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  45  * are no longer valid.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         __ceph_remove_cap(cap, false);
1236         if (!ci->i_auth_cap) {
1237                 struct ceph_cap_flush *cf;
1238                 struct ceph_mds_client *mdsc = fsc->mdsc;
1239
1240                 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
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 /*
1359  * wake up any threads waiting on this session's caps.  if the cap is
1360  * old (didn't get renewed on the client reconnect), remove it now.
1361  *
1362  * caller must hold s_mutex.
1363  */
1364 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1365                               void *arg)
1366 {
1367         struct ceph_inode_info *ci = ceph_inode(inode);
1368
1369         if (arg) {
1370                 spin_lock(&ci->i_ceph_lock);
1371                 ci->i_wanted_max_size = 0;
1372                 ci->i_requested_max_size = 0;
1373                 spin_unlock(&ci->i_ceph_lock);
1374         }
1375         wake_up_all(&ci->i_cap_wq);
1376         return 0;
1377 }
1378
1379 static void wake_up_session_caps(struct ceph_mds_session *session,
1380                                  int reconnect)
1381 {
1382         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1383         iterate_session_caps(session, wake_up_session_cb,
1384                              (void *)(unsigned long)reconnect);
1385 }
1386
1387 /*
1388  * Send periodic message to MDS renewing all currently held caps.  The
1389  * ack will reset the expiration for all caps from this session.
1390  *
1391  * caller holds s_mutex
1392  */
1393 static int send_renew_caps(struct ceph_mds_client *mdsc,
1394                            struct ceph_mds_session *session)
1395 {
1396         struct ceph_msg *msg;
1397         int state;
1398
1399         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1400             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1401                 pr_info("mds%d caps stale\n", session->s_mds);
1402         session->s_renew_requested = jiffies;
1403
1404         /* do not try to renew caps until a recovering mds has reconnected
1405          * with its clients. */
1406         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1407         if (state < CEPH_MDS_STATE_RECONNECT) {
1408                 dout("send_renew_caps ignoring mds%d (%s)\n",
1409                      session->s_mds, ceph_mds_state_name(state));
1410                 return 0;
1411         }
1412
1413         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1414                 ceph_mds_state_name(state));
1415         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1416                                  ++session->s_renew_seq);
1417         if (!msg)
1418                 return -ENOMEM;
1419         ceph_con_send(&session->s_con, msg);
1420         return 0;
1421 }
1422
1423 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1424                              struct ceph_mds_session *session, u64 seq)
1425 {
1426         struct ceph_msg *msg;
1427
1428         dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1429              session->s_mds, ceph_session_state_name(session->s_state), seq);
1430         msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1431         if (!msg)
1432                 return -ENOMEM;
1433         ceph_con_send(&session->s_con, msg);
1434         return 0;
1435 }
1436
1437
1438 /*
1439  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1440  *
1441  * Called under session->s_mutex
1442  */
1443 static void renewed_caps(struct ceph_mds_client *mdsc,
1444                          struct ceph_mds_session *session, int is_renew)
1445 {
1446         int was_stale;
1447         int wake = 0;
1448
1449         spin_lock(&session->s_cap_lock);
1450         was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1451
1452         session->s_cap_ttl = session->s_renew_requested +
1453                 mdsc->mdsmap->m_session_timeout*HZ;
1454
1455         if (was_stale) {
1456                 if (time_before(jiffies, session->s_cap_ttl)) {
1457                         pr_info("mds%d caps renewed\n", session->s_mds);
1458                         wake = 1;
1459                 } else {
1460                         pr_info("mds%d caps still stale\n", session->s_mds);
1461                 }
1462         }
1463         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1464              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1465              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1466         spin_unlock(&session->s_cap_lock);
1467
1468         if (wake)
1469                 wake_up_session_caps(session, 0);
1470 }
1471
1472 /*
1473  * send a session close request
1474  */
1475 static int request_close_session(struct ceph_mds_client *mdsc,
1476                                  struct ceph_mds_session *session)
1477 {
1478         struct ceph_msg *msg;
1479
1480         dout("request_close_session mds%d state %s seq %lld\n",
1481              session->s_mds, ceph_session_state_name(session->s_state),
1482              session->s_seq);
1483         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1484         if (!msg)
1485                 return -ENOMEM;
1486         ceph_con_send(&session->s_con, msg);
1487         return 1;
1488 }
1489
1490 /*
1491  * Called with s_mutex held.
1492  */
1493 static int __close_session(struct ceph_mds_client *mdsc,
1494                          struct ceph_mds_session *session)
1495 {
1496         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1497                 return 0;
1498         session->s_state = CEPH_MDS_SESSION_CLOSING;
1499         return request_close_session(mdsc, session);
1500 }
1501
1502 static bool drop_negative_children(struct dentry *dentry)
1503 {
1504         struct dentry *child;
1505         bool all_negative = true;
1506
1507         if (!d_is_dir(dentry))
1508                 goto out;
1509
1510         spin_lock(&dentry->d_lock);
1511         list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1512                 if (d_really_is_positive(child)) {
1513                         all_negative = false;
1514                         break;
1515                 }
1516         }
1517         spin_unlock(&dentry->d_lock);
1518
1519         if (all_negative)
1520                 shrink_dcache_parent(dentry);
1521 out:
1522         return all_negative;
1523 }
1524
1525 /*
1526  * Trim old(er) caps.
1527  *
1528  * Because we can't cache an inode without one or more caps, we do
1529  * this indirectly: if a cap is unused, we prune its aliases, at which
1530  * point the inode will hopefully get dropped to.
1531  *
1532  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1533  * memory pressure from the MDS, though, so it needn't be perfect.
1534  */
1535 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1536 {
1537         struct ceph_mds_session *session = arg;
1538         struct ceph_inode_info *ci = ceph_inode(inode);
1539         int used, wanted, oissued, mine;
1540
1541         if (session->s_trim_caps <= 0)
1542                 return -1;
1543
1544         spin_lock(&ci->i_ceph_lock);
1545         mine = cap->issued | cap->implemented;
1546         used = __ceph_caps_used(ci);
1547         wanted = __ceph_caps_file_wanted(ci);
1548         oissued = __ceph_caps_issued_other(ci, cap);
1549
1550         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1551              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1552              ceph_cap_string(used), ceph_cap_string(wanted));
1553         if (cap == ci->i_auth_cap) {
1554                 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1555                     !list_empty(&ci->i_cap_snaps))
1556                         goto out;
1557                 if ((used | wanted) & CEPH_CAP_ANY_WR)
1558                         goto out;
1559                 /* Note: it's possible that i_filelock_ref becomes non-zero
1560                  * after dropping auth caps. It doesn't hurt because reply
1561                  * of lock mds request will re-add auth caps. */
1562                 if (atomic_read(&ci->i_filelock_ref) > 0)
1563                         goto out;
1564         }
1565         /* The inode has cached pages, but it's no longer used.
1566          * we can safely drop it */
1567         if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1568             !(oissued & CEPH_CAP_FILE_CACHE)) {
1569           used = 0;
1570           oissued = 0;
1571         }
1572         if ((used | wanted) & ~oissued & mine)
1573                 goto out;   /* we need these caps */
1574
1575         if (oissued) {
1576                 /* we aren't the only cap.. just remove us */
1577                 __ceph_remove_cap(cap, true);
1578                 session->s_trim_caps--;
1579         } else {
1580                 struct dentry *dentry;
1581                 /* try dropping referring dentries */
1582                 spin_unlock(&ci->i_ceph_lock);
1583                 dentry = d_find_any_alias(inode);
1584                 if (dentry && drop_negative_children(dentry)) {
1585                         int count;
1586                         dput(dentry);
1587                         d_prune_aliases(inode);
1588                         count = atomic_read(&inode->i_count);
1589                         if (count == 1)
1590                                 session->s_trim_caps--;
1591                         dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1592                              inode, cap, count);
1593                 } else {
1594                         dput(dentry);
1595                 }
1596                 return 0;
1597         }
1598
1599 out:
1600         spin_unlock(&ci->i_ceph_lock);
1601         return 0;
1602 }
1603
1604 /*
1605  * Trim session cap count down to some max number.
1606  */
1607 int ceph_trim_caps(struct ceph_mds_client *mdsc,
1608                    struct ceph_mds_session *session,
1609                    int max_caps)
1610 {
1611         int trim_caps = session->s_nr_caps - max_caps;
1612
1613         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1614              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1615         if (trim_caps > 0) {
1616                 session->s_trim_caps = trim_caps;
1617                 iterate_session_caps(session, trim_caps_cb, session);
1618                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1619                      session->s_mds, session->s_nr_caps, max_caps,
1620                         trim_caps - session->s_trim_caps);
1621                 session->s_trim_caps = 0;
1622         }
1623
1624         ceph_send_cap_releases(mdsc, session);
1625         return 0;
1626 }
1627
1628 static int check_caps_flush(struct ceph_mds_client *mdsc,
1629                             u64 want_flush_tid)
1630 {
1631         int ret = 1;
1632
1633         spin_lock(&mdsc->cap_dirty_lock);
1634         if (!list_empty(&mdsc->cap_flush_list)) {
1635                 struct ceph_cap_flush *cf =
1636                         list_first_entry(&mdsc->cap_flush_list,
1637                                          struct ceph_cap_flush, g_list);
1638                 if (cf->tid <= want_flush_tid) {
1639                         dout("check_caps_flush still flushing tid "
1640                              "%llu <= %llu\n", cf->tid, want_flush_tid);
1641                         ret = 0;
1642                 }
1643         }
1644         spin_unlock(&mdsc->cap_dirty_lock);
1645         return ret;
1646 }
1647
1648 /*
1649  * flush all dirty inode data to disk.
1650  *
1651  * returns true if we've flushed through want_flush_tid
1652  */
1653 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1654                             u64 want_flush_tid)
1655 {
1656         dout("check_caps_flush want %llu\n", want_flush_tid);
1657
1658         wait_event(mdsc->cap_flushing_wq,
1659                    check_caps_flush(mdsc, want_flush_tid));
1660
1661         dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1662 }
1663
1664 /*
1665  * called under s_mutex
1666  */
1667 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1668                             struct ceph_mds_session *session)
1669 {
1670         struct ceph_msg *msg = NULL;
1671         struct ceph_mds_cap_release *head;
1672         struct ceph_mds_cap_item *item;
1673         struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
1674         struct ceph_cap *cap;
1675         LIST_HEAD(tmp_list);
1676         int num_cap_releases;
1677         __le32  barrier, *cap_barrier;
1678
1679         down_read(&osdc->lock);
1680         barrier = cpu_to_le32(osdc->epoch_barrier);
1681         up_read(&osdc->lock);
1682
1683         spin_lock(&session->s_cap_lock);
1684 again:
1685         list_splice_init(&session->s_cap_releases, &tmp_list);
1686         num_cap_releases = session->s_num_cap_releases;
1687         session->s_num_cap_releases = 0;
1688         spin_unlock(&session->s_cap_lock);
1689
1690         while (!list_empty(&tmp_list)) {
1691                 if (!msg) {
1692                         msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1693                                         PAGE_SIZE, GFP_NOFS, false);
1694                         if (!msg)
1695                                 goto out_err;
1696                         head = msg->front.iov_base;
1697                         head->num = cpu_to_le32(0);
1698                         msg->front.iov_len = sizeof(*head);
1699
1700                         msg->hdr.version = cpu_to_le16(2);
1701                         msg->hdr.compat_version = cpu_to_le16(1);
1702                 }
1703
1704                 cap = list_first_entry(&tmp_list, struct ceph_cap,
1705                                         session_caps);
1706                 list_del(&cap->session_caps);
1707                 num_cap_releases--;
1708
1709                 head = msg->front.iov_base;
1710                 le32_add_cpu(&head->num, 1);
1711                 item = msg->front.iov_base + msg->front.iov_len;
1712                 item->ino = cpu_to_le64(cap->cap_ino);
1713                 item->cap_id = cpu_to_le64(cap->cap_id);
1714                 item->migrate_seq = cpu_to_le32(cap->mseq);
1715                 item->seq = cpu_to_le32(cap->issue_seq);
1716                 msg->front.iov_len += sizeof(*item);
1717
1718                 ceph_put_cap(mdsc, cap);
1719
1720                 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1721                         // Append cap_barrier field
1722                         cap_barrier = msg->front.iov_base + msg->front.iov_len;
1723                         *cap_barrier = barrier;
1724                         msg->front.iov_len += sizeof(*cap_barrier);
1725
1726                         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1727                         dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1728                         ceph_con_send(&session->s_con, msg);
1729                         msg = NULL;
1730                 }
1731         }
1732
1733         BUG_ON(num_cap_releases != 0);
1734
1735         spin_lock(&session->s_cap_lock);
1736         if (!list_empty(&session->s_cap_releases))
1737                 goto again;
1738         spin_unlock(&session->s_cap_lock);
1739
1740         if (msg) {
1741                 // Append cap_barrier field
1742                 cap_barrier = msg->front.iov_base + msg->front.iov_len;
1743                 *cap_barrier = barrier;
1744                 msg->front.iov_len += sizeof(*cap_barrier);
1745
1746                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1747                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1748                 ceph_con_send(&session->s_con, msg);
1749         }
1750         return;
1751 out_err:
1752         pr_err("send_cap_releases mds%d, failed to allocate message\n",
1753                 session->s_mds);
1754         spin_lock(&session->s_cap_lock);
1755         list_splice(&tmp_list, &session->s_cap_releases);
1756         session->s_num_cap_releases += num_cap_releases;
1757         spin_unlock(&session->s_cap_lock);
1758 }
1759
1760 /*
1761  * requests
1762  */
1763
1764 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1765                                     struct inode *dir)
1766 {
1767         struct ceph_inode_info *ci = ceph_inode(dir);
1768         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1769         struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1770         size_t size = sizeof(struct ceph_mds_reply_dir_entry);
1771         int order, num_entries;
1772
1773         spin_lock(&ci->i_ceph_lock);
1774         num_entries = ci->i_files + ci->i_subdirs;
1775         spin_unlock(&ci->i_ceph_lock);
1776         num_entries = max(num_entries, 1);
1777         num_entries = min(num_entries, opt->max_readdir);
1778
1779         order = get_order(size * num_entries);
1780         while (order >= 0) {
1781                 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
1782                                                              __GFP_NOWARN,
1783                                                              order);
1784                 if (rinfo->dir_entries)
1785                         break;
1786                 order--;
1787         }
1788         if (!rinfo->dir_entries)
1789                 return -ENOMEM;
1790
1791         num_entries = (PAGE_SIZE << order) / size;
1792         num_entries = min(num_entries, opt->max_readdir);
1793
1794         rinfo->dir_buf_size = PAGE_SIZE << order;
1795         req->r_num_caps = num_entries + 1;
1796         req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1797         req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1798         return 0;
1799 }
1800
1801 /*
1802  * Create an mds request.
