fs/ceph/mds_client.c

   1 #include <linux/ceph/ceph_debug.h>
   2
   3 #include <linux/fs.h>
   4 #include <linux/wait.h>
   5 #include <linux/slab.h>
   6 #include <linux/gfp.h>
   7 #include <linux/sched.h>
   8 #include <linux/debugfs.h>
   9 #include <linux/seq_file.h>
  10 #include <linux/utsname.h>
  11
  12 #include "super.h"
  13 #include "mds_client.h"
  14
  15 #include <linux/ceph/ceph_features.h>
  16 #include <linux/ceph/messenger.h>
  17 #include <linux/ceph/decode.h>
  18 #include <linux/ceph/pagelist.h>
  19 #include <linux/ceph/auth.h>
  20 #include <linux/ceph/debugfs.h>
  21
  22 /*
  23  * A cluster of MDS (metadata server) daemons is responsible for
  24  * managing the file system namespace (the directory hierarchy and
  25  * inodes) and for coordinating shared access to storage.  Metadata is
  26  * partitioning hierarchically across a number of servers, and that
  27  * partition varies over time as the cluster adjusts the distribution
  28  * in order to balance load.
  29  *
  30  * The MDS client is primarily responsible to managing synchronous
  31  * metadata requests for operations like open, unlink, and so forth.
  32  * If there is a MDS failure, we find out about it when we (possibly
  33  * request and) receive a new MDS map, and can resubmit affected
  34  * requests.
  35  *
  36  * For the most part, though, we take advantage of a lossless
  37  * communications channel to the MDS, and do not need to worry about
  38  * timing out or resubmitting requests.
  39  *
  40  * We maintain a stateful "session" with each MDS we interact with.
  41  * Within each session, we sent periodic heartbeat messages to ensure
  42  * any capabilities or leases we have been issues remain valid.  If
  43  * the session times out and goes stale, our leases and capabilities
  44  * are no longer valid.
  45  */
  46
  47 struct ceph_reconnect_state {
  48         int nr_caps;
  49         struct ceph_pagelist *pagelist;
  50         bool flock;
  51 };
  52
  53 static void __wake_requests(struct ceph_mds_client *mdsc,
  54                             struct list_head *head);
  55
  56 static const struct ceph_connection_operations mds_con_ops;
  57
  58
  59 /*
  60  * mds reply parsing
  61  */
  62
  63 /*
  64  * parse individual inode info
  65  */
  66 static int parse_reply_info_in(void **p, void *end,
  67                                struct ceph_mds_reply_info_in *info,
  68                                u64 features)
  69 {
  70         int err = -EIO;
  71
  72         info->in = *p;
  73         *p += sizeof(struct ceph_mds_reply_inode) +
  74                 sizeof(*info->in->fragtree.splits) *
  75                 le32_to_cpu(info->in->fragtree.nsplits);
  76
  77         ceph_decode_32_safe(p, end, info->symlink_len, bad);
  78         ceph_decode_need(p, end, info->symlink_len, bad);
  79         info->symlink = *p;
  80         *p += info->symlink_len;
  81
  82         if (features & CEPH_FEATURE_DIRLAYOUTHASH)
  83                 ceph_decode_copy_safe(p, end, &info->dir_layout,
  84                                       sizeof(info->dir_layout), bad);
  85         else
  86                 memset(&info->dir_layout, 0, sizeof(info->dir_layout));
  87
  88         ceph_decode_32_safe(p, end, info->xattr_len, bad);
  89         ceph_decode_need(p, end, info->xattr_len, bad);
  90         info->xattr_data = *p;
  91         *p += info->xattr_len;
  92
  93         if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
  94                 ceph_decode_64_safe(p, end, info->inline_version, bad);
  95                 ceph_decode_32_safe(p, end, info->inline_len, bad);
  96                 ceph_decode_need(p, end, info->inline_len, bad);
  97                 info->inline_data = *p;
  98                 *p += info->inline_len;
  99         } else
 100                 info->inline_version = CEPH_INLINE_NONE;
 101
 102         return 0;
 103 bad:
 104         return err;
 105 }
 106
 107 /*
 108  * parse a normal reply, which may contain a (dir+)dentry and/or a
 109  * target inode.
 110  */
 111 static int parse_reply_info_trace(void **p, void *end,
 112                                   struct ceph_mds_reply_info_parsed *info,
 113                                   u64 features)
 114 {
 115         int err;
 116
 117         if (info->head->is_dentry) {
 118                 err = parse_reply_info_in(p, end, &info->diri, features);
 119                 if (err < 0)
 120                         goto out_bad;
 121
 122                 if (unlikely(*p + sizeof(*info->dirfrag) > end))
 123                         goto bad;
 124                 info->dirfrag = *p;
 125                 *p += sizeof(*info->dirfrag) +
 126                         sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
 127                 if (unlikely(*p > end))
 128                         goto bad;
 129
 130                 ceph_decode_32_safe(p, end, info->dname_len, bad);
 131                 ceph_decode_need(p, end, info->dname_len, bad);
 132                 info->dname = *p;
 133                 *p += info->dname_len;
 134                 info->dlease = *p;
 135                 *p += sizeof(*info->dlease);
 136         }
 137
 138         if (info->head->is_target) {
 139                 err = parse_reply_info_in(p, end, &info->targeti, features);
 140                 if (err < 0)
 141                         goto out_bad;
 142         }
 143
 144         if (unlikely(*p != end))
 145                 goto bad;
 146         return 0;
 147
 148 bad:
 149         err = -EIO;
 150 out_bad:
 151         pr_err("problem parsing mds trace %d\n", err);
 152         return err;
 153 }
 154
 155 /*
 156  * parse readdir results
 157  */
 158 static int parse_reply_info_dir(void **p, void *end,
 159                                 struct ceph_mds_reply_info_parsed *info,
 160                                 u64 features)
 161 {
 162         u32 num, i = 0;
 163         int err;
 164
 165         info->dir_dir = *p;
 166         if (*p + sizeof(*info->dir_dir) > end)
 167                 goto bad;
 168         *p += sizeof(*info->dir_dir) +
 169                 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
 170         if (*p > end)
 171                 goto bad;
 172
 173         ceph_decode_need(p, end, sizeof(num) + 2, bad);
 174         num = ceph_decode_32(p);
 175         info->dir_end = ceph_decode_8(p);
 176         info->dir_complete = ceph_decode_8(p);
 177         if (num == 0)
 178                 goto done;
 179
 180         BUG_ON(!info->dir_in);
 181         info->dir_dname = (void *)(info->dir_in + num);
 182         info->dir_dname_len = (void *)(info->dir_dname + num);
 183         info->dir_dlease = (void *)(info->dir_dname_len + num);
 184         if ((unsigned long)(info->dir_dlease + num) >
 185             (unsigned long)info->dir_in + info->dir_buf_size) {
 186                 pr_err("dir contents are larger than expected\n");
 187                 WARN_ON(1);
 188                 goto bad;
 189         }
 190
 191         info->dir_nr = num;
 192         while (num) {
 193                 /* dentry */
 194                 ceph_decode_need(p, end, sizeof(u32)*2, bad);
 195                 info->dir_dname_len[i] = ceph_decode_32(p);
 196                 ceph_decode_need(p, end, info->dir_dname_len[i], bad);
 197                 info->dir_dname[i] = *p;
 198                 *p += info->dir_dname_len[i];
 199                 dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
 200                      info->dir_dname[i]);
 201                 info->dir_dlease[i] = *p;
 202                 *p += sizeof(struct ceph_mds_reply_lease);
 203
 204                 /* inode */
 205                 err = parse_reply_info_in(p, end, &info->dir_in[i], features);
 206                 if (err < 0)
 207                         goto out_bad;
 208                 i++;
 209                 num--;
 210         }
 211
 212 done:
 213         if (*p != end)
 214                 goto bad;
 215         return 0;
 216
 217 bad:
 218         err = -EIO;
 219 out_bad:
 220         pr_err("problem parsing dir contents %d\n", err);
 221         return err;
 222 }
 223
 224 /*
 225  * parse fcntl F_GETLK results
 226  */
 227 static int parse_reply_info_filelock(void **p, void *end,
 228                                      struct ceph_mds_reply_info_parsed *info,
 229                                      u64 features)
 230 {
 231         if (*p + sizeof(*info->filelock_reply) > end)
 232                 goto bad;
 233
 234         info->filelock_reply = *p;
 235         *p += sizeof(*info->filelock_reply);
 236
 237         if (unlikely(*p != end))
 238                 goto bad;
 239         return 0;
 240
 241 bad:
 242         return -EIO;
 243 }
 244
 245 /*
 246  * parse create results
 247  */
 248 static int parse_reply_info_create(void **p, void *end,
 249                                   struct ceph_mds_reply_info_parsed *info,
 250                                   u64 features)
 251 {
 252         if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
 253                 if (*p == end) {
 254                         info->has_create_ino = false;
 255                 } else {
 256                         info->has_create_ino = true;
 257                         info->ino = ceph_decode_64(p);
 258                 }
 259         }
 260
 261         if (unlikely(*p != end))
 262                 goto bad;
 263         return 0;
 264
 265 bad:
 266         return -EIO;
 267 }
 268
 269 /*
 270  * parse extra results
 271  */
 272 static int parse_reply_info_extra(void **p, void *end,
 273                                   struct ceph_mds_reply_info_parsed *info,
 274                                   u64 features)
 275 {
 276         if (info->head->op == CEPH_MDS_OP_GETFILELOCK)
 277                 return parse_reply_info_filelock(p, end, info, features);
 278         else if (info->head->op == CEPH_MDS_OP_READDIR ||
 279                  info->head->op == CEPH_MDS_OP_LSSNAP)
 280                 return parse_reply_info_dir(p, end, info, features);
 281         else if (info->head->op == CEPH_MDS_OP_CREATE)
 282                 return parse_reply_info_create(p, end, info, features);
 283         else
 284                 return -EIO;
 285 }
 286
 287 /*
 288  * parse entire mds reply
 289  */
 290 static int parse_reply_info(struct ceph_msg *msg,
 291                             struct ceph_mds_reply_info_parsed *info,
 292                             u64 features)
 293 {
 294         void *p, *end;
 295         u32 len;
 296         int err;
 297
 298         info->head = msg->front.iov_base;
 299         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 300         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 301
 302         /* trace */
 303         ceph_decode_32_safe(&p, end, len, bad);
 304         if (len > 0) {
 305                 ceph_decode_need(&p, end, len, bad);
 306                 err = parse_reply_info_trace(&p, p+len, info, features);
 307                 if (err < 0)
 308                         goto out_bad;
 309         }
 310
 311         /* extra */
 312         ceph_decode_32_safe(&p, end, len, bad);
 313         if (len > 0) {
 314                 ceph_decode_need(&p, end, len, bad);
 315                 err = parse_reply_info_extra(&p, p+len, info, features);
 316                 if (err < 0)
 317                         goto out_bad;
 318         }
 319
 320         /* snap blob */
 321         ceph_decode_32_safe(&p, end, len, bad);
 322         info->snapblob_len = len;
 323         info->snapblob = p;
 324         p += len;
 325
 326         if (p != end)
 327                 goto bad;
 328         return 0;
 329
 330 bad:
 331         err = -EIO;
 332 out_bad:
 333         pr_err("mds parse_reply err %d\n", err);
 334         return err;
 335 }
 336
 337 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
 338 {
 339         if (!info->dir_in)
 340                 return;
 341         free_pages((unsigned long)info->dir_in, get_order(info->dir_buf_size));
 342 }
 343
 344
 345 /*
 346  * sessions
 347  */
 348 const char *ceph_session_state_name(int s)
 349 {
 350         switch (s) {
 351         case CEPH_MDS_SESSION_NEW: return "new";
 352         case CEPH_MDS_SESSION_OPENING: return "opening";
 353         case CEPH_MDS_SESSION_OPEN: return "open";
 354         case CEPH_MDS_SESSION_HUNG: return "hung";
 355         case CEPH_MDS_SESSION_CLOSING: return "closing";
 356         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
 357         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
 358         default: return "???";
 359         }
 360 }
 361
 362 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
 363 {
 364         if (atomic_inc_not_zero(&s->s_ref)) {
 365                 dout("mdsc get_session %p %d -> %d\n", s,
 366                      atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
 367                 return s;
 368         } else {
 369                 dout("mdsc get_session %p 0 -- FAIL", s);
 370                 return NULL;
 371         }
 372 }
 373
 374 void ceph_put_mds_session(struct ceph_mds_session *s)
 375 {
 376         dout("mdsc put_session %p %d -> %d\n", s,
 377              atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
 378         if (atomic_dec_and_test(&s->s_ref)) {
 379                 if (s->s_auth.authorizer)
 380                         ceph_auth_destroy_authorizer(
 381                                 s->s_mdsc->fsc->client->monc.auth,
 382                                 s->s_auth.authorizer);
 383                 kfree(s);
 384         }
 385 }
 386
 387 /*
 388  * called under mdsc->mutex
 389  */
 390 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
 391                                                    int mds)
 392 {
 393         struct ceph_mds_session *session;
 394
 395         if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
 396                 return NULL;
 397         session = mdsc->sessions[mds];
 398         dout("lookup_mds_session %p %d\n", session,
 399              atomic_read(&session->s_ref));
 400         get_session(session);
 401         return session;
 402 }
 403
 404 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
 405 {
 406         if (mds >= mdsc->max_sessions)
 407                 return false;
 408         return mdsc->sessions[mds];
 409 }
 410
 411 static int __verify_registered_session(struct ceph_mds_client *mdsc,
 412                                        struct ceph_mds_session *s)
 413 {
 414         if (s->s_mds >= mdsc->max_sessions ||
 415             mdsc->sessions[s->s_mds] != s)
 416                 return -ENOENT;
 417         return 0;
 418 }
 419
 420 /*
 421  * create+register a new session for given mds.
 422  * called under mdsc->mutex.
 423  */
 424 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 425                                                  int mds)
 426 {
 427         struct ceph_mds_session *s;
 428
 429         if (mds >= mdsc->mdsmap->m_max_mds)
 430                 return ERR_PTR(-EINVAL);
 431
 432         s = kzalloc(sizeof(*s), GFP_NOFS);
 433         if (!s)
 434                 return ERR_PTR(-ENOMEM);
 435         s->s_mdsc = mdsc;
 436         s->s_mds = mds;
 437         s->s_state = CEPH_MDS_SESSION_NEW;
 438         s->s_ttl = 0;
 439         s->s_seq = 0;
 440         mutex_init(&s->s_mutex);
 441
 442         ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
 443
 444         spin_lock_init(&s->s_gen_ttl_lock);
 445         s->s_cap_gen = 0;
 446         s->s_cap_ttl = jiffies - 1;
 447
 448         spin_lock_init(&s->s_cap_lock);
 449         s->s_renew_requested = 0;
 450         s->s_renew_seq = 0;
 451         INIT_LIST_HEAD(&s->s_caps);
 452         s->s_nr_caps = 0;
 453         s->s_trim_caps = 0;
 454         atomic_set(&s->s_ref, 1);
 455         INIT_LIST_HEAD(&s->s_waiting);
 456         INIT_LIST_HEAD(&s->s_unsafe);
 457         s->s_num_cap_releases = 0;
 458         s->s_cap_reconnect = 0;
 459         s->s_cap_iterator = NULL;
 460         INIT_LIST_HEAD(&s->s_cap_releases);
 461         INIT_LIST_HEAD(&s->s_cap_releases_done);
 462         INIT_LIST_HEAD(&s->s_cap_flushing);
 463         INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
 464
 465         dout("register_session mds%d\n", mds);
 466         if (mds >= mdsc->max_sessions) {
 467                 int newmax = 1 << get_count_order(mds+1);
 468                 struct ceph_mds_session **sa;
 469
 470                 dout("register_session realloc to %d\n", newmax);
 471                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 472                 if (sa == NULL)
 473                         goto fail_realloc;
 474                 if (mdsc->sessions) {
 475                         memcpy(sa, mdsc->sessions,
 476                                mdsc->max_sessions * sizeof(void *));
 477                         kfree(mdsc->sessions);
 478                 }
 479                 mdsc->sessions = sa;
 480                 mdsc->max_sessions = newmax;
 481         }
 482         mdsc->sessions[mds] = s;
 483         atomic_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
 484
 485         ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
 486                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
 487
 488         return s;
 489
 490 fail_realloc:
 491         kfree(s);
 492         return ERR_PTR(-ENOMEM);
 493 }
 494
 495 /*
 496  * called under mdsc->mutex
 497  */
 498 static void __unregister_session(struct ceph_mds_client *mdsc,
 499                                struct ceph_mds_session *s)
 500 {
 501         dout("__unregister_session mds%d %p\n", s->s_mds, s);
 502         BUG_ON(mdsc->sessions[s->s_mds] != s);
 503         mdsc->sessions[s->s_mds] = NULL;
 504         ceph_con_close(&s->s_con);
 505         ceph_put_mds_session(s);
 506 }
 507
 508 /*
 509  * drop session refs in request.
