fs/ceph/mds_client.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/ceph/ceph_debug.h>
   3
   4 #include <linux/fs.h>
   5 #include <linux/wait.h>
   6 #include <linux/slab.h>
   7 #include <linux/gfp.h>
   8 #include <linux/sched.h>
   9 #include <linux/debugfs.h>
  10 #include <linux/seq_file.h>
  11 #include <linux/ratelimit.h>
  12
  13 #include "super.h"
  14 #include "mds_client.h"
  15
  16 #include <linux/ceph/ceph_features.h>
  17 #include <linux/ceph/messenger.h>
  18 #include <linux/ceph/decode.h>
  19 #include <linux/ceph/pagelist.h>
  20 #include <linux/ceph/auth.h>
  21 #include <linux/ceph/debugfs.h>
  22
  23 /*
  24  * A cluster of MDS (metadata server) daemons is responsible for
  25  * managing the file system namespace (the directory hierarchy and
  26  * inodes) and for coordinating shared access to storage.  Metadata is
  27  * partitioning hierarchically across a number of servers, and that
  28  * partition varies over time as the cluster adjusts the distribution
  29  * in order to balance load.
  30  *
  31  * The MDS client is primarily responsible to managing synchronous
  32  * metadata requests for operations like open, unlink, and so forth.
  33  * If there is a MDS failure, we find out about it when we (possibly
  34  * request and) receive a new MDS map, and can resubmit affected
  35  * requests.
  36  *
  37  * For the most part, though, we take advantage of a lossless
  38  * communications channel to the MDS, and do not need to worry about
  39  * timing out or resubmitting requests.
  40  *
  41  * We maintain a stateful "session" with each MDS we interact with.
  42  * Within each session, we sent periodic heartbeat messages to ensure
  43  * any capabilities or leases we have been issues remain valid.  If
  44  * the session times out and goes stale, our leases and capabilities
  45  * are no longer valid.
  46  */
  47
  48 struct ceph_reconnect_state {
  49         int nr_caps;
  50         struct ceph_pagelist *pagelist;
  51         unsigned msg_version;
  52 };
  53
  54 static void __wake_requests(struct ceph_mds_client *mdsc,
  55                             struct list_head *head);
  56
  57 static const struct ceph_connection_operations mds_con_ops;
  58
  59
  60 /*
  61  * mds reply parsing
  62  */
  63
  64 /*
  65  * parse individual inode info
  66  */
  67 static int parse_reply_info_in(void **p, void *end,
  68                                struct ceph_mds_reply_info_in *info,
  69                                u64 features)
  70 {
  71         int err = -EIO;
  72
  73         info->in = *p;
  74         *p += sizeof(struct ceph_mds_reply_inode) +
  75                 sizeof(*info->in->fragtree.splits) *
  76                 le32_to_cpu(info->in->fragtree.nsplits);
  77
  78         ceph_decode_32_safe(p, end, info->symlink_len, bad);
  79         ceph_decode_need(p, end, info->symlink_len, bad);
  80         info->symlink = *p;
  81         *p += info->symlink_len;
  82
  83         ceph_decode_copy_safe(p, end, &info->dir_layout,
  84                               sizeof(info->dir_layout), bad);
  85         ceph_decode_32_safe(p, end, info->xattr_len, bad);
  86         ceph_decode_need(p, end, info->xattr_len, bad);
  87         info->xattr_data = *p;
  88         *p += info->xattr_len;
  89
  90         if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
  91                 ceph_decode_64_safe(p, end, info->inline_version, bad);
  92                 ceph_decode_32_safe(p, end, info->inline_len, bad);
  93                 ceph_decode_need(p, end, info->inline_len, bad);
  94                 info->inline_data = *p;
  95                 *p += info->inline_len;
  96         } else
  97                 info->inline_version = CEPH_INLINE_NONE;
  98
  99         if (features & CEPH_FEATURE_MDS_QUOTA) {
 100                 u8 struct_v, struct_compat;
 101                 u32 struct_len;
 102
 103                 /*
 104                  * both struct_v and struct_compat are expected to be >= 1
 105                  */
 106                 ceph_decode_8_safe(p, end, struct_v, bad);
 107                 ceph_decode_8_safe(p, end, struct_compat, bad);
 108                 if (!struct_v || !struct_compat)
 109                         goto bad;
 110                 ceph_decode_32_safe(p, end, struct_len, bad);
 111                 ceph_decode_need(p, end, struct_len, bad);
 112                 ceph_decode_64_safe(p, end, info->max_bytes, bad);
 113                 ceph_decode_64_safe(p, end, info->max_files, bad);
 114         } else {
 115                 info->max_bytes = 0;
 116                 info->max_files = 0;
 117         }
 118
 119         info->pool_ns_len = 0;
 120         info->pool_ns_data = NULL;
 121         if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
 122                 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
 123                 if (info->pool_ns_len > 0) {
 124                         ceph_decode_need(p, end, info->pool_ns_len, bad);
 125                         info->pool_ns_data = *p;
 126                         *p += info->pool_ns_len;
 127                 }
 128         }
 129
 130         return 0;
 131 bad:
 132         return err;
 133 }
 134
 135 /*
 136  * parse a normal reply, which may contain a (dir+)dentry and/or a
 137  * target inode.
 138  */
 139 static int parse_reply_info_trace(void **p, void *end,
 140                                   struct ceph_mds_reply_info_parsed *info,
 141                                   u64 features)
 142 {
 143         int err;
 144
 145         if (info->head->is_dentry) {
 146                 err = parse_reply_info_in(p, end, &info->diri, features);
 147                 if (err < 0)
 148                         goto out_bad;
 149
 150                 if (unlikely(*p + sizeof(*info->dirfrag) > end))
 151                         goto bad;
 152                 info->dirfrag = *p;
 153                 *p += sizeof(*info->dirfrag) +
 154                         sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
 155                 if (unlikely(*p > end))
 156                         goto bad;
 157
 158                 ceph_decode_32_safe(p, end, info->dname_len, bad);
 159                 ceph_decode_need(p, end, info->dname_len, bad);
 160                 info->dname = *p;
 161                 *p += info->dname_len;
 162                 info->dlease = *p;
 163                 *p += sizeof(*info->dlease);
 164         }
 165
 166         if (info->head->is_target) {
 167                 err = parse_reply_info_in(p, end, &info->targeti, features);
 168                 if (err < 0)
 169                         goto out_bad;
 170         }
 171
 172         if (unlikely(*p != end))
 173                 goto bad;
 174         return 0;
 175
 176 bad:
 177         err = -EIO;
 178 out_bad:
 179         pr_err("problem parsing mds trace %d\n", err);
 180         return err;
 181 }
 182
 183 /*
 184  * parse readdir results
 185  */
 186 static int parse_reply_info_dir(void **p, void *end,
 187                                 struct ceph_mds_reply_info_parsed *info,
 188                                 u64 features)
 189 {
 190         u32 num, i = 0;
 191         int err;
 192
 193         info->dir_dir = *p;
 194         if (*p + sizeof(*info->dir_dir) > end)
 195                 goto bad;
 196         *p += sizeof(*info->dir_dir) +
 197                 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
 198         if (*p > end)
 199                 goto bad;
 200
 201         ceph_decode_need(p, end, sizeof(num) + 2, bad);
 202         num = ceph_decode_32(p);
 203         {
 204                 u16 flags = ceph_decode_16(p);
 205                 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
 206                 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
 207                 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
 208                 info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
 209         }
 210         if (num == 0)
 211                 goto done;
 212
 213         BUG_ON(!info->dir_entries);
 214         if ((unsigned long)(info->dir_entries + num) >
 215             (unsigned long)info->dir_entries + info->dir_buf_size) {
 216                 pr_err("dir contents are larger than expected\n");
 217                 WARN_ON(1);
 218                 goto bad;
 219         }
 220
 221         info->dir_nr = num;
 222         while (num) {
 223                 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
 224                 /* dentry */
 225                 ceph_decode_need(p, end, sizeof(u32)*2, bad);
 226                 rde->name_len = ceph_decode_32(p);
 227                 ceph_decode_need(p, end, rde->name_len, bad);
 228                 rde->name = *p;
 229                 *p += rde->name_len;
 230                 dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
 231                 rde->lease = *p;
 232                 *p += sizeof(struct ceph_mds_reply_lease);
 233
 234                 /* inode */
 235                 err = parse_reply_info_in(p, end, &rde->inode, features);
 236                 if (err < 0)
 237                         goto out_bad;
 238                 /* ceph_readdir_prepopulate() will update it */
 239                 rde->offset = 0;
 240                 i++;
 241                 num--;
 242         }
 243
 244 done:
 245         if (*p != end)
 246                 goto bad;
 247         return 0;
 248
 249 bad:
 250         err = -EIO;
 251 out_bad:
 252         pr_err("problem parsing dir contents %d\n", err);
 253         return err;
 254 }
 255
 256 /*
 257  * parse fcntl F_GETLK results
 258  */
 259 static int parse_reply_info_filelock(void **p, void *end,
 260                                      struct ceph_mds_reply_info_parsed *info,
 261                                      u64 features)
 262 {
 263         if (*p + sizeof(*info->filelock_reply) > end)
 264                 goto bad;
 265
 266         info->filelock_reply = *p;
 267         *p += sizeof(*info->filelock_reply);
 268
 269         if (unlikely(*p != end))
 270                 goto bad;
 271         return 0;
 272
 273 bad:
 274         return -EIO;
 275 }
 276
 277 /*
 278  * parse create results
 279  */
 280 static int parse_reply_info_create(void **p, void *end,
 281                                   struct ceph_mds_reply_info_parsed *info,
 282                                   u64 features)
 283 {
 284         if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
 285                 if (*p == end) {
 286                         info->has_create_ino = false;
 287                 } else {
 288                         info->has_create_ino = true;
 289                         info->ino = ceph_decode_64(p);
 290                 }
 291         }
 292
 293         if (unlikely(*p != end))
 294                 goto bad;
 295         return 0;
 296
 297 bad:
 298         return -EIO;
 299 }
 300
 301 /*
 302  * parse extra results
 303  */
 304 static int parse_reply_info_extra(void **p, void *end,
 305                                   struct ceph_mds_reply_info_parsed *info,
 306                                   u64 features)
 307 {
 308         u32 op = le32_to_cpu(info->head->op);
 309
 310         if (op == CEPH_MDS_OP_GETFILELOCK)
 311                 return parse_reply_info_filelock(p, end, info, features);
 312         else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
 313                 return parse_reply_info_dir(p, end, info, features);
 314         else if (op == CEPH_MDS_OP_CREATE)
 315                 return parse_reply_info_create(p, end, info, features);
 316         else
 317                 return -EIO;
 318 }
 319
 320 /*
 321  * parse entire mds reply
 322  */
 323 static int parse_reply_info(struct ceph_msg *msg,
 324                             struct ceph_mds_reply_info_parsed *info,
 325                             u64 features)
 326 {
 327         void *p, *end;
 328         u32 len;
 329         int err;
 330
 331         info->head = msg->front.iov_base;
 332         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 333         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 334
 335         /* trace */
 336         ceph_decode_32_safe(&p, end, len, bad);
 337         if (len > 0) {
 338                 ceph_decode_need(&p, end, len, bad);
 339                 err = parse_reply_info_trace(&p, p+len, info, features);
 340                 if (err < 0)
 341                         goto out_bad;
 342         }
 343
 344         /* extra */
 345         ceph_decode_32_safe(&p, end, len, bad);
 346         if (len > 0) {
 347                 ceph_decode_need(&p, end, len, bad);
 348                 err = parse_reply_info_extra(&p, p+len, info, features);
 349                 if (err < 0)
 350                         goto out_bad;
 351         }
 352
 353         /* snap blob */
 354         ceph_decode_32_safe(&p, end, len, bad);
 355         info->snapblob_len = len;
 356         info->snapblob = p;
 357         p += len;
 358
 359         if (p != end)
 360                 goto bad;
 361         return 0;
 362
 363 bad:
 364         err = -EIO;
 365 out_bad:
 366         pr_err("mds parse_reply err %d\n", err);
 367         return err;
 368 }
 369
 370 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
 371 {
 372         if (!info->dir_entries)
 373                 return;
 374         free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
 375 }
 376
 377
 378 /*
 379  * sessions
 380  */
 381 const char *ceph_session_state_name(int s)
 382 {
 383         switch (s) {
 384         case CEPH_MDS_SESSION_NEW: return "new";
 385         case CEPH_MDS_SESSION_OPENING: return "opening";
 386         case CEPH_MDS_SESSION_OPEN: return "open";
 387         case CEPH_MDS_SESSION_HUNG: return "hung";
 388         case CEPH_MDS_SESSION_CLOSING: return "closing";
 389         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
 390         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
 391         case CEPH_MDS_SESSION_REJECTED: return "rejected";
 392         default: return "???";
 393         }
 394 }
 395
 396 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
 397 {
 398         if (refcount_inc_not_zero(&s->s_ref)) {
 399                 dout("mdsc get_session %p %d -> %d\n", s,
 400                      refcount_read(&s->s_ref)-1, refcount_read(&s->s_ref));
 401                 return s;
 402         } else {
 403                 dout("mdsc get_session %p 0 -- FAIL\n", s);
 404                 return NULL;
 405         }
 406 }
 407
 408 void ceph_put_mds_session(struct ceph_mds_session *s)
 409 {
 410         dout("mdsc put_session %p %d -> %d\n", s,
 411              refcount_read(&s->s_ref), refcount_read(&s->s_ref)-1);
 412         if (refcount_dec_and_test(&s->s_ref)) {
 413                 if (s->s_auth.authorizer)
 414                         ceph_auth_destroy_authorizer(s->s_auth.authorizer);
 415                 kfree(s);
 416         }
 417 }
 418
 419 /*
 420  * called under mdsc->mutex
 421  */
 422 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
 423                                                    int mds)
 424 {
 425         struct ceph_mds_session *session;
 426
 427         if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
 428                 return NULL;
 429         session = mdsc->sessions[mds];
 430         dout("lookup_mds_session %p %d\n", session,
 431              refcount_read(&session->s_ref));
 432         get_session(session);
 433         return session;
 434 }
 435
 436 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
 437 {
 438         if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
 439                 return false;
 440         else
 441                 return true;
 442 }
 443
 444 static int __verify_registered_session(struct ceph_mds_client *mdsc,
 445                                        struct ceph_mds_session *s)
 446 {
 447         if (s->s_mds >= mdsc->max_sessions ||
 448             mdsc->sessions[s->s_mds] != s)
 449                 return -ENOENT;
 450         return 0;
 451 }
 452
 453 /*
 454  * create+register a new session for given mds.
 455  * called under mdsc->mutex.
 456  */
 457 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 458                                                  int mds)
 459 {
 460         struct ceph_mds_session *s;
 461
 462         if (mds >= mdsc->mdsmap->m_num_mds)
 463                 return ERR_PTR(-EINVAL);
 464
 465         s = kzalloc(sizeof(*s), GFP_NOFS);
 466         if (!s)
 467                 return ERR_PTR(-ENOMEM);
 468
 469         if (mds >= mdsc->max_sessions) {
 470                 int newmax = 1 << get_count_order(mds + 1);
 471                 struct ceph_mds_session **sa;
 472
 473                 dout("%s: realloc to %d\n", __func__, newmax);
 474                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 475                 if (!sa)
 476                         goto fail_realloc;
 477                 if (mdsc->sessions) {
 478                         memcpy(sa, mdsc->sessions,
 479                                mdsc->max_sessions * sizeof(void *));
 480                         kfree(mdsc->sessions);
 481                 }
 482                 mdsc->sessions = sa;
 483                 mdsc->max_sessions = newmax;
 484         }
 485
 486         dout("%s: mds%d\n", __func__, mds);
 487         s->s_mdsc = mdsc;
 488         s->s_mds = mds;
 489         s->s_state = CEPH_MDS_SESSION_NEW;
 490         s->s_ttl = 0;
 491         s->s_seq = 0;
 492         mutex_init(&s->s_mutex);
 493
 494         ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
 495
 496         spin_lock_init(&s->s_gen_ttl_lock);
 497         s->s_cap_gen = 0;
 498         s->s_cap_ttl = jiffies - 1;
 499
 500         spin_lock_init(&s->s_cap_lock);
 501         s->s_renew_requested = 0;
 502         s->s_renew_seq = 0;
 503         INIT_LIST_HEAD(&s->s_caps);
 504         s->s_nr_caps = 0;
 505         s->s_trim_caps = 0;
 506         refcount_set(&s->s_ref, 1);
 507         INIT_LIST_HEAD(&s->s_waiting);
 508         INIT_LIST_HEAD(&s->s_unsafe);
 509         s->s_num_cap_releases = 0;
 510         s->s_cap_reconnect = 0;
 511         s->s_cap_iterator = NULL;
 512         INIT_LIST_HEAD(&s->s_cap_releases);
 513         INIT_LIST_HEAD(&s->s_cap_flushing);
 514
 515         mdsc->sessions[mds] = s;
 516         atomic_inc(&mdsc->num_sessions);
 517         refcount_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
 518
 519         ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
 520                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
 521
 522         return s;
 523
 524 fail_realloc:
 525         kfree(s);
 526         return ERR_PTR(-ENOMEM);
 527 }
 528
 529 /*
 530  * called under mdsc->mutex
 531  */
 532 static void __unregister_session(struct ceph_mds_client *mdsc,
 533                                struct ceph_mds_session *s)
 534 {
 535         dout("__unregister_session mds%d %p\n", s->s_mds, s);
 536         BUG_ON(mdsc->sessions[s->s_mds] != s);
 537         mdsc->sessions[s->s_mds] = NULL;
 538         ceph_con_close(&s->s_con);
 539         ceph_put_mds_session(s);
 540         atomic_dec(&mdsc->num_sessions);
 541 }
 542
 543 /*
 544  * drop session refs in request.
