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

   1 #include "ceph_debug.h"
   2
   3 #include <linux/wait.h>
   4 #include <linux/sched.h>
   5
   6 #include "mds_client.h"
   7 #include "mon_client.h"
   8 #include "super.h"
   9 #include "messenger.h"
  10 #include "decode.h"
  11 #include "auth.h"
  12 #include "pagelist.h"
  13
  14 /*
  15  * A cluster of MDS (metadata server) daemons is responsible for
  16  * managing the file system namespace (the directory hierarchy and
  17  * inodes) and for coordinating shared access to storage.  Metadata is
  18  * partitioning hierarchically across a number of servers, and that
  19  * partition varies over time as the cluster adjusts the distribution
  20  * in order to balance load.
  21  *
  22  * The MDS client is primarily responsible to managing synchronous
  23  * metadata requests for operations like open, unlink, and so forth.
  24  * If there is a MDS failure, we find out about it when we (possibly
  25  * request and) receive a new MDS map, and can resubmit affected
  26  * requests.
  27  *
  28  * For the most part, though, we take advantage of a lossless
  29  * communications channel to the MDS, and do not need to worry about
  30  * timing out or resubmitting requests.
  31  *
  32  * We maintain a stateful "session" with each MDS we interact with.
  33  * Within each session, we sent periodic heartbeat messages to ensure
  34  * any capabilities or leases we have been issues remain valid.  If
  35  * the session times out and goes stale, our leases and capabilities
  36  * are no longer valid.
  37  */
  38
  39 static void __wake_requests(struct ceph_mds_client *mdsc,
  40                             struct list_head *head);
  41
  42 const static struct ceph_connection_operations mds_con_ops;
  43
  44
  45 /*
  46  * mds reply parsing
  47  */
  48
  49 /*
  50  * parse individual inode info
  51  */
  52 static int parse_reply_info_in(void **p, void *end,
  53                                struct ceph_mds_reply_info_in *info)
  54 {
  55         int err = -EIO;
  56
  57         info->in = *p;
  58         *p += sizeof(struct ceph_mds_reply_inode) +
  59                 sizeof(*info->in->fragtree.splits) *
  60                 le32_to_cpu(info->in->fragtree.nsplits);
  61
  62         ceph_decode_32_safe(p, end, info->symlink_len, bad);
  63         ceph_decode_need(p, end, info->symlink_len, bad);
  64         info->symlink = *p;
  65         *p += info->symlink_len;
  66
  67         ceph_decode_32_safe(p, end, info->xattr_len, bad);
  68         ceph_decode_need(p, end, info->xattr_len, bad);
  69         info->xattr_data = *p;
  70         *p += info->xattr_len;
  71         return 0;
  72 bad:
  73         return err;
  74 }
  75
  76 /*
  77  * parse a normal reply, which may contain a (dir+)dentry and/or a
  78  * target inode.
  79  */
  80 static int parse_reply_info_trace(void **p, void *end,
  81                                   struct ceph_mds_reply_info_parsed *info)
  82 {
  83         int err;
  84
  85         if (info->head->is_dentry) {
  86                 err = parse_reply_info_in(p, end, &info->diri);
  87                 if (err < 0)
  88                         goto out_bad;
  89
  90                 if (unlikely(*p + sizeof(*info->dirfrag) > end))
  91                         goto bad;
  92                 info->dirfrag = *p;
  93                 *p += sizeof(*info->dirfrag) +
  94                         sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
  95                 if (unlikely(*p > end))
  96                         goto bad;
  97
  98                 ceph_decode_32_safe(p, end, info->dname_len, bad);
  99                 ceph_decode_need(p, end, info->dname_len, bad);
 100                 info->dname = *p;
 101                 *p += info->dname_len;
 102                 info->dlease = *p;
 103                 *p += sizeof(*info->dlease);
 104         }
 105
 106         if (info->head->is_target) {
 107                 err = parse_reply_info_in(p, end, &info->targeti);
 108                 if (err < 0)
 109                         goto out_bad;
 110         }
 111
 112         if (unlikely(*p != end))
 113                 goto bad;
 114         return 0;
 115
 116 bad:
 117         err = -EIO;
 118 out_bad:
 119         pr_err("problem parsing mds trace %d\n", err);
 120         return err;
 121 }
 122
 123 /*
 124  * parse readdir results
 125  */
 126 static int parse_reply_info_dir(void **p, void *end,
 127                                 struct ceph_mds_reply_info_parsed *info)
 128 {
 129         u32 num, i = 0;
 130         int err;
 131
 132         info->dir_dir = *p;
 133         if (*p + sizeof(*info->dir_dir) > end)
 134                 goto bad;
 135         *p += sizeof(*info->dir_dir) +
 136                 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
 137         if (*p > end)
 138                 goto bad;
 139
 140         ceph_decode_need(p, end, sizeof(num) + 2, bad);
 141         num = ceph_decode_32(p);
 142         info->dir_end = ceph_decode_8(p);
 143         info->dir_complete = ceph_decode_8(p);
 144         if (num == 0)
 145                 goto done;
 146
 147         /* alloc large array */
 148         info->dir_nr = num;
 149         info->dir_in = kcalloc(num, sizeof(*info->dir_in) +
 150                                sizeof(*info->dir_dname) +
 151                                sizeof(*info->dir_dname_len) +
 152                                sizeof(*info->dir_dlease),
 153                                GFP_NOFS);
 154         if (info->dir_in == NULL) {
 155                 err = -ENOMEM;
 156                 goto out_bad;
 157         }
 158         info->dir_dname = (void *)(info->dir_in + num);
 159         info->dir_dname_len = (void *)(info->dir_dname + num);
 160         info->dir_dlease = (void *)(info->dir_dname_len + num);
 161
 162         while (num) {
 163                 /* dentry */
 164                 ceph_decode_need(p, end, sizeof(u32)*2, bad);
 165                 info->dir_dname_len[i] = ceph_decode_32(p);
 166                 ceph_decode_need(p, end, info->dir_dname_len[i], bad);
 167                 info->dir_dname[i] = *p;
 168                 *p += info->dir_dname_len[i];
 169                 dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
 170                      info->dir_dname[i]);
 171                 info->dir_dlease[i] = *p;
 172                 *p += sizeof(struct ceph_mds_reply_lease);
 173
 174                 /* inode */
 175                 err = parse_reply_info_in(p, end, &info->dir_in[i]);
 176                 if (err < 0)
 177                         goto out_bad;
 178                 i++;
 179                 num--;
 180         }
 181
 182 done:
 183         if (*p != end)
 184                 goto bad;
 185         return 0;
 186
 187 bad:
 188         err = -EIO;
 189 out_bad:
 190         pr_err("problem parsing dir contents %d\n", err);
 191         return err;
 192 }
 193
 194 /*
 195  * parse entire mds reply
 196  */
 197 static int parse_reply_info(struct ceph_msg *msg,
 198                             struct ceph_mds_reply_info_parsed *info)
 199 {
 200         void *p, *end;
 201         u32 len;
 202         int err;
 203
 204         info->head = msg->front.iov_base;
 205         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 206         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 207
 208         /* trace */
 209         ceph_decode_32_safe(&p, end, len, bad);
 210         if (len > 0) {
 211                 err = parse_reply_info_trace(&p, p+len, info);
 212                 if (err < 0)
 213                         goto out_bad;
 214         }
 215
 216         /* dir content */
 217         ceph_decode_32_safe(&p, end, len, bad);
 218         if (len > 0) {
 219                 err = parse_reply_info_dir(&p, p+len, info);
 220                 if (err < 0)
 221                         goto out_bad;
 222         }
 223
 224         /* snap blob */
 225         ceph_decode_32_safe(&p, end, len, bad);
 226         info->snapblob_len = len;
 227         info->snapblob = p;
 228         p += len;
 229
 230         if (p != end)
 231                 goto bad;
 232         return 0;
 233
 234 bad:
 235         err = -EIO;
 236 out_bad:
 237         pr_err("mds parse_reply err %d\n", err);
 238         return err;
 239 }
 240
 241 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
 242 {
 243         kfree(info->dir_in);
 244 }
 245
 246
 247 /*
 248  * sessions
 249  */
 250 static const char *session_state_name(int s)
 251 {
 252         switch (s) {
 253         case CEPH_MDS_SESSION_NEW: return "new";
 254         case CEPH_MDS_SESSION_OPENING: return "opening";
 255         case CEPH_MDS_SESSION_OPEN: return "open";
 256         case CEPH_MDS_SESSION_HUNG: return "hung";
 257         case CEPH_MDS_SESSION_CLOSING: return "closing";
 258         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
 259         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
 260         default: return "???";
 261         }
 262 }
 263
 264 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
 265 {
 266         if (atomic_inc_not_zero(&s->s_ref)) {
 267                 dout("mdsc get_session %p %d -> %d\n", s,
 268                      atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
 269                 return s;
 270         } else {
 271                 dout("mdsc get_session %p 0 -- FAIL", s);
 272                 return NULL;
 273         }
 274 }
 275
 276 void ceph_put_mds_session(struct ceph_mds_session *s)
 277 {
 278         dout("mdsc put_session %p %d -> %d\n", s,
 279              atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
 280         if (atomic_dec_and_test(&s->s_ref)) {
 281                 if (s->s_authorizer)
 282                         s->s_mdsc->client->monc.auth->ops->destroy_authorizer(
 283                                 s->s_mdsc->client->monc.auth, s->s_authorizer);
 284                 kfree(s);
 285         }
 286 }
 287
 288 /*
 289  * called under mdsc->mutex
 290  */
 291 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
 292                                                    int mds)
 293 {
 294         struct ceph_mds_session *session;
 295
 296         if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
 297                 return NULL;
 298         session = mdsc->sessions[mds];
 299         dout("lookup_mds_session %p %d\n", session,
 300              atomic_read(&session->s_ref));
 301         get_session(session);
 302         return session;
 303 }
 304
 305 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
 306 {
 307         if (mds >= mdsc->max_sessions)
 308                 return false;
 309         return mdsc->sessions[mds];
 310 }
 311
 312 static int __verify_registered_session(struct ceph_mds_client *mdsc,
 313                                        struct ceph_mds_session *s)
 314 {
 315         if (s->s_mds >= mdsc->max_sessions ||
 316             mdsc->sessions[s->s_mds] != s)
 317                 return -ENOENT;
 318         return 0;
 319 }
 320
 321 /*
 322  * create+register a new session for given mds.
 323  * called under mdsc->mutex.
 324  */
 325 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 326                                                  int mds)
 327 {
 328         struct ceph_mds_session *s;
 329
 330         s = kzalloc(sizeof(*s), GFP_NOFS);
 331         if (!s)
 332                 return ERR_PTR(-ENOMEM);
 333         s->s_mdsc = mdsc;
 334         s->s_mds = mds;
 335         s->s_state = CEPH_MDS_SESSION_NEW;
 336         s->s_ttl = 0;
 337         s->s_seq = 0;
 338         mutex_init(&s->s_mutex);
 339
 340         ceph_con_init(mdsc->client->msgr, &s->s_con);
 341         s->s_con.private = s;
 342         s->s_con.ops = &mds_con_ops;
 343         s->s_con.peer_name.type = CEPH_ENTITY_TYPE_MDS;
 344         s->s_con.peer_name.num = cpu_to_le64(mds);
 345
 346         spin_lock_init(&s->s_cap_lock);
 347         s->s_cap_gen = 0;
 348         s->s_cap_ttl = 0;
 349         s->s_renew_requested = 0;
 350         s->s_renew_seq = 0;
 351         INIT_LIST_HEAD(&s->s_caps);
 352         s->s_nr_caps = 0;
 353         s->s_trim_caps = 0;
 354         atomic_set(&s->s_ref, 1);
 355         INIT_LIST_HEAD(&s->s_waiting);
 356         INIT_LIST_HEAD(&s->s_unsafe);
 357         s->s_num_cap_releases = 0;
 358         s->s_cap_iterator = NULL;
 359         INIT_LIST_HEAD(&s->s_cap_releases);
 360         INIT_LIST_HEAD(&s->s_cap_releases_done);
 361         INIT_LIST_HEAD(&s->s_cap_flushing);
 362         INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
 363
 364         dout("register_session mds%d\n", mds);
 365         if (mds >= mdsc->max_sessions) {
 366                 int newmax = 1 << get_count_order(mds+1);
 367                 struct ceph_mds_session **sa;
 368
 369                 dout("register_session realloc to %d\n", newmax);
 370                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 371                 if (sa == NULL)
 372                         goto fail_realloc;
 373                 if (mdsc->sessions) {
 374                         memcpy(sa, mdsc->sessions,
 375                                mdsc->max_sessions * sizeof(void *));
 376                         kfree(mdsc->sessions);
 377                 }
 378                 mdsc->sessions = sa;
 379                 mdsc->max_sessions = newmax;
 380         }
 381         mdsc->sessions[mds] = s;
 382         atomic_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
 383
 384         ceph_con_open(&s->s_con, ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
 385
 386         return s;
 387
 388 fail_realloc:
 389         kfree(s);
 390         return ERR_PTR(-ENOMEM);
 391 }
 392
 393 /*
 394  * called under mdsc->mutex
 395  */
 396 static void __unregister_session(struct ceph_mds_client *mdsc,
 397                                struct ceph_mds_session *s)
 398 {
 399         dout("__unregister_session mds%d %p\n", s->s_mds, s);
 400         BUG_ON(mdsc->sessions[s->s_mds] != s);
 401         mdsc->sessions[s->s_mds] = NULL;
 402         ceph_con_close(&s->s_con);
 403         ceph_put_mds_session(s);
 404 }
 405
 406 /*
 407  * drop session refs in request.
