inode *inode, *last_inode = NULL;
1114         struct ceph_cap *old_cap = NULL;
1115         int ret;
1116
1117         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1118         spin_lock(&session->s_cap_lock);
1119         p = session->s_caps.next;
1120         while (p != &session->s_caps) {
1121                 cap = list_entry(p, struct ceph_cap, session_caps);
1122                 inode = igrab(&cap->ci->vfs_inode);
1123                 if (!inode) {
1124                         p = p->next;
1125                         continue;
1126                 }
1127                 session->s_cap_iterator = cap;
1128                 spin_unlock(&session->s_cap_lock);
1129
1130                 if (last_inode) {
1131                         iput(last_inode);
1132                         last_inode = NULL;
1133                 }
1134                 if (old_cap) {
1135                         ceph_put_cap(session->s_mdsc, old_cap);
1136                         old_cap = NULL;
1137                 }
1138
1139                 ret = cb(inode, cap, arg);
1140                 last_inode = inode;
1141
1142                 spin_lock(&session->s_cap_lock);
1143                 p = p->next;
1144                 if (!cap->ci) {
1145                         dout("iterate_session_caps  finishing cap %p removal\n",
1146                              cap);
1147                         BUG_ON(cap->session != session);
1148                         cap->session = NULL;
1149                         list_del_init(&cap->session_caps);
1150                         session->s_nr_caps--;
1151                         if (cap->queue_release) {
1152                                 list_add_tail(&cap->session_caps,
1153                                               &session->s_cap_releases);
1154                                 session->s_num_cap_releases++;
1155                         } else {
1156                                 old_cap = cap;  /* put_cap it w/o locks held */
1157                         }
1158                 }
1159                 if (ret < 0)
1160                         goto out;
1161         }
1162         ret = 0;
1163 out:
1164         session->s_cap_iterator = NULL;
1165         spin_unlock(&session->s_cap_lock);
1166
1167         iput(last_inode);
1168         if (old_cap)
1169                 ceph_put_cap(session->s_mdsc, old_cap);
1170
1171         return ret;
1172 }
1173
1174 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1175                                   void *arg)
1176 {
1177         struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1178         struct ceph_inode_info *ci = ceph_inode(inode);
1179         LIST_HEAD(to_remove);
1180         bool drop = false;
1181         bool invalidate = false;
1182
1183         dout("removing cap %p, ci is %p, inode is %p\n",
1184              cap, ci, &ci->vfs_inode);
1185         spin_lock(&ci->i_ceph_lock);
1186         __ceph_remove_cap(cap, false);
1187         if (!ci->i_auth_cap) {
1188                 struct ceph_cap_flush *cf;
1189                 struct ceph_mds_client *mdsc = fsc->mdsc;
1190
1191                 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1192
1193                 if (ci->i_wrbuffer_ref > 0 &&
1194                     READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
1195                         invalidate = true;
1196
1197                 while (!list_empty(&ci->i_cap_flush_list)) {
1198                         cf = list_first_entry(&ci->i_cap_flush_list,
1199                                               struct ceph_cap_flush, i_list);
1200                         list_move(&cf->i_list, &to_remove);
1201                 }
1202
1203                 spin_lock(&mdsc->cap_dirty_lock);
1204
1205                 list_for_each_entry(cf, &to_remove, i_list)
1206                         list_del(&cf->g_list);
1207
1208                 if (!list_empty(&ci->i_dirty_item)) {
1209                         pr_warn_ratelimited(
1210                                 " dropping dirty %s state for %p %lld\n",
1211                                 ceph_cap_string(ci->i_dirty_caps),
1212                                 inode, ceph_ino(inode));
1213                         ci->i_dirty_caps = 0;
1214                         list_del_init(&ci->i_dirty_item);
1215                         drop = true;
1216                 }
1217                 if (!list_empty(&ci->i_flushing_item)) {
1218                         pr_warn_ratelimited(
1219                                 " dropping dirty+flushing %s state for %p %lld\n",
1220                                 ceph_cap_string(ci->i_flushing_caps),
1221                                 inode, ceph_ino(inode));
1222                         ci->i_flushing_caps = 0;
1223                         list_del_init(&ci->i_flushing_item);
1224                         mdsc->num_cap_flushing--;
1225                         drop = true;
1226                 }
1227                 spin_unlock(&mdsc->cap_dirty_lock);
1228
1229                 if (atomic_read(&ci->i_filelock_ref) > 0) {
1230                         /* make further file lock syscall return -EIO */
1231                         ci->i_ceph_flags |= CEPH_I_ERROR_FILELOCK;
1232                         pr_warn_ratelimited(" dropping file locks for %p %lld\n",
1233                                             inode, ceph_ino(inode));
1234                 }
1235
1236                 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1237                         list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1238                         ci->i_prealloc_cap_flush = NULL;
1239                 }
1240         }
1241         spin_unlock(&ci->i_ceph_lock);
1242         while (!list_empty(&to_remove)) {
1243                 struct ceph_cap_flush *cf;
1244                 cf = list_first_entry(&to_remove,
1245                                       struct ceph_cap_flush, i_list);
1246                 list_del(&cf->i_list);
1247                 ceph_free_cap_flush(cf);
1248         }
1249
1250         wake_up_all(&ci->i_cap_wq);
1251         if (invalidate)
1252                 ceph_queue_invalidate(inode);
1253         if (drop)
1254                 iput(inode);
1255         return 0;
1256 }
1257
1258 /*
1259  * caller must hold session s_mutex
1260  */
1261 static void remove_session_caps(struct ceph_mds_session *session)
1262 {
1263         struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1264         struct super_block *sb = fsc->sb;
1265         LIST_HEAD(dispose);
1266
1267         dout("remove_session_caps on %p\n", session);
1268         iterate_session_caps(session, remove_session_caps_cb, fsc);
1269
1270         wake_up_all(&fsc->mdsc->cap_flushing_wq);
1271
1272         spin_lock(&session->s_cap_lock);
1273         if (session->s_nr_caps > 0) {
1274                 struct inode *inode;
1275                 struct ceph_cap *cap, *prev = NULL;
1276                 struct ceph_vino vino;
1277                 /*
1278                  * iterate_session_caps() skips inodes that are being
1279                  * deleted, we need to wait until deletions are complete.
1280                  * __wait_on_freeing_inode() is designed for the job,
1281                  * but it is not exported, so use lookup inode function
1282                  * to access it.
1283                  */
1284                 while (!list_empty(&session->s_caps)) {
1285                         cap = list_entry(session->s_caps.next,
1286                                          struct ceph_cap, session_caps);
1287                         if (cap == prev)
1288                                 break;
1289                         prev = cap;
1290                         vino = cap->ci->i_vino;
1291                         spin_unlock(&session->s_cap_lock);
1292
1293                         inode = ceph_find_inode(sb, vino);
1294                         iput(inode);
1295
1296                         spin_lock(&session->s_cap_lock);
1297                 }
1298         }
1299
1300         // drop cap expires and unlock s_cap_lock
1301         detach_cap_releases(session, &dispose);
1302
1303         BUG_ON(session->s_nr_caps > 0);
1304         BUG_ON(!list_empty(&session->s_cap_flushing));
1305         spin_unlock(&session->s_cap_lock);
1306         dispose_cap_releases(session->s_mdsc, &dispose);
1307 }
1308
1309 /*
1310  * wake up any threads waiting on this session's caps.  if the cap is
1311  * old (didn't get renewed on the client reconnect), remove it now.
1312  *
1313  * caller must hold s_mutex.
1314  */
1315 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1316                               void *arg)
1317 {
1318         struct ceph_inode_info *ci = ceph_inode(inode);
1319
1320         if (arg) {
1321                 spin_lock(&ci->i_ceph_lock);
1322                 ci->i_wanted_max_size = 0;
1323                 ci->i_requested_max_size = 0;
1324                 spin_unlock(&ci->i_ceph_lock);
1325         }
1326         wake_up_all(&ci->i_cap_wq);
1327         return 0;
1328 }
1329
1330 static void wake_up_session_caps(struct ceph_mds_session *session,
1331                                  int reconnect)
1332 {
1333         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1334         iterate_session_caps(session, wake_up_session_cb,
1335                              (void *)(unsigned long)reconnect);
1336 }
1337
1338 /*
1339  * Send periodic message to MDS renewing all currently held caps.  The
1340  * ack will reset the expiration for all caps from this session.
1341  *
1342  * caller holds s_mutex
1343  */
1344 static int send_renew_caps(struct ceph_mds_client *mdsc,
1345                            struct ceph_mds_session *session)
1346 {
1347         struct ceph_msg *msg;
1348         int state;
1349
1350         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1351             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1352                 pr_info("mds%d caps stale\n", session->s_mds);
1353         session->s_renew_requested = jiffies;
1354
1355         /* do not try to renew caps until a recovering mds has reconnected
1356          * with its clients. */
1357         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1358         if (state < CEPH_MDS_STATE_RECONNECT) {
1359                 dout("send_renew_caps ignoring mds%d (%s)\n",
1360                      session->s_mds, ceph_mds_state_name(state));
1361                 return 0;
1362         }
1363
1364         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1365                 ceph_mds_state_name(state));
1366         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1367                                  ++session->s_renew_seq);
1368         if (!msg)
1369                 return -ENOMEM;
1370         ceph_con_send(&session->s_con, msg);
1371         return 0;
1372 }
1373
1374 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1375                              struct ceph_mds_session *session, u64 seq)
1376 {
1377         struct ceph_msg *msg;
1378
1379         dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1380              session->s_mds, ceph_session_state_name(session->s_state), seq);
1381         msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1382         if (!msg)
1383                 return -ENOMEM;
1384         ceph_con_send(&session->s_con, msg);
1385         return 0;
1386 }
1387
1388
1389 /*
1390  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1391  *
1392  * Called under session->s_mutex
1393  */
1394 static void renewed_caps(struct ceph_mds_client *mdsc,
1395                          struct ceph_mds_session *session, int is_renew)
1396 {
1397         int was_stale;
1398         int wake = 0;
1399
1400         spin_lock(&session->s_cap_lock);
1401         was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1402
1403         session->s_cap_ttl = session->s_renew_requested +
1404                 mdsc->mdsmap->m_session_timeout*HZ;
1405
1406         if (was_stale) {
1407                 if (time_before(jiffies, session->s_cap_ttl)) {
1408                         pr_info("mds%d caps renewed\n", session->s_mds);
1409                         wake = 1;
1410                 } else {
1411                         pr_info("mds%d caps still stale\n", session->s_mds);
1412                 }
1413         }
1414         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1415              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1416              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1417         spin_unlock(&session->s_cap_lock);
1418
1419         if (wake)
1420                 wake_up_session_caps(session, 0);
1421 }
1422
1423 /*
1424  * send a session close request
1425  */
1426 static int request_close_session(struct ceph_mds_client *mdsc,
1427                                  struct ceph_mds_session *session)
1428 {
1429         struct ceph_msg *msg;
1430
1431         dout("request_close_session mds%d state %s seq %lld\n",
1432              session->s_mds, ceph_session_state_name(session->s_state),
1433              session->s_seq);
1434         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1435         if (!msg)
1436                 return -ENOMEM;
1437         ceph_con_send(&session->s_con, msg);
1438         return 1;
1439 }
1440
1441 /*
1442  * Called with s_mutex held.
1443  */
1444 static int __close_session(struct ceph_mds_client *mdsc,
1445                          struct ceph_mds_session *session)
1446 {
1447         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1448                 return 0;
1449         session->s_state = CEPH_MDS_SESSION_CLOSING;
1450         return request_close_session(mdsc, session);
1451 }
1452
1453 static bool drop_negative_children(struct dentry *dentry)
1454 {
1455         struct dentry *child;
1456         bool all_negative = true;
1457
1458         if (!d_is_dir(dentry))
1459                 goto out;
1460
1461         spin_lock(&dentry->d_lock);
1462         list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1463                 if (d_really_is_positive(child)) {
1464                         all_negative = false;
1465                         break;
1466                 }
1467         }
1468         spin_unlock(&dentry->d_lock);
1469
1470         if (all_negative)
1471                 shrink_dcache_parent(dentry);
1472 out:
1473         return all_negative;
1474 }
1475
1476 /*
1477  * Trim old(er) caps.
1478  *
1479  * Because we can't cache an inode without one or more caps, we do
1480  * this indirectly: if a cap is unused, we prune its aliases, at which
1481  * point the inode will hopefully get dropped to.
1482  *
1483  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1484  * memory pressure from the MDS, though, so it needn't be perfect.
