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 #define RECONNECT_MAX_SIZE (INT_MAX - PAGE_SIZE)
  24
  25 /*
  26  * A cluster of MDS (metadata server) daemons is responsible for
  27  * managing the file system namespace (the directory hierarchy and
  28  * inodes) and for coordinating shared access to storage.  Metadata is
  29  * partitioning hierarchically across a number of servers, and that
  30  * partition varies over time as the cluster adjusts the distribution
  31  * in order to balance load.
  32  *
  33  * The MDS client is primarily responsible to managing synchronous
  34  * metadata requests for operations like open, unlink, and so forth.
  35  * If there is a MDS failure, we find out about it when we (possibly
  36  * request and) receive a new MDS map, and can resubmit affected
  37  * requests.
  38  *
  39  * For the most part, though, we take advantage of a lossless
  40  * communications channel to the MDS, and do not need to worry about
  41  * timing out or resubmitting requests.
  42  *
  43  * We maintain a stateful "session" with each MDS we interact with.
  44  * Within each session, we sent periodic heartbeat messages to ensure
  45  * any capabilities or leases we have been issues remain valid.  If
  46  * the session times out and goes stale, our leases and capabilities
  47  * are no longer valid.
  48  */
  49
  50 struct ceph_reconnect_state {
  51         struct ceph_mds_session *session;
  52         int nr_caps, nr_realms;
  53         struct ceph_pagelist *pagelist;
  54         unsigned msg_version;
  55         bool allow_multi;
  56 };
  57
  58 static void __wake_requests(struct ceph_mds_client *mdsc,
  59                             struct list_head *head);
  60 static void ceph_cap_release_work(struct work_struct *work);
  61 static void ceph_cap_reclaim_work(struct work_struct *work);
  62
  63 static const struct ceph_connection_operations mds_con_ops;
  64
  65
  66 /*
  67  * mds reply parsing
  68  */
  69
  70 static int parse_reply_info_quota(void **p, void *end,
  71                                   struct ceph_mds_reply_info_in *info)
  72 {
  73         u8 struct_v, struct_compat;
  74         u32 struct_len;
  75
  76         ceph_decode_8_safe(p, end, struct_v, bad);
  77         ceph_decode_8_safe(p, end, struct_compat, bad);
  78         /* struct_v is expected to be >= 1. we only
  79          * understand encoding with struct_compat == 1. */
  80         if (!struct_v || struct_compat != 1)
  81                 goto bad;
  82         ceph_decode_32_safe(p, end, struct_len, bad);
  83         ceph_decode_need(p, end, struct_len, bad);
  84         end = *p + struct_len;
  85         ceph_decode_64_safe(p, end, info->max_bytes, bad);
  86         ceph_decode_64_safe(p, end, info->max_files, bad);
  87         *p = end;
  88         return 0;
  89 bad:
  90         return -EIO;
  91 }
  92
  93 /*
  94  * parse individual inode info
  95  */
  96 static int parse_reply_info_in(void **p, void *end,
  97                                struct ceph_mds_reply_info_in *info,
  98                                u64 features)
  99 {
 100         int err = 0;
 101         u8 struct_v = 0;
 102
 103         if (features == (u64)-1) {
 104                 u32 struct_len;
 105                 u8 struct_compat;
 106                 ceph_decode_8_safe(p, end, struct_v, bad);
 107                 ceph_decode_8_safe(p, end, struct_compat, bad);
 108                 /* struct_v is expected to be >= 1. we only understand
 109                  * encoding with struct_compat == 1. */
 110                 if (!struct_v || struct_compat != 1)
 111                         goto bad;
 112                 ceph_decode_32_safe(p, end, struct_len, bad);
 113                 ceph_decode_need(p, end, struct_len, bad);
 114                 end = *p + struct_len;
 115         }
 116
 117         ceph_decode_need(p, end, sizeof(struct ceph_mds_reply_inode), bad);
 118         info->in = *p;
 119         *p += sizeof(struct ceph_mds_reply_inode) +
 120                 sizeof(*info->in->fragtree.splits) *
 121                 le32_to_cpu(info->in->fragtree.nsplits);
 122
 123         ceph_decode_32_safe(p, end, info->symlink_len, bad);
 124         ceph_decode_need(p, end, info->symlink_len, bad);
 125         info->symlink = *p;
 126         *p += info->symlink_len;
 127
 128         ceph_decode_copy_safe(p, end, &info->dir_layout,
 129                               sizeof(info->dir_layout), bad);
 130         ceph_decode_32_safe(p, end, info->xattr_len, bad);
 131         ceph_decode_need(p, end, info->xattr_len, bad);
 132         info->xattr_data = *p;
 133         *p += info->xattr_len;
 134
 135         if (features == (u64)-1) {
 136                 /* inline data */
 137                 ceph_decode_64_safe(p, end, info->inline_version, bad);
 138                 ceph_decode_32_safe(p, end, info->inline_len, bad);
 139                 ceph_decode_need(p, end, info->inline_len, bad);
 140                 info->inline_data = *p;
 141                 *p += info->inline_len;
 142                 /* quota */
 143                 err = parse_reply_info_quota(p, end, info);
 144                 if (err < 0)
 145                         goto out_bad;
 146                 /* pool namespace */
 147                 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
 148                 if (info->pool_ns_len > 0) {
 149                         ceph_decode_need(p, end, info->pool_ns_len, bad);
 150                         info->pool_ns_data = *p;
 151                         *p += info->pool_ns_len;
 152                 }
 153                 /* btime, change_attr */
 154                 {
 155                         struct ceph_timespec btime;
 156                         u64 change_attr;
 157                         ceph_decode_need(p, end, sizeof(btime), bad);
 158                         ceph_decode_copy(p, &btime, sizeof(btime));
 159                         ceph_decode_64_safe(p, end, change_attr, bad);
