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  37  * are no longer valid.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         dout("removing cap %p, ci is %p, inode is %p\n",
 806              cap, ci, &ci->vfs_inode);
 807         ceph_remove_cap(cap);
 808         return 0;
 809 }
 810
 811 /*
 812  * caller must hold session s_mutex
 813  */
 814 static void remove_session_caps(struct ceph_mds_session *session)
 815 {
 816         dout("remove_session_caps on %p\n", session);
 817         iterate_session_caps(session, remove_session_caps_cb, NULL);
 818         BUG_ON(session->s_nr_caps > 0);
 819         cleanup_cap_releases(session);
 820 }
 821
 822 /*
 823  * wake up any threads waiting on this session's caps.  if the cap is
 824  * old (didn't get renewed on the client reconnect), remove it now.
 825  *
 826  * caller must hold s_mutex.
 827  */
 828 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
 829                               void *arg)
 830 {
 831         struct ceph_inode_info *ci = ceph_inode(inode);
 832
 833         wake_up(&ci->i_cap_wq);
 834         if (arg) {
 835                 spin_lock(&inode->i_lock);
 836                 ci->i_wanted_max_size = 0;
 837                 ci->i_requested_max_size = 0;
 838                 spin_unlock(&inode->i_lock);
 839         }
 840         return 0;
 841 }
 842
 843 static void wake_up_session_caps(struct ceph_mds_session *session,
 844                                  int reconnect)
 845 {
 846         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
 847         iterate_session_caps(session, wake_up_session_cb,
 848                              (void *)(unsigned long)reconnect);
 849 }
 850
 851 /*
 852  * Send periodic message to MDS renewing all currently held caps.  The
 853  * ack will reset the expiration for all caps from this session.
 854  *
 855  * caller holds s_mutex
 856  */
 857 static int send_renew_caps(struct ceph_mds_client *mdsc,
 858                            struct ceph_mds_session *session)
 859 {
 860         struct ceph_msg *msg;
 861         int state;
 862
 863         if (time_after_eq(jiffies, session->s_cap_ttl) &&
 864             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
 865                 pr_info("mds%d caps stale\n", session->s_mds);
 866         session->s_renew_requested = jiffies;
 867
 868         /* do not try to renew caps until a recovering mds has reconnected
 869          * with its clients. */
 870         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
 871         if (state < CEPH_MDS_STATE_RECONNECT) {
 872                 dout("send_renew_caps ignoring mds%d (%s)\n",
 873                      session->s_mds, ceph_mds_state_name(state));
 874                 return 0;
 875         }
 876
 877         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
 878                 ceph_mds_state_name(state));
 879         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
 880                                  ++session->s_renew_seq);
 881         if (!msg)
 882                 return -ENOMEM;
 883         ceph_con_send(&session->s_con, msg);
 884         return 0;
 885 }
 886
 887 /*
 888  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
 889  *
 890  * Called under session->s_mutex
 891  */
 892 static void renewed_caps(struct ceph_mds_client *mdsc,
 893                          struct ceph_mds_session *session, int is_renew)
 894 {
 895         int was_stale;
 896         int wake = 0;
 897
 898         spin_lock(&session->s_cap_lock);
 899         was_stale = is_renew && (session->s_cap_ttl == 0 ||
 900                                  time_after_eq(jiffies, session->s_cap_ttl));
 901
 902         session->s_cap_ttl = session->s_renew_requested +
 903                 mdsc->mdsmap->m_session_timeout*HZ;
 904
 905         if (was_stale) {
 906                 if (time_before(jiffies, session->s_cap_ttl)) {
 907                         pr_info("mds%d caps renewed\n", session->s_mds);
 908                         wake = 1;
 909                 } else {
 910                         pr_info("mds%d caps still stale\n", session->s_mds);
 911                 }
 912         }
 913         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
 914              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
 915              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
 916         spin_unlock(&session->s_cap_lock);
 917
 918         if (wake)
 919                 wake_up_session_caps(session, 0);
 920 }
 921
 922 /*
 923  * send a session close request
 924  */
 925 static int request_close_session(struct ceph_mds_client *mdsc,
 926                                  struct ceph_mds_session *session)
 927 {
 928         struct ceph_msg *msg;
 929
 930         dout("request_close_session mds%d state %s seq %lld\n",
 931              session->s_mds, session_state_name(session->s_state),
 932              session->s_seq);
 933         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
 934         if (!msg)
 935                 return -ENOMEM;
 936         ceph_con_send(&session->s_con, msg);
 937         return 0;
 938 }
 939
 940 /*
 941  * Called with s_mutex held.
 942  */
 943 static int __close_session(struct ceph_mds_client *mdsc,
 944                          struct ceph_mds_session *session)
 945 {
 946         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
 947                 return 0;
 948         session->s_state = CEPH_MDS_SESSION_CLOSING;
 949         return request_close_session(mdsc, session);
 950 }
 951
 952 /*
 953  * Trim old(er) caps.
 954  *
 955  * Because we can't cache an inode without one or more caps, we do
 956  * this indirectly: if a cap is unused, we prune its aliases, at which
 957  * point the inode will hopefully get dropped to.
 958  *
 959  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
 960  * memory pressure from the MDS, though, so it needn't be perfect.
 961  */
 962 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
 963 {
 964         struct ceph_mds_session *session = arg;
 965         struct ceph_inode_info *ci = ceph_inode(inode);
 966         int used, oissued, mine;
 967
 968         if (session->s_trim_caps <= 0)
 969                 return -1;
 970
 971         spin_lock(&inode->i_lock);
 972         mine = cap->issued | cap->implemented;
 973         used = __ceph_caps_used(ci);
 974         oissued = __ceph_caps_issued_other(ci, cap);
 975
 976         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
 977              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
 978              ceph_cap_string(used));
 979         if (ci->i_dirty_caps)
 980                 goto out;   /* dirty caps */
 981         if ((used & ~oissued) & mine)
 982                 goto out;   /* we need these caps */
 983
 984         session->s_trim_caps--;
 985         if (oissued) {
 986                 /* we aren't the only cap.. just remove us */
 987                 __ceph_remove_cap(cap);
 988         } else {
 989                 /* try to drop referring dentries */
 990                 spin_unlock(&inode->i_lock);
 991                 d_prune_aliases(inode);
 992                 dout("trim_caps_cb %p cap %p  pruned, count now %d\n",
 993                      inode, cap, atomic_read(&inode->i_count));
 994                 return 0;
 995         }
 996
 997 out:
 998         spin_unlock(&inode->i_lock);
 999         return 0;
1000 }
1001
1002 /*
1003  * Trim session cap count down to some max number.
1004  */
1005 static int trim_caps(struct ceph_mds_client *mdsc,
1006                      struct ceph_mds_session *session,
1007                      int max_caps)
1008 {
1009         int trim_caps = session->s_nr_caps - max_caps;
1010
1011         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1012              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1013         if (trim_caps > 0) {
1014                 session->s_trim_caps = trim_caps;
1015                 iterate_session_caps(session, trim_caps_cb, session);
1016                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1017                      session->s_mds, session->s_nr_caps, max_caps,
1018                         trim_caps - session->s_trim_caps);
1019                 session->s_trim_caps = 0;
1020         }
1021         return 0;
1022 }
1023
1024 /*
1025  * Allocate cap_release messages.  If there is a partially full message
1026  * in the queue, try to allocate enough to cover it's remainder, so that
1027  * we can send it immediately.
1028  *
1029  * Called under s_mutex.
1030  */
1031 static int add_cap_releases(struct ceph_mds_client *mdsc,
1032                             struct ceph_mds_session *session,
1033                             int extra)
1034 {
1035         struct ceph_msg *msg;
1036         struct ceph_mds_cap_release *head;
1037         int err = -ENOMEM;
1038
1039         if (extra < 0)
1040                 extra = mdsc->client->mount_args->cap_release_safety;
1041
1042         spin_lock(&session->s_cap_lock);
1043
1044         if (!list_empty(&session->s_cap_releases)) {
1045                 msg = list_first_entry(&session->s_cap_releases,
1046                                        struct ceph_msg,
1047                                  list_head);
1048                 head = msg->front.iov_base;
1049                 extra += CEPH_CAPS_PER_RELEASE - le32_to_cpu(head->num);
1050         }
1051
1052         while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1053                 spin_unlock(&session->s_cap_lock);
1054                 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE);
1055                 if (!msg)
1056                         goto out_unlocked;
1057                 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1058                      (int)msg->front.iov_len);
1059                 head = msg->front.iov_base;
1060                 head->num = cpu_to_le32(0);
1061                 msg->front.iov_len = sizeof(*head);
1062                 spin_lock(&session->s_cap_lock);
1063                 list_add(&msg->list_head, &session->s_cap_releases);
1064                 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1065         }
1066
1067         if (!list_empty(&session->s_cap_releases)) {
1068                 msg = list_first_entry(&session->s_cap_releases,
1069                                        struct ceph_msg,
1070                                        list_head);
1071                 head = msg->front.iov_base;
1072                 if (head->num) {
1073                         dout(" queueing non-full %p (%d)\n", msg,
1074                              le32_to_cpu(head->num));
1075                         list_move_tail(&msg->list_head,
1076                                       &session->s_cap_releases_done);
1077                         session->s_num_cap_releases -=
1078                                 CEPH_CAPS_PER_RELEASE - le32_to_cpu(head->num);
1079                 }
1080         }
1081         err = 0;
1082         spin_unlock(&session->s_cap_lock);
1083 out_unlocked:
1084         return err;
1085 }
1086
1087 /*
1088  * flush all dirty inode data to disk.