1485  */
1486 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1487 {
1488         struct ceph_mds_session *session = arg;
1489         struct ceph_inode_info *ci = ceph_inode(inode);
1490         int used, wanted, oissued, mine;
1491
1492         if (session->s_trim_caps <= 0)
1493                 return -1;
1494
1495         spin_lock(&ci->i_ceph_lock);
1496         mine = cap->issued | cap->implemented;
1497         used = __ceph_caps_used(ci);
1498         wanted = __ceph_caps_file_wanted(ci);
1499         oissued = __ceph_caps_issued_other(ci, cap);
1500
1501         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1502              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1503              ceph_cap_string(used), ceph_cap_string(wanted));
1504         if (cap == ci->i_auth_cap) {
1505                 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1506                     !list_empty(&ci->i_cap_snaps))
1507                         goto out;
1508                 if ((used | wanted) & CEPH_CAP_ANY_WR)
1509                         goto out;
1510                 /* Note: it's possible that i_filelock_ref becomes non-zero
1511                  * after dropping auth caps. It doesn't hurt because reply
1512                  * of lock mds request will re-add auth caps. */
1513                 if (atomic_read(&ci->i_filelock_ref) > 0)
1514                         goto out;
1515         }
1516         /* The inode has cached pages, but it's no longer used.
1517          * we can safely drop it */
1518         if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1519             !(oissued & CEPH_CAP_FILE_CACHE)) {
1520           used = 0;
1521           oissued = 0;
1522         }
1523         if ((used | wanted) & ~oissued & mine)
1524                 goto out;   /* we need these caps */
1525
1526         if (oissued) {
1527                 /* we aren't the only cap.. just remove us */
1528                 __ceph_remove_cap(cap, true);
1529                 session->s_trim_caps--;
1530         } else {
1531                 struct dentry *dentry;
1532                 /* try dropping referring dentries */
1533                 spin_unlock(&ci->i_ceph_lock);
1534                 dentry = d_find_any_alias(inode);
1535                 if (dentry && drop_negative_children(dentry)) {
1536                         int count;
1537                         dput(dentry);
1538                         d_prune_aliases(inode);
1539                         count = atomic_read(&inode->i_count);
1540                         if (count == 1)
1541                                 session->s_trim_caps--;
1542                         dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1543                              inode, cap, count);
1544                 } else {
1545                         dput(dentry);
1546                 }
1547                 return 0;
1548         }
1549
1550 out:
1551         spin_unlock(&ci->i_ceph_lock);
1552         return 0;
1553 }
1554
1555 /*
1556  * Trim session cap count down to some max number.
1557  */
1558 int ceph_trim_caps(struct ceph_mds_client *mdsc,
1559                    struct ceph_mds_session *session,
1560                    int max_caps)
1561 {
1562         int trim_caps = session->s_nr_caps - max_caps;
1563
1564         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1565              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1566         if (trim_caps > 0) {
1567                 session->s_trim_caps = trim_caps;
1568                 iterate_session_caps(session, trim_caps_cb, session);
1569                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1570                      session->s_mds, session->s_nr_caps, max_caps,
1571                         trim_caps - session->s_trim_caps);
1572                 session->s_trim_caps = 0;
1573         }
1574
1575         ceph_send_cap_releases(mdsc, session);
1576         return 0;
1577 }
1578
1579 static int check_caps_flush(struct ceph_mds_client *mdsc,
1580                             u64 want_flush_tid)
1581 {
1582         int ret = 1;
1583
1584         spin_lock(&mdsc->cap_dirty_lock);
1585         if (!list_empty(&mdsc->cap_flush_list)) {
1586                 struct ceph_cap_flush *cf =
1587                         list_first_entry(&mdsc->cap_flush_list,
1588                                          struct ceph_cap_flush, g_list);
1589                 if (cf->tid <= want_flush_tid) {
1590                         dout("check_caps_flush still flushing tid "
1591                              "%llu <= %llu\n", cf->tid, want_flush_tid);
1592                         ret = 0;
1593                 }
1594         }
1595         spin_unlock(&mdsc->cap_dirty_lock);
1596         return ret;
1597 }
1598
1599 /*
1600  * flush all dirty inode data to disk.
1601  *
1602  * returns true if we've flushed through want_flush_tid
1603  */
1604 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1605                             u64 want_flush_tid)
1606 {
1607         dout("check_caps_flush want %llu\n", want_flush_tid);
1608
1609         wait_event(mdsc->cap_flushing_wq,
1610                    check_caps_flush(mdsc, want_flush_tid));
1611
1612         dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1613 }
1614
1615 /*
1616  * called under s_mutex
1617  */
1618 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1619                             struct ceph_mds_session *session)
1620 {
1621         struct ceph_msg *msg = NULL;
1622         struct ceph_mds_cap_release *head;
1623         struct ceph_mds_cap_item *item;
1624         struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
1625         struct ceph_cap *cap;
1626         LIST_HEAD(tmp_list);
1627         int num_cap_releases;
1628         __le32  barrier, *cap_barrier;
1629
1630         down_read(&osdc->lock);
1631         barrier = cpu_to_le32(osdc->epoch_barrier);
1632         up_read(&osdc->lock);
1633
1634         spin_lock(&session->s_cap_lock);
1635 again:
1636         list_splice_init(&session->s_cap_releases, &tmp_list);
1637         num_cap_releases = session->s_num_cap_releases;
1638         session->s_num_cap_releases = 0;
1639         spin_unlock(&session->s_cap_lock);
1640
1641         while (!list_empty(&tmp_list)) {
1642                 if (!msg) {
1643                         msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1644                                         PAGE_SIZE, GFP_NOFS, false);
1645                         if (!msg)
1646                                 goto out_err;
1647                         head = msg->front.iov_base;
1648                         head->num = cpu_to_le32(0);
1649                         msg->front.iov_len = sizeof(*head);
1650
1651                         msg->hdr.version = cpu_to_le16(2);
1652                         msg->hdr.compat_version = cpu_to_le16(1);
1653                 }
1654
1655                 cap = list_first_entry(&tmp_list, struct ceph_cap,
1656                                         session_caps);
1657                 list_del(&cap->session_caps);
1658                 num_cap_releases--;
1659
1660                 head = msg->front.iov_base;
1661                 le32_add_cpu(&head->num, 1);
1662                 item = msg->front.iov_base + msg->front.iov_len;
1663                 item->ino = cpu_to_le64(cap->cap_ino);
1664                 item->cap_id = cpu_to_le64(cap->cap_id);
1665                 item->migrate_seq = cpu_to_le32(cap->mseq);
1666                 item->seq = cpu_to_le32(cap->issue_seq);
1667                 msg->front.iov_len += sizeof(*item);
1668
1669                 ceph_put_cap(mdsc, cap);
1670
1671                 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1672                         // Append cap_barrier field
1673                         cap_barrier = msg->front.iov_base + msg->front.iov_len;
1674                         *cap_barrier = barrier;
1675                         msg->front.iov_len += sizeof(*cap_barrier);
1676
1677                         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1678                         dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1679                         ceph_con_send(&session->s_con, msg);
1680                         msg = NULL;
1681                 }
1682         }
1683
1684         BUG_ON(num_cap_releases != 0);
1685
1686         spin_lock(&session->s_cap_lock);
1687         if (!list_empty(&session->s_cap_releases))
1688                 goto again;
1689         spin_unlock(&session->s_cap_lock);
1690
1691         if (msg) {
1692                 // Append cap_barrier field
1693                 cap_barrier = msg->front.iov_base + msg->front.iov_len;
1694                 *cap_barrier = barrier;
1695                 msg->front.iov_len += sizeof(*cap_barrier);
1696
1697                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1698                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1699                 ceph_con_send(&session->s_con, msg);
1700         }
1701         return;
1702 out_err:
1703         pr_err("send_cap_releases mds%d, failed to allocate message\n",
1704                 session->s_mds);
1705         spin_lock(&session->s_cap_lock);
1706         list_splice(&tmp_list, &session->s_cap_releases);
1707         session->s_num_cap_releases += num_cap_releases;
1708         spin_unlock(&session->s_cap_lock);
1709 }
1710
1711 /*
1712  * requests
1713  */
1714
1715 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1716                                     struct inode *dir)
1717 {
1718         struct ceph_inode_info *ci = ceph_inode(dir);
1719         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1720         struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1721         size_t size = sizeof(struct ceph_mds_reply_dir_entry);
1722         int order, num_entries;
1723
1724         spin_lock(&ci->i_ceph_lock);
1725         num_entries = ci->i_files + ci->i_subdirs;
1726         spin_unlock(&ci->i_ceph_lock);
1727         num_entries = max(num_entries, 1);
1728         num_entries = min(num_entries, opt->max_readdir);
1729
1730         order = get_order(size * num_entries);
1731         while (order >= 0) {
1732                 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
1733                                                              __GFP_NOWARN,
1734                                                              order);
1735                 if (rinfo->dir_entries)
1736                         break;
1737                 order--;
1738         }
1739         if (!rinfo->dir_entries)
1740                 return -ENOMEM;
1741
1742         num_entries = (PAGE_SIZE << order) / size;
1743         num_entries = min(num_entries, opt->max_readdir);
1744
1745         rinfo->dir_buf_size = PAGE_SIZE << order;
1746         req->r_num_caps = num_entries + 1;
1747         req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1748         req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1749         return 0;
1750 }
1751
1752 /*
1753  * Create an mds request.
1754  */
1755 struct ceph_mds_request *
1756 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1757 {
1758         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1759
1760         if (!req)
1761                 return ERR_PTR(-ENOMEM);
1762
1763         mutex_init(&req->r_fill_mutex);
1764         req->r_mdsc = mdsc;
1765         req->r_started = jiffies;
1766         req->r_resend_mds = -1;
1767         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1768         INIT_LIST_HEAD(&req->r_unsafe_target_item);
1769         req->r_fmode = -1;
1770         kref_init(&req->r_kref);
1771         RB_CLEAR_NODE(&req->r_node);
1772         INIT_LIST_HEAD(&req->r_wait);
1773         init_completion(&req->r_completion);
1774         init_completion(&req->r_safe_completion);
1775         INIT_LIST_HEAD(&req->r_unsafe_item);
1776
1777         req->r_stamp = timespec_trunc(current_kernel_time(), mdsc->fsc->sb->s_time_gran);
1778
1779         req->r_op = op;
1780         req->r_direct_mode = mode;
1781         return req;
1782 }
1783
1784 /*
1785  * return oldest (lowest) request, tid in request tree, 0 if none.
1786  *
1787  * called under mdsc->mutex.
1788  */
1789 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1790 {
1791         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1792                 return NULL;
1793         return rb_entry(rb_first(&mdsc->request_tree),
1794                         struct ceph_mds_request, r_node);
1795 }
1796
1797 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1798 {
1799         return mdsc->oldest_tid;
1800 }
1801
1802 /*
1803  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1804  * on build_path_from_dentry in fs/cifs/dir.c.
1805  *
1806  * If @stop_on_nosnap, generate path relative to the first non-snapped
1807  * inode.
1808  *
1809  * Encode hidden .snap dirs as a double /, i.e.
1810  *   foo/.snap/bar -> foo//bar
1811  */
1812 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1813                            int stop_on_nosnap)
1814 {
1815         struct dentry *temp;
1816         char *path;
1817         int len, pos;
1818         unsigned seq;
1819
1820         if (!dentry)
1821                 return ERR_PTR(-EINVAL);
1822
1823 retry:
1824         len = 0;
1825         seq = read_seqbegin(&rename_lock);
1826         rcu_read_lock();
1827         for (temp = dentry; !IS_ROOT(temp);) {
1828                 struct inode *inode = d_inode(temp);
1829                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1830                         len++;  /* slash only */
1831                 else if (stop_on_nosnap && inode &&
1832                          ceph_snap(inode) == CEPH_NOSNAP)
1833                         break;
1834                 else
1835                         len += 1 + temp->d_name.len;
1836                 temp = temp->d_parent;
1837         }
1838         rcu_read_unlock();
1839         if (len)
1840                 len--;  /* no leading '/' */
1841
1842         path = kmalloc(len+1, GFP_NOFS);
1843         if (!path)
1844                 return ERR_PTR(-ENOMEM);
1845         pos = len;
1846         path[pos] = 0;  /* trailing null */
1847         rcu_read_lock();
1848         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1849                 struct inode *inode;
1850
1851                 spin_lock(&temp->d_lock);
1852                 inode = d_inode(temp);
1853                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1854                         dout("build_path path+%d: %p SNAPDIR\n",
1855                              pos, temp);
1856                 } else if (stop_on_nosnap && inode &&
1857                            ceph_snap(inode) == CEPH_NOSNAP) {
1858                         spin_unlock(&temp->d_lock);
1859                         break;
1860                 } else {
1861                         pos -= temp->d_name.len;
1862                         if (pos < 0) {
1863                                 spin_unlock(&temp->d_lock);
1864                                 break;
1865                         }
1866                         strncpy(path + pos, temp->d_name.name,
1867                                 temp->d_name.len);
1868                 }
1869                 spin_unlock(&temp->d_lock);
1870                 if (pos)
1871                         path[--pos] = '/';
1872                 temp = temp->d_parent;
1873         }
1874         rcu_read_unlock();
1875         if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1876                 pr_err("build_path did not end path lookup where "
1877                        "expected, namelen is %d, pos is %d\n", len, pos);
1878                 /* presumably this is only possible if racing with a
1879                    rename of one of the parent directories (we can not
1880                    lock the dentries above us to prevent this, but
1881                    retrying should be harmless) */
1882                 kfree(path);
1883                 goto retry;
1884         }
1885
1886         *base = ceph_ino(d_inode(temp));
1887         *plen = len;
1888         dout("build_path on %p %d built %llx '%.*s'\n",
1889              dentry, d_count(dentry), *base, len, path);
1890         return path;
1891 }
1892
1893 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
1894                              const char **ppath, int *ppathlen, u64 *pino,
1895                              int *pfreepath)
1896 {
1897         char *path;
1898
1899         rcu_read_lock();
1900         if (!dir)
1901                 dir = d_inode_rcu(dentry->d_parent);
1902         if (dir && ceph_snap(dir) == CEPH_NOSNAP) {
1903                 *pino = ceph_ino(dir);
1904                 rcu_read_unlock();
1905                 *ppath = dentry->d_name.name;
1906                 *ppathlen = dentry->d_name.len;
1907                 return 0;
1908         }
1909         rcu_read_unlock();
1910         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1911         if (IS_ERR(path))
1912                 return PTR_ERR(path);
1913         *ppath = path;
1914         *pfreepath = 1;
1915         return 0;
1916 }
1917
1918 static int build_inode_path(struct inode *inode,
1919                             const char **ppath, int *ppathlen, u64 *pino,
1920                             int *pfreepath)
1921 {
1922         struct dentry *dentry;
1923         char *path;
1924
1925         if (ceph_snap(inode) == CEPH_NOSNAP) {
1926                 *pino = ceph_ino(inode);
1927                 *ppathlen = 0;
1928                 return 0;
1929         }
1930         dentry = d_find_alias(inode);
1931         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1932         dput(dentry);
1933         if (IS_ERR(path))
1934                 return PTR_ERR(path);
1935         *ppath = path;
1936         *pfreepath = 1;
1937         return 0;
1938 }
1939
1940 /*
1941  * request arguments may be specified via an inode *, a dentry *, or
1942  * an explicit ino+path.
