fs/ceph/mds_client.c

   1 #include "ceph_debug.h"
   2
   3 #include <linux/wait.h>
   4 #include <linux/slab.h>
   5 #include <linux/sched.h>
   6
   7 #include "mds_client.h"
   8 #include "mon_client.h"
   9 #include "super.h"
  10 #include "messenger.h"
  11 #include "decode.h"
  12 #include "auth.h"
  13 #include "pagelist.h"
  14
  15 /*
  16  * A cluster of MDS (metadata server) daemons is responsible for
  17  * managing the file system namespace (the directory hierarchy and
  18  * inodes) and for coordinating shared access to storage.  Metadata is
  19  * partitioning hierarchically across a number of servers, and that
  20  * partition varies over time as the cluster adjusts the distribution
  21  * in order to balance load.
  22  *
  23  * The MDS client is primarily responsible to managing synchronous
  24  * metadata requests for operations like open, unlink, and so forth.
  25  * If there is a MDS failure, we find out about it when we (possibly
  26  * request and) receive a new MDS map, and can resubmit affected
  27  * requests.
  28  *
  29  * For the most part, though, we take advantage of a lossless
  30  * communications channel to the MDS, and do not need to worry about
  31  * timing out or resubmitting requests.
  32  *
  33  * We maintain a stateful "session" with each MDS we interact with.
  34  * Within each session, we sent periodic heartbeat messages to ensure
  35  * any capabilities or leases we have been issues remain valid.  If
  36  * the session times out and goes stale, our leases and capabilities
  37  * are no longer valid.
  38  */
  39
  40 static void __wake_requests(struct ceph_mds_client *mdsc,
  41                             struct list_head *head);
  42
  43 const static struct ceph_connection_operations mds_con_ops;
  44
  45
  46 /*
  47  * mds reply parsing
  48  */
  49
  50 /*
  51  * parse individual inode info
  52  */
  53 static int parse_reply_info_in(void **p, void *end,
  54                                struct ceph_mds_reply_info_in *info)
  55 {
  56         int err = -EIO;
  57
  58         info->in = *p;
  59         *p += sizeof(struct ceph_mds_reply_inode) +
  60                 sizeof(*info->in->fragtree.splits) *
  61                 le32_to_cpu(info->in->fragtree.nsplits);
  62
  63         ceph_decode_32_safe(p, end, info->symlink_len, bad);
  64         ceph_decode_need(p, end, info->symlink_len, bad);
  65         info->symlink = *p;
  66         *p += info->symlink_len;
  67
  68         ceph_decode_32_safe(p, end, info->xattr_len, bad);
  69         ceph_decode_need(p, end, info->xattr_len, bad);
  70         info->xattr_data = *p;
  71         *p += info->xattr_len;
  72         return 0;
  73 bad:
  74         return err;
  75 }
  76
  77 /*
  78  * parse a normal reply, which may contain a (dir+)dentry and/or a
  79  * target inode.
  80  */
  81 static int parse_reply_info_trace(void **p, void *end,
  82                                   struct ceph_mds_reply_info_parsed *info)
  83 {
  84         int err;
  85
  86         if (info->head->is_dentry) {
  87                 err = parse_reply_info_in(p, end, &info->diri);
  88                 if (err < 0)
  89                         goto out_bad;
  90
  91                 if (unlikely(*p + sizeof(*info->dirfrag) > end))
  92                         goto bad;
  93                 info->dirfrag = *p;
  94                 *p += sizeof(*info->dirfrag) +
  95                         sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
  96                 if (unlikely(*p > end))
  97                         goto bad;
  98
  99                 ceph_decode_32_safe(p, end, info->dname_len, bad);
 100                 ceph_decode_need(p, end, info->dname_len, bad);
 101                 info->dname = *p;
 102                 *p += info->dname_len;
 103                 info->dlease = *p;
 104                 *p += sizeof(*info->dlease);
 105         }
 106
 107         if (info->head->is_target) {
 108                 err = parse_reply_info_in(p, end, &info->targeti);
 109                 if (err < 0)
 110                         goto out_bad;
 111         }
 112
 113         if (unlikely(*p != end))
 114                 goto bad;
 115         return 0;
 116
 117 bad:
 118         err = -EIO;
 119 out_bad:
 120         pr_err("problem parsing mds trace %d\n", err);
 121         return err;
 122 }
 123
 124 /*
 125  * parse readdir results
 126  */
 127 static int parse_reply_info_dir(void **p, void *end,
 128                                 struct ceph_mds_reply_info_parsed *info)
 129 {
 130         u32 num, i = 0;
 131         int err;
 132
 133         info->dir_dir = *p;
 134         if (*p + sizeof(*info->dir_dir) > end)
 135                 goto bad;
 136         *p += sizeof(*info->dir_dir) +
 137                 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
 138         if (*p > end)
 139                 goto bad;
 140
 141         ceph_decode_need(p, end, sizeof(num) + 2, bad);
 142         num = ceph_decode_32(p);
 143         info->dir_end = ceph_decode_8(p);
 144         info->dir_complete = ceph_decode_8(p);
 145         if (num == 0)
 146                 goto done;
 147
 148         /* alloc large array */
 149         info->dir_nr = num;
 150         info->dir_in = kcalloc(num, sizeof(*info->dir_in) +
 151                                sizeof(*info->dir_dname) +
 152                                sizeof(*info->dir_dname_len) +
 153                                sizeof(*info->dir_dlease),
 154                                GFP_NOFS);
 155         if (info->dir_in == NULL) {
 156                 err = -ENOMEM;
 157                 goto out_bad;
 158         }
 159         info->dir_dname = (void *)(info->dir_in + num);
 160         info->dir_dname_len = (void *)(info->dir_dname + num);
 161         info->dir_dlease = (void *)(info->dir_dname_len + num);
 162
 163         while (num) {
 164                 /* dentry */
 165                 ceph_decode_need(p, end, sizeof(u32)*2, bad);
 166                 info->dir_dname_len[i] = ceph_decode_32(p);
 167                 ceph_decode_need(p, end, info->dir_dname_len[i], bad);
 168                 info->dir_dname[i] = *p;
 169                 *p += info->dir_dname_len[i];
 170                 dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
 171                      info->dir_dname[i]);
 172                 info->dir_dlease[i] = *p;
 173                 *p += sizeof(struct ceph_mds_reply_lease);
 174
 175                 /* inode */
 176                 err = parse_reply_info_in(p, end, &info->dir_in[i]);
 177                 if (err < 0)
 178                         goto out_bad;
 179                 i++;
 180                 num--;
 181         }
 182
 183 done:
 184         if (*p != end)
 185                 goto bad;
 186         return 0;
 187
 188 bad:
 189         err = -EIO;
 190 out_bad:
 191         pr_err("problem parsing dir contents %d\n", err);
 192         return err;
 193 }
 194
 195 /*
 196  * parse entire mds reply
 197  */
 198 static int parse_reply_info(struct ceph_msg *msg,
 199                             struct ceph_mds_reply_info_parsed *info)
 200 {
 201         void *p, *end;
 202         u32 len;
 203         int err;
 204
 205         info->head = msg->front.iov_base;
 206         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 207         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 208
 209         /* trace */
 210         ceph_decode_32_safe(&p, end, len, bad);
 211         if (len > 0) {
 212                 err = parse_reply_info_trace(&p, p+len, info);
 213                 if (err < 0)
 214                         goto out_bad;
 215         }
 216
 217         /* dir content */
 218         ceph_decode_32_safe(&p, end, len, bad);
 219         if (len > 0) {
 220                 err = parse_reply_info_dir(&p, p+len, info);
 221                 if (err < 0)
 222                         goto out_bad;
 223         }
 224
 225         /* snap blob */
 226         ceph_decode_32_safe(&p, end, len, bad);
 227         info->snapblob_len = len;
 228         info->snapblob = p;
 229         p += len;
 230
 231         if (p != end)
 232                 goto bad;
 233         return 0;
 234
 235 bad:
 236         err = -EIO;
 237 out_bad:
 238         pr_err("mds parse_reply err %d\n", err);
 239         return err;
 240 }
 241
 242 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
 243 {
 244         kfree(info->dir_in);
 245 }
 246
 247
 248 /*
 249  * sessions
 250  */
 251 static const char *session_state_name(int s)
 252 {
 253         switch (s) {
 254         case CEPH_MDS_SESSION_NEW: return "new";
 255         case CEPH_MDS_SESSION_OPENING: return "opening";
 256         case CEPH_MDS_SESSION_OPEN: return "open";
 257         case CEPH_MDS_SESSION_HUNG: return "hung";
 258         case CEPH_MDS_SESSION_CLOSING: return "closing";
 259         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
 260         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
 261         default: return "???";
 262         }
 263 }
 264
 265 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
 266 {
 267         if (atomic_inc_not_zero(&s->s_ref)) {
 268                 dout("mdsc get_session %p %d -> %d\n", s,
 269                      atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
 270                 return s;
 271         } else {
 272                 dout("mdsc get_session %p 0 -- FAIL", s);
 273                 return NULL;
 274         }
 275 }
 276
 277 void ceph_put_mds_session(struct ceph_mds_session *s)
 278 {
 279         dout("mdsc put_session %p %d -> %d\n", s,
 280              atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
 281         if (atomic_dec_and_test(&s->s_ref)) {
 282                 if (s->s_authorizer)
 283                         s->s_mdsc->client->monc.auth->ops->destroy_authorizer(
 284                                 s->s_mdsc->client->monc.auth, s->s_authorizer);
 285                 kfree(s);
 286         }
 287 }
 288
 289 /*
 290  * called under mdsc->mutex
 291  */
 292 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
 293                                                    int mds)
 294 {
 295         struct ceph_mds_session *session;
 296
 297         if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
 298                 return NULL;
 299         session = mdsc->sessions[mds];
 300         dout("lookup_mds_session %p %d\n", session,
 301              atomic_read(&session->s_ref));
 302         get_session(session);
 303         return session;
 304 }
 305
 306 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
 307 {
 308         if (mds >= mdsc->max_sessions)
 309                 return false;
 310         return mdsc->sessions[mds];
 311 }
 312
 313 static int __verify_registered_session(struct ceph_mds_client *mdsc,
 314                                        struct ceph_mds_session *s)
 315 {
 316         if (s->s_mds >= mdsc->max_sessions ||
 317             mdsc->sessions[s->s_mds] != s)
 318                 return -ENOENT;
 319         return 0;
 320 }
 321
 322 /*
 323  * create+register a new session for given mds.
 324  * called under mdsc->mutex.
 325  */
 326 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 327                                                  int mds)
 328 {
 329         struct ceph_mds_session *s;
 330
 331         s = kzalloc(sizeof(*s), GFP_NOFS);
 332         if (!s)
 333                 return ERR_PTR(-ENOMEM);
 334         s->s_mdsc = mdsc;
 335         s->s_mds = mds;
 336         s->s_state = CEPH_MDS_SESSION_NEW;
 337         s->s_ttl = 0;
 338         s->s_seq = 0;
 339         mutex_init(&s->s_mutex);
 340
 341         ceph_con_init(mdsc->client->msgr, &s->s_con);
 342         s->s_con.private = s;
 343         s->s_con.ops = &mds_con_ops;
 344         s->s_con.peer_name.type = CEPH_ENTITY_TYPE_MDS;
 345         s->s_con.peer_name.num = cpu_to_le64(mds);
 346
 347         spin_lock_init(&s->s_cap_lock);
 348         s->s_cap_gen = 0;
 349         s->s_cap_ttl = 0;
 350         s->s_renew_requested = 0;
 351         s->s_renew_seq = 0;
 352         INIT_LIST_HEAD(&s->s_caps);
 353         s->s_nr_caps = 0;
 354         s->s_trim_caps = 0;
 355         atomic_set(&s->s_ref, 1);
 356         INIT_LIST_HEAD(&s->s_waiting);
 357         INIT_LIST_HEAD(&s->s_unsafe);
 358         s->s_num_cap_releases = 0;
 359         s->s_cap_iterator = NULL;
 360         INIT_LIST_HEAD(&s->s_cap_releases);
 361         INIT_LIST_HEAD(&s->s_cap_releases_done);
 362         INIT_LIST_HEAD(&s->s_cap_flushing);
 363         INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
 364
 365         dout("register_session mds%d\n", mds);
 366         if (mds >= mdsc->max_sessions) {
 367                 int newmax = 1 << get_count_order(mds+1);
 368                 struct ceph_mds_session **sa;
 369
 370                 dout("register_session realloc to %d\n", newmax);
 371                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 372                 if (sa == NULL)
 373                         goto fail_realloc;
 374                 if (mdsc->sessions) {
 375                         memcpy(sa, mdsc->sessions,
 376                                mdsc->max_sessions * sizeof(void *));
 377                         kfree(mdsc->sessions);
 378                 }
 379                 mdsc->sessions = sa;
 380                 mdsc->max_sessions = newmax;
 381         }
 382         mdsc->sessions[mds] = s;
 383         atomic_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
 384
 385         ceph_con_open(&s->s_con, ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
 386
 387         return s;
 388
 389 fail_realloc:
 390         kfree(s);
 391         return ERR_PTR(-ENOMEM);
 392 }
 393
 394 /*
 395  * called under mdsc->mutex
 396  */
 397 static void __unregister_session(struct ceph_mds_client *mdsc,
 398                                struct ceph_mds_session *s)
 399 {
 400         dout("__unregister_session mds%d %p\n", s->s_mds, s);
 401         BUG_ON(mdsc->sessions[s->s_mds] != s);
 402         mdsc->sessions[s->s_mds] = NULL;
 403         ceph_con_close(&s->s_con);
 404         ceph_put_mds_session(s);
 405 }
 406
 407 /*
 408  * drop session refs in request.
