fs/ceph/mds_client.c

   1 #include <linux/ceph/ceph_debug.h>
   2
   3 #include <linux/fs.h>
   4 #include <linux/wait.h>
   5 #include <linux/slab.h>
   6 #include <linux/gfp.h>
   7 #include <linux/sched.h>
   8 #include <linux/debugfs.h>
   9 #include <linux/seq_file.h>
  10 #include <linux/utsname.h>
  11 #include <linux/ratelimit.h>
  12
  13 #include "super.h"
  14 #include "mds_client.h"
  15
  16 #include <linux/ceph/ceph_features.h>
  17 #include <linux/ceph/messenger.h>
  18 #include <linux/ceph/decode.h>
  19 #include <linux/ceph/pagelist.h>
  20 #include <linux/ceph/auth.h>
  21 #include <linux/ceph/debugfs.h>
  22
  23 /*
  24  * A cluster of MDS (metadata server) daemons is responsible for
  25  * managing the file system namespace (the directory hierarchy and
  26  * inodes) and for coordinating shared access to storage.  Metadata is
  27  * partitioning hierarchically across a number of servers, and that
  28  * partition varies over time as the cluster adjusts the distribution
  29  * in order to balance load.
  30  *
  31  * The MDS client is primarily responsible to managing synchronous
  32  * metadata requests for operations like open, unlink, and so forth.
  33  * If there is a MDS failure, we find out about it when we (possibly
  34  * request and) receive a new MDS map, and can resubmit affected
  35  * requests.
  36  *
  37  * For the most part, though, we take advantage of a lossless
  38  * communications channel to the MDS, and do not need to worry about
  39  * timing out or resubmitting requests.
  40  *
  41  * We maintain a stateful "session" with each MDS we interact with.
  42  * Within each session, we sent periodic heartbeat messages to ensure
  43  * any capabilities or leases we have been issues remain valid.  If
  44  * the session times out and goes stale, our leases and capabilities
  45  * are no longer valid.
  46  */
  47
  48 struct ceph_reconnect_state {
  49         int nr_caps;
  50         struct ceph_pagelist *pagelist;
  51         bool flock;
  52 };
  53
  54 static void __wake_requests(struct ceph_mds_client *mdsc,
  55                             struct list_head *head);
  56
  57 static const struct ceph_connection_operations mds_con_ops;
  58
  59
  60 /*
  61  * mds reply parsing
  62  */
  63
  64 /*
  65  * parse individual inode info
  66  */
  67 static int parse_reply_info_in(void **p, void *end,
  68                                struct ceph_mds_reply_info_in *info,
  69                                u64 features)
  70 {
  71         int err = -EIO;
  72
  73         info->in = *p;
  74         *p += sizeof(struct ceph_mds_reply_inode) +
  75                 sizeof(*info->in->fragtree.splits) *
  76                 le32_to_cpu(info->in->fragtree.nsplits);
  77
  78         ceph_decode_32_safe(p, end, info->symlink_len, bad);
  79         ceph_decode_need(p, end, info->symlink_len, bad);
  80         info->symlink = *p;
  81         *p += info->symlink_len;
  82
  83         if (features & CEPH_FEATURE_DIRLAYOUTHASH)
  84                 ceph_decode_copy_safe(p, end, &info->dir_layout,
  85                                       sizeof(info->dir_layout), bad);
  86         else
  87                 memset(&info->dir_layout, 0, sizeof(info->dir_layout));
  88
  89         ceph_decode_32_safe(p, end, info->xattr_len, bad);
  90         ceph_decode_need(p, end, info->xattr_len, bad);
  91         info->xattr_data = *p;
  92         *p += info->xattr_len;
  93
  94         if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
  95                 ceph_decode_64_safe(p, end, info->inline_version, bad);
  96                 ceph_decode_32_safe(p, end, info->inline_len, bad);
  97                 ceph_decode_need(p, end, info->inline_len, bad);
  98                 info->inline_data = *p;
  99                 *p += info->inline_len;
 100         } else
 101                 info->inline_version = CEPH_INLINE_NONE;
 102
 103         return 0;
 104 bad:
 105         return err;
 106 }
 107
 108 /*
 109  * parse a normal reply, which may contain a (dir+)dentry and/or a
 110  * target inode.
 111  */
 112 static int parse_reply_info_trace(void **p, void *end,
 113                                   struct ceph_mds_reply_info_parsed *info,
 114                                   u64 features)
 115 {
 116         int err;
 117
 118         if (info->head->is_dentry) {
 119                 err = parse_reply_info_in(p, end, &info->diri, features);
 120                 if (err < 0)
 121                         goto out_bad;
 122
 123                 if (unlikely(*p + sizeof(*info->dirfrag) > end))
 124                         goto bad;
 125                 info->dirfrag = *p;
 126                 *p += sizeof(*info->dirfrag) +
 127                         sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
 128                 if (unlikely(*p > end))
 129                         goto bad;
 130
 131                 ceph_decode_32_safe(p, end, info->dname_len, bad);
 132                 ceph_decode_need(p, end, info->dname_len, bad);
 133                 info->dname = *p;
 134                 *p += info->dname_len;
 135                 info->dlease = *p;
 136                 *p += sizeof(*info->dlease);
 137         }
 138
 139         if (info->head->is_target) {
 140                 err = parse_reply_info_in(p, end, &info->targeti, features);
 141                 if (err < 0)
 142                         goto out_bad;
 143         }
 144
 145         if (unlikely(*p != end))
 146                 goto bad;
 147         return 0;
 148
 149 bad:
 150         err = -EIO;
 151 out_bad:
 152         pr_err("problem parsing mds trace %d\n", err);
 153         return err;
 154 }
 155
 156 /*
 157  * parse readdir results
 158  */
 159 static int parse_reply_info_dir(void **p, void *end,
 160                                 struct ceph_mds_reply_info_parsed *info,
 161                                 u64 features)
 162 {
 163         u32 num, i = 0;
 164         int err;
 165
 166         info->dir_dir = *p;
 167         if (*p + sizeof(*info->dir_dir) > end)
 168                 goto bad;
 169         *p += sizeof(*info->dir_dir) +
 170                 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
 171         if (*p > end)
 172                 goto bad;
 173
 174         ceph_decode_need(p, end, sizeof(num) + 2, bad);
 175         num = ceph_decode_32(p);
 176         info->dir_end = ceph_decode_8(p);
 177         info->dir_complete = ceph_decode_8(p);
 178         if (num == 0)
 179                 goto done;
 180
 181         BUG_ON(!info->dir_in);
 182         info->dir_dname = (void *)(info->dir_in + num);
 183         info->dir_dname_len = (void *)(info->dir_dname + num);
 184         info->dir_dlease = (void *)(info->dir_dname_len + num);
 185         if ((unsigned long)(info->dir_dlease + num) >
 186             (unsigned long)info->dir_in + info->dir_buf_size) {
 187                 pr_err("dir contents are larger than expected\n");
 188                 WARN_ON(1);
 189                 goto bad;
 190         }
 191
 192         info->dir_nr = num;
 193         while (num) {
 194                 /* dentry */
 195                 ceph_decode_need(p, end, sizeof(u32)*2, bad);
 196                 info->dir_dname_len[i] = ceph_decode_32(p);
 197                 ceph_decode_need(p, end, info->dir_dname_len[i], bad);
 198                 info->dir_dname[i] = *p;
 199                 *p += info->dir_dname_len[i];
 200                 dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
 201                      info->dir_dname[i]);
 202                 info->dir_dlease[i] = *p;
 203                 *p += sizeof(struct ceph_mds_reply_lease);
 204
 205                 /* inode */
 206                 err = parse_reply_info_in(p, end, &info->dir_in[i], features);
 207                 if (err < 0)
 208                         goto out_bad;
 209                 i++;
 210                 num--;
 211         }
 212
 213 done:
 214         if (*p != end)
 215                 goto bad;
 216         return 0;
 217
 218 bad:
 219         err = -EIO;
 220 out_bad:
 221         pr_err("problem parsing dir contents %d\n", err);
 222         return err;
 223 }
 224
 225 /*
 226  * parse fcntl F_GETLK results
 227  */
 228 static int parse_reply_info_filelock(void **p, void *end,
 229                                      struct ceph_mds_reply_info_parsed *info,
 230                                      u64 features)
 231 {
 232         if (*p + sizeof(*info->filelock_reply) > end)
 233                 goto bad;
 234
 235         info->filelock_reply = *p;
 236         *p += sizeof(*info->filelock_reply);
 237
 238         if (unlikely(*p != end))
 239                 goto bad;
 240         return 0;
 241
 242 bad:
 243         return -EIO;
 244 }
 245
 246 /*
 247  * parse create results
 248  */
 249 static int parse_reply_info_create(void **p, void *end,
 250                                   struct ceph_mds_reply_info_parsed *info,
 251                                   u64 features)
 252 {
 253         if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
 254                 if (*p == end) {
 255                         info->has_create_ino = false;
 256                 } else {
 257                         info->has_create_ino = true;
 258                         info->ino = ceph_decode_64(p);
 259                 }
 260         }
 261
 262         if (unlikely(*p != end))
 263                 goto bad;
 264         return 0;
 265
 266 bad:
 267         return -EIO;
 268 }
 269
 270 /*
 271  * parse extra results
 272  */
 273 static int parse_reply_info_extra(void **p, void *end,
 274                                   struct ceph_mds_reply_info_parsed *info,
 275                                   u64 features)
 276 {
 277         u32 op = le32_to_cpu(info->head->op);
 278
 279         if (op == CEPH_MDS_OP_GETFILELOCK)
 280                 return parse_reply_info_filelock(p, end, info, features);
 281         else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
 282                 return parse_reply_info_dir(p, end, info, features);
 283         else if (op == CEPH_MDS_OP_CREATE)
 284                 return parse_reply_info_create(p, end, info, features);
 285         else
 286                 return -EIO;
 287 }
 288
 289 /*
 290  * parse entire mds reply
 291  */
 292 static int parse_reply_info(struct ceph_msg *msg,
 293                             struct ceph_mds_reply_info_parsed *info,
 294                             u64 features)
 295 {
 296         void *p, *end;
 297         u32 len;
 298         int err;
 299
 300         info->head = msg->front.iov_base;
 301         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 302         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 303
 304         /* trace */
 305         ceph_decode_32_safe(&p, end, len, bad);
 306         if (len > 0) {
 307                 ceph_decode_need(&p, end, len, bad);
 308                 err = parse_reply_info_trace(&p, p+len, info, features);
 309                 if (err < 0)
 310                         goto out_bad;
 311         }
 312
 313         /* extra */
 314         ceph_decode_32_safe(&p, end, len, bad);
 315         if (len > 0) {
 316                 ceph_decode_need(&p, end, len, bad);
 317                 err = parse_reply_info_extra(&p, p+len, info, features);
 318                 if (err < 0)
 319                         goto out_bad;
 320         }
 321
 322         /* snap blob */
 323         ceph_decode_32_safe(&p, end, len, bad);
 324         info->snapblob_len = len;
 325         info->snapblob = p;
 326         p += len;
 327
 328         if (p != end)
 329                 goto bad;
 330         return 0;
 331
 332 bad:
 333         err = -EIO;
 334 out_bad:
 335         pr_err("mds parse_reply err %d\n", err);
 336         return err;
 337 }
 338
 339 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
 340 {
 341         if (!info->dir_in)
 342                 return;
 343         free_pages((unsigned long)info->dir_in, get_order(info->dir_buf_size));
 344 }
 345
 346
 347 /*
 348  * sessions
 349  */
 350 const char *ceph_session_state_name(int s)
 351 {
 352         switch (s) {
 353         case CEPH_MDS_SESSION_NEW: return "new";
 354         case CEPH_MDS_SESSION_OPENING: return "opening";
 355         case CEPH_MDS_SESSION_OPEN: return "open";
 356         case CEPH_MDS_SESSION_HUNG: return "hung";
 357         case CEPH_MDS_SESSION_CLOSING: return "closing";
 358         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
 359         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
 360         default: return "???";
 361         }
 362 }
 363
 364 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
 365 {
 366         if (atomic_inc_not_zero(&s->s_ref)) {
 367                 dout("mdsc get_session %p %d -> %d\n", s,
 368                      atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
 369                 return s;
 370         } else {
 371                 dout("mdsc get_session %p 0 -- FAIL", s);
 372                 return NULL;
 373         }
 374 }
 375
 376 void ceph_put_mds_session(struct ceph_mds_session *s)
 377 {
 378         dout("mdsc put_session %p %d -> %d\n", s,
 379              atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
 380         if (atomic_dec_and_test(&s->s_ref)) {
 381                 if (s->s_auth.authorizer)
 382                         ceph_auth_destroy_authorizer(
 383                                 s->s_mdsc->fsc->client->monc.auth,
 384                                 s->s_auth.authorizer);
 385                 kfree(s);
 386         }
 387 }
 388
 389 /*
 390  * called under mdsc->mutex
 391  */
 392 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
 393                                                    int mds)
 394 {
 395         struct ceph_mds_session *session;
 396
 397         if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
 398                 return NULL;
 399         session = mdsc->sessions[mds];
 400         dout("lookup_mds_session %p %d\n", session,
 401              atomic_read(&session->s_ref));
 402         get_session(session);
 403         return session;
 404 }
 405
 406 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
 407 {
 408         if (mds >= mdsc->max_sessions)
 409                 return false;
 410         return mdsc->sessions[mds];
 411 }
 412
 413 static int __verify_registered_session(struct ceph_mds_client *mdsc,
 414                                        struct ceph_mds_session *s)
 415 {
 416         if (s->s_mds >= mdsc->max_sessions ||
 417             mdsc->sessions[s->s_mds] != s)
 418                 return -ENOENT;
 419         return 0;
 420 }
 421
 422 /*
 423  * create+register a new session for given mds.
 424  * called under mdsc->mutex.
 425  */
 426 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 427                                                  int mds)
 428 {
 429         struct ceph_mds_session *s;
 430
 431         if (mds >= mdsc->mdsmap->m_max_mds)
 432                 return ERR_PTR(-EINVAL);
 433
 434         s = kzalloc(sizeof(*s), GFP_NOFS);
 435         if (!s)
 436                 return ERR_PTR(-ENOMEM);
 437         s->s_mdsc = mdsc;
 438         s->s_mds = mds;
 439         s->s_state = CEPH_MDS_SESSION_NEW;
 440         s->s_ttl = 0;
 441         s->s_seq = 0;
 442         mutex_init(&s->s_mutex);
 443
 444         ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
 445
 446         spin_lock_init(&s->s_gen_ttl_lock);
 447         s->s_cap_gen = 0;
 448         s->s_cap_ttl = jiffies - 1;
 449
 450         spin_lock_init(&s->s_cap_lock);
 451         s->s_renew_requested = 0;
 452         s->s_renew_seq = 0;
 453         INIT_LIST_HEAD(&s->s_caps);
 454         s->s_nr_caps = 0;
 455         s->s_trim_caps = 0;
 456         atomic_set(&s->s_ref, 1);
 457         INIT_LIST_HEAD(&s->s_waiting);
 458         INIT_LIST_HEAD(&s->s_unsafe);
 459         s->s_num_cap_releases = 0;
 460         s->s_cap_reconnect = 0;
 461         s->s_cap_iterator = NULL;
 462         INIT_LIST_HEAD(&s->s_cap_releases);
 463         INIT_LIST_HEAD(&s->s_cap_flushing);
 464         INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
 465
 466         dout("register_session mds%d\n", mds);
 467         if (mds >= mdsc->max_sessions) {
 468                 int newmax = 1 << get_count_order(mds+1);
 469                 struct ceph_mds_session **sa;
 470
 471                 dout("register_session realloc to %d\n", newmax);
 472                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 473                 if (sa == NULL)
 474                         goto fail_realloc;
 475                 if (mdsc->sessions) {
 476                         memcpy(sa, mdsc->sessions,
 477                                mdsc->max_sessions * sizeof(void *));
 478                         kfree(mdsc->sessions);
 479                 }
 480                 mdsc->sessions = sa;
 481                 mdsc->max_sessions = newmax;
 482         }
 483         mdsc->sessions[mds] = s;
 484         atomic_inc(&mdsc->num_sessions);
 485         atomic_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
 486
 487         ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
 488                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
 489
 490         return s;
 491
 492 fail_realloc:
 493         kfree(s);
 494         return ERR_PTR(-ENOMEM);
 495 }
 496
 497 /*
 498  * called under mdsc->mutex
 499  */
 500 static void __unregister_session(struct ceph_mds_client *mdsc,
 501                                struct ceph_mds_session *s)
 502 {
 503         dout("__unregister_session mds%d %p\n", s->s_mds, s);
 504         BUG_ON(mdsc->sessions[s->s_mds] != s);
 505         mdsc->sessions[s->s_mds] = NULL;
 506         ceph_con_close(&s->s_con);
 507         ceph_put_mds_session(s);
 508         atomic_dec(&mdsc->num_sessions);
 509 }
 510
 511 /*
 512  * drop session refs in request.