1943  */
1944 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1945                                   struct inode *rdiri, const char *rpath,
1946                                   u64 rino, const char **ppath, int *pathlen,
1947                                   u64 *ino, int *freepath)
1948 {
1949         int r = 0;
1950
1951         if (rinode) {
1952                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1953                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1954                      ceph_snap(rinode));
1955         } else if (rdentry) {
1956                 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
1957                                         freepath);
1958                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1959                      *ppath);
1960         } else if (rpath || rino) {
1961                 *ino = rino;
1962                 *ppath = rpath;
1963                 *pathlen = rpath ? strlen(rpath) : 0;
1964                 dout(" path %.*s\n", *pathlen, rpath);
1965         }
1966
1967         return r;
1968 }
1969
1970 /*
1971  * called under mdsc->mutex
1972  */
1973 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1974                                                struct ceph_mds_request *req,
1975                                                int mds, bool drop_cap_releases)
1976 {
1977         struct ceph_msg *msg;
1978         struct ceph_mds_request_head *head;
1979         const char *path1 = NULL;
1980         const char *path2 = NULL;
1981         u64 ino1 = 0, ino2 = 0;
1982         int pathlen1 = 0, pathlen2 = 0;
1983         int freepath1 = 0, freepath2 = 0;
1984         int len;
1985         u16 releases;
1986         void *p, *end;
1987         int ret;
1988
1989         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1990                               req->r_parent, req->r_path1, req->r_ino1.ino,
1991                               &path1, &pathlen1, &ino1, &freepath1);
1992         if (ret < 0) {
1993                 msg = ERR_PTR(ret);
1994                 goto out;
1995         }
1996
1997         ret = set_request_path_attr(NULL, req->r_old_dentry,
1998                               req->r_old_dentry_dir,
1999                               req->r_path2, req->r_ino2.ino,
2000                               &path2, &pathlen2, &ino2, &freepath2);
2001         if (ret < 0) {
2002                 msg = ERR_PTR(ret);
2003                 goto out_free1;
2004         }
2005
2006         len = sizeof(*head) +
2007                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2008                 sizeof(struct ceph_timespec);
2009
2010         /* calculate (max) length for cap releases */
2011         len += sizeof(struct ceph_mds_request_release) *
2012                 (!!req->r_inode_drop + !!req->r_dentry_drop +
2013                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2014         if (req->r_dentry_drop)
2015                 len += req->r_dentry->d_name.len;
2016         if (req->r_old_dentry_drop)
2017                 len += req->r_old_dentry->d_name.len;
2018
2019         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
2020         if (!msg) {
2021                 msg = ERR_PTR(-ENOMEM);
2022                 goto out_free2;
2023         }
2024
2025         msg->hdr.version = cpu_to_le16(2);
2026         msg->hdr.tid = cpu_to_le64(req->r_tid);
2027
2028         head = msg->front.iov_base;
2029         p = msg->front.iov_base + sizeof(*head);
2030         end = msg->front.iov_base + msg->front.iov_len;
2031
2032         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2033         head->op = cpu_to_le32(req->r_op);
2034         head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
2035         head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
2036         head->args = req->r_args;
2037
2038         ceph_encode_filepath(&p, end, ino1, path1);
2039         ceph_encode_filepath(&p, end, ino2, path2);
2040
2041         /* make note of release offset, in case we need to replay */
2042         req->r_request_release_offset = p - msg->front.iov_base;
2043
2044         /* cap releases */
2045         releases = 0;
2046         if (req->r_inode_drop)
2047                 releases += ceph_encode_inode_release(&p,
2048                       req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2049                       mds, req->r_inode_drop, req->r_inode_unless, 0);
2050         if (req->r_dentry_drop)
2051                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
2052                                 req->r_parent, mds, req->r_dentry_drop,
2053                                 req->r_dentry_unless);
2054         if (req->r_old_dentry_drop)
2055                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2056                                 req->r_old_dentry_dir, mds,
2057                                 req->r_old_dentry_drop,
2058                                 req->r_old_dentry_unless);
2059         if (req->r_old_inode_drop)
2060                 releases += ceph_encode_inode_release(&p,
2061                       d_inode(req->r_old_dentry),
2062                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2063
2064         if (drop_cap_releases) {
2065                 releases = 0;
2066                 p = msg->front.iov_base + req->r_request_release_offset;
2067         }
2068
2069         head->num_releases = cpu_to_le16(releases);
2070
2071         /* time stamp */
2072         {
2073                 struct ceph_timespec ts;
2074                 ceph_encode_timespec(&ts, &req->r_stamp);
2075                 ceph_encode_copy(&p, &ts, sizeof(ts));
2076         }
2077
2078         BUG_ON(p > end);
2079         msg->front.iov_len = p - msg->front.iov_base;
2080         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2081
2082         if (req->r_pagelist) {
2083                 struct ceph_pagelist *pagelist = req->r_pagelist;
2084                 refcount_inc(&pagelist->refcnt);
2085                 ceph_msg_data_add_pagelist(msg, pagelist);
2086                 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2087         } else {
2088                 msg->hdr.data_len = 0;
2089         }
2090
2091         msg->hdr.data_off = cpu_to_le16(0);
2092
2093 out_free2:
2094         if (freepath2)
2095                 kfree((char *)path2);
2096 out_free1:
2097         if (freepath1)
2098                 kfree((char *)path1);
2099 out:
2100         return msg;
2101 }
2102
2103 /*
2104  * called under mdsc->mutex if error, under no mutex if
2105  * success.
2106  */
2107 static void complete_request(struct ceph_mds_client *mdsc,
2108                              struct ceph_mds_request *req)
2109 {
2110         if (req->r_callback)
2111                 req->r_callback(mdsc, req);
2112         else
2113                 complete_all(&req->r_completion);
2114 }
2115
2116 /*
2117  * called under mdsc->mutex
2118  */
2119 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2120                                   struct ceph_mds_request *req,
2121                                   int mds, bool drop_cap_releases)
2122 {
2123         struct ceph_mds_request_head *rhead;
2124         struct ceph_msg *msg;
2125         int flags = 0;
2126
2127         req->r_attempts++;
2128         if (req->r_inode) {
2129                 struct ceph_cap *cap =
2130                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2131
2132                 if (cap)
2133                         req->r_sent_on_mseq = cap->mseq;
2134                 else
2135                         req->r_sent_on_mseq = -1;
2136         }
2137         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2138              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2139
2140         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2141                 void *p;
2142                 /*
2143                  * Replay.  Do not regenerate message (and rebuild
2144                  * paths, etc.); just use the original message.
2145                  * Rebuilding paths will break for renames because
2146                  * d_move mangles the src name.
2147                  */
2148                 msg = req->r_request;
2149                 rhead = msg->front.iov_base;
2150
2151                 flags = le32_to_cpu(rhead->flags);
2152                 flags |= CEPH_MDS_FLAG_REPLAY;
2153                 rhead->flags = cpu_to_le32(flags);
2154
2155                 if (req->r_target_inode)
2156                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2157
2158                 rhead->num_retry = req->r_attempts - 1;
2159
2160                 /* remove cap/dentry releases from message */
2161                 rhead->num_releases = 0;
2162
2163                 /* time stamp */
2164                 p = msg->front.iov_base + req->r_request_release_offset;
2165                 {
2166                         struct ceph_timespec ts;
2167                         ceph_encode_timespec(&ts, &req->r_stamp);
2168                         ceph_encode_copy(&p, &ts, sizeof(ts));
2169                 }
2170
2171                 msg->front.iov_len = p - msg->front.iov_base;
2172                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2173                 return 0;
2174         }
2175
2176         if (req->r_request) {
2177                 ceph_msg_put(req->r_request);
2178                 req->r_request = NULL;
2179         }
2180         msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2181         if (IS_ERR(msg)) {
2182                 req->r_err = PTR_ERR(msg);
2183                 return PTR_ERR(msg);
2184         }
2185         req->r_request = msg;
2186
2187         rhead = msg->front.iov_base;
2188         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2189         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2190                 flags |= CEPH_MDS_FLAG_REPLAY;
2191         if (req->r_parent)
2192                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2193         rhead->flags = cpu_to_le32(flags);
2194         rhead->num_fwd = req->r_num_fwd;
2195         rhead->num_retry = req->r_attempts - 1;
2196         rhead->ino = 0;
2197
2198         dout(" r_parent = %p\n", req->r_parent);
2199         return 0;
2200 }
2201
2202 /*
2203  * send request, or put it on the appropriate wait list.
2204  */
2205 static int __do_request(struct ceph_mds_client *mdsc,
2206                         struct ceph_mds_request *req)
2207 {
2208         struct ceph_mds_session *session = NULL;
2209         int mds = -1;
2210         int err = 0;
2211
2212         if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2213                 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2214                         __unregister_request(mdsc, req);
2215                 goto out;
2216         }
2217
2218         if (req->r_timeout &&
2219             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2220                 dout("do_request timed out\n");
2221                 err = -EIO;
2222                 goto finish;
2223         }
2224         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2225                 dout("do_request forced umount\n");
2226                 err = -EIO;
2227                 goto finish;
2228         }
2229         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2230                 if (mdsc->mdsmap_err) {
2231                         err = mdsc->mdsmap_err;
2232                         dout("do_request mdsmap err %d\n", err);
2233                         goto finish;
2234                 }
2235                 if (mdsc->mdsmap->m_epoch == 0) {
2236                         dout("do_request no mdsmap, waiting for map\n");
2237                         list_add(&req->r_wait, &mdsc->waiting_for_map);
2238                         goto finish;
2239                 }
2240                 if (!(mdsc->fsc->mount_options->flags &
2241                       CEPH_MOUNT_OPT_MOUNTWAIT) &&
2242                     !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2243                         err = -ENOENT;
2244                         pr_info("probably no mds server is up\n");
2245                         goto finish;
2246                 }
2247         }
2248
2249         put_request_session(req);
2250
2251         mds = __choose_mds(mdsc, req);
2252         if (mds < 0 ||
2253             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2254                 dout("do_request no mds or not active, waiting for map\n");
2255                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2256                 goto out;
2257         }
2258
2259         /* get, open session */
2260         session = __ceph_lookup_mds_session(mdsc, mds);
2261         if (!session) {
2262                 session = register_session(mdsc, mds);
2263                 if (IS_ERR(session)) {
2264                         err = PTR_ERR(session);
2265                         goto finish;
2266                 }
2267         }
2268         req->r_session = get_session(session);
2269
2270         dout("do_request mds%d session %p state %s\n", mds, session,
2271              ceph_session_state_name(session->s_state));
2272         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2273             session->s_state != CEPH_MDS_SESSION_HUNG) {
2274                 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2275                         err = -EACCES;
2276                         goto out_session;
2277                 }
2278                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2279                     session->s_state == CEPH_MDS_SESSION_CLOSING)
2280                         __open_session(mdsc, session);
2281                 list_add(&req->r_wait, &session->s_waiting);
2282                 goto out_session;
2283         }
2284
2285         /* send request */
2286         req->r_resend_mds = -1;   /* forget any previous mds hint */
2287
2288         if (req->r_request_started == 0)   /* note request start time */
2289                 req->r_request_started = jiffies;
2290
2291         err = __prepare_send_request(mdsc, req, mds, false);
2292         if (!err) {
2293                 ceph_msg_get(req->r_request);
2294                 ceph_con_send(&session->s_con, req->r_request);
2295         }
2296
2297 out_session:
2298         ceph_put_mds_session(session);
2299 finish:
2300         if (err) {
2301                 dout("__do_request early error %d\n", err);
2302                 req->r_err = err;
2303                 complete_request(mdsc, req);
2304                 __unregister_request(mdsc, req);
2305         }
2306 out:
2307         return err;
2308 }
2309
2310 /*
2311  * called under mdsc->mutex
2312  */
2313 static void __wake_requests(struct ceph_mds_client *mdsc,
2314                             struct list_head *head)
2315 {
2316         struct ceph_mds_request *req;
2317         LIST_HEAD(tmp_list);
2318
2319         list_splice_init(head, &tmp_list);
2320
2321         while (!list_empty(&tmp_list)) {
2322                 req = list_entry(tmp_list.next,
2323                                  struct ceph_mds_request, r_wait);
2324                 list_del_init(&req->r_wait);
2325                 dout(" wake request %p tid %llu\n", req, req->r_tid);
2326                 __do_request(mdsc, req);
2327         }
2328 }
2329
2330 /*
2331  * Wake up threads with requests pending for @mds, so that they can
2332  * resubmit their requests to a possibly different mds.