 510  *
 511  * should be last request ref, or hold mdsc->mutex
 512  */
 513 static void put_request_session(struct ceph_mds_request *req)
 514 {
 515         if (req->r_session) {
 516                 ceph_put_mds_session(req->r_session);
 517                 req->r_session = NULL;
 518         }
 519 }
 520
 521 void ceph_mdsc_release_request(struct kref *kref)
 522 {
 523         struct ceph_mds_request *req = container_of(kref,
 524                                                     struct ceph_mds_request,
 525                                                     r_kref);
 526         destroy_reply_info(&req->r_reply_info);
 527         if (req->r_request)
 528                 ceph_msg_put(req->r_request);
 529         if (req->r_reply)
 530                 ceph_msg_put(req->r_reply);
 531         if (req->r_inode) {
 532                 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
 533                 iput(req->r_inode);
 534         }
 535         if (req->r_locked_dir)
 536                 ceph_put_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
 537         iput(req->r_target_inode);
 538         if (req->r_dentry)
 539                 dput(req->r_dentry);
 540         if (req->r_old_dentry)
 541                 dput(req->r_old_dentry);
 542         if (req->r_old_dentry_dir) {
 543                 /*
 544                  * track (and drop pins for) r_old_dentry_dir
 545                  * separately, since r_old_dentry's d_parent may have
 546                  * changed between the dir mutex being dropped and
 547                  * this request being freed.
 548                  */
 549                 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
 550                                   CEPH_CAP_PIN);
 551                 iput(req->r_old_dentry_dir);
 552         }
 553         kfree(req->r_path1);
 554         kfree(req->r_path2);
 555         if (req->r_pagelist)
 556                 ceph_pagelist_release(req->r_pagelist);
 557         put_request_session(req);
 558         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
 559         kfree(req);
 560 }
 561
 562 /*
 563  * lookup session, bump ref if found.
 564  *
 565  * called under mdsc->mutex.
 566  */
 567 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
 568                                              u64 tid)
 569 {
 570         struct ceph_mds_request *req;
 571         struct rb_node *n = mdsc->request_tree.rb_node;
 572
 573         while (n) {
 574                 req = rb_entry(n, struct ceph_mds_request, r_node);
 575                 if (tid < req->r_tid)
 576                         n = n->rb_left;
 577                 else if (tid > req->r_tid)
 578                         n = n->rb_right;
 579                 else {
 580                         ceph_mdsc_get_request(req);
 581                         return req;
 582                 }
 583         }
 584         return NULL;
 585 }
 586
 587 static void __insert_request(struct ceph_mds_client *mdsc,
 588                              struct ceph_mds_request *new)
 589 {
 590         struct rb_node **p = &mdsc->request_tree.rb_node;
 591         struct rb_node *parent = NULL;
 592         struct ceph_mds_request *req = NULL;
 593
 594         while (*p) {
 595                 parent = *p;
 596                 req = rb_entry(parent, struct ceph_mds_request, r_node);
 597                 if (new->r_tid < req->r_tid)
 598                         p = &(*p)->rb_left;
 599                 else if (new->r_tid > req->r_tid)
 600                         p = &(*p)->rb_right;
 601                 else
 602                         BUG();
 603         }
 604
 605         rb_link_node(&new->r_node, parent, p);
 606         rb_insert_color(&new->r_node, &mdsc->request_tree);
 607 }
 608
 609 /*
 610  * Register an in-flight request, and assign a tid.  Link to directory
 611  * are modifying (if any).
 612  *
 613  * Called under mdsc->mutex.
 614  */
 615 static void __register_request(struct ceph_mds_client *mdsc,
 616                                struct ceph_mds_request *req,
 617                                struct inode *dir)
 618 {
 619         req->r_tid = ++mdsc->last_tid;
 620         if (req->r_num_caps)
 621                 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
 622                                   req->r_num_caps);
 623         dout("__register_request %p tid %lld\n", req, req->r_tid);
 624         ceph_mdsc_get_request(req);
 625         __insert_request(mdsc, req);
 626
 627         req->r_uid = current_fsuid();
 628         req->r_gid = current_fsgid();
 629
 630         if (dir) {
 631                 struct ceph_inode_info *ci = ceph_inode(dir);
 632
 633                 ihold(dir);
 634                 spin_lock(&ci->i_unsafe_lock);
 635                 req->r_unsafe_dir = dir;
 636                 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
 637                 spin_unlock(&ci->i_unsafe_lock);
 638         }
 639 }
 640
 641 static void __unregister_request(struct ceph_mds_client *mdsc,
 642                                  struct ceph_mds_request *req)
 643 {
 644         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 645         rb_erase(&req->r_node, &mdsc->request_tree);
 646         RB_CLEAR_NODE(&req->r_node);
 647
 648         if (req->r_unsafe_dir) {
 649                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 650
 651                 spin_lock(&ci->i_unsafe_lock);
 652                 list_del_init(&req->r_unsafe_dir_item);
 653                 spin_unlock(&ci->i_unsafe_lock);
 654
 655                 iput(req->r_unsafe_dir);
 656                 req->r_unsafe_dir = NULL;
 657         }
 658
 659         complete_all(&req->r_safe_completion);
 660
 661         ceph_mdsc_put_request(req);
 662 }
 663
 664 /*
 665  * Choose mds to send request to next.  If there is a hint set in the
 666  * request (e.g., due to a prior forward hint from the mds), use that.
 667  * Otherwise, consult frag tree and/or caps to identify the
 668  * appropriate mds.  If all else fails, choose randomly.
 669  *
 670  * Called under mdsc->mutex.
 671  */
 672 static struct dentry *get_nonsnap_parent(struct dentry *dentry)
 673 {
 674         /*
 675          * we don't need to worry about protecting the d_parent access
 676          * here because we never renaming inside the snapped namespace
 677          * except to resplice to another snapdir, and either the old or new
 678          * result is a valid result.
 679          */
 680         while (!IS_ROOT(dentry) && ceph_snap(dentry->d_inode) != CEPH_NOSNAP)
 681                 dentry = dentry->d_parent;
 682         return dentry;
 683 }
 684
 685 static int __choose_mds(struct ceph_mds_client *mdsc,
 686                         struct ceph_mds_request *req)
 687 {
 688         struct inode *inode;
 689         struct ceph_inode_info *ci;
 690         struct ceph_cap *cap;
 691         int mode = req->r_direct_mode;
 692         int mds = -1;
 693         u32 hash = req->r_direct_hash;
 694         bool is_hash = req->r_direct_is_hash;
 695
 696         /*
 697          * is there a specific mds we should try?  ignore hint if we have
 698          * no session and the mds is not up (active or recovering).
 699          */
 700         if (req->r_resend_mds >= 0 &&
 701             (__have_session(mdsc, req->r_resend_mds) ||
 702              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 703                 dout("choose_mds using resend_mds mds%d\n",
 704                      req->r_resend_mds);
 705                 return req->r_resend_mds;
 706         }
 707
 708         if (mode == USE_RANDOM_MDS)
 709                 goto random;
 710
 711         inode = NULL;
 712         if (req->r_inode) {
 713                 inode = req->r_inode;
 714         } else if (req->r_dentry) {
 715                 /* ignore race with rename; old or new d_parent is okay */
 716                 struct dentry *parent = req->r_dentry->d_parent;
 717                 struct inode *dir = parent->d_inode;
 718
 719                 if (dir->i_sb != mdsc->fsc->sb) {
 720                         /* not this fs! */
 721                         inode = req->r_dentry->d_inode;
 722                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
 723                         /* direct snapped/virtual snapdir requests
 724                          * based on parent dir inode */
 725                         struct dentry *dn = get_nonsnap_parent(parent);
 726                         inode = dn->d_inode;
 727                         dout("__choose_mds using nonsnap parent %p\n", inode);
 728                 } else {
 729                         /* dentry target */
 730                         inode = req->r_dentry->d_inode;
 731                         if (!inode || mode == USE_AUTH_MDS) {
 732                                 /* dir + name */
 733                                 inode = dir;
 734                                 hash = ceph_dentry_hash(dir, req->r_dentry);
 735                                 is_hash = true;
 736                         }
 737                 }
 738         }
 739
 740         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 741              (int)hash, mode);
 742         if (!inode)
 743                 goto random;
 744         ci = ceph_inode(inode);
 745
 746         if (is_hash && S_ISDIR(inode->i_mode)) {
 747                 struct ceph_inode_frag frag;
 748                 int found;
 749
 750                 ceph_choose_frag(ci, hash, &frag, &found);
 751                 if (found) {
 752                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 753                                 u8 r;
 754
 755                                 /* choose a random replica */
 756                                 get_random_bytes(&r, 1);
 757                                 r %= frag.ndist;
 758                                 mds = frag.dist[r];
 759                                 dout("choose_mds %p %llx.%llx "
 760                                      "frag %u mds%d (%d/%d)\n",
 761                                      inode, ceph_vinop(inode),
 762                                      frag.frag, mds,
 763                                      (int)r, frag.ndist);
 764                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 765                                     CEPH_MDS_STATE_ACTIVE)
 766                                         return mds;
 767                         }
 768
 769                         /* since this file/dir wasn't known to be
 770                          * replicated, then we want to look for the
 771                          * authoritative mds. */
 772                         mode = USE_AUTH_MDS;
 773                         if (frag.mds >= 0) {
 774                                 /* choose auth mds */
 775                                 mds = frag.mds;
 776                                 dout("choose_mds %p %llx.%llx "
 777                                      "frag %u mds%d (auth)\n",
 778                                      inode, ceph_vinop(inode), frag.frag, mds);
 779                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 780                                     CEPH_MDS_STATE_ACTIVE)
 781                                         return mds;
 782                         }
 783                 }
 784         }
 785
 786         spin_lock(&ci->i_ceph_lock);
 787         cap = NULL;
 788         if (mode == USE_AUTH_MDS)
 789                 cap = ci->i_auth_cap;
 790         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
 791                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
 792         if (!cap) {
 793                 spin_unlock(&ci->i_ceph_lock);
 794                 goto random;
 795         }
 796         mds = cap->session->s_mds;
 797         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
 798              inode, ceph_vinop(inode), mds,
 799              cap == ci->i_auth_cap ? "auth " : "", cap);
 800         spin_unlock(&ci->i_ceph_lock);
 801         return mds;
 802
 803 random:
 804         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
 805         dout("choose_mds chose random mds%d\n", mds);
 806         return mds;
 807 }
 808
 809
 810 /*
 811  * session messages
 812  */
 813 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
 814 {
 815         struct ceph_msg *msg;
 816         struct ceph_mds_session_head *h;
 817
 818         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
 819                            false);
 820         if (!msg) {
 821                 pr_err("create_session_msg ENOMEM creating msg\n");
 822                 return NULL;
 823         }
 824         h = msg->front.iov_base;
 825         h->op = cpu_to_le32(op);
 826         h->seq = cpu_to_le64(seq);
 827
 828         return msg;
 829 }
 830
 831 /*
 832  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
 833  * to include additional client metadata fields.
 834  */
 835 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
 836 {
 837         struct ceph_msg *msg;
 838         struct ceph_mds_session_head *h;
 839         int i = -1;
 840         int metadata_bytes = 0;
 841         int metadata_key_count = 0;
 842         struct ceph_options *opt = mdsc->fsc->client->options;
 843         void *p;
 844
 845         const char* metadata[3][2] = {
 846                 {"hostname", utsname()->nodename},
 847                 {"entity_id", opt->name ? opt->name : ""},
 848                 {NULL, NULL}
 849         };
 850
 851         /* Calculate serialized length of metadata */
 852         metadata_bytes = 4;  /* map length */
 853         for (i = 0; metadata[i][0] != NULL; ++i) {
 854                 metadata_bytes += 8 + strlen(metadata[i][0]) +
 855                         strlen(metadata[i][1]);
 856                 metadata_key_count++;
 857         }
 858
 859         /* Allocate the message */
 860         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + metadata_bytes,
 861                            GFP_NOFS, false);
 862         if (!msg) {
 863                 pr_err("create_session_msg ENOMEM creating msg\n");
 864                 return NULL;
 865         }
 866         h = msg->front.iov_base;
 867         h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
 868         h->seq = cpu_to_le64(seq);
 869
 870         /*
 871          * Serialize client metadata into waiting buffer space, using
 872          * the format that userspace expects for map<string, string>
 873          *
 874          * ClientSession messages with metadata are v2
 875          */
 876         msg->hdr.version = cpu_to_le16(2);
 877         msg->hdr.compat_version = cpu_to_le16(1);
 878
 879         /* The write pointer, following the session_head structure */
 880         p = msg->front.iov_base + sizeof(*h);
 881
 882         /* Number of entries in the map */
 883         ceph_encode_32(&p, metadata_key_count);
 884
 885         /* Two length-prefixed strings for each entry in the map */
 886         for (i = 0; metadata[i][0] != NULL; ++i) {
 887                 size_t const key_len = strlen(metadata[i][0]);
 888                 size_t const val_len = strlen(metadata[i][1]);
 889
 890                 ceph_encode_32(&p, key_len);
 891                 memcpy(p, metadata[i][0], key_len);
 892                 p += key_len;
 893                 ceph_encode_32(&p, val_len);
 894                 memcpy(p, metadata[i][1], val_len);
 895                 p += val_len;
 896         }
 897
 898         return msg;
 899 }
 900
 901 /*
 902  * send session open request.
