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  45  * are no longer valid.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         spin_unlock(&ci->i_ceph_lock);
1295         while (!list_empty(&to_remove)) {
1296                 struct ceph_cap_flush *cf;
1297                 cf = list_first_entry(&to_remove,
1298                                       struct ceph_cap_flush, i_list);
1299                 list_del(&cf->i_list);
1300                 ceph_free_cap_flush(cf);
1301         }
1302
1303         wake_up_all(&ci->i_cap_wq);
1304         if (invalidate)
1305                 ceph_queue_invalidate(inode);
1306         if (drop)
1307                 iput(inode);
1308         return 0;
1309 }
1310
1311 /*
1312  * caller must hold session s_mutex
1313  */
1314 static void remove_session_caps(struct ceph_mds_session *session)
1315 {
1316         struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1317         struct super_block *sb = fsc->sb;
1318         LIST_HEAD(dispose);
1319
1320         dout("remove_session_caps on %p\n", session);
1321         iterate_session_caps(session, remove_session_caps_cb, fsc);
1322
1323         wake_up_all(&fsc->mdsc->cap_flushing_wq);
1324
1325         spin_lock(&session->s_cap_lock);
1326         if (session->s_nr_caps > 0) {
1327                 struct inode *inode;
1328                 struct ceph_cap *cap, *prev = NULL;
1329                 struct ceph_vino vino;
1330                 /*
1331                  * iterate_session_caps() skips inodes that are being
1332                  * deleted, we need to wait until deletions are complete.
1333                  * __wait_on_freeing_inode() is designed for the job,
1334                  * but it is not exported, so use lookup inode function
1335                  * to access it.
1336                  */
1337                 while (!list_empty(&session->s_caps)) {
1338                         cap = list_entry(session->s_caps.next,
1339                                          struct ceph_cap, session_caps);
1340                         if (cap == prev)
1341                                 break;
1342                         prev = cap;
1343                         vino = cap->ci->i_vino;
1344                         spin_unlock(&session->s_cap_lock);
1345
1346                         inode = ceph_find_inode(sb, vino);
1347                         iput(inode);
1348
1349                         spin_lock(&session->s_cap_lock);
1350                 }
1351         }
1352
1353         // drop cap expires and unlock s_cap_lock
1354         detach_cap_releases(session, &dispose);
1355
1356         BUG_ON(session->s_nr_caps > 0);
1357         BUG_ON(!list_empty(&session->s_cap_flushing));
1358         spin_unlock(&session->s_cap_lock);
1359         dispose_cap_releases(session->s_mdsc, &dispose);
1360 }
1361
1362 /*
1363  * wake up any threads waiting on this session's caps.  if the cap is
1364  * old (didn't get renewed on the client reconnect), remove it now.
1365  *
1366  * caller must hold s_mutex.
1367  */
1368 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1369                               void *arg)
1370 {
1371         struct ceph_inode_info *ci = ceph_inode(inode);
1372
1373         if (arg) {
1374                 spin_lock(&ci->i_ceph_lock);
1375                 ci->i_wanted_max_size = 0;
1376                 ci->i_requested_max_size = 0;
1377                 spin_unlock(&ci->i_ceph_lock);
1378         }
1379         wake_up_all(&ci->i_cap_wq);
1380         return 0;
1381 }
1382
1383 static void wake_up_session_caps(struct ceph_mds_session *session,
1384                                  int reconnect)
1385 {
1386         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1387         iterate_session_caps(session, wake_up_session_cb,
1388                              (void *)(unsigned long)reconnect);
1389 }
1390
1391 /*
1392  * Send periodic message to MDS renewing all currently held caps.  The
1393  * ack will reset the expiration for all caps from this session.
1394  *
1395  * caller holds s_mutex
1396  */
1397 static int send_renew_caps(struct ceph_mds_client *mdsc,
1398                            struct ceph_mds_session *session)
1399 {
1400         struct ceph_msg *msg;
1401         int state;
1402
1403         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1404             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1405                 pr_info("mds%d caps stale\n", session->s_mds);
1406         session->s_renew_requested = jiffies;
1407
1408         /* do not try to renew caps until a recovering mds has reconnected
1409          * with its clients. */
1410         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1411         if (state < CEPH_MDS_STATE_RECONNECT) {
1412                 dout("send_renew_caps ignoring mds%d (%s)\n",
1413                      session->s_mds, ceph_mds_state_name(state));
1414                 return 0;
1415         }
1416
1417         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1418                 ceph_mds_state_name(state));
1419         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1420                                  ++session->s_renew_seq);
1421         if (!msg)
1422                 return -ENOMEM;
1423         ceph_con_send(&session->s_con, msg);
1424         return 0;
1425 }
1426
1427 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1428                              struct ceph_mds_session *session, u64 seq)
1429 {
1430         struct ceph_msg *msg;
1431
1432         dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1433              session->s_mds, ceph_session_state_name(session->s_state), seq);
1434         msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1435         if (!msg)
1436                 return -ENOMEM;
1437         ceph_con_send(&session->s_con, msg);
1438         return 0;
1439 }
1440
1441
1442 /*
1443  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1444  *
1445  * Called under session->s_mutex
1446  */
1447 static void renewed_caps(struct ceph_mds_client *mdsc,
1448                          struct ceph_mds_session *session, int is_renew)
1449 {
1450         int was_stale;
1451         int wake = 0;
1452
1453         spin_lock(&session->s_cap_lock);
1454         was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1455
1456         session->s_cap_ttl = session->s_renew_requested +
1457                 mdsc->mdsmap->m_session_timeout*HZ;
1458
1459         if (was_stale) {
1460                 if (time_before(jiffies, session->s_cap_ttl)) {
1461                         pr_info("mds%d caps renewed\n", session->s_mds);
1462                         wake = 1;
1463                 } else {
1464                         pr_info("mds%d caps still stale\n", session->s_mds);
1465                 }
1466         }
1467         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1468              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1469              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1470         spin_unlock(&session->s_cap_lock);
1471
1472         if (wake)
1473                 wake_up_session_caps(session, 0);
1474 }
1475
1476 /*
1477  * send a session close request
1478  */
1479 static int request_close_session(struct ceph_mds_client *mdsc,
1480                                  struct ceph_mds_session *session)
1481 {
1482         struct ceph_msg *msg;
1483
1484         dout("request_close_session mds%d state %s seq %lld\n",
1485              session->s_mds, ceph_session_state_name(session->s_state),
1486              session->s_seq);
1487         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1488         if (!msg)
1489                 return -ENOMEM;
1490         ceph_con_send(&session->s_con, msg);
1491         return 1;
1492 }
1493
1494 /*
1495  * Called with s_mutex held.
1496  */
1497 static int __close_session(struct ceph_mds_client *mdsc,
1498                          struct ceph_mds_session *session)
1499 {
1500         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1501                 return 0;
1502         session->s_state = CEPH_MDS_SESSION_CLOSING;
1503         return request_close_session(mdsc, session);
1504 }
1505
1506 static bool drop_negative_children(struct dentry *dentry)
1507 {
1508         struct dentry *child;
1509         bool all_negative = true;
1510
1511         if (!d_is_dir(dentry))
1512                 goto out;
1513
1514         spin_lock(&dentry->d_lock);
1515         list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1516                 if (d_really_is_positive(child)) {
1517                         all_negative = false;
1518                         break;
1519                 }
1520         }
1521         spin_unlock(&dentry->d_lock);
1522
1523         if (all_negative)
1524                 shrink_dcache_parent(dentry);
1525 out:
1526         return all_negative;
1527 }
1528
1529 /*
1530  * Trim old(er) caps.
1531  *
1532  * Because we can't cache an inode without one or more caps, we do
1533  * this indirectly: if a cap is unused, we prune its aliases, at which
1534  * point the inode will hopefully get dropped to.
1535  *
1536  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1537  * memory pressure from the MDS, though, so it needn't be perfect.
1538  */
1539 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1540 {
1541         struct ceph_mds_session *session = arg;
1542         struct ceph_inode_info *ci = ceph_inode(inode);
1543         int used, wanted, oissued, mine;
1544
1545         if (session->s_trim_caps <= 0)
1546                 return -1;
1547
1548         spin_lock(&ci->i_ceph_lock);
1549         mine = cap->issued | cap->implemented;
1550         used = __ceph_caps_used(ci);
1551         wanted = __ceph_caps_file_wanted(ci);
1552         oissued = __ceph_caps_issued_other(ci, cap);
1553
1554         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1555              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1556              ceph_cap_string(used), ceph_cap_string(wanted));
1557         if (cap == ci->i_auth_cap) {
1558                 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1559                     !list_empty(&ci->i_cap_snaps))
1560                         goto out;
1561                 if ((used | wanted) & CEPH_CAP_ANY_WR)
1562                         goto out;
1563                 /* Note: it's possible that i_filelock_ref becomes non-zero
1564                  * after dropping auth caps. It doesn't hurt because reply
1565                  * of lock mds request will re-add auth caps. */
1566                 if (atomic_read(&ci->i_filelock_ref) > 0)
1567                         goto out;
1568         }
1569         /* The inode has cached pages, but it's no longer used.
1570          * we can safely drop it */
1571         if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1572             !(oissued & CEPH_CAP_FILE_CACHE)) {
1573           used = 0;
1574           oissued = 0;
1575         }
1576         if ((used | wanted) & ~oissued & mine)
1577                 goto out;   /* we need these caps */
1578
1579         if (oissued) {
1580                 /* we aren't the only cap.. just remove us */
1581                 __ceph_remove_cap(cap, true);
1582                 session->s_trim_caps--;
1583         } else {
1584                 struct dentry *dentry;
1585                 /* try dropping referring dentries */
1586                 spin_unlock(&ci->i_ceph_lock);
1587                 dentry = d_find_any_alias(inode);
1588                 if (dentry && drop_negative_children(dentry)) {
1589                         int count;
1590                         dput(dentry);
1591                         d_prune_aliases(inode);
1592                         count = atomic_read(&inode->i_count);
1593                         if (count == 1)
1594                                 session->s_trim_caps--;
1595                         dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1596                              inode, cap, count);
1597                 } else {
1598                         dput(dentry);
1599                 }
1600                 return 0;
1601         }
1602
1603 out:
1604         spin_unlock(&ci->i_ceph_lock);
1605         return 0;
1606 }
1607
1608 /*
1609  * Trim session cap count down to some max number.
1610  */
1611 int ceph_trim_caps(struct ceph_mds_client *mdsc,
1612                    struct ceph_mds_session *session,
1613                    int max_caps)
1614 {
1615         int trim_caps = session->s_nr_caps - max_caps;
1616
1617         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1618              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1619         if (trim_caps > 0) {
1620                 session->s_trim_caps = trim_caps;
1621                 iterate_session_caps(session, trim_caps_cb, session);
1622                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1623                      session->s_mds, session->s_nr_caps, max_caps,
1624                         trim_caps - session->s_trim_caps);
1625                 session->s_trim_caps = 0;
1626         }
1627
1628         ceph_send_cap_releases(mdsc, session);
1629         return 0;
1630 }
1631
1632 static int check_caps_flush(struct ceph_mds_client *mdsc,
1633                             u64 want_flush_tid)
1634 {
1635         int ret = 1;
1636
1637         spin_lock(&mdsc->cap_dirty_lock);
1638         if (!list_empty(&mdsc->cap_flush_list)) {
1639                 struct ceph_cap_flush *cf =
1640                         list_first_entry(&mdsc->cap_flush_list,
1641                                          struct ceph_cap_flush, g_list);
1642                 if (cf->tid <= want_flush_tid) {
1643                         dout("check_caps_flush still flushing tid "
1644                              "%llu <= %llu\n", cf->tid, want_flush_tid);
1645                         ret = 0;
1646                 }
1647         }
1648         spin_unlock(&mdsc->cap_dirty_lock);
1649         return ret;
1650 }
1651
1652 /*
1653  * flush all dirty inode data to disk.