 545  *
 546  * should be last request ref, or hold mdsc->mutex
 547  */
 548 static void put_request_session(struct ceph_mds_request *req)
 549 {
 550         if (req->r_session) {
 551                 ceph_put_mds_session(req->r_session);
 552                 req->r_session = NULL;
 553         }
 554 }
 555
 556 void ceph_mdsc_release_request(struct kref *kref)
 557 {
 558         struct ceph_mds_request *req = container_of(kref,
 559                                                     struct ceph_mds_request,
 560                                                     r_kref);
 561         destroy_reply_info(&req->r_reply_info);
 562         if (req->r_request)
 563                 ceph_msg_put(req->r_request);
 564         if (req->r_reply)
 565                 ceph_msg_put(req->r_reply);
 566         if (req->r_inode) {
 567                 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
 568                 iput(req->r_inode);
 569         }
 570         if (req->r_parent)
 571                 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
 572         iput(req->r_target_inode);
 573         if (req->r_dentry)
 574                 dput(req->r_dentry);
 575         if (req->r_old_dentry)
 576                 dput(req->r_old_dentry);
 577         if (req->r_old_dentry_dir) {
 578                 /*
 579                  * track (and drop pins for) r_old_dentry_dir
 580                  * separately, since r_old_dentry's d_parent may have
 581                  * changed between the dir mutex being dropped and
 582                  * this request being freed.
 583                  */
 584                 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
 585                                   CEPH_CAP_PIN);
 586                 iput(req->r_old_dentry_dir);
 587         }
 588         kfree(req->r_path1);
 589         kfree(req->r_path2);
 590         if (req->r_pagelist)
 591                 ceph_pagelist_release(req->r_pagelist);
 592         put_request_session(req);
 593         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
 594         kfree(req);
 595 }
 596
 597 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
 598
 599 /*
 600  * lookup session, bump ref if found.
 601  *
 602  * called under mdsc->mutex.
 603  */
 604 static struct ceph_mds_request *
 605 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
 606 {
 607         struct ceph_mds_request *req;
 608
 609         req = lookup_request(&mdsc->request_tree, tid);
 610         if (req)
 611                 ceph_mdsc_get_request(req);
 612
 613         return req;
 614 }
 615
 616 /*
 617  * Register an in-flight request, and assign a tid.  Link to directory
 618  * are modifying (if any).
 619  *
 620  * Called under mdsc->mutex.
 621  */
 622 static void __register_request(struct ceph_mds_client *mdsc,
 623                                struct ceph_mds_request *req,
 624                                struct inode *dir)
 625 {
 626         int ret = 0;
 627
 628         req->r_tid = ++mdsc->last_tid;
 629         if (req->r_num_caps) {
 630                 ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
 631                                         req->r_num_caps);
 632                 if (ret < 0) {
 633                         pr_err("__register_request %p "
 634                                "failed to reserve caps: %d\n", req, ret);
 635                         /* set req->r_err to fail early from __do_request */
 636                         req->r_err = ret;
 637                         return;
 638                 }
 639         }
 640         dout("__register_request %p tid %lld\n", req, req->r_tid);
 641         ceph_mdsc_get_request(req);
 642         insert_request(&mdsc->request_tree, req);
 643
 644         req->r_uid = current_fsuid();
 645         req->r_gid = current_fsgid();
 646
 647         if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
 648                 mdsc->oldest_tid = req->r_tid;
 649
 650         if (dir) {
 651                 ihold(dir);
 652                 req->r_unsafe_dir = dir;
 653         }
 654 }
 655
 656 static void __unregister_request(struct ceph_mds_client *mdsc,
 657                                  struct ceph_mds_request *req)
 658 {
 659         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 660
 661         /* Never leave an unregistered request on an unsafe list! */
 662         list_del_init(&req->r_unsafe_item);
 663
 664         if (req->r_tid == mdsc->oldest_tid) {
 665                 struct rb_node *p = rb_next(&req->r_node);
 666                 mdsc->oldest_tid = 0;
 667                 while (p) {
 668                         struct ceph_mds_request *next_req =
 669                                 rb_entry(p, struct ceph_mds_request, r_node);
 670                         if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
 671                                 mdsc->oldest_tid = next_req->r_tid;
 672                                 break;
 673                         }
 674                         p = rb_next(p);
 675                 }
 676         }
 677
 678         erase_request(&mdsc->request_tree, req);
 679
 680         if (req->r_unsafe_dir  &&
 681             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
 682                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 683                 spin_lock(&ci->i_unsafe_lock);
 684                 list_del_init(&req->r_unsafe_dir_item);
 685                 spin_unlock(&ci->i_unsafe_lock);
 686         }
 687         if (req->r_target_inode &&
 688             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
 689                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
 690                 spin_lock(&ci->i_unsafe_lock);
 691                 list_del_init(&req->r_unsafe_target_item);
 692                 spin_unlock(&ci->i_unsafe_lock);
 693         }
 694
 695         if (req->r_unsafe_dir) {
 696                 iput(req->r_unsafe_dir);
 697                 req->r_unsafe_dir = NULL;
 698         }
 699
 700         complete_all(&req->r_safe_completion);
 701
 702         ceph_mdsc_put_request(req);
 703 }
 704
 705 /*
 706  * Walk back up the dentry tree until we hit a dentry representing a
 707  * non-snapshot inode. We do this using the rcu_read_lock (which must be held
 708  * when calling this) to ensure that the objects won't disappear while we're
 709  * working with them. Once we hit a candidate dentry, we attempt to take a
 710  * reference to it, and return that as the result.
 711  */
 712 static struct inode *get_nonsnap_parent(struct dentry *dentry)
 713 {
 714         struct inode *inode = NULL;
 715
 716         while (dentry && !IS_ROOT(dentry)) {
 717                 inode = d_inode_rcu(dentry);
 718                 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
 719                         break;
 720                 dentry = dentry->d_parent;
 721         }
 722         if (inode)
 723                 inode = igrab(inode);
 724         return inode;
 725 }
 726
 727 /*
 728  * Choose mds to send request to next.  If there is a hint set in the
 729  * request (e.g., due to a prior forward hint from the mds), use that.
 730  * Otherwise, consult frag tree and/or caps to identify the
 731  * appropriate mds.  If all else fails, choose randomly.
 732  *
 733  * Called under mdsc->mutex.
 734  */
 735 static int __choose_mds(struct ceph_mds_client *mdsc,
 736                         struct ceph_mds_request *req)
 737 {
 738         struct inode *inode;
 739         struct ceph_inode_info *ci;
 740         struct ceph_cap *cap;
 741         int mode = req->r_direct_mode;
 742         int mds = -1;
 743         u32 hash = req->r_direct_hash;
 744         bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
 745
 746         /*
 747          * is there a specific mds we should try?  ignore hint if we have
 748          * no session and the mds is not up (active or recovering).
 749          */
 750         if (req->r_resend_mds >= 0 &&
 751             (__have_session(mdsc, req->r_resend_mds) ||
 752              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 753                 dout("choose_mds using resend_mds mds%d\n",
 754                      req->r_resend_mds);
 755                 return req->r_resend_mds;
 756         }
 757
 758         if (mode == USE_RANDOM_MDS)
 759                 goto random;
 760
 761         inode = NULL;
 762         if (req->r_inode) {
 763                 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
 764                         inode = req->r_inode;
 765                         ihold(inode);
 766                 } else {
 767                         /* req->r_dentry is non-null for LSSNAP request */
 768                         rcu_read_lock();
 769                         inode = get_nonsnap_parent(req->r_dentry);
 770                         rcu_read_unlock();
 771                         dout("__choose_mds using snapdir's parent %p\n", inode);
 772                 }
 773         } else if (req->r_dentry) {
 774                 /* ignore race with rename; old or new d_parent is okay */
 775                 struct dentry *parent;
 776                 struct inode *dir;
 777
 778                 rcu_read_lock();
 779                 parent = req->r_dentry->d_parent;
 780                 dir = req->r_parent ? : d_inode_rcu(parent);
 781
 782                 if (!dir || dir->i_sb != mdsc->fsc->sb) {
 783                         /*  not this fs or parent went negative */
 784                         inode = d_inode(req->r_dentry);
 785                         if (inode)
 786                                 ihold(inode);
 787                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
 788                         /* direct snapped/virtual snapdir requests
 789                          * based on parent dir inode */
 790                         inode = get_nonsnap_parent(parent);
 791                         dout("__choose_mds using nonsnap parent %p\n", inode);
 792                 } else {
 793                         /* dentry target */
 794                         inode = d_inode(req->r_dentry);
 795                         if (!inode || mode == USE_AUTH_MDS) {
 796                                 /* dir + name */
 797                                 inode = igrab(dir);
 798                                 hash = ceph_dentry_hash(dir, req->r_dentry);
 799                                 is_hash = true;
 800                         } else {
 801                                 ihold(inode);
 802                         }
 803                 }
 804                 rcu_read_unlock();
 805         }
 806
 807         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 808              (int)hash, mode);
 809         if (!inode)
 810                 goto random;
 811         ci = ceph_inode(inode);
 812
 813         if (is_hash && S_ISDIR(inode->i_mode)) {
 814                 struct ceph_inode_frag frag;
 815                 int found;
 816
 817                 ceph_choose_frag(ci, hash, &frag, &found);
 818                 if (found) {
 819                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 820                                 u8 r;
 821
 822                                 /* choose a random replica */
 823                                 get_random_bytes(&r, 1);
 824                                 r %= frag.ndist;
 825                                 mds = frag.dist[r];
 826                                 dout("choose_mds %p %llx.%llx "
 827                                      "frag %u mds%d (%d/%d)\n",
 828                                      inode, ceph_vinop(inode),
 829                                      frag.frag, mds,
 830                                      (int)r, frag.ndist);
 831                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 832                                     CEPH_MDS_STATE_ACTIVE)
 833                                         goto out;
 834                         }
 835
 836                         /* since this file/dir wasn't known to be
 837                          * replicated, then we want to look for the
 838                          * authoritative mds. */
 839                         mode = USE_AUTH_MDS;
 840                         if (frag.mds >= 0) {
 841                                 /* choose auth mds */
 842                                 mds = frag.mds;
 843                                 dout("choose_mds %p %llx.%llx "
 844                                      "frag %u mds%d (auth)\n",
 845                                      inode, ceph_vinop(inode), frag.frag, mds);
 846                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 847                                     CEPH_MDS_STATE_ACTIVE)
 848                                         goto out;
 849                         }
 850                 }
 851         }
 852
 853         spin_lock(&ci->i_ceph_lock);
 854         cap = NULL;
 855         if (mode == USE_AUTH_MDS)
 856                 cap = ci->i_auth_cap;
 857         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
 858                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
 859         if (!cap) {
 860                 spin_unlock(&ci->i_ceph_lock);
 861                 iput(inode);
 862                 goto random;
 863         }
 864         mds = cap->session->s_mds;
 865         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
 866              inode, ceph_vinop(inode), mds,
 867              cap == ci->i_auth_cap ? "auth " : "", cap);
 868         spin_unlock(&ci->i_ceph_lock);
 869 out:
 870         iput(inode);
 871         return mds;
 872
 873 random:
 874         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
 875         dout("choose_mds chose random mds%d\n", mds);
 876         return mds;
 877 }
 878
 879
 880 /*
 881  * session messages
 882  */
 883 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
 884 {
 885         struct ceph_msg *msg;
 886         struct ceph_mds_session_head *h;
 887
 888         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
 889                            false);
 890         if (!msg) {
 891                 pr_err("create_session_msg ENOMEM creating msg\n");
 892                 return NULL;
 893         }
 894         h = msg->front.iov_base;
 895         h->op = cpu_to_le32(op);
 896         h->seq = cpu_to_le64(seq);
 897
 898         return msg;
 899 }
 900
 901 static void encode_supported_features(void **p, void *end)
 902 {
 903         static const unsigned char bits[] = CEPHFS_FEATURES_CLIENT_SUPPORTED;
 904         static const size_t count = ARRAY_SIZE(bits);
 905
 906         if (count > 0) {
 907                 size_t i;
 908                 size_t size = ((size_t)bits[count - 1] + 64) / 64 * 8;
 909
 910                 BUG_ON(*p + 4 + size > end);
 911                 ceph_encode_32(p, size);
 912                 memset(*p, 0, size);
 913                 for (i = 0; i < count; i++)
 914                         ((unsigned char*)(*p))[i / 8] |= 1 << (bits[i] % 8);
 915                 *p += size;
 916         } else {
 917                 BUG_ON(*p + 4 > end);
 918                 ceph_encode_32(p, 0);
 919         }
 920 }
 921
 922 /*
 923  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
 924  * to include additional client metadata fields.
 925  */
 926 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
 927 {
 928         struct ceph_msg *msg;
 929         struct ceph_mds_session_head *h;
 930         int i = -1;
 931         int extra_bytes = 0;
 932         int metadata_key_count = 0;
 933         struct ceph_options *opt = mdsc->fsc->client->options;
 934         struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
 935         void *p, *end;
 936
 937         const char* metadata[][2] = {
 938                 {"hostname", mdsc->nodename},
 939                 {"kernel_version", init_utsname()->release},
 940                 {"entity_id", opt->name ? : ""},
 941                 {"root", fsopt->server_path ? : "/"},
 942                 {NULL, NULL}
 943         };
 944
 945         /* Calculate serialized length of metadata */
 946         extra_bytes = 4;  /* map length */
 947         for (i = 0; metadata[i][0]; ++i) {
 948                 extra_bytes += 8 + strlen(metadata[i][0]) +
 949                         strlen(metadata[i][1]);
 950                 metadata_key_count++;
 951         }
 952         /* supported feature */
 953         extra_bytes += 4 + 8;
 954
 955         /* Allocate the message */
 956         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + extra_bytes,
 957                            GFP_NOFS, false);
 958         if (!msg) {
 959                 pr_err("create_session_msg ENOMEM creating msg\n");
 960                 return NULL;
 961         }
 962         p = msg->front.iov_base;
 963         end = p + msg->front.iov_len;
 964
 965         h = p;
 966         h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
 967         h->seq = cpu_to_le64(seq);
 968
 969         /*
 970          * Serialize client metadata into waiting buffer space, using
 971          * the format that userspace expects for map<string, string>
 972          *
 973          * ClientSession messages with metadata are v2
 974          */
 975         msg->hdr.version = cpu_to_le16(3);
 976         msg->hdr.compat_version = cpu_to_le16(1);
 977
 978         /* The write pointer, following the session_head structure */
 979         p += sizeof(*h);
 980
 981         /* Number of entries in the map */
 982         ceph_encode_32(&p, metadata_key_count);
 983
 984         /* Two length-prefixed strings for each entry in the map */
 985         for (i = 0; metadata[i][0]; ++i) {
 986                 size_t const key_len = strlen(metadata[i][0]);
 987                 size_t const val_len = strlen(metadata[i][1]);
 988
 989                 ceph_encode_32(&p, key_len);
 990                 memcpy(p, metadata[i][0], key_len);
 991                 p += key_len;
 992                 ceph_encode_32(&p, val_len);
 993                 memcpy(p, metadata[i][1], val_len);
 994                 p += val_len;
 995         }
 996
 997         encode_supported_features(&p, end);
 998         msg->front.iov_len = p - msg->front.iov_base;
 999         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1000
1001         return msg;
1002 }
1003
1004 /*
1005  * send session open request.
1006  *
1007  * called under mdsc->mutex
1008  */
1009 static int __open_session(struct ceph_mds_client *mdsc,
1010                           struct ceph_mds_session *session)
1011 {
1012         struct ceph_msg *msg;
1013         int mstate;
1014         int mds = session->s_mds;
1015
1016         /* wait for mds to go active? */
1017         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
1018         dout("open_session to mds%d (%s)\n", mds,
1019              ceph_mds_state_name(mstate));
1020         session->s_state = CEPH_MDS_SESSION_OPENING;
1021         session->s_renew_requested = jiffies;
1022
1023         /* send connect message */
1024         msg = create_session_open_msg(mdsc, session->s_seq);
1025         if (!msg)
1026                 return -ENOMEM;
1027         ceph_con_send(&session->s_con, msg);
1028         return 0;
1029 }
1030
1031 /*
1032  * open sessions for any export targets for the given mds
1033  *
1034  * called under mdsc->mutex
1035  */
1036 static struct ceph_mds_session *
1037 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
1038 {
1039         struct ceph_mds_session *session;
1040
1041         session = __ceph_lookup_mds_session(mdsc, target);
1042         if (!session) {
1043                 session = register_session(mdsc, target);
1044                 if (IS_ERR(session))
1045                         return session;
1046         }
1047         if (session->s_state == CEPH_MDS_SESSION_NEW ||
1048             session->s_state == CEPH_MDS_SESSION_CLOSING)
1049                 __open_session(mdsc, session);
1050
1051         return session;
1052 }
1053
1054 struct ceph_mds_session *
1055 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1056 {
1057         struct ceph_mds_session *session;
1058
1059         dout("open_export_target_session to mds%d\n", target);
1060
1061         mutex_lock(&mdsc->mutex);
1062         session = __open_export_target_session(mdsc, target);
1063         mutex_unlock(&mdsc->mutex);
1064
1065         return session;
1066 }
1067
1068 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1069                                           struct ceph_mds_session *session)
1070 {
1071         struct ceph_mds_info *mi;
1072         struct ceph_mds_session *ts;
1073         int i, mds = session->s_mds;
1074
1075         if (mds >= mdsc->mdsmap->m_num_mds)
1076                 return;
1077
1078         mi = &mdsc->mdsmap->m_info[mds];
1079         dout("open_export_target_sessions for mds%d (%d targets)\n",
1080              session->s_mds, mi->num_export_targets);
1081
1082         for (i = 0; i < mi->num_export_targets; i++) {
1083                 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1084                 if (!IS_ERR(ts))
1085                         ceph_put_mds_session(ts);
1086         }
1087 }
1088
1089 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1090                                            struct ceph_mds_session *session)
1091 {
1092         mutex_lock(&mdsc->mutex);
1093         __open_export_target_sessions(mdsc, session);
1094         mutex_unlock(&mdsc->mutex);
1095 }
1096
1097 /*
1098  * session caps
1099  */
1100
1101 static void detach_cap_releases(struct ceph_mds_session *session,
1102                                 struct list_head *target)
1103 {
1104         lockdep_assert_held(&session->s_cap_lock);
1105
1106         list_splice_init(&session->s_cap_releases, target);
1107         session->s_num_cap_releases = 0;
1108         dout("dispose_cap_releases mds%d\n", session->s_mds);
1109 }
1110
1111 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1112                                  struct list_head *dispose)
1113 {
1114         while (!list_empty(dispose)) {
1115                 struct ceph_cap *cap;
1116                 /* zero out the in-progress message */
1117                 cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1118                 list_del(&cap->session_caps);
1119                 ceph_put_cap(mdsc, cap);
1120         }
1121 }
1122
1123 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1124                                      struct ceph_mds_session *session)
1125 {
1126         struct ceph_mds_request *req;
1127         struct rb_node *p;
1128
1129         dout("cleanup_session_requests mds%d\n", session->s_mds);
1130         mutex_lock(&mdsc->mutex);
1131         while (!list_empty(&session->s_unsafe)) {
1132                 req = list_first_entry(&session->s_unsafe,
1133                                        struct ceph_mds_request, r_unsafe_item);
1134                 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1135                                     req->r_tid);
1136                 __unregister_request(mdsc, req);
1137         }
1138         /* zero r_attempts, so kick_requests() will re-send requests */
1139         p = rb_first(&mdsc->request_tree);
1140         while (p) {
1141                 req = rb_entry(p, struct ceph_mds_request, r_node);
1142                 p = rb_next(p);
1143                 if (req->r_session &&
1144                     req->r_session->s_mds == session->s_mds)
1145                         req->r_attempts = 0;
1146         }
1147         mutex_unlock(&mdsc->mutex);
1148 }
1149
1150 /*
1151  * Helper to safely iterate over all caps associated with a session, with
1152  * special care taken to handle a racing __ceph_remove_cap().