 903  *
 904  * called under mdsc->mutex
 905  */
 906 static int __open_session(struct ceph_mds_client *mdsc,
 907                           struct ceph_mds_session *session)
 908 {
 909         struct ceph_msg *msg;
 910         int mstate;
 911         int mds = session->s_mds;
 912
 913         /* wait for mds to go active? */
 914         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
 915         dout("open_session to mds%d (%s)\n", mds,
 916              ceph_mds_state_name(mstate));
 917         session->s_state = CEPH_MDS_SESSION_OPENING;
 918         session->s_renew_requested = jiffies;
 919
 920         /* send connect message */
 921         msg = create_session_open_msg(mdsc, session->s_seq);
 922         if (!msg)
 923                 return -ENOMEM;
 924         ceph_con_send(&session->s_con, msg);
 925         return 0;
 926 }
 927
 928 /*
 929  * open sessions for any export targets for the given mds
 930  *
 931  * called under mdsc->mutex
 932  */
 933 static struct ceph_mds_session *
 934 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
 935 {
 936         struct ceph_mds_session *session;
 937
 938         session = __ceph_lookup_mds_session(mdsc, target);
 939         if (!session) {
 940                 session = register_session(mdsc, target);
 941                 if (IS_ERR(session))
 942                         return session;
 943         }
 944         if (session->s_state == CEPH_MDS_SESSION_NEW ||
 945             session->s_state == CEPH_MDS_SESSION_CLOSING)
 946                 __open_session(mdsc, session);
 947
 948         return session;
 949 }
 950
 951 struct ceph_mds_session *
 952 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
 953 {
 954         struct ceph_mds_session *session;
 955
 956         dout("open_export_target_session to mds%d\n", target);
 957
 958         mutex_lock(&mdsc->mutex);
 959         session = __open_export_target_session(mdsc, target);
 960         mutex_unlock(&mdsc->mutex);
 961
 962         return session;
 963 }
 964
 965 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
 966                                           struct ceph_mds_session *session)
 967 {
 968         struct ceph_mds_info *mi;
 969         struct ceph_mds_session *ts;
 970         int i, mds = session->s_mds;
 971
 972         if (mds >= mdsc->mdsmap->m_max_mds)
 973                 return;
 974
 975         mi = &mdsc->mdsmap->m_info[mds];
 976         dout("open_export_target_sessions for mds%d (%d targets)\n",
 977              session->s_mds, mi->num_export_targets);
 978
 979         for (i = 0; i < mi->num_export_targets; i++) {
 980                 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
 981                 if (!IS_ERR(ts))
 982                         ceph_put_mds_session(ts);
 983         }
 984 }
 985
 986 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
 987                                            struct ceph_mds_session *session)
 988 {
 989         mutex_lock(&mdsc->mutex);
 990         __open_export_target_sessions(mdsc, session);
 991         mutex_unlock(&mdsc->mutex);
 992 }
 993
 994 /*
 995  * session caps
 996  */
 997
 998 /*
 999  * Free preallocated cap messages assigned to this session
1000  */
1001 static void cleanup_cap_releases(struct ceph_mds_session *session)
1002 {
1003         struct ceph_msg *msg;
1004
1005         spin_lock(&session->s_cap_lock);
1006         while (!list_empty(&session->s_cap_releases)) {
1007                 msg = list_first_entry(&session->s_cap_releases,
1008                                        struct ceph_msg, list_head);
1009                 list_del_init(&msg->list_head);
1010                 ceph_msg_put(msg);
1011         }
1012         while (!list_empty(&session->s_cap_releases_done)) {
1013                 msg = list_first_entry(&session->s_cap_releases_done,
1014                                        struct ceph_msg, list_head);
1015                 list_del_init(&msg->list_head);
1016                 ceph_msg_put(msg);
1017         }
1018         spin_unlock(&session->s_cap_lock);
1019 }
1020
1021 /*
1022  * Helper to safely iterate over all caps associated with a session, with
1023  * special care taken to handle a racing __ceph_remove_cap().
1024  *
1025  * Caller must hold session s_mutex.
1026  */
1027 static int iterate_session_caps(struct ceph_mds_session *session,
1028                                  int (*cb)(struct inode *, struct ceph_cap *,
1029                                             void *), void *arg)
1030 {
1031         struct list_head *p;
1032         struct ceph_cap *cap;
1033         struct inode *inode, *last_inode = NULL;
1034         struct ceph_cap *old_cap = NULL;
1035         int ret;
1036
1037         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1038         spin_lock(&session->s_cap_lock);
1039         p = session->s_caps.next;
1040         while (p != &session->s_caps) {
1041                 cap = list_entry(p, struct ceph_cap, session_caps);
1042                 inode = igrab(&cap->ci->vfs_inode);
1043                 if (!inode) {
1044                         p = p->next;
1045                         continue;
1046                 }
1047                 session->s_cap_iterator = cap;
1048                 spin_unlock(&session->s_cap_lock);
1049
1050                 if (last_inode) {
1051                         iput(last_inode);
1052                         last_inode = NULL;
1053                 }
1054                 if (old_cap) {
1055                         ceph_put_cap(session->s_mdsc, old_cap);
1056                         old_cap = NULL;
1057                 }
1058
1059                 ret = cb(inode, cap, arg);
1060                 last_inode = inode;
1061
1062                 spin_lock(&session->s_cap_lock);
1063                 p = p->next;
1064                 if (cap->ci == NULL) {
1065                         dout("iterate_session_caps  finishing cap %p removal\n",
1066                              cap);
1067                         BUG_ON(cap->session != session);
1068                         list_del_init(&cap->session_caps);
1069                         session->s_nr_caps--;
1070                         cap->session = NULL;
1071                         old_cap = cap;  /* put_cap it w/o locks held */
1072                 }
1073                 if (ret < 0)
1074                         goto out;
1075         }
1076         ret = 0;
1077 out:
1078         session->s_cap_iterator = NULL;
1079         spin_unlock(&session->s_cap_lock);
1080
1081         iput(last_inode);
1082         if (old_cap)
1083                 ceph_put_cap(session->s_mdsc, old_cap);
1084
1085         return ret;
1086 }
1087
1088 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1089                                   void *arg)
1090 {
1091         struct ceph_inode_info *ci = ceph_inode(inode);
1092         int drop = 0;
1093
1094         dout("removing cap %p, ci is %p, inode is %p\n",
1095              cap, ci, &ci->vfs_inode);
1096         spin_lock(&ci->i_ceph_lock);
1097         __ceph_remove_cap(cap, false);
1098         if (!__ceph_is_any_real_caps(ci)) {
1099                 struct ceph_mds_client *mdsc =
1100                         ceph_sb_to_client(inode->i_sb)->mdsc;
1101
1102                 spin_lock(&mdsc->cap_dirty_lock);
1103                 if (!list_empty(&ci->i_dirty_item)) {
1104                         pr_info(" dropping dirty %s state for %p %lld\n",
1105                                 ceph_cap_string(ci->i_dirty_caps),
1106                                 inode, ceph_ino(inode));
1107                         ci->i_dirty_caps = 0;
1108                         list_del_init(&ci->i_dirty_item);
1109                         drop = 1;
1110                 }
1111                 if (!list_empty(&ci->i_flushing_item)) {
1112                         pr_info(" dropping dirty+flushing %s state for %p %lld\n",
1113                                 ceph_cap_string(ci->i_flushing_caps),
1114                                 inode, ceph_ino(inode));
1115                         ci->i_flushing_caps = 0;
1116                         list_del_init(&ci->i_flushing_item);
1117                         mdsc->num_cap_flushing--;
1118                         drop = 1;
1119                 }
1120                 if (drop && ci->i_wrbuffer_ref) {
1121                         pr_info(" dropping dirty data for %p %lld\n",
1122                                 inode, ceph_ino(inode));
1123                         ci->i_wrbuffer_ref = 0;
1124                         ci->i_wrbuffer_ref_head = 0;
1125                         drop++;
1126                 }
1127                 spin_unlock(&mdsc->cap_dirty_lock);
1128         }
1129         spin_unlock(&ci->i_ceph_lock);
1130         while (drop--)
1131                 iput(inode);
1132         return 0;
1133 }
1134
1135 /*
1136  * caller must hold session s_mutex
1137  */
1138 static void remove_session_caps(struct ceph_mds_session *session)
1139 {
1140         dout("remove_session_caps on %p\n", session);
1141         iterate_session_caps(session, remove_session_caps_cb, NULL);
1142
1143         spin_lock(&session->s_cap_lock);
1144         if (session->s_nr_caps > 0) {
1145                 struct super_block *sb = session->s_mdsc->fsc->sb;
1146                 struct inode *inode;
1147                 struct ceph_cap *cap, *prev = NULL;
1148                 struct ceph_vino vino;
1149                 /*
1150                  * iterate_session_caps() skips inodes that are being
1151                  * deleted, we need to wait until deletions are complete.
1152                  * __wait_on_freeing_inode() is designed for the job,
1153                  * but it is not exported, so use lookup inode function
1154                  * to access it.
1155                  */
1156                 while (!list_empty(&session->s_caps)) {
1157                         cap = list_entry(session->s_caps.next,
1158                                          struct ceph_cap, session_caps);
1159                         if (cap == prev)
1160                                 break;
1161                         prev = cap;
1162                         vino = cap->ci->i_vino;
1163                         spin_unlock(&session->s_cap_lock);
1164
1165                         inode = ceph_find_inode(sb, vino);
1166                         iput(inode);
1167
1168                         spin_lock(&session->s_cap_lock);
1169                 }
1170         }
1171         spin_unlock(&session->s_cap_lock);
1172
1173         BUG_ON(session->s_nr_caps > 0);
1174         BUG_ON(!list_empty(&session->s_cap_flushing));
1175         cleanup_cap_releases(session);
1176 }
1177
1178 /*
1179  * wake up any threads waiting on this session's caps.  if the cap is
1180  * old (didn't get renewed on the client reconnect), remove it now.
1181  *
1182  * caller must hold s_mutex.
1183  */
1184 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1185                               void *arg)
1186 {
1187         struct ceph_inode_info *ci = ceph_inode(inode);
1188
1189         wake_up_all(&ci->i_cap_wq);
1190         if (arg) {
1191                 spin_lock(&ci->i_ceph_lock);
1192                 ci->i_wanted_max_size = 0;
1193                 ci->i_requested_max_size = 0;
1194                 spin_unlock(&ci->i_ceph_lock);
1195         }
1196         return 0;
1197 }
1198
1199 static void wake_up_session_caps(struct ceph_mds_session *session,
1200                                  int reconnect)
1201 {
1202         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1203         iterate_session_caps(session, wake_up_session_cb,
1204                              (void *)(unsigned long)reconnect);
1205 }
1206
1207 /*
1208  * Send periodic message to MDS renewing all currently held caps.  The
1209  * ack will reset the expiration for all caps from this session.
1210  *
1211  * caller holds s_mutex
1212  */
1213 static int send_renew_caps(struct ceph_mds_client *mdsc,
1214                            struct ceph_mds_session *session)
1215 {
1216         struct ceph_msg *msg;
1217         int state;
1218
1219         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1220             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1221                 pr_info("mds%d caps stale\n", session->s_mds);
1222         session->s_renew_requested = jiffies;
1223
1224         /* do not try to renew caps until a recovering mds has reconnected
1225          * with its clients. */
1226         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1227         if (state < CEPH_MDS_STATE_RECONNECT) {
1228                 dout("send_renew_caps ignoring mds%d (%s)\n",
1229                      session->s_mds, ceph_mds_state_name(state));
1230                 return 0;
1231         }
1232
1233         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1234                 ceph_mds_state_name(state));
1235         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1236                                  ++session->s_renew_seq);
1237         if (!msg)
1238                 return -ENOMEM;
1239         ceph_con_send(&session->s_con, msg);
1240         return 0;
1241 }
1242
1243 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1244                              struct ceph_mds_session *session, u64 seq)
1245 {
1246         struct ceph_msg *msg;
1247
1248         dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1249              session->s_mds, ceph_session_state_name(session->s_state), seq);
1250         msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1251         if (!msg)
1252                 return -ENOMEM;
1253         ceph_con_send(&session->s_con, msg);
1254         return 0;
1255 }
1256
1257
1258 /*
1259  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1260  *
1261  * Called under session->s_mutex
1262  */
1263 static void renewed_caps(struct ceph_mds_client *mdsc,
1264                          struct ceph_mds_session *session, int is_renew)
1265 {
1266         int was_stale;
1267         int wake = 0;
1268
1269         spin_lock(&session->s_cap_lock);
1270         was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1271
1272         session->s_cap_ttl = session->s_renew_requested +
1273                 mdsc->mdsmap->m_session_timeout*HZ;
1274
1275         if (was_stale) {
1276                 if (time_before(jiffies, session->s_cap_ttl)) {
1277                         pr_info("mds%d caps renewed\n", session->s_mds);
1278                         wake = 1;
1279                 } else {
1280                         pr_info("mds%d caps still stale\n", session->s_mds);
1281                 }
1282         }
1283         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1284              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1285              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1286         spin_unlock(&session->s_cap_lock);
1287
1288         if (wake)
1289                 wake_up_session_caps(session, 0);
1290 }
1291
1292 /*
1293  * send a session close request
1294  */
1295 static int request_close_session(struct ceph_mds_client *mdsc,
1296                                  struct ceph_mds_session *session)
1297 {
1298         struct ceph_msg *msg;
1299
1300         dout("request_close_session mds%d state %s seq %lld\n",
1301              session->s_mds, ceph_session_state_name(session->s_state),
1302              session->s_seq);
1303         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1304         if (!msg)
1305                 return -ENOMEM;
1306         ceph_con_send(&session->s_con, msg);
1307         return 0;
1308 }
1309
1310 /*
1311  * Called with s_mutex held.
1312  */
1313 static int __close_session(struct ceph_mds_client *mdsc,
1314                          struct ceph_mds_session *session)
1315 {
1316         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1317                 return 0;
1318         session->s_state = CEPH_MDS_SESSION_CLOSING;
1319         return request_close_session(mdsc, session);
1320 }
1321
1322 /*
1323  * Trim old(er) caps.
1324  *
1325  * Because we can't cache an inode without one or more caps, we do
1326  * this indirectly: if a cap is unused, we prune its aliases, at which
1327  * point the inode will hopefully get dropped to.
1328  *
1329  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1330  * memory pressure from the MDS, though, so it needn't be perfect.
1331  */
1332 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1333 {
1334         struct ceph_mds_session *session = arg;
1335         struct ceph_inode_info *ci = ceph_inode(inode);
1336         int used, wanted, oissued, mine;
1337
1338         if (session->s_trim_caps <= 0)
1339                 return -1;
1340
1341         spin_lock(&ci->i_ceph_lock);
1342         mine = cap->issued | cap->implemented;
1343         used = __ceph_caps_used(ci);
1344         wanted = __ceph_caps_file_wanted(ci);
1345         oissued = __ceph_caps_issued_other(ci, cap);
1346
1347         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1348              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1349              ceph_cap_string(used), ceph_cap_string(wanted));
1350         if (cap == ci->i_auth_cap) {
1351                 if (ci->i_dirty_caps | ci->i_flushing_caps)
1352                         goto out;
1353                 if ((used | wanted) & CEPH_CAP_ANY_WR)
1354                         goto out;
1355         }
1356         if ((used | wanted) & ~oissued & mine)
1357                 goto out;   /* we need these caps */
1358
1359         session->s_trim_caps--;
1360         if (oissued) {
1361                 /* we aren't the only cap.. just remove us */
1362                 __ceph_remove_cap(cap, true);
1363         } else {
1364                 /* try to drop referring dentries */
1365                 spin_unlock(&ci->i_ceph_lock);
1366                 d_prune_aliases(inode);
1367                 dout("trim_caps_cb %p cap %p  pruned, count now %d\n",
1368                      inode, cap, atomic_read(&inode->i_count));
1369                 return 0;
1370         }
1371
1372 out:
1373         spin_unlock(&ci->i_ceph_lock);
1374         return 0;
1375 }
1376
1377 /*
1378  * Trim session cap count down to some max number.
1379  */
1380 static int trim_caps(struct ceph_mds_client *mdsc,
1381                      struct ceph_mds_session *session,
1382                      int max_caps)
1383 {
1384         int trim_caps = session->s_nr_caps - max_caps;
1385
1386         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1387              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1388         if (trim_caps > 0) {
1389                 session->s_trim_caps = trim_caps;
1390                 iterate_session_caps(session, trim_caps_cb, session);
1391                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1392                      session->s_mds, session->s_nr_caps, max_caps,
1393                         trim_caps - session->s_trim_caps);
1394                 session->s_trim_caps = 0;
1395         }
1396
1397         ceph_add_cap_releases(mdsc, session);
1398         ceph_send_cap_releases(mdsc, session);
1399         return 0;
1400 }
1401
1402 /*
1403  * Allocate cap_release messages.  If there is a partially full message
1404  * in the queue, try to allocate enough to cover it's remainder, so that
1405  * we can send it immediately.
1406  *
1407  * Called under s_mutex.