 408  *
 409  * should be last request ref, or hold mdsc->mutex
 410  */
 411 static void put_request_session(struct ceph_mds_request *req)
 412 {
 413         if (req->r_session) {
 414                 ceph_put_mds_session(req->r_session);
 415                 req->r_session = NULL;
 416         }
 417 }
 418
 419 void ceph_mdsc_release_request(struct kref *kref)
 420 {
 421         struct ceph_mds_request *req = container_of(kref,
 422                                                     struct ceph_mds_request,
 423                                                     r_kref);
 424         if (req->r_request)
 425                 ceph_msg_put(req->r_request);
 426         if (req->r_reply) {
 427                 ceph_msg_put(req->r_reply);
 428                 destroy_reply_info(&req->r_reply_info);
 429         }
 430         if (req->r_inode) {
 431                 ceph_put_cap_refs(ceph_inode(req->r_inode),
 432                                   CEPH_CAP_PIN);
 433                 iput(req->r_inode);
 434         }
 435         if (req->r_locked_dir)
 436                 ceph_put_cap_refs(ceph_inode(req->r_locked_dir),
 437                                   CEPH_CAP_PIN);
 438         if (req->r_target_inode)
 439                 iput(req->r_target_inode);
 440         if (req->r_dentry)
 441                 dput(req->r_dentry);
 442         if (req->r_old_dentry) {
 443                 ceph_put_cap_refs(
 444                         ceph_inode(req->r_old_dentry->d_parent->d_inode),
 445                         CEPH_CAP_PIN);
 446                 dput(req->r_old_dentry);
 447         }
 448         kfree(req->r_path1);
 449         kfree(req->r_path2);
 450         put_request_session(req);
 451         ceph_unreserve_caps(&req->r_caps_reservation);
 452         kfree(req);
 453 }
 454
 455 /*
 456  * lookup session, bump ref if found.
 457  *
 458  * called under mdsc->mutex.
 459  */
 460 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
 461                                              u64 tid)
 462 {
 463         struct ceph_mds_request *req;
 464         struct rb_node *n = mdsc->request_tree.rb_node;
 465
 466         while (n) {
 467                 req = rb_entry(n, struct ceph_mds_request, r_node);
 468                 if (tid < req->r_tid)
 469                         n = n->rb_left;
 470                 else if (tid > req->r_tid)
 471                         n = n->rb_right;
 472                 else {
 473                         ceph_mdsc_get_request(req);
 474                         return req;
 475                 }
 476         }
 477         return NULL;
 478 }
 479
 480 static void __insert_request(struct ceph_mds_client *mdsc,
 481                              struct ceph_mds_request *new)
 482 {
 483         struct rb_node **p = &mdsc->request_tree.rb_node;
 484         struct rb_node *parent = NULL;
 485         struct ceph_mds_request *req = NULL;
 486
 487         while (*p) {
 488                 parent = *p;
 489                 req = rb_entry(parent, struct ceph_mds_request, r_node);
 490                 if (new->r_tid < req->r_tid)
 491                         p = &(*p)->rb_left;
 492                 else if (new->r_tid > req->r_tid)
 493                         p = &(*p)->rb_right;
 494                 else
 495                         BUG();
 496         }
 497
 498         rb_link_node(&new->r_node, parent, p);
 499         rb_insert_color(&new->r_node, &mdsc->request_tree);
 500 }
 501
 502 /*
 503  * Register an in-flight request, and assign a tid.  Link to directory
 504  * are modifying (if any).
 505  *
 506  * Called under mdsc->mutex.
 507  */
 508 static void __register_request(struct ceph_mds_client *mdsc,
 509                                struct ceph_mds_request *req,
 510                                struct inode *dir)
 511 {
 512         req->r_tid = ++mdsc->last_tid;
 513         if (req->r_num_caps)
 514                 ceph_reserve_caps(&req->r_caps_reservation, req->r_num_caps);
 515         dout("__register_request %p tid %lld\n", req, req->r_tid);
 516         ceph_mdsc_get_request(req);
 517         __insert_request(mdsc, req);
 518
 519         if (dir) {
 520                 struct ceph_inode_info *ci = ceph_inode(dir);
 521
 522                 spin_lock(&ci->i_unsafe_lock);
 523                 req->r_unsafe_dir = dir;
 524                 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
 525                 spin_unlock(&ci->i_unsafe_lock);
 526         }
 527 }
 528
 529 static void __unregister_request(struct ceph_mds_client *mdsc,
 530                                  struct ceph_mds_request *req)
 531 {
 532         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 533         rb_erase(&req->r_node, &mdsc->request_tree);
 534         RB_CLEAR_NODE(&req->r_node);
 535         ceph_mdsc_put_request(req);
 536
 537         if (req->r_unsafe_dir) {
 538                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 539
 540                 spin_lock(&ci->i_unsafe_lock);
 541                 list_del_init(&req->r_unsafe_dir_item);
 542                 spin_unlock(&ci->i_unsafe_lock);
 543         }
 544 }
 545
 546 /*
 547  * Choose mds to send request to next.  If there is a hint set in the
 548  * request (e.g., due to a prior forward hint from the mds), use that.
 549  * Otherwise, consult frag tree and/or caps to identify the
 550  * appropriate mds.  If all else fails, choose randomly.
 551  *
 552  * Called under mdsc->mutex.
 553  */
 554 static int __choose_mds(struct ceph_mds_client *mdsc,
 555                         struct ceph_mds_request *req)
 556 {
 557         struct inode *inode;
 558         struct ceph_inode_info *ci;
 559         struct ceph_cap *cap;
 560         int mode = req->r_direct_mode;
 561         int mds = -1;
 562         u32 hash = req->r_direct_hash;
 563         bool is_hash = req->r_direct_is_hash;
 564
 565         /*
 566          * is there a specific mds we should try?  ignore hint if we have
 567          * no session and the mds is not up (active or recovering).
 568          */
 569         if (req->r_resend_mds >= 0 &&
 570             (__have_session(mdsc, req->r_resend_mds) ||
 571              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 572                 dout("choose_mds using resend_mds mds%d\n",
 573                      req->r_resend_mds);
 574                 return req->r_resend_mds;
 575         }
 576
 577         if (mode == USE_RANDOM_MDS)
 578                 goto random;
 579
 580         inode = NULL;
 581         if (req->r_inode) {
 582                 inode = req->r_inode;
 583         } else if (req->r_dentry) {
 584                 if (req->r_dentry->d_inode) {
 585                         inode = req->r_dentry->d_inode;
 586                 } else {
 587                         inode = req->r_dentry->d_parent->d_inode;
 588                         hash = req->r_dentry->d_name.hash;
 589                         is_hash = true;
 590                 }
 591         }
 592         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 593              (int)hash, mode);
 594         if (!inode)
 595                 goto random;
 596         ci = ceph_inode(inode);
 597
 598         if (is_hash && S_ISDIR(inode->i_mode)) {
 599                 struct ceph_inode_frag frag;
 600                 int found;
 601
 602                 ceph_choose_frag(ci, hash, &frag, &found);
 603                 if (found) {
 604                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 605                                 u8 r;
 606
 607                                 /* choose a random replica */
 608                                 get_random_bytes(&r, 1);
 609                                 r %= frag.ndist;
 610                                 mds = frag.dist[r];
 611                                 dout("choose_mds %p %llx.%llx "
 612                                      "frag %u mds%d (%d/%d)\n",
 613                                      inode, ceph_vinop(inode),
 614                                      frag.frag, frag.mds,
 615                                      (int)r, frag.ndist);
 616                                 return mds;
 617                         }
 618
 619                         /* since this file/dir wasn't known to be
 620                          * replicated, then we want to look for the
 621                          * authoritative mds. */
 622                         mode = USE_AUTH_MDS;
 623                         if (frag.mds >= 0) {
 624                                 /* choose auth mds */
 625                                 mds = frag.mds;
 626                                 dout("choose_mds %p %llx.%llx "
 627                                      "frag %u mds%d (auth)\n",
 628                                      inode, ceph_vinop(inode), frag.frag, mds);
 629                                 return mds;
 630                         }
 631                 }
 632         }
 633
 634         spin_lock(&inode->i_lock);
 635         cap = NULL;
 636         if (mode == USE_AUTH_MDS)
 637                 cap = ci->i_auth_cap;
 638         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
 639                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
 640         if (!cap) {
 641                 spin_unlock(&inode->i_lock);
 642                 goto random;
 643         }
 644         mds = cap->session->s_mds;
 645         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
 646              inode, ceph_vinop(inode), mds,
 647              cap == ci->i_auth_cap ? "auth " : "", cap);
 648         spin_unlock(&inode->i_lock);
 649         return mds;
 650
 651 random:
 652         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
 653         dout("choose_mds chose random mds%d\n", mds);
 654         return mds;
 655 }
 656
 657
 658 /*
 659  * session messages
 660  */
 661 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
 662 {
 663         struct ceph_msg *msg;
 664         struct ceph_mds_session_head *h;
 665
 666         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), 0, 0, NULL);
 667         if (IS_ERR(msg)) {
 668                 pr_err("create_session_msg ENOMEM creating msg\n");
 669                 return ERR_PTR(PTR_ERR(msg));
 670         }
 671         h = msg->front.iov_base;
 672         h->op = cpu_to_le32(op);
 673         h->seq = cpu_to_le64(seq);
 674         return msg;
 675 }
 676
 677 /*
 678  * send session open request.
 679  *
 680  * called under mdsc->mutex
 681  */
 682 static int __open_session(struct ceph_mds_client *mdsc,
 683                           struct ceph_mds_session *session)
 684 {
 685         struct ceph_msg *msg;
 686         int mstate;
 687         int mds = session->s_mds;
 688         int err = 0;
 689
 690         /* wait for mds to go active? */
 691         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
 692         dout("open_session to mds%d (%s)\n", mds,
 693              ceph_mds_state_name(mstate));
 694         session->s_state = CEPH_MDS_SESSION_OPENING;
 695         session->s_renew_requested = jiffies;
 696
 697         /* send connect message */
 698         msg = create_session_msg(CEPH_SESSION_REQUEST_OPEN, session->s_seq);
 699         if (IS_ERR(msg)) {
 700                 err = PTR_ERR(msg);
 701                 goto out;
 702         }
 703         ceph_con_send(&session->s_con, msg);
 704
 705 out:
 706         return 0;
 707 }
 708
 709 /*
 710  * session caps
 711  */
 712
 713 /*
 714  * Free preallocated cap messages assigned to this session
 715  */
 716 static void cleanup_cap_releases(struct ceph_mds_session *session)
 717 {
 718         struct ceph_msg *msg;
 719
 720         spin_lock(&session->s_cap_lock);
 721         while (!list_empty(&session->s_cap_releases)) {
 722                 msg = list_first_entry(&session->s_cap_releases,
 723                                        struct ceph_msg, list_head);
 724                 list_del_init(&msg->list_head);
 725                 ceph_msg_put(msg);
 726         }
 727         while (!list_empty(&session->s_cap_releases_done)) {
 728                 msg = list_first_entry(&session->s_cap_releases_done,
 729                                        struct ceph_msg, list_head);
 730                 list_del_init(&msg->list_head);
 731                 ceph_msg_put(msg);
 732         }
 733         spin_unlock(&session->s_cap_lock);
 734 }
 735
 736 /*
 737  * Helper to safely iterate over all caps associated with a session.