1089  *
1090  * returns true if we've flushed through want_flush_seq
1091  */
1092 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1093 {
1094         int mds, ret = 1;
1095
1096         dout("check_cap_flush want %lld\n", want_flush_seq);
1097         mutex_lock(&mdsc->mutex);
1098         for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1099                 struct ceph_mds_session *session = mdsc->sessions[mds];
1100
1101                 if (!session)
1102                         continue;
1103                 get_session(session);
1104                 mutex_unlock(&mdsc->mutex);
1105
1106                 mutex_lock(&session->s_mutex);
1107                 if (!list_empty(&session->s_cap_flushing)) {
1108                         struct ceph_inode_info *ci =
1109                                 list_entry(session->s_cap_flushing.next,
1110                                            struct ceph_inode_info,
1111                                            i_flushing_item);
1112                         struct inode *inode = &ci->vfs_inode;
1113
1114                         spin_lock(&inode->i_lock);
1115                         if (ci->i_cap_flush_seq <= want_flush_seq) {
1116                                 dout("check_cap_flush still flushing %p "
1117                                      "seq %lld <= %lld to mds%d\n", inode,
1118                                      ci->i_cap_flush_seq, want_flush_seq,
1119                                      session->s_mds);
1120                                 ret = 0;
1121                         }
1122                         spin_unlock(&inode->i_lock);
1123                 }
1124                 mutex_unlock(&session->s_mutex);
1125                 ceph_put_mds_session(session);
1126
1127                 if (!ret)
1128                         return ret;
1129                 mutex_lock(&mdsc->mutex);
1130         }
1131
1132         mutex_unlock(&mdsc->mutex);
1133         dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1134         return ret;
1135 }
1136
1137 /*
1138  * called under s_mutex
1139  */
1140 static void send_cap_releases(struct ceph_mds_client *mdsc,
1141                        struct ceph_mds_session *session)
1142 {
1143         struct ceph_msg *msg;
1144
1145         dout("send_cap_releases mds%d\n", session->s_mds);
1146         spin_lock(&session->s_cap_lock);
1147         while (!list_empty(&session->s_cap_releases_done)) {
1148                 msg = list_first_entry(&session->s_cap_releases_done,
1149                                  struct ceph_msg, list_head);
1150                 list_del_init(&msg->list_head);
1151                 spin_unlock(&session->s_cap_lock);
1152                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1153                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1154                 ceph_con_send(&session->s_con, msg);
1155                 spin_lock(&session->s_cap_lock);
1156         }
1157         spin_unlock(&session->s_cap_lock);
1158 }
1159
1160 static void discard_cap_releases(struct ceph_mds_client *mdsc,
1161                                  struct ceph_mds_session *session)
1162 {
1163         struct ceph_msg *msg;
1164         struct ceph_mds_cap_release *head;
1165         unsigned num;
1166
1167         dout("discard_cap_releases mds%d\n", session->s_mds);
1168         spin_lock(&session->s_cap_lock);
1169
1170         /* zero out the in-progress message */
1171         msg = list_first_entry(&session->s_cap_releases,
1172                                struct ceph_msg, list_head);
1173         head = msg->front.iov_base;
1174         num = le32_to_cpu(head->num);
1175         dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg, num);
1176         head->num = cpu_to_le32(0);
1177         session->s_num_cap_releases += num;
1178
1179         /* requeue completed messages */
1180         while (!list_empty(&session->s_cap_releases_done)) {
1181                 msg = list_first_entry(&session->s_cap_releases_done,
1182                                  struct ceph_msg, list_head);
1183                 list_del_init(&msg->list_head);
1184
1185                 head = msg->front.iov_base;
1186                 num = le32_to_cpu(head->num);
1187                 dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg,
1188                      num);
1189                 session->s_num_cap_releases += num;
1190                 head->num = cpu_to_le32(0);
1191                 msg->front.iov_len = sizeof(*head);
1192                 list_add(&msg->list_head, &session->s_cap_releases);
1193         }
1194
1195         spin_unlock(&session->s_cap_lock);
1196 }
1197
1198 /*
1199  * requests
1200  */
1201
1202 /*
1203  * Create an mds request.
1204  */
1205 struct ceph_mds_request *
1206 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1207 {
1208         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1209
1210         if (!req)
1211                 return ERR_PTR(-ENOMEM);
1212
1213         mutex_init(&req->r_fill_mutex);
1214         req->r_started = jiffies;
1215         req->r_resend_mds = -1;
1216         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1217         req->r_fmode = -1;
1218         kref_init(&req->r_kref);
1219         INIT_LIST_HEAD(&req->r_wait);
1220         init_completion(&req->r_completion);
1221         init_completion(&req->r_safe_completion);
1222         INIT_LIST_HEAD(&req->r_unsafe_item);
1223
1224         req->r_op = op;
1225         req->r_direct_mode = mode;
1226         return req;
1227 }
1228
1229 /*
1230  * return oldest (lowest) request, tid in request tree, 0 if none.
1231  *
1232  * called under mdsc->mutex.
1233  */
1234 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1235 {
1236         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1237                 return NULL;
1238         return rb_entry(rb_first(&mdsc->request_tree),
1239                         struct ceph_mds_request, r_node);
1240 }
1241
1242 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1243 {
1244         struct ceph_mds_request *req = __get_oldest_req(mdsc);
1245
1246         if (req)
1247                 return req->r_tid;
1248         return 0;
1249 }
1250
1251 /*
1252  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1253  * on build_path_from_dentry in fs/cifs/dir.c.
1254  *
1255  * If @stop_on_nosnap, generate path relative to the first non-snapped
1256  * inode.
1257  *
1258  * Encode hidden .snap dirs as a double /, i.e.
1259  *   foo/.snap/bar -> foo//bar
1260  */
1261 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1262                            int stop_on_nosnap)
1263 {
1264         struct dentry *temp;
1265         char *path;
1266         int len, pos;
1267
1268         if (dentry == NULL)
1269                 return ERR_PTR(-EINVAL);
1270
1271 retry:
1272         len = 0;
1273         for (temp = dentry; !IS_ROOT(temp);) {
1274                 struct inode *inode = temp->d_inode;
1275                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1276                         len++;  /* slash only */
1277                 else if (stop_on_nosnap && inode &&
1278                          ceph_snap(inode) == CEPH_NOSNAP)
1279                         break;
1280                 else
1281                         len += 1 + temp->d_name.len;
1282                 temp = temp->d_parent;
1283                 if (temp == NULL) {
1284                         pr_err("build_path_dentry corrupt dentry %p\n", dentry);
1285                         return ERR_PTR(-EINVAL);
1286                 }
1287         }
1288         if (len)
1289                 len--;  /* no leading '/' */
1290
1291         path = kmalloc(len+1, GFP_NOFS);
1292         if (path == NULL)
1293                 return ERR_PTR(-ENOMEM);
1294         pos = len;
1295         path[pos] = 0;  /* trailing null */
1296         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1297                 struct inode *inode = temp->d_inode;
1298
1299                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1300                         dout("build_path path+%d: %p SNAPDIR\n",
1301                              pos, temp);
1302                 } else if (stop_on_nosnap && inode &&
1303                            ceph_snap(inode) == CEPH_NOSNAP) {
1304                         break;
1305                 } else {
1306                         pos -= temp->d_name.len;
1307                         if (pos < 0)
1308                                 break;
1309                         strncpy(path + pos, temp->d_name.name,
1310                                 temp->d_name.len);
1311                 }
1312                 if (pos)
1313                         path[--pos] = '/';
1314                 temp = temp->d_parent;
1315                 if (temp == NULL) {
1316                         pr_err("build_path corrupt dentry\n");
1317                         kfree(path);
1318                         return ERR_PTR(-EINVAL);
1319                 }
1320         }
1321         if (pos != 0) {
1322                 pr_err("build_path did not end path lookup where "
1323                        "expected, namelen is %d, pos is %d\n", len, pos);
1324                 /* presumably this is only possible if racing with a
1325                    rename of one of the parent directories (we can not
1326                    lock the dentries above us to prevent this, but
1327                    retrying should be harmless) */
1328                 kfree(path);
1329                 goto retry;
1330         }
1331
1332         *base = ceph_ino(temp->d_inode);
1333         *plen = len;
1334         dout("build_path on %p %d built %llx '%.*s'\n",
1335              dentry, atomic_read(&dentry->d_count), *base, len, path);
1336         return path;
1337 }
1338
1339 static int build_dentry_path(struct dentry *dentry,
1340                              const char **ppath, int *ppathlen, u64 *pino,
1341                              int *pfreepath)
1342 {
1343         char *path;
1344
1345         if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1346                 *pino = ceph_ino(dentry->d_parent->d_inode);
1347                 *ppath = dentry->d_name.name;
1348                 *ppathlen = dentry->d_name.len;
1349                 return 0;
1350         }
1351         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1352         if (IS_ERR(path))
1353                 return PTR_ERR(path);
1354         *ppath = path;
1355         *pfreepath = 1;
1356         return 0;
1357 }
1358
1359 static int build_inode_path(struct inode *inode,
1360                             const char **ppath, int *ppathlen, u64 *pino,
1361                             int *pfreepath)
1362 {
1363         struct dentry *dentry;
1364         char *path;
1365
1366         if (ceph_snap(inode) == CEPH_NOSNAP) {
1367                 *pino = ceph_ino(inode);
1368                 *ppathlen = 0;
1369                 return 0;
1370         }
1371         dentry = d_find_alias(inode);
1372         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1373         dput(dentry);
1374         if (IS_ERR(path))
1375                 return PTR_ERR(path);
1376         *ppath = path;
1377         *pfreepath = 1;
1378         return 0;
1379 }
1380
1381 /*
1382  * request arguments may be specified via an inode *, a dentry *, or
1383  * an explicit ino+path.