1408  */
1409 int ceph_add_cap_releases(struct ceph_mds_client *mdsc,
1410                           struct ceph_mds_session *session)
1411 {
1412         struct ceph_msg *msg, *partial = NULL;
1413         struct ceph_mds_cap_release *head;
1414         int err = -ENOMEM;
1415         int extra = mdsc->fsc->mount_options->cap_release_safety;
1416         int num;
1417
1418         dout("add_cap_releases %p mds%d extra %d\n", session, session->s_mds,
1419              extra);
1420
1421         spin_lock(&session->s_cap_lock);
1422
1423         if (!list_empty(&session->s_cap_releases)) {
1424                 msg = list_first_entry(&session->s_cap_releases,
1425                                        struct ceph_msg,
1426                                  list_head);
1427                 head = msg->front.iov_base;
1428                 num = le32_to_cpu(head->num);
1429                 if (num) {
1430                         dout(" partial %p with (%d/%d)\n", msg, num,
1431                              (int)CEPH_CAPS_PER_RELEASE);
1432                         extra += CEPH_CAPS_PER_RELEASE - num;
1433                         partial = msg;
1434                 }
1435         }
1436         while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1437                 spin_unlock(&session->s_cap_lock);
1438                 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
1439                                    GFP_NOFS, false);
1440                 if (!msg)
1441                         goto out_unlocked;
1442                 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1443                      (int)msg->front.iov_len);
1444                 head = msg->front.iov_base;
1445                 head->num = cpu_to_le32(0);
1446                 msg->front.iov_len = sizeof(*head);
1447                 spin_lock(&session->s_cap_lock);
1448                 list_add(&msg->list_head, &session->s_cap_releases);
1449                 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1450         }
1451
1452         if (partial) {
1453                 head = partial->front.iov_base;
1454                 num = le32_to_cpu(head->num);
1455                 dout(" queueing partial %p with %d/%d\n", partial, num,
1456                      (int)CEPH_CAPS_PER_RELEASE);
1457                 list_move_tail(&partial->list_head,
1458                                &session->s_cap_releases_done);
1459                 session->s_num_cap_releases -= CEPH_CAPS_PER_RELEASE - num;
1460         }
1461         err = 0;
1462         spin_unlock(&session->s_cap_lock);
1463 out_unlocked:
1464         return err;
1465 }
1466
1467 /*
1468  * flush all dirty inode data to disk.
1469  *
1470  * returns true if we've flushed through want_flush_seq
1471  */
1472 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1473 {
1474         int mds, ret = 1;
1475
1476         dout("check_cap_flush want %lld\n", want_flush_seq);
1477         mutex_lock(&mdsc->mutex);
1478         for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1479                 struct ceph_mds_session *session = mdsc->sessions[mds];
1480
1481                 if (!session)
1482                         continue;
1483                 get_session(session);
1484                 mutex_unlock(&mdsc->mutex);
1485
1486                 mutex_lock(&session->s_mutex);
1487                 if (!list_empty(&session->s_cap_flushing)) {
1488                         struct ceph_inode_info *ci =
1489                                 list_entry(session->s_cap_flushing.next,
1490                                            struct ceph_inode_info,
1491                                            i_flushing_item);
1492                         struct inode *inode = &ci->vfs_inode;
1493
1494                         spin_lock(&ci->i_ceph_lock);
1495                         if (ci->i_cap_flush_seq <= want_flush_seq) {
1496                                 dout("check_cap_flush still flushing %p "
1497                                      "seq %lld <= %lld to mds%d\n", inode,
1498                                      ci->i_cap_flush_seq, want_flush_seq,
1499                                      session->s_mds);
1500                                 ret = 0;
1501                         }
1502                         spin_unlock(&ci->i_ceph_lock);
1503                 }
1504                 mutex_unlock(&session->s_mutex);
1505                 ceph_put_mds_session(session);
1506
1507                 if (!ret)
1508                         return ret;
1509                 mutex_lock(&mdsc->mutex);
1510         }
1511
1512         mutex_unlock(&mdsc->mutex);
1513         dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1514         return ret;
1515 }
1516
1517 /*
1518  * called under s_mutex
1519  */
1520 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1521                             struct ceph_mds_session *session)
1522 {
1523         struct ceph_msg *msg;
1524
1525         dout("send_cap_releases mds%d\n", session->s_mds);
1526         spin_lock(&session->s_cap_lock);
1527         while (!list_empty(&session->s_cap_releases_done)) {
1528                 msg = list_first_entry(&session->s_cap_releases_done,
1529                                  struct ceph_msg, list_head);
1530                 list_del_init(&msg->list_head);
1531                 spin_unlock(&session->s_cap_lock);
1532                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1533                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1534                 ceph_con_send(&session->s_con, msg);
1535                 spin_lock(&session->s_cap_lock);
1536         }
1537         spin_unlock(&session->s_cap_lock);
1538 }
1539
1540 static void discard_cap_releases(struct ceph_mds_client *mdsc,
1541                                  struct ceph_mds_session *session)
1542 {
1543         struct ceph_msg *msg;
1544         struct ceph_mds_cap_release *head;
1545         unsigned num;
1546
1547         dout("discard_cap_releases mds%d\n", session->s_mds);
1548
1549         if (!list_empty(&session->s_cap_releases)) {
1550                 /* zero out the in-progress message */
1551                 msg = list_first_entry(&session->s_cap_releases,
1552                                         struct ceph_msg, list_head);
1553                 head = msg->front.iov_base;
1554                 num = le32_to_cpu(head->num);
1555                 dout("discard_cap_releases mds%d %p %u\n",
1556                      session->s_mds, msg, num);
1557                 head->num = cpu_to_le32(0);
1558                 msg->front.iov_len = sizeof(*head);
1559                 session->s_num_cap_releases += num;
1560         }
1561
1562         /* requeue completed messages */
1563         while (!list_empty(&session->s_cap_releases_done)) {
1564                 msg = list_first_entry(&session->s_cap_releases_done,
1565                                  struct ceph_msg, list_head);
1566                 list_del_init(&msg->list_head);
1567
1568                 head = msg->front.iov_base;
1569                 num = le32_to_cpu(head->num);
1570                 dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg,
1571                      num);
1572                 session->s_num_cap_releases += num;
1573                 head->num = cpu_to_le32(0);
1574                 msg->front.iov_len = sizeof(*head);
1575                 list_add(&msg->list_head, &session->s_cap_releases);
1576         }
1577 }
1578
1579 /*
1580  * requests
1581  */
1582
1583 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1584                                     struct inode *dir)
1585 {
1586         struct ceph_inode_info *ci = ceph_inode(dir);
1587         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1588         struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1589         size_t size = sizeof(*rinfo->dir_in) + sizeof(*rinfo->dir_dname_len) +
1590                       sizeof(*rinfo->dir_dname) + sizeof(*rinfo->dir_dlease);
1591         int order, num_entries;
1592
1593         spin_lock(&ci->i_ceph_lock);
1594         num_entries = ci->i_files + ci->i_subdirs;
1595         spin_unlock(&ci->i_ceph_lock);
1596         num_entries = max(num_entries, 1);
1597         num_entries = min(num_entries, opt->max_readdir);
1598
1599         order = get_order(size * num_entries);
1600         while (order >= 0) {
1601                 rinfo->dir_in = (void*)__get_free_pages(GFP_NOFS | __GFP_NOWARN,
1602                                                         order);
1603                 if (rinfo->dir_in)
1604                         break;
1605                 order--;
1606         }
1607         if (!rinfo->dir_in)
1608                 return -ENOMEM;
1609
1610         num_entries = (PAGE_SIZE << order) / size;
1611         num_entries = min(num_entries, opt->max_readdir);
1612
1613         rinfo->dir_buf_size = PAGE_SIZE << order;
1614         req->r_num_caps = num_entries + 1;
1615         req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1616         req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1617         return 0;
1618 }
1619
1620 /*
1621  * Create an mds request.
1622  */
1623 struct ceph_mds_request *
1624 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1625 {
1626         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1627
1628         if (!req)
1629                 return ERR_PTR(-ENOMEM);
1630
1631         mutex_init(&req->r_fill_mutex);
1632         req->r_mdsc = mdsc;
1633         req->r_started = jiffies;
1634         req->r_resend_mds = -1;
1635         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1636         req->r_fmode = -1;
1637         kref_init(&req->r_kref);
1638         INIT_LIST_HEAD(&req->r_wait);
1639         init_completion(&req->r_completion);
1640         init_completion(&req->r_safe_completion);
1641         INIT_LIST_HEAD(&req->r_unsafe_item);
1642
1643         req->r_stamp = CURRENT_TIME;
1644
1645         req->r_op = op;
1646         req->r_direct_mode = mode;
1647         return req;
1648 }
1649
1650 /*
1651  * return oldest (lowest) request, tid in request tree, 0 if none.
1652  *
1653  * called under mdsc->mutex.
1654  */
1655 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1656 {
1657         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1658                 return NULL;
1659         return rb_entry(rb_first(&mdsc->request_tree),
1660                         struct ceph_mds_request, r_node);
1661 }
1662
1663 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1664 {
1665         struct ceph_mds_request *req = __get_oldest_req(mdsc);
1666
1667         if (req)
1668                 return req->r_tid;
1669         return 0;
1670 }
1671
1672 /*
1673  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1674  * on build_path_from_dentry in fs/cifs/dir.c.
1675  *
1676  * If @stop_on_nosnap, generate path relative to the first non-snapped
1677  * inode.
1678  *
1679  * Encode hidden .snap dirs as a double /, i.e.
1680  *   foo/.snap/bar -> foo//bar
1681  */
1682 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1683                            int stop_on_nosnap)
1684 {
1685         struct dentry *temp;
1686         char *path;
1687         int len, pos;
1688         unsigned seq;
1689
1690         if (dentry == NULL)
1691                 return ERR_PTR(-EINVAL);
1692
1693 retry:
1694         len = 0;
1695         seq = read_seqbegin(&rename_lock);
1696         rcu_read_lock();
1697         for (temp = dentry; !IS_ROOT(temp);) {
1698                 struct inode *inode = temp->d_inode;
1699                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1700                         len++;  /* slash only */
1701                 else if (stop_on_nosnap && inode &&
1702                          ceph_snap(inode) == CEPH_NOSNAP)
1703                         break;
1704                 else
1705                         len += 1 + temp->d_name.len;
1706                 temp = temp->d_parent;
1707         }
1708         rcu_read_unlock();
1709         if (len)
1710                 len--;  /* no leading '/' */
1711
1712         path = kmalloc(len+1, GFP_NOFS);
1713         if (path == NULL)
1714                 return ERR_PTR(-ENOMEM);
1715         pos = len;
1716         path[pos] = 0;  /* trailing null */
1717         rcu_read_lock();
1718         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1719                 struct inode *inode;
1720
1721                 spin_lock(&temp->d_lock);
1722                 inode = temp->d_inode;
1723                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1724                         dout("build_path path+%d: %p SNAPDIR\n",
1725                              pos, temp);
1726                 } else if (stop_on_nosnap && inode &&
1727                            ceph_snap(inode) == CEPH_NOSNAP) {
1728                         spin_unlock(&temp->d_lock);
1729                         break;
1730                 } else {
1731                         pos -= temp->d_name.len;
1732                         if (pos < 0) {
1733                                 spin_unlock(&temp->d_lock);
1734                                 break;
1735                         }
1736                         strncpy(path + pos, temp->d_name.name,
1737                                 temp->d_name.len);
1738                 }
1739                 spin_unlock(&temp->d_lock);
1740                 if (pos)
1741                         path[--pos] = '/';
1742                 temp = temp->d_parent;
1743         }
1744         rcu_read_unlock();
1745         if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1746                 pr_err("build_path did not end path lookup where "
1747                        "expected, namelen is %d, pos is %d\n", len, pos);
1748                 /* presumably this is only possible if racing with a
1749                    rename of one of the parent directories (we can not
1750                    lock the dentries above us to prevent this, but
1751                    retrying should be harmless) */
1752                 kfree(path);
1753                 goto retry;
1754         }
1755
1756         *base = ceph_ino(temp->d_inode);
1757         *plen = len;
1758         dout("build_path on %p %d built %llx '%.*s'\n",
1759              dentry, d_count(dentry), *base, len, path);
1760         return path;
1761 }
1762
1763 static int build_dentry_path(struct dentry *dentry,
1764                              const char **ppath, int *ppathlen, u64 *pino,
1765                              int *pfreepath)
1766 {
1767         char *path;
1768
1769         if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1770                 *pino = ceph_ino(dentry->d_parent->d_inode);
1771                 *ppath = dentry->d_name.name;
1772                 *ppathlen = dentry->d_name.len;
1773                 return 0;
1774         }
1775         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1776         if (IS_ERR(path))
1777                 return PTR_ERR(path);
1778         *ppath = path;
1779         *pfreepath = 1;
1780         return 0;
1781 }
1782
1783 static int build_inode_path(struct inode *inode,
1784                             const char **ppath, int *ppathlen, u64 *pino,
1785                             int *pfreepath)
1786 {
1787         struct dentry *dentry;
1788         char *path;
1789
1790         if (ceph_snap(inode) == CEPH_NOSNAP) {
1791                 *pino = ceph_ino(inode);
1792                 *ppathlen = 0;
1793                 return 0;
1794         }
1795         dentry = d_find_alias(inode);
1796         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1797         dput(dentry);
1798         if (IS_ERR(path))
1799                 return PTR_ERR(path);
1800         *ppath = path;
1801         *pfreepath = 1;
1802         return 0;
1803 }
1804
1805 /*
1806  * request arguments may be specified via an inode *, a dentry *, or
1807  * an explicit ino+path.
1808  */
1809 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1810                                   const char *rpath, u64 rino,
1811                                   const char **ppath, int *pathlen,
1812                                   u64 *ino, int *freepath)
1813 {
1814         int r = 0;
1815
1816         if (rinode) {
1817                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1818                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1819                      ceph_snap(rinode));
1820         } else if (rdentry) {
1821                 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1822                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1823                      *ppath);
1824         } else if (rpath || rino) {
1825                 *ino = rino;
1826                 *ppath = rpath;
1827                 *pathlen = rpath ? strlen(rpath) : 0;
1828                 dout(" path %.*s\n", *pathlen, rpath);
1829         }
1830
1831         return r;
1832 }
1833
1834 /*
1835  * called under mdsc->mutex
1836  */
1837 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1838                                                struct ceph_mds_request *req,
1839                                                int mds)
1840 {
1841         struct ceph_msg *msg;
1842         struct ceph_mds_request_head *head;
1843         const char *path1 = NULL;
1844         const char *path2 = NULL;
1845         u64 ino1 = 0, ino2 = 0;
1846         int pathlen1 = 0, pathlen2 = 0;
1847         int freepath1 = 0, freepath2 = 0;
1848         int len;
1849         u16 releases;
1850         void *p, *end;
1851         int ret;
1852
1853         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1854                               req->r_path1, req->r_ino1.ino,
1855                               &path1, &pathlen1, &ino1, &freepath1);
1856         if (ret < 0) {
1857                 msg = ERR_PTR(ret);
1858                 goto out;
1859         }
1860
1861         ret = set_request_path_attr(NULL, req->r_old_dentry,
1862                               req->r_path2, req->r_ino2.ino,
1863                               &path2, &pathlen2, &ino2, &freepath2);
1864         if (ret < 0) {
1865                 msg = ERR_PTR(ret);
1866                 goto out_free1;
1867         }
1868
1869         len = sizeof(*head) +
1870                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
1871                 sizeof(struct timespec);
1872
1873         /* calculate (max) length for cap releases */
1874         len += sizeof(struct ceph_mds_request_release) *
1875                 (!!req->r_inode_drop + !!req->r_dentry_drop +
1876                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1877         if (req->r_dentry_drop)
1878                 len += req->r_dentry->d_name.len;
1879         if (req->r_old_dentry_drop)
1880                 len += req->r_old_dentry->d_name.len;
1881
1882         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1883         if (!msg) {
1884                 msg = ERR_PTR(-ENOMEM);
1885                 goto out_free2;
1886         }
1887
1888         msg->hdr.version = cpu_to_le16(2);
1889         msg->hdr.tid = cpu_to_le64(req->r_tid);
1890
1891         head = msg->front.iov_base;
1892         p = msg->front.iov_base + sizeof(*head);
1893         end = msg->front.iov_base + msg->front.iov_len;
1894
1895         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1896         head->op = cpu_to_le32(req->r_op);
1897         head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1898         head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1899         head->args = req->r_args;
1900
1901         ceph_encode_filepath(&p, end, ino1, path1);
1902         ceph_encode_filepath(&p, end, ino2, path2);
1903
1904         /* make note of release offset, in case we need to replay */
1905         req->r_request_release_offset = p - msg->front.iov_base;
1906
1907         /* cap releases */
1908         releases = 0;
1909         if (req->r_inode_drop)
1910                 releases += ceph_encode_inode_release(&p,
1911                       req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1912                       mds, req->r_inode_drop, req->r_inode_unless, 0);
1913         if (req->r_dentry_drop)
1914                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1915                        mds, req->r_dentry_drop, req->r_dentry_unless);
1916         if (req->r_old_dentry_drop)
1917                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1918                        mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1919         if (req->r_old_inode_drop)
1920                 releases += ceph_encode_inode_release(&p,
1921                       req->r_old_dentry->d_inode,
1922                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1923         head->num_releases = cpu_to_le16(releases);
1924
1925         /* time stamp */
1926         ceph_encode_copy(&p, &req->r_stamp, sizeof(req->r_stamp));
1927
1928         BUG_ON(p > end);
1929         msg->front.iov_len = p - msg->front.iov_base;
1930         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1931
1932         if (req->r_pagelist) {
1933                 struct ceph_pagelist *pagelist = req->r_pagelist;
1934                 atomic_inc(&pagelist->refcnt);
1935                 ceph_msg_data_add_pagelist(msg, pagelist);
1936                 msg->hdr.data_len = cpu_to_le32(pagelist->length);
1937         } else {
1938                 msg->hdr.data_len = 0;
1939         }
1940
1941         msg->hdr.data_off = cpu_to_le16(0);
1942
1943 out_free2:
1944         if (freepath2)
1945                 kfree((char *)path2);
1946 out_free1:
1947         if (freepath1)
1948                 kfree((char *)path1);
1949 out:
1950         return msg;
1951 }
1952
1953 /*
1954  * called under mdsc->mutex if error, under no mutex if
1955  * success.