 513  *
 514  * should be last request ref, or hold mdsc->mutex
 515  */
 516 static void put_request_session(struct ceph_mds_request *req)
 517 {
 518         if (req->r_session) {
 519                 ceph_put_mds_session(req->r_session);
 520                 req->r_session = NULL;
 521         }
 522 }
 523
 524 void ceph_mdsc_release_request(struct kref *kref)
 525 {
 526         struct ceph_mds_request *req = container_of(kref,
 527                                                     struct ceph_mds_request,
 528                                                     r_kref);
 529         destroy_reply_info(&req->r_reply_info);
 530         if (req->r_request)
 531                 ceph_msg_put(req->r_request);
 532         if (req->r_reply)
 533                 ceph_msg_put(req->r_reply);
 534         if (req->r_inode) {
 535                 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
 536                 iput(req->r_inode);
 537         }
 538         if (req->r_locked_dir)
 539                 ceph_put_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
 540         iput(req->r_target_inode);
 541         if (req->r_dentry)
 542                 dput(req->r_dentry);
 543         if (req->r_old_dentry)
 544                 dput(req->r_old_dentry);
 545         if (req->r_old_dentry_dir) {
 546                 /*
 547                  * track (and drop pins for) r_old_dentry_dir
 548                  * separately, since r_old_dentry's d_parent may have
 549                  * changed between the dir mutex being dropped and
 550                  * this request being freed.
 551                  */
 552                 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
 553                                   CEPH_CAP_PIN);
 554                 iput(req->r_old_dentry_dir);
 555         }
 556         kfree(req->r_path1);
 557         kfree(req->r_path2);
 558         if (req->r_pagelist)
 559                 ceph_pagelist_release(req->r_pagelist);
 560         put_request_session(req);
 561         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
 562         kfree(req);
 563 }
 564
 565 /*
 566  * lookup session, bump ref if found.
 567  *
 568  * called under mdsc->mutex.
 569  */
 570 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
 571                                              u64 tid)
 572 {
 573         struct ceph_mds_request *req;
 574         struct rb_node *n = mdsc->request_tree.rb_node;
 575
 576         while (n) {
 577                 req = rb_entry(n, struct ceph_mds_request, r_node);
 578                 if (tid < req->r_tid)
 579                         n = n->rb_left;
 580                 else if (tid > req->r_tid)
 581                         n = n->rb_right;
 582                 else {
 583                         ceph_mdsc_get_request(req);
 584                         return req;
 585                 }
 586         }
 587         return NULL;
 588 }
 589
 590 static void __insert_request(struct ceph_mds_client *mdsc,
 591                              struct ceph_mds_request *new)
 592 {
 593         struct rb_node **p = &mdsc->request_tree.rb_node;
 594         struct rb_node *parent = NULL;
 595         struct ceph_mds_request *req = NULL;
 596
 597         while (*p) {
 598                 parent = *p;
 599                 req = rb_entry(parent, struct ceph_mds_request, r_node);
 600                 if (new->r_tid < req->r_tid)
 601                         p = &(*p)->rb_left;
 602                 else if (new->r_tid > req->r_tid)
 603                         p = &(*p)->rb_right;
 604                 else
 605                         BUG();
 606         }
 607
 608         rb_link_node(&new->r_node, parent, p);
 609         rb_insert_color(&new->r_node, &mdsc->request_tree);
 610 }
 611
 612 /*
 613  * Register an in-flight request, and assign a tid.  Link to directory
 614  * are modifying (if any).
 615  *
 616  * Called under mdsc->mutex.
 617  */
 618 static void __register_request(struct ceph_mds_client *mdsc,
 619                                struct ceph_mds_request *req,
 620                                struct inode *dir)
 621 {
 622         req->r_tid = ++mdsc->last_tid;
 623         if (req->r_num_caps)
 624                 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
 625                                   req->r_num_caps);
 626         dout("__register_request %p tid %lld\n", req, req->r_tid);
 627         ceph_mdsc_get_request(req);
 628         __insert_request(mdsc, req);
 629
 630         req->r_uid = current_fsuid();
 631         req->r_gid = current_fsgid();
 632
 633         if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
 634                 mdsc->oldest_tid = req->r_tid;
 635
 636         if (dir) {
 637                 ihold(dir);
 638                 req->r_unsafe_dir = dir;
 639         }
 640 }
 641
 642 static void __unregister_request(struct ceph_mds_client *mdsc,
 643                                  struct ceph_mds_request *req)
 644 {
 645         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 646
 647         /* Never leave an unregistered request on an unsafe list! */
 648         list_del_init(&req->r_unsafe_item);
 649
 650         if (req->r_tid == mdsc->oldest_tid) {
 651                 struct rb_node *p = rb_next(&req->r_node);
 652                 mdsc->oldest_tid = 0;
 653                 while (p) {
 654                         struct ceph_mds_request *next_req =
 655                                 rb_entry(p, struct ceph_mds_request, r_node);
 656                         if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
 657                                 mdsc->oldest_tid = next_req->r_tid;
 658                                 break;
 659                         }
 660                         p = rb_next(p);
 661                 }
 662         }
 663
 664         rb_erase(&req->r_node, &mdsc->request_tree);
 665         RB_CLEAR_NODE(&req->r_node);
 666
 667         if (req->r_unsafe_dir && req->r_got_unsafe) {
 668                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 669                 spin_lock(&ci->i_unsafe_lock);
 670                 list_del_init(&req->r_unsafe_dir_item);
 671                 spin_unlock(&ci->i_unsafe_lock);
 672         }
 673         if (req->r_target_inode && req->r_got_unsafe) {
 674                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
 675                 spin_lock(&ci->i_unsafe_lock);
 676                 list_del_init(&req->r_unsafe_target_item);
 677                 spin_unlock(&ci->i_unsafe_lock);
 678         }
 679
 680         if (req->r_unsafe_dir) {
 681                 iput(req->r_unsafe_dir);
 682                 req->r_unsafe_dir = NULL;
 683         }
 684
 685         complete_all(&req->r_safe_completion);
 686
 687         ceph_mdsc_put_request(req);
 688 }
 689
 690 /*
 691  * Choose mds to send request to next.  If there is a hint set in the
 692  * request (e.g., due to a prior forward hint from the mds), use that.
 693  * Otherwise, consult frag tree and/or caps to identify the
 694  * appropriate mds.  If all else fails, choose randomly.
 695  *
 696  * Called under mdsc->mutex.
 697  */
 698 static struct dentry *get_nonsnap_parent(struct dentry *dentry)
 699 {
 700         /*
 701          * we don't need to worry about protecting the d_parent access
 702          * here because we never renaming inside the snapped namespace
 703          * except to resplice to another snapdir, and either the old or new
 704          * result is a valid result.
 705          */
 706         while (!IS_ROOT(dentry) && ceph_snap(d_inode(dentry)) != CEPH_NOSNAP)
 707                 dentry = dentry->d_parent;
 708         return dentry;
 709 }
 710
 711 static int __choose_mds(struct ceph_mds_client *mdsc,
 712                         struct ceph_mds_request *req)
 713 {
 714         struct inode *inode;
 715         struct ceph_inode_info *ci;
 716         struct ceph_cap *cap;
 717         int mode = req->r_direct_mode;
 718         int mds = -1;
 719         u32 hash = req->r_direct_hash;
 720         bool is_hash = req->r_direct_is_hash;
 721
 722         /*
 723          * is there a specific mds we should try?  ignore hint if we have
 724          * no session and the mds is not up (active or recovering).
 725          */
 726         if (req->r_resend_mds >= 0 &&
 727             (__have_session(mdsc, req->r_resend_mds) ||
 728              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 729                 dout("choose_mds using resend_mds mds%d\n",
 730                      req->r_resend_mds);
 731                 return req->r_resend_mds;
 732         }
 733
 734         if (mode == USE_RANDOM_MDS)
 735                 goto random;
 736
 737         inode = NULL;
 738         if (req->r_inode) {
 739                 inode = req->r_inode;
 740         } else if (req->r_dentry) {
 741                 /* ignore race with rename; old or new d_parent is okay */
 742                 struct dentry *parent = req->r_dentry->d_parent;
 743                 struct inode *dir = d_inode(parent);
 744
 745                 if (dir->i_sb != mdsc->fsc->sb) {
 746                         /* not this fs! */
 747                         inode = d_inode(req->r_dentry);
 748                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
 749                         /* direct snapped/virtual snapdir requests
 750                          * based on parent dir inode */
 751                         struct dentry *dn = get_nonsnap_parent(parent);
 752                         inode = d_inode(dn);
 753                         dout("__choose_mds using nonsnap parent %p\n", inode);
 754                 } else {
 755                         /* dentry target */
 756                         inode = d_inode(req->r_dentry);
 757                         if (!inode || mode == USE_AUTH_MDS) {
 758                                 /* dir + name */
 759                                 inode = dir;
 760                                 hash = ceph_dentry_hash(dir, req->r_dentry);
 761                                 is_hash = true;
 762                         }
 763                 }
 764         }
 765
 766         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 767              (int)hash, mode);
 768         if (!inode)
 769                 goto random;
 770         ci = ceph_inode(inode);
 771
 772         if (is_hash && S_ISDIR(inode->i_mode)) {
 773                 struct ceph_inode_frag frag;
 774                 int found;
 775
 776                 ceph_choose_frag(ci, hash, &frag, &found);
 777                 if (found) {
 778                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 779                                 u8 r;
 780
 781                                 /* choose a random replica */
 782                                 get_random_bytes(&r, 1);
 783                                 r %= frag.ndist;
 784                                 mds = frag.dist[r];
 785                                 dout("choose_mds %p %llx.%llx "
 786                                      "frag %u mds%d (%d/%d)\n",
 787                                      inode, ceph_vinop(inode),
 788                                      frag.frag, mds,
 789                                      (int)r, frag.ndist);
 790                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 791                                     CEPH_MDS_STATE_ACTIVE)
 792                                         return mds;
 793                         }
 794
 795                         /* since this file/dir wasn't known to be
 796                          * replicated, then we want to look for the
 797                          * authoritative mds. */
 798                         mode = USE_AUTH_MDS;
 799                         if (frag.mds >= 0) {
 800                                 /* choose auth mds */
 801                                 mds = frag.mds;
 802                                 dout("choose_mds %p %llx.%llx "
 803                                      "frag %u mds%d (auth)\n",
 804                                      inode, ceph_vinop(inode), frag.frag, mds);
 805                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 806                                     CEPH_MDS_STATE_ACTIVE)
 807                                         return mds;
 808                         }
 809                 }
 810         }
 811
 812         spin_lock(&ci->i_ceph_lock);
 813         cap = NULL;
 814         if (mode == USE_AUTH_MDS)
 815                 cap = ci->i_auth_cap;
 816         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
 817                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
 818         if (!cap) {
 819                 spin_unlock(&ci->i_ceph_lock);
 820                 goto random;
 821         }
 822         mds = cap->session->s_mds;
 823         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
 824              inode, ceph_vinop(inode), mds,
 825              cap == ci->i_auth_cap ? "auth " : "", cap);
 826         spin_unlock(&ci->i_ceph_lock);
 827         return mds;
 828
 829 random:
 830         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
 831         dout("choose_mds chose random mds%d\n", mds);
 832         return mds;
 833 }
 834
 835
 836 /*
 837  * session messages
 838  */
 839 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
 840 {
 841         struct ceph_msg *msg;
 842         struct ceph_mds_session_head *h;
 843
 844         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
 845                            false);
 846         if (!msg) {
 847                 pr_err("create_session_msg ENOMEM creating msg\n");
 848                 return NULL;
 849         }
 850         h = msg->front.iov_base;
 851         h->op = cpu_to_le32(op);
 852         h->seq = cpu_to_le64(seq);
 853
 854         return msg;
 855 }
 856
 857 /*
 858  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
 859  * to include additional client metadata fields.
 860  */
 861 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
 862 {
 863         struct ceph_msg *msg;
 864         struct ceph_mds_session_head *h;
 865         int i = -1;
 866         int metadata_bytes = 0;
 867         int metadata_key_count = 0;
 868         struct ceph_options *opt = mdsc->fsc->client->options;
 869         void *p;
 870
 871         const char* metadata[][2] = {
 872                 {"hostname", utsname()->nodename},
 873                 {"kernel_version", utsname()->release},
 874                 {"entity_id", opt->name ? opt->name : ""},
 875                 {NULL, NULL}
 876         };
 877
 878         /* Calculate serialized length of metadata */
 879         metadata_bytes = 4;  /* map length */
 880         for (i = 0; metadata[i][0] != NULL; ++i) {
 881                 metadata_bytes += 8 + strlen(metadata[i][0]) +
 882                         strlen(metadata[i][1]);
 883                 metadata_key_count++;
 884         }
 885
 886         /* Allocate the message */
 887         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + metadata_bytes,
 888                            GFP_NOFS, false);
 889         if (!msg) {
 890                 pr_err("create_session_msg ENOMEM creating msg\n");
 891                 return NULL;
 892         }
 893         h = msg->front.iov_base;
 894         h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
 895         h->seq = cpu_to_le64(seq);
 896
 897         /*
 898          * Serialize client metadata into waiting buffer space, using
 899          * the format that userspace expects for map<string, string>
 900          *
 901          * ClientSession messages with metadata are v2
 902          */
 903         msg->hdr.version = cpu_to_le16(2);
 904         msg->hdr.compat_version = cpu_to_le16(1);
 905
 906         /* The write pointer, following the session_head structure */
 907         p = msg->front.iov_base + sizeof(*h);
 908
 909         /* Number of entries in the map */
 910         ceph_encode_32(&p, metadata_key_count);
 911
 912         /* Two length-prefixed strings for each entry in the map */
 913         for (i = 0; metadata[i][0] != NULL; ++i) {
 914                 size_t const key_len = strlen(metadata[i][0]);
 915                 size_t const val_len = strlen(metadata[i][1]);
 916
 917                 ceph_encode_32(&p, key_len);
 918                 memcpy(p, metadata[i][0], key_len);
 919                 p += key_len;
 920                 ceph_encode_32(&p, val_len);
 921                 memcpy(p, metadata[i][1], val_len);
 922                 p += val_len;
 923         }
 924
 925         return msg;
 926 }
 927
 928 /*
 929  * send session open request.
 930  *
 931  * called under mdsc->mutex
 932  */
 933 static int __open_session(struct ceph_mds_client *mdsc,
 934                           struct ceph_mds_session *session)
 935 {
 936         struct ceph_msg *msg;
 937         int mstate;
 938         int mds = session->s_mds;
 939
 940         /* wait for mds to go active? */
 941         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
 942         dout("open_session to mds%d (%s)\n", mds,
 943              ceph_mds_state_name(mstate));
 944         session->s_state = CEPH_MDS_SESSION_OPENING;
 945         session->s_renew_requested = jiffies;
 946
 947         /* send connect message */
 948         msg = create_session_open_msg(mdsc, session->s_seq);
 949         if (!msg)
 950                 return -ENOMEM;
 951         ceph_con_send(&session->s_con, msg);
 952         return 0;
 953 }
 954
 955 /*
 956  * open sessions for any export targets for the given mds
 957  *
 958  * called under mdsc->mutex
 959  */
 960 static struct ceph_mds_session *
 961 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
 962 {
 963         struct ceph_mds_session *session;
 964
 965         session = __ceph_lookup_mds_session(mdsc, target);
 966         if (!session) {
 967                 session = register_session(mdsc, target);
 968                 if (IS_ERR(session))
 969                         return session;
 970         }
 971         if (session->s_state == CEPH_MDS_SESSION_NEW ||
 972             session->s_state == CEPH_MDS_SESSION_CLOSING)
 973                 __open_session(mdsc, session);
 974
 975         return session;
 976 }
 977
 978 struct ceph_mds_session *
 979 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
 980 {
 981         struct ceph_mds_session *session;
 982
 983         dout("open_export_target_session to mds%d\n", target);
 984
 985         mutex_lock(&mdsc->mutex);
 986         session = __open_export_target_session(mdsc, target);
 987         mutex_unlock(&mdsc->mutex);
 988
 989         return session;
 990 }
 991
 992 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
 993                                           struct ceph_mds_session *session)
 994 {
 995         struct ceph_mds_info *mi;
 996         struct ceph_mds_session *ts;
 997         int i, mds = session->s_mds;
 998
 999         if (mds >= mdsc->mdsmap->m_max_mds)
1000                 return;
1001
1002         mi = &mdsc->mdsmap->m_info[mds];
1003         dout("open_export_target_sessions for mds%d (%d targets)\n",
1004              session->s_mds, mi->num_export_targets);
1005
1006         for (i = 0; i < mi->num_export_targets; i++) {
1007                 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1008                 if (!IS_ERR(ts))
1009                         ceph_put_mds_session(ts);
1010         }
1011 }
1012
1013 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1014                                            struct ceph_mds_session *session)
1015 {
1016         mutex_lock(&mdsc->mutex);
1017         __open_export_target_sessions(mdsc, session);
1018         mutex_unlock(&mdsc->mutex);
1019 }
1020
1021 /*
1022  * session caps
1023  */
1024
1025 /* caller holds s_cap_lock, we drop it */
1026 static void cleanup_cap_releases(struct ceph_mds_client *mdsc,
1027                                  struct ceph_mds_session *session)
1028         __releases(session->s_cap_lock)
1029 {
1030         LIST_HEAD(tmp_list);
1031         list_splice_init(&session->s_cap_releases, &tmp_list);
1032         session->s_num_cap_releases = 0;
1033         spin_unlock(&session->s_cap_lock);
1034
1035         dout("cleanup_cap_releases mds%d\n", session->s_mds);
1036         while (!list_empty(&tmp_list)) {
1037                 struct ceph_cap *cap;
1038                 /* zero out the in-progress message */
1039                 cap = list_first_entry(&tmp_list,
1040                                         struct ceph_cap, session_caps);
1041                 list_del(&cap->session_caps);
1042                 ceph_put_cap(mdsc, cap);
1043         }
1044 }
1045
1046 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1047                                      struct ceph_mds_session *session)
1048 {
1049         struct ceph_mds_request *req;
1050         struct rb_node *p;
1051
1052         dout("cleanup_session_requests mds%d\n", session->s_mds);
1053         mutex_lock(&mdsc->mutex);
1054         while (!list_empty(&session->s_unsafe)) {
1055                 req = list_first_entry(&session->s_unsafe,
1056                                        struct ceph_mds_request, r_unsafe_item);
1057                 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1058                                     req->r_tid);
1059                 __unregister_request(mdsc, req);
1060         }
1061         /* zero r_attempts, so kick_requests() will re-send requests */
1062         p = rb_first(&mdsc->request_tree);
1063         while (p) {
1064                 req = rb_entry(p, struct ceph_mds_request, r_node);
1065                 p = rb_next(p);
1066                 if (req->r_session &&
1067                     req->r_session->s_mds == session->s_mds)
1068                         req->r_attempts = 0;
1069         }
1070         mutex_unlock(&mdsc->mutex);
1071 }
1072
1073 /*
1074  * Helper to safely iterate over all caps associated with a session, with
1075  * special care taken to handle a racing __ceph_remove_cap().
