fs/ceph/mds_client.c

   1 #include <linux/ceph/ceph_debug.h>
   2
   3 #include <linux/fs.h>
   4 #include <linux/wait.h>
   5 #include <linux/slab.h>
   6 #include <linux/gfp.h>
   7 #include <linux/sched.h>
   8 #include <linux/debugfs.h>
   9 #include <linux/seq_file.h>
  10 #include <linux/utsname.h>
  11 #include <linux/ratelimit.h>
  12
  13 #include "super.h"
  14 #include "mds_client.h"
  15
  16 #include <linux/ceph/ceph_features.h>
  17 #include <linux/ceph/messenger.h>
  18 #include <linux/ceph/decode.h>
  19 #include <linux/ceph/pagelist.h>
  20 #include <linux/ceph/auth.h>
  21 #include <linux/ceph/debugfs.h>
  22
  23 /*
  24  * A cluster of MDS (metadata server) daemons is responsible for
  25  * managing the file system namespace (the directory hierarchy and
  26  * inodes) and for coordinating shared access to storage.  Metadata is
  27  * partitioning hierarchically across a number of servers, and that
  28  * partition varies over time as the cluster adjusts the distribution
  29  * in order to balance load.
  30  *
  31  * The MDS client is primarily responsible to managing synchronous
  32  * metadata requests for operations like open, unlink, and so forth.
  33  * If there is a MDS failure, we find out about it when we (possibly
  34  * request and) receive a new MDS map, and can resubmit affected
  35  * requests.
  36  *
  37  * For the most part, though, we take advantage of a lossless
  38  * communications channel to the MDS, and do not need to worry about
  39  * timing out or resubmitting requests.
  40  *
  41  * We maintain a stateful "session" with each MDS we interact with.
  42  * Within each session, we sent periodic heartbeat messages to ensure
  43  * any capabilities or leases we have been issues remain valid.  If
  44  * the session times out and goes stale, our leases and capabilities
  45  * are no longer valid.
  46  */
  47
  48 struct ceph_reconnect_state {
  49         int nr_caps;
  50         struct ceph_pagelist *pagelist;
  51         unsigned msg_version;
  52 };
  53
  54 static void __wake_requests(struct ceph_mds_client *mdsc,
  55                             struct list_head *head);
  56
  57 static const struct ceph_connection_operations mds_con_ops;
  58
  59
  60 /*
  61  * mds reply parsing
  62  */
  63
  64 /*
  65  * parse individual inode info
  66  */
  67 static int parse_reply_info_in(void **p, void *end,
  68                                struct ceph_mds_reply_info_in *info,
  69                                u64 features)
  70 {
  71         int err = -EIO;
  72
  73         info->in = *p;
  74         *p += sizeof(struct ceph_mds_reply_inode) +
  75                 sizeof(*info->in->fragtree.splits) *
  76                 le32_to_cpu(info->in->fragtree.nsplits);
  77
  78         ceph_decode_32_safe(p, end, info->symlink_len, bad);
  79         ceph_decode_need(p, end, info->symlink_len, bad);
  80         info->symlink = *p;
  81         *p += info->symlink_len;
  82
  83         if (features & CEPH_FEATURE_DIRLAYOUTHASH)
  84                 ceph_decode_copy_safe(p, end, &info->dir_layout,
  85                                       sizeof(info->dir_layout), bad);
  86         else
  87                 memset(&info->dir_layout, 0, sizeof(info->dir_layout));
  88
  89         ceph_decode_32_safe(p, end, info->xattr_len, bad);
  90         ceph_decode_need(p, end, info->xattr_len, bad);
  91         info->xattr_data = *p;
  92         *p += info->xattr_len;
  93
  94         if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
  95                 ceph_decode_64_safe(p, end, info->inline_version, bad);
  96                 ceph_decode_32_safe(p, end, info->inline_len, bad);
  97                 ceph_decode_need(p, end, info->inline_len, bad);
  98                 info->inline_data = *p;
  99                 *p += info->inline_len;
 100         } else
 101                 info->inline_version = CEPH_INLINE_NONE;
 102
 103         info->pool_ns_len = 0;
 104         info->pool_ns_data = NULL;
 105         if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
 106                 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
 107                 if (info->pool_ns_len > 0) {
 108                         ceph_decode_need(p, end, info->pool_ns_len, bad);
 109                         info->pool_ns_data = *p;
 110                         *p += info->pool_ns_len;
 111                 }
 112         }
 113
 114         return 0;
 115 bad:
 116         return err;
 117 }
 118
 119 /*
 120  * parse a normal reply, which may contain a (dir+)dentry and/or a
 121  * target inode.
 122  */
 123 static int parse_reply_info_trace(void **p, void *end,
 124                                   struct ceph_mds_reply_info_parsed *info,
 125                                   u64 features)
 126 {
 127         int err;
 128
 129         if (info->head->is_dentry) {
 130                 err = parse_reply_info_in(p, end, &info->diri, features);
 131                 if (err < 0)
 132                         goto out_bad;
 133
 134                 if (unlikely(*p + sizeof(*info->dirfrag) > end))
 135                         goto bad;
 136                 info->dirfrag = *p;
 137                 *p += sizeof(*info->dirfrag) +
 138                         sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
 139                 if (unlikely(*p > end))
 140                         goto bad;
 141
 142                 ceph_decode_32_safe(p, end, info->dname_len, bad);
 143                 ceph_decode_need(p, end, info->dname_len, bad);
 144                 info->dname = *p;
 145                 *p += info->dname_len;
 146                 info->dlease = *p;
 147                 *p += sizeof(*info->dlease);
 148         }
 149
 150         if (info->head->is_target) {
 151                 err = parse_reply_info_in(p, end, &info->targeti, features);
 152                 if (err < 0)
 153                         goto out_bad;
 154         }
 155
 156         if (unlikely(*p != end))
 157                 goto bad;
 158         return 0;
 159
 160 bad:
 161         err = -EIO;
 162 out_bad:
 163         pr_err("problem parsing mds trace %d\n", err);
 164         return err;
 165 }
 166
 167 /*
 168  * parse readdir results
 169  */
 170 static int parse_reply_info_dir(void **p, void *end,
 171                                 struct ceph_mds_reply_info_parsed *info,
 172                                 u64 features)
 173 {
 174         u32 num, i = 0;
 175         int err;
 176
 177         info->dir_dir = *p;
 178         if (*p + sizeof(*info->dir_dir) > end)
 179                 goto bad;
 180         *p += sizeof(*info->dir_dir) +
 181                 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
 182         if (*p > end)
 183                 goto bad;
 184
 185         ceph_decode_need(p, end, sizeof(num) + 2, bad);
 186         num = ceph_decode_32(p);
 187         {
 188                 u16 flags = ceph_decode_16(p);
 189                 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
 190                 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
 191                 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
 192         }
 193         if (num == 0)
 194                 goto done;
 195
 196         BUG_ON(!info->dir_entries);
 197         if ((unsigned long)(info->dir_entries + num) >
 198             (unsigned long)info->dir_entries + info->dir_buf_size) {
 199                 pr_err("dir contents are larger than expected\n");
 200                 WARN_ON(1);
 201                 goto bad;
 202         }
 203
 204         info->dir_nr = num;
 205         while (num) {
 206                 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
 207                 /* dentry */
 208                 ceph_decode_need(p, end, sizeof(u32)*2, bad);
 209                 rde->name_len = ceph_decode_32(p);
 210                 ceph_decode_need(p, end, rde->name_len, bad);
 211                 rde->name = *p;
 212                 *p += rde->name_len;
 213                 dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
 214                 rde->lease = *p;
 215                 *p += sizeof(struct ceph_mds_reply_lease);
 216
 217                 /* inode */
 218                 err = parse_reply_info_in(p, end, &rde->inode, features);
 219                 if (err < 0)
 220                         goto out_bad;
 221                 /* ceph_readdir_prepopulate() will update it */
 222                 rde->offset = 0;
 223                 i++;
 224                 num--;
 225         }
 226
 227 done:
 228         if (*p != end)
 229                 goto bad;
 230         return 0;
 231
 232 bad:
 233         err = -EIO;
 234 out_bad:
 235         pr_err("problem parsing dir contents %d\n", err);
 236         return err;
 237 }
 238
 239 /*
 240  * parse fcntl F_GETLK results
 241  */
 242 static int parse_reply_info_filelock(void **p, void *end,
 243                                      struct ceph_mds_reply_info_parsed *info,
 244                                      u64 features)
 245 {
 246         if (*p + sizeof(*info->filelock_reply) > end)
 247                 goto bad;
 248
 249         info->filelock_reply = *p;
 250         *p += sizeof(*info->filelock_reply);
 251
 252         if (unlikely(*p != end))
 253                 goto bad;
 254         return 0;
 255
 256 bad:
 257         return -EIO;
 258 }
 259
 260 /*
 261  * parse create results
 262  */
 263 static int parse_reply_info_create(void **p, void *end,
 264                                   struct ceph_mds_reply_info_parsed *info,
 265                                   u64 features)
 266 {
 267         if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
 268                 if (*p == end) {
 269                         info->has_create_ino = false;
 270                 } else {
 271                         info->has_create_ino = true;
 272                         info->ino = ceph_decode_64(p);
 273                 }
 274         }
 275
 276         if (unlikely(*p != end))
 277                 goto bad;
 278         return 0;
 279
 280 bad:
 281         return -EIO;
 282 }
 283
 284 /*
 285  * parse extra results
 286  */
 287 static int parse_reply_info_extra(void **p, void *end,
 288                                   struct ceph_mds_reply_info_parsed *info,
 289                                   u64 features)
 290 {
 291         u32 op = le32_to_cpu(info->head->op);
 292
 293         if (op == CEPH_MDS_OP_GETFILELOCK)
 294                 return parse_reply_info_filelock(p, end, info, features);
 295         else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
 296                 return parse_reply_info_dir(p, end, info, features);
 297         else if (op == CEPH_MDS_OP_CREATE)
 298                 return parse_reply_info_create(p, end, info, features);
 299         else
 300                 return -EIO;
 301 }
 302
 303 /*
 304  * parse entire mds reply
 305  */
 306 static int parse_reply_info(struct ceph_msg *msg,
 307                             struct ceph_mds_reply_info_parsed *info,
 308                             u64 features)
 309 {
 310         void *p, *end;
 311         u32 len;
 312         int err;
 313
 314         info->head = msg->front.iov_base;
 315         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 316         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 317
 318         /* trace */
 319         ceph_decode_32_safe(&p, end, len, bad);
 320         if (len > 0) {
 321                 ceph_decode_need(&p, end, len, bad);
 322                 err = parse_reply_info_trace(&p, p+len, info, features);
 323                 if (err < 0)
 324                         goto out_bad;
 325         }
 326
 327         /* extra */
 328         ceph_decode_32_safe(&p, end, len, bad);
 329         if (len > 0) {
 330                 ceph_decode_need(&p, end, len, bad);
 331                 err = parse_reply_info_extra(&p, p+len, info, features);
 332                 if (err < 0)
 333                         goto out_bad;
 334         }
 335
 336         /* snap blob */
 337         ceph_decode_32_safe(&p, end, len, bad);
 338         info->snapblob_len = len;
 339         info->snapblob = p;
 340         p += len;
 341
 342         if (p != end)
 343                 goto bad;
 344         return 0;
 345
 346 bad:
 347         err = -EIO;
 348 out_bad:
 349         pr_err("mds parse_reply err %d\n", err);
 350         return err;
 351 }
 352
 353 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
 354 {
 355         if (!info->dir_entries)
 356                 return;
 357         free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
 358 }
 359
 360
 361 /*
 362  * sessions
 363  */
 364 const char *ceph_session_state_name(int s)
 365 {
 366         switch (s) {
 367         case CEPH_MDS_SESSION_NEW: return "new";
 368         case CEPH_MDS_SESSION_OPENING: return "opening";
 369         case CEPH_MDS_SESSION_OPEN: return "open";
 370         case CEPH_MDS_SESSION_HUNG: return "hung";
 371         case CEPH_MDS_SESSION_CLOSING: return "closing";
 372         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
 373         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
 374         case CEPH_MDS_SESSION_REJECTED: return "rejected";
 375         default: return "???";
 376         }
 377 }
 378
 379 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
 380 {
 381         if (atomic_inc_not_zero(&s->s_ref)) {
 382                 dout("mdsc get_session %p %d -> %d\n", s,
 383                      atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
 384                 return s;
 385         } else {
 386                 dout("mdsc get_session %p 0 -- FAIL", s);
 387                 return NULL;
 388         }
 389 }
 390
 391 void ceph_put_mds_session(struct ceph_mds_session *s)
 392 {
 393         dout("mdsc put_session %p %d -> %d\n", s,
 394              atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
 395         if (atomic_dec_and_test(&s->s_ref)) {
 396                 if (s->s_auth.authorizer)
 397                         ceph_auth_destroy_authorizer(s->s_auth.authorizer);
 398                 kfree(s);
 399         }
 400 }
 401
 402 /*
 403  * called under mdsc->mutex
 404  */
 405 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
 406                                                    int mds)
 407 {
 408         struct ceph_mds_session *session;
 409
 410         if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
 411                 return NULL;
 412         session = mdsc->sessions[mds];
 413         dout("lookup_mds_session %p %d\n", session,
 414              atomic_read(&session->s_ref));
 415         get_session(session);
 416         return session;
 417 }
 418
 419 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
 420 {
 421         if (mds >= mdsc->max_sessions)
 422                 return false;
 423         return mdsc->sessions[mds];
 424 }
 425
 426 static int __verify_registered_session(struct ceph_mds_client *mdsc,
 427                                        struct ceph_mds_session *s)
 428 {
 429         if (s->s_mds >= mdsc->max_sessions ||
 430             mdsc->sessions[s->s_mds] != s)
 431                 return -ENOENT;
 432         return 0;
 433 }
 434
 435 /*
 436  * create+register a new session for given mds.
 437  * called under mdsc->mutex.
 438  */
 439 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 440                                                  int mds)
 441 {
 442         struct ceph_mds_session *s;
 443
 444         if (mds >= mdsc->mdsmap->m_max_mds)
 445                 return ERR_PTR(-EINVAL);
 446
 447         s = kzalloc(sizeof(*s), GFP_NOFS);
 448         if (!s)
 449                 return ERR_PTR(-ENOMEM);
 450         s->s_mdsc = mdsc;
 451         s->s_mds = mds;
 452         s->s_state = CEPH_MDS_SESSION_NEW;
 453         s->s_ttl = 0;
 454         s->s_seq = 0;
 455         mutex_init(&s->s_mutex);
 456
 457         ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
 458
 459         spin_lock_init(&s->s_gen_ttl_lock);
 460         s->s_cap_gen = 0;
 461         s->s_cap_ttl = jiffies - 1;
 462
 463         spin_lock_init(&s->s_cap_lock);
 464         s->s_renew_requested = 0;
 465         s->s_renew_seq = 0;
 466         INIT_LIST_HEAD(&s->s_caps);
 467         s->s_nr_caps = 0;
 468         s->s_trim_caps = 0;
 469         atomic_set(&s->s_ref, 1);
 470         INIT_LIST_HEAD(&s->s_waiting);
 471         INIT_LIST_HEAD(&s->s_unsafe);
 472         s->s_num_cap_releases = 0;
 473         s->s_cap_reconnect = 0;
 474         s->s_cap_iterator = NULL;
 475         INIT_LIST_HEAD(&s->s_cap_releases);
 476         INIT_LIST_HEAD(&s->s_cap_flushing);
 477
 478         dout("register_session mds%d\n", mds);
 479         if (mds >= mdsc->max_sessions) {
 480                 int newmax = 1 << get_count_order(mds+1);
 481                 struct ceph_mds_session **sa;
 482
 483                 dout("register_session realloc to %d\n", newmax);
 484                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 485                 if (sa == NULL)
 486                         goto fail_realloc;
 487                 if (mdsc->sessions) {
 488                         memcpy(sa, mdsc->sessions,
 489                                mdsc->max_sessions * sizeof(void *));
 490                         kfree(mdsc->sessions);
 491                 }
 492                 mdsc->sessions = sa;
 493                 mdsc->max_sessions = newmax;
 494         }
 495         mdsc->sessions[mds] = s;
 496         atomic_inc(&mdsc->num_sessions);
 497         atomic_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
 498
 499         ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
 500                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
 501
 502         return s;
 503
 504 fail_realloc:
 505         kfree(s);
 506         return ERR_PTR(-ENOMEM);
 507 }
 508
 509 /*
 510  * called under mdsc->mutex
 511  */
 512 static void __unregister_session(struct ceph_mds_client *mdsc,
 513                                struct ceph_mds_session *s)
 514 {
 515         dout("__unregister_session mds%d %p\n", s->s_mds, s);
 516         BUG_ON(mdsc->sessions[s->s_mds] != s);
 517         mdsc->sessions[s->s_mds] = NULL;
 518         ceph_con_close(&s->s_con);
 519         ceph_put_mds_session(s);
 520         atomic_dec(&mdsc->num_sessions);
 521 }
 522
 523 /*
 524  * drop session refs in request.