2333  */
2334 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2335 {
2336         struct ceph_mds_request *req;
2337         struct rb_node *p = rb_first(&mdsc->request_tree);
2338
2339         dout("kick_requests mds%d\n", mds);
2340         while (p) {
2341                 req = rb_entry(p, struct ceph_mds_request, r_node);
2342                 p = rb_next(p);
2343                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2344                         continue;
2345                 if (req->r_attempts > 0)
2346                         continue; /* only new requests */
2347                 if (req->r_session &&
2348                     req->r_session->s_mds == mds) {
2349                         dout(" kicking tid %llu\n", req->r_tid);
2350                         list_del_init(&req->r_wait);
2351                         __do_request(mdsc, req);
2352                 }
2353         }
2354 }
2355
2356 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2357                               struct ceph_mds_request *req)
2358 {
2359         dout("submit_request on %p\n", req);
2360         mutex_lock(&mdsc->mutex);
2361         __register_request(mdsc, req, NULL);
2362         __do_request(mdsc, req);
2363         mutex_unlock(&mdsc->mutex);
2364 }
2365
2366 /*
2367  * Synchrously perform an mds request.  Take care of all of the
2368  * session setup, forwarding, retry details.
2369  */
2370 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2371                          struct inode *dir,
2372                          struct ceph_mds_request *req)
2373 {
2374         int err;
2375
2376         dout("do_request on %p\n", req);
2377
2378         /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2379         if (req->r_inode)
2380                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2381         if (req->r_parent)
2382                 ceph_get_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
2383         if (req->r_old_dentry_dir)
2384                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2385                                   CEPH_CAP_PIN);
2386
2387         /* issue */
2388         mutex_lock(&mdsc->mutex);
2389         __register_request(mdsc, req, dir);
2390         __do_request(mdsc, req);
2391
2392         if (req->r_err) {
2393                 err = req->r_err;
2394                 goto out;
2395         }
2396
2397         /* wait */
2398         mutex_unlock(&mdsc->mutex);
2399         dout("do_request waiting\n");
2400         if (!req->r_timeout && req->r_wait_for_completion) {
2401                 err = req->r_wait_for_completion(mdsc, req);
2402         } else {
2403                 long timeleft = wait_for_completion_killable_timeout(
2404                                         &req->r_completion,
2405                                         ceph_timeout_jiffies(req->r_timeout));
2406                 if (timeleft > 0)
2407                         err = 0;
2408                 else if (!timeleft)
2409                         err = -EIO;  /* timed out */
2410                 else
2411                         err = timeleft;  /* killed */
2412         }
2413         dout("do_request waited, got %d\n", err);
2414         mutex_lock(&mdsc->mutex);
2415
2416         /* only abort if we didn't race with a real reply */
2417         if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2418                 err = le32_to_cpu(req->r_reply_info.head->result);
2419         } else if (err < 0) {
2420                 dout("aborted request %lld with %d\n", req->r_tid, err);
2421
2422                 /*
2423                  * ensure we aren't running concurrently with
2424                  * ceph_fill_trace or ceph_readdir_prepopulate, which
2425                  * rely on locks (dir mutex) held by our caller.
2426                  */
2427                 mutex_lock(&req->r_fill_mutex);
2428                 req->r_err = err;
2429                 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2430                 mutex_unlock(&req->r_fill_mutex);
2431
2432                 if (req->r_parent &&
2433                     (req->r_op & CEPH_MDS_OP_WRITE))
2434                         ceph_invalidate_dir_request(req);
2435         } else {
2436                 err = req->r_err;
2437         }
2438
2439 out:
2440         mutex_unlock(&mdsc->mutex);
2441         dout("do_request %p done, result %d\n", req, err);
2442         return err;
2443 }
2444
2445 /*
2446  * Invalidate dir's completeness, dentry lease state on an aborted MDS
2447  * namespace request.
2448  */
2449 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2450 {
2451         struct inode *dir = req->r_parent;
2452         struct inode *old_dir = req->r_old_dentry_dir;
2453
2454         dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
2455
2456         ceph_dir_clear_complete(dir);
2457         if (old_dir)
2458                 ceph_dir_clear_complete(old_dir);
2459         if (req->r_dentry)
2460                 ceph_invalidate_dentry_lease(req->r_dentry);
2461         if (req->r_old_dentry)
2462                 ceph_invalidate_dentry_lease(req->r_old_dentry);
2463 }
2464
2465 /*
2466  * Handle mds reply.
2467  *
2468  * We take the session mutex and parse and process the reply immediately.
2469  * This preserves the logical ordering of replies, capabilities, etc., sent
2470  * by the MDS as they are applied to our local cache.
2471  */
2472 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2473 {
2474         struct ceph_mds_client *mdsc = session->s_mdsc;
2475         struct ceph_mds_request *req;
2476         struct ceph_mds_reply_head *head = msg->front.iov_base;
2477         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2478         struct ceph_snap_realm *realm;
2479         u64 tid;
2480         int err, result;
2481         int mds = session->s_mds;
2482
2483         if (msg->front.iov_len < sizeof(*head)) {
2484                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2485                 ceph_msg_dump(msg);
2486                 return;
2487         }
2488
2489         /* get request, session */
2490         tid = le64_to_cpu(msg->hdr.tid);
2491         mutex_lock(&mdsc->mutex);
2492         req = lookup_get_request(mdsc, tid);
2493         if (!req) {
2494                 dout("handle_reply on unknown tid %llu\n", tid);
2495                 mutex_unlock(&mdsc->mutex);
2496                 return;
2497         }
2498         dout("handle_reply %p\n", req);
2499
2500         /* correct session? */
2501         if (req->r_session != session) {
2502                 pr_err("mdsc_handle_reply got %llu on session mds%d"
2503                        " not mds%d\n", tid, session->s_mds,
2504                        req->r_session ? req->r_session->s_mds : -1);
2505                 mutex_unlock(&mdsc->mutex);
2506                 goto out;
2507         }
2508
2509         /* dup? */
2510         if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
2511             (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
2512                 pr_warn("got a dup %s reply on %llu from mds%d\n",
2513                            head->safe ? "safe" : "unsafe", tid, mds);
2514                 mutex_unlock(&mdsc->mutex);
2515                 goto out;
2516         }
2517         if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
2518                 pr_warn("got unsafe after safe on %llu from mds%d\n",
2519                            tid, mds);
2520                 mutex_unlock(&mdsc->mutex);
2521                 goto out;
2522         }
2523
2524         result = le32_to_cpu(head->result);
2525
2526         /*
2527          * Handle an ESTALE
2528          * if we're not talking to the authority, send to them
2529          * if the authority has changed while we weren't looking,
2530          * send to new authority
2531          * Otherwise we just have to return an ESTALE
2532          */
2533         if (result == -ESTALE) {
2534                 dout("got ESTALE on request %llu\n", req->r_tid);
2535                 req->r_resend_mds = -1;
2536                 if (req->r_direct_mode != USE_AUTH_MDS) {
2537                         dout("not using auth, setting for that now\n");
2538                         req->r_direct_mode = USE_AUTH_MDS;
2539                         __do_request(mdsc, req);
2540                         mutex_unlock(&mdsc->mutex);
2541                         goto out;
2542                 } else  {
2543                         int mds = __choose_mds(mdsc, req);
2544                         if (mds >= 0 && mds != req->r_session->s_mds) {
2545                                 dout("but auth changed, so resending\n");
2546                                 __do_request(mdsc, req);
2547                                 mutex_unlock(&mdsc->mutex);
2548                                 goto out;
2549                         }
2550                 }
2551                 dout("have to return ESTALE on request %llu\n", req->r_tid);
2552         }
2553
2554
2555         if (head->safe) {
2556                 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
2557                 __unregister_request(mdsc, req);
2558
2559                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2560                         /*
2561                          * We already handled the unsafe response, now do the
2562                          * cleanup.  No need to examine the response; the MDS
2563                          * doesn't include any result info in the safe
2564                          * response.  And even if it did, there is nothing
2565                          * useful we could do with a revised return value.
2566                          */
2567                         dout("got safe reply %llu, mds%d\n", tid, mds);
2568
2569                         /* last unsafe request during umount? */
2570                         if (mdsc->stopping && !__get_oldest_req(mdsc))
2571                                 complete_all(&mdsc->safe_umount_waiters);
2572                         mutex_unlock(&mdsc->mutex);
2573                         goto out;
2574                 }
2575         } else {
2576                 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
2577                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2578                 if (req->r_unsafe_dir) {
2579                         struct ceph_inode_info *ci =
2580                                         ceph_inode(req->r_unsafe_dir);
2581                         spin_lock(&ci->i_unsafe_lock);
2582                         list_add_tail(&req->r_unsafe_dir_item,
2583                                       &ci->i_unsafe_dirops);
2584                         spin_unlock(&ci->i_unsafe_lock);
2585                 }
2586         }
2587
2588         dout("handle_reply tid %lld result %d\n", tid, result);
2589         rinfo = &req->r_reply_info;
2590         err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2591         mutex_unlock(&mdsc->mutex);
2592
2593         mutex_lock(&session->s_mutex);
2594         if (err < 0) {
2595                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2596                 ceph_msg_dump(msg);
2597                 goto out_err;
2598         }
2599
2600         /* snap trace */
2601         realm = NULL;
2602         if (rinfo->snapblob_len) {
2603                 down_write(&mdsc->snap_rwsem);
2604                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2605                                 rinfo->snapblob + rinfo->snapblob_len,
2606                                 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2607                                 &realm);
2608                 downgrade_write(&mdsc->snap_rwsem);
2609         } else {
2610                 down_read(&mdsc->snap_rwsem);
2611         }
2612
2613         /* insert trace into our cache */
2614         mutex_lock(&req->r_fill_mutex);
2615         current->journal_info = req;
2616         err = ceph_fill_trace(mdsc->fsc->sb, req);
2617         if (err == 0) {
2618                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2619                                     req->r_op == CEPH_MDS_OP_LSSNAP))
2620                         ceph_readdir_prepopulate(req, req->r_session);
2621                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2622         }
2623         current->journal_info = NULL;
2624         mutex_unlock(&req->r_fill_mutex);
2625
2626         up_read(&mdsc->snap_rwsem);
2627         if (realm)
2628                 ceph_put_snap_realm(mdsc, realm);
2629
2630         if (err == 0 && req->r_target_inode &&
2631             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2632                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
2633                 spin_lock(&ci->i_unsafe_lock);
2634                 list_add_tail(&req->r_unsafe_target_item, &ci->i_unsafe_iops);
2635                 spin_unlock(&ci->i_unsafe_lock);
2636         }
2637 out_err:
2638         mutex_lock(&mdsc->mutex);
2639         if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2640                 if (err) {
2641                         req->r_err = err;
2642                 } else {
2643                         req->r_reply =  ceph_msg_get(msg);
2644                         set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
2645                 }
2646         } else {
2647                 dout("reply arrived after request %lld was aborted\n", tid);
2648         }
2649         mutex_unlock(&mdsc->mutex);
2650
2651         mutex_unlock(&session->s_mutex);
2652
2653         /* kick calling process */
2654         complete_request(mdsc, req);
2655 out:
2656         ceph_mdsc_put_request(req);
2657         return;
2658 }
2659
2660
2661
2662 /*
2663  * handle mds notification that our request has been forwarded.