1803  */
1804 struct ceph_mds_request *
1805 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1806 {
1807         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1808         struct timespec64 ts;
1809
1810         if (!req)
1811                 return ERR_PTR(-ENOMEM);
1812
1813         mutex_init(&req->r_fill_mutex);
1814         req->r_mdsc = mdsc;
1815         req->r_started = jiffies;
1816         req->r_resend_mds = -1;
1817         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1818         INIT_LIST_HEAD(&req->r_unsafe_target_item);
1819         req->r_fmode = -1;
1820         kref_init(&req->r_kref);
1821         RB_CLEAR_NODE(&req->r_node);
1822         INIT_LIST_HEAD(&req->r_wait);
1823         init_completion(&req->r_completion);
1824         init_completion(&req->r_safe_completion);
1825         INIT_LIST_HEAD(&req->r_unsafe_item);
1826
1827         ktime_get_coarse_real_ts64(&ts);
1828         req->r_stamp = timespec64_trunc(ts, mdsc->fsc->sb->s_time_gran);
1829
1830         req->r_op = op;
1831         req->r_direct_mode = mode;
1832         return req;
1833 }
1834
1835 /*
1836  * return oldest (lowest) request, tid in request tree, 0 if none.
1837  *
1838  * called under mdsc->mutex.
1839  */
1840 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1841 {
1842         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1843                 return NULL;
1844         return rb_entry(rb_first(&mdsc->request_tree),
1845                         struct ceph_mds_request, r_node);
1846 }
1847
1848 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1849 {
1850         return mdsc->oldest_tid;
1851 }
1852
1853 /*
1854  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1855  * on build_path_from_dentry in fs/cifs/dir.c.
1856  *
1857  * If @stop_on_nosnap, generate path relative to the first non-snapped
1858  * inode.
1859  *
1860  * Encode hidden .snap dirs as a double /, i.e.
1861  *   foo/.snap/bar -> foo//bar
1862  */
1863 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1864                            int stop_on_nosnap)
1865 {
1866         struct dentry *temp;
1867         char *path;
1868         int len, pos;
1869         unsigned seq;
1870
1871         if (!dentry)
1872                 return ERR_PTR(-EINVAL);
1873
1874 retry:
1875         len = 0;
1876         seq = read_seqbegin(&rename_lock);
1877         rcu_read_lock();
1878         for (temp = dentry; !IS_ROOT(temp);) {
1879                 struct inode *inode = d_inode(temp);
1880                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1881                         len++;  /* slash only */
1882                 else if (stop_on_nosnap && inode &&
1883                          ceph_snap(inode) == CEPH_NOSNAP)
1884                         break;
1885                 else
1886                         len += 1 + temp->d_name.len;
1887                 temp = temp->d_parent;
1888         }
1889         rcu_read_unlock();
1890         if (len)
1891                 len--;  /* no leading '/' */
1892
1893         path = kmalloc(len+1, GFP_NOFS);
1894         if (!path)
1895                 return ERR_PTR(-ENOMEM);
1896         pos = len;
1897         path[pos] = 0;  /* trailing null */
1898         rcu_read_lock();
1899         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1900                 struct inode *inode;
1901
1902                 spin_lock(&temp->d_lock);
1903                 inode = d_inode(temp);
1904                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1905                         dout("build_path path+%d: %p SNAPDIR\n",
1906                              pos, temp);
1907                 } else if (stop_on_nosnap && inode &&
1908                            ceph_snap(inode) == CEPH_NOSNAP) {
1909                         spin_unlock(&temp->d_lock);
1910                         break;
1911                 } else {
1912                         pos -= temp->d_name.len;
1913                         if (pos < 0) {
1914                                 spin_unlock(&temp->d_lock);
1915                                 break;
1916                         }
1917                         strncpy(path + pos, temp->d_name.name,
1918                                 temp->d_name.len);
1919                 }
1920                 spin_unlock(&temp->d_lock);
1921                 if (pos)
1922                         path[--pos] = '/';
1923                 temp = temp->d_parent;
1924         }
1925         rcu_read_unlock();
1926         if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1927                 pr_err("build_path did not end path lookup where "
1928                        "expected, namelen is %d, pos is %d\n", len, pos);
1929                 /* presumably this is only possible if racing with a
1930                    rename of one of the parent directories (we can not
1931                    lock the dentries above us to prevent this, but
1932                    retrying should be harmless) */
1933                 kfree(path);
1934                 goto retry;
1935         }
1936
1937         *base = ceph_ino(d_inode(temp));
1938         *plen = len;
1939         dout("build_path on %p %d built %llx '%.*s'\n",
1940              dentry, d_count(dentry), *base, len, path);
1941         return path;
1942 }
1943
1944 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
1945                              const char **ppath, int *ppathlen, u64 *pino,
1946                              int *pfreepath)
1947 {
1948         char *path;
1949
1950         rcu_read_lock();
1951         if (!dir)
1952                 dir = d_inode_rcu(dentry->d_parent);
1953         if (dir && ceph_snap(dir) == CEPH_NOSNAP) {
1954                 *pino = ceph_ino(dir);
1955                 rcu_read_unlock();
1956                 *ppath = dentry->d_name.name;
1957                 *ppathlen = dentry->d_name.len;
1958                 return 0;
1959         }
1960         rcu_read_unlock();
1961         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1962         if (IS_ERR(path))
1963                 return PTR_ERR(path);
1964         *ppath = path;
1965         *pfreepath = 1;
1966         return 0;
1967 }
1968
1969 static int build_inode_path(struct inode *inode,
1970                             const char **ppath, int *ppathlen, u64 *pino,
1971                             int *pfreepath)
1972 {
1973         struct dentry *dentry;
1974         char *path;
1975
1976         if (ceph_snap(inode) == CEPH_NOSNAP) {
1977                 *pino = ceph_ino(inode);
1978                 *ppathlen = 0;
1979                 return 0;
1980         }
1981         dentry = d_find_alias(inode);
1982         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1983         dput(dentry);
1984         if (IS_ERR(path))
1985                 return PTR_ERR(path);
1986         *ppath = path;
1987         *pfreepath = 1;
1988         return 0;
1989 }
1990
1991 /*
1992  * request arguments may be specified via an inode *, a dentry *, or
1993  * an explicit ino+path.
1994  */
1995 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1996                                   struct inode *rdiri, const char *rpath,
1997                                   u64 rino, const char **ppath, int *pathlen,
1998                                   u64 *ino, int *freepath)
1999 {
2000         int r = 0;
2001
2002         if (rinode) {
2003                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2004                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2005                      ceph_snap(rinode));
2006         } else if (rdentry) {
2007                 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
2008                                         freepath);
2009                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
2010                      *ppath);
2011         } else if (rpath || rino) {
2012                 *ino = rino;
2013                 *ppath = rpath;
2014                 *pathlen = rpath ? strlen(rpath) : 0;
2015                 dout(" path %.*s\n", *pathlen, rpath);
2016         }
2017
2018         return r;
2019 }
2020
2021 /*
2022  * called under mdsc->mutex
2023  */
2024 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
2025                                                struct ceph_mds_request *req,
2026                                                int mds, bool drop_cap_releases)
2027 {
2028         struct ceph_msg *msg;
2029         struct ceph_mds_request_head *head;
2030         const char *path1 = NULL;
2031         const char *path2 = NULL;
2032         u64 ino1 = 0, ino2 = 0;
2033         int pathlen1 = 0, pathlen2 = 0;
2034         int freepath1 = 0, freepath2 = 0;
2035         int len;
2036         u16 releases;
2037         void *p, *end;
2038         int ret;
2039
2040         ret = set_request_path_attr(req->r_inode, req->r_dentry,
2041                               req->r_parent, req->r_path1, req->r_ino1.ino,
2042                               &path1, &pathlen1, &ino1, &freepath1);
2043         if (ret < 0) {
2044                 msg = ERR_PTR(ret);
2045                 goto out;
2046         }
2047
2048         ret = set_request_path_attr(NULL, req->r_old_dentry,
2049                               req->r_old_dentry_dir,
2050                               req->r_path2, req->r_ino2.ino,
2051                               &path2, &pathlen2, &ino2, &freepath2);
2052         if (ret < 0) {
2053                 msg = ERR_PTR(ret);
2054                 goto out_free1;
2055         }
2056
2057         len = sizeof(*head) +
2058                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2059                 sizeof(struct ceph_timespec);
2060
2061         /* calculate (max) length for cap releases */
2062         len += sizeof(struct ceph_mds_request_release) *
2063                 (!!req->r_inode_drop + !!req->r_dentry_drop +
2064                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2065         if (req->r_dentry_drop)
2066                 len += req->r_dentry->d_name.len;
2067         if (req->r_old_dentry_drop)
2068                 len += req->r_old_dentry->d_name.len;
2069
2070         msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
2071         if (!msg) {
2072                 msg = ERR_PTR(-ENOMEM);
2073                 goto out_free2;
2074         }
2075
2076         msg->hdr.version = cpu_to_le16(2);
2077         msg->hdr.tid = cpu_to_le64(req->r_tid);
2078
2079         head = msg->front.iov_base;
2080         p = msg->front.iov_base + sizeof(*head);
2081         end = msg->front.iov_base + msg->front.iov_len;
2082
2083         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2084         head->op = cpu_to_le32(req->r_op);
2085         head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
2086         head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
2087         head->args = req->r_args;
2088
2089         ceph_encode_filepath(&p, end, ino1, path1);
2090         ceph_encode_filepath(&p, end, ino2, path2);
2091
2092         /* make note of release offset, in case we need to replay */
2093         req->r_request_release_offset = p - msg->front.iov_base;
2094
2095         /* cap releases */
2096         releases = 0;
2097         if (req->r_inode_drop)
2098                 releases += ceph_encode_inode_release(&p,
2099                       req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2100                       mds, req->r_inode_drop, req->r_inode_unless, 0);
2101         if (req->r_dentry_drop)
2102                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
2103                                 req->r_parent, mds, req->r_dentry_drop,
2104                                 req->r_dentry_unless);
2105         if (req->r_old_dentry_drop)
2106                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2107                                 req->r_old_dentry_dir, mds,
2108                                 req->r_old_dentry_drop,
2109                                 req->r_old_dentry_unless);
2110         if (req->r_old_inode_drop)
2111                 releases += ceph_encode_inode_release(&p,
2112                       d_inode(req->r_old_dentry),
2113                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2114
2115         if (drop_cap_releases) {
2116                 releases = 0;
2117                 p = msg->front.iov_base + req->r_request_release_offset;
2118         }
2119
2120         head->num_releases = cpu_to_le16(releases);
2121
2122         /* time stamp */
2123         {
2124                 struct ceph_timespec ts;
2125                 ceph_encode_timespec64(&ts, &req->r_stamp);
2126                 ceph_encode_copy(&p, &ts, sizeof(ts));
2127         }
2128
2129         BUG_ON(p > end);
2130         msg->front.iov_len = p - msg->front.iov_base;
2131         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2132
2133         if (req->r_pagelist) {
2134                 struct ceph_pagelist *pagelist = req->r_pagelist;
2135                 ceph_msg_data_add_pagelist(msg, pagelist);
2136                 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2137         } else {
2138                 msg->hdr.data_len = 0;
2139         }
2140
2141         msg->hdr.data_off = cpu_to_le16(0);
2142
2143 out_free2:
2144         if (freepath2)
2145                 kfree((char *)path2);
2146 out_free1:
2147         if (freepath1)
2148                 kfree((char *)path1);
2149 out:
2150         return msg;
2151 }
2152
2153 /*
2154  * called under mdsc->mutex if error, under no mutex if
2155  * success.