1384  */
1385 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1386                                   const char *rpath, u64 rino,
1387                                   const char **ppath, int *pathlen,
1388                                   u64 *ino, int *freepath)
1389 {
1390         int r = 0;
1391
1392         if (rinode) {
1393                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1394                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1395                      ceph_snap(rinode));
1396         } else if (rdentry) {
1397                 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1398                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1399                      *ppath);
1400         } else if (rpath) {
1401                 *ino = rino;
1402                 *ppath = rpath;
1403                 *pathlen = strlen(rpath);
1404                 dout(" path %.*s\n", *pathlen, rpath);
1405         }
1406
1407         return r;
1408 }
1409
1410 /*
1411  * called under mdsc->mutex
1412  */
1413 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1414                                                struct ceph_mds_request *req,
1415                                                int mds)
1416 {
1417         struct ceph_msg *msg;
1418         struct ceph_mds_request_head *head;
1419         const char *path1 = NULL;
1420         const char *path2 = NULL;
1421         u64 ino1 = 0, ino2 = 0;
1422         int pathlen1 = 0, pathlen2 = 0;
1423         int freepath1 = 0, freepath2 = 0;
1424         int len;
1425         u16 releases;
1426         void *p, *end;
1427         int ret;
1428
1429         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1430                               req->r_path1, req->r_ino1.ino,
1431                               &path1, &pathlen1, &ino1, &freepath1);
1432         if (ret < 0) {
1433                 msg = ERR_PTR(ret);
1434                 goto out;
1435         }
1436
1437         ret = set_request_path_attr(NULL, req->r_old_dentry,
1438                               req->r_path2, req->r_ino2.ino,
1439                               &path2, &pathlen2, &ino2, &freepath2);
1440         if (ret < 0) {
1441                 msg = ERR_PTR(ret);
1442                 goto out_free1;
1443         }
1444
1445         len = sizeof(*head) +
1446                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64));
1447
1448         /* calculate (max) length for cap releases */
1449         len += sizeof(struct ceph_mds_request_release) *
1450                 (!!req->r_inode_drop + !!req->r_dentry_drop +
1451                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1452         if (req->r_dentry_drop)
1453                 len += req->r_dentry->d_name.len;
1454         if (req->r_old_dentry_drop)
1455                 len += req->r_old_dentry->d_name.len;
1456
1457         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len);
1458         if (!msg) {
1459                 msg = ERR_PTR(-ENOMEM);
1460                 goto out_free2;
1461         }
1462
1463         msg->hdr.tid = cpu_to_le64(req->r_tid);
1464
1465         head = msg->front.iov_base;
1466         p = msg->front.iov_base + sizeof(*head);
1467         end = msg->front.iov_base + msg->front.iov_len;
1468
1469         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1470         head->op = cpu_to_le32(req->r_op);
1471         head->caller_uid = cpu_to_le32(current_fsuid());
1472         head->caller_gid = cpu_to_le32(current_fsgid());
1473         head->args = req->r_args;
1474
1475         ceph_encode_filepath(&p, end, ino1, path1);
1476         ceph_encode_filepath(&p, end, ino2, path2);
1477
1478         /* cap releases */
1479         releases = 0;
1480         if (req->r_inode_drop)
1481                 releases += ceph_encode_inode_release(&p,
1482                       req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1483                       mds, req->r_inode_drop, req->r_inode_unless, 0);
1484         if (req->r_dentry_drop)
1485                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1486                        mds, req->r_dentry_drop, req->r_dentry_unless);
1487         if (req->r_old_dentry_drop)
1488                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1489                        mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1490         if (req->r_old_inode_drop)
1491                 releases += ceph_encode_inode_release(&p,
1492                       req->r_old_dentry->d_inode,
1493                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1494         head->num_releases = cpu_to_le16(releases);
1495
1496         BUG_ON(p > end);
1497         msg->front.iov_len = p - msg->front.iov_base;
1498         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1499
1500         msg->pages = req->r_pages;
1501         msg->nr_pages = req->r_num_pages;
1502         msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1503         msg->hdr.data_off = cpu_to_le16(0);
1504
1505 out_free2:
1506         if (freepath2)
1507                 kfree((char *)path2);
1508 out_free1:
1509         if (freepath1)
1510                 kfree((char *)path1);
1511 out:
1512         return msg;
1513 }
1514
1515 /*
1516  * called under mdsc->mutex if error, under no mutex if
1517  * success.
1518  */
1519 static void complete_request(struct ceph_mds_client *mdsc,
1520                              struct ceph_mds_request *req)
1521 {
1522         if (req->r_callback)
1523                 req->r_callback(mdsc, req);
1524         else
1525                 complete(&req->r_completion);
1526 }
1527
1528 /*
1529  * called under mdsc->mutex
1530  */
1531 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1532                                   struct ceph_mds_request *req,
1533                                   int mds)
1534 {
1535         struct ceph_mds_request_head *rhead;
1536         struct ceph_msg *msg;
1537         int flags = 0;
1538
1539         req->r_mds = mds;
1540         req->r_attempts++;
1541         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1542              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1543
1544         if (req->r_request) {
1545                 ceph_msg_put(req->r_request);
1546                 req->r_request = NULL;
1547         }
1548         msg = create_request_message(mdsc, req, mds);
1549         if (IS_ERR(msg)) {
1550                 req->r_err = PTR_ERR(msg);
1551                 complete_request(mdsc, req);
1552                 return PTR_ERR(msg);
1553         }
1554         req->r_request = msg;
1555
1556         rhead = msg->front.iov_base;
1557         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1558         if (req->r_got_unsafe)
1559                 flags |= CEPH_MDS_FLAG_REPLAY;
1560         if (req->r_locked_dir)
1561                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1562         rhead->flags = cpu_to_le32(flags);
1563         rhead->num_fwd = req->r_num_fwd;
1564         rhead->num_retry = req->r_attempts - 1;
1565
1566         dout(" r_locked_dir = %p\n", req->r_locked_dir);
1567
1568         if (req->r_target_inode && req->r_got_unsafe)
1569                 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1570         else
1571                 rhead->ino = 0;
1572         return 0;
1573 }
1574
1575 /*
1576  * send request, or put it on the appropriate wait list.
1577  */
1578 static int __do_request(struct ceph_mds_client *mdsc,
1579                         struct ceph_mds_request *req)
1580 {
1581         struct ceph_mds_session *session = NULL;
1582         int mds = -1;
1583         int err = -EAGAIN;
1584
1585         if (req->r_err || req->r_got_result)
1586                 goto out;
1587
1588         if (req->r_timeout &&
1589             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1590                 dout("do_request timed out\n");
1591                 err = -EIO;
1592                 goto finish;
1593         }
1594
1595         mds = __choose_mds(mdsc, req);
1596         if (mds < 0 ||
1597             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1598                 dout("do_request no mds or not active, waiting for map\n");
1599                 list_add(&req->r_wait, &mdsc->waiting_for_map);
1600                 goto out;
1601         }
1602
1603         /* get, open session */
1604         session = __ceph_lookup_mds_session(mdsc, mds);
1605         if (!session) {
1606                 session = register_session(mdsc, mds);
1607                 if (IS_ERR(session)) {
1608                         err = PTR_ERR(session);
1609                         goto finish;
1610                 }
1611         }
1612         dout("do_request mds%d session %p state %s\n", mds, session,
1613              session_state_name(session->s_state));
1614         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1615             session->s_state != CEPH_MDS_SESSION_HUNG) {
1616                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1617                     session->s_state == CEPH_MDS_SESSION_CLOSING)
1618                         __open_session(mdsc, session);
1619                 list_add(&req->r_wait, &session->s_waiting);
1620                 goto out_session;
1621         }
1622
1623         /* send request */
1624         req->r_session = get_session(session);
1625         req->r_resend_mds = -1;   /* forget any previous mds hint */
1626
1627         if (req->r_request_started == 0)   /* note request start time */
1628                 req->r_request_started = jiffies;
1629
1630         err = __prepare_send_request(mdsc, req, mds);
1631         if (!err) {
1632                 ceph_msg_get(req->r_request);
1633                 ceph_con_send(&session->s_con, req->r_request);
1634         }
1635
1636 out_session:
1637         ceph_put_mds_session(session);
1638 out:
1639         return err;
1640
1641 finish:
1642         req->r_err = err;
1643         complete_request(mdsc, req);
1644         goto out;
1645 }
1646
1647 /*
1648  * called under mdsc->mutex
1649  */
1650 static void __wake_requests(struct ceph_mds_client *mdsc,
1651                             struct list_head *head)
1652 {
1653         struct ceph_mds_request *req, *nreq;
1654
1655         list_for_each_entry_safe(req, nreq, head, r_wait) {
1656                 list_del_init(&req->r_wait);
1657                 __do_request(mdsc, req);
1658         }
1659 }
1660
1661 /*
1662  * Wake up threads with requests pending for @mds, so that they can
1663  * resubmit their requests to a possibly different mds.
1664  */
1665 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
1666 {
1667         struct ceph_mds_request *req;
1668         struct rb_node *p;
1669
1670         dout("kick_requests mds%d\n", mds);
1671         for (p = rb_first(&mdsc->request_tree); p; p = rb_next(p)) {
1672                 req = rb_entry(p, struct ceph_mds_request, r_node);
1673                 if (req->r_got_unsafe)
1674                         continue;
1675                 if (req->r_session &&
1676                     req->r_session->s_mds == mds) {
1677                         dout(" kicking tid %llu\n", req->r_tid);
1678                         put_request_session(req);
1679                         __do_request(mdsc, req);
1680                 }
1681         }
1682 }
1683
1684 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
1685                               struct ceph_mds_request *req)
1686 {
1687         dout("submit_request on %p\n", req);
1688         mutex_lock(&mdsc->mutex);
1689         __register_request(mdsc, req, NULL);
1690         __do_request(mdsc, req);
1691         mutex_unlock(&mdsc->mutex);
1692 }
1693
1694 /*
1695  * Synchrously perform an mds request.  Take care of all of the
1696  * session setup, forwarding, retry details.