1153  *
1154  * Caller must hold session s_mutex.
1155  */
1156 static int iterate_session_caps(struct ceph_mds_session *session,
1157                                  int (*cb)(struct inode *, struct ceph_cap *,
1158                                             void *), void *arg)
1159 {
1160         struct list_head *p;
1161         struct ceph_cap *cap;
1162         struct inode *inode, *last_inode = NULL;
1163         struct ceph_cap *old_cap = NULL;
1164         int ret;
1165
1166         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1167         spin_lock(&session->s_cap_lock);
1168         p = session->s_caps.next;
1169         while (p != &session->s_caps) {
1170                 cap = list_entry(p, struct ceph_cap, session_caps);
1171                 inode = igrab(&cap->ci->vfs_inode);
1172                 if (!inode) {
1173                         p = p->next;
1174                         continue;
1175                 }
1176                 session->s_cap_iterator = cap;
1177                 spin_unlock(&session->s_cap_lock);
1178
1179                 if (last_inode) {
1180                         iput(last_inode);
1181                         last_inode = NULL;
1182                 }
1183                 if (old_cap) {
1184                         ceph_put_cap(session->s_mdsc, old_cap);
1185                         old_cap = NULL;
1186                 }
1187
1188                 ret = cb(inode, cap, arg);
1189                 last_inode = inode;
1190
1191                 spin_lock(&session->s_cap_lock);
1192                 p = p->next;
1193                 if (!cap->ci) {
1194                         dout("iterate_session_caps  finishing cap %p removal\n",
1195                              cap);
1196                         BUG_ON(cap->session != session);
1197                         cap->session = NULL;
1198                         list_del_init(&cap->session_caps);
1199                         session->s_nr_caps--;
1200                         if (cap->queue_release) {
1201                                 list_add_tail(&cap->session_caps,
1202                                               &session->s_cap_releases);
1203                                 session->s_num_cap_releases++;
1204                         } else {
1205                                 old_cap = cap;  /* put_cap it w/o locks held */
1206                         }
1207                 }
1208                 if (ret < 0)
1209                         goto out;
1210         }
1211         ret = 0;
1212 out:
1213         session->s_cap_iterator = NULL;
1214         spin_unlock(&session->s_cap_lock);
1215
1216         iput(last_inode);
1217         if (old_cap)
1218                 ceph_put_cap(session->s_mdsc, old_cap);
1219
1220         return ret;
1221 }
1222
1223 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1224                                   void *arg)
1225 {
1226         struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1227         struct ceph_inode_info *ci = ceph_inode(inode);
1228         LIST_HEAD(to_remove);
1229         bool drop = false;
1230         bool invalidate = false;
1231
1232         dout("removing cap %p, ci is %p, inode is %p\n",
1233              cap, ci, &ci->vfs_inode);
1234         spin_lock(&ci->i_ceph_lock);
1235         if (cap->mds_wanted | cap->issued)
1236                 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1237         __ceph_remove_cap(cap, false);
1238         if (!ci->i_auth_cap) {
1239                 struct ceph_cap_flush *cf;
1240                 struct ceph_mds_client *mdsc = fsc->mdsc;
1241
1242                 if (ci->i_wrbuffer_ref > 0 &&
1243                     READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
1244                         invalidate = true;
1245
1246                 while (!list_empty(&ci->i_cap_flush_list)) {
1247                         cf = list_first_entry(&ci->i_cap_flush_list,
1248                                               struct ceph_cap_flush, i_list);
1249                         list_move(&cf->i_list, &to_remove);
1250                 }
1251
1252                 spin_lock(&mdsc->cap_dirty_lock);
1253
1254                 list_for_each_entry(cf, &to_remove, i_list)
1255                         list_del(&cf->g_list);
1256
1257                 if (!list_empty(&ci->i_dirty_item)) {
1258                         pr_warn_ratelimited(
1259                                 " dropping dirty %s state for %p %lld\n",
1260                                 ceph_cap_string(ci->i_dirty_caps),
1261                                 inode, ceph_ino(inode));
1262                         ci->i_dirty_caps = 0;
1263                         list_del_init(&ci->i_dirty_item);
1264                         drop = true;
1265                 }
1266                 if (!list_empty(&ci->i_flushing_item)) {
1267                         pr_warn_ratelimited(
1268                                 " dropping dirty+flushing %s state for %p %lld\n",
1269                                 ceph_cap_string(ci->i_flushing_caps),
1270                                 inode, ceph_ino(inode));
1271                         ci->i_flushing_caps = 0;
1272                         list_del_init(&ci->i_flushing_item);
1273                         mdsc->num_cap_flushing--;
1274                         drop = true;
1275                 }
1276                 spin_unlock(&mdsc->cap_dirty_lock);
1277
1278                 if (atomic_read(&ci->i_filelock_ref) > 0) {
1279                         /* make further file lock syscall return -EIO */
1280                         ci->i_ceph_flags |= CEPH_I_ERROR_FILELOCK;
1281                         pr_warn_ratelimited(" dropping file locks for %p %lld\n",
1282                                             inode, ceph_ino(inode));
1283                 }
1284
1285                 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1286                         list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1287                         ci->i_prealloc_cap_flush = NULL;
1288                 }
1289         }
1290         spin_unlock(&ci->i_ceph_lock);
1291         while (!list_empty(&to_remove)) {
1292                 struct ceph_cap_flush *cf;
1293                 cf = list_first_entry(&to_remove,
1294                                       struct ceph_cap_flush, i_list);
1295                 list_del(&cf->i_list);
1296                 ceph_free_cap_flush(cf);
1297         }
1298
1299         wake_up_all(&ci->i_cap_wq);
1300         if (invalidate)
1301                 ceph_queue_invalidate(inode);
1302         if (drop)
1303                 iput(inode);
1304         return 0;
1305 }
1306
1307 /*
1308  * caller must hold session s_mutex
1309  */
1310 static void remove_session_caps(struct ceph_mds_session *session)
1311 {
1312         struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1313         struct super_block *sb = fsc->sb;
1314         LIST_HEAD(dispose);
1315
1316         dout("remove_session_caps on %p\n", session);
1317         iterate_session_caps(session, remove_session_caps_cb, fsc);
1318
1319         wake_up_all(&fsc->mdsc->cap_flushing_wq);
1320
1321         spin_lock(&session->s_cap_lock);
1322         if (session->s_nr_caps > 0) {
1323                 struct inode *inode;
1324                 struct ceph_cap *cap, *prev = NULL;
1325                 struct ceph_vino vino;
1326                 /*
1327                  * iterate_session_caps() skips inodes that are being
1328                  * deleted, we need to wait until deletions are complete.
1329                  * __wait_on_freeing_inode() is designed for the job,
1330                  * but it is not exported, so use lookup inode function
1331                  * to access it.
1332                  */
1333                 while (!list_empty(&session->s_caps)) {
1334                         cap = list_entry(session->s_caps.next,
1335                                          struct ceph_cap, session_caps);
1336                         if (cap == prev)
1337                                 break;
1338                         prev = cap;
1339                         vino = cap->ci->i_vino;
1340                         spin_unlock(&session->s_cap_lock);
1341
1342                         inode = ceph_find_inode(sb, vino);
1343                         iput(inode);
1344
1345                         spin_lock(&session->s_cap_lock);
1346                 }
1347         }
1348
1349         // drop cap expires and unlock s_cap_lock
1350         detach_cap_releases(session, &dispose);
1351
1352         BUG_ON(session->s_nr_caps > 0);
1353         BUG_ON(!list_empty(&session->s_cap_flushing));
1354         spin_unlock(&session->s_cap_lock);
1355         dispose_cap_releases(session->s_mdsc, &dispose);
1356 }
1357
1358 enum {
1359         RECONNECT,
1360         RENEWCAPS,
1361         FORCE_RO,
1362 };
1363
1364 /*
1365  * wake up any threads waiting on this session's caps.  if the cap is
1366  * old (didn't get renewed on the client reconnect), remove it now.
1367  *
1368  * caller must hold s_mutex.
1369  */
1370 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1371                               void *arg)
1372 {
1373         struct ceph_inode_info *ci = ceph_inode(inode);
1374         unsigned long ev = (unsigned long)arg;
1375
1376         if (ev == RECONNECT) {
1377                 spin_lock(&ci->i_ceph_lock);
1378                 ci->i_wanted_max_size = 0;
1379                 ci->i_requested_max_size = 0;
1380                 spin_unlock(&ci->i_ceph_lock);
1381         } else if (ev == RENEWCAPS) {
1382                 if (cap->cap_gen < cap->session->s_cap_gen) {
1383                         /* mds did not re-issue stale cap */
1384                         spin_lock(&ci->i_ceph_lock);
1385                         cap->issued = cap->implemented = CEPH_CAP_PIN;
1386                         /* make sure mds knows what we want */
1387                         if (__ceph_caps_file_wanted(ci) & ~cap->mds_wanted)
1388                                 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1389                         spin_unlock(&ci->i_ceph_lock);
1390                 }
1391         } else if (ev == FORCE_RO) {
1392         }
1393         wake_up_all(&ci->i_cap_wq);
1394         return 0;
1395 }
1396
1397 static void wake_up_session_caps(struct ceph_mds_session *session, int ev)
1398 {
1399         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1400         iterate_session_caps(session, wake_up_session_cb,
1401                              (void *)(unsigned long)ev);
1402 }
1403
1404 /*
1405  * Send periodic message to MDS renewing all currently held caps.  The
1406  * ack will reset the expiration for all caps from this session.
1407  *
1408  * caller holds s_mutex
1409  */
1410 static int send_renew_caps(struct ceph_mds_client *mdsc,
1411                            struct ceph_mds_session *session)
1412 {
1413         struct ceph_msg *msg;
1414         int state;
1415
1416         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1417             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1418                 pr_info("mds%d caps stale\n", session->s_mds);
1419         session->s_renew_requested = jiffies;
1420
1421         /* do not try to renew caps until a recovering mds has reconnected
1422          * with its clients. */
1423         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1424         if (state < CEPH_MDS_STATE_RECONNECT) {
1425                 dout("send_renew_caps ignoring mds%d (%s)\n",
1426                      session->s_mds, ceph_mds_state_name(state));
1427                 return 0;
1428         }
1429
1430         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1431                 ceph_mds_state_name(state));
1432         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1433                                  ++session->s_renew_seq);
1434         if (!msg)
1435                 return -ENOMEM;
1436         ceph_con_send(&session->s_con, msg);
1437         return 0;
1438 }
1439
1440 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1441                              struct ceph_mds_session *session, u64 seq)
1442 {
1443         struct ceph_msg *msg;
1444
1445         dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1446              session->s_mds, ceph_session_state_name(session->s_state), seq);
1447         msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1448         if (!msg)
1449                 return -ENOMEM;
1450         ceph_con_send(&session->s_con, msg);
1451         return 0;
1452 }
1453
1454
1455 /*
1456  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1457  *
1458  * Called under session->s_mutex
1459  */
1460 static void renewed_caps(struct ceph_mds_client *mdsc,
1461                          struct ceph_mds_session *session, int is_renew)
1462 {
1463         int was_stale;
1464         int wake = 0;
1465
1466         spin_lock(&session->s_cap_lock);
1467         was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1468
1469         session->s_cap_ttl = session->s_renew_requested +
1470                 mdsc->mdsmap->m_session_timeout*HZ;
1471
1472         if (was_stale) {
1473                 if (time_before(jiffies, session->s_cap_ttl)) {
1474                         pr_info("mds%d caps renewed\n", session->s_mds);
1475                         wake = 1;
1476                 } else {
1477                         pr_info("mds%d caps still stale\n", session->s_mds);
1478                 }
1479         }
1480         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1481              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1482              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1483         spin_unlock(&session->s_cap_lock);
1484
1485         if (wake)
1486                 wake_up_session_caps(session, RENEWCAPS);
1487 }
1488
1489 /*
1490  * send a session close request
1491  */
1492 static int request_close_session(struct ceph_mds_client *mdsc,
1493                                  struct ceph_mds_session *session)
1494 {
1495         struct ceph_msg *msg;
1496
1497         dout("request_close_session mds%d state %s seq %lld\n",
1498              session->s_mds, ceph_session_state_name(session->s_state),
1499              session->s_seq);
1500         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1501         if (!msg)
1502                 return -ENOMEM;
1503         ceph_con_send(&session->s_con, msg);
1504         return 1;
1505 }
1506
1507 /*
1508  * Called with s_mutex held.
1509  */
1510 static int __close_session(struct ceph_mds_client *mdsc,
1511                          struct ceph_mds_session *session)
1512 {
1513         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1514                 return 0;
1515         session->s_state = CEPH_MDS_SESSION_CLOSING;
1516         return request_close_session(mdsc, session);
1517 }
1518
1519 static bool drop_negative_children(struct dentry *dentry)
1520 {
1521         struct dentry *child;
1522         bool all_negative = true;
1523
1524         if (!d_is_dir(dentry))
1525                 goto out;
1526
1527         spin_lock(&dentry->d_lock);
1528         list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1529                 if (d_really_is_positive(child)) {
1530                         all_negative = false;
1531                         break;
1532                 }
1533         }
1534         spin_unlock(&dentry->d_lock);
1535
1536         if (all_negative)
1537                 shrink_dcache_parent(dentry);
1538 out:
1539         return all_negative;
1540 }
1541
1542 /*
1543  * Trim old(er) caps.
1544  *
1545  * Because we can't cache an inode without one or more caps, we do
1546  * this indirectly: if a cap is unused, we prune its aliases, at which
1547  * point the inode will hopefully get dropped to.
1548  *
1549  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1550  * memory pressure from the MDS, though, so it needn't be perfect.
1551  */
1552 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1553 {
1554         struct ceph_mds_session *session = arg;
1555         struct ceph_inode_info *ci = ceph_inode(inode);
1556         int used, wanted, oissued, mine;
1557
1558         if (session->s_trim_caps <= 0)
1559                 return -1;
1560
1561         spin_lock(&ci->i_ceph_lock);
1562         mine = cap->issued | cap->implemented;
1563         used = __ceph_caps_used(ci);
1564         wanted = __ceph_caps_file_wanted(ci);
1565         oissued = __ceph_caps_issued_other(ci, cap);
1566
1567         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1568              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1569              ceph_cap_string(used), ceph_cap_string(wanted));
1570         if (cap == ci->i_auth_cap) {
1571                 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1572                     !list_empty(&ci->i_cap_snaps))
1573                         goto out;
1574                 if ((used | wanted) & CEPH_CAP_ANY_WR)
1575                         goto out;
1576                 /* Note: it's possible that i_filelock_ref becomes non-zero
1577                  * after dropping auth caps. It doesn't hurt because reply
1578                  * of lock mds request will re-add auth caps. */
1579                 if (atomic_read(&ci->i_filelock_ref) > 0)
1580                         goto out;
1581         }
1582         /* The inode has cached pages, but it's no longer used.
1583          * we can safely drop it */
1584         if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1585             !(oissued & CEPH_CAP_FILE_CACHE)) {
1586           used = 0;
1587           oissued = 0;
1588         }
1589         if ((used | wanted) & ~oissued & mine)
1590                 goto out;   /* we need these caps */
1591
1592         if (oissued) {
1593                 /* we aren't the only cap.. just remove us */
1594                 __ceph_remove_cap(cap, true);
1595                 session->s_trim_caps--;
1596         } else {
1597                 struct dentry *dentry;
1598                 /* try dropping referring dentries */
1599                 spin_unlock(&ci->i_ceph_lock);
1600                 dentry = d_find_any_alias(inode);
1601                 if (dentry && drop_negative_children(dentry)) {
1602                         int count;
1603                         dput(dentry);
1604                         d_prune_aliases(inode);
1605                         count = atomic_read(&inode->i_count);
1606                         if (count == 1)
1607                                 session->s_trim_caps--;
1608                         dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1609                              inode, cap, count);
1610                 } else {
1611                         dput(dentry);
1612                 }
1613                 return 0;
1614         }
1615
1616 out:
1617         spin_unlock(&ci->i_ceph_lock);
1618         return 0;
1619 }
1620
1621 /*
1622  * Trim session cap count down to some max number.