 738  *
 739  * caller must hold session s_mutex
 740  */
 741 static int iterate_session_caps(struct ceph_mds_session *session,
 742                                  int (*cb)(struct inode *, struct ceph_cap *,
 743                                             void *), void *arg)
 744 {
 745         struct list_head *p;
 746         struct ceph_cap *cap;
 747         struct inode *inode, *last_inode = NULL;
 748         struct ceph_cap *old_cap = NULL;
 749         int ret;
 750
 751         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
 752         spin_lock(&session->s_cap_lock);
 753         p = session->s_caps.next;
 754         while (p != &session->s_caps) {
 755                 cap = list_entry(p, struct ceph_cap, session_caps);
 756                 inode = igrab(&cap->ci->vfs_inode);
 757                 if (!inode) {
 758                         p = p->next;
 759                         continue;
 760                 }
 761                 session->s_cap_iterator = cap;
 762                 spin_unlock(&session->s_cap_lock);
 763
 764                 if (last_inode) {
 765                         iput(last_inode);
 766                         last_inode = NULL;
 767                 }
 768                 if (old_cap) {
 769                         ceph_put_cap(old_cap);
 770                         old_cap = NULL;
 771                 }
 772
 773                 ret = cb(inode, cap, arg);
 774                 last_inode = inode;
 775
 776                 spin_lock(&session->s_cap_lock);
 777                 p = p->next;
 778                 if (cap->ci == NULL) {
 779                         dout("iterate_session_caps  finishing cap %p removal\n",
 780                              cap);
 781                         BUG_ON(cap->session != session);
 782                         list_del_init(&cap->session_caps);
 783                         session->s_nr_caps--;
 784                         cap->session = NULL;
 785                         old_cap = cap;  /* put_cap it w/o locks held */
 786                 }
 787                 if (ret < 0)
 788                         goto out;
 789         }
 790         ret = 0;
 791 out:
 792         session->s_cap_iterator = NULL;
 793         spin_unlock(&session->s_cap_lock);
 794
 795         if (last_inode)
 796                 iput(last_inode);
 797         if (old_cap)
 798                 ceph_put_cap(old_cap);
 799
 800         return ret;
 801 }
 802
 803 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
 804                                    void *arg)
 805 {
 806         struct ceph_inode_info *ci = ceph_inode(inode);
 807         dout("removing cap %p, ci is %p, inode is %p\n",
 808              cap, ci, &ci->vfs_inode);
 809         ceph_remove_cap(cap);
 810         return 0;
 811 }
 812
 813 /*
 814  * caller must hold session s_mutex
 815  */
 816 static void remove_session_caps(struct ceph_mds_session *session)
 817 {
 818         dout("remove_session_caps on %p\n", session);
 819         iterate_session_caps(session, remove_session_caps_cb, NULL);
 820         BUG_ON(session->s_nr_caps > 0);
 821         cleanup_cap_releases(session);
 822 }
 823
 824 /*
 825  * wake up any threads waiting on this session's caps.  if the cap is
 826  * old (didn't get renewed on the client reconnect), remove it now.
 827  *
 828  * caller must hold s_mutex.
 829  */
 830 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
 831                               void *arg)
 832 {
 833         struct ceph_inode_info *ci = ceph_inode(inode);
 834
 835         wake_up(&ci->i_cap_wq);
 836         if (arg) {
 837                 spin_lock(&inode->i_lock);
 838                 ci->i_wanted_max_size = 0;
 839                 ci->i_requested_max_size = 0;
 840                 spin_unlock(&inode->i_lock);
 841         }
 842         return 0;
 843 }
 844
 845 static void wake_up_session_caps(struct ceph_mds_session *session,
 846                                  int reconnect)
 847 {
 848         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
 849         iterate_session_caps(session, wake_up_session_cb,
 850                              (void *)(unsigned long)reconnect);
 851 }
 852
 853 /*
 854  * Send periodic message to MDS renewing all currently held caps.  The
 855  * ack will reset the expiration for all caps from this session.
 856  *
 857  * caller holds s_mutex
 858  */
 859 static int send_renew_caps(struct ceph_mds_client *mdsc,
 860                            struct ceph_mds_session *session)
 861 {
 862         struct ceph_msg *msg;
 863         int state;
 864
 865         if (time_after_eq(jiffies, session->s_cap_ttl) &&
 866             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
 867                 pr_info("mds%d caps stale\n", session->s_mds);
 868         session->s_renew_requested = jiffies;
 869
 870         /* do not try to renew caps until a recovering mds has reconnected
 871          * with its clients. */
 872         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
 873         if (state < CEPH_MDS_STATE_RECONNECT) {
 874                 dout("send_renew_caps ignoring mds%d (%s)\n",
 875                      session->s_mds, ceph_mds_state_name(state));
 876                 return 0;
 877         }
 878
 879         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
 880                 ceph_mds_state_name(state));
 881         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
 882                                  ++session->s_renew_seq);
 883         if (IS_ERR(msg))
 884                 return PTR_ERR(msg);
 885         ceph_con_send(&session->s_con, msg);
 886         return 0;
 887 }
 888
 889 /*
 890  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
 891  *
 892  * Called under session->s_mutex
 893  */
 894 static void renewed_caps(struct ceph_mds_client *mdsc,
 895                          struct ceph_mds_session *session, int is_renew)
 896 {
 897         int was_stale;
 898         int wake = 0;
 899
 900         spin_lock(&session->s_cap_lock);
 901         was_stale = is_renew && (session->s_cap_ttl == 0 ||
 902                                  time_after_eq(jiffies, session->s_cap_ttl));
 903
 904         session->s_cap_ttl = session->s_renew_requested +
 905                 mdsc->mdsmap->m_session_timeout*HZ;
 906
 907         if (was_stale) {
 908                 if (time_before(jiffies, session->s_cap_ttl)) {
 909                         pr_info("mds%d caps renewed\n", session->s_mds);
 910                         wake = 1;
 911                 } else {
 912                         pr_info("mds%d caps still stale\n", session->s_mds);
 913                 }
 914         }
 915         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
 916              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
 917              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
 918         spin_unlock(&session->s_cap_lock);
 919
 920         if (wake)
 921                 wake_up_session_caps(session, 0);
 922 }
 923
 924 /*
 925  * send a session close request
 926  */
 927 static int request_close_session(struct ceph_mds_client *mdsc,
 928                                  struct ceph_mds_session *session)
 929 {
 930         struct ceph_msg *msg;
 931         int err = 0;
 932
 933         dout("request_close_session mds%d state %s seq %lld\n",
 934              session->s_mds, session_state_name(session->s_state),
 935              session->s_seq);
 936         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
 937         if (IS_ERR(msg))
 938                 err = PTR_ERR(msg);
 939         else
 940                 ceph_con_send(&session->s_con, msg);
 941         return err;
 942 }
 943
 944 /*
 945  * Called with s_mutex held.
 946  */
 947 static int __close_session(struct ceph_mds_client *mdsc,
 948                          struct ceph_mds_session *session)
 949 {
 950         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
 951                 return 0;
 952         session->s_state = CEPH_MDS_SESSION_CLOSING;
 953         return request_close_session(mdsc, session);
 954 }
 955
 956 /*
 957  * Trim old(er) caps.
 958  *
 959  * Because we can't cache an inode without one or more caps, we do
 960  * this indirectly: if a cap is unused, we prune its aliases, at which
 961  * point the inode will hopefully get dropped to.
 962  *
 963  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
 964  * memory pressure from the MDS, though, so it needn't be perfect.
 965  */
 966 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
 967 {
 968         struct ceph_mds_session *session = arg;
 969         struct ceph_inode_info *ci = ceph_inode(inode);
 970         int used, oissued, mine;
 971
 972         if (session->s_trim_caps <= 0)
 973                 return -1;
 974
 975         spin_lock(&inode->i_lock);
 976         mine = cap->issued | cap->implemented;
 977         used = __ceph_caps_used(ci);
 978         oissued = __ceph_caps_issued_other(ci, cap);
 979
 980         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
 981              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
 982              ceph_cap_string(used));
 983         if (ci->i_dirty_caps)
 984                 goto out;   /* dirty caps */
 985         if ((used & ~oissued) & mine)
 986                 goto out;   /* we need these caps */
 987
 988         session->s_trim_caps--;
 989         if (oissued) {
 990                 /* we aren't the only cap.. just remove us */
 991                 __ceph_remove_cap(cap);
 992         } else {
 993                 /* try to drop referring dentries */
 994                 spin_unlock(&inode->i_lock);
 995                 d_prune_aliases(inode);
 996                 dout("trim_caps_cb %p cap %p  pruned, count now %d\n",
 997                      inode, cap, atomic_read(&inode->i_count));
 998                 return 0;
 999         }
1000
1001 out:
1002         spin_unlock(&inode->i_lock);
1003         return 0;
1004 }
1005
1006 /*
1007  * Trim session cap count down to some max number.
1008  */
1009 static int trim_caps(struct ceph_mds_client *mdsc,
1010                      struct ceph_mds_session *session,
1011                      int max_caps)
1012 {
1013         int trim_caps = session->s_nr_caps - max_caps;
1014
1015         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1016              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1017         if (trim_caps > 0) {
1018                 session->s_trim_caps = trim_caps;
1019                 iterate_session_caps(session, trim_caps_cb, session);
1020                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1021                      session->s_mds, session->s_nr_caps, max_caps,
1022                         trim_caps - session->s_trim_caps);
1023                 session->s_trim_caps = 0;
1024         }
1025         return 0;
1026 }
1027
1028 /*
1029  * Allocate cap_release messages.  If there is a partially full message
1030  * in the queue, try to allocate enough to cover it's remainder, so that
1031  * we can send it immediately.
1032  *
1033  * Called under s_mutex.
1034  */
1035 static int add_cap_releases(struct ceph_mds_client *mdsc,
1036                             struct ceph_mds_session *session,
1037                             int extra)
1038 {
1039         struct ceph_msg *msg;
1040         struct ceph_mds_cap_release *head;
1041         int err = -ENOMEM;
1042
1043         if (extra < 0)
1044                 extra = mdsc->client->mount_args->cap_release_safety;
1045
1046         spin_lock(&session->s_cap_lock);
1047
1048         if (!list_empty(&session->s_cap_releases)) {
1049                 msg = list_first_entry(&session->s_cap_releases,
1050                                        struct ceph_msg,
1051                                  list_head);
1052                 head = msg->front.iov_base;
1053                 extra += CEPH_CAPS_PER_RELEASE - le32_to_cpu(head->num);
1054         }
1055
1056         while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1057                 spin_unlock(&session->s_cap_lock);
1058                 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
1059                                    0, 0, NULL);
1060                 if (!msg)
1061                         goto out_unlocked;
1062                 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1063                      (int)msg->front.iov_len);
1064                 head = msg->front.iov_base;
1065                 head->num = cpu_to_le32(0);
1066                 msg->front.iov_len = sizeof(*head);
1067                 spin_lock(&session->s_cap_lock);
1068                 list_add(&msg->list_head, &session->s_cap_releases);
1069                 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1070         }
1071
1072         if (!list_empty(&session->s_cap_releases)) {
1073                 msg = list_first_entry(&session->s_cap_releases,
1074                                        struct ceph_msg,
1075                                        list_head);
1076                 head = msg->front.iov_base;
1077                 if (head->num) {
1078                         dout(" queueing non-full %p (%d)\n", msg,
1079                              le32_to_cpu(head->num));
1080                         list_move_tail(&msg->list_head,
1081                                       &session->s_cap_releases_done);
1082                         session->s_num_cap_releases -=
1083                                 CEPH_CAPS_PER_RELEASE - le32_to_cpu(head->num);
1084                 }
1085         }
1086         err = 0;
1087         spin_unlock(&session->s_cap_lock);
1088 out_unlocked:
1089         return err;
1090 }
1091
1092 /*
1093  * flush all dirty inode data to disk.
1094  *
1095  * returns true if we've flushed through want_flush_seq
1096  */
1097 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1098 {
1099         int mds, ret = 1;
1100
1101         dout("check_cap_flush want %lld\n", want_flush_seq);
1102         mutex_lock(&mdsc->mutex);
1103         for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1104                 struct ceph_mds_session *session = mdsc->sessions[mds];
1105
1106                 if (!session)
1107                         continue;
1108                 get_session(session);
1109                 mutex_unlock(&mdsc->mutex);
1110
1111                 mutex_lock(&session->s_mutex);
1112                 if (!list_empty(&session->s_cap_flushing)) {
1113                         struct ceph_inode_info *ci =
1114                                 list_entry(session->s_cap_flushing.next,
1115                                            struct ceph_inode_info,
1116                                            i_flushing_item);
1117                         struct inode *inode = &ci->vfs_inode;
1118
1119                         spin_lock(&inode->i_lock);
1120                         if (ci->i_cap_flush_seq <= want_flush_seq) {
1121                                 dout("check_cap_flush still flushing %p "
1122                                      "seq %lld <= %lld to mds%d\n", inode,
1123                                      ci->i_cap_flush_seq, want_flush_seq,
1124                                      session->s_mds);
1125                                 ret = 0;
1126                         }
1127                         spin_unlock(&inode->i_lock);
1128                 }
1129                 mutex_unlock(&session->s_mutex);
1130                 ceph_put_mds_session(session);
1131
1132                 if (!ret)
1133                         return ret;
1134                 mutex_lock(&mdsc->mutex);
1135         }
1136
1137         mutex_unlock(&mdsc->mutex);
1138         dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1139         return ret;
1140 }
1141
1142 /*
1143  * called under s_mutex
1144  */
1145 static void send_cap_releases(struct ceph_mds_client *mdsc,
1146                        struct ceph_mds_session *session)
1147 {
1148         struct ceph_msg *msg;
1149
1150         dout("send_cap_releases mds%d\n", session->s_mds);
1151         while (1) {
1152                 spin_lock(&session->s_cap_lock);
1153                 if (list_empty(&session->s_cap_releases_done))
1154                         break;
1155                 msg = list_first_entry(&session->s_cap_releases_done,
1156                                  struct ceph_msg, list_head);
1157                 list_del_init(&msg->list_head);
1158                 spin_unlock(&session->s_cap_lock);
1159                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1160                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1161                 ceph_con_send(&session->s_con, msg);
1162         }
1163         spin_unlock(&session->s_cap_lock);
1164 }
1165
1166 /*
1167  * requests
1168  */
1169
1170 /*
1171  * Create an mds request.