1697  */
1698 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
1699                          struct inode *dir,
1700                          struct ceph_mds_request *req)
1701 {
1702         int err;
1703
1704         dout("do_request on %p\n", req);
1705
1706         /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
1707         if (req->r_inode)
1708                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
1709         if (req->r_locked_dir)
1710                 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
1711         if (req->r_old_dentry)
1712                 ceph_get_cap_refs(
1713                         ceph_inode(req->r_old_dentry->d_parent->d_inode),
1714                         CEPH_CAP_PIN);
1715
1716         /* issue */
1717         mutex_lock(&mdsc->mutex);
1718         __register_request(mdsc, req, dir);
1719         __do_request(mdsc, req);
1720
1721         if (req->r_err) {
1722                 err = req->r_err;
1723                 __unregister_request(mdsc, req);
1724                 dout("do_request early error %d\n", err);
1725                 goto out;
1726         }
1727
1728         /* wait */
1729         mutex_unlock(&mdsc->mutex);
1730         dout("do_request waiting\n");
1731         if (req->r_timeout) {
1732                 err = (long)wait_for_completion_interruptible_timeout(
1733                         &req->r_completion, req->r_timeout);
1734                 if (err == 0)
1735                         err = -EIO;
1736         } else {
1737                 err = wait_for_completion_interruptible(&req->r_completion);
1738         }
1739         dout("do_request waited, got %d\n", err);
1740         mutex_lock(&mdsc->mutex);
1741
1742         /* only abort if we didn't race with a real reply */
1743         if (req->r_got_result) {
1744                 err = le32_to_cpu(req->r_reply_info.head->result);
1745         } else if (err < 0) {
1746                 dout("aborted request %lld with %d\n", req->r_tid, err);
1747
1748                 /*
1749                  * ensure we aren't running concurrently with
1750                  * ceph_fill_trace or ceph_readdir_prepopulate, which
1751                  * rely on locks (dir mutex) held by our caller.
1752                  */
1753                 mutex_lock(&req->r_fill_mutex);
1754                 req->r_err = err;
1755                 req->r_aborted = true;
1756                 mutex_unlock(&req->r_fill_mutex);
1757
1758                 if (req->r_locked_dir &&
1759                     (req->r_op & CEPH_MDS_OP_WRITE)) {
1760                         struct ceph_inode_info *ci =
1761                                 ceph_inode(req->r_locked_dir);
1762
1763                         dout("aborted, clearing I_COMPLETE on %p, leases\n",
1764                              req->r_locked_dir);
1765                         spin_lock(&req->r_locked_dir->i_lock);
1766                         ci->i_ceph_flags &= ~CEPH_I_COMPLETE;
1767                         ci->i_release_count++;
1768                         spin_unlock(&req->r_locked_dir->i_lock);
1769
1770                         if (req->r_dentry)
1771                                 ceph_invalidate_dentry_lease(req->r_dentry);
1772                         if (req->r_old_dentry)
1773                                 ceph_invalidate_dentry_lease(req->r_old_dentry);
1774                 }
1775         } else {
1776                 err = req->r_err;
1777         }
1778
1779 out:
1780         mutex_unlock(&mdsc->mutex);
1781         dout("do_request %p done, result %d\n", req, err);
1782         return err;
1783 }
1784
1785 /*
1786  * Handle mds reply.
1787  *
1788  * We take the session mutex and parse and process the reply immediately.
1789  * This preserves the logical ordering of replies, capabilities, etc., sent
1790  * by the MDS as they are applied to our local cache.
1791  */
1792 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
1793 {
1794         struct ceph_mds_client *mdsc = session->s_mdsc;
1795         struct ceph_mds_request *req;
1796         struct ceph_mds_reply_head *head = msg->front.iov_base;
1797         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
1798         u64 tid;
1799         int err, result;
1800         int mds = session->s_mds;
1801
1802         if (msg->front.iov_len < sizeof(*head)) {
1803                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
1804                 ceph_msg_dump(msg);
1805                 return;
1806         }
1807
1808         /* get request, session */
1809         tid = le64_to_cpu(msg->hdr.tid);
1810         mutex_lock(&mdsc->mutex);
1811         req = __lookup_request(mdsc, tid);
1812         if (!req) {
1813                 dout("handle_reply on unknown tid %llu\n", tid);
1814                 mutex_unlock(&mdsc->mutex);
1815                 return;
1816         }
1817         dout("handle_reply %p\n", req);
1818
1819         /* correct session? */
1820         if (req->r_session != session) {
1821                 pr_err("mdsc_handle_reply got %llu on session mds%d"
1822                        " not mds%d\n", tid, session->s_mds,
1823                        req->r_session ? req->r_session->s_mds : -1);
1824                 mutex_unlock(&mdsc->mutex);
1825                 goto out;
1826         }
1827
1828         /* dup? */
1829         if ((req->r_got_unsafe && !head->safe) ||
1830             (req->r_got_safe && head->safe)) {
1831                 pr_warning("got a dup %s reply on %llu from mds%d\n",
1832                            head->safe ? "safe" : "unsafe", tid, mds);
1833                 mutex_unlock(&mdsc->mutex);
1834                 goto out;
1835         }
1836
1837         result = le32_to_cpu(head->result);
1838
1839         /*
1840          * Tolerate 2 consecutive ESTALEs from the same mds.
1841          * FIXME: we should be looking at the cap migrate_seq.
1842          */
1843         if (result == -ESTALE) {
1844                 req->r_direct_mode = USE_AUTH_MDS;
1845                 req->r_num_stale++;
1846                 if (req->r_num_stale <= 2) {
1847                         __do_request(mdsc, req);
1848                         mutex_unlock(&mdsc->mutex);
1849                         goto out;
1850                 }
1851         } else {
1852                 req->r_num_stale = 0;
1853         }
1854
1855         if (head->safe) {
1856                 req->r_got_safe = true;
1857                 __unregister_request(mdsc, req);
1858                 complete(&req->r_safe_completion);
1859
1860                 if (req->r_got_unsafe) {
1861                         /*
1862                          * We already handled the unsafe response, now do the
1863                          * cleanup.  No need to examine the response; the MDS
1864                          * doesn't include any result info in the safe
1865                          * response.  And even if it did, there is nothing
1866                          * useful we could do with a revised return value.
1867                          */
1868                         dout("got safe reply %llu, mds%d\n", tid, mds);
1869                         list_del_init(&req->r_unsafe_item);
1870
1871                         /* last unsafe request during umount? */
1872                         if (mdsc->stopping && !__get_oldest_req(mdsc))
1873                                 complete(&mdsc->safe_umount_waiters);
1874                         mutex_unlock(&mdsc->mutex);
1875                         goto out;
1876                 }
1877         } else {
1878                 req->r_got_unsafe = true;
1879                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
1880         }
1881
1882         dout("handle_reply tid %lld result %d\n", tid, result);
1883         rinfo = &req->r_reply_info;
1884         err = parse_reply_info(msg, rinfo);
1885         mutex_unlock(&mdsc->mutex);
1886
1887         mutex_lock(&session->s_mutex);
1888         if (err < 0) {
1889                 pr_err("mdsc_handle_reply got corrupt reply mds%d\n", mds);
1890                 ceph_msg_dump(msg);
1891                 goto out_err;
1892         }
1893
1894         /* snap trace */
1895         if (rinfo->snapblob_len) {
1896                 down_write(&mdsc->snap_rwsem);
1897                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
1898                                rinfo->snapblob + rinfo->snapblob_len,
1899                                le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
1900                 downgrade_write(&mdsc->snap_rwsem);
1901         } else {
1902                 down_read(&mdsc->snap_rwsem);
1903         }
1904
1905         /* insert trace into our cache */
1906         mutex_lock(&req->r_fill_mutex);
1907         err = ceph_fill_trace(mdsc->client->sb, req, req->r_session);
1908         if (err == 0) {
1909                 if (result == 0 && rinfo->dir_nr)
1910                         ceph_readdir_prepopulate(req, req->r_session);
1911                 ceph_unreserve_caps(&req->r_caps_reservation);
1912         }
1913         mutex_unlock(&req->r_fill_mutex);
1914
1915         up_read(&mdsc->snap_rwsem);
1916 out_err:
1917         mutex_lock(&mdsc->mutex);
1918         if (!req->r_aborted) {
1919                 if (err) {
1920                         req->r_err = err;
1921                 } else {
1922                         req->r_reply = msg;
1923                         ceph_msg_get(msg);
1924                         req->r_got_result = true;
1925                 }
1926         } else {
1927                 dout("reply arrived after request %lld was aborted\n", tid);
1928         }
1929         mutex_unlock(&mdsc->mutex);
1930
1931         add_cap_releases(mdsc, req->r_session, -1);
1932         mutex_unlock(&session->s_mutex);
1933
1934         /* kick calling process */
1935         complete_request(mdsc, req);
1936 out:
1937         ceph_mdsc_put_request(req);
1938         return;
1939 }
1940
1941
1942
1943 /*
1944  * handle mds notification that our request has been forwarded.