2664  */
2665 static void handle_forward(struct ceph_mds_client *mdsc,
2666                            struct ceph_mds_session *session,
2667                            struct ceph_msg *msg)
2668 {
2669         struct ceph_mds_request *req;
2670         u64 tid = le64_to_cpu(msg->hdr.tid);
2671         u32 next_mds;
2672         u32 fwd_seq;
2673         int err = -EINVAL;
2674         void *p = msg->front.iov_base;
2675         void *end = p + msg->front.iov_len;
2676
2677         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2678         next_mds = ceph_decode_32(&p);
2679         fwd_seq = ceph_decode_32(&p);
2680
2681         mutex_lock(&mdsc->mutex);
2682         req = lookup_get_request(mdsc, tid);
2683         if (!req) {
2684                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2685                 goto out;  /* dup reply? */
2686         }
2687
2688         if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2689                 dout("forward tid %llu aborted, unregistering\n", tid);
2690                 __unregister_request(mdsc, req);
2691         } else if (fwd_seq <= req->r_num_fwd) {
2692                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2693                      tid, next_mds, req->r_num_fwd, fwd_seq);
2694         } else {
2695                 /* resend. forward race not possible; mds would drop */
2696                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2697                 BUG_ON(req->r_err);
2698                 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
2699                 req->r_attempts = 0;
2700                 req->r_num_fwd = fwd_seq;
2701                 req->r_resend_mds = next_mds;
2702                 put_request_session(req);
2703                 __do_request(mdsc, req);
2704         }
2705         ceph_mdsc_put_request(req);
2706 out:
2707         mutex_unlock(&mdsc->mutex);
2708         return;
2709
2710 bad:
2711         pr_err("mdsc_handle_forward decode error err=%d\n", err);
2712 }
2713
2714 /*
2715  * handle a mds session control message
2716  */
2717 static void handle_session(struct ceph_mds_session *session,
2718                            struct ceph_msg *msg)
2719 {
2720         struct ceph_mds_client *mdsc = session->s_mdsc;
2721         u32 op;
2722         u64 seq;
2723         int mds = session->s_mds;
2724         struct ceph_mds_session_head *h = msg->front.iov_base;
2725         int wake = 0;
2726
2727         /* decode */
2728         if (msg->front.iov_len != sizeof(*h))
2729                 goto bad;
2730         op = le32_to_cpu(h->op);
2731         seq = le64_to_cpu(h->seq);
2732
2733         mutex_lock(&mdsc->mutex);
2734         if (op == CEPH_SESSION_CLOSE) {
2735                 get_session(session);
2736                 __unregister_session(mdsc, session);
2737         }
2738         /* FIXME: this ttl calculation is generous */
2739         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2740         mutex_unlock(&mdsc->mutex);
2741
2742         mutex_lock(&session->s_mutex);
2743
2744         dout("handle_session mds%d %s %p state %s seq %llu\n",
2745              mds, ceph_session_op_name(op), session,
2746              ceph_session_state_name(session->s_state), seq);
2747
2748         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2749                 session->s_state = CEPH_MDS_SESSION_OPEN;
2750                 pr_info("mds%d came back\n", session->s_mds);
2751         }
2752
2753         switch (op) {
2754         case CEPH_SESSION_OPEN:
2755                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2756                         pr_info("mds%d reconnect success\n", session->s_mds);
2757                 session->s_state = CEPH_MDS_SESSION_OPEN;
2758                 renewed_caps(mdsc, session, 0);
2759                 wake = 1;
2760                 if (mdsc->stopping)
2761                         __close_session(mdsc, session);
2762                 break;
2763
2764         case CEPH_SESSION_RENEWCAPS:
2765                 if (session->s_renew_seq == seq)
2766                         renewed_caps(mdsc, session, 1);
2767                 break;
2768
2769         case CEPH_SESSION_CLOSE:
2770                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2771                         pr_info("mds%d reconnect denied\n", session->s_mds);
2772                 cleanup_session_requests(mdsc, session);
2773                 remove_session_caps(session);
2774                 wake = 2; /* for good measure */
2775                 wake_up_all(&mdsc->session_close_wq);
2776                 break;
2777
2778         case CEPH_SESSION_STALE:
2779                 pr_info("mds%d caps went stale, renewing\n",
2780                         session->s_mds);
2781                 spin_lock(&session->s_gen_ttl_lock);
2782                 session->s_cap_gen++;
2783                 session->s_cap_ttl = jiffies - 1;
2784                 spin_unlock(&session->s_gen_ttl_lock);
2785                 send_renew_caps(mdsc, session);
2786                 break;
2787
2788         case CEPH_SESSION_RECALL_STATE:
2789                 ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2790                 break;
2791
2792         case CEPH_SESSION_FLUSHMSG:
2793                 send_flushmsg_ack(mdsc, session, seq);
2794                 break;
2795
2796         case CEPH_SESSION_FORCE_RO:
2797                 dout("force_session_readonly %p\n", session);
2798                 spin_lock(&session->s_cap_lock);
2799                 session->s_readonly = true;
2800                 spin_unlock(&session->s_cap_lock);
2801                 wake_up_session_caps(session, 0);
2802                 break;
2803
2804         case CEPH_SESSION_REJECT:
2805                 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
2806                 pr_info("mds%d rejected session\n", session->s_mds);
2807                 session->s_state = CEPH_MDS_SESSION_REJECTED;
2808                 cleanup_session_requests(mdsc, session);
2809                 remove_session_caps(session);
2810                 wake = 2; /* for good measure */
2811                 break;
2812
2813         default:
2814                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2815                 WARN_ON(1);
2816         }
2817
2818         mutex_unlock(&session->s_mutex);
2819         if (wake) {
2820                 mutex_lock(&mdsc->mutex);
2821                 __wake_requests(mdsc, &session->s_waiting);
2822                 if (wake == 2)
2823                         kick_requests(mdsc, mds);
2824                 mutex_unlock(&mdsc->mutex);
2825         }
2826         if (op == CEPH_SESSION_CLOSE)
2827                 ceph_put_mds_session(session);
2828         return;
2829
2830 bad:
2831         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2832                (int)msg->front.iov_len);
2833         ceph_msg_dump(msg);
2834         return;
2835 }
2836
2837
2838 /*
2839  * called under session->mutex.
2840  */
2841 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2842                                    struct ceph_mds_session *session)
2843 {
2844         struct ceph_mds_request *req, *nreq;
2845         struct rb_node *p;
2846         int err;
2847
2848         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2849
2850         mutex_lock(&mdsc->mutex);
2851         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2852                 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2853                 if (!err) {
2854                         ceph_msg_get(req->r_request);
2855                         ceph_con_send(&session->s_con, req->r_request);
2856                 }
2857         }
2858
2859         /*
2860          * also re-send old requests when MDS enters reconnect stage. So that MDS
2861          * can process completed request in clientreplay stage.
2862          */
2863         p = rb_first(&mdsc->request_tree);
2864         while (p) {
2865                 req = rb_entry(p, struct ceph_mds_request, r_node);
2866                 p = rb_next(p);
2867                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2868                         continue;
2869                 if (req->r_attempts == 0)
2870                         continue; /* only old requests */
2871                 if (req->r_session &&
2872                     req->r_session->s_mds == session->s_mds) {
2873                         err = __prepare_send_request(mdsc, req,
2874                                                      session->s_mds, true);
2875                         if (!err) {
2876                                 ceph_msg_get(req->r_request);
2877                                 ceph_con_send(&session->s_con, req->r_request);
2878                         }
2879                 }
2880         }
2881         mutex_unlock(&mdsc->mutex);
2882 }
2883
2884 /*
2885  * Encode information about a cap for a reconnect with the MDS.
2886  */
2887 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2888                           void *arg)
2889 {
2890         union {
2891                 struct ceph_mds_cap_reconnect v2;
2892                 struct ceph_mds_cap_reconnect_v1 v1;
2893         } rec;
2894         struct ceph_inode_info *ci = cap->ci;
2895         struct ceph_reconnect_state *recon_state = arg;
2896         struct ceph_pagelist *pagelist = recon_state->pagelist;
2897         char *path;
2898         int pathlen, err;
2899         u64 pathbase;
2900         u64 snap_follows;
2901         struct dentry *dentry;
2902
2903         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2904              inode, ceph_vinop(inode), cap, cap->cap_id,
2905              ceph_cap_string(cap->issued));
2906         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2907         if (err)
2908                 return err;
2909
2910         dentry = d_find_alias(inode);
2911         if (dentry) {
2912                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2913                 if (IS_ERR(path)) {
2914                         err = PTR_ERR(path);
2915                         goto out_dput;
2916                 }
2917         } else {
2918                 path = NULL;
2919                 pathlen = 0;
2920                 pathbase = 0;
2921         }
2922
2923         spin_lock(&ci->i_ceph_lock);
2924         cap->seq = 0;        /* reset cap seq */
2925         cap->issue_seq = 0;  /* and issue_seq */
2926         cap->mseq = 0;       /* and migrate_seq */
2927         cap->cap_gen = cap->session->s_cap_gen;
2928
2929         if (recon_state->msg_version >= 2) {
2930                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2931                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2932                 rec.v2.issued = cpu_to_le32(cap->issued);
2933                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2934                 rec.v2.pathbase = cpu_to_le64(pathbase);
2935                 rec.v2.flock_len = (__force __le32)
2936                         ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
2937         } else {
2938                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2939                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2940                 rec.v1.issued = cpu_to_le32(cap->issued);
2941                 rec.v1.size = cpu_to_le64(inode->i_size);
2942                 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2943                 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2944                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2945                 rec.v1.pathbase = cpu_to_le64(pathbase);
2946         }
2947
2948         if (list_empty(&ci->i_cap_snaps)) {
2949                 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
2950         } else {
2951                 struct ceph_cap_snap *capsnap =
2952                         list_first_entry(&ci->i_cap_snaps,
2953                                          struct ceph_cap_snap, ci_item);
2954                 snap_follows = capsnap->follows;
2955         }
2956         spin_unlock(&ci->i_ceph_lock);
2957
2958         if (recon_state->msg_version >= 2) {
2959                 int num_fcntl_locks, num_flock_locks;
2960                 struct ceph_filelock *flocks = NULL;
2961                 size_t struct_len, total_len = 0;
2962                 u8 struct_v = 0;
2963
2964 encode_again:
2965                 if (rec.v2.flock_len) {
2966                         ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2967                 } else {
2968                         num_fcntl_locks = 0;
2969                         num_flock_locks = 0;
2970                 }
2971                 if (num_fcntl_locks + num_flock_locks > 0) {
2972                         flocks = kmalloc((num_fcntl_locks + num_flock_locks) *
2973                                          sizeof(struct ceph_filelock), GFP_NOFS);
2974                         if (!flocks) {
2975                                 err = -ENOMEM;
2976                                 goto out_free;
2977                         }
2978                         err = ceph_encode_locks_to_buffer(inode, flocks,
2979                                                           num_fcntl_locks,
2980                                                           num_flock_locks);
2981                         if (err) {
2982                                 kfree(flocks);
2983                                 flocks = NULL;
2984                                 if (err == -ENOSPC)
2985                                         goto encode_again;
2986                                 goto out_free;
2987                         }
2988                 } else {
2989                         kfree(flocks);
2990                         flocks = NULL;
2991                 }
2992
2993                 if (recon_state->msg_version >= 3) {
2994                         /* version, compat_version and struct_len */
2995                         total_len = 2 * sizeof(u8) + sizeof(u32);
2996                         struct_v = 2;
2997                 }
2998                 /*
2999                  * number of encoded locks is stable, so copy to pagelist
3000                  */
3001                 struct_len = 2 * sizeof(u32) +
3002                             (num_fcntl_locks + num_flock_locks) *
3003                             sizeof(struct ceph_filelock);
3004                 rec.v2.flock_len = cpu_to_le32(struct_len);
3005
3006                 struct_len += sizeof(rec.v2);
3007                 struct_len += sizeof(u32) + pathlen;
3008
3009                 if (struct_v >= 2)
3010                         struct_len += sizeof(u64); /* snap_follows */
3011
3012                 total_len += struct_len;
3013                 err = ceph_pagelist_reserve(pagelist, total_len);
3014
3015                 if (!err) {
3016                         if (recon_state->msg_version >= 3) {
3017                                 ceph_pagelist_encode_8(pagelist, struct_v);
3018                                 ceph_pagelist_encode_8(pagelist, 1);
3019                                 ceph_pagelist_encode_32(pagelist, struct_len);
3020                         }
3021                         ceph_pagelist_encode_string(pagelist, path, pathlen);
3022                         ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
3023                         ceph_locks_to_pagelist(flocks, pagelist,
3024                                                num_fcntl_locks,
3025                                                num_flock_locks);
3026                         if (struct_v >= 2)
3027                                 ceph_pagelist_encode_64(pagelist, snap_follows);
3028                 }
3029                 kfree(flocks);
3030         } else {
3031                 size_t size = sizeof(u32) + pathlen + sizeof(rec.v1);
3032                 err = ceph_pagelist_reserve(pagelist, size);
3033                 if (!err) {
3034                         ceph_pagelist_encode_string(pagelist, path, pathlen);
3035                         ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
3036                 }
3037         }
3038
3039         recon_state->nr_caps++;
3040 out_free:
3041         kfree(path);
3042 out_dput:
3043         dput(dentry);
3044         return err;
3045 }
3046
3047
3048 /*
3049  * If an MDS fails and recovers, clients need to reconnect in order to
3050  * reestablish shared state.  This includes all caps issued through
3051  * this session _and_ the snap_realm hierarchy.  Because it's not
3052  * clear which snap realms the mds cares about, we send everything we
3053  * know about.. that ensures we'll then get any new info the
3054  * recovering MDS might have.