1654  *
1655  * returns true if we've flushed through want_flush_tid
1656  */
1657 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1658                             u64 want_flush_tid)
1659 {
1660         dout("check_caps_flush want %llu\n", want_flush_tid);
1661
1662         wait_event(mdsc->cap_flushing_wq,
1663                    check_caps_flush(mdsc, want_flush_tid));
1664
1665         dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1666 }
1667
1668 /*
1669  * called under s_mutex
1670  */
1671 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1672                             struct ceph_mds_session *session)
1673 {
1674         struct ceph_msg *msg = NULL;
1675         struct ceph_mds_cap_release *head;
1676         struct ceph_mds_cap_item *item;
1677         struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
1678         struct ceph_cap *cap;
1679         LIST_HEAD(tmp_list);
1680         int num_cap_releases;
1681         __le32  barrier, *cap_barrier;
1682
1683         down_read(&osdc->lock);
1684         barrier = cpu_to_le32(osdc->epoch_barrier);
1685         up_read(&osdc->lock);
1686
1687         spin_lock(&session->s_cap_lock);
1688 again:
1689         list_splice_init(&session->s_cap_releases, &tmp_list);
1690         num_cap_releases = session->s_num_cap_releases;
1691         session->s_num_cap_releases = 0;
1692         spin_unlock(&session->s_cap_lock);
1693
1694         while (!list_empty(&tmp_list)) {
1695                 if (!msg) {
1696                         msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1697                                         PAGE_SIZE, GFP_NOFS, false);
1698                         if (!msg)
1699                                 goto out_err;
1700                         head = msg->front.iov_base;
1701                         head->num = cpu_to_le32(0);
1702                         msg->front.iov_len = sizeof(*head);
1703
1704                         msg->hdr.version = cpu_to_le16(2);
1705                         msg->hdr.compat_version = cpu_to_le16(1);
1706                 }
1707
1708                 cap = list_first_entry(&tmp_list, struct ceph_cap,
1709                                         session_caps);
1710                 list_del(&cap->session_caps);
1711                 num_cap_releases--;
1712
1713                 head = msg->front.iov_base;
1714                 le32_add_cpu(&head->num, 1);
1715                 item = msg->front.iov_base + msg->front.iov_len;
1716                 item->ino = cpu_to_le64(cap->cap_ino);
1717                 item->cap_id = cpu_to_le64(cap->cap_id);
1718                 item->migrate_seq = cpu_to_le32(cap->mseq);
1719                 item->seq = cpu_to_le32(cap->issue_seq);
1720                 msg->front.iov_len += sizeof(*item);
1721
1722                 ceph_put_cap(mdsc, cap);
1723
1724                 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1725                         // Append cap_barrier field
1726                         cap_barrier = msg->front.iov_base + msg->front.iov_len;
1727                         *cap_barrier = barrier;
1728                         msg->front.iov_len += sizeof(*cap_barrier);
1729
1730                         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1731                         dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1732                         ceph_con_send(&session->s_con, msg);
1733                         msg = NULL;
1734                 }
1735         }
1736
1737         BUG_ON(num_cap_releases != 0);
1738
1739         spin_lock(&session->s_cap_lock);
1740         if (!list_empty(&session->s_cap_releases))
1741                 goto again;
1742         spin_unlock(&session->s_cap_lock);
1743
1744         if (msg) {
1745                 // Append cap_barrier field
1746                 cap_barrier = msg->front.iov_base + msg->front.iov_len;
1747                 *cap_barrier = barrier;
1748                 msg->front.iov_len += sizeof(*cap_barrier);
1749
1750                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1751                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1752                 ceph_con_send(&session->s_con, msg);
1753         }
1754         return;
1755 out_err:
1756         pr_err("send_cap_releases mds%d, failed to allocate message\n",
1757                 session->s_mds);
1758         spin_lock(&session->s_cap_lock);
1759         list_splice(&tmp_list, &session->s_cap_releases);
1760         session->s_num_cap_releases += num_cap_releases;
1761         spin_unlock(&session->s_cap_lock);
1762 }
1763
1764 /*
1765  * requests
1766  */
1767
1768 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1769                                     struct inode *dir)
1770 {
1771         struct ceph_inode_info *ci = ceph_inode(dir);
1772         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1773         struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1774         size_t size = sizeof(struct ceph_mds_reply_dir_entry);
1775         int order, num_entries;
1776
1777         spin_lock(&ci->i_ceph_lock);
1778         num_entries = ci->i_files + ci->i_subdirs;
1779         spin_unlock(&ci->i_ceph_lock);
1780         num_entries = max(num_entries, 1);
1781         num_entries = min(num_entries, opt->max_readdir);
1782
1783         order = get_order(size * num_entries);
1784         while (order >= 0) {
1785                 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
1786                                                              __GFP_NOWARN,
1787                                                              order);
1788                 if (rinfo->dir_entries)
1789                         break;
1790                 order--;
1791         }
1792         if (!rinfo->dir_entries)
1793                 return -ENOMEM;
1794
1795         num_entries = (PAGE_SIZE << order) / size;
1796         num_entries = min(num_entries, opt->max_readdir);
1797
1798         rinfo->dir_buf_size = PAGE_SIZE << order;
1799         req->r_num_caps = num_entries + 1;
1800         req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1801         req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1802         return 0;
1803 }
1804
1805 /*
1806  * Create an mds request.
1807  */
1808 struct ceph_mds_request *
1809 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1810 {
1811         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1812         struct timespec64 ts;
1813
1814         if (!req)
1815                 return ERR_PTR(-ENOMEM);
1816
1817         mutex_init(&req->r_fill_mutex);
1818         req->r_mdsc = mdsc;
1819         req->r_started = jiffies;
1820         req->r_resend_mds = -1;
1821         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1822         INIT_LIST_HEAD(&req->r_unsafe_target_item);
1823         req->r_fmode = -1;
1824         kref_init(&req->r_kref);
1825         RB_CLEAR_NODE(&req->r_node);
1826         INIT_LIST_HEAD(&req->r_wait);
1827         init_completion(&req->r_completion);
1828         init_completion(&req->r_safe_completion);
1829         INIT_LIST_HEAD(&req->r_unsafe_item);
1830
1831         ktime_get_coarse_real_ts64(&ts);
1832         req->r_stamp = timespec64_trunc(ts, mdsc->fsc->sb->s_time_gran);
1833
1834         req->r_op = op;
1835         req->r_direct_mode = mode;
1836         return req;
1837 }
1838
1839 /*
1840  * return oldest (lowest) request, tid in request tree, 0 if none.
1841  *
1842  * called under mdsc->mutex.
1843  */
1844 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1845 {
1846         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1847                 return NULL;
1848         return rb_entry(rb_first(&mdsc->request_tree),
1849                         struct ceph_mds_request, r_node);
1850 }
1851
1852 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1853 {
1854         return mdsc->oldest_tid;
1855 }
1856
1857 /*
1858  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1859  * on build_path_from_dentry in fs/cifs/dir.c.
1860  *
1861  * If @stop_on_nosnap, generate path relative to the first non-snapped
1862  * inode.
1863  *
1864  * Encode hidden .snap dirs as a double /, i.e.
1865  *   foo/.snap/bar -> foo//bar
1866  */
1867 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1868                            int stop_on_nosnap)
1869 {
1870         struct dentry *temp;
1871         char *path;
1872         int len, pos;
1873         unsigned seq;
1874
1875         if (!dentry)
1876                 return ERR_PTR(-EINVAL);
1877
1878 retry:
1879         len = 0;
1880         seq = read_seqbegin(&rename_lock);
1881         rcu_read_lock();
1882         for (temp = dentry; !IS_ROOT(temp);) {
1883                 struct inode *inode = d_inode(temp);
1884                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1885                         len++;  /* slash only */
1886                 else if (stop_on_nosnap && inode &&
1887                          ceph_snap(inode) == CEPH_NOSNAP)
1888                         break;
1889                 else
1890                         len += 1 + temp->d_name.len;
1891                 temp = temp->d_parent;
1892         }
1893         rcu_read_unlock();
1894         if (len)
1895                 len--;  /* no leading '/' */
1896
1897         path = kmalloc(len+1, GFP_NOFS);
1898         if (!path)
1899                 return ERR_PTR(-ENOMEM);
1900         pos = len;
1901         path[pos] = 0;  /* trailing null */
1902         rcu_read_lock();
1903         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1904                 struct inode *inode;
1905
1906                 spin_lock(&temp->d_lock);
1907                 inode = d_inode(temp);
1908                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1909                         dout("build_path path+%d: %p SNAPDIR\n",
1910                              pos, temp);
1911                 } else if (stop_on_nosnap && inode &&
1912                            ceph_snap(inode) == CEPH_NOSNAP) {
1913                         spin_unlock(&temp->d_lock);
1914                         break;
1915                 } else {
1916                         pos -= temp->d_name.len;
1917                         if (pos < 0) {
1918                                 spin_unlock(&temp->d_lock);
1919                                 break;
1920                         }
1921                         strncpy(path + pos, temp->d_name.name,
1922                                 temp->d_name.len);
1923                 }
1924                 spin_unlock(&temp->d_lock);
1925                 if (pos)
1926                         path[--pos] = '/';
1927                 temp = temp->d_parent;
1928         }
1929         rcu_read_unlock();
1930         if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1931                 pr_err("build_path did not end path lookup where "
1932                        "expected, namelen is %d, pos is %d\n", len, pos);
1933                 /* presumably this is only possible if racing with a
1934                    rename of one of the parent directories (we can not
1935                    lock the dentries above us to prevent this, but
1936                    retrying should be harmless) */
1937                 kfree(path);
1938                 goto retry;
1939         }
1940
1941         *base = ceph_ino(d_inode(temp));
1942         *plen = len;
1943         dout("build_path on %p %d built %llx '%.*s'\n",
1944              dentry, d_count(dentry), *base, len, path);
1945         return path;
1946 }
1947
1948 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
1949                              const char **ppath, int *ppathlen, u64 *pino,
1950                              int *pfreepath)
1951 {
1952         char *path;
1953
1954         rcu_read_lock();
1955         if (!dir)
1956                 dir = d_inode_rcu(dentry->d_parent);
1957         if (dir && ceph_snap(dir) == CEPH_NOSNAP) {
1958                 *pino = ceph_ino(dir);
1959                 rcu_read_unlock();
1960                 *ppath = dentry->d_name.name;
1961                 *ppathlen = dentry->d_name.len;
1962                 return 0;
1963         }
1964         rcu_read_unlock();
1965         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1966         if (IS_ERR(path))
1967                 return PTR_ERR(path);
1968         *ppath = path;
1969         *pfreepath = 1;
1970         return 0;
1971 }
1972
1973 static int build_inode_path(struct inode *inode,
1974                             const char **ppath, int *ppathlen, u64 *pino,
1975                             int *pfreepath)
1976 {
1977         struct dentry *dentry;
1978         char *path;
1979
1980         if (ceph_snap(inode) == CEPH_NOSNAP) {
1981                 *pino = ceph_ino(inode);
1982                 *ppathlen = 0;
1983                 return 0;
1984         }
1985         dentry = d_find_alias(inode);
1986         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1987         dput(dentry);
1988         if (IS_ERR(path))
1989                 return PTR_ERR(path);
1990         *ppath = path;
1991         *pfreepath = 1;
1992         return 0;
1993 }
1994
1995 /*
1996  * request arguments may be specified via an inode *, a dentry *, or
1997  * an explicit ino+path.