1623  */
1624 int ceph_trim_caps(struct ceph_mds_client *mdsc,
1625                    struct ceph_mds_session *session,
1626                    int max_caps)
1627 {
1628         int trim_caps = session->s_nr_caps - max_caps;
1629
1630         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1631              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1632         if (trim_caps > 0) {
1633                 session->s_trim_caps = trim_caps;
1634                 iterate_session_caps(session, trim_caps_cb, session);
1635                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1636                      session->s_mds, session->s_nr_caps, max_caps,
1637                         trim_caps - session->s_trim_caps);
1638                 session->s_trim_caps = 0;
1639         }
1640
1641         ceph_send_cap_releases(mdsc, session);
1642         return 0;
1643 }
1644
1645 static int check_caps_flush(struct ceph_mds_client *mdsc,
1646                             u64 want_flush_tid)
1647 {
1648         int ret = 1;
1649
1650         spin_lock(&mdsc->cap_dirty_lock);
1651         if (!list_empty(&mdsc->cap_flush_list)) {
1652                 struct ceph_cap_flush *cf =
1653                         list_first_entry(&mdsc->cap_flush_list,
1654                                          struct ceph_cap_flush, g_list);
1655                 if (cf->tid <= want_flush_tid) {
1656                         dout("check_caps_flush still flushing tid "
1657                              "%llu <= %llu\n", cf->tid, want_flush_tid);
1658                         ret = 0;
1659                 }
1660         }
1661         spin_unlock(&mdsc->cap_dirty_lock);
1662         return ret;
1663 }
1664
1665 /*
1666  * flush all dirty inode data to disk.
1667  *
1668  * returns true if we've flushed through want_flush_tid
1669  */
1670 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1671                             u64 want_flush_tid)
1672 {
1673         dout("check_caps_flush want %llu\n", want_flush_tid);
1674
1675         wait_event(mdsc->cap_flushing_wq,
1676                    check_caps_flush(mdsc, want_flush_tid));
1677
1678         dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1679 }
1680
1681 /*
1682  * called under s_mutex
1683  */
1684 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1685                             struct ceph_mds_session *session)
1686 {
1687         struct ceph_msg *msg = NULL;
1688         struct ceph_mds_cap_release *head;
1689         struct ceph_mds_cap_item *item;
1690         struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
1691         struct ceph_cap *cap;
1692         LIST_HEAD(tmp_list);
1693         int num_cap_releases;
1694         __le32  barrier, *cap_barrier;
1695
1696         down_read(&osdc->lock);
1697         barrier = cpu_to_le32(osdc->epoch_barrier);
1698         up_read(&osdc->lock);
1699
1700         spin_lock(&session->s_cap_lock);
1701 again:
1702         list_splice_init(&session->s_cap_releases, &tmp_list);
1703         num_cap_releases = session->s_num_cap_releases;
1704         session->s_num_cap_releases = 0;
1705         spin_unlock(&session->s_cap_lock);
1706
1707         while (!list_empty(&tmp_list)) {
1708                 if (!msg) {
1709                         msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1710                                         PAGE_SIZE, GFP_NOFS, false);
1711                         if (!msg)
1712                                 goto out_err;
1713                         head = msg->front.iov_base;
1714                         head->num = cpu_to_le32(0);
1715                         msg->front.iov_len = sizeof(*head);
1716
1717                         msg->hdr.version = cpu_to_le16(2);
1718                         msg->hdr.compat_version = cpu_to_le16(1);
1719                 }
1720
1721                 cap = list_first_entry(&tmp_list, struct ceph_cap,
1722                                         session_caps);
1723                 list_del(&cap->session_caps);
1724                 num_cap_releases--;
1725
1726                 head = msg->front.iov_base;
1727                 le32_add_cpu(&head->num, 1);
1728                 item = msg->front.iov_base + msg->front.iov_len;
1729                 item->ino = cpu_to_le64(cap->cap_ino);
1730                 item->cap_id = cpu_to_le64(cap->cap_id);
1731                 item->migrate_seq = cpu_to_le32(cap->mseq);
1732                 item->seq = cpu_to_le32(cap->issue_seq);
1733                 msg->front.iov_len += sizeof(*item);
1734
1735                 ceph_put_cap(mdsc, cap);
1736
1737                 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1738                         // Append cap_barrier field
1739                         cap_barrier = msg->front.iov_base + msg->front.iov_len;
1740                         *cap_barrier = barrier;
1741                         msg->front.iov_len += sizeof(*cap_barrier);
1742
1743                         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1744                         dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1745                         ceph_con_send(&session->s_con, msg);
1746                         msg = NULL;
1747                 }
1748         }
1749
1750         BUG_ON(num_cap_releases != 0);
1751
1752         spin_lock(&session->s_cap_lock);
1753         if (!list_empty(&session->s_cap_releases))
1754                 goto again;
1755         spin_unlock(&session->s_cap_lock);
1756
1757         if (msg) {
1758                 // Append cap_barrier field
1759                 cap_barrier = msg->front.iov_base + msg->front.iov_len;
1760                 *cap_barrier = barrier;
1761                 msg->front.iov_len += sizeof(*cap_barrier);
1762
1763                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1764                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1765                 ceph_con_send(&session->s_con, msg);
1766         }
1767         return;
1768 out_err:
1769         pr_err("send_cap_releases mds%d, failed to allocate message\n",
1770                 session->s_mds);
1771         spin_lock(&session->s_cap_lock);
1772         list_splice(&tmp_list, &session->s_cap_releases);
1773         session->s_num_cap_releases += num_cap_releases;
1774         spin_unlock(&session->s_cap_lock);
1775 }
1776
1777 /*
1778  * requests
1779  */
1780
1781 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1782                                     struct inode *dir)
1783 {
1784         struct ceph_inode_info *ci = ceph_inode(dir);
1785         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1786         struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1787         size_t size = sizeof(struct ceph_mds_reply_dir_entry);
1788         int order, num_entries;
1789
1790         spin_lock(&ci->i_ceph_lock);
1791         num_entries = ci->i_files + ci->i_subdirs;
1792         spin_unlock(&ci->i_ceph_lock);
1793         num_entries = max(num_entries, 1);
1794         num_entries = min(num_entries, opt->max_readdir);
1795
1796         order = get_order(size * num_entries);
1797         while (order >= 0) {
1798                 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
1799                                                              __GFP_NOWARN,
1800                                                              order);
1801                 if (rinfo->dir_entries)
1802                         break;
1803                 order--;
1804         }
1805         if (!rinfo->dir_entries)
1806                 return -ENOMEM;
1807
1808         num_entries = (PAGE_SIZE << order) / size;
1809         num_entries = min(num_entries, opt->max_readdir);
1810
1811         rinfo->dir_buf_size = PAGE_SIZE << order;
1812         req->r_num_caps = num_entries + 1;
1813         req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1814         req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1815         return 0;
1816 }
1817
1818 /*
1819  * Create an mds request.
1820  */
1821 struct ceph_mds_request *
1822 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1823 {
1824         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1825         struct timespec64 ts;
1826
1827         if (!req)
1828                 return ERR_PTR(-ENOMEM);
1829
1830         mutex_init(&req->r_fill_mutex);
1831         req->r_mdsc = mdsc;
1832         req->r_started = jiffies;
1833         req->r_resend_mds = -1;
1834         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1835         INIT_LIST_HEAD(&req->r_unsafe_target_item);
1836         req->r_fmode = -1;
1837         kref_init(&req->r_kref);
1838         RB_CLEAR_NODE(&req->r_node);
1839         INIT_LIST_HEAD(&req->r_wait);
1840         init_completion(&req->r_completion);
1841         init_completion(&req->r_safe_completion);
1842         INIT_LIST_HEAD(&req->r_unsafe_item);
1843
1844         ktime_get_coarse_real_ts64(&ts);
1845         req->r_stamp = timespec64_trunc(ts, mdsc->fsc->sb->s_time_gran);
1846
1847         req->r_op = op;
1848         req->r_direct_mode = mode;
1849         return req;
1850 }
1851
1852 /*
1853  * return oldest (lowest) request, tid in request tree, 0 if none.
1854  *
1855  * called under mdsc->mutex.
1856  */
1857 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1858 {
1859         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1860                 return NULL;
1861         return rb_entry(rb_first(&mdsc->request_tree),
1862                         struct ceph_mds_request, r_node);
1863 }
1864
1865 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1866 {
1867         return mdsc->oldest_tid;
1868 }
1869
1870 /*
1871  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1872  * on build_path_from_dentry in fs/cifs/dir.c.
1873  *
1874  * If @stop_on_nosnap, generate path relative to the first non-snapped
1875  * inode.
1876  *
1877  * Encode hidden .snap dirs as a double /, i.e.
1878  *   foo/.snap/bar -> foo//bar
1879  */
1880 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1881                            int stop_on_nosnap)
1882 {
1883         struct dentry *temp;
1884         char *path;
1885         int len, pos;
1886         unsigned seq;
1887
1888         if (!dentry)
1889                 return ERR_PTR(-EINVAL);
1890
1891 retry:
1892         len = 0;
1893         seq = read_seqbegin(&rename_lock);
1894         rcu_read_lock();
1895         for (temp = dentry; !IS_ROOT(temp);) {
1896                 struct inode *inode = d_inode(temp);
1897                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1898                         len++;  /* slash only */
1899                 else if (stop_on_nosnap && inode &&
1900                          ceph_snap(inode) == CEPH_NOSNAP)
1901                         break;
1902                 else
1903                         len += 1 + temp->d_name.len;
1904                 temp = temp->d_parent;
1905         }
1906         rcu_read_unlock();
1907         if (len)
1908                 len--;  /* no leading '/' */
1909
1910         path = kmalloc(len+1, GFP_NOFS);
1911         if (!path)
1912                 return ERR_PTR(-ENOMEM);
1913         pos = len;
1914         path[pos] = 0;  /* trailing null */
1915         rcu_read_lock();
1916         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1917                 struct inode *inode;
1918
1919                 spin_lock(&temp->d_lock);
1920                 inode = d_inode(temp);
1921                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1922                         dout("build_path path+%d: %p SNAPDIR\n",
1923                              pos, temp);
1924                 } else if (stop_on_nosnap && inode &&
1925                            ceph_snap(inode) == CEPH_NOSNAP) {
1926                         spin_unlock(&temp->d_lock);
1927                         break;
1928                 } else {
1929                         pos -= temp->d_name.len;
1930                         if (pos < 0) {
1931                                 spin_unlock(&temp->d_lock);
1932                                 break;
1933                         }
1934                         strncpy(path + pos, temp->d_name.name,
1935                                 temp->d_name.len);
1936                 }
1937                 spin_unlock(&temp->d_lock);
1938                 if (pos)
1939                         path[--pos] = '/';
1940                 temp = temp->d_parent;
1941         }
1942         rcu_read_unlock();
1943         if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1944                 pr_err("build_path did not end path lookup where "
1945                        "expected, namelen is %d, pos is %d\n", len, pos);
1946                 /* presumably this is only possible if racing with a
1947                    rename of one of the parent directories (we can not
1948                    lock the dentries above us to prevent this, but
1949                    retrying should be harmless) */
1950                 kfree(path);
1951                 goto retry;
1952         }
1953
1954         *base = ceph_ino(d_inode(temp));
1955         *plen = len;
1956         dout("build_path on %p %d built %llx '%.*s'\n",
1957              dentry, d_count(dentry), *base, len, path);
1958         return path;
1959 }
1960
1961 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
1962                              const char **ppath, int *ppathlen, u64 *pino,
1963                              int *pfreepath)
1964 {
1965         char *path;
1966
1967         rcu_read_lock();
1968         if (!dir)
1969                 dir = d_inode_rcu(dentry->d_parent);
1970         if (dir && ceph_snap(dir) == CEPH_NOSNAP) {
1971                 *pino = ceph_ino(dir);
1972                 rcu_read_unlock();
1973                 *ppath = dentry->d_name.name;
1974                 *ppathlen = dentry->d_name.len;
1975                 return 0;
1976         }
1977         rcu_read_unlock();
1978         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1979         if (IS_ERR(path))
1980                 return PTR_ERR(path);
1981         *ppath = path;
1982         *pfreepath = 1;
1983         return 0;
1984 }
1985
1986 static int build_inode_path(struct inode *inode,
1987                             const char **ppath, int *ppathlen, u64 *pino,
1988                             int *pfreepath)
1989 {
1990         struct dentry *dentry;
1991         char *path;
1992
1993         if (ceph_snap(inode) == CEPH_NOSNAP) {
1994                 *pino = ceph_ino(inode);
1995                 *ppathlen = 0;
1996                 return 0;
1997         }
1998         dentry = d_find_alias(inode);
1999         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2000         dput(dentry);
2001         if (IS_ERR(path))
2002                 return PTR_ERR(path);
2003         *ppath = path;
2004         *pfreepath = 1;
2005         return 0;
2006 }
2007
2008 /*
2009  * request arguments may be specified via an inode *, a dentry *, or
2010  * an explicit ino+path.
2011  */
2012 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
2013                                   struct inode *rdiri, const char *rpath,
2014                                   u64 rino, const char **ppath, int *pathlen,
2015                                   u64 *ino, int *freepath)
2016 {
2017         int r = 0;
2018
2019         if (rinode) {
2020                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2021                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2022                      ceph_snap(rinode));
2023         } else if (rdentry) {
2024                 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
2025                                         freepath);
2026                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
2027                      *ppath);
2028         } else if (rpath || rino) {
2029                 *ino = rino;
2030                 *ppath = rpath;
2031                 *pathlen = rpath ? strlen(rpath) : 0;
2032                 dout(" path %.*s\n", *pathlen, rpath);
2033         }
2034
2035         return r;
2036 }
2037
2038 /*
2039  * called under mdsc->mutex
2040  */
2041 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
2042                                                struct ceph_mds_request *req,
2043                                                int mds, bool drop_cap_releases)
2044 {
2045         struct ceph_msg *msg;
2046         struct ceph_mds_request_head *head;
2047         const char *path1 = NULL;
2048         const char *path2 = NULL;
2049         u64 ino1 = 0, ino2 = 0;
2050         int pathlen1 = 0, pathlen2 = 0;
2051         int freepath1 = 0, freepath2 = 0;
2052         int len;
2053         u16 releases;
2054         void *p, *end;
2055         int ret;
2056
2057         ret = set_request_path_attr(req->r_inode, req->r_dentry,
2058                               req->r_parent, req->r_path1, req->r_ino1.ino,
2059                               &path1, &pathlen1, &ino1, &freepath1);
2060         if (ret < 0) {
2061                 msg = ERR_PTR(ret);
2062                 goto out;
2063         }
2064
2065         ret = set_request_path_attr(NULL, req->r_old_dentry,
2066                               req->r_old_dentry_dir,
2067                               req->r_path2, req->r_ino2.ino,
2068                               &path2, &pathlen2, &ino2, &freepath2);
2069         if (ret < 0) {
2070                 msg = ERR_PTR(ret);
2071                 goto out_free1;
2072         }
2073
2074         len = sizeof(*head) +
2075                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2076                 sizeof(struct ceph_timespec);
2077
2078         /* calculate (max) length for cap releases */
2079         len += sizeof(struct ceph_mds_request_release) *
2080                 (!!req->r_inode_drop + !!req->r_dentry_drop +
2081                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2082         if (req->r_dentry_drop)
2083                 len += req->r_dentry->d_name.len;
2084         if (req->r_old_dentry_drop)
2085                 len += req->r_old_dentry->d_name.len;
2086
2087         msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
2088         if (!msg) {
2089                 msg = ERR_PTR(-ENOMEM);
2090                 goto out_free2;
2091         }
2092
2093         msg->hdr.version = cpu_to_le16(2);
2094         msg->hdr.tid = cpu_to_le64(req->r_tid);
2095
2096         head = msg->front.iov_base;
2097         p = msg->front.iov_base + sizeof(*head);
2098         end = msg->front.iov_base + msg->front.iov_len;
2099
2100         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2101         head->op = cpu_to_le32(req->r_op);
2102         head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
2103         head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
2104         head->args = req->r_args;
2105
2106         ceph_encode_filepath(&p, end, ino1, path1);
2107         ceph_encode_filepath(&p, end, ino2, path2);
2108
2109         /* make note of release offset, in case we need to replay */
2110         req->r_request_release_offset = p - msg->front.iov_base;
2111
2112         /* cap releases */
2113         releases = 0;
2114         if (req->r_inode_drop)
2115                 releases += ceph_encode_inode_release(&p,
2116                       req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2117                       mds, req->r_inode_drop, req->r_inode_unless, 0);
2118         if (req->r_dentry_drop)
2119                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
2120                                 req->r_parent, mds, req->r_dentry_drop,
2121                                 req->r_dentry_unless);
2122         if (req->r_old_dentry_drop)
2123                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2124                                 req->r_old_dentry_dir, mds,
2125                                 req->r_old_dentry_drop,
2126                                 req->r_old_dentry_unless);
2127         if (req->r_old_inode_drop)
2128                 releases += ceph_encode_inode_release(&p,
2129                       d_inode(req->r_old_dentry),
2130                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2131
2132         if (drop_cap_releases) {
2133                 releases = 0;
2134                 p = msg->front.iov_base + req->r_request_release_offset;
2135         }
2136
2137         head->num_releases = cpu_to_le16(releases);
2138
2139         /* time stamp */
2140         {
2141                 struct ceph_timespec ts;
2142                 ceph_encode_timespec64(&ts, &req->r_stamp);
2143                 ceph_encode_copy(&p, &ts, sizeof(ts));
2144         }
2145
2146         BUG_ON(p > end);
2147         msg->front.iov_len = p - msg->front.iov_base;
2148         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2149
2150         if (req->r_pagelist) {
2151                 struct ceph_pagelist *pagelist = req->r_pagelist;
2152                 ceph_msg_data_add_pagelist(msg, pagelist);
2153                 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2154         } else {
2155                 msg->hdr.data_len = 0;
2156         }
2157
2158         msg->hdr.data_off = cpu_to_le16(0);
2159
2160 out_free2:
2161         if (freepath2)
2162                 kfree((char *)path2);
2163 out_free1:
2164         if (freepath1)
2165                 kfree((char *)path1);
2166 out:
2167         return msg;
2168 }
2169
2170 /*
2171  * called under mdsc->mutex if error, under no mutex if
2172  * success.