2156  */
2157 static void complete_request(struct ceph_mds_client *mdsc,
2158                              struct ceph_mds_request *req)
2159 {
2160         if (req->r_callback)
2161                 req->r_callback(mdsc, req);
2162         else
2163                 complete_all(&req->r_completion);
2164 }
2165
2166 /*
2167  * called under mdsc->mutex
2168  */
2169 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2170                                   struct ceph_mds_request *req,
2171                                   int mds, bool drop_cap_releases)
2172 {
2173         struct ceph_mds_request_head *rhead;
2174         struct ceph_msg *msg;
2175         int flags = 0;
2176
2177         req->r_attempts++;
2178         if (req->r_inode) {
2179                 struct ceph_cap *cap =
2180                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2181
2182                 if (cap)
2183                         req->r_sent_on_mseq = cap->mseq;
2184                 else
2185                         req->r_sent_on_mseq = -1;
2186         }
2187         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2188              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2189
2190         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2191                 void *p;
2192                 /*
2193                  * Replay.  Do not regenerate message (and rebuild
2194                  * paths, etc.); just use the original message.
2195                  * Rebuilding paths will break for renames because
2196                  * d_move mangles the src name.
2197                  */
2198                 msg = req->r_request;
2199                 rhead = msg->front.iov_base;
2200
2201                 flags = le32_to_cpu(rhead->flags);
2202                 flags |= CEPH_MDS_FLAG_REPLAY;
2203                 rhead->flags = cpu_to_le32(flags);
2204
2205                 if (req->r_target_inode)
2206                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2207
2208                 rhead->num_retry = req->r_attempts - 1;
2209
2210                 /* remove cap/dentry releases from message */
2211                 rhead->num_releases = 0;
2212
2213                 /* time stamp */
2214                 p = msg->front.iov_base + req->r_request_release_offset;
2215                 {
2216                         struct ceph_timespec ts;
2217                         ceph_encode_timespec64(&ts, &req->r_stamp);
2218                         ceph_encode_copy(&p, &ts, sizeof(ts));
2219                 }
2220
2221                 msg->front.iov_len = p - msg->front.iov_base;
2222                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2223                 return 0;
2224         }
2225
2226         if (req->r_request) {
2227                 ceph_msg_put(req->r_request);
2228                 req->r_request = NULL;
2229         }
2230         msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2231         if (IS_ERR(msg)) {
2232                 req->r_err = PTR_ERR(msg);
2233                 return PTR_ERR(msg);
2234         }
2235         req->r_request = msg;
2236
2237         rhead = msg->front.iov_base;
2238         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2239         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2240                 flags |= CEPH_MDS_FLAG_REPLAY;
2241         if (req->r_parent)
2242                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2243         rhead->flags = cpu_to_le32(flags);
2244         rhead->num_fwd = req->r_num_fwd;
2245         rhead->num_retry = req->r_attempts - 1;
2246         rhead->ino = 0;
2247
2248         dout(" r_parent = %p\n", req->r_parent);
2249         return 0;
2250 }
2251
2252 /*
2253  * send request, or put it on the appropriate wait list.
2254  */
2255 static void __do_request(struct ceph_mds_client *mdsc,
2256                         struct ceph_mds_request *req)
2257 {
2258         struct ceph_mds_session *session = NULL;
2259         int mds = -1;
2260         int err = 0;
2261
2262         if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2263                 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2264                         __unregister_request(mdsc, req);
2265                 return;
2266         }
2267
2268         if (req->r_timeout &&
2269             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2270                 dout("do_request timed out\n");
2271                 err = -EIO;
2272                 goto finish;
2273         }
2274         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2275                 dout("do_request forced umount\n");
2276                 err = -EIO;
2277                 goto finish;
2278         }
2279         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2280                 if (mdsc->mdsmap_err) {
2281                         err = mdsc->mdsmap_err;
2282                         dout("do_request mdsmap err %d\n", err);
2283                         goto finish;
2284                 }
2285                 if (mdsc->mdsmap->m_epoch == 0) {
2286                         dout("do_request no mdsmap, waiting for map\n");
2287                         list_add(&req->r_wait, &mdsc->waiting_for_map);
2288                         return;
2289                 }
2290                 if (!(mdsc->fsc->mount_options->flags &
2291                       CEPH_MOUNT_OPT_MOUNTWAIT) &&
2292                     !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2293                         err = -ENOENT;
2294                         pr_info("probably no mds server is up\n");
2295                         goto finish;
2296                 }
2297         }
2298
2299         put_request_session(req);
2300
2301         mds = __choose_mds(mdsc, req);
2302         if (mds < 0 ||
2303             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2304                 dout("do_request no mds or not active, waiting for map\n");
2305                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2306                 return;
2307         }
2308
2309         /* get, open session */
2310         session = __ceph_lookup_mds_session(mdsc, mds);
2311         if (!session) {
2312                 session = register_session(mdsc, mds);
2313                 if (IS_ERR(session)) {
2314                         err = PTR_ERR(session);
2315                         goto finish;
2316                 }
2317         }
2318         req->r_session = get_session(session);
2319
2320         dout("do_request mds%d session %p state %s\n", mds, session,
2321              ceph_session_state_name(session->s_state));
2322         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2323             session->s_state != CEPH_MDS_SESSION_HUNG) {
2324                 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2325                         err = -EACCES;
2326                         goto out_session;
2327                 }
2328                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2329                     session->s_state == CEPH_MDS_SESSION_CLOSING)
2330                         __open_session(mdsc, session);
2331                 list_add(&req->r_wait, &session->s_waiting);
2332                 goto out_session;
2333         }
2334
2335         /* send request */
2336         req->r_resend_mds = -1;   /* forget any previous mds hint */
2337
2338         if (req->r_request_started == 0)   /* note request start time */
2339                 req->r_request_started = jiffies;
2340
2341         err = __prepare_send_request(mdsc, req, mds, false);
2342         if (!err) {
2343                 ceph_msg_get(req->r_request);
2344                 ceph_con_send(&session->s_con, req->r_request);
2345         }
2346
2347 out_session:
2348         ceph_put_mds_session(session);
2349 finish:
2350         if (err) {
2351                 dout("__do_request early error %d\n", err);
2352                 req->r_err = err;
2353                 complete_request(mdsc, req);
2354                 __unregister_request(mdsc, req);
2355         }
2356         return;
2357 }
2358
2359 /*
2360  * called under mdsc->mutex
2361  */
2362 static void __wake_requests(struct ceph_mds_client *mdsc,
2363                             struct list_head *head)
2364 {
2365         struct ceph_mds_request *req;
2366         LIST_HEAD(tmp_list);
2367
2368         list_splice_init(head, &tmp_list);
2369
2370         while (!list_empty(&tmp_list)) {
2371                 req = list_entry(tmp_list.next,
2372                                  struct ceph_mds_request, r_wait);
2373                 list_del_init(&req->r_wait);
2374                 dout(" wake request %p tid %llu\n", req, req->r_tid);
2375                 __do_request(mdsc, req);
2376         }
2377 }
2378
2379 /*
2380  * Wake up threads with requests pending for @mds, so that they can
2381  * resubmit their requests to a possibly different mds.
2382  */
2383 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2384 {
2385         struct ceph_mds_request *req;
2386         struct rb_node *p = rb_first(&mdsc->request_tree);
2387
2388         dout("kick_requests mds%d\n", mds);
2389         while (p) {
2390                 req = rb_entry(p, struct ceph_mds_request, r_node);
2391                 p = rb_next(p);
2392                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2393                         continue;
2394                 if (req->r_attempts > 0)
2395                         continue; /* only new requests */
2396                 if (req->r_session &&
2397                     req->r_session->s_mds == mds) {
2398                         dout(" kicking tid %llu\n", req->r_tid);
2399                         list_del_init(&req->r_wait);
2400                         __do_request(mdsc, req);
2401                 }
2402         }
2403 }
2404
2405 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2406                               struct ceph_mds_request *req)
2407 {
2408         dout("submit_request on %p\n", req);
2409         mutex_lock(&mdsc->mutex);
2410         __register_request(mdsc, req, NULL);
2411         __do_request(mdsc, req);
2412         mutex_unlock(&mdsc->mutex);
2413 }
2414
2415 /*
2416  * Synchrously perform an mds request.  Take care of all of the
2417  * session setup, forwarding, retry details.
2418  */
2419 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2420                          struct inode *dir,
2421                          struct ceph_mds_request *req)
2422 {
2423         int err;
2424
2425         dout("do_request on %p\n", req);
2426
2427         /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2428         if (req->r_inode)
2429                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2430         if (req->r_parent)
2431                 ceph_get_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
2432         if (req->r_old_dentry_dir)
2433                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2434                                   CEPH_CAP_PIN);
2435
2436         /* issue */
2437         mutex_lock(&mdsc->mutex);
2438         __register_request(mdsc, req, dir);
2439         __do_request(mdsc, req);
2440
2441         if (req->r_err) {
2442                 err = req->r_err;
2443                 goto out;
2444         }
2445
2446         /* wait */
2447         mutex_unlock(&mdsc->mutex);
2448         dout("do_request waiting\n");
2449         if (!req->r_timeout && req->r_wait_for_completion) {
2450                 err = req->r_wait_for_completion(mdsc, req);
2451         } else {
2452                 long timeleft = wait_for_completion_killable_timeout(
2453                                         &req->r_completion,
2454                                         ceph_timeout_jiffies(req->r_timeout));
2455                 if (timeleft > 0)
2456                         err = 0;
2457                 else if (!timeleft)
2458                         err = -EIO;  /* timed out */
2459                 else
2460                         err = timeleft;  /* killed */
2461         }
2462         dout("do_request waited, got %d\n", err);
2463         mutex_lock(&mdsc->mutex);
2464
2465         /* only abort if we didn't race with a real reply */
2466         if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2467                 err = le32_to_cpu(req->r_reply_info.head->result);
2468         } else if (err < 0) {
2469                 dout("aborted request %lld with %d\n", req->r_tid, err);
2470
2471                 /*
2472                  * ensure we aren't running concurrently with
2473                  * ceph_fill_trace or ceph_readdir_prepopulate, which
2474                  * rely on locks (dir mutex) held by our caller.
2475                  */
2476                 mutex_lock(&req->r_fill_mutex);
2477                 req->r_err = err;
2478                 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2479                 mutex_unlock(&req->r_fill_mutex);
2480
2481                 if (req->r_parent &&
2482                     (req->r_op & CEPH_MDS_OP_WRITE))
2483                         ceph_invalidate_dir_request(req);
2484         } else {
2485                 err = req->r_err;
2486         }
2487
2488 out:
2489         mutex_unlock(&mdsc->mutex);
2490         dout("do_request %p done, result %d\n", req, err);
2491         return err;
2492 }
2493
2494 /*
2495  * Invalidate dir's completeness, dentry lease state on an aborted MDS
2496  * namespace request.
