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 #include <linux/smp_lock.h>
   7
   8 #include "mds_client.h"
   9 #include "mon_client.h"
  10 #include "super.h"
  11 #include "messenger.h"
  12 #include "decode.h"
  13 #include "auth.h"
  14 #include "pagelist.h"
  15
  16 /*
  17  * A cluster of MDS (metadata server) daemons is responsible for
  18  * managing the file system namespace (the directory hierarchy and
  19  * inodes) and for coordinating shared access to storage.  Metadata is
  20  * partitioning hierarchically across a number of servers, and that
  21  * partition varies over time as the cluster adjusts the distribution
  22  * in order to balance load.
  23  *
  24  * The MDS client is primarily responsible to managing synchronous
  25  * metadata requests for operations like open, unlink, and so forth.
  26  * If there is a MDS failure, we find out about it when we (possibly
  27  * request and) receive a new MDS map, and can resubmit affected
  28  * requests.
  29  *
  30  * For the most part, though, we take advantage of a lossless
  31  * communications channel to the MDS, and do not need to worry about
  32  * timing out or resubmitting requests.
  33  *
  34  * We maintain a stateful "session" with each MDS we interact with.
  35  * Within each session, we sent periodic heartbeat messages to ensure
  36  * any capabilities or leases we have been issues remain valid.  If
  37  * the session times out and goes stale, our leases and capabilities
  38  * are no longer valid.
  39  */
  40
  41 struct ceph_reconnect_state {
  42         struct ceph_pagelist *pagelist;
  43         bool flock;
  44 };
  45
  46 static void __wake_requests(struct ceph_mds_client *mdsc,
  47                             struct list_head *head);
  48
  49 static const struct ceph_connection_operations mds_con_ops;
  50
  51
  52 /*
  53  * mds reply parsing
  54  */
  55
  56 /*
  57  * parse individual inode info
  58  */
  59 static int parse_reply_info_in(void **p, void *end,
  60                                struct ceph_mds_reply_info_in *info)
  61 {
  62         int err = -EIO;
  63
  64         info->in = *p;
  65         *p += sizeof(struct ceph_mds_reply_inode) +
  66                 sizeof(*info->in->fragtree.splits) *
  67                 le32_to_cpu(info->in->fragtree.nsplits);
  68
  69         ceph_decode_32_safe(p, end, info->symlink_len, bad);
  70         ceph_decode_need(p, end, info->symlink_len, bad);
  71         info->symlink = *p;
  72         *p += info->symlink_len;
  73
  74         ceph_decode_32_safe(p, end, info->xattr_len, bad);
  75         ceph_decode_need(p, end, info->xattr_len, bad);
  76         info->xattr_data = *p;
  77         *p += info->xattr_len;
  78         return 0;
  79 bad:
  80         return err;
  81 }
  82
  83 /*
  84  * parse a normal reply, which may contain a (dir+)dentry and/or a
  85  * target inode.
  86  */
  87 static int parse_reply_info_trace(void **p, void *end,
  88                                   struct ceph_mds_reply_info_parsed *info)
  89 {
  90         int err;
  91
  92         if (info->head->is_dentry) {
  93                 err = parse_reply_info_in(p, end, &info->diri);
  94                 if (err < 0)
  95                         goto out_bad;
  96
  97                 if (unlikely(*p + sizeof(*info->dirfrag) > end))
  98                         goto bad;
  99                 info->dirfrag = *p;
 100                 *p += sizeof(*info->dirfrag) +
 101                         sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
 102                 if (unlikely(*p > end))
 103                         goto bad;
 104
 105                 ceph_decode_32_safe(p, end, info->dname_len, bad);
 106                 ceph_decode_need(p, end, info->dname_len, bad);
 107                 info->dname = *p;
 108                 *p += info->dname_len;
 109                 info->dlease = *p;
 110                 *p += sizeof(*info->dlease);
 111         }
 112
 113         if (info->head->is_target) {
 114                 err = parse_reply_info_in(p, end, &info->targeti);
 115                 if (err < 0)
 116                         goto out_bad;
 117         }
 118
 119         if (unlikely(*p != end))
 120                 goto bad;
 121         return 0;
 122
 123 bad:
 124         err = -EIO;
 125 out_bad:
 126         pr_err("problem parsing mds trace %d\n", err);
 127         return err;
 128 }
 129
 130 /*
 131  * parse readdir results
 132  */
 133 static int parse_reply_info_dir(void **p, void *end,
 134                                 struct ceph_mds_reply_info_parsed *info)
 135 {
 136         u32 num, i = 0;
 137         int err;
 138
 139         info->dir_dir = *p;
 140         if (*p + sizeof(*info->dir_dir) > end)
 141                 goto bad;
 142         *p += sizeof(*info->dir_dir) +
 143                 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
 144         if (*p > end)
 145                 goto bad;
 146
 147         ceph_decode_need(p, end, sizeof(num) + 2, bad);
 148         num = ceph_decode_32(p);
 149         info->dir_end = ceph_decode_8(p);
 150         info->dir_complete = ceph_decode_8(p);
 151         if (num == 0)
 152                 goto done;
 153
 154         /* alloc large array */
 155         info->dir_nr = num;
 156         info->dir_in = kcalloc(num, sizeof(*info->dir_in) +
 157                                sizeof(*info->dir_dname) +
 158                                sizeof(*info->dir_dname_len) +
 159                                sizeof(*info->dir_dlease),
 160                                GFP_NOFS);
 161         if (info->dir_in == NULL) {
 162                 err = -ENOMEM;
 163                 goto out_bad;
 164         }
 165         info->dir_dname = (void *)(info->dir_in + num);
 166         info->dir_dname_len = (void *)(info->dir_dname + num);
 167         info->dir_dlease = (void *)(info->dir_dname_len + num);
 168
 169         while (num) {
 170                 /* dentry */
 171                 ceph_decode_need(p, end, sizeof(u32)*2, bad);
 172                 info->dir_dname_len[i] = ceph_decode_32(p);
 173                 ceph_decode_need(p, end, info->dir_dname_len[i], bad);
 174                 info->dir_dname[i] = *p;
 175                 *p += info->dir_dname_len[i];
 176                 dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
 177                      info->dir_dname[i]);
 178                 info->dir_dlease[i] = *p;
 179                 *p += sizeof(struct ceph_mds_reply_lease);
 180
 181                 /* inode */
 182                 err = parse_reply_info_in(p, end, &info->dir_in[i]);
 183                 if (err < 0)
 184                         goto out_bad;
 185                 i++;
 186                 num--;
 187         }
 188
 189 done:
 190         if (*p != end)
 191                 goto bad;
 192         return 0;
 193
 194 bad:
 195         err = -EIO;
 196 out_bad:
 197         pr_err("problem parsing dir contents %d\n", err);
 198         return err;
 199 }
 200
 201 /*
 202  * parse entire mds reply
 203  */
 204 static int parse_reply_info(struct ceph_msg *msg,
 205                             struct ceph_mds_reply_info_parsed *info)
 206 {
 207         void *p, *end;
 208         u32 len;
 209         int err;
 210
 211         info->head = msg->front.iov_base;
 212         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 213         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 214
 215         /* trace */
 216         ceph_decode_32_safe(&p, end, len, bad);
 217         if (len > 0) {
 218                 err = parse_reply_info_trace(&p, p+len, info);
 219                 if (err < 0)
 220                         goto out_bad;
 221         }
 222
 223         /* dir content */
 224         ceph_decode_32_safe(&p, end, len, bad);
 225         if (len > 0) {
 226                 err = parse_reply_info_dir(&p, p+len, info);
 227                 if (err < 0)
 228                         goto out_bad;
 229         }
 230
 231         /* snap blob */
 232         ceph_decode_32_safe(&p, end, len, bad);
 233         info->snapblob_len = len;
 234         info->snapblob = p;
 235         p += len;
 236
 237         if (p != end)
 238                 goto bad;
 239         return 0;
 240
 241 bad:
 242         err = -EIO;
 243 out_bad:
 244         pr_err("mds parse_reply err %d\n", err);
 245         return err;
 246 }
 247
 248 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
 249 {
 250         kfree(info->dir_in);
 251 }
 252
 253
 254 /*
 255  * sessions
 256  */
 257 static const char *session_state_name(int s)
 258 {
 259         switch (s) {
 260         case CEPH_MDS_SESSION_NEW: return "new";
 261         case CEPH_MDS_SESSION_OPENING: return "opening";
 262         case CEPH_MDS_SESSION_OPEN: return "open";
 263         case CEPH_MDS_SESSION_HUNG: return "hung";
 264         case CEPH_MDS_SESSION_CLOSING: return "closing";
 265         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
 266         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
 267         default: return "???";
 268         }
 269 }
 270
 271 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
 272 {
 273         if (atomic_inc_not_zero(&s->s_ref)) {
 274                 dout("mdsc get_session %p %d -> %d\n", s,
 275                      atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
 276                 return s;
 277         } else {
 278                 dout("mdsc get_session %p 0 -- FAIL", s);
 279                 return NULL;
 280         }
 281 }
 282
 283 void ceph_put_mds_session(struct ceph_mds_session *s)
 284 {
 285         dout("mdsc put_session %p %d -> %d\n", s,
 286              atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
 287         if (atomic_dec_and_test(&s->s_ref)) {
 288                 if (s->s_authorizer)
 289                         s->s_mdsc->client->monc.auth->ops->destroy_authorizer(
 290                                 s->s_mdsc->client->monc.auth, s->s_authorizer);
 291                 kfree(s);
 292         }
 293 }
 294
 295 /*
 296  * called under mdsc->mutex
 297  */
 298 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
 299                                                    int mds)
 300 {
 301         struct ceph_mds_session *session;
 302
 303         if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
 304                 return NULL;
 305         session = mdsc->sessions[mds];
 306         dout("lookup_mds_session %p %d\n", session,
 307              atomic_read(&session->s_ref));
 308         get_session(session);
 309         return session;
 310 }
 311
 312 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
 313 {
 314         if (mds >= mdsc->max_sessions)
 315                 return false;
 316         return mdsc->sessions[mds];
 317 }
 318
 319 static int __verify_registered_session(struct ceph_mds_client *mdsc,
 320                                        struct ceph_mds_session *s)
 321 {
 322         if (s->s_mds >= mdsc->max_sessions ||
 323             mdsc->sessions[s->s_mds] != s)
 324                 return -ENOENT;
 325         return 0;
 326 }
 327
 328 /*
 329  * create+register a new session for given mds.
 330  * called under mdsc->mutex.
 331  */
 332 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 333                                                  int mds)
 334 {
 335         struct ceph_mds_session *s;
 336
 337         s = kzalloc(sizeof(*s), GFP_NOFS);
 338         if (!s)
 339                 return ERR_PTR(-ENOMEM);
 340         s->s_mdsc = mdsc;
 341         s->s_mds = mds;
 342         s->s_state = CEPH_MDS_SESSION_NEW;
 343         s->s_ttl = 0;
 344         s->s_seq = 0;
 345         mutex_init(&s->s_mutex);
 346
 347         ceph_con_init(mdsc->client->msgr, &s->s_con);
 348         s->s_con.private = s;
 349         s->s_con.ops = &mds_con_ops;
 350         s->s_con.peer_name.type = CEPH_ENTITY_TYPE_MDS;
 351         s->s_con.peer_name.num = cpu_to_le64(mds);
 352
 353         spin_lock_init(&s->s_cap_lock);
 354         s->s_cap_gen = 0;
 355         s->s_cap_ttl = 0;
 356         s->s_renew_requested = 0;
 357         s->s_renew_seq = 0;
 358         INIT_LIST_HEAD(&s->s_caps);
 359         s->s_nr_caps = 0;
 360         s->s_trim_caps = 0;
 361         atomic_set(&s->s_ref, 1);
 362         INIT_LIST_HEAD(&s->s_waiting);
 363         INIT_LIST_HEAD(&s->s_unsafe);
 364         s->s_num_cap_releases = 0;
 365         s->s_cap_iterator = NULL;
 366         INIT_LIST_HEAD(&s->s_cap_releases);
 367         INIT_LIST_HEAD(&s->s_cap_releases_done);
 368         INIT_LIST_HEAD(&s->s_cap_flushing);
 369         INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
 370
 371         dout("register_session mds%d\n", mds);
 372         if (mds >= mdsc->max_sessions) {
 373                 int newmax = 1 << get_count_order(mds+1);
 374                 struct ceph_mds_session **sa;
 375
 376                 dout("register_session realloc to %d\n", newmax);
 377                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 378                 if (sa == NULL)
 379                         goto fail_realloc;
 380                 if (mdsc->sessions) {
 381                         memcpy(sa, mdsc->sessions,
 382                                mdsc->max_sessions * sizeof(void *));
 383                         kfree(mdsc->sessions);
 384                 }
 385                 mdsc->sessions = sa;
 386                 mdsc->max_sessions = newmax;
 387         }
 388         mdsc->sessions[mds] = s;
 389         atomic_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
 390
 391         ceph_con_open(&s->s_con, ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
 392
 393         return s;
 394
 395 fail_realloc:
 396         kfree(s);
 397         return ERR_PTR(-ENOMEM);
 398 }
 399
 400 /*
 401  * called under mdsc->mutex
 402  */
 403 static void __unregister_session(struct ceph_mds_client *mdsc,
 404                                struct ceph_mds_session *s)
 405 {
 406         dout("__unregister_session mds%d %p\n", s->s_mds, s);
 407         BUG_ON(mdsc->sessions[s->s_mds] != s);
 408         mdsc->sessions[s->s_mds] = NULL;
 409         ceph_con_close(&s->s_con);
 410         ceph_put_mds_session(s);
 411 }
 412
 413 /*
 414  * drop session refs in request.