1998  */
1999 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
2000                                   struct inode *rdiri, const char *rpath,
2001                                   u64 rino, const char **ppath, int *pathlen,
2002                                   u64 *ino, int *freepath)
2003 {
2004         int r = 0;
2005
2006         if (rinode) {
2007                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2008                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2009                      ceph_snap(rinode));
2010         } else if (rdentry) {
2011                 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
2012                                         freepath);
2013                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
2014                      *ppath);
2015         } else if (rpath || rino) {
2016                 *ino = rino;
2017                 *ppath = rpath;
2018                 *pathlen = rpath ? strlen(rpath) : 0;
2019                 dout(" path %.*s\n", *pathlen, rpath);
2020         }
2021
2022         return r;
2023 }
2024
2025 /*
2026  * called under mdsc->mutex
2027  */
2028 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
2029                                                struct ceph_mds_request *req,
2030                                                int mds, bool drop_cap_releases)
2031 {
2032         struct ceph_msg *msg;
2033         struct ceph_mds_request_head *head;
2034         const char *path1 = NULL;
2035         const char *path2 = NULL;
2036         u64 ino1 = 0, ino2 = 0;
2037         int pathlen1 = 0, pathlen2 = 0;
2038         int freepath1 = 0, freepath2 = 0;
2039         int len;
2040         u16 releases;
2041         void *p, *end;
2042         int ret;
2043
2044         ret = set_request_path_attr(req->r_inode, req->r_dentry,
2045                               req->r_parent, req->r_path1, req->r_ino1.ino,
2046                               &path1, &pathlen1, &ino1, &freepath1);
2047         if (ret < 0) {
2048                 msg = ERR_PTR(ret);
2049                 goto out;
2050         }
2051
2052         ret = set_request_path_attr(NULL, req->r_old_dentry,
2053                               req->r_old_dentry_dir,
2054                               req->r_path2, req->r_ino2.ino,
2055                               &path2, &pathlen2, &ino2, &freepath2);
2056         if (ret < 0) {
2057                 msg = ERR_PTR(ret);
2058                 goto out_free1;
2059         }
2060
2061         len = sizeof(*head) +
2062                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2063                 sizeof(struct ceph_timespec);
2064
2065         /* calculate (max) length for cap releases */
2066         len += sizeof(struct ceph_mds_request_release) *
2067                 (!!req->r_inode_drop + !!req->r_dentry_drop +
2068                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2069         if (req->r_dentry_drop)
2070                 len += req->r_dentry->d_name.len;
2071         if (req->r_old_dentry_drop)
2072                 len += req->r_old_dentry->d_name.len;
2073
2074         msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
2075         if (!msg) {
2076                 msg = ERR_PTR(-ENOMEM);
2077                 goto out_free2;
2078         }
2079
2080         msg->hdr.version = cpu_to_le16(2);
2081         msg->hdr.tid = cpu_to_le64(req->r_tid);
2082
2083         head = msg->front.iov_base;
2084         p = msg->front.iov_base + sizeof(*head);
2085         end = msg->front.iov_base + msg->front.iov_len;
2086
2087         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2088         head->op = cpu_to_le32(req->r_op);
2089         head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
2090         head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
2091         head->args = req->r_args;
2092
2093         ceph_encode_filepath(&p, end, ino1, path1);
2094         ceph_encode_filepath(&p, end, ino2, path2);
2095
2096         /* make note of release offset, in case we need to replay */
2097         req->r_request_release_offset = p - msg->front.iov_base;
2098
2099         /* cap releases */
2100         releases = 0;
2101         if (req->r_inode_drop)
2102                 releases += ceph_encode_inode_release(&p,
2103                       req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2104                       mds, req->r_inode_drop, req->r_inode_unless, 0);
2105         if (req->r_dentry_drop)
2106                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
2107                                 req->r_parent, mds, req->r_dentry_drop,
2108                                 req->r_dentry_unless);
2109         if (req->r_old_dentry_drop)
2110                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2111                                 req->r_old_dentry_dir, mds,
2112                                 req->r_old_dentry_drop,
2113                                 req->r_old_dentry_unless);
2114         if (req->r_old_inode_drop)
2115                 releases += ceph_encode_inode_release(&p,
2116                       d_inode(req->r_old_dentry),
2117                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2118
2119         if (drop_cap_releases) {
2120                 releases = 0;
2121                 p = msg->front.iov_base + req->r_request_release_offset;
2122         }
2123
2124         head->num_releases = cpu_to_le16(releases);
2125
2126         /* time stamp */
2127         {
2128                 struct ceph_timespec ts;
2129                 ceph_encode_timespec64(&ts, &req->r_stamp);
2130                 ceph_encode_copy(&p, &ts, sizeof(ts));
2131         }
2132
2133         BUG_ON(p > end);
2134         msg->front.iov_len = p - msg->front.iov_base;
2135         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2136
2137         if (req->r_pagelist) {
2138                 struct ceph_pagelist *pagelist = req->r_pagelist;
2139                 ceph_msg_data_add_pagelist(msg, pagelist);
2140                 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2141         } else {
2142                 msg->hdr.data_len = 0;
2143         }
2144
2145         msg->hdr.data_off = cpu_to_le16(0);
2146
2147 out_free2:
2148         if (freepath2)
2149                 kfree((char *)path2);
2150 out_free1:
2151         if (freepath1)
2152                 kfree((char *)path1);
2153 out:
2154         return msg;
2155 }
2156
2157 /*
2158  * called under mdsc->mutex if error, under no mutex if
2159  * success.
2160  */
2161 static void complete_request(struct ceph_mds_client *mdsc,
2162                              struct ceph_mds_request *req)
2163 {
2164         if (req->r_callback)
2165                 req->r_callback(mdsc, req);
2166         else
2167                 complete_all(&req->r_completion);
2168 }
2169
2170 /*
2171  * called under mdsc->mutex
2172  */
2173 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2174                                   struct ceph_mds_request *req,
2175                                   int mds, bool drop_cap_releases)
2176 {
2177         struct ceph_mds_request_head *rhead;
2178         struct ceph_msg *msg;
2179         int flags = 0;
2180
2181         req->r_attempts++;
2182         if (req->r_inode) {
2183                 struct ceph_cap *cap =
2184                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2185
2186                 if (cap)
2187                         req->r_sent_on_mseq = cap->mseq;
2188                 else
2189                         req->r_sent_on_mseq = -1;
2190         }
2191         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2192              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2193
2194         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2195                 void *p;
2196                 /*
2197                  * Replay.  Do not regenerate message (and rebuild
2198                  * paths, etc.); just use the original message.
2199                  * Rebuilding paths will break for renames because
2200                  * d_move mangles the src name.
2201                  */
2202                 msg = req->r_request;
2203                 rhead = msg->front.iov_base;
2204
2205                 flags = le32_to_cpu(rhead->flags);
2206                 flags |= CEPH_MDS_FLAG_REPLAY;
2207                 rhead->flags = cpu_to_le32(flags);
2208
2209                 if (req->r_target_inode)
2210                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2211
2212                 rhead->num_retry = req->r_attempts - 1;
2213
2214                 /* remove cap/dentry releases from message */
2215                 rhead->num_releases = 0;
2216
2217                 /* time stamp */
2218                 p = msg->front.iov_base + req->r_request_release_offset;
2219                 {
2220                         struct ceph_timespec ts;
2221                         ceph_encode_timespec64(&ts, &req->r_stamp);
2222                         ceph_encode_copy(&p, &ts, sizeof(ts));
2223                 }
2224
2225                 msg->front.iov_len = p - msg->front.iov_base;
2226                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2227                 return 0;
2228         }
2229
2230         if (req->r_request) {
2231                 ceph_msg_put(req->r_request);
2232                 req->r_request = NULL;
2233         }
2234         msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2235         if (IS_ERR(msg)) {
2236                 req->r_err = PTR_ERR(msg);
2237                 return PTR_ERR(msg);
2238         }
2239         req->r_request = msg;
2240
2241         rhead = msg->front.iov_base;
2242         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2243         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2244                 flags |= CEPH_MDS_FLAG_REPLAY;
2245         if (req->r_parent)
2246                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2247         rhead->flags = cpu_to_le32(flags);
2248         rhead->num_fwd = req->r_num_fwd;
2249         rhead->num_retry = req->r_attempts - 1;
2250         rhead->ino = 0;
2251
2252         dout(" r_parent = %p\n", req->r_parent);
2253         return 0;
2254 }
2255
2256 /*
2257  * send request, or put it on the appropriate wait list.
2258  */
2259 static void __do_request(struct ceph_mds_client *mdsc,
2260                         struct ceph_mds_request *req)
2261 {
2262         struct ceph_mds_session *session = NULL;
2263         int mds = -1;
2264         int err = 0;
2265
2266         if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2267                 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2268                         __unregister_request(mdsc, req);
2269                 return;
2270         }
2271
2272         if (req->r_timeout &&
2273             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2274                 dout("do_request timed out\n");
2275                 err = -EIO;
2276                 goto finish;
2277         }
2278         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2279                 dout("do_request forced umount\n");
2280                 err = -EIO;
2281                 goto finish;
2282         }
2283         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2284                 if (mdsc->mdsmap_err) {
2285                         err = mdsc->mdsmap_err;
2286                         dout("do_request mdsmap err %d\n", err);
2287                         goto finish;
2288                 }
2289                 if (mdsc->mdsmap->m_epoch == 0) {
2290                         dout("do_request no mdsmap, waiting for map\n");
2291                         list_add(&req->r_wait, &mdsc->waiting_for_map);
2292                         return;
2293                 }
2294                 if (!(mdsc->fsc->mount_options->flags &
2295                       CEPH_MOUNT_OPT_MOUNTWAIT) &&
2296                     !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2297                         err = -ENOENT;
2298                         pr_info("probably no mds server is up\n");
2299                         goto finish;
2300                 }
2301         }
2302
2303         put_request_session(req);
2304
2305         mds = __choose_mds(mdsc, req);
2306         if (mds < 0 ||
2307             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2308                 dout("do_request no mds or not active, waiting for map\n");
2309                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2310                 return;
2311         }
2312
2313         /* get, open session */
2314         session = __ceph_lookup_mds_session(mdsc, mds);
2315         if (!session) {
2316                 session = register_session(mdsc, mds);
2317                 if (IS_ERR(session)) {
2318                         err = PTR_ERR(session);
2319                         goto finish;
2320                 }
2321         }
2322         req->r_session = get_session(session);
2323
2324         dout("do_request mds%d session %p state %s\n", mds, session,
2325              ceph_session_state_name(session->s_state));
2326         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2327             session->s_state != CEPH_MDS_SESSION_HUNG) {
2328                 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2329                         err = -EACCES;
2330                         goto out_session;
2331                 }
2332                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2333                     session->s_state == CEPH_MDS_SESSION_CLOSING)
2334                         __open_session(mdsc, session);
2335                 list_add(&req->r_wait, &session->s_waiting);
2336                 goto out_session;
2337         }
2338
2339         /* send request */
2340         req->r_resend_mds = -1;   /* forget any previous mds hint */
2341
2342         if (req->r_request_started == 0)   /* note request start time */
2343                 req->r_request_started = jiffies;
2344
2345         err = __prepare_send_request(mdsc, req, mds, false);
2346         if (!err) {
2347                 ceph_msg_get(req->r_request);
2348                 ceph_con_send(&session->s_con, req->r_request);
2349         }
2350
2351 out_session:
2352         ceph_put_mds_session(session);
2353 finish:
2354         if (err) {
2355                 dout("__do_request early error %d\n", err);
2356                 req->r_err = err;
2357                 complete_request(mdsc, req);
2358                 __unregister_request(mdsc, req);
2359         }
2360         return;
2361 }
2362
2363 /*
2364  * called under mdsc->mutex
2365  */
2366 static void __wake_requests(struct ceph_mds_client *mdsc,
2367                             struct list_head *head)
2368 {
2369         struct ceph_mds_request *req;
2370         LIST_HEAD(tmp_list);
2371
2372         list_splice_init(head, &tmp_list);
2373
2374         while (!list_empty(&tmp_list)) {
2375                 req = list_entry(tmp_list.next,
2376                                  struct ceph_mds_request, r_wait);
2377                 list_del_init(&req->r_wait);
2378                 dout(" wake request %p tid %llu\n", req, req->r_tid);
2379                 __do_request(mdsc, req);
2380         }
2381 }
2382
2383 /*
2384  * Wake up threads with requests pending for @mds, so that they can
2385  * resubmit their requests to a possibly different mds.
2386  */
2387 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2388 {
2389         struct ceph_mds_request *req;
2390         struct rb_node *p = rb_first(&mdsc->request_tree);
2391
2392         dout("kick_requests mds%d\n", mds);
2393         while (p) {
2394                 req = rb_entry(p, struct ceph_mds_request, r_node);
2395                 p = rb_next(p);
2396                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2397                         continue;
2398                 if (req->r_attempts > 0)
2399                         continue; /* only new requests */
2400                 if (req->r_session &&
2401                     req->r_session->s_mds == mds) {
2402                         dout(" kicking tid %llu\n", req->r_tid);
2403                         list_del_init(&req->r_wait);
2404                         __do_request(mdsc, req);
2405                 }
2406         }
2407 }
2408
2409 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2410                               struct ceph_mds_request *req)
2411 {
2412         dout("submit_request on %p\n", req);
2413         mutex_lock(&mdsc->mutex);
2414         __register_request(mdsc, req, NULL);
2415         __do_request(mdsc, req);
2416         mutex_unlock(&mdsc->mutex);
2417 }
2418
2419 /*
2420  * Synchrously perform an mds request.  Take care of all of the
2421  * session setup, forwarding, retry details.