 409  *
 410  * should be last request ref, or hold mdsc->mutex
 411  */
 412 static void put_request_session(struct ceph_mds_request *req)
 413 {
 414         if (req->r_session) {
 415                 ceph_put_mds_session(req->r_session);
 416                 req->r_session = NULL;
 417         }
 418 }
 419
 420 void ceph_mdsc_release_request(struct kref *kref)
 421 {
 422         struct ceph_mds_request *req = container_of(kref,
 423                                                     struct ceph_mds_request,
 424                                                     r_kref);
 425         if (req->r_request)
 426                 ceph_msg_put(req->r_request);
 427         if (req->r_reply) {
 428                 ceph_msg_put(req->r_reply);
 429                 destroy_reply_info(&req->r_reply_info);
 430         }
 431         if (req->r_inode) {
 432                 ceph_put_cap_refs(ceph_inode(req->r_inode),
 433                                   CEPH_CAP_PIN);
 434                 iput(req->r_inode);
 435         }
 436         if (req->r_locked_dir)
 437                 ceph_put_cap_refs(ceph_inode(req->r_locked_dir),
 438                                   CEPH_CAP_PIN);
 439         if (req->r_target_inode)
 440                 iput(req->r_target_inode);
 441         if (req->r_dentry)
 442                 dput(req->r_dentry);
 443         if (req->r_old_dentry) {
 444                 ceph_put_cap_refs(
 445                         ceph_inode(req->r_old_dentry->d_parent->d_inode),
 446                         CEPH_CAP_PIN);
 447                 dput(req->r_old_dentry);
 448         }
 449         kfree(req->r_path1);
 450         kfree(req->r_path2);
 451         put_request_session(req);
 452         ceph_unreserve_caps(&req->r_caps_reservation);
 453         kfree(req);
 454 }
 455
 456 /*
 457  * lookup session, bump ref if found.
 458  *
 459  * called under mdsc->mutex.
 460  */
 461 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
 462                                              u64 tid)
 463 {
 464         struct ceph_mds_request *req;
 465         struct rb_node *n = mdsc->request_tree.rb_node;
 466
 467         while (n) {
 468                 req = rb_entry(n, struct ceph_mds_request, r_node);
 469                 if (tid < req->r_tid)
 470                         n = n->rb_left;
 471                 else if (tid > req->r_tid)
 472                         n = n->rb_right;
 473                 else {
 474                         ceph_mdsc_get_request(req);
 475                         return req;
 476                 }
 477         }
 478         return NULL;
 479 }
 480
 481 static void __insert_request(struct ceph_mds_client *mdsc,
 482                              struct ceph_mds_request *new)
 483 {
 484         struct rb_node **p = &mdsc->request_tree.rb_node;
 485         struct rb_node *parent = NULL;
 486         struct ceph_mds_request *req = NULL;
 487
 488         while (*p) {
 489                 parent = *p;
 490                 req = rb_entry(parent, struct ceph_mds_request, r_node);
 491                 if (new->r_tid < req->r_tid)
 492                         p = &(*p)->rb_left;
 493                 else if (new->r_tid > req->r_tid)
 494                         p = &(*p)->rb_right;
 495                 else
 496                         BUG();
 497         }
 498
 499         rb_link_node(&new->r_node, parent, p);
 500         rb_insert_color(&new->r_node, &mdsc->request_tree);
 501 }
 502
 503 /*
 504  * Register an in-flight request, and assign a tid.  Link to directory
 505  * are modifying (if any).
 506  *
 507  * Called under mdsc->mutex.
 508  */
 509 static void __register_request(struct ceph_mds_client *mdsc,
 510                                struct ceph_mds_request *req,
 511                                struct inode *dir)
 512 {
 513         req->r_tid = ++mdsc->last_tid;
 514         if (req->r_num_caps)
 515                 ceph_reserve_caps(&req->r_caps_reservation, req->r_num_caps);
 516         dout("__register_request %p tid %lld\n", req, req->r_tid);
 517         ceph_mdsc_get_request(req);
 518         __insert_request(mdsc, req);
 519
 520         if (dir) {
 521                 struct ceph_inode_info *ci = ceph_inode(dir);
 522
 523                 spin_lock(&ci->i_unsafe_lock);
 524                 req->r_unsafe_dir = dir;
 525                 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
 526                 spin_unlock(&ci->i_unsafe_lock);
 527         }
 528 }
 529
 530 static void __unregister_request(struct ceph_mds_client *mdsc,
 531                                  struct ceph_mds_request *req)
 532 {
 533         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 534         rb_erase(&req->r_node, &mdsc->request_tree);
 535         RB_CLEAR_NODE(&req->r_node);
 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         ceph_mdsc_put_request(req);
 546 }
 547
 548 /*
 549  * Choose mds to send request to next.  If there is a hint set in the
 550  * request (e.g., due to a prior forward hint from the mds), use that.
 551  * Otherwise, consult frag tree and/or caps to identify the
 552  * appropriate mds.  If all else fails, choose randomly.
 553  *
 554  * Called under mdsc->mutex.
 555  */
 556 static int __choose_mds(struct ceph_mds_client *mdsc,
 557                         struct ceph_mds_request *req)
 558 {
 559         struct inode *inode;
 560         struct ceph_inode_info *ci;
 561         struct ceph_cap *cap;
 562         int mode = req->r_direct_mode;
 563         int mds = -1;
 564         u32 hash = req->r_direct_hash;
 565         bool is_hash = req->r_direct_is_hash;
 566
 567         /*
 568          * is there a specific mds we should try?  ignore hint if we have
 569          * no session and the mds is not up (active or recovering).
 570          */
 571         if (req->r_resend_mds >= 0 &&
 572             (__have_session(mdsc, req->r_resend_mds) ||
 573              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 574                 dout("choose_mds using resend_mds mds%d\n",
 575                      req->r_resend_mds);
 576                 return req->r_resend_mds;
 577         }
 578
 579         if (mode == USE_RANDOM_MDS)
 580                 goto random;
 581
 582         inode = NULL;
 583         if (req->r_inode) {
 584                 inode = req->r_inode;
 585         } else if (req->r_dentry) {
 586                 if (req->r_dentry->d_inode) {
 587                         inode = req->r_dentry->d_inode;
 588                 } else {
 589                         inode = req->r_dentry->d_parent->d_inode;
 590                         hash = req->r_dentry->d_name.hash;
 591                         is_hash = true;
 592                 }
 593         }
 594         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 595              (int)hash, mode);
 596         if (!inode)
 597                 goto random;
 598         ci = ceph_inode(inode);
 599
 600         if (is_hash && S_ISDIR(inode->i_mode)) {
 601                 struct ceph_inode_frag frag;
 602                 int found;
 603
 604                 ceph_choose_frag(ci, hash, &frag, &found);
 605                 if (found) {
 606                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 607                                 u8 r;
 608
 609                                 /* choose a random replica */
 610                                 get_random_bytes(&r, 1);
 611                                 r %= frag.ndist;
 612                                 mds = frag.dist[r];
 613                                 dout("choose_mds %p %llx.%llx "
 614                                      "frag %u mds%d (%d/%d)\n",
 615                                      inode, ceph_vinop(inode),
 616                                      frag.frag, frag.mds,
 617                                      (int)r, frag.ndist);
 618                                 return mds;
 619                         }
 620
 621                         /* since this file/dir wasn't known to be
 622                          * replicated, then we want to look for the
 623                          * authoritative mds. */
 624                         mode = USE_AUTH_MDS;
 625                         if (frag.mds >= 0) {
 626                                 /* choose auth mds */
 627                                 mds = frag.mds;
 628                                 dout("choose_mds %p %llx.%llx "
 629                                      "frag %u mds%d (auth)\n",
 630                                      inode, ceph_vinop(inode), frag.frag, mds);
 631                                 return mds;
 632                         }
 633                 }
 634         }
 635
 636         spin_lock(&inode->i_lock);
 637         cap = NULL;
 638         if (mode == USE_AUTH_MDS)
 639                 cap = ci->i_auth_cap;
 640         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
 641                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
 642         if (!cap) {
 643                 spin_unlock(&inode->i_lock);
 644                 goto random;
 645         }
 646         mds = cap->session->s_mds;
 647         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
 648              inode, ceph_vinop(inode), mds,
 649              cap == ci->i_auth_cap ? "auth " : "", cap);
 650         spin_unlock(&inode->i_lock);
 651         return mds;
 652
 653 random:
 654         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
 655         dout("choose_mds chose random mds%d\n", mds);
 656         return mds;
 657 }
 658
 659
 660 /*
 661  * session messages
 662  */
 663 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
 664 {
 665         struct ceph_msg *msg;
 666         struct ceph_mds_session_head *h;
 667
 668         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h));
 669         if (!msg) {
 670                 pr_err("create_session_msg ENOMEM creating msg\n");
 671                 return NULL;
 672         }
 673         h = msg->front.iov_base;
 674         h->op = cpu_to_le32(op);
 675         h->seq = cpu_to_le64(seq);
 676         return msg;
 677 }
 678
 679 /*
 680  * send session open request.
 681  *
 682  * called under mdsc->mutex
 683  */
 684 static int __open_session(struct ceph_mds_client *mdsc,
 685                           struct ceph_mds_session *session)
 686 {
 687         struct ceph_msg *msg;
 688         int mstate;
 689         int mds = session->s_mds;
 690
 691         /* wait for mds to go active? */
 692         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
 693         dout("open_session to mds%d (%s)\n", mds,
 694              ceph_mds_state_name(mstate));
 695         session->s_state = CEPH_MDS_SESSION_OPENING;
 696         session->s_renew_requested = jiffies;
 697
 698         /* send connect message */
 699         msg = create_session_msg(CEPH_SESSION_REQUEST_OPEN, session->s_seq);
 700         if (!msg)
 701                 return -ENOMEM;
 702         ceph_con_send(&session->s_con, msg);
 703         return 0;
 704 }
 705
 706 /*
 707  * session caps
 708  */
 709
 710 /*
 711  * Free preallocated cap messages assigned to this session
 712  */
 713 static void cleanup_cap_releases(struct ceph_mds_session *session)
 714 {
 715         struct ceph_msg *msg;
 716
 717         spin_lock(&session->s_cap_lock);
 718         while (!list_empty(&session->s_cap_releases)) {
 719                 msg = list_first_entry(&session->s_cap_releases,
 720                                        struct ceph_msg, list_head);
 721                 list_del_init(&msg->list_head);
 722                 ceph_msg_put(msg);
 723         }
 724         while (!list_empty(&session->s_cap_releases_done)) {
 725                 msg = list_first_entry(&session->s_cap_releases_done,
 726                                        struct ceph_msg, list_head);
 727                 list_del_init(&msg->list_head);
 728                 ceph_msg_put(msg);
 729         }
 730         spin_unlock(&session->s_cap_lock);
 731 }
 732
 733 /*
 734  * Helper to safely iterate over all caps associated with a session, with
 735  * special care taken to handle a racing __ceph_remove_cap().
 736  *
 737  * Caller must hold session s_mutex.