1172  */
1173 struct ceph_mds_request *
1174 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1175 {
1176         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1177
1178         if (!req)
1179                 return ERR_PTR(-ENOMEM);
1180
1181         req->r_started = jiffies;
1182         req->r_resend_mds = -1;
1183         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1184         req->r_fmode = -1;
1185         kref_init(&req->r_kref);
1186         INIT_LIST_HEAD(&req->r_wait);
1187         init_completion(&req->r_completion);
1188         init_completion(&req->r_safe_completion);
1189         INIT_LIST_HEAD(&req->r_unsafe_item);
1190
1191         req->r_op = op;
1192         req->r_direct_mode = mode;
1193         return req;
1194 }
1195
1196 /*
1197  * return oldest (lowest) request, tid in request tree, 0 if none.
1198  *
1199  * called under mdsc->mutex.
1200  */
1201 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1202 {
1203         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1204                 return NULL;
1205         return rb_entry(rb_first(&mdsc->request_tree),
1206                         struct ceph_mds_request, r_node);
1207 }
1208
1209 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1210 {
1211         struct ceph_mds_request *req = __get_oldest_req(mdsc);
1212
1213         if (req)
1214                 return req->r_tid;
1215         return 0;
1216 }
1217
1218 /*
1219  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1220  * on build_path_from_dentry in fs/cifs/dir.c.
1221  *
1222  * If @stop_on_nosnap, generate path relative to the first non-snapped
1223  * inode.
1224  *
1225  * Encode hidden .snap dirs as a double /, i.e.
1226  *   foo/.snap/bar -> foo//bar
1227  */
1228 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1229                            int stop_on_nosnap)
1230 {
1231         struct dentry *temp;
1232         char *path;
1233         int len, pos;
1234
1235         if (dentry == NULL)
1236                 return ERR_PTR(-EINVAL);
1237
1238 retry:
1239         len = 0;
1240         for (temp = dentry; !IS_ROOT(temp);) {
1241                 struct inode *inode = temp->d_inode;
1242                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1243                         len++;  /* slash only */
1244                 else if (stop_on_nosnap && inode &&
1245                          ceph_snap(inode) == CEPH_NOSNAP)
1246                         break;
1247                 else
1248                         len += 1 + temp->d_name.len;
1249                 temp = temp->d_parent;
1250                 if (temp == NULL) {
1251                         pr_err("build_path_dentry corrupt dentry %p\n", dentry);
1252                         return ERR_PTR(-EINVAL);
1253                 }
1254         }
1255         if (len)
1256                 len--;  /* no leading '/' */
1257
1258         path = kmalloc(len+1, GFP_NOFS);
1259         if (path == NULL)
1260                 return ERR_PTR(-ENOMEM);
1261         pos = len;
1262         path[pos] = 0;  /* trailing null */
1263         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1264                 struct inode *inode = temp->d_inode;
1265
1266                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1267                         dout("build_path_dentry path+%d: %p SNAPDIR\n",
1268                              pos, temp);
1269                 } else if (stop_on_nosnap && inode &&
1270                            ceph_snap(inode) == CEPH_NOSNAP) {
1271                         break;
1272                 } else {
1273                         pos -= temp->d_name.len;
1274                         if (pos < 0)
1275                                 break;
1276                         strncpy(path + pos, temp->d_name.name,
1277                                 temp->d_name.len);
1278                         dout("build_path_dentry path+%d: %p '%.*s'\n",
1279                              pos, temp, temp->d_name.len, path + pos);
1280                 }
1281                 if (pos)
1282                         path[--pos] = '/';
1283                 temp = temp->d_parent;
1284                 if (temp == NULL) {
1285                         pr_err("build_path_dentry corrupt dentry\n");
1286                         kfree(path);
1287                         return ERR_PTR(-EINVAL);
1288                 }
1289         }
1290         if (pos != 0) {
1291                 pr_err("build_path_dentry did not end path lookup where "
1292                        "expected, namelen is %d, pos is %d\n", len, pos);
1293                 /* presumably this is only possible if racing with a
1294                    rename of one of the parent directories (we can not
1295                    lock the dentries above us to prevent this, but
1296                    retrying should be harmless) */
1297                 kfree(path);
1298                 goto retry;
1299         }
1300
1301         *base = ceph_ino(temp->d_inode);
1302         *plen = len;
1303         dout("build_path_dentry on %p %d built %llx '%.*s'\n",
1304              dentry, atomic_read(&dentry->d_count), *base, len, path);
1305         return path;
1306 }
1307
1308 static int build_dentry_path(struct dentry *dentry,
1309                              const char **ppath, int *ppathlen, u64 *pino,
1310                              int *pfreepath)
1311 {
1312         char *path;
1313
1314         if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1315                 *pino = ceph_ino(dentry->d_parent->d_inode);
1316                 *ppath = dentry->d_name.name;
1317                 *ppathlen = dentry->d_name.len;
1318                 return 0;
1319         }
1320         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1321         if (IS_ERR(path))
1322                 return PTR_ERR(path);
1323         *ppath = path;
1324         *pfreepath = 1;
1325         return 0;
1326 }
1327
1328 static int build_inode_path(struct inode *inode,
1329                             const char **ppath, int *ppathlen, u64 *pino,
1330                             int *pfreepath)
1331 {
1332         struct dentry *dentry;
1333         char *path;
1334
1335         if (ceph_snap(inode) == CEPH_NOSNAP) {
1336                 *pino = ceph_ino(inode);
1337                 *ppathlen = 0;
1338                 return 0;
1339         }
1340         dentry = d_find_alias(inode);
1341         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1342         dput(dentry);
1343         if (IS_ERR(path))
1344                 return PTR_ERR(path);
1345         *ppath = path;
1346         *pfreepath = 1;
1347         return 0;
1348 }
1349
1350 /*
1351  * request arguments may be specified via an inode *, a dentry *, or
1352  * an explicit ino+path.
1353  */
1354 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1355                                   const char *rpath, u64 rino,
1356                                   const char **ppath, int *pathlen,
1357                                   u64 *ino, int *freepath)
1358 {
1359         int r = 0;
1360
1361         if (rinode) {
1362                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1363                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1364                      ceph_snap(rinode));
1365         } else if (rdentry) {
1366                 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1367                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1368                      *ppath);
1369         } else if (rpath) {
1370                 *ino = rino;
1371                 *ppath = rpath;
1372                 *pathlen = strlen(rpath);
1373                 dout(" path %.*s\n", *pathlen, rpath);
1374         }
1375
1376         return r;
1377 }
1378
1379 /*
1380  * called under mdsc->mutex
1381  */
1382 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1383                                                struct ceph_mds_request *req,
1384                                                int mds)
1385 {
1386         struct ceph_msg *msg;
1387         struct ceph_mds_request_head *head;
1388         const char *path1 = NULL;
1389         const char *path2 = NULL;
1390         u64 ino1 = 0, ino2 = 0;
1391         int pathlen1 = 0, pathlen2 = 0;
1392         int freepath1 = 0, freepath2 = 0;
1393         int len;
1394         u16 releases;
1395         void *p, *end;
1396         int ret;
1397
1398         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1399                               req->r_path1, req->r_ino1.ino,
1400                               &path1, &pathlen1, &ino1, &freepath1);
1401         if (ret < 0) {
1402                 msg = ERR_PTR(ret);
1403                 goto out;
1404         }
1405
1406         ret = set_request_path_attr(NULL, req->r_old_dentry,
1407                               req->r_path2, req->r_ino2.ino,
1408                               &path2, &pathlen2, &ino2, &freepath2);
1409         if (ret < 0) {
1410                 msg = ERR_PTR(ret);
1411                 goto out_free1;
1412         }
1413
1414         len = sizeof(*head) +
1415                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64));
1416
1417         /* calculate (max) length for cap releases */
1418         len += sizeof(struct ceph_mds_request_release) *
1419                 (!!req->r_inode_drop + !!req->r_dentry_drop +
1420                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1421         if (req->r_dentry_drop)
1422                 len += req->r_dentry->d_name.len;
1423         if (req->r_old_dentry_drop)
1424                 len += req->r_old_dentry->d_name.len;
1425
1426         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, 0, 0, NULL);
1427         if (IS_ERR(msg))
1428                 goto out_free2;
1429
1430         msg->hdr.tid = cpu_to_le64(req->r_tid);
1431
1432         head = msg->front.iov_base;
1433         p = msg->front.iov_base + sizeof(*head);
1434         end = msg->front.iov_base + msg->front.iov_len;
1435
1436         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1437         head->op = cpu_to_le32(req->r_op);
1438         head->caller_uid = cpu_to_le32(current_fsuid());
1439         head->caller_gid = cpu_to_le32(current_fsgid());
1440         head->args = req->r_args;
1441
1442         ceph_encode_filepath(&p, end, ino1, path1);
1443         ceph_encode_filepath(&p, end, ino2, path2);
1444
1445         /* cap releases */
1446         releases = 0;
1447         if (req->r_inode_drop)
1448                 releases += ceph_encode_inode_release(&p,
1449                       req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1450                       mds, req->r_inode_drop, req->r_inode_unless, 0);
1451         if (req->r_dentry_drop)
1452                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1453                        mds, req->r_dentry_drop, req->r_dentry_unless);
1454         if (req->r_old_dentry_drop)
1455                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1456                        mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1457         if (req->r_old_inode_drop)
1458                 releases += ceph_encode_inode_release(&p,
1459                       req->r_old_dentry->d_inode,
1460                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1461         head->num_releases = cpu_to_le16(releases);
1462
1463         BUG_ON(p > end);
1464         msg->front.iov_len = p - msg->front.iov_base;
1465         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1466
1467         msg->pages = req->r_pages;
1468         msg->nr_pages = req->r_num_pages;
1469         msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1470         msg->hdr.data_off = cpu_to_le16(0);
1471
1472 out_free2:
1473         if (freepath2)
1474                 kfree((char *)path2);
1475 out_free1:
1476         if (freepath1)
1477                 kfree((char *)path1);
1478 out:
1479         return msg;
1480 }
1481
1482 /*
1483  * called under mdsc->mutex if error, under no mutex if
1484  * success.
1485  */
1486 static void complete_request(struct ceph_mds_client *mdsc,
1487                              struct ceph_mds_request *req)
1488 {
1489         if (req->r_callback)
1490                 req->r_callback(mdsc, req);
1491         else
1492                 complete(&req->r_completion);
1493 }
1494
1495 /*
1496  * called under mdsc->mutex
1497  */
1498 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1499                                   struct ceph_mds_request *req,
1500                                   int mds)
1501 {
1502         struct ceph_mds_request_head *rhead;
1503         struct ceph_msg *msg;
1504         int flags = 0;
1505
1506         req->r_mds = mds;
1507         req->r_attempts++;
1508         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1509              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1510
1511         if (req->r_request) {
1512                 ceph_msg_put(req->r_request);
1513                 req->r_request = NULL;
1514         }
1515         msg = create_request_message(mdsc, req, mds);
1516         if (IS_ERR(msg)) {
1517                 req->r_reply = ERR_PTR(PTR_ERR(msg));
1518                 complete_request(mdsc, req);
1519                 return -PTR_ERR(msg);
1520         }
1521         req->r_request = msg;
1522
1523         rhead = msg->front.iov_base;
1524         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1525         if (req->r_got_unsafe)
1526                 flags |= CEPH_MDS_FLAG_REPLAY;
1527         if (req->r_locked_dir)
1528                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1529         rhead->flags = cpu_to_le32(flags);
1530         rhead->num_fwd = req->r_num_fwd;
1531         rhead->num_retry = req->r_attempts - 1;
1532
1533         dout(" r_locked_dir = %p\n", req->r_locked_dir);
1534
1535         if (req->r_target_inode && req->r_got_unsafe)
1536                 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1537         else
1538                 rhead->ino = 0;
1539         return 0;
1540 }
1541
1542 /*
1543  * send request, or put it on the appropriate wait list.
