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  45  * are no longer valid.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         if (features & CEPH_FEATURE_MDS_QUOTA) {
 104                 u8 struct_v, struct_compat;
 105                 u32 struct_len;
 106
 107                 /*
 108                  * both struct_v and struct_compat are expected to be >= 1
 109                  */
 110                 ceph_decode_8_safe(p, end, struct_v, bad);
 111                 ceph_decode_8_safe(p, end, struct_compat, bad);
 112                 if (!struct_v || !struct_compat)
 113                         goto bad;
 114                 ceph_decode_32_safe(p, end, struct_len, bad);
 115                 ceph_decode_need(p, end, struct_len, bad);
 116                 ceph_decode_64_safe(p, end, info->max_bytes, bad);
 117                 ceph_decode_64_safe(p, end, info->max_files, bad);
 118         } else {
 119                 info->max_bytes = 0;
 120                 info->max_files = 0;
 121         }
 122
 123         info->pool_ns_len = 0;
 124         info->pool_ns_data = NULL;
 125         if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
 126                 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
 127                 if (info->pool_ns_len > 0) {
 128                         ceph_decode_need(p, end, info->pool_ns_len, bad);
 129                         info->pool_ns_data = *p;
 130                         *p += info->pool_ns_len;
 131                 }
 132         }
 133
 134         return 0;
 135 bad:
 136         return err;
 137 }
 138
 139 /*
 140  * parse a normal reply, which may contain a (dir+)dentry and/or a
 141  * target inode.
 142  */
 143 static int parse_reply_info_trace(void **p, void *end,
 144                                   struct ceph_mds_reply_info_parsed *info,
 145                                   u64 features)
 146 {
 147         int err;
 148
 149         if (info->head->is_dentry) {
 150                 err = parse_reply_info_in(p, end, &info->diri, features);
 151                 if (err < 0)
 152                         goto out_bad;
 153
 154                 if (unlikely(*p + sizeof(*info->dirfrag) > end))
 155                         goto bad;
 156                 info->dirfrag = *p;
 157                 *p += sizeof(*info->dirfrag) +
 158                         sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
 159                 if (unlikely(*p > end))
 160                         goto bad;
 161
 162                 ceph_decode_32_safe(p, end, info->dname_len, bad);
 163                 ceph_decode_need(p, end, info->dname_len, bad);
 164                 info->dname = *p;
 165                 *p += info->dname_len;
 166                 info->dlease = *p;
 167                 *p += sizeof(*info->dlease);
 168         }
 169
 170         if (info->head->is_target) {
 171                 err = parse_reply_info_in(p, end, &info->targeti, features);
 172                 if (err < 0)
 173                         goto out_bad;
 174         }
 175
 176         if (unlikely(*p != end))
 177                 goto bad;
 178         return 0;
 179
 180 bad:
 181         err = -EIO;
 182 out_bad:
 183         pr_err("problem parsing mds trace %d\n", err);
 184         return err;
 185 }
 186
 187 /*
 188  * parse readdir results
 189  */
 190 static int parse_reply_info_dir(void **p, void *end,
 191                                 struct ceph_mds_reply_info_parsed *info,
 192                                 u64 features)
 193 {
 194         u32 num, i = 0;
 195         int err;
 196
 197         info->dir_dir = *p;
 198         if (*p + sizeof(*info->dir_dir) > end)
 199                 goto bad;
 200         *p += sizeof(*info->dir_dir) +
 201                 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
 202         if (*p > end)
 203                 goto bad;
 204
 205         ceph_decode_need(p, end, sizeof(num) + 2, bad);
 206         num = ceph_decode_32(p);
 207         {
 208                 u16 flags = ceph_decode_16(p);
 209                 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
 210                 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
 211                 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
 212                 info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
 213         }
 214         if (num == 0)
 215                 goto done;
 216
 217         BUG_ON(!info->dir_entries);
 218         if ((unsigned long)(info->dir_entries + num) >
 219             (unsigned long)info->dir_entries + info->dir_buf_size) {
 220                 pr_err("dir contents are larger than expected\n");
 221                 WARN_ON(1);
 222                 goto bad;
 223         }
 224
 225         info->dir_nr = num;
 226         while (num) {
 227                 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
 228                 /* dentry */
 229                 ceph_decode_need(p, end, sizeof(u32)*2, bad);
 230                 rde->name_len = ceph_decode_32(p);
 231                 ceph_decode_need(p, end, rde->name_len, bad);
 232                 rde->name = *p;
 233                 *p += rde->name_len;
 234                 dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
 235                 rde->lease = *p;
 236                 *p += sizeof(struct ceph_mds_reply_lease);
 237
 238                 /* inode */
 239                 err = parse_reply_info_in(p, end, &rde->inode, features);
 240                 if (err < 0)
 241                         goto out_bad;
 242                 /* ceph_readdir_prepopulate() will update it */
 243                 rde->offset = 0;
 244                 i++;
 245                 num--;
 246         }
 247
 248 done:
 249         if (*p != end)
 250                 goto bad;
 251         return 0;
 252
 253 bad:
 254         err = -EIO;
 255 out_bad:
 256         pr_err("problem parsing dir contents %d\n", err);
 257         return err;
 258 }
 259
 260 /*
 261  * parse fcntl F_GETLK results
 262  */
 263 static int parse_reply_info_filelock(void **p, void *end,
 264                                      struct ceph_mds_reply_info_parsed *info,
 265                                      u64 features)
 266 {
 267         if (*p + sizeof(*info->filelock_reply) > end)
 268                 goto bad;
 269
 270         info->filelock_reply = *p;
 271         *p += sizeof(*info->filelock_reply);
 272
 273         if (unlikely(*p != end))
 274                 goto bad;
 275         return 0;
 276
 277 bad:
 278         return -EIO;
 279 }
 280
 281 /*
 282  * parse create results
 283  */
 284 static int parse_reply_info_create(void **p, void *end,
 285                                   struct ceph_mds_reply_info_parsed *info,
 286                                   u64 features)
 287 {
 288         if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
 289                 if (*p == end) {
 290                         info->has_create_ino = false;
 291                 } else {
 292                         info->has_create_ino = true;
 293                         info->ino = ceph_decode_64(p);
 294                 }
 295         }
 296
 297         if (unlikely(*p != end))
 298                 goto bad;
 299         return 0;
 300
 301 bad:
 302         return -EIO;
 303 }
 304
 305 /*
 306  * parse extra results
 307  */
 308 static int parse_reply_info_extra(void **p, void *end,
 309                                   struct ceph_mds_reply_info_parsed *info,
 310                                   u64 features)
 311 {
 312         u32 op = le32_to_cpu(info->head->op);
 313
 314         if (op == CEPH_MDS_OP_GETFILELOCK)
 315                 return parse_reply_info_filelock(p, end, info, features);
 316         else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
 317                 return parse_reply_info_dir(p, end, info, features);
 318         else if (op == CEPH_MDS_OP_CREATE)
 319                 return parse_reply_info_create(p, end, info, features);
 320         else
 321                 return -EIO;
 322 }
 323
 324 /*
 325  * parse entire mds reply
 326  */
 327 static int parse_reply_info(struct ceph_msg *msg,
 328                             struct ceph_mds_reply_info_parsed *info,
 329                             u64 features)
 330 {
 331         void *p, *end;
 332         u32 len;
 333         int err;
 334
 335         info->head = msg->front.iov_base;
 336         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 337         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 338
 339         /* trace */
 340         ceph_decode_32_safe(&p, end, len, bad);
 341         if (len > 0) {
 342                 ceph_decode_need(&p, end, len, bad);
 343                 err = parse_reply_info_trace(&p, p+len, info, features);
 344                 if (err < 0)
 345                         goto out_bad;
 346         }
 347
 348         /* extra */
 349         ceph_decode_32_safe(&p, end, len, bad);
 350         if (len > 0) {
 351                 ceph_decode_need(&p, end, len, bad);
 352                 err = parse_reply_info_extra(&p, p+len, info, features);
 353                 if (err < 0)
 354                         goto out_bad;
 355         }
 356
 357         /* snap blob */
 358         ceph_decode_32_safe(&p, end, len, bad);
 359         info->snapblob_len = len;
 360         info->snapblob = p;
 361         p += len;
 362
 363         if (p != end)
 364                 goto bad;
 365         return 0;
 366
 367 bad:
 368         err = -EIO;
 369 out_bad:
 370         pr_err("mds parse_reply err %d\n", err);
 371         return err;
 372 }
 373
 374 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
 375 {
 376         if (!info->dir_entries)
 377                 return;
 378         free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
 379 }
 380
 381
 382 /*
 383  * sessions
 384  */
 385 const char *ceph_session_state_name(int s)
 386 {
 387         switch (s) {
 388         case CEPH_MDS_SESSION_NEW: return "new";
 389         case CEPH_MDS_SESSION_OPENING: return "opening";
 390         case CEPH_MDS_SESSION_OPEN: return "open";
 391         case CEPH_MDS_SESSION_HUNG: return "hung";
 392         case CEPH_MDS_SESSION_CLOSING: return "closing";
 393         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
 394         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
 395         case CEPH_MDS_SESSION_REJECTED: return "rejected";
 396         default: return "???";
 397         }
 398 }
 399
 400 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
 401 {
 402         if (refcount_inc_not_zero(&s->s_ref)) {
 403                 dout("mdsc get_session %p %d -> %d\n", s,
 404                      refcount_read(&s->s_ref)-1, refcount_read(&s->s_ref));
 405                 return s;
 406         } else {
 407                 dout("mdsc get_session %p 0 -- FAIL\n", s);
 408                 return NULL;
 409         }
 410 }
 411
 412 void ceph_put_mds_session(struct ceph_mds_session *s)
 413 {
 414         dout("mdsc put_session %p %d -> %d\n", s,
 415              refcount_read(&s->s_ref), refcount_read(&s->s_ref)-1);
 416         if (refcount_dec_and_test(&s->s_ref)) {
 417                 if (s->s_auth.authorizer)
 418                         ceph_auth_destroy_authorizer(s->s_auth.authorizer);
 419                 kfree(s);
 420         }
 421 }
 422
 423 /*
 424  * called under mdsc->mutex
 425  */
 426 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
 427                                                    int mds)
 428 {
 429         struct ceph_mds_session *session;
 430
 431         if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
 432                 return NULL;
 433         session = mdsc->sessions[mds];
 434         dout("lookup_mds_session %p %d\n", session,
 435              refcount_read(&session->s_ref));
 436         get_session(session);
 437         return session;
 438 }
 439
 440 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
 441 {
 442         if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
 443                 return false;
 444         else
 445                 return true;
 446 }
 447
 448 static int __verify_registered_session(struct ceph_mds_client *mdsc,
 449                                        struct ceph_mds_session *s)
 450 {
 451         if (s->s_mds >= mdsc->max_sessions ||
 452             mdsc->sessions[s->s_mds] != s)
 453                 return -ENOENT;
 454         return 0;
 455 }
 456
 457 /*
 458  * create+register a new session for given mds.
 459  * called under mdsc->mutex.
 460  */
 461 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 462                                                  int mds)
 463 {
 464         struct ceph_mds_session *s;
 465
 466         if (mds >= mdsc->mdsmap->m_num_mds)
 467                 return ERR_PTR(-EINVAL);
 468
 469         s = kzalloc(sizeof(*s), GFP_NOFS);
 470         if (!s)
 471                 return ERR_PTR(-ENOMEM);
 472
 473         if (mds >= mdsc->max_sessions) {
 474                 int newmax = 1 << get_count_order(mds + 1);
 475                 struct ceph_mds_session **sa;
 476
 477                 dout("%s: realloc to %d\n", __func__, newmax);
 478                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 479                 if (!sa)
 480                         goto fail_realloc;
 481                 if (mdsc->sessions) {
 482                         memcpy(sa, mdsc->sessions,
 483                                mdsc->max_sessions * sizeof(void *));
 484                         kfree(mdsc->sessions);
 485                 }
 486                 mdsc->sessions = sa;
 487                 mdsc->max_sessions = newmax;
 488         }
 489
 490         dout("%s: mds%d\n", __func__, mds);
 491         s->s_mdsc = mdsc;
 492         s->s_mds = mds;
 493         s->s_state = CEPH_MDS_SESSION_NEW;
 494         s->s_ttl = 0;
 495         s->s_seq = 0;
 496         mutex_init(&s->s_mutex);
 497
 498         ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
 499
 500         spin_lock_init(&s->s_gen_ttl_lock);
 501         s->s_cap_gen = 0;
 502         s->s_cap_ttl = jiffies - 1;
 503
 504         spin_lock_init(&s->s_cap_lock);
 505         s->s_renew_requested = 0;
 506         s->s_renew_seq = 0;
 507         INIT_LIST_HEAD(&s->s_caps);
 508         s->s_nr_caps = 0;
 509         s->s_trim_caps = 0;
 510         refcount_set(&s->s_ref, 1);
 511         INIT_LIST_HEAD(&s->s_waiting);
 512         INIT_LIST_HEAD(&s->s_unsafe);
 513         s->s_num_cap_releases = 0;
 514         s->s_cap_reconnect = 0;
 515         s->s_cap_iterator = NULL;
 516         INIT_LIST_HEAD(&s->s_cap_releases);
 517         INIT_LIST_HEAD(&s->s_cap_flushing);
 518
 519         mdsc->sessions[mds] = s;
 520         atomic_inc(&mdsc->num_sessions);
 521         refcount_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
 522
 523         ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
 524                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
 525
 526         return s;
 527
 528 fail_realloc:
 529         kfree(s);
 530         return ERR_PTR(-ENOMEM);
 531 }
 532
 533 /*
 534  * called under mdsc->mutex
 535  */
 536 static void __unregister_session(struct ceph_mds_client *mdsc,
 537                                struct ceph_mds_session *s)
 538 {
 539         dout("__unregister_session mds%d %p\n", s->s_mds, s);
 540         BUG_ON(mdsc->sessions[s->s_mds] != s);
 541         mdsc->sessions[s->s_mds] = NULL;
 542         ceph_con_close(&s->s_con);
 543         ceph_put_mds_session(s);
 544         atomic_dec(&mdsc->num_sessions);
 545 }
 546
 547 /*
 548  * drop session refs in request.
 549  *
 550  * should be last request ref, or hold mdsc->mutex
 551  */
 552 static void put_request_session(struct ceph_mds_request *req)
 553 {
 554         if (req->r_session) {
 555                 ceph_put_mds_session(req->r_session);
 556                 req->r_session = NULL;
 557         }
 558 }
 559
 560 void ceph_mdsc_release_request(struct kref *kref)
 561 {
 562         struct ceph_mds_request *req = container_of(kref,
 563                                                     struct ceph_mds_request,
 564                                                     r_kref);
 565         destroy_reply_info(&req->r_reply_info);
 566         if (req->r_request)
 567                 ceph_msg_put(req->r_request);
 568         if (req->r_reply)
 569                 ceph_msg_put(req->r_reply);
 570         if (req->r_inode) {
 571                 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
 572                 iput(req->r_inode);
 573         }
 574         if (req->r_parent)
 575                 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
 576         iput(req->r_target_inode);
 577         if (req->r_dentry)
 578                 dput(req->r_dentry);
 579         if (req->r_old_dentry)
 580                 dput(req->r_old_dentry);
 581         if (req->r_old_dentry_dir) {
 582                 /*
 583                  * track (and drop pins for) r_old_dentry_dir
 584                  * separately, since r_old_dentry's d_parent may have
 585                  * changed between the dir mutex being dropped and
 586                  * this request being freed.
 587                  */
 588                 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
 589                                   CEPH_CAP_PIN);
 590                 iput(req->r_old_dentry_dir);
 591         }
 592         kfree(req->r_path1);
 593         kfree(req->r_path2);
 594         if (req->r_pagelist)
 595                 ceph_pagelist_release(req->r_pagelist);
 596         put_request_session(req);
 597         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
 598         kfree(req);
 599 }
 600
 601 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
 602
 603 /*
 604  * lookup session, bump ref if found.
 605  *
 606  * called under mdsc->mutex.
 607  */
 608 static struct ceph_mds_request *
 609 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
 610 {
 611         struct ceph_mds_request *req;
 612
 613         req = lookup_request(&mdsc->request_tree, tid);
 614         if (req)
 615                 ceph_mdsc_get_request(req);
 616
 617         return req;
 618 }
 619
 620 /*
 621  * Register an in-flight request, and assign a tid.  Link to directory
 622  * are modifying (if any).
 623  *
 624  * Called under mdsc->mutex.