2497  */
2498 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2499 {
2500         struct inode *dir = req->r_parent;
2501         struct inode *old_dir = req->r_old_dentry_dir;
2502
2503         dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
2504
2505         ceph_dir_clear_complete(dir);
2506         if (old_dir)
2507                 ceph_dir_clear_complete(old_dir);
2508         if (req->r_dentry)
2509                 ceph_invalidate_dentry_lease(req->r_dentry);
2510         if (req->r_old_dentry)
2511                 ceph_invalidate_dentry_lease(req->r_old_dentry);
2512 }
2513
2514 /*
2515  * Handle mds reply.
2516  *
2517  * We take the session mutex and parse and process the reply immediately.
2518  * This preserves the logical ordering of replies, capabilities, etc., sent
2519  * by the MDS as they are applied to our local cache.
2520  */
2521 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2522 {
2523         struct ceph_mds_client *mdsc = session->s_mdsc;
2524         struct ceph_mds_request *req;
2525         struct ceph_mds_reply_head *head = msg->front.iov_base;
2526         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2527         struct ceph_snap_realm *realm;
2528         u64 tid;
2529         int err, result;
2530         int mds = session->s_mds;
2531
2532         if (msg->front.iov_len < sizeof(*head)) {
2533                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2534                 ceph_msg_dump(msg);
2535                 return;
2536         }
2537
2538         /* get request, session */
2539         tid = le64_to_cpu(msg->hdr.tid);
2540         mutex_lock(&mdsc->mutex);
2541         req = lookup_get_request(mdsc, tid);
2542         if (!req) {
2543                 dout("handle_reply on unknown tid %llu\n", tid);
2544                 mutex_unlock(&mdsc->mutex);
2545                 return;
2546         }
2547         dout("handle_reply %p\n", req);
2548
2549         /* correct session? */
2550         if (req->r_session != session) {
2551                 pr_err("mdsc_handle_reply got %llu on session mds%d"
2552                        " not mds%d\n", tid, session->s_mds,
2553                        req->r_session ? req->r_session->s_mds : -1);
2554                 mutex_unlock(&mdsc->mutex);
2555                 goto out;
2556         }
2557
2558         /* dup? */
2559         if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
2560             (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
2561                 pr_warn("got a dup %s reply on %llu from mds%d\n",
2562                            head->safe ? "safe" : "unsafe", tid, mds);
2563                 mutex_unlock(&mdsc->mutex);
2564                 goto out;
2565         }
2566         if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
2567                 pr_warn("got unsafe after safe on %llu from mds%d\n",
2568                            tid, mds);
2569                 mutex_unlock(&mdsc->mutex);
2570                 goto out;
2571         }
2572
2573         result = le32_to_cpu(head->result);
2574
2575         /*
2576          * Handle an ESTALE
2577          * if we're not talking to the authority, send to them
2578          * if the authority has changed while we weren't looking,
2579          * send to new authority
2580          * Otherwise we just have to return an ESTALE
2581          */
2582         if (result == -ESTALE) {
2583                 dout("got ESTALE on request %llu\n", req->r_tid);
2584                 req->r_resend_mds = -1;
2585                 if (req->r_direct_mode != USE_AUTH_MDS) {
2586                         dout("not using auth, setting for that now\n");
2587                         req->r_direct_mode = USE_AUTH_MDS;
2588                         __do_request(mdsc, req);
2589                         mutex_unlock(&mdsc->mutex);
2590                         goto out;
2591                 } else  {
2592                         int mds = __choose_mds(mdsc, req);
2593                         if (mds >= 0 && mds != req->r_session->s_mds) {
2594                                 dout("but auth changed, so resending\n");
2595                                 __do_request(mdsc, req);
2596                                 mutex_unlock(&mdsc->mutex);
2597                                 goto out;
2598                         }
2599                 }
2600                 dout("have to return ESTALE on request %llu\n", req->r_tid);
2601         }
2602
2603
2604         if (head->safe) {
2605                 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
2606                 __unregister_request(mdsc, req);
2607
2608                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2609                         /*
2610                          * We already handled the unsafe response, now do the
2611                          * cleanup.  No need to examine the response; the MDS
2612                          * doesn't include any result info in the safe
2613                          * response.  And even if it did, there is nothing
2614                          * useful we could do with a revised return value.
2615                          */
2616                         dout("got safe reply %llu, mds%d\n", tid, mds);
2617
2618                         /* last unsafe request during umount? */
2619                         if (mdsc->stopping && !__get_oldest_req(mdsc))
2620                                 complete_all(&mdsc->safe_umount_waiters);
2621                         mutex_unlock(&mdsc->mutex);
2622                         goto out;
2623                 }
2624         } else {
2625                 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
2626                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2627                 if (req->r_unsafe_dir) {
2628                         struct ceph_inode_info *ci =
2629                                         ceph_inode(req->r_unsafe_dir);
2630                         spin_lock(&ci->i_unsafe_lock);
2631                         list_add_tail(&req->r_unsafe_dir_item,
2632                                       &ci->i_unsafe_dirops);
2633                         spin_unlock(&ci->i_unsafe_lock);
2634                 }
2635         }
2636
2637         dout("handle_reply tid %lld result %d\n", tid, result);
2638         rinfo = &req->r_reply_info;
2639         err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2640         mutex_unlock(&mdsc->mutex);
2641
2642         mutex_lock(&session->s_mutex);
2643         if (err < 0) {
2644                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2645                 ceph_msg_dump(msg);
2646                 goto out_err;
2647         }
2648
2649         /* snap trace */
2650         realm = NULL;
2651         if (rinfo->snapblob_len) {
2652                 down_write(&mdsc->snap_rwsem);
2653                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2654                                 rinfo->snapblob + rinfo->snapblob_len,
2655                                 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2656                                 &realm);
2657                 downgrade_write(&mdsc->snap_rwsem);
2658         } else {
2659                 down_read(&mdsc->snap_rwsem);
2660         }
2661
2662         /* insert trace into our cache */
2663         mutex_lock(&req->r_fill_mutex);
2664         current->journal_info = req;
2665         err = ceph_fill_trace(mdsc->fsc->sb, req);
2666         if (err == 0) {
2667                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2668                                     req->r_op == CEPH_MDS_OP_LSSNAP))
2669                         ceph_readdir_prepopulate(req, req->r_session);
2670                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2671         }
2672         current->journal_info = NULL;
2673         mutex_unlock(&req->r_fill_mutex);
2674
2675         up_read(&mdsc->snap_rwsem);
2676         if (realm)
2677                 ceph_put_snap_realm(mdsc, realm);
2678
2679         if (err == 0 && req->r_target_inode &&
2680             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2681                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
2682                 spin_lock(&ci->i_unsafe_lock);
2683                 list_add_tail(&req->r_unsafe_target_item, &ci->i_unsafe_iops);
2684                 spin_unlock(&ci->i_unsafe_lock);
2685         }
2686 out_err:
2687         mutex_lock(&mdsc->mutex);
2688         if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2689                 if (err) {
2690                         req->r_err = err;
2691                 } else {
2692                         req->r_reply =  ceph_msg_get(msg);
2693                         set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
2694                 }
2695         } else {
2696                 dout("reply arrived after request %lld was aborted\n", tid);
2697         }
2698         mutex_unlock(&mdsc->mutex);
2699
2700         mutex_unlock(&session->s_mutex);
2701
2702         /* kick calling process */
2703         complete_request(mdsc, req);
2704 out:
2705         ceph_mdsc_put_request(req);
2706         return;
2707 }
2708
2709
2710
2711 /*
2712  * handle mds notification that our request has been forwarded.
2713  */
2714 static void handle_forward(struct ceph_mds_client *mdsc,
2715                            struct ceph_mds_session *session,
2716                            struct ceph_msg *msg)
2717 {
2718         struct ceph_mds_request *req;
2719         u64 tid = le64_to_cpu(msg->hdr.tid);
2720         u32 next_mds;
2721         u32 fwd_seq;
2722         int err = -EINVAL;
2723         void *p = msg->front.iov_base;
2724         void *end = p + msg->front.iov_len;
2725
2726         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2727         next_mds = ceph_decode_32(&p);
2728         fwd_seq = ceph_decode_32(&p);
2729
2730         mutex_lock(&mdsc->mutex);
2731         req = lookup_get_request(mdsc, tid);
2732         if (!req) {
2733                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2734                 goto out;  /* dup reply? */
2735         }
2736
2737         if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2738                 dout("forward tid %llu aborted, unregistering\n", tid);
2739                 __unregister_request(mdsc, req);
2740         } else if (fwd_seq <= req->r_num_fwd) {
2741                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2742                      tid, next_mds, req->r_num_fwd, fwd_seq);
2743         } else {
2744                 /* resend. forward race not possible; mds would drop */
2745                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2746                 BUG_ON(req->r_err);
2747                 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
2748                 req->r_attempts = 0;
2749                 req->r_num_fwd = fwd_seq;
2750                 req->r_resend_mds = next_mds;
2751                 put_request_session(req);
2752                 __do_request(mdsc, req);
2753         }
2754         ceph_mdsc_put_request(req);
2755 out:
2756         mutex_unlock(&mdsc->mutex);
2757         return;
2758
2759 bad:
2760         pr_err("mdsc_handle_forward decode error err=%d\n", err);
2761 }
2762
2763 /*
2764  * handle a mds session control message
2765  */
2766 static void handle_session(struct ceph_mds_session *session,
2767                            struct ceph_msg *msg)
2768 {
2769         struct ceph_mds_client *mdsc = session->s_mdsc;
2770         u32 op;
2771         u64 seq;
2772         int mds = session->s_mds;
2773         struct ceph_mds_session_head *h = msg->front.iov_base;
2774         int wake = 0;
2775
2776         /* decode */
2777         if (msg->front.iov_len < sizeof(*h))
2778                 goto bad;
2779         op = le32_to_cpu(h->op);
2780         seq = le64_to_cpu(h->seq);
2781
2782         mutex_lock(&mdsc->mutex);
2783         if (op == CEPH_SESSION_CLOSE) {
2784                 get_session(session);
2785                 __unregister_session(mdsc, session);
2786         }
2787         /* FIXME: this ttl calculation is generous */
2788         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2789         mutex_unlock(&mdsc->mutex);
2790
2791         mutex_lock(&session->s_mutex);
2792
2793         dout("handle_session mds%d %s %p state %s seq %llu\n",
2794              mds, ceph_session_op_name(op), session,
2795              ceph_session_state_name(session->s_state), seq);
2796
2797         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2798                 session->s_state = CEPH_MDS_SESSION_OPEN;
2799                 pr_info("mds%d came back\n", session->s_mds);
2800         }
2801
2802         switch (op) {
2803         case CEPH_SESSION_OPEN:
2804                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2805                         pr_info("mds%d reconnect success\n", session->s_mds);
2806                 session->s_state = CEPH_MDS_SESSION_OPEN;
2807                 renewed_caps(mdsc, session, 0);
2808                 wake = 1;
2809                 if (mdsc->stopping)
2810                         __close_session(mdsc, session);
2811                 break;
2812
2813         case CEPH_SESSION_RENEWCAPS:
2814                 if (session->s_renew_seq == seq)
2815                         renewed_caps(mdsc, session, 1);
2816                 break;
2817
2818         case CEPH_SESSION_CLOSE:
2819                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2820                         pr_info("mds%d reconnect denied\n", session->s_mds);
2821                 cleanup_session_requests(mdsc, session);
2822                 remove_session_caps(session);
2823                 wake = 2; /* for good measure */
2824                 wake_up_all(&mdsc->session_close_wq);
2825                 break;
2826
2827         case CEPH_SESSION_STALE:
2828                 pr_info("mds%d caps went stale, renewing\n",
2829                         session->s_mds);
2830                 spin_lock(&session->s_gen_ttl_lock);
2831                 session->s_cap_gen++;
2832                 session->s_cap_ttl = jiffies - 1;
2833                 spin_unlock(&session->s_gen_ttl_lock);
2834                 send_renew_caps(mdsc, session);
2835                 break;
2836
2837         case CEPH_SESSION_RECALL_STATE:
2838                 ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2839                 break;
2840
2841         case CEPH_SESSION_FLUSHMSG:
2842                 send_flushmsg_ack(mdsc, session, seq);
2843                 break;
2844
2845         case CEPH_SESSION_FORCE_RO:
2846                 dout("force_session_readonly %p\n", session);
2847                 spin_lock(&session->s_cap_lock);
2848                 session->s_readonly = true;
2849                 spin_unlock(&session->s_cap_lock);
2850                 wake_up_session_caps(session, 0);
2851                 break;
2852
2853         case CEPH_SESSION_REJECT:
2854                 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
2855                 pr_info("mds%d rejected session\n", session->s_mds);
2856                 session->s_state = CEPH_MDS_SESSION_REJECTED;
2857                 cleanup_session_requests(mdsc, session);
2858                 remove_session_caps(session);
2859                 wake = 2; /* for good measure */
2860                 break;
2861
2862         default:
2863                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2864                 WARN_ON(1);
2865         }
2866
2867         mutex_unlock(&session->s_mutex);
2868         if (wake) {
2869                 mutex_lock(&mdsc->mutex);
2870                 __wake_requests(mdsc, &session->s_waiting);
2871                 if (wake == 2)
2872                         kick_requests(mdsc, mds);
2873                 mutex_unlock(&mdsc->mutex);
2874         }
2875         if (op == CEPH_SESSION_CLOSE)
2876                 ceph_put_mds_session(session);
2877         return;
2878
2879 bad:
2880         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2881                (int)msg->front.iov_len);
2882         ceph_msg_dump(msg);
2883         return;
2884 }
2885
2886
2887 /*
2888  * called under session->mutex.