1544  */
1545 static int __do_request(struct ceph_mds_client *mdsc,
1546                         struct ceph_mds_request *req)
1547 {
1548         struct ceph_mds_session *session = NULL;
1549         int mds = -1;
1550         int err = -EAGAIN;
1551
1552         if (req->r_reply)
1553                 goto out;
1554
1555         if (req->r_timeout &&
1556             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1557                 dout("do_request timed out\n");
1558                 err = -EIO;
1559                 goto finish;
1560         }
1561
1562         mds = __choose_mds(mdsc, req);
1563         if (mds < 0 ||
1564             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1565                 dout("do_request no mds or not active, waiting for map\n");
1566                 list_add(&req->r_wait, &mdsc->waiting_for_map);
1567                 goto out;
1568         }
1569
1570         /* get, open session */
1571         session = __ceph_lookup_mds_session(mdsc, mds);
1572         if (!session) {
1573                 session = register_session(mdsc, mds);
1574                 if (IS_ERR(session)) {
1575                         err = PTR_ERR(session);
1576                         goto finish;
1577                 }
1578         }
1579         dout("do_request mds%d session %p state %s\n", mds, session,
1580              session_state_name(session->s_state));
1581         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1582             session->s_state != CEPH_MDS_SESSION_HUNG) {
1583                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1584                     session->s_state == CEPH_MDS_SESSION_CLOSING)
1585                         __open_session(mdsc, session);
1586                 list_add(&req->r_wait, &session->s_waiting);
1587                 goto out_session;
1588         }
1589
1590         /* send request */
1591         req->r_session = get_session(session);
1592         req->r_resend_mds = -1;   /* forget any previous mds hint */
1593
1594         if (req->r_request_started == 0)   /* note request start time */
1595                 req->r_request_started = jiffies;
1596
1597         err = __prepare_send_request(mdsc, req, mds);
1598         if (!err) {
1599                 ceph_msg_get(req->r_request);
1600                 ceph_con_send(&session->s_con, req->r_request);
1601         }
1602
1603 out_session:
1604         ceph_put_mds_session(session);
1605 out:
1606         return err;
1607
1608 finish:
1609         req->r_reply = ERR_PTR(err);
1610         complete_request(mdsc, req);
1611         goto out;
1612 }
1613
1614 /*
1615  * called under mdsc->mutex
1616  */
1617 static void __wake_requests(struct ceph_mds_client *mdsc,
1618                             struct list_head *head)
1619 {
1620         struct ceph_mds_request *req, *nreq;
1621
1622         list_for_each_entry_safe(req, nreq, head, r_wait) {
1623                 list_del_init(&req->r_wait);
1624                 __do_request(mdsc, req);
1625         }
1626 }
1627
1628 /*
1629  * Wake up threads with requests pending for @mds, so that they can
1630  * resubmit their requests to a possibly different mds.  If @all is set,
1631  * wake up if their requests has been forwarded to @mds, too.
1632  */
1633 static void kick_requests(struct ceph_mds_client *mdsc, int mds, int all)
1634 {
1635         struct ceph_mds_request *req;
1636         struct rb_node *p;
1637
1638         dout("kick_requests mds%d\n", mds);
1639         for (p = rb_first(&mdsc->request_tree); p; p = rb_next(p)) {
1640                 req = rb_entry(p, struct ceph_mds_request, r_node);
1641                 if (req->r_got_unsafe)
1642                         continue;
1643                 if (req->r_session &&
1644                     req->r_session->s_mds == mds) {
1645                         dout(" kicking tid %llu\n", req->r_tid);
1646                         put_request_session(req);
1647                         __do_request(mdsc, req);
1648                 }
1649         }
1650 }
1651
1652 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
1653                               struct ceph_mds_request *req)
1654 {
1655         dout("submit_request on %p\n", req);
1656         mutex_lock(&mdsc->mutex);
1657         __register_request(mdsc, req, NULL);
1658         __do_request(mdsc, req);
1659         mutex_unlock(&mdsc->mutex);
1660 }
1661
1662 /*
1663  * Synchrously perform an mds request.  Take care of all of the
1664  * session setup, forwarding, retry details.
1665  */
1666 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
1667                          struct inode *dir,
1668                          struct ceph_mds_request *req)
1669 {
1670         int err;
1671
1672         dout("do_request on %p\n", req);
1673
1674         /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
1675         if (req->r_inode)
1676                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
1677         if (req->r_locked_dir)
1678                 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
1679         if (req->r_old_dentry)
1680                 ceph_get_cap_refs(
1681                         ceph_inode(req->r_old_dentry->d_parent->d_inode),
1682                         CEPH_CAP_PIN);
1683
1684         /* issue */
1685         mutex_lock(&mdsc->mutex);
1686         __register_request(mdsc, req, dir);
1687         __do_request(mdsc, req);
1688
1689         /* wait */
1690         if (!req->r_reply) {
1691                 mutex_unlock(&mdsc->mutex);
1692                 if (req->r_timeout) {
1693                         err = (long)wait_for_completion_interruptible_timeout(
1694                                 &req->r_completion, req->r_timeout);
1695                         if (err == 0)
1696                                 req->r_reply = ERR_PTR(-EIO);
1697                         else if (err < 0)
1698                                 req->r_reply = ERR_PTR(err);
1699                 } else {
1700                         err = wait_for_completion_interruptible(
1701                                 &req->r_completion);
1702                         if (err)
1703                                 req->r_reply = ERR_PTR(err);
1704                 }
1705                 mutex_lock(&mdsc->mutex);
1706         }
1707
1708         if (IS_ERR(req->r_reply)) {
1709                 err = PTR_ERR(req->r_reply);
1710                 req->r_reply = NULL;
1711
1712                 if (err == -ERESTARTSYS) {
1713                         /* aborted */
1714                         req->r_aborted = true;
1715
1716                         if (req->r_locked_dir &&
1717                             (req->r_op & CEPH_MDS_OP_WRITE)) {
1718                                 struct ceph_inode_info *ci =
1719                                         ceph_inode(req->r_locked_dir);
1720
1721                                 dout("aborted, clearing I_COMPLETE on %p\n",
1722                                      req->r_locked_dir);
1723                                 spin_lock(&req->r_locked_dir->i_lock);
1724                                 ci->i_ceph_flags &= ~CEPH_I_COMPLETE;
1725                                 ci->i_release_count++;
1726                                 spin_unlock(&req->r_locked_dir->i_lock);
1727                         }
1728                 } else {
1729                         /* clean up this request */
1730                         __unregister_request(mdsc, req);
1731                         if (!list_empty(&req->r_unsafe_item))
1732                                 list_del_init(&req->r_unsafe_item);
1733                         complete(&req->r_safe_completion);
1734                 }
1735         } else if (req->r_err) {
1736                 err = req->r_err;
1737         } else {
1738                 err = le32_to_cpu(req->r_reply_info.head->result);
1739         }
1740         mutex_unlock(&mdsc->mutex);
1741
1742         dout("do_request %p done, result %d\n", req, err);
1743         return err;
1744 }
1745
1746 /*
1747  * Handle mds reply.
1748  *
1749  * We take the session mutex and parse and process the reply immediately.
1750  * This preserves the logical ordering of replies, capabilities, etc., sent
1751  * by the MDS as they are applied to our local cache.
1752  */
1753 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
1754 {
1755         struct ceph_mds_client *mdsc = session->s_mdsc;
1756         struct ceph_mds_request *req;
1757         struct ceph_mds_reply_head *head = msg->front.iov_base;
1758         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
1759         u64 tid;
1760         int err, result;
1761         int mds = session->s_mds;
1762
1763         if (msg->front.iov_len < sizeof(*head)) {
1764                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
1765                 ceph_msg_dump(msg);
1766                 return;
1767         }
1768
1769         /* get request, session */
1770         tid = le64_to_cpu(msg->hdr.tid);
1771         mutex_lock(&mdsc->mutex);
1772         req = __lookup_request(mdsc, tid);
1773         if (!req) {
1774                 dout("handle_reply on unknown tid %llu\n", tid);
1775                 mutex_unlock(&mdsc->mutex);
1776                 return;
1777         }
1778         dout("handle_reply %p\n", req);
1779
1780         /* correct session? */
1781         if (req->r_session != session) {
1782                 pr_err("mdsc_handle_reply got %llu on session mds%d"
1783                        " not mds%d\n", tid, session->s_mds,
1784                        req->r_session ? req->r_session->s_mds : -1);
1785                 mutex_unlock(&mdsc->mutex);
1786                 goto out;
1787         }
1788
1789         /* dup? */
1790         if ((req->r_got_unsafe && !head->safe) ||
1791             (req->r_got_safe && head->safe)) {
1792                 pr_warning("got a dup %s reply on %llu from mds%d\n",
1793                            head->safe ? "safe" : "unsafe", tid, mds);
1794                 mutex_unlock(&mdsc->mutex);
1795                 goto out;
1796         }
1797
1798         result = le32_to_cpu(head->result);
1799
1800         /*
1801          * Tolerate 2 consecutive ESTALEs from the same mds.
1802          * FIXME: we should be looking at the cap migrate_seq.
1803          */
1804         if (result == -ESTALE) {
1805                 req->r_direct_mode = USE_AUTH_MDS;
1806                 req->r_num_stale++;
1807                 if (req->r_num_stale <= 2) {
1808                         __do_request(mdsc, req);
1809                         mutex_unlock(&mdsc->mutex);
1810                         goto out;
1811                 }
1812         } else {
1813                 req->r_num_stale = 0;
1814         }
1815
1816         if (head->safe) {
1817                 req->r_got_safe = true;
1818                 __unregister_request(mdsc, req);
1819                 complete(&req->r_safe_completion);
1820
1821                 if (req->r_got_unsafe) {
1822                         /*
1823                          * We already handled the unsafe response, now do the
1824                          * cleanup.  No need to examine the response; the MDS
1825                          * doesn't include any result info in the safe
1826                          * response.  And even if it did, there is nothing
1827                          * useful we could do with a revised return value.
1828                          */
1829                         dout("got safe reply %llu, mds%d\n", tid, mds);
1830                         list_del_init(&req->r_unsafe_item);
1831
1832                         /* last unsafe request during umount? */
1833                         if (mdsc->stopping && !__get_oldest_req(mdsc))
1834                                 complete(&mdsc->safe_umount_waiters);
1835                         mutex_unlock(&mdsc->mutex);
1836                         goto out;
1837                 }
1838         }
1839
1840         BUG_ON(req->r_reply);
1841
1842         if (!head->safe) {
1843                 req->r_got_unsafe = true;
1844                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
1845         }
1846
1847         dout("handle_reply tid %lld result %d\n", tid, result);
1848         rinfo = &req->r_reply_info;
1849         err = parse_reply_info(msg, rinfo);
1850         mutex_unlock(&mdsc->mutex);
1851
1852         mutex_lock(&session->s_mutex);
1853         if (err < 0) {
1854                 pr_err("mdsc_handle_reply got corrupt reply mds%d\n", mds);
1855                 ceph_msg_dump(msg);
1856                 goto out_err;
1857         }
1858
1859         /* snap trace */
1860         if (rinfo->snapblob_len) {
1861                 down_write(&mdsc->snap_rwsem);
1862                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
1863                                rinfo->snapblob + rinfo->snapblob_len,
1864                                le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
1865                 downgrade_write(&mdsc->snap_rwsem);
1866         } else {
1867                 down_read(&mdsc->snap_rwsem);
1868         }
1869
1870         /* insert trace into our cache */
1871         err = ceph_fill_trace(mdsc->client->sb, req, req->r_session);
1872         if (err == 0) {
1873                 if (result == 0 && rinfo->dir_nr)
1874                         ceph_readdir_prepopulate(req, req->r_session);
1875                 ceph_unreserve_caps(&req->r_caps_reservation);
1876         }
1877
1878         up_read(&mdsc->snap_rwsem);
1879 out_err:
1880         if (err) {
1881                 req->r_err = err;
1882         } else {
1883                 req->r_reply = msg;
1884                 ceph_msg_get(msg);
1885         }
1886
1887         add_cap_releases(mdsc, req->r_session, -1);
1888         mutex_unlock(&session->s_mutex);
1889
1890         /* kick calling process */
1891         complete_request(mdsc, req);
1892 out:
1893         ceph_mdsc_put_request(req);
1894         return;
1895 }
1896
1897
1898
1899 /*
1900  * handle mds notification that our request has been forwarded.