 525  *
 526  * should be last request ref, or hold mdsc->mutex
 527  */
 528 static void put_request_session(struct ceph_mds_request *req)
 529 {
 530         if (req->r_session) {
 531                 ceph_put_mds_session(req->r_session);
 532                 req->r_session = NULL;
 533         }
 534 }
 535
 536 void ceph_mdsc_release_request(struct kref *kref)
 537 {
 538         struct ceph_mds_request *req = container_of(kref,
 539                                                     struct ceph_mds_request,
 540                                                     r_kref);
 541         destroy_reply_info(&req->r_reply_info);
 542         if (req->r_request)
 543                 ceph_msg_put(req->r_request);
 544         if (req->r_reply)
 545                 ceph_msg_put(req->r_reply);
 546         if (req->r_inode) {
 547                 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
 548                 iput(req->r_inode);
 549         }
 550         if (req->r_parent)
 551                 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
 552         iput(req->r_target_inode);
 553         if (req->r_dentry)
 554                 dput(req->r_dentry);
 555         if (req->r_old_dentry)
 556                 dput(req->r_old_dentry);
 557         if (req->r_old_dentry_dir) {
 558                 /*
 559                  * track (and drop pins for) r_old_dentry_dir
 560                  * separately, since r_old_dentry's d_parent may have
 561                  * changed between the dir mutex being dropped and
 562                  * this request being freed.
 563                  */
 564                 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
 565                                   CEPH_CAP_PIN);
 566                 iput(req->r_old_dentry_dir);
 567         }
 568         kfree(req->r_path1);
 569         kfree(req->r_path2);
 570         if (req->r_pagelist)
 571                 ceph_pagelist_release(req->r_pagelist);
 572         put_request_session(req);
 573         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
 574         kfree(req);
 575 }
 576
 577 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
 578
 579 /*
 580  * lookup session, bump ref if found.
 581  *
 582  * called under mdsc->mutex.
 583  */
 584 static struct ceph_mds_request *
 585 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
 586 {
 587         struct ceph_mds_request *req;
 588
 589         req = lookup_request(&mdsc->request_tree, tid);
 590         if (req)
 591                 ceph_mdsc_get_request(req);
 592
 593         return req;
 594 }
 595
 596 /*
 597  * Register an in-flight request, and assign a tid.  Link to directory
 598  * are modifying (if any).
 599  *
 600  * Called under mdsc->mutex.
 601  */
 602 static void __register_request(struct ceph_mds_client *mdsc,
 603                                struct ceph_mds_request *req,
 604                                struct inode *dir)
 605 {
 606         req->r_tid = ++mdsc->last_tid;
 607         if (req->r_num_caps)
 608                 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
 609                                   req->r_num_caps);
 610         dout("__register_request %p tid %lld\n", req, req->r_tid);
 611         ceph_mdsc_get_request(req);
 612         insert_request(&mdsc->request_tree, req);
 613
 614         req->r_uid = current_fsuid();
 615         req->r_gid = current_fsgid();
 616
 617         if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
 618                 mdsc->oldest_tid = req->r_tid;
 619
 620         if (dir) {
 621                 ihold(dir);
 622                 req->r_unsafe_dir = dir;
 623         }
 624 }
 625
 626 static void __unregister_request(struct ceph_mds_client *mdsc,
 627                                  struct ceph_mds_request *req)
 628 {
 629         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 630
 631         /* Never leave an unregistered request on an unsafe list! */
 632         list_del_init(&req->r_unsafe_item);
 633
 634         if (req->r_tid == mdsc->oldest_tid) {
 635                 struct rb_node *p = rb_next(&req->r_node);
 636                 mdsc->oldest_tid = 0;
 637                 while (p) {
 638                         struct ceph_mds_request *next_req =
 639                                 rb_entry(p, struct ceph_mds_request, r_node);
 640                         if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
 641                                 mdsc->oldest_tid = next_req->r_tid;
 642                                 break;
 643                         }
 644                         p = rb_next(p);
 645                 }
 646         }
 647
 648         erase_request(&mdsc->request_tree, req);
 649
 650         if (req->r_unsafe_dir  &&
 651             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
 652                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 653                 spin_lock(&ci->i_unsafe_lock);
 654                 list_del_init(&req->r_unsafe_dir_item);
 655                 spin_unlock(&ci->i_unsafe_lock);
 656         }
 657         if (req->r_target_inode &&
 658             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
 659                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
 660                 spin_lock(&ci->i_unsafe_lock);
 661                 list_del_init(&req->r_unsafe_target_item);
 662                 spin_unlock(&ci->i_unsafe_lock);
 663         }
 664
 665         if (req->r_unsafe_dir) {
 666                 iput(req->r_unsafe_dir);
 667                 req->r_unsafe_dir = NULL;
 668         }
 669
 670         complete_all(&req->r_safe_completion);
 671
 672         ceph_mdsc_put_request(req);
 673 }
 674
 675 /*
 676  * Walk back up the dentry tree until we hit a dentry representing a
 677  * non-snapshot inode. We do this using the rcu_read_lock (which must be held
 678  * when calling this) to ensure that the objects won't disappear while we're
 679  * working with them. Once we hit a candidate dentry, we attempt to take a
 680  * reference to it, and return that as the result.
 681  */
 682 static struct inode *get_nonsnap_parent(struct dentry *dentry)
 683 {
 684         struct inode *inode = NULL;
 685
 686         while (dentry && !IS_ROOT(dentry)) {
 687                 inode = d_inode_rcu(dentry);
 688                 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
 689                         break;
 690                 dentry = dentry->d_parent;
 691         }
 692         if (inode)
 693                 inode = igrab(inode);
 694         return inode;
 695 }
 696
 697 /*
 698  * Choose mds to send request to next.  If there is a hint set in the
 699  * request (e.g., due to a prior forward hint from the mds), use that.
 700  * Otherwise, consult frag tree and/or caps to identify the
 701  * appropriate mds.  If all else fails, choose randomly.
 702  *
 703  * Called under mdsc->mutex.
 704  */
 705 static int __choose_mds(struct ceph_mds_client *mdsc,
 706                         struct ceph_mds_request *req)
 707 {
 708         struct inode *inode;
 709         struct ceph_inode_info *ci;
 710         struct ceph_cap *cap;
 711         int mode = req->r_direct_mode;
 712         int mds = -1;
 713         u32 hash = req->r_direct_hash;
 714         bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
 715
 716         /*
 717          * is there a specific mds we should try?  ignore hint if we have
 718          * no session and the mds is not up (active or recovering).
 719          */
 720         if (req->r_resend_mds >= 0 &&
 721             (__have_session(mdsc, req->r_resend_mds) ||
 722              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 723                 dout("choose_mds using resend_mds mds%d\n",
 724                      req->r_resend_mds);
 725                 return req->r_resend_mds;
 726         }
 727
 728         if (mode == USE_RANDOM_MDS)
 729                 goto random;
 730
 731         inode = NULL;
 732         if (req->r_inode) {
 733                 inode = req->r_inode;
 734                 ihold(inode);
 735         } else if (req->r_dentry) {
 736                 /* ignore race with rename; old or new d_parent is okay */
 737                 struct dentry *parent;
 738                 struct inode *dir;
 739
 740                 rcu_read_lock();
 741                 parent = req->r_dentry->d_parent;
 742                 dir = req->r_parent ? : d_inode_rcu(parent);
 743
 744                 if (!dir || dir->i_sb != mdsc->fsc->sb) {
 745                         /*  not this fs or parent went negative */
 746                         inode = d_inode(req->r_dentry);
 747                         if (inode)
 748                                 ihold(inode);
 749                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
 750                         /* direct snapped/virtual snapdir requests
 751                          * based on parent dir inode */
 752                         inode = get_nonsnap_parent(parent);
 753                         dout("__choose_mds using nonsnap parent %p\n", inode);
 754                 } else {
 755                         /* dentry target */
 756                         inode = d_inode(req->r_dentry);
 757                         if (!inode || mode == USE_AUTH_MDS) {
 758                                 /* dir + name */
 759                                 inode = igrab(dir);
 760                                 hash = ceph_dentry_hash(dir, req->r_dentry);
 761                                 is_hash = true;
 762                         } else {
 763                                 ihold(inode);
 764                         }
 765                 }
 766                 rcu_read_unlock();
 767         }
 768
 769         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 770              (int)hash, mode);
 771         if (!inode)
 772                 goto random;
 773         ci = ceph_inode(inode);
 774
 775         if (is_hash && S_ISDIR(inode->i_mode)) {
 776                 struct ceph_inode_frag frag;
 777                 int found;
 778
 779                 ceph_choose_frag(ci, hash, &frag, &found);
 780                 if (found) {
 781                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 782                                 u8 r;
 783
 784                                 /* choose a random replica */
 785                                 get_random_bytes(&r, 1);
 786                                 r %= frag.ndist;
 787                                 mds = frag.dist[r];
 788                                 dout("choose_mds %p %llx.%llx "
 789                                      "frag %u mds%d (%d/%d)\n",
 790                                      inode, ceph_vinop(inode),
 791                                      frag.frag, mds,
 792                                      (int)r, frag.ndist);
 793                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 794                                     CEPH_MDS_STATE_ACTIVE)
 795                                         goto out;
 796                         }
 797
 798                         /* since this file/dir wasn't known to be
 799                          * replicated, then we want to look for the
 800                          * authoritative mds. */
 801                         mode = USE_AUTH_MDS;
 802                         if (frag.mds >= 0) {
 803                                 /* choose auth mds */
 804                                 mds = frag.mds;
 805                                 dout("choose_mds %p %llx.%llx "
 806                                      "frag %u mds%d (auth)\n",
 807                                      inode, ceph_vinop(inode), frag.frag, mds);
 808                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 809                                     CEPH_MDS_STATE_ACTIVE)
 810                                         goto out;
 811                         }
 812                 }
 813         }
 814
 815         spin_lock(&ci->i_ceph_lock);
 816         cap = NULL;
 817         if (mode == USE_AUTH_MDS)
 818                 cap = ci->i_auth_cap;
 819         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
 820                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
 821         if (!cap) {
 822                 spin_unlock(&ci->i_ceph_lock);
 823                 iput(inode);
 824                 goto random;
 825         }
 826         mds = cap->session->s_mds;
 827         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
 828              inode, ceph_vinop(inode), mds,
 829              cap == ci->i_auth_cap ? "auth " : "", cap);
 830         spin_unlock(&ci->i_ceph_lock);
 831 out:
 832         iput(inode);
 833         return mds;
 834
 835 random:
 836         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
 837         dout("choose_mds chose random mds%d\n", mds);
 838         return mds;
 839 }
 840
 841
 842 /*
 843  * session messages
 844  */
 845 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
 846 {
 847         struct ceph_msg *msg;
 848         struct ceph_mds_session_head *h;
 849
 850         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
 851                            false);
 852         if (!msg) {
 853                 pr_err("create_session_msg ENOMEM creating msg\n");
 854                 return NULL;
 855         }
 856         h = msg->front.iov_base;
 857         h->op = cpu_to_le32(op);
 858         h->seq = cpu_to_le64(seq);
 859
 860         return msg;
 861 }
 862
 863 /*
 864  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
 865  * to include additional client metadata fields.
 866  */
 867 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
 868 {
 869         struct ceph_msg *msg;
 870         struct ceph_mds_session_head *h;
 871         int i = -1;
 872         int metadata_bytes = 0;
 873         int metadata_key_count = 0;
 874         struct ceph_options *opt = mdsc->fsc->client->options;
 875         struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
 876         void *p;
 877
 878         const char* metadata[][2] = {
 879                 {"hostname", utsname()->nodename},
 880                 {"kernel_version", utsname()->release},
 881                 {"entity_id", opt->name ? : ""},
 882                 {"root", fsopt->server_path ? : "/"},
 883                 {NULL, NULL}
 884         };
 885
 886         /* Calculate serialized length of metadata */
 887         metadata_bytes = 4;  /* map length */
 888         for (i = 0; metadata[i][0] != NULL; ++i) {
 889                 metadata_bytes += 8 + strlen(metadata[i][0]) +
 890                         strlen(metadata[i][1]);
 891                 metadata_key_count++;
 892         }
 893
 894         /* Allocate the message */
 895         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + metadata_bytes,
 896                            GFP_NOFS, false);
 897         if (!msg) {
 898                 pr_err("create_session_msg ENOMEM creating msg\n");
 899                 return NULL;
 900         }
 901         h = msg->front.iov_base;
 902         h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
 903         h->seq = cpu_to_le64(seq);
 904
 905         /*
 906          * Serialize client metadata into waiting buffer space, using
 907          * the format that userspace expects for map<string, string>
 908          *
 909          * ClientSession messages with metadata are v2
 910          */
 911         msg->hdr.version = cpu_to_le16(2);
 912         msg->hdr.compat_version = cpu_to_le16(1);
 913
 914         /* The write pointer, following the session_head structure */
 915         p = msg->front.iov_base + sizeof(*h);
 916
 917         /* Number of entries in the map */
 918         ceph_encode_32(&p, metadata_key_count);
 919
 920         /* Two length-prefixed strings for each entry in the map */
 921         for (i = 0; metadata[i][0] != NULL; ++i) {
 922                 size_t const key_len = strlen(metadata[i][0]);
 923                 size_t const val_len = strlen(metadata[i][1]);
 924
 925                 ceph_encode_32(&p, key_len);
 926                 memcpy(p, metadata[i][0], key_len);
 927                 p += key_len;
 928                 ceph_encode_32(&p, val_len);
 929                 memcpy(p, metadata[i][1], val_len);
 930                 p += val_len;
 931         }
 932
 933         return msg;
 934 }
 935
 936 /*
 937  * send session open request.
 938  *
 939  * called under mdsc->mutex
 940  */
 941 static int __open_session(struct ceph_mds_client *mdsc,
 942                           struct ceph_mds_session *session)
 943 {
 944         struct ceph_msg *msg;
 945         int mstate;
 946         int mds = session->s_mds;
 947
 948         /* wait for mds to go active? */
 949         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
 950         dout("open_session to mds%d (%s)\n", mds,
 951              ceph_mds_state_name(mstate));
 952         session->s_state = CEPH_MDS_SESSION_OPENING;
 953         session->s_renew_requested = jiffies;
 954
 955         /* send connect message */
 956         msg = create_session_open_msg(mdsc, session->s_seq);
 957         if (!msg)
 958                 return -ENOMEM;
 959         ceph_con_send(&session->s_con, msg);
 960         return 0;
 961 }
 962
 963 /*
 964  * open sessions for any export targets for the given mds
 965  *
 966  * called under mdsc->mutex
 967  */
 968 static struct ceph_mds_session *
 969 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
 970 {
 971         struct ceph_mds_session *session;
 972
 973         session = __ceph_lookup_mds_session(mdsc, target);
 974         if (!session) {
 975                 session = register_session(mdsc, target);
 976                 if (IS_ERR(session))
 977                         return session;
 978         }
 979         if (session->s_state == CEPH_MDS_SESSION_NEW ||
 980             session->s_state == CEPH_MDS_SESSION_CLOSING)
 981                 __open_session(mdsc, session);
 982
 983         return session;
 984 }
 985
 986 struct ceph_mds_session *
 987 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
 988 {
 989         struct ceph_mds_session *session;
 990
 991         dout("open_export_target_session to mds%d\n", target);
 992
 993         mutex_lock(&mdsc->mutex);
 994         session = __open_export_target_session(mdsc, target);
 995         mutex_unlock(&mdsc->mutex);
 996
 997         return session;
 998 }
 999
1000 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1001                                           struct ceph_mds_session *session)
1002 {
1003         struct ceph_mds_info *mi;
1004         struct ceph_mds_session *ts;
1005         int i, mds = session->s_mds;
1006
1007         if (mds >= mdsc->mdsmap->m_max_mds)
1008                 return;
1009
1010         mi = &mdsc->mdsmap->m_info[mds];
1011         dout("open_export_target_sessions for mds%d (%d targets)\n",
1012              session->s_mds, mi->num_export_targets);
1013
1014         for (i = 0; i < mi->num_export_targets; i++) {
1015                 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1016                 if (!IS_ERR(ts))
1017                         ceph_put_mds_session(ts);
1018         }
1019 }
1020
1021 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1022                                            struct ceph_mds_session *session)
1023 {
1024         mutex_lock(&mdsc->mutex);
1025         __open_export_target_sessions(mdsc, session);
1026         mutex_unlock(&mdsc->mutex);
1027 }
1028
1029 /*
1030  * session caps
1031  */
1032
1033 /* caller holds s_cap_lock, we drop it */
1034 static void cleanup_cap_releases(struct ceph_mds_client *mdsc,
1035                                  struct ceph_mds_session *session)
1036         __releases(session->s_cap_lock)
1037 {
1038         LIST_HEAD(tmp_list);
1039         list_splice_init(&session->s_cap_releases, &tmp_list);
1040         session->s_num_cap_releases = 0;
1041         spin_unlock(&session->s_cap_lock);
1042
1043         dout("cleanup_cap_releases mds%d\n", session->s_mds);
1044         while (!list_empty(&tmp_list)) {
1045                 struct ceph_cap *cap;
1046                 /* zero out the in-progress message */
1047                 cap = list_first_entry(&tmp_list,
1048                                         struct ceph_cap, session_caps);
1049                 list_del(&cap->session_caps);
1050                 ceph_put_cap(mdsc, cap);
1051         }
1052 }
1053
1054 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1055                                      struct ceph_mds_session *session)
1056 {
1057         struct ceph_mds_request *req;
1058         struct rb_node *p;
1059
1060         dout("cleanup_session_requests mds%d\n", session->s_mds);
1061         mutex_lock(&mdsc->mutex);
1062         while (!list_empty(&session->s_unsafe)) {
1063                 req = list_first_entry(&session->s_unsafe,
1064                                        struct ceph_mds_request, r_unsafe_item);
1065                 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1066                                     req->r_tid);
1067                 __unregister_request(mdsc, req);
1068         }
1069         /* zero r_attempts, so kick_requests() will re-send requests */
1070         p = rb_first(&mdsc->request_tree);
1071         while (p) {
1072                 req = rb_entry(p, struct ceph_mds_request, r_node);
1073                 p = rb_next(p);
1074                 if (req->r_session &&
1075                     req->r_session->s_mds == session->s_mds)
1076                         req->r_attempts = 0;
1077         }
1078         mutex_unlock(&mdsc->mutex);
1079 }
1080
1081 /*
1082  * Helper to safely iterate over all caps associated with a session, with
1083  * special care taken to handle a racing __ceph_remove_cap().