2422  */
2423 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2424                          struct inode *dir,
2425                          struct ceph_mds_request *req)
2426 {
2427         int err;
2428
2429         dout("do_request on %p\n", req);
2430
2431         /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2432         if (req->r_inode)
2433                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2434         if (req->r_parent)
2435                 ceph_get_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
2436         if (req->r_old_dentry_dir)
2437                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2438                                   CEPH_CAP_PIN);
2439
2440         /* issue */
2441         mutex_lock(&mdsc->mutex);
2442         __register_request(mdsc, req, dir);
2443         __do_request(mdsc, req);
2444
2445         if (req->r_err) {
2446                 err = req->r_err;
2447                 goto out;
2448         }
2449
2450         /* wait */
2451         mutex_unlock(&mdsc->mutex);
2452         dout("do_request waiting\n");
2453         if (!req->r_timeout && req->r_wait_for_completion) {
2454                 err = req->r_wait_for_completion(mdsc, req);
2455         } else {
2456                 long timeleft = wait_for_completion_killable_timeout(
2457                                         &req->r_completion,
2458                                         ceph_timeout_jiffies(req->r_timeout));
2459                 if (timeleft > 0)
2460                         err = 0;
2461                 else if (!timeleft)
2462                         err = -EIO;  /* timed out */
2463                 else
2464                         err = timeleft;  /* killed */
2465         }
2466         dout("do_request waited, got %d\n", err);
2467         mutex_lock(&mdsc->mutex);
2468
2469         /* only abort if we didn't race with a real reply */
2470         if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2471                 err = le32_to_cpu(req->r_reply_info.head->result);
2472         } else if (err < 0) {
2473                 dout("aborted request %lld with %d\n", req->r_tid, err);
2474
2475                 /*
2476                  * ensure we aren't running concurrently with
2477                  * ceph_fill_trace or ceph_readdir_prepopulate, which
2478                  * rely on locks (dir mutex) held by our caller.
2479                  */
2480                 mutex_lock(&req->r_fill_mutex);
2481                 req->r_err = err;
2482                 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2483                 mutex_unlock(&req->r_fill_mutex);
2484
2485                 if (req->r_parent &&
2486                     (req->r_op & CEPH_MDS_OP_WRITE))
2487                         ceph_invalidate_dir_request(req);
2488         } else {
2489                 err = req->r_err;
2490         }
2491
2492 out:
2493         mutex_unlock(&mdsc->mutex);
2494         dout("do_request %p done, result %d\n", req, err);
2495         return err;
2496 }
2497
2498 /*
2499  * Invalidate dir's completeness, dentry lease state on an aborted MDS
2500  * namespace request.
2501  */
2502 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2503 {
2504         struct inode *dir = req->r_parent;
2505         struct inode *old_dir = req->r_old_dentry_dir;
2506
2507         dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
2508
2509         ceph_dir_clear_complete(dir);
2510         if (old_dir)
2511                 ceph_dir_clear_complete(old_dir);
2512         if (req->r_dentry)
2513                 ceph_invalidate_dentry_lease(req->r_dentry);
2514         if (req->r_old_dentry)
2515                 ceph_invalidate_dentry_lease(req->r_old_dentry);
2516 }
2517
2518 /*
2519  * Handle mds reply.
2520  *
2521  * We take the session mutex and parse and process the reply immediately.
2522  * This preserves the logical ordering of replies, capabilities, etc., sent
2523  * by the MDS as they are applied to our local cache.
2524  */
2525 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2526 {
2527         struct ceph_mds_client *mdsc = session->s_mdsc;
2528         struct ceph_mds_request *req;
2529         struct ceph_mds_reply_head *head = msg->front.iov_base;
2530         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2531         struct ceph_snap_realm *realm;
2532         u64 tid;
2533         int err, result;
2534         int mds = session->s_mds;
2535
2536         if (msg->front.iov_len < sizeof(*head)) {
2537                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2538                 ceph_msg_dump(msg);
2539                 return;
2540         }
2541
2542         /* get request, session */
2543         tid = le64_to_cpu(msg->hdr.tid);
2544         mutex_lock(&mdsc->mutex);
2545         req = lookup_get_request(mdsc, tid);
2546         if (!req) {
2547                 dout("handle_reply on unknown tid %llu\n", tid);
2548                 mutex_unlock(&mdsc->mutex);
2549                 return;
2550         }
2551         dout("handle_reply %p\n", req);
2552
2553         /* correct session? */
2554         if (req->r_session != session) {
2555                 pr_err("mdsc_handle_reply got %llu on session mds%d"
2556                        " not mds%d\n", tid, session->s_mds,
2557                        req->r_session ? req->r_session->s_mds : -1);
2558                 mutex_unlock(&mdsc->mutex);
2559                 goto out;
2560         }
2561
2562         /* dup? */
2563         if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
2564             (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
2565                 pr_warn("got a dup %s reply on %llu from mds%d\n",
2566                            head->safe ? "safe" : "unsafe", tid, mds);
2567                 mutex_unlock(&mdsc->mutex);
2568                 goto out;
2569         }
2570         if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
2571                 pr_warn("got unsafe after safe on %llu from mds%d\n",
2572                            tid, mds);
2573                 mutex_unlock(&mdsc->mutex);
2574                 goto out;
2575         }
2576
2577         result = le32_to_cpu(head->result);
2578
2579         /*
2580          * Handle an ESTALE
2581          * if we're not talking to the authority, send to them
2582          * if the authority has changed while we weren't looking,
2583          * send to new authority
2584          * Otherwise we just have to return an ESTALE
2585          */
2586         if (result == -ESTALE) {
2587                 dout("got ESTALE on request %llu\n", req->r_tid);
2588                 req->r_resend_mds = -1;
2589                 if (req->r_direct_mode != USE_AUTH_MDS) {
2590                         dout("not using auth, setting for that now\n");
2591                         req->r_direct_mode = USE_AUTH_MDS;
2592                         __do_request(mdsc, req);
2593                         mutex_unlock(&mdsc->mutex);
2594                         goto out;
2595                 } else  {
2596                         int mds = __choose_mds(mdsc, req);
2597                         if (mds >= 0 && mds != req->r_session->s_mds) {
2598                                 dout("but auth changed, so resending\n");
2599                                 __do_request(mdsc, req);
2600                                 mutex_unlock(&mdsc->mutex);
2601                                 goto out;
2602                         }
2603                 }
2604                 dout("have to return ESTALE on request %llu\n", req->r_tid);
2605         }
2606
2607
2608         if (head->safe) {
2609                 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
2610                 __unregister_request(mdsc, req);
2611
2612                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2613                         /*
2614                          * We already handled the unsafe response, now do the
2615                          * cleanup.  No need to examine the response; the MDS
2616                          * doesn't include any result info in the safe
2617                          * response.  And even if it did, there is nothing
2618                          * useful we could do with a revised return value.
2619                          */
2620                         dout("got safe reply %llu, mds%d\n", tid, mds);
2621
2622                         /* last unsafe request during umount? */
2623                         if (mdsc->stopping && !__get_oldest_req(mdsc))
2624                                 complete_all(&mdsc->safe_umount_waiters);
2625                         mutex_unlock(&mdsc->mutex);
2626                         goto out;
2627                 }
2628         } else {
2629                 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
2630                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2631                 if (req->r_unsafe_dir) {
2632                         struct ceph_inode_info *ci =
2633                                         ceph_inode(req->r_unsafe_dir);
2634                         spin_lock(&ci->i_unsafe_lock);
2635                         list_add_tail(&req->r_unsafe_dir_item,
2636                                       &ci->i_unsafe_dirops);
2637                         spin_unlock(&ci->i_unsafe_lock);
2638                 }
2639         }
2640
2641         dout("handle_reply tid %lld result %d\n", tid, result);
2642         rinfo = &req->r_reply_info;
2643         err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2644         mutex_unlock(&mdsc->mutex);
2645
2646         mutex_lock(&session->s_mutex);
2647         if (err < 0) {
2648                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2649                 ceph_msg_dump(msg);
2650                 goto out_err;
2651         }
2652
2653         /* snap trace */
2654         realm = NULL;
2655         if (rinfo->snapblob_len) {
2656                 down_write(&mdsc->snap_rwsem);
2657                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2658                                 rinfo->snapblob + rinfo->snapblob_len,
2659                                 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2660                                 &realm);
2661                 downgrade_write(&mdsc->snap_rwsem);
2662         } else {
2663                 down_read(&mdsc->snap_rwsem);
2664         }
2665
2666         /* insert trace into our cache */
2667         mutex_lock(&req->r_fill_mutex);
2668         current->journal_info = req;
2669         err = ceph_fill_trace(mdsc->fsc->sb, req);
2670         if (err == 0) {
2671                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2672                                     req->r_op == CEPH_MDS_OP_LSSNAP))
2673                         ceph_readdir_prepopulate(req, req->r_session);
2674                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2675         }
2676         current->journal_info = NULL;
2677         mutex_unlock(&req->r_fill_mutex);
2678
2679         up_read(&mdsc->snap_rwsem);
2680         if (realm)
2681                 ceph_put_snap_realm(mdsc, realm);
2682
2683         if (err == 0 && req->r_target_inode &&
2684             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2685                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
2686                 spin_lock(&ci->i_unsafe_lock);
2687                 list_add_tail(&req->r_unsafe_target_item, &ci->i_unsafe_iops);
2688                 spin_unlock(&ci->i_unsafe_lock);
2689         }
2690 out_err:
2691         mutex_lock(&mdsc->mutex);
2692         if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2693                 if (err) {
2694                         req->r_err = err;
2695                 } else {
2696                         req->r_reply =  ceph_msg_get(msg);
2697                         set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
2698                 }
2699         } else {
2700                 dout("reply arrived after request %lld was aborted\n", tid);
2701         }
2702         mutex_unlock(&mdsc->mutex);
2703
2704         mutex_unlock(&session->s_mutex);
2705
2706         /* kick calling process */
2707         complete_request(mdsc, req);
2708 out:
2709         ceph_mdsc_put_request(req);
2710         return;
2711 }
2712
2713
2714
2715 /*
2716  * handle mds notification that our request has been forwarded.
2717  */
2718 static void handle_forward(struct ceph_mds_client *mdsc,
2719                            struct ceph_mds_session *session,
2720                            struct ceph_msg *msg)
2721 {
2722         struct ceph_mds_request *req;
2723         u64 tid = le64_to_cpu(msg->hdr.tid);
2724         u32 next_mds;
2725         u32 fwd_seq;
2726         int err = -EINVAL;
2727         void *p = msg->front.iov_base;
2728         void *end = p + msg->front.iov_len;
2729
2730         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2731         next_mds = ceph_decode_32(&p);
2732         fwd_seq = ceph_decode_32(&p);
2733
2734         mutex_lock(&mdsc->mutex);
2735         req = lookup_get_request(mdsc, tid);
2736         if (!req) {
2737                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2738                 goto out;  /* dup reply? */
2739         }
2740
2741         if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2742                 dout("forward tid %llu aborted, unregistering\n", tid);
2743                 __unregister_request(mdsc, req);
2744         } else if (fwd_seq <= req->r_num_fwd) {
2745                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2746                      tid, next_mds, req->r_num_fwd, fwd_seq);
2747         } else {
2748                 /* resend. forward race not possible; mds would drop */
2749                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2750                 BUG_ON(req->r_err);
2751                 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
2752                 req->r_attempts = 0;
2753                 req->r_num_fwd = fwd_seq;
2754                 req->r_resend_mds = next_mds;
2755                 put_request_session(req);
2756                 __do_request(mdsc, req);
2757         }
2758         ceph_mdsc_put_request(req);
2759 out:
2760         mutex_unlock(&mdsc->mutex);
2761         return;
2762
2763 bad:
2764         pr_err("mdsc_handle_forward decode error err=%d\n", err);
2765 }
2766
2767 /*
2768  * handle a mds session control message
2769  */
2770 static void handle_session(struct ceph_mds_session *session,
2771                            struct ceph_msg *msg)
2772 {
2773         struct ceph_mds_client *mdsc = session->s_mdsc;
2774         u32 op;
2775         u64 seq;
2776         int mds = session->s_mds;
2777         struct ceph_mds_session_head *h = msg->front.iov_base;
2778         int wake = 0;
2779
2780         /* decode */
2781         if (msg->front.iov_len < sizeof(*h))
2782                 goto bad;
2783         op = le32_to_cpu(h->op);
2784         seq = le64_to_cpu(h->seq);
2785
2786         mutex_lock(&mdsc->mutex);
2787         if (op == CEPH_SESSION_CLOSE) {
2788                 get_session(session);
2789                 __unregister_session(mdsc, session);
2790         }
2791         /* FIXME: this ttl calculation is generous */
2792         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2793         mutex_unlock(&mdsc->mutex);
2794
2795         mutex_lock(&session->s_mutex);
2796
2797         dout("handle_session mds%d %s %p state %s seq %llu\n",
2798              mds, ceph_session_op_name(op), session,
2799              ceph_session_state_name(session->s_state), seq);
2800
2801         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2802                 session->s_state = CEPH_MDS_SESSION_OPEN;
2803                 pr_info("mds%d came back\n", session->s_mds);
2804         }
2805
2806         switch (op) {
2807         case CEPH_SESSION_OPEN:
2808                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2809                         pr_info("mds%d reconnect success\n", session->s_mds);
2810                 session->s_state = CEPH_MDS_SESSION_OPEN;
2811                 renewed_caps(mdsc, session, 0);
2812                 wake = 1;
2813                 if (mdsc->stopping)
2814                         __close_session(mdsc, session);
2815                 break;
2816
2817         case CEPH_SESSION_RENEWCAPS:
2818                 if (session->s_renew_seq == seq)
2819                         renewed_caps(mdsc, session, 1);
2820                 break;
2821
2822         case CEPH_SESSION_CLOSE:
2823                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2824                         pr_info("mds%d reconnect denied\n", session->s_mds);
2825                 cleanup_session_requests(mdsc, session);
2826                 remove_session_caps(session);
2827                 wake = 2; /* for good measure */
2828                 wake_up_all(&mdsc->session_close_wq);
2829                 break;
2830
2831         case CEPH_SESSION_STALE:
2832                 pr_info("mds%d caps went stale, renewing\n",
2833                         session->s_mds);
2834                 spin_lock(&session->s_gen_ttl_lock);
2835                 session->s_cap_gen++;
2836                 session->s_cap_ttl = jiffies - 1;
2837                 spin_unlock(&session->s_gen_ttl_lock);
2838                 send_renew_caps(mdsc, session);
2839                 break;
2840
2841         case CEPH_SESSION_RECALL_STATE:
2842                 ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2843                 break;
2844
2845         case CEPH_SESSION_FLUSHMSG:
2846                 send_flushmsg_ack(mdsc, session, seq);
2847                 break;
2848
2849         case CEPH_SESSION_FORCE_RO:
2850                 dout("force_session_readonly %p\n", session);
2851                 spin_lock(&session->s_cap_lock);
2852                 session->s_readonly = true;
2853                 spin_unlock(&session->s_cap_lock);
2854                 wake_up_session_caps(session, 0);
2855                 break;
2856
2857         case CEPH_SESSION_REJECT:
2858                 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
2859                 pr_info("mds%d rejected session\n", session->s_mds);
2860                 session->s_state = CEPH_MDS_SESSION_REJECTED;
2861                 cleanup_session_requests(mdsc, session);
2862                 remove_session_caps(session);
2863                 wake = 2; /* for good measure */
2864                 break;
2865
2866         default:
2867                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2868                 WARN_ON(1);
2869         }
2870
2871         mutex_unlock(&session->s_mutex);
2872         if (wake) {
2873                 mutex_lock(&mdsc->mutex);
2874                 __wake_requests(mdsc, &session->s_waiting);
2875                 if (wake == 2)
2876                         kick_requests(mdsc, mds);
2877                 mutex_unlock(&mdsc->mutex);
2878         }
2879         if (op == CEPH_SESSION_CLOSE)
2880                 ceph_put_mds_session(session);
2881         return;
2882
2883 bad:
2884         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2885                (int)msg->front.iov_len);
2886         ceph_msg_dump(msg);
2887         return;
2888 }
2889
2890
2891 /*
2892  * called under session->mutex.