 625  */
 626 static void __register_request(struct ceph_mds_client *mdsc,
 627                                struct ceph_mds_request *req,
 628                                struct inode *dir)
 629 {
 630         int ret = 0;
 631
 632         req->r_tid = ++mdsc->last_tid;
 633         if (req->r_num_caps) {
 634                 ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
 635                                         req->r_num_caps);
 636                 if (ret < 0) {
 637                         pr_err("__register_request %p "
 638                                "failed to reserve caps: %d\n", req, ret);
 639                         /* set req->r_err to fail early from __do_request */
 640                         req->r_err = ret;
 641                         return;
 642                 }
 643         }
 644         dout("__register_request %p tid %lld\n", req, req->r_tid);
 645         ceph_mdsc_get_request(req);
 646         insert_request(&mdsc->request_tree, req);
 647
 648         req->r_uid = current_fsuid();
 649         req->r_gid = current_fsgid();
 650
 651         if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
 652                 mdsc->oldest_tid = req->r_tid;
 653
 654         if (dir) {
 655                 ihold(dir);
 656                 req->r_unsafe_dir = dir;
 657         }
 658 }
 659
 660 static void __unregister_request(struct ceph_mds_client *mdsc,
 661                                  struct ceph_mds_request *req)
 662 {
 663         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 664
 665         /* Never leave an unregistered request on an unsafe list! */
 666         list_del_init(&req->r_unsafe_item);
 667
 668         if (req->r_tid == mdsc->oldest_tid) {
 669                 struct rb_node *p = rb_next(&req->r_node);
 670                 mdsc->oldest_tid = 0;
 671                 while (p) {
 672                         struct ceph_mds_request *next_req =
 673                                 rb_entry(p, struct ceph_mds_request, r_node);
 674                         if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
 675                                 mdsc->oldest_tid = next_req->r_tid;
 676                                 break;
 677                         }
 678                         p = rb_next(p);
 679                 }
 680         }
 681
 682         erase_request(&mdsc->request_tree, req);
 683
 684         if (req->r_unsafe_dir  &&
 685             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
 686                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 687                 spin_lock(&ci->i_unsafe_lock);
 688                 list_del_init(&req->r_unsafe_dir_item);
 689                 spin_unlock(&ci->i_unsafe_lock);
 690         }
 691         if (req->r_target_inode &&
 692             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
 693                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
 694                 spin_lock(&ci->i_unsafe_lock);
 695                 list_del_init(&req->r_unsafe_target_item);
 696                 spin_unlock(&ci->i_unsafe_lock);
 697         }
 698
 699         if (req->r_unsafe_dir) {
 700                 iput(req->r_unsafe_dir);
 701                 req->r_unsafe_dir = NULL;
 702         }
 703
 704         complete_all(&req->r_safe_completion);
 705
 706         ceph_mdsc_put_request(req);
 707 }
 708
 709 /*
 710  * Walk back up the dentry tree until we hit a dentry representing a
 711  * non-snapshot inode. We do this using the rcu_read_lock (which must be held
 712  * when calling this) to ensure that the objects won't disappear while we're
 713  * working with them. Once we hit a candidate dentry, we attempt to take a
 714  * reference to it, and return that as the result.
 715  */
 716 static struct inode *get_nonsnap_parent(struct dentry *dentry)
 717 {
 718         struct inode *inode = NULL;
 719
 720         while (dentry && !IS_ROOT(dentry)) {
 721                 inode = d_inode_rcu(dentry);
 722                 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
 723                         break;
 724                 dentry = dentry->d_parent;
 725         }
 726         if (inode)
 727                 inode = igrab(inode);
 728         return inode;
 729 }
 730
 731 /*
 732  * Choose mds to send request to next.  If there is a hint set in the
 733  * request (e.g., due to a prior forward hint from the mds), use that.
 734  * Otherwise, consult frag tree and/or caps to identify the
 735  * appropriate mds.  If all else fails, choose randomly.
 736  *
 737  * Called under mdsc->mutex.
 738  */
 739 static int __choose_mds(struct ceph_mds_client *mdsc,
 740                         struct ceph_mds_request *req)
 741 {
 742         struct inode *inode;
 743         struct ceph_inode_info *ci;
 744         struct ceph_cap *cap;
 745         int mode = req->r_direct_mode;
 746         int mds = -1;
 747         u32 hash = req->r_direct_hash;
 748         bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
 749
 750         /*
 751          * is there a specific mds we should try?  ignore hint if we have
 752          * no session and the mds is not up (active or recovering).
 753          */
 754         if (req->r_resend_mds >= 0 &&
 755             (__have_session(mdsc, req->r_resend_mds) ||
 756              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 757                 dout("choose_mds using resend_mds mds%d\n",
 758                      req->r_resend_mds);
 759                 return req->r_resend_mds;
 760         }
 761
 762         if (mode == USE_RANDOM_MDS)
 763                 goto random;
 764
 765         inode = NULL;
 766         if (req->r_inode) {
 767                 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
 768                         inode = req->r_inode;
 769                         ihold(inode);
 770                 } else {
 771                         /* req->r_dentry is non-null for LSSNAP request */
 772                         rcu_read_lock();
 773                         inode = get_nonsnap_parent(req->r_dentry);
 774                         rcu_read_unlock();
 775                         dout("__choose_mds using snapdir's parent %p\n", inode);
 776                 }
 777         } else if (req->r_dentry) {
 778                 /* ignore race with rename; old or new d_parent is okay */
 779                 struct dentry *parent;
 780                 struct inode *dir;
 781
 782                 rcu_read_lock();
 783                 parent = req->r_dentry->d_parent;
 784                 dir = req->r_parent ? : d_inode_rcu(parent);
 785
 786                 if (!dir || dir->i_sb != mdsc->fsc->sb) {
 787                         /*  not this fs or parent went negative */
 788                         inode = d_inode(req->r_dentry);
 789                         if (inode)
 790                                 ihold(inode);
 791                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
 792                         /* direct snapped/virtual snapdir requests
 793                          * based on parent dir inode */
 794                         inode = get_nonsnap_parent(parent);
 795                         dout("__choose_mds using nonsnap parent %p\n", inode);
 796                 } else {
 797                         /* dentry target */
 798                         inode = d_inode(req->r_dentry);
 799                         if (!inode || mode == USE_AUTH_MDS) {
 800                                 /* dir + name */
 801                                 inode = igrab(dir);
 802                                 hash = ceph_dentry_hash(dir, req->r_dentry);
 803                                 is_hash = true;
 804                         } else {
 805                                 ihold(inode);
 806                         }
 807                 }
 808                 rcu_read_unlock();
 809         }
 810
 811         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 812              (int)hash, mode);
 813         if (!inode)
 814                 goto random;
 815         ci = ceph_inode(inode);
 816
 817         if (is_hash && S_ISDIR(inode->i_mode)) {
 818                 struct ceph_inode_frag frag;
 819                 int found;
 820
 821                 ceph_choose_frag(ci, hash, &frag, &found);
 822                 if (found) {
 823                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 824                                 u8 r;
 825
 826                                 /* choose a random replica */
 827                                 get_random_bytes(&r, 1);
 828                                 r %= frag.ndist;
 829                                 mds = frag.dist[r];
 830                                 dout("choose_mds %p %llx.%llx "
 831                                      "frag %u mds%d (%d/%d)\n",
 832                                      inode, ceph_vinop(inode),
 833                                      frag.frag, mds,
 834                                      (int)r, frag.ndist);
 835                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 836                                     CEPH_MDS_STATE_ACTIVE)
 837                                         goto out;
 838                         }
 839
 840                         /* since this file/dir wasn't known to be
 841                          * replicated, then we want to look for the
 842                          * authoritative mds. */
 843                         mode = USE_AUTH_MDS;
 844                         if (frag.mds >= 0) {
 845                                 /* choose auth mds */
 846                                 mds = frag.mds;
 847                                 dout("choose_mds %p %llx.%llx "
 848                                      "frag %u mds%d (auth)\n",
 849                                      inode, ceph_vinop(inode), frag.frag, mds);
 850                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 851                                     CEPH_MDS_STATE_ACTIVE)
 852                                         goto out;
 853                         }
 854                 }
 855         }
 856
 857         spin_lock(&ci->i_ceph_lock);
 858         cap = NULL;
 859         if (mode == USE_AUTH_MDS)
 860                 cap = ci->i_auth_cap;
 861         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
 862                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
 863         if (!cap) {
 864                 spin_unlock(&ci->i_ceph_lock);
 865                 iput(inode);
 866                 goto random;
 867         }
 868         mds = cap->session->s_mds;
 869         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
 870              inode, ceph_vinop(inode), mds,
 871              cap == ci->i_auth_cap ? "auth " : "", cap);
 872         spin_unlock(&ci->i_ceph_lock);
 873 out:
 874         iput(inode);
 875         return mds;
 876
 877 random:
 878         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
 879         dout("choose_mds chose random mds%d\n", mds);
 880         return mds;
 881 }
 882
 883
 884 /*
 885  * session messages
 886  */
 887 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
 888 {
 889         struct ceph_msg *msg;
 890         struct ceph_mds_session_head *h;
 891
 892         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
 893                            false);
 894         if (!msg) {
 895                 pr_err("create_session_msg ENOMEM creating msg\n");
 896                 return NULL;
 897         }
 898         h = msg->front.iov_base;
 899         h->op = cpu_to_le32(op);
 900         h->seq = cpu_to_le64(seq);
 901
 902         return msg;
 903 }
 904
 905 static void encode_supported_features(void **p, void *end)
 906 {
 907         static const unsigned char bits[] = CEPHFS_FEATURES_CLIENT_SUPPORTED;
 908         static const size_t count = ARRAY_SIZE(bits);
 909
 910         if (count > 0) {
 911                 size_t i;
 912                 size_t size = ((size_t)bits[count - 1] + 64) / 64 * 8;
 913
 914                 BUG_ON(*p + 4 + size > end);
 915                 ceph_encode_32(p, size);
 916                 memset(*p, 0, size);
 917                 for (i = 0; i < count; i++)
 918                         ((unsigned char*)(*p))[i / 8] |= 1 << (bits[i] % 8);
 919                 *p += size;
 920         } else {
 921                 BUG_ON(*p + 4 > end);
 922                 ceph_encode_32(p, 0);
 923         }
 924 }
 925
 926 /*
 927  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
 928  * to include additional client metadata fields.
 929  */
 930 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
 931 {
 932         struct ceph_msg *msg;
 933         struct ceph_mds_session_head *h;
 934         int i = -1;
 935         int extra_bytes = 0;
 936         int metadata_key_count = 0;
 937         struct ceph_options *opt = mdsc->fsc->client->options;
 938         struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
 939         void *p, *end;
 940
 941         const char* metadata[][2] = {
 942                 {"hostname", mdsc->nodename},
 943                 {"kernel_version", init_utsname()->release},
 944                 {"entity_id", opt->name ? : ""},
 945                 {"root", fsopt->server_path ? : "/"},
 946                 {NULL, NULL}
 947         };
 948
 949         /* Calculate serialized length of metadata */
 950         extra_bytes = 4;  /* map length */
 951         for (i = 0; metadata[i][0]; ++i) {
 952                 extra_bytes += 8 + strlen(metadata[i][0]) +
 953                         strlen(metadata[i][1]);
 954                 metadata_key_count++;
 955         }
 956         /* supported feature */
 957         extra_bytes += 4 + 8;
 958
 959         /* Allocate the message */
 960         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + extra_bytes,
 961                            GFP_NOFS, false);
 962         if (!msg) {
 963                 pr_err("create_session_msg ENOMEM creating msg\n");
 964                 return NULL;
 965         }
 966         p = msg->front.iov_base;
 967         end = p + msg->front.iov_len;
 968
 969         h = p;
 970         h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
 971         h->seq = cpu_to_le64(seq);
 972
 973         /*
 974          * Serialize client metadata into waiting buffer space, using
 975          * the format that userspace expects for map<string, string>
 976          *
 977          * ClientSession messages with metadata are v2
 978          */
 979         msg->hdr.version = cpu_to_le16(3);
 980         msg->hdr.compat_version = cpu_to_le16(1);
 981
 982         /* The write pointer, following the session_head structure */
 983         p += sizeof(*h);
 984
 985         /* Number of entries in the map */
 986         ceph_encode_32(&p, metadata_key_count);
 987
 988         /* Two length-prefixed strings for each entry in the map */
 989         for (i = 0; metadata[i][0]; ++i) {
 990                 size_t const key_len = strlen(metadata[i][0]);
 991                 size_t const val_len = strlen(metadata[i][1]);
 992
 993                 ceph_encode_32(&p, key_len);
 994                 memcpy(p, metadata[i][0], key_len);
 995                 p += key_len;
 996                 ceph_encode_32(&p, val_len);
 997                 memcpy(p, metadata[i][1], val_len);
 998                 p += val_len;
 999         }
1000
1001         encode_supported_features(&p, end);
1002         msg->front.iov_len = p - msg->front.iov_base;
1003         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1004
1005         return msg;
1006 }
1007
1008 /*
1009  * send session open request.
1010  *
1011  * called under mdsc->mutex
1012  */
1013 static int __open_session(struct ceph_mds_client *mdsc,
1014                           struct ceph_mds_session *session)
1015 {
1016         struct ceph_msg *msg;
1017         int mstate;
1018         int mds = session->s_mds;
1019
1020         /* wait for mds to go active? */
1021         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
1022         dout("open_session to mds%d (%s)\n", mds,
1023              ceph_mds_state_name(mstate));
1024         session->s_state = CEPH_MDS_SESSION_OPENING;
1025         session->s_renew_requested = jiffies;
1026
1027         /* send connect message */
1028         msg = create_session_open_msg(mdsc, session->s_seq);
1029         if (!msg)
1030                 return -ENOMEM;
1031         ceph_con_send(&session->s_con, msg);
1032         return 0;
1033 }
1034
1035 /*
1036  * open sessions for any export targets for the given mds
1037  *
1038  * called under mdsc->mutex
1039  */
1040 static struct ceph_mds_session *
1041 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
1042 {
1043         struct ceph_mds_session *session;
1044
1045         session = __ceph_lookup_mds_session(mdsc, target);
1046         if (!session) {
1047                 session = register_session(mdsc, target);
1048                 if (IS_ERR(session))
1049                         return session;
1050         }
1051         if (session->s_state == CEPH_MDS_SESSION_NEW ||
1052             session->s_state == CEPH_MDS_SESSION_CLOSING)
1053                 __open_session(mdsc, session);
1054
1055         return session;
1056 }
1057
1058 struct ceph_mds_session *
1059 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1060 {
1061         struct ceph_mds_session *session;
1062
1063         dout("open_export_target_session to mds%d\n", target);
1064
1065         mutex_lock(&mdsc->mutex);
1066         session = __open_export_target_session(mdsc, target);
1067         mutex_unlock(&mdsc->mutex);
1068
1069         return session;
1070 }
1071
1072 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1073                                           struct ceph_mds_session *session)
1074 {
1075         struct ceph_mds_info *mi;
1076         struct ceph_mds_session *ts;
1077         int i, mds = session->s_mds;
1078
1079         if (mds >= mdsc->mdsmap->m_num_mds)
1080                 return;
1081
1082         mi = &mdsc->mdsmap->m_info[mds];
1083         dout("open_export_target_sessions for mds%d (%d targets)\n",
1084              session->s_mds, mi->num_export_targets);
1085
1086         for (i = 0; i < mi->num_export_targets; i++) {
1087                 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1088                 if (!IS_ERR(ts))
1089                         ceph_put_mds_session(ts);
1090         }
1091 }
1092
1093 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1094                                            struct ceph_mds_session *session)
1095 {
1096         mutex_lock(&mdsc->mutex);
1097         __open_export_target_sessions(mdsc, session);
1098         mutex_unlock(&mdsc->mutex);
1099 }
1100
1101 /*
1102  * session caps
1103  */
1104
1105 static void detach_cap_releases(struct ceph_mds_session *session,
1106                                 struct list_head *target)
1107 {
1108         lockdep_assert_held(&session->s_cap_lock);
1109
1110         list_splice_init(&session->s_cap_releases, target);
1111         session->s_num_cap_releases = 0;
1112         dout("dispose_cap_releases mds%d\n", session->s_mds);
1113 }
1114
1115 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1116                                  struct list_head *dispose)
1117 {
1118         while (!list_empty(dispose)) {
1119                 struct ceph_cap *cap;
1120                 /* zero out the in-progress message */
1121                 cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1122                 list_del(&cap->session_caps);
1123                 ceph_put_cap(mdsc, cap);
1124         }
1125 }
1126
1127 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1128                                      struct ceph_mds_session *session)
1129 {
1130         struct ceph_mds_request *req;
1131         struct rb_node *p;
1132
1133         dout("cleanup_session_requests mds%d\n", session->s_mds);
1134         mutex_lock(&mdsc->mutex);
1135         while (!list_empty(&session->s_unsafe)) {
1136                 req = list_first_entry(&session->s_unsafe,
1137                                        struct ceph_mds_request, r_unsafe_item);
1138                 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1139                                     req->r_tid);
1140                 __unregister_request(mdsc, req);
1141         }
1142         /* zero r_attempts, so kick_requests() will re-send requests */
1143         p = rb_first(&mdsc->request_tree);
1144         while (p) {
1145                 req = rb_entry(p, struct ceph_mds_request, r_node);
1146                 p = rb_next(p);
1147                 if (req->r_session &&
1148                     req->r_session->s_mds == session->s_mds)
1149                         req->r_attempts = 0;
1150         }
1151         mutex_unlock(&mdsc->mutex);
1152 }
1153
1154 /*
1155  * Helper to safely iterate over all caps associated with a session, with
1156  * special care taken to handle a racing __ceph_remove_cap().
1157  *
1158  * Caller must hold session s_mutex.