 738  */
 739 static int iterate_session_caps(struct ceph_mds_session *session,
 740                                  int (*cb)(struct inode *, struct ceph_cap *,
 741                                             void *), void *arg)
 742 {
 743         struct list_head *p;
 744         struct ceph_cap *cap;
 745         struct inode *inode, *last_inode = NULL;
 746         struct ceph_cap *old_cap = NULL;
 747         int ret;
 748
 749         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
 750         spin_lock(&session->s_cap_lock);
 751         p = session->s_caps.next;
 752         while (p != &session->s_caps) {
 753                 cap = list_entry(p, struct ceph_cap, session_caps);
 754                 inode = igrab(&cap->ci->vfs_inode);
 755                 if (!inode) {
 756                         p = p->next;
 757                         continue;
 758                 }
 759                 session->s_cap_iterator = cap;
 760                 spin_unlock(&session->s_cap_lock);
 761
 762                 if (last_inode) {
 763                         iput(last_inode);
 764                         last_inode = NULL;
 765                 }
 766                 if (old_cap) {
 767                         ceph_put_cap(old_cap);
 768                         old_cap = NULL;
 769                 }
 770
 771                 ret = cb(inode, cap, arg);
 772                 last_inode = inode;
 773
 774                 spin_lock(&session->s_cap_lock);
 775                 p = p->next;
 776                 if (cap->ci == NULL) {
 777                         dout("iterate_session_caps  finishing cap %p removal\n",
 778                              cap);
 779                         BUG_ON(cap->session != session);
 780                         list_del_init(&cap->session_caps);
 781                         session->s_nr_caps--;
 782                         cap->session = NULL;
 783                         old_cap = cap;  /* put_cap it w/o locks held */
 784                 }
 785                 if (ret < 0)
 786                         goto out;
 787         }
 788         ret = 0;
 789 out:
 790         session->s_cap_iterator = NULL;
 791         spin_unlock(&session->s_cap_lock);
 792
 793         if (last_inode)
 794                 iput(last_inode);
 795         if (old_cap)
 796                 ceph_put_cap(old_cap);
 797
 798         return ret;
 799 }
 800
 801 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
 802                                   void *arg)
 803 {
 804         struct ceph_inode_info *ci = ceph_inode(inode);
 805         int drop = 0;
 806
 807         dout("removing cap %p, ci is %p, inode is %p\n",
 808              cap, ci, &ci->vfs_inode);
 809         spin_lock(&inode->i_lock);
 810         __ceph_remove_cap(cap);
 811         if (!__ceph_is_any_real_caps(ci)) {
 812                 struct ceph_mds_client *mdsc =
 813                         &ceph_sb_to_client(inode->i_sb)->mdsc;
 814
 815                 spin_lock(&mdsc->cap_dirty_lock);
 816                 if (!list_empty(&ci->i_dirty_item)) {
 817                         pr_info(" dropping dirty %s state for %p %lld\n",
 818                                 ceph_cap_string(ci->i_dirty_caps),
 819                                 inode, ceph_ino(inode));
 820                         ci->i_dirty_caps = 0;
 821                         list_del_init(&ci->i_dirty_item);
 822                         drop = 1;
 823                 }
 824                 if (!list_empty(&ci->i_flushing_item)) {
 825                         pr_info(" dropping dirty+flushing %s state for %p %lld\n",
 826                                 ceph_cap_string(ci->i_flushing_caps),
 827                                 inode, ceph_ino(inode));
 828                         ci->i_flushing_caps = 0;
 829                         list_del_init(&ci->i_flushing_item);
 830                         mdsc->num_cap_flushing--;
 831                         drop = 1;
 832                 }
 833                 if (drop && ci->i_wrbuffer_ref) {
 834                         pr_info(" dropping dirty data for %p %lld\n",
 835                                 inode, ceph_ino(inode));
 836                         ci->i_wrbuffer_ref = 0;
 837                         ci->i_wrbuffer_ref_head = 0;
 838                         drop++;
 839                 }
 840                 spin_unlock(&mdsc->cap_dirty_lock);
 841         }
 842         spin_unlock(&inode->i_lock);
 843         while (drop--)
 844                 iput(inode);
 845         return 0;
 846 }
 847
 848 /*
 849  * caller must hold session s_mutex
 850  */
 851 static void remove_session_caps(struct ceph_mds_session *session)
 852 {
 853         dout("remove_session_caps on %p\n", session);
 854         iterate_session_caps(session, remove_session_caps_cb, NULL);
 855         BUG_ON(session->s_nr_caps > 0);
 856         BUG_ON(!list_empty(&session->s_cap_flushing));
 857         cleanup_cap_releases(session);
 858 }
 859
 860 /*
 861  * wake up any threads waiting on this session's caps.  if the cap is
 862  * old (didn't get renewed on the client reconnect), remove it now.
 863  *
 864  * caller must hold s_mutex.
 865  */
 866 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
 867                               void *arg)
 868 {
 869         struct ceph_inode_info *ci = ceph_inode(inode);
 870
 871         wake_up(&ci->i_cap_wq);
 872         if (arg) {
 873                 spin_lock(&inode->i_lock);
 874                 ci->i_wanted_max_size = 0;
 875                 ci->i_requested_max_size = 0;
 876                 spin_unlock(&inode->i_lock);
 877         }
 878         return 0;
 879 }
 880
 881 static void wake_up_session_caps(struct ceph_mds_session *session,
 882                                  int reconnect)
 883 {
 884         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
 885         iterate_session_caps(session, wake_up_session_cb,
 886                              (void *)(unsigned long)reconnect);
 887 }
 888
 889 /*
 890  * Send periodic message to MDS renewing all currently held caps.  The
 891  * ack will reset the expiration for all caps from this session.
 892  *
 893  * caller holds s_mutex
 894  */
 895 static int send_renew_caps(struct ceph_mds_client *mdsc,
 896                            struct ceph_mds_session *session)
 897 {
 898         struct ceph_msg *msg;
 899         int state;
 900
 901         if (time_after_eq(jiffies, session->s_cap_ttl) &&
 902             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
 903                 pr_info("mds%d caps stale\n", session->s_mds);
 904         session->s_renew_requested = jiffies;
 905
 906         /* do not try to renew caps until a recovering mds has reconnected
 907          * with its clients. */
 908         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
 909         if (state < CEPH_MDS_STATE_RECONNECT) {
 910                 dout("send_renew_caps ignoring mds%d (%s)\n",
 911                      session->s_mds, ceph_mds_state_name(state));
 912                 return 0;
 913         }
 914
 915         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
 916                 ceph_mds_state_name(state));
 917         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
 918                                  ++session->s_renew_seq);
 919         if (!msg)
 920                 return -ENOMEM;
 921         ceph_con_send(&session->s_con, msg);
 922         return 0;
 923 }
 924
 925 /*
 926  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
 927  *
 928  * Called under session->s_mutex
 929  */
 930 static void renewed_caps(struct ceph_mds_client *mdsc,
 931                          struct ceph_mds_session *session, int is_renew)
 932 {
 933         int was_stale;
 934         int wake = 0;
 935
 936         spin_lock(&session->s_cap_lock);
 937         was_stale = is_renew && (session->s_cap_ttl == 0 ||
 938                                  time_after_eq(jiffies, session->s_cap_ttl));
 939
 940         session->s_cap_ttl = session->s_renew_requested +
 941                 mdsc->mdsmap->m_session_timeout*HZ;
 942
 943         if (was_stale) {
 944                 if (time_before(jiffies, session->s_cap_ttl)) {
 945                         pr_info("mds%d caps renewed\n", session->s_mds);
 946                         wake = 1;
 947                 } else {
 948                         pr_info("mds%d caps still stale\n", session->s_mds);
 949                 }
 950         }
 951         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
 952              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
 953              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
 954         spin_unlock(&session->s_cap_lock);
 955
 956         if (wake)
 957                 wake_up_session_caps(session, 0);
 958 }
 959
 960 /*
 961  * send a session close request
 962  */
 963 static int request_close_session(struct ceph_mds_client *mdsc,
 964                                  struct ceph_mds_session *session)
 965 {
 966         struct ceph_msg *msg;
 967
 968         dout("request_close_session mds%d state %s seq %lld\n",
 969              session->s_mds, session_state_name(session->s_state),
 970              session->s_seq);
 971         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
 972         if (!msg)
 973                 return -ENOMEM;
 974         ceph_con_send(&session->s_con, msg);
 975         return 0;
 976 }
 977
 978 /*
 979  * Called with s_mutex held.
 980  */
 981 static int __close_session(struct ceph_mds_client *mdsc,
 982                          struct ceph_mds_session *session)
 983 {
 984         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
 985                 return 0;
 986         session->s_state = CEPH_MDS_SESSION_CLOSING;
 987         return request_close_session(mdsc, session);
 988 }
 989
 990 /*
 991  * Trim old(er) caps.
 992  *
 993  * Because we can't cache an inode without one or more caps, we do
 994  * this indirectly: if a cap is unused, we prune its aliases, at which
 995  * point the inode will hopefully get dropped to.
 996  *
 997  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
 998  * memory pressure from the MDS, though, so it needn't be perfect.
 999  */
1000 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1001 {
1002         struct ceph_mds_session *session = arg;
1003         struct ceph_inode_info *ci = ceph_inode(inode);
1004         int used, oissued, mine;
1005
1006         if (session->s_trim_caps <= 0)
1007                 return -1;
1008
1009         spin_lock(&inode->i_lock);
1010         mine = cap->issued | cap->implemented;
1011         used = __ceph_caps_used(ci);
1012         oissued = __ceph_caps_issued_other(ci, cap);
1013
1014         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
1015              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1016              ceph_cap_string(used));
1017         if (ci->i_dirty_caps)
1018                 goto out;   /* dirty caps */
1019         if ((used & ~oissued) & mine)
1020                 goto out;   /* we need these caps */
1021
1022         session->s_trim_caps--;
1023         if (oissued) {
1024                 /* we aren't the only cap.. just remove us */
1025                 __ceph_remove_cap(cap);
1026         } else {
1027                 /* try to drop referring dentries */
1028                 spin_unlock(&inode->i_lock);
1029                 d_prune_aliases(inode);
1030                 dout("trim_caps_cb %p cap %p  pruned, count now %d\n",
1031                      inode, cap, atomic_read(&inode->i_count));
1032                 return 0;
1033         }
1034
1035 out:
1036         spin_unlock(&inode->i_lock);
1037         return 0;
1038 }
1039
1040 /*
1041  * Trim session cap count down to some max number.
1042  */
1043 static int trim_caps(struct ceph_mds_client *mdsc,
1044                      struct ceph_mds_session *session,
1045                      int max_caps)
1046 {
1047         int trim_caps = session->s_nr_caps - max_caps;
1048
1049         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1050              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1051         if (trim_caps > 0) {
1052                 session->s_trim_caps = trim_caps;
1053                 iterate_session_caps(session, trim_caps_cb, session);
1054                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1055                      session->s_mds, session->s_nr_caps, max_caps,
1056                         trim_caps - session->s_trim_caps);
1057                 session->s_trim_caps = 0;
1058         }
1059         return 0;
1060 }
1061
1062 /*
1063  * Allocate cap_release messages.  If there is a partially full message
1064  * in the queue, try to allocate enough to cover it's remainder, so that
1065  * we can send it immediately.
1066  *
1067  * Called under s_mutex.
1068  */
1069 static int add_cap_releases(struct ceph_mds_client *mdsc,
1070                             struct ceph_mds_session *session,
1071                             int extra)
1072 {
1073         struct ceph_msg *msg;
1074         struct ceph_mds_cap_release *head;
1075         int err = -ENOMEM;
1076
1077         if (extra < 0)
1078                 extra = mdsc->client->mount_args->cap_release_safety;
1079
1080         spin_lock(&session->s_cap_lock);
1081
1082         if (!list_empty(&session->s_cap_releases)) {
1083                 msg = list_first_entry(&session->s_cap_releases,
1084                                        struct ceph_msg,
1085                                  list_head);
1086                 head = msg->front.iov_base;
1087                 extra += CEPH_CAPS_PER_RELEASE - le32_to_cpu(head->num);
1088         }
1089
1090         while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1091                 spin_unlock(&session->s_cap_lock);
1092                 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE);
1093                 if (!msg)
1094                         goto out_unlocked;
1095                 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1096                      (int)msg->front.iov_len);
1097                 head = msg->front.iov_base;
1098                 head->num = cpu_to_le32(0);
1099                 msg->front.iov_len = sizeof(*head);
1100                 spin_lock(&session->s_cap_lock);
1101                 list_add(&msg->list_head, &session->s_cap_releases);
1102                 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1103         }
1104
1105         if (!list_empty(&session->s_cap_releases)) {
1106                 msg = list_first_entry(&session->s_cap_releases,
1107                                        struct ceph_msg,
1108                                        list_head);
1109                 head = msg->front.iov_base;
1110                 if (head->num) {
1111                         dout(" queueing non-full %p (%d)\n", msg,
1112                              le32_to_cpu(head->num));
1113                         list_move_tail(&msg->list_head,
1114                                       &session->s_cap_releases_done);
1115                         session->s_num_cap_releases -=
1116                                 CEPH_CAPS_PER_RELEASE - le32_to_cpu(head->num);
1117                 }
1118         }
1119         err = 0;
1120         spin_unlock(&session->s_cap_lock);
1121 out_unlocked:
1122         return err;
1123 }
1124
1125 /*
1126  * flush all dirty inode data to disk.