1956  */
1957 static void complete_request(struct ceph_mds_client *mdsc,
1958                              struct ceph_mds_request *req)
1959 {
1960         if (req->r_callback)
1961                 req->r_callback(mdsc, req);
1962         else
1963                 complete_all(&req->r_completion);
1964 }
1965
1966 /*
1967  * called under mdsc->mutex
1968  */
1969 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1970                                   struct ceph_mds_request *req,
1971                                   int mds)
1972 {
1973         struct ceph_mds_request_head *rhead;
1974         struct ceph_msg *msg;
1975         int flags = 0;
1976
1977         req->r_attempts++;
1978         if (req->r_inode) {
1979                 struct ceph_cap *cap =
1980                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
1981
1982                 if (cap)
1983                         req->r_sent_on_mseq = cap->mseq;
1984                 else
1985                         req->r_sent_on_mseq = -1;
1986         }
1987         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1988              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1989
1990         if (req->r_got_unsafe) {
1991                 void *p;
1992                 /*
1993                  * Replay.  Do not regenerate message (and rebuild
1994                  * paths, etc.); just use the original message.
1995                  * Rebuilding paths will break for renames because
1996                  * d_move mangles the src name.
1997                  */
1998                 msg = req->r_request;
1999                 rhead = msg->front.iov_base;
2000
2001                 flags = le32_to_cpu(rhead->flags);
2002                 flags |= CEPH_MDS_FLAG_REPLAY;
2003                 rhead->flags = cpu_to_le32(flags);
2004
2005                 if (req->r_target_inode)
2006                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2007
2008                 rhead->num_retry = req->r_attempts - 1;
2009
2010                 /* remove cap/dentry releases from message */
2011                 rhead->num_releases = 0;
2012
2013                 /* time stamp */
2014                 p = msg->front.iov_base + req->r_request_release_offset;
2015                 ceph_encode_copy(&p, &req->r_stamp, sizeof(req->r_stamp));
2016
2017                 msg->front.iov_len = p - msg->front.iov_base;
2018                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2019                 return 0;
2020         }
2021
2022         if (req->r_request) {
2023                 ceph_msg_put(req->r_request);
2024                 req->r_request = NULL;
2025         }
2026         msg = create_request_message(mdsc, req, mds);
2027         if (IS_ERR(msg)) {
2028                 req->r_err = PTR_ERR(msg);
2029                 complete_request(mdsc, req);
2030                 return PTR_ERR(msg);
2031         }
2032         req->r_request = msg;
2033
2034         rhead = msg->front.iov_base;
2035         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2036         if (req->r_got_unsafe)
2037                 flags |= CEPH_MDS_FLAG_REPLAY;
2038         if (req->r_locked_dir)
2039                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2040         rhead->flags = cpu_to_le32(flags);
2041         rhead->num_fwd = req->r_num_fwd;
2042         rhead->num_retry = req->r_attempts - 1;
2043         rhead->ino = 0;
2044
2045         dout(" r_locked_dir = %p\n", req->r_locked_dir);
2046         return 0;
2047 }
2048
2049 /*
2050  * send request, or put it on the appropriate wait list.
2051  */
2052 static int __do_request(struct ceph_mds_client *mdsc,
2053                         struct ceph_mds_request *req)
2054 {
2055         struct ceph_mds_session *session = NULL;
2056         int mds = -1;
2057         int err = -EAGAIN;
2058
2059         if (req->r_err || req->r_got_result) {
2060                 if (req->r_aborted)
2061                         __unregister_request(mdsc, req);
2062                 goto out;
2063         }
2064
2065         if (req->r_timeout &&
2066             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2067                 dout("do_request timed out\n");
2068                 err = -EIO;
2069                 goto finish;
2070         }
2071
2072         put_request_session(req);
2073
2074         mds = __choose_mds(mdsc, req);
2075         if (mds < 0 ||
2076             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2077                 dout("do_request no mds or not active, waiting for map\n");
2078                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2079                 goto out;
2080         }
2081
2082         /* get, open session */
2083         session = __ceph_lookup_mds_session(mdsc, mds);
2084         if (!session) {
2085                 session = register_session(mdsc, mds);
2086                 if (IS_ERR(session)) {
2087                         err = PTR_ERR(session);
2088                         goto finish;
2089                 }
2090         }
2091         req->r_session = get_session(session);
2092
2093         dout("do_request mds%d session %p state %s\n", mds, session,
2094              ceph_session_state_name(session->s_state));
2095         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2096             session->s_state != CEPH_MDS_SESSION_HUNG) {
2097                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2098                     session->s_state == CEPH_MDS_SESSION_CLOSING)
2099                         __open_session(mdsc, session);
2100                 list_add(&req->r_wait, &session->s_waiting);
2101                 goto out_session;
2102         }
2103
2104         /* send request */
2105         req->r_resend_mds = -1;   /* forget any previous mds hint */
2106
2107         if (req->r_request_started == 0)   /* note request start time */
2108                 req->r_request_started = jiffies;
2109
2110         err = __prepare_send_request(mdsc, req, mds);
2111         if (!err) {
2112                 ceph_msg_get(req->r_request);
2113                 ceph_con_send(&session->s_con, req->r_request);
2114         }
2115
2116 out_session:
2117         ceph_put_mds_session(session);
2118 out:
2119         return err;
2120
2121 finish:
2122         req->r_err = err;
2123         complete_request(mdsc, req);
2124         goto out;
2125 }
2126
2127 /*
2128  * called under mdsc->mutex
2129  */
2130 static void __wake_requests(struct ceph_mds_client *mdsc,
2131                             struct list_head *head)
2132 {
2133         struct ceph_mds_request *req;
2134         LIST_HEAD(tmp_list);
2135
2136         list_splice_init(head, &tmp_list);
2137
2138         while (!list_empty(&tmp_list)) {
2139                 req = list_entry(tmp_list.next,
2140                                  struct ceph_mds_request, r_wait);
2141                 list_del_init(&req->r_wait);
2142                 dout(" wake request %p tid %llu\n", req, req->r_tid);
2143                 __do_request(mdsc, req);
2144         }
2145 }
2146
2147 /*
2148  * Wake up threads with requests pending for @mds, so that they can
2149  * resubmit their requests to a possibly different mds.
2150  */
2151 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2152 {
2153         struct ceph_mds_request *req;
2154         struct rb_node *p = rb_first(&mdsc->request_tree);
2155
2156         dout("kick_requests mds%d\n", mds);
2157         while (p) {
2158                 req = rb_entry(p, struct ceph_mds_request, r_node);
2159                 p = rb_next(p);
2160                 if (req->r_got_unsafe)
2161                         continue;
2162                 if (req->r_session &&
2163                     req->r_session->s_mds == mds) {
2164                         dout(" kicking tid %llu\n", req->r_tid);
2165                         list_del_init(&req->r_wait);
2166                         __do_request(mdsc, req);
2167                 }
2168         }
2169 }
2170
2171 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2172                               struct ceph_mds_request *req)
2173 {
2174         dout("submit_request on %p\n", req);
2175         mutex_lock(&mdsc->mutex);
2176         __register_request(mdsc, req, NULL);
2177         __do_request(mdsc, req);
2178         mutex_unlock(&mdsc->mutex);
2179 }
2180
2181 /*
2182  * Synchrously perform an mds request.  Take care of all of the
2183  * session setup, forwarding, retry details.
2184  */
2185 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2186                          struct inode *dir,
2187                          struct ceph_mds_request *req)
2188 {
2189         int err;
2190
2191         dout("do_request on %p\n", req);
2192
2193         /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
2194         if (req->r_inode)
2195                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2196         if (req->r_locked_dir)
2197                 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
2198         if (req->r_old_dentry_dir)
2199                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2200                                   CEPH_CAP_PIN);
2201
2202         /* issue */
2203         mutex_lock(&mdsc->mutex);
2204         __register_request(mdsc, req, dir);
2205         __do_request(mdsc, req);
2206
2207         if (req->r_err) {
2208                 err = req->r_err;
2209                 __unregister_request(mdsc, req);
2210                 dout("do_request early error %d\n", err);
2211                 goto out;
2212         }
2213
2214         /* wait */
2215         mutex_unlock(&mdsc->mutex);
2216         dout("do_request waiting\n");
2217         if (req->r_timeout) {
2218                 err = (long)wait_for_completion_killable_timeout(
2219                         &req->r_completion, req->r_timeout);
2220                 if (err == 0)
2221                         err = -EIO;
2222         } else if (req->r_wait_for_completion) {
2223                 err = req->r_wait_for_completion(mdsc, req);
2224         } else {
2225                 err = wait_for_completion_killable(&req->r_completion);
2226         }
2227         dout("do_request waited, got %d\n", err);
2228         mutex_lock(&mdsc->mutex);
2229
2230         /* only abort if we didn't race with a real reply */
2231         if (req->r_got_result) {
2232                 err = le32_to_cpu(req->r_reply_info.head->result);
2233         } else if (err < 0) {
2234                 dout("aborted request %lld with %d\n", req->r_tid, err);
2235
2236                 /*
2237                  * ensure we aren't running concurrently with
2238                  * ceph_fill_trace or ceph_readdir_prepopulate, which
2239                  * rely on locks (dir mutex) held by our caller.
2240                  */
2241                 mutex_lock(&req->r_fill_mutex);
2242                 req->r_err = err;
2243                 req->r_aborted = true;
2244                 mutex_unlock(&req->r_fill_mutex);
2245
2246                 if (req->r_locked_dir &&
2247                     (req->r_op & CEPH_MDS_OP_WRITE))
2248                         ceph_invalidate_dir_request(req);
2249         } else {
2250                 err = req->r_err;
2251         }
2252
2253 out:
2254         mutex_unlock(&mdsc->mutex);
2255         dout("do_request %p done, result %d\n", req, err);
2256         return err;
2257 }
2258
2259 /*
2260  * Invalidate dir's completeness, dentry lease state on an aborted MDS
2261  * namespace request.
2262  */
2263 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2264 {
2265         struct inode *inode = req->r_locked_dir;
2266
2267         dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2268
2269         ceph_dir_clear_complete(inode);
2270         if (req->r_dentry)
2271                 ceph_invalidate_dentry_lease(req->r_dentry);
2272         if (req->r_old_dentry)
2273                 ceph_invalidate_dentry_lease(req->r_old_dentry);
2274 }
2275
2276 /*
2277  * Handle mds reply.
2278  *
2279  * We take the session mutex and parse and process the reply immediately.
2280  * This preserves the logical ordering of replies, capabilities, etc., sent
2281  * by the MDS as they are applied to our local cache.
2282  */
2283 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2284 {
2285         struct ceph_mds_client *mdsc = session->s_mdsc;
2286         struct ceph_mds_request *req;
2287         struct ceph_mds_reply_head *head = msg->front.iov_base;
2288         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2289         u64 tid;
2290         int err, result;
2291         int mds = session->s_mds;
2292
2293         if (msg->front.iov_len < sizeof(*head)) {
2294                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2295                 ceph_msg_dump(msg);
2296                 return;
2297         }
2298
2299         /* get request, session */
2300         tid = le64_to_cpu(msg->hdr.tid);
2301         mutex_lock(&mdsc->mutex);
2302         req = __lookup_request(mdsc, tid);
2303         if (!req) {
2304                 dout("handle_reply on unknown tid %llu\n", tid);
2305                 mutex_unlock(&mdsc->mutex);
2306                 return;
2307         }
2308         dout("handle_reply %p\n", req);
2309
2310         /* correct session? */
2311         if (req->r_session != session) {
2312                 pr_err("mdsc_handle_reply got %llu on session mds%d"
2313                        " not mds%d\n", tid, session->s_mds,
2314                        req->r_session ? req->r_session->s_mds : -1);
2315                 mutex_unlock(&mdsc->mutex);
2316                 goto out;
2317         }
2318
2319         /* dup? */
2320         if ((req->r_got_unsafe && !head->safe) ||
2321             (req->r_got_safe && head->safe)) {
2322                 pr_warn("got a dup %s reply on %llu from mds%d\n",
2323                            head->safe ? "safe" : "unsafe", tid, mds);
2324                 mutex_unlock(&mdsc->mutex);
2325                 goto out;
2326         }
2327         if (req->r_got_safe && !head->safe) {
2328                 pr_warn("got unsafe after safe on %llu from mds%d\n",
2329                            tid, mds);
2330                 mutex_unlock(&mdsc->mutex);
2331                 goto out;
2332         }
2333
2334         result = le32_to_cpu(head->result);
2335
2336         /*
2337          * Handle an ESTALE
2338          * if we're not talking to the authority, send to them
2339          * if the authority has changed while we weren't looking,
2340          * send to new authority
2341          * Otherwise we just have to return an ESTALE
2342          */
2343         if (result == -ESTALE) {
2344                 dout("got ESTALE on request %llu", req->r_tid);
2345                 req->r_resend_mds = -1;
2346                 if (req->r_direct_mode != USE_AUTH_MDS) {
2347                         dout("not using auth, setting for that now");
2348                         req->r_direct_mode = USE_AUTH_MDS;
2349                         __do_request(mdsc, req);
2350                         mutex_unlock(&mdsc->mutex);
2351                         goto out;
2352                 } else  {
2353                         int mds = __choose_mds(mdsc, req);
2354                         if (mds >= 0 && mds != req->r_session->s_mds) {
2355                                 dout("but auth changed, so resending");
2356                                 __do_request(mdsc, req);
2357                                 mutex_unlock(&mdsc->mutex);
2358                                 goto out;
2359                         }
2360                 }
2361                 dout("have to return ESTALE on request %llu", req->r_tid);
2362         }
2363
2364
2365         if (head->safe) {
2366                 req->r_got_safe = true;
2367                 __unregister_request(mdsc, req);
2368
2369                 if (req->r_got_unsafe) {
2370                         /*
2371                          * We already handled the unsafe response, now do the
2372                          * cleanup.  No need to examine the response; the MDS
2373                          * doesn't include any result info in the safe
2374                          * response.  And even if it did, there is nothing
2375                          * useful we could do with a revised return value.