1945  */
1946 static void handle_forward(struct ceph_mds_client *mdsc,
1947                            struct ceph_mds_session *session,
1948                            struct ceph_msg *msg)
1949 {
1950         struct ceph_mds_request *req;
1951         u64 tid = le64_to_cpu(msg->hdr.tid);
1952         u32 next_mds;
1953         u32 fwd_seq;
1954         int err = -EINVAL;
1955         void *p = msg->front.iov_base;
1956         void *end = p + msg->front.iov_len;
1957
1958         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1959         next_mds = ceph_decode_32(&p);
1960         fwd_seq = ceph_decode_32(&p);
1961
1962         mutex_lock(&mdsc->mutex);
1963         req = __lookup_request(mdsc, tid);
1964         if (!req) {
1965                 dout("forward %llu to mds%d - req dne\n", tid, next_mds);
1966                 goto out;  /* dup reply? */
1967         }
1968
1969         if (fwd_seq <= req->r_num_fwd) {
1970                 dout("forward %llu to mds%d - old seq %d <= %d\n",
1971                      tid, next_mds, req->r_num_fwd, fwd_seq);
1972         } else {
1973                 /* resend. forward race not possible; mds would drop */
1974                 dout("forward %llu to mds%d (we resend)\n", tid, next_mds);
1975                 req->r_num_fwd = fwd_seq;
1976                 req->r_resend_mds = next_mds;
1977                 put_request_session(req);
1978                 __do_request(mdsc, req);
1979         }
1980         ceph_mdsc_put_request(req);
1981 out:
1982         mutex_unlock(&mdsc->mutex);
1983         return;
1984
1985 bad:
1986         pr_err("mdsc_handle_forward decode error err=%d\n", err);
1987 }
1988
1989 /*
1990  * handle a mds session control message
1991  */
1992 static void handle_session(struct ceph_mds_session *session,
1993                            struct ceph_msg *msg)
1994 {
1995         struct ceph_mds_client *mdsc = session->s_mdsc;
1996         u32 op;
1997         u64 seq;
1998         int mds = session->s_mds;
1999         struct ceph_mds_session_head *h = msg->front.iov_base;
2000         int wake = 0;
2001
2002         /* decode */
2003         if (msg->front.iov_len != sizeof(*h))
2004                 goto bad;
2005         op = le32_to_cpu(h->op);
2006         seq = le64_to_cpu(h->seq);
2007
2008         mutex_lock(&mdsc->mutex);
2009         if (op == CEPH_SESSION_CLOSE)
2010                 __unregister_session(mdsc, session);
2011         /* FIXME: this ttl calculation is generous */
2012         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2013         mutex_unlock(&mdsc->mutex);
2014
2015         mutex_lock(&session->s_mutex);
2016
2017         dout("handle_session mds%d %s %p state %s seq %llu\n",
2018              mds, ceph_session_op_name(op), session,
2019              session_state_name(session->s_state), seq);
2020
2021         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2022                 session->s_state = CEPH_MDS_SESSION_OPEN;
2023                 pr_info("mds%d came back\n", session->s_mds);
2024         }
2025
2026         switch (op) {
2027         case CEPH_SESSION_OPEN:
2028                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2029                         pr_info("mds%d reconnect success\n", session->s_mds);
2030                 session->s_state = CEPH_MDS_SESSION_OPEN;
2031                 renewed_caps(mdsc, session, 0);
2032                 wake = 1;
2033                 if (mdsc->stopping)
2034                         __close_session(mdsc, session);
2035                 break;
2036
2037         case CEPH_SESSION_RENEWCAPS:
2038                 if (session->s_renew_seq == seq)
2039                         renewed_caps(mdsc, session, 1);
2040                 break;
2041
2042         case CEPH_SESSION_CLOSE:
2043                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2044                         pr_info("mds%d reconnect denied\n", session->s_mds);
2045                 remove_session_caps(session);
2046                 wake = 1; /* for good measure */
2047                 complete(&mdsc->session_close_waiters);
2048                 kick_requests(mdsc, mds);
2049                 break;
2050
2051         case CEPH_SESSION_STALE:
2052                 pr_info("mds%d caps went stale, renewing\n",
2053                         session->s_mds);
2054                 spin_lock(&session->s_cap_lock);
2055                 session->s_cap_gen++;
2056                 session->s_cap_ttl = 0;
2057                 spin_unlock(&session->s_cap_lock);
2058                 send_renew_caps(mdsc, session);
2059                 break;
2060
2061         case CEPH_SESSION_RECALL_STATE:
2062                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2063                 break;
2064
2065         default:
2066                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2067                 WARN_ON(1);
2068         }
2069
2070         mutex_unlock(&session->s_mutex);
2071         if (wake) {
2072                 mutex_lock(&mdsc->mutex);
2073                 __wake_requests(mdsc, &session->s_waiting);
2074                 mutex_unlock(&mdsc->mutex);
2075         }
2076         return;
2077
2078 bad:
2079         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2080                (int)msg->front.iov_len);
2081         ceph_msg_dump(msg);
2082         return;
2083 }
2084
2085
2086 /*
2087  * called under session->mutex.
2088  */
2089 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2090                                    struct ceph_mds_session *session)
2091 {
2092         struct ceph_mds_request *req, *nreq;
2093         int err;
2094
2095         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2096
2097         mutex_lock(&mdsc->mutex);
2098         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2099                 err = __prepare_send_request(mdsc, req, session->s_mds);
2100                 if (!err) {
2101                         ceph_msg_get(req->r_request);
2102                         ceph_con_send(&session->s_con, req->r_request);
2103                 }
2104         }
2105         mutex_unlock(&mdsc->mutex);
2106 }
2107
2108 /*
2109  * Encode information about a cap for a reconnect with the MDS.
2110  */
2111 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2112                           void *arg)
2113 {
2114         struct ceph_mds_cap_reconnect rec;
2115         struct ceph_inode_info *ci;
2116         struct ceph_pagelist *pagelist = arg;
2117         char *path;
2118         int pathlen, err;
2119         u64 pathbase;
2120         struct dentry *dentry;
2121
2122         ci = cap->ci;
2123
2124         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2125              inode, ceph_vinop(inode), cap, cap->cap_id,
2126              ceph_cap_string(cap->issued));
2127         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2128         if (err)
2129                 return err;
2130
2131         dentry = d_find_alias(inode);
2132         if (dentry) {
2133                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2134                 if (IS_ERR(path)) {
2135                         err = PTR_ERR(path);
2136                         BUG_ON(err);
2137                 }
2138         } else {
2139                 path = NULL;
2140                 pathlen = 0;
2141         }
2142         err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2143         if (err)
2144                 goto out;
2145
2146         spin_lock(&inode->i_lock);
2147         cap->seq = 0;        /* reset cap seq */
2148         cap->issue_seq = 0;  /* and issue_seq */
2149         rec.cap_id = cpu_to_le64(cap->cap_id);
2150         rec.pathbase = cpu_to_le64(pathbase);
2151         rec.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2152         rec.issued = cpu_to_le32(cap->issued);
2153         rec.size = cpu_to_le64(inode->i_size);
2154         ceph_encode_timespec(&rec.mtime, &inode->i_mtime);
2155         ceph_encode_timespec(&rec.atime, &inode->i_atime);
2156         rec.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2157         spin_unlock(&inode->i_lock);
2158
2159         err = ceph_pagelist_append(pagelist, &rec, sizeof(rec));
2160
2161 out:
2162         kfree(path);
2163         dput(dentry);
2164         return err;
2165 }
2166
2167
2168 /*
2169  * If an MDS fails and recovers, clients need to reconnect in order to
2170  * reestablish shared state.  This includes all caps issued through
2171  * this session _and_ the snap_realm hierarchy.  Because it's not
2172  * clear which snap realms the mds cares about, we send everything we
2173  * know about.. that ensures we'll then get any new info the
2174  * recovering MDS might have.
2175  *
2176  * This is a relatively heavyweight operation, but it's rare.
2177  *
2178  * called with mdsc->mutex held.
2179  */
2180 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2181                                struct ceph_mds_session *session)
2182 {
2183         struct ceph_msg *reply;
2184         struct rb_node *p;
2185         int mds = session->s_mds;
2186         int err = -ENOMEM;
2187         struct ceph_pagelist *pagelist;
2188
2189         pr_info("mds%d reconnect start\n", mds);
2190
2191         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2192         if (!pagelist)
2193                 goto fail_nopagelist;
2194         ceph_pagelist_init(pagelist);
2195
2196         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0);
2197         if (!reply)
2198                 goto fail_nomsg;
2199
2200         mutex_lock(&session->s_mutex);
2201         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2202         session->s_seq = 0;
2203
2204         ceph_con_open(&session->s_con,
2205                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2206
2207         /* replay unsafe requests */
2208         replay_unsafe_requests(mdsc, session);
2209
2210         down_read(&mdsc->snap_rwsem);
2211
2212         dout("session %p state %s\n", session,
2213              session_state_name(session->s_state));
2214
2215         /* drop old cap expires; we're about to reestablish that state */
2216         discard_cap_releases(mdsc, session);
2217
2218         /* traverse this session's caps */
2219         err = ceph_pagelist_encode_32(pagelist, session->s_nr_caps);
2220         if (err)
2221                 goto fail;
2222         err = iterate_session_caps(session, encode_caps_cb, pagelist);
2223         if (err < 0)
2224                 goto fail;
2225
2226         /*
2227          * snaprealms.  we provide mds with the ino, seq (version), and
2228          * parent for all of our realms.  If the mds has any newer info,
2229          * it will tell us.