2889  */
2890 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2891                                    struct ceph_mds_session *session)
2892 {
2893         struct ceph_mds_request *req, *nreq;
2894         struct rb_node *p;
2895         int err;
2896
2897         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2898
2899         mutex_lock(&mdsc->mutex);
2900         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2901                 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2902                 if (!err) {
2903                         ceph_msg_get(req->r_request);
2904                         ceph_con_send(&session->s_con, req->r_request);
2905                 }
2906         }
2907
2908         /*
2909          * also re-send old requests when MDS enters reconnect stage. So that MDS
2910          * can process completed request in clientreplay stage.
2911          */
2912         p = rb_first(&mdsc->request_tree);
2913         while (p) {
2914                 req = rb_entry(p, struct ceph_mds_request, r_node);
2915                 p = rb_next(p);
2916                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2917                         continue;
2918                 if (req->r_attempts == 0)
2919                         continue; /* only old requests */
2920                 if (req->r_session &&
2921                     req->r_session->s_mds == session->s_mds) {
2922                         err = __prepare_send_request(mdsc, req,
2923                                                      session->s_mds, true);
2924                         if (!err) {
2925                                 ceph_msg_get(req->r_request);
2926                                 ceph_con_send(&session->s_con, req->r_request);
2927                         }
2928                 }
2929         }
2930         mutex_unlock(&mdsc->mutex);
2931 }
2932
2933 /*
2934  * Encode information about a cap for a reconnect with the MDS.
2935  */
2936 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2937                           void *arg)
2938 {
2939         union {
2940                 struct ceph_mds_cap_reconnect v2;
2941                 struct ceph_mds_cap_reconnect_v1 v1;
2942         } rec;
2943         struct ceph_inode_info *ci = cap->ci;
2944         struct ceph_reconnect_state *recon_state = arg;
2945         struct ceph_pagelist *pagelist = recon_state->pagelist;
2946         char *path;
2947         int pathlen, err;
2948         u64 pathbase;
2949         u64 snap_follows;
2950         struct dentry *dentry;
2951
2952         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2953              inode, ceph_vinop(inode), cap, cap->cap_id,
2954              ceph_cap_string(cap->issued));
2955         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2956         if (err)
2957                 return err;
2958
2959         dentry = d_find_alias(inode);
2960         if (dentry) {
2961                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2962                 if (IS_ERR(path)) {
2963                         err = PTR_ERR(path);
2964                         goto out_dput;
2965                 }
2966         } else {
2967                 path = NULL;
2968                 pathlen = 0;
2969                 pathbase = 0;
2970         }
2971
2972         spin_lock(&ci->i_ceph_lock);
2973         cap->seq = 0;        /* reset cap seq */
2974         cap->issue_seq = 0;  /* and issue_seq */
2975         cap->mseq = 0;       /* and migrate_seq */
2976         cap->cap_gen = cap->session->s_cap_gen;
2977
2978         if (recon_state->msg_version >= 2) {
2979                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2980                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2981                 rec.v2.issued = cpu_to_le32(cap->issued);
2982                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2983                 rec.v2.pathbase = cpu_to_le64(pathbase);
2984                 rec.v2.flock_len = (__force __le32)
2985                         ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
2986         } else {
2987                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2988                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2989                 rec.v1.issued = cpu_to_le32(cap->issued);
2990                 rec.v1.size = cpu_to_le64(inode->i_size);
2991                 ceph_encode_timespec64(&rec.v1.mtime, &inode->i_mtime);
2992                 ceph_encode_timespec64(&rec.v1.atime, &inode->i_atime);
2993                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2994                 rec.v1.pathbase = cpu_to_le64(pathbase);
2995         }
2996
2997         if (list_empty(&ci->i_cap_snaps)) {
2998                 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
2999         } else {
3000                 struct ceph_cap_snap *capsnap =
3001                         list_first_entry(&ci->i_cap_snaps,
3002                                          struct ceph_cap_snap, ci_item);
3003                 snap_follows = capsnap->follows;
3004         }
3005         spin_unlock(&ci->i_ceph_lock);
3006
3007         if (recon_state->msg_version >= 2) {
3008                 int num_fcntl_locks, num_flock_locks;
3009                 struct ceph_filelock *flocks = NULL;
3010                 size_t struct_len, total_len = 0;
3011                 u8 struct_v = 0;
3012
3013 encode_again:
3014                 if (rec.v2.flock_len) {
3015                         ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
3016                 } else {
3017                         num_fcntl_locks = 0;
3018                         num_flock_locks = 0;
3019                 }
3020                 if (num_fcntl_locks + num_flock_locks > 0) {
3021                         flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
3022                                                sizeof(struct ceph_filelock),
3023                                                GFP_NOFS);
3024                         if (!flocks) {
3025                                 err = -ENOMEM;
3026                                 goto out_free;
3027                         }
3028                         err = ceph_encode_locks_to_buffer(inode, flocks,
3029                                                           num_fcntl_locks,
3030                                                           num_flock_locks);
3031                         if (err) {
3032                                 kfree(flocks);
3033                                 flocks = NULL;
3034                                 if (err == -ENOSPC)
3035                                         goto encode_again;
3036                                 goto out_free;
3037                         }
3038                 } else {
3039                         kfree(flocks);
3040                         flocks = NULL;
3041                 }
3042
3043                 if (recon_state->msg_version >= 3) {
3044                         /* version, compat_version and struct_len */
3045                         total_len = 2 * sizeof(u8) + sizeof(u32);
3046                         struct_v = 2;
3047                 }
3048                 /*
3049                  * number of encoded locks is stable, so copy to pagelist
3050                  */
3051                 struct_len = 2 * sizeof(u32) +
3052                             (num_fcntl_locks + num_flock_locks) *
3053                             sizeof(struct ceph_filelock);
3054                 rec.v2.flock_len = cpu_to_le32(struct_len);
3055
3056                 struct_len += sizeof(rec.v2);
3057                 struct_len += sizeof(u32) + pathlen;
3058
3059                 if (struct_v >= 2)
3060                         struct_len += sizeof(u64); /* snap_follows */
3061
3062                 total_len += struct_len;
3063                 err = ceph_pagelist_reserve(pagelist, total_len);
3064
3065                 if (!err) {
3066                         if (recon_state->msg_version >= 3) {
3067                                 ceph_pagelist_encode_8(pagelist, struct_v);
3068                                 ceph_pagelist_encode_8(pagelist, 1);
3069                                 ceph_pagelist_encode_32(pagelist, struct_len);
3070                         }
3071                         ceph_pagelist_encode_string(pagelist, path, pathlen);
3072                         ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
3073                         ceph_locks_to_pagelist(flocks, pagelist,
3074                                                num_fcntl_locks,
3075                                                num_flock_locks);
3076                         if (struct_v >= 2)
3077                                 ceph_pagelist_encode_64(pagelist, snap_follows);
3078                 }
3079                 kfree(flocks);
3080         } else {
3081                 size_t size = sizeof(u32) + pathlen + sizeof(rec.v1);
3082                 err = ceph_pagelist_reserve(pagelist, size);
3083                 if (!err) {
3084                         ceph_pagelist_encode_string(pagelist, path, pathlen);
3085                         ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
3086                 }
3087         }
3088
3089         recon_state->nr_caps++;
3090 out_free:
3091         kfree(path);
3092 out_dput:
3093         dput(dentry);
3094         return err;
3095 }
3096
3097
3098 /*
3099  * If an MDS fails and recovers, clients need to reconnect in order to
3100  * reestablish shared state.  This includes all caps issued through
3101  * this session _and_ the snap_realm hierarchy.  Because it's not
3102  * clear which snap realms the mds cares about, we send everything we
3103  * know about.. that ensures we'll then get any new info the
3104  * recovering MDS might have.
3105  *
3106  * This is a relatively heavyweight operation, but it's rare.
3107  *
3108  * called with mdsc->mutex held.
3109  */
3110 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3111                                struct ceph_mds_session *session)
3112 {
3113         struct ceph_msg *reply;
3114         struct rb_node *p;
3115         int mds = session->s_mds;
3116         int err = -ENOMEM;
3117         int s_nr_caps;
3118         struct ceph_pagelist *pagelist;
3119         struct ceph_reconnect_state recon_state;
3120         LIST_HEAD(dispose);
3121
3122         pr_info("mds%d reconnect start\n", mds);
3123
3124         pagelist = ceph_pagelist_alloc(GFP_NOFS);
3125         if (!pagelist)
3126                 goto fail_nopagelist;
3127
3128         reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3129         if (!reply)
3130                 goto fail_nomsg;
3131
3132         mutex_lock(&session->s_mutex);
3133         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3134         session->s_seq = 0;
3135
3136         dout("session %p state %s\n", session,
3137              ceph_session_state_name(session->s_state));
3138
3139         spin_lock(&session->s_gen_ttl_lock);
3140         session->s_cap_gen++;
3141         spin_unlock(&session->s_gen_ttl_lock);
3142
3143         spin_lock(&session->s_cap_lock);
3144         /* don't know if session is readonly */
3145         session->s_readonly = 0;
3146         /*
3147          * notify __ceph_remove_cap() that we are composing cap reconnect.
3148          * If a cap get released before being added to the cap reconnect,
3149          * __ceph_remove_cap() should skip queuing cap release.
3150          */
3151         session->s_cap_reconnect = 1;
3152         /* drop old cap expires; we're about to reestablish that state */
3153         detach_cap_releases(session, &dispose);
3154         spin_unlock(&session->s_cap_lock);
3155         dispose_cap_releases(mdsc, &dispose);
3156
3157         /* trim unused caps to reduce MDS's cache rejoin time */
3158         if (mdsc->fsc->sb->s_root)
3159                 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3160
3161         ceph_con_close(&session->s_con);
3162         ceph_con_open(&session->s_con,
3163                       CEPH_ENTITY_TYPE_MDS, mds,
3164                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3165
3166         /* replay unsafe requests */
3167         replay_unsafe_requests(mdsc, session);
3168
3169         down_read(&mdsc->snap_rwsem);
3170
3171         /* traverse this session's caps */
3172         s_nr_caps = session->s_nr_caps;
3173         err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
3174         if (err)
3175                 goto fail;
3176
3177         recon_state.nr_caps = 0;
3178         recon_state.pagelist = pagelist;
3179         if (session->s_con.peer_features & CEPH_FEATURE_MDSENC)
3180                 recon_state.msg_version = 3;
3181         else
3182                 recon_state.msg_version = 2;
3183         err = iterate_session_caps(session, encode_caps_cb, &recon_state);
3184         if (err < 0)
3185                 goto fail;
3186
3187         spin_lock(&session->s_cap_lock);
3188         session->s_cap_reconnect = 0;
3189         spin_unlock(&session->s_cap_lock);
3190
3191         /*
3192          * snaprealms.  we provide mds with the ino, seq (version), and
3193          * parent for all of our realms.  If the mds has any newer info,
3194          * it will tell us.