2173  */
2174 static void complete_request(struct ceph_mds_client *mdsc,
2175                              struct ceph_mds_request *req)
2176 {
2177         if (req->r_callback)
2178                 req->r_callback(mdsc, req);
2179         else
2180                 complete_all(&req->r_completion);
2181 }
2182
2183 /*
2184  * called under mdsc->mutex
2185  */
2186 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2187                                   struct ceph_mds_request *req,
2188                                   int mds, bool drop_cap_releases)
2189 {
2190         struct ceph_mds_request_head *rhead;
2191         struct ceph_msg *msg;
2192         int flags = 0;
2193
2194         req->r_attempts++;
2195         if (req->r_inode) {
2196                 struct ceph_cap *cap =
2197                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2198
2199                 if (cap)
2200                         req->r_sent_on_mseq = cap->mseq;
2201                 else
2202                         req->r_sent_on_mseq = -1;
2203         }
2204         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2205              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2206
2207         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2208                 void *p;
2209                 /*
2210                  * Replay.  Do not regenerate message (and rebuild
2211                  * paths, etc.); just use the original message.
2212                  * Rebuilding paths will break for renames because
2213                  * d_move mangles the src name.
2214                  */
2215                 msg = req->r_request;
2216                 rhead = msg->front.iov_base;
2217
2218                 flags = le32_to_cpu(rhead->flags);
2219                 flags |= CEPH_MDS_FLAG_REPLAY;
2220                 rhead->flags = cpu_to_le32(flags);
2221
2222                 if (req->r_target_inode)
2223                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2224
2225                 rhead->num_retry = req->r_attempts - 1;
2226
2227                 /* remove cap/dentry releases from message */
2228                 rhead->num_releases = 0;
2229
2230                 /* time stamp */
2231                 p = msg->front.iov_base + req->r_request_release_offset;
2232                 {
2233                         struct ceph_timespec ts;
2234                         ceph_encode_timespec64(&ts, &req->r_stamp);
2235                         ceph_encode_copy(&p, &ts, sizeof(ts));
2236                 }
2237
2238                 msg->front.iov_len = p - msg->front.iov_base;
2239                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2240                 return 0;
2241         }
2242
2243         if (req->r_request) {
2244                 ceph_msg_put(req->r_request);
2245                 req->r_request = NULL;
2246         }
2247         msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2248         if (IS_ERR(msg)) {
2249                 req->r_err = PTR_ERR(msg);
2250                 return PTR_ERR(msg);
2251         }
2252         req->r_request = msg;
2253
2254         rhead = msg->front.iov_base;
2255         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2256         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2257                 flags |= CEPH_MDS_FLAG_REPLAY;
2258         if (req->r_parent)
2259                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2260         rhead->flags = cpu_to_le32(flags);
2261         rhead->num_fwd = req->r_num_fwd;
2262         rhead->num_retry = req->r_attempts - 1;
2263         rhead->ino = 0;
2264
2265         dout(" r_parent = %p\n", req->r_parent);
2266         return 0;
2267 }
2268
2269 /*
2270  * send request, or put it on the appropriate wait list.
2271  */
2272 static void __do_request(struct ceph_mds_client *mdsc,
2273                         struct ceph_mds_request *req)
2274 {
2275         struct ceph_mds_session *session = NULL;
2276         int mds = -1;
2277         int err = 0;
2278
2279         if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2280                 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2281                         __unregister_request(mdsc, req);
2282                 return;
2283         }
2284
2285         if (req->r_timeout &&
2286             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2287                 dout("do_request timed out\n");
2288                 err = -EIO;
2289                 goto finish;
2290         }
2291         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2292                 dout("do_request forced umount\n");
2293                 err = -EIO;
2294                 goto finish;
2295         }
2296         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2297                 if (mdsc->mdsmap_err) {
2298                         err = mdsc->mdsmap_err;
2299                         dout("do_request mdsmap err %d\n", err);
2300                         goto finish;
2301                 }
2302                 if (mdsc->mdsmap->m_epoch == 0) {
2303                         dout("do_request no mdsmap, waiting for map\n");
2304                         list_add(&req->r_wait, &mdsc->waiting_for_map);
2305                         return;
2306                 }
2307                 if (!(mdsc->fsc->mount_options->flags &
2308                       CEPH_MOUNT_OPT_MOUNTWAIT) &&
2309                     !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2310                         err = -ENOENT;
2311                         pr_info("probably no mds server is up\n");
2312                         goto finish;
2313                 }
2314         }
2315
2316         put_request_session(req);
2317
2318         mds = __choose_mds(mdsc, req);
2319         if (mds < 0 ||
2320             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2321                 dout("do_request no mds or not active, waiting for map\n");
2322                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2323                 return;
2324         }
2325
2326         /* get, open session */
2327         session = __ceph_lookup_mds_session(mdsc, mds);
2328         if (!session) {
2329                 session = register_session(mdsc, mds);
2330                 if (IS_ERR(session)) {
2331                         err = PTR_ERR(session);
2332                         goto finish;
2333                 }
2334         }
2335         req->r_session = get_session(session);
2336
2337         dout("do_request mds%d session %p state %s\n", mds, session,
2338              ceph_session_state_name(session->s_state));
2339         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2340             session->s_state != CEPH_MDS_SESSION_HUNG) {
2341                 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2342                         err = -EACCES;
2343                         goto out_session;
2344                 }
2345                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2346                     session->s_state == CEPH_MDS_SESSION_CLOSING)
2347                         __open_session(mdsc, session);
2348                 list_add(&req->r_wait, &session->s_waiting);
2349                 goto out_session;
2350         }
2351
2352         /* send request */
2353         req->r_resend_mds = -1;   /* forget any previous mds hint */
2354
2355         if (req->r_request_started == 0)   /* note request start time */
2356                 req->r_request_started = jiffies;
2357
2358         err = __prepare_send_request(mdsc, req, mds, false);
2359         if (!err) {
2360                 ceph_msg_get(req->r_request);
2361                 ceph_con_send(&session->s_con, req->r_request);
2362         }
2363
2364 out_session:
2365         ceph_put_mds_session(session);
2366 finish:
2367         if (err) {
2368                 dout("__do_request early error %d\n", err);
2369                 req->r_err = err;
2370                 complete_request(mdsc, req);
2371                 __unregister_request(mdsc, req);
2372         }
2373         return;
2374 }
2375
2376 /*
2377  * called under mdsc->mutex
2378  */
2379 static void __wake_requests(struct ceph_mds_client *mdsc,
2380                             struct list_head *head)
2381 {
2382         struct ceph_mds_request *req;
2383         LIST_HEAD(tmp_list);
2384
2385         list_splice_init(head, &tmp_list);
2386
2387         while (!list_empty(&tmp_list)) {
2388                 req = list_entry(tmp_list.next,
2389                                  struct ceph_mds_request, r_wait);
2390                 list_del_init(&req->r_wait);
2391                 dout(" wake request %p tid %llu\n", req, req->r_tid);
2392                 __do_request(mdsc, req);
2393         }
2394 }
2395
2396 /*
2397  * Wake up threads with requests pending for @mds, so that they can
2398  * resubmit their requests to a possibly different mds.
2399  */
2400 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2401 {
2402         struct ceph_mds_request *req;
2403         struct rb_node *p = rb_first(&mdsc->request_tree);
2404
2405         dout("kick_requests mds%d\n", mds);
2406         while (p) {
2407                 req = rb_entry(p, struct ceph_mds_request, r_node);
2408                 p = rb_next(p);
2409                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2410                         continue;
2411                 if (req->r_attempts > 0)
2412                         continue; /* only new requests */
2413                 if (req->r_session &&
2414                     req->r_session->s_mds == mds) {
2415                         dout(" kicking tid %llu\n", req->r_tid);
2416                         list_del_init(&req->r_wait);
2417                         __do_request(mdsc, req);
2418                 }
2419         }
2420 }
2421
2422 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2423                               struct ceph_mds_request *req)
2424 {
2425         dout("submit_request on %p\n", req);
2426         mutex_lock(&mdsc->mutex);
2427         __register_request(mdsc, req, NULL);
2428         __do_request(mdsc, req);
2429         mutex_unlock(&mdsc->mutex);
2430 }
2431
2432 /*
2433  * Synchrously perform an mds request.  Take care of all of the
2434  * session setup, forwarding, retry details.
2435  */
2436 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2437                          struct inode *dir,
2438                          struct ceph_mds_request *req)
2439 {
2440         int err;
2441
2442         dout("do_request on %p\n", req);
2443
2444         /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2445         if (req->r_inode)
2446                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2447         if (req->r_parent)
2448                 ceph_get_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
2449         if (req->r_old_dentry_dir)
2450                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2451                                   CEPH_CAP_PIN);
2452
2453         /* issue */
2454         mutex_lock(&mdsc->mutex);
2455         __register_request(mdsc, req, dir);
2456         __do_request(mdsc, req);
2457
2458         if (req->r_err) {
2459                 err = req->r_err;
2460                 goto out;
2461         }
2462
2463         /* wait */
2464         mutex_unlock(&mdsc->mutex);
2465         dout("do_request waiting\n");
2466         if (!req->r_timeout && req->r_wait_for_completion) {
2467                 err = req->r_wait_for_completion(mdsc, req);
2468         } else {
2469                 long timeleft = wait_for_completion_killable_timeout(
2470                                         &req->r_completion,
2471                                         ceph_timeout_jiffies(req->r_timeout));
2472                 if (timeleft > 0)
2473                         err = 0;
2474                 else if (!timeleft)
2475                         err = -EIO;  /* timed out */
2476                 else
2477                         err = timeleft;  /* killed */
2478         }
2479         dout("do_request waited, got %d\n", err);
2480         mutex_lock(&mdsc->mutex);
2481
2482         /* only abort if we didn't race with a real reply */
2483         if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2484                 err = le32_to_cpu(req->r_reply_info.head->result);
2485         } else if (err < 0) {
2486                 dout("aborted request %lld with %d\n", req->r_tid, err);
2487
2488                 /*
2489                  * ensure we aren't running concurrently with
2490                  * ceph_fill_trace or ceph_readdir_prepopulate, which
2491                  * rely on locks (dir mutex) held by our caller.
2492                  */
2493                 mutex_lock(&req->r_fill_mutex);
2494                 req->r_err = err;
2495                 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2496                 mutex_unlock(&req->r_fill_mutex);
2497
2498                 if (req->r_parent &&
2499                     (req->r_op & CEPH_MDS_OP_WRITE))
2500                         ceph_invalidate_dir_request(req);
2501         } else {
2502                 err = req->r_err;
2503         }
2504
2505 out:
2506         mutex_unlock(&mdsc->mutex);
2507         dout("do_request %p done, result %d\n", req, err);
2508         return err;
2509 }
2510
2511 /*
2512  * Invalidate dir's completeness, dentry lease state on an aborted MDS
2513  * namespace request.
2514  */
2515 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2516 {
2517         struct inode *dir = req->r_parent;
2518         struct inode *old_dir = req->r_old_dentry_dir;
2519
2520         dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
2521
2522         ceph_dir_clear_complete(dir);
2523         if (old_dir)
2524                 ceph_dir_clear_complete(old_dir);
2525         if (req->r_dentry)
2526                 ceph_invalidate_dentry_lease(req->r_dentry);
2527         if (req->r_old_dentry)
2528                 ceph_invalidate_dentry_lease(req->r_old_dentry);
2529 }
2530
2531 /*
2532  * Handle mds reply.
2533  *
2534  * We take the session mutex and parse and process the reply immediately.
2535  * This preserves the logical ordering of replies, capabilities, etc., sent
2536  * by the MDS as they are applied to our local cache.
2537  */
2538 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2539 {
2540         struct ceph_mds_client *mdsc = session->s_mdsc;
2541         struct ceph_mds_request *req;
2542         struct ceph_mds_reply_head *head = msg->front.iov_base;
2543         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2544         struct ceph_snap_realm *realm;
2545         u64 tid;
2546         int err, result;
2547         int mds = session->s_mds;
2548
2549         if (msg->front.iov_len < sizeof(*head)) {
2550                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2551                 ceph_msg_dump(msg);
2552                 return;
2553         }
2554
2555         /* get request, session */
2556         tid = le64_to_cpu(msg->hdr.tid);
2557         mutex_lock(&mdsc->mutex);
2558         req = lookup_get_request(mdsc, tid);
2559         if (!req) {
2560                 dout("handle_reply on unknown tid %llu\n", tid);
2561                 mutex_unlock(&mdsc->mutex);
2562                 return;
2563         }
2564         dout("handle_reply %p\n", req);
2565
2566         /* correct session? */
2567         if (req->r_session != session) {
2568                 pr_err("mdsc_handle_reply got %llu on session mds%d"
2569                        " not mds%d\n", tid, session->s_mds,
2570                        req->r_session ? req->r_session->s_mds : -1);
2571                 mutex_unlock(&mdsc->mutex);
2572                 goto out;
2573         }
2574
2575         /* dup? */
2576         if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
2577             (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
2578                 pr_warn("got a dup %s reply on %llu from mds%d\n",
2579                            head->safe ? "safe" : "unsafe", tid, mds);
2580                 mutex_unlock(&mdsc->mutex);
2581                 goto out;
2582         }
2583         if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
2584                 pr_warn("got unsafe after safe on %llu from mds%d\n",
2585                            tid, mds);
2586                 mutex_unlock(&mdsc->mutex);
2587                 goto out;
2588         }
2589
2590         result = le32_to_cpu(head->result);
2591
2592         /*
2593          * Handle an ESTALE
2594          * if we're not talking to the authority, send to them
2595          * if the authority has changed while we weren't looking,
2596          * send to new authority
2597          * Otherwise we just have to return an ESTALE
2598          */
2599         if (result == -ESTALE) {
2600                 dout("got ESTALE on request %llu\n", req->r_tid);
2601                 req->r_resend_mds = -1;
2602                 if (req->r_direct_mode != USE_AUTH_MDS) {
2603                         dout("not using auth, setting for that now\n");
2604                         req->r_direct_mode = USE_AUTH_MDS;
2605                         __do_request(mdsc, req);
2606                         mutex_unlock(&mdsc->mutex);
2607                         goto out;
2608                 } else  {
2609                         int mds = __choose_mds(mdsc, req);
2610                         if (mds >= 0 && mds != req->r_session->s_mds) {
2611                                 dout("but auth changed, so resending\n");
2612                                 __do_request(mdsc, req);
2613                                 mutex_unlock(&mdsc->mutex);
2614                                 goto out;
2615                         }
2616                 }
2617                 dout("have to return ESTALE on request %llu\n", req->r_tid);
2618         }
2619
2620
2621         if (head->safe) {
2622                 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
2623                 __unregister_request(mdsc, req);
2624
2625                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2626                         /*
2627                          * We already handled the unsafe response, now do the
2628                          * cleanup.  No need to examine the response; the MDS
2629                          * doesn't include any result info in the safe
2630                          * response.  And even if it did, there is nothing
2631                          * useful we could do with a revised return value.
2632                          */
2633                         dout("got safe reply %llu, mds%d\n", tid, mds);
2634
2635                         /* last unsafe request during umount? */
2636                         if (mdsc->stopping && !__get_oldest_req(mdsc))
2637                                 complete_all(&mdsc->safe_umount_waiters);
2638                         mutex_unlock(&mdsc->mutex);
2639                         goto out;
2640                 }
2641         } else {
2642                 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
2643                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2644                 if (req->r_unsafe_dir) {
2645                         struct ceph_inode_info *ci =
2646                                         ceph_inode(req->r_unsafe_dir);
2647                         spin_lock(&ci->i_unsafe_lock);
2648                         list_add_tail(&req->r_unsafe_dir_item,
2649                                       &ci->i_unsafe_dirops);
2650                         spin_unlock(&ci->i_unsafe_lock);
2651                 }
2652         }
2653
2654         dout("handle_reply tid %lld result %d\n", tid, result);
2655         rinfo = &req->r_reply_info;
2656         err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2657         mutex_unlock(&mdsc->mutex);
2658
2659         mutex_lock(&session->s_mutex);
2660         if (err < 0) {
2661                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2662                 ceph_msg_dump(msg);
2663                 goto out_err;
2664         }
2665
2666         /* snap trace */
2667         realm = NULL;
2668         if (rinfo->snapblob_len) {
2669                 down_write(&mdsc->snap_rwsem);
2670                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2671                                 rinfo->snapblob + rinfo->snapblob_len,
2672                                 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2673                                 &realm);
2674                 downgrade_write(&mdsc->snap_rwsem);
2675         } else {
2676                 down_read(&mdsc->snap_rwsem);
2677         }
2678
2679         /* insert trace into our cache */
2680         mutex_lock(&req->r_fill_mutex);
2681         current->journal_info = req;
2682         err = ceph_fill_trace(mdsc->fsc->sb, req);
2683         if (err == 0) {
2684                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2685                                     req->r_op == CEPH_MDS_OP_LSSNAP))
2686                         ceph_readdir_prepopulate(req, req->r_session);
2687                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2688         }
2689         current->journal_info = NULL;
2690         mutex_unlock(&req->r_fill_mutex);
2691
2692         up_read(&mdsc->snap_rwsem);
2693         if (realm)
2694                 ceph_put_snap_realm(mdsc, realm);
2695
2696         if (err == 0 && req->r_target_inode &&
2697             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2698                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
2699                 spin_lock(&ci->i_unsafe_lock);
2700                 list_add_tail(&req->r_unsafe_target_item, &ci->i_unsafe_iops);
2701                 spin_unlock(&ci->i_unsafe_lock);
2702         }
2703 out_err:
2704         mutex_lock(&mdsc->mutex);
2705         if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2706                 if (err) {
2707                         req->r_err = err;
2708                 } else {
2709                         req->r_reply =  ceph_msg_get(msg);
2710                         set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
2711                 }
2712         } else {
2713                 dout("reply arrived after request %lld was aborted\n", tid);
2714         }
2715         mutex_unlock(&mdsc->mutex);
2716
2717         mutex_unlock(&session->s_mutex);
2718
2719         /* kick calling process */
2720         complete_request(mdsc, req);
2721 out:
2722         ceph_mdsc_put_request(req);
2723         return;
2724 }
2725
2726
2727
2728 /*
2729  * handle mds notification that our request has been forwarded.
