fs/ceph/mds_client.c

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  44  * are no longer valid.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