1159  */
1160 static int iterate_session_caps(struct ceph_mds_session *session,
1161                                  int (*cb)(struct inode *, struct ceph_cap *,
1162                                             void *), void *arg)
1163 {
1164         struct list_head *p;
1165         struct ceph_cap *cap;
1166         struct inode *inode, *last_inode = NULL;
1167         struct ceph_cap *old_cap = NULL;
1168         int ret;
1169
1170         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1171         spin_lock(&session->s_cap_lock);
1172         p = session->s_caps.next;
1173         while (p != &session->s_caps) {
1174                 cap = list_entry(p, struct ceph_cap, session_caps);
1175                 inode = igrab(&cap->ci->vfs_inode);
1176                 if (!inode) {
1177                         p = p->next;
1178                         continue;
1179                 }
1180                 session->s_cap_iterator = cap;
1181                 spin_unlock(&session->s_cap_lock);
1182
1183                 if (last_inode) {
1184                         iput(last_inode);
1185                         last_inode = NULL;
1186                 }
1187                 if (old_cap) {
1188                         ceph_put_cap(session->s_mdsc, old_cap);
1189                         old_cap = NULL;
1190                 }
1191
1192                 ret = cb(inode, cap, arg);
1193                 last_inode = inode;
1194
1195                 spin_lock(&session->s_cap_lock);
1196                 p = p->next;
1197                 if (!cap->ci) {
1198                         dout("iterate_session_caps  finishing cap %p removal\n",
1199                              cap);
1200                         BUG_ON(cap->session != session);
1201                         cap->session = NULL;
1202                         list_del_init(&cap->session_caps);
1203                         session->s_nr_caps--;
1204                         if (cap->queue_release) {
1205                                 list_add_tail(&cap->session_caps,
1206                                               &session->s_cap_releases);
1207                                 session->s_num_cap_releases++;
1208                         } else {
1209                                 old_cap = cap;  /* put_cap it w/o locks held */
1210                         }
1211                 }
1212                 if (ret < 0)
1213                         goto out;
1214         }
1215         ret = 0;
1216 out:
1217         session->s_cap_iterator = NULL;
1218         spin_unlock(&session->s_cap_lock);
1219
1220         iput(last_inode);
1221         if (old_cap)
1222                 ceph_put_cap(session->s_mdsc, old_cap);
1223
1224         return ret;
1225 }
1226
1227 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1228                                   void *arg)
1229 {
1230         struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1231         struct ceph_inode_info *ci = ceph_inode(inode);
1232         LIST_HEAD(to_remove);
1233         bool drop = false;
1234         bool invalidate = false;
1235
1236         dout("removing cap %p, ci is %p, inode is %p\n",
1237              cap, ci, &ci->vfs_inode);
1238         spin_lock(&ci->i_ceph_lock);
1239         __ceph_remove_cap(cap, false);
1240         if (!ci->i_auth_cap) {
1241                 struct ceph_cap_flush *cf;
1242                 struct ceph_mds_client *mdsc = fsc->mdsc;
1243
1244                 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1245
1246                 if (ci->i_wrbuffer_ref > 0 &&
1247                     READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
1248                         invalidate = true;
1249
1250                 while (!list_empty(&ci->i_cap_flush_list)) {
1251                         cf = list_first_entry(&ci->i_cap_flush_list,
1252                                               struct ceph_cap_flush, i_list);
1253                         list_move(&cf->i_list, &to_remove);
1254                 }
1255
1256                 spin_lock(&mdsc->cap_dirty_lock);
1257
1258                 list_for_each_entry(cf, &to_remove, i_list)
1259                         list_del(&cf->g_list);
1260
1261                 if (!list_empty(&ci->i_dirty_item)) {
1262                         pr_warn_ratelimited(
1263                                 " dropping dirty %s state for %p %lld\n",
1264                                 ceph_cap_string(ci->i_dirty_caps),
1265                                 inode, ceph_ino(inode));
1266                         ci->i_dirty_caps = 0;
1267                         list_del_init(&ci->i_dirty_item);
1268                         drop = true;
1269                 }
1270                 if (!list_empty(&ci->i_flushing_item)) {
1271                         pr_warn_ratelimited(
1272                                 " dropping dirty+flushing %s state for %p %lld\n",
1273                                 ceph_cap_string(ci->i_flushing_caps),
1274                                 inode, ceph_ino(inode));
1275                         ci->i_flushing_caps = 0;
1276                         list_del_init(&ci->i_flushing_item);
1277                         mdsc->num_cap_flushing--;
1278                         drop = true;
1279                 }
1280                 spin_unlock(&mdsc->cap_dirty_lock);
1281
1282                 if (atomic_read(&ci->i_filelock_ref) > 0) {
1283                         /* make further file lock syscall return -EIO */
1284                         ci->i_ceph_flags |= CEPH_I_ERROR_FILELOCK;
1285                         pr_warn_ratelimited(" dropping file locks for %p %lld\n",
1286                                             inode, ceph_ino(inode));
1287                 }
1288
1289                 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1290                         list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1291                         ci->i_prealloc_cap_flush = NULL;
1292                 }
1293
1294                if (drop &&
1295                   ci->i_wrbuffer_ref_head == 0 &&
1296                   ci->i_wr_ref == 0 &&
1297                   ci->i_dirty_caps == 0 &&
1298                   ci->i_flushing_caps == 0) {
1299                       ceph_put_snap_context(ci->i_head_snapc);
1300                       ci->i_head_snapc = NULL;
1301                }
1302         }
1303         spin_unlock(&ci->i_ceph_lock);
1304         while (!list_empty(&to_remove)) {
1305                 struct ceph_cap_flush *cf;
1306                 cf = list_first_entry(&to_remove,
1307                                       struct ceph_cap_flush, i_list);
1308                 list_del(&cf->i_list);
1309                 ceph_free_cap_flush(cf);
1310         }
1311
1312         wake_up_all(&ci->i_cap_wq);
1313         if (invalidate)
1314                 ceph_queue_invalidate(inode);
1315         if (drop)
1316                 iput(inode);
1317         return 0;
1318 }
1319
1320 /*
1321  * caller must hold session s_mutex
1322  */
1323 static void remove_session_caps(struct ceph_mds_session *session)
1324 {
1325         struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1326         struct super_block *sb = fsc->sb;
1327         LIST_HEAD(dispose);
1328
1329         dout("remove_session_caps on %p\n", session);
1330         iterate_session_caps(session, remove_session_caps_cb, fsc);
1331
1332         wake_up_all(&fsc->mdsc->cap_flushing_wq);
1333
1334         spin_lock(&session->s_cap_lock);
1335         if (session->s_nr_caps > 0) {
1336                 struct inode *inode;
1337                 struct ceph_cap *cap, *prev = NULL;
1338                 struct ceph_vino vino;
1339                 /*
1340                  * iterate_session_caps() skips inodes that are being
1341                  * deleted, we need to wait until deletions are complete.
1342                  * __wait_on_freeing_inode() is designed for the job,
1343                  * but it is not exported, so use lookup inode function
1344                  * to access it.
1345                  */
1346                 while (!list_empty(&session->s_caps)) {
1347                         cap = list_entry(session->s_caps.next,
1348                                          struct ceph_cap, session_caps);
1349                         if (cap == prev)
1350                                 break;
1351                         prev = cap;
1352                         vino = cap->ci->i_vino;
1353                         spin_unlock(&session->s_cap_lock);
1354
1355                         inode = ceph_find_inode(sb, vino);
1356                         iput(inode);
1357
1358                         spin_lock(&session->s_cap_lock);
1359                 }
1360         }
1361
1362         // drop cap expires and unlock s_cap_lock
1363         detach_cap_releases(session, &dispose);
1364
1365         BUG_ON(session->s_nr_caps > 0);
1366         BUG_ON(!list_empty(&session->s_cap_flushing));
1367         spin_unlock(&session->s_cap_lock);
1368         dispose_cap_releases(session->s_mdsc, &dispose);
1369 }
1370
1371 /*
1372  * wake up any threads waiting on this session's caps.  if the cap is
1373  * old (didn't get renewed on the client reconnect), remove it now.
1374  *
1375  * caller must hold s_mutex.
1376  */
1377 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1378                               void *arg)
1379 {
1380         struct ceph_inode_info *ci = ceph_inode(inode);
1381
1382         if (arg) {
1383                 spin_lock(&ci->i_ceph_lock);
1384                 ci->i_wanted_max_size = 0;
1385                 ci->i_requested_max_size = 0;
1386                 spin_unlock(&ci->i_ceph_lock);
1387         }
1388         wake_up_all(&ci->i_cap_wq);
1389         return 0;
1390 }
1391
1392 static void wake_up_session_caps(struct ceph_mds_session *session,
1393                                  int reconnect)
1394 {
1395         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1396         iterate_session_caps(session, wake_up_session_cb,
1397                              (void *)(unsigned long)reconnect);
1398 }
1399
1400 /*
1401  * Send periodic message to MDS renewing all currently held caps.  The
1402  * ack will reset the expiration for all caps from this session.
1403  *
1404  * caller holds s_mutex
1405  */
1406 static int send_renew_caps(struct ceph_mds_client *mdsc,
1407                            struct ceph_mds_session *session)
1408 {
1409         struct ceph_msg *msg;
1410         int state;
1411
1412         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1413             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1414                 pr_info("mds%d caps stale\n", session->s_mds);
1415         session->s_renew_requested = jiffies;
1416
1417         /* do not try to renew caps until a recovering mds has reconnected
1418          * with its clients. */
1419         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1420         if (state < CEPH_MDS_STATE_RECONNECT) {
1421                 dout("send_renew_caps ignoring mds%d (%s)\n",
1422                      session->s_mds, ceph_mds_state_name(state));
1423                 return 0;
1424         }
1425
1426         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1427                 ceph_mds_state_name(state));
1428         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1429                                  ++session->s_renew_seq);
1430         if (!msg)
1431                 return -ENOMEM;
1432         ceph_con_send(&session->s_con, msg);
1433         return 0;
1434 }
1435
1436 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1437                              struct ceph_mds_session *session, u64 seq)
1438 {
1439         struct ceph_msg *msg;
1440
1441         dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1442              session->s_mds, ceph_session_state_name(session->s_state), seq);
1443         msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1444         if (!msg)
1445                 return -ENOMEM;
1446         ceph_con_send(&session->s_con, msg);
1447         return 0;
1448 }
1449
1450
1451 /*
1452  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1453  *
1454  * Called under session->s_mutex
1455  */
1456 static void renewed_caps(struct ceph_mds_client *mdsc,
1457                          struct ceph_mds_session *session, int is_renew)
1458 {
1459         int was_stale;
1460         int wake = 0;
1461
1462         spin_lock(&session->s_cap_lock);
1463         was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1464
1465         session->s_cap_ttl = session->s_renew_requested +
1466                 mdsc->mdsmap->m_session_timeout*HZ;
1467
1468         if (was_stale) {
1469                 if (time_before(jiffies, session->s_cap_ttl)) {
1470                         pr_info("mds%d caps renewed\n", session->s_mds);
1471                         wake = 1;
1472                 } else {
1473                         pr_info("mds%d caps still stale\n", session->s_mds);
1474                 }
1475         }
1476         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1477              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1478              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1479         spin_unlock(&session->s_cap_lock);
1480
1481         if (wake)
1482                 wake_up_session_caps(session, 0);
1483 }
1484
1485 /*
1486  * send a session close request
1487  */
1488 static int request_close_session(struct ceph_mds_client *mdsc,
1489                                  struct ceph_mds_session *session)
1490 {
1491         struct ceph_msg *msg;
1492
1493         dout("request_close_session mds%d state %s seq %lld\n",
1494              session->s_mds, ceph_session_state_name(session->s_state),
1495              session->s_seq);
1496         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1497         if (!msg)
1498                 return -ENOMEM;
1499         ceph_con_send(&session->s_con, msg);
1500         return 1;
1501 }
1502
1503 /*
1504  * Called with s_mutex held.
1505  */
1506 static int __close_session(struct ceph_mds_client *mdsc,
1507                          struct ceph_mds_session *session)
1508 {
1509         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1510                 return 0;
1511         session->s_state = CEPH_MDS_SESSION_CLOSING;
1512         return request_close_session(mdsc, session);
1513 }
1514
1515 static bool drop_negative_children(struct dentry *dentry)
1516 {
1517         struct dentry *child;
1518         bool all_negative = true;
1519
1520         if (!d_is_dir(dentry))
1521                 goto out;
1522
1523         spin_lock(&dentry->d_lock);
1524         list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1525                 if (d_really_is_positive(child)) {
1526                         all_negative = false;
1527                         break;
1528                 }
1529         }
1530         spin_unlock(&dentry->d_lock);
1531
1532         if (all_negative)
1533                 shrink_dcache_parent(dentry);
1534 out:
1535         return all_negative;
1536 }
1537
1538 /*
1539  * Trim old(er) caps.
1540  *
1541  * Because we can't cache an inode without one or more caps, we do
1542  * this indirectly: if a cap is unused, we prune its aliases, at which
1543  * point the inode will hopefully get dropped to.
1544  *
1545  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1546  * memory pressure from the MDS, though, so it needn't be perfect.
1547  */
1548 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1549 {
1550         struct ceph_mds_session *session = arg;
1551         struct ceph_inode_info *ci = ceph_inode(inode);
1552         int used, wanted, oissued, mine;
1553
1554         if (session->s_trim_caps <= 0)
1555                 return -1;
1556
1557         spin_lock(&ci->i_ceph_lock);
1558         mine = cap->issued | cap->implemented;
1559         used = __ceph_caps_used(ci);
1560         wanted = __ceph_caps_file_wanted(ci);
1561         oissued = __ceph_caps_issued_other(ci, cap);
1562
1563         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1564              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1565              ceph_cap_string(used), ceph_cap_string(wanted));
1566         if (cap == ci->i_auth_cap) {
1567                 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1568                     !list_empty(&ci->i_cap_snaps))
1569                         goto out;
1570                 if ((used | wanted) & CEPH_CAP_ANY_WR)
1571                         goto out;
1572                 /* Note: it's possible that i_filelock_ref becomes non-zero
1573                  * after dropping auth caps. It doesn't hurt because reply
1574                  * of lock mds request will re-add auth caps. */
1575                 if (atomic_read(&ci->i_filelock_ref) > 0)
1576                         goto out;
1577         }
1578         /* The inode has cached pages, but it's no longer used.
1579          * we can safely drop it */
1580         if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1581             !(oissued & CEPH_CAP_FILE_CACHE)) {
1582           used = 0;
1583           oissued = 0;
1584         }
1585         if ((used | wanted) & ~oissued & mine)
1586                 goto out;   /* we need these caps */
1587
1588         if (oissued) {
1589                 /* we aren't the only cap.. just remove us */
1590                 __ceph_remove_cap(cap, true);
1591                 session->s_trim_caps--;
1592         } else {
1593                 struct dentry *dentry;
1594                 /* try dropping referring dentries */
1595                 spin_unlock(&ci->i_ceph_lock);
1596                 dentry = d_find_any_alias(inode);
1597                 if (dentry && drop_negative_children(dentry)) {
1598                         int count;
1599                         dput(dentry);
1600                         d_prune_aliases(inode);
1601                         count = atomic_read(&inode->i_count);
1602                         if (count == 1)
1603                                 session->s_trim_caps--;
1604                         dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1605                              inode, cap, count);
1606                 } else {
1607                         dput(dentry);
1608                 }
1609                 return 0;
1610         }
1611
1612 out:
1613         spin_unlock(&ci->i_ceph_lock);
1614         return 0;
1615 }
1616
1617 /*
1618  * Trim session cap count down to some max number.
1619  */
1620 int ceph_trim_caps(struct ceph_mds_client *mdsc,
1621                    struct ceph_mds_session *session,
1622                    int max_caps)
1623 {
1624         int trim_caps = session->s_nr_caps - max_caps;
1625
1626         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1627              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1628         if (trim_caps > 0) {
1629                 session->s_trim_caps = trim_caps;
1630                 iterate_session_caps(session, trim_caps_cb, session);
1631                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1632                      session->s_mds, session->s_nr_caps, max_caps,
1633                         trim_caps - session->s_trim_caps);
1634                 session->s_trim_caps = 0;
1635         }
1636
1637         ceph_send_cap_releases(mdsc, session);
1638         return 0;
1639 }
1640
1641 static int check_caps_flush(struct ceph_mds_client *mdsc,
1642                             u64 want_flush_tid)
1643 {
1644         int ret = 1;
1645
1646         spin_lock(&mdsc->cap_dirty_lock);
1647         if (!list_empty(&mdsc->cap_flush_list)) {
1648                 struct ceph_cap_flush *cf =
1649                         list_first_entry(&mdsc->cap_flush_list,
1650                                          struct ceph_cap_flush, g_list);
1651                 if (cf->tid <= want_flush_tid) {
1652                         dout("check_caps_flush still flushing tid "
1653                              "%llu <= %llu\n", cf->tid, want_flush_tid);
1654                         ret = 0;
1655                 }
1656         }
1657         spin_unlock(&mdsc->cap_dirty_lock);
1658         return ret;
1659 }
1660
1661 /*
1662  * flush all dirty inode data to disk.