3055  *
3056  * This is a relatively heavyweight operation, but it's rare.
3057  *
3058  * called with mdsc->mutex held.
3059  */
3060 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3061                                struct ceph_mds_session *session)
3062 {
3063         struct ceph_msg *reply;
3064         struct rb_node *p;
3065         int mds = session->s_mds;
3066         int err = -ENOMEM;
3067         int s_nr_caps;
3068         struct ceph_pagelist *pagelist;
3069         struct ceph_reconnect_state recon_state;
3070         LIST_HEAD(dispose);
3071
3072         pr_info("mds%d reconnect start\n", mds);
3073
3074         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
3075         if (!pagelist)
3076                 goto fail_nopagelist;
3077         ceph_pagelist_init(pagelist);
3078
3079         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
3080         if (!reply)
3081                 goto fail_nomsg;
3082
3083         mutex_lock(&session->s_mutex);
3084         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3085         session->s_seq = 0;
3086
3087         dout("session %p state %s\n", session,
3088              ceph_session_state_name(session->s_state));
3089
3090         spin_lock(&session->s_gen_ttl_lock);
3091         session->s_cap_gen++;
3092         spin_unlock(&session->s_gen_ttl_lock);
3093
3094         spin_lock(&session->s_cap_lock);
3095         /* don't know if session is readonly */
3096         session->s_readonly = 0;
3097         /*
3098          * notify __ceph_remove_cap() that we are composing cap reconnect.
3099          * If a cap get released before being added to the cap reconnect,
3100          * __ceph_remove_cap() should skip queuing cap release.
3101          */
3102         session->s_cap_reconnect = 1;
3103         /* drop old cap expires; we're about to reestablish that state */
3104         detach_cap_releases(session, &dispose);
3105         spin_unlock(&session->s_cap_lock);
3106         dispose_cap_releases(mdsc, &dispose);
3107
3108         /* trim unused caps to reduce MDS's cache rejoin time */
3109         if (mdsc->fsc->sb->s_root)
3110                 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3111
3112         ceph_con_close(&session->s_con);
3113         ceph_con_open(&session->s_con,
3114                       CEPH_ENTITY_TYPE_MDS, mds,
3115                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3116
3117         /* replay unsafe requests */
3118         replay_unsafe_requests(mdsc, session);
3119
3120         down_read(&mdsc->snap_rwsem);
3121
3122         /* traverse this session's caps */
3123         s_nr_caps = session->s_nr_caps;
3124         err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
3125         if (err)
3126                 goto fail;
3127
3128         recon_state.nr_caps = 0;
3129         recon_state.pagelist = pagelist;
3130         if (session->s_con.peer_features & CEPH_FEATURE_MDSENC)
3131                 recon_state.msg_version = 3;
3132         else if (session->s_con.peer_features & CEPH_FEATURE_FLOCK)
3133                 recon_state.msg_version = 2;
3134         else
3135                 recon_state.msg_version = 1;
3136         err = iterate_session_caps(session, encode_caps_cb, &recon_state);
3137         if (err < 0)
3138                 goto fail;
3139
3140         spin_lock(&session->s_cap_lock);
3141         session->s_cap_reconnect = 0;
3142         spin_unlock(&session->s_cap_lock);
3143
3144         /*
3145          * snaprealms.  we provide mds with the ino, seq (version), and
3146          * parent for all of our realms.  If the mds has any newer info,
3147          * it will tell us.
3148          */
3149         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3150                 struct ceph_snap_realm *realm =
3151                         rb_entry(p, struct ceph_snap_realm, node);
3152                 struct ceph_mds_snaprealm_reconnect sr_rec;
3153
3154                 dout(" adding snap realm %llx seq %lld parent %llx\n",
3155                      realm->ino, realm->seq, realm->parent_ino);
3156                 sr_rec.ino = cpu_to_le64(realm->ino);
3157                 sr_rec.seq = cpu_to_le64(realm->seq);
3158                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3159                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3160                 if (err)
3161                         goto fail;
3162         }
3163
3164         reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3165
3166         /* raced with cap release? */
3167         if (s_nr_caps != recon_state.nr_caps) {
3168                 struct page *page = list_first_entry(&pagelist->head,
3169                                                      struct page, lru);
3170                 __le32 *addr = kmap_atomic(page);
3171                 *addr = cpu_to_le32(recon_state.nr_caps);
3172                 kunmap_atomic(addr);
3173         }
3174
3175         reply->hdr.data_len = cpu_to_le32(pagelist->length);
3176         ceph_msg_data_add_pagelist(reply, pagelist);
3177
3178         ceph_early_kick_flushing_caps(mdsc, session);
3179
3180         ceph_con_send(&session->s_con, reply);
3181
3182         mutex_unlock(&session->s_mutex);
3183
3184         mutex_lock(&mdsc->mutex);
3185         __wake_requests(mdsc, &session->s_waiting);
3186         mutex_unlock(&mdsc->mutex);
3187
3188         up_read(&mdsc->snap_rwsem);
3189         return;
3190
3191 fail:
3192         ceph_msg_put(reply);
3193         up_read(&mdsc->snap_rwsem);
3194         mutex_unlock(&session->s_mutex);
3195 fail_nomsg:
3196         ceph_pagelist_release(pagelist);
3197 fail_nopagelist:
3198         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3199         return;
3200 }
3201
3202
3203 /*
3204  * compare old and new mdsmaps, kicking requests
3205  * and closing out old connections as necessary
3206  *
3207  * called under mdsc->mutex.
3208  */
3209 static void check_new_map(struct ceph_mds_client *mdsc,
3210                           struct ceph_mdsmap *newmap,
3211                           struct ceph_mdsmap *oldmap)
3212 {
3213         int i;
3214         int oldstate, newstate;
3215         struct ceph_mds_session *s;
3216
3217         dout("check_new_map new %u old %u\n",
3218              newmap->m_epoch, oldmap->m_epoch);
3219
3220         for (i = 0; i < oldmap->m_num_mds && i < mdsc->max_sessions; i++) {
3221                 if (!mdsc->sessions[i])
3222                         continue;
3223                 s = mdsc->sessions[i];
3224                 oldstate = ceph_mdsmap_get_state(oldmap, i);
3225                 newstate = ceph_mdsmap_get_state(newmap, i);
3226
3227                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3228                      i, ceph_mds_state_name(oldstate),
3229                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3230                      ceph_mds_state_name(newstate),
3231                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3232                      ceph_session_state_name(s->s_state));
3233
3234                 if (i >= newmap->m_num_mds ||
3235                     memcmp(ceph_mdsmap_get_addr(oldmap, i),
3236                            ceph_mdsmap_get_addr(newmap, i),
3237                            sizeof(struct ceph_entity_addr))) {
3238                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3239                                 /* the session never opened, just close it
3240                                  * out now */
3241                                 get_session(s);
3242                                 __unregister_session(mdsc, s);
3243                                 __wake_requests(mdsc, &s->s_waiting);
3244                                 ceph_put_mds_session(s);
3245                         } else if (i >= newmap->m_num_mds) {
3246                                 /* force close session for stopped mds */
3247                                 get_session(s);
3248                                 __unregister_session(mdsc, s);
3249                                 __wake_requests(mdsc, &s->s_waiting);
3250                                 kick_requests(mdsc, i);
3251                                 mutex_unlock(&mdsc->mutex);
3252
3253                                 mutex_lock(&s->s_mutex);
3254                                 cleanup_session_requests(mdsc, s);
3255                                 remove_session_caps(s);
3256                                 mutex_unlock(&s->s_mutex);
3257
3258                                 ceph_put_mds_session(s);
3259
3260                                 mutex_lock(&mdsc->mutex);
3261                         } else {
3262                                 /* just close it */
3263                                 mutex_unlock(&mdsc->mutex);
3264                                 mutex_lock(&s->s_mutex);
3265                                 mutex_lock(&mdsc->mutex);
3266                                 ceph_con_close(&s->s_con);
3267                                 mutex_unlock(&s->s_mutex);
3268                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3269                         }
3270                 } else if (oldstate == newstate) {
3271                         continue;  /* nothing new with this mds */
3272                 }
3273
3274                 /*
3275                  * send reconnect?
3276                  */
3277                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3278                     newstate >= CEPH_MDS_STATE_RECONNECT) {
3279                         mutex_unlock(&mdsc->mutex);
3280                         send_mds_reconnect(mdsc, s);
3281                         mutex_lock(&mdsc->mutex);
3282                 }
3283
3284                 /*
3285                  * kick request on any mds that has gone active.
3286                  */
3287                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3288                     newstate >= CEPH_MDS_STATE_ACTIVE) {
3289                         if (oldstate != CEPH_MDS_STATE_CREATING &&
3290                             oldstate != CEPH_MDS_STATE_STARTING)
3291                                 pr_info("mds%d recovery completed\n", s->s_mds);
3292                         kick_requests(mdsc, i);
3293                         ceph_kick_flushing_caps(mdsc, s);
3294                         wake_up_session_caps(s, 1);
3295                 }
3296         }
3297
3298         for (i = 0; i < newmap->m_num_mds && i < mdsc->max_sessions; i++) {
3299                 s = mdsc->sessions[i];
3300                 if (!s)
3301                         continue;
3302                 if (!ceph_mdsmap_is_laggy(newmap, i))
3303                         continue;
3304                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3305                     s->s_state == CEPH_MDS_SESSION_HUNG ||
3306                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
3307                         dout(" connecting to export targets of laggy mds%d\n",
3308                              i);
3309                         __open_export_target_sessions(mdsc, s);
3310                 }
3311         }
3312 }
3313
3314
3315
3316 /*
3317  * leases
3318  */
3319
3320 /*
3321  * caller must hold session s_mutex, dentry->d_lock
3322  */
3323 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3324 {
3325         struct ceph_dentry_info *di = ceph_dentry(dentry);
3326
3327         ceph_put_mds_session(di->lease_session);
3328         di->lease_session = NULL;
3329 }
3330
3331 static void handle_lease(struct ceph_mds_client *mdsc,
3332                          struct ceph_mds_session *session,
3333                          struct ceph_msg *msg)
3334 {
3335         struct super_block *sb = mdsc->fsc->sb;
3336         struct inode *inode;
3337         struct dentry *parent, *dentry;
3338         struct ceph_dentry_info *di;
3339         int mds = session->s_mds;
3340         struct ceph_mds_lease *h = msg->front.iov_base;
3341         u32 seq;
3342         struct ceph_vino vino;
3343         struct qstr dname;
3344         int release = 0;
3345
3346         dout("handle_lease from mds%d\n", mds);
3347
3348         /* decode */
3349         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3350                 goto bad;
3351         vino.ino = le64_to_cpu(h->ino);
3352         vino.snap = CEPH_NOSNAP;
3353         seq = le32_to_cpu(h->seq);
3354         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3355         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3356         if (dname.len != get_unaligned_le32(h+1))
3357                 goto bad;
3358
3359         /* lookup inode */
3360         inode = ceph_find_inode(sb, vino);
3361         dout("handle_lease %s, ino %llx %p %.*s\n",
3362              ceph_lease_op_name(h->action), vino.ino, inode,
3363              dname.len, dname.name);
3364
3365         mutex_lock(&session->s_mutex);
3366         session->s_seq++;
3367
3368         if (!inode) {
3369                 dout("handle_lease no inode %llx\n", vino.ino);
3370                 goto release;
3371         }
3372
3373         /* dentry */
3374         parent = d_find_alias(inode);
3375         if (!parent) {
3376                 dout("no parent dentry on inode %p\n", inode);
3377                 WARN_ON(1);
3378                 goto release;  /* hrm... */
3379         }
3380         dname.hash = full_name_hash(parent, dname.name, dname.len);
3381         dentry = d_lookup(parent, &dname);
3382         dput(parent);
3383         if (!dentry)
3384                 goto release;
3385
3386         spin_lock(&dentry->d_lock);
3387         di = ceph_dentry(dentry);
3388         switch (h->action) {
3389         case CEPH_MDS_LEASE_REVOKE:
3390                 if (di->lease_session == session) {
3391                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3392                                 h->seq = cpu_to_le32(di->lease_seq);
3393                         __ceph_mdsc_drop_dentry_lease(dentry);
3394                 }
3395                 release = 1;
3396                 break;
3397
3398         case CEPH_MDS_LEASE_RENEW:
3399                 if (di->lease_session == session &&
3400                     di->lease_gen == session->s_cap_gen &&
3401                     di->lease_renew_from &&
3402                     di->lease_renew_after == 0) {
3403                         unsigned long duration =
3404                                 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3405
3406                         di->lease_seq = seq;
3407                         di->time = di->lease_renew_from + duration;
3408                         di->lease_renew_after = di->lease_renew_from +
3409                                 (duration >> 1);
3410                         di->lease_renew_from = 0;
3411                 }
3412                 break;
3413         }
3414         spin_unlock(&dentry->d_lock);
3415         dput(dentry);
3416
3417         if (!release)
3418                 goto out;
3419
3420 release:
3421         /* let's just reuse the same message */
3422         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3423         ceph_msg_get(msg);
3424         ceph_con_send(&session->s_con, msg);
3425
3426 out:
3427         iput(inode);
3428         mutex_unlock(&session->s_mutex);
3429         return;
3430
3431 bad:
3432         pr_err("corrupt lease message\n");
3433         ceph_msg_dump(msg);
3434 }
3435
3436 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3437                               struct inode *inode,
3438                               struct dentry *dentry, char action,
3439                               u32 seq)
3440 {
3441         struct ceph_msg *msg;
3442         struct ceph_mds_lease *lease;
3443         int len = sizeof(*lease) + sizeof(u32);
3444         int dnamelen = 0;
3445
3446         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3447              inode, dentry, ceph_lease_op_name(action), session->s_mds);
3448         dnamelen = dentry->d_name.len;
3449         len += dnamelen;
3450
3451         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3452         if (!msg)
3453                 return;
3454         lease = msg->front.iov_base;
3455         lease->action = action;
3456         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3457         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3458         lease->seq = cpu_to_le32(seq);
3459         put_unaligned_le32(dnamelen, lease + 1);
3460         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3461
3462         /*
3463          * if this is a preemptive lease RELEASE, no need to
3464          * flush request stream, since the actual request will
3465          * soon follow.