2230          */
2231         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2232                 struct ceph_snap_realm *realm =
2233                         rb_entry(p, struct ceph_snap_realm, node);
2234                 struct ceph_mds_snaprealm_reconnect sr_rec;
2235
2236                 dout(" adding snap realm %llx seq %lld parent %llx\n",
2237                      realm->ino, realm->seq, realm->parent_ino);
2238                 sr_rec.ino = cpu_to_le64(realm->ino);
2239                 sr_rec.seq = cpu_to_le64(realm->seq);
2240                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2241                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2242                 if (err)
2243                         goto fail;
2244         }
2245
2246         reply->pagelist = pagelist;
2247         reply->hdr.data_len = cpu_to_le32(pagelist->length);
2248         reply->nr_pages = calc_pages_for(0, pagelist->length);
2249         ceph_con_send(&session->s_con, reply);
2250
2251         mutex_unlock(&session->s_mutex);
2252
2253         mutex_lock(&mdsc->mutex);
2254         __wake_requests(mdsc, &session->s_waiting);
2255         mutex_unlock(&mdsc->mutex);
2256
2257         up_read(&mdsc->snap_rwsem);
2258         return;
2259
2260 fail:
2261         ceph_msg_put(reply);
2262         up_read(&mdsc->snap_rwsem);
2263         mutex_unlock(&session->s_mutex);
2264 fail_nomsg:
2265         ceph_pagelist_release(pagelist);
2266         kfree(pagelist);
2267 fail_nopagelist:
2268         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
2269         return;
2270 }
2271
2272
2273 /*
2274  * compare old and new mdsmaps, kicking requests
2275  * and closing out old connections as necessary
2276  *
2277  * called under mdsc->mutex.
2278  */
2279 static void check_new_map(struct ceph_mds_client *mdsc,
2280                           struct ceph_mdsmap *newmap,
2281                           struct ceph_mdsmap *oldmap)
2282 {
2283         int i;
2284         int oldstate, newstate;
2285         struct ceph_mds_session *s;
2286
2287         dout("check_new_map new %u old %u\n",
2288              newmap->m_epoch, oldmap->m_epoch);
2289
2290         for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2291                 if (mdsc->sessions[i] == NULL)
2292                         continue;
2293                 s = mdsc->sessions[i];
2294                 oldstate = ceph_mdsmap_get_state(oldmap, i);
2295                 newstate = ceph_mdsmap_get_state(newmap, i);
2296
2297                 dout("check_new_map mds%d state %s -> %s (session %s)\n",
2298                      i, ceph_mds_state_name(oldstate),
2299                      ceph_mds_state_name(newstate),
2300                      session_state_name(s->s_state));
2301
2302                 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
2303                            ceph_mdsmap_get_addr(newmap, i),
2304                            sizeof(struct ceph_entity_addr))) {
2305                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2306                                 /* the session never opened, just close it
2307                                  * out now */
2308                                 __wake_requests(mdsc, &s->s_waiting);
2309                                 __unregister_session(mdsc, s);
2310                         } else {
2311                                 /* just close it */
2312                                 mutex_unlock(&mdsc->mutex);
2313                                 mutex_lock(&s->s_mutex);
2314                                 mutex_lock(&mdsc->mutex);
2315                                 ceph_con_close(&s->s_con);
2316                                 mutex_unlock(&s->s_mutex);
2317                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2318                         }
2319
2320                         /* kick any requests waiting on the recovering mds */
2321                         kick_requests(mdsc, i);
2322                 } else if (oldstate == newstate) {
2323                         continue;  /* nothing new with this mds */
2324                 }
2325
2326                 /*
2327                  * send reconnect?
2328                  */
2329                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2330                     newstate >= CEPH_MDS_STATE_RECONNECT) {
2331                         mutex_unlock(&mdsc->mutex);
2332                         send_mds_reconnect(mdsc, s);
2333                         mutex_lock(&mdsc->mutex);
2334                 }
2335
2336                 /*
2337                  * kick request on any mds that has gone active.
2338                  */
2339                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2340                     newstate >= CEPH_MDS_STATE_ACTIVE) {
2341                         if (oldstate != CEPH_MDS_STATE_CREATING &&
2342                             oldstate != CEPH_MDS_STATE_STARTING)
2343                                 pr_info("mds%d recovery completed\n", s->s_mds);
2344                         kick_requests(mdsc, i);
2345                         ceph_kick_flushing_caps(mdsc, s);
2346                         wake_up_session_caps(s, 1);
2347                 }
2348         }
2349 }
2350
2351
2352
2353 /*
2354  * leases
2355  */
2356
2357 /*
2358  * caller must hold session s_mutex, dentry->d_lock
2359  */
2360 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2361 {
2362         struct ceph_dentry_info *di = ceph_dentry(dentry);
2363
2364         ceph_put_mds_session(di->lease_session);
2365         di->lease_session = NULL;
2366 }
2367
2368 static void handle_lease(struct ceph_mds_client *mdsc,
2369                          struct ceph_mds_session *session,
2370                          struct ceph_msg *msg)
2371 {
2372         struct super_block *sb = mdsc->client->sb;
2373         struct inode *inode;
2374         struct ceph_inode_info *ci;
2375         struct dentry *parent, *dentry;
2376         struct ceph_dentry_info *di;
2377         int mds = session->s_mds;
2378         struct ceph_mds_lease *h = msg->front.iov_base;
2379         struct ceph_vino vino;
2380         int mask;
2381         struct qstr dname;
2382         int release = 0;
2383
2384         dout("handle_lease from mds%d\n", mds);
2385
2386         /* decode */
2387         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2388                 goto bad;
2389         vino.ino = le64_to_cpu(h->ino);
2390         vino.snap = CEPH_NOSNAP;
2391         mask = le16_to_cpu(h->mask);
2392         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2393         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2394         if (dname.len != get_unaligned_le32(h+1))
2395                 goto bad;
2396
2397         mutex_lock(&session->s_mutex);
2398         session->s_seq++;
2399
2400         /* lookup inode */
2401         inode = ceph_find_inode(sb, vino);
2402         dout("handle_lease '%s', mask %d, ino %llx %p\n",
2403              ceph_lease_op_name(h->action), mask, vino.ino, inode);
2404         if (inode == NULL) {
2405                 dout("handle_lease no inode %llx\n", vino.ino);
2406                 goto release;
2407         }
2408         ci = ceph_inode(inode);
2409
2410         /* dentry */
2411         parent = d_find_alias(inode);
2412         if (!parent) {
2413                 dout("no parent dentry on inode %p\n", inode);
2414                 WARN_ON(1);
2415                 goto release;  /* hrm... */
2416         }
2417         dname.hash = full_name_hash(dname.name, dname.len);
2418         dentry = d_lookup(parent, &dname);
2419         dput(parent);
2420         if (!dentry)
2421                 goto release;
2422
2423         spin_lock(&dentry->d_lock);
2424         di = ceph_dentry(dentry);
2425         switch (h->action) {
2426         case CEPH_MDS_LEASE_REVOKE:
2427                 if (di && di->lease_session == session) {
2428                         h->seq = cpu_to_le32(di->lease_seq);
2429                         __ceph_mdsc_drop_dentry_lease(dentry);
2430                 }
2431                 release = 1;
2432                 break;
2433
2434         case CEPH_MDS_LEASE_RENEW:
2435                 if (di && di->lease_session == session &&
2436                     di->lease_gen == session->s_cap_gen &&
2437                     di->lease_renew_from &&
2438                     di->lease_renew_after == 0) {
2439                         unsigned long duration =
2440                                 le32_to_cpu(h->duration_ms) * HZ / 1000;
2441
2442                         di->lease_seq = le32_to_cpu(h->seq);
2443                         dentry->d_time = di->lease_renew_from + duration;
2444                         di->lease_renew_after = di->lease_renew_from +
2445                                 (duration >> 1);
2446                         di->lease_renew_from = 0;
2447                 }
2448                 break;
2449         }
2450         spin_unlock(&dentry->d_lock);
2451         dput(dentry);
2452
2453         if (!release)
2454                 goto out;
2455
2456 release:
2457         /* let's just reuse the same message */
2458         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2459         ceph_msg_get(msg);
2460         ceph_con_send(&session->s_con, msg);
2461
2462 out:
2463         iput(inode);
2464         mutex_unlock(&session->s_mutex);
2465         return;
2466
2467 bad:
2468         pr_err("corrupt lease message\n");
2469         ceph_msg_dump(msg);
2470 }
2471
2472 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
2473                               struct inode *inode,
2474                               struct dentry *dentry, char action,
2475                               u32 seq)
2476 {
2477         struct ceph_msg *msg;
2478         struct ceph_mds_lease *lease;
2479         int len = sizeof(*lease) + sizeof(u32);
2480         int dnamelen = 0;
2481
2482         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2483              inode, dentry, ceph_lease_op_name(action), session->s_mds);
2484         dnamelen = dentry->d_name.len;
2485         len += dnamelen;
2486
2487         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len);
2488         if (!msg)
2489                 return;
2490         lease = msg->front.iov_base;
2491         lease->action = action;
2492         lease->mask = cpu_to_le16(CEPH_LOCK_DN);
2493         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
2494         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
2495         lease->seq = cpu_to_le32(seq);
2496         put_unaligned_le32(dnamelen, lease + 1);
2497         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
2498
2499         /*
2500          * if this is a preemptive lease RELEASE, no need to
2501          * flush request stream, since the actual request will
2502          * soon follow.
2503          */
2504         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
2505
2506         ceph_con_send(&session->s_con, msg);
2507 }
2508
2509 /*
2510  * Preemptively release a lease we expect to invalidate anyway.
2511  * Pass @inode always, @dentry is optional.