 415  *
 416  * should be last request ref, or hold mdsc->mutex
 417  */
 418 static void put_request_session(struct ceph_mds_request *req)
 419 {
 420         if (req->r_session) {
 421                 ceph_put_mds_session(req->r_session);
 422                 req->r_session = NULL;
 423         }
 424 }
 425
 426 void ceph_mdsc_release_request(struct kref *kref)
 427 {
 428         struct ceph_mds_request *req = container_of(kref,
 429                                                     struct ceph_mds_request,
 430                                                     r_kref);
 431         if (req->r_request)
 432                 ceph_msg_put(req->r_request);
 433         if (req->r_reply) {
 434                 ceph_msg_put(req->r_reply);
 435                 destroy_reply_info(&req->r_reply_info);
 436         }
 437         if (req->r_inode) {
 438                 ceph_put_cap_refs(ceph_inode(req->r_inode),
 439                                   CEPH_CAP_PIN);
 440                 iput(req->r_inode);
 441         }
 442         if (req->r_locked_dir)
 443                 ceph_put_cap_refs(ceph_inode(req->r_locked_dir),
 444                                   CEPH_CAP_PIN);
 445         if (req->r_target_inode)
 446                 iput(req->r_target_inode);
 447         if (req->r_dentry)
 448                 dput(req->r_dentry);
 449         if (req->r_old_dentry) {
 450                 ceph_put_cap_refs(
 451                         ceph_inode(req->r_old_dentry->d_parent->d_inode),
 452                         CEPH_CAP_PIN);
 453                 dput(req->r_old_dentry);
 454         }
 455         kfree(req->r_path1);
 456         kfree(req->r_path2);
 457         put_request_session(req);
 458         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
 459         kfree(req);
 460 }
 461
 462 /*
 463  * lookup session, bump ref if found.
 464  *
 465  * called under mdsc->mutex.
 466  */
 467 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
 468                                              u64 tid)
 469 {
 470         struct ceph_mds_request *req;
 471         struct rb_node *n = mdsc->request_tree.rb_node;
 472
 473         while (n) {
 474                 req = rb_entry(n, struct ceph_mds_request, r_node);
 475                 if (tid < req->r_tid)
 476                         n = n->rb_left;
 477                 else if (tid > req->r_tid)
 478                         n = n->rb_right;
 479                 else {
 480                         ceph_mdsc_get_request(req);
 481                         return req;
 482                 }
 483         }
 484         return NULL;
 485 }
 486
 487 static void __insert_request(struct ceph_mds_client *mdsc,
 488                              struct ceph_mds_request *new)
 489 {
 490         struct rb_node **p = &mdsc->request_tree.rb_node;
 491         struct rb_node *parent = NULL;
 492         struct ceph_mds_request *req = NULL;
 493
 494         while (*p) {
 495                 parent = *p;
 496                 req = rb_entry(parent, struct ceph_mds_request, r_node);
 497                 if (new->r_tid < req->r_tid)
 498                         p = &(*p)->rb_left;
 499                 else if (new->r_tid > req->r_tid)
 500                         p = &(*p)->rb_right;
 501                 else
 502                         BUG();
 503         }
 504
 505         rb_link_node(&new->r_node, parent, p);
 506         rb_insert_color(&new->r_node, &mdsc->request_tree);
 507 }
 508
 509 /*
 510  * Register an in-flight request, and assign a tid.  Link to directory
 511  * are modifying (if any).
 512  *
 513  * Called under mdsc->mutex.
 514  */
 515 static void __register_request(struct ceph_mds_client *mdsc,
 516                                struct ceph_mds_request *req,
 517                                struct inode *dir)
 518 {
 519         req->r_tid = ++mdsc->last_tid;
 520         if (req->r_num_caps)
 521                 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
 522                                   req->r_num_caps);
 523         dout("__register_request %p tid %lld\n", req, req->r_tid);
 524         ceph_mdsc_get_request(req);
 525         __insert_request(mdsc, req);
 526
 527         if (dir) {
 528                 struct ceph_inode_info *ci = ceph_inode(dir);
 529
 530                 spin_lock(&ci->i_unsafe_lock);
 531                 req->r_unsafe_dir = dir;
 532                 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
 533                 spin_unlock(&ci->i_unsafe_lock);
 534         }
 535 }
 536
 537 static void __unregister_request(struct ceph_mds_client *mdsc,
 538                                  struct ceph_mds_request *req)
 539 {
 540         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 541         rb_erase(&req->r_node, &mdsc->request_tree);
 542         RB_CLEAR_NODE(&req->r_node);
 543
 544         if (req->r_unsafe_dir) {
 545                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 546
 547                 spin_lock(&ci->i_unsafe_lock);
 548                 list_del_init(&req->r_unsafe_dir_item);
 549                 spin_unlock(&ci->i_unsafe_lock);
 550         }
 551
 552         ceph_mdsc_put_request(req);
 553 }
 554
 555 /*
 556  * Choose mds to send request to next.  If there is a hint set in the
 557  * request (e.g., due to a prior forward hint from the mds), use that.
 558  * Otherwise, consult frag tree and/or caps to identify the
 559  * appropriate mds.  If all else fails, choose randomly.
 560  *
 561  * Called under mdsc->mutex.
 562  */
 563 struct dentry *get_nonsnap_parent(struct dentry *dentry)
 564 {
 565         while (!IS_ROOT(dentry) && ceph_snap(dentry->d_inode) != CEPH_NOSNAP)
 566                 dentry = dentry->d_parent;
 567         return dentry;
 568 }
 569
 570 static int __choose_mds(struct ceph_mds_client *mdsc,
 571                         struct ceph_mds_request *req)
 572 {
 573         struct inode *inode;
 574         struct ceph_inode_info *ci;
 575         struct ceph_cap *cap;
 576         int mode = req->r_direct_mode;
 577         int mds = -1;
 578         u32 hash = req->r_direct_hash;
 579         bool is_hash = req->r_direct_is_hash;
 580
 581         /*
 582          * is there a specific mds we should try?  ignore hint if we have
 583          * no session and the mds is not up (active or recovering).
 584          */
 585         if (req->r_resend_mds >= 0 &&
 586             (__have_session(mdsc, req->r_resend_mds) ||
 587              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 588                 dout("choose_mds using resend_mds mds%d\n",
 589                      req->r_resend_mds);
 590                 return req->r_resend_mds;
 591         }
 592
 593         if (mode == USE_RANDOM_MDS)
 594                 goto random;
 595
 596         inode = NULL;
 597         if (req->r_inode) {
 598                 inode = req->r_inode;
 599         } else if (req->r_dentry) {
 600                 struct inode *dir = req->r_dentry->d_parent->d_inode;
 601
 602                 if (dir->i_sb != mdsc->client->sb) {
 603                         /* not this fs! */
 604                         inode = req->r_dentry->d_inode;
 605                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
 606                         /* direct snapped/virtual snapdir requests
 607                          * based on parent dir inode */
 608                         struct dentry *dn =
 609                                 get_nonsnap_parent(req->r_dentry->d_parent);
 610                         inode = dn->d_inode;
 611                         dout("__choose_mds using nonsnap parent %p\n", inode);
 612                 } else if (req->r_dentry->d_inode) {
 613                         /* dentry target */
 614                         inode = req->r_dentry->d_inode;
 615                 } else {
 616                         /* dir + name */
 617                         inode = dir;
 618                         hash = req->r_dentry->d_name.hash;
 619                         is_hash = true;
 620                 }
 621         }
 622
 623         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 624              (int)hash, mode);
 625         if (!inode)
 626                 goto random;
 627         ci = ceph_inode(inode);
 628
 629         if (is_hash && S_ISDIR(inode->i_mode)) {
 630                 struct ceph_inode_frag frag;
 631                 int found;
 632
 633                 ceph_choose_frag(ci, hash, &frag, &found);
 634                 if (found) {
 635                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 636                                 u8 r;
 637
 638                                 /* choose a random replica */
 639                                 get_random_bytes(&r, 1);
 640                                 r %= frag.ndist;
 641                                 mds = frag.dist[r];
 642                                 dout("choose_mds %p %llx.%llx "
 643                                      "frag %u mds%d (%d/%d)\n",
 644                                      inode, ceph_vinop(inode),
 645                                      frag.frag, frag.mds,
 646                                      (int)r, frag.ndist);
 647                                 return mds;
 648                         }
 649
 650                         /* since this file/dir wasn't known to be
 651                          * replicated, then we want to look for the
 652                          * authoritative mds. */
 653                         mode = USE_AUTH_MDS;
 654                         if (frag.mds >= 0) {
 655                                 /* choose auth mds */
 656                                 mds = frag.mds;
 657                                 dout("choose_mds %p %llx.%llx "
 658                                      "frag %u mds%d (auth)\n",
 659                                      inode, ceph_vinop(inode), frag.frag, mds);
 660                                 return mds;
 661                         }
 662                 }
 663         }
 664
 665         spin_lock(&inode->i_lock);
 666         cap = NULL;
 667         if (mode == USE_AUTH_MDS)
 668                 cap = ci->i_auth_cap;
 669         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
 670                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
 671         if (!cap) {
 672                 spin_unlock(&inode->i_lock);
 673                 goto random;
 674         }
 675         mds = cap->session->s_mds;
 676         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
 677              inode, ceph_vinop(inode), mds,
 678              cap == ci->i_auth_cap ? "auth " : "", cap);
 679         spin_unlock(&inode->i_lock);
 680         return mds;
 681
 682 random:
 683         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
 684         dout("choose_mds chose random mds%d\n", mds);
 685         return mds;
 686 }
 687
 688
 689 /*
 690  * session messages
 691  */
 692 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
 693 {
 694         struct ceph_msg *msg;
 695         struct ceph_mds_session_head *h;
 696
 697         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS);
 698         if (!msg) {
 699                 pr_err("create_session_msg ENOMEM creating msg\n");
 700                 return NULL;
 701         }
 702         h = msg->front.iov_base;
 703         h->op = cpu_to_le32(op);
 704         h->seq = cpu_to_le64(seq);
 705         return msg;
 706 }
 707
 708 /*
 709  * send session open request.
 710  *
 711  * called under mdsc->mutex
 712  */
 713 static int __open_session(struct ceph_mds_client *mdsc,
 714                           struct ceph_mds_session *session)
 715 {
 716         struct ceph_msg *msg;
 717         int mstate;
 718         int mds = session->s_mds;
 719
 720         /* wait for mds to go active? */
 721         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
 722         dout("open_session to mds%d (%s)\n", mds,
 723              ceph_mds_state_name(mstate));
 724         session->s_state = CEPH_MDS_SESSION_OPENING;
 725         session->s_renew_requested = jiffies;
 726
 727         /* send connect message */
 728         msg = create_session_msg(CEPH_SESSION_REQUEST_OPEN, session->s_seq);
 729         if (!msg)
 730                 return -ENOMEM;
 731         ceph_con_send(&session->s_con, msg);
 732         return 0;
 733 }
 734
 735 /*
 736  * open sessions for any export targets for the given mds
 737  *
 738  * called under mdsc->mutex
 739  */
 740 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
 741                                           struct ceph_mds_session *session)
 742 {
 743         struct ceph_mds_info *mi;
 744         struct ceph_mds_session *ts;
 745         int i, mds = session->s_mds;
 746         int target;
 747
 748         if (mds >= mdsc->mdsmap->m_max_mds)
 749                 return;
 750         mi = &mdsc->mdsmap->m_info[mds];
 751         dout("open_export_target_sessions for mds%d (%d targets)\n",
 752              session->s_mds, mi->num_export_targets);
 753
 754         for (i = 0; i < mi->num_export_targets; i++) {
 755                 target = mi->export_targets[i];
 756                 ts = __ceph_lookup_mds_session(mdsc, target);
 757                 if (!ts) {
 758                         ts = register_session(mdsc, target);
 759                         if (IS_ERR(ts))
 760                                 return;
 761                 }
 762                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
 763                     session->s_state == CEPH_MDS_SESSION_CLOSING)
 764                         __open_session(mdsc, session);
 765                 else
 766                         dout(" mds%d target mds%d %p is %s\n", session->s_mds,
 767                              i, ts, session_state_name(ts->s_state));
 768                 ceph_put_mds_session(ts);
 769         }
 770 }
 771
 772 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
 773                                            struct ceph_mds_session *session)
 774 {
 775         mutex_lock(&mdsc->mutex);
 776         __open_export_target_sessions(mdsc, session);
 777         mutex_unlock(&mdsc->mutex);
 778 }
 779
 780 /*
 781  * session caps
 782  */
 783
 784 /*
 785  * Free preallocated cap messages assigned to this session
 786  */
 787 static void cleanup_cap_releases(struct ceph_mds_session *session)
 788 {
 789         struct ceph_msg *msg;
 790
 791         spin_lock(&session->s_cap_lock);
 792         while (!list_empty(&session->s_cap_releases)) {
 793                 msg = list_first_entry(&session->s_cap_releases,
 794                                        struct ceph_msg, list_head);
 795                 list_del_init(&msg->list_head);
 796                 ceph_msg_put(msg);
 797         }
 798         while (!list_empty(&session->s_cap_releases_done)) {
 799                 msg = list_first_entry(&session->s_cap_releases_done,
 800                                        struct ceph_msg, list_head);
 801                 list_del_init(&msg->list_head);
 802                 ceph_msg_put(msg);
 803         }
 804         spin_unlock(&session->s_cap_lock);
 805 }
 806
 807 /*
 808  * Helper to safely iterate over all caps associated with a session, with
 809  * special care taken to handle a racing __ceph_remove_cap().
 810  *
 811  * Caller must hold session s_mutex.