3195          */
3196         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3197                 struct ceph_snap_realm *realm =
3198                         rb_entry(p, struct ceph_snap_realm, node);
3199                 struct ceph_mds_snaprealm_reconnect sr_rec;
3200
3201                 dout(" adding snap realm %llx seq %lld parent %llx\n",
3202                      realm->ino, realm->seq, realm->parent_ino);
3203                 sr_rec.ino = cpu_to_le64(realm->ino);
3204                 sr_rec.seq = cpu_to_le64(realm->seq);
3205                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3206                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3207                 if (err)
3208                         goto fail;
3209         }
3210
3211         reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3212
3213         /* raced with cap release? */
3214         if (s_nr_caps != recon_state.nr_caps) {
3215                 struct page *page = list_first_entry(&pagelist->head,
3216                                                      struct page, lru);
3217                 __le32 *addr = kmap_atomic(page);
3218                 *addr = cpu_to_le32(recon_state.nr_caps);
3219                 kunmap_atomic(addr);
3220         }
3221
3222         reply->hdr.data_len = cpu_to_le32(pagelist->length);
3223         ceph_msg_data_add_pagelist(reply, pagelist);
3224
3225         ceph_early_kick_flushing_caps(mdsc, session);
3226
3227         ceph_con_send(&session->s_con, reply);
3228
3229         mutex_unlock(&session->s_mutex);
3230
3231         mutex_lock(&mdsc->mutex);
3232         __wake_requests(mdsc, &session->s_waiting);
3233         mutex_unlock(&mdsc->mutex);
3234
3235         up_read(&mdsc->snap_rwsem);
3236         ceph_pagelist_release(pagelist);
3237         return;
3238
3239 fail:
3240         ceph_msg_put(reply);
3241         up_read(&mdsc->snap_rwsem);
3242         mutex_unlock(&session->s_mutex);
3243 fail_nomsg:
3244         ceph_pagelist_release(pagelist);
3245 fail_nopagelist:
3246         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3247         return;
3248 }
3249
3250
3251 /*
3252  * compare old and new mdsmaps, kicking requests
3253  * and closing out old connections as necessary
3254  *
3255  * called under mdsc->mutex.
3256  */
3257 static void check_new_map(struct ceph_mds_client *mdsc,
3258                           struct ceph_mdsmap *newmap,
3259                           struct ceph_mdsmap *oldmap)
3260 {
3261         int i;
3262         int oldstate, newstate;
3263         struct ceph_mds_session *s;
3264
3265         dout("check_new_map new %u old %u\n",
3266              newmap->m_epoch, oldmap->m_epoch);
3267
3268         for (i = 0; i < oldmap->m_num_mds && i < mdsc->max_sessions; i++) {
3269                 if (!mdsc->sessions[i])
3270                         continue;
3271                 s = mdsc->sessions[i];
3272                 oldstate = ceph_mdsmap_get_state(oldmap, i);
3273                 newstate = ceph_mdsmap_get_state(newmap, i);
3274
3275                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3276                      i, ceph_mds_state_name(oldstate),
3277                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3278                      ceph_mds_state_name(newstate),
3279                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3280                      ceph_session_state_name(s->s_state));
3281
3282                 if (i >= newmap->m_num_mds ||
3283                     memcmp(ceph_mdsmap_get_addr(oldmap, i),
3284                            ceph_mdsmap_get_addr(newmap, i),
3285                            sizeof(struct ceph_entity_addr))) {
3286                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3287                                 /* the session never opened, just close it
3288                                  * out now */
3289                                 get_session(s);
3290                                 __unregister_session(mdsc, s);
3291                                 __wake_requests(mdsc, &s->s_waiting);
3292                                 ceph_put_mds_session(s);
3293                         } else if (i >= newmap->m_num_mds) {
3294                                 /* force close session for stopped mds */
3295                                 get_session(s);
3296                                 __unregister_session(mdsc, s);
3297                                 __wake_requests(mdsc, &s->s_waiting);
3298                                 kick_requests(mdsc, i);
3299                                 mutex_unlock(&mdsc->mutex);
3300
3301                                 mutex_lock(&s->s_mutex);
3302                                 cleanup_session_requests(mdsc, s);
3303                                 remove_session_caps(s);
3304                                 mutex_unlock(&s->s_mutex);
3305
3306                                 ceph_put_mds_session(s);
3307
3308                                 mutex_lock(&mdsc->mutex);
3309                         } else {
3310                                 /* just close it */
3311                                 mutex_unlock(&mdsc->mutex);
3312                                 mutex_lock(&s->s_mutex);
3313                                 mutex_lock(&mdsc->mutex);
3314                                 ceph_con_close(&s->s_con);
3315                                 mutex_unlock(&s->s_mutex);
3316                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3317                         }
3318                 } else if (oldstate == newstate) {
3319                         continue;  /* nothing new with this mds */
3320                 }
3321
3322                 /*
3323                  * send reconnect?
3324                  */
3325                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3326                     newstate >= CEPH_MDS_STATE_RECONNECT) {
3327                         mutex_unlock(&mdsc->mutex);
3328                         send_mds_reconnect(mdsc, s);
3329                         mutex_lock(&mdsc->mutex);
3330                 }
3331
3332                 /*
3333                  * kick request on any mds that has gone active.
3334                  */
3335                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3336                     newstate >= CEPH_MDS_STATE_ACTIVE) {
3337                         if (oldstate != CEPH_MDS_STATE_CREATING &&
3338                             oldstate != CEPH_MDS_STATE_STARTING)
3339                                 pr_info("mds%d recovery completed\n", s->s_mds);
3340                         kick_requests(mdsc, i);
3341                         ceph_kick_flushing_caps(mdsc, s);
3342                         wake_up_session_caps(s, 1);
3343                 }
3344         }
3345
3346         for (i = 0; i < newmap->m_num_mds && i < mdsc->max_sessions; i++) {
3347                 s = mdsc->sessions[i];
3348                 if (!s)
3349                         continue;
3350                 if (!ceph_mdsmap_is_laggy(newmap, i))
3351                         continue;
3352                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3353                     s->s_state == CEPH_MDS_SESSION_HUNG ||
3354                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
3355                         dout(" connecting to export targets of laggy mds%d\n",
3356                              i);
3357                         __open_export_target_sessions(mdsc, s);
3358                 }
3359         }
3360 }
3361
3362
3363
3364 /*
3365  * leases
3366  */
3367
3368 /*
3369  * caller must hold session s_mutex, dentry->d_lock
3370  */
3371 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3372 {
3373         struct ceph_dentry_info *di = ceph_dentry(dentry);
3374
3375         ceph_put_mds_session(di->lease_session);
3376         di->lease_session = NULL;
3377 }
3378
3379 static void handle_lease(struct ceph_mds_client *mdsc,
3380                          struct ceph_mds_session *session,
3381                          struct ceph_msg *msg)
3382 {
3383         struct super_block *sb = mdsc->fsc->sb;
3384         struct inode *inode;
3385         struct dentry *parent, *dentry;
3386         struct ceph_dentry_info *di;
3387         int mds = session->s_mds;
3388         struct ceph_mds_lease *h = msg->front.iov_base;
3389         u32 seq;
3390         struct ceph_vino vino;
3391         struct qstr dname;
3392         int release = 0;
3393
3394         dout("handle_lease from mds%d\n", mds);
3395
3396         /* decode */
3397         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3398                 goto bad;
3399         vino.ino = le64_to_cpu(h->ino);
3400         vino.snap = CEPH_NOSNAP;
3401         seq = le32_to_cpu(h->seq);
3402         dname.len = get_unaligned_le32(h + 1);
3403         if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
3404                 goto bad;
3405         dname.name = (void *)(h + 1) + sizeof(u32);
3406
3407         /* lookup inode */
3408         inode = ceph_find_inode(sb, vino);
3409         dout("handle_lease %s, ino %llx %p %.*s\n",
3410              ceph_lease_op_name(h->action), vino.ino, inode,
3411              dname.len, dname.name);
3412
3413         mutex_lock(&session->s_mutex);
3414         session->s_seq++;
3415
3416         if (!inode) {
3417                 dout("handle_lease no inode %llx\n", vino.ino);
3418                 goto release;
3419         }
3420
3421         /* dentry */
3422         parent = d_find_alias(inode);
3423         if (!parent) {
3424                 dout("no parent dentry on inode %p\n", inode);
3425                 WARN_ON(1);
3426                 goto release;  /* hrm... */
3427         }
3428         dname.hash = full_name_hash(parent, dname.name, dname.len);
3429         dentry = d_lookup(parent, &dname);
3430         dput(parent);
3431         if (!dentry)
3432                 goto release;
3433
3434         spin_lock(&dentry->d_lock);
3435         di = ceph_dentry(dentry);
3436         switch (h->action) {
3437         case CEPH_MDS_LEASE_REVOKE:
3438                 if (di->lease_session == session) {
3439                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3440                                 h->seq = cpu_to_le32(di->lease_seq);
3441                         __ceph_mdsc_drop_dentry_lease(dentry);
3442                 }
3443                 release = 1;
3444                 break;
3445
3446         case CEPH_MDS_LEASE_RENEW:
3447                 if (di->lease_session == session &&
3448                     di->lease_gen == session->s_cap_gen &&
3449                     di->lease_renew_from &&
3450                     di->lease_renew_after == 0) {
3451                         unsigned long duration =
3452                                 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3453
3454                         di->lease_seq = seq;
3455                         di->time = di->lease_renew_from + duration;
3456                         di->lease_renew_after = di->lease_renew_from +
3457                                 (duration >> 1);
3458                         di->lease_renew_from = 0;
3459                 }
3460                 break;
3461         }
3462         spin_unlock(&dentry->d_lock);
3463         dput(dentry);
3464
3465         if (!release)
3466                 goto out;
3467
3468 release:
3469         /* let's just reuse the same message */
3470         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3471         ceph_msg_get(msg);
3472         ceph_con_send(&session->s_con, msg);
3473
3474 out:
3475         iput(inode);
3476         mutex_unlock(&session->s_mutex);
3477         return;
3478
3479 bad:
3480         pr_err("corrupt lease message\n");
3481         ceph_msg_dump(msg);
3482 }
3483
3484 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3485                               struct inode *inode,
3486                               struct dentry *dentry, char action,
3487                               u32 seq)
3488 {
3489         struct ceph_msg *msg;
3490         struct ceph_mds_lease *lease;
3491         int len = sizeof(*lease) + sizeof(u32);
3492         int dnamelen = 0;
3493
3494         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3495              inode, dentry, ceph_lease_op_name(action), session->s_mds);
3496         dnamelen = dentry->d_name.len;
3497         len += dnamelen;
3498
3499         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3500         if (!msg)
3501                 return;
3502         lease = msg->front.iov_base;
3503         lease->action = action;
3504         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3505         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3506         lease->seq = cpu_to_le32(seq);
3507         put_unaligned_le32(dnamelen, lease + 1);
3508         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3509
3510         /*
3511          * if this is a preemptive lease RELEASE, no need to
3512          * flush request stream, since the actual request will
3513          * soon follow.