1084  *
1085  * Caller must hold session s_mutex.
1086  */
1087 static int iterate_session_caps(struct ceph_mds_session *session,
1088                                  int (*cb)(struct inode *, struct ceph_cap *,
1089                                             void *), void *arg)
1090 {
1091         struct list_head *p;
1092         struct ceph_cap *cap;
1093         struct inode *inode, *last_inode = NULL;
1094         struct ceph_cap *old_cap = NULL;
1095         int ret;
1096
1097         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1098         spin_lock(&session->s_cap_lock);
1099         p = session->s_caps.next;
1100         while (p != &session->s_caps) {
1101                 cap = list_entry(p, struct ceph_cap, session_caps);
1102                 inode = igrab(&cap->ci->vfs_inode);
1103                 if (!inode) {
1104                         p = p->next;
1105                         continue;
1106                 }
1107                 session->s_cap_iterator = cap;
1108                 spin_unlock(&session->s_cap_lock);
1109
1110                 if (last_inode) {
1111                         iput(last_inode);
1112                         last_inode = NULL;
1113                 }
1114                 if (old_cap) {
1115                         ceph_put_cap(session->s_mdsc, old_cap);
1116                         old_cap = NULL;
1117                 }
1118
1119                 ret = cb(inode, cap, arg);
1120                 last_inode = inode;
1121
1122                 spin_lock(&session->s_cap_lock);
1123                 p = p->next;
1124                 if (cap->ci == NULL) {
1125                         dout("iterate_session_caps  finishing cap %p removal\n",
1126                              cap);
1127                         BUG_ON(cap->session != session);
1128                         cap->session = NULL;
1129                         list_del_init(&cap->session_caps);
1130                         session->s_nr_caps--;
1131                         if (cap->queue_release) {
1132                                 list_add_tail(&cap->session_caps,
1133                                               &session->s_cap_releases);
1134                                 session->s_num_cap_releases++;
1135                         } else {
1136                                 old_cap = cap;  /* put_cap it w/o locks held */
1137                         }
1138                 }
1139                 if (ret < 0)
1140                         goto out;
1141         }
1142         ret = 0;
1143 out:
1144         session->s_cap_iterator = NULL;
1145         spin_unlock(&session->s_cap_lock);
1146
1147         iput(last_inode);
1148         if (old_cap)
1149                 ceph_put_cap(session->s_mdsc, old_cap);
1150
1151         return ret;
1152 }
1153
1154 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1155                                   void *arg)
1156 {
1157         struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1158         struct ceph_inode_info *ci = ceph_inode(inode);
1159         LIST_HEAD(to_remove);
1160         bool drop = false;
1161         bool invalidate = false;
1162
1163         dout("removing cap %p, ci is %p, inode is %p\n",
1164              cap, ci, &ci->vfs_inode);
1165         spin_lock(&ci->i_ceph_lock);
1166         __ceph_remove_cap(cap, false);
1167         if (!ci->i_auth_cap) {
1168                 struct ceph_cap_flush *cf;
1169                 struct ceph_mds_client *mdsc = fsc->mdsc;
1170
1171                 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1172
1173                 if (ci->i_wrbuffer_ref > 0 &&
1174                     READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
1175                         invalidate = true;
1176
1177                 while (!list_empty(&ci->i_cap_flush_list)) {
1178                         cf = list_first_entry(&ci->i_cap_flush_list,
1179                                               struct ceph_cap_flush, i_list);
1180                         list_move(&cf->i_list, &to_remove);
1181                 }
1182
1183                 spin_lock(&mdsc->cap_dirty_lock);
1184
1185                 list_for_each_entry(cf, &to_remove, i_list)
1186                         list_del(&cf->g_list);
1187
1188                 if (!list_empty(&ci->i_dirty_item)) {
1189                         pr_warn_ratelimited(
1190                                 " dropping dirty %s state for %p %lld\n",
1191                                 ceph_cap_string(ci->i_dirty_caps),
1192                                 inode, ceph_ino(inode));
1193                         ci->i_dirty_caps = 0;
1194                         list_del_init(&ci->i_dirty_item);
1195                         drop = true;
1196                 }
1197                 if (!list_empty(&ci->i_flushing_item)) {
1198                         pr_warn_ratelimited(
1199                                 " dropping dirty+flushing %s state for %p %lld\n",
1200                                 ceph_cap_string(ci->i_flushing_caps),
1201                                 inode, ceph_ino(inode));
1202                         ci->i_flushing_caps = 0;
1203                         list_del_init(&ci->i_flushing_item);
1204                         mdsc->num_cap_flushing--;
1205                         drop = true;
1206                 }
1207                 spin_unlock(&mdsc->cap_dirty_lock);
1208
1209                 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1210                         list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1211                         ci->i_prealloc_cap_flush = NULL;
1212                 }
1213         }
1214         spin_unlock(&ci->i_ceph_lock);
1215         while (!list_empty(&to_remove)) {
1216                 struct ceph_cap_flush *cf;
1217                 cf = list_first_entry(&to_remove,
1218                                       struct ceph_cap_flush, i_list);
1219                 list_del(&cf->i_list);
1220                 ceph_free_cap_flush(cf);
1221         }
1222
1223         wake_up_all(&ci->i_cap_wq);
1224         if (invalidate)
1225                 ceph_queue_invalidate(inode);
1226         if (drop)
1227                 iput(inode);
1228         return 0;
1229 }
1230
1231 /*
1232  * caller must hold session s_mutex
1233  */
1234 static void remove_session_caps(struct ceph_mds_session *session)
1235 {
1236         struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1237         struct super_block *sb = fsc->sb;
1238         dout("remove_session_caps on %p\n", session);
1239         iterate_session_caps(session, remove_session_caps_cb, fsc);
1240
1241         wake_up_all(&fsc->mdsc->cap_flushing_wq);
1242
1243         spin_lock(&session->s_cap_lock);
1244         if (session->s_nr_caps > 0) {
1245                 struct inode *inode;
1246                 struct ceph_cap *cap, *prev = NULL;
1247                 struct ceph_vino vino;
1248                 /*
1249                  * iterate_session_caps() skips inodes that are being
1250                  * deleted, we need to wait until deletions are complete.
1251                  * __wait_on_freeing_inode() is designed for the job,
1252                  * but it is not exported, so use lookup inode function
1253                  * to access it.
1254                  */
1255                 while (!list_empty(&session->s_caps)) {
1256                         cap = list_entry(session->s_caps.next,
1257                                          struct ceph_cap, session_caps);
1258                         if (cap == prev)
1259                                 break;
1260                         prev = cap;
1261                         vino = cap->ci->i_vino;
1262                         spin_unlock(&session->s_cap_lock);
1263
1264                         inode = ceph_find_inode(sb, vino);
1265                         iput(inode);
1266
1267                         spin_lock(&session->s_cap_lock);
1268                 }
1269         }
1270
1271         // drop cap expires and unlock s_cap_lock
1272         cleanup_cap_releases(session->s_mdsc, session);
1273
1274         BUG_ON(session->s_nr_caps > 0);
1275         BUG_ON(!list_empty(&session->s_cap_flushing));
1276 }
1277
1278 /*
1279  * wake up any threads waiting on this session's caps.  if the cap is
1280  * old (didn't get renewed on the client reconnect), remove it now.
1281  *
1282  * caller must hold s_mutex.
1283  */
1284 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1285                               void *arg)
1286 {
1287         struct ceph_inode_info *ci = ceph_inode(inode);
1288
1289         if (arg) {
1290                 spin_lock(&ci->i_ceph_lock);
1291                 ci->i_wanted_max_size = 0;
1292                 ci->i_requested_max_size = 0;
1293                 spin_unlock(&ci->i_ceph_lock);
1294         }
1295         wake_up_all(&ci->i_cap_wq);
1296         return 0;
1297 }
1298
1299 static void wake_up_session_caps(struct ceph_mds_session *session,
1300                                  int reconnect)
1301 {
1302         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1303         iterate_session_caps(session, wake_up_session_cb,
1304                              (void *)(unsigned long)reconnect);
1305 }
1306
1307 /*
1308  * Send periodic message to MDS renewing all currently held caps.  The
1309  * ack will reset the expiration for all caps from this session.
1310  *
1311  * caller holds s_mutex
1312  */
1313 static int send_renew_caps(struct ceph_mds_client *mdsc,
1314                            struct ceph_mds_session *session)
1315 {
1316         struct ceph_msg *msg;
1317         int state;
1318
1319         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1320             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1321                 pr_info("mds%d caps stale\n", session->s_mds);
1322         session->s_renew_requested = jiffies;
1323
1324         /* do not try to renew caps until a recovering mds has reconnected
1325          * with its clients. */
1326         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1327         if (state < CEPH_MDS_STATE_RECONNECT) {
1328                 dout("send_renew_caps ignoring mds%d (%s)\n",
1329                      session->s_mds, ceph_mds_state_name(state));
1330                 return 0;
1331         }
1332
1333         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1334                 ceph_mds_state_name(state));
1335         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1336                                  ++session->s_renew_seq);
1337         if (!msg)
1338                 return -ENOMEM;
1339         ceph_con_send(&session->s_con, msg);
1340         return 0;
1341 }
1342
1343 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1344                              struct ceph_mds_session *session, u64 seq)
1345 {
1346         struct ceph_msg *msg;
1347
1348         dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1349              session->s_mds, ceph_session_state_name(session->s_state), seq);
1350         msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1351         if (!msg)
1352                 return -ENOMEM;
1353         ceph_con_send(&session->s_con, msg);
1354         return 0;
1355 }
1356
1357
1358 /*
1359  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1360  *
1361  * Called under session->s_mutex
1362  */
1363 static void renewed_caps(struct ceph_mds_client *mdsc,
1364                          struct ceph_mds_session *session, int is_renew)
1365 {
1366         int was_stale;
1367         int wake = 0;
1368
1369         spin_lock(&session->s_cap_lock);
1370         was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1371
1372         session->s_cap_ttl = session->s_renew_requested +
1373                 mdsc->mdsmap->m_session_timeout*HZ;
1374
1375         if (was_stale) {
1376                 if (time_before(jiffies, session->s_cap_ttl)) {
1377                         pr_info("mds%d caps renewed\n", session->s_mds);
1378                         wake = 1;
1379                 } else {
1380                         pr_info("mds%d caps still stale\n", session->s_mds);
1381                 }
1382         }
1383         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1384              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1385              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1386         spin_unlock(&session->s_cap_lock);
1387
1388         if (wake)
1389                 wake_up_session_caps(session, 0);
1390 }
1391
1392 /*
1393  * send a session close request
1394  */
1395 static int request_close_session(struct ceph_mds_client *mdsc,
1396                                  struct ceph_mds_session *session)
1397 {
1398         struct ceph_msg *msg;
1399
1400         dout("request_close_session mds%d state %s seq %lld\n",
1401              session->s_mds, ceph_session_state_name(session->s_state),
1402              session->s_seq);
1403         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1404         if (!msg)
1405                 return -ENOMEM;
1406         ceph_con_send(&session->s_con, msg);
1407         return 1;
1408 }
1409
1410 /*
1411  * Called with s_mutex held.
1412  */
1413 static int __close_session(struct ceph_mds_client *mdsc,
1414                          struct ceph_mds_session *session)
1415 {
1416         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1417                 return 0;
1418         session->s_state = CEPH_MDS_SESSION_CLOSING;
1419         return request_close_session(mdsc, session);
1420 }
1421
1422 /*
1423  * Trim old(er) caps.
1424  *
1425  * Because we can't cache an inode without one or more caps, we do
1426  * this indirectly: if a cap is unused, we prune its aliases, at which
1427  * point the inode will hopefully get dropped to.
1428  *
1429  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1430  * memory pressure from the MDS, though, so it needn't be perfect.
1431  */
1432 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1433 {
1434         struct ceph_mds_session *session = arg;
1435         struct ceph_inode_info *ci = ceph_inode(inode);
1436         int used, wanted, oissued, mine;
1437
1438         if (session->s_trim_caps <= 0)
1439                 return -1;
1440
1441         spin_lock(&ci->i_ceph_lock);
1442         mine = cap->issued | cap->implemented;
1443         used = __ceph_caps_used(ci);
1444         wanted = __ceph_caps_file_wanted(ci);
1445         oissued = __ceph_caps_issued_other(ci, cap);
1446
1447         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1448              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1449              ceph_cap_string(used), ceph_cap_string(wanted));
1450         if (cap == ci->i_auth_cap) {
1451                 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1452                     !list_empty(&ci->i_cap_snaps))
1453                         goto out;
1454                 if ((used | wanted) & CEPH_CAP_ANY_WR)
1455                         goto out;
1456         }
1457         /* The inode has cached pages, but it's no longer used.
1458          * we can safely drop it */
1459         if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1460             !(oissued & CEPH_CAP_FILE_CACHE)) {
1461           used = 0;
1462           oissued = 0;
1463         }
1464         if ((used | wanted) & ~oissued & mine)
1465                 goto out;   /* we need these caps */
1466
1467         session->s_trim_caps--;
1468         if (oissued) {
1469                 /* we aren't the only cap.. just remove us */
1470                 __ceph_remove_cap(cap, true);
1471         } else {
1472                 /* try dropping referring dentries */
1473                 spin_unlock(&ci->i_ceph_lock);
1474                 d_prune_aliases(inode);
1475                 dout("trim_caps_cb %p cap %p  pruned, count now %d\n",
1476                      inode, cap, atomic_read(&inode->i_count));
1477                 return 0;
1478         }
1479
1480 out:
1481         spin_unlock(&ci->i_ceph_lock);
1482         return 0;
1483 }
1484
1485 /*
1486  * Trim session cap count down to some max number.
1487  */
1488 static int trim_caps(struct ceph_mds_client *mdsc,
1489                      struct ceph_mds_session *session,
1490                      int max_caps)
1491 {
1492         int trim_caps = session->s_nr_caps - max_caps;
1493
1494         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1495              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1496         if (trim_caps > 0) {
1497                 session->s_trim_caps = trim_caps;
1498                 iterate_session_caps(session, trim_caps_cb, session);
1499                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1500                      session->s_mds, session->s_nr_caps, max_caps,
1501                         trim_caps - session->s_trim_caps);
1502                 session->s_trim_caps = 0;
1503         }
1504
1505         ceph_send_cap_releases(mdsc, session);
1506         return 0;
1507 }
1508
1509 static int check_caps_flush(struct ceph_mds_client *mdsc,
1510                             u64 want_flush_tid)
1511 {
1512         int ret = 1;
1513
1514         spin_lock(&mdsc->cap_dirty_lock);
1515         if (!list_empty(&mdsc->cap_flush_list)) {
1516                 struct ceph_cap_flush *cf =
1517                         list_first_entry(&mdsc->cap_flush_list,
1518                                          struct ceph_cap_flush, g_list);
1519                 if (cf->tid <= want_flush_tid) {
1520                         dout("check_caps_flush still flushing tid "
1521                              "%llu <= %llu\n", cf->tid, want_flush_tid);
1522                         ret = 0;
1523                 }
1524         }
1525         spin_unlock(&mdsc->cap_dirty_lock);
1526         return ret;
1527 }
1528
1529 /*
1530  * flush all dirty inode data to disk.