1663  *
1664  * returns true if we've flushed through want_flush_tid
1665  */
1666 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1667                             u64 want_flush_tid)
1668 {
1669         dout("check_caps_flush want %llu\n", want_flush_tid);
1670
1671         wait_event(mdsc->cap_flushing_wq,
1672                    check_caps_flush(mdsc, want_flush_tid));
1673
1674         dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1675 }
1676
1677 /*
1678  * called under s_mutex
1679  */
1680 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1681                             struct ceph_mds_session *session)
1682 {
1683         struct ceph_msg *msg = NULL;
1684         struct ceph_mds_cap_release *head;
1685         struct ceph_mds_cap_item *item;
1686         struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
1687         struct ceph_cap *cap;
1688         LIST_HEAD(tmp_list);
1689         int num_cap_releases;
1690         __le32  barrier, *cap_barrier;
1691
1692         down_read(&osdc->lock);
1693         barrier = cpu_to_le32(osdc->epoch_barrier);
1694         up_read(&osdc->lock);
1695
1696         spin_lock(&session->s_cap_lock);
1697 again:
1698         list_splice_init(&session->s_cap_releases, &tmp_list);
1699         num_cap_releases = session->s_num_cap_releases;
1700         session->s_num_cap_releases = 0;
1701         spin_unlock(&session->s_cap_lock);
1702
1703         while (!list_empty(&tmp_list)) {
1704                 if (!msg) {
1705                         msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1706                                         PAGE_SIZE, GFP_NOFS, false);
1707                         if (!msg)
1708                                 goto out_err;
1709                         head = msg->front.iov_base;
1710                         head->num = cpu_to_le32(0);
1711                         msg->front.iov_len = sizeof(*head);
1712
1713                         msg->hdr.version = cpu_to_le16(2);
1714                         msg->hdr.compat_version = cpu_to_le16(1);
1715                 }
1716
1717                 cap = list_first_entry(&tmp_list, struct ceph_cap,
1718                                         session_caps);
1719                 list_del(&cap->session_caps);
1720                 num_cap_releases--;
1721
1722                 head = msg->front.iov_base;
1723                 le32_add_cpu(&head->num, 1);
1724                 item = msg->front.iov_base + msg->front.iov_len;
1725                 item->ino = cpu_to_le64(cap->cap_ino);
1726                 item->cap_id = cpu_to_le64(cap->cap_id);
1727                 item->migrate_seq = cpu_to_le32(cap->mseq);
1728                 item->seq = cpu_to_le32(cap->issue_seq);
1729                 msg->front.iov_len += sizeof(*item);
1730
1731                 ceph_put_cap(mdsc, cap);
1732
1733                 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1734                         // Append cap_barrier field
1735                         cap_barrier = msg->front.iov_base + msg->front.iov_len;
1736                         *cap_barrier = barrier;
1737                         msg->front.iov_len += sizeof(*cap_barrier);
1738
1739                         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1740                         dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1741                         ceph_con_send(&session->s_con, msg);
1742                         msg = NULL;
1743                 }
1744         }
1745
1746         BUG_ON(num_cap_releases != 0);
1747
1748         spin_lock(&session->s_cap_lock);
1749         if (!list_empty(&session->s_cap_releases))
1750                 goto again;
1751         spin_unlock(&session->s_cap_lock);
1752
1753         if (msg) {
1754                 // Append cap_barrier field
1755                 cap_barrier = msg->front.iov_base + msg->front.iov_len;
1756                 *cap_barrier = barrier;
1757                 msg->front.iov_len += sizeof(*cap_barrier);
1758
1759                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1760                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1761                 ceph_con_send(&session->s_con, msg);
1762         }
1763         return;
1764 out_err:
1765         pr_err("send_cap_releases mds%d, failed to allocate message\n",
1766                 session->s_mds);
1767         spin_lock(&session->s_cap_lock);
1768         list_splice(&tmp_list, &session->s_cap_releases);
1769         session->s_num_cap_releases += num_cap_releases;
1770         spin_unlock(&session->s_cap_lock);
1771 }
1772
1773 /*
1774  * requests
1775  */
1776
1777 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1778                                     struct inode *dir)
1779 {
1780         struct ceph_inode_info *ci = ceph_inode(dir);
1781         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1782         struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1783         size_t size = sizeof(struct ceph_mds_reply_dir_entry);
1784         int order, num_entries;
1785
1786         spin_lock(&ci->i_ceph_lock);
1787         num_entries = ci->i_files + ci->i_subdirs;
1788         spin_unlock(&ci->i_ceph_lock);
1789         num_entries = max(num_entries, 1);
1790         num_entries = min(num_entries, opt->max_readdir);
1791
1792         order = get_order(size * num_entries);
1793         while (order >= 0) {
1794                 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
1795                                                              __GFP_NOWARN,
1796                                                              order);
1797                 if (rinfo->dir_entries)
1798                         break;
1799                 order--;
1800         }
1801         if (!rinfo->dir_entries)
1802                 return -ENOMEM;
1803
1804         num_entries = (PAGE_SIZE << order) / size;
1805         num_entries = min(num_entries, opt->max_readdir);
1806
1807         rinfo->dir_buf_size = PAGE_SIZE << order;
1808         req->r_num_caps = num_entries + 1;
1809         req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1810         req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1811         return 0;
1812 }
1813
1814 /*
1815  * Create an mds request.
1816  */
1817 struct ceph_mds_request *
1818 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1819 {
1820         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1821         struct timespec64 ts;
1822
1823         if (!req)
1824                 return ERR_PTR(-ENOMEM);
1825
1826         mutex_init(&req->r_fill_mutex);
1827         req->r_mdsc = mdsc;
1828         req->r_started = jiffies;
1829         req->r_resend_mds = -1;
1830         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1831         INIT_LIST_HEAD(&req->r_unsafe_target_item);
1832         req->r_fmode = -1;
1833         kref_init(&req->r_kref);
1834         RB_CLEAR_NODE(&req->r_node);
1835         INIT_LIST_HEAD(&req->r_wait);
1836         init_completion(&req->r_completion);
1837         init_completion(&req->r_safe_completion);
1838         INIT_LIST_HEAD(&req->r_unsafe_item);
1839
1840         ktime_get_coarse_real_ts64(&ts);
1841         req->r_stamp = timespec64_trunc(ts, mdsc->fsc->sb->s_time_gran);
1842
1843         req->r_op = op;
1844         req->r_direct_mode = mode;
1845         return req;
1846 }
1847
1848 /*
1849  * return oldest (lowest) request, tid in request tree, 0 if none.
1850  *
1851  * called under mdsc->mutex.
1852  */
1853 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1854 {
1855         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1856                 return NULL;
1857         return rb_entry(rb_first(&mdsc->request_tree),
1858                         struct ceph_mds_request, r_node);
1859 }
1860
1861 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1862 {
1863         return mdsc->oldest_tid;
1864 }
1865
1866 /*
1867  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1868  * on build_path_from_dentry in fs/cifs/dir.c.
1869  *
1870  * If @stop_on_nosnap, generate path relative to the first non-snapped
1871  * inode.
1872  *
1873  * Encode hidden .snap dirs as a double /, i.e.
1874  *   foo/.snap/bar -> foo//bar
1875  */
1876 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1877                            int stop_on_nosnap)
1878 {
1879         struct dentry *temp;
1880         char *path;
1881         int len, pos;
1882         unsigned seq;
1883
1884         if (!dentry)
1885                 return ERR_PTR(-EINVAL);
1886
1887 retry:
1888         len = 0;
1889         seq = read_seqbegin(&rename_lock);
1890         rcu_read_lock();
1891         for (temp = dentry; !IS_ROOT(temp);) {
1892                 struct inode *inode = d_inode(temp);
1893                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1894                         len++;  /* slash only */
1895                 else if (stop_on_nosnap && inode &&
1896                          ceph_snap(inode) == CEPH_NOSNAP)
1897                         break;
1898                 else
1899                         len += 1 + temp->d_name.len;
1900                 temp = temp->d_parent;
1901         }
1902         rcu_read_unlock();
1903         if (len)
1904                 len--;  /* no leading '/' */
1905
1906         path = kmalloc(len+1, GFP_NOFS);
1907         if (!path)
1908                 return ERR_PTR(-ENOMEM);
1909         pos = len;
1910         path[pos] = 0;  /* trailing null */
1911         rcu_read_lock();
1912         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1913                 struct inode *inode;
1914
1915                 spin_lock(&temp->d_lock);
1916                 inode = d_inode(temp);
1917                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1918                         dout("build_path path+%d: %p SNAPDIR\n",
1919                              pos, temp);
1920                 } else if (stop_on_nosnap && inode &&
1921                            ceph_snap(inode) == CEPH_NOSNAP) {
1922                         spin_unlock(&temp->d_lock);
1923                         break;
1924                 } else {
1925                         pos -= temp->d_name.len;
1926                         if (pos < 0) {
1927                                 spin_unlock(&temp->d_lock);
1928                                 break;
1929                         }
1930                         strncpy(path + pos, temp->d_name.name,
1931                                 temp->d_name.len);
1932                 }
1933                 spin_unlock(&temp->d_lock);
1934                 if (pos)
1935                         path[--pos] = '/';
1936                 temp = temp->d_parent;
1937         }
1938         rcu_read_unlock();
1939         if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1940                 pr_err("build_path did not end path lookup where "
1941                        "expected, namelen is %d, pos is %d\n", len, pos);
1942                 /* presumably this is only possible if racing with a
1943                    rename of one of the parent directories (we can not
1944                    lock the dentries above us to prevent this, but
1945                    retrying should be harmless) */
1946                 kfree(path);
1947                 goto retry;
1948         }
1949
1950         *base = ceph_ino(d_inode(temp));
1951         *plen = len;
1952         dout("build_path on %p %d built %llx '%.*s'\n",
1953              dentry, d_count(dentry), *base, len, path);
1954         return path;
1955 }
1956
1957 /* Duplicate the dentry->d_name.name safely */
1958 static int clone_dentry_name(struct dentry *dentry, const char **ppath,
1959                              int *ppathlen)
1960 {
1961         u32 len;
1962         char *name;
1963
1964 retry:
1965         len = READ_ONCE(dentry->d_name.len);
1966         name = kmalloc(len + 1, GFP_NOFS);
1967         if (!name)
1968                 return -ENOMEM;
1969
1970         spin_lock(&dentry->d_lock);
1971         if (dentry->d_name.len != len) {
1972                 spin_unlock(&dentry->d_lock);
1973                 kfree(name);
1974                 goto retry;
1975         }
1976         memcpy(name, dentry->d_name.name, len);
1977         spin_unlock(&dentry->d_lock);
1978
1979         name[len] = '\0';
1980         *ppath = name;
1981         *ppathlen = len;
1982         return 0;
1983 }
1984
1985 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
1986                              const char **ppath, int *ppathlen, u64 *pino,
1987                              bool *pfreepath, bool parent_locked)
1988 {
1989         int ret;
1990         char *path;
1991
1992         rcu_read_lock();
1993         if (!dir)
1994                 dir = d_inode_rcu(dentry->d_parent);
1995         if (dir && ceph_snap(dir) == CEPH_NOSNAP) {
1996                 *pino = ceph_ino(dir);
1997                 rcu_read_unlock();
1998                 if (parent_locked) {
1999                         *ppath = dentry->d_name.name;
2000                         *ppathlen = dentry->d_name.len;
2001                 } else {
2002                         ret = clone_dentry_name(dentry, ppath, ppathlen);
2003                         if (ret)
2004                                 return ret;
2005                         *pfreepath = true;
2006                 }
2007                 return 0;
2008         }
2009         rcu_read_unlock();
2010         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2011         if (IS_ERR(path))
2012                 return PTR_ERR(path);
2013         *ppath = path;
2014         *pfreepath = true;
2015         return 0;
2016 }
2017
2018 static int build_inode_path(struct inode *inode,
2019                             const char **ppath, int *ppathlen, u64 *pino,
2020                             bool *pfreepath)
2021 {
2022         struct dentry *dentry;
2023         char *path;
2024
2025         if (ceph_snap(inode) == CEPH_NOSNAP) {
2026                 *pino = ceph_ino(inode);
2027                 *ppathlen = 0;
2028                 return 0;
2029         }
2030         dentry = d_find_alias(inode);
2031         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2032         dput(dentry);
2033         if (IS_ERR(path))
2034                 return PTR_ERR(path);
2035         *ppath = path;
2036         *pfreepath = true;
2037         return 0;
2038 }
2039
2040 /*
2041  * request arguments may be specified via an inode *, a dentry *, or
2042  * an explicit ino+path.
2043  */
2044 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
2045                                   struct inode *rdiri, const char *rpath,
2046                                   u64 rino, const char **ppath, int *pathlen,
2047                                   u64 *ino, bool *freepath, bool parent_locked)
2048 {
2049         int r = 0;
2050
2051         if (rinode) {
2052                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2053                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2054                      ceph_snap(rinode));
2055         } else if (rdentry) {
2056                 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
2057                                         freepath, parent_locked);
2058                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
2059                      *ppath);
2060         } else if (rpath || rino) {
2061                 *ino = rino;
2062                 *ppath = rpath;
2063                 *pathlen = rpath ? strlen(rpath) : 0;
2064                 dout(" path %.*s\n", *pathlen, rpath);
2065         }
2066
2067         return r;
2068 }
2069
2070 /*
2071  * called under mdsc->mutex
2072  */
2073 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
2074                                                struct ceph_mds_request *req,
2075                                                int mds, bool drop_cap_releases)
2076 {
2077         struct ceph_msg *msg;
2078         struct ceph_mds_request_head *head;
2079         const char *path1 = NULL;
2080         const char *path2 = NULL;
2081         u64 ino1 = 0, ino2 = 0;
2082         int pathlen1 = 0, pathlen2 = 0;
2083         bool freepath1 = false, freepath2 = false;
2084         int len;
2085         u16 releases;
2086         void *p, *end;
2087         int ret;
2088
2089         ret = set_request_path_attr(req->r_inode, req->r_dentry,
2090                               req->r_parent, req->r_path1, req->r_ino1.ino,
2091                               &path1, &pathlen1, &ino1, &freepath1,
2092                               test_bit(CEPH_MDS_R_PARENT_LOCKED,
2093                                         &req->r_req_flags));
2094         if (ret < 0) {
2095                 msg = ERR_PTR(ret);
2096                 goto out;
2097         }
2098
2099         /* If r_old_dentry is set, then assume that its parent is locked */
2100         ret = set_request_path_attr(NULL, req->r_old_dentry,
2101                               req->r_old_dentry_dir,
2102                               req->r_path2, req->r_ino2.ino,
2103                               &path2, &pathlen2, &ino2, &freepath2, true);
2104         if (ret < 0) {
2105                 msg = ERR_PTR(ret);
2106                 goto out_free1;
2107         }
2108
2109         len = sizeof(*head) +
2110                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2111                 sizeof(struct ceph_timespec);
2112
2113         /* calculate (max) length for cap releases */
2114         len += sizeof(struct ceph_mds_request_release) *
2115                 (!!req->r_inode_drop + !!req->r_dentry_drop +
2116                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2117         if (req->r_dentry_drop)
2118                 len += req->r_dentry->d_name.len;
2119         if (req->r_old_dentry_drop)
2120                 len += req->r_old_dentry->d_name.len;
2121
2122         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
2123         if (!msg) {
2124                 msg = ERR_PTR(-ENOMEM);
2125                 goto out_free2;
2126         }
2127
2128         msg->hdr.version = cpu_to_le16(2);
2129         msg->hdr.tid = cpu_to_le64(req->r_tid);
2130
2131         head = msg->front.iov_base;
2132         p = msg->front.iov_base + sizeof(*head);
2133         end = msg->front.iov_base + msg->front.iov_len;
2134
2135         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2136         head->op = cpu_to_le32(req->r_op);
2137         head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
2138         head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
2139         head->args = req->r_args;
2140
2141         ceph_encode_filepath(&p, end, ino1, path1);
2142         ceph_encode_filepath(&p, end, ino2, path2);
2143
2144         /* make note of release offset, in case we need to replay */
2145         req->r_request_release_offset = p - msg->front.iov_base;
2146
2147         /* cap releases */
2148         releases = 0;
2149         if (req->r_inode_drop)
2150                 releases += ceph_encode_inode_release(&p,
2151                       req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2152                       mds, req->r_inode_drop, req->r_inode_unless, 0);
2153         if (req->r_dentry_drop)
2154                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
2155                                 req->r_parent, mds, req->r_dentry_drop,
2156                                 req->r_dentry_unless);
2157         if (req->r_old_dentry_drop)
2158                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2159                                 req->r_old_dentry_dir, mds,
2160                                 req->r_old_dentry_drop,
2161                                 req->r_old_dentry_unless);
2162         if (req->r_old_inode_drop)
2163                 releases += ceph_encode_inode_release(&p,
2164                       d_inode(req->r_old_dentry),
2165                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2166
2167         if (drop_cap_releases) {
2168                 releases = 0;
2169                 p = msg->front.iov_base + req->r_request_release_offset;
2170         }
2171
2172         head->num_releases = cpu_to_le16(releases);
2173
2174         /* time stamp */
2175         {
2176                 struct ceph_timespec ts;
2177                 ceph_encode_timespec64(&ts, &req->r_stamp);
2178                 ceph_encode_copy(&p, &ts, sizeof(ts));
2179         }
2180
2181         BUG_ON(p > end);
2182         msg->front.iov_len = p - msg->front.iov_base;
2183         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2184
2185         if (req->r_pagelist) {
2186                 struct ceph_pagelist *pagelist = req->r_pagelist;
2187                 refcount_inc(&pagelist->refcnt);
2188                 ceph_msg_data_add_pagelist(msg, pagelist);
2189                 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2190         } else {
2191                 msg->hdr.data_len = 0;
2192         }
2193
2194         msg->hdr.data_off = cpu_to_le16(0);
2195
2196 out_free2:
2197         if (freepath2)
2198                 kfree((char *)path2);
2199 out_free1:
2200         if (freepath1)
2201                 kfree((char *)path1);
2202 out:
2203         return msg;
2204 }
2205
2206 /*
2207  * called under mdsc->mutex if error, under no mutex if
2208  * success.
2209  */
2210 static void complete_request(struct ceph_mds_client *mdsc,
2211                              struct ceph_mds_request *req)
2212 {
2213         if (req->r_callback)
2214                 req->r_callback(mdsc, req);
2215         else
2216                 complete_all(&req->r_completion);
2217 }
2218
2219 /*
2220  * called under mdsc->mutex
2221  */
2222 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2223                                   struct ceph_mds_request *req,
2224                                   int mds, bool drop_cap_releases)
2225 {
2226         struct ceph_mds_request_head *rhead;
2227         struct ceph_msg *msg;
2228         int flags = 0;
2229
2230         req->r_attempts++;
2231         if (req->r_inode) {
2232                 struct ceph_cap *cap =
2233                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2234
2235                 if (cap)
2236                         req->r_sent_on_mseq = cap->mseq;
2237                 else
2238                         req->r_sent_on_mseq = -1;
2239         }
2240         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2241              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2242
2243         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2244                 void *p;
2245                 /*
2246                  * Replay.  Do not regenerate message (and rebuild
2247                  * paths, etc.); just use the original message.
2248                  * Rebuilding paths will break for renames because
2249                  * d_move mangles the src name.