1076  *
1077  * Caller must hold session s_mutex.
1078  */
1079 static int iterate_session_caps(struct ceph_mds_session *session,
1080                                  int (*cb)(struct inode *, struct ceph_cap *,
1081                                             void *), void *arg)
1082 {
1083         struct list_head *p;
1084         struct ceph_cap *cap;
1085         struct inode *inode, *last_inode = NULL;
1086         struct ceph_cap *old_cap = NULL;
1087         int ret;
1088
1089         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1090         spin_lock(&session->s_cap_lock);
1091         p = session->s_caps.next;
1092         while (p != &session->s_caps) {
1093                 cap = list_entry(p, struct ceph_cap, session_caps);
1094                 inode = igrab(&cap->ci->vfs_inode);
1095                 if (!inode) {
1096                         p = p->next;
1097                         continue;
1098                 }
1099                 session->s_cap_iterator = cap;
1100                 spin_unlock(&session->s_cap_lock);
1101
1102                 if (last_inode) {
1103                         iput(last_inode);
1104                         last_inode = NULL;
1105                 }
1106                 if (old_cap) {
1107                         ceph_put_cap(session->s_mdsc, old_cap);
1108                         old_cap = NULL;
1109                 }
1110
1111                 ret = cb(inode, cap, arg);
1112                 last_inode = inode;
1113
1114                 spin_lock(&session->s_cap_lock);
1115                 p = p->next;
1116                 if (cap->ci == NULL) {
1117                         dout("iterate_session_caps  finishing cap %p removal\n",
1118                              cap);
1119                         BUG_ON(cap->session != session);
1120                         cap->session = NULL;
1121                         list_del_init(&cap->session_caps);
1122                         session->s_nr_caps--;
1123                         if (cap->queue_release) {
1124                                 list_add_tail(&cap->session_caps,
1125                                               &session->s_cap_releases);
1126                                 session->s_num_cap_releases++;
1127                         } else {
1128                                 old_cap = cap;  /* put_cap it w/o locks held */
1129                         }
1130                 }
1131                 if (ret < 0)
1132                         goto out;
1133         }
1134         ret = 0;
1135 out:
1136         session->s_cap_iterator = NULL;
1137         spin_unlock(&session->s_cap_lock);
1138
1139         iput(last_inode);
1140         if (old_cap)
1141                 ceph_put_cap(session->s_mdsc, old_cap);
1142
1143         return ret;
1144 }
1145
1146 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1147                                   void *arg)
1148 {
1149         struct ceph_inode_info *ci = ceph_inode(inode);
1150         LIST_HEAD(to_remove);
1151         int drop = 0;
1152
1153         dout("removing cap %p, ci is %p, inode is %p\n",
1154              cap, ci, &ci->vfs_inode);
1155         spin_lock(&ci->i_ceph_lock);
1156         __ceph_remove_cap(cap, false);
1157         if (!ci->i_auth_cap) {
1158                 struct ceph_cap_flush *cf;
1159                 struct ceph_mds_client *mdsc =
1160                         ceph_sb_to_client(inode->i_sb)->mdsc;
1161
1162                 while (true) {
1163                         struct rb_node *n = rb_first(&ci->i_cap_flush_tree);
1164                         if (!n)
1165                                 break;
1166                         cf = rb_entry(n, struct ceph_cap_flush, i_node);
1167                         rb_erase(&cf->i_node, &ci->i_cap_flush_tree);
1168                         list_add(&cf->list, &to_remove);
1169                 }
1170
1171                 spin_lock(&mdsc->cap_dirty_lock);
1172
1173                 list_for_each_entry(cf, &to_remove, list)
1174                         rb_erase(&cf->g_node, &mdsc->cap_flush_tree);
1175
1176                 if (!list_empty(&ci->i_dirty_item)) {
1177                         pr_warn_ratelimited(
1178                                 " dropping dirty %s state for %p %lld\n",
1179                                 ceph_cap_string(ci->i_dirty_caps),
1180                                 inode, ceph_ino(inode));
1181                         ci->i_dirty_caps = 0;
1182                         list_del_init(&ci->i_dirty_item);
1183                         drop = 1;
1184                 }
1185                 if (!list_empty(&ci->i_flushing_item)) {
1186                         pr_warn_ratelimited(
1187                                 " dropping dirty+flushing %s state for %p %lld\n",
1188                                 ceph_cap_string(ci->i_flushing_caps),
1189                                 inode, ceph_ino(inode));
1190                         ci->i_flushing_caps = 0;
1191                         list_del_init(&ci->i_flushing_item);
1192                         mdsc->num_cap_flushing--;
1193                         drop = 1;
1194                 }
1195                 spin_unlock(&mdsc->cap_dirty_lock);
1196
1197                 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1198                         list_add(&ci->i_prealloc_cap_flush->list, &to_remove);
1199                         ci->i_prealloc_cap_flush = NULL;
1200                 }
1201
1202                if (drop &&
1203                   ci->i_wrbuffer_ref_head == 0 &&
1204                   ci->i_wr_ref == 0 &&
1205                   ci->i_dirty_caps == 0 &&
1206                   ci->i_flushing_caps == 0) {
1207                       ceph_put_snap_context(ci->i_head_snapc);
1208                       ci->i_head_snapc = NULL;
1209                }
1210         }
1211         spin_unlock(&ci->i_ceph_lock);
1212         while (!list_empty(&to_remove)) {
1213                 struct ceph_cap_flush *cf;
1214                 cf = list_first_entry(&to_remove,
1215                                       struct ceph_cap_flush, list);
1216                 list_del(&cf->list);
1217                 ceph_free_cap_flush(cf);
1218         }
1219         while (drop--)
1220                 iput(inode);
1221         return 0;
1222 }
1223
1224 /*
1225  * caller must hold session s_mutex
1226  */
1227 static void remove_session_caps(struct ceph_mds_session *session)
1228 {
1229         dout("remove_session_caps on %p\n", session);
1230         iterate_session_caps(session, remove_session_caps_cb, NULL);
1231
1232         spin_lock(&session->s_cap_lock);
1233         if (session->s_nr_caps > 0) {
1234                 struct super_block *sb = session->s_mdsc->fsc->sb;
1235                 struct inode *inode;
1236                 struct ceph_cap *cap, *prev = NULL;
1237                 struct ceph_vino vino;
1238                 /*
1239                  * iterate_session_caps() skips inodes that are being
1240                  * deleted, we need to wait until deletions are complete.
1241                  * __wait_on_freeing_inode() is designed for the job,
1242                  * but it is not exported, so use lookup inode function
1243                  * to access it.
1244                  */
1245                 while (!list_empty(&session->s_caps)) {
1246                         cap = list_entry(session->s_caps.next,
1247                                          struct ceph_cap, session_caps);
1248                         if (cap == prev)
1249                                 break;
1250                         prev = cap;
1251                         vino = cap->ci->i_vino;
1252                         spin_unlock(&session->s_cap_lock);
1253
1254                         inode = ceph_find_inode(sb, vino);
1255                         iput(inode);
1256
1257                         spin_lock(&session->s_cap_lock);
1258                 }
1259         }
1260
1261         // drop cap expires and unlock s_cap_lock
1262         cleanup_cap_releases(session->s_mdsc, session);
1263
1264         BUG_ON(session->s_nr_caps > 0);
1265         BUG_ON(!list_empty(&session->s_cap_flushing));
1266 }
1267
1268 /*
1269  * wake up any threads waiting on this session's caps.  if the cap is
1270  * old (didn't get renewed on the client reconnect), remove it now.
1271  *
1272  * caller must hold s_mutex.
1273  */
1274 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1275                               void *arg)
1276 {
1277         struct ceph_inode_info *ci = ceph_inode(inode);
1278
1279         wake_up_all(&ci->i_cap_wq);
1280         if (arg) {
1281                 spin_lock(&ci->i_ceph_lock);
1282                 ci->i_wanted_max_size = 0;
1283                 ci->i_requested_max_size = 0;
1284                 spin_unlock(&ci->i_ceph_lock);
1285         }
1286         return 0;
1287 }
1288
1289 static void wake_up_session_caps(struct ceph_mds_session *session,
1290                                  int reconnect)
1291 {
1292         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1293         iterate_session_caps(session, wake_up_session_cb,
1294                              (void *)(unsigned long)reconnect);
1295 }
1296
1297 /*
1298  * Send periodic message to MDS renewing all currently held caps.  The
1299  * ack will reset the expiration for all caps from this session.
1300  *
1301  * caller holds s_mutex
1302  */
1303 static int send_renew_caps(struct ceph_mds_client *mdsc,
1304                            struct ceph_mds_session *session)
1305 {
1306         struct ceph_msg *msg;
1307         int state;
1308
1309         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1310             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1311                 pr_info("mds%d caps stale\n", session->s_mds);
1312         session->s_renew_requested = jiffies;
1313
1314         /* do not try to renew caps until a recovering mds has reconnected
1315          * with its clients. */
1316         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1317         if (state < CEPH_MDS_STATE_RECONNECT) {
1318                 dout("send_renew_caps ignoring mds%d (%s)\n",
1319                      session->s_mds, ceph_mds_state_name(state));
1320                 return 0;
1321         }
1322
1323         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1324                 ceph_mds_state_name(state));
1325         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1326                                  ++session->s_renew_seq);
1327         if (!msg)
1328                 return -ENOMEM;
1329         ceph_con_send(&session->s_con, msg);
1330         return 0;
1331 }
1332
1333 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1334                              struct ceph_mds_session *session, u64 seq)
1335 {
1336         struct ceph_msg *msg;
1337
1338         dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1339              session->s_mds, ceph_session_state_name(session->s_state), seq);
1340         msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1341         if (!msg)
1342                 return -ENOMEM;
1343         ceph_con_send(&session->s_con, msg);
1344         return 0;
1345 }
1346
1347
1348 /*
1349  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1350  *
1351  * Called under session->s_mutex
1352  */
1353 static void renewed_caps(struct ceph_mds_client *mdsc,
1354                          struct ceph_mds_session *session, int is_renew)
1355 {
1356         int was_stale;
1357         int wake = 0;
1358
1359         spin_lock(&session->s_cap_lock);
1360         was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1361
1362         session->s_cap_ttl = session->s_renew_requested +
1363                 mdsc->mdsmap->m_session_timeout*HZ;
1364
1365         if (was_stale) {
1366                 if (time_before(jiffies, session->s_cap_ttl)) {
1367                         pr_info("mds%d caps renewed\n", session->s_mds);
1368                         wake = 1;
1369                 } else {
1370                         pr_info("mds%d caps still stale\n", session->s_mds);
1371                 }
1372         }
1373         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1374              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1375              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1376         spin_unlock(&session->s_cap_lock);
1377
1378         if (wake)
1379                 wake_up_session_caps(session, 0);
1380 }
1381
1382 /*
1383  * send a session close request
1384  */
1385 static int request_close_session(struct ceph_mds_client *mdsc,
1386                                  struct ceph_mds_session *session)
1387 {
1388         struct ceph_msg *msg;
1389
1390         dout("request_close_session mds%d state %s seq %lld\n",
1391              session->s_mds, ceph_session_state_name(session->s_state),
1392              session->s_seq);
1393         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1394         if (!msg)
1395                 return -ENOMEM;
1396         ceph_con_send(&session->s_con, msg);
1397         return 0;
1398 }
1399
1400 /*
1401  * Called with s_mutex held.
1402  */
1403 static int __close_session(struct ceph_mds_client *mdsc,
1404                          struct ceph_mds_session *session)
1405 {
1406         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1407                 return 0;
1408         session->s_state = CEPH_MDS_SESSION_CLOSING;
1409         return request_close_session(mdsc, session);
1410 }
1411
1412 static bool drop_negative_children(struct dentry *dentry)
1413 {
1414         struct dentry *child;
1415         bool all_negative = true;
1416
1417         if (!d_is_dir(dentry))
1418                 goto out;
1419
1420         spin_lock(&dentry->d_lock);
1421         list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1422                 if (d_really_is_positive(child)) {
1423                         all_negative = false;
1424                         break;
1425                 }
1426         }
1427         spin_unlock(&dentry->d_lock);
1428
1429         if (all_negative)
1430                 shrink_dcache_parent(dentry);
1431 out:
1432         return all_negative;
1433 }
1434
1435 /*
1436  * Trim old(er) caps.
1437  *
1438  * Because we can't cache an inode without one or more caps, we do
1439  * this indirectly: if a cap is unused, we prune its aliases, at which
1440  * point the inode will hopefully get dropped to.
1441  *
1442  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1443  * memory pressure from the MDS, though, so it needn't be perfect.
1444  */
1445 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1446 {
1447         struct ceph_mds_session *session = arg;
1448         struct ceph_inode_info *ci = ceph_inode(inode);
1449         int used, wanted, oissued, mine;
1450
1451         if (session->s_trim_caps <= 0)
1452                 return -1;
1453
1454         spin_lock(&ci->i_ceph_lock);
1455         mine = cap->issued | cap->implemented;
1456         used = __ceph_caps_used(ci);
1457         wanted = __ceph_caps_file_wanted(ci);
1458         oissued = __ceph_caps_issued_other(ci, cap);
1459
1460         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1461              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1462              ceph_cap_string(used), ceph_cap_string(wanted));
1463         if (cap == ci->i_auth_cap) {
1464                 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1465                     !list_empty(&ci->i_cap_snaps))
1466                         goto out;
1467                 if ((used | wanted) & CEPH_CAP_ANY_WR)
1468                         goto out;
1469         }
1470         /* The inode has cached pages, but it's no longer used.
1471          * we can safely drop it */
1472         if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1473             !(oissued & CEPH_CAP_FILE_CACHE)) {
1474           used = 0;
1475           oissued = 0;
1476         }
1477         if ((used | wanted) & ~oissued & mine)
1478                 goto out;   /* we need these caps */
1479
1480         if (oissued) {
1481                 /* we aren't the only cap.. just remove us */
1482                 __ceph_remove_cap(cap, true);
1483                 session->s_trim_caps--;
1484         } else {
1485                 struct dentry *dentry;
1486                 /* try dropping referring dentries */
1487                 spin_unlock(&ci->i_ceph_lock);
1488                 dentry = d_find_any_alias(inode);
1489                 if (dentry && drop_negative_children(dentry)) {
1490                         int count;
1491                         dput(dentry);
1492                         d_prune_aliases(inode);
1493                         count = atomic_read(&inode->i_count);
1494                         if (count == 1)
1495                                 session->s_trim_caps--;
1496                         dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1497                              inode, cap, count);
1498                 } else {
1499                         dput(dentry);
1500                 }
1501                 return 0;
1502         }
1503
1504 out:
1505         spin_unlock(&ci->i_ceph_lock);
1506         return 0;
1507 }
1508
1509 /*
1510  * Trim session cap count down to some max number.