2376                          */
2377                         dout("got safe reply %llu, mds%d\n", tid, mds);
2378                         list_del_init(&req->r_unsafe_item);
2379
2380                         /* last unsafe request during umount? */
2381                         if (mdsc->stopping && !__get_oldest_req(mdsc))
2382                                 complete_all(&mdsc->safe_umount_waiters);
2383                         mutex_unlock(&mdsc->mutex);
2384                         goto out;
2385                 }
2386         } else {
2387                 req->r_got_unsafe = true;
2388                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2389         }
2390
2391         dout("handle_reply tid %lld result %d\n", tid, result);
2392         rinfo = &req->r_reply_info;
2393         err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2394         mutex_unlock(&mdsc->mutex);
2395
2396         mutex_lock(&session->s_mutex);
2397         if (err < 0) {
2398                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2399                 ceph_msg_dump(msg);
2400                 goto out_err;
2401         }
2402
2403         /* snap trace */
2404         if (rinfo->snapblob_len) {
2405                 down_write(&mdsc->snap_rwsem);
2406                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2407                                rinfo->snapblob + rinfo->snapblob_len,
2408                                le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
2409                 downgrade_write(&mdsc->snap_rwsem);
2410         } else {
2411                 down_read(&mdsc->snap_rwsem);
2412         }
2413
2414         /* insert trace into our cache */
2415         mutex_lock(&req->r_fill_mutex);
2416         err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2417         if (err == 0) {
2418                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2419                                     req->r_op == CEPH_MDS_OP_LSSNAP))
2420                         ceph_readdir_prepopulate(req, req->r_session);
2421                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2422         }
2423         mutex_unlock(&req->r_fill_mutex);
2424
2425         up_read(&mdsc->snap_rwsem);
2426 out_err:
2427         mutex_lock(&mdsc->mutex);
2428         if (!req->r_aborted) {
2429                 if (err) {
2430                         req->r_err = err;
2431                 } else {
2432                         req->r_reply = msg;
2433                         ceph_msg_get(msg);
2434                         req->r_got_result = true;
2435                 }
2436         } else {
2437                 dout("reply arrived after request %lld was aborted\n", tid);
2438         }
2439         mutex_unlock(&mdsc->mutex);
2440
2441         ceph_add_cap_releases(mdsc, req->r_session);
2442         mutex_unlock(&session->s_mutex);
2443
2444         /* kick calling process */
2445         complete_request(mdsc, req);
2446 out:
2447         ceph_mdsc_put_request(req);
2448         return;
2449 }
2450
2451
2452
2453 /*
2454  * handle mds notification that our request has been forwarded.
2455  */
2456 static void handle_forward(struct ceph_mds_client *mdsc,
2457                            struct ceph_mds_session *session,
2458                            struct ceph_msg *msg)
2459 {
2460         struct ceph_mds_request *req;
2461         u64 tid = le64_to_cpu(msg->hdr.tid);
2462         u32 next_mds;
2463         u32 fwd_seq;
2464         int err = -EINVAL;
2465         void *p = msg->front.iov_base;
2466         void *end = p + msg->front.iov_len;
2467
2468         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2469         next_mds = ceph_decode_32(&p);
2470         fwd_seq = ceph_decode_32(&p);
2471
2472         mutex_lock(&mdsc->mutex);
2473         req = __lookup_request(mdsc, tid);
2474         if (!req) {
2475                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2476                 goto out;  /* dup reply? */
2477         }
2478
2479         if (req->r_aborted) {
2480                 dout("forward tid %llu aborted, unregistering\n", tid);
2481                 __unregister_request(mdsc, req);
2482         } else if (fwd_seq <= req->r_num_fwd) {
2483                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2484                      tid, next_mds, req->r_num_fwd, fwd_seq);
2485         } else {
2486                 /* resend. forward race not possible; mds would drop */
2487                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2488                 BUG_ON(req->r_err);
2489                 BUG_ON(req->r_got_result);
2490                 req->r_num_fwd = fwd_seq;
2491                 req->r_resend_mds = next_mds;
2492                 put_request_session(req);
2493                 __do_request(mdsc, req);
2494         }
2495         ceph_mdsc_put_request(req);
2496 out:
2497         mutex_unlock(&mdsc->mutex);
2498         return;
2499
2500 bad:
2501         pr_err("mdsc_handle_forward decode error err=%d\n", err);
2502 }
2503
2504 /*
2505  * handle a mds session control message
2506  */
2507 static void handle_session(struct ceph_mds_session *session,
2508                            struct ceph_msg *msg)
2509 {
2510         struct ceph_mds_client *mdsc = session->s_mdsc;
2511         u32 op;
2512         u64 seq;
2513         int mds = session->s_mds;
2514         struct ceph_mds_session_head *h = msg->front.iov_base;
2515         int wake = 0;
2516
2517         /* decode */
2518         if (msg->front.iov_len != sizeof(*h))
2519                 goto bad;
2520         op = le32_to_cpu(h->op);
2521         seq = le64_to_cpu(h->seq);
2522
2523         mutex_lock(&mdsc->mutex);
2524         if (op == CEPH_SESSION_CLOSE)
2525                 __unregister_session(mdsc, session);
2526         /* FIXME: this ttl calculation is generous */
2527         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2528         mutex_unlock(&mdsc->mutex);
2529
2530         mutex_lock(&session->s_mutex);
2531
2532         dout("handle_session mds%d %s %p state %s seq %llu\n",
2533              mds, ceph_session_op_name(op), session,
2534              ceph_session_state_name(session->s_state), seq);
2535
2536         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2537                 session->s_state = CEPH_MDS_SESSION_OPEN;
2538                 pr_info("mds%d came back\n", session->s_mds);
2539         }
2540
2541         switch (op) {
2542         case CEPH_SESSION_OPEN:
2543                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2544                         pr_info("mds%d reconnect success\n", session->s_mds);
2545                 session->s_state = CEPH_MDS_SESSION_OPEN;
2546                 renewed_caps(mdsc, session, 0);
2547                 wake = 1;
2548                 if (mdsc->stopping)
2549                         __close_session(mdsc, session);
2550                 break;
2551
2552         case CEPH_SESSION_RENEWCAPS:
2553                 if (session->s_renew_seq == seq)
2554                         renewed_caps(mdsc, session, 1);
2555                 break;
2556
2557         case CEPH_SESSION_CLOSE:
2558                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2559                         pr_info("mds%d reconnect denied\n", session->s_mds);
2560                 remove_session_caps(session);
2561                 wake = 2; /* for good measure */
2562                 wake_up_all(&mdsc->session_close_wq);
2563                 break;
2564
2565         case CEPH_SESSION_STALE:
2566                 pr_info("mds%d caps went stale, renewing\n",
2567                         session->s_mds);
2568                 spin_lock(&session->s_gen_ttl_lock);
2569                 session->s_cap_gen++;
2570                 session->s_cap_ttl = jiffies - 1;
2571                 spin_unlock(&session->s_gen_ttl_lock);
2572                 send_renew_caps(mdsc, session);
2573                 break;
2574
2575         case CEPH_SESSION_RECALL_STATE:
2576                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2577                 break;
2578
2579         case CEPH_SESSION_FLUSHMSG:
2580                 send_flushmsg_ack(mdsc, session, seq);
2581                 break;
2582
2583         default:
2584                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2585                 WARN_ON(1);
2586         }
2587
2588         mutex_unlock(&session->s_mutex);
2589         if (wake) {
2590                 mutex_lock(&mdsc->mutex);
2591                 __wake_requests(mdsc, &session->s_waiting);
2592                 if (wake == 2)
2593                         kick_requests(mdsc, mds);
2594                 mutex_unlock(&mdsc->mutex);
2595         }
2596         return;
2597
2598 bad:
2599         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2600                (int)msg->front.iov_len);
2601         ceph_msg_dump(msg);
2602         return;
2603 }
2604
2605
2606 /*
2607  * called under session->mutex.
2608  */
2609 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2610                                    struct ceph_mds_session *session)
2611 {
2612         struct ceph_mds_request *req, *nreq;
2613         int err;
2614
2615         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2616
2617         mutex_lock(&mdsc->mutex);
2618         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2619                 err = __prepare_send_request(mdsc, req, session->s_mds);
2620                 if (!err) {
2621                         ceph_msg_get(req->r_request);
2622                         ceph_con_send(&session->s_con, req->r_request);
2623                 }
2624         }
2625         mutex_unlock(&mdsc->mutex);
2626 }
2627
2628 /*
2629  * Encode information about a cap for a reconnect with the MDS.
2630  */
2631 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2632                           void *arg)
2633 {
2634         union {
2635                 struct ceph_mds_cap_reconnect v2;
2636                 struct ceph_mds_cap_reconnect_v1 v1;
2637         } rec;
2638         size_t reclen;
2639         struct ceph_inode_info *ci;
2640         struct ceph_reconnect_state *recon_state = arg;
2641         struct ceph_pagelist *pagelist = recon_state->pagelist;
2642         char *path;
2643         int pathlen, err;
2644         u64 pathbase;
2645         struct dentry *dentry;
2646
2647         ci = cap->ci;
2648
2649         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2650              inode, ceph_vinop(inode), cap, cap->cap_id,
2651              ceph_cap_string(cap->issued));
2652         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2653         if (err)
2654                 return err;
2655
2656         dentry = d_find_alias(inode);
2657         if (dentry) {
2658                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2659                 if (IS_ERR(path)) {
2660                         err = PTR_ERR(path);
2661                         goto out_dput;
2662                 }
2663         } else {
2664                 path = NULL;
2665                 pathlen = 0;
2666         }
2667         err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2668         if (err)
2669                 goto out_free;
2670
2671         spin_lock(&ci->i_ceph_lock);
2672         cap->seq = 0;        /* reset cap seq */
2673         cap->issue_seq = 0;  /* and issue_seq */
2674         cap->mseq = 0;       /* and migrate_seq */
2675         cap->cap_gen = cap->session->s_cap_gen;
2676
2677         if (recon_state->flock) {
2678                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2679                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2680                 rec.v2.issued = cpu_to_le32(cap->issued);
2681                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2682                 rec.v2.pathbase = cpu_to_le64(pathbase);
2683                 rec.v2.flock_len = 0;
2684                 reclen = sizeof(rec.v2);
2685         } else {
2686                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2687                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2688                 rec.v1.issued = cpu_to_le32(cap->issued);
2689                 rec.v1.size = cpu_to_le64(inode->i_size);
2690                 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2691                 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2692                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2693                 rec.v1.pathbase = cpu_to_le64(pathbase);
2694                 reclen = sizeof(rec.v1);
2695         }
2696         spin_unlock(&ci->i_ceph_lock);
2697
2698         if (recon_state->flock) {
2699                 int num_fcntl_locks, num_flock_locks;
2700                 struct ceph_filelock *flocks;
2701
2702 encode_again:
2703                 spin_lock(&inode->i_lock);
2704                 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2705                 spin_unlock(&inode->i_lock);
2706                 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2707                                  sizeof(struct ceph_filelock), GFP_NOFS);
2708                 if (!flocks) {
2709                         err = -ENOMEM;
2710                         goto out_free;
2711                 }
2712                 spin_lock(&inode->i_lock);
2713                 err = ceph_encode_locks_to_buffer(inode, flocks,
2714                                                   num_fcntl_locks,
2715                                                   num_flock_locks);
2716                 spin_unlock(&inode->i_lock);
2717                 if (err) {
2718                         kfree(flocks);
2719                         if (err == -ENOSPC)
2720                                 goto encode_again;
2721                         goto out_free;
2722                 }
2723                 /*
2724                  * number of encoded locks is stable, so copy to pagelist
2725                  */
2726                 rec.v2.flock_len = cpu_to_le32(2*sizeof(u32) +
2727                                     (num_fcntl_locks+num_flock_locks) *
2728                                     sizeof(struct ceph_filelock));
2729                 err = ceph_pagelist_append(pagelist, &rec, reclen);
2730                 if (!err)
2731                         err = ceph_locks_to_pagelist(flocks, pagelist,
2732                                                      num_fcntl_locks,
2733                                                      num_flock_locks);
2734                 kfree(flocks);
2735         } else {
2736                 err = ceph_pagelist_append(pagelist, &rec, reclen);
2737         }
2738
2739         recon_state->nr_caps++;
2740 out_free:
2741         kfree(path);
2742 out_dput:
2743         dput(dentry);
2744         return err;
2745 }
2746
2747
2748 /*
2749  * If an MDS fails and recovers, clients need to reconnect in order to
2750  * reestablish shared state.  This includes all caps issued through
2751  * this session _and_ the snap_realm hierarchy.  Because it's not
2752  * clear which snap realms the mds cares about, we send everything we
2753  * know about.. that ensures we'll then get any new info the
2754  * recovering MDS might have.
2755  *
2756  * This is a relatively heavyweight operation, but it's rare.
2757  *
2758  * called with mdsc->mutex held.
2759  */
2760 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2761                                struct ceph_mds_session *session)
2762 {
2763         struct ceph_msg *reply;
2764         struct rb_node *p;
2765         int mds = session->s_mds;
2766         int err = -ENOMEM;
2767         int s_nr_caps;
2768         struct ceph_pagelist *pagelist;
2769         struct ceph_reconnect_state recon_state;
2770
2771         pr_info("mds%d reconnect start\n", mds);
2772
2773         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2774         if (!pagelist)
2775                 goto fail_nopagelist;
2776         ceph_pagelist_init(pagelist);
2777
2778         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2779         if (!reply)
2780                 goto fail_nomsg;
2781
2782         mutex_lock(&session->s_mutex);
2783         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2784         session->s_seq = 0;
2785
2786         dout("session %p state %s\n", session,
2787              ceph_session_state_name(session->s_state));
2788
2789         spin_lock(&session->s_gen_ttl_lock);
2790         session->s_cap_gen++;
2791         spin_unlock(&session->s_gen_ttl_lock);
2792
2793         spin_lock(&session->s_cap_lock);
2794         /*
2795          * notify __ceph_remove_cap() that we are composing cap reconnect.
2796          * If a cap get released before being added to the cap reconnect,
2797          * __ceph_remove_cap() should skip queuing cap release.
2798          */
2799         session->s_cap_reconnect = 1;
2800         /* drop old cap expires; we're about to reestablish that state */
2801         discard_cap_releases(mdsc, session);
2802         spin_unlock(&session->s_cap_lock);
2803
2804         /* trim unused caps to reduce MDS's cache rejoin time */
2805         shrink_dcache_parent(mdsc->fsc->sb->s_root);
2806
2807         ceph_con_close(&session->s_con);
2808         ceph_con_open(&session->s_con,
2809                       CEPH_ENTITY_TYPE_MDS, mds,
2810                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2811
2812         /* replay unsafe requests */
2813         replay_unsafe_requests(mdsc, session);
2814
2815         down_read(&mdsc->snap_rwsem);
2816
2817         /* traverse this session's caps */
2818         s_nr_caps = session->s_nr_caps;
2819         err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
2820         if (err)
2821                 goto fail;
2822
2823         recon_state.nr_caps = 0;
2824         recon_state.pagelist = pagelist;
2825         recon_state.flock = session->s_con.peer_features & CEPH_FEATURE_FLOCK;
2826         err = iterate_session_caps(session, encode_caps_cb, &recon_state);
2827         if (err < 0)
2828                 goto fail;
2829
2830         spin_lock(&session->s_cap_lock);
2831         session->s_cap_reconnect = 0;
2832         spin_unlock(&session->s_cap_lock);
2833
2834         /*
2835          * snaprealms.  we provide mds with the ino, seq (version), and
2836          * parent for all of our realms.  If the mds has any newer info,
2837          * it will tell us.
