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  43  * are no longer valid.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