2250                  */
2251                 msg = req->r_request;
2252                 rhead = msg->front.iov_base;
2253
2254                 flags = le32_to_cpu(rhead->flags);
2255                 flags |= CEPH_MDS_FLAG_REPLAY;
2256                 rhead->flags = cpu_to_le32(flags);
2257
2258                 if (req->r_target_inode)
2259                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2260
2261                 rhead->num_retry = req->r_attempts - 1;
2262
2263                 /* remove cap/dentry releases from message */
2264                 rhead->num_releases = 0;
2265
2266                 /* time stamp */
2267                 p = msg->front.iov_base + req->r_request_release_offset;
2268                 {
2269                         struct ceph_timespec ts;
2270                         ceph_encode_timespec64(&ts, &req->r_stamp);
2271                         ceph_encode_copy(&p, &ts, sizeof(ts));
2272                 }
2273
2274                 msg->front.iov_len = p - msg->front.iov_base;
2275                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2276                 return 0;
2277         }
2278
2279         if (req->r_request) {
2280                 ceph_msg_put(req->r_request);
2281                 req->r_request = NULL;
2282         }
2283         msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2284         if (IS_ERR(msg)) {
2285                 req->r_err = PTR_ERR(msg);
2286                 return PTR_ERR(msg);
2287         }
2288         req->r_request = msg;
2289
2290         rhead = msg->front.iov_base;
2291         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2292         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2293                 flags |= CEPH_MDS_FLAG_REPLAY;
2294         if (req->r_parent)
2295                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2296         rhead->flags = cpu_to_le32(flags);
2297         rhead->num_fwd = req->r_num_fwd;
2298         rhead->num_retry = req->r_attempts - 1;
2299         rhead->ino = 0;
2300
2301         dout(" r_parent = %p\n", req->r_parent);
2302         return 0;
2303 }
2304
2305 /*
2306  * send request, or put it on the appropriate wait list.
2307  */
2308 static void __do_request(struct ceph_mds_client *mdsc,
2309                         struct ceph_mds_request *req)
2310 {
2311         struct ceph_mds_session *session = NULL;
2312         int mds = -1;
2313         int err = 0;
2314
2315         if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2316                 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2317                         __unregister_request(mdsc, req);
2318                 return;
2319         }
2320
2321         if (req->r_timeout &&
2322             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2323                 dout("do_request timed out\n");
2324                 err = -EIO;
2325                 goto finish;
2326         }
2327         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2328                 dout("do_request forced umount\n");
2329                 err = -EIO;
2330                 goto finish;
2331         }
2332         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2333                 if (mdsc->mdsmap_err) {
2334                         err = mdsc->mdsmap_err;
2335                         dout("do_request mdsmap err %d\n", err);
2336                         goto finish;
2337                 }
2338                 if (mdsc->mdsmap->m_epoch == 0) {
2339                         dout("do_request no mdsmap, waiting for map\n");
2340                         list_add(&req->r_wait, &mdsc->waiting_for_map);
2341                         return;
2342                 }
2343                 if (!(mdsc->fsc->mount_options->flags &
2344                       CEPH_MOUNT_OPT_MOUNTWAIT) &&
2345                     !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2346                         err = -EHOSTUNREACH;
2347                         goto finish;
2348                 }
2349         }
2350
2351         put_request_session(req);
2352
2353         mds = __choose_mds(mdsc, req);
2354         if (mds < 0 ||
2355             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2356                 dout("do_request no mds or not active, waiting for map\n");
2357                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2358                 return;
2359         }
2360
2361         /* get, open session */
2362         session = __ceph_lookup_mds_session(mdsc, mds);
2363         if (!session) {
2364                 session = register_session(mdsc, mds);
2365                 if (IS_ERR(session)) {
2366                         err = PTR_ERR(session);
2367                         goto finish;
2368                 }
2369         }
2370         req->r_session = get_session(session);
2371
2372         dout("do_request mds%d session %p state %s\n", mds, session,
2373              ceph_session_state_name(session->s_state));
2374         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2375             session->s_state != CEPH_MDS_SESSION_HUNG) {
2376                 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2377                         err = -EACCES;
2378                         goto out_session;
2379                 }
2380                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2381                     session->s_state == CEPH_MDS_SESSION_CLOSING)
2382                         __open_session(mdsc, session);
2383                 list_add(&req->r_wait, &session->s_waiting);
2384                 goto out_session;
2385         }
2386
2387         /* send request */
2388         req->r_resend_mds = -1;   /* forget any previous mds hint */
2389
2390         if (req->r_request_started == 0)   /* note request start time */
2391                 req->r_request_started = jiffies;
2392
2393         err = __prepare_send_request(mdsc, req, mds, false);
2394         if (!err) {
2395                 ceph_msg_get(req->r_request);
2396                 ceph_con_send(&session->s_con, req->r_request);
2397         }
2398
2399 out_session:
2400         ceph_put_mds_session(session);
2401 finish:
2402         if (err) {
2403                 dout("__do_request early error %d\n", err);
2404                 req->r_err = err;
2405                 complete_request(mdsc, req);
2406                 __unregister_request(mdsc, req);
2407         }
2408         return;
2409 }
2410
2411 /*
2412  * called under mdsc->mutex
2413  */
2414 static void __wake_requests(struct ceph_mds_client *mdsc,
2415                             struct list_head *head)
2416 {
2417         struct ceph_mds_request *req;
2418         LIST_HEAD(tmp_list);
2419
2420         list_splice_init(head, &tmp_list);
2421
2422         while (!list_empty(&tmp_list)) {
2423                 req = list_entry(tmp_list.next,
2424                                  struct ceph_mds_request, r_wait);
2425                 list_del_init(&req->r_wait);
2426                 dout(" wake request %p tid %llu\n", req, req->r_tid);
2427                 __do_request(mdsc, req);
2428         }
2429 }
2430
2431 /*
2432  * Wake up threads with requests pending for @mds, so that they can
2433  * resubmit their requests to a possibly different mds.
2434  */
2435 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2436 {
2437         struct ceph_mds_request *req;
2438         struct rb_node *p = rb_first(&mdsc->request_tree);
2439
2440         dout("kick_requests mds%d\n", mds);
2441         while (p) {
2442                 req = rb_entry(p, struct ceph_mds_request, r_node);
2443                 p = rb_next(p);
2444                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2445                         continue;
2446                 if (req->r_attempts > 0)
2447                         continue; /* only new requests */
2448                 if (req->r_session &&
2449                     req->r_session->s_mds == mds) {
2450                         dout(" kicking tid %llu\n", req->r_tid);
2451                         list_del_init(&req->r_wait);
2452                         __do_request(mdsc, req);
2453                 }
2454         }
2455 }
2456
2457 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2458                               struct ceph_mds_request *req)
2459 {
2460         dout("submit_request on %p\n", req);
2461         mutex_lock(&mdsc->mutex);
2462         __register_request(mdsc, req, NULL);
2463         __do_request(mdsc, req);
2464         mutex_unlock(&mdsc->mutex);
2465 }
2466
2467 /*
2468  * Synchrously perform an mds request.  Take care of all of the
2469  * session setup, forwarding, retry details.
2470  */
2471 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2472                          struct inode *dir,
2473                          struct ceph_mds_request *req)
2474 {
2475         int err;
2476
2477         dout("do_request on %p\n", req);
2478
2479         /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2480         if (req->r_inode)
2481                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2482         if (req->r_parent)
2483                 ceph_get_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
2484         if (req->r_old_dentry_dir)
2485                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2486                                   CEPH_CAP_PIN);
2487
2488         /* issue */
2489         mutex_lock(&mdsc->mutex);
2490         __register_request(mdsc, req, dir);
2491         __do_request(mdsc, req);
2492
2493         if (req->r_err) {
2494                 err = req->r_err;
2495                 goto out;
2496         }
2497
2498         /* wait */
2499         mutex_unlock(&mdsc->mutex);
2500         dout("do_request waiting\n");
2501         if (!req->r_timeout && req->r_wait_for_completion) {
2502                 err = req->r_wait_for_completion(mdsc, req);
2503         } else {
2504                 long timeleft = wait_for_completion_killable_timeout(
2505                                         &req->r_completion,
2506                                         ceph_timeout_jiffies(req->r_timeout));
2507                 if (timeleft > 0)
2508                         err = 0;
2509                 else if (!timeleft)
2510                         err = -EIO;  /* timed out */
2511                 else
2512                         err = timeleft;  /* killed */
2513         }
2514         dout("do_request waited, got %d\n", err);
2515         mutex_lock(&mdsc->mutex);
2516
2517         /* only abort if we didn't race with a real reply */
2518         if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2519                 err = le32_to_cpu(req->r_reply_info.head->result);
2520         } else if (err < 0) {
2521                 dout("aborted request %lld with %d\n", req->r_tid, err);
2522
2523                 /*
2524                  * ensure we aren't running concurrently with
2525                  * ceph_fill_trace or ceph_readdir_prepopulate, which
2526                  * rely on locks (dir mutex) held by our caller.
2527                  */
2528                 mutex_lock(&req->r_fill_mutex);
2529                 req->r_err = err;
2530                 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2531                 mutex_unlock(&req->r_fill_mutex);
2532
2533                 if (req->r_parent &&
2534                     (req->r_op & CEPH_MDS_OP_WRITE))
2535                         ceph_invalidate_dir_request(req);
2536         } else {
2537                 err = req->r_err;
2538         }
2539
2540 out:
2541         mutex_unlock(&mdsc->mutex);
2542         dout("do_request %p done, result %d\n", req, err);
2543         return err;
2544 }
2545
2546 /*
2547  * Invalidate dir's completeness, dentry lease state on an aborted MDS
2548  * namespace request.
2549  */
2550 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2551 {
2552         struct inode *dir = req->r_parent;
2553         struct inode *old_dir = req->r_old_dentry_dir;
2554
2555         dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
2556
2557         ceph_dir_clear_complete(dir);
2558         if (old_dir)
2559                 ceph_dir_clear_complete(old_dir);
2560         if (req->r_dentry)
2561                 ceph_invalidate_dentry_lease(req->r_dentry);
2562         if (req->r_old_dentry)
2563                 ceph_invalidate_dentry_lease(req->r_old_dentry);
2564 }
2565
2566 /*
2567  * Handle mds reply.
2568  *
2569  * We take the session mutex and parse and process the reply immediately.
2570  * This preserves the logical ordering of replies, capabilities, etc., sent
2571  * by the MDS as they are applied to our local cache.
2572  */
2573 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2574 {
2575         struct ceph_mds_client *mdsc = session->s_mdsc;
2576         struct ceph_mds_request *req;
2577         struct ceph_mds_reply_head *head = msg->front.iov_base;
2578         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2579         struct ceph_snap_realm *realm;
2580         u64 tid;
2581         int err, result;
2582         int mds = session->s_mds;
2583
2584         if (msg->front.iov_len < sizeof(*head)) {
2585                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2586                 ceph_msg_dump(msg);
2587                 return;
2588         }
2589
2590         /* get request, session */
2591         tid = le64_to_cpu(msg->hdr.tid);
2592         mutex_lock(&mdsc->mutex);
2593         req = lookup_get_request(mdsc, tid);
2594         if (!req) {
2595                 dout("handle_reply on unknown tid %llu\n", tid);
2596                 mutex_unlock(&mdsc->mutex);
2597                 return;
2598         }
2599         dout("handle_reply %p\n", req);
2600
2601         /* correct session? */
2602         if (req->r_session != session) {
2603                 pr_err("mdsc_handle_reply got %llu on session mds%d"
2604                        " not mds%d\n", tid, session->s_mds,
2605                        req->r_session ? req->r_session->s_mds : -1);
2606                 mutex_unlock(&mdsc->mutex);
2607                 goto out;
2608         }
2609
2610         /* dup? */
2611         if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
2612             (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
2613                 pr_warn("got a dup %s reply on %llu from mds%d\n",
2614                            head->safe ? "safe" : "unsafe", tid, mds);
2615                 mutex_unlock(&mdsc->mutex);
2616                 goto out;
2617         }
2618         if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
2619                 pr_warn("got unsafe after safe on %llu from mds%d\n",
2620                            tid, mds);
2621                 mutex_unlock(&mdsc->mutex);
2622                 goto out;
2623         }
2624
2625         result = le32_to_cpu(head->result);
2626
2627         /*
2628          * Handle an ESTALE
2629          * if we're not talking to the authority, send to them
2630          * if the authority has changed while we weren't looking,
2631          * send to new authority
2632          * Otherwise we just have to return an ESTALE
2633          */
2634         if (result == -ESTALE) {
2635                 dout("got ESTALE on request %llu\n", req->r_tid);
2636                 req->r_resend_mds = -1;
2637                 if (req->r_direct_mode != USE_AUTH_MDS) {
2638                         dout("not using auth, setting for that now\n");
2639                         req->r_direct_mode = USE_AUTH_MDS;
2640                         __do_request(mdsc, req);
2641                         mutex_unlock(&mdsc->mutex);
2642                         goto out;
2643                 } else  {
2644                         int mds = __choose_mds(mdsc, req);
2645                         if (mds >= 0 && mds != req->r_session->s_mds) {
2646                                 dout("but auth changed, so resending\n");
2647                                 __do_request(mdsc, req);
2648                                 mutex_unlock(&mdsc->mutex);
2649                                 goto out;
2650                         }
2651                 }
2652                 dout("have to return ESTALE on request %llu\n", req->r_tid);
2653         }
2654
2655
2656         if (head->safe) {
2657                 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
2658                 __unregister_request(mdsc, req);
2659
2660                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2661                         /*
2662                          * We already handled the unsafe response, now do the
2663                          * cleanup.  No need to examine the response; the MDS
2664                          * doesn't include any result info in the safe
2665                          * response.  And even if it did, there is nothing
2666                          * useful we could do with a revised return value.
2667                          */
2668                         dout("got safe reply %llu, mds%d\n", tid, mds);
2669
2670                         /* last unsafe request during umount? */
2671                         if (mdsc->stopping && !__get_oldest_req(mdsc))
2672                                 complete_all(&mdsc->safe_umount_waiters);
2673                         mutex_unlock(&mdsc->mutex);
2674                         goto out;
2675                 }
2676         } else {
2677                 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
2678                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2679                 if (req->r_unsafe_dir) {
2680                         struct ceph_inode_info *ci =
2681                                         ceph_inode(req->r_unsafe_dir);
2682                         spin_lock(&ci->i_unsafe_lock);
2683                         list_add_tail(&req->r_unsafe_dir_item,
2684                                       &ci->i_unsafe_dirops);
2685                         spin_unlock(&ci->i_unsafe_lock);
2686                 }
2687         }
2688
2689         dout("handle_reply tid %lld result %d\n", tid, result);
2690         rinfo = &req->r_reply_info;
2691         err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2692         mutex_unlock(&mdsc->mutex);
2693
2694         mutex_lock(&session->s_mutex);
2695         if (err < 0) {
2696                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2697                 ceph_msg_dump(msg);
2698                 goto out_err;
2699         }
2700
2701         /* snap trace */
2702         realm = NULL;
2703         if (rinfo->snapblob_len) {
2704                 down_write(&mdsc->snap_rwsem);
2705                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2706                                 rinfo->snapblob + rinfo->snapblob_len,
2707                                 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2708                                 &realm);
2709                 downgrade_write(&mdsc->snap_rwsem);
2710         } else {
2711                 down_read(&mdsc->snap_rwsem);
2712         }
2713
2714         /* insert trace into our cache */
2715         mutex_lock(&req->r_fill_mutex);
2716         current->journal_info = req;
2717         err = ceph_fill_trace(mdsc->fsc->sb, req);
2718         if (err == 0) {
2719                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2720                                     req->r_op == CEPH_MDS_OP_LSSNAP))
2721                         ceph_readdir_prepopulate(req, req->r_session);
2722                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2723         }
2724         current->journal_info = NULL;
2725         mutex_unlock(&req->r_fill_mutex);
2726
2727         up_read(&mdsc->snap_rwsem);
2728         if (realm)
2729                 ceph_put_snap_realm(mdsc, realm);
2730
2731         if (err == 0 && req->r_target_inode &&
2732             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2733                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
2734                 spin_lock(&ci->i_unsafe_lock);
2735                 list_add_tail(&req->r_unsafe_target_item, &ci->i_unsafe_iops);
2736                 spin_unlock(&ci->i_unsafe_lock);
2737         }
2738 out_err:
2739         mutex_lock(&mdsc->mutex);
2740         if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2741                 if (err) {
2742                         req->r_err = err;
2743                 } else {
2744                         req->r_reply =  ceph_msg_get(msg);
2745                         set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
2746                 }
2747         } else {
2748                 dout("reply arrived after request %lld was aborted\n", tid);
2749         }
2750         mutex_unlock(&mdsc->mutex);
2751
2752         mutex_unlock(&session->s_mutex);
2753
2754         /* kick calling process */
2755         complete_request(mdsc, req);
2756 out:
2757         ceph_mdsc_put_request(req);
2758         return;
2759 }
2760
2761
2762
2763 /*
2764  * handle mds notification that our request has been forwarded.