1901  */
1902 static void handle_forward(struct ceph_mds_client *mdsc,
1903                            struct ceph_mds_session *session,
1904                            struct ceph_msg *msg)
1905 {
1906         struct ceph_mds_request *req;
1907         u64 tid = le64_to_cpu(msg->hdr.tid);
1908         u32 next_mds;
1909         u32 fwd_seq;
1910         int err = -EINVAL;
1911         void *p = msg->front.iov_base;
1912         void *end = p + msg->front.iov_len;
1913
1914         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1915         next_mds = ceph_decode_32(&p);
1916         fwd_seq = ceph_decode_32(&p);
1917
1918         mutex_lock(&mdsc->mutex);
1919         req = __lookup_request(mdsc, tid);
1920         if (!req) {
1921                 dout("forward %llu to mds%d - req dne\n", tid, next_mds);
1922                 goto out;  /* dup reply? */
1923         }
1924
1925         if (fwd_seq <= req->r_num_fwd) {
1926                 dout("forward %llu to mds%d - old seq %d <= %d\n",
1927                      tid, next_mds, req->r_num_fwd, fwd_seq);
1928         } else {
1929                 /* resend. forward race not possible; mds would drop */
1930                 dout("forward %llu to mds%d (we resend)\n", tid, next_mds);
1931                 req->r_num_fwd = fwd_seq;
1932                 req->r_resend_mds = next_mds;
1933                 put_request_session(req);
1934                 __do_request(mdsc, req);
1935         }
1936         ceph_mdsc_put_request(req);
1937 out:
1938         mutex_unlock(&mdsc->mutex);
1939         return;
1940
1941 bad:
1942         pr_err("mdsc_handle_forward decode error err=%d\n", err);
1943 }
1944
1945 /*
1946  * handle a mds session control message
1947  */
1948 static void handle_session(struct ceph_mds_session *session,
1949                            struct ceph_msg *msg)
1950 {
1951         struct ceph_mds_client *mdsc = session->s_mdsc;
1952         u32 op;
1953         u64 seq;
1954         int mds = session->s_mds;
1955         struct ceph_mds_session_head *h = msg->front.iov_base;
1956         int wake = 0;
1957
1958         /* decode */
1959         if (msg->front.iov_len != sizeof(*h))
1960                 goto bad;
1961         op = le32_to_cpu(h->op);
1962         seq = le64_to_cpu(h->seq);
1963
1964         mutex_lock(&mdsc->mutex);
1965         if (op == CEPH_SESSION_CLOSE)
1966                 __unregister_session(mdsc, session);
1967         /* FIXME: this ttl calculation is generous */
1968         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
1969         mutex_unlock(&mdsc->mutex);
1970
1971         mutex_lock(&session->s_mutex);
1972
1973         dout("handle_session mds%d %s %p state %s seq %llu\n",
1974              mds, ceph_session_op_name(op), session,
1975              session_state_name(session->s_state), seq);
1976
1977         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
1978                 session->s_state = CEPH_MDS_SESSION_OPEN;
1979                 pr_info("mds%d came back\n", session->s_mds);
1980         }
1981
1982         switch (op) {
1983         case CEPH_SESSION_OPEN:
1984                 session->s_state = CEPH_MDS_SESSION_OPEN;
1985                 renewed_caps(mdsc, session, 0);
1986                 wake = 1;
1987                 if (mdsc->stopping)
1988                         __close_session(mdsc, session);
1989                 break;
1990
1991         case CEPH_SESSION_RENEWCAPS:
1992                 if (session->s_renew_seq == seq)
1993                         renewed_caps(mdsc, session, 1);
1994                 break;
1995
1996         case CEPH_SESSION_CLOSE:
1997                 remove_session_caps(session);
1998                 wake = 1; /* for good measure */
1999                 complete(&mdsc->session_close_waiters);
2000                 kick_requests(mdsc, mds, 0);      /* cur only */
2001                 break;
2002
2003         case CEPH_SESSION_STALE:
2004                 pr_info("mds%d caps went stale, renewing\n",
2005                         session->s_mds);
2006                 spin_lock(&session->s_cap_lock);
2007                 session->s_cap_gen++;
2008                 session->s_cap_ttl = 0;
2009                 spin_unlock(&session->s_cap_lock);
2010                 send_renew_caps(mdsc, session);
2011                 break;
2012
2013         case CEPH_SESSION_RECALL_STATE:
2014                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2015                 break;
2016
2017         default:
2018                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2019                 WARN_ON(1);
2020         }
2021
2022         mutex_unlock(&session->s_mutex);
2023         if (wake) {
2024                 mutex_lock(&mdsc->mutex);
2025                 __wake_requests(mdsc, &session->s_waiting);
2026                 mutex_unlock(&mdsc->mutex);
2027         }
2028         return;
2029
2030 bad:
2031         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2032                (int)msg->front.iov_len);
2033         ceph_msg_dump(msg);
2034         return;
2035 }
2036
2037
2038 /*
2039  * called under session->mutex.
2040  */
2041 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2042                                    struct ceph_mds_session *session)
2043 {
2044         struct ceph_mds_request *req, *nreq;
2045         int err;
2046
2047         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2048
2049         mutex_lock(&mdsc->mutex);
2050         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2051                 err = __prepare_send_request(mdsc, req, session->s_mds);
2052                 if (!err) {
2053                         ceph_msg_get(req->r_request);
2054                         ceph_con_send(&session->s_con, req->r_request);
2055                 }
2056         }
2057         mutex_unlock(&mdsc->mutex);
2058 }
2059
2060 /*
2061  * Encode information about a cap for a reconnect with the MDS.
2062  */
2063 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2064                           void *arg)
2065 {
2066         struct ceph_mds_cap_reconnect rec;
2067         struct ceph_inode_info *ci;
2068         struct ceph_pagelist *pagelist = arg;
2069         char *path;
2070         int pathlen, err;
2071         u64 pathbase;
2072         struct dentry *dentry;
2073
2074         ci = cap->ci;
2075
2076         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2077              inode, ceph_vinop(inode), cap, cap->cap_id,
2078              ceph_cap_string(cap->issued));
2079         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2080         if (err)
2081                 return err;
2082
2083         dentry = d_find_alias(inode);
2084         if (dentry) {
2085                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2086                 if (IS_ERR(path)) {
2087                         err = PTR_ERR(path);
2088                         BUG_ON(err);
2089                 }
2090         } else {
2091                 path = NULL;
2092                 pathlen = 0;
2093         }
2094         err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2095         if (err)
2096                 goto out;
2097
2098         spin_lock(&inode->i_lock);
2099         cap->seq = 0;        /* reset cap seq */
2100         cap->issue_seq = 0;  /* and issue_seq */
2101         rec.cap_id = cpu_to_le64(cap->cap_id);
2102         rec.pathbase = cpu_to_le64(pathbase);
2103         rec.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2104         rec.issued = cpu_to_le32(cap->issued);
2105         rec.size = cpu_to_le64(inode->i_size);
2106         ceph_encode_timespec(&rec.mtime, &inode->i_mtime);
2107         ceph_encode_timespec(&rec.atime, &inode->i_atime);
2108         rec.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2109         spin_unlock(&inode->i_lock);
2110
2111         err = ceph_pagelist_append(pagelist, &rec, sizeof(rec));
2112
2113 out:
2114         kfree(path);
2115         dput(dentry);
2116         return err;
2117 }
2118
2119
2120 /*
2121  * If an MDS fails and recovers, clients need to reconnect in order to
2122  * reestablish shared state.  This includes all caps issued through
2123  * this session _and_ the snap_realm hierarchy.  Because it's not
2124  * clear which snap realms the mds cares about, we send everything we
2125  * know about.. that ensures we'll then get any new info the
2126  * recovering MDS might have.
2127  *
2128  * This is a relatively heavyweight operation, but it's rare.
2129  *
2130  * called with mdsc->mutex held.
2131  */
2132 static void send_mds_reconnect(struct ceph_mds_client *mdsc, int mds)
2133 {
2134         struct ceph_mds_session *session = NULL;
2135         struct ceph_msg *reply;
2136         struct rb_node *p;
2137         int err;
2138         struct ceph_pagelist *pagelist;
2139
2140         pr_info("reconnect to recovering mds%d\n", mds);
2141
2142         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2143         if (!pagelist)
2144                 goto fail_nopagelist;
2145         ceph_pagelist_init(pagelist);
2146
2147         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, 0, 0, NULL);
2148         if (IS_ERR(reply)) {
2149                 err = PTR_ERR(reply);
2150                 goto fail_nomsg;
2151         }
2152
2153         /* find session */
2154         session = __ceph_lookup_mds_session(mdsc, mds);
2155         mutex_unlock(&mdsc->mutex);    /* drop lock for duration */
2156
2157         if (session) {
2158                 mutex_lock(&session->s_mutex);
2159
2160                 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2161                 session->s_seq = 0;
2162
2163                 ceph_con_open(&session->s_con,
2164                               ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2165
2166                 /* replay unsafe requests */
2167                 replay_unsafe_requests(mdsc, session);
2168         } else {
2169                 dout("no session for mds%d, will send short reconnect\n",
2170                      mds);
2171         }
2172
2173         down_read(&mdsc->snap_rwsem);
2174
2175         if (!session)
2176                 goto send;
2177         dout("session %p state %s\n", session,
2178              session_state_name(session->s_state));
2179
2180         /* traverse this session's caps */
2181         err = ceph_pagelist_encode_32(pagelist, session->s_nr_caps);
2182         if (err)
2183                 goto fail;
2184         err = iterate_session_caps(session, encode_caps_cb, pagelist);
2185         if (err < 0)
2186                 goto out;
2187
2188         /*
2189          * snaprealms.  we provide mds with the ino, seq (version), and
2190          * parent for all of our realms.  If the mds has any newer info,
2191          * it will tell us.
2192          */
2193         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2194                 struct ceph_snap_realm *realm =
2195                         rb_entry(p, struct ceph_snap_realm, node);
2196                 struct ceph_mds_snaprealm_reconnect sr_rec;
2197
2198                 dout(" adding snap realm %llx seq %lld parent %llx\n",
2199                      realm->ino, realm->seq, realm->parent_ino);
2200                 sr_rec.ino = cpu_to_le64(realm->ino);
2201                 sr_rec.seq = cpu_to_le64(realm->seq);
2202                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2203                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2204                 if (err)
2205                         goto fail;
2206         }
2207
2208 send:
2209         reply->pagelist = pagelist;
2210         reply->hdr.data_len = cpu_to_le32(pagelist->length);
2211         reply->nr_pages = calc_pages_for(0, pagelist->length);
2212         ceph_con_send(&session->s_con, reply);
2213
2214         if (session) {
2215                 session->s_state = CEPH_MDS_SESSION_OPEN;
2216                 __wake_requests(mdsc, &session->s_waiting);
2217         }
2218
2219 out:
2220         up_read(&mdsc->snap_rwsem);
2221         if (session) {
2222                 mutex_unlock(&session->s_mutex);
2223                 ceph_put_mds_session(session);
2224         }
2225         mutex_lock(&mdsc->mutex);
2226         return;
2227
2228 fail:
2229         ceph_msg_put(reply);
2230 fail_nomsg:
2231         ceph_pagelist_release(pagelist);
2232         kfree(pagelist);
2233 fail_nopagelist:
2234         pr_err("ENOMEM preparing reconnect for mds%d\n", mds);
2235         goto out;
2236 }
2237
2238
2239 /*
2240  * compare old and new mdsmaps, kicking requests
2241  * and closing out old connections as necessary
2242  *
2243  * called under mdsc->mutex.
2244  */
2245 static void check_new_map(struct ceph_mds_client *mdsc,
2246                           struct ceph_mdsmap *newmap,
2247                           struct ceph_mdsmap *oldmap)
2248 {
2249         int i;
2250         int oldstate, newstate;
2251         struct ceph_mds_session *s;
2252
2253         dout("check_new_map new %u old %u\n",
2254              newmap->m_epoch, oldmap->m_epoch);
2255
2256         for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2257                 if (mdsc->sessions[i] == NULL)
2258                         continue;
2259                 s = mdsc->sessions[i];
2260                 oldstate = ceph_mdsmap_get_state(oldmap, i);
2261                 newstate = ceph_mdsmap_get_state(newmap, i);
2262
2263                 dout("check_new_map mds%d state %s -> %s (session %s)\n",
2264                      i, ceph_mds_state_name(oldstate),
2265                      ceph_mds_state_name(newstate),
2266                      session_state_name(s->s_state));
2267
2268                 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
2269                            ceph_mdsmap_get_addr(newmap, i),
2270                            sizeof(struct ceph_entity_addr))) {
2271                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2272                                 /* the session never opened, just close it
2273                                  * out now */
2274                                 __wake_requests(mdsc, &s->s_waiting);
2275                                 __unregister_session(mdsc, s);
2276                         } else {
2277                                 /* just close it */
2278                                 mutex_unlock(&mdsc->mutex);
2279                                 mutex_lock(&s->s_mutex);
2280                                 mutex_lock(&mdsc->mutex);
2281                                 ceph_con_close(&s->s_con);
2282                                 mutex_unlock(&s->s_mutex);
2283                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2284                         }
2285
2286                         /* kick any requests waiting on the recovering mds */
2287                         kick_requests(mdsc, i, 1);
2288                 } else if (oldstate == newstate) {
2289                         continue;  /* nothing new with this mds */
2290                 }
2291
2292                 /*
2293                  * send reconnect?
2294                  */
2295                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2296                     newstate >= CEPH_MDS_STATE_RECONNECT)
2297                         send_mds_reconnect(mdsc, i);
2298
2299                 /*
2300                  * kick requests on any mds that has gone active.