1511  */
1512 static int trim_caps(struct ceph_mds_client *mdsc,
1513                      struct ceph_mds_session *session,
1514                      int max_caps)
1515 {
1516         int trim_caps = session->s_nr_caps - max_caps;
1517
1518         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1519              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1520         if (trim_caps > 0) {
1521                 session->s_trim_caps = trim_caps;
1522                 iterate_session_caps(session, trim_caps_cb, session);
1523                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1524                      session->s_mds, session->s_nr_caps, max_caps,
1525                         trim_caps - session->s_trim_caps);
1526                 session->s_trim_caps = 0;
1527         }
1528
1529         ceph_send_cap_releases(mdsc, session);
1530         return 0;
1531 }
1532
1533 static int check_capsnap_flush(struct ceph_inode_info *ci,
1534                                u64 want_snap_seq)
1535 {
1536         int ret = 1;
1537         spin_lock(&ci->i_ceph_lock);
1538         if (want_snap_seq > 0 && !list_empty(&ci->i_cap_snaps)) {
1539                 struct ceph_cap_snap *capsnap =
1540                         list_first_entry(&ci->i_cap_snaps,
1541                                          struct ceph_cap_snap, ci_item);
1542                 ret = capsnap->follows >= want_snap_seq;
1543         }
1544         spin_unlock(&ci->i_ceph_lock);
1545         return ret;
1546 }
1547
1548 static int check_caps_flush(struct ceph_mds_client *mdsc,
1549                             u64 want_flush_tid)
1550 {
1551         struct rb_node *n;
1552         struct ceph_cap_flush *cf;
1553         int ret = 1;
1554
1555         spin_lock(&mdsc->cap_dirty_lock);
1556         n = rb_first(&mdsc->cap_flush_tree);
1557         cf = n ? rb_entry(n, struct ceph_cap_flush, g_node) : NULL;
1558         if (cf && cf->tid <= want_flush_tid) {
1559                 dout("check_caps_flush still flushing tid %llu <= %llu\n",
1560                      cf->tid, want_flush_tid);
1561                 ret = 0;
1562         }
1563         spin_unlock(&mdsc->cap_dirty_lock);
1564         return ret;
1565 }
1566
1567 /*
1568  * flush all dirty inode data to disk.
1569  *
1570  * returns true if we've flushed through want_flush_tid
1571  */
1572 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1573                             u64 want_flush_tid, u64 want_snap_seq)
1574 {
1575         int mds;
1576
1577         dout("check_caps_flush want %llu snap want %llu\n",
1578              want_flush_tid, want_snap_seq);
1579         mutex_lock(&mdsc->mutex);
1580         for (mds = 0; mds < mdsc->max_sessions; ) {
1581                 struct ceph_mds_session *session = mdsc->sessions[mds];
1582                 struct inode *inode = NULL;
1583
1584                 if (!session) {
1585                         mds++;
1586                         continue;
1587                 }
1588                 get_session(session);
1589                 mutex_unlock(&mdsc->mutex);
1590
1591                 mutex_lock(&session->s_mutex);
1592                 if (!list_empty(&session->s_cap_snaps_flushing)) {
1593                         struct ceph_cap_snap *capsnap =
1594                                 list_first_entry(&session->s_cap_snaps_flushing,
1595                                                  struct ceph_cap_snap,
1596                                                  flushing_item);
1597                         struct ceph_inode_info *ci = capsnap->ci;
1598                         if (!check_capsnap_flush(ci, want_snap_seq)) {
1599                                 dout("check_cap_flush still flushing snap %p "
1600                                      "follows %lld <= %lld to mds%d\n",
1601                                      &ci->vfs_inode, capsnap->follows,
1602                                      want_snap_seq, mds);
1603                                 inode = igrab(&ci->vfs_inode);
1604                         }
1605                 }
1606                 mutex_unlock(&session->s_mutex);
1607                 ceph_put_mds_session(session);
1608
1609                 if (inode) {
1610                         wait_event(mdsc->cap_flushing_wq,
1611                                    check_capsnap_flush(ceph_inode(inode),
1612                                                        want_snap_seq));
1613                         iput(inode);
1614                 } else {
1615                         mds++;
1616                 }
1617
1618                 mutex_lock(&mdsc->mutex);
1619         }
1620         mutex_unlock(&mdsc->mutex);
1621
1622         wait_event(mdsc->cap_flushing_wq,
1623                    check_caps_flush(mdsc, want_flush_tid));
1624
1625         dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1626 }
1627
1628 /*
1629  * called under s_mutex
1630  */
1631 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1632                             struct ceph_mds_session *session)
1633 {
1634         struct ceph_msg *msg = NULL;
1635         struct ceph_mds_cap_release *head;
1636         struct ceph_mds_cap_item *item;
1637         struct ceph_cap *cap;
1638         LIST_HEAD(tmp_list);
1639         int num_cap_releases;
1640
1641         spin_lock(&session->s_cap_lock);
1642 again:
1643         list_splice_init(&session->s_cap_releases, &tmp_list);
1644         num_cap_releases = session->s_num_cap_releases;
1645         session->s_num_cap_releases = 0;
1646         spin_unlock(&session->s_cap_lock);
1647
1648         while (!list_empty(&tmp_list)) {
1649                 if (!msg) {
1650                         msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1651                                         PAGE_CACHE_SIZE, GFP_NOFS, false);
1652                         if (!msg)
1653                                 goto out_err;
1654                         head = msg->front.iov_base;
1655                         head->num = cpu_to_le32(0);
1656                         msg->front.iov_len = sizeof(*head);
1657                 }
1658                 cap = list_first_entry(&tmp_list, struct ceph_cap,
1659                                         session_caps);
1660                 list_del(&cap->session_caps);
1661                 num_cap_releases--;
1662
1663                 head = msg->front.iov_base;
1664                 le32_add_cpu(&head->num, 1);
1665                 item = msg->front.iov_base + msg->front.iov_len;
1666                 item->ino = cpu_to_le64(cap->cap_ino);
1667                 item->cap_id = cpu_to_le64(cap->cap_id);
1668                 item->migrate_seq = cpu_to_le32(cap->mseq);
1669                 item->seq = cpu_to_le32(cap->issue_seq);
1670                 msg->front.iov_len += sizeof(*item);
1671
1672                 ceph_put_cap(mdsc, cap);
1673
1674                 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1675                         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1676                         dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1677                         ceph_con_send(&session->s_con, msg);
1678                         msg = NULL;
1679                 }
1680         }
1681
1682         BUG_ON(num_cap_releases != 0);
1683
1684         spin_lock(&session->s_cap_lock);
1685         if (!list_empty(&session->s_cap_releases))
1686                 goto again;
1687         spin_unlock(&session->s_cap_lock);
1688
1689         if (msg) {
1690                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1691                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1692                 ceph_con_send(&session->s_con, msg);
1693         }
1694         return;
1695 out_err:
1696         pr_err("send_cap_releases mds%d, failed to allocate message\n",
1697                 session->s_mds);
1698         spin_lock(&session->s_cap_lock);
1699         list_splice(&tmp_list, &session->s_cap_releases);
1700         session->s_num_cap_releases += num_cap_releases;
1701         spin_unlock(&session->s_cap_lock);
1702 }
1703
1704 /*
1705  * requests
1706  */
1707
1708 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1709                                     struct inode *dir)
1710 {
1711         struct ceph_inode_info *ci = ceph_inode(dir);
1712         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1713         struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1714         size_t size = sizeof(*rinfo->dir_in) + sizeof(*rinfo->dir_dname_len) +
1715                       sizeof(*rinfo->dir_dname) + sizeof(*rinfo->dir_dlease);
1716         int order, num_entries;
1717
1718         spin_lock(&ci->i_ceph_lock);
1719         num_entries = ci->i_files + ci->i_subdirs;
1720         spin_unlock(&ci->i_ceph_lock);
1721         num_entries = max(num_entries, 1);
1722         num_entries = min(num_entries, opt->max_readdir);
1723
1724         order = get_order(size * num_entries);
1725         while (order >= 0) {
1726                 rinfo->dir_in = (void*)__get_free_pages(GFP_KERNEL |
1727                                                         __GFP_NOWARN,
1728                                                         order);
1729                 if (rinfo->dir_in)
1730                         break;
1731                 order--;
1732         }
1733         if (!rinfo->dir_in)
1734                 return -ENOMEM;
1735
1736         num_entries = (PAGE_SIZE << order) / size;
1737         num_entries = min(num_entries, opt->max_readdir);
1738
1739         rinfo->dir_buf_size = PAGE_SIZE << order;
1740         req->r_num_caps = num_entries + 1;
1741         req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1742         req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1743         return 0;
1744 }
1745
1746 /*
1747  * Create an mds request.
1748  */
1749 struct ceph_mds_request *
1750 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1751 {
1752         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1753
1754         if (!req)
1755                 return ERR_PTR(-ENOMEM);
1756
1757         mutex_init(&req->r_fill_mutex);
1758         req->r_mdsc = mdsc;
1759         req->r_started = jiffies;
1760         req->r_resend_mds = -1;
1761         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1762         INIT_LIST_HEAD(&req->r_unsafe_target_item);
1763         req->r_fmode = -1;
1764         kref_init(&req->r_kref);
1765         INIT_LIST_HEAD(&req->r_wait);
1766         init_completion(&req->r_completion);
1767         init_completion(&req->r_safe_completion);
1768         INIT_LIST_HEAD(&req->r_unsafe_item);
1769
1770         req->r_stamp = CURRENT_TIME;
1771
1772         req->r_op = op;
1773         req->r_direct_mode = mode;
1774         return req;
1775 }
1776
1777 /*
1778  * return oldest (lowest) request, tid in request tree, 0 if none.
1779  *
1780  * called under mdsc->mutex.
1781  */
1782 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1783 {
1784         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1785                 return NULL;
1786         return rb_entry(rb_first(&mdsc->request_tree),
1787                         struct ceph_mds_request, r_node);
1788 }
1789
1790 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1791 {
1792         return mdsc->oldest_tid;
1793 }
1794
1795 /*
1796  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1797  * on build_path_from_dentry in fs/cifs/dir.c.
1798  *
1799  * If @stop_on_nosnap, generate path relative to the first non-snapped
1800  * inode.
1801  *
1802  * Encode hidden .snap dirs as a double /, i.e.
1803  *   foo/.snap/bar -> foo//bar
1804  */
1805 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1806                            int stop_on_nosnap)
1807 {
1808         struct dentry *temp;
1809         char *path;
1810         int len, pos;
1811         unsigned seq;
1812
1813         if (dentry == NULL)
1814                 return ERR_PTR(-EINVAL);
1815
1816 retry:
1817         len = 0;
1818         seq = read_seqbegin(&rename_lock);
1819         rcu_read_lock();
1820         for (temp = dentry; !IS_ROOT(temp);) {
1821                 struct inode *inode = d_inode(temp);
1822                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1823                         len++;  /* slash only */
1824                 else if (stop_on_nosnap && inode &&
1825                          ceph_snap(inode) == CEPH_NOSNAP)
1826                         break;
1827                 else
1828                         len += 1 + temp->d_name.len;
1829                 temp = temp->d_parent;
1830         }
1831         rcu_read_unlock();
1832         if (len)
1833                 len--;  /* no leading '/' */
1834
1835         path = kmalloc(len+1, GFP_NOFS);
1836         if (path == NULL)
1837                 return ERR_PTR(-ENOMEM);
1838         pos = len;
1839         path[pos] = 0;  /* trailing null */
1840         rcu_read_lock();
1841         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1842                 struct inode *inode;
1843
1844                 spin_lock(&temp->d_lock);
1845                 inode = d_inode(temp);
1846                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1847                         dout("build_path path+%d: %p SNAPDIR\n",
1848                              pos, temp);
1849                 } else if (stop_on_nosnap && inode &&
1850                            ceph_snap(inode) == CEPH_NOSNAP) {
1851                         spin_unlock(&temp->d_lock);
1852                         break;
1853                 } else {
1854                         pos -= temp->d_name.len;
1855                         if (pos < 0) {
1856                                 spin_unlock(&temp->d_lock);
1857                                 break;
1858                         }
1859                         strncpy(path + pos, temp->d_name.name,
1860                                 temp->d_name.len);
1861                 }
1862                 spin_unlock(&temp->d_lock);
1863                 if (pos)
1864                         path[--pos] = '/';
1865                 temp = temp->d_parent;
1866         }
1867         rcu_read_unlock();
1868         if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1869                 pr_err("build_path did not end path lookup where "
1870                        "expected, namelen is %d, pos is %d\n", len, pos);
1871                 /* presumably this is only possible if racing with a
1872                    rename of one of the parent directories (we can not
1873                    lock the dentries above us to prevent this, but
1874                    retrying should be harmless) */
1875                 kfree(path);
1876                 goto retry;
1877         }
1878
1879         *base = ceph_ino(d_inode(temp));
1880         *plen = len;
1881         dout("build_path on %p %d built %llx '%.*s'\n",
1882              dentry, d_count(dentry), *base, len, path);
1883         return path;
1884 }
1885
1886 static int build_dentry_path(struct dentry *dentry,
1887                              const char **ppath, int *ppathlen, u64 *pino,
1888                              int *pfreepath)
1889 {
1890         char *path;
1891         struct inode *dir;
1892
1893         rcu_read_lock();
1894         dir = d_inode_rcu(dentry->d_parent);
1895         if (dir && ceph_snap(dir) == CEPH_NOSNAP) {
1896                 *pino = ceph_ino(dir);
1897                 rcu_read_unlock();
1898                 *ppath = dentry->d_name.name;
1899                 *ppathlen = dentry->d_name.len;
1900                 return 0;
1901         }
1902         rcu_read_unlock();
1903         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1904         if (IS_ERR(path))
1905                 return PTR_ERR(path);
1906         *ppath = path;
1907         *pfreepath = 1;
1908         return 0;
1909 }
1910
1911 static int build_inode_path(struct inode *inode,
1912                             const char **ppath, int *ppathlen, u64 *pino,
1913                             int *pfreepath)
1914 {
1915         struct dentry *dentry;
1916         char *path;
1917
1918         if (ceph_snap(inode) == CEPH_NOSNAP) {
1919                 *pino = ceph_ino(inode);
1920                 *ppathlen = 0;
1921                 return 0;
1922         }
1923         dentry = d_find_alias(inode);
1924         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1925         dput(dentry);
1926         if (IS_ERR(path))
1927                 return PTR_ERR(path);
1928         *ppath = path;
1929         *pfreepath = 1;
1930         return 0;
1931 }
1932
1933 /*
1934  * request arguments may be specified via an inode *, a dentry *, or
1935  * an explicit ino+path.
1936  */
1937 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1938                                   const char *rpath, u64 rino,
1939                                   const char **ppath, int *pathlen,
1940                                   u64 *ino, int *freepath)
1941 {
1942         int r = 0;
1943
1944         if (rinode) {
1945                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1946                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1947                      ceph_snap(rinode));
1948         } else if (rdentry) {
1949                 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1950                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1951                      *ppath);
1952         } else if (rpath || rino) {
1953                 *ino = rino;
1954                 *ppath = rpath;
1955                 *pathlen = rpath ? strlen(rpath) : 0;
1956                 dout(" path %.*s\n", *pathlen, rpath);
1957         }
1958
1959         return r;
1960 }
1961
1962 /*
1963  * called under mdsc->mutex
1964  */
1965 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1966                                                struct ceph_mds_request *req,
1967                                                int mds, bool drop_cap_releases)
1968 {
1969         struct ceph_msg *msg;
1970         struct ceph_mds_request_head *head;
1971         const char *path1 = NULL;
1972         const char *path2 = NULL;
1973         u64 ino1 = 0, ino2 = 0;
1974         int pathlen1 = 0, pathlen2 = 0;
1975         int freepath1 = 0, freepath2 = 0;
1976         int len;
1977         u16 releases;
1978         void *p, *end;
1979         int ret;
1980
1981         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1982                               req->r_path1, req->r_ino1.ino,
1983                               &path1, &pathlen1, &ino1, &freepath1);
1984         if (ret < 0) {
1985                 msg = ERR_PTR(ret);
1986                 goto out;
1987         }
1988
1989         ret = set_request_path_attr(NULL, req->r_old_dentry,
1990                               req->r_path2, req->r_ino2.ino,
1991                               &path2, &pathlen2, &ino2, &freepath2);
1992         if (ret < 0) {
1993                 msg = ERR_PTR(ret);
1994                 goto out_free1;
1995         }
1996
1997         len = sizeof(*head) +
1998                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
1999                 sizeof(struct ceph_timespec);
2000
2001         /* calculate (max) length for cap releases */
2002         len += sizeof(struct ceph_mds_request_release) *
2003                 (!!req->r_inode_drop + !!req->r_dentry_drop +
2004                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2005         if (req->r_dentry_drop)
2006                 len += req->r_dentry->d_name.len;
2007         if (req->r_old_dentry_drop)
2008                 len += req->r_old_dentry->d_name.len;
2009
2010         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
2011         if (!msg) {
2012                 msg = ERR_PTR(-ENOMEM);
2013                 goto out_free2;
2014         }
2015
2016         msg->hdr.version = cpu_to_le16(2);
2017         msg->hdr.tid = cpu_to_le64(req->r_tid);
2018
2019         head = msg->front.iov_base;
2020         p = msg->front.iov_base + sizeof(*head);
2021         end = msg->front.iov_base + msg->front.iov_len;
2022
2023         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2024         head->op = cpu_to_le32(req->r_op);
2025         head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
2026         head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
2027         head->args = req->r_args;
2028
2029         ceph_encode_filepath(&p, end, ino1, path1);
2030         ceph_encode_filepath(&p, end, ino2, path2);
2031
2032         /* make note of release offset, in case we need to replay */
2033         req->r_request_release_offset = p - msg->front.iov_base;
2034
2035         /* cap releases */
2036         releases = 0;
2037         if (req->r_inode_drop)
2038                 releases += ceph_encode_inode_release(&p,
2039                       req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2040                       mds, req->r_inode_drop, req->r_inode_unless, 0);
2041         if (req->r_dentry_drop)
2042                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
2043                        mds, req->r_dentry_drop, req->r_dentry_unless);
2044         if (req->r_old_dentry_drop)
2045                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2046                        mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
2047         if (req->r_old_inode_drop)
2048                 releases += ceph_encode_inode_release(&p,
2049                       d_inode(req->r_old_dentry),
2050                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2051
2052         if (drop_cap_releases) {
2053                 releases = 0;
2054                 p = msg->front.iov_base + req->r_request_release_offset;
2055         }
2056
2057         head->num_releases = cpu_to_le16(releases);
2058
2059         /* time stamp */
2060         {
2061                 struct ceph_timespec ts;
2062                 ceph_encode_timespec(&ts, &req->r_stamp);
2063                 ceph_encode_copy(&p, &ts, sizeof(ts));
2064         }
2065
2066         BUG_ON(p > end);
2067         msg->front.iov_len = p - msg->front.iov_base;
2068         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2069
2070         if (req->r_pagelist) {
2071                 struct ceph_pagelist *pagelist = req->r_pagelist;
2072                 atomic_inc(&pagelist->refcnt);
2073                 ceph_msg_data_add_pagelist(msg, pagelist);
2074                 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2075         } else {
2076                 msg->hdr.data_len = 0;
2077         }
2078
2079         msg->hdr.data_off = cpu_to_le16(0);
2080
2081 out_free2:
2082         if (freepath2)
2083                 kfree((char *)path2);
2084 out_free1:
2085         if (freepath1)
2086                 kfree((char *)path1);
2087 out:
2088         return msg;
2089 }
2090
2091 /*
2092  * called under mdsc->mutex if error, under no mutex if
2093  * success.