2512  */
2513 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2514                              struct dentry *dentry, int mask)
2515 {
2516         struct ceph_dentry_info *di;
2517         struct ceph_mds_session *session;
2518         u32 seq;
2519
2520         BUG_ON(inode == NULL);
2521         BUG_ON(dentry == NULL);
2522         BUG_ON(mask != CEPH_LOCK_DN);
2523
2524         /* is dentry lease valid? */
2525         spin_lock(&dentry->d_lock);
2526         di = ceph_dentry(dentry);
2527         if (!di || !di->lease_session ||
2528             di->lease_session->s_mds < 0 ||
2529             di->lease_gen != di->lease_session->s_cap_gen ||
2530             !time_before(jiffies, dentry->d_time)) {
2531                 dout("lease_release inode %p dentry %p -- "
2532                      "no lease on %d\n",
2533                      inode, dentry, mask);
2534                 spin_unlock(&dentry->d_lock);
2535                 return;
2536         }
2537
2538         /* we do have a lease on this dentry; note mds and seq */
2539         session = ceph_get_mds_session(di->lease_session);
2540         seq = di->lease_seq;
2541         __ceph_mdsc_drop_dentry_lease(dentry);
2542         spin_unlock(&dentry->d_lock);
2543
2544         dout("lease_release inode %p dentry %p mask %d to mds%d\n",
2545              inode, dentry, mask, session->s_mds);
2546         ceph_mdsc_lease_send_msg(session, inode, dentry,
2547                                  CEPH_MDS_LEASE_RELEASE, seq);
2548         ceph_put_mds_session(session);
2549 }
2550
2551 /*
2552  * drop all leases (and dentry refs) in preparation for umount
2553  */
2554 static void drop_leases(struct ceph_mds_client *mdsc)
2555 {
2556         int i;
2557
2558         dout("drop_leases\n");
2559         mutex_lock(&mdsc->mutex);
2560         for (i = 0; i < mdsc->max_sessions; i++) {
2561                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2562                 if (!s)
2563                         continue;
2564                 mutex_unlock(&mdsc->mutex);
2565                 mutex_lock(&s->s_mutex);
2566                 mutex_unlock(&s->s_mutex);
2567                 ceph_put_mds_session(s);
2568                 mutex_lock(&mdsc->mutex);
2569         }
2570         mutex_unlock(&mdsc->mutex);
2571 }
2572
2573
2574
2575 /*
2576  * delayed work -- periodically trim expired leases, renew caps with mds
2577  */
2578 static void schedule_delayed(struct ceph_mds_client *mdsc)
2579 {
2580         int delay = 5;
2581         unsigned hz = round_jiffies_relative(HZ * delay);
2582         schedule_delayed_work(&mdsc->delayed_work, hz);
2583 }
2584
2585 static void delayed_work(struct work_struct *work)
2586 {
2587         int i;
2588         struct ceph_mds_client *mdsc =
2589                 container_of(work, struct ceph_mds_client, delayed_work.work);
2590         int renew_interval;
2591         int renew_caps;
2592
2593         dout("mdsc delayed_work\n");
2594         ceph_check_delayed_caps(mdsc);
2595
2596         mutex_lock(&mdsc->mutex);
2597         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
2598         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
2599                                    mdsc->last_renew_caps);
2600         if (renew_caps)
2601                 mdsc->last_renew_caps = jiffies;
2602
2603         for (i = 0; i < mdsc->max_sessions; i++) {
2604                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2605                 if (s == NULL)
2606                         continue;
2607                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
2608                         dout("resending session close request for mds%d\n",
2609                              s->s_mds);
2610                         request_close_session(mdsc, s);
2611                         ceph_put_mds_session(s);
2612                         continue;
2613                 }
2614                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
2615                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
2616                                 s->s_state = CEPH_MDS_SESSION_HUNG;
2617                                 pr_info("mds%d hung\n", s->s_mds);
2618                         }
2619                 }
2620                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
2621                         /* this mds is failed or recovering, just wait */
2622                         ceph_put_mds_session(s);
2623                         continue;
2624                 }
2625                 mutex_unlock(&mdsc->mutex);
2626
2627                 mutex_lock(&s->s_mutex);
2628                 if (renew_caps)
2629                         send_renew_caps(mdsc, s);
2630                 else
2631                         ceph_con_keepalive(&s->s_con);
2632                 add_cap_releases(mdsc, s, -1);
2633                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2634                     s->s_state == CEPH_MDS_SESSION_HUNG)
2635                         send_cap_releases(mdsc, s);
2636                 mutex_unlock(&s->s_mutex);
2637                 ceph_put_mds_session(s);
2638
2639                 mutex_lock(&mdsc->mutex);
2640         }
2641         mutex_unlock(&mdsc->mutex);
2642
2643         schedule_delayed(mdsc);
2644 }
2645
2646
2647 int ceph_mdsc_init(struct ceph_mds_client *mdsc, struct ceph_client *client)
2648 {
2649         mdsc->client = client;
2650         mutex_init(&mdsc->mutex);
2651         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
2652         if (mdsc->mdsmap == NULL)
2653                 return -ENOMEM;
2654
2655         init_completion(&mdsc->safe_umount_waiters);
2656         init_completion(&mdsc->session_close_waiters);
2657         INIT_LIST_HEAD(&mdsc->waiting_for_map);
2658         mdsc->sessions = NULL;
2659         mdsc->max_sessions = 0;
2660         mdsc->stopping = 0;
2661         init_rwsem(&mdsc->snap_rwsem);
2662         mdsc->snap_realms = RB_ROOT;
2663         INIT_LIST_HEAD(&mdsc->snap_empty);
2664         spin_lock_init(&mdsc->snap_empty_lock);
2665         mdsc->last_tid = 0;
2666         mdsc->request_tree = RB_ROOT;
2667         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
2668         mdsc->last_renew_caps = jiffies;
2669         INIT_LIST_HEAD(&mdsc->cap_delay_list);
2670         spin_lock_init(&mdsc->cap_delay_lock);
2671         INIT_LIST_HEAD(&mdsc->snap_flush_list);
2672         spin_lock_init(&mdsc->snap_flush_lock);
2673         mdsc->cap_flush_seq = 0;
2674         INIT_LIST_HEAD(&mdsc->cap_dirty);
2675         mdsc->num_cap_flushing = 0;
2676         spin_lock_init(&mdsc->cap_dirty_lock);
2677         init_waitqueue_head(&mdsc->cap_flushing_wq);
2678         spin_lock_init(&mdsc->dentry_lru_lock);
2679         INIT_LIST_HEAD(&mdsc->dentry_lru);
2680
2681         return 0;
2682 }
2683
2684 /*
2685  * Wait for safe replies on open mds requests.  If we time out, drop
2686  * all requests from the tree to avoid dangling dentry refs.
2687  */
2688 static void wait_requests(struct ceph_mds_client *mdsc)
2689 {
2690         struct ceph_mds_request *req;
2691         struct ceph_client *client = mdsc->client;
2692
2693         mutex_lock(&mdsc->mutex);
2694         if (__get_oldest_req(mdsc)) {
2695                 mutex_unlock(&mdsc->mutex);
2696
2697                 dout("wait_requests waiting for requests\n");
2698                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
2699                                     client->mount_args->mount_timeout * HZ);
2700
2701                 /* tear down remaining requests */
2702                 mutex_lock(&mdsc->mutex);
2703                 while ((req = __get_oldest_req(mdsc))) {
2704                         dout("wait_requests timed out on tid %llu\n",
2705                              req->r_tid);
2706                         __unregister_request(mdsc, req);
2707                 }
2708         }
2709         mutex_unlock(&mdsc->mutex);
2710         dout("wait_requests done\n");
2711 }
2712
2713 /*
2714  * called before mount is ro, and before dentries are torn down.
2715  * (hmm, does this still race with new lookups?)
2716  */
2717 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
2718 {
2719         dout("pre_umount\n");
2720         mdsc->stopping = 1;
2721
2722         drop_leases(mdsc);
2723         ceph_flush_dirty_caps(mdsc);
2724         wait_requests(mdsc);
2725 }
2726
2727 /*
2728  * wait for all write mds requests to flush.
2729  */
2730 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
2731 {
2732         struct ceph_mds_request *req = NULL, *nextreq;
2733         struct rb_node *n;
2734
2735         mutex_lock(&mdsc->mutex);
2736         dout("wait_unsafe_requests want %lld\n", want_tid);
2737 restart:
2738         req = __get_oldest_req(mdsc);
2739         while (req && req->r_tid <= want_tid) {
2740                 /* find next request */
2741                 n = rb_next(&req->r_node);
2742                 if (n)
2743                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
2744                 else
2745                         nextreq = NULL;
2746                 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
2747                         /* write op */
2748                         ceph_mdsc_get_request(req);
2749                         if (nextreq)
2750                                 ceph_mdsc_get_request(nextreq);
2751                         mutex_unlock(&mdsc->mutex);
2752                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
2753                              req->r_tid, want_tid);
2754                         wait_for_completion(&req->r_safe_completion);
2755                         mutex_lock(&mdsc->mutex);
2756                         ceph_mdsc_put_request(req);
2757                         if (!nextreq)
2758                                 break;  /* next dne before, so we're done! */
2759                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
2760                                 /* next request was removed from tree */
2761                                 ceph_mdsc_put_request(nextreq);
2762                                 goto restart;
2763                         }
2764                         ceph_mdsc_put_request(nextreq);  /* won't go away */
2765                 }
2766                 req = nextreq;
2767         }
2768         mutex_unlock(&mdsc->mutex);
2769         dout("wait_unsafe_requests done\n");
2770 }
2771
2772 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
2773 {
2774         u64 want_tid, want_flush;
2775
2776         if (mdsc->client->mount_state == CEPH_MOUNT_SHUTDOWN)
2777                 return;
2778
2779         dout("sync\n");
2780         mutex_lock(&mdsc->mutex);
2781         want_tid = mdsc->last_tid;
2782         want_flush = mdsc->cap_flush_seq;
2783         mutex_unlock(&mdsc->mutex);
2784         dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
2785
2786         ceph_flush_dirty_caps(mdsc);
2787
2788         wait_unsafe_requests(mdsc, want_tid);
2789         wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
2790 }
2791
2792
2793 /*
2794  * called after sb is ro.