 812  */
 813 static int iterate_session_caps(struct ceph_mds_session *session,
 814                                  int (*cb)(struct inode *, struct ceph_cap *,
 815                                             void *), void *arg)
 816 {
 817         struct list_head *p;
 818         struct ceph_cap *cap;
 819         struct inode *inode, *last_inode = NULL;
 820         struct ceph_cap *old_cap = NULL;
 821         int ret;
 822
 823         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
 824         spin_lock(&session->s_cap_lock);
 825         p = session->s_caps.next;
 826         while (p != &session->s_caps) {
 827                 cap = list_entry(p, struct ceph_cap, session_caps);
 828                 inode = igrab(&cap->ci->vfs_inode);
 829                 if (!inode) {
 830                         p = p->next;
 831                         continue;
 832                 }
 833                 session->s_cap_iterator = cap;
 834                 spin_unlock(&session->s_cap_lock);
 835
 836                 if (last_inode) {
 837                         iput(last_inode);
 838                         last_inode = NULL;
 839                 }
 840                 if (old_cap) {
 841                         ceph_put_cap(session->s_mdsc, old_cap);
 842                         old_cap = NULL;
 843                 }
 844
 845                 ret = cb(inode, cap, arg);
 846                 last_inode = inode;
 847
 848                 spin_lock(&session->s_cap_lock);
 849                 p = p->next;
 850                 if (cap->ci == NULL) {
 851                         dout("iterate_session_caps  finishing cap %p removal\n",
 852                              cap);
 853                         BUG_ON(cap->session != session);
 854                         list_del_init(&cap->session_caps);
 855                         session->s_nr_caps--;
 856                         cap->session = NULL;
 857                         old_cap = cap;  /* put_cap it w/o locks held */
 858                 }
 859                 if (ret < 0)
 860                         goto out;
 861         }
 862         ret = 0;
 863 out:
 864         session->s_cap_iterator = NULL;
 865         spin_unlock(&session->s_cap_lock);
 866
 867         if (last_inode)
 868                 iput(last_inode);
 869         if (old_cap)
 870                 ceph_put_cap(session->s_mdsc, old_cap);
 871
 872         return ret;
 873 }
 874
 875 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
 876                                   void *arg)
 877 {
 878         struct ceph_inode_info *ci = ceph_inode(inode);
 879         int drop = 0;
 880
 881         dout("removing cap %p, ci is %p, inode is %p\n",
 882              cap, ci, &ci->vfs_inode);
 883         spin_lock(&inode->i_lock);
 884         __ceph_remove_cap(cap);
 885         if (!__ceph_is_any_real_caps(ci)) {
 886                 struct ceph_mds_client *mdsc =
 887                         &ceph_sb_to_client(inode->i_sb)->mdsc;
 888
 889                 spin_lock(&mdsc->cap_dirty_lock);
 890                 if (!list_empty(&ci->i_dirty_item)) {
 891                         pr_info(" dropping dirty %s state for %p %lld\n",
 892                                 ceph_cap_string(ci->i_dirty_caps),
 893                                 inode, ceph_ino(inode));
 894                         ci->i_dirty_caps = 0;
 895                         list_del_init(&ci->i_dirty_item);
 896                         drop = 1;
 897                 }
 898                 if (!list_empty(&ci->i_flushing_item)) {
 899                         pr_info(" dropping dirty+flushing %s state for %p %lld\n",
 900                                 ceph_cap_string(ci->i_flushing_caps),
 901                                 inode, ceph_ino(inode));
 902                         ci->i_flushing_caps = 0;
 903                         list_del_init(&ci->i_flushing_item);
 904                         mdsc->num_cap_flushing--;
 905                         drop = 1;
 906                 }
 907                 if (drop && ci->i_wrbuffer_ref) {
 908                         pr_info(" dropping dirty data for %p %lld\n",
 909                                 inode, ceph_ino(inode));
 910                         ci->i_wrbuffer_ref = 0;
 911                         ci->i_wrbuffer_ref_head = 0;
 912                         drop++;
 913                 }
 914                 spin_unlock(&mdsc->cap_dirty_lock);
 915         }
 916         spin_unlock(&inode->i_lock);
 917         while (drop--)
 918                 iput(inode);
 919         return 0;
 920 }
 921
 922 /*
 923  * caller must hold session s_mutex
 924  */
 925 static void remove_session_caps(struct ceph_mds_session *session)
 926 {
 927         dout("remove_session_caps on %p\n", session);
 928         iterate_session_caps(session, remove_session_caps_cb, NULL);
 929         BUG_ON(session->s_nr_caps > 0);
 930         BUG_ON(!list_empty(&session->s_cap_flushing));
 931         cleanup_cap_releases(session);
 932 }
 933
 934 /*
 935  * wake up any threads waiting on this session's caps.  if the cap is
 936  * old (didn't get renewed on the client reconnect), remove it now.
 937  *
 938  * caller must hold s_mutex.
 939  */
 940 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
 941                               void *arg)
 942 {
 943         struct ceph_inode_info *ci = ceph_inode(inode);
 944
 945         wake_up_all(&ci->i_cap_wq);
 946         if (arg) {
 947                 spin_lock(&inode->i_lock);
 948                 ci->i_wanted_max_size = 0;
 949                 ci->i_requested_max_size = 0;
 950                 spin_unlock(&inode->i_lock);
 951         }
 952         return 0;
 953 }
 954
 955 static void wake_up_session_caps(struct ceph_mds_session *session,
 956                                  int reconnect)
 957 {
 958         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
 959         iterate_session_caps(session, wake_up_session_cb,
 960                              (void *)(unsigned long)reconnect);
 961 }
 962
 963 /*
 964  * Send periodic message to MDS renewing all currently held caps.  The
 965  * ack will reset the expiration for all caps from this session.
 966  *
 967  * caller holds s_mutex
 968  */
 969 static int send_renew_caps(struct ceph_mds_client *mdsc,
 970                            struct ceph_mds_session *session)
 971 {
 972         struct ceph_msg *msg;
 973         int state;
 974
 975         if (time_after_eq(jiffies, session->s_cap_ttl) &&
 976             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
 977                 pr_info("mds%d caps stale\n", session->s_mds);
 978         session->s_renew_requested = jiffies;
 979
 980         /* do not try to renew caps until a recovering mds has reconnected
 981          * with its clients. */
 982         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
 983         if (state < CEPH_MDS_STATE_RECONNECT) {
 984                 dout("send_renew_caps ignoring mds%d (%s)\n",
 985                      session->s_mds, ceph_mds_state_name(state));
 986                 return 0;
 987         }
 988
 989         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
 990                 ceph_mds_state_name(state));
 991         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
 992                                  ++session->s_renew_seq);
 993         if (!msg)
 994                 return -ENOMEM;
 995         ceph_con_send(&session->s_con, msg);
 996         return 0;
 997 }
 998
 999 /*
1000  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1001  *
1002  * Called under session->s_mutex
1003  */
1004 static void renewed_caps(struct ceph_mds_client *mdsc,
1005                          struct ceph_mds_session *session, int is_renew)
1006 {
1007         int was_stale;
1008         int wake = 0;
1009
1010         spin_lock(&session->s_cap_lock);
1011         was_stale = is_renew && (session->s_cap_ttl == 0 ||
1012                                  time_after_eq(jiffies, session->s_cap_ttl));
1013
1014         session->s_cap_ttl = session->s_renew_requested +
1015                 mdsc->mdsmap->m_session_timeout*HZ;
1016
1017         if (was_stale) {
1018                 if (time_before(jiffies, session->s_cap_ttl)) {
1019                         pr_info("mds%d caps renewed\n", session->s_mds);
1020                         wake = 1;
1021                 } else {
1022                         pr_info("mds%d caps still stale\n", session->s_mds);
1023                 }
1024         }
1025         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1026              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1027              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1028         spin_unlock(&session->s_cap_lock);
1029
1030         if (wake)
1031                 wake_up_session_caps(session, 0);
1032 }
1033
1034 /*
1035  * send a session close request
1036  */
1037 static int request_close_session(struct ceph_mds_client *mdsc,
1038                                  struct ceph_mds_session *session)
1039 {
1040         struct ceph_msg *msg;
1041
1042         dout("request_close_session mds%d state %s seq %lld\n",
1043              session->s_mds, session_state_name(session->s_state),
1044              session->s_seq);
1045         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1046         if (!msg)
1047                 return -ENOMEM;
1048         ceph_con_send(&session->s_con, msg);
1049         return 0;
1050 }
1051
1052 /*
1053  * Called with s_mutex held.
1054  */
1055 static int __close_session(struct ceph_mds_client *mdsc,
1056                          struct ceph_mds_session *session)
1057 {
1058         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1059                 return 0;
1060         session->s_state = CEPH_MDS_SESSION_CLOSING;
1061         return request_close_session(mdsc, session);
1062 }
1063
1064 /*
1065  * Trim old(er) caps.
1066  *
1067  * Because we can't cache an inode without one or more caps, we do
1068  * this indirectly: if a cap is unused, we prune its aliases, at which
1069  * point the inode will hopefully get dropped to.
1070  *
1071  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1072  * memory pressure from the MDS, though, so it needn't be perfect.
1073  */
1074 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1075 {
1076         struct ceph_mds_session *session = arg;
1077         struct ceph_inode_info *ci = ceph_inode(inode);
1078         int used, oissued, mine;
1079
1080         if (session->s_trim_caps <= 0)
1081                 return -1;
1082
1083         spin_lock(&inode->i_lock);
1084         mine = cap->issued | cap->implemented;
1085         used = __ceph_caps_used(ci);
1086         oissued = __ceph_caps_issued_other(ci, cap);
1087
1088         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
1089              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1090              ceph_cap_string(used));
1091         if (ci->i_dirty_caps)
1092                 goto out;   /* dirty caps */
1093         if ((used & ~oissued) & mine)
1094                 goto out;   /* we need these caps */
1095
1096         session->s_trim_caps--;
1097         if (oissued) {
1098                 /* we aren't the only cap.. just remove us */
1099                 __ceph_remove_cap(cap);
1100         } else {
1101                 /* try to drop referring dentries */
1102                 spin_unlock(&inode->i_lock);
1103                 d_prune_aliases(inode);
1104                 dout("trim_caps_cb %p cap %p  pruned, count now %d\n",
1105                      inode, cap, atomic_read(&inode->i_count));
1106                 return 0;
1107         }
1108
1109 out:
1110         spin_unlock(&inode->i_lock);
1111         return 0;
1112 }
1113
1114 /*
1115  * Trim session cap count down to some max number.
1116  */
1117 static int trim_caps(struct ceph_mds_client *mdsc,
1118                      struct ceph_mds_session *session,
1119                      int max_caps)
1120 {
1121         int trim_caps = session->s_nr_caps - max_caps;
1122
1123         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1124              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1125         if (trim_caps > 0) {
1126                 session->s_trim_caps = trim_caps;
1127                 iterate_session_caps(session, trim_caps_cb, session);
1128                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1129                      session->s_mds, session->s_nr_caps, max_caps,
1130                         trim_caps - session->s_trim_caps);
1131                 session->s_trim_caps = 0;
1132         }
1133         return 0;
1134 }
1135
1136 /*
1137  * Allocate cap_release messages.  If there is a partially full message
1138  * in the queue, try to allocate enough to cover it's remainder, so that
1139  * we can send it immediately.
1140  *
1141  * Called under s_mutex.
1142  */
1143 int ceph_add_cap_releases(struct ceph_mds_client *mdsc,
1144                           struct ceph_mds_session *session)
1145 {
1146         struct ceph_msg *msg, *partial = NULL;
1147         struct ceph_mds_cap_release *head;
1148         int err = -ENOMEM;
1149         int extra = mdsc->client->mount_args->cap_release_safety;
1150         int num;
1151
1152         dout("add_cap_releases %p mds%d extra %d\n", session, session->s_mds,
1153              extra);
1154
1155         spin_lock(&session->s_cap_lock);
1156
1157         if (!list_empty(&session->s_cap_releases)) {
1158                 msg = list_first_entry(&session->s_cap_releases,
1159                                        struct ceph_msg,
1160                                  list_head);
1161                 head = msg->front.iov_base;
1162                 num = le32_to_cpu(head->num);
1163                 if (num) {
1164                         dout(" partial %p with (%d/%d)\n", msg, num,
1165                              (int)CEPH_CAPS_PER_RELEASE);
1166                         extra += CEPH_CAPS_PER_RELEASE - num;
1167                         partial = msg;
1168                 }
1169         }
1170         while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1171                 spin_unlock(&session->s_cap_lock);
1172                 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
1173                                    GFP_NOFS);
1174                 if (!msg)
1175                         goto out_unlocked;
1176                 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1177                      (int)msg->front.iov_len);
1178                 head = msg->front.iov_base;
1179                 head->num = cpu_to_le32(0);
1180                 msg->front.iov_len = sizeof(*head);
1181                 spin_lock(&session->s_cap_lock);
1182                 list_add(&msg->list_head, &session->s_cap_releases);
1183                 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1184         }
1185
1186         if (partial) {
1187                 head = partial->front.iov_base;
1188                 num = le32_to_cpu(head->num);
1189                 dout(" queueing partial %p with %d/%d\n", partial, num,
1190                      (int)CEPH_CAPS_PER_RELEASE);
1191                 list_move_tail(&partial->list_head,
1192                                &session->s_cap_releases_done);
1193                 session->s_num_cap_releases -= CEPH_CAPS_PER_RELEASE - num;
1194         }
1195         err = 0;
1196         spin_unlock(&session->s_cap_lock);
1197 out_unlocked:
1198         return err;
1199 }
1200
1201 /*
1202  * flush all dirty inode data to disk.
