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  42  * are no longer valid.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