1127  *
1128  * returns true if we've flushed through want_flush_seq
1129  */
1130 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1131 {
1132         int mds, ret = 1;
1133
1134         dout("check_cap_flush want %lld\n", want_flush_seq);
1135         mutex_lock(&mdsc->mutex);
1136         for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1137                 struct ceph_mds_session *session = mdsc->sessions[mds];
1138
1139                 if (!session)
1140                         continue;
1141                 get_session(session);
1142                 mutex_unlock(&mdsc->mutex);
1143
1144                 mutex_lock(&session->s_mutex);
1145                 if (!list_empty(&session->s_cap_flushing)) {
1146                         struct ceph_inode_info *ci =
1147                                 list_entry(session->s_cap_flushing.next,
1148                                            struct ceph_inode_info,
1149                                            i_flushing_item);
1150                         struct inode *inode = &ci->vfs_inode;
1151
1152                         spin_lock(&inode->i_lock);
1153                         if (ci->i_cap_flush_seq <= want_flush_seq) {
1154                                 dout("check_cap_flush still flushing %p "
1155                                      "seq %lld <= %lld to mds%d\n", inode,
1156                                      ci->i_cap_flush_seq, want_flush_seq,
1157                                      session->s_mds);
1158                                 ret = 0;
1159                         }
1160                         spin_unlock(&inode->i_lock);
1161                 }
1162                 mutex_unlock(&session->s_mutex);
1163                 ceph_put_mds_session(session);
1164
1165                 if (!ret)
1166                         return ret;
1167                 mutex_lock(&mdsc->mutex);
1168         }
1169
1170         mutex_unlock(&mdsc->mutex);
1171         dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1172         return ret;
1173 }
1174
1175 /*
1176  * called under s_mutex
1177  */
1178 static void send_cap_releases(struct ceph_mds_client *mdsc,
1179                        struct ceph_mds_session *session)
1180 {
1181         struct ceph_msg *msg;
1182
1183         dout("send_cap_releases mds%d\n", session->s_mds);
1184         spin_lock(&session->s_cap_lock);
1185         while (!list_empty(&session->s_cap_releases_done)) {
1186                 msg = list_first_entry(&session->s_cap_releases_done,
1187                                  struct ceph_msg, list_head);
1188                 list_del_init(&msg->list_head);
1189                 spin_unlock(&session->s_cap_lock);
1190                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1191                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1192                 ceph_con_send(&session->s_con, msg);
1193                 spin_lock(&session->s_cap_lock);
1194         }
1195         spin_unlock(&session->s_cap_lock);
1196 }
1197
1198 static void discard_cap_releases(struct ceph_mds_client *mdsc,
1199                                  struct ceph_mds_session *session)
1200 {
1201         struct ceph_msg *msg;
1202         struct ceph_mds_cap_release *head;
1203         unsigned num;
1204
1205         dout("discard_cap_releases mds%d\n", session->s_mds);
1206         spin_lock(&session->s_cap_lock);
1207
1208         /* zero out the in-progress message */
1209         msg = list_first_entry(&session->s_cap_releases,
1210                                struct ceph_msg, list_head);
1211         head = msg->front.iov_base;
1212         num = le32_to_cpu(head->num);
1213         dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg, num);
1214         head->num = cpu_to_le32(0);
1215         session->s_num_cap_releases += num;
1216
1217         /* requeue completed messages */
1218         while (!list_empty(&session->s_cap_releases_done)) {
1219                 msg = list_first_entry(&session->s_cap_releases_done,
1220                                  struct ceph_msg, list_head);
1221                 list_del_init(&msg->list_head);
1222
1223                 head = msg->front.iov_base;
1224                 num = le32_to_cpu(head->num);
1225                 dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg,
1226                      num);
1227                 session->s_num_cap_releases += num;
1228                 head->num = cpu_to_le32(0);
1229                 msg->front.iov_len = sizeof(*head);
1230                 list_add(&msg->list_head, &session->s_cap_releases);
1231         }
1232
1233         spin_unlock(&session->s_cap_lock);
1234 }
1235
1236 /*
1237  * requests
1238  */
1239
1240 /*
1241  * Create an mds request.
1242  */
1243 struct ceph_mds_request *
1244 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1245 {
1246         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1247
1248         if (!req)
1249                 return ERR_PTR(-ENOMEM);
1250
1251         mutex_init(&req->r_fill_mutex);
1252         req->r_started = jiffies;
1253         req->r_resend_mds = -1;
1254         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1255         req->r_fmode = -1;
1256         kref_init(&req->r_kref);
1257         INIT_LIST_HEAD(&req->r_wait);
1258         init_completion(&req->r_completion);
1259         init_completion(&req->r_safe_completion);
1260         INIT_LIST_HEAD(&req->r_unsafe_item);
1261
1262         req->r_op = op;
1263         req->r_direct_mode = mode;
1264         return req;
1265 }
1266
1267 /*
1268  * return oldest (lowest) request, tid in request tree, 0 if none.
1269  *
1270  * called under mdsc->mutex.
1271  */
1272 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1273 {
1274         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1275                 return NULL;
1276         return rb_entry(rb_first(&mdsc->request_tree),
1277                         struct ceph_mds_request, r_node);
1278 }
1279
1280 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1281 {
1282         struct ceph_mds_request *req = __get_oldest_req(mdsc);
1283
1284         if (req)
1285                 return req->r_tid;
1286         return 0;
1287 }
1288
1289 /*
1290  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1291  * on build_path_from_dentry in fs/cifs/dir.c.
1292  *
1293  * If @stop_on_nosnap, generate path relative to the first non-snapped
1294  * inode.
1295  *
1296  * Encode hidden .snap dirs as a double /, i.e.
1297  *   foo/.snap/bar -> foo//bar
1298  */
1299 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1300                            int stop_on_nosnap)
1301 {
1302         struct dentry *temp;
1303         char *path;
1304         int len, pos;
1305
1306         if (dentry == NULL)
1307                 return ERR_PTR(-EINVAL);
1308
1309 retry:
1310         len = 0;
1311         for (temp = dentry; !IS_ROOT(temp);) {
1312                 struct inode *inode = temp->d_inode;
1313                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1314                         len++;  /* slash only */
1315                 else if (stop_on_nosnap && inode &&
1316                          ceph_snap(inode) == CEPH_NOSNAP)
1317                         break;
1318                 else
1319                         len += 1 + temp->d_name.len;
1320                 temp = temp->d_parent;
1321                 if (temp == NULL) {
1322                         pr_err("build_path corrupt dentry %p\n", dentry);
1323                         return ERR_PTR(-EINVAL);
1324                 }
1325         }
1326         if (len)
1327                 len--;  /* no leading '/' */
1328
1329         path = kmalloc(len+1, GFP_NOFS);
1330         if (path == NULL)
1331                 return ERR_PTR(-ENOMEM);
1332         pos = len;
1333         path[pos] = 0;  /* trailing null */
1334         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1335                 struct inode *inode = temp->d_inode;
1336
1337                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1338                         dout("build_path path+%d: %p SNAPDIR\n",
1339                              pos, temp);
1340                 } else if (stop_on_nosnap && inode &&
1341                            ceph_snap(inode) == CEPH_NOSNAP) {
1342                         break;
1343                 } else {
1344                         pos -= temp->d_name.len;
1345                         if (pos < 0)
1346                                 break;
1347                         strncpy(path + pos, temp->d_name.name,
1348                                 temp->d_name.len);
1349                 }
1350                 if (pos)
1351                         path[--pos] = '/';
1352                 temp = temp->d_parent;
1353                 if (temp == NULL) {
1354                         pr_err("build_path corrupt dentry\n");
1355                         kfree(path);
1356                         return ERR_PTR(-EINVAL);
1357                 }
1358         }
1359         if (pos != 0) {
1360                 pr_err("build_path did not end path lookup where "
1361                        "expected, namelen is %d, pos is %d\n", len, pos);
1362                 /* presumably this is only possible if racing with a
1363                    rename of one of the parent directories (we can not
1364                    lock the dentries above us to prevent this, but
1365                    retrying should be harmless) */
1366                 kfree(path);
1367                 goto retry;
1368         }
1369
1370         *base = ceph_ino(temp->d_inode);
1371         *plen = len;
1372         dout("build_path on %p %d built %llx '%.*s'\n",
1373              dentry, atomic_read(&dentry->d_count), *base, len, path);
1374         return path;
1375 }
1376
1377 static int build_dentry_path(struct dentry *dentry,
1378                              const char **ppath, int *ppathlen, u64 *pino,
1379                              int *pfreepath)
1380 {
1381         char *path;
1382
1383         if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1384                 *pino = ceph_ino(dentry->d_parent->d_inode);
1385                 *ppath = dentry->d_name.name;
1386                 *ppathlen = dentry->d_name.len;
1387                 return 0;
1388         }
1389         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1390         if (IS_ERR(path))
1391                 return PTR_ERR(path);
1392         *ppath = path;
1393         *pfreepath = 1;
1394         return 0;
1395 }
1396
1397 static int build_inode_path(struct inode *inode,
1398                             const char **ppath, int *ppathlen, u64 *pino,
1399                             int *pfreepath)
1400 {
1401         struct dentry *dentry;
1402         char *path;
1403
1404         if (ceph_snap(inode) == CEPH_NOSNAP) {
1405                 *pino = ceph_ino(inode);
1406                 *ppathlen = 0;
1407                 return 0;
1408         }
1409         dentry = d_find_alias(inode);
1410         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1411         dput(dentry);
1412         if (IS_ERR(path))
1413                 return PTR_ERR(path);
1414         *ppath = path;
1415         *pfreepath = 1;
1416         return 0;
1417 }
1418
1419 /*
1420  * request arguments may be specified via an inode *, a dentry *, or
1421  * an explicit ino+path.
1422  */
1423 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1424                                   const char *rpath, u64 rino,
1425                                   const char **ppath, int *pathlen,
1426                                   u64 *ino, int *freepath)
1427 {
1428         int r = 0;
1429
1430         if (rinode) {
1431                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1432                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1433                      ceph_snap(rinode));
1434         } else if (rdentry) {
1435                 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1436                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1437                      *ppath);
1438         } else if (rpath) {
1439                 *ino = rino;
1440                 *ppath = rpath;
1441                 *pathlen = strlen(rpath);
1442                 dout(" path %.*s\n", *pathlen, rpath);
1443         }
1444
1445         return r;
1446 }
1447
1448 /*
1449  * called under mdsc->mutex
1450  */
1451 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1452                                                struct ceph_mds_request *req,
1453                                                int mds)
1454 {
1455         struct ceph_msg *msg;
1456         struct ceph_mds_request_head *head;
1457         const char *path1 = NULL;
1458         const char *path2 = NULL;
1459         u64 ino1 = 0, ino2 = 0;
1460         int pathlen1 = 0, pathlen2 = 0;
1461         int freepath1 = 0, freepath2 = 0;
1462         int len;
1463         u16 releases;
1464         void *p, *end;
1465         int ret;
1466
1467         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1468                               req->r_path1, req->r_ino1.ino,
1469                               &path1, &pathlen1, &ino1, &freepath1);
1470         if (ret < 0) {
1471                 msg = ERR_PTR(ret);
1472                 goto out;
1473         }
1474
1475         ret = set_request_path_attr(NULL, req->r_old_dentry,
1476                               req->r_path2, req->r_ino2.ino,
1477                               &path2, &pathlen2, &ino2, &freepath2);
1478         if (ret < 0) {
1479                 msg = ERR_PTR(ret);
1480                 goto out_free1;
1481         }
1482
1483         len = sizeof(*head) +
1484                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64));
1485
1486         /* calculate (max) length for cap releases */
1487         len += sizeof(struct ceph_mds_request_release) *
1488                 (!!req->r_inode_drop + !!req->r_dentry_drop +
1489                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1490         if (req->r_dentry_drop)
1491                 len += req->r_dentry->d_name.len;
1492         if (req->r_old_dentry_drop)
1493                 len += req->r_old_dentry->d_name.len;
1494
1495         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len);
1496         if (!msg) {
1497                 msg = ERR_PTR(-ENOMEM);
1498                 goto out_free2;
1499         }
1500
1501         msg->hdr.tid = cpu_to_le64(req->r_tid);
1502
1503         head = msg->front.iov_base;
1504         p = msg->front.iov_base + sizeof(*head);
1505         end = msg->front.iov_base + msg->front.iov_len;
1506
1507         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1508         head->op = cpu_to_le32(req->r_op);
1509         head->caller_uid = cpu_to_le32(current_fsuid());
1510         head->caller_gid = cpu_to_le32(current_fsgid());
1511         head->args = req->r_args;
1512
1513         ceph_encode_filepath(&p, end, ino1, path1);
1514         ceph_encode_filepath(&p, end, ino2, path2);
1515
1516         /* cap releases */
1517         releases = 0;
1518         if (req->r_inode_drop)
1519                 releases += ceph_encode_inode_release(&p,
1520                       req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1521                       mds, req->r_inode_drop, req->r_inode_unless, 0);
1522         if (req->r_dentry_drop)
1523                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1524                        mds, req->r_dentry_drop, req->r_dentry_unless);
1525         if (req->r_old_dentry_drop)
1526                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1527                        mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1528         if (req->r_old_inode_drop)
1529                 releases += ceph_encode_inode_release(&p,
1530                       req->r_old_dentry->d_inode,
1531                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1532         head->num_releases = cpu_to_le16(releases);
1533
1534         BUG_ON(p > end);
1535         msg->front.iov_len = p - msg->front.iov_base;
1536         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1537
1538         msg->pages = req->r_pages;
1539         msg->nr_pages = req->r_num_pages;
1540         msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1541         msg->hdr.data_off = cpu_to_le16(0);
1542
1543 out_free2:
1544         if (freepath2)
1545                 kfree((char *)path2);
1546 out_free1:
1547         if (freepath1)
1548                 kfree((char *)path1);
1549 out:
1550         return msg;
1551 }
1552
1553 /*
1554  * called under mdsc->mutex if error, under no mutex if
1555  * success.