1203  *
1204  * returns true if we've flushed through want_flush_seq
1205  */
1206 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1207 {
1208         int mds, ret = 1;
1209
1210         dout("check_cap_flush want %lld\n", want_flush_seq);
1211         mutex_lock(&mdsc->mutex);
1212         for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1213                 struct ceph_mds_session *session = mdsc->sessions[mds];
1214
1215                 if (!session)
1216                         continue;
1217                 get_session(session);
1218                 mutex_unlock(&mdsc->mutex);
1219
1220                 mutex_lock(&session->s_mutex);
1221                 if (!list_empty(&session->s_cap_flushing)) {
1222                         struct ceph_inode_info *ci =
1223                                 list_entry(session->s_cap_flushing.next,
1224                                            struct ceph_inode_info,
1225                                            i_flushing_item);
1226                         struct inode *inode = &ci->vfs_inode;
1227
1228                         spin_lock(&inode->i_lock);
1229                         if (ci->i_cap_flush_seq <= want_flush_seq) {
1230                                 dout("check_cap_flush still flushing %p "
1231                                      "seq %lld <= %lld to mds%d\n", inode,
1232                                      ci->i_cap_flush_seq, want_flush_seq,
1233                                      session->s_mds);
1234                                 ret = 0;
1235                         }
1236                         spin_unlock(&inode->i_lock);
1237                 }
1238                 mutex_unlock(&session->s_mutex);
1239                 ceph_put_mds_session(session);
1240
1241                 if (!ret)
1242                         return ret;
1243                 mutex_lock(&mdsc->mutex);
1244         }
1245
1246         mutex_unlock(&mdsc->mutex);
1247         dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1248         return ret;
1249 }
1250
1251 /*
1252  * called under s_mutex
1253  */
1254 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1255                             struct ceph_mds_session *session)
1256 {
1257         struct ceph_msg *msg;
1258
1259         dout("send_cap_releases mds%d\n", session->s_mds);
1260         spin_lock(&session->s_cap_lock);
1261         while (!list_empty(&session->s_cap_releases_done)) {
1262                 msg = list_first_entry(&session->s_cap_releases_done,
1263                                  struct ceph_msg, list_head);
1264                 list_del_init(&msg->list_head);
1265                 spin_unlock(&session->s_cap_lock);
1266                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1267                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1268                 ceph_con_send(&session->s_con, msg);
1269                 spin_lock(&session->s_cap_lock);
1270         }
1271         spin_unlock(&session->s_cap_lock);
1272 }
1273
1274 static void discard_cap_releases(struct ceph_mds_client *mdsc,
1275                                  struct ceph_mds_session *session)
1276 {
1277         struct ceph_msg *msg;
1278         struct ceph_mds_cap_release *head;
1279         unsigned num;
1280
1281         dout("discard_cap_releases mds%d\n", session->s_mds);
1282         spin_lock(&session->s_cap_lock);
1283
1284         /* zero out the in-progress message */
1285         msg = list_first_entry(&session->s_cap_releases,
1286                                struct ceph_msg, list_head);
1287         head = msg->front.iov_base;
1288         num = le32_to_cpu(head->num);
1289         dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg, num);
1290         head->num = cpu_to_le32(0);
1291         session->s_num_cap_releases += num;
1292
1293         /* requeue completed messages */
1294         while (!list_empty(&session->s_cap_releases_done)) {
1295                 msg = list_first_entry(&session->s_cap_releases_done,
1296                                  struct ceph_msg, list_head);
1297                 list_del_init(&msg->list_head);
1298
1299                 head = msg->front.iov_base;
1300                 num = le32_to_cpu(head->num);
1301                 dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg,
1302                      num);
1303                 session->s_num_cap_releases += num;
1304                 head->num = cpu_to_le32(0);
1305                 msg->front.iov_len = sizeof(*head);
1306                 list_add(&msg->list_head, &session->s_cap_releases);
1307         }
1308
1309         spin_unlock(&session->s_cap_lock);
1310 }
1311
1312 /*
1313  * requests
1314  */
1315
1316 /*
1317  * Create an mds request.
1318  */
1319 struct ceph_mds_request *
1320 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1321 {
1322         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1323
1324         if (!req)
1325                 return ERR_PTR(-ENOMEM);
1326
1327         mutex_init(&req->r_fill_mutex);
1328         req->r_mdsc = mdsc;
1329         req->r_started = jiffies;
1330         req->r_resend_mds = -1;
1331         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1332         req->r_fmode = -1;
1333         kref_init(&req->r_kref);
1334         INIT_LIST_HEAD(&req->r_wait);
1335         init_completion(&req->r_completion);
1336         init_completion(&req->r_safe_completion);
1337         INIT_LIST_HEAD(&req->r_unsafe_item);
1338
1339         req->r_op = op;
1340         req->r_direct_mode = mode;
1341         return req;
1342 }
1343
1344 /*
1345  * return oldest (lowest) request, tid in request tree, 0 if none.
1346  *
1347  * called under mdsc->mutex.
1348  */
1349 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1350 {
1351         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1352                 return NULL;
1353         return rb_entry(rb_first(&mdsc->request_tree),
1354                         struct ceph_mds_request, r_node);
1355 }
1356
1357 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1358 {
1359         struct ceph_mds_request *req = __get_oldest_req(mdsc);
1360
1361         if (req)
1362                 return req->r_tid;
1363         return 0;
1364 }
1365
1366 /*
1367  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1368  * on build_path_from_dentry in fs/cifs/dir.c.
1369  *
1370  * If @stop_on_nosnap, generate path relative to the first non-snapped
1371  * inode.
1372  *
1373  * Encode hidden .snap dirs as a double /, i.e.
1374  *   foo/.snap/bar -> foo//bar
1375  */
1376 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1377                            int stop_on_nosnap)
1378 {
1379         struct dentry *temp;
1380         char *path;
1381         int len, pos;
1382
1383         if (dentry == NULL)
1384                 return ERR_PTR(-EINVAL);
1385
1386 retry:
1387         len = 0;
1388         for (temp = dentry; !IS_ROOT(temp);) {
1389                 struct inode *inode = temp->d_inode;
1390                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1391                         len++;  /* slash only */
1392                 else if (stop_on_nosnap && inode &&
1393                          ceph_snap(inode) == CEPH_NOSNAP)
1394                         break;
1395                 else
1396                         len += 1 + temp->d_name.len;
1397                 temp = temp->d_parent;
1398                 if (temp == NULL) {
1399                         pr_err("build_path corrupt dentry %p\n", dentry);
1400                         return ERR_PTR(-EINVAL);
1401                 }
1402         }
1403         if (len)
1404                 len--;  /* no leading '/' */
1405
1406         path = kmalloc(len+1, GFP_NOFS);
1407         if (path == NULL)
1408                 return ERR_PTR(-ENOMEM);
1409         pos = len;
1410         path[pos] = 0;  /* trailing null */
1411         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1412                 struct inode *inode = temp->d_inode;
1413
1414                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1415                         dout("build_path path+%d: %p SNAPDIR\n",
1416                              pos, temp);
1417                 } else if (stop_on_nosnap && inode &&
1418                            ceph_snap(inode) == CEPH_NOSNAP) {
1419                         break;
1420                 } else {
1421                         pos -= temp->d_name.len;
1422                         if (pos < 0)
1423                                 break;
1424                         strncpy(path + pos, temp->d_name.name,
1425                                 temp->d_name.len);
1426                 }
1427                 if (pos)
1428                         path[--pos] = '/';
1429                 temp = temp->d_parent;
1430                 if (temp == NULL) {
1431                         pr_err("build_path corrupt dentry\n");
1432                         kfree(path);
1433                         return ERR_PTR(-EINVAL);
1434                 }
1435         }
1436         if (pos != 0) {
1437                 pr_err("build_path did not end path lookup where "
1438                        "expected, namelen is %d, pos is %d\n", len, pos);
1439                 /* presumably this is only possible if racing with a
1440                    rename of one of the parent directories (we can not
1441                    lock the dentries above us to prevent this, but
1442                    retrying should be harmless) */
1443                 kfree(path);
1444                 goto retry;
1445         }
1446
1447         *base = ceph_ino(temp->d_inode);
1448         *plen = len;
1449         dout("build_path on %p %d built %llx '%.*s'\n",
1450              dentry, atomic_read(&dentry->d_count), *base, len, path);
1451         return path;
1452 }
1453
1454 static int build_dentry_path(struct dentry *dentry,
1455                              const char **ppath, int *ppathlen, u64 *pino,
1456                              int *pfreepath)
1457 {
1458         char *path;
1459
1460         if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1461                 *pino = ceph_ino(dentry->d_parent->d_inode);
1462                 *ppath = dentry->d_name.name;
1463                 *ppathlen = dentry->d_name.len;
1464                 return 0;
1465         }
1466         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1467         if (IS_ERR(path))
1468                 return PTR_ERR(path);
1469         *ppath = path;
1470         *pfreepath = 1;
1471         return 0;
1472 }
1473
1474 static int build_inode_path(struct inode *inode,
1475                             const char **ppath, int *ppathlen, u64 *pino,
1476                             int *pfreepath)
1477 {
1478         struct dentry *dentry;
1479         char *path;
1480
1481         if (ceph_snap(inode) == CEPH_NOSNAP) {
1482                 *pino = ceph_ino(inode);
1483                 *ppathlen = 0;
1484                 return 0;
1485         }
1486         dentry = d_find_alias(inode);
1487         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1488         dput(dentry);
1489         if (IS_ERR(path))
1490                 return PTR_ERR(path);
1491         *ppath = path;
1492         *pfreepath = 1;
1493         return 0;
1494 }
1495
1496 /*
1497  * request arguments may be specified via an inode *, a dentry *, or
1498  * an explicit ino+path.
1499  */
1500 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1501                                   const char *rpath, u64 rino,
1502                                   const char **ppath, int *pathlen,
1503                                   u64 *ino, int *freepath)
1504 {
1505         int r = 0;
1506
1507         if (rinode) {
1508                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1509                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1510                      ceph_snap(rinode));
1511         } else if (rdentry) {
1512                 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1513                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1514                      *ppath);
1515         } else if (rpath) {
1516                 *ino = rino;
1517                 *ppath = rpath;
1518                 *pathlen = strlen(rpath);
1519                 dout(" path %.*s\n", *pathlen, rpath);
1520         }
1521
1522         return r;
1523 }
1524
1525 /*
1526  * called under mdsc->mutex
1527  */
1528 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1529                                                struct ceph_mds_request *req,
1530                                                int mds)
1531 {
1532         struct ceph_msg *msg;
1533         struct ceph_mds_request_head *head;
1534         const char *path1 = NULL;
1535         const char *path2 = NULL;
1536         u64 ino1 = 0, ino2 = 0;
1537         int pathlen1 = 0, pathlen2 = 0;
1538         int freepath1 = 0, freepath2 = 0;
1539         int len;
1540         u16 releases;
1541         void *p, *end;
1542         int ret;
1543
1544         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1545                               req->r_path1, req->r_ino1.ino,
1546                               &path1, &pathlen1, &ino1, &freepath1);
1547         if (ret < 0) {
1548                 msg = ERR_PTR(ret);
1549                 goto out;
1550         }
1551
1552         ret = set_request_path_attr(NULL, req->r_old_dentry,
1553                               req->r_path2, req->r_ino2.ino,
1554                               &path2, &pathlen2, &ino2, &freepath2);
1555         if (ret < 0) {
1556                 msg = ERR_PTR(ret);
1557                 goto out_free1;
1558         }
1559
1560         len = sizeof(*head) +
1561                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64));
1562
1563         /* calculate (max) length for cap releases */
1564         len += sizeof(struct ceph_mds_request_release) *
1565                 (!!req->r_inode_drop + !!req->r_dentry_drop +
1566                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1567         if (req->r_dentry_drop)
1568                 len += req->r_dentry->d_name.len;
1569         if (req->r_old_dentry_drop)
1570                 len += req->r_old_dentry->d_name.len;
1571
1572         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS);
1573         if (!msg) {
1574                 msg = ERR_PTR(-ENOMEM);
1575                 goto out_free2;
1576         }
1577
1578         msg->hdr.tid = cpu_to_le64(req->r_tid);
1579
1580         head = msg->front.iov_base;
1581         p = msg->front.iov_base + sizeof(*head);
1582         end = msg->front.iov_base + msg->front.iov_len;
1583
1584         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1585         head->op = cpu_to_le32(req->r_op);
1586         head->caller_uid = cpu_to_le32(current_fsuid());
1587         head->caller_gid = cpu_to_le32(current_fsgid());
1588         head->args = req->r_args;
1589
1590         ceph_encode_filepath(&p, end, ino1, path1);
1591         ceph_encode_filepath(&p, end, ino2, path2);
1592
1593         /* make note of release offset, in case we need to replay */
1594         req->r_request_release_offset = p - msg->front.iov_base;
1595
1596         /* cap releases */
1597         releases = 0;
1598         if (req->r_inode_drop)
1599                 releases += ceph_encode_inode_release(&p,
1600                       req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1601                       mds, req->r_inode_drop, req->r_inode_unless, 0);
1602         if (req->r_dentry_drop)
1603                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1604                        mds, req->r_dentry_drop, req->r_dentry_unless);
1605         if (req->r_old_dentry_drop)
1606                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1607                        mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1608         if (req->r_old_inode_drop)
1609                 releases += ceph_encode_inode_release(&p,
1610                       req->r_old_dentry->d_inode,
1611                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1612         head->num_releases = cpu_to_le16(releases);
1613
1614         BUG_ON(p > end);
1615         msg->front.iov_len = p - msg->front.iov_base;
1616         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1617
1618         msg->pages = req->r_pages;
1619         msg->nr_pages = req->r_num_pages;
1620         msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1621         msg->hdr.data_off = cpu_to_le16(0);
1622
1623 out_free2:
1624         if (freepath2)
1625                 kfree((char *)path2);
1626 out_free1:
1627         if (freepath1)
1628                 kfree((char *)path1);
1629 out:
1630         return msg;
1631 }
1632
1633 /*
1634  * called under mdsc->mutex if error, under no mutex if
1635  * success.
1636  */
1637 static void complete_request(struct ceph_mds_client *mdsc,
1638                              struct ceph_mds_request *req)
1639 {
1640         if (req->r_callback)
1641                 req->r_callback(mdsc, req);
1642         else
1643                 complete_all(&req->r_completion);
1644 }
1645
1646 /*
1647  * called under mdsc->mutex
1648  */
1649 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1650                                   struct ceph_mds_request *req,
1651                                   int mds)
1652 {
1653         struct ceph_mds_request_head *rhead;
1654         struct ceph_msg *msg;
1655         int flags = 0;
1656
1657         req->r_mds = mds;
1658         req->r_attempts++;
1659         if (req->r_inode) {
1660                 struct ceph_cap *cap =
1661                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
1662
1663                 if (cap)
1664                         req->r_sent_on_mseq = cap->mseq;
1665                 else
1666                         req->r_sent_on_mseq = -1;
1667         }
1668         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1669              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1670
1671         if (req->r_got_unsafe) {
1672                 /*
1673                  * Replay.  Do not regenerate message (and rebuild
1674                  * paths, etc.); just use the original message.
1675                  * Rebuilding paths will break for renames because
1676                  * d_move mangles the src name.