1531  *
1532  * returns true if we've flushed through want_flush_tid
1533  */
1534 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1535                             u64 want_flush_tid)
1536 {
1537         dout("check_caps_flush want %llu\n", want_flush_tid);
1538
1539         wait_event(mdsc->cap_flushing_wq,
1540                    check_caps_flush(mdsc, want_flush_tid));
1541
1542         dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1543 }
1544
1545 /*
1546  * called under s_mutex
1547  */
1548 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1549                             struct ceph_mds_session *session)
1550 {
1551         struct ceph_msg *msg = NULL;
1552         struct ceph_mds_cap_release *head;
1553         struct ceph_mds_cap_item *item;
1554         struct ceph_cap *cap;
1555         LIST_HEAD(tmp_list);
1556         int num_cap_releases;
1557
1558         spin_lock(&session->s_cap_lock);
1559 again:
1560         list_splice_init(&session->s_cap_releases, &tmp_list);
1561         num_cap_releases = session->s_num_cap_releases;
1562         session->s_num_cap_releases = 0;
1563         spin_unlock(&session->s_cap_lock);
1564
1565         while (!list_empty(&tmp_list)) {
1566                 if (!msg) {
1567                         msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1568                                         PAGE_SIZE, GFP_NOFS, false);
1569                         if (!msg)
1570                                 goto out_err;
1571                         head = msg->front.iov_base;
1572                         head->num = cpu_to_le32(0);
1573                         msg->front.iov_len = sizeof(*head);
1574                 }
1575                 cap = list_first_entry(&tmp_list, struct ceph_cap,
1576                                         session_caps);
1577                 list_del(&cap->session_caps);
1578                 num_cap_releases--;
1579
1580                 head = msg->front.iov_base;
1581                 le32_add_cpu(&head->num, 1);
1582                 item = msg->front.iov_base + msg->front.iov_len;
1583                 item->ino = cpu_to_le64(cap->cap_ino);
1584                 item->cap_id = cpu_to_le64(cap->cap_id);
1585                 item->migrate_seq = cpu_to_le32(cap->mseq);
1586                 item->seq = cpu_to_le32(cap->issue_seq);
1587                 msg->front.iov_len += sizeof(*item);
1588
1589                 ceph_put_cap(mdsc, cap);
1590
1591                 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1592                         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1593                         dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1594                         ceph_con_send(&session->s_con, msg);
1595                         msg = NULL;
1596                 }
1597         }
1598
1599         BUG_ON(num_cap_releases != 0);
1600
1601         spin_lock(&session->s_cap_lock);
1602         if (!list_empty(&session->s_cap_releases))
1603                 goto again;
1604         spin_unlock(&session->s_cap_lock);
1605
1606         if (msg) {
1607                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1608                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1609                 ceph_con_send(&session->s_con, msg);
1610         }
1611         return;
1612 out_err:
1613         pr_err("send_cap_releases mds%d, failed to allocate message\n",
1614                 session->s_mds);
1615         spin_lock(&session->s_cap_lock);
1616         list_splice(&tmp_list, &session->s_cap_releases);
1617         session->s_num_cap_releases += num_cap_releases;
1618         spin_unlock(&session->s_cap_lock);
1619 }
1620
1621 /*
1622  * requests
1623  */
1624
1625 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1626                                     struct inode *dir)
1627 {
1628         struct ceph_inode_info *ci = ceph_inode(dir);
1629         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1630         struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1631         size_t size = sizeof(struct ceph_mds_reply_dir_entry);
1632         int order, num_entries;
1633
1634         spin_lock(&ci->i_ceph_lock);
1635         num_entries = ci->i_files + ci->i_subdirs;
1636         spin_unlock(&ci->i_ceph_lock);
1637         num_entries = max(num_entries, 1);
1638         num_entries = min(num_entries, opt->max_readdir);
1639
1640         order = get_order(size * num_entries);
1641         while (order >= 0) {
1642                 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
1643                                                              __GFP_NOWARN,
1644                                                              order);
1645                 if (rinfo->dir_entries)
1646                         break;
1647                 order--;
1648         }
1649         if (!rinfo->dir_entries)
1650                 return -ENOMEM;
1651
1652         num_entries = (PAGE_SIZE << order) / size;
1653         num_entries = min(num_entries, opt->max_readdir);
1654
1655         rinfo->dir_buf_size = PAGE_SIZE << order;
1656         req->r_num_caps = num_entries + 1;
1657         req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1658         req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1659         return 0;
1660 }
1661
1662 /*
1663  * Create an mds request.
1664  */
1665 struct ceph_mds_request *
1666 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1667 {
1668         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1669
1670         if (!req)
1671                 return ERR_PTR(-ENOMEM);
1672
1673         mutex_init(&req->r_fill_mutex);
1674         req->r_mdsc = mdsc;
1675         req->r_started = jiffies;
1676         req->r_resend_mds = -1;
1677         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1678         INIT_LIST_HEAD(&req->r_unsafe_target_item);
1679         req->r_fmode = -1;
1680         kref_init(&req->r_kref);
1681         RB_CLEAR_NODE(&req->r_node);
1682         INIT_LIST_HEAD(&req->r_wait);
1683         init_completion(&req->r_completion);
1684         init_completion(&req->r_safe_completion);
1685         INIT_LIST_HEAD(&req->r_unsafe_item);
1686
1687         req->r_stamp = current_fs_time(mdsc->fsc->sb);
1688
1689         req->r_op = op;
1690         req->r_direct_mode = mode;
1691         return req;
1692 }
1693
1694 /*
1695  * return oldest (lowest) request, tid in request tree, 0 if none.
1696  *
1697  * called under mdsc->mutex.
1698  */
1699 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1700 {
1701         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1702                 return NULL;
1703         return rb_entry(rb_first(&mdsc->request_tree),
1704                         struct ceph_mds_request, r_node);
1705 }
1706
1707 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1708 {
1709         return mdsc->oldest_tid;
1710 }
1711
1712 /*
1713  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1714  * on build_path_from_dentry in fs/cifs/dir.c.
1715  *
1716  * If @stop_on_nosnap, generate path relative to the first non-snapped
1717  * inode.
1718  *
1719  * Encode hidden .snap dirs as a double /, i.e.
1720  *   foo/.snap/bar -> foo//bar
1721  */
1722 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1723                            int stop_on_nosnap)
1724 {
1725         struct dentry *temp;
1726         char *path;
1727         int len, pos;
1728         unsigned seq;
1729
1730         if (dentry == NULL)
1731                 return ERR_PTR(-EINVAL);
1732
1733 retry:
1734         len = 0;
1735         seq = read_seqbegin(&rename_lock);
1736         rcu_read_lock();
1737         for (temp = dentry; !IS_ROOT(temp);) {
1738                 struct inode *inode = d_inode(temp);
1739                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1740                         len++;  /* slash only */
1741                 else if (stop_on_nosnap && inode &&
1742                          ceph_snap(inode) == CEPH_NOSNAP)
1743                         break;
1744                 else
1745                         len += 1 + temp->d_name.len;
1746                 temp = temp->d_parent;
1747         }
1748         rcu_read_unlock();
1749         if (len)
1750                 len--;  /* no leading '/' */
1751
1752         path = kmalloc(len+1, GFP_NOFS);
1753         if (path == NULL)
1754                 return ERR_PTR(-ENOMEM);
1755         pos = len;
1756         path[pos] = 0;  /* trailing null */
1757         rcu_read_lock();
1758         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1759                 struct inode *inode;
1760
1761                 spin_lock(&temp->d_lock);
1762                 inode = d_inode(temp);
1763                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1764                         dout("build_path path+%d: %p SNAPDIR\n",
1765                              pos, temp);
1766                 } else if (stop_on_nosnap && inode &&
1767                            ceph_snap(inode) == CEPH_NOSNAP) {
1768                         spin_unlock(&temp->d_lock);
1769                         break;
1770                 } else {
1771                         pos -= temp->d_name.len;
1772                         if (pos < 0) {
1773                                 spin_unlock(&temp->d_lock);
1774                                 break;
1775                         }
1776                         strncpy(path + pos, temp->d_name.name,
1777                                 temp->d_name.len);
1778                 }
1779                 spin_unlock(&temp->d_lock);
1780                 if (pos)
1781                         path[--pos] = '/';
1782                 temp = temp->d_parent;
1783         }
1784         rcu_read_unlock();
1785         if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1786                 pr_err("build_path did not end path lookup where "
1787                        "expected, namelen is %d, pos is %d\n", len, pos);
1788                 /* presumably this is only possible if racing with a
1789                    rename of one of the parent directories (we can not
1790                    lock the dentries above us to prevent this, but
1791                    retrying should be harmless) */
1792                 kfree(path);
1793                 goto retry;
1794         }
1795
1796         *base = ceph_ino(d_inode(temp));
1797         *plen = len;
1798         dout("build_path on %p %d built %llx '%.*s'\n",
1799              dentry, d_count(dentry), *base, len, path);
1800         return path;
1801 }
1802
1803 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
1804                              const char **ppath, int *ppathlen, u64 *pino,
1805                              int *pfreepath)
1806 {
1807         char *path;
1808
1809         rcu_read_lock();
1810         if (!dir)
1811                 dir = d_inode_rcu(dentry->d_parent);
1812         if (dir && ceph_snap(dir) == CEPH_NOSNAP) {
1813                 *pino = ceph_ino(dir);
1814                 rcu_read_unlock();
1815                 *ppath = dentry->d_name.name;
1816                 *ppathlen = dentry->d_name.len;
1817                 return 0;
1818         }
1819         rcu_read_unlock();
1820         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1821         if (IS_ERR(path))
1822                 return PTR_ERR(path);
1823         *ppath = path;
1824         *pfreepath = 1;
1825         return 0;
1826 }
1827
1828 static int build_inode_path(struct inode *inode,
1829                             const char **ppath, int *ppathlen, u64 *pino,
1830                             int *pfreepath)
1831 {
1832         struct dentry *dentry;
1833         char *path;
1834
1835         if (ceph_snap(inode) == CEPH_NOSNAP) {
1836                 *pino = ceph_ino(inode);
1837                 *ppathlen = 0;
1838                 return 0;
1839         }
1840         dentry = d_find_alias(inode);
1841         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1842         dput(dentry);
1843         if (IS_ERR(path))
1844                 return PTR_ERR(path);
1845         *ppath = path;
1846         *pfreepath = 1;
1847         return 0;
1848 }
1849
1850 /*
1851  * request arguments may be specified via an inode *, a dentry *, or
1852  * an explicit ino+path.
1853  */
1854 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1855                                   struct inode *rdiri, const char *rpath,
1856                                   u64 rino, const char **ppath, int *pathlen,
1857                                   u64 *ino, int *freepath)
1858 {
1859         int r = 0;
1860
1861         if (rinode) {
1862                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1863                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1864                      ceph_snap(rinode));
1865         } else if (rdentry) {
1866                 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
1867                                         freepath);
1868                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1869                      *ppath);
1870         } else if (rpath || rino) {
1871                 *ino = rino;
1872                 *ppath = rpath;
1873                 *pathlen = rpath ? strlen(rpath) : 0;
1874                 dout(" path %.*s\n", *pathlen, rpath);
1875         }
1876
1877         return r;
1878 }
1879
1880 /*
1881  * called under mdsc->mutex
1882  */
1883 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1884                                                struct ceph_mds_request *req,
1885                                                int mds, bool drop_cap_releases)
1886 {
1887         struct ceph_msg *msg;
1888         struct ceph_mds_request_head *head;
1889         const char *path1 = NULL;
1890         const char *path2 = NULL;
1891         u64 ino1 = 0, ino2 = 0;
1892         int pathlen1 = 0, pathlen2 = 0;
1893         int freepath1 = 0, freepath2 = 0;
1894         int len;
1895         u16 releases;
1896         void *p, *end;
1897         int ret;
1898
1899         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1900                               req->r_parent, req->r_path1, req->r_ino1.ino,
1901                               &path1, &pathlen1, &ino1, &freepath1);
1902         if (ret < 0) {
1903                 msg = ERR_PTR(ret);
1904                 goto out;
1905         }
1906
1907         ret = set_request_path_attr(NULL, req->r_old_dentry,
1908                               req->r_old_dentry_dir,
1909                               req->r_path2, req->r_ino2.ino,
1910                               &path2, &pathlen2, &ino2, &freepath2);
1911         if (ret < 0) {
1912                 msg = ERR_PTR(ret);
1913                 goto out_free1;
1914         }
1915
1916         len = sizeof(*head) +
1917                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
1918                 sizeof(struct ceph_timespec);
1919
1920         /* calculate (max) length for cap releases */
1921         len += sizeof(struct ceph_mds_request_release) *
1922                 (!!req->r_inode_drop + !!req->r_dentry_drop +
1923                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1924         if (req->r_dentry_drop)
1925                 len += req->r_dentry->d_name.len;
1926         if (req->r_old_dentry_drop)
1927                 len += req->r_old_dentry->d_name.len;
1928
1929         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1930         if (!msg) {
1931                 msg = ERR_PTR(-ENOMEM);
1932                 goto out_free2;
1933         }
1934
1935         msg->hdr.version = cpu_to_le16(2);
1936         msg->hdr.tid = cpu_to_le64(req->r_tid);
1937
1938         head = msg->front.iov_base;
1939         p = msg->front.iov_base + sizeof(*head);
1940         end = msg->front.iov_base + msg->front.iov_len;
1941
1942         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1943         head->op = cpu_to_le32(req->r_op);
1944         head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1945         head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1946         head->args = req->r_args;
1947
1948         ceph_encode_filepath(&p, end, ino1, path1);
1949         ceph_encode_filepath(&p, end, ino2, path2);
1950
1951         /* make note of release offset, in case we need to replay */
1952         req->r_request_release_offset = p - msg->front.iov_base;
1953
1954         /* cap releases */
1955         releases = 0;
1956         if (req->r_inode_drop)
1957                 releases += ceph_encode_inode_release(&p,
1958                       req->r_inode ? req->r_inode : d_inode(req->r_dentry),
1959                       mds, req->r_inode_drop, req->r_inode_unless, 0);
1960         if (req->r_dentry_drop)
1961                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1962                                 req->r_parent, mds, req->r_dentry_drop,
1963                                 req->r_dentry_unless);
1964         if (req->r_old_dentry_drop)
1965                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1966                                 req->r_old_dentry_dir, mds,
1967                                 req->r_old_dentry_drop,
1968                                 req->r_old_dentry_unless);
1969         if (req->r_old_inode_drop)
1970                 releases += ceph_encode_inode_release(&p,
1971                       d_inode(req->r_old_dentry),
1972                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1973
1974         if (drop_cap_releases) {
1975                 releases = 0;
1976                 p = msg->front.iov_base + req->r_request_release_offset;
1977         }
1978
1979         head->num_releases = cpu_to_le16(releases);
1980
1981         /* time stamp */
1982         {
1983                 struct ceph_timespec ts;
1984                 ceph_encode_timespec(&ts, &req->r_stamp);
1985                 ceph_encode_copy(&p, &ts, sizeof(ts));
1986         }
1987
1988         BUG_ON(p > end);
1989         msg->front.iov_len = p - msg->front.iov_base;
1990         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1991
1992         if (req->r_pagelist) {
1993                 struct ceph_pagelist *pagelist = req->r_pagelist;
1994                 atomic_inc(&pagelist->refcnt);
1995                 ceph_msg_data_add_pagelist(msg, pagelist);
1996                 msg->hdr.data_len = cpu_to_le32(pagelist->length);
1997         } else {
1998                 msg->hdr.data_len = 0;
1999         }
2000
2001         msg->hdr.data_off = cpu_to_le16(0);
2002
2003 out_free2:
2004         if (freepath2)
2005                 kfree((char *)path2);
2006 out_free1:
2007         if (freepath1)
2008                 kfree((char *)path1);
2009 out:
2010         return msg;
2011 }
2012
2013 /*
2014  * called under mdsc->mutex if error, under no mutex if
2015  * success.
2016  */
2017 static void complete_request(struct ceph_mds_client *mdsc,
2018                              struct ceph_mds_request *req)
2019 {
2020         if (req->r_callback)
2021                 req->r_callback(mdsc, req);
2022         else
2023                 complete_all(&req->r_completion);
2024 }
2025
2026 /*
2027  * called under mdsc->mutex
2028  */
2029 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2030                                   struct ceph_mds_request *req,
2031                                   int mds, bool drop_cap_releases)
2032 {
2033         struct ceph_mds_request_head *rhead;
2034         struct ceph_msg *msg;
2035         int flags = 0;
2036
2037         req->r_attempts++;
2038         if (req->r_inode) {
2039                 struct ceph_cap *cap =
2040                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2041
2042                 if (cap)
2043                         req->r_sent_on_mseq = cap->mseq;
2044                 else
2045                         req->r_sent_on_mseq = -1;
2046         }
2047         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2048              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2049
2050         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2051                 void *p;
2052                 /*
2053                  * Replay.  Do not regenerate message (and rebuild
2054                  * paths, etc.); just use the original message.