2765  */
2766 static void handle_forward(struct ceph_mds_client *mdsc,
2767                            struct ceph_mds_session *session,
2768                            struct ceph_msg *msg)
2769 {
2770         struct ceph_mds_request *req;
2771         u64 tid = le64_to_cpu(msg->hdr.tid);
2772         u32 next_mds;
2773         u32 fwd_seq;
2774         int err = -EINVAL;
2775         void *p = msg->front.iov_base;
2776         void *end = p + msg->front.iov_len;
2777
2778         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2779         next_mds = ceph_decode_32(&p);
2780         fwd_seq = ceph_decode_32(&p);
2781
2782         mutex_lock(&mdsc->mutex);
2783         req = lookup_get_request(mdsc, tid);
2784         if (!req) {
2785                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2786                 goto out;  /* dup reply? */
2787         }
2788
2789         if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2790                 dout("forward tid %llu aborted, unregistering\n", tid);
2791                 __unregister_request(mdsc, req);
2792         } else if (fwd_seq <= req->r_num_fwd) {
2793                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2794                      tid, next_mds, req->r_num_fwd, fwd_seq);
2795         } else {
2796                 /* resend. forward race not possible; mds would drop */
2797                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2798                 BUG_ON(req->r_err);
2799                 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
2800                 req->r_attempts = 0;
2801                 req->r_num_fwd = fwd_seq;
2802                 req->r_resend_mds = next_mds;
2803                 put_request_session(req);
2804                 __do_request(mdsc, req);
2805         }
2806         ceph_mdsc_put_request(req);
2807 out:
2808         mutex_unlock(&mdsc->mutex);
2809         return;
2810
2811 bad:
2812         pr_err("mdsc_handle_forward decode error err=%d\n", err);
2813 }
2814
2815 /*
2816  * handle a mds session control message
2817  */
2818 static void handle_session(struct ceph_mds_session *session,
2819                            struct ceph_msg *msg)
2820 {
2821         struct ceph_mds_client *mdsc = session->s_mdsc;
2822         u32 op;
2823         u64 seq;
2824         int mds = session->s_mds;
2825         struct ceph_mds_session_head *h = msg->front.iov_base;
2826         int wake = 0;
2827
2828         /* decode */
2829         if (msg->front.iov_len < sizeof(*h))
2830                 goto bad;
2831         op = le32_to_cpu(h->op);
2832         seq = le64_to_cpu(h->seq);
2833
2834         mutex_lock(&mdsc->mutex);
2835         if (op == CEPH_SESSION_CLOSE) {
2836                 get_session(session);
2837                 __unregister_session(mdsc, session);
2838         }
2839         /* FIXME: this ttl calculation is generous */
2840         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2841         mutex_unlock(&mdsc->mutex);
2842
2843         mutex_lock(&session->s_mutex);
2844
2845         dout("handle_session mds%d %s %p state %s seq %llu\n",
2846              mds, ceph_session_op_name(op), session,
2847              ceph_session_state_name(session->s_state), seq);
2848
2849         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2850                 session->s_state = CEPH_MDS_SESSION_OPEN;
2851                 pr_info("mds%d came back\n", session->s_mds);
2852         }
2853
2854         switch (op) {
2855         case CEPH_SESSION_OPEN:
2856                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2857                         pr_info("mds%d reconnect success\n", session->s_mds);
2858                 session->s_state = CEPH_MDS_SESSION_OPEN;
2859                 renewed_caps(mdsc, session, 0);
2860                 wake = 1;
2861                 if (mdsc->stopping)
2862                         __close_session(mdsc, session);
2863                 break;
2864
2865         case CEPH_SESSION_RENEWCAPS:
2866                 if (session->s_renew_seq == seq)
2867                         renewed_caps(mdsc, session, 1);
2868                 break;
2869
2870         case CEPH_SESSION_CLOSE:
2871                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2872                         pr_info("mds%d reconnect denied\n", session->s_mds);
2873                 cleanup_session_requests(mdsc, session);
2874                 remove_session_caps(session);
2875                 wake = 2; /* for good measure */
2876                 wake_up_all(&mdsc->session_close_wq);
2877                 break;
2878
2879         case CEPH_SESSION_STALE:
2880                 pr_info("mds%d caps went stale, renewing\n",
2881                         session->s_mds);
2882                 spin_lock(&session->s_gen_ttl_lock);
2883                 session->s_cap_gen++;
2884                 session->s_cap_ttl = jiffies - 1;
2885                 spin_unlock(&session->s_gen_ttl_lock);
2886                 send_renew_caps(mdsc, session);
2887                 break;
2888
2889         case CEPH_SESSION_RECALL_STATE:
2890                 ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2891                 break;
2892
2893         case CEPH_SESSION_FLUSHMSG:
2894                 send_flushmsg_ack(mdsc, session, seq);
2895                 break;
2896
2897         case CEPH_SESSION_FORCE_RO:
2898                 dout("force_session_readonly %p\n", session);
2899                 spin_lock(&session->s_cap_lock);
2900                 session->s_readonly = true;
2901                 spin_unlock(&session->s_cap_lock);
2902                 wake_up_session_caps(session, 0);
2903                 break;
2904
2905         case CEPH_SESSION_REJECT:
2906                 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
2907                 pr_info("mds%d rejected session\n", session->s_mds);
2908                 session->s_state = CEPH_MDS_SESSION_REJECTED;
2909                 cleanup_session_requests(mdsc, session);
2910                 remove_session_caps(session);
2911                 wake = 2; /* for good measure */
2912                 break;
2913
2914         default:
2915                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2916                 WARN_ON(1);
2917         }
2918
2919         mutex_unlock(&session->s_mutex);
2920         if (wake) {
2921                 mutex_lock(&mdsc->mutex);
2922                 __wake_requests(mdsc, &session->s_waiting);
2923                 if (wake == 2)
2924                         kick_requests(mdsc, mds);
2925                 mutex_unlock(&mdsc->mutex);
2926         }
2927         if (op == CEPH_SESSION_CLOSE)
2928                 ceph_put_mds_session(session);
2929         return;
2930
2931 bad:
2932         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2933                (int)msg->front.iov_len);
2934         ceph_msg_dump(msg);
2935         return;
2936 }
2937
2938
2939 /*
2940  * called under session->mutex.
2941  */
2942 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2943                                    struct ceph_mds_session *session)
2944 {
2945         struct ceph_mds_request *req, *nreq;
2946         struct rb_node *p;
2947         int err;
2948
2949         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2950
2951         mutex_lock(&mdsc->mutex);
2952         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2953                 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2954                 if (!err) {
2955                         ceph_msg_get(req->r_request);
2956                         ceph_con_send(&session->s_con, req->r_request);
2957                 }
2958         }
2959
2960         /*
2961          * also re-send old requests when MDS enters reconnect stage. So that MDS
2962          * can process completed request in clientreplay stage.
2963          */
2964         p = rb_first(&mdsc->request_tree);
2965         while (p) {
2966                 req = rb_entry(p, struct ceph_mds_request, r_node);
2967                 p = rb_next(p);
2968                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2969                         continue;
2970                 if (req->r_attempts == 0)
2971                         continue; /* only old requests */
2972                 if (req->r_session &&
2973                     req->r_session->s_mds == session->s_mds) {
2974                         err = __prepare_send_request(mdsc, req,
2975                                                      session->s_mds, true);
2976                         if (!err) {
2977                                 ceph_msg_get(req->r_request);
2978                                 ceph_con_send(&session->s_con, req->r_request);
2979                         }
2980                 }
2981         }
2982         mutex_unlock(&mdsc->mutex);
2983 }
2984
2985 /*
2986  * Encode information about a cap for a reconnect with the MDS.
2987  */
2988 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2989                           void *arg)
2990 {
2991         union {
2992                 struct ceph_mds_cap_reconnect v2;
2993                 struct ceph_mds_cap_reconnect_v1 v1;
2994         } rec;
2995         struct ceph_inode_info *ci = cap->ci;
2996         struct ceph_reconnect_state *recon_state = arg;
2997         struct ceph_pagelist *pagelist = recon_state->pagelist;
2998         char *path;
2999         int pathlen, err;
3000         u64 pathbase;
3001         u64 snap_follows;
3002         struct dentry *dentry;
3003
3004         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
3005              inode, ceph_vinop(inode), cap, cap->cap_id,
3006              ceph_cap_string(cap->issued));
3007         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3008         if (err)
3009                 return err;
3010
3011         dentry = d_find_alias(inode);
3012         if (dentry) {
3013                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
3014                 if (IS_ERR(path)) {
3015                         err = PTR_ERR(path);
3016                         goto out_dput;
3017                 }
3018         } else {
3019                 path = NULL;
3020                 pathlen = 0;
3021                 pathbase = 0;
3022         }
3023
3024         spin_lock(&ci->i_ceph_lock);
3025         cap->seq = 0;        /* reset cap seq */
3026         cap->issue_seq = 0;  /* and issue_seq */
3027         cap->mseq = 0;       /* and migrate_seq */
3028         cap->cap_gen = cap->session->s_cap_gen;
3029
3030         if (recon_state->msg_version >= 2) {
3031                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
3032                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3033                 rec.v2.issued = cpu_to_le32(cap->issued);
3034                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3035                 rec.v2.pathbase = cpu_to_le64(pathbase);
3036                 rec.v2.flock_len = (__force __le32)
3037                         ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
3038         } else {
3039                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
3040                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3041                 rec.v1.issued = cpu_to_le32(cap->issued);
3042                 rec.v1.size = cpu_to_le64(inode->i_size);
3043                 ceph_encode_timespec64(&rec.v1.mtime, &inode->i_mtime);
3044                 ceph_encode_timespec64(&rec.v1.atime, &inode->i_atime);
3045                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3046                 rec.v1.pathbase = cpu_to_le64(pathbase);
3047         }
3048
3049         if (list_empty(&ci->i_cap_snaps)) {
3050                 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
3051         } else {
3052                 struct ceph_cap_snap *capsnap =
3053                         list_first_entry(&ci->i_cap_snaps,
3054                                          struct ceph_cap_snap, ci_item);
3055                 snap_follows = capsnap->follows;
3056         }
3057         spin_unlock(&ci->i_ceph_lock);
3058
3059         if (recon_state->msg_version >= 2) {
3060                 int num_fcntl_locks, num_flock_locks;
3061                 struct ceph_filelock *flocks = NULL;
3062                 size_t struct_len, total_len = 0;
3063                 u8 struct_v = 0;
3064
3065 encode_again:
3066                 if (rec.v2.flock_len) {
3067                         ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
3068                 } else {
3069                         num_fcntl_locks = 0;
3070                         num_flock_locks = 0;
3071                 }
3072                 if (num_fcntl_locks + num_flock_locks > 0) {
3073                         flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
3074                                                sizeof(struct ceph_filelock),
3075                                                GFP_NOFS);
3076                         if (!flocks) {
3077                                 err = -ENOMEM;
3078                                 goto out_free;
3079                         }
3080                         err = ceph_encode_locks_to_buffer(inode, flocks,
3081                                                           num_fcntl_locks,
3082                                                           num_flock_locks);
3083                         if (err) {
3084                                 kfree(flocks);
3085                                 flocks = NULL;
3086                                 if (err == -ENOSPC)
3087                                         goto encode_again;
3088                                 goto out_free;
3089                         }
3090                 } else {
3091                         kfree(flocks);
3092                         flocks = NULL;
3093                 }
3094
3095                 if (recon_state->msg_version >= 3) {
3096                         /* version, compat_version and struct_len */
3097                         total_len = 2 * sizeof(u8) + sizeof(u32);
3098                         struct_v = 2;
3099                 }
3100                 /*
3101                  * number of encoded locks is stable, so copy to pagelist
3102                  */
3103                 struct_len = 2 * sizeof(u32) +
3104                             (num_fcntl_locks + num_flock_locks) *
3105                             sizeof(struct ceph_filelock);
3106                 rec.v2.flock_len = cpu_to_le32(struct_len);
3107
3108                 struct_len += sizeof(rec.v2);
3109                 struct_len += sizeof(u32) + pathlen;
3110
3111                 if (struct_v >= 2)
3112                         struct_len += sizeof(u64); /* snap_follows */
3113
3114                 total_len += struct_len;
3115                 err = ceph_pagelist_reserve(pagelist, total_len);
3116
3117                 if (!err) {
3118                         if (recon_state->msg_version >= 3) {
3119                                 ceph_pagelist_encode_8(pagelist, struct_v);
3120                                 ceph_pagelist_encode_8(pagelist, 1);
3121                                 ceph_pagelist_encode_32(pagelist, struct_len);
3122                         }
3123                         ceph_pagelist_encode_string(pagelist, path, pathlen);
3124                         ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
3125                         ceph_locks_to_pagelist(flocks, pagelist,
3126                                                num_fcntl_locks,
3127                                                num_flock_locks);
3128                         if (struct_v >= 2)
3129                                 ceph_pagelist_encode_64(pagelist, snap_follows);
3130                 }
3131                 kfree(flocks);
3132         } else {
3133                 size_t size = sizeof(u32) + pathlen + sizeof(rec.v1);
3134                 err = ceph_pagelist_reserve(pagelist, size);
3135                 if (!err) {
3136                         ceph_pagelist_encode_string(pagelist, path, pathlen);
3137                         ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
3138                 }
3139         }
3140
3141         recon_state->nr_caps++;
3142 out_free:
3143         kfree(path);
3144 out_dput:
3145         dput(dentry);
3146         return err;
3147 }
3148
3149
3150 /*
3151  * If an MDS fails and recovers, clients need to reconnect in order to
3152  * reestablish shared state.  This includes all caps issued through
3153  * this session _and_ the snap_realm hierarchy.  Because it's not
3154  * clear which snap realms the mds cares about, we send everything we
3155  * know about.. that ensures we'll then get any new info the
3156  * recovering MDS might have.
3157  *
3158  * This is a relatively heavyweight operation, but it's rare.
3159  *
3160  * called with mdsc->mutex held.
3161  */
3162 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3163                                struct ceph_mds_session *session)
3164 {
3165         struct ceph_msg *reply;
3166         struct rb_node *p;
3167         int mds = session->s_mds;
3168         int err = -ENOMEM;
3169         int s_nr_caps;
3170         struct ceph_pagelist *pagelist;
3171         struct ceph_reconnect_state recon_state;
3172         LIST_HEAD(dispose);
3173
3174         pr_info("mds%d reconnect start\n", mds);
3175
3176         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
3177         if (!pagelist)
3178                 goto fail_nopagelist;
3179         ceph_pagelist_init(pagelist);
3180
3181         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
3182         if (!reply)
3183                 goto fail_nomsg;
3184
3185         mutex_lock(&session->s_mutex);
3186         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3187         session->s_seq = 0;
3188
3189         dout("session %p state %s\n", session,
3190              ceph_session_state_name(session->s_state));
3191
3192         spin_lock(&session->s_gen_ttl_lock);
3193         session->s_cap_gen++;
3194         spin_unlock(&session->s_gen_ttl_lock);
3195
3196         spin_lock(&session->s_cap_lock);
3197         /* don't know if session is readonly */
3198         session->s_readonly = 0;
3199         /*
3200          * notify __ceph_remove_cap() that we are composing cap reconnect.
3201          * If a cap get released before being added to the cap reconnect,
3202          * __ceph_remove_cap() should skip queuing cap release.
3203          */
3204         session->s_cap_reconnect = 1;
3205         /* drop old cap expires; we're about to reestablish that state */
3206         detach_cap_releases(session, &dispose);
3207         spin_unlock(&session->s_cap_lock);
3208         dispose_cap_releases(mdsc, &dispose);
3209
3210         /* trim unused caps to reduce MDS's cache rejoin time */
3211         if (mdsc->fsc->sb->s_root)
3212                 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3213
3214         ceph_con_close(&session->s_con);
3215         ceph_con_open(&session->s_con,
3216                       CEPH_ENTITY_TYPE_MDS, mds,
3217                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3218
3219         /* replay unsafe requests */
3220         replay_unsafe_requests(mdsc, session);
3221
3222         down_read(&mdsc->snap_rwsem);
3223
3224         /* traverse this session's caps */
3225         s_nr_caps = session->s_nr_caps;
3226         err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
3227         if (err)
3228                 goto fail;
3229
3230         recon_state.nr_caps = 0;
3231         recon_state.pagelist = pagelist;
3232         if (session->s_con.peer_features & CEPH_FEATURE_MDSENC)
3233                 recon_state.msg_version = 3;
3234         else if (session->s_con.peer_features & CEPH_FEATURE_FLOCK)
3235                 recon_state.msg_version = 2;
3236         else
3237                 recon_state.msg_version = 1;
3238         err = iterate_session_caps(session, encode_caps_cb, &recon_state);
3239         if (err < 0)
3240                 goto fail;
3241
3242         spin_lock(&session->s_cap_lock);
3243         session->s_cap_reconnect = 0;
3244         spin_unlock(&session->s_cap_lock);
3245
3246         /*
3247          * snaprealms.  we provide mds with the ino, seq (version), and
3248          * parent for all of our realms.  If the mds has any newer info,
3249          * it will tell us.
3250          */
3251         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3252                 struct ceph_snap_realm *realm =
3253                         rb_entry(p, struct ceph_snap_realm, node);
3254                 struct ceph_mds_snaprealm_reconnect sr_rec;
3255
3256                 dout(" adding snap realm %llx seq %lld parent %llx\n",
3257                      realm->ino, realm->seq, realm->parent_ino);
3258                 sr_rec.ino = cpu_to_le64(realm->ino);
3259                 sr_rec.seq = cpu_to_le64(realm->seq);
3260                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3261                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3262                 if (err)
3263                         goto fail;
3264         }
3265
3266         reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3267
3268         /* raced with cap release? */
3269         if (s_nr_caps != recon_state.nr_caps) {
3270                 struct page *page = list_first_entry(&pagelist->head,
3271                                                      struct page, lru);
3272                 __le32 *addr = kmap_atomic(page);
3273                 *addr = cpu_to_le32(recon_state.nr_caps);
3274                 kunmap_atomic(addr);
3275         }
3276
3277         reply->hdr.data_len = cpu_to_le32(pagelist->length);
3278         ceph_msg_data_add_pagelist(reply, pagelist);
3279
3280         ceph_early_kick_flushing_caps(mdsc, session);
3281
3282         ceph_con_send(&session->s_con, reply);
3283
3284         mutex_unlock(&session->s_mutex);
3285
3286         mutex_lock(&mdsc->mutex);
3287         __wake_requests(mdsc, &session->s_waiting);
3288         mutex_unlock(&mdsc->mutex);
3289
3290         up_read(&mdsc->snap_rwsem);
3291         return;
3292
3293 fail:
3294         ceph_msg_put(reply);
3295         up_read(&mdsc->snap_rwsem);
3296         mutex_unlock(&session->s_mutex);
3297 fail_nomsg:
3298         ceph_pagelist_release(pagelist);
3299 fail_nopagelist:
3300         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3301         return;
3302 }
3303
3304
3305 /*
3306  * compare old and new mdsmaps, kicking requests
3307  * and closing out old connections as necessary
3308  *
3309  * called under mdsc->mutex.