1677                  */
1678                 msg = req->r_request;
1679                 rhead = msg->front.iov_base;
1680
1681                 flags = le32_to_cpu(rhead->flags);
1682                 flags |= CEPH_MDS_FLAG_REPLAY;
1683                 rhead->flags = cpu_to_le32(flags);
1684
1685                 if (req->r_target_inode)
1686                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1687
1688                 rhead->num_retry = req->r_attempts - 1;
1689
1690                 /* remove cap/dentry releases from message */
1691                 rhead->num_releases = 0;
1692                 msg->hdr.front_len = cpu_to_le32(req->r_request_release_offset);
1693                 msg->front.iov_len = req->r_request_release_offset;
1694                 return 0;
1695         }
1696
1697         if (req->r_request) {
1698                 ceph_msg_put(req->r_request);
1699                 req->r_request = NULL;
1700         }
1701         msg = create_request_message(mdsc, req, mds);
1702         if (IS_ERR(msg)) {
1703                 req->r_err = PTR_ERR(msg);
1704                 complete_request(mdsc, req);
1705                 return PTR_ERR(msg);
1706         }
1707         req->r_request = msg;
1708
1709         rhead = msg->front.iov_base;
1710         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1711         if (req->r_got_unsafe)
1712                 flags |= CEPH_MDS_FLAG_REPLAY;
1713         if (req->r_locked_dir)
1714                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1715         rhead->flags = cpu_to_le32(flags);
1716         rhead->num_fwd = req->r_num_fwd;
1717         rhead->num_retry = req->r_attempts - 1;
1718         rhead->ino = 0;
1719
1720         dout(" r_locked_dir = %p\n", req->r_locked_dir);
1721         return 0;
1722 }
1723
1724 /*
1725  * send request, or put it on the appropriate wait list.
1726  */
1727 static int __do_request(struct ceph_mds_client *mdsc,
1728                         struct ceph_mds_request *req)
1729 {
1730         struct ceph_mds_session *session = NULL;
1731         int mds = -1;
1732         int err = -EAGAIN;
1733
1734         if (req->r_err || req->r_got_result)
1735                 goto out;
1736
1737         if (req->r_timeout &&
1738             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1739                 dout("do_request timed out\n");
1740                 err = -EIO;
1741                 goto finish;
1742         }
1743
1744         mds = __choose_mds(mdsc, req);
1745         if (mds < 0 ||
1746             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1747                 dout("do_request no mds or not active, waiting for map\n");
1748                 list_add(&req->r_wait, &mdsc->waiting_for_map);
1749                 goto out;
1750         }
1751
1752         /* get, open session */
1753         session = __ceph_lookup_mds_session(mdsc, mds);
1754         if (!session) {
1755                 session = register_session(mdsc, mds);
1756                 if (IS_ERR(session)) {
1757                         err = PTR_ERR(session);
1758                         goto finish;
1759                 }
1760         }
1761         dout("do_request mds%d session %p state %s\n", mds, session,
1762              session_state_name(session->s_state));
1763         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1764             session->s_state != CEPH_MDS_SESSION_HUNG) {
1765                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1766                     session->s_state == CEPH_MDS_SESSION_CLOSING)
1767                         __open_session(mdsc, session);
1768                 list_add(&req->r_wait, &session->s_waiting);
1769                 goto out_session;
1770         }
1771
1772         /* send request */
1773         req->r_session = get_session(session);
1774         req->r_resend_mds = -1;   /* forget any previous mds hint */
1775
1776         if (req->r_request_started == 0)   /* note request start time */
1777                 req->r_request_started = jiffies;
1778
1779         err = __prepare_send_request(mdsc, req, mds);
1780         if (!err) {
1781                 ceph_msg_get(req->r_request);
1782                 ceph_con_send(&session->s_con, req->r_request);
1783         }
1784
1785 out_session:
1786         ceph_put_mds_session(session);
1787 out:
1788         return err;
1789
1790 finish:
1791         req->r_err = err;
1792         complete_request(mdsc, req);
1793         goto out;
1794 }
1795
1796 /*
1797  * called under mdsc->mutex
1798  */
1799 static void __wake_requests(struct ceph_mds_client *mdsc,
1800                             struct list_head *head)
1801 {
1802         struct ceph_mds_request *req, *nreq;
1803
1804         list_for_each_entry_safe(req, nreq, head, r_wait) {
1805                 list_del_init(&req->r_wait);
1806                 __do_request(mdsc, req);
1807         }
1808 }
1809
1810 /*
1811  * Wake up threads with requests pending for @mds, so that they can
1812  * resubmit their requests to a possibly different mds.
1813  */
1814 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
1815 {
1816         struct ceph_mds_request *req;
1817         struct rb_node *p;
1818
1819         dout("kick_requests mds%d\n", mds);
1820         for (p = rb_first(&mdsc->request_tree); p; p = rb_next(p)) {
1821                 req = rb_entry(p, struct ceph_mds_request, r_node);
1822                 if (req->r_got_unsafe)
1823                         continue;
1824                 if (req->r_session &&
1825                     req->r_session->s_mds == mds) {
1826                         dout(" kicking tid %llu\n", req->r_tid);
1827                         put_request_session(req);
1828                         __do_request(mdsc, req);
1829                 }
1830         }
1831 }
1832
1833 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
1834                               struct ceph_mds_request *req)
1835 {
1836         dout("submit_request on %p\n", req);
1837         mutex_lock(&mdsc->mutex);
1838         __register_request(mdsc, req, NULL);
1839         __do_request(mdsc, req);
1840         mutex_unlock(&mdsc->mutex);
1841 }
1842
1843 /*
1844  * Synchrously perform an mds request.  Take care of all of the
1845  * session setup, forwarding, retry details.
1846  */
1847 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
1848                          struct inode *dir,
1849                          struct ceph_mds_request *req)
1850 {
1851         int err;
1852
1853         dout("do_request on %p\n", req);
1854
1855         /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
1856         if (req->r_inode)
1857                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
1858         if (req->r_locked_dir)
1859                 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
1860         if (req->r_old_dentry)
1861                 ceph_get_cap_refs(
1862                         ceph_inode(req->r_old_dentry->d_parent->d_inode),
1863                         CEPH_CAP_PIN);
1864
1865         /* issue */
1866         mutex_lock(&mdsc->mutex);
1867         __register_request(mdsc, req, dir);
1868         __do_request(mdsc, req);
1869
1870         if (req->r_err) {
1871                 err = req->r_err;
1872                 __unregister_request(mdsc, req);
1873                 dout("do_request early error %d\n", err);
1874                 goto out;
1875         }
1876
1877         /* wait */
1878         mutex_unlock(&mdsc->mutex);
1879         dout("do_request waiting\n");
1880         if (req->r_timeout) {
1881                 err = (long)wait_for_completion_killable_timeout(
1882                         &req->r_completion, req->r_timeout);
1883                 if (err == 0)
1884                         err = -EIO;
1885         } else {
1886                 err = wait_for_completion_killable(&req->r_completion);
1887         }
1888         dout("do_request waited, got %d\n", err);
1889         mutex_lock(&mdsc->mutex);
1890
1891         /* only abort if we didn't race with a real reply */
1892         if (req->r_got_result) {
1893                 err = le32_to_cpu(req->r_reply_info.head->result);
1894         } else if (err < 0) {
1895                 dout("aborted request %lld with %d\n", req->r_tid, err);
1896
1897                 /*
1898                  * ensure we aren't running concurrently with
1899                  * ceph_fill_trace or ceph_readdir_prepopulate, which
1900                  * rely on locks (dir mutex) held by our caller.
1901                  */
1902                 mutex_lock(&req->r_fill_mutex);
1903                 req->r_err = err;
1904                 req->r_aborted = true;
1905                 mutex_unlock(&req->r_fill_mutex);
1906
1907                 if (req->r_locked_dir &&
1908                     (req->r_op & CEPH_MDS_OP_WRITE))
1909                         ceph_invalidate_dir_request(req);
1910         } else {
1911                 err = req->r_err;
1912         }
1913
1914 out:
1915         mutex_unlock(&mdsc->mutex);
1916         dout("do_request %p done, result %d\n", req, err);
1917         return err;
1918 }
1919
1920 /*
1921  * Invalidate dir I_COMPLETE, dentry lease state on an aborted MDS
1922  * namespace request.
1923  */
1924 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
1925 {
1926         struct inode *inode = req->r_locked_dir;
1927         struct ceph_inode_info *ci = ceph_inode(inode);
1928
1929         dout("invalidate_dir_request %p (I_COMPLETE, lease(s))\n", inode);
1930         spin_lock(&inode->i_lock);
1931         ci->i_ceph_flags &= ~CEPH_I_COMPLETE;
1932         ci->i_release_count++;
1933         spin_unlock(&inode->i_lock);
1934
1935         if (req->r_dentry)
1936                 ceph_invalidate_dentry_lease(req->r_dentry);
1937         if (req->r_old_dentry)
1938                 ceph_invalidate_dentry_lease(req->r_old_dentry);
1939 }
1940
1941 /*
1942  * Handle mds reply.
1943  *
1944  * We take the session mutex and parse and process the reply immediately.
1945  * This preserves the logical ordering of replies, capabilities, etc., sent
1946  * by the MDS as they are applied to our local cache.
1947  */
1948 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
1949 {
1950         struct ceph_mds_client *mdsc = session->s_mdsc;
1951         struct ceph_mds_request *req;
1952         struct ceph_mds_reply_head *head = msg->front.iov_base;
1953         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
1954         u64 tid;
1955         int err, result;
1956         int mds = session->s_mds;
1957
1958         if (msg->front.iov_len < sizeof(*head)) {
1959                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
1960                 ceph_msg_dump(msg);
1961                 return;
1962         }
1963
1964         /* get request, session */
1965         tid = le64_to_cpu(msg->hdr.tid);
1966         mutex_lock(&mdsc->mutex);
1967         req = __lookup_request(mdsc, tid);
1968         if (!req) {
1969                 dout("handle_reply on unknown tid %llu\n", tid);
1970                 mutex_unlock(&mdsc->mutex);
1971                 return;
1972         }
1973         dout("handle_reply %p\n", req);
1974
1975         /* correct session? */
1976         if (req->r_session != session) {
1977                 pr_err("mdsc_handle_reply got %llu on session mds%d"
1978                        " not mds%d\n", tid, session->s_mds,
1979                        req->r_session ? req->r_session->s_mds : -1);
1980                 mutex_unlock(&mdsc->mutex);
1981                 goto out;
1982         }
1983
1984         /* dup? */
1985         if ((req->r_got_unsafe && !head->safe) ||
1986             (req->r_got_safe && head->safe)) {
1987                 pr_warning("got a dup %s reply on %llu from mds%d\n",
1988                            head->safe ? "safe" : "unsafe", tid, mds);
1989                 mutex_unlock(&mdsc->mutex);
1990                 goto out;
1991         }
1992         if (req->r_got_safe && !head->safe) {
1993                 pr_warning("got unsafe after safe on %llu from mds%d\n",
1994                            tid, mds);
1995                 mutex_unlock(&mdsc->mutex);
1996                 goto out;
1997         }
1998
1999         result = le32_to_cpu(head->result);
2000
2001         /*
2002          * Handle an ESTALE
2003          * if we're not talking to the authority, send to them
2004          * if the authority has changed while we weren't looking,
2005          * send to new authority
2006          * Otherwise we just have to return an ESTALE
2007          */
2008         if (result == -ESTALE) {
2009                 dout("got ESTALE on request %llu", req->r_tid);
2010                 if (!req->r_inode) {
2011                         /* do nothing; not an authority problem */
2012                 } else if (req->r_direct_mode != USE_AUTH_MDS) {
2013                         dout("not using auth, setting for that now");
2014                         req->r_direct_mode = USE_AUTH_MDS;
2015                         __do_request(mdsc, req);
2016                         mutex_unlock(&mdsc->mutex);
2017                         goto out;
2018                 } else  {
2019                         struct ceph_inode_info *ci = ceph_inode(req->r_inode);
2020                         struct ceph_cap *cap =
2021                                 ceph_get_cap_for_mds(ci, req->r_mds);;
2022
2023                         dout("already using auth");
2024                         if ((!cap || cap != ci->i_auth_cap) ||
2025                             (cap->mseq != req->r_sent_on_mseq)) {
2026                                 dout("but cap changed, so resending");
2027                                 __do_request(mdsc, req);
2028                                 mutex_unlock(&mdsc->mutex);
2029                                 goto out;
2030                         }
2031                 }
2032                 dout("have to return ESTALE on request %llu", req->r_tid);
2033         }
2034
2035
2036         if (head->safe) {
2037                 req->r_got_safe = true;
2038                 __unregister_request(mdsc, req);
2039                 complete_all(&req->r_safe_completion);
2040
2041                 if (req->r_got_unsafe) {
2042                         /*
2043                          * We already handled the unsafe response, now do the
2044                          * cleanup.  No need to examine the response; the MDS
2045                          * doesn't include any result info in the safe
2046                          * response.  And even if it did, there is nothing
2047                          * useful we could do with a revised return value.
2048                          */
2049                         dout("got safe reply %llu, mds%d\n", tid, mds);
2050                         list_del_init(&req->r_unsafe_item);
2051
2052                         /* last unsafe request during umount? */
2053                         if (mdsc->stopping && !__get_oldest_req(mdsc))
2054                                 complete_all(&mdsc->safe_umount_waiters);
2055                         mutex_unlock(&mdsc->mutex);
2056                         goto out;
2057                 }
2058         } else {
2059                 req->r_got_unsafe = true;
2060                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2061         }
2062
2063         dout("handle_reply tid %lld result %d\n", tid, result);
2064         rinfo = &req->r_reply_info;
2065         err = parse_reply_info(msg, rinfo);
2066         mutex_unlock(&mdsc->mutex);
2067
2068         mutex_lock(&session->s_mutex);
2069         if (err < 0) {
2070                 pr_err("mdsc_handle_reply got corrupt reply mds%d\n", mds);
2071                 ceph_msg_dump(msg);
2072                 goto out_err;
2073         }
2074
2075         /* snap trace */
2076         if (rinfo->snapblob_len) {
2077                 down_write(&mdsc->snap_rwsem);
2078                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2079                                rinfo->snapblob + rinfo->snapblob_len,
2080                                le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
2081                 downgrade_write(&mdsc->snap_rwsem);
2082         } else {
2083                 down_read(&mdsc->snap_rwsem);
2084         }
2085
2086         /* insert trace into our cache */
2087         mutex_lock(&req->r_fill_mutex);
2088         err = ceph_fill_trace(mdsc->client->sb, req, req->r_session);
2089         if (err == 0) {
2090                 if (result == 0 && rinfo->dir_nr)
2091                         ceph_readdir_prepopulate(req, req->r_session);
2092                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2093         }
2094         mutex_unlock(&req->r_fill_mutex);
2095
2096         up_read(&mdsc->snap_rwsem);
2097 out_err:
2098         mutex_lock(&mdsc->mutex);
2099         if (!req->r_aborted) {
2100                 if (err) {
2101                         req->r_err = err;
2102                 } else {
2103                         req->r_reply = msg;
2104                         ceph_msg_get(msg);
2105                         req->r_got_result = true;
2106                 }
2107         } else {
2108                 dout("reply arrived after request %lld was aborted\n", tid);
2109         }
2110         mutex_unlock(&mdsc->mutex);
2111
2112         ceph_add_cap_releases(mdsc, req->r_session);
2113         mutex_unlock(&session->s_mutex);
2114
2115         /* kick calling process */
2116         complete_request(mdsc, req);
2117 out:
2118         ceph_mdsc_put_request(req);
2119         return;
2120 }
2121
2122
2123
2124 /*
2125  * handle mds notification that our request has been forwarded.