2730  */
2731 static void handle_forward(struct ceph_mds_client *mdsc,
2732                            struct ceph_mds_session *session,
2733                            struct ceph_msg *msg)
2734 {
2735         struct ceph_mds_request *req;
2736         u64 tid = le64_to_cpu(msg->hdr.tid);
2737         u32 next_mds;
2738         u32 fwd_seq;
2739         int err = -EINVAL;
2740         void *p = msg->front.iov_base;
2741         void *end = p + msg->front.iov_len;
2742
2743         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2744         next_mds = ceph_decode_32(&p);
2745         fwd_seq = ceph_decode_32(&p);
2746
2747         mutex_lock(&mdsc->mutex);
2748         req = lookup_get_request(mdsc, tid);
2749         if (!req) {
2750                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2751                 goto out;  /* dup reply? */
2752         }
2753
2754         if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2755                 dout("forward tid %llu aborted, unregistering\n", tid);
2756                 __unregister_request(mdsc, req);
2757         } else if (fwd_seq <= req->r_num_fwd) {
2758                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2759                      tid, next_mds, req->r_num_fwd, fwd_seq);
2760         } else {
2761                 /* resend. forward race not possible; mds would drop */
2762                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2763                 BUG_ON(req->r_err);
2764                 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
2765                 req->r_attempts = 0;
2766                 req->r_num_fwd = fwd_seq;
2767                 req->r_resend_mds = next_mds;
2768                 put_request_session(req);
2769                 __do_request(mdsc, req);
2770         }
2771         ceph_mdsc_put_request(req);
2772 out:
2773         mutex_unlock(&mdsc->mutex);
2774         return;
2775
2776 bad:
2777         pr_err("mdsc_handle_forward decode error err=%d\n", err);
2778 }
2779
2780 static int __decode_and_drop_session_metadata(void **p, void *end)
2781 {
2782         /* map<string,string> */
2783         u32 n;
2784         ceph_decode_32_safe(p, end, n, bad);
2785         while (n-- > 0) {
2786                 u32 len;
2787                 ceph_decode_32_safe(p, end, len, bad);
2788                 ceph_decode_need(p, end, len, bad);
2789                 *p += len;
2790                 ceph_decode_32_safe(p, end, len, bad);
2791                 ceph_decode_need(p, end, len, bad);
2792                 *p += len;
2793         }
2794         return 0;
2795 bad:
2796         return -1;
2797 }
2798
2799 /*
2800  * handle a mds session control message
2801  */
2802 static void handle_session(struct ceph_mds_session *session,
2803                            struct ceph_msg *msg)
2804 {
2805         struct ceph_mds_client *mdsc = session->s_mdsc;
2806         int mds = session->s_mds;
2807         int msg_version = le16_to_cpu(msg->hdr.version);
2808         void *p = msg->front.iov_base;
2809         void *end = p + msg->front.iov_len;
2810         struct ceph_mds_session_head *h;
2811         u32 op;
2812         u64 seq;
2813         unsigned long features = 0;
2814         int wake = 0;
2815
2816         /* decode */
2817         ceph_decode_need(&p, end, sizeof(*h), bad);
2818         h = p;
2819         p += sizeof(*h);
2820
2821         op = le32_to_cpu(h->op);
2822         seq = le64_to_cpu(h->seq);
2823
2824         if (msg_version >= 3) {
2825                 u32 len;
2826                 /* version >= 2, metadata */
2827                 if (__decode_and_drop_session_metadata(&p, end) < 0)
2828                         goto bad;
2829                 /* version >= 3, feature bits */
2830                 ceph_decode_32_safe(&p, end, len, bad);
2831                 ceph_decode_need(&p, end, len, bad);
2832                 memcpy(&features, p, min_t(size_t, len, sizeof(features)));
2833                 p += len;
2834         }
2835
2836         mutex_lock(&mdsc->mutex);
2837         if (op == CEPH_SESSION_CLOSE) {
2838                 get_session(session);
2839                 __unregister_session(mdsc, session);
2840         }
2841         /* FIXME: this ttl calculation is generous */
2842         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2843         mutex_unlock(&mdsc->mutex);
2844
2845         mutex_lock(&session->s_mutex);
2846
2847         dout("handle_session mds%d %s %p state %s seq %llu\n",
2848              mds, ceph_session_op_name(op), session,
2849              ceph_session_state_name(session->s_state), seq);
2850
2851         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2852                 session->s_state = CEPH_MDS_SESSION_OPEN;
2853                 pr_info("mds%d came back\n", session->s_mds);
2854         }
2855
2856         switch (op) {
2857         case CEPH_SESSION_OPEN:
2858                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2859                         pr_info("mds%d reconnect success\n", session->s_mds);
2860                 session->s_state = CEPH_MDS_SESSION_OPEN;
2861                 session->s_features = features;
2862                 renewed_caps(mdsc, session, 0);
2863                 wake = 1;
2864                 if (mdsc->stopping)
2865                         __close_session(mdsc, session);
2866                 break;
2867
2868         case CEPH_SESSION_RENEWCAPS:
2869                 if (session->s_renew_seq == seq)
2870                         renewed_caps(mdsc, session, 1);
2871                 break;
2872
2873         case CEPH_SESSION_CLOSE:
2874                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2875                         pr_info("mds%d reconnect denied\n", session->s_mds);
2876                 cleanup_session_requests(mdsc, session);
2877                 remove_session_caps(session);
2878                 wake = 2; /* for good measure */
2879                 wake_up_all(&mdsc->session_close_wq);
2880                 break;
2881
2882         case CEPH_SESSION_STALE:
2883                 pr_info("mds%d caps went stale, renewing\n",
2884                         session->s_mds);
2885                 spin_lock(&session->s_gen_ttl_lock);
2886                 session->s_cap_gen++;
2887                 session->s_cap_ttl = jiffies - 1;
2888                 spin_unlock(&session->s_gen_ttl_lock);
2889                 send_renew_caps(mdsc, session);
2890                 break;
2891
2892         case CEPH_SESSION_RECALL_STATE:
2893                 ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2894                 break;
2895
2896         case CEPH_SESSION_FLUSHMSG:
2897                 send_flushmsg_ack(mdsc, session, seq);
2898                 break;
2899
2900         case CEPH_SESSION_FORCE_RO:
2901                 dout("force_session_readonly %p\n", session);
2902                 spin_lock(&session->s_cap_lock);
2903                 session->s_readonly = true;
2904                 spin_unlock(&session->s_cap_lock);
2905                 wake_up_session_caps(session, FORCE_RO);
2906                 break;
2907
2908         case CEPH_SESSION_REJECT:
2909                 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
2910                 pr_info("mds%d rejected session\n", session->s_mds);
2911                 session->s_state = CEPH_MDS_SESSION_REJECTED;
2912                 cleanup_session_requests(mdsc, session);
2913                 remove_session_caps(session);
2914                 wake = 2; /* for good measure */
2915                 break;
2916
2917         default:
2918                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2919                 WARN_ON(1);
2920         }
2921
2922         mutex_unlock(&session->s_mutex);
2923         if (wake) {
2924                 mutex_lock(&mdsc->mutex);
2925                 __wake_requests(mdsc, &session->s_waiting);
2926                 if (wake == 2)
2927                         kick_requests(mdsc, mds);
2928                 mutex_unlock(&mdsc->mutex);
2929         }
2930         if (op == CEPH_SESSION_CLOSE)
2931                 ceph_put_mds_session(session);
2932         return;
2933
2934 bad:
2935         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2936                (int)msg->front.iov_len);
2937         ceph_msg_dump(msg);
2938         return;
2939 }
2940
2941
2942 /*
2943  * called under session->mutex.
2944  */
2945 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2946                                    struct ceph_mds_session *session)
2947 {
2948         struct ceph_mds_request *req, *nreq;
2949         struct rb_node *p;
2950         int err;
2951
2952         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2953
2954         mutex_lock(&mdsc->mutex);
2955         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2956                 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2957                 if (!err) {
2958                         ceph_msg_get(req->r_request);
2959                         ceph_con_send(&session->s_con, req->r_request);
2960                 }
2961         }
2962
2963         /*
2964          * also re-send old requests when MDS enters reconnect stage. So that MDS
2965          * can process completed request in clientreplay stage.
2966          */
2967         p = rb_first(&mdsc->request_tree);
2968         while (p) {
2969                 req = rb_entry(p, struct ceph_mds_request, r_node);
2970                 p = rb_next(p);
2971                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2972                         continue;
2973                 if (req->r_attempts == 0)
2974                         continue; /* only old requests */
2975                 if (req->r_session &&
2976                     req->r_session->s_mds == session->s_mds) {
2977                         err = __prepare_send_request(mdsc, req,
2978                                                      session->s_mds, true);
2979                         if (!err) {
2980                                 ceph_msg_get(req->r_request);
2981                                 ceph_con_send(&session->s_con, req->r_request);
2982                         }
2983                 }
2984         }
2985         mutex_unlock(&mdsc->mutex);
2986 }
2987
2988 /*
2989  * Encode information about a cap for a reconnect with the MDS.
2990  */
2991 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2992                           void *arg)
2993 {
2994         union {
2995                 struct ceph_mds_cap_reconnect v2;
2996                 struct ceph_mds_cap_reconnect_v1 v1;
2997         } rec;
2998         struct ceph_inode_info *ci = cap->ci;
2999         struct ceph_reconnect_state *recon_state = arg;
3000         struct ceph_pagelist *pagelist = recon_state->pagelist;
3001         int err;
3002         u64 snap_follows;
3003
3004         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
3005              inode, ceph_vinop(inode), cap, cap->cap_id,
3006              ceph_cap_string(cap->issued));
3007         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3008         if (err)
3009                 return err;
3010
3011         spin_lock(&ci->i_ceph_lock);
3012         cap->seq = 0;        /* reset cap seq */
3013         cap->issue_seq = 0;  /* and issue_seq */
3014         cap->mseq = 0;       /* and migrate_seq */
3015         cap->cap_gen = cap->session->s_cap_gen;
3016
3017         if (recon_state->msg_version >= 2) {
3018                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
3019                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3020                 rec.v2.issued = cpu_to_le32(cap->issued);
3021                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3022                 rec.v2.pathbase = 0;
3023                 rec.v2.flock_len = (__force __le32)
3024                         ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
3025         } else {
3026                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
3027                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3028                 rec.v1.issued = cpu_to_le32(cap->issued);
3029                 rec.v1.size = cpu_to_le64(inode->i_size);
3030                 ceph_encode_timespec64(&rec.v1.mtime, &inode->i_mtime);
3031                 ceph_encode_timespec64(&rec.v1.atime, &inode->i_atime);
3032                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3033                 rec.v1.pathbase = 0;
3034         }
3035
3036         if (list_empty(&ci->i_cap_snaps)) {
3037                 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
3038         } else {
3039                 struct ceph_cap_snap *capsnap =
3040                         list_first_entry(&ci->i_cap_snaps,
3041                                          struct ceph_cap_snap, ci_item);
3042                 snap_follows = capsnap->follows;
3043         }
3044         spin_unlock(&ci->i_ceph_lock);
3045
3046         if (recon_state->msg_version >= 2) {
3047                 int num_fcntl_locks, num_flock_locks;
3048                 struct ceph_filelock *flocks = NULL;
3049                 size_t struct_len, total_len = 0;
3050                 u8 struct_v = 0;
3051
3052 encode_again:
3053                 if (rec.v2.flock_len) {
3054                         ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
3055                 } else {
3056                         num_fcntl_locks = 0;
3057                         num_flock_locks = 0;
3058                 }
3059                 if (num_fcntl_locks + num_flock_locks > 0) {
3060                         flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
3061                                                sizeof(struct ceph_filelock),
3062                                                GFP_NOFS);
3063                         if (!flocks) {
3064                                 err = -ENOMEM;
3065                                 goto out_err;
3066                         }
3067                         err = ceph_encode_locks_to_buffer(inode, flocks,
3068                                                           num_fcntl_locks,
3069                                                           num_flock_locks);
3070                         if (err) {
3071                                 kfree(flocks);
3072                                 flocks = NULL;
3073                                 if (err == -ENOSPC)
3074                                         goto encode_again;
3075                                 goto out_err;
3076                         }
3077                 } else {
3078                         kfree(flocks);
3079                         flocks = NULL;
3080                 }
3081
3082                 if (recon_state->msg_version >= 3) {
3083                         /* version, compat_version and struct_len */
3084                         total_len = 2 * sizeof(u8) + sizeof(u32);
3085                         struct_v = 2;
3086                 }
3087                 /*
3088                  * number of encoded locks is stable, so copy to pagelist
3089                  */
3090                 struct_len = 2 * sizeof(u32) +
3091                             (num_fcntl_locks + num_flock_locks) *
3092                             sizeof(struct ceph_filelock);
3093                 rec.v2.flock_len = cpu_to_le32(struct_len);
3094
3095                 struct_len += sizeof(u32) + sizeof(rec.v2);
3096
3097                 if (struct_v >= 2)
3098                         struct_len += sizeof(u64); /* snap_follows */
3099
3100                 total_len += struct_len;
3101                 err = ceph_pagelist_reserve(pagelist, total_len);
3102                 if (err) {
3103                         kfree(flocks);
3104                         goto out_err;
3105                 }
3106
3107                 if (recon_state->msg_version >= 3) {
3108                         ceph_pagelist_encode_8(pagelist, struct_v);
3109                         ceph_pagelist_encode_8(pagelist, 1);
3110                         ceph_pagelist_encode_32(pagelist, struct_len);
3111                 }
3112                 ceph_pagelist_encode_string(pagelist, NULL, 0);
3113                 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
3114                 ceph_locks_to_pagelist(flocks, pagelist,
3115                                        num_fcntl_locks, num_flock_locks);
3116                 if (struct_v >= 2)
3117                         ceph_pagelist_encode_64(pagelist, snap_follows);
3118
3119                 kfree(flocks);
3120         } else {
3121                 u64 pathbase = 0;
3122                 int pathlen = 0;
3123                 char *path = NULL;
3124                 struct dentry *dentry;
3125
3126                 dentry = d_find_alias(inode);
3127                 if (dentry) {
3128                         path = ceph_mdsc_build_path(dentry,
3129                                                 &pathlen, &pathbase, 0);
3130                         dput(dentry);
3131                         if (IS_ERR(path)) {
3132                                 err = PTR_ERR(path);
3133                                 goto out_err;
3134                         }
3135                         rec.v1.pathbase = cpu_to_le64(pathbase);
3136                 }
3137
3138                 err = ceph_pagelist_reserve(pagelist,
3139                                 pathlen + sizeof(u32) + sizeof(rec.v1));
3140                 if (err) {
3141                         kfree(path);
3142                         goto out_err;
3143                 }
3144
3145                 ceph_pagelist_encode_string(pagelist, path, pathlen);
3146                 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
3147
3148                 kfree(path);
3149         }
3150
3151         recon_state->nr_caps++;
3152 out_err:
3153         return err;
3154 }
3155
3156
3157 /*
3158  * If an MDS fails and recovers, clients need to reconnect in order to
3159  * reestablish shared state.  This includes all caps issued through
3160  * this session _and_ the snap_realm hierarchy.  Because it's not
3161  * clear which snap realms the mds cares about, we send everything we
3162  * know about.. that ensures we'll then get any new info the
3163  * recovering MDS might have.
3164  *
3165  * This is a relatively heavyweight operation, but it's rare.
3166  *
3167  * called with mdsc->mutex held.
3168  */
3169 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3170                                struct ceph_mds_session *session)
3171 {
3172         struct ceph_msg *reply;
3173         struct rb_node *p;
3174         int mds = session->s_mds;
3175         int err = -ENOMEM;
3176         int s_nr_caps;
3177         struct ceph_pagelist *pagelist;
3178         struct ceph_reconnect_state recon_state;
3179         LIST_HEAD(dispose);
3180
3181         pr_info("mds%d reconnect start\n", mds);
3182
3183         pagelist = ceph_pagelist_alloc(GFP_NOFS);
3184         if (!pagelist)
3185                 goto fail_nopagelist;
3186
3187         reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3188         if (!reply)
3189                 goto fail_nomsg;
3190
3191         mutex_lock(&session->s_mutex);
3192         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3193         session->s_seq = 0;
3194
3195         dout("session %p state %s\n", session,
3196              ceph_session_state_name(session->s_state));
3197
3198         spin_lock(&session->s_gen_ttl_lock);
3199         session->s_cap_gen++;
3200         spin_unlock(&session->s_gen_ttl_lock);
3201
3202         spin_lock(&session->s_cap_lock);
3203         /* don't know if session is readonly */
3204         session->s_readonly = 0;
3205         /*
3206          * notify __ceph_remove_cap() that we are composing cap reconnect.
3207          * If a cap get released before being added to the cap reconnect,
3208          * __ceph_remove_cap() should skip queuing cap release.
3209          */
3210         session->s_cap_reconnect = 1;
3211         /* drop old cap expires; we're about to reestablish that state */
3212         detach_cap_releases(session, &dispose);
3213         spin_unlock(&session->s_cap_lock);
3214         dispose_cap_releases(mdsc, &dispose);
3215
3216         /* trim unused caps to reduce MDS's cache rejoin time */
3217         if (mdsc->fsc->sb->s_root)
3218                 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3219
3220         ceph_con_close(&session->s_con);
3221         ceph_con_open(&session->s_con,
3222                       CEPH_ENTITY_TYPE_MDS, mds,
3223                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3224
3225         /* replay unsafe requests */
3226         replay_unsafe_requests(mdsc, session);
3227
3228         down_read(&mdsc->snap_rwsem);
3229
3230         /* traverse this session's caps */
3231         s_nr_caps = session->s_nr_caps;
3232         err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
3233         if (err)
3234                 goto fail;
3235
3236         recon_state.nr_caps = 0;
3237         recon_state.pagelist = pagelist;
3238         if (session->s_con.peer_features & CEPH_FEATURE_MDSENC)
3239                 recon_state.msg_version = 3;
3240         else
3241                 recon_state.msg_version = 2;
3242         err = iterate_session_caps(session, encode_caps_cb, &recon_state);
3243         if (err < 0)
3244                 goto fail;
3245
3246         spin_lock(&session->s_cap_lock);
3247         session->s_cap_reconnect = 0;
3248         spin_unlock(&session->s_cap_lock);
3249
3250         /*
3251          * snaprealms.  we provide mds with the ino, seq (version), and
3252          * parent for all of our realms.  If the mds has any newer info,
3253          * it will tell us.