2301                  *
2302                  * kick requests on cur or forwarder: we may have sent
2303                  * the request to mds1, mds1 told us it forwarded it
2304                  * to mds2, but then we learn mds1 failed and can't be
2305                  * sure it successfully forwarded our request before
2306                  * it died.
2307                  */
2308                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2309                     newstate >= CEPH_MDS_STATE_ACTIVE) {
2310                         pr_info("mds%d reconnect completed\n", s->s_mds);
2311                         kick_requests(mdsc, i, 1);
2312                         ceph_kick_flushing_caps(mdsc, s);
2313                         wake_up_session_caps(s, 1);
2314                 }
2315         }
2316 }
2317
2318
2319
2320 /*
2321  * leases
2322  */
2323
2324 /*
2325  * caller must hold session s_mutex, dentry->d_lock
2326  */
2327 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2328 {
2329         struct ceph_dentry_info *di = ceph_dentry(dentry);
2330
2331         ceph_put_mds_session(di->lease_session);
2332         di->lease_session = NULL;
2333 }
2334
2335 static void handle_lease(struct ceph_mds_client *mdsc,
2336                          struct ceph_mds_session *session,
2337                          struct ceph_msg *msg)
2338 {
2339         struct super_block *sb = mdsc->client->sb;
2340         struct inode *inode;
2341         struct ceph_inode_info *ci;
2342         struct dentry *parent, *dentry;
2343         struct ceph_dentry_info *di;
2344         int mds = session->s_mds;
2345         struct ceph_mds_lease *h = msg->front.iov_base;
2346         struct ceph_vino vino;
2347         int mask;
2348         struct qstr dname;
2349         int release = 0;
2350
2351         dout("handle_lease from mds%d\n", mds);
2352
2353         /* decode */
2354         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2355                 goto bad;
2356         vino.ino = le64_to_cpu(h->ino);
2357         vino.snap = CEPH_NOSNAP;
2358         mask = le16_to_cpu(h->mask);
2359         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2360         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2361         if (dname.len != get_unaligned_le32(h+1))
2362                 goto bad;
2363
2364         mutex_lock(&session->s_mutex);
2365         session->s_seq++;
2366
2367         /* lookup inode */
2368         inode = ceph_find_inode(sb, vino);
2369         dout("handle_lease '%s', mask %d, ino %llx %p\n",
2370              ceph_lease_op_name(h->action), mask, vino.ino, inode);
2371         if (inode == NULL) {
2372                 dout("handle_lease no inode %llx\n", vino.ino);
2373                 goto release;
2374         }
2375         ci = ceph_inode(inode);
2376
2377         /* dentry */
2378         parent = d_find_alias(inode);
2379         if (!parent) {
2380                 dout("no parent dentry on inode %p\n", inode);
2381                 WARN_ON(1);
2382                 goto release;  /* hrm... */
2383         }
2384         dname.hash = full_name_hash(dname.name, dname.len);
2385         dentry = d_lookup(parent, &dname);
2386         dput(parent);
2387         if (!dentry)
2388                 goto release;
2389
2390         spin_lock(&dentry->d_lock);
2391         di = ceph_dentry(dentry);
2392         switch (h->action) {
2393         case CEPH_MDS_LEASE_REVOKE:
2394                 if (di && di->lease_session == session) {
2395                         h->seq = cpu_to_le32(di->lease_seq);
2396                         __ceph_mdsc_drop_dentry_lease(dentry);
2397                 }
2398                 release = 1;
2399                 break;
2400
2401         case CEPH_MDS_LEASE_RENEW:
2402                 if (di && di->lease_session == session &&
2403                     di->lease_gen == session->s_cap_gen &&
2404                     di->lease_renew_from &&
2405                     di->lease_renew_after == 0) {
2406                         unsigned long duration =
2407                                 le32_to_cpu(h->duration_ms) * HZ / 1000;
2408
2409                         di->lease_seq = le32_to_cpu(h->seq);
2410                         dentry->d_time = di->lease_renew_from + duration;
2411                         di->lease_renew_after = di->lease_renew_from +
2412                                 (duration >> 1);
2413                         di->lease_renew_from = 0;
2414                 }
2415                 break;
2416         }
2417         spin_unlock(&dentry->d_lock);
2418         dput(dentry);
2419
2420         if (!release)
2421                 goto out;
2422
2423 release:
2424         /* let's just reuse the same message */
2425         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2426         ceph_msg_get(msg);
2427         ceph_con_send(&session->s_con, msg);
2428
2429 out:
2430         iput(inode);
2431         mutex_unlock(&session->s_mutex);
2432         return;
2433
2434 bad:
2435         pr_err("corrupt lease message\n");
2436         ceph_msg_dump(msg);
2437 }
2438
2439 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
2440                               struct inode *inode,
2441                               struct dentry *dentry, char action,
2442                               u32 seq)
2443 {
2444         struct ceph_msg *msg;
2445         struct ceph_mds_lease *lease;
2446         int len = sizeof(*lease) + sizeof(u32);
2447         int dnamelen = 0;
2448
2449         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2450              inode, dentry, ceph_lease_op_name(action), session->s_mds);
2451         dnamelen = dentry->d_name.len;
2452         len += dnamelen;
2453
2454         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, 0, 0, NULL);
2455         if (IS_ERR(msg))
2456                 return;
2457         lease = msg->front.iov_base;
2458         lease->action = action;
2459         lease->mask = cpu_to_le16(CEPH_LOCK_DN);
2460         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
2461         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
2462         lease->seq = cpu_to_le32(seq);
2463         put_unaligned_le32(dnamelen, lease + 1);
2464         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
2465
2466         /*
2467          * if this is a preemptive lease RELEASE, no need to
2468          * flush request stream, since the actual request will
2469          * soon follow.
2470          */
2471         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
2472
2473         ceph_con_send(&session->s_con, msg);
2474 }
2475
2476 /*
2477  * Preemptively release a lease we expect to invalidate anyway.
2478  * Pass @inode always, @dentry is optional.
2479  */
2480 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2481                              struct dentry *dentry, int mask)
2482 {
2483         struct ceph_dentry_info *di;
2484         struct ceph_mds_session *session;
2485         u32 seq;
2486
2487         BUG_ON(inode == NULL);
2488         BUG_ON(dentry == NULL);
2489         BUG_ON(mask != CEPH_LOCK_DN);
2490
2491         /* is dentry lease valid? */
2492         spin_lock(&dentry->d_lock);
2493         di = ceph_dentry(dentry);
2494         if (!di || !di->lease_session ||
2495             di->lease_session->s_mds < 0 ||
2496             di->lease_gen != di->lease_session->s_cap_gen ||
2497             !time_before(jiffies, dentry->d_time)) {
2498                 dout("lease_release inode %p dentry %p -- "
2499                      "no lease on %d\n",
2500                      inode, dentry, mask);
2501                 spin_unlock(&dentry->d_lock);
2502                 return;
2503         }
2504
2505         /* we do have a lease on this dentry; note mds and seq */
2506         session = ceph_get_mds_session(di->lease_session);
2507         seq = di->lease_seq;
2508         __ceph_mdsc_drop_dentry_lease(dentry);
2509         spin_unlock(&dentry->d_lock);
2510
2511         dout("lease_release inode %p dentry %p mask %d to mds%d\n",
2512              inode, dentry, mask, session->s_mds);
2513         ceph_mdsc_lease_send_msg(session, inode, dentry,
2514                                  CEPH_MDS_LEASE_RELEASE, seq);
2515         ceph_put_mds_session(session);
2516 }
2517
2518 /*
2519  * drop all leases (and dentry refs) in preparation for umount
2520  */
2521 static void drop_leases(struct ceph_mds_client *mdsc)
2522 {
2523         int i;
2524
2525         dout("drop_leases\n");
2526         mutex_lock(&mdsc->mutex);
2527         for (i = 0; i < mdsc->max_sessions; i++) {
2528                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2529                 if (!s)
2530                         continue;
2531                 mutex_unlock(&mdsc->mutex);
2532                 mutex_lock(&s->s_mutex);
2533                 mutex_unlock(&s->s_mutex);
2534                 ceph_put_mds_session(s);
2535                 mutex_lock(&mdsc->mutex);
2536         }
2537         mutex_unlock(&mdsc->mutex);
2538 }
2539
2540
2541
2542 /*
2543  * delayed work -- periodically trim expired leases, renew caps with mds
2544  */
2545 static void schedule_delayed(struct ceph_mds_client *mdsc)
2546 {
2547         int delay = 5;
2548         unsigned hz = round_jiffies_relative(HZ * delay);
2549         schedule_delayed_work(&mdsc->delayed_work, hz);
2550 }
2551
2552 static void delayed_work(struct work_struct *work)
2553 {
2554         int i;
2555         struct ceph_mds_client *mdsc =
2556                 container_of(work, struct ceph_mds_client, delayed_work.work);
2557         int renew_interval;
2558         int renew_caps;
2559
2560         dout("mdsc delayed_work\n");
2561         ceph_check_delayed_caps(mdsc);
2562
2563         mutex_lock(&mdsc->mutex);
2564         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
2565         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
2566                                    mdsc->last_renew_caps);
2567         if (renew_caps)
2568                 mdsc->last_renew_caps = jiffies;
2569
2570         for (i = 0; i < mdsc->max_sessions; i++) {
2571                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2572                 if (s == NULL)
2573                         continue;
2574                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
2575                         dout("resending session close request for mds%d\n",
2576                              s->s_mds);
2577                         request_close_session(mdsc, s);
2578                         ceph_put_mds_session(s);
2579                         continue;
2580                 }
2581                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
2582                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
2583                                 s->s_state = CEPH_MDS_SESSION_HUNG;
2584                                 pr_info("mds%d hung\n", s->s_mds);
2585                         }
2586                 }
2587                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
2588                         /* this mds is failed or recovering, just wait */
2589                         ceph_put_mds_session(s);
2590                         continue;
2591                 }
2592                 mutex_unlock(&mdsc->mutex);
2593
2594                 mutex_lock(&s->s_mutex);
2595                 if (renew_caps)
2596                         send_renew_caps(mdsc, s);
2597                 else
2598                         ceph_con_keepalive(&s->s_con);
2599                 add_cap_releases(mdsc, s, -1);
2600                 send_cap_releases(mdsc, s);
2601                 mutex_unlock(&s->s_mutex);
2602                 ceph_put_mds_session(s);
2603
2604                 mutex_lock(&mdsc->mutex);
2605         }
2606         mutex_unlock(&mdsc->mutex);
2607
2608         schedule_delayed(mdsc);
2609 }
2610
2611
2612 int ceph_mdsc_init(struct ceph_mds_client *mdsc, struct ceph_client *client)
2613 {
2614         mdsc->client = client;
2615         mutex_init(&mdsc->mutex);
2616         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
2617         init_completion(&mdsc->safe_umount_waiters);
2618         init_completion(&mdsc->session_close_waiters);
2619         INIT_LIST_HEAD(&mdsc->waiting_for_map);
2620         mdsc->sessions = NULL;
2621         mdsc->max_sessions = 0;
2622         mdsc->stopping = 0;
2623         init_rwsem(&mdsc->snap_rwsem);
2624         mdsc->snap_realms = RB_ROOT;
2625         INIT_LIST_HEAD(&mdsc->snap_empty);
2626         spin_lock_init(&mdsc->snap_empty_lock);
2627         mdsc->last_tid = 0;
2628         mdsc->request_tree = RB_ROOT;
2629         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
2630         mdsc->last_renew_caps = jiffies;
2631         INIT_LIST_HEAD(&mdsc->cap_delay_list);
2632         spin_lock_init(&mdsc->cap_delay_lock);
2633         INIT_LIST_HEAD(&mdsc->snap_flush_list);
2634         spin_lock_init(&mdsc->snap_flush_lock);
2635         mdsc->cap_flush_seq = 0;
2636         INIT_LIST_HEAD(&mdsc->cap_dirty);
2637         mdsc->num_cap_flushing = 0;
2638         spin_lock_init(&mdsc->cap_dirty_lock);
2639         init_waitqueue_head(&mdsc->cap_flushing_wq);
2640         spin_lock_init(&mdsc->dentry_lru_lock);
2641         INIT_LIST_HEAD(&mdsc->dentry_lru);
2642         return 0;
2643 }
2644
2645 /*
2646  * Wait for safe replies on open mds requests.  If we time out, drop
2647  * all requests from the tree to avoid dangling dentry refs.