2126  */
2127 static void handle_forward(struct ceph_mds_client *mdsc,
2128                            struct ceph_mds_session *session,
2129                            struct ceph_msg *msg)
2130 {
2131         struct ceph_mds_request *req;
2132         u64 tid = le64_to_cpu(msg->hdr.tid);
2133         u32 next_mds;
2134         u32 fwd_seq;
2135         int err = -EINVAL;
2136         void *p = msg->front.iov_base;
2137         void *end = p + msg->front.iov_len;
2138
2139         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2140         next_mds = ceph_decode_32(&p);
2141         fwd_seq = ceph_decode_32(&p);
2142
2143         mutex_lock(&mdsc->mutex);
2144         req = __lookup_request(mdsc, tid);
2145         if (!req) {
2146                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2147                 goto out;  /* dup reply? */
2148         }
2149
2150         if (req->r_aborted) {
2151                 dout("forward tid %llu aborted, unregistering\n", tid);
2152                 __unregister_request(mdsc, req);
2153         } else if (fwd_seq <= req->r_num_fwd) {
2154                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2155                      tid, next_mds, req->r_num_fwd, fwd_seq);
2156         } else {
2157                 /* resend. forward race not possible; mds would drop */
2158                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2159                 BUG_ON(req->r_err);
2160                 BUG_ON(req->r_got_result);
2161                 req->r_num_fwd = fwd_seq;
2162                 req->r_resend_mds = next_mds;
2163                 put_request_session(req);
2164                 __do_request(mdsc, req);
2165         }
2166         ceph_mdsc_put_request(req);
2167 out:
2168         mutex_unlock(&mdsc->mutex);
2169         return;
2170
2171 bad:
2172         pr_err("mdsc_handle_forward decode error err=%d\n", err);
2173 }
2174
2175 /*
2176  * handle a mds session control message
2177  */
2178 static void handle_session(struct ceph_mds_session *session,
2179                            struct ceph_msg *msg)
2180 {
2181         struct ceph_mds_client *mdsc = session->s_mdsc;
2182         u32 op;
2183         u64 seq;
2184         int mds = session->s_mds;
2185         struct ceph_mds_session_head *h = msg->front.iov_base;
2186         int wake = 0;
2187
2188         /* decode */
2189         if (msg->front.iov_len != sizeof(*h))
2190                 goto bad;
2191         op = le32_to_cpu(h->op);
2192         seq = le64_to_cpu(h->seq);
2193
2194         mutex_lock(&mdsc->mutex);
2195         if (op == CEPH_SESSION_CLOSE)
2196                 __unregister_session(mdsc, session);
2197         /* FIXME: this ttl calculation is generous */
2198         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2199         mutex_unlock(&mdsc->mutex);
2200
2201         mutex_lock(&session->s_mutex);
2202
2203         dout("handle_session mds%d %s %p state %s seq %llu\n",
2204              mds, ceph_session_op_name(op), session,
2205              session_state_name(session->s_state), seq);
2206
2207         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2208                 session->s_state = CEPH_MDS_SESSION_OPEN;
2209                 pr_info("mds%d came back\n", session->s_mds);
2210         }
2211
2212         switch (op) {
2213         case CEPH_SESSION_OPEN:
2214                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2215                         pr_info("mds%d reconnect success\n", session->s_mds);
2216                 session->s_state = CEPH_MDS_SESSION_OPEN;
2217                 renewed_caps(mdsc, session, 0);
2218                 wake = 1;
2219                 if (mdsc->stopping)
2220                         __close_session(mdsc, session);
2221                 break;
2222
2223         case CEPH_SESSION_RENEWCAPS:
2224                 if (session->s_renew_seq == seq)
2225                         renewed_caps(mdsc, session, 1);
2226                 break;
2227
2228         case CEPH_SESSION_CLOSE:
2229                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2230                         pr_info("mds%d reconnect denied\n", session->s_mds);
2231                 remove_session_caps(session);
2232                 wake = 1; /* for good measure */
2233                 wake_up_all(&mdsc->session_close_wq);
2234                 kick_requests(mdsc, mds);
2235                 break;
2236
2237         case CEPH_SESSION_STALE:
2238                 pr_info("mds%d caps went stale, renewing\n",
2239                         session->s_mds);
2240                 spin_lock(&session->s_cap_lock);
2241                 session->s_cap_gen++;
2242                 session->s_cap_ttl = 0;
2243                 spin_unlock(&session->s_cap_lock);
2244                 send_renew_caps(mdsc, session);
2245                 break;
2246
2247         case CEPH_SESSION_RECALL_STATE:
2248                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2249                 break;
2250
2251         default:
2252                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2253                 WARN_ON(1);
2254         }
2255
2256         mutex_unlock(&session->s_mutex);
2257         if (wake) {
2258                 mutex_lock(&mdsc->mutex);
2259                 __wake_requests(mdsc, &session->s_waiting);
2260                 mutex_unlock(&mdsc->mutex);
2261         }
2262         return;
2263
2264 bad:
2265         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2266                (int)msg->front.iov_len);
2267         ceph_msg_dump(msg);
2268         return;
2269 }
2270
2271
2272 /*
2273  * called under session->mutex.
2274  */
2275 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2276                                    struct ceph_mds_session *session)
2277 {
2278         struct ceph_mds_request *req, *nreq;
2279         int err;
2280
2281         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2282
2283         mutex_lock(&mdsc->mutex);
2284         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2285                 err = __prepare_send_request(mdsc, req, session->s_mds);
2286                 if (!err) {
2287                         ceph_msg_get(req->r_request);
2288                         ceph_con_send(&session->s_con, req->r_request);
2289                 }
2290         }
2291         mutex_unlock(&mdsc->mutex);
2292 }
2293
2294 /*
2295  * Encode information about a cap for a reconnect with the MDS.
2296  */
2297 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2298                           void *arg)
2299 {
2300         union {
2301                 struct ceph_mds_cap_reconnect v2;
2302                 struct ceph_mds_cap_reconnect_v1 v1;
2303         } rec;
2304         size_t reclen;
2305         struct ceph_inode_info *ci;
2306         struct ceph_reconnect_state *recon_state = arg;
2307         struct ceph_pagelist *pagelist = recon_state->pagelist;
2308         char *path;
2309         int pathlen, err;
2310         u64 pathbase;
2311         struct dentry *dentry;
2312
2313         ci = cap->ci;
2314
2315         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2316              inode, ceph_vinop(inode), cap, cap->cap_id,
2317              ceph_cap_string(cap->issued));
2318         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2319         if (err)
2320                 return err;
2321
2322         dentry = d_find_alias(inode);
2323         if (dentry) {
2324                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2325                 if (IS_ERR(path)) {
2326                         err = PTR_ERR(path);
2327                         goto out_dput;
2328                 }
2329         } else {
2330                 path = NULL;
2331                 pathlen = 0;
2332         }
2333         err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2334         if (err)
2335                 goto out_free;
2336
2337         spin_lock(&inode->i_lock);
2338         cap->seq = 0;        /* reset cap seq */
2339         cap->issue_seq = 0;  /* and issue_seq */
2340
2341         if (recon_state->flock) {
2342                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2343                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2344                 rec.v2.issued = cpu_to_le32(cap->issued);
2345                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2346                 rec.v2.pathbase = cpu_to_le64(pathbase);
2347                 rec.v2.flock_len = 0;
2348                 reclen = sizeof(rec.v2);
2349         } else {
2350                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2351                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2352                 rec.v1.issued = cpu_to_le32(cap->issued);
2353                 rec.v1.size = cpu_to_le64(inode->i_size);
2354                 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2355                 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2356                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2357                 rec.v1.pathbase = cpu_to_le64(pathbase);
2358                 reclen = sizeof(rec.v1);
2359         }
2360         spin_unlock(&inode->i_lock);
2361
2362         if (recon_state->flock) {
2363                 int num_fcntl_locks, num_flock_locks;
2364
2365                 lock_kernel();
2366                 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2367                 rec.v2.flock_len = (2*sizeof(u32) +
2368                                     (num_fcntl_locks+num_flock_locks) *
2369                                     sizeof(struct ceph_filelock));
2370
2371                 err = ceph_pagelist_append(pagelist, &rec, reclen);
2372                 if (!err)
2373                         err = ceph_encode_locks(inode, pagelist,
2374                                                 num_fcntl_locks,
2375                                                 num_flock_locks);
2376                 unlock_kernel();
2377         } else {
2378                 err = ceph_pagelist_append(pagelist, &rec, reclen);
2379         }
2380
2381 out_free:
2382         kfree(path);
2383 out_dput:
2384         dput(dentry);
2385         return err;
2386 }
2387
2388
2389 /*
2390  * If an MDS fails and recovers, clients need to reconnect in order to
2391  * reestablish shared state.  This includes all caps issued through
2392  * this session _and_ the snap_realm hierarchy.  Because it's not
2393  * clear which snap realms the mds cares about, we send everything we
2394  * know about.. that ensures we'll then get any new info the
2395  * recovering MDS might have.
2396  *
2397  * This is a relatively heavyweight operation, but it's rare.
2398  *
2399  * called with mdsc->mutex held.
2400  */
2401 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2402                                struct ceph_mds_session *session)
2403 {
2404         struct ceph_msg *reply;
2405         struct rb_node *p;
2406         int mds = session->s_mds;
2407         int err = -ENOMEM;
2408         struct ceph_pagelist *pagelist;
2409         struct ceph_reconnect_state recon_state;
2410
2411         pr_info("mds%d reconnect start\n", mds);
2412
2413         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2414         if (!pagelist)
2415                 goto fail_nopagelist;
2416         ceph_pagelist_init(pagelist);
2417
2418         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS);
2419         if (!reply)
2420                 goto fail_nomsg;
2421
2422         mutex_lock(&session->s_mutex);
2423         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2424         session->s_seq = 0;
2425
2426         ceph_con_open(&session->s_con,
2427                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2428
2429         /* replay unsafe requests */
2430         replay_unsafe_requests(mdsc, session);
2431
2432         down_read(&mdsc->snap_rwsem);
2433
2434         dout("session %p state %s\n", session,
2435              session_state_name(session->s_state));
2436
2437         /* drop old cap expires; we're about to reestablish that state */
2438         discard_cap_releases(mdsc, session);
2439
2440         /* traverse this session's caps */
2441         err = ceph_pagelist_encode_32(pagelist, session->s_nr_caps);
2442         if (err)
2443                 goto fail;
2444
2445         recon_state.pagelist = pagelist;
2446         recon_state.flock = session->s_con.peer_features & CEPH_FEATURE_FLOCK;
2447         err = iterate_session_caps(session, encode_caps_cb, &recon_state);
2448         if (err < 0)
2449                 goto fail;
2450
2451         /*
2452          * snaprealms.  we provide mds with the ino, seq (version), and
2453          * parent for all of our realms.  If the mds has any newer info,
2454          * it will tell us.
2455          */
2456         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2457                 struct ceph_snap_realm *realm =
2458                         rb_entry(p, struct ceph_snap_realm, node);
2459                 struct ceph_mds_snaprealm_reconnect sr_rec;
2460
2461                 dout(" adding snap realm %llx seq %lld parent %llx\n",
2462                      realm->ino, realm->seq, realm->parent_ino);
2463                 sr_rec.ino = cpu_to_le64(realm->ino);
2464                 sr_rec.seq = cpu_to_le64(realm->seq);
2465                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2466                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2467                 if (err)
2468                         goto fail;
2469         }
2470
2471         reply->pagelist = pagelist;
2472         if (recon_state.flock)
2473                 reply->hdr.version = cpu_to_le16(2);
2474         reply->hdr.data_len = cpu_to_le32(pagelist->length);
2475         reply->nr_pages = calc_pages_for(0, pagelist->length);
2476         ceph_con_send(&session->s_con, reply);
2477
2478         mutex_unlock(&session->s_mutex);
2479
2480         mutex_lock(&mdsc->mutex);
2481         __wake_requests(mdsc, &session->s_waiting);
2482         mutex_unlock(&mdsc->mutex);
2483
2484         up_read(&mdsc->snap_rwsem);
2485         return;
2486
2487 fail:
2488         ceph_msg_put(reply);
2489         up_read(&mdsc->snap_rwsem);
2490         mutex_unlock(&session->s_mutex);
2491 fail_nomsg:
2492         ceph_pagelist_release(pagelist);
2493         kfree(pagelist);
2494 fail_nopagelist:
2495         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
2496         return;
2497 }
2498
2499
2500 /*
2501  * compare old and new mdsmaps, kicking requests
2502  * and closing out old connections as necessary
2503  *
2504  * called under mdsc->mutex.