2094  */
2095 static void complete_request(struct ceph_mds_client *mdsc,
2096                              struct ceph_mds_request *req)
2097 {
2098         if (req->r_callback)
2099                 req->r_callback(mdsc, req);
2100         else
2101                 complete_all(&req->r_completion);
2102 }
2103
2104 /*
2105  * called under mdsc->mutex
2106  */
2107 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2108                                   struct ceph_mds_request *req,
2109                                   int mds, bool drop_cap_releases)
2110 {
2111         struct ceph_mds_request_head *rhead;
2112         struct ceph_msg *msg;
2113         int flags = 0;
2114
2115         req->r_attempts++;
2116         if (req->r_inode) {
2117                 struct ceph_cap *cap =
2118                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2119
2120                 if (cap)
2121                         req->r_sent_on_mseq = cap->mseq;
2122                 else
2123                         req->r_sent_on_mseq = -1;
2124         }
2125         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2126              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2127
2128         if (req->r_got_unsafe) {
2129                 void *p;
2130                 /*
2131                  * Replay.  Do not regenerate message (and rebuild
2132                  * paths, etc.); just use the original message.
2133                  * Rebuilding paths will break for renames because
2134                  * d_move mangles the src name.
2135                  */
2136                 msg = req->r_request;
2137                 rhead = msg->front.iov_base;
2138
2139                 flags = le32_to_cpu(rhead->flags);
2140                 flags |= CEPH_MDS_FLAG_REPLAY;
2141                 rhead->flags = cpu_to_le32(flags);
2142
2143                 if (req->r_target_inode)
2144                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2145
2146                 rhead->num_retry = req->r_attempts - 1;
2147
2148                 /* remove cap/dentry releases from message */
2149                 rhead->num_releases = 0;
2150
2151                 /* time stamp */
2152                 p = msg->front.iov_base + req->r_request_release_offset;
2153                 {
2154                         struct ceph_timespec ts;
2155                         ceph_encode_timespec(&ts, &req->r_stamp);
2156                         ceph_encode_copy(&p, &ts, sizeof(ts));
2157                 }
2158
2159                 msg->front.iov_len = p - msg->front.iov_base;
2160                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2161                 return 0;
2162         }
2163
2164         if (req->r_request) {
2165                 ceph_msg_put(req->r_request);
2166                 req->r_request = NULL;
2167         }
2168         msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2169         if (IS_ERR(msg)) {
2170                 req->r_err = PTR_ERR(msg);
2171                 return PTR_ERR(msg);
2172         }
2173         req->r_request = msg;
2174
2175         rhead = msg->front.iov_base;
2176         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2177         if (req->r_got_unsafe)
2178                 flags |= CEPH_MDS_FLAG_REPLAY;
2179         if (req->r_locked_dir)
2180                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2181         rhead->flags = cpu_to_le32(flags);
2182         rhead->num_fwd = req->r_num_fwd;
2183         rhead->num_retry = req->r_attempts - 1;
2184         rhead->ino = 0;
2185
2186         dout(" r_locked_dir = %p\n", req->r_locked_dir);
2187         return 0;
2188 }
2189
2190 /*
2191  * send request, or put it on the appropriate wait list.
2192  */
2193 static int __do_request(struct ceph_mds_client *mdsc,
2194                         struct ceph_mds_request *req)
2195 {
2196         struct ceph_mds_session *session = NULL;
2197         int mds = -1;
2198         int err = 0;
2199
2200         if (req->r_err || req->r_got_result) {
2201                 if (req->r_aborted)
2202                         __unregister_request(mdsc, req);
2203                 goto out;
2204         }
2205
2206         if (req->r_timeout &&
2207             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2208                 dout("do_request timed out\n");
2209                 err = -EIO;
2210                 goto finish;
2211         }
2212         if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2213                 dout("do_request forced umount\n");
2214                 err = -EIO;
2215                 goto finish;
2216         }
2217
2218         put_request_session(req);
2219
2220         mds = __choose_mds(mdsc, req);
2221         if (mds < 0 ||
2222             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2223                 dout("do_request no mds or not active, waiting for map\n");
2224                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2225                 goto out;
2226         }
2227
2228         /* get, open session */
2229         session = __ceph_lookup_mds_session(mdsc, mds);
2230         if (!session) {
2231                 session = register_session(mdsc, mds);
2232                 if (IS_ERR(session)) {
2233                         err = PTR_ERR(session);
2234                         goto finish;
2235                 }
2236         }
2237         req->r_session = get_session(session);
2238
2239         dout("do_request mds%d session %p state %s\n", mds, session,
2240              ceph_session_state_name(session->s_state));
2241         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2242             session->s_state != CEPH_MDS_SESSION_HUNG) {
2243                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2244                     session->s_state == CEPH_MDS_SESSION_CLOSING)
2245                         __open_session(mdsc, session);
2246                 list_add(&req->r_wait, &session->s_waiting);
2247                 goto out_session;
2248         }
2249
2250         /* send request */
2251         req->r_resend_mds = -1;   /* forget any previous mds hint */
2252
2253         if (req->r_request_started == 0)   /* note request start time */
2254                 req->r_request_started = jiffies;
2255
2256         err = __prepare_send_request(mdsc, req, mds, false);
2257         if (!err) {
2258                 ceph_msg_get(req->r_request);
2259                 ceph_con_send(&session->s_con, req->r_request);
2260         }
2261
2262 out_session:
2263         ceph_put_mds_session(session);
2264 finish:
2265         if (err) {
2266                 dout("__do_request early error %d\n", err);
2267                 req->r_err = err;
2268                 complete_request(mdsc, req);
2269                 __unregister_request(mdsc, req);
2270         }
2271 out:
2272         return err;
2273 }
2274
2275 /*
2276  * called under mdsc->mutex
2277  */
2278 static void __wake_requests(struct ceph_mds_client *mdsc,
2279                             struct list_head *head)
2280 {
2281         struct ceph_mds_request *req;
2282         LIST_HEAD(tmp_list);
2283
2284         list_splice_init(head, &tmp_list);
2285
2286         while (!list_empty(&tmp_list)) {
2287                 req = list_entry(tmp_list.next,
2288                                  struct ceph_mds_request, r_wait);
2289                 list_del_init(&req->r_wait);
2290                 dout(" wake request %p tid %llu\n", req, req->r_tid);
2291                 __do_request(mdsc, req);
2292         }
2293 }
2294
2295 /*
2296  * Wake up threads with requests pending for @mds, so that they can
2297  * resubmit their requests to a possibly different mds.
2298  */
2299 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2300 {
2301         struct ceph_mds_request *req;
2302         struct rb_node *p = rb_first(&mdsc->request_tree);
2303
2304         dout("kick_requests mds%d\n", mds);
2305         while (p) {
2306                 req = rb_entry(p, struct ceph_mds_request, r_node);
2307                 p = rb_next(p);
2308                 if (req->r_got_unsafe)
2309                         continue;
2310                 if (req->r_attempts > 0)
2311                         continue; /* only new requests */
2312                 if (req->r_session &&
2313                     req->r_session->s_mds == mds) {
2314                         dout(" kicking tid %llu\n", req->r_tid);
2315                         list_del_init(&req->r_wait);
2316                         __do_request(mdsc, req);
2317                 }
2318         }
2319 }
2320
2321 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2322                               struct ceph_mds_request *req)
2323 {
2324         dout("submit_request on %p\n", req);
2325         mutex_lock(&mdsc->mutex);
2326         __register_request(mdsc, req, NULL);
2327         __do_request(mdsc, req);
2328         mutex_unlock(&mdsc->mutex);
2329 }
2330
2331 /*
2332  * Synchrously perform an mds request.  Take care of all of the
2333  * session setup, forwarding, retry details.
2334  */
2335 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2336                          struct inode *dir,
2337                          struct ceph_mds_request *req)
2338 {
2339         int err;
2340
2341         dout("do_request on %p\n", req);
2342
2343         /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
2344         if (req->r_inode)
2345                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2346         if (req->r_locked_dir)
2347                 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
2348         if (req->r_old_dentry_dir)
2349                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2350                                   CEPH_CAP_PIN);
2351
2352         /* issue */
2353         mutex_lock(&mdsc->mutex);
2354         __register_request(mdsc, req, dir);
2355         __do_request(mdsc, req);
2356
2357         if (req->r_err) {
2358                 err = req->r_err;
2359                 goto out;
2360         }
2361
2362         /* wait */
2363         mutex_unlock(&mdsc->mutex);
2364         dout("do_request waiting\n");
2365         if (!req->r_timeout && req->r_wait_for_completion) {
2366                 err = req->r_wait_for_completion(mdsc, req);
2367         } else {
2368                 long timeleft = wait_for_completion_killable_timeout(
2369                                         &req->r_completion,
2370                                         ceph_timeout_jiffies(req->r_timeout));
2371                 if (timeleft > 0)
2372                         err = 0;
2373                 else if (!timeleft)
2374                         err = -EIO;  /* timed out */
2375                 else
2376                         err = timeleft;  /* killed */
2377         }
2378         dout("do_request waited, got %d\n", err);
2379         mutex_lock(&mdsc->mutex);
2380
2381         /* only abort if we didn't race with a real reply */
2382         if (req->r_got_result) {
2383                 err = le32_to_cpu(req->r_reply_info.head->result);
2384         } else if (err < 0) {
2385                 dout("aborted request %lld with %d\n", req->r_tid, err);
2386
2387                 /*
2388                  * ensure we aren't running concurrently with
2389                  * ceph_fill_trace or ceph_readdir_prepopulate, which
2390                  * rely on locks (dir mutex) held by our caller.
2391                  */
2392                 mutex_lock(&req->r_fill_mutex);
2393                 req->r_err = err;
2394                 req->r_aborted = true;
2395                 mutex_unlock(&req->r_fill_mutex);
2396
2397                 if (req->r_locked_dir &&
2398                     (req->r_op & CEPH_MDS_OP_WRITE))
2399                         ceph_invalidate_dir_request(req);
2400         } else {
2401                 err = req->r_err;
2402         }
2403
2404 out:
2405         mutex_unlock(&mdsc->mutex);
2406         dout("do_request %p done, result %d\n", req, err);
2407         return err;
2408 }
2409
2410 /*
2411  * Invalidate dir's completeness, dentry lease state on an aborted MDS
2412  * namespace request.
2413  */
2414 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2415 {
2416         struct inode *inode = req->r_locked_dir;
2417
2418         dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2419
2420         ceph_dir_clear_complete(inode);
2421         if (req->r_dentry)
2422                 ceph_invalidate_dentry_lease(req->r_dentry);
2423         if (req->r_old_dentry)
2424                 ceph_invalidate_dentry_lease(req->r_old_dentry);
2425 }
2426
2427 /*
2428  * Handle mds reply.
2429  *
2430  * We take the session mutex and parse and process the reply immediately.
2431  * This preserves the logical ordering of replies, capabilities, etc., sent
2432  * by the MDS as they are applied to our local cache.
2433  */
2434 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2435 {
2436         struct ceph_mds_client *mdsc = session->s_mdsc;
2437         struct ceph_mds_request *req;
2438         struct ceph_mds_reply_head *head = msg->front.iov_base;
2439         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2440         struct ceph_snap_realm *realm;
2441         u64 tid;
2442         int err, result;
2443         int mds = session->s_mds;
2444
2445         if (msg->front.iov_len < sizeof(*head)) {
2446                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2447                 ceph_msg_dump(msg);
2448                 return;
2449         }
2450
2451         /* get request, session */
2452         tid = le64_to_cpu(msg->hdr.tid);
2453         mutex_lock(&mdsc->mutex);
2454         req = __lookup_request(mdsc, tid);
2455         if (!req) {
2456                 dout("handle_reply on unknown tid %llu\n", tid);
2457                 mutex_unlock(&mdsc->mutex);
2458                 return;
2459         }
2460         dout("handle_reply %p\n", req);
2461
2462         /* correct session? */
2463         if (req->r_session != session) {
2464                 pr_err("mdsc_handle_reply got %llu on session mds%d"
2465                        " not mds%d\n", tid, session->s_mds,
2466                        req->r_session ? req->r_session->s_mds : -1);
2467                 mutex_unlock(&mdsc->mutex);
2468                 goto out;
2469         }
2470
2471         /* dup? */
2472         if ((req->r_got_unsafe && !head->safe) ||
2473             (req->r_got_safe && head->safe)) {
2474                 pr_warn("got a dup %s reply on %llu from mds%d\n",
2475                            head->safe ? "safe" : "unsafe", tid, mds);
2476                 mutex_unlock(&mdsc->mutex);
2477                 goto out;
2478         }
2479         if (req->r_got_safe) {
2480                 pr_warn("got unsafe after safe on %llu from mds%d\n",
2481                            tid, mds);
2482                 mutex_unlock(&mdsc->mutex);
2483                 goto out;
2484         }
2485
2486         result = le32_to_cpu(head->result);
2487
2488         /*
2489          * Handle an ESTALE
2490          * if we're not talking to the authority, send to them
2491          * if the authority has changed while we weren't looking,
2492          * send to new authority
2493          * Otherwise we just have to return an ESTALE
2494          */
2495         if (result == -ESTALE) {
2496                 dout("got ESTALE on request %llu", req->r_tid);
2497                 req->r_resend_mds = -1;
2498                 if (req->r_direct_mode != USE_AUTH_MDS) {
2499                         dout("not using auth, setting for that now");
2500                         req->r_direct_mode = USE_AUTH_MDS;
2501                         __do_request(mdsc, req);
2502                         mutex_unlock(&mdsc->mutex);
2503                         goto out;
2504                 } else  {
2505                         int mds = __choose_mds(mdsc, req);
2506                         if (mds >= 0 && mds != req->r_session->s_mds) {
2507                                 dout("but auth changed, so resending");
2508                                 __do_request(mdsc, req);
2509                                 mutex_unlock(&mdsc->mutex);
2510                                 goto out;
2511                         }
2512                 }
2513                 dout("have to return ESTALE on request %llu", req->r_tid);
2514         }
2515
2516
2517         if (head->safe) {
2518                 req->r_got_safe = true;
2519                 __unregister_request(mdsc, req);
2520
2521                 if (req->r_got_unsafe) {
2522                         /*
2523                          * We already handled the unsafe response, now do the
2524                          * cleanup.  No need to examine the response; the MDS
2525                          * doesn't include any result info in the safe
2526                          * response.  And even if it did, there is nothing
2527                          * useful we could do with a revised return value.