3514          */
3515         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3516
3517         ceph_con_send(&session->s_con, msg);
3518 }
3519
3520 /*
3521  * lock unlock sessions, to wait ongoing session activities
3522  */
3523 static void lock_unlock_sessions(struct ceph_mds_client *mdsc)
3524 {
3525         int i;
3526
3527         mutex_lock(&mdsc->mutex);
3528         for (i = 0; i < mdsc->max_sessions; i++) {
3529                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3530                 if (!s)
3531                         continue;
3532                 mutex_unlock(&mdsc->mutex);
3533                 mutex_lock(&s->s_mutex);
3534                 mutex_unlock(&s->s_mutex);
3535                 ceph_put_mds_session(s);
3536                 mutex_lock(&mdsc->mutex);
3537         }
3538         mutex_unlock(&mdsc->mutex);
3539 }
3540
3541
3542
3543 /*
3544  * delayed work -- periodically trim expired leases, renew caps with mds
3545  */
3546 static void schedule_delayed(struct ceph_mds_client *mdsc)
3547 {
3548         int delay = 5;
3549         unsigned hz = round_jiffies_relative(HZ * delay);
3550         schedule_delayed_work(&mdsc->delayed_work, hz);
3551 }
3552
3553 static void delayed_work(struct work_struct *work)
3554 {
3555         int i;
3556         struct ceph_mds_client *mdsc =
3557                 container_of(work, struct ceph_mds_client, delayed_work.work);
3558         int renew_interval;
3559         int renew_caps;
3560
3561         dout("mdsc delayed_work\n");
3562         ceph_check_delayed_caps(mdsc);
3563
3564         mutex_lock(&mdsc->mutex);
3565         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3566         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3567                                    mdsc->last_renew_caps);
3568         if (renew_caps)
3569                 mdsc->last_renew_caps = jiffies;
3570
3571         for (i = 0; i < mdsc->max_sessions; i++) {
3572                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3573                 if (!s)
3574                         continue;
3575                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3576                         dout("resending session close request for mds%d\n",
3577                              s->s_mds);
3578                         request_close_session(mdsc, s);
3579                         ceph_put_mds_session(s);
3580                         continue;
3581                 }
3582                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3583                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3584                                 s->s_state = CEPH_MDS_SESSION_HUNG;
3585                                 pr_info("mds%d hung\n", s->s_mds);
3586                         }
3587                 }
3588                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3589                         /* this mds is failed or recovering, just wait */
3590                         ceph_put_mds_session(s);
3591                         continue;
3592                 }
3593                 mutex_unlock(&mdsc->mutex);
3594
3595                 mutex_lock(&s->s_mutex);
3596                 if (renew_caps)
3597                         send_renew_caps(mdsc, s);
3598                 else
3599                         ceph_con_keepalive(&s->s_con);
3600                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3601                     s->s_state == CEPH_MDS_SESSION_HUNG)
3602                         ceph_send_cap_releases(mdsc, s);
3603                 mutex_unlock(&s->s_mutex);
3604                 ceph_put_mds_session(s);
3605
3606                 mutex_lock(&mdsc->mutex);
3607         }
3608         mutex_unlock(&mdsc->mutex);
3609
3610         schedule_delayed(mdsc);
3611 }
3612
3613 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3614
3615 {
3616         struct ceph_mds_client *mdsc;
3617
3618         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3619         if (!mdsc)
3620                 return -ENOMEM;
3621         mdsc->fsc = fsc;
3622         mutex_init(&mdsc->mutex);
3623         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3624         if (!mdsc->mdsmap) {
3625                 kfree(mdsc);
3626                 return -ENOMEM;
3627         }
3628
3629         fsc->mdsc = mdsc;
3630         init_completion(&mdsc->safe_umount_waiters);
3631         init_waitqueue_head(&mdsc->session_close_wq);
3632         INIT_LIST_HEAD(&mdsc->waiting_for_map);
3633         mdsc->sessions = NULL;
3634         atomic_set(&mdsc->num_sessions, 0);
3635         mdsc->max_sessions = 0;
3636         mdsc->stopping = 0;
3637         atomic64_set(&mdsc->quotarealms_count, 0);
3638         mdsc->last_snap_seq = 0;
3639         init_rwsem(&mdsc->snap_rwsem);
3640         mdsc->snap_realms = RB_ROOT;
3641         INIT_LIST_HEAD(&mdsc->snap_empty);
3642         spin_lock_init(&mdsc->snap_empty_lock);
3643         mdsc->last_tid = 0;
3644         mdsc->oldest_tid = 0;
3645         mdsc->request_tree = RB_ROOT;
3646         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3647         mdsc->last_renew_caps = jiffies;
3648         INIT_LIST_HEAD(&mdsc->cap_delay_list);
3649         spin_lock_init(&mdsc->cap_delay_lock);
3650         INIT_LIST_HEAD(&mdsc->snap_flush_list);
3651         spin_lock_init(&mdsc->snap_flush_lock);
3652         mdsc->last_cap_flush_tid = 1;
3653         INIT_LIST_HEAD(&mdsc->cap_flush_list);
3654         INIT_LIST_HEAD(&mdsc->cap_dirty);
3655         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3656         mdsc->num_cap_flushing = 0;
3657         spin_lock_init(&mdsc->cap_dirty_lock);
3658         init_waitqueue_head(&mdsc->cap_flushing_wq);
3659         spin_lock_init(&mdsc->dentry_lru_lock);
3660         INIT_LIST_HEAD(&mdsc->dentry_lru);
3661
3662         ceph_caps_init(mdsc);
3663         ceph_adjust_min_caps(mdsc, fsc->min_caps);
3664
3665         init_rwsem(&mdsc->pool_perm_rwsem);
3666         mdsc->pool_perm_tree = RB_ROOT;
3667
3668         strscpy(mdsc->nodename, utsname()->nodename,
3669                 sizeof(mdsc->nodename));
3670         return 0;
3671 }
3672
3673 /*
3674  * Wait for safe replies on open mds requests.  If we time out, drop
3675  * all requests from the tree to avoid dangling dentry refs.
3676  */
3677 static void wait_requests(struct ceph_mds_client *mdsc)
3678 {
3679         struct ceph_options *opts = mdsc->fsc->client->options;
3680         struct ceph_mds_request *req;
3681
3682         mutex_lock(&mdsc->mutex);
3683         if (__get_oldest_req(mdsc)) {
3684                 mutex_unlock(&mdsc->mutex);
3685
3686                 dout("wait_requests waiting for requests\n");
3687                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3688                                     ceph_timeout_jiffies(opts->mount_timeout));
3689
3690                 /* tear down remaining requests */
3691                 mutex_lock(&mdsc->mutex);
3692                 while ((req = __get_oldest_req(mdsc))) {
3693                         dout("wait_requests timed out on tid %llu\n",
3694                              req->r_tid);
3695                         __unregister_request(mdsc, req);
3696                 }
3697         }
3698         mutex_unlock(&mdsc->mutex);
3699         dout("wait_requests done\n");
3700 }
3701
3702 /*
3703  * called before mount is ro, and before dentries are torn down.
3704  * (hmm, does this still race with new lookups?)
3705  */
3706 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3707 {
3708         dout("pre_umount\n");
3709         mdsc->stopping = 1;
3710
3711         lock_unlock_sessions(mdsc);
3712         ceph_flush_dirty_caps(mdsc);
3713         wait_requests(mdsc);
3714
3715         /*
3716          * wait for reply handlers to drop their request refs and
3717          * their inode/dcache refs
3718          */
3719         ceph_msgr_flush();
3720 }
3721
3722 /*
3723  * wait for all write mds requests to flush.
3724  */
3725 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3726 {
3727         struct ceph_mds_request *req = NULL, *nextreq;
3728         struct rb_node *n;
3729
3730         mutex_lock(&mdsc->mutex);
3731         dout("wait_unsafe_requests want %lld\n", want_tid);
3732 restart:
3733         req = __get_oldest_req(mdsc);
3734         while (req && req->r_tid <= want_tid) {
3735                 /* find next request */
3736                 n = rb_next(&req->r_node);
3737                 if (n)
3738                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3739                 else
3740                         nextreq = NULL;
3741                 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3742                     (req->r_op & CEPH_MDS_OP_WRITE)) {
3743                         /* write op */
3744                         ceph_mdsc_get_request(req);
3745                         if (nextreq)
3746                                 ceph_mdsc_get_request(nextreq);
3747                         mutex_unlock(&mdsc->mutex);
3748                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
3749                              req->r_tid, want_tid);
3750                         wait_for_completion(&req->r_safe_completion);
3751                         mutex_lock(&mdsc->mutex);
3752                         ceph_mdsc_put_request(req);
3753                         if (!nextreq)
3754                                 break;  /* next dne before, so we're done! */
3755                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
3756                                 /* next request was removed from tree */
3757                                 ceph_mdsc_put_request(nextreq);
3758                                 goto restart;
3759                         }
3760                         ceph_mdsc_put_request(nextreq);  /* won't go away */
3761                 }
3762                 req = nextreq;
3763         }
3764         mutex_unlock(&mdsc->mutex);
3765         dout("wait_unsafe_requests done\n");
3766 }
3767
3768 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3769 {
3770         u64 want_tid, want_flush;
3771
3772         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3773                 return;
3774
3775         dout("sync\n");
3776         mutex_lock(&mdsc->mutex);
3777         want_tid = mdsc->last_tid;
3778         mutex_unlock(&mdsc->mutex);
3779
3780         ceph_flush_dirty_caps(mdsc);
3781         spin_lock(&mdsc->cap_dirty_lock);
3782         want_flush = mdsc->last_cap_flush_tid;
3783         if (!list_empty(&mdsc->cap_flush_list)) {
3784                 struct ceph_cap_flush *cf =
3785                         list_last_entry(&mdsc->cap_flush_list,
3786                                         struct ceph_cap_flush, g_list);
3787                 cf->wake = true;
3788         }
3789         spin_unlock(&mdsc->cap_dirty_lock);
3790
3791         dout("sync want tid %lld flush_seq %lld\n",
3792              want_tid, want_flush);
3793
3794         wait_unsafe_requests(mdsc, want_tid);
3795         wait_caps_flush(mdsc, want_flush);
3796 }
3797
3798 /*
3799  * true if all sessions are closed, or we force unmount
3800  */
3801 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
3802 {
3803         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3804                 return true;
3805         return atomic_read(&mdsc->num_sessions) <= skipped;
3806 }
3807
3808 /*
3809  * called after sb is ro.