1556  */
1557 static void complete_request(struct ceph_mds_client *mdsc,
1558                              struct ceph_mds_request *req)
1559 {
1560         if (req->r_callback)
1561                 req->r_callback(mdsc, req);
1562         else
1563                 complete(&req->r_completion);
1564 }
1565
1566 /*
1567  * called under mdsc->mutex
1568  */
1569 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1570                                   struct ceph_mds_request *req,
1571                                   int mds)
1572 {
1573         struct ceph_mds_request_head *rhead;
1574         struct ceph_msg *msg;
1575         int flags = 0;
1576
1577         req->r_mds = mds;
1578         req->r_attempts++;
1579         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1580              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1581
1582         if (req->r_request) {
1583                 ceph_msg_put(req->r_request);
1584                 req->r_request = NULL;
1585         }
1586         msg = create_request_message(mdsc, req, mds);
1587         if (IS_ERR(msg)) {
1588                 req->r_err = PTR_ERR(msg);
1589                 complete_request(mdsc, req);
1590                 return PTR_ERR(msg);
1591         }
1592         req->r_request = msg;
1593
1594         rhead = msg->front.iov_base;
1595         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1596         if (req->r_got_unsafe)
1597                 flags |= CEPH_MDS_FLAG_REPLAY;
1598         if (req->r_locked_dir)
1599                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1600         rhead->flags = cpu_to_le32(flags);
1601         rhead->num_fwd = req->r_num_fwd;
1602         rhead->num_retry = req->r_attempts - 1;
1603
1604         dout(" r_locked_dir = %p\n", req->r_locked_dir);
1605
1606         if (req->r_target_inode && req->r_got_unsafe)
1607                 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1608         else
1609                 rhead->ino = 0;
1610         return 0;
1611 }
1612
1613 /*
1614  * send request, or put it on the appropriate wait list.
1615  */
1616 static int __do_request(struct ceph_mds_client *mdsc,
1617                         struct ceph_mds_request *req)
1618 {
1619         struct ceph_mds_session *session = NULL;
1620         int mds = -1;
1621         int err = -EAGAIN;
1622
1623         if (req->r_err || req->r_got_result)
1624                 goto out;
1625
1626         if (req->r_timeout &&
1627             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1628                 dout("do_request timed out\n");
1629                 err = -EIO;
1630                 goto finish;
1631         }
1632
1633         mds = __choose_mds(mdsc, req);
1634         if (mds < 0 ||
1635             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1636                 dout("do_request no mds or not active, waiting for map\n");
1637                 list_add(&req->r_wait, &mdsc->waiting_for_map);
1638                 goto out;
1639         }
1640
1641         /* get, open session */
1642         session = __ceph_lookup_mds_session(mdsc, mds);
1643         if (!session) {
1644                 session = register_session(mdsc, mds);
1645                 if (IS_ERR(session)) {
1646                         err = PTR_ERR(session);
1647                         goto finish;
1648                 }
1649         }
1650         dout("do_request mds%d session %p state %s\n", mds, session,
1651              session_state_name(session->s_state));
1652         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1653             session->s_state != CEPH_MDS_SESSION_HUNG) {
1654                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1655                     session->s_state == CEPH_MDS_SESSION_CLOSING)
1656                         __open_session(mdsc, session);
1657                 list_add(&req->r_wait, &session->s_waiting);
1658                 goto out_session;
1659         }
1660
1661         /* send request */
1662         req->r_session = get_session(session);
1663         req->r_resend_mds = -1;   /* forget any previous mds hint */
1664
1665         if (req->r_request_started == 0)   /* note request start time */
1666                 req->r_request_started = jiffies;
1667
1668         err = __prepare_send_request(mdsc, req, mds);
1669         if (!err) {
1670                 ceph_msg_get(req->r_request);
1671                 ceph_con_send(&session->s_con, req->r_request);
1672         }
1673
1674 out_session:
1675         ceph_put_mds_session(session);
1676 out:
1677         return err;
1678
1679 finish:
1680         req->r_err = err;
1681         complete_request(mdsc, req);
1682         goto out;
1683 }
1684
1685 /*
1686  * called under mdsc->mutex
1687  */
1688 static void __wake_requests(struct ceph_mds_client *mdsc,
1689                             struct list_head *head)
1690 {
1691         struct ceph_mds_request *req, *nreq;
1692
1693         list_for_each_entry_safe(req, nreq, head, r_wait) {
1694                 list_del_init(&req->r_wait);
1695                 __do_request(mdsc, req);
1696         }
1697 }
1698
1699 /*
1700  * Wake up threads with requests pending for @mds, so that they can
1701  * resubmit their requests to a possibly different mds.
1702  */
1703 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
1704 {
1705         struct ceph_mds_request *req;
1706         struct rb_node *p;
1707
1708         dout("kick_requests mds%d\n", mds);
1709         for (p = rb_first(&mdsc->request_tree); p; p = rb_next(p)) {
1710                 req = rb_entry(p, struct ceph_mds_request, r_node);
1711                 if (req->r_got_unsafe)
1712                         continue;
1713                 if (req->r_session &&
1714                     req->r_session->s_mds == mds) {
1715                         dout(" kicking tid %llu\n", req->r_tid);
1716                         put_request_session(req);
1717                         __do_request(mdsc, req);
1718                 }
1719         }
1720 }
1721
1722 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
1723                               struct ceph_mds_request *req)
1724 {
1725         dout("submit_request on %p\n", req);
1726         mutex_lock(&mdsc->mutex);
1727         __register_request(mdsc, req, NULL);
1728         __do_request(mdsc, req);
1729         mutex_unlock(&mdsc->mutex);
1730 }
1731
1732 /*
1733  * Synchrously perform an mds request.  Take care of all of the
1734  * session setup, forwarding, retry details.
1735  */
1736 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
1737                          struct inode *dir,
1738                          struct ceph_mds_request *req)
1739 {
1740         int err;
1741
1742         dout("do_request on %p\n", req);
1743
1744         /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
1745         if (req->r_inode)
1746                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
1747         if (req->r_locked_dir)
1748                 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
1749         if (req->r_old_dentry)
1750                 ceph_get_cap_refs(
1751                         ceph_inode(req->r_old_dentry->d_parent->d_inode),
1752                         CEPH_CAP_PIN);
1753
1754         /* issue */
1755         mutex_lock(&mdsc->mutex);
1756         __register_request(mdsc, req, dir);
1757         __do_request(mdsc, req);
1758
1759         if (req->r_err) {
1760                 err = req->r_err;
1761                 __unregister_request(mdsc, req);
1762                 dout("do_request early error %d\n", err);
1763                 goto out;
1764         }
1765
1766         /* wait */
1767         mutex_unlock(&mdsc->mutex);
1768         dout("do_request waiting\n");
1769         if (req->r_timeout) {
1770                 err = (long)wait_for_completion_interruptible_timeout(
1771                         &req->r_completion, req->r_timeout);
1772                 if (err == 0)
1773                         err = -EIO;
1774         } else {
1775                 err = wait_for_completion_interruptible(&req->r_completion);
1776         }
1777         dout("do_request waited, got %d\n", err);
1778         mutex_lock(&mdsc->mutex);
1779
1780         /* only abort if we didn't race with a real reply */
1781         if (req->r_got_result) {
1782                 err = le32_to_cpu(req->r_reply_info.head->result);
1783         } else if (err < 0) {
1784                 dout("aborted request %lld with %d\n", req->r_tid, err);
1785
1786                 /*
1787                  * ensure we aren't running concurrently with
1788                  * ceph_fill_trace or ceph_readdir_prepopulate, which
1789                  * rely on locks (dir mutex) held by our caller.
1790                  */
1791                 mutex_lock(&req->r_fill_mutex);
1792                 req->r_err = err;
1793                 req->r_aborted = true;
1794                 mutex_unlock(&req->r_fill_mutex);
1795
1796                 if (req->r_locked_dir &&
1797                     (req->r_op & CEPH_MDS_OP_WRITE)) {
1798                         struct ceph_inode_info *ci =
1799                                 ceph_inode(req->r_locked_dir);
1800
1801                         dout("aborted, clearing I_COMPLETE on %p, leases\n",
1802                              req->r_locked_dir);
1803                         spin_lock(&req->r_locked_dir->i_lock);
1804                         ci->i_ceph_flags &= ~CEPH_I_COMPLETE;
1805                         ci->i_release_count++;
1806                         spin_unlock(&req->r_locked_dir->i_lock);
1807
1808                         if (req->r_dentry)
1809                                 ceph_invalidate_dentry_lease(req->r_dentry);
1810                         if (req->r_old_dentry)
1811                                 ceph_invalidate_dentry_lease(req->r_old_dentry);
1812                 }
1813         } else {
1814                 err = req->r_err;
1815         }
1816
1817 out:
1818         mutex_unlock(&mdsc->mutex);
1819         dout("do_request %p done, result %d\n", req, err);
1820         return err;
1821 }
1822
1823 /*
1824  * Handle mds reply.
1825  *
1826  * We take the session mutex and parse and process the reply immediately.
1827  * This preserves the logical ordering of replies, capabilities, etc., sent
1828  * by the MDS as they are applied to our local cache.
1829  */
1830 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
1831 {
1832         struct ceph_mds_client *mdsc = session->s_mdsc;
1833         struct ceph_mds_request *req;
1834         struct ceph_mds_reply_head *head = msg->front.iov_base;
1835         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
1836         u64 tid;
1837         int err, result;
1838         int mds = session->s_mds;
1839
1840         if (msg->front.iov_len < sizeof(*head)) {
1841                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
1842                 ceph_msg_dump(msg);
1843                 return;
1844         }
1845
1846         /* get request, session */
1847         tid = le64_to_cpu(msg->hdr.tid);
1848         mutex_lock(&mdsc->mutex);
1849         req = __lookup_request(mdsc, tid);
1850         if (!req) {
1851                 dout("handle_reply on unknown tid %llu\n", tid);
1852                 mutex_unlock(&mdsc->mutex);
1853                 return;
1854         }
1855         dout("handle_reply %p\n", req);
1856
1857         /* correct session? */
1858         if (req->r_session != session) {
1859                 pr_err("mdsc_handle_reply got %llu on session mds%d"
1860                        " not mds%d\n", tid, session->s_mds,
1861                        req->r_session ? req->r_session->s_mds : -1);
1862                 mutex_unlock(&mdsc->mutex);
1863                 goto out;
1864         }
1865
1866         /* dup? */
1867         if ((req->r_got_unsafe && !head->safe) ||
1868             (req->r_got_safe && head->safe)) {
1869                 pr_warning("got a dup %s reply on %llu from mds%d\n",
1870                            head->safe ? "safe" : "unsafe", tid, mds);
1871                 mutex_unlock(&mdsc->mutex);
1872                 goto out;
1873         }
1874
1875         result = le32_to_cpu(head->result);
1876
1877         /*
1878          * Tolerate 2 consecutive ESTALEs from the same mds.
1879          * FIXME: we should be looking at the cap migrate_seq.
1880          */
1881         if (result == -ESTALE) {
1882                 req->r_direct_mode = USE_AUTH_MDS;
1883                 req->r_num_stale++;
1884                 if (req->r_num_stale <= 2) {
1885                         __do_request(mdsc, req);
1886                         mutex_unlock(&mdsc->mutex);
1887                         goto out;
1888                 }
1889         } else {
1890                 req->r_num_stale = 0;
1891         }
1892
1893         if (head->safe) {
1894                 req->r_got_safe = true;
1895                 __unregister_request(mdsc, req);
1896                 complete(&req->r_safe_completion);
1897
1898                 if (req->r_got_unsafe) {
1899                         /*
1900                          * We already handled the unsafe response, now do the
1901                          * cleanup.  No need to examine the response; the MDS
1902                          * doesn't include any result info in the safe
1903                          * response.  And even if it did, there is nothing
1904                          * useful we could do with a revised return value.
1905                          */
1906                         dout("got safe reply %llu, mds%d\n", tid, mds);
1907                         list_del_init(&req->r_unsafe_item);
1908
1909                         /* last unsafe request during umount? */
1910                         if (mdsc->stopping && !__get_oldest_req(mdsc))
1911                                 complete(&mdsc->safe_umount_waiters);
1912                         mutex_unlock(&mdsc->mutex);
1913                         goto out;
1914                 }
1915         } else {
1916                 req->r_got_unsafe = true;
1917                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
1918         }
1919
1920         dout("handle_reply tid %lld result %d\n", tid, result);
1921         rinfo = &req->r_reply_info;
1922         err = parse_reply_info(msg, rinfo);
1923         mutex_unlock(&mdsc->mutex);
1924
1925         mutex_lock(&session->s_mutex);
1926         if (err < 0) {
1927                 pr_err("mdsc_handle_reply got corrupt reply mds%d\n", mds);
1928                 ceph_msg_dump(msg);
1929                 goto out_err;
1930         }
1931
1932         /* snap trace */
1933         if (rinfo->snapblob_len) {
1934                 down_write(&mdsc->snap_rwsem);
1935                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
1936                                rinfo->snapblob + rinfo->snapblob_len,
1937                                le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
1938                 downgrade_write(&mdsc->snap_rwsem);
1939         } else {
1940                 down_read(&mdsc->snap_rwsem);
1941         }
1942
1943         /* insert trace into our cache */
1944         mutex_lock(&req->r_fill_mutex);
1945         err = ceph_fill_trace(mdsc->client->sb, req, req->r_session);
1946         if (err == 0) {
1947                 if (result == 0 && rinfo->dir_nr)
1948                         ceph_readdir_prepopulate(req, req->r_session);
1949                 ceph_unreserve_caps(&req->r_caps_reservation);
1950         }
1951         mutex_unlock(&req->r_fill_mutex);
1952
1953         up_read(&mdsc->snap_rwsem);
1954 out_err:
1955         mutex_lock(&mdsc->mutex);
1956         if (!req->r_aborted) {
1957                 if (err) {
1958                         req->r_err = err;
1959                 } else {
1960                         req->r_reply = msg;
1961                         ceph_msg_get(msg);
1962                         req->r_got_result = true;
1963                 }
1964         } else {
1965                 dout("reply arrived after request %lld was aborted\n", tid);
1966         }
1967         mutex_unlock(&mdsc->mutex);
1968
1969         add_cap_releases(mdsc, req->r_session, -1);
1970         mutex_unlock(&session->s_mutex);
1971
1972         /* kick calling process */
1973         complete_request(mdsc, req);
1974 out:
1975         ceph_mdsc_put_request(req);
1976         return;
1977 }
1978
1979
1980
1981 /*
1982  * handle mds notification that our request has been forwarded.