2505  */
2506 static void check_new_map(struct ceph_mds_client *mdsc,
2507                           struct ceph_mdsmap *newmap,
2508                           struct ceph_mdsmap *oldmap)
2509 {
2510         int i;
2511         int oldstate, newstate;
2512         struct ceph_mds_session *s;
2513
2514         dout("check_new_map new %u old %u\n",
2515              newmap->m_epoch, oldmap->m_epoch);
2516
2517         for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2518                 if (mdsc->sessions[i] == NULL)
2519                         continue;
2520                 s = mdsc->sessions[i];
2521                 oldstate = ceph_mdsmap_get_state(oldmap, i);
2522                 newstate = ceph_mdsmap_get_state(newmap, i);
2523
2524                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
2525                      i, ceph_mds_state_name(oldstate),
2526                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
2527                      ceph_mds_state_name(newstate),
2528                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
2529                      session_state_name(s->s_state));
2530
2531                 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
2532                            ceph_mdsmap_get_addr(newmap, i),
2533                            sizeof(struct ceph_entity_addr))) {
2534                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2535                                 /* the session never opened, just close it
2536                                  * out now */
2537                                 __wake_requests(mdsc, &s->s_waiting);
2538                                 __unregister_session(mdsc, s);
2539                         } else {
2540                                 /* just close it */
2541                                 mutex_unlock(&mdsc->mutex);
2542                                 mutex_lock(&s->s_mutex);
2543                                 mutex_lock(&mdsc->mutex);
2544                                 ceph_con_close(&s->s_con);
2545                                 mutex_unlock(&s->s_mutex);
2546                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2547                         }
2548
2549                         /* kick any requests waiting on the recovering mds */
2550                         kick_requests(mdsc, i);
2551                 } else if (oldstate == newstate) {
2552                         continue;  /* nothing new with this mds */
2553                 }
2554
2555                 /*
2556                  * send reconnect?
2557                  */
2558                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2559                     newstate >= CEPH_MDS_STATE_RECONNECT) {
2560                         mutex_unlock(&mdsc->mutex);
2561                         send_mds_reconnect(mdsc, s);
2562                         mutex_lock(&mdsc->mutex);
2563                 }
2564
2565                 /*
2566                  * kick request on any mds that has gone active.
2567                  */
2568                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2569                     newstate >= CEPH_MDS_STATE_ACTIVE) {
2570                         if (oldstate != CEPH_MDS_STATE_CREATING &&
2571                             oldstate != CEPH_MDS_STATE_STARTING)
2572                                 pr_info("mds%d recovery completed\n", s->s_mds);
2573                         kick_requests(mdsc, i);
2574                         ceph_kick_flushing_caps(mdsc, s);
2575                         wake_up_session_caps(s, 1);
2576                 }
2577         }
2578
2579         for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
2580                 s = mdsc->sessions[i];
2581                 if (!s)
2582                         continue;
2583                 if (!ceph_mdsmap_is_laggy(newmap, i))
2584                         continue;
2585                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2586                     s->s_state == CEPH_MDS_SESSION_HUNG ||
2587                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
2588                         dout(" connecting to export targets of laggy mds%d\n",
2589                              i);
2590                         __open_export_target_sessions(mdsc, s);
2591                 }
2592         }
2593 }
2594
2595
2596
2597 /*
2598  * leases
2599  */
2600
2601 /*
2602  * caller must hold session s_mutex, dentry->d_lock
2603  */
2604 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2605 {
2606         struct ceph_dentry_info *di = ceph_dentry(dentry);
2607
2608         ceph_put_mds_session(di->lease_session);
2609         di->lease_session = NULL;
2610 }
2611
2612 static void handle_lease(struct ceph_mds_client *mdsc,
2613                          struct ceph_mds_session *session,
2614                          struct ceph_msg *msg)
2615 {
2616         struct super_block *sb = mdsc->client->sb;
2617         struct inode *inode;
2618         struct ceph_inode_info *ci;
2619         struct dentry *parent, *dentry;
2620         struct ceph_dentry_info *di;
2621         int mds = session->s_mds;
2622         struct ceph_mds_lease *h = msg->front.iov_base;
2623         u32 seq;
2624         struct ceph_vino vino;
2625         int mask;
2626         struct qstr dname;
2627         int release = 0;
2628
2629         dout("handle_lease from mds%d\n", mds);
2630
2631         /* decode */
2632         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2633                 goto bad;
2634         vino.ino = le64_to_cpu(h->ino);
2635         vino.snap = CEPH_NOSNAP;
2636         mask = le16_to_cpu(h->mask);
2637         seq = le32_to_cpu(h->seq);
2638         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2639         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2640         if (dname.len != get_unaligned_le32(h+1))
2641                 goto bad;
2642
2643         mutex_lock(&session->s_mutex);
2644         session->s_seq++;
2645
2646         /* lookup inode */
2647         inode = ceph_find_inode(sb, vino);
2648         dout("handle_lease %s, mask %d, ino %llx %p %.*s\n",
2649              ceph_lease_op_name(h->action), mask, vino.ino, inode,
2650              dname.len, dname.name);
2651         if (inode == NULL) {
2652                 dout("handle_lease no inode %llx\n", vino.ino);
2653                 goto release;
2654         }
2655         ci = ceph_inode(inode);
2656
2657         /* dentry */
2658         parent = d_find_alias(inode);
2659         if (!parent) {
2660                 dout("no parent dentry on inode %p\n", inode);
2661                 WARN_ON(1);
2662                 goto release;  /* hrm... */
2663         }
2664         dname.hash = full_name_hash(dname.name, dname.len);
2665         dentry = d_lookup(parent, &dname);
2666         dput(parent);
2667         if (!dentry)
2668                 goto release;
2669
2670         spin_lock(&dentry->d_lock);
2671         di = ceph_dentry(dentry);
2672         switch (h->action) {
2673         case CEPH_MDS_LEASE_REVOKE:
2674                 if (di && di->lease_session == session) {
2675                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
2676                                 h->seq = cpu_to_le32(di->lease_seq);
2677                         __ceph_mdsc_drop_dentry_lease(dentry);
2678                 }
2679                 release = 1;
2680                 break;
2681
2682         case CEPH_MDS_LEASE_RENEW:
2683                 if (di && di->lease_session == session &&
2684                     di->lease_gen == session->s_cap_gen &&
2685                     di->lease_renew_from &&
2686                     di->lease_renew_after == 0) {
2687                         unsigned long duration =
2688                                 le32_to_cpu(h->duration_ms) * HZ / 1000;
2689
2690                         di->lease_seq = seq;
2691                         dentry->d_time = di->lease_renew_from + duration;
2692                         di->lease_renew_after = di->lease_renew_from +
2693                                 (duration >> 1);
2694                         di->lease_renew_from = 0;
2695                 }
2696                 break;
2697         }
2698         spin_unlock(&dentry->d_lock);
2699         dput(dentry);
2700
2701         if (!release)
2702                 goto out;
2703
2704 release:
2705         /* let's just reuse the same message */
2706         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2707         ceph_msg_get(msg);
2708         ceph_con_send(&session->s_con, msg);
2709
2710 out:
2711         iput(inode);
2712         mutex_unlock(&session->s_mutex);
2713         return;
2714
2715 bad:
2716         pr_err("corrupt lease message\n");
2717         ceph_msg_dump(msg);
2718 }
2719
2720 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
2721                               struct inode *inode,
2722                               struct dentry *dentry, char action,
2723                               u32 seq)
2724 {
2725         struct ceph_msg *msg;
2726         struct ceph_mds_lease *lease;
2727         int len = sizeof(*lease) + sizeof(u32);
2728         int dnamelen = 0;
2729
2730         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2731              inode, dentry, ceph_lease_op_name(action), session->s_mds);
2732         dnamelen = dentry->d_name.len;
2733         len += dnamelen;
2734
2735         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS);
2736         if (!msg)
2737                 return;
2738         lease = msg->front.iov_base;
2739         lease->action = action;
2740         lease->mask = cpu_to_le16(1);
2741         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
2742         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
2743         lease->seq = cpu_to_le32(seq);
2744         put_unaligned_le32(dnamelen, lease + 1);
2745         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
2746
2747         /*
2748          * if this is a preemptive lease RELEASE, no need to
2749          * flush request stream, since the actual request will
2750          * soon follow.
2751          */
2752         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
2753
2754         ceph_con_send(&session->s_con, msg);
2755 }
2756
2757 /*
2758  * Preemptively release a lease we expect to invalidate anyway.
2759  * Pass @inode always, @dentry is optional.
2760  */
2761 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2762                              struct dentry *dentry, int mask)
2763 {
2764         struct ceph_dentry_info *di;
2765         struct ceph_mds_session *session;
2766         u32 seq;
2767
2768         BUG_ON(inode == NULL);
2769         BUG_ON(dentry == NULL);
2770         BUG_ON(mask == 0);
2771
2772         /* is dentry lease valid? */
2773         spin_lock(&dentry->d_lock);
2774         di = ceph_dentry(dentry);
2775         if (!di || !di->lease_session ||
2776             di->lease_session->s_mds < 0 ||
2777             di->lease_gen != di->lease_session->s_cap_gen ||
2778             !time_before(jiffies, dentry->d_time)) {
2779                 dout("lease_release inode %p dentry %p -- "
2780                      "no lease on %d\n",
2781                      inode, dentry, mask);
2782                 spin_unlock(&dentry->d_lock);
2783                 return;
2784         }
2785
2786         /* we do have a lease on this dentry; note mds and seq */
2787         session = ceph_get_mds_session(di->lease_session);
2788         seq = di->lease_seq;
2789         __ceph_mdsc_drop_dentry_lease(dentry);
2790         spin_unlock(&dentry->d_lock);
2791
2792         dout("lease_release inode %p dentry %p mask %d to mds%d\n",
2793              inode, dentry, mask, session->s_mds);
2794         ceph_mdsc_lease_send_msg(session, inode, dentry,
2795                                  CEPH_MDS_LEASE_RELEASE, seq);
2796         ceph_put_mds_session(session);
2797 }
2798
2799 /*
2800  * drop all leases (and dentry refs) in preparation for umount
2801  */
2802 static void drop_leases(struct ceph_mds_client *mdsc)
2803 {
2804         int i;
2805
2806         dout("drop_leases\n");
2807         mutex_lock(&mdsc->mutex);
2808         for (i = 0; i < mdsc->max_sessions; i++) {
2809                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2810                 if (!s)
2811                         continue;
2812                 mutex_unlock(&mdsc->mutex);
2813                 mutex_lock(&s->s_mutex);
2814                 mutex_unlock(&s->s_mutex);
2815                 ceph_put_mds_session(s);
2816                 mutex_lock(&mdsc->mutex);
2817         }
2818         mutex_unlock(&mdsc->mutex);
2819 }
2820
2821
2822
2823 /*
2824  * delayed work -- periodically trim expired leases, renew caps with mds
2825  */
2826 static void schedule_delayed(struct ceph_mds_client *mdsc)
2827 {
2828         int delay = 5;
2829         unsigned hz = round_jiffies_relative(HZ * delay);
2830         schedule_delayed_work(&mdsc->delayed_work, hz);
2831 }
2832
2833 static void delayed_work(struct work_struct *work)
2834 {
2835         int i;
2836         struct ceph_mds_client *mdsc =
2837                 container_of(work, struct ceph_mds_client, delayed_work.work);
2838         int renew_interval;
2839         int renew_caps;
2840
2841         dout("mdsc delayed_work\n");
2842         ceph_check_delayed_caps(mdsc);
2843
2844         mutex_lock(&mdsc->mutex);
2845         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
2846         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
2847                                    mdsc->last_renew_caps);
2848         if (renew_caps)
2849                 mdsc->last_renew_caps = jiffies;
2850
2851         for (i = 0; i < mdsc->max_sessions; i++) {
2852                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2853                 if (s == NULL)
2854                         continue;
2855                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
2856                         dout("resending session close request for mds%d\n",
2857                              s->s_mds);
2858                         request_close_session(mdsc, s);
2859                         ceph_put_mds_session(s);
2860                         continue;
2861                 }
2862                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
2863                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
2864                                 s->s_state = CEPH_MDS_SESSION_HUNG;
2865                                 pr_info("mds%d hung\n", s->s_mds);
2866                         }
2867                 }
2868                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
2869                         /* this mds is failed or recovering, just wait */
2870                         ceph_put_mds_session(s);
2871                         continue;
2872                 }
2873                 mutex_unlock(&mdsc->mutex);
2874
2875                 mutex_lock(&s->s_mutex);
2876                 if (renew_caps)
2877                         send_renew_caps(mdsc, s);
2878                 else
2879                         ceph_con_keepalive(&s->s_con);
2880                 ceph_add_cap_releases(mdsc, s);
2881                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2882                     s->s_state == CEPH_MDS_SESSION_HUNG)
2883                         ceph_send_cap_releases(mdsc, s);
2884                 mutex_unlock(&s->s_mutex);
2885                 ceph_put_mds_session(s);
2886
2887                 mutex_lock(&mdsc->mutex);
2888         }
2889         mutex_unlock(&mdsc->mutex);
2890
2891         schedule_delayed(mdsc);
2892 }
2893
2894
2895 int ceph_mdsc_init(struct ceph_mds_client *mdsc, struct ceph_client *client)
2896 {
2897         mdsc->client = client;
2898         mutex_init(&mdsc->mutex);
2899         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
2900         if (mdsc->mdsmap == NULL)
2901                 return -ENOMEM;
2902
2903         init_completion(&mdsc->safe_umount_waiters);
2904         init_waitqueue_head(&mdsc->session_close_wq);
2905         INIT_LIST_HEAD(&mdsc->waiting_for_map);
2906         mdsc->sessions = NULL;
2907         mdsc->max_sessions = 0;
2908         mdsc->stopping = 0;
2909         init_rwsem(&mdsc->snap_rwsem);
2910         mdsc->snap_realms = RB_ROOT;
2911         INIT_LIST_HEAD(&mdsc->snap_empty);
2912         spin_lock_init(&mdsc->snap_empty_lock);
2913         mdsc->last_tid = 0;
2914         mdsc->request_tree = RB_ROOT;
2915         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
2916         mdsc->last_renew_caps = jiffies;
2917         INIT_LIST_HEAD(&mdsc->cap_delay_list);
2918         spin_lock_init(&mdsc->cap_delay_lock);
2919         INIT_LIST_HEAD(&mdsc->snap_flush_list);
2920         spin_lock_init(&mdsc->snap_flush_lock);
2921         mdsc->cap_flush_seq = 0;
2922         INIT_LIST_HEAD(&mdsc->cap_dirty);
2923         mdsc->num_cap_flushing = 0;
2924         spin_lock_init(&mdsc->cap_dirty_lock);
2925         init_waitqueue_head(&mdsc->cap_flushing_wq);
2926         spin_lock_init(&mdsc->dentry_lru_lock);
2927         INIT_LIST_HEAD(&mdsc->dentry_lru);
2928
2929         ceph_caps_init(mdsc);
2930         ceph_adjust_min_caps(mdsc, client->min_caps);
2931
2932         return 0;
2933 }
2934
2935 /*
2936  * Wait for safe replies on open mds requests.  If we time out, drop
2937  * all requests from the tree to avoid dangling dentry refs.
