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  38  * are no longer valid.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