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  42  * are no longer valid.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