2055                  * Rebuilding paths will break for renames because
2056                  * d_move mangles the src name.
2057                  */
2058                 msg = req->r_request;
2059                 rhead = msg->front.iov_base;
2060
2061                 flags = le32_to_cpu(rhead->flags);
2062                 flags |= CEPH_MDS_FLAG_REPLAY;
2063                 rhead->flags = cpu_to_le32(flags);
2064
2065                 if (req->r_target_inode)
2066                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2067
2068                 rhead->num_retry = req->r_attempts - 1;
2069
2070                 /* remove cap/dentry releases from message */
2071                 rhead->num_releases = 0;
2072
2073                 /* time stamp */
2074                 p = msg->front.iov_base + req->r_request_release_offset;
2075                 {
2076                         struct ceph_timespec ts;
2077                         ceph_encode_timespec(&ts, &req->r_stamp);
2078                         ceph_encode_copy(&p, &ts, sizeof(ts));
2079                 }
2080
2081                 msg->front.iov_len = p - msg->front.iov_base;
2082                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2083                 return 0;
2084         }
2085
2086         if (req->r_request) {
2087                 ceph_msg_put(req->r_request);
2088                 req->r_request = NULL;
2089         }
2090         msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2091         if (IS_ERR(msg)) {
2092                 req->r_err = PTR_ERR(msg);
2093                 return PTR_ERR(msg);
2094         }
2095         req->r_request = msg;
2096
2097         rhead = msg->front.iov_base;
2098         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2099         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2100                 flags |= CEPH_MDS_FLAG_REPLAY;
2101         if (req->r_parent)
2102                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2103         rhead->flags = cpu_to_le32(flags);
2104         rhead->num_fwd = req->r_num_fwd;
2105         rhead->num_retry = req->r_attempts - 1;
2106         rhead->ino = 0;
2107
2108         dout(" r_parent = %p\n", req->r_parent);
2109         return 0;
2110 }
2111
2112 /*
2113  * send request, or put it on the appropriate wait list.
2114  */
2115 static int __do_request(struct ceph_mds_client *mdsc,
2116                         struct ceph_mds_request *req)
2117 {
2118         struct ceph_mds_session *session = NULL;
2119         int mds = -1;
2120         int err = 0;
2121
2122         if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2123                 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2124                         __unregister_request(mdsc, req);
2125                 goto out;
2126         }
2127
2128         if (req->r_timeout &&
2129             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2130                 dout("do_request timed out\n");
2131                 err = -EIO;
2132                 goto finish;
2133         }
2134         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2135                 dout("do_request forced umount\n");
2136                 err = -EIO;
2137                 goto finish;
2138         }
2139         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2140                 if (mdsc->mdsmap_err) {
2141                         err = mdsc->mdsmap_err;
2142                         dout("do_request mdsmap err %d\n", err);
2143                         goto finish;
2144                 }
2145                 if (mdsc->mdsmap->m_epoch == 0) {
2146                         dout("do_request no mdsmap, waiting for map\n");
2147                         list_add(&req->r_wait, &mdsc->waiting_for_map);
2148                         goto finish;
2149                 }
2150                 if (!(mdsc->fsc->mount_options->flags &
2151                       CEPH_MOUNT_OPT_MOUNTWAIT) &&
2152                     !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2153                         err = -ENOENT;
2154                         pr_info("probably no mds server is up\n");
2155                         goto finish;
2156                 }
2157         }
2158
2159         put_request_session(req);
2160
2161         mds = __choose_mds(mdsc, req);
2162         if (mds < 0 ||
2163             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2164                 dout("do_request no mds or not active, waiting for map\n");
2165                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2166                 goto out;
2167         }
2168
2169         /* get, open session */
2170         session = __ceph_lookup_mds_session(mdsc, mds);
2171         if (!session) {
2172                 session = register_session(mdsc, mds);
2173                 if (IS_ERR(session)) {
2174                         err = PTR_ERR(session);
2175                         goto finish;
2176                 }
2177         }
2178         req->r_session = get_session(session);
2179
2180         dout("do_request mds%d session %p state %s\n", mds, session,
2181              ceph_session_state_name(session->s_state));
2182         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2183             session->s_state != CEPH_MDS_SESSION_HUNG) {
2184                 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2185                         err = -EACCES;
2186                         goto out_session;
2187                 }
2188                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2189                     session->s_state == CEPH_MDS_SESSION_CLOSING)
2190                         __open_session(mdsc, session);
2191                 list_add(&req->r_wait, &session->s_waiting);
2192                 goto out_session;
2193         }
2194
2195         /* send request */
2196         req->r_resend_mds = -1;   /* forget any previous mds hint */
2197
2198         if (req->r_request_started == 0)   /* note request start time */
2199                 req->r_request_started = jiffies;
2200
2201         err = __prepare_send_request(mdsc, req, mds, false);
2202         if (!err) {
2203                 ceph_msg_get(req->r_request);
2204                 ceph_con_send(&session->s_con, req->r_request);
2205         }
2206
2207 out_session:
2208         ceph_put_mds_session(session);
2209 finish:
2210         if (err) {
2211                 dout("__do_request early error %d\n", err);
2212                 req->r_err = err;
2213                 complete_request(mdsc, req);
2214                 __unregister_request(mdsc, req);
2215         }
2216 out:
2217         return err;
2218 }
2219
2220 /*
2221  * called under mdsc->mutex
2222  */
2223 static void __wake_requests(struct ceph_mds_client *mdsc,
2224                             struct list_head *head)
2225 {
2226         struct ceph_mds_request *req;
2227         LIST_HEAD(tmp_list);
2228
2229         list_splice_init(head, &tmp_list);
2230
2231         while (!list_empty(&tmp_list)) {
2232                 req = list_entry(tmp_list.next,
2233                                  struct ceph_mds_request, r_wait);
2234                 list_del_init(&req->r_wait);
2235                 dout(" wake request %p tid %llu\n", req, req->r_tid);
2236                 __do_request(mdsc, req);
2237         }
2238 }
2239
2240 /*
2241  * Wake up threads with requests pending for @mds, so that they can
2242  * resubmit their requests to a possibly different mds.
2243  */
2244 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2245 {
2246         struct ceph_mds_request *req;
2247         struct rb_node *p = rb_first(&mdsc->request_tree);
2248
2249         dout("kick_requests mds%d\n", mds);
2250         while (p) {
2251                 req = rb_entry(p, struct ceph_mds_request, r_node);
2252                 p = rb_next(p);
2253                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2254                         continue;
2255                 if (req->r_attempts > 0)
2256                         continue; /* only new requests */
2257                 if (req->r_session &&
2258                     req->r_session->s_mds == mds) {
2259                         dout(" kicking tid %llu\n", req->r_tid);
2260                         list_del_init(&req->r_wait);
2261                         __do_request(mdsc, req);
2262                 }
2263         }
2264 }
2265
2266 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2267                               struct ceph_mds_request *req)
2268 {
2269         dout("submit_request on %p\n", req);
2270         mutex_lock(&mdsc->mutex);
2271         __register_request(mdsc, req, NULL);
2272         __do_request(mdsc, req);
2273         mutex_unlock(&mdsc->mutex);
2274 }
2275
2276 /*
2277  * Synchrously perform an mds request.  Take care of all of the
2278  * session setup, forwarding, retry details.
2279  */
2280 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2281                          struct inode *dir,
2282                          struct ceph_mds_request *req)
2283 {
2284         int err;
2285
2286         dout("do_request on %p\n", req);
2287
2288         /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2289         if (req->r_inode)
2290                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2291         if (req->r_parent)
2292                 ceph_get_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
2293         if (req->r_old_dentry_dir)
2294                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2295                                   CEPH_CAP_PIN);
2296
2297         /* issue */
2298         mutex_lock(&mdsc->mutex);
2299         __register_request(mdsc, req, dir);
2300         __do_request(mdsc, req);
2301
2302         if (req->r_err) {
2303                 err = req->r_err;
2304                 goto out;
2305         }
2306
2307         /* wait */
2308         mutex_unlock(&mdsc->mutex);
2309         dout("do_request waiting\n");
2310         if (!req->r_timeout && req->r_wait_for_completion) {
2311                 err = req->r_wait_for_completion(mdsc, req);
2312         } else {
2313                 long timeleft = wait_for_completion_killable_timeout(
2314                                         &req->r_completion,
2315                                         ceph_timeout_jiffies(req->r_timeout));
2316                 if (timeleft > 0)
2317                         err = 0;
2318                 else if (!timeleft)
2319                         err = -EIO;  /* timed out */
2320                 else
2321                         err = timeleft;  /* killed */
2322         }
2323         dout("do_request waited, got %d\n", err);
2324         mutex_lock(&mdsc->mutex);
2325
2326         /* only abort if we didn't race with a real reply */
2327         if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2328                 err = le32_to_cpu(req->r_reply_info.head->result);
2329         } else if (err < 0) {
2330                 dout("aborted request %lld with %d\n", req->r_tid, err);
2331
2332                 /*
2333                  * ensure we aren't running concurrently with
2334                  * ceph_fill_trace or ceph_readdir_prepopulate, which
2335                  * rely on locks (dir mutex) held by our caller.
2336                  */
2337                 mutex_lock(&req->r_fill_mutex);
2338                 req->r_err = err;
2339                 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2340                 mutex_unlock(&req->r_fill_mutex);
2341
2342                 if (req->r_parent &&
2343                     (req->r_op & CEPH_MDS_OP_WRITE))
2344                         ceph_invalidate_dir_request(req);
2345         } else {
2346                 err = req->r_err;
2347         }
2348
2349 out:
2350         mutex_unlock(&mdsc->mutex);
2351         dout("do_request %p done, result %d\n", req, err);
2352         return err;
2353 }
2354
2355 /*
2356  * Invalidate dir's completeness, dentry lease state on an aborted MDS
2357  * namespace request.
2358  */
2359 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2360 {
2361         struct inode *inode = req->r_parent;
2362
2363         dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2364
2365         ceph_dir_clear_complete(inode);
2366         if (req->r_dentry)
2367                 ceph_invalidate_dentry_lease(req->r_dentry);
2368         if (req->r_old_dentry)
2369                 ceph_invalidate_dentry_lease(req->r_old_dentry);
2370 }
2371
2372 /*
2373  * Handle mds reply.
2374  *
2375  * We take the session mutex and parse and process the reply immediately.
2376  * This preserves the logical ordering of replies, capabilities, etc., sent
2377  * by the MDS as they are applied to our local cache.
2378  */
2379 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2380 {
2381         struct ceph_mds_client *mdsc = session->s_mdsc;
2382         struct ceph_mds_request *req;
2383         struct ceph_mds_reply_head *head = msg->front.iov_base;
2384         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2385         struct ceph_snap_realm *realm;
2386         u64 tid;
2387         int err, result;
2388         int mds = session->s_mds;
2389
2390         if (msg->front.iov_len < sizeof(*head)) {
2391                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2392                 ceph_msg_dump(msg);
2393                 return;
2394         }
2395
2396         /* get request, session */
2397         tid = le64_to_cpu(msg->hdr.tid);
2398         mutex_lock(&mdsc->mutex);
2399         req = lookup_get_request(mdsc, tid);
2400         if (!req) {
2401                 dout("handle_reply on unknown tid %llu\n", tid);
2402                 mutex_unlock(&mdsc->mutex);
2403                 return;
2404         }
2405         dout("handle_reply %p\n", req);
2406
2407         /* correct session? */
2408         if (req->r_session != session) {
2409                 pr_err("mdsc_handle_reply got %llu on session mds%d"
2410                        " not mds%d\n", tid, session->s_mds,
2411                        req->r_session ? req->r_session->s_mds : -1);
2412                 mutex_unlock(&mdsc->mutex);
2413                 goto out;
2414         }
2415
2416         /* dup? */
2417         if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
2418             (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
2419                 pr_warn("got a dup %s reply on %llu from mds%d\n",
2420                            head->safe ? "safe" : "unsafe", tid, mds);
2421                 mutex_unlock(&mdsc->mutex);
2422                 goto out;
2423         }
2424         if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
2425                 pr_warn("got unsafe after safe on %llu from mds%d\n",
2426                            tid, mds);
2427                 mutex_unlock(&mdsc->mutex);
2428                 goto out;
2429         }
2430
2431         result = le32_to_cpu(head->result);
2432
2433         /*
2434          * Handle an ESTALE
2435          * if we're not talking to the authority, send to them
2436          * if the authority has changed while we weren't looking,
2437          * send to new authority
2438          * Otherwise we just have to return an ESTALE
2439          */
2440         if (result == -ESTALE) {
2441                 dout("got ESTALE on request %llu", req->r_tid);
2442                 req->r_resend_mds = -1;
2443                 if (req->r_direct_mode != USE_AUTH_MDS) {
2444                         dout("not using auth, setting for that now");
2445                         req->r_direct_mode = USE_AUTH_MDS;
2446                         __do_request(mdsc, req);
2447                         mutex_unlock(&mdsc->mutex);
2448                         goto out;
2449                 } else  {
2450                         int mds = __choose_mds(mdsc, req);
2451                         if (mds >= 0 && mds != req->r_session->s_mds) {
2452                                 dout("but auth changed, so resending");
2453                                 __do_request(mdsc, req);
2454                                 mutex_unlock(&mdsc->mutex);
2455                                 goto out;
2456                         }
2457                 }
2458                 dout("have to return ESTALE on request %llu", req->r_tid);
2459         }
2460
2461
2462         if (head->safe) {
2463                 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
2464                 __unregister_request(mdsc, req);
2465
2466                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2467                         /*
2468                          * We already handled the unsafe response, now do the
2469                          * cleanup.  No need to examine the response; the MDS
2470                          * doesn't include any result info in the safe
2471                          * response.  And even if it did, there is nothing
2472                          * useful we could do with a revised return value.
2473                          */
2474                         dout("got safe reply %llu, mds%d\n", tid, mds);
2475
2476                         /* last unsafe request during umount? */
2477                         if (mdsc->stopping && !__get_oldest_req(mdsc))
2478                                 complete_all(&mdsc->safe_umount_waiters);
2479                         mutex_unlock(&mdsc->mutex);
2480                         goto out;
2481                 }
2482         } else {
2483                 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
2484                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2485                 if (req->r_unsafe_dir) {
2486                         struct ceph_inode_info *ci =
2487                                         ceph_inode(req->r_unsafe_dir);
2488                         spin_lock(&ci->i_unsafe_lock);
2489                         list_add_tail(&req->r_unsafe_dir_item,
2490                                       &ci->i_unsafe_dirops);
2491                         spin_unlock(&ci->i_unsafe_lock);
2492                 }
2493         }
2494
2495         dout("handle_reply tid %lld result %d\n", tid, result);
2496         rinfo = &req->r_reply_info;
2497         err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2498         mutex_unlock(&mdsc->mutex);
2499
2500         mutex_lock(&session->s_mutex);
2501         if (err < 0) {
2502                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2503                 ceph_msg_dump(msg);
2504                 goto out_err;
2505         }
2506
2507         /* snap trace */
2508         realm = NULL;
2509         if (rinfo->snapblob_len) {
2510                 down_write(&mdsc->snap_rwsem);
2511                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2512                                 rinfo->snapblob + rinfo->snapblob_len,
2513                                 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2514                                 &realm);
2515                 downgrade_write(&mdsc->snap_rwsem);
2516         } else {
2517                 down_read(&mdsc->snap_rwsem);
2518         }
2519
2520         /* insert trace into our cache */
2521         mutex_lock(&req->r_fill_mutex);
2522         current->journal_info = req;
2523         err = ceph_fill_trace(mdsc->fsc->sb, req);
2524         if (err == 0) {
2525                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2526                                     req->r_op == CEPH_MDS_OP_LSSNAP))
2527                         ceph_readdir_prepopulate(req, req->r_session);
2528                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2529         }
2530         current->journal_info = NULL;
2531         mutex_unlock(&req->r_fill_mutex);
2532
2533         up_read(&mdsc->snap_rwsem);
2534         if (realm)
2535                 ceph_put_snap_realm(mdsc, realm);
2536
2537         if (err == 0 && req->r_target_inode &&
2538             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2539                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
2540                 spin_lock(&ci->i_unsafe_lock);
2541                 list_add_tail(&req->r_unsafe_target_item, &ci->i_unsafe_iops);
2542                 spin_unlock(&ci->i_unsafe_lock);
2543         }
2544 out_err:
2545         mutex_lock(&mdsc->mutex);
2546         if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2547                 if (err) {
2548                         req->r_err = err;
2549                 } else {
2550                         req->r_reply =  ceph_msg_get(msg);
2551                         set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
2552                 }
2553         } else {
2554                 dout("reply arrived after request %lld was aborted\n", tid);
2555         }
2556         mutex_unlock(&mdsc->mutex);
2557
2558         mutex_unlock(&session->s_mutex);
2559
2560         /* kick calling process */
2561         complete_request(mdsc, req);
2562 out:
2563         ceph_mdsc_put_request(req);
2564         return;
2565 }
2566
2567
2568
2569 /*
2570  * handle mds notification that our request has been forwarded.