2938  */
2939 static void wait_requests(struct ceph_mds_client *mdsc)
2940 {
2941         struct ceph_mds_request *req;
2942         struct ceph_client *client = mdsc->client;
2943
2944         mutex_lock(&mdsc->mutex);
2945         if (__get_oldest_req(mdsc)) {
2946                 mutex_unlock(&mdsc->mutex);
2947
2948                 dout("wait_requests waiting for requests\n");
2949                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
2950                                     client->mount_args->mount_timeout * HZ);
2951
2952                 /* tear down remaining requests */
2953                 mutex_lock(&mdsc->mutex);
2954                 while ((req = __get_oldest_req(mdsc))) {
2955                         dout("wait_requests timed out on tid %llu\n",
2956                              req->r_tid);
2957                         __unregister_request(mdsc, req);
2958                 }
2959         }
2960         mutex_unlock(&mdsc->mutex);
2961         dout("wait_requests done\n");
2962 }
2963
2964 /*
2965  * called before mount is ro, and before dentries are torn down.
2966  * (hmm, does this still race with new lookups?)
2967  */
2968 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
2969 {
2970         dout("pre_umount\n");
2971         mdsc->stopping = 1;
2972
2973         drop_leases(mdsc);
2974         ceph_flush_dirty_caps(mdsc);
2975         wait_requests(mdsc);
2976
2977         /*
2978          * wait for reply handlers to drop their request refs and
2979          * their inode/dcache refs
2980          */
2981         ceph_msgr_flush();
2982 }
2983
2984 /*
2985  * wait for all write mds requests to flush.
2986  */
2987 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
2988 {
2989         struct ceph_mds_request *req = NULL, *nextreq;
2990         struct rb_node *n;
2991
2992         mutex_lock(&mdsc->mutex);
2993         dout("wait_unsafe_requests want %lld\n", want_tid);
2994 restart:
2995         req = __get_oldest_req(mdsc);
2996         while (req && req->r_tid <= want_tid) {
2997                 /* find next request */
2998                 n = rb_next(&req->r_node);
2999                 if (n)
3000                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3001                 else
3002                         nextreq = NULL;
3003                 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
3004                         /* write op */
3005                         ceph_mdsc_get_request(req);
3006                         if (nextreq)
3007                                 ceph_mdsc_get_request(nextreq);
3008                         mutex_unlock(&mdsc->mutex);
3009                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
3010                              req->r_tid, want_tid);
3011                         wait_for_completion(&req->r_safe_completion);
3012                         mutex_lock(&mdsc->mutex);
3013                         ceph_mdsc_put_request(req);
3014                         if (!nextreq)
3015                                 break;  /* next dne before, so we're done! */
3016                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
3017                                 /* next request was removed from tree */
3018                                 ceph_mdsc_put_request(nextreq);
3019                                 goto restart;
3020                         }
3021                         ceph_mdsc_put_request(nextreq);  /* won't go away */
3022                 }
3023                 req = nextreq;
3024         }
3025         mutex_unlock(&mdsc->mutex);
3026         dout("wait_unsafe_requests done\n");
3027 }
3028
3029 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3030 {
3031         u64 want_tid, want_flush;
3032
3033         if (mdsc->client->mount_state == CEPH_MOUNT_SHUTDOWN)
3034                 return;
3035
3036         dout("sync\n");
3037         mutex_lock(&mdsc->mutex);
3038         want_tid = mdsc->last_tid;
3039         want_flush = mdsc->cap_flush_seq;
3040         mutex_unlock(&mdsc->mutex);
3041         dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
3042
3043         ceph_flush_dirty_caps(mdsc);
3044
3045         wait_unsafe_requests(mdsc, want_tid);
3046         wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
3047 }
3048
3049 /*
3050  * true if all sessions are closed, or we force unmount
3051  */
3052 bool done_closing_sessions(struct ceph_mds_client *mdsc)
3053 {
3054         int i, n = 0;
3055
3056         if (mdsc->client->mount_state == CEPH_MOUNT_SHUTDOWN)
3057                 return true;
3058
3059         mutex_lock(&mdsc->mutex);
3060         for (i = 0; i < mdsc->max_sessions; i++)
3061                 if (mdsc->sessions[i])
3062                         n++;
3063         mutex_unlock(&mdsc->mutex);
3064         return n == 0;
3065 }
3066
3067 /*
3068  * called after sb is ro.
3069  */
3070 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3071 {
3072         struct ceph_mds_session *session;
3073         int i;
3074         struct ceph_client *client = mdsc->client;
3075         unsigned long timeout = client->mount_args->mount_timeout * HZ;
3076
3077         dout("close_sessions\n");
3078
3079         /* close sessions */
3080         mutex_lock(&mdsc->mutex);
3081         for (i = 0; i < mdsc->max_sessions; i++) {
3082                 session = __ceph_lookup_mds_session(mdsc, i);
3083                 if (!session)
3084                         continue;
3085                 mutex_unlock(&mdsc->mutex);
3086                 mutex_lock(&session->s_mutex);
3087                 __close_session(mdsc, session);
3088                 mutex_unlock(&session->s_mutex);
3089                 ceph_put_mds_session(session);
3090                 mutex_lock(&mdsc->mutex);
3091         }
3092         mutex_unlock(&mdsc->mutex);
3093
3094         dout("waiting for sessions to close\n");
3095         wait_event_timeout(mdsc->session_close_wq, done_closing_sessions(mdsc),
3096                            timeout);
3097
3098         /* tear down remaining sessions */
3099         mutex_lock(&mdsc->mutex);
3100         for (i = 0; i < mdsc->max_sessions; i++) {
3101                 if (mdsc->sessions[i]) {
3102                         session = get_session(mdsc->sessions[i]);
3103                         __unregister_session(mdsc, session);
3104                         mutex_unlock(&mdsc->mutex);
3105                         mutex_lock(&session->s_mutex);
3106                         remove_session_caps(session);
3107                         mutex_unlock(&session->s_mutex);
3108                         ceph_put_mds_session(session);
3109                         mutex_lock(&mdsc->mutex);
3110                 }
3111         }
3112         WARN_ON(!list_empty(&mdsc->cap_delay_list));
3113         mutex_unlock(&mdsc->mutex);
3114
3115         ceph_cleanup_empty_realms(mdsc);
3116
3117         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3118
3119         dout("stopped\n");
3120 }
3121
3122 void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3123 {
3124         dout("stop\n");
3125         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3126         if (mdsc->mdsmap)
3127                 ceph_mdsmap_destroy(mdsc->mdsmap);
3128         kfree(mdsc->sessions);
3129         ceph_caps_finalize(mdsc);
3130 }
3131
3132
3133 /*
3134  * handle mds map update.
3135  */
3136 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3137 {
3138         u32 epoch;
3139         u32 maplen;
3140         void *p = msg->front.iov_base;
3141         void *end = p + msg->front.iov_len;
3142         struct ceph_mdsmap *newmap, *oldmap;
3143         struct ceph_fsid fsid;
3144         int err = -EINVAL;
3145
3146         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3147         ceph_decode_copy(&p, &fsid, sizeof(fsid));
3148         if (ceph_check_fsid(mdsc->client, &fsid) < 0)
3149                 return;
3150         epoch = ceph_decode_32(&p);
3151         maplen = ceph_decode_32(&p);
3152         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3153
3154         /* do we need it? */
3155         ceph_monc_got_mdsmap(&mdsc->client->monc, epoch);
3156         mutex_lock(&mdsc->mutex);
3157         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3158                 dout("handle_map epoch %u <= our %u\n",
3159                      epoch, mdsc->mdsmap->m_epoch);
3160                 mutex_unlock(&mdsc->mutex);
3161                 return;
3162         }
3163
3164         newmap = ceph_mdsmap_decode(&p, end);
3165         if (IS_ERR(newmap)) {
3166                 err = PTR_ERR(newmap);
3167                 goto bad_unlock;
3168         }
3169
3170         /* swap into place */
3171         if (mdsc->mdsmap) {
3172                 oldmap = mdsc->mdsmap;
3173                 mdsc->mdsmap = newmap;
3174                 check_new_map(mdsc, newmap, oldmap);
3175                 ceph_mdsmap_destroy(oldmap);
3176         } else {
3177                 mdsc->mdsmap = newmap;  /* first mds map */
3178         }
3179         mdsc->client->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3180
3181         __wake_requests(mdsc, &mdsc->waiting_for_map);
3182
3183         mutex_unlock(&mdsc->mutex);
3184         schedule_delayed(mdsc);
3185         return;
3186
3187 bad_unlock:
3188         mutex_unlock(&mdsc->mutex);
3189 bad:
3190         pr_err("error decoding mdsmap %d\n", err);
3191         return;
3192 }
3193
3194 static struct ceph_connection *con_get(struct ceph_connection *con)
3195 {
3196         struct ceph_mds_session *s = con->private;
3197
3198         if (get_session(s)) {
3199                 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3200                 return con;
3201         }
3202         dout("mdsc con_get %p FAIL\n", s);
3203         return NULL;
3204 }
3205
3206 static void con_put(struct ceph_connection *con)
3207 {
3208         struct ceph_mds_session *s = con->private;
3209
3210         ceph_put_mds_session(s);
3211         dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref));
3212 }
3213
3214 /*
3215  * if the client is unresponsive for long enough, the mds will kill
3216  * the session entirely.
3217  */
3218 static void peer_reset(struct ceph_connection *con)
3219 {
3220         struct ceph_mds_session *s = con->private;
3221         struct ceph_mds_client *mdsc = s->s_mdsc;
3222
3223         pr_warning("mds%d closed our session\n", s->s_mds);
3224         send_mds_reconnect(mdsc, s);
3225 }
3226
3227 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3228 {
3229         struct ceph_mds_session *s = con->private;
3230         struct ceph_mds_client *mdsc = s->s_mdsc;
3231         int type = le16_to_cpu(msg->hdr.type);
3232
3233         mutex_lock(&mdsc->mutex);
3234         if (__verify_registered_session(mdsc, s) < 0) {
3235                 mutex_unlock(&mdsc->mutex);
3236                 goto out;
3237         }
3238         mutex_unlock(&mdsc->mutex);
3239
3240         switch (type) {
3241         case CEPH_MSG_MDS_MAP:
3242                 ceph_mdsc_handle_map(mdsc, msg);
3243                 break;
3244         case CEPH_MSG_CLIENT_SESSION:
3245                 handle_session(s, msg);
3246                 break;
3247         case CEPH_MSG_CLIENT_REPLY:
3248                 handle_reply(s, msg);
3249                 break;
3250         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3251                 handle_forward(mdsc, s, msg);
3252                 break;
3253         case CEPH_MSG_CLIENT_CAPS:
3254                 ceph_handle_caps(s, msg);
3255                 break;
3256         case CEPH_MSG_CLIENT_SNAP:
3257                 ceph_handle_snap(mdsc, s, msg);
3258                 break;
3259         case CEPH_MSG_CLIENT_LEASE:
3260                 handle_lease(mdsc, s, msg);
3261                 break;
3262
3263         default:
3264                 pr_err("received unknown message type %d %s\n", type,
3265                        ceph_msg_type_name(type));
3266         }
3267 out:
3268         ceph_msg_put(msg);
3269 }
3270
3271 /*
3272  * authentication
3273  */
3274 static int get_authorizer(struct ceph_connection *con,
3275                           void **buf, int *len, int *proto,
3276                           void **reply_buf, int *reply_len, int force_new)
3277 {
3278         struct ceph_mds_session *s = con->private;
3279         struct ceph_mds_client *mdsc = s->s_mdsc;
3280         struct ceph_auth_client *ac = mdsc->client->monc.auth;
3281         int ret = 0;
3282
3283         if (force_new && s->s_authorizer) {
3284                 ac->ops->destroy_authorizer(ac, s->s_authorizer);
3285                 s->s_authorizer = NULL;
3286         }
3287         if (s->s_authorizer == NULL) {
3288                 if (ac->ops->create_authorizer) {
3289                         ret = ac->ops->create_authorizer(
3290                                 ac, CEPH_ENTITY_TYPE_MDS,
3291                                 &s->s_authorizer,
3292                                 &s->s_authorizer_buf,
3293                                 &s->s_authorizer_buf_len,
3294                                 &s->s_authorizer_reply_buf,
3295                                 &s->s_authorizer_reply_buf_len);
3296                         if (ret)
3297                                 return ret;
3298                 }
3299         }
3300
3301         *proto = ac->protocol;
3302         *buf = s->s_authorizer_buf;
3303         *len = s->s_authorizer_buf_len;
3304         *reply_buf = s->s_authorizer_reply_buf;
3305         *reply_len = s->s_authorizer_reply_buf_len;
3306         return 0;
3307 }
3308
3309
3310 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3311 {
3312         struct ceph_mds_session *s = con->private;
3313         struct ceph_mds_client *mdsc = s->s_mdsc;
3314         struct ceph_auth_client *ac = mdsc->client->monc.auth;
3315
3316         return ac->ops->verify_authorizer_reply(ac, s->s_authorizer, len);
3317 }
3318
3319 static int invalidate_authorizer(struct ceph_connection *con)
3320 {
3321         struct ceph_mds_session *s = con->private;
3322         struct ceph_mds_client *mdsc = s->s_mdsc;
3323         struct ceph_auth_client *ac = mdsc->client->monc.auth;
3324
3325         if (ac->ops->invalidate_authorizer)
3326                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3327
3328         return ceph_monc_validate_auth(&mdsc->client->monc);
3329 }
3330
3331 static const struct ceph_connection_operations mds_con_ops = {
3332         .get = con_get,
3333         .put = con_put,
3334         .dispatch = dispatch,
3335         .get_authorizer = get_authorizer,
3336         .verify_authorizer_reply = verify_authorizer_reply,
3337         .invalidate_authorizer = invalidate_authorizer,
3338         .peer_reset = peer_reset,
3339 };
3340
3341
3342
3343
3344 /* eof */