1983  */
1984 static void handle_forward(struct ceph_mds_client *mdsc,
1985                            struct ceph_mds_session *session,
1986                            struct ceph_msg *msg)
1987 {
1988         struct ceph_mds_request *req;
1989         u64 tid = le64_to_cpu(msg->hdr.tid);
1990         u32 next_mds;
1991         u32 fwd_seq;
1992         int err = -EINVAL;
1993         void *p = msg->front.iov_base;
1994         void *end = p + msg->front.iov_len;
1995
1996         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1997         next_mds = ceph_decode_32(&p);
1998         fwd_seq = ceph_decode_32(&p);
1999
2000         mutex_lock(&mdsc->mutex);
2001         req = __lookup_request(mdsc, tid);
2002         if (!req) {
2003                 dout("forward %llu to mds%d - req dne\n", tid, next_mds);
2004                 goto out;  /* dup reply? */
2005         }
2006
2007         if (fwd_seq <= req->r_num_fwd) {
2008                 dout("forward %llu to mds%d - old seq %d <= %d\n",
2009                      tid, next_mds, req->r_num_fwd, fwd_seq);
2010         } else {
2011                 /* resend. forward race not possible; mds would drop */
2012                 dout("forward %llu to mds%d (we resend)\n", tid, next_mds);
2013                 req->r_num_fwd = fwd_seq;
2014                 req->r_resend_mds = next_mds;
2015                 put_request_session(req);
2016                 __do_request(mdsc, req);
2017         }
2018         ceph_mdsc_put_request(req);
2019 out:
2020         mutex_unlock(&mdsc->mutex);
2021         return;
2022
2023 bad:
2024         pr_err("mdsc_handle_forward decode error err=%d\n", err);
2025 }
2026
2027 /*
2028  * handle a mds session control message
2029  */
2030 static void handle_session(struct ceph_mds_session *session,
2031                            struct ceph_msg *msg)
2032 {
2033         struct ceph_mds_client *mdsc = session->s_mdsc;
2034         u32 op;
2035         u64 seq;
2036         int mds = session->s_mds;
2037         struct ceph_mds_session_head *h = msg->front.iov_base;
2038         int wake = 0;
2039
2040         /* decode */
2041         if (msg->front.iov_len != sizeof(*h))
2042                 goto bad;
2043         op = le32_to_cpu(h->op);
2044         seq = le64_to_cpu(h->seq);
2045
2046         mutex_lock(&mdsc->mutex);
2047         if (op == CEPH_SESSION_CLOSE)
2048                 __unregister_session(mdsc, session);
2049         /* FIXME: this ttl calculation is generous */
2050         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2051         mutex_unlock(&mdsc->mutex);
2052
2053         mutex_lock(&session->s_mutex);
2054
2055         dout("handle_session mds%d %s %p state %s seq %llu\n",
2056              mds, ceph_session_op_name(op), session,
2057              session_state_name(session->s_state), seq);
2058
2059         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2060                 session->s_state = CEPH_MDS_SESSION_OPEN;
2061                 pr_info("mds%d came back\n", session->s_mds);
2062         }
2063
2064         switch (op) {
2065         case CEPH_SESSION_OPEN:
2066                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2067                         pr_info("mds%d reconnect success\n", session->s_mds);
2068                 session->s_state = CEPH_MDS_SESSION_OPEN;
2069                 renewed_caps(mdsc, session, 0);
2070                 wake = 1;
2071                 if (mdsc->stopping)
2072                         __close_session(mdsc, session);
2073                 break;
2074
2075         case CEPH_SESSION_RENEWCAPS:
2076                 if (session->s_renew_seq == seq)
2077                         renewed_caps(mdsc, session, 1);
2078                 break;
2079
2080         case CEPH_SESSION_CLOSE:
2081                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2082                         pr_info("mds%d reconnect denied\n", session->s_mds);
2083                 remove_session_caps(session);
2084                 wake = 1; /* for good measure */
2085                 complete(&mdsc->session_close_waiters);
2086                 kick_requests(mdsc, mds);
2087                 break;
2088
2089         case CEPH_SESSION_STALE:
2090                 pr_info("mds%d caps went stale, renewing\n",
2091                         session->s_mds);
2092                 spin_lock(&session->s_cap_lock);
2093                 session->s_cap_gen++;
2094                 session->s_cap_ttl = 0;
2095                 spin_unlock(&session->s_cap_lock);
2096                 send_renew_caps(mdsc, session);
2097                 break;
2098
2099         case CEPH_SESSION_RECALL_STATE:
2100                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2101                 break;
2102
2103         default:
2104                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2105                 WARN_ON(1);
2106         }
2107
2108         mutex_unlock(&session->s_mutex);
2109         if (wake) {
2110                 mutex_lock(&mdsc->mutex);
2111                 __wake_requests(mdsc, &session->s_waiting);
2112                 mutex_unlock(&mdsc->mutex);
2113         }
2114         return;
2115
2116 bad:
2117         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2118                (int)msg->front.iov_len);
2119         ceph_msg_dump(msg);
2120         return;
2121 }
2122
2123
2124 /*
2125  * called under session->mutex.
2126  */
2127 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2128                                    struct ceph_mds_session *session)
2129 {
2130         struct ceph_mds_request *req, *nreq;
2131         int err;
2132
2133         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2134
2135         mutex_lock(&mdsc->mutex);
2136         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2137                 err = __prepare_send_request(mdsc, req, session->s_mds);
2138                 if (!err) {
2139                         ceph_msg_get(req->r_request);
2140                         ceph_con_send(&session->s_con, req->r_request);
2141                 }
2142         }
2143         mutex_unlock(&mdsc->mutex);
2144 }
2145
2146 /*
2147  * Encode information about a cap for a reconnect with the MDS.
2148  */
2149 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2150                           void *arg)
2151 {
2152         struct ceph_mds_cap_reconnect rec;
2153         struct ceph_inode_info *ci;
2154         struct ceph_pagelist *pagelist = arg;
2155         char *path;
2156         int pathlen, err;
2157         u64 pathbase;
2158         struct dentry *dentry;
2159
2160         ci = cap->ci;
2161
2162         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2163              inode, ceph_vinop(inode), cap, cap->cap_id,
2164              ceph_cap_string(cap->issued));
2165         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2166         if (err)
2167                 return err;
2168
2169         dentry = d_find_alias(inode);
2170         if (dentry) {
2171                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2172                 if (IS_ERR(path)) {
2173                         err = PTR_ERR(path);
2174                         BUG_ON(err);
2175                 }
2176         } else {
2177                 path = NULL;
2178                 pathlen = 0;
2179         }
2180         err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2181         if (err)
2182                 goto out;
2183
2184         spin_lock(&inode->i_lock);
2185         cap->seq = 0;        /* reset cap seq */
2186         cap->issue_seq = 0;  /* and issue_seq */
2187         rec.cap_id = cpu_to_le64(cap->cap_id);
2188         rec.pathbase = cpu_to_le64(pathbase);
2189         rec.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2190         rec.issued = cpu_to_le32(cap->issued);
2191         rec.size = cpu_to_le64(inode->i_size);
2192         ceph_encode_timespec(&rec.mtime, &inode->i_mtime);
2193         ceph_encode_timespec(&rec.atime, &inode->i_atime);
2194         rec.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2195         spin_unlock(&inode->i_lock);
2196
2197         err = ceph_pagelist_append(pagelist, &rec, sizeof(rec));
2198
2199 out:
2200         kfree(path);
2201         dput(dentry);
2202         return err;
2203 }
2204
2205
2206 /*
2207  * If an MDS fails and recovers, clients need to reconnect in order to
2208  * reestablish shared state.  This includes all caps issued through
2209  * this session _and_ the snap_realm hierarchy.  Because it's not
2210  * clear which snap realms the mds cares about, we send everything we
2211  * know about.. that ensures we'll then get any new info the
2212  * recovering MDS might have.
2213  *
2214  * This is a relatively heavyweight operation, but it's rare.
2215  *
2216  * called with mdsc->mutex held.
2217  */
2218 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2219                                struct ceph_mds_session *session)
2220 {
2221         struct ceph_msg *reply;
2222         struct rb_node *p;
2223         int mds = session->s_mds;
2224         int err = -ENOMEM;
2225         struct ceph_pagelist *pagelist;
2226
2227         pr_info("mds%d reconnect start\n", mds);
2228
2229         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2230         if (!pagelist)
2231                 goto fail_nopagelist;
2232         ceph_pagelist_init(pagelist);
2233
2234         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0);
2235         if (!reply)
2236                 goto fail_nomsg;
2237
2238         mutex_lock(&session->s_mutex);
2239         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2240         session->s_seq = 0;
2241
2242         ceph_con_open(&session->s_con,
2243                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2244
2245         /* replay unsafe requests */
2246         replay_unsafe_requests(mdsc, session);
2247
2248         down_read(&mdsc->snap_rwsem);
2249
2250         dout("session %p state %s\n", session,
2251              session_state_name(session->s_state));
2252
2253         /* drop old cap expires; we're about to reestablish that state */
2254         discard_cap_releases(mdsc, session);
2255
2256         /* traverse this session's caps */
2257         err = ceph_pagelist_encode_32(pagelist, session->s_nr_caps);
2258         if (err)
2259                 goto fail;
2260         err = iterate_session_caps(session, encode_caps_cb, pagelist);
2261         if (err < 0)
2262                 goto fail;
2263
2264         /*
2265          * snaprealms.  we provide mds with the ino, seq (version), and
2266          * parent for all of our realms.  If the mds has any newer info,
2267          * it will tell us.
2268          */
2269         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2270                 struct ceph_snap_realm *realm =
2271                         rb_entry(p, struct ceph_snap_realm, node);
2272                 struct ceph_mds_snaprealm_reconnect sr_rec;
2273
2274                 dout(" adding snap realm %llx seq %lld parent %llx\n",
2275                      realm->ino, realm->seq, realm->parent_ino);
2276                 sr_rec.ino = cpu_to_le64(realm->ino);
2277                 sr_rec.seq = cpu_to_le64(realm->seq);
2278                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2279                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2280                 if (err)
2281                         goto fail;
2282         }
2283
2284         reply->pagelist = pagelist;
2285         reply->hdr.data_len = cpu_to_le32(pagelist->length);
2286         reply->nr_pages = calc_pages_for(0, pagelist->length);
2287         ceph_con_send(&session->s_con, reply);
2288
2289         mutex_unlock(&session->s_mutex);
2290
2291         mutex_lock(&mdsc->mutex);
2292         __wake_requests(mdsc, &session->s_waiting);
2293         mutex_unlock(&mdsc->mutex);
2294
2295         up_read(&mdsc->snap_rwsem);
2296         return;
2297
2298 fail:
2299         ceph_msg_put(reply);
2300         up_read(&mdsc->snap_rwsem);
2301         mutex_unlock(&session->s_mutex);
2302 fail_nomsg:
2303         ceph_pagelist_release(pagelist);
2304         kfree(pagelist);
2305 fail_nopagelist:
2306         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
2307         return;
2308 }
2309
2310
2311 /*
2312  * compare old and new mdsmaps, kicking requests
2313  * and closing out old connections as necessary
2314  *
2315  * called under mdsc->mutex.
2316  */
2317 static void check_new_map(struct ceph_mds_client *mdsc,
2318                           struct ceph_mdsmap *newmap,
2319                           struct ceph_mdsmap *oldmap)
2320 {
2321         int i;
2322         int oldstate, newstate;
2323         struct ceph_mds_session *s;
2324
2325         dout("check_new_map new %u old %u\n",
2326              newmap->m_epoch, oldmap->m_epoch);
2327
2328         for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2329                 if (mdsc->sessions[i] == NULL)
2330                         continue;
2331                 s = mdsc->sessions[i];
2332                 oldstate = ceph_mdsmap_get_state(oldmap, i);
2333                 newstate = ceph_mdsmap_get_state(newmap, i);
2334
2335                 dout("check_new_map mds%d state %s -> %s (session %s)\n",
2336                      i, ceph_mds_state_name(oldstate),
2337                      ceph_mds_state_name(newstate),
2338                      session_state_name(s->s_state));
2339
2340                 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
2341                            ceph_mdsmap_get_addr(newmap, i),
2342                            sizeof(struct ceph_entity_addr))) {
2343                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2344                                 /* the session never opened, just close it
2345                                  * out now */
2346                                 __wake_requests(mdsc, &s->s_waiting);
2347                                 __unregister_session(mdsc, s);
2348                         } else {
2349                                 /* just close it */
2350                                 mutex_unlock(&mdsc->mutex);
2351                                 mutex_lock(&s->s_mutex);
2352                                 mutex_lock(&mdsc->mutex);
2353                                 ceph_con_close(&s->s_con);
2354                                 mutex_unlock(&s->s_mutex);
2355                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2356                         }
2357
2358                         /* kick any requests waiting on the recovering mds */
2359                         kick_requests(mdsc, i);
2360                 } else if (oldstate == newstate) {
2361                         continue;  /* nothing new with this mds */
2362                 }
2363
2364                 /*
2365                  * send reconnect?