2571  */
2572 static void handle_forward(struct ceph_mds_client *mdsc,
2573                            struct ceph_mds_session *session,
2574                            struct ceph_msg *msg)
2575 {
2576         struct ceph_mds_request *req;
2577         u64 tid = le64_to_cpu(msg->hdr.tid);
2578         u32 next_mds;
2579         u32 fwd_seq;
2580         int err = -EINVAL;
2581         void *p = msg->front.iov_base;
2582         void *end = p + msg->front.iov_len;
2583
2584         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2585         next_mds = ceph_decode_32(&p);
2586         fwd_seq = ceph_decode_32(&p);
2587
2588         mutex_lock(&mdsc->mutex);
2589         req = lookup_get_request(mdsc, tid);
2590         if (!req) {
2591                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2592                 goto out;  /* dup reply? */
2593         }
2594
2595         if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2596                 dout("forward tid %llu aborted, unregistering\n", tid);
2597                 __unregister_request(mdsc, req);
2598         } else if (fwd_seq <= req->r_num_fwd) {
2599                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2600                      tid, next_mds, req->r_num_fwd, fwd_seq);
2601         } else {
2602                 /* resend. forward race not possible; mds would drop */
2603                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2604                 BUG_ON(req->r_err);
2605                 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
2606                 req->r_attempts = 0;
2607                 req->r_num_fwd = fwd_seq;
2608                 req->r_resend_mds = next_mds;
2609                 put_request_session(req);
2610                 __do_request(mdsc, req);
2611         }
2612         ceph_mdsc_put_request(req);
2613 out:
2614         mutex_unlock(&mdsc->mutex);
2615         return;
2616
2617 bad:
2618         pr_err("mdsc_handle_forward decode error err=%d\n", err);
2619 }
2620
2621 /*
2622  * handle a mds session control message
2623  */
2624 static void handle_session(struct ceph_mds_session *session,
2625                            struct ceph_msg *msg)
2626 {
2627         struct ceph_mds_client *mdsc = session->s_mdsc;
2628         u32 op;
2629         u64 seq;
2630         int mds = session->s_mds;
2631         struct ceph_mds_session_head *h = msg->front.iov_base;
2632         int wake = 0;
2633
2634         /* decode */
2635         if (msg->front.iov_len != sizeof(*h))
2636                 goto bad;
2637         op = le32_to_cpu(h->op);
2638         seq = le64_to_cpu(h->seq);
2639
2640         mutex_lock(&mdsc->mutex);
2641         if (op == CEPH_SESSION_CLOSE)
2642                 __unregister_session(mdsc, session);
2643         /* FIXME: this ttl calculation is generous */
2644         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2645         mutex_unlock(&mdsc->mutex);
2646
2647         mutex_lock(&session->s_mutex);
2648
2649         dout("handle_session mds%d %s %p state %s seq %llu\n",
2650              mds, ceph_session_op_name(op), session,
2651              ceph_session_state_name(session->s_state), seq);
2652
2653         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2654                 session->s_state = CEPH_MDS_SESSION_OPEN;
2655                 pr_info("mds%d came back\n", session->s_mds);
2656         }
2657
2658         switch (op) {
2659         case CEPH_SESSION_OPEN:
2660                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2661                         pr_info("mds%d reconnect success\n", session->s_mds);
2662                 session->s_state = CEPH_MDS_SESSION_OPEN;
2663                 renewed_caps(mdsc, session, 0);
2664                 wake = 1;
2665                 if (mdsc->stopping)
2666                         __close_session(mdsc, session);
2667                 break;
2668
2669         case CEPH_SESSION_RENEWCAPS:
2670                 if (session->s_renew_seq == seq)
2671                         renewed_caps(mdsc, session, 1);
2672                 break;
2673
2674         case CEPH_SESSION_CLOSE:
2675                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2676                         pr_info("mds%d reconnect denied\n", session->s_mds);
2677                 cleanup_session_requests(mdsc, session);
2678                 remove_session_caps(session);
2679                 wake = 2; /* for good measure */
2680                 wake_up_all(&mdsc->session_close_wq);
2681                 break;
2682
2683         case CEPH_SESSION_STALE:
2684                 pr_info("mds%d caps went stale, renewing\n",
2685                         session->s_mds);
2686                 spin_lock(&session->s_gen_ttl_lock);
2687                 session->s_cap_gen++;
2688                 session->s_cap_ttl = jiffies - 1;
2689                 spin_unlock(&session->s_gen_ttl_lock);
2690                 send_renew_caps(mdsc, session);
2691                 break;
2692
2693         case CEPH_SESSION_RECALL_STATE:
2694                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2695                 break;
2696
2697         case CEPH_SESSION_FLUSHMSG:
2698                 send_flushmsg_ack(mdsc, session, seq);
2699                 break;
2700
2701         case CEPH_SESSION_FORCE_RO:
2702                 dout("force_session_readonly %p\n", session);
2703                 spin_lock(&session->s_cap_lock);
2704                 session->s_readonly = true;
2705                 spin_unlock(&session->s_cap_lock);
2706                 wake_up_session_caps(session, 0);
2707                 break;
2708
2709         case CEPH_SESSION_REJECT:
2710                 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
2711                 pr_info("mds%d rejected session\n", session->s_mds);
2712                 session->s_state = CEPH_MDS_SESSION_REJECTED;
2713                 cleanup_session_requests(mdsc, session);
2714                 remove_session_caps(session);
2715                 wake = 2; /* for good measure */
2716                 break;
2717
2718         default:
2719                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2720                 WARN_ON(1);
2721         }
2722
2723         mutex_unlock(&session->s_mutex);
2724         if (wake) {
2725                 mutex_lock(&mdsc->mutex);
2726                 __wake_requests(mdsc, &session->s_waiting);
2727                 if (wake == 2)
2728                         kick_requests(mdsc, mds);
2729                 mutex_unlock(&mdsc->mutex);
2730         }
2731         return;
2732
2733 bad:
2734         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2735                (int)msg->front.iov_len);
2736         ceph_msg_dump(msg);
2737         return;
2738 }
2739
2740
2741 /*
2742  * called under session->mutex.
2743  */
2744 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2745                                    struct ceph_mds_session *session)
2746 {
2747         struct ceph_mds_request *req, *nreq;
2748         struct rb_node *p;
2749         int err;
2750
2751         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2752
2753         mutex_lock(&mdsc->mutex);
2754         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2755                 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2756                 if (!err) {
2757                         ceph_msg_get(req->r_request);
2758                         ceph_con_send(&session->s_con, req->r_request);
2759                 }
2760         }
2761
2762         /*
2763          * also re-send old requests when MDS enters reconnect stage. So that MDS
2764          * can process completed request in clientreplay stage.
2765          */
2766         p = rb_first(&mdsc->request_tree);
2767         while (p) {
2768                 req = rb_entry(p, struct ceph_mds_request, r_node);
2769                 p = rb_next(p);
2770                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2771                         continue;
2772                 if (req->r_attempts == 0)
2773                         continue; /* only old requests */
2774                 if (req->r_session &&
2775                     req->r_session->s_mds == session->s_mds) {
2776                         err = __prepare_send_request(mdsc, req,
2777                                                      session->s_mds, true);
2778                         if (!err) {
2779                                 ceph_msg_get(req->r_request);
2780                                 ceph_con_send(&session->s_con, req->r_request);
2781                         }
2782                 }
2783         }
2784         mutex_unlock(&mdsc->mutex);
2785 }
2786
2787 /*
2788  * Encode information about a cap for a reconnect with the MDS.
2789  */
2790 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2791                           void *arg)
2792 {
2793         union {
2794                 struct ceph_mds_cap_reconnect v2;
2795                 struct ceph_mds_cap_reconnect_v1 v1;
2796         } rec;
2797         struct ceph_inode_info *ci;
2798         struct ceph_reconnect_state *recon_state = arg;
2799         struct ceph_pagelist *pagelist = recon_state->pagelist;
2800         char *path;
2801         int pathlen, err;
2802         u64 pathbase;
2803         u64 snap_follows;
2804         struct dentry *dentry;
2805
2806         ci = cap->ci;
2807
2808         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2809              inode, ceph_vinop(inode), cap, cap->cap_id,
2810              ceph_cap_string(cap->issued));
2811         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2812         if (err)
2813                 return err;
2814
2815         dentry = d_find_alias(inode);
2816         if (dentry) {
2817                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2818                 if (IS_ERR(path)) {
2819                         err = PTR_ERR(path);
2820                         goto out_dput;
2821                 }
2822         } else {
2823                 path = NULL;
2824                 pathlen = 0;
2825                 pathbase = 0;
2826         }
2827
2828         spin_lock(&ci->i_ceph_lock);
2829         cap->seq = 0;        /* reset cap seq */
2830         cap->issue_seq = 0;  /* and issue_seq */
2831         cap->mseq = 0;       /* and migrate_seq */
2832         cap->cap_gen = cap->session->s_cap_gen;
2833
2834         if (recon_state->msg_version >= 2) {
2835                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2836                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2837                 rec.v2.issued = cpu_to_le32(cap->issued);
2838                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2839                 rec.v2.pathbase = cpu_to_le64(pathbase);
2840                 rec.v2.flock_len = 0;
2841         } else {
2842                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2843                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2844                 rec.v1.issued = cpu_to_le32(cap->issued);
2845                 rec.v1.size = cpu_to_le64(inode->i_size);
2846                 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2847                 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2848                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2849                 rec.v1.pathbase = cpu_to_le64(pathbase);
2850         }
2851
2852         if (list_empty(&ci->i_cap_snaps)) {
2853                 snap_follows = 0;
2854         } else {
2855                 struct ceph_cap_snap *capsnap =
2856                         list_first_entry(&ci->i_cap_snaps,
2857                                          struct ceph_cap_snap, ci_item);
2858                 snap_follows = capsnap->follows;
2859         }
2860         spin_unlock(&ci->i_ceph_lock);
2861
2862         if (recon_state->msg_version >= 2) {
2863                 int num_fcntl_locks, num_flock_locks;
2864                 struct ceph_filelock *flocks;
2865                 size_t struct_len, total_len = 0;
2866                 u8 struct_v = 0;
2867
2868 encode_again:
2869                 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2870                 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2871                                  sizeof(struct ceph_filelock), GFP_NOFS);
2872                 if (!flocks) {
2873                         err = -ENOMEM;
2874                         goto out_free;
2875                 }
2876                 err = ceph_encode_locks_to_buffer(inode, flocks,
2877                                                   num_fcntl_locks,
2878                                                   num_flock_locks);
2879                 if (err) {
2880                         kfree(flocks);
2881                         if (err == -ENOSPC)
2882                                 goto encode_again;
2883                         goto out_free;
2884                 }
2885
2886                 if (recon_state->msg_version >= 3) {
2887                         /* version, compat_version and struct_len */
2888                         total_len = 2 * sizeof(u8) + sizeof(u32);
2889                         struct_v = 2;
2890                 }
2891                 /*
2892                  * number of encoded locks is stable, so copy to pagelist
2893                  */
2894                 struct_len = 2 * sizeof(u32) +
2895                             (num_fcntl_locks + num_flock_locks) *
2896                             sizeof(struct ceph_filelock);
2897                 rec.v2.flock_len = cpu_to_le32(struct_len);
2898
2899                 struct_len += sizeof(rec.v2);
2900                 struct_len += sizeof(u32) + pathlen;
2901
2902                 if (struct_v >= 2)
2903                         struct_len += sizeof(u64); /* snap_follows */
2904
2905                 total_len += struct_len;
2906                 err = ceph_pagelist_reserve(pagelist, total_len);
2907
2908                 if (!err) {
2909                         if (recon_state->msg_version >= 3) {
2910                                 ceph_pagelist_encode_8(pagelist, struct_v);
2911                                 ceph_pagelist_encode_8(pagelist, 1);
2912                                 ceph_pagelist_encode_32(pagelist, struct_len);
2913                         }
2914                         ceph_pagelist_encode_string(pagelist, path, pathlen);
2915                         ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
2916                         ceph_locks_to_pagelist(flocks, pagelist,
2917                                                num_fcntl_locks,
2918                                                num_flock_locks);
2919                         if (struct_v >= 2)
2920                                 ceph_pagelist_encode_64(pagelist, snap_follows);
2921                 }
2922                 kfree(flocks);
2923         } else {
2924                 size_t size = sizeof(u32) + pathlen + sizeof(rec.v1);
2925                 err = ceph_pagelist_reserve(pagelist, size);
2926                 if (!err) {
2927                         ceph_pagelist_encode_string(pagelist, path, pathlen);
2928                         ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
2929                 }
2930         }
2931
2932         recon_state->nr_caps++;
2933 out_free:
2934         kfree(path);
2935 out_dput:
2936         dput(dentry);
2937         return err;
2938 }
2939
2940
2941 /*
2942  * If an MDS fails and recovers, clients need to reconnect in order to
2943  * reestablish shared state.  This includes all caps issued through
2944  * this session _and_ the snap_realm hierarchy.  Because it's not
2945  * clear which snap realms the mds cares about, we send everything we
2946  * know about.. that ensures we'll then get any new info the
2947  * recovering MDS might have.
2948  *
2949  * This is a relatively heavyweight operation, but it's rare.
2950  *
2951  * called with mdsc->mutex held.
2952  */
2953 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2954                                struct ceph_mds_session *session)
2955 {
2956         struct ceph_msg *reply;
2957         struct rb_node *p;
2958         int mds = session->s_mds;
2959         int err = -ENOMEM;
2960         int s_nr_caps;
2961         struct ceph_pagelist *pagelist;
2962         struct ceph_reconnect_state recon_state;
2963
2964         pr_info("mds%d reconnect start\n", mds);
2965
2966         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2967         if (!pagelist)
2968                 goto fail_nopagelist;
2969         ceph_pagelist_init(pagelist);
2970
2971         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2972         if (!reply)
2973                 goto fail_nomsg;
2974
2975         mutex_lock(&session->s_mutex);
2976         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2977         session->s_seq = 0;
2978
2979         dout("session %p state %s\n", session,
2980              ceph_session_state_name(session->s_state));
2981
2982         spin_lock(&session->s_gen_ttl_lock);
2983         session->s_cap_gen++;
2984         spin_unlock(&session->s_gen_ttl_lock);
2985
2986         spin_lock(&session->s_cap_lock);
2987         /* don't know if session is readonly */
2988         session->s_readonly = 0;
2989         /*
2990          * notify __ceph_remove_cap() that we are composing cap reconnect.
2991          * If a cap get released before being added to the cap reconnect,
2992          * __ceph_remove_cap() should skip queuing cap release.
2993          */
2994         session->s_cap_reconnect = 1;
2995         /* drop old cap expires; we're about to reestablish that state */
2996         cleanup_cap_releases(mdsc, session);
2997
2998         /* trim unused caps to reduce MDS's cache rejoin time */
2999         if (mdsc->fsc->sb->s_root)
3000                 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3001
3002         ceph_con_close(&session->s_con);
3003         ceph_con_open(&session->s_con,
3004                       CEPH_ENTITY_TYPE_MDS, mds,
3005                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3006
3007         /* replay unsafe requests */
3008         replay_unsafe_requests(mdsc, session);
3009
3010         down_read(&mdsc->snap_rwsem);
3011
3012         /* traverse this session's caps */
3013         s_nr_caps = session->s_nr_caps;
3014         err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
3015         if (err)
3016                 goto fail;
3017
3018         recon_state.nr_caps = 0;
3019         recon_state.pagelist = pagelist;
3020         if (session->s_con.peer_features & CEPH_FEATURE_MDSENC)
3021                 recon_state.msg_version = 3;
3022         else if (session->s_con.peer_features & CEPH_FEATURE_FLOCK)
3023                 recon_state.msg_version = 2;
3024         else
3025                 recon_state.msg_version = 1;
3026         err = iterate_session_caps(session, encode_caps_cb, &recon_state);
3027         if (err < 0)
3028                 goto fail;
3029
3030         spin_lock(&session->s_cap_lock);
3031         session->s_cap_reconnect = 0;
3032         spin_unlock(&session->s_cap_lock);
3033
3034         /*
3035          * snaprealms.  we provide mds with the ino, seq (version), and
3036          * parent for all of our realms.  If the mds has any newer info,
3037          * it will tell us.