2893  */
2894 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2895                                    struct ceph_mds_session *session)
2896 {
2897         struct ceph_mds_request *req, *nreq;
2898         struct rb_node *p;
2899         int err;
2900
2901         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2902
2903         mutex_lock(&mdsc->mutex);
2904         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2905                 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2906                 if (!err) {
2907                         ceph_msg_get(req->r_request);
2908                         ceph_con_send(&session->s_con, req->r_request);
2909                 }
2910         }
2911
2912         /*
2913          * also re-send old requests when MDS enters reconnect stage. So that MDS
2914          * can process completed request in clientreplay stage.
2915          */
2916         p = rb_first(&mdsc->request_tree);
2917         while (p) {
2918                 req = rb_entry(p, struct ceph_mds_request, r_node);
2919                 p = rb_next(p);
2920                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2921                         continue;
2922                 if (req->r_attempts == 0)
2923                         continue; /* only old requests */
2924                 if (req->r_session &&
2925                     req->r_session->s_mds == session->s_mds) {
2926                         err = __prepare_send_request(mdsc, req,
2927                                                      session->s_mds, true);
2928                         if (!err) {
2929                                 ceph_msg_get(req->r_request);
2930                                 ceph_con_send(&session->s_con, req->r_request);
2931                         }
2932                 }
2933         }
2934         mutex_unlock(&mdsc->mutex);
2935 }
2936
2937 /*
2938  * Encode information about a cap for a reconnect with the MDS.
2939  */
2940 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2941                           void *arg)
2942 {
2943         union {
2944                 struct ceph_mds_cap_reconnect v2;
2945                 struct ceph_mds_cap_reconnect_v1 v1;
2946         } rec;
2947         struct ceph_inode_info *ci = cap->ci;
2948         struct ceph_reconnect_state *recon_state = arg;
2949         struct ceph_pagelist *pagelist = recon_state->pagelist;
2950         char *path;
2951         int pathlen, err;
2952         u64 pathbase;
2953         u64 snap_follows;
2954         struct dentry *dentry;
2955
2956         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2957              inode, ceph_vinop(inode), cap, cap->cap_id,
2958              ceph_cap_string(cap->issued));
2959         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2960         if (err)
2961                 return err;
2962
2963         dentry = d_find_alias(inode);
2964         if (dentry) {
2965                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2966                 if (IS_ERR(path)) {
2967                         err = PTR_ERR(path);
2968                         goto out_dput;
2969                 }
2970         } else {
2971                 path = NULL;
2972                 pathlen = 0;
2973                 pathbase = 0;
2974         }
2975
2976         spin_lock(&ci->i_ceph_lock);
2977         cap->seq = 0;        /* reset cap seq */
2978         cap->issue_seq = 0;  /* and issue_seq */
2979         cap->mseq = 0;       /* and migrate_seq */
2980         cap->cap_gen = cap->session->s_cap_gen;
2981
2982         if (recon_state->msg_version >= 2) {
2983                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2984                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2985                 rec.v2.issued = cpu_to_le32(cap->issued);
2986                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2987                 rec.v2.pathbase = cpu_to_le64(pathbase);
2988                 rec.v2.flock_len = (__force __le32)
2989                         ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
2990         } else {
2991                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2992                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2993                 rec.v1.issued = cpu_to_le32(cap->issued);
2994                 rec.v1.size = cpu_to_le64(inode->i_size);
2995                 ceph_encode_timespec64(&rec.v1.mtime, &inode->i_mtime);
2996                 ceph_encode_timespec64(&rec.v1.atime, &inode->i_atime);
2997                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2998                 rec.v1.pathbase = cpu_to_le64(pathbase);
2999         }
3000
3001         if (list_empty(&ci->i_cap_snaps)) {
3002                 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
3003         } else {
3004                 struct ceph_cap_snap *capsnap =
3005                         list_first_entry(&ci->i_cap_snaps,
3006                                          struct ceph_cap_snap, ci_item);
3007                 snap_follows = capsnap->follows;
3008         }
3009         spin_unlock(&ci->i_ceph_lock);
3010
3011         if (recon_state->msg_version >= 2) {
3012                 int num_fcntl_locks, num_flock_locks;
3013                 struct ceph_filelock *flocks = NULL;
3014                 size_t struct_len, total_len = 0;
3015                 u8 struct_v = 0;
3016
3017 encode_again:
3018                 if (rec.v2.flock_len) {
3019                         ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
3020                 } else {
3021                         num_fcntl_locks = 0;
3022                         num_flock_locks = 0;
3023                 }
3024                 if (num_fcntl_locks + num_flock_locks > 0) {
3025                         flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
3026                                                sizeof(struct ceph_filelock),
3027                                                GFP_NOFS);
3028                         if (!flocks) {
3029                                 err = -ENOMEM;
3030                                 goto out_free;
3031                         }
3032                         err = ceph_encode_locks_to_buffer(inode, flocks,
3033                                                           num_fcntl_locks,
3034                                                           num_flock_locks);
3035                         if (err) {
3036                                 kfree(flocks);
3037                                 flocks = NULL;
3038                                 if (err == -ENOSPC)
3039                                         goto encode_again;
3040                                 goto out_free;
3041                         }
3042                 } else {
3043                         kfree(flocks);
3044                         flocks = NULL;
3045                 }
3046
3047                 if (recon_state->msg_version >= 3) {
3048                         /* version, compat_version and struct_len */
3049                         total_len = 2 * sizeof(u8) + sizeof(u32);
3050                         struct_v = 2;
3051                 }
3052                 /*
3053                  * number of encoded locks is stable, so copy to pagelist
3054                  */
3055                 struct_len = 2 * sizeof(u32) +
3056                             (num_fcntl_locks + num_flock_locks) *
3057                             sizeof(struct ceph_filelock);
3058                 rec.v2.flock_len = cpu_to_le32(struct_len);
3059
3060                 struct_len += sizeof(rec.v2);
3061                 struct_len += sizeof(u32) + pathlen;
3062
3063                 if (struct_v >= 2)
3064                         struct_len += sizeof(u64); /* snap_follows */
3065
3066                 total_len += struct_len;
3067                 err = ceph_pagelist_reserve(pagelist, total_len);
3068
3069                 if (!err) {
3070                         if (recon_state->msg_version >= 3) {
3071                                 ceph_pagelist_encode_8(pagelist, struct_v);
3072                                 ceph_pagelist_encode_8(pagelist, 1);
3073                                 ceph_pagelist_encode_32(pagelist, struct_len);
3074                         }
3075                         ceph_pagelist_encode_string(pagelist, path, pathlen);
3076                         ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
3077                         ceph_locks_to_pagelist(flocks, pagelist,
3078                                                num_fcntl_locks,
3079                                                num_flock_locks);
3080                         if (struct_v >= 2)
3081                                 ceph_pagelist_encode_64(pagelist, snap_follows);
3082                 }
3083                 kfree(flocks);
3084         } else {
3085                 size_t size = sizeof(u32) + pathlen + sizeof(rec.v1);
3086                 err = ceph_pagelist_reserve(pagelist, size);
3087                 if (!err) {
3088                         ceph_pagelist_encode_string(pagelist, path, pathlen);
3089                         ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
3090                 }
3091         }
3092
3093         recon_state->nr_caps++;
3094 out_free:
3095         kfree(path);
3096 out_dput:
3097         dput(dentry);
3098         return err;
3099 }
3100
3101
3102 /*
3103  * If an MDS fails and recovers, clients need to reconnect in order to
3104  * reestablish shared state.  This includes all caps issued through
3105  * this session _and_ the snap_realm hierarchy.  Because it's not
3106  * clear which snap realms the mds cares about, we send everything we
3107  * know about.. that ensures we'll then get any new info the
3108  * recovering MDS might have.
3109  *
3110  * This is a relatively heavyweight operation, but it's rare.
3111  *
3112  * called with mdsc->mutex held.
3113  */
3114 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3115                                struct ceph_mds_session *session)
3116 {
3117         struct ceph_msg *reply;
3118         struct rb_node *p;
3119         int mds = session->s_mds;
3120         int err = -ENOMEM;
3121         int s_nr_caps;
3122         struct ceph_pagelist *pagelist;
3123         struct ceph_reconnect_state recon_state;
3124         LIST_HEAD(dispose);
3125
3126         pr_info("mds%d reconnect start\n", mds);
3127
3128         pagelist = ceph_pagelist_alloc(GFP_NOFS);
3129         if (!pagelist)
3130                 goto fail_nopagelist;
3131
3132         reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3133         if (!reply)
3134                 goto fail_nomsg;
3135
3136         mutex_lock(&session->s_mutex);
3137         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3138         session->s_seq = 0;
3139
3140         dout("session %p state %s\n", session,
3141              ceph_session_state_name(session->s_state));
3142
3143         spin_lock(&session->s_gen_ttl_lock);
3144         session->s_cap_gen++;
3145         spin_unlock(&session->s_gen_ttl_lock);
3146
3147         spin_lock(&session->s_cap_lock);
3148         /* don't know if session is readonly */
3149         session->s_readonly = 0;
3150         /*
3151          * notify __ceph_remove_cap() that we are composing cap reconnect.
3152          * If a cap get released before being added to the cap reconnect,
3153          * __ceph_remove_cap() should skip queuing cap release.
