Pileus Git - ~andy/linux/blob - fs/ceph/mds_client.c

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  42  * are no longer valid.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