3810  */
3811 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3812 {
3813         struct ceph_options *opts = mdsc->fsc->client->options;
3814         struct ceph_mds_session *session;
3815         int i;
3816         int skipped = 0;
3817
3818         dout("close_sessions\n");
3819
3820         /* close sessions */
3821         mutex_lock(&mdsc->mutex);
3822         for (i = 0; i < mdsc->max_sessions; i++) {
3823                 session = __ceph_lookup_mds_session(mdsc, i);
3824                 if (!session)
3825                         continue;
3826                 mutex_unlock(&mdsc->mutex);
3827                 mutex_lock(&session->s_mutex);
3828                 if (__close_session(mdsc, session) <= 0)
3829                         skipped++;
3830                 mutex_unlock(&session->s_mutex);
3831                 ceph_put_mds_session(session);
3832                 mutex_lock(&mdsc->mutex);
3833         }
3834         mutex_unlock(&mdsc->mutex);
3835
3836         dout("waiting for sessions to close\n");
3837         wait_event_timeout(mdsc->session_close_wq,
3838                            done_closing_sessions(mdsc, skipped),
3839                            ceph_timeout_jiffies(opts->mount_timeout));
3840
3841         /* tear down remaining sessions */
3842         mutex_lock(&mdsc->mutex);
3843         for (i = 0; i < mdsc->max_sessions; i++) {
3844                 if (mdsc->sessions[i]) {
3845                         session = get_session(mdsc->sessions[i]);
3846                         __unregister_session(mdsc, session);
3847                         mutex_unlock(&mdsc->mutex);
3848                         mutex_lock(&session->s_mutex);
3849                         remove_session_caps(session);
3850                         mutex_unlock(&session->s_mutex);
3851                         ceph_put_mds_session(session);
3852                         mutex_lock(&mdsc->mutex);
3853                 }
3854         }
3855         WARN_ON(!list_empty(&mdsc->cap_delay_list));
3856         mutex_unlock(&mdsc->mutex);
3857
3858         ceph_cleanup_empty_realms(mdsc);
3859
3860         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3861
3862         dout("stopped\n");
3863 }
3864
3865 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
3866 {
3867         struct ceph_mds_session *session;
3868         int mds;
3869
3870         dout("force umount\n");
3871
3872         mutex_lock(&mdsc->mutex);
3873         for (mds = 0; mds < mdsc->max_sessions; mds++) {
3874                 session = __ceph_lookup_mds_session(mdsc, mds);
3875                 if (!session)
3876                         continue;
3877                 mutex_unlock(&mdsc->mutex);
3878                 mutex_lock(&session->s_mutex);
3879                 __close_session(mdsc, session);
3880                 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
3881                         cleanup_session_requests(mdsc, session);
3882                         remove_session_caps(session);
3883                 }
3884                 mutex_unlock(&session->s_mutex);
3885                 ceph_put_mds_session(session);
3886                 mutex_lock(&mdsc->mutex);
3887                 kick_requests(mdsc, mds);
3888         }
3889         __wake_requests(mdsc, &mdsc->waiting_for_map);
3890         mutex_unlock(&mdsc->mutex);
3891 }
3892
3893 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3894 {
3895         dout("stop\n");
3896         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3897         if (mdsc->mdsmap)
3898                 ceph_mdsmap_destroy(mdsc->mdsmap);
3899         kfree(mdsc->sessions);
3900         ceph_caps_finalize(mdsc);
3901         ceph_pool_perm_destroy(mdsc);
3902 }
3903
3904 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3905 {
3906         struct ceph_mds_client *mdsc = fsc->mdsc;
3907         dout("mdsc_destroy %p\n", mdsc);
3908
3909         if (!mdsc)
3910                 return;
3911
3912         /* flush out any connection work with references to us */
3913         ceph_msgr_flush();
3914
3915         ceph_mdsc_stop(mdsc);
3916
3917         fsc->mdsc = NULL;
3918         kfree(mdsc);
3919         dout("mdsc_destroy %p done\n", mdsc);
3920 }
3921
3922 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3923 {
3924         struct ceph_fs_client *fsc = mdsc->fsc;
3925         const char *mds_namespace = fsc->mount_options->mds_namespace;
3926         void *p = msg->front.iov_base;
3927         void *end = p + msg->front.iov_len;
3928         u32 epoch;
3929         u32 map_len;
3930         u32 num_fs;
3931         u32 mount_fscid = (u32)-1;
3932         u8 struct_v, struct_cv;
3933         int err = -EINVAL;
3934
3935         ceph_decode_need(&p, end, sizeof(u32), bad);
3936         epoch = ceph_decode_32(&p);
3937
3938         dout("handle_fsmap epoch %u\n", epoch);
3939
3940         ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3941         struct_v = ceph_decode_8(&p);
3942         struct_cv = ceph_decode_8(&p);
3943         map_len = ceph_decode_32(&p);
3944
3945         ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
3946         p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
3947
3948         num_fs = ceph_decode_32(&p);
3949         while (num_fs-- > 0) {
3950                 void *info_p, *info_end;
3951                 u32 info_len;
3952                 u8 info_v, info_cv;
3953                 u32 fscid, namelen;
3954
3955                 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3956                 info_v = ceph_decode_8(&p);
3957                 info_cv = ceph_decode_8(&p);
3958                 info_len = ceph_decode_32(&p);
3959                 ceph_decode_need(&p, end, info_len, bad);
3960                 info_p = p;
3961                 info_end = p + info_len;
3962                 p = info_end;
3963
3964                 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
3965                 fscid = ceph_decode_32(&info_p);
3966                 namelen = ceph_decode_32(&info_p);
3967                 ceph_decode_need(&info_p, info_end, namelen, bad);
3968
3969                 if (mds_namespace &&
3970                     strlen(mds_namespace) == namelen &&
3971                     !strncmp(mds_namespace, (char *)info_p, namelen)) {
3972                         mount_fscid = fscid;
3973                         break;
3974                 }
3975         }
3976
3977         ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
3978         if (mount_fscid != (u32)-1) {
3979                 fsc->client->monc.fs_cluster_id = mount_fscid;
3980                 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
3981                                    0, true);
3982                 ceph_monc_renew_subs(&fsc->client->monc);
3983         } else {
3984                 err = -ENOENT;
3985                 goto err_out;
3986         }
3987         return;
3988
3989 bad:
3990         pr_err("error decoding fsmap\n");
3991 err_out:
3992         mutex_lock(&mdsc->mutex);
3993         mdsc->mdsmap_err = err;
3994         __wake_requests(mdsc, &mdsc->waiting_for_map);
3995         mutex_unlock(&mdsc->mutex);
3996 }
3997
3998 /*
3999  * handle mds map update.
4000  */
4001 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4002 {
4003         u32 epoch;
4004         u32 maplen;
4005         void *p = msg->front.iov_base;
4006         void *end = p + msg->front.iov_len;
4007         struct ceph_mdsmap *newmap, *oldmap;
4008         struct ceph_fsid fsid;
4009         int err = -EINVAL;
4010
4011         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
4012         ceph_decode_copy(&p, &fsid, sizeof(fsid));
4013         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
4014                 return;
4015         epoch = ceph_decode_32(&p);
4016         maplen = ceph_decode_32(&p);
4017         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
4018
4019         /* do we need it? */
4020         mutex_lock(&mdsc->mutex);
4021         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
4022                 dout("handle_map epoch %u <= our %u\n",
4023                      epoch, mdsc->mdsmap->m_epoch);
4024                 mutex_unlock(&mdsc->mutex);
4025                 return;
4026         }
4027
4028         newmap = ceph_mdsmap_decode(&p, end);
4029         if (IS_ERR(newmap)) {
4030                 err = PTR_ERR(newmap);
4031                 goto bad_unlock;
4032         }
4033
4034         /* swap into place */
4035         if (mdsc->mdsmap) {
4036                 oldmap = mdsc->mdsmap;
4037                 mdsc->mdsmap = newmap;
4038                 check_new_map(mdsc, newmap, oldmap);
4039                 ceph_mdsmap_destroy(oldmap);
4040         } else {
4041                 mdsc->mdsmap = newmap;  /* first mds map */
4042         }
4043         mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
4044                                         MAX_LFS_FILESIZE);
4045
4046         __wake_requests(mdsc, &mdsc->waiting_for_map);
4047         ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
4048                           mdsc->mdsmap->m_epoch);
4049
4050         mutex_unlock(&mdsc->mutex);
4051         schedule_delayed(mdsc);
4052         return;
4053
4054 bad_unlock:
4055         mutex_unlock(&mdsc->mutex);
4056 bad:
4057         pr_err("error decoding mdsmap %d\n", err);
4058         return;
4059 }
4060
4061 static struct ceph_connection *con_get(struct ceph_connection *con)
4062 {
4063         struct ceph_mds_session *s = con->private;
4064
4065         if (get_session(s)) {
4066                 dout("mdsc con_get %p ok (%d)\n", s, refcount_read(&s->s_ref));
4067                 return con;
4068         }
4069         dout("mdsc con_get %p FAIL\n", s);
4070         return NULL;
4071 }
4072
4073 static void con_put(struct ceph_connection *con)
4074 {
4075         struct ceph_mds_session *s = con->private;
4076
4077         dout("mdsc con_put %p (%d)\n", s, refcount_read(&s->s_ref) - 1);
4078         ceph_put_mds_session(s);
4079 }
4080
4081 /*
4082  * if the client is unresponsive for long enough, the mds will kill
4083  * the session entirely.
4084  */
4085 static void peer_reset(struct ceph_connection *con)
4086 {
4087         struct ceph_mds_session *s = con->private;
4088         struct ceph_mds_client *mdsc = s->s_mdsc;
4089
4090         pr_warn("mds%d closed our session\n", s->s_mds);
4091         send_mds_reconnect(mdsc, s);
4092 }
4093
4094 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
4095 {
4096         struct ceph_mds_session *s = con->private;
4097         struct ceph_mds_client *mdsc = s->s_mdsc;
4098         int type = le16_to_cpu(msg->hdr.type);
4099
4100         mutex_lock(&mdsc->mutex);
4101         if (__verify_registered_session(mdsc, s) < 0) {
4102                 mutex_unlock(&mdsc->mutex);
4103                 goto out;
4104         }
4105         mutex_unlock(&mdsc->mutex);
4106
4107         switch (type) {
4108         case CEPH_MSG_MDS_MAP:
4109                 ceph_mdsc_handle_mdsmap(mdsc, msg);
4110                 break;
4111         case CEPH_MSG_FS_MAP_USER:
4112                 ceph_mdsc_handle_fsmap(mdsc, msg);
4113                 break;
4114         case CEPH_MSG_CLIENT_SESSION:
4115                 handle_session(s, msg);
4116                 break;
4117         case CEPH_MSG_CLIENT_REPLY:
4118                 handle_reply(s, msg);
4119                 break;
4120         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
4121                 handle_forward(mdsc, s, msg);
4122                 break;
4123         case CEPH_MSG_CLIENT_CAPS:
4124                 ceph_handle_caps(s, msg);
4125                 break;
4126         case CEPH_MSG_CLIENT_SNAP:
4127                 ceph_handle_snap(mdsc, s, msg);
4128                 break;
4129         case CEPH_MSG_CLIENT_LEASE:
4130                 handle_lease(mdsc, s, msg);
4131                 break;
4132         case CEPH_MSG_CLIENT_QUOTA:
4133                 ceph_handle_quota(mdsc, s, msg);
4134                 break;
4135
4136         default:
4137                 pr_err("received unknown message type %d %s\n", type,
4138                        ceph_msg_type_name(type));
4139         }
4140 out:
4141         ceph_msg_put(msg);
4142 }
4143
4144 /*
4145  * authentication
4146  */
4147
4148 /*
4149  * Note: returned pointer is the address of a structure that's
4150  * managed separately.  Caller must *not* attempt to free it.
4151  */
4152 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
4153                                         int *proto, int force_new)
4154 {
4155         struct ceph_mds_session *s = con->private;
4156         struct ceph_mds_client *mdsc = s->s_mdsc;
4157         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4158         struct ceph_auth_handshake *auth = &s->s_auth;
4159
4160         if (force_new && auth->authorizer) {
4161                 ceph_auth_destroy_authorizer(auth->authorizer);
4162                 auth->authorizer = NULL;
4163         }
4164         if (!auth->authorizer) {
4165                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4166                                                       auth);
4167                 if (ret)
4168                         return ERR_PTR(ret);
4169         } else {
4170                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4171                                                       auth);
4172                 if (ret)
4173                         return ERR_PTR(ret);
4174         }
4175         *proto = ac->protocol;
4176
4177         return auth;
4178 }
4179
4180 static int add_authorizer_challenge(struct ceph_connection *con,
4181                                     void *challenge_buf, int challenge_buf_len)
4182 {
4183         struct ceph_mds_session *s = con->private;
4184         struct ceph_mds_client *mdsc = s->s_mdsc;
4185         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4186
4187         return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
4188                                             challenge_buf, challenge_buf_len);
4189 }
4190
4191 static int verify_authorizer_reply(struct ceph_connection *con)
4192 {
4193         struct ceph_mds_session *s = con->private;
4194         struct ceph_mds_client *mdsc = s->s_mdsc;
4195         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4196
4197         return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
4198 }
4199
4200 static int invalidate_authorizer(struct ceph_connection *con)
4201 {
4202         struct ceph_mds_session *s = con->private;
4203         struct ceph_mds_client *mdsc = s->s_mdsc;
4204         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4205
4206         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
4207
4208         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
4209 }
4210
4211 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
4212                                 struct ceph_msg_header *hdr, int *skip)
4213 {
4214         struct ceph_msg *msg;
4215         int type = (int) le16_to_cpu(hdr->type);
4216         int front_len = (int) le32_to_cpu(hdr->front_len);
4217
4218         if (con->in_msg)
4219                 return con->in_msg;
4220
4221         *skip = 0;
4222         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
4223         if (!msg) {
4224                 pr_err("unable to allocate msg type %d len %d\n",
4225                        type, front_len);
4226                 return NULL;
4227         }
4228
4229         return msg;
4230 }
4231
4232 static int mds_sign_message(struct ceph_msg *msg)
4233 {
4234        struct ceph_mds_session *s = msg->con->private;
4235        struct ceph_auth_handshake *auth = &s->s_auth;
4236
4237        return ceph_auth_sign_message(auth, msg);
4238 }
4239
4240 static int mds_check_message_signature(struct ceph_msg *msg)
4241 {
4242        struct ceph_mds_session *s = msg->con->private;
4243        struct ceph_auth_handshake *auth = &s->s_auth;
4244
4245        return ceph_auth_check_message_signature(auth, msg);
4246 }
4247
4248 static const struct ceph_connection_operations mds_con_ops = {
4249         .get = con_get,
4250         .put = con_put,
4251         .dispatch = dispatch,
4252         .get_authorizer = get_authorizer,
4253         .add_authorizer_challenge = add_authorizer_challenge,
4254         .verify_authorizer_reply = verify_authorizer_reply,
4255         .invalidate_authorizer = invalidate_authorizer,
4256         .peer_reset = peer_reset,
4257         .alloc_msg = mds_alloc_msg,
4258         .sign_message = mds_sign_message,
4259         .check_message_signature = mds_check_message_signature,
4260 };
4261
4262 /* eof */