3154          */
3155         session->s_cap_reconnect = 1;
3156         /* drop old cap expires; we're about to reestablish that state */
3157         detach_cap_releases(session, &dispose);
3158         spin_unlock(&session->s_cap_lock);
3159         dispose_cap_releases(mdsc, &dispose);
3160
3161         /* trim unused caps to reduce MDS's cache rejoin time */
3162         if (mdsc->fsc->sb->s_root)
3163                 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3164
3165         ceph_con_close(&session->s_con);
3166         ceph_con_open(&session->s_con,
3167                       CEPH_ENTITY_TYPE_MDS, mds,
3168                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3169
3170         /* replay unsafe requests */
3171         replay_unsafe_requests(mdsc, session);
3172
3173         down_read(&mdsc->snap_rwsem);
3174
3175         /* traverse this session's caps */
3176         s_nr_caps = session->s_nr_caps;
3177         err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
3178         if (err)
3179                 goto fail;
3180
3181         recon_state.nr_caps = 0;
3182         recon_state.pagelist = pagelist;
3183         if (session->s_con.peer_features & CEPH_FEATURE_MDSENC)
3184                 recon_state.msg_version = 3;
3185         else if (session->s_con.peer_features & CEPH_FEATURE_FLOCK)
3186                 recon_state.msg_version = 2;
3187         else
3188                 recon_state.msg_version = 1;
3189         err = iterate_session_caps(session, encode_caps_cb, &recon_state);
3190         if (err < 0)
3191                 goto fail;
3192
3193         spin_lock(&session->s_cap_lock);
3194         session->s_cap_reconnect = 0;
3195         spin_unlock(&session->s_cap_lock);
3196
3197         /*
3198          * snaprealms.  we provide mds with the ino, seq (version), and
3199          * parent for all of our realms.  If the mds has any newer info,
3200          * it will tell us.
3201          */
3202         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3203                 struct ceph_snap_realm *realm =
3204                         rb_entry(p, struct ceph_snap_realm, node);
3205                 struct ceph_mds_snaprealm_reconnect sr_rec;
3206
3207                 dout(" adding snap realm %llx seq %lld parent %llx\n",
3208                      realm->ino, realm->seq, realm->parent_ino);
3209                 sr_rec.ino = cpu_to_le64(realm->ino);
3210                 sr_rec.seq = cpu_to_le64(realm->seq);
3211                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3212                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3213                 if (err)
3214                         goto fail;
3215         }
3216
3217         reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3218
3219         /* raced with cap release? */
3220         if (s_nr_caps != recon_state.nr_caps) {
3221                 struct page *page = list_first_entry(&pagelist->head,
3222                                                      struct page, lru);
3223                 __le32 *addr = kmap_atomic(page);
3224                 *addr = cpu_to_le32(recon_state.nr_caps);
3225                 kunmap_atomic(addr);
3226         }
3227
3228         reply->hdr.data_len = cpu_to_le32(pagelist->length);
3229         ceph_msg_data_add_pagelist(reply, pagelist);
3230
3231         ceph_early_kick_flushing_caps(mdsc, session);
3232
3233         ceph_con_send(&session->s_con, reply);
3234
3235         mutex_unlock(&session->s_mutex);
3236
3237         mutex_lock(&mdsc->mutex);
3238         __wake_requests(mdsc, &session->s_waiting);
3239         mutex_unlock(&mdsc->mutex);
3240
3241         up_read(&mdsc->snap_rwsem);
3242         ceph_pagelist_release(pagelist);
3243         return;
3244
3245 fail:
3246         ceph_msg_put(reply);
3247         up_read(&mdsc->snap_rwsem);
3248         mutex_unlock(&session->s_mutex);
3249 fail_nomsg:
3250         ceph_pagelist_release(pagelist);
3251 fail_nopagelist:
3252         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3253         return;
3254 }
3255
3256
3257 /*
3258  * compare old and new mdsmaps, kicking requests
3259  * and closing out old connections as necessary
3260  *
3261  * called under mdsc->mutex.
3262  */
3263 static void check_new_map(struct ceph_mds_client *mdsc,
3264                           struct ceph_mdsmap *newmap,
3265                           struct ceph_mdsmap *oldmap)
3266 {
3267         int i;
3268         int oldstate, newstate;
3269         struct ceph_mds_session *s;
3270
3271         dout("check_new_map new %u old %u\n",
3272              newmap->m_epoch, oldmap->m_epoch);
3273
3274         for (i = 0; i < oldmap->m_num_mds && i < mdsc->max_sessions; i++) {
3275                 if (!mdsc->sessions[i])
3276                         continue;
3277                 s = mdsc->sessions[i];
3278                 oldstate = ceph_mdsmap_get_state(oldmap, i);
3279                 newstate = ceph_mdsmap_get_state(newmap, i);
3280
3281                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3282                      i, ceph_mds_state_name(oldstate),
3283                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3284                      ceph_mds_state_name(newstate),
3285                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3286                      ceph_session_state_name(s->s_state));
3287
3288                 if (i >= newmap->m_num_mds ||
3289                     memcmp(ceph_mdsmap_get_addr(oldmap, i),
3290                            ceph_mdsmap_get_addr(newmap, i),
3291                            sizeof(struct ceph_entity_addr))) {
3292                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3293                                 /* the session never opened, just close it
3294                                  * out now */
3295                                 get_session(s);
3296                                 __unregister_session(mdsc, s);
3297                                 __wake_requests(mdsc, &s->s_waiting);
3298                                 ceph_put_mds_session(s);
3299                         } else if (i >= newmap->m_num_mds) {
3300                                 /* force close session for stopped mds */
3301                                 get_session(s);
3302                                 __unregister_session(mdsc, s);
3303                                 __wake_requests(mdsc, &s->s_waiting);
3304                                 kick_requests(mdsc, i);
3305                                 mutex_unlock(&mdsc->mutex);
3306
3307                                 mutex_lock(&s->s_mutex);
3308                                 cleanup_session_requests(mdsc, s);
3309                                 remove_session_caps(s);
3310                                 mutex_unlock(&s->s_mutex);
3311
3312                                 ceph_put_mds_session(s);
3313
3314                                 mutex_lock(&mdsc->mutex);
3315                         } else {
3316                                 /* just close it */
3317                                 mutex_unlock(&mdsc->mutex);
3318                                 mutex_lock(&s->s_mutex);
3319                                 mutex_lock(&mdsc->mutex);
3320                                 ceph_con_close(&s->s_con);
3321                                 mutex_unlock(&s->s_mutex);
3322                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3323                         }
3324                 } else if (oldstate == newstate) {
3325                         continue;  /* nothing new with this mds */
3326                 }
3327
3328                 /*
3329                  * send reconnect?
3330                  */
3331                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3332                     newstate >= CEPH_MDS_STATE_RECONNECT) {
3333                         mutex_unlock(&mdsc->mutex);
3334                         send_mds_reconnect(mdsc, s);
3335                         mutex_lock(&mdsc->mutex);
3336                 }
3337
3338                 /*
3339                  * kick request on any mds that has gone active.
3340                  */
3341                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3342                     newstate >= CEPH_MDS_STATE_ACTIVE) {
3343                         if (oldstate != CEPH_MDS_STATE_CREATING &&
3344                             oldstate != CEPH_MDS_STATE_STARTING)
3345                                 pr_info("mds%d recovery completed\n", s->s_mds);
3346                         kick_requests(mdsc, i);
3347                         ceph_kick_flushing_caps(mdsc, s);
3348                         wake_up_session_caps(s, 1);
3349                 }
3350         }
3351
3352         for (i = 0; i < newmap->m_num_mds && i < mdsc->max_sessions; i++) {
3353                 s = mdsc->sessions[i];
3354                 if (!s)
3355                         continue;
3356                 if (!ceph_mdsmap_is_laggy(newmap, i))
3357                         continue;
3358                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3359                     s->s_state == CEPH_MDS_SESSION_HUNG ||
3360                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
3361                         dout(" connecting to export targets of laggy mds%d\n",
3362                              i);
3363                         __open_export_target_sessions(mdsc, s);
3364                 }
3365         }
3366 }
3367
3368
3369
3370 /*
3371  * leases
3372  */
3373
3374 /*
3375  * caller must hold session s_mutex, dentry->d_lock
3376  */
3377 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3378 {
3379         struct ceph_dentry_info *di = ceph_dentry(dentry);
3380
3381         ceph_put_mds_session(di->lease_session);
3382         di->lease_session = NULL;
3383 }
3384
3385 static void handle_lease(struct ceph_mds_client *mdsc,
3386                          struct ceph_mds_session *session,
3387                          struct ceph_msg *msg)
3388 {
3389         struct super_block *sb = mdsc->fsc->sb;
3390         struct inode *inode;
3391         struct dentry *parent, *dentry;
3392         struct ceph_dentry_info *di;
3393         int mds = session->s_mds;
3394         struct ceph_mds_lease *h = msg->front.iov_base;
3395         u32 seq;
3396         struct ceph_vino vino;
3397         struct qstr dname;
3398         int release = 0;
3399
3400         dout("handle_lease from mds%d\n", mds);
3401
3402         /* decode */
3403         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3404                 goto bad;
3405         vino.ino = le64_to_cpu(h->ino);
3406         vino.snap = CEPH_NOSNAP;
3407         seq = le32_to_cpu(h->seq);
3408         dname.len = get_unaligned_le32(h + 1);
3409         if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
3410                 goto bad;
3411         dname.name = (void *)(h + 1) + sizeof(u32);
3412
3413         /* lookup inode */
3414         inode = ceph_find_inode(sb, vino);
3415         dout("handle_lease %s, ino %llx %p %.*s\n",
3416              ceph_lease_op_name(h->action), vino.ino, inode,
3417              dname.len, dname.name);
3418
3419         mutex_lock(&session->s_mutex);
3420         session->s_seq++;
3421
3422         if (!inode) {
3423                 dout("handle_lease no inode %llx\n", vino.ino);
3424                 goto release;
3425         }
3426
3427         /* dentry */
3428         parent = d_find_alias(inode);
3429         if (!parent) {
3430                 dout("no parent dentry on inode %p\n", inode);
3431                 WARN_ON(1);
3432                 goto release;  /* hrm... */
3433         }
3434         dname.hash = full_name_hash(parent, dname.name, dname.len);
3435         dentry = d_lookup(parent, &dname);
3436         dput(parent);
3437         if (!dentry)
3438                 goto release;
3439
3440         spin_lock(&dentry->d_lock);
3441         di = ceph_dentry(dentry);
3442         switch (h->action) {
3443         case CEPH_MDS_LEASE_REVOKE:
3444                 if (di->lease_session == session) {
3445                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3446                                 h->seq = cpu_to_le32(di->lease_seq);
3447                         __ceph_mdsc_drop_dentry_lease(dentry);
3448                 }
3449                 release = 1;
3450                 break;
3451
3452         case CEPH_MDS_LEASE_RENEW:
3453                 if (di->lease_session == session &&
3454                     di->lease_gen == session->s_cap_gen &&
3455                     di->lease_renew_from &&
3456                     di->lease_renew_after == 0) {
3457                         unsigned long duration =
3458                                 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3459
3460                         di->lease_seq = seq;
3461                         di->time = di->lease_renew_from + duration;
3462                         di->lease_renew_after = di->lease_renew_from +
3463                                 (duration >> 1);
3464                         di->lease_renew_from = 0;
3465                 }
3466                 break;
3467         }
3468         spin_unlock(&dentry->d_lock);
3469         dput(dentry);
3470
3471         if (!release)
3472                 goto out;
3473
3474 release:
3475         /* let's just reuse the same message */
3476         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3477         ceph_msg_get(msg);
3478         ceph_con_send(&session->s_con, msg);
3479
3480 out:
3481         iput(inode);
3482         mutex_unlock(&session->s_mutex);
3483         return;
3484
3485 bad:
3486         pr_err("corrupt lease message\n");
3487         ceph_msg_dump(msg);
3488 }
3489
3490 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3491                               struct inode *inode,
3492                               struct dentry *dentry, char action,
3493                               u32 seq)
3494 {
3495         struct ceph_msg *msg;
3496         struct ceph_mds_lease *lease;
3497         int len = sizeof(*lease) + sizeof(u32);
3498         int dnamelen = 0;
3499
3500         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3501              inode, dentry, ceph_lease_op_name(action), session->s_mds);
3502         dnamelen = dentry->d_name.len;
3503         len += dnamelen;
3504
3505         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3506         if (!msg)
3507                 return;
3508         lease = msg->front.iov_base;
3509         lease->action = action;
3510         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3511         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3512         lease->seq = cpu_to_le32(seq);
3513         put_unaligned_le32(dnamelen, lease + 1);
3514         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3515
3516         /*
3517          * if this is a preemptive lease RELEASE, no need to
3518          * flush request stream, since the actual request will
3519          * soon follow.