3310  */
3311 static void check_new_map(struct ceph_mds_client *mdsc,
3312                           struct ceph_mdsmap *newmap,
3313                           struct ceph_mdsmap *oldmap)
3314 {
3315         int i;
3316         int oldstate, newstate;
3317         struct ceph_mds_session *s;
3318
3319         dout("check_new_map new %u old %u\n",
3320              newmap->m_epoch, oldmap->m_epoch);
3321
3322         for (i = 0; i < oldmap->m_num_mds && i < mdsc->max_sessions; i++) {
3323                 if (!mdsc->sessions[i])
3324                         continue;
3325                 s = mdsc->sessions[i];
3326                 oldstate = ceph_mdsmap_get_state(oldmap, i);
3327                 newstate = ceph_mdsmap_get_state(newmap, i);
3328
3329                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3330                      i, ceph_mds_state_name(oldstate),
3331                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3332                      ceph_mds_state_name(newstate),
3333                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3334                      ceph_session_state_name(s->s_state));
3335
3336                 if (i >= newmap->m_num_mds ||
3337                     memcmp(ceph_mdsmap_get_addr(oldmap, i),
3338                            ceph_mdsmap_get_addr(newmap, i),
3339                            sizeof(struct ceph_entity_addr))) {
3340                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3341                                 /* the session never opened, just close it
3342                                  * out now */
3343                                 get_session(s);
3344                                 __unregister_session(mdsc, s);
3345                                 __wake_requests(mdsc, &s->s_waiting);
3346                                 ceph_put_mds_session(s);
3347                         } else if (i >= newmap->m_num_mds) {
3348                                 /* force close session for stopped mds */
3349                                 get_session(s);
3350                                 __unregister_session(mdsc, s);
3351                                 __wake_requests(mdsc, &s->s_waiting);
3352                                 kick_requests(mdsc, i);
3353                                 mutex_unlock(&mdsc->mutex);
3354
3355                                 mutex_lock(&s->s_mutex);
3356                                 cleanup_session_requests(mdsc, s);
3357                                 remove_session_caps(s);
3358                                 mutex_unlock(&s->s_mutex);
3359
3360                                 ceph_put_mds_session(s);
3361
3362                                 mutex_lock(&mdsc->mutex);
3363                         } else {
3364                                 /* just close it */
3365                                 mutex_unlock(&mdsc->mutex);
3366                                 mutex_lock(&s->s_mutex);
3367                                 mutex_lock(&mdsc->mutex);
3368                                 ceph_con_close(&s->s_con);
3369                                 mutex_unlock(&s->s_mutex);
3370                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3371                         }
3372                 } else if (oldstate == newstate) {
3373                         continue;  /* nothing new with this mds */
3374                 }
3375
3376                 /*
3377                  * send reconnect?
3378                  */
3379                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3380                     newstate >= CEPH_MDS_STATE_RECONNECT) {
3381                         mutex_unlock(&mdsc->mutex);
3382                         send_mds_reconnect(mdsc, s);
3383                         mutex_lock(&mdsc->mutex);
3384                 }
3385
3386                 /*
3387                  * kick request on any mds that has gone active.
3388                  */
3389                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3390                     newstate >= CEPH_MDS_STATE_ACTIVE) {
3391                         if (oldstate != CEPH_MDS_STATE_CREATING &&
3392                             oldstate != CEPH_MDS_STATE_STARTING)
3393                                 pr_info("mds%d recovery completed\n", s->s_mds);
3394                         kick_requests(mdsc, i);
3395                         ceph_kick_flushing_caps(mdsc, s);
3396                         wake_up_session_caps(s, 1);
3397                 }
3398         }
3399
3400         for (i = 0; i < newmap->m_num_mds && i < mdsc->max_sessions; i++) {
3401                 s = mdsc->sessions[i];
3402                 if (!s)
3403                         continue;
3404                 if (!ceph_mdsmap_is_laggy(newmap, i))
3405                         continue;
3406                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3407                     s->s_state == CEPH_MDS_SESSION_HUNG ||
3408                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
3409                         dout(" connecting to export targets of laggy mds%d\n",
3410                              i);
3411                         __open_export_target_sessions(mdsc, s);
3412                 }
3413         }
3414 }
3415
3416
3417
3418 /*
3419  * leases
3420  */
3421
3422 /*
3423  * caller must hold session s_mutex, dentry->d_lock
3424  */
3425 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3426 {
3427         struct ceph_dentry_info *di = ceph_dentry(dentry);
3428
3429         ceph_put_mds_session(di->lease_session);
3430         di->lease_session = NULL;
3431 }
3432
3433 static void handle_lease(struct ceph_mds_client *mdsc,
3434                          struct ceph_mds_session *session,
3435                          struct ceph_msg *msg)
3436 {
3437         struct super_block *sb = mdsc->fsc->sb;
3438         struct inode *inode;
3439         struct dentry *parent, *dentry;
3440         struct ceph_dentry_info *di;
3441         int mds = session->s_mds;
3442         struct ceph_mds_lease *h = msg->front.iov_base;
3443         u32 seq;
3444         struct ceph_vino vino;
3445         struct qstr dname;
3446         int release = 0;
3447
3448         dout("handle_lease from mds%d\n", mds);
3449
3450         /* decode */
3451         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3452                 goto bad;
3453         vino.ino = le64_to_cpu(h->ino);
3454         vino.snap = CEPH_NOSNAP;
3455         seq = le32_to_cpu(h->seq);
3456         dname.len = get_unaligned_le32(h + 1);
3457         if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
3458                 goto bad;
3459         dname.name = (void *)(h + 1) + sizeof(u32);
3460
3461         /* lookup inode */
3462         inode = ceph_find_inode(sb, vino);
3463         dout("handle_lease %s, ino %llx %p %.*s\n",
3464              ceph_lease_op_name(h->action), vino.ino, inode,
3465              dname.len, dname.name);
3466
3467         mutex_lock(&session->s_mutex);
3468         session->s_seq++;
3469
3470         if (!inode) {
3471                 dout("handle_lease no inode %llx\n", vino.ino);
3472                 goto release;
3473         }
3474
3475         /* dentry */
3476         parent = d_find_alias(inode);
3477         if (!parent) {
3478                 dout("no parent dentry on inode %p\n", inode);
3479                 WARN_ON(1);
3480                 goto release;  /* hrm... */
3481         }
3482         dname.hash = full_name_hash(parent, dname.name, dname.len);
3483         dentry = d_lookup(parent, &dname);
3484         dput(parent);
3485         if (!dentry)
3486                 goto release;
3487
3488         spin_lock(&dentry->d_lock);
3489         di = ceph_dentry(dentry);
3490         switch (h->action) {
3491         case CEPH_MDS_LEASE_REVOKE:
3492                 if (di->lease_session == session) {
3493                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3494                                 h->seq = cpu_to_le32(di->lease_seq);
3495                         __ceph_mdsc_drop_dentry_lease(dentry);
3496                 }
3497                 release = 1;
3498                 break;
3499
3500         case CEPH_MDS_LEASE_RENEW:
3501                 if (di->lease_session == session &&
3502                     di->lease_gen == session->s_cap_gen &&
3503                     di->lease_renew_from &&
3504                     di->lease_renew_after == 0) {
3505                         unsigned long duration =
3506                                 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3507
3508                         di->lease_seq = seq;
3509                         di->time = di->lease_renew_from + duration;
3510                         di->lease_renew_after = di->lease_renew_from +
3511                                 (duration >> 1);
3512                         di->lease_renew_from = 0;
3513                 }
3514                 break;
3515         }
3516         spin_unlock(&dentry->d_lock);
3517         dput(dentry);
3518
3519         if (!release)
3520                 goto out;
3521
3522 release:
3523         /* let's just reuse the same message */
3524         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3525         ceph_msg_get(msg);
3526         ceph_con_send(&session->s_con, msg);
3527
3528 out:
3529         iput(inode);
3530         mutex_unlock(&session->s_mutex);
3531         return;
3532
3533 bad:
3534         pr_err("corrupt lease message\n");
3535         ceph_msg_dump(msg);
3536 }
3537
3538 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3539                               struct inode *inode,
3540                               struct dentry *dentry, char action,
3541                               u32 seq)
3542 {
3543         struct ceph_msg *msg;
3544         struct ceph_mds_lease *lease;
3545         int len = sizeof(*lease) + sizeof(u32);
3546         int dnamelen = 0;
3547
3548         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3549              inode, dentry, ceph_lease_op_name(action), session->s_mds);
3550         dnamelen = dentry->d_name.len;
3551         len += dnamelen;
3552
3553         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3554         if (!msg)
3555                 return;
3556         lease = msg->front.iov_base;
3557         lease->action = action;
3558         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3559         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3560         lease->seq = cpu_to_le32(seq);
3561         put_unaligned_le32(dnamelen, lease + 1);
3562         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3563
3564         /*
3565          * if this is a preemptive lease RELEASE, no need to
3566          * flush request stream, since the actual request will
3567          * soon follow.
3568          */
3569         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3570
3571         ceph_con_send(&session->s_con, msg);
3572 }
3573
3574 /*
3575  * lock unlock sessions, to wait ongoing session activities
3576  */
3577 static void lock_unlock_sessions(struct ceph_mds_client *mdsc)
3578 {
3579         int i;
3580
3581         mutex_lock(&mdsc->mutex);
3582         for (i = 0; i < mdsc->max_sessions; i++) {
3583                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3584                 if (!s)
3585                         continue;
3586                 mutex_unlock(&mdsc->mutex);
3587                 mutex_lock(&s->s_mutex);
3588                 mutex_unlock(&s->s_mutex);
3589                 ceph_put_mds_session(s);
3590                 mutex_lock(&mdsc->mutex);
3591         }
3592         mutex_unlock(&mdsc->mutex);
3593 }
3594
3595
3596
3597 /*
3598  * delayed work -- periodically trim expired leases, renew caps with mds
3599  */
3600 static void schedule_delayed(struct ceph_mds_client *mdsc)
3601 {
3602         int delay = 5;
3603         unsigned hz = round_jiffies_relative(HZ * delay);
3604         schedule_delayed_work(&mdsc->delayed_work, hz);
3605 }
3606
3607 static void delayed_work(struct work_struct *work)
3608 {
3609         int i;
3610         struct ceph_mds_client *mdsc =
3611                 container_of(work, struct ceph_mds_client, delayed_work.work);
3612         int renew_interval;
3613         int renew_caps;
3614
3615         dout("mdsc delayed_work\n");
3616         ceph_check_delayed_caps(mdsc);
3617
3618         if (mdsc->stopping)
3619                 return;
3620
3621         mutex_lock(&mdsc->mutex);
3622         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3623         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3624                                    mdsc->last_renew_caps);
3625         if (renew_caps)
3626                 mdsc->last_renew_caps = jiffies;
3627
3628         for (i = 0; i < mdsc->max_sessions; i++) {
3629                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3630                 if (!s)
3631                         continue;
3632                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3633                         dout("resending session close request for mds%d\n",
3634                              s->s_mds);
3635                         request_close_session(mdsc, s);
3636                         ceph_put_mds_session(s);
3637                         continue;
3638                 }
3639                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3640                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3641                                 s->s_state = CEPH_MDS_SESSION_HUNG;
3642                                 pr_info("mds%d hung\n", s->s_mds);
3643                         }
3644                 }
3645                 if (s->s_state == CEPH_MDS_SESSION_NEW ||
3646                     s->s_state == CEPH_MDS_SESSION_RESTARTING ||
3647                     s->s_state == CEPH_MDS_SESSION_REJECTED) {
3648                         /* this mds is failed or recovering, just wait */
3649                         ceph_put_mds_session(s);
3650                         continue;
3651                 }
3652                 mutex_unlock(&mdsc->mutex);
3653
3654                 mutex_lock(&s->s_mutex);
3655                 if (renew_caps)
3656                         send_renew_caps(mdsc, s);
3657                 else
3658                         ceph_con_keepalive(&s->s_con);
3659                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3660                     s->s_state == CEPH_MDS_SESSION_HUNG)
3661                         ceph_send_cap_releases(mdsc, s);
3662                 mutex_unlock(&s->s_mutex);
3663                 ceph_put_mds_session(s);
3664
3665                 mutex_lock(&mdsc->mutex);
3666         }
3667         mutex_unlock(&mdsc->mutex);
3668
3669         schedule_delayed(mdsc);
3670 }
3671
3672 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3673
3674 {
3675         struct ceph_mds_client *mdsc;
3676
3677         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3678         if (!mdsc)
3679                 return -ENOMEM;
3680         mdsc->fsc = fsc;
3681         mutex_init(&mdsc->mutex);
3682         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3683         if (!mdsc->mdsmap) {
3684                 kfree(mdsc);
3685                 return -ENOMEM;
3686         }
3687
3688         init_completion(&mdsc->safe_umount_waiters);
3689         init_waitqueue_head(&mdsc->session_close_wq);
3690         INIT_LIST_HEAD(&mdsc->waiting_for_map);
3691         mdsc->sessions = NULL;
3692         atomic_set(&mdsc->num_sessions, 0);
3693         mdsc->max_sessions = 0;
3694         mdsc->stopping = 0;
3695         atomic64_set(&mdsc->quotarealms_count, 0);
3696         mdsc->last_snap_seq = 0;
3697         init_rwsem(&mdsc->snap_rwsem);
3698         mdsc->snap_realms = RB_ROOT;
3699         INIT_LIST_HEAD(&mdsc->snap_empty);
3700         spin_lock_init(&mdsc->snap_empty_lock);
3701         mdsc->last_tid = 0;
3702         mdsc->oldest_tid = 0;
3703         mdsc->request_tree = RB_ROOT;
3704         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3705         mdsc->last_renew_caps = jiffies;
3706         INIT_LIST_HEAD(&mdsc->cap_delay_list);
3707         spin_lock_init(&mdsc->cap_delay_lock);
3708         INIT_LIST_HEAD(&mdsc->snap_flush_list);
3709         spin_lock_init(&mdsc->snap_flush_lock);
3710         mdsc->last_cap_flush_tid = 1;
3711         INIT_LIST_HEAD(&mdsc->cap_flush_list);
3712         INIT_LIST_HEAD(&mdsc->cap_dirty);
3713         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3714         mdsc->num_cap_flushing = 0;
3715         spin_lock_init(&mdsc->cap_dirty_lock);
3716         init_waitqueue_head(&mdsc->cap_flushing_wq);
3717         spin_lock_init(&mdsc->dentry_lru_lock);
3718         INIT_LIST_HEAD(&mdsc->dentry_lru);
3719
3720         ceph_caps_init(mdsc);
3721         ceph_adjust_min_caps(mdsc, fsc->min_caps);
3722
3723         init_rwsem(&mdsc->pool_perm_rwsem);
3724         mdsc->pool_perm_tree = RB_ROOT;
3725
3726         strscpy(mdsc->nodename, utsname()->nodename,
3727                 sizeof(mdsc->nodename));
3728
3729         fsc->mdsc = mdsc;
3730         return 0;
3731 }
3732
3733 /*
3734  * Wait for safe replies on open mds requests.  If we time out, drop
3735  * all requests from the tree to avoid dangling dentry refs.
3736  */
3737 static void wait_requests(struct ceph_mds_client *mdsc)
3738 {
3739         struct ceph_options *opts = mdsc->fsc->client->options;
3740         struct ceph_mds_request *req;
3741
3742         mutex_lock(&mdsc->mutex);
3743         if (__get_oldest_req(mdsc)) {
3744                 mutex_unlock(&mdsc->mutex);
3745
3746                 dout("wait_requests waiting for requests\n");
3747                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3748                                     ceph_timeout_jiffies(opts->mount_timeout));
3749
3750                 /* tear down remaining requests */
3751                 mutex_lock(&mdsc->mutex);
3752                 while ((req = __get_oldest_req(mdsc))) {
3753                         dout("wait_requests timed out on tid %llu\n",
3754                              req->r_tid);
3755                         __unregister_request(mdsc, req);
3756                 }
3757         }
3758         mutex_unlock(&mdsc->mutex);
3759         dout("wait_requests done\n");
3760 }
3761
3762 /*
3763  * called before mount is ro, and before dentries are torn down.
3764  * (hmm, does this still race with new lookups?)
3765  */
3766 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3767 {
3768         dout("pre_umount\n");
3769         mdsc->stopping = 1;
3770
3771         lock_unlock_sessions(mdsc);
3772         ceph_flush_dirty_caps(mdsc);
3773         wait_requests(mdsc);
3774
3775         /*
3776          * wait for reply handlers to drop their request refs and
3777          * their inode/dcache refs
3778          */
3779         ceph_msgr_flush();
3780 }
3781
3782 /*
3783  * wait for all write mds requests to flush.