2648  */
2649 static void wait_requests(struct ceph_mds_client *mdsc)
2650 {
2651         struct ceph_mds_request *req;
2652         struct ceph_client *client = mdsc->client;
2653
2654         mutex_lock(&mdsc->mutex);
2655         if (__get_oldest_req(mdsc)) {
2656                 mutex_unlock(&mdsc->mutex);
2657
2658                 dout("wait_requests waiting for requests\n");
2659                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
2660                                     client->mount_args->mount_timeout * HZ);
2661
2662                 /* tear down remaining requests */
2663                 mutex_lock(&mdsc->mutex);
2664                 while ((req = __get_oldest_req(mdsc))) {
2665                         dout("wait_requests timed out on tid %llu\n",
2666                              req->r_tid);
2667                         __unregister_request(mdsc, req);
2668                 }
2669         }
2670         mutex_unlock(&mdsc->mutex);
2671         dout("wait_requests done\n");
2672 }
2673
2674 /*
2675  * called before mount is ro, and before dentries are torn down.
2676  * (hmm, does this still race with new lookups?)
2677  */
2678 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
2679 {
2680         dout("pre_umount\n");
2681         mdsc->stopping = 1;
2682
2683         drop_leases(mdsc);
2684         ceph_flush_dirty_caps(mdsc);
2685         wait_requests(mdsc);
2686 }
2687
2688 /*
2689  * wait for all write mds requests to flush.
2690  */
2691 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
2692 {
2693         struct ceph_mds_request *req = NULL, *nextreq;
2694         struct rb_node *n;
2695
2696         mutex_lock(&mdsc->mutex);
2697         dout("wait_unsafe_requests want %lld\n", want_tid);
2698 restart:
2699         req = __get_oldest_req(mdsc);
2700         while (req && req->r_tid <= want_tid) {
2701                 /* find next request */
2702                 n = rb_next(&req->r_node);
2703                 if (n)
2704                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
2705                 else
2706                         nextreq = NULL;
2707                 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
2708                         /* write op */
2709                         ceph_mdsc_get_request(req);
2710                         if (nextreq)
2711                                 ceph_mdsc_get_request(nextreq);
2712                         mutex_unlock(&mdsc->mutex);
2713                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
2714                              req->r_tid, want_tid);
2715                         wait_for_completion(&req->r_safe_completion);
2716                         mutex_lock(&mdsc->mutex);
2717                         ceph_mdsc_put_request(req);
2718                         if (!nextreq)
2719                                 break;  /* next dne before, so we're done! */
2720                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
2721                                 /* next request was removed from tree */
2722                                 ceph_mdsc_put_request(nextreq);
2723                                 goto restart;
2724                         }
2725                         ceph_mdsc_put_request(nextreq);  /* won't go away */
2726                 }
2727                 req = nextreq;
2728         }
2729         mutex_unlock(&mdsc->mutex);
2730         dout("wait_unsafe_requests done\n");
2731 }
2732
2733 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
2734 {
2735         u64 want_tid, want_flush;
2736
2737         dout("sync\n");
2738         mutex_lock(&mdsc->mutex);
2739         want_tid = mdsc->last_tid;
2740         want_flush = mdsc->cap_flush_seq;
2741         mutex_unlock(&mdsc->mutex);
2742         dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
2743
2744         ceph_flush_dirty_caps(mdsc);
2745
2746         wait_unsafe_requests(mdsc, want_tid);
2747         wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
2748 }
2749
2750
2751 /*
2752  * called after sb is ro.
2753  */
2754 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
2755 {
2756         struct ceph_mds_session *session;
2757         int i;
2758         int n;
2759         struct ceph_client *client = mdsc->client;
2760         unsigned long started, timeout = client->mount_args->mount_timeout * HZ;
2761
2762         dout("close_sessions\n");
2763
2764         mutex_lock(&mdsc->mutex);
2765
2766         /* close sessions */
2767         started = jiffies;
2768         while (time_before(jiffies, started + timeout)) {
2769                 dout("closing sessions\n");
2770                 n = 0;
2771                 for (i = 0; i < mdsc->max_sessions; i++) {
2772                         session = __ceph_lookup_mds_session(mdsc, i);
2773                         if (!session)
2774                                 continue;
2775                         mutex_unlock(&mdsc->mutex);
2776                         mutex_lock(&session->s_mutex);
2777                         __close_session(mdsc, session);
2778                         mutex_unlock(&session->s_mutex);
2779                         ceph_put_mds_session(session);
2780                         mutex_lock(&mdsc->mutex);
2781                         n++;
2782                 }
2783                 if (n == 0)
2784                         break;
2785
2786                 if (client->mount_state == CEPH_MOUNT_SHUTDOWN)
2787                         break;
2788
2789                 dout("waiting for sessions to close\n");
2790                 mutex_unlock(&mdsc->mutex);
2791                 wait_for_completion_timeout(&mdsc->session_close_waiters,
2792                                             timeout);
2793                 mutex_lock(&mdsc->mutex);
2794         }
2795
2796         /* tear down remaining sessions */
2797         for (i = 0; i < mdsc->max_sessions; i++) {
2798                 if (mdsc->sessions[i]) {
2799                         session = get_session(mdsc->sessions[i]);
2800                         __unregister_session(mdsc, session);
2801                         mutex_unlock(&mdsc->mutex);
2802                         mutex_lock(&session->s_mutex);
2803                         remove_session_caps(session);
2804                         mutex_unlock(&session->s_mutex);
2805                         ceph_put_mds_session(session);
2806                         mutex_lock(&mdsc->mutex);
2807                 }
2808         }
2809
2810         WARN_ON(!list_empty(&mdsc->cap_delay_list));
2811
2812         mutex_unlock(&mdsc->mutex);
2813
2814         ceph_cleanup_empty_realms(mdsc);
2815
2816         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2817
2818         dout("stopped\n");
2819 }
2820
2821 void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
2822 {
2823         dout("stop\n");
2824         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2825         if (mdsc->mdsmap)
2826                 ceph_mdsmap_destroy(mdsc->mdsmap);
2827         kfree(mdsc->sessions);
2828 }
2829
2830
2831 /*
2832  * handle mds map update.
2833  */
2834 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
2835 {
2836         u32 epoch;
2837         u32 maplen;
2838         void *p = msg->front.iov_base;
2839         void *end = p + msg->front.iov_len;
2840         struct ceph_mdsmap *newmap, *oldmap;
2841         struct ceph_fsid fsid;
2842         int err = -EINVAL;
2843
2844         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
2845         ceph_decode_copy(&p, &fsid, sizeof(fsid));
2846         if (ceph_check_fsid(mdsc->client, &fsid) < 0)
2847                 return;
2848         epoch = ceph_decode_32(&p);
2849         maplen = ceph_decode_32(&p);
2850         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
2851
2852         /* do we need it? */
2853         ceph_monc_got_mdsmap(&mdsc->client->monc, epoch);
2854         mutex_lock(&mdsc->mutex);
2855         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
2856                 dout("handle_map epoch %u <= our %u\n",
2857                      epoch, mdsc->mdsmap->m_epoch);
2858                 mutex_unlock(&mdsc->mutex);
2859                 return;
2860         }
2861
2862         newmap = ceph_mdsmap_decode(&p, end);
2863         if (IS_ERR(newmap)) {
2864                 err = PTR_ERR(newmap);
2865                 goto bad_unlock;
2866         }
2867
2868         /* swap into place */
2869         if (mdsc->mdsmap) {
2870                 oldmap = mdsc->mdsmap;
2871                 mdsc->mdsmap = newmap;
2872                 check_new_map(mdsc, newmap, oldmap);
2873                 ceph_mdsmap_destroy(oldmap);
2874         } else {
2875                 mdsc->mdsmap = newmap;  /* first mds map */
2876         }
2877         mdsc->client->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
2878
2879         __wake_requests(mdsc, &mdsc->waiting_for_map);
2880
2881         mutex_unlock(&mdsc->mutex);
2882         schedule_delayed(mdsc);
2883         return;
2884
2885 bad_unlock:
2886         mutex_unlock(&mdsc->mutex);
2887 bad:
2888         pr_err("error decoding mdsmap %d\n", err);
2889         return;
2890 }
2891
2892 static struct ceph_connection *con_get(struct ceph_connection *con)
2893 {
2894         struct ceph_mds_session *s = con->private;
2895
2896         if (get_session(s)) {
2897                 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
2898                 return con;
2899         }
2900         dout("mdsc con_get %p FAIL\n", s);
2901         return NULL;
2902 }
2903
2904 static void con_put(struct ceph_connection *con)
2905 {
2906         struct ceph_mds_session *s = con->private;
2907
2908         ceph_put_mds_session(s);
2909         dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref));
2910 }
2911
2912 /*
2913  * if the client is unresponsive for long enough, the mds will kill
2914  * the session entirely.
2915  */
2916 static void peer_reset(struct ceph_connection *con)
2917 {
2918         struct ceph_mds_session *s = con->private;
2919
2920         pr_err("mds%d gave us the boot.  IMPLEMENT RECONNECT.\n",
2921                s->s_mds);
2922 }
2923
2924 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2925 {
2926         struct ceph_mds_session *s = con->private;
2927         struct ceph_mds_client *mdsc = s->s_mdsc;
2928         int type = le16_to_cpu(msg->hdr.type);
2929
2930         mutex_lock(&mdsc->mutex);
2931         if (__verify_registered_session(mdsc, s) < 0) {
2932                 mutex_unlock(&mdsc->mutex);
2933                 goto out;
2934         }
2935         mutex_unlock(&mdsc->mutex);
2936
2937         switch (type) {
2938         case CEPH_MSG_MDS_MAP:
2939                 ceph_mdsc_handle_map(mdsc, msg);
2940                 break;
2941         case CEPH_MSG_CLIENT_SESSION:
2942                 handle_session(s, msg);
2943                 break;
2944         case CEPH_MSG_CLIENT_REPLY:
2945                 handle_reply(s, msg);
2946                 break;
2947         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
2948                 handle_forward(mdsc, s, msg);
2949                 break;
2950         case CEPH_MSG_CLIENT_CAPS:
2951                 ceph_handle_caps(s, msg);
2952                 break;
2953         case CEPH_MSG_CLIENT_SNAP:
2954                 ceph_handle_snap(mdsc, s, msg);
2955                 break;
2956         case CEPH_MSG_CLIENT_LEASE:
2957                 handle_lease(mdsc, s, msg);
2958                 break;
2959
2960         default:
2961                 pr_err("received unknown message type %d %s\n", type,
2962                        ceph_msg_type_name(type));
2963         }
2964 out:
2965         ceph_msg_put(msg);
2966 }
2967
2968 /*
2969  * authentication
2970  */
2971 static int get_authorizer(struct ceph_connection *con,
2972                           void **buf, int *len, int *proto,
2973                           void **reply_buf, int *reply_len, int force_new)
2974 {
2975         struct ceph_mds_session *s = con->private;
2976         struct ceph_mds_client *mdsc = s->s_mdsc;
2977         struct ceph_auth_client *ac = mdsc->client->monc.auth;
2978         int ret = 0;
2979
2980         if (force_new && s->s_authorizer) {
2981                 ac->ops->destroy_authorizer(ac, s->s_authorizer);
2982                 s->s_authorizer = NULL;
2983         }
2984         if (s->s_authorizer == NULL) {
2985                 if (ac->ops->create_authorizer) {
2986                         ret = ac->ops->create_authorizer(
2987                                 ac, CEPH_ENTITY_TYPE_MDS,
2988                                 &s->s_authorizer,
2989                                 &s->s_authorizer_buf,
2990                                 &s->s_authorizer_buf_len,
2991                                 &s->s_authorizer_reply_buf,
2992                                 &s->s_authorizer_reply_buf_len);
2993                         if (ret)
2994                                 return ret;
2995                 }
2996         }
2997
2998         *proto = ac->protocol;
2999         *buf = s->s_authorizer_buf;
3000         *len = s->s_authorizer_buf_len;
3001         *reply_buf = s->s_authorizer_reply_buf;
3002         *reply_len = s->s_authorizer_reply_buf_len;
3003         return 0;
3004 }
3005
3006
3007 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3008 {
3009         struct ceph_mds_session *s = con->private;
3010         struct ceph_mds_client *mdsc = s->s_mdsc;
3011         struct ceph_auth_client *ac = mdsc->client->monc.auth;
3012
3013         return ac->ops->verify_authorizer_reply(ac, s->s_authorizer, len);
3014 }
3015
3016 static int invalidate_authorizer(struct ceph_connection *con)
3017 {
3018         struct ceph_mds_session *s = con->private;
3019         struct ceph_mds_client *mdsc = s->s_mdsc;
3020         struct ceph_auth_client *ac = mdsc->client->monc.auth;
3021
3022         if (ac->ops->invalidate_authorizer)
3023                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3024
3025         return ceph_monc_validate_auth(&mdsc->client->monc);
3026 }
3027
3028 const static struct ceph_connection_operations mds_con_ops = {
3029         .get = con_get,
3030         .put = con_put,
3031         .dispatch = dispatch,
3032         .get_authorizer = get_authorizer,
3033         .verify_authorizer_reply = verify_authorizer_reply,
3034         .invalidate_authorizer = invalidate_authorizer,
3035         .peer_reset = peer_reset,
3036 };
3037
3038
3039
3040
3041 /* eof */