2366                  */
2367                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2368                     newstate >= CEPH_MDS_STATE_RECONNECT) {
2369                         mutex_unlock(&mdsc->mutex);
2370                         send_mds_reconnect(mdsc, s);
2371                         mutex_lock(&mdsc->mutex);
2372                 }
2373
2374                 /*
2375                  * kick request on any mds that has gone active.
2376                  */
2377                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2378                     newstate >= CEPH_MDS_STATE_ACTIVE) {
2379                         if (oldstate != CEPH_MDS_STATE_CREATING &&
2380                             oldstate != CEPH_MDS_STATE_STARTING)
2381                                 pr_info("mds%d recovery completed\n", s->s_mds);
2382                         kick_requests(mdsc, i);
2383                         ceph_kick_flushing_caps(mdsc, s);
2384                         wake_up_session_caps(s, 1);
2385                 }
2386         }
2387 }
2388
2389
2390
2391 /*
2392  * leases
2393  */
2394
2395 /*
2396  * caller must hold session s_mutex, dentry->d_lock
2397  */
2398 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2399 {
2400         struct ceph_dentry_info *di = ceph_dentry(dentry);
2401
2402         ceph_put_mds_session(di->lease_session);
2403         di->lease_session = NULL;
2404 }
2405
2406 static void handle_lease(struct ceph_mds_client *mdsc,
2407                          struct ceph_mds_session *session,
2408                          struct ceph_msg *msg)
2409 {
2410         struct super_block *sb = mdsc->client->sb;
2411         struct inode *inode;
2412         struct ceph_inode_info *ci;
2413         struct dentry *parent, *dentry;
2414         struct ceph_dentry_info *di;
2415         int mds = session->s_mds;
2416         struct ceph_mds_lease *h = msg->front.iov_base;
2417         struct ceph_vino vino;
2418         int mask;
2419         struct qstr dname;
2420         int release = 0;
2421
2422         dout("handle_lease from mds%d\n", mds);
2423
2424         /* decode */
2425         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2426                 goto bad;
2427         vino.ino = le64_to_cpu(h->ino);
2428         vino.snap = CEPH_NOSNAP;
2429         mask = le16_to_cpu(h->mask);
2430         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2431         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2432         if (dname.len != get_unaligned_le32(h+1))
2433                 goto bad;
2434
2435         mutex_lock(&session->s_mutex);
2436         session->s_seq++;
2437
2438         /* lookup inode */
2439         inode = ceph_find_inode(sb, vino);
2440         dout("handle_lease '%s', mask %d, ino %llx %p\n",
2441              ceph_lease_op_name(h->action), mask, vino.ino, inode);
2442         if (inode == NULL) {
2443                 dout("handle_lease no inode %llx\n", vino.ino);
2444                 goto release;
2445         }
2446         ci = ceph_inode(inode);
2447
2448         /* dentry */
2449         parent = d_find_alias(inode);
2450         if (!parent) {
2451                 dout("no parent dentry on inode %p\n", inode);
2452                 WARN_ON(1);
2453                 goto release;  /* hrm... */
2454         }
2455         dname.hash = full_name_hash(dname.name, dname.len);
2456         dentry = d_lookup(parent, &dname);
2457         dput(parent);
2458         if (!dentry)
2459                 goto release;
2460
2461         spin_lock(&dentry->d_lock);
2462         di = ceph_dentry(dentry);
2463         switch (h->action) {
2464         case CEPH_MDS_LEASE_REVOKE:
2465                 if (di && di->lease_session == session) {
2466                         h->seq = cpu_to_le32(di->lease_seq);
2467                         __ceph_mdsc_drop_dentry_lease(dentry);
2468                 }
2469                 release = 1;
2470                 break;
2471
2472         case CEPH_MDS_LEASE_RENEW:
2473                 if (di && di->lease_session == session &&
2474                     di->lease_gen == session->s_cap_gen &&
2475                     di->lease_renew_from &&
2476                     di->lease_renew_after == 0) {
2477                         unsigned long duration =
2478                                 le32_to_cpu(h->duration_ms) * HZ / 1000;
2479
2480                         di->lease_seq = le32_to_cpu(h->seq);
2481                         dentry->d_time = di->lease_renew_from + duration;
2482                         di->lease_renew_after = di->lease_renew_from +
2483                                 (duration >> 1);
2484                         di->lease_renew_from = 0;
2485                 }
2486                 break;
2487         }
2488         spin_unlock(&dentry->d_lock);
2489         dput(dentry);
2490
2491         if (!release)
2492                 goto out;
2493
2494 release:
2495         /* let's just reuse the same message */
2496         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2497         ceph_msg_get(msg);
2498         ceph_con_send(&session->s_con, msg);
2499
2500 out:
2501         iput(inode);
2502         mutex_unlock(&session->s_mutex);
2503         return;
2504
2505 bad:
2506         pr_err("corrupt lease message\n");
2507         ceph_msg_dump(msg);
2508 }
2509
2510 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
2511                               struct inode *inode,
2512                               struct dentry *dentry, char action,
2513                               u32 seq)
2514 {
2515         struct ceph_msg *msg;
2516         struct ceph_mds_lease *lease;
2517         int len = sizeof(*lease) + sizeof(u32);
2518         int dnamelen = 0;
2519
2520         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2521              inode, dentry, ceph_lease_op_name(action), session->s_mds);
2522         dnamelen = dentry->d_name.len;
2523         len += dnamelen;
2524
2525         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len);
2526         if (!msg)
2527                 return;
2528         lease = msg->front.iov_base;
2529         lease->action = action;
2530         lease->mask = cpu_to_le16(CEPH_LOCK_DN);
2531         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
2532         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
2533         lease->seq = cpu_to_le32(seq);
2534         put_unaligned_le32(dnamelen, lease + 1);
2535         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
2536
2537         /*
2538          * if this is a preemptive lease RELEASE, no need to
2539          * flush request stream, since the actual request will
2540          * soon follow.
2541          */
2542         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
2543
2544         ceph_con_send(&session->s_con, msg);
2545 }
2546
2547 /*
2548  * Preemptively release a lease we expect to invalidate anyway.
2549  * Pass @inode always, @dentry is optional.
2550  */
2551 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2552                              struct dentry *dentry, int mask)
2553 {
2554         struct ceph_dentry_info *di;
2555         struct ceph_mds_session *session;
2556         u32 seq;
2557
2558         BUG_ON(inode == NULL);
2559         BUG_ON(dentry == NULL);
2560         BUG_ON(mask != CEPH_LOCK_DN);
2561
2562         /* is dentry lease valid? */
2563         spin_lock(&dentry->d_lock);
2564         di = ceph_dentry(dentry);
2565         if (!di || !di->lease_session ||
2566             di->lease_session->s_mds < 0 ||
2567             di->lease_gen != di->lease_session->s_cap_gen ||
2568             !time_before(jiffies, dentry->d_time)) {
2569                 dout("lease_release inode %p dentry %p -- "
2570                      "no lease on %d\n",
2571                      inode, dentry, mask);
2572                 spin_unlock(&dentry->d_lock);
2573                 return;
2574         }
2575
2576         /* we do have a lease on this dentry; note mds and seq */
2577         session = ceph_get_mds_session(di->lease_session);
2578         seq = di->lease_seq;
2579         __ceph_mdsc_drop_dentry_lease(dentry);
2580         spin_unlock(&dentry->d_lock);
2581
2582         dout("lease_release inode %p dentry %p mask %d to mds%d\n",
2583              inode, dentry, mask, session->s_mds);
2584         ceph_mdsc_lease_send_msg(session, inode, dentry,
2585                                  CEPH_MDS_LEASE_RELEASE, seq);
2586         ceph_put_mds_session(session);
2587 }
2588
2589 /*
2590  * drop all leases (and dentry refs) in preparation for umount
2591  */
2592 static void drop_leases(struct ceph_mds_client *mdsc)
2593 {
2594         int i;
2595
2596         dout("drop_leases\n");
2597         mutex_lock(&mdsc->mutex);
2598         for (i = 0; i < mdsc->max_sessions; i++) {
2599                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2600                 if (!s)
2601                         continue;
2602                 mutex_unlock(&mdsc->mutex);
2603                 mutex_lock(&s->s_mutex);
2604                 mutex_unlock(&s->s_mutex);
2605                 ceph_put_mds_session(s);
2606                 mutex_lock(&mdsc->mutex);
2607         }
2608         mutex_unlock(&mdsc->mutex);
2609 }
2610
2611
2612
2613 /*
2614  * delayed work -- periodically trim expired leases, renew caps with mds
2615  */
2616 static void schedule_delayed(struct ceph_mds_client *mdsc)
2617 {
2618         int delay = 5;
2619         unsigned hz = round_jiffies_relative(HZ * delay);
2620         schedule_delayed_work(&mdsc->delayed_work, hz);
2621 }
2622
2623 static void delayed_work(struct work_struct *work)
2624 {
2625         int i;
2626         struct ceph_mds_client *mdsc =
2627                 container_of(work, struct ceph_mds_client, delayed_work.work);
2628         int renew_interval;
2629         int renew_caps;
2630
2631         dout("mdsc delayed_work\n");
2632         ceph_check_delayed_caps(mdsc);
2633
2634         mutex_lock(&mdsc->mutex);
2635         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
2636         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
2637                                    mdsc->last_renew_caps);
2638         if (renew_caps)
2639                 mdsc->last_renew_caps = jiffies;
2640
2641         for (i = 0; i < mdsc->max_sessions; i++) {
2642                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2643                 if (s == NULL)
2644                         continue;
2645                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
2646                         dout("resending session close request for mds%d\n",
2647                              s->s_mds);
2648                         request_close_session(mdsc, s);
2649                         ceph_put_mds_session(s);
2650                         continue;
2651                 }
2652                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
2653                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
2654                                 s->s_state = CEPH_MDS_SESSION_HUNG;
2655                                 pr_info("mds%d hung\n", s->s_mds);
2656                         }
2657                 }
2658                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
2659                         /* this mds is failed or recovering, just wait */
2660                         ceph_put_mds_session(s);
2661                         continue;
2662                 }
2663                 mutex_unlock(&mdsc->mutex);
2664
2665                 mutex_lock(&s->s_mutex);
2666                 if (renew_caps)
2667                         send_renew_caps(mdsc, s);
2668                 else
2669                         ceph_con_keepalive(&s->s_con);
2670                 add_cap_releases(mdsc, s, -1);
2671                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2672                     s->s_state == CEPH_MDS_SESSION_HUNG)
2673                         send_cap_releases(mdsc, s);
2674                 mutex_unlock(&s->s_mutex);
2675                 ceph_put_mds_session(s);
2676
2677                 mutex_lock(&mdsc->mutex);
2678         }
2679         mutex_unlock(&mdsc->mutex);
2680
2681         schedule_delayed(mdsc);
2682 }
2683
2684
2685 int ceph_mdsc_init(struct ceph_mds_client *mdsc, struct ceph_client *client)
2686 {
2687         mdsc->client = client;
2688         mutex_init(&mdsc->mutex);
2689         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
2690         if (mdsc->mdsmap == NULL)
2691                 return -ENOMEM;
2692
2693         init_completion(&mdsc->safe_umount_waiters);
2694         init_completion(&mdsc->session_close_waiters);
2695         INIT_LIST_HEAD(&mdsc->waiting_for_map);
2696         mdsc->sessions = NULL;
2697         mdsc->max_sessions = 0;
2698         mdsc->stopping = 0;
2699         init_rwsem(&mdsc->snap_rwsem);
2700         mdsc->snap_realms = RB_ROOT;
2701         INIT_LIST_HEAD(&mdsc->snap_empty);
2702         spin_lock_init(&mdsc->snap_empty_lock);
2703         mdsc->last_tid = 0;
2704         mdsc->request_tree = RB_ROOT;
2705         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
2706         mdsc->last_renew_caps = jiffies;
2707         INIT_LIST_HEAD(&mdsc->cap_delay_list);
2708         spin_lock_init(&mdsc->cap_delay_lock);
2709         INIT_LIST_HEAD(&mdsc->snap_flush_list);
2710         spin_lock_init(&mdsc->snap_flush_lock);
2711         mdsc->cap_flush_seq = 0;
2712         INIT_LIST_HEAD(&mdsc->cap_dirty);
2713         mdsc->num_cap_flushing = 0;
2714         spin_lock_init(&mdsc->cap_dirty_lock);
2715         init_waitqueue_head(&mdsc->cap_flushing_wq);
2716         spin_lock_init(&mdsc->dentry_lru_lock);
2717         INIT_LIST_HEAD(&mdsc->dentry_lru);
2718
2719         return 0;
2720 }
2721
2722 /*
2723  * Wait for safe replies on open mds requests.  If we time out, drop
2724  * all requests from the tree to avoid dangling dentry refs.