3784  */
3785 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3786 {
3787         struct ceph_mds_request *req = NULL, *nextreq;
3788         struct rb_node *n;
3789
3790         mutex_lock(&mdsc->mutex);
3791         dout("wait_unsafe_requests want %lld\n", want_tid);
3792 restart:
3793         req = __get_oldest_req(mdsc);
3794         while (req && req->r_tid <= want_tid) {
3795                 /* find next request */
3796                 n = rb_next(&req->r_node);
3797                 if (n)
3798                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3799                 else
3800                         nextreq = NULL;
3801                 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3802                     (req->r_op & CEPH_MDS_OP_WRITE)) {
3803                         /* write op */
3804                         ceph_mdsc_get_request(req);
3805                         if (nextreq)
3806                                 ceph_mdsc_get_request(nextreq);
3807                         mutex_unlock(&mdsc->mutex);
3808                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
3809                              req->r_tid, want_tid);
3810                         wait_for_completion(&req->r_safe_completion);
3811                         mutex_lock(&mdsc->mutex);
3812                         ceph_mdsc_put_request(req);
3813                         if (!nextreq)
3814                                 break;  /* next dne before, so we're done! */
3815                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
3816                                 /* next request was removed from tree */
3817                                 ceph_mdsc_put_request(nextreq);
3818                                 goto restart;
3819                         }
3820                         ceph_mdsc_put_request(nextreq);  /* won't go away */
3821                 }
3822                 req = nextreq;
3823         }
3824         mutex_unlock(&mdsc->mutex);
3825         dout("wait_unsafe_requests done\n");
3826 }
3827
3828 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3829 {
3830         u64 want_tid, want_flush;
3831
3832         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3833                 return;
3834
3835         dout("sync\n");
3836         mutex_lock(&mdsc->mutex);
3837         want_tid = mdsc->last_tid;
3838         mutex_unlock(&mdsc->mutex);
3839
3840         ceph_flush_dirty_caps(mdsc);
3841         spin_lock(&mdsc->cap_dirty_lock);
3842         want_flush = mdsc->last_cap_flush_tid;
3843         if (!list_empty(&mdsc->cap_flush_list)) {
3844                 struct ceph_cap_flush *cf =
3845                         list_last_entry(&mdsc->cap_flush_list,
3846                                         struct ceph_cap_flush, g_list);
3847                 cf->wake = true;
3848         }
3849         spin_unlock(&mdsc->cap_dirty_lock);
3850
3851         dout("sync want tid %lld flush_seq %lld\n",
3852              want_tid, want_flush);
3853
3854         wait_unsafe_requests(mdsc, want_tid);
3855         wait_caps_flush(mdsc, want_flush);
3856 }
3857
3858 /*
3859  * true if all sessions are closed, or we force unmount
3860  */
3861 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
3862 {
3863         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3864                 return true;
3865         return atomic_read(&mdsc->num_sessions) <= skipped;
3866 }
3867
3868 /*
3869  * called after sb is ro.
3870  */
3871 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3872 {
3873         struct ceph_options *opts = mdsc->fsc->client->options;
3874         struct ceph_mds_session *session;
3875         int i;
3876         int skipped = 0;
3877
3878         dout("close_sessions\n");
3879
3880         /* close sessions */
3881         mutex_lock(&mdsc->mutex);
3882         for (i = 0; i < mdsc->max_sessions; i++) {
3883                 session = __ceph_lookup_mds_session(mdsc, i);
3884                 if (!session)
3885                         continue;
3886                 mutex_unlock(&mdsc->mutex);
3887                 mutex_lock(&session->s_mutex);
3888                 if (__close_session(mdsc, session) <= 0)
3889                         skipped++;
3890                 mutex_unlock(&session->s_mutex);
3891                 ceph_put_mds_session(session);
3892                 mutex_lock(&mdsc->mutex);
3893         }
3894         mutex_unlock(&mdsc->mutex);
3895
3896         dout("waiting for sessions to close\n");
3897         wait_event_timeout(mdsc->session_close_wq,
3898                            done_closing_sessions(mdsc, skipped),
3899                            ceph_timeout_jiffies(opts->mount_timeout));
3900
3901         /* tear down remaining sessions */
3902         mutex_lock(&mdsc->mutex);
3903         for (i = 0; i < mdsc->max_sessions; i++) {
3904                 if (mdsc->sessions[i]) {
3905                         session = get_session(mdsc->sessions[i]);
3906                         __unregister_session(mdsc, session);
3907                         mutex_unlock(&mdsc->mutex);
3908                         mutex_lock(&session->s_mutex);
3909                         remove_session_caps(session);
3910                         mutex_unlock(&session->s_mutex);
3911                         ceph_put_mds_session(session);
3912                         mutex_lock(&mdsc->mutex);
3913                 }
3914         }
3915         WARN_ON(!list_empty(&mdsc->cap_delay_list));
3916         mutex_unlock(&mdsc->mutex);
3917
3918         ceph_cleanup_empty_realms(mdsc);
3919
3920         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3921
3922         dout("stopped\n");
3923 }
3924
3925 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
3926 {
3927         struct ceph_mds_session *session;
3928         int mds;
3929
3930         dout("force umount\n");
3931
3932         mutex_lock(&mdsc->mutex);
3933         for (mds = 0; mds < mdsc->max_sessions; mds++) {
3934                 session = __ceph_lookup_mds_session(mdsc, mds);
3935                 if (!session)
3936                         continue;
3937                 mutex_unlock(&mdsc->mutex);
3938                 mutex_lock(&session->s_mutex);
3939                 __close_session(mdsc, session);
3940                 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
3941                         cleanup_session_requests(mdsc, session);
3942                         remove_session_caps(session);
3943                 }
3944                 mutex_unlock(&session->s_mutex);
3945                 ceph_put_mds_session(session);
3946                 mutex_lock(&mdsc->mutex);
3947                 kick_requests(mdsc, mds);
3948         }
3949         __wake_requests(mdsc, &mdsc->waiting_for_map);
3950         mutex_unlock(&mdsc->mutex);
3951 }
3952
3953 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3954 {
3955         dout("stop\n");
3956         /*
3957          * Make sure the delayed work stopped before releasing
3958          * the resources.
3959          *
3960          * Because the cancel_delayed_work_sync() will only
3961          * guarantee that the work finishes executing. But the
3962          * delayed work will re-arm itself again after that.
3963          */
3964         flush_delayed_work(&mdsc->delayed_work);
3965
3966         if (mdsc->mdsmap)
3967                 ceph_mdsmap_destroy(mdsc->mdsmap);
3968         kfree(mdsc->sessions);
3969         ceph_caps_finalize(mdsc);
3970         ceph_pool_perm_destroy(mdsc);
3971 }
3972
3973 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3974 {
3975         struct ceph_mds_client *mdsc = fsc->mdsc;
3976         dout("mdsc_destroy %p\n", mdsc);
3977
3978         if (!mdsc)
3979                 return;
3980
3981         /* flush out any connection work with references to us */
3982         ceph_msgr_flush();
3983
3984         ceph_mdsc_stop(mdsc);
3985
3986         fsc->mdsc = NULL;
3987         kfree(mdsc);
3988         dout("mdsc_destroy %p done\n", mdsc);
3989 }
3990
3991 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3992 {
3993         struct ceph_fs_client *fsc = mdsc->fsc;
3994         const char *mds_namespace = fsc->mount_options->mds_namespace;
3995         void *p = msg->front.iov_base;
3996         void *end = p + msg->front.iov_len;
3997         u32 epoch;
3998         u32 map_len;
3999         u32 num_fs;
4000         u32 mount_fscid = (u32)-1;
4001         u8 struct_v, struct_cv;
4002         int err = -EINVAL;
4003
4004         ceph_decode_need(&p, end, sizeof(u32), bad);
4005         epoch = ceph_decode_32(&p);
4006
4007         dout("handle_fsmap epoch %u\n", epoch);
4008
4009         ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4010         struct_v = ceph_decode_8(&p);
4011         struct_cv = ceph_decode_8(&p);
4012         map_len = ceph_decode_32(&p);
4013
4014         ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
4015         p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
4016
4017         num_fs = ceph_decode_32(&p);
4018         while (num_fs-- > 0) {
4019                 void *info_p, *info_end;
4020                 u32 info_len;
4021                 u8 info_v, info_cv;
4022                 u32 fscid, namelen;
4023
4024                 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4025                 info_v = ceph_decode_8(&p);
4026                 info_cv = ceph_decode_8(&p);
4027                 info_len = ceph_decode_32(&p);
4028                 ceph_decode_need(&p, end, info_len, bad);
4029                 info_p = p;
4030                 info_end = p + info_len;
4031                 p = info_end;
4032
4033                 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
4034                 fscid = ceph_decode_32(&info_p);
4035                 namelen = ceph_decode_32(&info_p);
4036                 ceph_decode_need(&info_p, info_end, namelen, bad);
4037
4038                 if (mds_namespace &&
4039                     strlen(mds_namespace) == namelen &&
4040                     !strncmp(mds_namespace, (char *)info_p, namelen)) {
4041                         mount_fscid = fscid;
4042                         break;
4043                 }
4044         }
4045
4046         ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
4047         if (mount_fscid != (u32)-1) {
4048                 fsc->client->monc.fs_cluster_id = mount_fscid;
4049                 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
4050                                    0, true);
4051                 ceph_monc_renew_subs(&fsc->client->monc);
4052         } else {
4053                 err = -ENOENT;
4054                 goto err_out;
4055         }
4056         return;
4057
4058 bad:
4059         pr_err("error decoding fsmap\n");
4060 err_out:
4061         mutex_lock(&mdsc->mutex);
4062         mdsc->mdsmap_err = err;
4063         __wake_requests(mdsc, &mdsc->waiting_for_map);
4064         mutex_unlock(&mdsc->mutex);
4065 }
4066
4067 /*
4068  * handle mds map update.
4069  */
4070 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4071 {
4072         u32 epoch;
4073         u32 maplen;
4074         void *p = msg->front.iov_base;
4075         void *end = p + msg->front.iov_len;
4076         struct ceph_mdsmap *newmap, *oldmap;
4077         struct ceph_fsid fsid;
4078         int err = -EINVAL;
4079
4080         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
4081         ceph_decode_copy(&p, &fsid, sizeof(fsid));
4082         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
4083                 return;
4084         epoch = ceph_decode_32(&p);
4085         maplen = ceph_decode_32(&p);
4086         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
4087
4088         /* do we need it? */
4089         mutex_lock(&mdsc->mutex);
4090         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
4091                 dout("handle_map epoch %u <= our %u\n",
4092                      epoch, mdsc->mdsmap->m_epoch);
4093                 mutex_unlock(&mdsc->mutex);
4094                 return;
4095         }
4096
4097         newmap = ceph_mdsmap_decode(&p, end);
4098         if (IS_ERR(newmap)) {
4099                 err = PTR_ERR(newmap);
4100                 goto bad_unlock;
4101         }
4102
4103         /* swap into place */
4104         if (mdsc->mdsmap) {
4105                 oldmap = mdsc->mdsmap;
4106                 mdsc->mdsmap = newmap;
4107                 check_new_map(mdsc, newmap, oldmap);
4108                 ceph_mdsmap_destroy(oldmap);
4109         } else {
4110                 mdsc->mdsmap = newmap;  /* first mds map */
4111         }
4112         mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
4113                                         MAX_LFS_FILESIZE);
4114
4115         __wake_requests(mdsc, &mdsc->waiting_for_map);
4116         ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
4117                           mdsc->mdsmap->m_epoch);
4118
4119         mutex_unlock(&mdsc->mutex);
4120         schedule_delayed(mdsc);
4121         return;
4122
4123 bad_unlock:
4124         mutex_unlock(&mdsc->mutex);
4125 bad:
4126         pr_err("error decoding mdsmap %d\n", err);
4127         return;
4128 }
4129
4130 static struct ceph_connection *con_get(struct ceph_connection *con)
4131 {
4132         struct ceph_mds_session *s = con->private;
4133
4134         if (get_session(s)) {
4135                 dout("mdsc con_get %p ok (%d)\n", s, refcount_read(&s->s_ref));
4136                 return con;
4137         }
4138         dout("mdsc con_get %p FAIL\n", s);
4139         return NULL;
4140 }
4141
4142 static void con_put(struct ceph_connection *con)
4143 {
4144         struct ceph_mds_session *s = con->private;
4145
4146         dout("mdsc con_put %p (%d)\n", s, refcount_read(&s->s_ref) - 1);
4147         ceph_put_mds_session(s);
4148 }
4149
4150 /*
4151  * if the client is unresponsive for long enough, the mds will kill
4152  * the session entirely.
4153  */
4154 static void peer_reset(struct ceph_connection *con)
4155 {
4156         struct ceph_mds_session *s = con->private;
4157         struct ceph_mds_client *mdsc = s->s_mdsc;
4158
4159         pr_warn("mds%d closed our session\n", s->s_mds);
4160         send_mds_reconnect(mdsc, s);
4161 }
4162
4163 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
4164 {
4165         struct ceph_mds_session *s = con->private;
4166         struct ceph_mds_client *mdsc = s->s_mdsc;
4167         int type = le16_to_cpu(msg->hdr.type);
4168
4169         mutex_lock(&mdsc->mutex);
4170         if (__verify_registered_session(mdsc, s) < 0) {
4171                 mutex_unlock(&mdsc->mutex);
4172                 goto out;
4173         }
4174         mutex_unlock(&mdsc->mutex);
4175
4176         switch (type) {
4177         case CEPH_MSG_MDS_MAP:
4178                 ceph_mdsc_handle_mdsmap(mdsc, msg);
4179                 break;
4180         case CEPH_MSG_FS_MAP_USER:
4181                 ceph_mdsc_handle_fsmap(mdsc, msg);
4182                 break;
4183         case CEPH_MSG_CLIENT_SESSION:
4184                 handle_session(s, msg);
4185                 break;
4186         case CEPH_MSG_CLIENT_REPLY:
4187                 handle_reply(s, msg);
4188                 break;
4189         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
4190                 handle_forward(mdsc, s, msg);
4191                 break;
4192         case CEPH_MSG_CLIENT_CAPS:
4193                 ceph_handle_caps(s, msg);
4194                 break;
4195         case CEPH_MSG_CLIENT_SNAP:
4196                 ceph_handle_snap(mdsc, s, msg);
4197                 break;
4198         case CEPH_MSG_CLIENT_LEASE:
4199                 handle_lease(mdsc, s, msg);
4200                 break;
4201         case CEPH_MSG_CLIENT_QUOTA:
4202                 ceph_handle_quota(mdsc, s, msg);
4203                 break;
4204
4205         default:
4206                 pr_err("received unknown message type %d %s\n", type,
4207                        ceph_msg_type_name(type));
4208         }
4209 out:
4210         ceph_msg_put(msg);
4211 }
4212
4213 /*
4214  * authentication
4215  */
4216
4217 /*
4218  * Note: returned pointer is the address of a structure that's
4219  * managed separately.  Caller must *not* attempt to free it.
4220  */
4221 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
4222                                         int *proto, int force_new)
4223 {
4224         struct ceph_mds_session *s = con->private;
4225         struct ceph_mds_client *mdsc = s->s_mdsc;
4226         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4227         struct ceph_auth_handshake *auth = &s->s_auth;
4228
4229         if (force_new && auth->authorizer) {
4230                 ceph_auth_destroy_authorizer(auth->authorizer);
4231                 auth->authorizer = NULL;
4232         }
4233         if (!auth->authorizer) {
4234                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4235                                                       auth);
4236                 if (ret)
4237                         return ERR_PTR(ret);
4238         } else {
4239                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4240                                                       auth);
4241                 if (ret)
4242                         return ERR_PTR(ret);
4243         }
4244         *proto = ac->protocol;
4245
4246         return auth;
4247 }
4248
4249 static int add_authorizer_challenge(struct ceph_connection *con,
4250                                     void *challenge_buf, int challenge_buf_len)
4251 {
4252         struct ceph_mds_session *s = con->private;
4253         struct ceph_mds_client *mdsc = s->s_mdsc;
4254         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4255
4256         return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
4257                                             challenge_buf, challenge_buf_len);
4258 }
4259
4260 static int verify_authorizer_reply(struct ceph_connection *con)
4261 {
4262         struct ceph_mds_session *s = con->private;
4263         struct ceph_mds_client *mdsc = s->s_mdsc;
4264         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4265
4266         return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
4267 }
4268
4269 static int invalidate_authorizer(struct ceph_connection *con)
4270 {
4271         struct ceph_mds_session *s = con->private;
4272         struct ceph_mds_client *mdsc = s->s_mdsc;
4273         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4274
4275         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
4276
4277         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
4278 }
4279
4280 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
4281                                 struct ceph_msg_header *hdr, int *skip)
4282 {
4283         struct ceph_msg *msg;
4284         int type = (int) le16_to_cpu(hdr->type);
4285         int front_len = (int) le32_to_cpu(hdr->front_len);
4286
4287         if (con->in_msg)
4288                 return con->in_msg;
4289
4290         *skip = 0;
4291         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
4292         if (!msg) {
4293                 pr_err("unable to allocate msg type %d len %d\n",
4294                        type, front_len);
4295                 return NULL;
4296         }
4297
4298         return msg;
4299 }
4300
4301 static int mds_sign_message(struct ceph_msg *msg)
4302 {
4303        struct ceph_mds_session *s = msg->con->private;
4304        struct ceph_auth_handshake *auth = &s->s_auth;
4305
4306        return ceph_auth_sign_message(auth, msg);
4307 }
4308
4309 static int mds_check_message_signature(struct ceph_msg *msg)
4310 {
4311        struct ceph_mds_session *s = msg->con->private;
4312        struct ceph_auth_handshake *auth = &s->s_auth;
4313
4314        return ceph_auth_check_message_signature(auth, msg);
4315 }
4316
4317 static const struct ceph_connection_operations mds_con_ops = {
4318         .get = con_get,
4319         .put = con_put,
4320         .dispatch = dispatch,
4321         .get_authorizer = get_authorizer,
4322         .add_authorizer_challenge = add_authorizer_challenge,
4323         .verify_authorizer_reply = verify_authorizer_reply,
4324         .invalidate_authorizer = invalidate_authorizer,
4325         .peer_reset = peer_reset,
4326         .alloc_msg = mds_alloc_msg,
4327         .sign_message = mds_sign_message,
4328         .check_message_signature = mds_check_message_signature,
4329 };
4330
4331 /* eof */