3254          */
3255         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3256                 struct ceph_snap_realm *realm =
3257                         rb_entry(p, struct ceph_snap_realm, node);
3258                 struct ceph_mds_snaprealm_reconnect sr_rec;
3259
3260                 dout(" adding snap realm %llx seq %lld parent %llx\n",
3261                      realm->ino, realm->seq, realm->parent_ino);
3262                 sr_rec.ino = cpu_to_le64(realm->ino);
3263                 sr_rec.seq = cpu_to_le64(realm->seq);
3264                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3265                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3266                 if (err)
3267                         goto fail;
3268         }
3269
3270         reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3271
3272         /* raced with cap release? */
3273         if (s_nr_caps != recon_state.nr_caps) {
3274                 struct page *page = list_first_entry(&pagelist->head,
3275                                                      struct page, lru);
3276                 __le32 *addr = kmap_atomic(page);
3277                 *addr = cpu_to_le32(recon_state.nr_caps);
3278                 kunmap_atomic(addr);
3279         }
3280
3281         reply->hdr.data_len = cpu_to_le32(pagelist->length);
3282         ceph_msg_data_add_pagelist(reply, pagelist);
3283
3284         ceph_early_kick_flushing_caps(mdsc, session);
3285
3286         ceph_con_send(&session->s_con, reply);
3287
3288         mutex_unlock(&session->s_mutex);
3289
3290         mutex_lock(&mdsc->mutex);
3291         __wake_requests(mdsc, &session->s_waiting);
3292         mutex_unlock(&mdsc->mutex);
3293
3294         up_read(&mdsc->snap_rwsem);
3295         ceph_pagelist_release(pagelist);
3296         return;
3297
3298 fail:
3299         ceph_msg_put(reply);
3300         up_read(&mdsc->snap_rwsem);
3301         mutex_unlock(&session->s_mutex);
3302 fail_nomsg:
3303         ceph_pagelist_release(pagelist);
3304 fail_nopagelist:
3305         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3306         return;
3307 }
3308
3309
3310 /*
3311  * compare old and new mdsmaps, kicking requests
3312  * and closing out old connections as necessary
3313  *
3314  * called under mdsc->mutex.
3315  */
3316 static void check_new_map(struct ceph_mds_client *mdsc,
3317                           struct ceph_mdsmap *newmap,
3318                           struct ceph_mdsmap *oldmap)
3319 {
3320         int i;
3321         int oldstate, newstate;
3322         struct ceph_mds_session *s;
3323
3324         dout("check_new_map new %u old %u\n",
3325              newmap->m_epoch, oldmap->m_epoch);
3326
3327         for (i = 0; i < oldmap->m_num_mds && i < mdsc->max_sessions; i++) {
3328                 if (!mdsc->sessions[i])
3329                         continue;
3330                 s = mdsc->sessions[i];
3331                 oldstate = ceph_mdsmap_get_state(oldmap, i);
3332                 newstate = ceph_mdsmap_get_state(newmap, i);
3333
3334                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3335                      i, ceph_mds_state_name(oldstate),
3336                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3337                      ceph_mds_state_name(newstate),
3338                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3339                      ceph_session_state_name(s->s_state));
3340
3341                 if (i >= newmap->m_num_mds ||
3342                     memcmp(ceph_mdsmap_get_addr(oldmap, i),
3343                            ceph_mdsmap_get_addr(newmap, i),
3344                            sizeof(struct ceph_entity_addr))) {
3345                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3346                                 /* the session never opened, just close it
3347                                  * out now */
3348                                 get_session(s);
3349                                 __unregister_session(mdsc, s);
3350                                 __wake_requests(mdsc, &s->s_waiting);
3351                                 ceph_put_mds_session(s);
3352                         } else if (i >= newmap->m_num_mds) {
3353                                 /* force close session for stopped mds */
3354                                 get_session(s);
3355                                 __unregister_session(mdsc, s);
3356                                 __wake_requests(mdsc, &s->s_waiting);
3357                                 kick_requests(mdsc, i);
3358                                 mutex_unlock(&mdsc->mutex);
3359
3360                                 mutex_lock(&s->s_mutex);
3361                                 cleanup_session_requests(mdsc, s);
3362                                 remove_session_caps(s);
3363                                 mutex_unlock(&s->s_mutex);
3364
3365                                 ceph_put_mds_session(s);
3366
3367                                 mutex_lock(&mdsc->mutex);
3368                         } else {
3369                                 /* just close it */
3370                                 mutex_unlock(&mdsc->mutex);
3371                                 mutex_lock(&s->s_mutex);
3372                                 mutex_lock(&mdsc->mutex);
3373                                 ceph_con_close(&s->s_con);
3374                                 mutex_unlock(&s->s_mutex);
3375                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3376                         }
3377                 } else if (oldstate == newstate) {
3378                         continue;  /* nothing new with this mds */
3379                 }
3380
3381                 /*
3382                  * send reconnect?
3383                  */
3384                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3385                     newstate >= CEPH_MDS_STATE_RECONNECT) {
3386                         mutex_unlock(&mdsc->mutex);
3387                         send_mds_reconnect(mdsc, s);
3388                         mutex_lock(&mdsc->mutex);
3389                 }
3390
3391                 /*
3392                  * kick request on any mds that has gone active.
3393                  */
3394                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3395                     newstate >= CEPH_MDS_STATE_ACTIVE) {
3396                         if (oldstate != CEPH_MDS_STATE_CREATING &&
3397                             oldstate != CEPH_MDS_STATE_STARTING)
3398                                 pr_info("mds%d recovery completed\n", s->s_mds);
3399                         kick_requests(mdsc, i);
3400                         ceph_kick_flushing_caps(mdsc, s);
3401                         wake_up_session_caps(s, RECONNECT);
3402                 }
3403         }
3404
3405         for (i = 0; i < newmap->m_num_mds && i < mdsc->max_sessions; i++) {
3406                 s = mdsc->sessions[i];
3407                 if (!s)
3408                         continue;
3409                 if (!ceph_mdsmap_is_laggy(newmap, i))
3410                         continue;
3411                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3412                     s->s_state == CEPH_MDS_SESSION_HUNG ||
3413                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
3414                         dout(" connecting to export targets of laggy mds%d\n",
3415                              i);
3416                         __open_export_target_sessions(mdsc, s);
3417                 }
3418         }
3419 }
3420
3421
3422
3423 /*
3424  * leases
3425  */
3426
3427 /*
3428  * caller must hold session s_mutex, dentry->d_lock
3429  */
3430 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3431 {
3432         struct ceph_dentry_info *di = ceph_dentry(dentry);
3433
3434         ceph_put_mds_session(di->lease_session);
3435         di->lease_session = NULL;
3436 }
3437
3438 static void handle_lease(struct ceph_mds_client *mdsc,
3439                          struct ceph_mds_session *session,
3440                          struct ceph_msg *msg)
3441 {
3442         struct super_block *sb = mdsc->fsc->sb;
3443         struct inode *inode;
3444         struct dentry *parent, *dentry;
3445         struct ceph_dentry_info *di;
3446         int mds = session->s_mds;
3447         struct ceph_mds_lease *h = msg->front.iov_base;
3448         u32 seq;
3449         struct ceph_vino vino;
3450         struct qstr dname;
3451         int release = 0;
3452
3453         dout("handle_lease from mds%d\n", mds);
3454
3455         /* decode */
3456         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3457                 goto bad;
3458         vino.ino = le64_to_cpu(h->ino);
3459         vino.snap = CEPH_NOSNAP;
3460         seq = le32_to_cpu(h->seq);
3461         dname.len = get_unaligned_le32(h + 1);
3462         if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
3463                 goto bad;
3464         dname.name = (void *)(h + 1) + sizeof(u32);
3465
3466         /* lookup inode */
3467         inode = ceph_find_inode(sb, vino);
3468         dout("handle_lease %s, ino %llx %p %.*s\n",
3469              ceph_lease_op_name(h->action), vino.ino, inode,
3470              dname.len, dname.name);
3471
3472         mutex_lock(&session->s_mutex);
3473         session->s_seq++;
3474
3475         if (!inode) {
3476                 dout("handle_lease no inode %llx\n", vino.ino);
3477                 goto release;
3478         }
3479
3480         /* dentry */
3481         parent = d_find_alias(inode);
3482         if (!parent) {
3483                 dout("no parent dentry on inode %p\n", inode);
3484                 WARN_ON(1);
3485                 goto release;  /* hrm... */
3486         }
3487         dname.hash = full_name_hash(parent, dname.name, dname.len);
3488         dentry = d_lookup(parent, &dname);
3489         dput(parent);
3490         if (!dentry)
3491                 goto release;
3492
3493         spin_lock(&dentry->d_lock);
3494         di = ceph_dentry(dentry);
3495         switch (h->action) {
3496         case CEPH_MDS_LEASE_REVOKE:
3497                 if (di->lease_session == session) {
3498                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3499                                 h->seq = cpu_to_le32(di->lease_seq);
3500                         __ceph_mdsc_drop_dentry_lease(dentry);
3501                 }
3502                 release = 1;
3503                 break;
3504
3505         case CEPH_MDS_LEASE_RENEW:
3506                 if (di->lease_session == session &&
3507                     di->lease_gen == session->s_cap_gen &&
3508                     di->lease_renew_from &&
3509                     di->lease_renew_after == 0) {
3510                         unsigned long duration =
3511                                 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3512
3513                         di->lease_seq = seq;
3514                         di->time = di->lease_renew_from + duration;
3515                         di->lease_renew_after = di->lease_renew_from +
3516                                 (duration >> 1);
3517                         di->lease_renew_from = 0;
3518                 }
3519                 break;
3520         }
3521         spin_unlock(&dentry->d_lock);
3522         dput(dentry);
3523
3524         if (!release)
3525                 goto out;
3526
3527 release:
3528         /* let's just reuse the same message */
3529         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3530         ceph_msg_get(msg);
3531         ceph_con_send(&session->s_con, msg);
3532
3533 out:
3534         iput(inode);
3535         mutex_unlock(&session->s_mutex);
3536         return;
3537
3538 bad:
3539         pr_err("corrupt lease message\n");
3540         ceph_msg_dump(msg);
3541 }
3542
3543 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3544                               struct inode *inode,
3545                               struct dentry *dentry, char action,
3546                               u32 seq)
3547 {
3548         struct ceph_msg *msg;
3549         struct ceph_mds_lease *lease;
3550         int len = sizeof(*lease) + sizeof(u32);
3551         int dnamelen = 0;
3552
3553         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3554              inode, dentry, ceph_lease_op_name(action), session->s_mds);
3555         dnamelen = dentry->d_name.len;
3556         len += dnamelen;
3557
3558         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3559         if (!msg)
3560                 return;
3561         lease = msg->front.iov_base;
3562         lease->action = action;
3563         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3564         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3565         lease->seq = cpu_to_le32(seq);
3566         put_unaligned_le32(dnamelen, lease + 1);
3567         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3568
3569         /*
3570          * if this is a preemptive lease RELEASE, no need to
3571          * flush request stream, since the actual request will
3572          * soon follow.
3573          */
3574         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3575
3576         ceph_con_send(&session->s_con, msg);
3577 }
3578
3579 /*
3580  * lock unlock sessions, to wait ongoing session activities
3581  */
3582 static void lock_unlock_sessions(struct ceph_mds_client *mdsc)
3583 {
3584         int i;
3585
3586         mutex_lock(&mdsc->mutex);
3587         for (i = 0; i < mdsc->max_sessions; i++) {
3588                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3589                 if (!s)
3590                         continue;
3591                 mutex_unlock(&mdsc->mutex);
3592                 mutex_lock(&s->s_mutex);
3593                 mutex_unlock(&s->s_mutex);
3594                 ceph_put_mds_session(s);
3595                 mutex_lock(&mdsc->mutex);
3596         }
3597         mutex_unlock(&mdsc->mutex);
3598 }
3599
3600
3601
3602 /*
3603  * delayed work -- periodically trim expired leases, renew caps with mds
3604  */
3605 static void schedule_delayed(struct ceph_mds_client *mdsc)
3606 {
3607         int delay = 5;
3608         unsigned hz = round_jiffies_relative(HZ * delay);
3609         schedule_delayed_work(&mdsc->delayed_work, hz);
3610 }
3611
3612 static void delayed_work(struct work_struct *work)
3613 {
3614         int i;
3615         struct ceph_mds_client *mdsc =
3616                 container_of(work, struct ceph_mds_client, delayed_work.work);
3617         int renew_interval;
3618         int renew_caps;
3619
3620         dout("mdsc delayed_work\n");
3621         ceph_check_delayed_caps(mdsc);
3622
3623         mutex_lock(&mdsc->mutex);
3624         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3625         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3626                                    mdsc->last_renew_caps);
3627         if (renew_caps)
3628                 mdsc->last_renew_caps = jiffies;
3629
3630         for (i = 0; i < mdsc->max_sessions; i++) {
3631                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3632                 if (!s)
3633                         continue;
3634                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3635                         dout("resending session close request for mds%d\n",
3636                              s->s_mds);
3637                         request_close_session(mdsc, s);
3638                         ceph_put_mds_session(s);
3639                         continue;
3640                 }
3641                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3642                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3643                                 s->s_state = CEPH_MDS_SESSION_HUNG;
3644                                 pr_info("mds%d hung\n", s->s_mds);
3645                         }
3646                 }
3647                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3648                         /* this mds is failed or recovering, just wait */
3649                         ceph_put_mds_session(s);
3650                         continue;
3651                 }
3652                 mutex_unlock(&mdsc->mutex);
3653
3654                 mutex_lock(&s->s_mutex);
3655                 if (renew_caps)
3656                         send_renew_caps(mdsc, s);
3657                 else
3658                         ceph_con_keepalive(&s->s_con);
3659                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3660                     s->s_state == CEPH_MDS_SESSION_HUNG)
3661                         ceph_send_cap_releases(mdsc, s);
3662                 mutex_unlock(&s->s_mutex);
3663                 ceph_put_mds_session(s);
3664
3665                 mutex_lock(&mdsc->mutex);
3666         }
3667         mutex_unlock(&mdsc->mutex);
3668
3669         schedule_delayed(mdsc);
3670 }
3671
3672 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3673
3674 {
3675         struct ceph_mds_client *mdsc;
3676
3677         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3678         if (!mdsc)
3679                 return -ENOMEM;
3680         mdsc->fsc = fsc;
3681         mutex_init(&mdsc->mutex);
3682         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3683         if (!mdsc->mdsmap) {
3684                 kfree(mdsc);
3685                 return -ENOMEM;
3686         }
3687
3688         fsc->mdsc = mdsc;
3689         init_completion(&mdsc->safe_umount_waiters);
3690         init_waitqueue_head(&mdsc->session_close_wq);
3691         INIT_LIST_HEAD(&mdsc->waiting_for_map);
3692         mdsc->sessions = NULL;
3693         atomic_set(&mdsc->num_sessions, 0);
3694         mdsc->max_sessions = 0;
3695         mdsc->stopping = 0;
3696         atomic64_set(&mdsc->quotarealms_count, 0);
3697         mdsc->last_snap_seq = 0;
3698         init_rwsem(&mdsc->snap_rwsem);
3699         mdsc->snap_realms = RB_ROOT;
3700         INIT_LIST_HEAD(&mdsc->snap_empty);
3701         spin_lock_init(&mdsc->snap_empty_lock);
3702         mdsc->last_tid = 0;
3703         mdsc->oldest_tid = 0;
3704         mdsc->request_tree = RB_ROOT;
3705         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3706         mdsc->last_renew_caps = jiffies;
3707         INIT_LIST_HEAD(&mdsc->cap_delay_list);
3708         spin_lock_init(&mdsc->cap_delay_lock);
3709         INIT_LIST_HEAD(&mdsc->snap_flush_list);
3710         spin_lock_init(&mdsc->snap_flush_lock);
3711         mdsc->last_cap_flush_tid = 1;
3712         INIT_LIST_HEAD(&mdsc->cap_flush_list);
3713         INIT_LIST_HEAD(&mdsc->cap_dirty);
3714         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3715         mdsc->num_cap_flushing = 0;
3716         spin_lock_init(&mdsc->cap_dirty_lock);
3717         init_waitqueue_head(&mdsc->cap_flushing_wq);
3718         spin_lock_init(&mdsc->dentry_lru_lock);
3719         INIT_LIST_HEAD(&mdsc->dentry_lru);
3720
3721         ceph_caps_init(mdsc);
3722         ceph_adjust_min_caps(mdsc, fsc->min_caps);
3723
3724         init_rwsem(&mdsc->pool_perm_rwsem);
3725         mdsc->pool_perm_tree = RB_ROOT;
3726
3727         strscpy(mdsc->nodename, utsname()->nodename,
3728                 sizeof(mdsc->nodename));
3729         return 0;
3730 }
3731
3732 /*
3733  * Wait for safe replies on open mds requests.  If we time out, drop
3734  * all requests from the tree to avoid dangling dentry refs.
3735  */
3736 static void wait_requests(struct ceph_mds_client *mdsc)
3737 {
3738         struct ceph_options *opts = mdsc->fsc->client->options;
3739         struct ceph_mds_request *req;
3740
3741         mutex_lock(&mdsc->mutex);
3742         if (__get_oldest_req(mdsc)) {
3743                 mutex_unlock(&mdsc->mutex);
3744
3745                 dout("wait_requests waiting for requests\n");
3746                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3747                                     ceph_timeout_jiffies(opts->mount_timeout));
3748
3749                 /* tear down remaining requests */
3750                 mutex_lock(&mdsc->mutex);
3751                 while ((req = __get_oldest_req(mdsc))) {
3752                         dout("wait_requests timed out on tid %llu\n",
3753                              req->r_tid);
3754                         __unregister_request(mdsc, req);
3755                 }
3756         }
3757         mutex_unlock(&mdsc->mutex);
3758         dout("wait_requests done\n");
3759 }
3760
3761 /*
3762  * called before mount is ro, and before dentries are torn down.
3763  * (hmm, does this still race with new lookups?)
3764  */
3765 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3766 {
3767         dout("pre_umount\n");
3768         mdsc->stopping = 1;
3769
3770         lock_unlock_sessions(mdsc);
3771         ceph_flush_dirty_caps(mdsc);
3772         wait_requests(mdsc);
3773
3774         /*
3775          * wait for reply handlers to drop their request refs and
3776          * their inode/dcache refs
3777          */
3778         ceph_msgr_flush();
3779 }
3780
3781 /*
3782  * wait for all write mds requests to flush.