3466          */
3467         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3468
3469         ceph_con_send(&session->s_con, msg);
3470 }
3471
3472 /*
3473  * drop all leases (and dentry refs) in preparation for umount
3474  */
3475 static void drop_leases(struct ceph_mds_client *mdsc)
3476 {
3477         int i;
3478
3479         dout("drop_leases\n");
3480         mutex_lock(&mdsc->mutex);
3481         for (i = 0; i < mdsc->max_sessions; i++) {
3482                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3483                 if (!s)
3484                         continue;
3485                 mutex_unlock(&mdsc->mutex);
3486                 mutex_lock(&s->s_mutex);
3487                 mutex_unlock(&s->s_mutex);
3488                 ceph_put_mds_session(s);
3489                 mutex_lock(&mdsc->mutex);
3490         }
3491         mutex_unlock(&mdsc->mutex);
3492 }
3493
3494
3495
3496 /*
3497  * delayed work -- periodically trim expired leases, renew caps with mds
3498  */
3499 static void schedule_delayed(struct ceph_mds_client *mdsc)
3500 {
3501         int delay = 5;
3502         unsigned hz = round_jiffies_relative(HZ * delay);
3503         schedule_delayed_work(&mdsc->delayed_work, hz);
3504 }
3505
3506 static void delayed_work(struct work_struct *work)
3507 {
3508         int i;
3509         struct ceph_mds_client *mdsc =
3510                 container_of(work, struct ceph_mds_client, delayed_work.work);
3511         int renew_interval;
3512         int renew_caps;
3513
3514         dout("mdsc delayed_work\n");
3515         ceph_check_delayed_caps(mdsc);
3516
3517         mutex_lock(&mdsc->mutex);
3518         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3519         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3520                                    mdsc->last_renew_caps);
3521         if (renew_caps)
3522                 mdsc->last_renew_caps = jiffies;
3523
3524         for (i = 0; i < mdsc->max_sessions; i++) {
3525                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3526                 if (!s)
3527                         continue;
3528                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3529                         dout("resending session close request for mds%d\n",
3530                              s->s_mds);
3531                         request_close_session(mdsc, s);
3532                         ceph_put_mds_session(s);
3533                         continue;
3534                 }
3535                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3536                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3537                                 s->s_state = CEPH_MDS_SESSION_HUNG;
3538                                 pr_info("mds%d hung\n", s->s_mds);
3539                         }
3540                 }
3541                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3542                         /* this mds is failed or recovering, just wait */
3543                         ceph_put_mds_session(s);
3544                         continue;
3545                 }
3546                 mutex_unlock(&mdsc->mutex);
3547
3548                 mutex_lock(&s->s_mutex);
3549                 if (renew_caps)
3550                         send_renew_caps(mdsc, s);
3551                 else
3552                         ceph_con_keepalive(&s->s_con);
3553                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3554                     s->s_state == CEPH_MDS_SESSION_HUNG)
3555                         ceph_send_cap_releases(mdsc, s);
3556                 mutex_unlock(&s->s_mutex);
3557                 ceph_put_mds_session(s);
3558
3559                 mutex_lock(&mdsc->mutex);
3560         }
3561         mutex_unlock(&mdsc->mutex);
3562
3563         schedule_delayed(mdsc);
3564 }
3565
3566 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3567
3568 {
3569         struct ceph_mds_client *mdsc;
3570
3571         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3572         if (!mdsc)
3573                 return -ENOMEM;
3574         mdsc->fsc = fsc;
3575         fsc->mdsc = mdsc;
3576         mutex_init(&mdsc->mutex);
3577         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3578         if (!mdsc->mdsmap) {
3579                 kfree(mdsc);
3580                 return -ENOMEM;
3581         }
3582
3583         init_completion(&mdsc->safe_umount_waiters);
3584         init_waitqueue_head(&mdsc->session_close_wq);
3585         INIT_LIST_HEAD(&mdsc->waiting_for_map);
3586         mdsc->sessions = NULL;
3587         atomic_set(&mdsc->num_sessions, 0);
3588         mdsc->max_sessions = 0;
3589         mdsc->stopping = 0;
3590         mdsc->last_snap_seq = 0;
3591         init_rwsem(&mdsc->snap_rwsem);
3592         mdsc->snap_realms = RB_ROOT;
3593         INIT_LIST_HEAD(&mdsc->snap_empty);
3594         spin_lock_init(&mdsc->snap_empty_lock);
3595         mdsc->last_tid = 0;
3596         mdsc->oldest_tid = 0;
3597         mdsc->request_tree = RB_ROOT;
3598         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3599         mdsc->last_renew_caps = jiffies;
3600         INIT_LIST_HEAD(&mdsc->cap_delay_list);
3601         spin_lock_init(&mdsc->cap_delay_lock);
3602         INIT_LIST_HEAD(&mdsc->snap_flush_list);
3603         spin_lock_init(&mdsc->snap_flush_lock);
3604         mdsc->last_cap_flush_tid = 1;
3605         INIT_LIST_HEAD(&mdsc->cap_flush_list);
3606         INIT_LIST_HEAD(&mdsc->cap_dirty);
3607         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3608         mdsc->num_cap_flushing = 0;
3609         spin_lock_init(&mdsc->cap_dirty_lock);
3610         init_waitqueue_head(&mdsc->cap_flushing_wq);
3611         spin_lock_init(&mdsc->dentry_lru_lock);
3612         INIT_LIST_HEAD(&mdsc->dentry_lru);
3613
3614         ceph_caps_init(mdsc);
3615         ceph_adjust_min_caps(mdsc, fsc->min_caps);
3616
3617         init_rwsem(&mdsc->pool_perm_rwsem);
3618         mdsc->pool_perm_tree = RB_ROOT;
3619
3620         strncpy(mdsc->nodename, utsname()->nodename,
3621                 sizeof(mdsc->nodename) - 1);
3622         return 0;
3623 }
3624
3625 /*
3626  * Wait for safe replies on open mds requests.  If we time out, drop
3627  * all requests from the tree to avoid dangling dentry refs.
3628  */
3629 static void wait_requests(struct ceph_mds_client *mdsc)
3630 {
3631         struct ceph_options *opts = mdsc->fsc->client->options;
3632         struct ceph_mds_request *req;
3633
3634         mutex_lock(&mdsc->mutex);
3635         if (__get_oldest_req(mdsc)) {
3636                 mutex_unlock(&mdsc->mutex);
3637
3638                 dout("wait_requests waiting for requests\n");
3639                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3640                                     ceph_timeout_jiffies(opts->mount_timeout));
3641
3642                 /* tear down remaining requests */
3643                 mutex_lock(&mdsc->mutex);
3644                 while ((req = __get_oldest_req(mdsc))) {
3645                         dout("wait_requests timed out on tid %llu\n",
3646                              req->r_tid);
3647                         __unregister_request(mdsc, req);
3648                 }
3649         }
3650         mutex_unlock(&mdsc->mutex);
3651         dout("wait_requests done\n");
3652 }
3653
3654 /*
3655  * called before mount is ro, and before dentries are torn down.
3656  * (hmm, does this still race with new lookups?)
3657  */
3658 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3659 {
3660         dout("pre_umount\n");
3661         mdsc->stopping = 1;
3662
3663         drop_leases(mdsc);
3664         ceph_flush_dirty_caps(mdsc);
3665         wait_requests(mdsc);
3666
3667         /*
3668          * wait for reply handlers to drop their request refs and
3669          * their inode/dcache refs
3670          */
3671         ceph_msgr_flush();
3672 }
3673
3674 /*
3675  * wait for all write mds requests to flush.
3676  */
3677 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3678 {
3679         struct ceph_mds_request *req = NULL, *nextreq;
3680         struct rb_node *n;
3681
3682         mutex_lock(&mdsc->mutex);
3683         dout("wait_unsafe_requests want %lld\n", want_tid);
3684 restart:
3685         req = __get_oldest_req(mdsc);
3686         while (req && req->r_tid <= want_tid) {
3687                 /* find next request */
3688                 n = rb_next(&req->r_node);
3689                 if (n)
3690                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3691                 else
3692                         nextreq = NULL;
3693                 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3694                     (req->r_op & CEPH_MDS_OP_WRITE)) {
3695                         /* write op */
3696                         ceph_mdsc_get_request(req);
3697                         if (nextreq)
3698                                 ceph_mdsc_get_request(nextreq);
3699                         mutex_unlock(&mdsc->mutex);
3700                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
3701                              req->r_tid, want_tid);
3702                         wait_for_completion(&req->r_safe_completion);
3703                         mutex_lock(&mdsc->mutex);
3704                         ceph_mdsc_put_request(req);
3705                         if (!nextreq)
3706                                 break;  /* next dne before, so we're done! */
3707                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
3708                                 /* next request was removed from tree */
3709                                 ceph_mdsc_put_request(nextreq);
3710                                 goto restart;
3711                         }
3712                         ceph_mdsc_put_request(nextreq);  /* won't go away */
3713                 }
3714                 req = nextreq;
3715         }
3716         mutex_unlock(&mdsc->mutex);
3717         dout("wait_unsafe_requests done\n");
3718 }
3719
3720 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3721 {
3722         u64 want_tid, want_flush;
3723
3724         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3725                 return;
3726
3727         dout("sync\n");
3728         mutex_lock(&mdsc->mutex);
3729         want_tid = mdsc->last_tid;
3730         mutex_unlock(&mdsc->mutex);
3731
3732         ceph_flush_dirty_caps(mdsc);
3733         spin_lock(&mdsc->cap_dirty_lock);
3734         want_flush = mdsc->last_cap_flush_tid;
3735         if (!list_empty(&mdsc->cap_flush_list)) {
3736                 struct ceph_cap_flush *cf =
3737                         list_last_entry(&mdsc->cap_flush_list,
3738                                         struct ceph_cap_flush, g_list);
3739                 cf->wake = true;
3740         }
3741         spin_unlock(&mdsc->cap_dirty_lock);
3742
3743         dout("sync want tid %lld flush_seq %lld\n",
3744              want_tid, want_flush);
3745
3746         wait_unsafe_requests(mdsc, want_tid);
3747         wait_caps_flush(mdsc, want_flush);
3748 }
3749
3750 /*
3751  * true if all sessions are closed, or we force unmount
3752  */
3753 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
3754 {
3755         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3756                 return true;
3757         return atomic_read(&mdsc->num_sessions) <= skipped;
3758 }
3759
3760 /*
3761  * called after sb is ro.
