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  42  * are no longer valid.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