3783  */
3784 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3785 {
3786         struct ceph_mds_request *req = NULL, *nextreq;
3787         struct rb_node *n;
3788
3789         mutex_lock(&mdsc->mutex);
3790         dout("wait_unsafe_requests want %lld\n", want_tid);
3791 restart:
3792         req = __get_oldest_req(mdsc);
3793         while (req && req->r_tid <= want_tid) {
3794                 /* find next request */
3795                 n = rb_next(&req->r_node);
3796                 if (n)
3797                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3798                 else
3799                         nextreq = NULL;
3800                 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3801                     (req->r_op & CEPH_MDS_OP_WRITE)) {
3802                         /* write op */
3803                         ceph_mdsc_get_request(req);
3804                         if (nextreq)
3805                                 ceph_mdsc_get_request(nextreq);
3806                         mutex_unlock(&mdsc->mutex);
3807                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
3808                              req->r_tid, want_tid);
3809                         wait_for_completion(&req->r_safe_completion);
3810                         mutex_lock(&mdsc->mutex);
3811                         ceph_mdsc_put_request(req);
3812                         if (!nextreq)
3813                                 break;  /* next dne before, so we're done! */
3814                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
3815                                 /* next request was removed from tree */
3816                                 ceph_mdsc_put_request(nextreq);
3817                                 goto restart;
3818                         }
3819                         ceph_mdsc_put_request(nextreq);  /* won't go away */
3820                 }
3821                 req = nextreq;
3822         }
3823         mutex_unlock(&mdsc->mutex);
3824         dout("wait_unsafe_requests done\n");
3825 }
3826
3827 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3828 {
3829         u64 want_tid, want_flush;
3830
3831         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3832                 return;
3833
3834         dout("sync\n");
3835         mutex_lock(&mdsc->mutex);
3836         want_tid = mdsc->last_tid;
3837         mutex_unlock(&mdsc->mutex);
3838
3839         ceph_flush_dirty_caps(mdsc);
3840         spin_lock(&mdsc->cap_dirty_lock);
3841         want_flush = mdsc->last_cap_flush_tid;
3842         if (!list_empty(&mdsc->cap_flush_list)) {
3843                 struct ceph_cap_flush *cf =
3844                         list_last_entry(&mdsc->cap_flush_list,
3845                                         struct ceph_cap_flush, g_list);
3846                 cf->wake = true;
3847         }
3848         spin_unlock(&mdsc->cap_dirty_lock);
3849
3850         dout("sync want tid %lld flush_seq %lld\n",
3851              want_tid, want_flush);
3852
3853         wait_unsafe_requests(mdsc, want_tid);
3854         wait_caps_flush(mdsc, want_flush);
3855 }
3856
3857 /*
3858  * true if all sessions are closed, or we force unmount
3859  */
3860 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
3861 {
3862         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3863                 return true;
3864         return atomic_read(&mdsc->num_sessions) <= skipped;
3865 }
3866
3867 /*
3868  * called after sb is ro.
3869  */
3870 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3871 {
3872         struct ceph_options *opts = mdsc->fsc->client->options;
3873         struct ceph_mds_session *session;
3874         int i;
3875         int skipped = 0;
3876
3877         dout("close_sessions\n");
3878
3879         /* close sessions */
3880         mutex_lock(&mdsc->mutex);
3881         for (i = 0; i < mdsc->max_sessions; i++) {
3882                 session = __ceph_lookup_mds_session(mdsc, i);
3883                 if (!session)
3884                         continue;
3885                 mutex_unlock(&mdsc->mutex);
3886                 mutex_lock(&session->s_mutex);
3887                 if (__close_session(mdsc, session) <= 0)
3888                         skipped++;
3889                 mutex_unlock(&session->s_mutex);
3890                 ceph_put_mds_session(session);
3891                 mutex_lock(&mdsc->mutex);
3892         }
3893         mutex_unlock(&mdsc->mutex);
3894
3895         dout("waiting for sessions to close\n");
3896         wait_event_timeout(mdsc->session_close_wq,
3897                            done_closing_sessions(mdsc, skipped),
3898                            ceph_timeout_jiffies(opts->mount_timeout));
3899
3900         /* tear down remaining sessions */
3901         mutex_lock(&mdsc->mutex);
3902         for (i = 0; i < mdsc->max_sessions; i++) {
3903                 if (mdsc->sessions[i]) {
3904                         session = get_session(mdsc->sessions[i]);
3905                         __unregister_session(mdsc, session);
3906                         mutex_unlock(&mdsc->mutex);
3907                         mutex_lock(&session->s_mutex);
3908                         remove_session_caps(session);
3909                         mutex_unlock(&session->s_mutex);
3910                         ceph_put_mds_session(session);
3911                         mutex_lock(&mdsc->mutex);
3912                 }
3913         }
3914         WARN_ON(!list_empty(&mdsc->cap_delay_list));
3915         mutex_unlock(&mdsc->mutex);
3916
3917         ceph_cleanup_empty_realms(mdsc);
3918
3919         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3920
3921         dout("stopped\n");
3922 }
3923
3924 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
3925 {
3926         struct ceph_mds_session *session;
3927         int mds;
3928
3929         dout("force umount\n");
3930
3931         mutex_lock(&mdsc->mutex);
3932         for (mds = 0; mds < mdsc->max_sessions; mds++) {
3933                 session = __ceph_lookup_mds_session(mdsc, mds);
3934                 if (!session)
3935                         continue;
3936                 mutex_unlock(&mdsc->mutex);
3937                 mutex_lock(&session->s_mutex);
3938                 __close_session(mdsc, session);
3939                 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
3940                         cleanup_session_requests(mdsc, session);
3941                         remove_session_caps(session);
3942                 }
3943                 mutex_unlock(&session->s_mutex);
3944                 ceph_put_mds_session(session);
3945                 mutex_lock(&mdsc->mutex);
3946                 kick_requests(mdsc, mds);
3947         }
3948         __wake_requests(mdsc, &mdsc->waiting_for_map);
3949         mutex_unlock(&mdsc->mutex);
3950 }
3951
3952 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3953 {
3954         dout("stop\n");
3955         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3956         if (mdsc->mdsmap)
3957                 ceph_mdsmap_destroy(mdsc->mdsmap);
3958         kfree(mdsc->sessions);
3959         ceph_caps_finalize(mdsc);
3960         ceph_pool_perm_destroy(mdsc);
3961 }
3962
3963 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3964 {
3965         struct ceph_mds_client *mdsc = fsc->mdsc;
3966         dout("mdsc_destroy %p\n", mdsc);
3967
3968         if (!mdsc)
3969                 return;
3970
3971         /* flush out any connection work with references to us */
3972         ceph_msgr_flush();
3973
3974         ceph_mdsc_stop(mdsc);
3975
3976         fsc->mdsc = NULL;
3977         kfree(mdsc);
3978         dout("mdsc_destroy %p done\n", mdsc);
3979 }
3980
3981 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3982 {
3983         struct ceph_fs_client *fsc = mdsc->fsc;
3984         const char *mds_namespace = fsc->mount_options->mds_namespace;
3985         void *p = msg->front.iov_base;
3986         void *end = p + msg->front.iov_len;
3987         u32 epoch;
3988         u32 map_len;
3989         u32 num_fs;
3990         u32 mount_fscid = (u32)-1;
3991         u8 struct_v, struct_cv;
3992         int err = -EINVAL;
3993
3994         ceph_decode_need(&p, end, sizeof(u32), bad);
3995         epoch = ceph_decode_32(&p);
3996
3997         dout("handle_fsmap epoch %u\n", epoch);
3998
3999         ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4000         struct_v = ceph_decode_8(&p);
4001         struct_cv = ceph_decode_8(&p);
4002         map_len = ceph_decode_32(&p);
4003
4004         ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
4005         p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
4006
4007         num_fs = ceph_decode_32(&p);
4008         while (num_fs-- > 0) {
4009                 void *info_p, *info_end;
4010                 u32 info_len;
4011                 u8 info_v, info_cv;
4012                 u32 fscid, namelen;
4013
4014                 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4015                 info_v = ceph_decode_8(&p);
4016                 info_cv = ceph_decode_8(&p);
4017                 info_len = ceph_decode_32(&p);
4018                 ceph_decode_need(&p, end, info_len, bad);
4019                 info_p = p;
4020                 info_end = p + info_len;
4021                 p = info_end;
4022
4023                 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
4024                 fscid = ceph_decode_32(&info_p);
4025                 namelen = ceph_decode_32(&info_p);
4026                 ceph_decode_need(&info_p, info_end, namelen, bad);
4027
4028                 if (mds_namespace &&
4029                     strlen(mds_namespace) == namelen &&
4030                     !strncmp(mds_namespace, (char *)info_p, namelen)) {
4031                         mount_fscid = fscid;
4032                         break;
4033                 }
4034         }
4035
4036         ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
4037         if (mount_fscid != (u32)-1) {
4038                 fsc->client->monc.fs_cluster_id = mount_fscid;
4039                 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
4040                                    0, true);
4041                 ceph_monc_renew_subs(&fsc->client->monc);
4042         } else {
4043                 err = -ENOENT;
4044                 goto err_out;
4045         }
4046         return;
4047
4048 bad:
4049         pr_err("error decoding fsmap\n");
4050 err_out:
4051         mutex_lock(&mdsc->mutex);
4052         mdsc->mdsmap_err = err;
4053         __wake_requests(mdsc, &mdsc->waiting_for_map);
4054         mutex_unlock(&mdsc->mutex);
4055 }
4056
4057 /*
4058  * handle mds map update.
4059  */
4060 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4061 {
4062         u32 epoch;
4063         u32 maplen;
4064         void *p = msg->front.iov_base;
4065         void *end = p + msg->front.iov_len;
4066         struct ceph_mdsmap *newmap, *oldmap;
4067         struct ceph_fsid fsid;
4068         int err = -EINVAL;
4069
4070         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
4071         ceph_decode_copy(&p, &fsid, sizeof(fsid));
4072         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
4073                 return;
4074         epoch = ceph_decode_32(&p);
4075         maplen = ceph_decode_32(&p);
4076         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
4077
4078         /* do we need it? */
4079         mutex_lock(&mdsc->mutex);
4080         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
4081                 dout("handle_map epoch %u <= our %u\n",
4082                      epoch, mdsc->mdsmap->m_epoch);
4083                 mutex_unlock(&mdsc->mutex);
4084                 return;
4085         }
4086
4087         newmap = ceph_mdsmap_decode(&p, end);
4088         if (IS_ERR(newmap)) {
4089                 err = PTR_ERR(newmap);
4090                 goto bad_unlock;
4091         }
4092
4093         /* swap into place */
4094         if (mdsc->mdsmap) {
4095                 oldmap = mdsc->mdsmap;
4096                 mdsc->mdsmap = newmap;
4097                 check_new_map(mdsc, newmap, oldmap);
4098                 ceph_mdsmap_destroy(oldmap);
4099         } else {
4100                 mdsc->mdsmap = newmap;  /* first mds map */
4101         }
4102         mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
4103                                         MAX_LFS_FILESIZE);
4104
4105         __wake_requests(mdsc, &mdsc->waiting_for_map);
4106         ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
4107                           mdsc->mdsmap->m_epoch);
4108
4109         mutex_unlock(&mdsc->mutex);
4110         schedule_delayed(mdsc);
4111         return;
4112
4113 bad_unlock:
4114         mutex_unlock(&mdsc->mutex);
4115 bad:
4116         pr_err("error decoding mdsmap %d\n", err);
4117         return;
4118 }
4119
4120 static struct ceph_connection *con_get(struct ceph_connection *con)
4121 {
4122         struct ceph_mds_session *s = con->private;
4123
4124         if (get_session(s)) {
4125                 dout("mdsc con_get %p ok (%d)\n", s, refcount_read(&s->s_ref));
4126                 return con;
4127         }
4128         dout("mdsc con_get %p FAIL\n", s);
4129         return NULL;
4130 }
4131
4132 static void con_put(struct ceph_connection *con)
4133 {
4134         struct ceph_mds_session *s = con->private;
4135
4136         dout("mdsc con_put %p (%d)\n", s, refcount_read(&s->s_ref) - 1);
4137         ceph_put_mds_session(s);
4138 }
4139
4140 /*
4141  * if the client is unresponsive for long enough, the mds will kill
4142  * the session entirely.
4143  */
4144 static void peer_reset(struct ceph_connection *con)
4145 {
4146         struct ceph_mds_session *s = con->private;
4147         struct ceph_mds_client *mdsc = s->s_mdsc;
4148
4149         pr_warn("mds%d closed our session\n", s->s_mds);
4150         send_mds_reconnect(mdsc, s);
4151 }
4152
4153 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
4154 {
4155         struct ceph_mds_session *s = con->private;
4156         struct ceph_mds_client *mdsc = s->s_mdsc;
4157         int type = le16_to_cpu(msg->hdr.type);
4158
4159         mutex_lock(&mdsc->mutex);
4160         if (__verify_registered_session(mdsc, s) < 0) {
4161                 mutex_unlock(&mdsc->mutex);
4162                 goto out;
4163         }
4164         mutex_unlock(&mdsc->mutex);
4165
4166         switch (type) {
4167         case CEPH_MSG_MDS_MAP:
4168                 ceph_mdsc_handle_mdsmap(mdsc, msg);
4169                 break;
4170         case CEPH_MSG_FS_MAP_USER:
4171                 ceph_mdsc_handle_fsmap(mdsc, msg);
4172                 break;
4173         case CEPH_MSG_CLIENT_SESSION:
4174                 handle_session(s, msg);
4175                 break;
4176         case CEPH_MSG_CLIENT_REPLY:
4177                 handle_reply(s, msg);
4178                 break;
4179         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
4180                 handle_forward(mdsc, s, msg);
4181                 break;
4182         case CEPH_MSG_CLIENT_CAPS:
4183                 ceph_handle_caps(s, msg);
4184                 break;
4185         case CEPH_MSG_CLIENT_SNAP:
4186                 ceph_handle_snap(mdsc, s, msg);
4187                 break;
4188         case CEPH_MSG_CLIENT_LEASE:
4189                 handle_lease(mdsc, s, msg);
4190                 break;
4191         case CEPH_MSG_CLIENT_QUOTA:
4192                 ceph_handle_quota(mdsc, s, msg);
4193                 break;
4194
4195         default:
4196                 pr_err("received unknown message type %d %s\n", type,
4197                        ceph_msg_type_name(type));
4198         }
4199 out:
4200         ceph_msg_put(msg);
4201 }
4202
4203 /*
4204  * authentication
4205  */
4206
4207 /*
4208  * Note: returned pointer is the address of a structure that's
4209  * managed separately.  Caller must *not* attempt to free it.
4210  */
4211 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
4212                                         int *proto, int force_new)
4213 {
4214         struct ceph_mds_session *s = con->private;
4215         struct ceph_mds_client *mdsc = s->s_mdsc;
4216         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4217         struct ceph_auth_handshake *auth = &s->s_auth;
4218
4219         if (force_new && auth->authorizer) {
4220                 ceph_auth_destroy_authorizer(auth->authorizer);
4221                 auth->authorizer = NULL;
4222         }
4223         if (!auth->authorizer) {
4224                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4225                                                       auth);
4226                 if (ret)
4227                         return ERR_PTR(ret);
4228         } else {
4229                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4230                                                       auth);
4231                 if (ret)
4232                         return ERR_PTR(ret);
4233         }
4234         *proto = ac->protocol;
4235
4236         return auth;
4237 }
4238
4239 static int add_authorizer_challenge(struct ceph_connection *con,
4240                                     void *challenge_buf, int challenge_buf_len)
4241 {
4242         struct ceph_mds_session *s = con->private;
4243         struct ceph_mds_client *mdsc = s->s_mdsc;
4244         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4245
4246         return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
4247                                             challenge_buf, challenge_buf_len);
4248 }
4249
4250 static int verify_authorizer_reply(struct ceph_connection *con)
4251 {
4252         struct ceph_mds_session *s = con->private;
4253         struct ceph_mds_client *mdsc = s->s_mdsc;
4254         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4255
4256         return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
4257 }
4258
4259 static int invalidate_authorizer(struct ceph_connection *con)
4260 {
4261         struct ceph_mds_session *s = con->private;
4262         struct ceph_mds_client *mdsc = s->s_mdsc;
4263         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4264
4265         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
4266
4267         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
4268 }
4269
4270 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
4271                                 struct ceph_msg_header *hdr, int *skip)
4272 {
4273         struct ceph_msg *msg;
4274         int type = (int) le16_to_cpu(hdr->type);
4275         int front_len = (int) le32_to_cpu(hdr->front_len);
4276
4277         if (con->in_msg)
4278                 return con->in_msg;
4279
4280         *skip = 0;
4281         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
4282         if (!msg) {
4283                 pr_err("unable to allocate msg type %d len %d\n",
4284                        type, front_len);
4285                 return NULL;
4286         }
4287
4288         return msg;
4289 }
4290
4291 static int mds_sign_message(struct ceph_msg *msg)
4292 {
4293        struct ceph_mds_session *s = msg->con->private;
4294        struct ceph_auth_handshake *auth = &s->s_auth;
4295
4296        return ceph_auth_sign_message(auth, msg);
4297 }
4298
4299 static int mds_check_message_signature(struct ceph_msg *msg)
4300 {
4301        struct ceph_mds_session *s = msg->con->private;
4302        struct ceph_auth_handshake *auth = &s->s_auth;
4303
4304        return ceph_auth_check_message_signature(auth, msg);
4305 }
4306
4307 static const struct ceph_connection_operations mds_con_ops = {
4308         .get = con_get,
4309         .put = con_put,
4310         .dispatch = dispatch,
4311         .get_authorizer = get_authorizer,
4312         .add_authorizer_challenge = add_authorizer_challenge,
4313         .verify_authorizer_reply = verify_authorizer_reply,
4314         .invalidate_authorizer = invalidate_authorizer,
4315         .peer_reset = peer_reset,
4316         .alloc_msg = mds_alloc_msg,
4317         .sign_message = mds_sign_message,
4318         .check_message_signature = mds_check_message_signature,
4319 };
4320
4321 /* eof */