2528                          */
2529                         dout("got safe reply %llu, mds%d\n", tid, mds);
2530
2531                         /* last unsafe request during umount? */
2532                         if (mdsc->stopping && !__get_oldest_req(mdsc))
2533                                 complete_all(&mdsc->safe_umount_waiters);
2534                         mutex_unlock(&mdsc->mutex);
2535                         goto out;
2536                 }
2537         } else {
2538                 req->r_got_unsafe = true;
2539                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2540                 if (req->r_unsafe_dir) {
2541                         struct ceph_inode_info *ci =
2542                                         ceph_inode(req->r_unsafe_dir);
2543                         spin_lock(&ci->i_unsafe_lock);
2544                         list_add_tail(&req->r_unsafe_dir_item,
2545                                       &ci->i_unsafe_dirops);
2546                         spin_unlock(&ci->i_unsafe_lock);
2547                 }
2548         }
2549
2550         dout("handle_reply tid %lld result %d\n", tid, result);
2551         rinfo = &req->r_reply_info;
2552         err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2553         mutex_unlock(&mdsc->mutex);
2554
2555         mutex_lock(&session->s_mutex);
2556         if (err < 0) {
2557                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2558                 ceph_msg_dump(msg);
2559                 goto out_err;
2560         }
2561
2562         /* snap trace */
2563         realm = NULL;
2564         if (rinfo->snapblob_len) {
2565                 down_write(&mdsc->snap_rwsem);
2566                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2567                                 rinfo->snapblob + rinfo->snapblob_len,
2568                                 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2569                                 &realm);
2570                 downgrade_write(&mdsc->snap_rwsem);
2571         } else {
2572                 down_read(&mdsc->snap_rwsem);
2573         }
2574
2575         /* insert trace into our cache */
2576         mutex_lock(&req->r_fill_mutex);
2577         err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2578         if (err == 0) {
2579                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2580                                     req->r_op == CEPH_MDS_OP_LSSNAP))
2581                         ceph_readdir_prepopulate(req, req->r_session);
2582                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2583         }
2584         mutex_unlock(&req->r_fill_mutex);
2585
2586         up_read(&mdsc->snap_rwsem);
2587         if (realm)
2588                 ceph_put_snap_realm(mdsc, realm);
2589
2590         if (err == 0 && req->r_got_unsafe && req->r_target_inode) {
2591                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
2592                 spin_lock(&ci->i_unsafe_lock);
2593                 list_add_tail(&req->r_unsafe_target_item, &ci->i_unsafe_iops);
2594                 spin_unlock(&ci->i_unsafe_lock);
2595         }
2596 out_err:
2597         mutex_lock(&mdsc->mutex);
2598         if (!req->r_aborted) {
2599                 if (err) {
2600                         req->r_err = err;
2601                 } else {
2602                         req->r_reply =  ceph_msg_get(msg);
2603                         req->r_got_result = true;
2604                 }
2605         } else {
2606                 dout("reply arrived after request %lld was aborted\n", tid);
2607         }
2608         mutex_unlock(&mdsc->mutex);
2609
2610         mutex_unlock(&session->s_mutex);
2611
2612         /* kick calling process */
2613         complete_request(mdsc, req);
2614 out:
2615         ceph_mdsc_put_request(req);
2616         return;
2617 }
2618
2619
2620
2621 /*
2622  * handle mds notification that our request has been forwarded.
2623  */
2624 static void handle_forward(struct ceph_mds_client *mdsc,
2625                            struct ceph_mds_session *session,
2626                            struct ceph_msg *msg)
2627 {
2628         struct ceph_mds_request *req;
2629         u64 tid = le64_to_cpu(msg->hdr.tid);
2630         u32 next_mds;
2631         u32 fwd_seq;
2632         int err = -EINVAL;
2633         void *p = msg->front.iov_base;
2634         void *end = p + msg->front.iov_len;
2635
2636         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2637         next_mds = ceph_decode_32(&p);
2638         fwd_seq = ceph_decode_32(&p);
2639
2640         mutex_lock(&mdsc->mutex);
2641         req = __lookup_request(mdsc, tid);
2642         if (!req) {
2643                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2644                 goto out;  /* dup reply? */
2645         }
2646
2647         if (req->r_aborted) {
2648                 dout("forward tid %llu aborted, unregistering\n", tid);
2649                 __unregister_request(mdsc, req);
2650         } else if (fwd_seq <= req->r_num_fwd) {
2651                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2652                      tid, next_mds, req->r_num_fwd, fwd_seq);
2653         } else {
2654                 /* resend. forward race not possible; mds would drop */
2655                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2656                 BUG_ON(req->r_err);
2657                 BUG_ON(req->r_got_result);
2658                 req->r_attempts = 0;
2659                 req->r_num_fwd = fwd_seq;
2660                 req->r_resend_mds = next_mds;
2661                 put_request_session(req);
2662                 __do_request(mdsc, req);
2663         }
2664         ceph_mdsc_put_request(req);
2665 out:
2666         mutex_unlock(&mdsc->mutex);
2667         return;
2668
2669 bad:
2670         pr_err("mdsc_handle_forward decode error err=%d\n", err);
2671 }
2672
2673 /*
2674  * handle a mds session control message
2675  */
2676 static void handle_session(struct ceph_mds_session *session,
2677                            struct ceph_msg *msg)
2678 {
2679         struct ceph_mds_client *mdsc = session->s_mdsc;
2680         u32 op;
2681         u64 seq;
2682         int mds = session->s_mds;
2683         struct ceph_mds_session_head *h = msg->front.iov_base;
2684         int wake = 0;
2685
2686         /* decode */
2687         if (msg->front.iov_len != sizeof(*h))
2688                 goto bad;
2689         op = le32_to_cpu(h->op);
2690         seq = le64_to_cpu(h->seq);
2691
2692         mutex_lock(&mdsc->mutex);
2693         if (op == CEPH_SESSION_CLOSE)
2694                 __unregister_session(mdsc, session);
2695         /* FIXME: this ttl calculation is generous */
2696         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2697         mutex_unlock(&mdsc->mutex);
2698
2699         mutex_lock(&session->s_mutex);
2700
2701         dout("handle_session mds%d %s %p state %s seq %llu\n",
2702              mds, ceph_session_op_name(op), session,
2703              ceph_session_state_name(session->s_state), seq);
2704
2705         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2706                 session->s_state = CEPH_MDS_SESSION_OPEN;
2707                 pr_info("mds%d came back\n", session->s_mds);
2708         }
2709
2710         switch (op) {
2711         case CEPH_SESSION_OPEN:
2712                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2713                         pr_info("mds%d reconnect success\n", session->s_mds);
2714                 session->s_state = CEPH_MDS_SESSION_OPEN;
2715                 renewed_caps(mdsc, session, 0);
2716                 wake = 1;
2717                 if (mdsc->stopping)
2718                         __close_session(mdsc, session);
2719                 break;
2720
2721         case CEPH_SESSION_RENEWCAPS:
2722                 if (session->s_renew_seq == seq)
2723                         renewed_caps(mdsc, session, 1);
2724                 break;
2725
2726         case CEPH_SESSION_CLOSE:
2727                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2728                         pr_info("mds%d reconnect denied\n", session->s_mds);
2729                 cleanup_session_requests(mdsc, session);
2730                 remove_session_caps(session);
2731                 wake = 2; /* for good measure */
2732                 wake_up_all(&mdsc->session_close_wq);
2733                 break;
2734
2735         case CEPH_SESSION_STALE:
2736                 pr_info("mds%d caps went stale, renewing\n",
2737                         session->s_mds);
2738                 spin_lock(&session->s_gen_ttl_lock);
2739                 session->s_cap_gen++;
2740                 session->s_cap_ttl = jiffies - 1;
2741                 spin_unlock(&session->s_gen_ttl_lock);
2742                 send_renew_caps(mdsc, session);
2743                 break;
2744
2745         case CEPH_SESSION_RECALL_STATE:
2746                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2747                 break;
2748
2749         case CEPH_SESSION_FLUSHMSG:
2750                 send_flushmsg_ack(mdsc, session, seq);
2751                 break;
2752
2753         case CEPH_SESSION_FORCE_RO:
2754                 dout("force_session_readonly %p\n", session);
2755                 spin_lock(&session->s_cap_lock);
2756                 session->s_readonly = true;
2757                 spin_unlock(&session->s_cap_lock);
2758                 wake_up_session_caps(session, 0);
2759                 break;
2760
2761         default:
2762                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2763                 WARN_ON(1);
2764         }
2765
2766         mutex_unlock(&session->s_mutex);
2767         if (wake) {
2768                 mutex_lock(&mdsc->mutex);
2769                 __wake_requests(mdsc, &session->s_waiting);
2770                 if (wake == 2)
2771                         kick_requests(mdsc, mds);
2772                 mutex_unlock(&mdsc->mutex);
2773         }
2774         return;
2775
2776 bad:
2777         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2778                (int)msg->front.iov_len);
2779         ceph_msg_dump(msg);
2780         return;
2781 }
2782
2783
2784 /*
2785  * called under session->mutex.
2786  */
2787 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2788                                    struct ceph_mds_session *session)
2789 {
2790         struct ceph_mds_request *req, *nreq;
2791         struct rb_node *p;
2792         int err;
2793
2794         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2795
2796         mutex_lock(&mdsc->mutex);
2797         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2798                 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2799                 if (!err) {
2800                         ceph_msg_get(req->r_request);
2801                         ceph_con_send(&session->s_con, req->r_request);
2802                 }
2803         }
2804
2805         /*
2806          * also re-send old requests when MDS enters reconnect stage. So that MDS
2807          * can process completed request in clientreplay stage.
2808          */
2809         p = rb_first(&mdsc->request_tree);
2810         while (p) {
2811                 req = rb_entry(p, struct ceph_mds_request, r_node);
2812                 p = rb_next(p);
2813                 if (req->r_got_unsafe)
2814                         continue;
2815                 if (req->r_attempts == 0)
2816                         continue; /* only old requests */
2817                 if (req->r_session &&
2818                     req->r_session->s_mds == session->s_mds) {
2819                         err = __prepare_send_request(mdsc, req,
2820                                                      session->s_mds, true);
2821                         if (!err) {
2822                                 ceph_msg_get(req->r_request);
2823                                 ceph_con_send(&session->s_con, req->r_request);
2824                         }
2825                 }
2826         }
2827         mutex_unlock(&mdsc->mutex);
2828 }
2829
2830 /*
2831  * Encode information about a cap for a reconnect with the MDS.
2832  */
2833 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2834                           void *arg)
2835 {
2836         union {
2837                 struct ceph_mds_cap_reconnect v2;
2838                 struct ceph_mds_cap_reconnect_v1 v1;
2839         } rec;
2840         size_t reclen;
2841         struct ceph_inode_info *ci;
2842         struct ceph_reconnect_state *recon_state = arg;
2843         struct ceph_pagelist *pagelist = recon_state->pagelist;
2844         char *path;
2845         int pathlen, err;
2846         u64 pathbase;
2847         struct dentry *dentry;
2848
2849         ci = cap->ci;
2850
2851         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2852              inode, ceph_vinop(inode), cap, cap->cap_id,
2853              ceph_cap_string(cap->issued));
2854         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2855         if (err)
2856                 return err;
2857
2858         dentry = d_find_alias(inode);
2859         if (dentry) {
2860                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2861                 if (IS_ERR(path)) {
2862                         err = PTR_ERR(path);
2863                         goto out_dput;
2864                 }
2865         } else {
2866                 path = NULL;
2867                 pathlen = 0;
2868         }
2869         err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2870         if (err)
2871                 goto out_free;
2872
2873         spin_lock(&ci->i_ceph_lock);
2874         cap->seq = 0;        /* reset cap seq */
2875         cap->issue_seq = 0;  /* and issue_seq */
2876         cap->mseq = 0;       /* and migrate_seq */
2877         cap->cap_gen = cap->session->s_cap_gen;
2878
2879         if (recon_state->flock) {
2880                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2881                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2882                 rec.v2.issued = cpu_to_le32(cap->issued);
2883                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2884                 rec.v2.pathbase = cpu_to_le64(pathbase);
2885                 rec.v2.flock_len = 0;
2886                 reclen = sizeof(rec.v2);
2887         } else {
2888                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2889                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2890                 rec.v1.issued = cpu_to_le32(cap->issued);
2891                 rec.v1.size = cpu_to_le64(inode->i_size);
2892                 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2893                 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2894                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2895                 rec.v1.pathbase = cpu_to_le64(pathbase);
2896                 reclen = sizeof(rec.v1);
2897         }
2898         spin_unlock(&ci->i_ceph_lock);
2899
2900         if (recon_state->flock) {
2901                 int num_fcntl_locks, num_flock_locks;
2902                 struct ceph_filelock *flocks;
2903
2904 encode_again:
2905                 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2906                 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2907                                  sizeof(struct ceph_filelock), GFP_NOFS);
2908                 if (!flocks) {
2909                         err = -ENOMEM;
2910                         goto out_free;
2911                 }
2912                 err = ceph_encode_locks_to_buffer(inode, flocks,
2913                                                   num_fcntl_locks,
2914                                                   num_flock_locks);
2915                 if (err) {
2916                         kfree(flocks);
2917                         if (err == -ENOSPC)
2918                                 goto encode_again;
2919                         goto out_free;
2920                 }
2921                 /*
2922                  * number of encoded locks is stable, so copy to pagelist
2923                  */
2924                 rec.v2.flock_len = cpu_to_le32(2*sizeof(u32) +
2925                                     (num_fcntl_locks+num_flock_locks) *
2926                                     sizeof(struct ceph_filelock));
2927                 err = ceph_pagelist_append(pagelist, &rec, reclen);
2928                 if (!err)
2929                         err = ceph_locks_to_pagelist(flocks, pagelist,
2930                                                      num_fcntl_locks,
2931                                                      num_flock_locks);
2932                 kfree(flocks);
2933         } else {
2934                 err = ceph_pagelist_append(pagelist, &rec, reclen);
2935         }
2936
2937         recon_state->nr_caps++;
2938 out_free:
2939         kfree(path);
2940 out_dput:
2941         dput(dentry);
2942         return err;
2943 }
2944
2945
2946 /*
2947  * If an MDS fails and recovers, clients need to reconnect in order to
2948  * reestablish shared state.  This includes all caps issued through
2949  * this session _and_ the snap_realm hierarchy.  Because it's not
2950  * clear which snap realms the mds cares about, we send everything we
2951  * know about.. that ensures we'll then get any new info the
2952  * recovering MDS might have.
2953  *
2954  * This is a relatively heavyweight operation, but it's rare.
2955  *
2956  * called with mdsc->mutex held.
2957  */
2958 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2959                                struct ceph_mds_session *session)
2960 {
2961         struct ceph_msg *reply;
2962         struct rb_node *p;
2963         int mds = session->s_mds;
2964         int err = -ENOMEM;
2965         int s_nr_caps;
2966         struct ceph_pagelist *pagelist;
2967         struct ceph_reconnect_state recon_state;
2968
2969         pr_info("mds%d reconnect start\n", mds);
2970
2971         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2972         if (!pagelist)
2973                 goto fail_nopagelist;
2974         ceph_pagelist_init(pagelist);
2975
2976         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2977         if (!reply)
2978                 goto fail_nomsg;
2979
2980         mutex_lock(&session->s_mutex);
2981         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2982         session->s_seq = 0;
2983
2984         dout("session %p state %s\n", session,
2985              ceph_session_state_name(session->s_state));
2986
2987         spin_lock(&session->s_gen_ttl_lock);
2988         session->s_cap_gen++;
2989         spin_unlock(&session->s_gen_ttl_lock);
2990
2991         spin_lock(&session->s_cap_lock);
2992         /* don't know if session is readonly */
2993         session->s_readonly = 0;
2994         /*
2995          * notify __ceph_remove_cap() that we are composing cap reconnect.
2996          * If a cap get released before being added to the cap reconnect,
2997          * __ceph_remove_cap() should skip queuing cap release.
2998          */
2999         session->s_cap_reconnect = 1;
3000         /* drop old cap expires; we're about to reestablish that state */
3001         cleanup_cap_releases(mdsc, session);
3002
3003         /* trim unused caps to reduce MDS's cache rejoin time */
3004         if (mdsc->fsc->sb->s_root)
3005                 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3006
3007         ceph_con_close(&session->s_con);
3008         ceph_con_open(&session->s_con,
3009                       CEPH_ENTITY_TYPE_MDS, mds,
3010                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3011
3012         /* replay unsafe requests */
3013         replay_unsafe_requests(mdsc, session);
3014
3015         down_read(&mdsc->snap_rwsem);
3016
3017         /* traverse this session's caps */
3018         s_nr_caps = session->s_nr_caps;
3019         err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
3020         if (err)
3021                 goto fail;
3022
3023         recon_state.nr_caps = 0;
3024         recon_state.pagelist = pagelist;
3025         recon_state.flock = session->s_con.peer_features & CEPH_FEATURE_FLOCK;
3026         err = iterate_session_caps(session, encode_caps_cb, &recon_state);
3027         if (err < 0)
3028                 goto fail;
3029
3030         spin_lock(&session->s_cap_lock);
3031         session->s_cap_reconnect = 0;
3032         spin_unlock(&session->s_cap_lock);
3033
3034         /*
3035          * snaprealms.  we provide mds with the ino, seq (version), and
3036          * parent for all of our realms.  If the mds has any newer info,
3037          * it will tell us.