2795  */
2796 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
2797 {
2798         struct ceph_mds_session *session;
2799         int i;
2800         int n;
2801         struct ceph_client *client = mdsc->client;
2802         unsigned long started, timeout = client->mount_args->mount_timeout * HZ;
2803
2804         dout("close_sessions\n");
2805
2806         mutex_lock(&mdsc->mutex);
2807
2808         /* close sessions */
2809         started = jiffies;
2810         while (time_before(jiffies, started + timeout)) {
2811                 dout("closing sessions\n");
2812                 n = 0;
2813                 for (i = 0; i < mdsc->max_sessions; i++) {
2814                         session = __ceph_lookup_mds_session(mdsc, i);
2815                         if (!session)
2816                                 continue;
2817                         mutex_unlock(&mdsc->mutex);
2818                         mutex_lock(&session->s_mutex);
2819                         __close_session(mdsc, session);
2820                         mutex_unlock(&session->s_mutex);
2821                         ceph_put_mds_session(session);
2822                         mutex_lock(&mdsc->mutex);
2823                         n++;
2824                 }
2825                 if (n == 0)
2826                         break;
2827
2828                 if (client->mount_state == CEPH_MOUNT_SHUTDOWN)
2829                         break;
2830
2831                 dout("waiting for sessions to close\n");
2832                 mutex_unlock(&mdsc->mutex);
2833                 wait_for_completion_timeout(&mdsc->session_close_waiters,
2834                                             timeout);
2835                 mutex_lock(&mdsc->mutex);
2836         }
2837
2838         /* tear down remaining sessions */
2839         for (i = 0; i < mdsc->max_sessions; i++) {
2840                 if (mdsc->sessions[i]) {
2841                         session = get_session(mdsc->sessions[i]);
2842                         __unregister_session(mdsc, session);
2843                         mutex_unlock(&mdsc->mutex);
2844                         mutex_lock(&session->s_mutex);
2845                         remove_session_caps(session);
2846                         mutex_unlock(&session->s_mutex);
2847                         ceph_put_mds_session(session);
2848                         mutex_lock(&mdsc->mutex);
2849                 }
2850         }
2851
2852         WARN_ON(!list_empty(&mdsc->cap_delay_list));
2853
2854         mutex_unlock(&mdsc->mutex);
2855
2856         ceph_cleanup_empty_realms(mdsc);
2857
2858         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2859
2860         dout("stopped\n");
2861 }
2862
2863 void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
2864 {
2865         dout("stop\n");
2866         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2867         if (mdsc->mdsmap)
2868                 ceph_mdsmap_destroy(mdsc->mdsmap);
2869         kfree(mdsc->sessions);
2870 }
2871
2872
2873 /*
2874  * handle mds map update.
2875  */
2876 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
2877 {
2878         u32 epoch;
2879         u32 maplen;
2880         void *p = msg->front.iov_base;
2881         void *end = p + msg->front.iov_len;
2882         struct ceph_mdsmap *newmap, *oldmap;
2883         struct ceph_fsid fsid;
2884         int err = -EINVAL;
2885
2886         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
2887         ceph_decode_copy(&p, &fsid, sizeof(fsid));
2888         if (ceph_check_fsid(mdsc->client, &fsid) < 0)
2889                 return;
2890         epoch = ceph_decode_32(&p);
2891         maplen = ceph_decode_32(&p);
2892         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
2893
2894         /* do we need it? */
2895         ceph_monc_got_mdsmap(&mdsc->client->monc, epoch);
2896         mutex_lock(&mdsc->mutex);
2897         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
2898                 dout("handle_map epoch %u <= our %u\n",
2899                      epoch, mdsc->mdsmap->m_epoch);
2900                 mutex_unlock(&mdsc->mutex);
2901                 return;
2902         }
2903
2904         newmap = ceph_mdsmap_decode(&p, end);
2905         if (IS_ERR(newmap)) {
2906                 err = PTR_ERR(newmap);
2907                 goto bad_unlock;
2908         }
2909
2910         /* swap into place */
2911         if (mdsc->mdsmap) {
2912                 oldmap = mdsc->mdsmap;
2913                 mdsc->mdsmap = newmap;
2914                 check_new_map(mdsc, newmap, oldmap);
2915                 ceph_mdsmap_destroy(oldmap);
2916         } else {
2917                 mdsc->mdsmap = newmap;  /* first mds map */
2918         }
2919         mdsc->client->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
2920
2921         __wake_requests(mdsc, &mdsc->waiting_for_map);
2922
2923         mutex_unlock(&mdsc->mutex);
2924         schedule_delayed(mdsc);
2925         return;
2926
2927 bad_unlock:
2928         mutex_unlock(&mdsc->mutex);
2929 bad:
2930         pr_err("error decoding mdsmap %d\n", err);
2931         return;
2932 }
2933
2934 static struct ceph_connection *con_get(struct ceph_connection *con)
2935 {
2936         struct ceph_mds_session *s = con->private;
2937
2938         if (get_session(s)) {
2939                 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
2940                 return con;
2941         }
2942         dout("mdsc con_get %p FAIL\n", s);
2943         return NULL;
2944 }
2945
2946 static void con_put(struct ceph_connection *con)
2947 {
2948         struct ceph_mds_session *s = con->private;
2949
2950         ceph_put_mds_session(s);
2951         dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref));
2952 }
2953
2954 /*
2955  * if the client is unresponsive for long enough, the mds will kill
2956  * the session entirely.
2957  */
2958 static void peer_reset(struct ceph_connection *con)
2959 {
2960         struct ceph_mds_session *s = con->private;
2961
2962         pr_err("mds%d gave us the boot.  IMPLEMENT RECONNECT.\n",
2963                s->s_mds);
2964 }
2965
2966 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2967 {
2968         struct ceph_mds_session *s = con->private;
2969         struct ceph_mds_client *mdsc = s->s_mdsc;
2970         int type = le16_to_cpu(msg->hdr.type);
2971
2972         mutex_lock(&mdsc->mutex);
2973         if (__verify_registered_session(mdsc, s) < 0) {
2974                 mutex_unlock(&mdsc->mutex);
2975                 goto out;
2976         }
2977         mutex_unlock(&mdsc->mutex);
2978
2979         switch (type) {
2980         case CEPH_MSG_MDS_MAP:
2981                 ceph_mdsc_handle_map(mdsc, msg);
2982                 break;
2983         case CEPH_MSG_CLIENT_SESSION:
2984                 handle_session(s, msg);
2985                 break;
2986         case CEPH_MSG_CLIENT_REPLY:
2987                 handle_reply(s, msg);
2988                 break;
2989         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
2990                 handle_forward(mdsc, s, msg);
2991                 break;
2992         case CEPH_MSG_CLIENT_CAPS:
2993                 ceph_handle_caps(s, msg);
2994                 break;
2995         case CEPH_MSG_CLIENT_SNAP:
2996                 ceph_handle_snap(mdsc, s, msg);
2997                 break;
2998         case CEPH_MSG_CLIENT_LEASE:
2999                 handle_lease(mdsc, s, msg);
3000                 break;
3001
3002         default:
3003                 pr_err("received unknown message type %d %s\n", type,
3004                        ceph_msg_type_name(type));
3005         }
3006 out:
3007         ceph_msg_put(msg);
3008 }
3009
3010 /*
3011  * authentication
3012  */
3013 static int get_authorizer(struct ceph_connection *con,
3014                           void **buf, int *len, int *proto,
3015                           void **reply_buf, int *reply_len, int force_new)
3016 {
3017         struct ceph_mds_session *s = con->private;
3018         struct ceph_mds_client *mdsc = s->s_mdsc;
3019         struct ceph_auth_client *ac = mdsc->client->monc.auth;
3020         int ret = 0;
3021
3022         if (force_new && s->s_authorizer) {
3023                 ac->ops->destroy_authorizer(ac, s->s_authorizer);
3024                 s->s_authorizer = NULL;
3025         }
3026         if (s->s_authorizer == NULL) {
3027                 if (ac->ops->create_authorizer) {
3028                         ret = ac->ops->create_authorizer(
3029                                 ac, CEPH_ENTITY_TYPE_MDS,
3030                                 &s->s_authorizer,
3031                                 &s->s_authorizer_buf,
3032                                 &s->s_authorizer_buf_len,
3033                                 &s->s_authorizer_reply_buf,
3034                                 &s->s_authorizer_reply_buf_len);
3035                         if (ret)
3036                                 return ret;
3037                 }
3038         }
3039
3040         *proto = ac->protocol;
3041         *buf = s->s_authorizer_buf;
3042         *len = s->s_authorizer_buf_len;
3043         *reply_buf = s->s_authorizer_reply_buf;
3044         *reply_len = s->s_authorizer_reply_buf_len;
3045         return 0;
3046 }
3047
3048
3049 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3050 {
3051         struct ceph_mds_session *s = con->private;
3052         struct ceph_mds_client *mdsc = s->s_mdsc;
3053         struct ceph_auth_client *ac = mdsc->client->monc.auth;
3054
3055         return ac->ops->verify_authorizer_reply(ac, s->s_authorizer, len);
3056 }
3057
3058 static int invalidate_authorizer(struct ceph_connection *con)
3059 {
3060         struct ceph_mds_session *s = con->private;
3061         struct ceph_mds_client *mdsc = s->s_mdsc;
3062         struct ceph_auth_client *ac = mdsc->client->monc.auth;
3063
3064         if (ac->ops->invalidate_authorizer)
3065                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3066
3067         return ceph_monc_validate_auth(&mdsc->client->monc);
3068 }
3069
3070 const static struct ceph_connection_operations mds_con_ops = {
3071         .get = con_get,
3072         .put = con_put,
3073         .dispatch = dispatch,
3074         .get_authorizer = get_authorizer,
3075         .verify_authorizer_reply = verify_authorizer_reply,
3076         .invalidate_authorizer = invalidate_authorizer,
3077         .peer_reset = peer_reset,
3078 };
3079
3080
3081
3082
3083 /* eof */