3038          */
3039         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3040                 struct ceph_snap_realm *realm =
3041                         rb_entry(p, struct ceph_snap_realm, node);
3042                 struct ceph_mds_snaprealm_reconnect sr_rec;
3043
3044                 dout(" adding snap realm %llx seq %lld parent %llx\n",
3045                      realm->ino, realm->seq, realm->parent_ino);
3046                 sr_rec.ino = cpu_to_le64(realm->ino);
3047                 sr_rec.seq = cpu_to_le64(realm->seq);
3048                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3049                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3050                 if (err)
3051                         goto fail;
3052         }
3053
3054         reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3055
3056         /* raced with cap release? */
3057         if (s_nr_caps != recon_state.nr_caps) {
3058                 struct page *page = list_first_entry(&pagelist->head,
3059                                                      struct page, lru);
3060                 __le32 *addr = kmap_atomic(page);
3061                 *addr = cpu_to_le32(recon_state.nr_caps);
3062                 kunmap_atomic(addr);
3063         }
3064
3065         reply->hdr.data_len = cpu_to_le32(pagelist->length);
3066         ceph_msg_data_add_pagelist(reply, pagelist);
3067
3068         ceph_early_kick_flushing_caps(mdsc, session);
3069
3070         ceph_con_send(&session->s_con, reply);
3071
3072         mutex_unlock(&session->s_mutex);
3073
3074         mutex_lock(&mdsc->mutex);
3075         __wake_requests(mdsc, &session->s_waiting);
3076         mutex_unlock(&mdsc->mutex);
3077
3078         up_read(&mdsc->snap_rwsem);
3079         return;
3080
3081 fail:
3082         ceph_msg_put(reply);
3083         up_read(&mdsc->snap_rwsem);
3084         mutex_unlock(&session->s_mutex);
3085 fail_nomsg:
3086         ceph_pagelist_release(pagelist);
3087 fail_nopagelist:
3088         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3089         return;
3090 }
3091
3092
3093 /*
3094  * compare old and new mdsmaps, kicking requests
3095  * and closing out old connections as necessary
3096  *
3097  * called under mdsc->mutex.
3098  */
3099 static void check_new_map(struct ceph_mds_client *mdsc,
3100                           struct ceph_mdsmap *newmap,
3101                           struct ceph_mdsmap *oldmap)
3102 {
3103         int i;
3104         int oldstate, newstate;
3105         struct ceph_mds_session *s;
3106
3107         dout("check_new_map new %u old %u\n",
3108              newmap->m_epoch, oldmap->m_epoch);
3109
3110         for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
3111                 if (mdsc->sessions[i] == NULL)
3112                         continue;
3113                 s = mdsc->sessions[i];
3114                 oldstate = ceph_mdsmap_get_state(oldmap, i);
3115                 newstate = ceph_mdsmap_get_state(newmap, i);
3116
3117                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3118                      i, ceph_mds_state_name(oldstate),
3119                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3120                      ceph_mds_state_name(newstate),
3121                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3122                      ceph_session_state_name(s->s_state));
3123
3124                 if (i >= newmap->m_max_mds ||
3125                     memcmp(ceph_mdsmap_get_addr(oldmap, i),
3126                            ceph_mdsmap_get_addr(newmap, i),
3127                            sizeof(struct ceph_entity_addr))) {
3128                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3129                                 /* the session never opened, just close it
3130                                  * out now */
3131                                 __wake_requests(mdsc, &s->s_waiting);
3132                                 __unregister_session(mdsc, s);
3133                         } else {
3134                                 /* just close it */
3135                                 mutex_unlock(&mdsc->mutex);
3136                                 mutex_lock(&s->s_mutex);
3137                                 mutex_lock(&mdsc->mutex);
3138                                 ceph_con_close(&s->s_con);
3139                                 mutex_unlock(&s->s_mutex);
3140                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3141                         }
3142                 } else if (oldstate == newstate) {
3143                         continue;  /* nothing new with this mds */
3144                 }
3145
3146                 /*
3147                  * send reconnect?
3148                  */
3149                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3150                     newstate >= CEPH_MDS_STATE_RECONNECT) {
3151                         mutex_unlock(&mdsc->mutex);
3152                         send_mds_reconnect(mdsc, s);
3153                         mutex_lock(&mdsc->mutex);
3154                 }
3155
3156                 /*
3157                  * kick request on any mds that has gone active.
3158                  */
3159                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3160                     newstate >= CEPH_MDS_STATE_ACTIVE) {
3161                         if (oldstate != CEPH_MDS_STATE_CREATING &&
3162                             oldstate != CEPH_MDS_STATE_STARTING)
3163                                 pr_info("mds%d recovery completed\n", s->s_mds);
3164                         kick_requests(mdsc, i);
3165                         ceph_kick_flushing_caps(mdsc, s);
3166                         wake_up_session_caps(s, 1);
3167                 }
3168         }
3169
3170         for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
3171                 s = mdsc->sessions[i];
3172                 if (!s)
3173                         continue;
3174                 if (!ceph_mdsmap_is_laggy(newmap, i))
3175                         continue;
3176                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3177                     s->s_state == CEPH_MDS_SESSION_HUNG ||
3178                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
3179                         dout(" connecting to export targets of laggy mds%d\n",
3180                              i);
3181                         __open_export_target_sessions(mdsc, s);
3182                 }
3183         }
3184 }
3185
3186
3187
3188 /*
3189  * leases
3190  */
3191
3192 /*
3193  * caller must hold session s_mutex, dentry->d_lock
3194  */
3195 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3196 {
3197         struct ceph_dentry_info *di = ceph_dentry(dentry);
3198
3199         ceph_put_mds_session(di->lease_session);
3200         di->lease_session = NULL;
3201 }
3202
3203 static void handle_lease(struct ceph_mds_client *mdsc,
3204                          struct ceph_mds_session *session,
3205                          struct ceph_msg *msg)
3206 {
3207         struct super_block *sb = mdsc->fsc->sb;
3208         struct inode *inode;
3209         struct dentry *parent, *dentry;
3210         struct ceph_dentry_info *di;
3211         int mds = session->s_mds;
3212         struct ceph_mds_lease *h = msg->front.iov_base;
3213         u32 seq;
3214         struct ceph_vino vino;
3215         struct qstr dname;
3216         int release = 0;
3217
3218         dout("handle_lease from mds%d\n", mds);
3219
3220         /* decode */
3221         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3222                 goto bad;
3223         vino.ino = le64_to_cpu(h->ino);
3224         vino.snap = CEPH_NOSNAP;
3225         seq = le32_to_cpu(h->seq);
3226         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3227         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3228         if (dname.len != get_unaligned_le32(h+1))
3229                 goto bad;
3230
3231         /* lookup inode */
3232         inode = ceph_find_inode(sb, vino);
3233         dout("handle_lease %s, ino %llx %p %.*s\n",
3234              ceph_lease_op_name(h->action), vino.ino, inode,
3235              dname.len, dname.name);
3236
3237         mutex_lock(&session->s_mutex);
3238         session->s_seq++;
3239
3240         if (inode == NULL) {
3241                 dout("handle_lease no inode %llx\n", vino.ino);
3242                 goto release;
3243         }
3244
3245         /* dentry */
3246         parent = d_find_alias(inode);
3247         if (!parent) {
3248                 dout("no parent dentry on inode %p\n", inode);
3249                 WARN_ON(1);
3250                 goto release;  /* hrm... */
3251         }
3252         dname.hash = full_name_hash(parent, dname.name, dname.len);
3253         dentry = d_lookup(parent, &dname);
3254         dput(parent);
3255         if (!dentry)
3256                 goto release;
3257
3258         spin_lock(&dentry->d_lock);
3259         di = ceph_dentry(dentry);
3260         switch (h->action) {
3261         case CEPH_MDS_LEASE_REVOKE:
3262                 if (di->lease_session == session) {
3263                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3264                                 h->seq = cpu_to_le32(di->lease_seq);
3265                         __ceph_mdsc_drop_dentry_lease(dentry);
3266                 }
3267                 release = 1;
3268                 break;
3269
3270         case CEPH_MDS_LEASE_RENEW:
3271                 if (di->lease_session == session &&
3272                     di->lease_gen == session->s_cap_gen &&
3273                     di->lease_renew_from &&
3274                     di->lease_renew_after == 0) {
3275                         unsigned long duration =
3276                                 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3277
3278                         di->lease_seq = seq;
3279                         di->time = di->lease_renew_from + duration;
3280                         di->lease_renew_after = di->lease_renew_from +
3281                                 (duration >> 1);
3282                         di->lease_renew_from = 0;
3283                 }
3284                 break;
3285         }
3286         spin_unlock(&dentry->d_lock);
3287         dput(dentry);
3288
3289         if (!release)
3290                 goto out;
3291
3292 release:
3293         /* let's just reuse the same message */
3294         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3295         ceph_msg_get(msg);
3296         ceph_con_send(&session->s_con, msg);
3297
3298 out:
3299         iput(inode);
3300         mutex_unlock(&session->s_mutex);
3301         return;
3302
3303 bad:
3304         pr_err("corrupt lease message\n");
3305         ceph_msg_dump(msg);
3306 }
3307
3308 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3309                               struct inode *inode,
3310                               struct dentry *dentry, char action,
3311                               u32 seq)
3312 {
3313         struct ceph_msg *msg;
3314         struct ceph_mds_lease *lease;
3315         int len = sizeof(*lease) + sizeof(u32);
3316         int dnamelen = 0;
3317
3318         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3319              inode, dentry, ceph_lease_op_name(action), session->s_mds);
3320         dnamelen = dentry->d_name.len;
3321         len += dnamelen;
3322
3323         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3324         if (!msg)
3325                 return;
3326         lease = msg->front.iov_base;
3327         lease->action = action;
3328         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3329         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3330         lease->seq = cpu_to_le32(seq);
3331         put_unaligned_le32(dnamelen, lease + 1);
3332         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3333
3334         /*
3335          * if this is a preemptive lease RELEASE, no need to
3336          * flush request stream, since the actual request will
3337          * soon follow.
3338          */
3339         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3340
3341         ceph_con_send(&session->s_con, msg);
3342 }
3343
3344 /*
3345  * drop all leases (and dentry refs) in preparation for umount
3346  */
3347 static void drop_leases(struct ceph_mds_client *mdsc)
3348 {
3349         int i;
3350
3351         dout("drop_leases\n");
3352         mutex_lock(&mdsc->mutex);
3353         for (i = 0; i < mdsc->max_sessions; i++) {
3354                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3355                 if (!s)
3356                         continue;
3357                 mutex_unlock(&mdsc->mutex);
3358                 mutex_lock(&s->s_mutex);
3359                 mutex_unlock(&s->s_mutex);
3360                 ceph_put_mds_session(s);
3361                 mutex_lock(&mdsc->mutex);
3362         }
3363         mutex_unlock(&mdsc->mutex);
3364 }
3365
3366
3367
3368 /*
3369  * delayed work -- periodically trim expired leases, renew caps with mds
3370  */
3371 static void schedule_delayed(struct ceph_mds_client *mdsc)
3372 {
3373         int delay = 5;
3374         unsigned hz = round_jiffies_relative(HZ * delay);
3375         schedule_delayed_work(&mdsc->delayed_work, hz);
3376 }
3377
3378 static void delayed_work(struct work_struct *work)
3379 {
3380         int i;
3381         struct ceph_mds_client *mdsc =
3382                 container_of(work, struct ceph_mds_client, delayed_work.work);
3383         int renew_interval;
3384         int renew_caps;
3385
3386         dout("mdsc delayed_work\n");
3387         ceph_check_delayed_caps(mdsc);
3388
3389         mutex_lock(&mdsc->mutex);
3390         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3391         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3392                                    mdsc->last_renew_caps);
3393         if (renew_caps)
3394                 mdsc->last_renew_caps = jiffies;
3395
3396         for (i = 0; i < mdsc->max_sessions; i++) {
3397                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3398                 if (s == NULL)
3399                         continue;
3400                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3401                         dout("resending session close request for mds%d\n",
3402                              s->s_mds);
3403                         request_close_session(mdsc, s);
3404                         ceph_put_mds_session(s);
3405                         continue;
3406                 }
3407                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3408                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3409                                 s->s_state = CEPH_MDS_SESSION_HUNG;
3410                                 pr_info("mds%d hung\n", s->s_mds);
3411                         }
3412                 }
3413                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3414                         /* this mds is failed or recovering, just wait */
3415                         ceph_put_mds_session(s);
3416                         continue;
3417                 }
3418                 mutex_unlock(&mdsc->mutex);
3419
3420                 mutex_lock(&s->s_mutex);
3421                 if (renew_caps)
3422                         send_renew_caps(mdsc, s);
3423                 else
3424                         ceph_con_keepalive(&s->s_con);
3425                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3426                     s->s_state == CEPH_MDS_SESSION_HUNG)
3427                         ceph_send_cap_releases(mdsc, s);
3428                 mutex_unlock(&s->s_mutex);
3429                 ceph_put_mds_session(s);
3430
3431                 mutex_lock(&mdsc->mutex);
3432         }
3433         mutex_unlock(&mdsc->mutex);
3434
3435         schedule_delayed(mdsc);
3436 }
3437
3438 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3439
3440 {
3441         struct ceph_mds_client *mdsc;
3442
3443         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3444         if (!mdsc)
3445                 return -ENOMEM;
3446         mdsc->fsc = fsc;
3447         fsc->mdsc = mdsc;
3448         mutex_init(&mdsc->mutex);
3449         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3450         if (mdsc->mdsmap == NULL) {
3451                 kfree(mdsc);
3452                 return -ENOMEM;
3453         }
3454
3455         init_completion(&mdsc->safe_umount_waiters);
3456         init_waitqueue_head(&mdsc->session_close_wq);
3457         INIT_LIST_HEAD(&mdsc->waiting_for_map);
3458         mdsc->sessions = NULL;
3459         atomic_set(&mdsc->num_sessions, 0);
3460         mdsc->max_sessions = 0;
3461         mdsc->stopping = 0;
3462         mdsc->last_snap_seq = 0;
3463         init_rwsem(&mdsc->snap_rwsem);
3464         mdsc->snap_realms = RB_ROOT;
3465         INIT_LIST_HEAD(&mdsc->snap_empty);
3466         spin_lock_init(&mdsc->snap_empty_lock);
3467         mdsc->last_tid = 0;
3468         mdsc->oldest_tid = 0;
3469         mdsc->request_tree = RB_ROOT;
3470         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3471         mdsc->last_renew_caps = jiffies;
3472         INIT_LIST_HEAD(&mdsc->cap_delay_list);
3473         spin_lock_init(&mdsc->cap_delay_lock);
3474         INIT_LIST_HEAD(&mdsc->snap_flush_list);
3475         spin_lock_init(&mdsc->snap_flush_lock);
3476         mdsc->last_cap_flush_tid = 1;
3477         INIT_LIST_HEAD(&mdsc->cap_flush_list);
3478         INIT_LIST_HEAD(&mdsc->cap_dirty);
3479         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3480         mdsc->num_cap_flushing = 0;
3481         spin_lock_init(&mdsc->cap_dirty_lock);
3482         init_waitqueue_head(&mdsc->cap_flushing_wq);
3483         spin_lock_init(&mdsc->dentry_lru_lock);
3484         INIT_LIST_HEAD(&mdsc->dentry_lru);
3485
3486         ceph_caps_init(mdsc);
3487         ceph_adjust_min_caps(mdsc, fsc->min_caps);
3488
3489         init_rwsem(&mdsc->pool_perm_rwsem);
3490         mdsc->pool_perm_tree = RB_ROOT;
3491
3492         return 0;
3493 }
3494
3495 /*
3496  * Wait for safe replies on open mds requests.  If we time out, drop
3497  * all requests from the tree to avoid dangling dentry refs.
3498  */
3499 static void wait_requests(struct ceph_mds_client *mdsc)
3500 {
3501         struct ceph_options *opts = mdsc->fsc->client->options;
3502         struct ceph_mds_request *req;
3503
3504         mutex_lock(&mdsc->mutex);
3505         if (__get_oldest_req(mdsc)) {
3506                 mutex_unlock(&mdsc->mutex);
3507
3508                 dout("wait_requests waiting for requests\n");
3509                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3510                                     ceph_timeout_jiffies(opts->mount_timeout));
3511
3512                 /* tear down remaining requests */
3513                 mutex_lock(&mdsc->mutex);
3514                 while ((req = __get_oldest_req(mdsc))) {
3515                         dout("wait_requests timed out on tid %llu\n",
3516                              req->r_tid);
3517                         __unregister_request(mdsc, req);
3518                 }
3519         }
3520         mutex_unlock(&mdsc->mutex);
3521         dout("wait_requests done\n");
3522 }
3523
3524 /*
3525  * called before mount is ro, and before dentries are torn down.
3526  * (hmm, does this still race with new lookups?)
3527  */
3528 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3529 {
3530         dout("pre_umount\n");
3531         mdsc->stopping = 1;
3532
3533         drop_leases(mdsc);
3534         ceph_flush_dirty_caps(mdsc);
3535         wait_requests(mdsc);
3536
3537         /*
3538          * wait for reply handlers to drop their request refs and
3539          * their inode/dcache refs
3540          */
3541         ceph_msgr_flush();
3542 }
3543
3544 /*
3545  * wait for all write mds requests to flush.