3038          */
3039         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3040                 struct ceph_snap_realm *realm =
3041                         rb_entry(p, struct ceph_snap_realm, node);
3042                 struct ceph_mds_snaprealm_reconnect sr_rec;
3043
3044                 dout(" adding snap realm %llx seq %lld parent %llx\n",
3045                      realm->ino, realm->seq, realm->parent_ino);
3046                 sr_rec.ino = cpu_to_le64(realm->ino);
3047                 sr_rec.seq = cpu_to_le64(realm->seq);
3048                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3049                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3050                 if (err)
3051                         goto fail;
3052         }
3053
3054         if (recon_state.flock)
3055                 reply->hdr.version = cpu_to_le16(2);
3056
3057         /* raced with cap release? */
3058         if (s_nr_caps != recon_state.nr_caps) {
3059                 struct page *page = list_first_entry(&pagelist->head,
3060                                                      struct page, lru);
3061                 __le32 *addr = kmap_atomic(page);
3062                 *addr = cpu_to_le32(recon_state.nr_caps);
3063                 kunmap_atomic(addr);
3064         }
3065
3066         reply->hdr.data_len = cpu_to_le32(pagelist->length);
3067         ceph_msg_data_add_pagelist(reply, pagelist);
3068
3069         ceph_early_kick_flushing_caps(mdsc, session);
3070
3071         ceph_con_send(&session->s_con, reply);
3072
3073         mutex_unlock(&session->s_mutex);
3074
3075         mutex_lock(&mdsc->mutex);
3076         __wake_requests(mdsc, &session->s_waiting);
3077         mutex_unlock(&mdsc->mutex);
3078
3079         up_read(&mdsc->snap_rwsem);
3080         return;
3081
3082 fail:
3083         ceph_msg_put(reply);
3084         up_read(&mdsc->snap_rwsem);
3085         mutex_unlock(&session->s_mutex);
3086 fail_nomsg:
3087         ceph_pagelist_release(pagelist);
3088 fail_nopagelist:
3089         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3090         return;
3091 }
3092
3093
3094 /*
3095  * compare old and new mdsmaps, kicking requests
3096  * and closing out old connections as necessary
3097  *
3098  * called under mdsc->mutex.
3099  */
3100 static void check_new_map(struct ceph_mds_client *mdsc,
3101                           struct ceph_mdsmap *newmap,
3102                           struct ceph_mdsmap *oldmap)
3103 {
3104         int i;
3105         int oldstate, newstate;
3106         struct ceph_mds_session *s;
3107
3108         dout("check_new_map new %u old %u\n",
3109              newmap->m_epoch, oldmap->m_epoch);
3110
3111         for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
3112                 if (mdsc->sessions[i] == NULL)
3113                         continue;
3114                 s = mdsc->sessions[i];
3115                 oldstate = ceph_mdsmap_get_state(oldmap, i);
3116                 newstate = ceph_mdsmap_get_state(newmap, i);
3117
3118                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3119                      i, ceph_mds_state_name(oldstate),
3120                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3121                      ceph_mds_state_name(newstate),
3122                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3123                      ceph_session_state_name(s->s_state));
3124
3125                 if (i >= newmap->m_max_mds ||
3126                     memcmp(ceph_mdsmap_get_addr(oldmap, i),
3127                            ceph_mdsmap_get_addr(newmap, i),
3128                            sizeof(struct ceph_entity_addr))) {
3129                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3130                                 /* the session never opened, just close it
3131                                  * out now */
3132                                 __wake_requests(mdsc, &s->s_waiting);
3133                                 __unregister_session(mdsc, s);
3134                         } else {
3135                                 /* just close it */
3136                                 mutex_unlock(&mdsc->mutex);
3137                                 mutex_lock(&s->s_mutex);
3138                                 mutex_lock(&mdsc->mutex);
3139                                 ceph_con_close(&s->s_con);
3140                                 mutex_unlock(&s->s_mutex);
3141                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3142                         }
3143                 } else if (oldstate == newstate) {
3144                         continue;  /* nothing new with this mds */
3145                 }
3146
3147                 /*
3148                  * send reconnect?
3149                  */
3150                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3151                     newstate >= CEPH_MDS_STATE_RECONNECT) {
3152                         mutex_unlock(&mdsc->mutex);
3153                         send_mds_reconnect(mdsc, s);
3154                         mutex_lock(&mdsc->mutex);
3155                 }
3156
3157                 /*
3158                  * kick request on any mds that has gone active.
3159                  */
3160                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3161                     newstate >= CEPH_MDS_STATE_ACTIVE) {
3162                         if (oldstate != CEPH_MDS_STATE_CREATING &&
3163                             oldstate != CEPH_MDS_STATE_STARTING)
3164                                 pr_info("mds%d recovery completed\n", s->s_mds);
3165                         kick_requests(mdsc, i);
3166                         ceph_kick_flushing_caps(mdsc, s);
3167                         wake_up_session_caps(s, 1);
3168                 }
3169         }
3170
3171         for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
3172                 s = mdsc->sessions[i];
3173                 if (!s)
3174                         continue;
3175                 if (!ceph_mdsmap_is_laggy(newmap, i))
3176                         continue;
3177                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3178                     s->s_state == CEPH_MDS_SESSION_HUNG ||
3179                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
3180                         dout(" connecting to export targets of laggy mds%d\n",
3181                              i);
3182                         __open_export_target_sessions(mdsc, s);
3183                 }
3184         }
3185 }
3186
3187
3188
3189 /*
3190  * leases
3191  */
3192
3193 /*
3194  * caller must hold session s_mutex, dentry->d_lock
3195  */
3196 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3197 {
3198         struct ceph_dentry_info *di = ceph_dentry(dentry);
3199
3200         ceph_put_mds_session(di->lease_session);
3201         di->lease_session = NULL;
3202 }
3203
3204 static void handle_lease(struct ceph_mds_client *mdsc,
3205                          struct ceph_mds_session *session,
3206                          struct ceph_msg *msg)
3207 {
3208         struct super_block *sb = mdsc->fsc->sb;
3209         struct inode *inode;
3210         struct dentry *parent, *dentry;
3211         struct ceph_dentry_info *di;
3212         int mds = session->s_mds;
3213         struct ceph_mds_lease *h = msg->front.iov_base;
3214         u32 seq;
3215         struct ceph_vino vino;
3216         struct qstr dname;
3217         int release = 0;
3218
3219         dout("handle_lease from mds%d\n", mds);
3220
3221         /* decode */
3222         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3223                 goto bad;
3224         vino.ino = le64_to_cpu(h->ino);
3225         vino.snap = CEPH_NOSNAP;
3226         seq = le32_to_cpu(h->seq);
3227         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3228         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3229         if (dname.len != get_unaligned_le32(h+1))
3230                 goto bad;
3231
3232         /* lookup inode */
3233         inode = ceph_find_inode(sb, vino);
3234         dout("handle_lease %s, ino %llx %p %.*s\n",
3235              ceph_lease_op_name(h->action), vino.ino, inode,
3236              dname.len, dname.name);
3237
3238         mutex_lock(&session->s_mutex);
3239         session->s_seq++;
3240
3241         if (inode == NULL) {
3242                 dout("handle_lease no inode %llx\n", vino.ino);
3243                 goto release;
3244         }
3245
3246         /* dentry */
3247         parent = d_find_alias(inode);
3248         if (!parent) {
3249                 dout("no parent dentry on inode %p\n", inode);
3250                 WARN_ON(1);
3251                 goto release;  /* hrm... */
3252         }
3253         dname.hash = full_name_hash(dname.name, dname.len);
3254         dentry = d_lookup(parent, &dname);
3255         dput(parent);
3256         if (!dentry)
3257                 goto release;
3258
3259         spin_lock(&dentry->d_lock);
3260         di = ceph_dentry(dentry);
3261         switch (h->action) {
3262         case CEPH_MDS_LEASE_REVOKE:
3263                 if (di->lease_session == session) {
3264                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3265                                 h->seq = cpu_to_le32(di->lease_seq);
3266                         __ceph_mdsc_drop_dentry_lease(dentry);
3267                 }
3268                 release = 1;
3269                 break;
3270
3271         case CEPH_MDS_LEASE_RENEW:
3272                 if (di->lease_session == session &&
3273                     di->lease_gen == session->s_cap_gen &&
3274                     di->lease_renew_from &&
3275                     di->lease_renew_after == 0) {
3276                         unsigned long duration =
3277                                 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3278
3279                         di->lease_seq = seq;
3280                         dentry->d_time = di->lease_renew_from + duration;
3281                         di->lease_renew_after = di->lease_renew_from +
3282                                 (duration >> 1);
3283                         di->lease_renew_from = 0;
3284                 }
3285                 break;
3286         }
3287         spin_unlock(&dentry->d_lock);
3288         dput(dentry);
3289
3290         if (!release)
3291                 goto out;
3292
3293 release:
3294         /* let's just reuse the same message */
3295         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3296         ceph_msg_get(msg);
3297         ceph_con_send(&session->s_con, msg);
3298
3299 out:
3300         iput(inode);
3301         mutex_unlock(&session->s_mutex);
3302         return;
3303
3304 bad:
3305         pr_err("corrupt lease message\n");
3306         ceph_msg_dump(msg);
3307 }
3308
3309 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3310                               struct inode *inode,
3311                               struct dentry *dentry, char action,
3312                               u32 seq)
3313 {
3314         struct ceph_msg *msg;
3315         struct ceph_mds_lease *lease;
3316         int len = sizeof(*lease) + sizeof(u32);
3317         int dnamelen = 0;
3318
3319         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3320              inode, dentry, ceph_lease_op_name(action), session->s_mds);
3321         dnamelen = dentry->d_name.len;
3322         len += dnamelen;
3323
3324         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3325         if (!msg)
3326                 return;
3327         lease = msg->front.iov_base;
3328         lease->action = action;
3329         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3330         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3331         lease->seq = cpu_to_le32(seq);
3332         put_unaligned_le32(dnamelen, lease + 1);
3333         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3334
3335         /*
3336          * if this is a preemptive lease RELEASE, no need to
3337          * flush request stream, since the actual request will
3338          * soon follow.
3339          */
3340         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3341
3342         ceph_con_send(&session->s_con, msg);
3343 }
3344
3345 /*
3346  * Preemptively release a lease we expect to invalidate anyway.
3347  * Pass @inode always, @dentry is optional.
3348  */
3349 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
3350                              struct dentry *dentry)
3351 {
3352         struct ceph_dentry_info *di;
3353         struct ceph_mds_session *session;
3354         u32 seq;
3355
3356         BUG_ON(inode == NULL);
3357         BUG_ON(dentry == NULL);
3358
3359         /* is dentry lease valid? */
3360         spin_lock(&dentry->d_lock);
3361         di = ceph_dentry(dentry);
3362         if (!di || !di->lease_session ||
3363             di->lease_session->s_mds < 0 ||
3364             di->lease_gen != di->lease_session->s_cap_gen ||
3365             !time_before(jiffies, dentry->d_time)) {
3366                 dout("lease_release inode %p dentry %p -- "
3367                      "no lease\n",
3368                      inode, dentry);
3369                 spin_unlock(&dentry->d_lock);
3370                 return;
3371         }
3372
3373         /* we do have a lease on this dentry; note mds and seq */
3374         session = ceph_get_mds_session(di->lease_session);
3375         seq = di->lease_seq;
3376         __ceph_mdsc_drop_dentry_lease(dentry);
3377         spin_unlock(&dentry->d_lock);
3378
3379         dout("lease_release inode %p dentry %p to mds%d\n",
3380              inode, dentry, session->s_mds);
3381         ceph_mdsc_lease_send_msg(session, inode, dentry,
3382                                  CEPH_MDS_LEASE_RELEASE, seq);
3383         ceph_put_mds_session(session);
3384 }
3385
3386 /*
3387  * drop all leases (and dentry refs) in preparation for umount
3388  */
3389 static void drop_leases(struct ceph_mds_client *mdsc)
3390 {
3391         int i;
3392
3393         dout("drop_leases\n");
3394         mutex_lock(&mdsc->mutex);
3395         for (i = 0; i < mdsc->max_sessions; i++) {
3396                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3397                 if (!s)
3398                         continue;
3399                 mutex_unlock(&mdsc->mutex);
3400                 mutex_lock(&s->s_mutex);
3401                 mutex_unlock(&s->s_mutex);
3402                 ceph_put_mds_session(s);
3403                 mutex_lock(&mdsc->mutex);
3404         }
3405         mutex_unlock(&mdsc->mutex);
3406 }
3407
3408
3409
3410 /*
3411  * delayed work -- periodically trim expired leases, renew caps with mds
3412  */
3413 static void schedule_delayed(struct ceph_mds_client *mdsc)
3414 {
3415         int delay = 5;
3416         unsigned hz = round_jiffies_relative(HZ * delay);
3417         schedule_delayed_work(&mdsc->delayed_work, hz);
3418 }
3419
3420 static void delayed_work(struct work_struct *work)
3421 {
3422         int i;
3423         struct ceph_mds_client *mdsc =
3424                 container_of(work, struct ceph_mds_client, delayed_work.work);
3425         int renew_interval;
3426         int renew_caps;
3427
3428         dout("mdsc delayed_work\n");
3429         ceph_check_delayed_caps(mdsc);
3430
3431         if (mdsc->stopping)
3432                 return;
3433
3434         mutex_lock(&mdsc->mutex);
3435         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3436         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3437                                    mdsc->last_renew_caps);
3438         if (renew_caps)
3439                 mdsc->last_renew_caps = jiffies;
3440
3441         for (i = 0; i < mdsc->max_sessions; i++) {
3442                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3443                 if (s == NULL)
3444                         continue;
3445                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3446                         dout("resending session close request for mds%d\n",
3447                              s->s_mds);
3448                         request_close_session(mdsc, s);
3449                         ceph_put_mds_session(s);
3450                         continue;
3451                 }
3452                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3453                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3454                                 s->s_state = CEPH_MDS_SESSION_HUNG;
3455                                 pr_info("mds%d hung\n", s->s_mds);
3456                         }
3457                 }
3458                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3459                         /* this mds is failed or recovering, just wait */
3460                         ceph_put_mds_session(s);
3461                         continue;
3462                 }
3463                 mutex_unlock(&mdsc->mutex);
3464
3465                 mutex_lock(&s->s_mutex);
3466                 if (renew_caps)
3467                         send_renew_caps(mdsc, s);
3468                 else
3469                         ceph_con_keepalive(&s->s_con);
3470                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3471                     s->s_state == CEPH_MDS_SESSION_HUNG)
3472                         ceph_send_cap_releases(mdsc, s);
3473                 mutex_unlock(&s->s_mutex);
3474                 ceph_put_mds_session(s);
3475
3476                 mutex_lock(&mdsc->mutex);
3477         }
3478         mutex_unlock(&mdsc->mutex);
3479
3480         schedule_delayed(mdsc);
3481 }
3482
3483 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3484
3485 {
3486         struct ceph_mds_client *mdsc;
3487
3488         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3489         if (!mdsc)
3490                 return -ENOMEM;
3491         mdsc->fsc = fsc;
3492         fsc->mdsc = mdsc;
3493         mutex_init(&mdsc->mutex);
3494         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3495         if (mdsc->mdsmap == NULL) {
3496                 kfree(mdsc);
3497                 return -ENOMEM;
3498         }
3499
3500         init_completion(&mdsc->safe_umount_waiters);
3501         init_waitqueue_head(&mdsc->session_close_wq);
3502         INIT_LIST_HEAD(&mdsc->waiting_for_map);
3503         mdsc->sessions = NULL;
3504         atomic_set(&mdsc->num_sessions, 0);
3505         mdsc->max_sessions = 0;
3506         mdsc->stopping = 0;
3507         mdsc->last_snap_seq = 0;
3508         init_rwsem(&mdsc->snap_rwsem);
3509         mdsc->snap_realms = RB_ROOT;
3510         INIT_LIST_HEAD(&mdsc->snap_empty);
3511         spin_lock_init(&mdsc->snap_empty_lock);
3512         mdsc->last_tid = 0;
3513         mdsc->oldest_tid = 0;
3514         mdsc->request_tree = RB_ROOT;
3515         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3516         mdsc->last_renew_caps = jiffies;
3517         INIT_LIST_HEAD(&mdsc->cap_delay_list);
3518         spin_lock_init(&mdsc->cap_delay_lock);
3519         INIT_LIST_HEAD(&mdsc->snap_flush_list);
3520         spin_lock_init(&mdsc->snap_flush_lock);
3521         mdsc->last_cap_flush_tid = 1;
3522         mdsc->cap_flush_tree = RB_ROOT;
3523         INIT_LIST_HEAD(&mdsc->cap_dirty);
3524         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3525         mdsc->num_cap_flushing = 0;
3526         spin_lock_init(&mdsc->cap_dirty_lock);
3527         init_waitqueue_head(&mdsc->cap_flushing_wq);
3528         spin_lock_init(&mdsc->dentry_lru_lock);
3529         INIT_LIST_HEAD(&mdsc->dentry_lru);
3530
3531         ceph_caps_init(mdsc);
3532         ceph_adjust_min_caps(mdsc, fsc->min_caps);
3533
3534         init_rwsem(&mdsc->pool_perm_rwsem);
3535         mdsc->pool_perm_tree = RB_ROOT;
3536
3537         return 0;
3538 }
3539
3540 /*
3541  * Wait for safe replies on open mds requests.  If we time out, drop
3542  * all requests from the tree to avoid dangling dentry refs.