3762  */
3763 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3764 {
3765         struct ceph_options *opts = mdsc->fsc->client->options;
3766         struct ceph_mds_session *session;
3767         int i;
3768         int skipped = 0;
3769
3770         dout("close_sessions\n");
3771
3772         /* close sessions */
3773         mutex_lock(&mdsc->mutex);
3774         for (i = 0; i < mdsc->max_sessions; i++) {
3775                 session = __ceph_lookup_mds_session(mdsc, i);
3776                 if (!session)
3777                         continue;
3778                 mutex_unlock(&mdsc->mutex);
3779                 mutex_lock(&session->s_mutex);
3780                 if (__close_session(mdsc, session) <= 0)
3781                         skipped++;
3782                 mutex_unlock(&session->s_mutex);
3783                 ceph_put_mds_session(session);
3784                 mutex_lock(&mdsc->mutex);
3785         }
3786         mutex_unlock(&mdsc->mutex);
3787
3788         dout("waiting for sessions to close\n");
3789         wait_event_timeout(mdsc->session_close_wq,
3790                            done_closing_sessions(mdsc, skipped),
3791                            ceph_timeout_jiffies(opts->mount_timeout));
3792
3793         /* tear down remaining sessions */
3794         mutex_lock(&mdsc->mutex);
3795         for (i = 0; i < mdsc->max_sessions; i++) {
3796                 if (mdsc->sessions[i]) {
3797                         session = get_session(mdsc->sessions[i]);
3798                         __unregister_session(mdsc, session);
3799                         mutex_unlock(&mdsc->mutex);
3800                         mutex_lock(&session->s_mutex);
3801                         remove_session_caps(session);
3802                         mutex_unlock(&session->s_mutex);
3803                         ceph_put_mds_session(session);
3804                         mutex_lock(&mdsc->mutex);
3805                 }
3806         }
3807         WARN_ON(!list_empty(&mdsc->cap_delay_list));
3808         mutex_unlock(&mdsc->mutex);
3809
3810         ceph_cleanup_empty_realms(mdsc);
3811
3812         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3813
3814         dout("stopped\n");
3815 }
3816
3817 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
3818 {
3819         struct ceph_mds_session *session;
3820         int mds;
3821
3822         dout("force umount\n");
3823
3824         mutex_lock(&mdsc->mutex);
3825         for (mds = 0; mds < mdsc->max_sessions; mds++) {
3826                 session = __ceph_lookup_mds_session(mdsc, mds);
3827                 if (!session)
3828                         continue;
3829                 mutex_unlock(&mdsc->mutex);
3830                 mutex_lock(&session->s_mutex);
3831                 __close_session(mdsc, session);
3832                 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
3833                         cleanup_session_requests(mdsc, session);
3834                         remove_session_caps(session);
3835                 }
3836                 mutex_unlock(&session->s_mutex);
3837                 ceph_put_mds_session(session);
3838                 mutex_lock(&mdsc->mutex);
3839                 kick_requests(mdsc, mds);
3840         }
3841         __wake_requests(mdsc, &mdsc->waiting_for_map);
3842         mutex_unlock(&mdsc->mutex);
3843 }
3844
3845 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3846 {
3847         dout("stop\n");
3848         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3849         if (mdsc->mdsmap)
3850                 ceph_mdsmap_destroy(mdsc->mdsmap);
3851         kfree(mdsc->sessions);
3852         ceph_caps_finalize(mdsc);
3853         ceph_pool_perm_destroy(mdsc);
3854 }
3855
3856 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3857 {
3858         struct ceph_mds_client *mdsc = fsc->mdsc;
3859         dout("mdsc_destroy %p\n", mdsc);
3860
3861         /* flush out any connection work with references to us */
3862         ceph_msgr_flush();
3863
3864         ceph_mdsc_stop(mdsc);
3865
3866         fsc->mdsc = NULL;
3867         kfree(mdsc);
3868         dout("mdsc_destroy %p done\n", mdsc);
3869 }
3870
3871 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3872 {
3873         struct ceph_fs_client *fsc = mdsc->fsc;
3874         const char *mds_namespace = fsc->mount_options->mds_namespace;
3875         void *p = msg->front.iov_base;
3876         void *end = p + msg->front.iov_len;
3877         u32 epoch;
3878         u32 map_len;
3879         u32 num_fs;
3880         u32 mount_fscid = (u32)-1;
3881         u8 struct_v, struct_cv;
3882         int err = -EINVAL;
3883
3884         ceph_decode_need(&p, end, sizeof(u32), bad);
3885         epoch = ceph_decode_32(&p);
3886
3887         dout("handle_fsmap epoch %u\n", epoch);
3888
3889         ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3890         struct_v = ceph_decode_8(&p);
3891         struct_cv = ceph_decode_8(&p);
3892         map_len = ceph_decode_32(&p);
3893
3894         ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
3895         p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
3896
3897         num_fs = ceph_decode_32(&p);
3898         while (num_fs-- > 0) {
3899                 void *info_p, *info_end;
3900                 u32 info_len;
3901                 u8 info_v, info_cv;
3902                 u32 fscid, namelen;
3903
3904                 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3905                 info_v = ceph_decode_8(&p);
3906                 info_cv = ceph_decode_8(&p);
3907                 info_len = ceph_decode_32(&p);
3908                 ceph_decode_need(&p, end, info_len, bad);
3909                 info_p = p;
3910                 info_end = p + info_len;
3911                 p = info_end;
3912
3913                 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
3914                 fscid = ceph_decode_32(&info_p);
3915                 namelen = ceph_decode_32(&info_p);
3916                 ceph_decode_need(&info_p, info_end, namelen, bad);
3917
3918                 if (mds_namespace &&
3919                     strlen(mds_namespace) == namelen &&
3920                     !strncmp(mds_namespace, (char *)info_p, namelen)) {
3921                         mount_fscid = fscid;
3922                         break;
3923                 }
3924         }
3925
3926         ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
3927         if (mount_fscid != (u32)-1) {
3928                 fsc->client->monc.fs_cluster_id = mount_fscid;
3929                 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
3930                                    0, true);
3931                 ceph_monc_renew_subs(&fsc->client->monc);
3932         } else {
3933                 err = -ENOENT;
3934                 goto err_out;
3935         }
3936         return;
3937
3938 bad:
3939         pr_err("error decoding fsmap\n");
3940 err_out:
3941         mutex_lock(&mdsc->mutex);
3942         mdsc->mdsmap_err = err;
3943         __wake_requests(mdsc, &mdsc->waiting_for_map);
3944         mutex_unlock(&mdsc->mutex);
3945 }
3946
3947 /*
3948  * handle mds map update.
3949  */
3950 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3951 {
3952         u32 epoch;
3953         u32 maplen;
3954         void *p = msg->front.iov_base;
3955         void *end = p + msg->front.iov_len;
3956         struct ceph_mdsmap *newmap, *oldmap;
3957         struct ceph_fsid fsid;
3958         int err = -EINVAL;
3959
3960         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3961         ceph_decode_copy(&p, &fsid, sizeof(fsid));
3962         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3963                 return;
3964         epoch = ceph_decode_32(&p);
3965         maplen = ceph_decode_32(&p);
3966         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3967
3968         /* do we need it? */
3969         mutex_lock(&mdsc->mutex);
3970         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3971                 dout("handle_map epoch %u <= our %u\n",
3972                      epoch, mdsc->mdsmap->m_epoch);
3973                 mutex_unlock(&mdsc->mutex);
3974                 return;
3975         }
3976
3977         newmap = ceph_mdsmap_decode(&p, end);
3978         if (IS_ERR(newmap)) {
3979                 err = PTR_ERR(newmap);
3980                 goto bad_unlock;
3981         }
3982
3983         /* swap into place */
3984         if (mdsc->mdsmap) {
3985                 oldmap = mdsc->mdsmap;
3986                 mdsc->mdsmap = newmap;
3987                 check_new_map(mdsc, newmap, oldmap);
3988                 ceph_mdsmap_destroy(oldmap);
3989         } else {
3990                 mdsc->mdsmap = newmap;  /* first mds map */
3991         }
3992         mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3993
3994         __wake_requests(mdsc, &mdsc->waiting_for_map);
3995         ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
3996                           mdsc->mdsmap->m_epoch);
3997
3998         mutex_unlock(&mdsc->mutex);
3999         schedule_delayed(mdsc);
4000         return;
4001
4002 bad_unlock:
4003         mutex_unlock(&mdsc->mutex);
4004 bad:
4005         pr_err("error decoding mdsmap %d\n", err);
4006         return;
4007 }
4008
4009 static struct ceph_connection *con_get(struct ceph_connection *con)
4010 {
4011         struct ceph_mds_session *s = con->private;
4012
4013         if (get_session(s)) {
4014                 dout("mdsc con_get %p ok (%d)\n", s, refcount_read(&s->s_ref));
4015                 return con;
4016         }
4017         dout("mdsc con_get %p FAIL\n", s);
4018         return NULL;
4019 }
4020
4021 static void con_put(struct ceph_connection *con)
4022 {
4023         struct ceph_mds_session *s = con->private;
4024
4025         dout("mdsc con_put %p (%d)\n", s, refcount_read(&s->s_ref) - 1);
4026         ceph_put_mds_session(s);
4027 }
4028
4029 /*
4030  * if the client is unresponsive for long enough, the mds will kill
4031  * the session entirely.
4032  */
4033 static void peer_reset(struct ceph_connection *con)
4034 {
4035         struct ceph_mds_session *s = con->private;
4036         struct ceph_mds_client *mdsc = s->s_mdsc;
4037
4038         pr_warn("mds%d closed our session\n", s->s_mds);
4039         send_mds_reconnect(mdsc, s);
4040 }
4041
4042 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
4043 {
4044         struct ceph_mds_session *s = con->private;
4045         struct ceph_mds_client *mdsc = s->s_mdsc;
4046         int type = le16_to_cpu(msg->hdr.type);
4047
4048         mutex_lock(&mdsc->mutex);
4049         if (__verify_registered_session(mdsc, s) < 0) {
4050                 mutex_unlock(&mdsc->mutex);
4051                 goto out;
4052         }
4053         mutex_unlock(&mdsc->mutex);
4054
4055         switch (type) {
4056         case CEPH_MSG_MDS_MAP:
4057                 ceph_mdsc_handle_mdsmap(mdsc, msg);
4058                 break;
4059         case CEPH_MSG_FS_MAP_USER:
4060                 ceph_mdsc_handle_fsmap(mdsc, msg);
4061                 break;
4062         case CEPH_MSG_CLIENT_SESSION:
4063                 handle_session(s, msg);
4064                 break;
4065         case CEPH_MSG_CLIENT_REPLY:
4066                 handle_reply(s, msg);
4067                 break;
4068         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
4069                 handle_forward(mdsc, s, msg);
4070                 break;
4071         case CEPH_MSG_CLIENT_CAPS:
4072                 ceph_handle_caps(s, msg);
4073                 break;
4074         case CEPH_MSG_CLIENT_SNAP:
4075                 ceph_handle_snap(mdsc, s, msg);
4076                 break;
4077         case CEPH_MSG_CLIENT_LEASE:
4078                 handle_lease(mdsc, s, msg);
4079                 break;
4080
4081         default:
4082                 pr_err("received unknown message type %d %s\n", type,
4083                        ceph_msg_type_name(type));
4084         }
4085 out:
4086         ceph_msg_put(msg);
4087 }
4088
4089 /*
4090  * authentication
4091  */
4092
4093 /*
4094  * Note: returned pointer is the address of a structure that's
4095  * managed separately.  Caller must *not* attempt to free it.
4096  */
4097 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
4098                                         int *proto, int force_new)
4099 {
4100         struct ceph_mds_session *s = con->private;
4101         struct ceph_mds_client *mdsc = s->s_mdsc;
4102         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4103         struct ceph_auth_handshake *auth = &s->s_auth;
4104
4105         if (force_new && auth->authorizer) {
4106                 ceph_auth_destroy_authorizer(auth->authorizer);
4107                 auth->authorizer = NULL;
4108         }
4109         if (!auth->authorizer) {
4110                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4111                                                       auth);
4112                 if (ret)
4113                         return ERR_PTR(ret);
4114         } else {
4115                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4116                                                       auth);
4117                 if (ret)
4118                         return ERR_PTR(ret);
4119         }
4120         *proto = ac->protocol;
4121
4122         return auth;
4123 }
4124
4125
4126 static int verify_authorizer_reply(struct ceph_connection *con)
4127 {
4128         struct ceph_mds_session *s = con->private;
4129         struct ceph_mds_client *mdsc = s->s_mdsc;
4130         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4131
4132         return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
4133 }
4134
4135 static int invalidate_authorizer(struct ceph_connection *con)
4136 {
4137         struct ceph_mds_session *s = con->private;
4138         struct ceph_mds_client *mdsc = s->s_mdsc;
4139         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4140
4141         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
4142
4143         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
4144 }
4145
4146 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
4147                                 struct ceph_msg_header *hdr, int *skip)
4148 {
4149         struct ceph_msg *msg;
4150         int type = (int) le16_to_cpu(hdr->type);
4151         int front_len = (int) le32_to_cpu(hdr->front_len);
4152
4153         if (con->in_msg)
4154                 return con->in_msg;
4155
4156         *skip = 0;
4157         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
4158         if (!msg) {
4159                 pr_err("unable to allocate msg type %d len %d\n",
4160                        type, front_len);
4161                 return NULL;
4162         }
4163
4164         return msg;
4165 }
4166
4167 static int mds_sign_message(struct ceph_msg *msg)
4168 {
4169        struct ceph_mds_session *s = msg->con->private;
4170        struct ceph_auth_handshake *auth = &s->s_auth;
4171
4172        return ceph_auth_sign_message(auth, msg);
4173 }
4174
4175 static int mds_check_message_signature(struct ceph_msg *msg)
4176 {
4177        struct ceph_mds_session *s = msg->con->private;
4178        struct ceph_auth_handshake *auth = &s->s_auth;
4179
4180        return ceph_auth_check_message_signature(auth, msg);
4181 }
4182
4183 static const struct ceph_connection_operations mds_con_ops = {
4184         .get = con_get,
4185         .put = con_put,
4186         .dispatch = dispatch,
4187         .get_authorizer = get_authorizer,
4188         .verify_authorizer_reply = verify_authorizer_reply,
4189         .invalidate_authorizer = invalidate_authorizer,
4190         .peer_reset = peer_reset,
4191         .alloc_msg = mds_alloc_msg,
4192         .sign_message = mds_sign_message,
4193         .check_message_signature = mds_check_message_signature,
4194 };
4195
4196 /* eof */