3546  */
3547 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3548 {
3549         struct ceph_mds_request *req = NULL, *nextreq;
3550         struct rb_node *n;
3551
3552         mutex_lock(&mdsc->mutex);
3553         dout("wait_unsafe_requests want %lld\n", want_tid);
3554 restart:
3555         req = __get_oldest_req(mdsc);
3556         while (req && req->r_tid <= want_tid) {
3557                 /* find next request */
3558                 n = rb_next(&req->r_node);
3559                 if (n)
3560                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3561                 else
3562                         nextreq = NULL;
3563                 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3564                     (req->r_op & CEPH_MDS_OP_WRITE)) {
3565                         /* write op */
3566                         ceph_mdsc_get_request(req);
3567                         if (nextreq)
3568                                 ceph_mdsc_get_request(nextreq);
3569                         mutex_unlock(&mdsc->mutex);
3570                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
3571                              req->r_tid, want_tid);
3572                         wait_for_completion(&req->r_safe_completion);
3573                         mutex_lock(&mdsc->mutex);
3574                         ceph_mdsc_put_request(req);
3575                         if (!nextreq)
3576                                 break;  /* next dne before, so we're done! */
3577                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
3578                                 /* next request was removed from tree */
3579                                 ceph_mdsc_put_request(nextreq);
3580                                 goto restart;
3581                         }
3582                         ceph_mdsc_put_request(nextreq);  /* won't go away */
3583                 }
3584                 req = nextreq;
3585         }
3586         mutex_unlock(&mdsc->mutex);
3587         dout("wait_unsafe_requests done\n");
3588 }
3589
3590 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3591 {
3592         u64 want_tid, want_flush;
3593
3594         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3595                 return;
3596
3597         dout("sync\n");
3598         mutex_lock(&mdsc->mutex);
3599         want_tid = mdsc->last_tid;
3600         mutex_unlock(&mdsc->mutex);
3601
3602         ceph_flush_dirty_caps(mdsc);
3603         spin_lock(&mdsc->cap_dirty_lock);
3604         want_flush = mdsc->last_cap_flush_tid;
3605         if (!list_empty(&mdsc->cap_flush_list)) {
3606                 struct ceph_cap_flush *cf =
3607                         list_last_entry(&mdsc->cap_flush_list,
3608                                         struct ceph_cap_flush, g_list);
3609                 cf->wake = true;
3610         }
3611         spin_unlock(&mdsc->cap_dirty_lock);
3612
3613         dout("sync want tid %lld flush_seq %lld\n",
3614              want_tid, want_flush);
3615
3616         wait_unsafe_requests(mdsc, want_tid);
3617         wait_caps_flush(mdsc, want_flush);
3618 }
3619
3620 /*
3621  * true if all sessions are closed, or we force unmount
3622  */
3623 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
3624 {
3625         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3626                 return true;
3627         return atomic_read(&mdsc->num_sessions) <= skipped;
3628 }
3629
3630 /*
3631  * called after sb is ro.
3632  */
3633 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3634 {
3635         struct ceph_options *opts = mdsc->fsc->client->options;
3636         struct ceph_mds_session *session;
3637         int i;
3638         int skipped = 0;
3639
3640         dout("close_sessions\n");
3641
3642         /* close sessions */
3643         mutex_lock(&mdsc->mutex);
3644         for (i = 0; i < mdsc->max_sessions; i++) {
3645                 session = __ceph_lookup_mds_session(mdsc, i);
3646                 if (!session)
3647                         continue;
3648                 mutex_unlock(&mdsc->mutex);
3649                 mutex_lock(&session->s_mutex);
3650                 if (__close_session(mdsc, session) <= 0)
3651                         skipped++;
3652                 mutex_unlock(&session->s_mutex);
3653                 ceph_put_mds_session(session);
3654                 mutex_lock(&mdsc->mutex);
3655         }
3656         mutex_unlock(&mdsc->mutex);
3657
3658         dout("waiting for sessions to close\n");
3659         wait_event_timeout(mdsc->session_close_wq,
3660                            done_closing_sessions(mdsc, skipped),
3661                            ceph_timeout_jiffies(opts->mount_timeout));
3662
3663         /* tear down remaining sessions */
3664         mutex_lock(&mdsc->mutex);
3665         for (i = 0; i < mdsc->max_sessions; i++) {
3666                 if (mdsc->sessions[i]) {
3667                         session = get_session(mdsc->sessions[i]);
3668                         __unregister_session(mdsc, session);
3669                         mutex_unlock(&mdsc->mutex);
3670                         mutex_lock(&session->s_mutex);
3671                         remove_session_caps(session);
3672                         mutex_unlock(&session->s_mutex);
3673                         ceph_put_mds_session(session);
3674                         mutex_lock(&mdsc->mutex);
3675                 }
3676         }
3677         WARN_ON(!list_empty(&mdsc->cap_delay_list));
3678         mutex_unlock(&mdsc->mutex);
3679
3680         ceph_cleanup_empty_realms(mdsc);
3681
3682         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3683
3684         dout("stopped\n");
3685 }
3686
3687 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
3688 {
3689         struct ceph_mds_session *session;
3690         int mds;
3691
3692         dout("force umount\n");
3693
3694         mutex_lock(&mdsc->mutex);
3695         for (mds = 0; mds < mdsc->max_sessions; mds++) {
3696                 session = __ceph_lookup_mds_session(mdsc, mds);
3697                 if (!session)
3698                         continue;
3699                 mutex_unlock(&mdsc->mutex);
3700                 mutex_lock(&session->s_mutex);
3701                 __close_session(mdsc, session);
3702                 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
3703                         cleanup_session_requests(mdsc, session);
3704                         remove_session_caps(session);
3705                 }
3706                 mutex_unlock(&session->s_mutex);
3707                 ceph_put_mds_session(session);
3708                 mutex_lock(&mdsc->mutex);
3709                 kick_requests(mdsc, mds);
3710         }
3711         __wake_requests(mdsc, &mdsc->waiting_for_map);
3712         mutex_unlock(&mdsc->mutex);
3713 }
3714
3715 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3716 {
3717         dout("stop\n");
3718         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3719         if (mdsc->mdsmap)
3720                 ceph_mdsmap_destroy(mdsc->mdsmap);
3721         kfree(mdsc->sessions);
3722         ceph_caps_finalize(mdsc);
3723         ceph_pool_perm_destroy(mdsc);
3724 }
3725
3726 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3727 {
3728         struct ceph_mds_client *mdsc = fsc->mdsc;
3729
3730         dout("mdsc_destroy %p\n", mdsc);
3731         ceph_mdsc_stop(mdsc);
3732
3733         /* flush out any connection work with references to us */
3734         ceph_msgr_flush();
3735
3736         fsc->mdsc = NULL;
3737         kfree(mdsc);
3738         dout("mdsc_destroy %p done\n", mdsc);
3739 }
3740
3741 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3742 {
3743         struct ceph_fs_client *fsc = mdsc->fsc;
3744         const char *mds_namespace = fsc->mount_options->mds_namespace;
3745         void *p = msg->front.iov_base;
3746         void *end = p + msg->front.iov_len;
3747         u32 epoch;
3748         u32 map_len;
3749         u32 num_fs;
3750         u32 mount_fscid = (u32)-1;
3751         u8 struct_v, struct_cv;
3752         int err = -EINVAL;
3753
3754         ceph_decode_need(&p, end, sizeof(u32), bad);
3755         epoch = ceph_decode_32(&p);
3756
3757         dout("handle_fsmap epoch %u\n", epoch);
3758
3759         ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3760         struct_v = ceph_decode_8(&p);
3761         struct_cv = ceph_decode_8(&p);
3762         map_len = ceph_decode_32(&p);
3763
3764         ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
3765         p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
3766
3767         num_fs = ceph_decode_32(&p);
3768         while (num_fs-- > 0) {
3769                 void *info_p, *info_end;
3770                 u32 info_len;
3771                 u8 info_v, info_cv;
3772                 u32 fscid, namelen;
3773
3774                 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3775                 info_v = ceph_decode_8(&p);
3776                 info_cv = ceph_decode_8(&p);
3777                 info_len = ceph_decode_32(&p);
3778                 ceph_decode_need(&p, end, info_len, bad);
3779                 info_p = p;
3780                 info_end = p + info_len;
3781                 p = info_end;
3782
3783                 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
3784                 fscid = ceph_decode_32(&info_p);
3785                 namelen = ceph_decode_32(&info_p);
3786                 ceph_decode_need(&info_p, info_end, namelen, bad);
3787
3788                 if (mds_namespace &&
3789                     strlen(mds_namespace) == namelen &&
3790                     !strncmp(mds_namespace, (char *)info_p, namelen)) {
3791                         mount_fscid = fscid;
3792                         break;
3793                 }
3794         }
3795
3796         ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
3797         if (mount_fscid != (u32)-1) {
3798                 fsc->client->monc.fs_cluster_id = mount_fscid;
3799                 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
3800                                    0, true);
3801                 ceph_monc_renew_subs(&fsc->client->monc);
3802         } else {
3803                 err = -ENOENT;
3804                 goto err_out;
3805         }
3806         return;
3807 bad:
3808         pr_err("error decoding fsmap\n");
3809 err_out:
3810         mutex_lock(&mdsc->mutex);
3811         mdsc->mdsmap_err = -ENOENT;
3812         __wake_requests(mdsc, &mdsc->waiting_for_map);
3813         mutex_unlock(&mdsc->mutex);
3814         return;
3815 }
3816
3817 /*
3818  * handle mds map update.
3819  */
3820 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3821 {
3822         u32 epoch;
3823         u32 maplen;
3824         void *p = msg->front.iov_base;
3825         void *end = p + msg->front.iov_len;
3826         struct ceph_mdsmap *newmap, *oldmap;
3827         struct ceph_fsid fsid;
3828         int err = -EINVAL;
3829
3830         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3831         ceph_decode_copy(&p, &fsid, sizeof(fsid));
3832         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3833                 return;
3834         epoch = ceph_decode_32(&p);
3835         maplen = ceph_decode_32(&p);
3836         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3837
3838         /* do we need it? */
3839         mutex_lock(&mdsc->mutex);
3840         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3841                 dout("handle_map epoch %u <= our %u\n",
3842                      epoch, mdsc->mdsmap->m_epoch);
3843                 mutex_unlock(&mdsc->mutex);
3844                 return;
3845         }
3846
3847         newmap = ceph_mdsmap_decode(&p, end);
3848         if (IS_ERR(newmap)) {
3849                 err = PTR_ERR(newmap);
3850                 goto bad_unlock;
3851         }
3852
3853         /* swap into place */
3854         if (mdsc->mdsmap) {
3855                 oldmap = mdsc->mdsmap;
3856                 mdsc->mdsmap = newmap;
3857                 check_new_map(mdsc, newmap, oldmap);
3858                 ceph_mdsmap_destroy(oldmap);
3859         } else {
3860                 mdsc->mdsmap = newmap;  /* first mds map */
3861         }
3862         mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3863
3864         __wake_requests(mdsc, &mdsc->waiting_for_map);
3865         ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
3866                           mdsc->mdsmap->m_epoch);
3867
3868         mutex_unlock(&mdsc->mutex);
3869         schedule_delayed(mdsc);
3870         return;
3871
3872 bad_unlock:
3873         mutex_unlock(&mdsc->mutex);
3874 bad:
3875         pr_err("error decoding mdsmap %d\n", err);
3876         return;
3877 }
3878
3879 static struct ceph_connection *con_get(struct ceph_connection *con)
3880 {
3881         struct ceph_mds_session *s = con->private;
3882
3883         if (get_session(s)) {
3884                 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3885                 return con;
3886         }
3887         dout("mdsc con_get %p FAIL\n", s);
3888         return NULL;
3889 }
3890
3891 static void con_put(struct ceph_connection *con)
3892 {
3893         struct ceph_mds_session *s = con->private;
3894
3895         dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3896         ceph_put_mds_session(s);
3897 }
3898
3899 /*
3900  * if the client is unresponsive for long enough, the mds will kill
3901  * the session entirely.
3902  */
3903 static void peer_reset(struct ceph_connection *con)
3904 {
3905         struct ceph_mds_session *s = con->private;
3906         struct ceph_mds_client *mdsc = s->s_mdsc;
3907
3908         pr_warn("mds%d closed our session\n", s->s_mds);
3909         send_mds_reconnect(mdsc, s);
3910 }
3911
3912 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3913 {
3914         struct ceph_mds_session *s = con->private;
3915         struct ceph_mds_client *mdsc = s->s_mdsc;
3916         int type = le16_to_cpu(msg->hdr.type);
3917
3918         mutex_lock(&mdsc->mutex);
3919         if (__verify_registered_session(mdsc, s) < 0) {
3920                 mutex_unlock(&mdsc->mutex);
3921                 goto out;
3922         }
3923         mutex_unlock(&mdsc->mutex);
3924
3925         switch (type) {
3926         case CEPH_MSG_MDS_MAP:
3927                 ceph_mdsc_handle_mdsmap(mdsc, msg);
3928                 break;
3929         case CEPH_MSG_FS_MAP_USER:
3930                 ceph_mdsc_handle_fsmap(mdsc, msg);
3931                 break;
3932         case CEPH_MSG_CLIENT_SESSION:
3933                 handle_session(s, msg);
3934                 break;
3935         case CEPH_MSG_CLIENT_REPLY:
3936                 handle_reply(s, msg);
3937                 break;
3938         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3939                 handle_forward(mdsc, s, msg);
3940                 break;
3941         case CEPH_MSG_CLIENT_CAPS:
3942                 ceph_handle_caps(s, msg);
3943                 break;
3944         case CEPH_MSG_CLIENT_SNAP:
3945                 ceph_handle_snap(mdsc, s, msg);
3946                 break;
3947         case CEPH_MSG_CLIENT_LEASE:
3948                 handle_lease(mdsc, s, msg);
3949                 break;
3950
3951         default:
3952                 pr_err("received unknown message type %d %s\n", type,
3953                        ceph_msg_type_name(type));
3954         }
3955 out:
3956         ceph_msg_put(msg);
3957 }
3958
3959 /*
3960  * authentication
3961  */
3962
3963 /*
3964  * Note: returned pointer is the address of a structure that's
3965  * managed separately.  Caller must *not* attempt to free it.
3966  */
3967 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
3968                                         int *proto, int force_new)
3969 {
3970         struct ceph_mds_session *s = con->private;
3971         struct ceph_mds_client *mdsc = s->s_mdsc;
3972         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3973         struct ceph_auth_handshake *auth = &s->s_auth;
3974
3975         if (force_new && auth->authorizer) {
3976                 ceph_auth_destroy_authorizer(auth->authorizer);
3977                 auth->authorizer = NULL;
3978         }
3979         if (!auth->authorizer) {
3980                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3981                                                       auth);
3982                 if (ret)
3983                         return ERR_PTR(ret);
3984         } else {
3985                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3986                                                       auth);
3987                 if (ret)
3988                         return ERR_PTR(ret);
3989         }
3990         *proto = ac->protocol;
3991
3992         return auth;
3993 }
3994
3995
3996 static int verify_authorizer_reply(struct ceph_connection *con)
3997 {
3998         struct ceph_mds_session *s = con->private;
3999         struct ceph_mds_client *mdsc = s->s_mdsc;
4000         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4001
4002         return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
4003 }
4004
4005 static int invalidate_authorizer(struct ceph_connection *con)
4006 {
4007         struct ceph_mds_session *s = con->private;
4008         struct ceph_mds_client *mdsc = s->s_mdsc;
4009         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4010
4011         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
4012
4013         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
4014 }
4015
4016 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
4017                                 struct ceph_msg_header *hdr, int *skip)
4018 {
4019         struct ceph_msg *msg;
4020         int type = (int) le16_to_cpu(hdr->type);
4021         int front_len = (int) le32_to_cpu(hdr->front_len);
4022
4023         if (con->in_msg)
4024                 return con->in_msg;
4025
4026         *skip = 0;
4027         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
4028         if (!msg) {
4029                 pr_err("unable to allocate msg type %d len %d\n",
4030                        type, front_len);
4031                 return NULL;
4032         }
4033
4034         return msg;
4035 }
4036
4037 static int mds_sign_message(struct ceph_msg *msg)
4038 {
4039        struct ceph_mds_session *s = msg->con->private;
4040        struct ceph_auth_handshake *auth = &s->s_auth;
4041
4042        return ceph_auth_sign_message(auth, msg);
4043 }
4044
4045 static int mds_check_message_signature(struct ceph_msg *msg)
4046 {
4047        struct ceph_mds_session *s = msg->con->private;
4048        struct ceph_auth_handshake *auth = &s->s_auth;
4049
4050        return ceph_auth_check_message_signature(auth, msg);
4051 }
4052
4053 static const struct ceph_connection_operations mds_con_ops = {
4054         .get = con_get,
4055         .put = con_put,
4056         .dispatch = dispatch,
4057         .get_authorizer = get_authorizer,
4058         .verify_authorizer_reply = verify_authorizer_reply,
4059         .invalidate_authorizer = invalidate_authorizer,
4060         .peer_reset = peer_reset,
4061         .alloc_msg = mds_alloc_msg,
4062         .sign_message = mds_sign_message,
4063         .check_message_signature = mds_check_message_signature,
4064 };
4065
4066 /* eof */