3543  */
3544 static void wait_requests(struct ceph_mds_client *mdsc)
3545 {
3546         struct ceph_options *opts = mdsc->fsc->client->options;
3547         struct ceph_mds_request *req;
3548
3549         mutex_lock(&mdsc->mutex);
3550         if (__get_oldest_req(mdsc)) {
3551                 mutex_unlock(&mdsc->mutex);
3552
3553                 dout("wait_requests waiting for requests\n");
3554                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3555                                     ceph_timeout_jiffies(opts->mount_timeout));
3556
3557                 /* tear down remaining requests */
3558                 mutex_lock(&mdsc->mutex);
3559                 while ((req = __get_oldest_req(mdsc))) {
3560                         dout("wait_requests timed out on tid %llu\n",
3561                              req->r_tid);
3562                         __unregister_request(mdsc, req);
3563                 }
3564         }
3565         mutex_unlock(&mdsc->mutex);
3566         dout("wait_requests done\n");
3567 }
3568
3569 /*
3570  * called before mount is ro, and before dentries are torn down.
3571  * (hmm, does this still race with new lookups?)
3572  */
3573 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3574 {
3575         dout("pre_umount\n");
3576         mdsc->stopping = 1;
3577
3578         drop_leases(mdsc);
3579         ceph_flush_dirty_caps(mdsc);
3580         wait_requests(mdsc);
3581
3582         /*
3583          * wait for reply handlers to drop their request refs and
3584          * their inode/dcache refs
3585          */
3586         ceph_msgr_flush();
3587 }
3588
3589 /*
3590  * wait for all write mds requests to flush.
3591  */
3592 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3593 {
3594         struct ceph_mds_request *req = NULL, *nextreq;
3595         struct rb_node *n;
3596
3597         mutex_lock(&mdsc->mutex);
3598         dout("wait_unsafe_requests want %lld\n", want_tid);
3599 restart:
3600         req = __get_oldest_req(mdsc);
3601         while (req && req->r_tid <= want_tid) {
3602                 /* find next request */
3603                 n = rb_next(&req->r_node);
3604                 if (n)
3605                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3606                 else
3607                         nextreq = NULL;
3608                 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3609                     (req->r_op & CEPH_MDS_OP_WRITE)) {
3610                         /* write op */
3611                         ceph_mdsc_get_request(req);
3612                         if (nextreq)
3613                                 ceph_mdsc_get_request(nextreq);
3614                         mutex_unlock(&mdsc->mutex);
3615                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
3616                              req->r_tid, want_tid);
3617                         wait_for_completion(&req->r_safe_completion);
3618                         mutex_lock(&mdsc->mutex);
3619                         ceph_mdsc_put_request(req);
3620                         if (!nextreq)
3621                                 break;  /* next dne before, so we're done! */
3622                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
3623                                 /* next request was removed from tree */
3624                                 ceph_mdsc_put_request(nextreq);
3625                                 goto restart;
3626                         }
3627                         ceph_mdsc_put_request(nextreq);  /* won't go away */
3628                 }
3629                 req = nextreq;
3630         }
3631         mutex_unlock(&mdsc->mutex);
3632         dout("wait_unsafe_requests done\n");
3633 }
3634
3635 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3636 {
3637         u64 want_tid, want_flush, want_snap;
3638
3639         if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3640                 return;
3641
3642         dout("sync\n");
3643         mutex_lock(&mdsc->mutex);
3644         want_tid = mdsc->last_tid;
3645         mutex_unlock(&mdsc->mutex);
3646
3647         ceph_flush_dirty_caps(mdsc);
3648         spin_lock(&mdsc->cap_dirty_lock);
3649         want_flush = mdsc->last_cap_flush_tid;
3650         spin_unlock(&mdsc->cap_dirty_lock);
3651
3652         down_read(&mdsc->snap_rwsem);
3653         want_snap = mdsc->last_snap_seq;
3654         up_read(&mdsc->snap_rwsem);
3655
3656         dout("sync want tid %lld flush_seq %lld snap_seq %lld\n",
3657              want_tid, want_flush, want_snap);
3658
3659         wait_unsafe_requests(mdsc, want_tid);
3660         wait_caps_flush(mdsc, want_flush, want_snap);
3661 }
3662
3663 /*
3664  * true if all sessions are closed, or we force unmount
3665  */
3666 static bool done_closing_sessions(struct ceph_mds_client *mdsc)
3667 {
3668         if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3669                 return true;
3670         return atomic_read(&mdsc->num_sessions) == 0;
3671 }
3672
3673 /*
3674  * called after sb is ro.
3675  */
3676 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3677 {
3678         struct ceph_options *opts = mdsc->fsc->client->options;
3679         struct ceph_mds_session *session;
3680         int i;
3681
3682         dout("close_sessions\n");
3683
3684         /* close sessions */
3685         mutex_lock(&mdsc->mutex);
3686         for (i = 0; i < mdsc->max_sessions; i++) {
3687                 session = __ceph_lookup_mds_session(mdsc, i);
3688                 if (!session)
3689                         continue;
3690                 mutex_unlock(&mdsc->mutex);
3691                 mutex_lock(&session->s_mutex);
3692                 __close_session(mdsc, session);
3693                 mutex_unlock(&session->s_mutex);
3694                 ceph_put_mds_session(session);
3695                 mutex_lock(&mdsc->mutex);
3696         }
3697         mutex_unlock(&mdsc->mutex);
3698
3699         dout("waiting for sessions to close\n");
3700         wait_event_timeout(mdsc->session_close_wq, done_closing_sessions(mdsc),
3701                            ceph_timeout_jiffies(opts->mount_timeout));
3702
3703         /* tear down remaining sessions */
3704         mutex_lock(&mdsc->mutex);
3705         for (i = 0; i < mdsc->max_sessions; i++) {
3706                 if (mdsc->sessions[i]) {
3707                         session = get_session(mdsc->sessions[i]);
3708                         __unregister_session(mdsc, session);
3709                         mutex_unlock(&mdsc->mutex);
3710                         mutex_lock(&session->s_mutex);
3711                         remove_session_caps(session);
3712                         mutex_unlock(&session->s_mutex);
3713                         ceph_put_mds_session(session);
3714                         mutex_lock(&mdsc->mutex);
3715                 }
3716         }
3717         WARN_ON(!list_empty(&mdsc->cap_delay_list));
3718         mutex_unlock(&mdsc->mutex);
3719
3720         ceph_cleanup_empty_realms(mdsc);
3721
3722         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3723
3724         dout("stopped\n");
3725 }
3726
3727 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
3728 {
3729         struct ceph_mds_session *session;
3730         int mds;
3731
3732         dout("force umount\n");
3733
3734         mutex_lock(&mdsc->mutex);
3735         for (mds = 0; mds < mdsc->max_sessions; mds++) {
3736                 session = __ceph_lookup_mds_session(mdsc, mds);
3737                 if (!session)
3738                         continue;
3739                 mutex_unlock(&mdsc->mutex);
3740                 mutex_lock(&session->s_mutex);
3741                 __close_session(mdsc, session);
3742                 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
3743                         cleanup_session_requests(mdsc, session);
3744                         remove_session_caps(session);
3745                 }
3746                 mutex_unlock(&session->s_mutex);
3747                 ceph_put_mds_session(session);
3748                 mutex_lock(&mdsc->mutex);
3749                 kick_requests(mdsc, mds);
3750         }
3751         __wake_requests(mdsc, &mdsc->waiting_for_map);
3752         mutex_unlock(&mdsc->mutex);
3753 }
3754
3755 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3756 {
3757         dout("stop\n");
3758         /*
3759          * Make sure the delayed work stopped before releasing
3760          * the resources.
3761          *
3762          * Because the cancel_delayed_work_sync() will only
3763          * guarantee that the work finishes executing. But the
3764          * delayed work will re-arm itself again after that.
3765          */
3766         flush_delayed_work(&mdsc->delayed_work);
3767
3768         if (mdsc->mdsmap)
3769                 ceph_mdsmap_destroy(mdsc->mdsmap);
3770         kfree(mdsc->sessions);
3771         ceph_caps_finalize(mdsc);
3772         ceph_pool_perm_destroy(mdsc);
3773 }
3774
3775 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3776 {
3777         struct ceph_mds_client *mdsc = fsc->mdsc;
3778
3779         dout("mdsc_destroy %p\n", mdsc);
3780         ceph_mdsc_stop(mdsc);
3781
3782         /* flush out any connection work with references to us */
3783         ceph_msgr_flush();
3784
3785         fsc->mdsc = NULL;
3786         kfree(mdsc);
3787         dout("mdsc_destroy %p done\n", mdsc);
3788 }
3789
3790
3791 /*
3792  * handle mds map update.
3793  */
3794 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3795 {
3796         u32 epoch;
3797         u32 maplen;
3798         void *p = msg->front.iov_base;
3799         void *end = p + msg->front.iov_len;
3800         struct ceph_mdsmap *newmap, *oldmap;
3801         struct ceph_fsid fsid;
3802         int err = -EINVAL;
3803
3804         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3805         ceph_decode_copy(&p, &fsid, sizeof(fsid));
3806         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3807                 return;
3808         epoch = ceph_decode_32(&p);
3809         maplen = ceph_decode_32(&p);
3810         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3811
3812         /* do we need it? */
3813         ceph_monc_got_mdsmap(&mdsc->fsc->client->monc, epoch);
3814         mutex_lock(&mdsc->mutex);
3815         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3816                 dout("handle_map epoch %u <= our %u\n",
3817                      epoch, mdsc->mdsmap->m_epoch);
3818                 mutex_unlock(&mdsc->mutex);
3819                 return;
3820         }
3821
3822         newmap = ceph_mdsmap_decode(&p, end);
3823         if (IS_ERR(newmap)) {
3824                 err = PTR_ERR(newmap);
3825                 goto bad_unlock;
3826         }
3827
3828         /* swap into place */
3829         if (mdsc->mdsmap) {
3830                 oldmap = mdsc->mdsmap;
3831                 mdsc->mdsmap = newmap;
3832                 check_new_map(mdsc, newmap, oldmap);
3833                 ceph_mdsmap_destroy(oldmap);
3834         } else {
3835                 mdsc->mdsmap = newmap;  /* first mds map */
3836         }
3837         mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3838
3839         __wake_requests(mdsc, &mdsc->waiting_for_map);
3840
3841         mutex_unlock(&mdsc->mutex);
3842         schedule_delayed(mdsc);
3843         return;
3844
3845 bad_unlock:
3846         mutex_unlock(&mdsc->mutex);
3847 bad:
3848         pr_err("error decoding mdsmap %d\n", err);
3849         return;
3850 }
3851
3852 static struct ceph_connection *con_get(struct ceph_connection *con)
3853 {
3854         struct ceph_mds_session *s = con->private;
3855
3856         if (get_session(s)) {
3857                 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3858                 return con;
3859         }
3860         dout("mdsc con_get %p FAIL\n", s);
3861         return NULL;
3862 }
3863
3864 static void con_put(struct ceph_connection *con)
3865 {
3866         struct ceph_mds_session *s = con->private;
3867
3868         dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3869         ceph_put_mds_session(s);
3870 }
3871
3872 /*
3873  * if the client is unresponsive for long enough, the mds will kill
3874  * the session entirely.
3875  */
3876 static void peer_reset(struct ceph_connection *con)
3877 {
3878         struct ceph_mds_session *s = con->private;
3879         struct ceph_mds_client *mdsc = s->s_mdsc;
3880
3881         pr_warn("mds%d closed our session\n", s->s_mds);
3882         send_mds_reconnect(mdsc, s);
3883 }
3884
3885 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3886 {
3887         struct ceph_mds_session *s = con->private;
3888         struct ceph_mds_client *mdsc = s->s_mdsc;
3889         int type = le16_to_cpu(msg->hdr.type);
3890
3891         mutex_lock(&mdsc->mutex);
3892         if (__verify_registered_session(mdsc, s) < 0) {
3893                 mutex_unlock(&mdsc->mutex);
3894                 goto out;
3895         }
3896         mutex_unlock(&mdsc->mutex);
3897
3898         switch (type) {
3899         case CEPH_MSG_MDS_MAP:
3900                 ceph_mdsc_handle_map(mdsc, msg);
3901                 break;
3902         case CEPH_MSG_CLIENT_SESSION:
3903                 handle_session(s, msg);
3904                 break;
3905         case CEPH_MSG_CLIENT_REPLY:
3906                 handle_reply(s, msg);
3907                 break;
3908         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3909                 handle_forward(mdsc, s, msg);
3910                 break;
3911         case CEPH_MSG_CLIENT_CAPS:
3912                 ceph_handle_caps(s, msg);
3913                 break;
3914         case CEPH_MSG_CLIENT_SNAP:
3915                 ceph_handle_snap(mdsc, s, msg);
3916                 break;
3917         case CEPH_MSG_CLIENT_LEASE:
3918                 handle_lease(mdsc, s, msg);
3919                 break;
3920
3921         default:
3922                 pr_err("received unknown message type %d %s\n", type,
3923                        ceph_msg_type_name(type));
3924         }
3925 out:
3926         ceph_msg_put(msg);
3927 }
3928
3929 /*
3930  * authentication
3931  */
3932
3933 /*
3934  * Note: returned pointer is the address of a structure that's
3935  * managed separately.  Caller must *not* attempt to free it.
3936  */
3937 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
3938                                         int *proto, int force_new)
3939 {
3940         struct ceph_mds_session *s = con->private;
3941         struct ceph_mds_client *mdsc = s->s_mdsc;
3942         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3943         struct ceph_auth_handshake *auth = &s->s_auth;
3944
3945         if (force_new && auth->authorizer) {
3946                 ceph_auth_destroy_authorizer(ac, auth->authorizer);
3947                 auth->authorizer = NULL;
3948         }
3949         if (!auth->authorizer) {
3950                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3951                                                       auth);
3952                 if (ret)
3953                         return ERR_PTR(ret);
3954         } else {
3955                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3956                                                       auth);
3957                 if (ret)
3958                         return ERR_PTR(ret);
3959         }
3960         *proto = ac->protocol;
3961
3962         return auth;
3963 }
3964
3965
3966 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3967 {
3968         struct ceph_mds_session *s = con->private;
3969         struct ceph_mds_client *mdsc = s->s_mdsc;
3970         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3971
3972         return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer, len);
3973 }
3974
3975 static int invalidate_authorizer(struct ceph_connection *con)
3976 {
3977         struct ceph_mds_session *s = con->private;
3978         struct ceph_mds_client *mdsc = s->s_mdsc;
3979         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3980
3981         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3982
3983         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
3984 }
3985
3986 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
3987                                 struct ceph_msg_header *hdr, int *skip)
3988 {
3989         struct ceph_msg *msg;
3990         int type = (int) le16_to_cpu(hdr->type);
3991         int front_len = (int) le32_to_cpu(hdr->front_len);
3992
3993         if (con->in_msg)
3994                 return con->in_msg;
3995
3996         *skip = 0;
3997         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
3998         if (!msg) {
3999                 pr_err("unable to allocate msg type %d len %d\n",
4000                        type, front_len);
4001                 return NULL;
4002         }
4003
4004         return msg;
4005 }
4006
4007 static int mds_sign_message(struct ceph_msg *msg)
4008 {
4009        struct ceph_mds_session *s = msg->con->private;
4010        struct ceph_auth_handshake *auth = &s->s_auth;
4011
4012        return ceph_auth_sign_message(auth, msg);
4013 }
4014
4015 static int mds_check_message_signature(struct ceph_msg *msg)
4016 {
4017        struct ceph_mds_session *s = msg->con->private;
4018        struct ceph_auth_handshake *auth = &s->s_auth;
4019
4020        return ceph_auth_check_message_signature(auth, msg);
4021 }
4022
4023 static const struct ceph_connection_operations mds_con_ops = {
4024         .get = con_get,
4025         .put = con_put,
4026         .dispatch = dispatch,
4027         .get_authorizer = get_authorizer,
4028         .verify_authorizer_reply = verify_authorizer_reply,
4029         .invalidate_authorizer = invalidate_authorizer,
4030         .peer_reset = peer_reset,
4031         .alloc_msg = mds_alloc_msg,
4032         .sign_message = mds_sign_message,
4033         .check_message_signature = mds_check_message_signature,
4034 };
4035
4036 /* eof */