2838          */
2839         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2840                 struct ceph_snap_realm *realm =
2841                         rb_entry(p, struct ceph_snap_realm, node);
2842                 struct ceph_mds_snaprealm_reconnect sr_rec;
2843
2844                 dout(" adding snap realm %llx seq %lld parent %llx\n",
2845                      realm->ino, realm->seq, realm->parent_ino);
2846                 sr_rec.ino = cpu_to_le64(realm->ino);
2847                 sr_rec.seq = cpu_to_le64(realm->seq);
2848                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2849                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2850                 if (err)
2851                         goto fail;
2852         }
2853
2854         if (recon_state.flock)
2855                 reply->hdr.version = cpu_to_le16(2);
2856
2857         /* raced with cap release? */
2858         if (s_nr_caps != recon_state.nr_caps) {
2859                 struct page *page = list_first_entry(&pagelist->head,
2860                                                      struct page, lru);
2861                 __le32 *addr = kmap_atomic(page);
2862                 *addr = cpu_to_le32(recon_state.nr_caps);
2863                 kunmap_atomic(addr);
2864         }
2865
2866         reply->hdr.data_len = cpu_to_le32(pagelist->length);
2867         ceph_msg_data_add_pagelist(reply, pagelist);
2868         ceph_con_send(&session->s_con, reply);
2869
2870         mutex_unlock(&session->s_mutex);
2871
2872         mutex_lock(&mdsc->mutex);
2873         __wake_requests(mdsc, &session->s_waiting);
2874         mutex_unlock(&mdsc->mutex);
2875
2876         up_read(&mdsc->snap_rwsem);
2877         return;
2878
2879 fail:
2880         ceph_msg_put(reply);
2881         up_read(&mdsc->snap_rwsem);
2882         mutex_unlock(&session->s_mutex);
2883 fail_nomsg:
2884         ceph_pagelist_release(pagelist);
2885 fail_nopagelist:
2886         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
2887         return;
2888 }
2889
2890
2891 /*
2892  * compare old and new mdsmaps, kicking requests
2893  * and closing out old connections as necessary
2894  *
2895  * called under mdsc->mutex.
2896  */
2897 static void check_new_map(struct ceph_mds_client *mdsc,
2898                           struct ceph_mdsmap *newmap,
2899                           struct ceph_mdsmap *oldmap)
2900 {
2901         int i;
2902         int oldstate, newstate;
2903         struct ceph_mds_session *s;
2904
2905         dout("check_new_map new %u old %u\n",
2906              newmap->m_epoch, oldmap->m_epoch);
2907
2908         for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2909                 if (mdsc->sessions[i] == NULL)
2910                         continue;
2911                 s = mdsc->sessions[i];
2912                 oldstate = ceph_mdsmap_get_state(oldmap, i);
2913                 newstate = ceph_mdsmap_get_state(newmap, i);
2914
2915                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
2916                      i, ceph_mds_state_name(oldstate),
2917                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
2918                      ceph_mds_state_name(newstate),
2919                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
2920                      ceph_session_state_name(s->s_state));
2921
2922                 if (i >= newmap->m_max_mds ||
2923                     memcmp(ceph_mdsmap_get_addr(oldmap, i),
2924                            ceph_mdsmap_get_addr(newmap, i),
2925                            sizeof(struct ceph_entity_addr))) {
2926                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2927                                 /* the session never opened, just close it
2928                                  * out now */
2929                                 __wake_requests(mdsc, &s->s_waiting);
2930                                 __unregister_session(mdsc, s);
2931                         } else {
2932                                 /* just close it */
2933                                 mutex_unlock(&mdsc->mutex);
2934                                 mutex_lock(&s->s_mutex);
2935                                 mutex_lock(&mdsc->mutex);
2936                                 ceph_con_close(&s->s_con);
2937                                 mutex_unlock(&s->s_mutex);
2938                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2939                         }
2940
2941                         /* kick any requests waiting on the recovering mds */
2942                         kick_requests(mdsc, i);
2943                 } else if (oldstate == newstate) {
2944                         continue;  /* nothing new with this mds */
2945                 }
2946
2947                 /*
2948                  * send reconnect?
2949                  */
2950                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2951                     newstate >= CEPH_MDS_STATE_RECONNECT) {
2952                         mutex_unlock(&mdsc->mutex);
2953                         send_mds_reconnect(mdsc, s);
2954                         mutex_lock(&mdsc->mutex);
2955                 }
2956
2957                 /*
2958                  * kick request on any mds that has gone active.
2959                  */
2960                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2961                     newstate >= CEPH_MDS_STATE_ACTIVE) {
2962                         if (oldstate != CEPH_MDS_STATE_CREATING &&
2963                             oldstate != CEPH_MDS_STATE_STARTING)
2964                                 pr_info("mds%d recovery completed\n", s->s_mds);
2965                         kick_requests(mdsc, i);
2966                         ceph_kick_flushing_caps(mdsc, s);
2967                         wake_up_session_caps(s, 1);
2968                 }
2969         }
2970
2971         for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
2972                 s = mdsc->sessions[i];
2973                 if (!s)
2974                         continue;
2975                 if (!ceph_mdsmap_is_laggy(newmap, i))
2976                         continue;
2977                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2978                     s->s_state == CEPH_MDS_SESSION_HUNG ||
2979                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
2980                         dout(" connecting to export targets of laggy mds%d\n",
2981                              i);
2982                         __open_export_target_sessions(mdsc, s);
2983                 }
2984         }
2985 }
2986
2987
2988
2989 /*
2990  * leases
2991  */
2992
2993 /*
2994  * caller must hold session s_mutex, dentry->d_lock
2995  */
2996 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2997 {
2998         struct ceph_dentry_info *di = ceph_dentry(dentry);
2999
3000         ceph_put_mds_session(di->lease_session);
3001         di->lease_session = NULL;
3002 }
3003
3004 static void handle_lease(struct ceph_mds_client *mdsc,
3005                          struct ceph_mds_session *session,
3006                          struct ceph_msg *msg)
3007 {
3008         struct super_block *sb = mdsc->fsc->sb;
3009         struct inode *inode;
3010         struct dentry *parent, *dentry;
3011         struct ceph_dentry_info *di;
3012         int mds = session->s_mds;
3013         struct ceph_mds_lease *h = msg->front.iov_base;
3014         u32 seq;
3015         struct ceph_vino vino;
3016         struct qstr dname;
3017         int release = 0;
3018
3019         dout("handle_lease from mds%d\n", mds);
3020
3021         /* decode */
3022         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3023                 goto bad;
3024         vino.ino = le64_to_cpu(h->ino);
3025         vino.snap = CEPH_NOSNAP;
3026         seq = le32_to_cpu(h->seq);
3027         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3028         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3029         if (dname.len != get_unaligned_le32(h+1))
3030                 goto bad;
3031
3032         /* lookup inode */
3033         inode = ceph_find_inode(sb, vino);
3034         dout("handle_lease %s, ino %llx %p %.*s\n",
3035              ceph_lease_op_name(h->action), vino.ino, inode,
3036              dname.len, dname.name);
3037
3038         mutex_lock(&session->s_mutex);
3039         session->s_seq++;
3040
3041         if (inode == NULL) {
3042                 dout("handle_lease no inode %llx\n", vino.ino);
3043                 goto release;
3044         }
3045
3046         /* dentry */
3047         parent = d_find_alias(inode);
3048         if (!parent) {
3049                 dout("no parent dentry on inode %p\n", inode);
3050                 WARN_ON(1);
3051                 goto release;  /* hrm... */
3052         }
3053         dname.hash = full_name_hash(dname.name, dname.len);
3054         dentry = d_lookup(parent, &dname);
3055         dput(parent);
3056         if (!dentry)
3057                 goto release;
3058
3059         spin_lock(&dentry->d_lock);
3060         di = ceph_dentry(dentry);
3061         switch (h->action) {
3062         case CEPH_MDS_LEASE_REVOKE:
3063                 if (di->lease_session == session) {
3064                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3065                                 h->seq = cpu_to_le32(di->lease_seq);
3066                         __ceph_mdsc_drop_dentry_lease(dentry);
3067                 }
3068                 release = 1;
3069                 break;
3070
3071         case CEPH_MDS_LEASE_RENEW:
3072                 if (di->lease_session == session &&
3073                     di->lease_gen == session->s_cap_gen &&
3074                     di->lease_renew_from &&
3075                     di->lease_renew_after == 0) {
3076                         unsigned long duration =
3077                                 le32_to_cpu(h->duration_ms) * HZ / 1000;
3078
3079                         di->lease_seq = seq;
3080                         dentry->d_time = di->lease_renew_from + duration;
3081                         di->lease_renew_after = di->lease_renew_from +
3082                                 (duration >> 1);
3083                         di->lease_renew_from = 0;
3084                 }
3085                 break;
3086         }
3087         spin_unlock(&dentry->d_lock);
3088         dput(dentry);
3089
3090         if (!release)
3091                 goto out;
3092
3093 release:
3094         /* let's just reuse the same message */
3095         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3096         ceph_msg_get(msg);
3097         ceph_con_send(&session->s_con, msg);
3098
3099 out:
3100         iput(inode);
3101         mutex_unlock(&session->s_mutex);
3102         return;
3103
3104 bad:
3105         pr_err("corrupt lease message\n");
3106         ceph_msg_dump(msg);
3107 }
3108
3109 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3110                               struct inode *inode,
3111                               struct dentry *dentry, char action,
3112                               u32 seq)
3113 {
3114         struct ceph_msg *msg;
3115         struct ceph_mds_lease *lease;
3116         int len = sizeof(*lease) + sizeof(u32);
3117         int dnamelen = 0;
3118
3119         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3120              inode, dentry, ceph_lease_op_name(action), session->s_mds);
3121         dnamelen = dentry->d_name.len;
3122         len += dnamelen;
3123
3124         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3125         if (!msg)
3126                 return;
3127         lease = msg->front.iov_base;
3128         lease->action = action;
3129         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3130         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3131         lease->seq = cpu_to_le32(seq);
3132         put_unaligned_le32(dnamelen, lease + 1);
3133         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3134
3135         /*
3136          * if this is a preemptive lease RELEASE, no need to
3137          * flush request stream, since the actual request will
3138          * soon follow.
3139          */
3140         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3141
3142         ceph_con_send(&session->s_con, msg);
3143 }
3144
3145 /*
3146  * Preemptively release a lease we expect to invalidate anyway.
3147  * Pass @inode always, @dentry is optional.
3148  */
3149 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
3150                              struct dentry *dentry)
3151 {
3152         struct ceph_dentry_info *di;
3153         struct ceph_mds_session *session;
3154         u32 seq;
3155
3156         BUG_ON(inode == NULL);
3157         BUG_ON(dentry == NULL);
3158
3159         /* is dentry lease valid? */
3160         spin_lock(&dentry->d_lock);
3161         di = ceph_dentry(dentry);
3162         if (!di || !di->lease_session ||
3163             di->lease_session->s_mds < 0 ||
3164             di->lease_gen != di->lease_session->s_cap_gen ||
3165             !time_before(jiffies, dentry->d_time)) {
3166                 dout("lease_release inode %p dentry %p -- "
3167                      "no lease\n",
3168                      inode, dentry);
3169                 spin_unlock(&dentry->d_lock);
3170                 return;
3171         }
3172
3173         /* we do have a lease on this dentry; note mds and seq */
3174         session = ceph_get_mds_session(di->lease_session);
3175         seq = di->lease_seq;
3176         __ceph_mdsc_drop_dentry_lease(dentry);
3177         spin_unlock(&dentry->d_lock);
3178
3179         dout("lease_release inode %p dentry %p to mds%d\n",
3180              inode, dentry, session->s_mds);
3181         ceph_mdsc_lease_send_msg(session, inode, dentry,
3182                                  CEPH_MDS_LEASE_RELEASE, seq);
3183         ceph_put_mds_session(session);
3184 }
3185
3186 /*
3187  * drop all leases (and dentry refs) in preparation for umount
3188  */
3189 static void drop_leases(struct ceph_mds_client *mdsc)
3190 {
3191         int i;
3192
3193         dout("drop_leases\n");
3194         mutex_lock(&mdsc->mutex);
3195         for (i = 0; i < mdsc->max_sessions; i++) {
3196                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3197                 if (!s)
3198                         continue;
3199                 mutex_unlock(&mdsc->mutex);
3200                 mutex_lock(&s->s_mutex);
3201                 mutex_unlock(&s->s_mutex);
3202                 ceph_put_mds_session(s);
3203                 mutex_lock(&mdsc->mutex);
3204         }
3205         mutex_unlock(&mdsc->mutex);
3206 }
3207
3208
3209
3210 /*
3211  * delayed work -- periodically trim expired leases, renew caps with mds
3212  */
3213 static void schedule_delayed(struct ceph_mds_client *mdsc)
3214 {
3215         int delay = 5;
3216         unsigned hz = round_jiffies_relative(HZ * delay);
3217         schedule_delayed_work(&mdsc->delayed_work, hz);
3218 }
3219
3220 static void delayed_work(struct work_struct *work)
3221 {
3222         int i;
3223         struct ceph_mds_client *mdsc =
3224                 container_of(work, struct ceph_mds_client, delayed_work.work);
3225         int renew_interval;
3226         int renew_caps;
3227
3228         dout("mdsc delayed_work\n");
3229         ceph_check_delayed_caps(mdsc);
3230
3231         mutex_lock(&mdsc->mutex);
3232         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3233         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3234                                    mdsc->last_renew_caps);
3235         if (renew_caps)
3236                 mdsc->last_renew_caps = jiffies;
3237
3238         for (i = 0; i < mdsc->max_sessions; i++) {
3239                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3240                 if (s == NULL)
3241                         continue;
3242                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3243                         dout("resending session close request for mds%d\n",
3244                              s->s_mds);
3245                         request_close_session(mdsc, s);
3246                         ceph_put_mds_session(s);
3247                         continue;
3248                 }
3249                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3250                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3251                                 s->s_state = CEPH_MDS_SESSION_HUNG;
3252                                 pr_info("mds%d hung\n", s->s_mds);
3253                         }
3254                 }
3255                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3256                         /* this mds is failed or recovering, just wait */
3257                         ceph_put_mds_session(s);
3258                         continue;
3259                 }
3260                 mutex_unlock(&mdsc->mutex);
3261
3262                 mutex_lock(&s->s_mutex);
3263                 if (renew_caps)
3264                         send_renew_caps(mdsc, s);
3265                 else
3266                         ceph_con_keepalive(&s->s_con);
3267                 ceph_add_cap_releases(mdsc, s);
3268                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3269                     s->s_state == CEPH_MDS_SESSION_HUNG)
3270                         ceph_send_cap_releases(mdsc, s);
3271                 mutex_unlock(&s->s_mutex);
3272                 ceph_put_mds_session(s);
3273
3274                 mutex_lock(&mdsc->mutex);
3275         }
3276         mutex_unlock(&mdsc->mutex);
3277
3278         schedule_delayed(mdsc);
3279 }
3280
3281 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3282
3283 {
3284         struct ceph_mds_client *mdsc;
3285
3286         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3287         if (!mdsc)
3288                 return -ENOMEM;
3289         mdsc->fsc = fsc;
3290         fsc->mdsc = mdsc;
3291         mutex_init(&mdsc->mutex);
3292         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3293         if (mdsc->mdsmap == NULL) {
3294                 kfree(mdsc);
3295                 return -ENOMEM;
3296         }
3297
3298         init_completion(&mdsc->safe_umount_waiters);
3299         init_waitqueue_head(&mdsc->session_close_wq);
3300         INIT_LIST_HEAD(&mdsc->waiting_for_map);
3301         mdsc->sessions = NULL;
3302         mdsc->max_sessions = 0;
3303         mdsc->stopping = 0;
3304         init_rwsem(&mdsc->snap_rwsem);
3305         mdsc->snap_realms = RB_ROOT;
3306         INIT_LIST_HEAD(&mdsc->snap_empty);
3307         spin_lock_init(&mdsc->snap_empty_lock);
3308         mdsc->last_tid = 0;
3309         mdsc->request_tree = RB_ROOT;
3310         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3311         mdsc->last_renew_caps = jiffies;
3312         INIT_LIST_HEAD(&mdsc->cap_delay_list);
3313         spin_lock_init(&mdsc->cap_delay_lock);
3314         INIT_LIST_HEAD(&mdsc->snap_flush_list);
3315         spin_lock_init(&mdsc->snap_flush_lock);
3316         mdsc->cap_flush_seq = 0;
3317         INIT_LIST_HEAD(&mdsc->cap_dirty);
3318         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3319         mdsc->num_cap_flushing = 0;
3320         spin_lock_init(&mdsc->cap_dirty_lock);
3321         init_waitqueue_head(&mdsc->cap_flushing_wq);
3322         spin_lock_init(&mdsc->dentry_lru_lock);
3323         INIT_LIST_HEAD(&mdsc->dentry_lru);
3324
3325         ceph_caps_init(mdsc);
3326         ceph_adjust_min_caps(mdsc, fsc->min_caps);
3327
3328         return 0;
3329 }
3330
3331 /*
3332  * Wait for safe replies on open mds requests.  If we time out, drop
3333  * all requests from the tree to avoid dangling dentry refs.