2725  */
2726 static void wait_requests(struct ceph_mds_client *mdsc)
2727 {
2728         struct ceph_mds_request *req;
2729         struct ceph_client *client = mdsc->client;
2730
2731         mutex_lock(&mdsc->mutex);
2732         if (__get_oldest_req(mdsc)) {
2733                 mutex_unlock(&mdsc->mutex);
2734
2735                 dout("wait_requests waiting for requests\n");
2736                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
2737                                     client->mount_args->mount_timeout * HZ);
2738
2739                 /* tear down remaining requests */
2740                 mutex_lock(&mdsc->mutex);
2741                 while ((req = __get_oldest_req(mdsc))) {
2742                         dout("wait_requests timed out on tid %llu\n",
2743                              req->r_tid);
2744                         __unregister_request(mdsc, req);
2745                 }
2746         }
2747         mutex_unlock(&mdsc->mutex);
2748         dout("wait_requests done\n");
2749 }
2750
2751 /*
2752  * called before mount is ro, and before dentries are torn down.
2753  * (hmm, does this still race with new lookups?)
2754  */
2755 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
2756 {
2757         dout("pre_umount\n");
2758         mdsc->stopping = 1;
2759
2760         drop_leases(mdsc);
2761         ceph_flush_dirty_caps(mdsc);
2762         wait_requests(mdsc);
2763 }
2764
2765 /*
2766  * wait for all write mds requests to flush.
2767  */
2768 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
2769 {
2770         struct ceph_mds_request *req = NULL, *nextreq;
2771         struct rb_node *n;
2772
2773         mutex_lock(&mdsc->mutex);
2774         dout("wait_unsafe_requests want %lld\n", want_tid);
2775 restart:
2776         req = __get_oldest_req(mdsc);
2777         while (req && req->r_tid <= want_tid) {
2778                 /* find next request */
2779                 n = rb_next(&req->r_node);
2780                 if (n)
2781                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
2782                 else
2783                         nextreq = NULL;
2784                 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
2785                         /* write op */
2786                         ceph_mdsc_get_request(req);
2787                         if (nextreq)
2788                                 ceph_mdsc_get_request(nextreq);
2789                         mutex_unlock(&mdsc->mutex);
2790                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
2791                              req->r_tid, want_tid);
2792                         wait_for_completion(&req->r_safe_completion);
2793                         mutex_lock(&mdsc->mutex);
2794                         ceph_mdsc_put_request(req);
2795                         if (!nextreq)
2796                                 break;  /* next dne before, so we're done! */
2797                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
2798                                 /* next request was removed from tree */
2799                                 ceph_mdsc_put_request(nextreq);
2800                                 goto restart;
2801                         }
2802                         ceph_mdsc_put_request(nextreq);  /* won't go away */
2803                 }
2804                 req = nextreq;
2805         }
2806         mutex_unlock(&mdsc->mutex);
2807         dout("wait_unsafe_requests done\n");
2808 }
2809
2810 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
2811 {
2812         u64 want_tid, want_flush;
2813
2814         if (mdsc->client->mount_state == CEPH_MOUNT_SHUTDOWN)
2815                 return;
2816
2817         dout("sync\n");
2818         mutex_lock(&mdsc->mutex);
2819         want_tid = mdsc->last_tid;
2820         want_flush = mdsc->cap_flush_seq;
2821         mutex_unlock(&mdsc->mutex);
2822         dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
2823
2824         ceph_flush_dirty_caps(mdsc);
2825
2826         wait_unsafe_requests(mdsc, want_tid);
2827         wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
2828 }
2829
2830
2831 /*
2832  * called after sb is ro.
2833  */
2834 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
2835 {
2836         struct ceph_mds_session *session;
2837         int i;
2838         int n;
2839         struct ceph_client *client = mdsc->client;
2840         unsigned long started, timeout = client->mount_args->mount_timeout * HZ;
2841
2842         dout("close_sessions\n");
2843
2844         mutex_lock(&mdsc->mutex);
2845
2846         /* close sessions */
2847         started = jiffies;
2848         while (time_before(jiffies, started + timeout)) {
2849                 dout("closing sessions\n");
2850                 n = 0;
2851                 for (i = 0; i < mdsc->max_sessions; i++) {
2852                         session = __ceph_lookup_mds_session(mdsc, i);
2853                         if (!session)
2854                                 continue;
2855                         mutex_unlock(&mdsc->mutex);
2856                         mutex_lock(&session->s_mutex);
2857                         __close_session(mdsc, session);
2858                         mutex_unlock(&session->s_mutex);
2859                         ceph_put_mds_session(session);
2860                         mutex_lock(&mdsc->mutex);
2861                         n++;
2862                 }
2863                 if (n == 0)
2864                         break;
2865
2866                 if (client->mount_state == CEPH_MOUNT_SHUTDOWN)
2867                         break;
2868
2869                 dout("waiting for sessions to close\n");
2870                 mutex_unlock(&mdsc->mutex);
2871                 wait_for_completion_timeout(&mdsc->session_close_waiters,
2872                                             timeout);
2873                 mutex_lock(&mdsc->mutex);
2874         }
2875
2876         /* tear down remaining sessions */
2877         for (i = 0; i < mdsc->max_sessions; i++) {
2878                 if (mdsc->sessions[i]) {
2879                         session = get_session(mdsc->sessions[i]);
2880                         __unregister_session(mdsc, session);
2881                         mutex_unlock(&mdsc->mutex);
2882                         mutex_lock(&session->s_mutex);
2883                         remove_session_caps(session);
2884                         mutex_unlock(&session->s_mutex);
2885                         ceph_put_mds_session(session);
2886                         mutex_lock(&mdsc->mutex);
2887                 }
2888         }
2889
2890         WARN_ON(!list_empty(&mdsc->cap_delay_list));
2891
2892         mutex_unlock(&mdsc->mutex);
2893
2894         ceph_cleanup_empty_realms(mdsc);
2895
2896         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2897
2898         dout("stopped\n");
2899 }
2900
2901 void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
2902 {
2903         dout("stop\n");
2904         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2905         if (mdsc->mdsmap)
2906                 ceph_mdsmap_destroy(mdsc->mdsmap);
2907         kfree(mdsc->sessions);
2908 }
2909
2910
2911 /*
2912  * handle mds map update.
2913  */
2914 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
2915 {
2916         u32 epoch;
2917         u32 maplen;
2918         void *p = msg->front.iov_base;
2919         void *end = p + msg->front.iov_len;
2920         struct ceph_mdsmap *newmap, *oldmap;
2921         struct ceph_fsid fsid;
2922         int err = -EINVAL;
2923
2924         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
2925         ceph_decode_copy(&p, &fsid, sizeof(fsid));
2926         if (ceph_check_fsid(mdsc->client, &fsid) < 0)
2927                 return;
2928         epoch = ceph_decode_32(&p);
2929         maplen = ceph_decode_32(&p);
2930         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
2931
2932         /* do we need it? */
2933         ceph_monc_got_mdsmap(&mdsc->client->monc, epoch);
2934         mutex_lock(&mdsc->mutex);
2935         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
2936                 dout("handle_map epoch %u <= our %u\n",
2937                      epoch, mdsc->mdsmap->m_epoch);
2938                 mutex_unlock(&mdsc->mutex);
2939                 return;
2940         }
2941
2942         newmap = ceph_mdsmap_decode(&p, end);
2943         if (IS_ERR(newmap)) {
2944                 err = PTR_ERR(newmap);
2945                 goto bad_unlock;
2946         }
2947
2948         /* swap into place */
2949         if (mdsc->mdsmap) {
2950                 oldmap = mdsc->mdsmap;
2951                 mdsc->mdsmap = newmap;
2952                 check_new_map(mdsc, newmap, oldmap);
2953                 ceph_mdsmap_destroy(oldmap);
2954         } else {
2955                 mdsc->mdsmap = newmap;  /* first mds map */
2956         }
2957         mdsc->client->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
2958
2959         __wake_requests(mdsc, &mdsc->waiting_for_map);
2960
2961         mutex_unlock(&mdsc->mutex);
2962         schedule_delayed(mdsc);
2963         return;
2964
2965 bad_unlock:
2966         mutex_unlock(&mdsc->mutex);
2967 bad:
2968         pr_err("error decoding mdsmap %d\n", err);
2969         return;
2970 }
2971
2972 static struct ceph_connection *con_get(struct ceph_connection *con)
2973 {
2974         struct ceph_mds_session *s = con->private;
2975
2976         if (get_session(s)) {
2977                 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
2978                 return con;
2979         }
2980         dout("mdsc con_get %p FAIL\n", s);
2981         return NULL;
2982 }
2983
2984 static void con_put(struct ceph_connection *con)
2985 {
2986         struct ceph_mds_session *s = con->private;
2987
2988         ceph_put_mds_session(s);
2989         dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref));
2990 }
2991
2992 /*
2993  * if the client is unresponsive for long enough, the mds will kill
2994  * the session entirely.
2995  */
2996 static void peer_reset(struct ceph_connection *con)
2997 {
2998         struct ceph_mds_session *s = con->private;
2999         struct ceph_mds_client *mdsc = s->s_mdsc;
3000
3001         pr_warning("mds%d closed our session\n", s->s_mds);
3002         send_mds_reconnect(mdsc, s);
3003 }
3004
3005 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3006 {
3007         struct ceph_mds_session *s = con->private;
3008         struct ceph_mds_client *mdsc = s->s_mdsc;
3009         int type = le16_to_cpu(msg->hdr.type);
3010
3011         mutex_lock(&mdsc->mutex);
3012         if (__verify_registered_session(mdsc, s) < 0) {
3013                 mutex_unlock(&mdsc->mutex);
3014                 goto out;
3015         }
3016         mutex_unlock(&mdsc->mutex);
3017
3018         switch (type) {
3019         case CEPH_MSG_MDS_MAP:
3020                 ceph_mdsc_handle_map(mdsc, msg);
3021                 break;
3022         case CEPH_MSG_CLIENT_SESSION:
3023                 handle_session(s, msg);
3024                 break;
3025         case CEPH_MSG_CLIENT_REPLY:
3026                 handle_reply(s, msg);
3027                 break;
3028         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3029                 handle_forward(mdsc, s, msg);
3030                 break;
3031         case CEPH_MSG_CLIENT_CAPS:
3032                 ceph_handle_caps(s, msg);
3033                 break;
3034         case CEPH_MSG_CLIENT_SNAP:
3035                 ceph_handle_snap(mdsc, s, msg);
3036                 break;
3037         case CEPH_MSG_CLIENT_LEASE:
3038                 handle_lease(mdsc, s, msg);
3039                 break;
3040
3041         default:
3042                 pr_err("received unknown message type %d %s\n", type,
3043                        ceph_msg_type_name(type));
3044         }
3045 out:
3046         ceph_msg_put(msg);
3047 }
3048
3049 /*
3050  * authentication
3051  */
3052 static int get_authorizer(struct ceph_connection *con,
3053                           void **buf, int *len, int *proto,
3054                           void **reply_buf, int *reply_len, int force_new)
3055 {
3056         struct ceph_mds_session *s = con->private;
3057         struct ceph_mds_client *mdsc = s->s_mdsc;
3058         struct ceph_auth_client *ac = mdsc->client->monc.auth;
3059         int ret = 0;
3060
3061         if (force_new && s->s_authorizer) {
3062                 ac->ops->destroy_authorizer(ac, s->s_authorizer);
3063                 s->s_authorizer = NULL;
3064         }
3065         if (s->s_authorizer == NULL) {
3066                 if (ac->ops->create_authorizer) {
3067                         ret = ac->ops->create_authorizer(
3068                                 ac, CEPH_ENTITY_TYPE_MDS,
3069                                 &s->s_authorizer,
3070                                 &s->s_authorizer_buf,
3071                                 &s->s_authorizer_buf_len,
3072                                 &s->s_authorizer_reply_buf,
3073                                 &s->s_authorizer_reply_buf_len);
3074                         if (ret)
3075                                 return ret;
3076                 }
3077         }
3078
3079         *proto = ac->protocol;
3080         *buf = s->s_authorizer_buf;
3081         *len = s->s_authorizer_buf_len;
3082         *reply_buf = s->s_authorizer_reply_buf;
3083         *reply_len = s->s_authorizer_reply_buf_len;
3084         return 0;
3085 }
3086
3087
3088 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3089 {
3090         struct ceph_mds_session *s = con->private;
3091         struct ceph_mds_client *mdsc = s->s_mdsc;
3092         struct ceph_auth_client *ac = mdsc->client->monc.auth;
3093
3094         return ac->ops->verify_authorizer_reply(ac, s->s_authorizer, len);
3095 }
3096
3097 static int invalidate_authorizer(struct ceph_connection *con)
3098 {
3099         struct ceph_mds_session *s = con->private;
3100         struct ceph_mds_client *mdsc = s->s_mdsc;
3101         struct ceph_auth_client *ac = mdsc->client->monc.auth;
3102
3103         if (ac->ops->invalidate_authorizer)
3104                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3105
3106         return ceph_monc_validate_auth(&mdsc->client->monc);
3107 }
3108
3109 const static struct ceph_connection_operations mds_con_ops = {
3110         .get = con_get,
3111         .put = con_put,
3112         .dispatch = dispatch,
3113         .get_authorizer = get_authorizer,
3114         .verify_authorizer_reply = verify_authorizer_reply,
3115         .invalidate_authorizer = invalidate_authorizer,
3116         .peer_reset = peer_reset,
3117 };
3118
3119
3120
3121
3122 /* eof */