3520          */
3521         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3522
3523         ceph_con_send(&session->s_con, msg);
3524 }
3525
3526 /*
3527  * lock unlock sessions, to wait ongoing session activities
3528  */
3529 static void lock_unlock_sessions(struct ceph_mds_client *mdsc)
3530 {
3531         int i;
3532
3533         mutex_lock(&mdsc->mutex);
3534         for (i = 0; i < mdsc->max_sessions; i++) {
3535                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3536                 if (!s)
3537                         continue;
3538                 mutex_unlock(&mdsc->mutex);
3539                 mutex_lock(&s->s_mutex);
3540                 mutex_unlock(&s->s_mutex);
3541                 ceph_put_mds_session(s);
3542                 mutex_lock(&mdsc->mutex);
3543         }
3544         mutex_unlock(&mdsc->mutex);
3545 }
3546
3547
3548
3549 /*
3550  * delayed work -- periodically trim expired leases, renew caps with mds
3551  */
3552 static void schedule_delayed(struct ceph_mds_client *mdsc)
3553 {
3554         int delay = 5;
3555         unsigned hz = round_jiffies_relative(HZ * delay);
3556         schedule_delayed_work(&mdsc->delayed_work, hz);
3557 }
3558
3559 static void delayed_work(struct work_struct *work)
3560 {
3561         int i;
3562         struct ceph_mds_client *mdsc =
3563                 container_of(work, struct ceph_mds_client, delayed_work.work);
3564         int renew_interval;
3565         int renew_caps;
3566
3567         dout("mdsc delayed_work\n");
3568         ceph_check_delayed_caps(mdsc);
3569
3570         mutex_lock(&mdsc->mutex);
3571         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3572         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3573                                    mdsc->last_renew_caps);
3574         if (renew_caps)
3575                 mdsc->last_renew_caps = jiffies;
3576
3577         for (i = 0; i < mdsc->max_sessions; i++) {
3578                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3579                 if (!s)
3580                         continue;
3581                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3582                         dout("resending session close request for mds%d\n",
3583                              s->s_mds);
3584                         request_close_session(mdsc, s);
3585                         ceph_put_mds_session(s);
3586                         continue;
3587                 }
3588                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3589                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3590                                 s->s_state = CEPH_MDS_SESSION_HUNG;
3591                                 pr_info("mds%d hung\n", s->s_mds);
3592                         }
3593                 }
3594                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3595                         /* this mds is failed or recovering, just wait */
3596                         ceph_put_mds_session(s);
3597                         continue;
3598                 }
3599                 mutex_unlock(&mdsc->mutex);
3600
3601                 mutex_lock(&s->s_mutex);
3602                 if (renew_caps)
3603                         send_renew_caps(mdsc, s);
3604                 else
3605                         ceph_con_keepalive(&s->s_con);
3606                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3607                     s->s_state == CEPH_MDS_SESSION_HUNG)
3608                         ceph_send_cap_releases(mdsc, s);
3609                 mutex_unlock(&s->s_mutex);
3610                 ceph_put_mds_session(s);
3611
3612                 mutex_lock(&mdsc->mutex);
3613         }
3614         mutex_unlock(&mdsc->mutex);
3615
3616         schedule_delayed(mdsc);
3617 }
3618
3619 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3620
3621 {
3622         struct ceph_mds_client *mdsc;
3623
3624         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3625         if (!mdsc)
3626                 return -ENOMEM;
3627         mdsc->fsc = fsc;
3628         mutex_init(&mdsc->mutex);
3629         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3630         if (!mdsc->mdsmap) {
3631                 kfree(mdsc);
3632                 return -ENOMEM;
3633         }
3634
3635         fsc->mdsc = mdsc;
3636         init_completion(&mdsc->safe_umount_waiters);
3637         init_waitqueue_head(&mdsc->session_close_wq);
3638         INIT_LIST_HEAD(&mdsc->waiting_for_map);
3639         mdsc->sessions = NULL;
3640         atomic_set(&mdsc->num_sessions, 0);
3641         mdsc->max_sessions = 0;
3642         mdsc->stopping = 0;
3643         atomic64_set(&mdsc->quotarealms_count, 0);
3644         mdsc->last_snap_seq = 0;
3645         init_rwsem(&mdsc->snap_rwsem);
3646         mdsc->snap_realms = RB_ROOT;
3647         INIT_LIST_HEAD(&mdsc->snap_empty);
3648         spin_lock_init(&mdsc->snap_empty_lock);
3649         mdsc->last_tid = 0;
3650         mdsc->oldest_tid = 0;
3651         mdsc->request_tree = RB_ROOT;
3652         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3653         mdsc->last_renew_caps = jiffies;
3654         INIT_LIST_HEAD(&mdsc->cap_delay_list);
3655         spin_lock_init(&mdsc->cap_delay_lock);
3656         INIT_LIST_HEAD(&mdsc->snap_flush_list);
3657         spin_lock_init(&mdsc->snap_flush_lock);
3658         mdsc->last_cap_flush_tid = 1;
3659         INIT_LIST_HEAD(&mdsc->cap_flush_list);
3660         INIT_LIST_HEAD(&mdsc->cap_dirty);
3661         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3662         mdsc->num_cap_flushing = 0;
3663         spin_lock_init(&mdsc->cap_dirty_lock);
3664         init_waitqueue_head(&mdsc->cap_flushing_wq);
3665         spin_lock_init(&mdsc->dentry_lru_lock);
3666         INIT_LIST_HEAD(&mdsc->dentry_lru);
3667
3668         ceph_caps_init(mdsc);
3669         ceph_adjust_min_caps(mdsc, fsc->min_caps);
3670
3671         init_rwsem(&mdsc->pool_perm_rwsem);
3672         mdsc->pool_perm_tree = RB_ROOT;
3673
3674         strscpy(mdsc->nodename, utsname()->nodename,
3675                 sizeof(mdsc->nodename));
3676         return 0;
3677 }
3678
3679 /*
3680  * Wait for safe replies on open mds requests.  If we time out, drop
3681  * all requests from the tree to avoid dangling dentry refs.
3682  */
3683 static void wait_requests(struct ceph_mds_client *mdsc)
3684 {
3685         struct ceph_options *opts = mdsc->fsc->client->options;
3686         struct ceph_mds_request *req;
3687
3688         mutex_lock(&mdsc->mutex);
3689         if (__get_oldest_req(mdsc)) {
3690                 mutex_unlock(&mdsc->mutex);
3691
3692                 dout("wait_requests waiting for requests\n");
3693                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3694                                     ceph_timeout_jiffies(opts->mount_timeout));
3695
3696                 /* tear down remaining requests */
3697                 mutex_lock(&mdsc->mutex);
3698                 while ((req = __get_oldest_req(mdsc))) {
3699                         dout("wait_requests timed out on tid %llu\n",
3700                              req->r_tid);
3701                         __unregister_request(mdsc, req);
3702                 }
3703         }
3704         mutex_unlock(&mdsc->mutex);
3705         dout("wait_requests done\n");
3706 }
3707
3708 /*
3709  * called before mount is ro, and before dentries are torn down.
3710  * (hmm, does this still race with new lookups?)
3711  */
3712 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3713 {
3714         dout("pre_umount\n");
3715         mdsc->stopping = 1;
3716
3717         lock_unlock_sessions(mdsc);
3718         ceph_flush_dirty_caps(mdsc);
3719         wait_requests(mdsc);
3720
3721         /*
3722          * wait for reply handlers to drop their request refs and
3723          * their inode/dcache refs
3724          */
3725         ceph_msgr_flush();
3726 }
3727
3728 /*
3729  * wait for all write mds requests to flush.
3730  */
3731 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3732 {
3733         struct ceph_mds_request *req = NULL, *nextreq;
3734         struct rb_node *n;
3735
3736         mutex_lock(&mdsc->mutex);
3737         dout("wait_unsafe_requests want %lld\n", want_tid);
3738 restart:
3739         req = __get_oldest_req(mdsc);
3740         while (req && req->r_tid <= want_tid) {
3741                 /* find next request */
3742                 n = rb_next(&req->r_node);
3743                 if (n)
3744                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3745                 else
3746                         nextreq = NULL;
3747                 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3748                     (req->r_op & CEPH_MDS_OP_WRITE)) {
3749                         /* write op */
3750                         ceph_mdsc_get_request(req);
3751                         if (nextreq)
3752                                 ceph_mdsc_get_request(nextreq);
3753                         mutex_unlock(&mdsc->mutex);
3754                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
3755                              req->r_tid, want_tid);
3756                         wait_for_completion(&req->r_safe_completion);
3757                         mutex_lock(&mdsc->mutex);
3758                         ceph_mdsc_put_request(req);
3759                         if (!nextreq)
3760                                 break;  /* next dne before, so we're done! */
3761                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
3762                                 /* next request was removed from tree */
3763                                 ceph_mdsc_put_request(nextreq);
3764                                 goto restart;
3765                         }
3766                         ceph_mdsc_put_request(nextreq);  /* won't go away */
3767                 }
3768                 req = nextreq;
3769         }
3770         mutex_unlock(&mdsc->mutex);
3771         dout("wait_unsafe_requests done\n");
3772 }
3773
3774 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3775 {
3776         u64 want_tid, want_flush;
3777
3778         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3779                 return;
3780
3781         dout("sync\n");
3782         mutex_lock(&mdsc->mutex);
3783         want_tid = mdsc->last_tid;
3784         mutex_unlock(&mdsc->mutex);
3785
3786         ceph_flush_dirty_caps(mdsc);
3787         spin_lock(&mdsc->cap_dirty_lock);
3788         want_flush = mdsc->last_cap_flush_tid;
3789         if (!list_empty(&mdsc->cap_flush_list)) {
3790                 struct ceph_cap_flush *cf =
3791                         list_last_entry(&mdsc->cap_flush_list,
3792                                         struct ceph_cap_flush, g_list);
3793                 cf->wake = true;
3794         }
3795         spin_unlock(&mdsc->cap_dirty_lock);
3796
3797         dout("sync want tid %lld flush_seq %lld\n",
3798              want_tid, want_flush);
3799
3800         wait_unsafe_requests(mdsc, want_tid);
3801         wait_caps_flush(mdsc, want_flush);
3802 }
3803
3804 /*
3805  * true if all sessions are closed, or we force unmount
3806  */
3807 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
3808 {
3809         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3810                 return true;
3811         return atomic_read(&mdsc->num_sessions) <= skipped;
3812 }
3813
3814 /*
3815  * called after sb is ro.