3334  */
3335 static void wait_requests(struct ceph_mds_client *mdsc)
3336 {
3337         struct ceph_mds_request *req;
3338         struct ceph_fs_client *fsc = mdsc->fsc;
3339
3340         mutex_lock(&mdsc->mutex);
3341         if (__get_oldest_req(mdsc)) {
3342                 mutex_unlock(&mdsc->mutex);
3343
3344                 dout("wait_requests waiting for requests\n");
3345                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3346                                     fsc->client->options->mount_timeout * HZ);
3347
3348                 /* tear down remaining requests */
3349                 mutex_lock(&mdsc->mutex);
3350                 while ((req = __get_oldest_req(mdsc))) {
3351                         dout("wait_requests timed out on tid %llu\n",
3352                              req->r_tid);
3353                         __unregister_request(mdsc, req);
3354                 }
3355         }
3356         mutex_unlock(&mdsc->mutex);
3357         dout("wait_requests done\n");
3358 }
3359
3360 /*
3361  * called before mount is ro, and before dentries are torn down.
3362  * (hmm, does this still race with new lookups?)
3363  */
3364 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3365 {
3366         dout("pre_umount\n");
3367         mdsc->stopping = 1;
3368
3369         drop_leases(mdsc);
3370         ceph_flush_dirty_caps(mdsc);
3371         wait_requests(mdsc);
3372
3373         /*
3374          * wait for reply handlers to drop their request refs and
3375          * their inode/dcache refs
3376          */
3377         ceph_msgr_flush();
3378 }
3379
3380 /*
3381  * wait for all write mds requests to flush.
3382  */
3383 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3384 {
3385         struct ceph_mds_request *req = NULL, *nextreq;
3386         struct rb_node *n;
3387
3388         mutex_lock(&mdsc->mutex);
3389         dout("wait_unsafe_requests want %lld\n", want_tid);
3390 restart:
3391         req = __get_oldest_req(mdsc);
3392         while (req && req->r_tid <= want_tid) {
3393                 /* find next request */
3394                 n = rb_next(&req->r_node);
3395                 if (n)
3396                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3397                 else
3398                         nextreq = NULL;
3399                 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
3400                         /* write op */
3401                         ceph_mdsc_get_request(req);
3402                         if (nextreq)
3403                                 ceph_mdsc_get_request(nextreq);
3404                         mutex_unlock(&mdsc->mutex);
3405                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
3406                              req->r_tid, want_tid);
3407                         wait_for_completion(&req->r_safe_completion);
3408                         mutex_lock(&mdsc->mutex);
3409                         ceph_mdsc_put_request(req);
3410                         if (!nextreq)
3411                                 break;  /* next dne before, so we're done! */
3412                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
3413                                 /* next request was removed from tree */
3414                                 ceph_mdsc_put_request(nextreq);
3415                                 goto restart;
3416                         }
3417                         ceph_mdsc_put_request(nextreq);  /* won't go away */
3418                 }
3419                 req = nextreq;
3420         }
3421         mutex_unlock(&mdsc->mutex);
3422         dout("wait_unsafe_requests done\n");
3423 }
3424
3425 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3426 {
3427         u64 want_tid, want_flush;
3428
3429         if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3430                 return;
3431
3432         dout("sync\n");
3433         mutex_lock(&mdsc->mutex);
3434         want_tid = mdsc->last_tid;
3435         want_flush = mdsc->cap_flush_seq;
3436         mutex_unlock(&mdsc->mutex);
3437         dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
3438
3439         ceph_flush_dirty_caps(mdsc);
3440
3441         wait_unsafe_requests(mdsc, want_tid);
3442         wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
3443 }
3444
3445 /*
3446  * true if all sessions are closed, or we force unmount
3447  */
3448 static bool done_closing_sessions(struct ceph_mds_client *mdsc)
3449 {
3450         int i, n = 0;
3451
3452         if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3453                 return true;
3454
3455         mutex_lock(&mdsc->mutex);
3456         for (i = 0; i < mdsc->max_sessions; i++)
3457                 if (mdsc->sessions[i])
3458                         n++;
3459         mutex_unlock(&mdsc->mutex);
3460         return n == 0;
3461 }
3462
3463 /*
3464  * called after sb is ro.
3465  */
3466 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3467 {
3468         struct ceph_mds_session *session;
3469         int i;
3470         struct ceph_fs_client *fsc = mdsc->fsc;
3471         unsigned long timeout = fsc->client->options->mount_timeout * HZ;
3472
3473         dout("close_sessions\n");
3474
3475         /* close sessions */
3476         mutex_lock(&mdsc->mutex);
3477         for (i = 0; i < mdsc->max_sessions; i++) {
3478                 session = __ceph_lookup_mds_session(mdsc, i);
3479                 if (!session)
3480                         continue;
3481                 mutex_unlock(&mdsc->mutex);
3482                 mutex_lock(&session->s_mutex);
3483                 __close_session(mdsc, session);
3484                 mutex_unlock(&session->s_mutex);
3485                 ceph_put_mds_session(session);
3486                 mutex_lock(&mdsc->mutex);
3487         }
3488         mutex_unlock(&mdsc->mutex);
3489
3490         dout("waiting for sessions to close\n");
3491         wait_event_timeout(mdsc->session_close_wq, done_closing_sessions(mdsc),
3492                            timeout);
3493
3494         /* tear down remaining sessions */
3495         mutex_lock(&mdsc->mutex);
3496         for (i = 0; i < mdsc->max_sessions; i++) {
3497                 if (mdsc->sessions[i]) {
3498                         session = get_session(mdsc->sessions[i]);
3499                         __unregister_session(mdsc, session);
3500                         mutex_unlock(&mdsc->mutex);
3501                         mutex_lock(&session->s_mutex);
3502                         remove_session_caps(session);
3503                         mutex_unlock(&session->s_mutex);
3504                         ceph_put_mds_session(session);
3505                         mutex_lock(&mdsc->mutex);
3506                 }
3507         }
3508         WARN_ON(!list_empty(&mdsc->cap_delay_list));
3509         mutex_unlock(&mdsc->mutex);
3510
3511         ceph_cleanup_empty_realms(mdsc);
3512
3513         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3514
3515         dout("stopped\n");
3516 }
3517
3518 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3519 {
3520         dout("stop\n");
3521         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3522         if (mdsc->mdsmap)
3523                 ceph_mdsmap_destroy(mdsc->mdsmap);
3524         kfree(mdsc->sessions);
3525         ceph_caps_finalize(mdsc);
3526 }
3527
3528 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3529 {
3530         struct ceph_mds_client *mdsc = fsc->mdsc;
3531
3532         dout("mdsc_destroy %p\n", mdsc);
3533         ceph_mdsc_stop(mdsc);
3534
3535         /* flush out any connection work with references to us */
3536         ceph_msgr_flush();
3537
3538         fsc->mdsc = NULL;
3539         kfree(mdsc);
3540         dout("mdsc_destroy %p done\n", mdsc);
3541 }
3542
3543
3544 /*
3545  * handle mds map update.
3546  */
3547 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3548 {
3549         u32 epoch;
3550         u32 maplen;
3551         void *p = msg->front.iov_base;
3552         void *end = p + msg->front.iov_len;
3553         struct ceph_mdsmap *newmap, *oldmap;
3554         struct ceph_fsid fsid;
3555         int err = -EINVAL;
3556
3557         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3558         ceph_decode_copy(&p, &fsid, sizeof(fsid));
3559         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3560                 return;
3561         epoch = ceph_decode_32(&p);
3562         maplen = ceph_decode_32(&p);
3563         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3564
3565         /* do we need it? */
3566         ceph_monc_got_mdsmap(&mdsc->fsc->client->monc, epoch);
3567         mutex_lock(&mdsc->mutex);
3568         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3569                 dout("handle_map epoch %u <= our %u\n",
3570                      epoch, mdsc->mdsmap->m_epoch);
3571                 mutex_unlock(&mdsc->mutex);
3572                 return;
3573         }
3574
3575         newmap = ceph_mdsmap_decode(&p, end);
3576         if (IS_ERR(newmap)) {
3577                 err = PTR_ERR(newmap);
3578                 goto bad_unlock;
3579         }
3580
3581         /* swap into place */
3582         if (mdsc->mdsmap) {
3583                 oldmap = mdsc->mdsmap;
3584                 mdsc->mdsmap = newmap;
3585                 check_new_map(mdsc, newmap, oldmap);
3586                 ceph_mdsmap_destroy(oldmap);
3587         } else {
3588                 mdsc->mdsmap = newmap;  /* first mds map */
3589         }
3590         mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3591
3592         __wake_requests(mdsc, &mdsc->waiting_for_map);
3593
3594         mutex_unlock(&mdsc->mutex);
3595         schedule_delayed(mdsc);
3596         return;
3597
3598 bad_unlock:
3599         mutex_unlock(&mdsc->mutex);
3600 bad:
3601         pr_err("error decoding mdsmap %d\n", err);
3602         return;
3603 }
3604
3605 static struct ceph_connection *con_get(struct ceph_connection *con)
3606 {
3607         struct ceph_mds_session *s = con->private;
3608
3609         if (get_session(s)) {
3610                 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3611                 return con;
3612         }
3613         dout("mdsc con_get %p FAIL\n", s);
3614         return NULL;
3615 }
3616
3617 static void con_put(struct ceph_connection *con)
3618 {
3619         struct ceph_mds_session *s = con->private;
3620
3621         dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3622         ceph_put_mds_session(s);
3623 }
3624
3625 /*
3626  * if the client is unresponsive for long enough, the mds will kill
3627  * the session entirely.
3628  */
3629 static void peer_reset(struct ceph_connection *con)
3630 {
3631         struct ceph_mds_session *s = con->private;
3632         struct ceph_mds_client *mdsc = s->s_mdsc;
3633
3634         pr_warn("mds%d closed our session\n", s->s_mds);
3635         send_mds_reconnect(mdsc, s);
3636 }
3637
3638 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3639 {
3640         struct ceph_mds_session *s = con->private;
3641         struct ceph_mds_client *mdsc = s->s_mdsc;
3642         int type = le16_to_cpu(msg->hdr.type);
3643
3644         mutex_lock(&mdsc->mutex);
3645         if (__verify_registered_session(mdsc, s) < 0) {
3646                 mutex_unlock(&mdsc->mutex);
3647                 goto out;
3648         }
3649         mutex_unlock(&mdsc->mutex);
3650
3651         switch (type) {
3652         case CEPH_MSG_MDS_MAP:
3653                 ceph_mdsc_handle_map(mdsc, msg);
3654                 break;
3655         case CEPH_MSG_CLIENT_SESSION:
3656                 handle_session(s, msg);
3657                 break;
3658         case CEPH_MSG_CLIENT_REPLY:
3659                 handle_reply(s, msg);
3660                 break;
3661         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3662                 handle_forward(mdsc, s, msg);
3663                 break;
3664         case CEPH_MSG_CLIENT_CAPS:
3665                 ceph_handle_caps(s, msg);
3666                 break;
3667         case CEPH_MSG_CLIENT_SNAP:
3668                 ceph_handle_snap(mdsc, s, msg);
3669                 break;
3670         case CEPH_MSG_CLIENT_LEASE:
3671                 handle_lease(mdsc, s, msg);
3672                 break;
3673
3674         default:
3675                 pr_err("received unknown message type %d %s\n", type,
3676                        ceph_msg_type_name(type));
3677         }
3678 out:
3679         ceph_msg_put(msg);
3680 }
3681
3682 /*
3683  * authentication
3684  */
3685
3686 /*
3687  * Note: returned pointer is the address of a structure that's
3688  * managed separately.  Caller must *not* attempt to free it.
3689  */
3690 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
3691                                         int *proto, int force_new)
3692 {
3693         struct ceph_mds_session *s = con->private;
3694         struct ceph_mds_client *mdsc = s->s_mdsc;
3695         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3696         struct ceph_auth_handshake *auth = &s->s_auth;
3697
3698         if (force_new && auth->authorizer) {
3699                 ceph_auth_destroy_authorizer(ac, auth->authorizer);
3700                 auth->authorizer = NULL;
3701         }
3702         if (!auth->authorizer) {
3703                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3704                                                       auth);
3705                 if (ret)
3706                         return ERR_PTR(ret);
3707         } else {
3708                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3709                                                       auth);
3710                 if (ret)
3711                         return ERR_PTR(ret);
3712         }
3713         *proto = ac->protocol;
3714
3715         return auth;
3716 }
3717
3718
3719 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3720 {
3721         struct ceph_mds_session *s = con->private;
3722         struct ceph_mds_client *mdsc = s->s_mdsc;
3723         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3724
3725         return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer, len);
3726 }
3727
3728 static int invalidate_authorizer(struct ceph_connection *con)
3729 {
3730         struct ceph_mds_session *s = con->private;
3731         struct ceph_mds_client *mdsc = s->s_mdsc;
3732         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3733
3734         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3735
3736         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
3737 }
3738
3739 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
3740                                 struct ceph_msg_header *hdr, int *skip)
3741 {
3742         struct ceph_msg *msg;
3743         int type = (int) le16_to_cpu(hdr->type);
3744         int front_len = (int) le32_to_cpu(hdr->front_len);
3745
3746         if (con->in_msg)
3747                 return con->in_msg;
3748
3749         *skip = 0;
3750         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
3751         if (!msg) {
3752                 pr_err("unable to allocate msg type %d len %d\n",
3753                        type, front_len);
3754                 return NULL;
3755         }
3756
3757         return msg;
3758 }
3759
3760 static int sign_message(struct ceph_connection *con, struct ceph_msg *msg)
3761 {
3762        struct ceph_mds_session *s = con->private;
3763        struct ceph_auth_handshake *auth = &s->s_auth;
3764        return ceph_auth_sign_message(auth, msg);
3765 }
3766
3767 static int check_message_signature(struct ceph_connection *con, struct ceph_msg *msg)
3768 {
3769        struct ceph_mds_session *s = con->private;
3770        struct ceph_auth_handshake *auth = &s->s_auth;
3771        return ceph_auth_check_message_signature(auth, msg);
3772 }
3773
3774 static const struct ceph_connection_operations mds_con_ops = {
3775         .get = con_get,
3776         .put = con_put,
3777         .dispatch = dispatch,
3778         .get_authorizer = get_authorizer,
3779         .verify_authorizer_reply = verify_authorizer_reply,
3780         .invalidate_authorizer = invalidate_authorizer,
3781         .peer_reset = peer_reset,
3782         .alloc_msg = mds_alloc_msg,
3783         .sign_message = sign_message,
3784         .check_message_signature = check_message_signature,
3785 };
3786
3787 /* eof */