3816  */
3817 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3818 {
3819         struct ceph_options *opts = mdsc->fsc->client->options;
3820         struct ceph_mds_session *session;
3821         int i;
3822         int skipped = 0;
3823
3824         dout("close_sessions\n");
3825
3826         /* close sessions */
3827         mutex_lock(&mdsc->mutex);
3828         for (i = 0; i < mdsc->max_sessions; i++) {
3829                 session = __ceph_lookup_mds_session(mdsc, i);
3830                 if (!session)
3831                         continue;
3832                 mutex_unlock(&mdsc->mutex);
3833                 mutex_lock(&session->s_mutex);
3834                 if (__close_session(mdsc, session) <= 0)
3835                         skipped++;
3836                 mutex_unlock(&session->s_mutex);
3837                 ceph_put_mds_session(session);
3838                 mutex_lock(&mdsc->mutex);
3839         }
3840         mutex_unlock(&mdsc->mutex);
3841
3842         dout("waiting for sessions to close\n");
3843         wait_event_timeout(mdsc->session_close_wq,
3844                            done_closing_sessions(mdsc, skipped),
3845                            ceph_timeout_jiffies(opts->mount_timeout));
3846
3847         /* tear down remaining sessions */
3848         mutex_lock(&mdsc->mutex);
3849         for (i = 0; i < mdsc->max_sessions; i++) {
3850                 if (mdsc->sessions[i]) {
3851                         session = get_session(mdsc->sessions[i]);
3852                         __unregister_session(mdsc, session);
3853                         mutex_unlock(&mdsc->mutex);
3854                         mutex_lock(&session->s_mutex);
3855                         remove_session_caps(session);
3856                         mutex_unlock(&session->s_mutex);
3857                         ceph_put_mds_session(session);
3858                         mutex_lock(&mdsc->mutex);
3859                 }
3860         }
3861         WARN_ON(!list_empty(&mdsc->cap_delay_list));
3862         mutex_unlock(&mdsc->mutex);
3863
3864         ceph_cleanup_empty_realms(mdsc);
3865
3866         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3867
3868         dout("stopped\n");
3869 }
3870
3871 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
3872 {
3873         struct ceph_mds_session *session;
3874         int mds;
3875
3876         dout("force umount\n");
3877
3878         mutex_lock(&mdsc->mutex);
3879         for (mds = 0; mds < mdsc->max_sessions; mds++) {
3880                 session = __ceph_lookup_mds_session(mdsc, mds);
3881                 if (!session)
3882                         continue;
3883                 mutex_unlock(&mdsc->mutex);
3884                 mutex_lock(&session->s_mutex);
3885                 __close_session(mdsc, session);
3886                 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
3887                         cleanup_session_requests(mdsc, session);
3888                         remove_session_caps(session);
3889                 }
3890                 mutex_unlock(&session->s_mutex);
3891                 ceph_put_mds_session(session);
3892                 mutex_lock(&mdsc->mutex);
3893                 kick_requests(mdsc, mds);
3894         }
3895         __wake_requests(mdsc, &mdsc->waiting_for_map);
3896         mutex_unlock(&mdsc->mutex);
3897 }
3898
3899 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3900 {
3901         dout("stop\n");
3902         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3903         if (mdsc->mdsmap)
3904                 ceph_mdsmap_destroy(mdsc->mdsmap);
3905         kfree(mdsc->sessions);
3906         ceph_caps_finalize(mdsc);
3907         ceph_pool_perm_destroy(mdsc);
3908 }
3909
3910 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3911 {
3912         struct ceph_mds_client *mdsc = fsc->mdsc;
3913         dout("mdsc_destroy %p\n", mdsc);
3914
3915         if (!mdsc)
3916                 return;
3917
3918         /* flush out any connection work with references to us */
3919         ceph_msgr_flush();
3920
3921         ceph_mdsc_stop(mdsc);
3922
3923         fsc->mdsc = NULL;
3924         kfree(mdsc);
3925         dout("mdsc_destroy %p done\n", mdsc);
3926 }
3927
3928 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3929 {
3930         struct ceph_fs_client *fsc = mdsc->fsc;
3931         const char *mds_namespace = fsc->mount_options->mds_namespace;
3932         void *p = msg->front.iov_base;
3933         void *end = p + msg->front.iov_len;
3934         u32 epoch;
3935         u32 map_len;
3936         u32 num_fs;
3937         u32 mount_fscid = (u32)-1;
3938         u8 struct_v, struct_cv;
3939         int err = -EINVAL;
3940
3941         ceph_decode_need(&p, end, sizeof(u32), bad);
3942         epoch = ceph_decode_32(&p);
3943
3944         dout("handle_fsmap epoch %u\n", epoch);
3945
3946         ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3947         struct_v = ceph_decode_8(&p);
3948         struct_cv = ceph_decode_8(&p);
3949         map_len = ceph_decode_32(&p);
3950
3951         ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
3952         p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
3953
3954         num_fs = ceph_decode_32(&p);
3955         while (num_fs-- > 0) {
3956                 void *info_p, *info_end;
3957                 u32 info_len;
3958                 u8 info_v, info_cv;
3959                 u32 fscid, namelen;
3960
3961                 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3962                 info_v = ceph_decode_8(&p);
3963                 info_cv = ceph_decode_8(&p);
3964                 info_len = ceph_decode_32(&p);
3965                 ceph_decode_need(&p, end, info_len, bad);
3966                 info_p = p;
3967                 info_end = p + info_len;
3968                 p = info_end;
3969
3970                 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
3971                 fscid = ceph_decode_32(&info_p);
3972                 namelen = ceph_decode_32(&info_p);
3973                 ceph_decode_need(&info_p, info_end, namelen, bad);
3974
3975                 if (mds_namespace &&
3976                     strlen(mds_namespace) == namelen &&
3977                     !strncmp(mds_namespace, (char *)info_p, namelen)) {
3978                         mount_fscid = fscid;
3979                         break;
3980                 }
3981         }
3982
3983         ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
3984         if (mount_fscid != (u32)-1) {
3985                 fsc->client->monc.fs_cluster_id = mount_fscid;
3986                 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
3987                                    0, true);
3988                 ceph_monc_renew_subs(&fsc->client->monc);
3989         } else {
3990                 err = -ENOENT;
3991                 goto err_out;
3992         }
3993         return;
3994
3995 bad:
3996         pr_err("error decoding fsmap\n");
3997 err_out:
3998         mutex_lock(&mdsc->mutex);
3999         mdsc->mdsmap_err = err;
4000         __wake_requests(mdsc, &mdsc->waiting_for_map);
4001         mutex_unlock(&mdsc->mutex);
4002 }
4003
4004 /*
4005  * handle mds map update.
4006  */
4007 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4008 {
4009         u32 epoch;
4010         u32 maplen;
4011         void *p = msg->front.iov_base;
4012         void *end = p + msg->front.iov_len;
4013         struct ceph_mdsmap *newmap, *oldmap;
4014         struct ceph_fsid fsid;
4015         int err = -EINVAL;
4016
4017         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
4018         ceph_decode_copy(&p, &fsid, sizeof(fsid));
4019         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
4020                 return;
4021         epoch = ceph_decode_32(&p);
4022         maplen = ceph_decode_32(&p);
4023         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
4024
4025         /* do we need it? */
4026         mutex_lock(&mdsc->mutex);
4027         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
4028                 dout("handle_map epoch %u <= our %u\n",
4029                      epoch, mdsc->mdsmap->m_epoch);
4030                 mutex_unlock(&mdsc->mutex);
4031                 return;
4032         }
4033
4034         newmap = ceph_mdsmap_decode(&p, end);
4035         if (IS_ERR(newmap)) {
4036                 err = PTR_ERR(newmap);
4037                 goto bad_unlock;
4038         }
4039
4040         /* swap into place */
4041         if (mdsc->mdsmap) {
4042                 oldmap = mdsc->mdsmap;
4043                 mdsc->mdsmap = newmap;
4044                 check_new_map(mdsc, newmap, oldmap);
4045                 ceph_mdsmap_destroy(oldmap);
4046         } else {
4047                 mdsc->mdsmap = newmap;  /* first mds map */
4048         }
4049         mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
4050                                         MAX_LFS_FILESIZE);
4051
4052         __wake_requests(mdsc, &mdsc->waiting_for_map);
4053         ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
4054                           mdsc->mdsmap->m_epoch);
4055
4056         mutex_unlock(&mdsc->mutex);
4057         schedule_delayed(mdsc);
4058         return;
4059
4060 bad_unlock:
4061         mutex_unlock(&mdsc->mutex);
4062 bad:
4063         pr_err("error decoding mdsmap %d\n", err);
4064         return;
4065 }
4066
4067 static struct ceph_connection *con_get(struct ceph_connection *con)
4068 {
4069         struct ceph_mds_session *s = con->private;
4070
4071         if (get_session(s)) {
4072                 dout("mdsc con_get %p ok (%d)\n", s, refcount_read(&s->s_ref));
4073                 return con;
4074         }
4075         dout("mdsc con_get %p FAIL\n", s);
4076         return NULL;
4077 }
4078
4079 static void con_put(struct ceph_connection *con)
4080 {
4081         struct ceph_mds_session *s = con->private;
4082
4083         dout("mdsc con_put %p (%d)\n", s, refcount_read(&s->s_ref) - 1);
4084         ceph_put_mds_session(s);
4085 }
4086
4087 /*
4088  * if the client is unresponsive for long enough, the mds will kill
4089  * the session entirely.
4090  */
4091 static void peer_reset(struct ceph_connection *con)
4092 {
4093         struct ceph_mds_session *s = con->private;
4094         struct ceph_mds_client *mdsc = s->s_mdsc;
4095
4096         pr_warn("mds%d closed our session\n", s->s_mds);
4097         send_mds_reconnect(mdsc, s);
4098 }
4099
4100 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
4101 {
4102         struct ceph_mds_session *s = con->private;
4103         struct ceph_mds_client *mdsc = s->s_mdsc;
4104         int type = le16_to_cpu(msg->hdr.type);
4105
4106         mutex_lock(&mdsc->mutex);
4107         if (__verify_registered_session(mdsc, s) < 0) {
4108                 mutex_unlock(&mdsc->mutex);
4109                 goto out;
4110         }
4111         mutex_unlock(&mdsc->mutex);
4112
4113         switch (type) {
4114         case CEPH_MSG_MDS_MAP:
4115                 ceph_mdsc_handle_mdsmap(mdsc, msg);
4116                 break;
4117         case CEPH_MSG_FS_MAP_USER:
4118                 ceph_mdsc_handle_fsmap(mdsc, msg);
4119                 break;
4120         case CEPH_MSG_CLIENT_SESSION:
4121                 handle_session(s, msg);
4122                 break;
4123         case CEPH_MSG_CLIENT_REPLY:
4124                 handle_reply(s, msg);
4125                 break;
4126         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
4127                 handle_forward(mdsc, s, msg);
4128                 break;
4129         case CEPH_MSG_CLIENT_CAPS:
4130                 ceph_handle_caps(s, msg);
4131                 break;
4132         case CEPH_MSG_CLIENT_SNAP:
4133                 ceph_handle_snap(mdsc, s, msg);
4134                 break;
4135         case CEPH_MSG_CLIENT_LEASE:
4136                 handle_lease(mdsc, s, msg);
4137                 break;
4138         case CEPH_MSG_CLIENT_QUOTA:
4139                 ceph_handle_quota(mdsc, s, msg);
4140                 break;
4141
4142         default:
4143                 pr_err("received unknown message type %d %s\n", type,
4144                        ceph_msg_type_name(type));
4145         }
4146 out:
4147         ceph_msg_put(msg);
4148 }
4149
4150 /*
4151  * authentication
4152  */
4153
4154 /*
4155  * Note: returned pointer is the address of a structure that's
4156  * managed separately.  Caller must *not* attempt to free it.
4157  */
4158 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
4159                                         int *proto, int force_new)
4160 {
4161         struct ceph_mds_session *s = con->private;
4162         struct ceph_mds_client *mdsc = s->s_mdsc;
4163         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4164         struct ceph_auth_handshake *auth = &s->s_auth;
4165
4166         if (force_new && auth->authorizer) {
4167                 ceph_auth_destroy_authorizer(auth->authorizer);
4168                 auth->authorizer = NULL;
4169         }
4170         if (!auth->authorizer) {
4171                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4172                                                       auth);
4173                 if (ret)
4174                         return ERR_PTR(ret);
4175         } else {
4176                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4177                                                       auth);
4178                 if (ret)
4179                         return ERR_PTR(ret);
4180         }
4181         *proto = ac->protocol;
4182
4183         return auth;
4184 }
4185
4186 static int add_authorizer_challenge(struct ceph_connection *con,
4187                                     void *challenge_buf, int challenge_buf_len)
4188 {
4189         struct ceph_mds_session *s = con->private;
4190         struct ceph_mds_client *mdsc = s->s_mdsc;
4191         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4192
4193         return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
4194                                             challenge_buf, challenge_buf_len);
4195 }
4196
4197 static int verify_authorizer_reply(struct ceph_connection *con)
4198 {
4199         struct ceph_mds_session *s = con->private;
4200         struct ceph_mds_client *mdsc = s->s_mdsc;
4201         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4202
4203         return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
4204 }
4205
4206 static int invalidate_authorizer(struct ceph_connection *con)
4207 {
4208         struct ceph_mds_session *s = con->private;
4209         struct ceph_mds_client *mdsc = s->s_mdsc;
4210         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4211
4212         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
4213
4214         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
4215 }
4216
4217 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
4218                                 struct ceph_msg_header *hdr, int *skip)
4219 {
4220         struct ceph_msg *msg;
4221         int type = (int) le16_to_cpu(hdr->type);
4222         int front_len = (int) le32_to_cpu(hdr->front_len);
4223
4224         if (con->in_msg)
4225                 return con->in_msg;
4226
4227         *skip = 0;
4228         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
4229         if (!msg) {
4230                 pr_err("unable to allocate msg type %d len %d\n",
4231                        type, front_len);
4232                 return NULL;
4233         }
4234
4235         return msg;
4236 }
4237
4238 static int mds_sign_message(struct ceph_msg *msg)
4239 {
4240        struct ceph_mds_session *s = msg->con->private;
4241        struct ceph_auth_handshake *auth = &s->s_auth;
4242
4243        return ceph_auth_sign_message(auth, msg);
4244 }
4245
4246 static int mds_check_message_signature(struct ceph_msg *msg)
4247 {
4248        struct ceph_mds_session *s = msg->con->private;
4249        struct ceph_auth_handshake *auth = &s->s_auth;
4250
4251        return ceph_auth_check_message_signature(auth, msg);
4252 }
4253
4254 static const struct ceph_connection_operations mds_con_ops = {
4255         .get = con_get,
4256         .put = con_put,
4257         .dispatch = dispatch,
4258         .get_authorizer = get_authorizer,
4259         .add_authorizer_challenge = add_authorizer_challenge,
4260         .verify_authorizer_reply = verify_authorizer_reply,
4261         .invalidate_authorizer = invalidate_authorizer,
4262         .peer_reset = peer_reset,
4263         .alloc_msg = mds_alloc_msg,
4264         .sign_message = mds_sign_message,
4265         .check_message_signature = mds_check_message_signature,
4266 };
4267
4268 /* eof */