 160                 }
 161
 162                 /* dir pin */
 163                 if (struct_v >= 2) {
 164                         ceph_decode_32_safe(p, end, info->dir_pin, bad);
 165                 } else {
 166                         info->dir_pin = -ENODATA;
 167                 }
 168
 169                 *p = end;
 170         } else {
 171                 if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
 172                         ceph_decode_64_safe(p, end, info->inline_version, bad);
 173                         ceph_decode_32_safe(p, end, info->inline_len, bad);
 174                         ceph_decode_need(p, end, info->inline_len, bad);
 175                         info->inline_data = *p;
 176                         *p += info->inline_len;
 177                 } else
 178                         info->inline_version = CEPH_INLINE_NONE;
 179
 180                 if (features & CEPH_FEATURE_MDS_QUOTA) {
 181                         err = parse_reply_info_quota(p, end, info);
 182                         if (err < 0)
 183                                 goto out_bad;
 184                 } else {
 185                         info->max_bytes = 0;
 186                         info->max_files = 0;
 187                 }
 188
 189                 info->pool_ns_len = 0;
 190                 info->pool_ns_data = NULL;
 191                 if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
 192                         ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
 193                         if (info->pool_ns_len > 0) {
 194                                 ceph_decode_need(p, end, info->pool_ns_len, bad);
 195                                 info->pool_ns_data = *p;
 196                                 *p += info->pool_ns_len;
 197                         }
 198                 }
 199
 200                 info->dir_pin = -ENODATA;
 201         }
 202         return 0;
 203 bad:
 204         err = -EIO;
 205 out_bad:
 206         return err;
 207 }
 208
 209 static int parse_reply_info_dir(void **p, void *end,
 210                                 struct ceph_mds_reply_dirfrag **dirfrag,
 211                                 u64 features)
 212 {
 213         if (features == (u64)-1) {
 214                 u8 struct_v, struct_compat;
 215                 u32 struct_len;
 216                 ceph_decode_8_safe(p, end, struct_v, bad);
 217                 ceph_decode_8_safe(p, end, struct_compat, bad);
 218                 /* struct_v is expected to be >= 1. we only understand
 219                  * encoding whose struct_compat == 1. */
 220                 if (!struct_v || struct_compat != 1)
 221                         goto bad;
 222                 ceph_decode_32_safe(p, end, struct_len, bad);
 223                 ceph_decode_need(p, end, struct_len, bad);
 224                 end = *p + struct_len;
 225         }
 226
 227         ceph_decode_need(p, end, sizeof(**dirfrag), bad);
 228         *dirfrag = *p;
 229         *p += sizeof(**dirfrag) + sizeof(u32) * le32_to_cpu((*dirfrag)->ndist);
 230         if (unlikely(*p > end))
 231                 goto bad;
 232         if (features == (u64)-1)
 233                 *p = end;
 234         return 0;
 235 bad:
 236         return -EIO;
 237 }
 238
 239 static int parse_reply_info_lease(void **p, void *end,
 240                                   struct ceph_mds_reply_lease **lease,
 241                                   u64 features)
 242 {
 243         if (features == (u64)-1) {
 244                 u8 struct_v, struct_compat;
 245                 u32 struct_len;
 246                 ceph_decode_8_safe(p, end, struct_v, bad);
 247                 ceph_decode_8_safe(p, end, struct_compat, bad);
 248                 /* struct_v is expected to be >= 1. we only understand
 249                  * encoding whose struct_compat == 1. */
 250                 if (!struct_v || struct_compat != 1)
 251                         goto bad;
 252                 ceph_decode_32_safe(p, end, struct_len, bad);
 253                 ceph_decode_need(p, end, struct_len, bad);
 254                 end = *p + struct_len;
 255         }
 256
 257         ceph_decode_need(p, end, sizeof(**lease), bad);
 258         *lease = *p;
 259         *p += sizeof(**lease);
 260         if (features == (u64)-1)
 261                 *p = end;
 262         return 0;
 263 bad:
 264         return -EIO;
 265 }
 266
 267 /*
 268  * parse a normal reply, which may contain a (dir+)dentry and/or a
 269  * target inode.
 270  */
 271 static int parse_reply_info_trace(void **p, void *end,
 272                                   struct ceph_mds_reply_info_parsed *info,
 273                                   u64 features)
 274 {
 275         int err;
 276
 277         if (info->head->is_dentry) {
 278                 err = parse_reply_info_in(p, end, &info->diri, features);
 279                 if (err < 0)
 280                         goto out_bad;
 281
 282                 err = parse_reply_info_dir(p, end, &info->dirfrag, features);
 283                 if (err < 0)
 284                         goto out_bad;
 285
 286                 ceph_decode_32_safe(p, end, info->dname_len, bad);
 287                 ceph_decode_need(p, end, info->dname_len, bad);
 288                 info->dname = *p;
 289                 *p += info->dname_len;
 290
 291                 err = parse_reply_info_lease(p, end, &info->dlease, features);
 292                 if (err < 0)
 293                         goto out_bad;
 294         }
 295
 296         if (info->head->is_target) {
 297                 err = parse_reply_info_in(p, end, &info->targeti, features);
 298                 if (err < 0)
 299                         goto out_bad;
 300         }
 301
 302         if (unlikely(*p != end))
 303                 goto bad;
 304         return 0;
 305
 306 bad:
 307         err = -EIO;
 308 out_bad:
 309         pr_err("problem parsing mds trace %d\n", err);
 310         return err;
 311 }
 312
 313 /*
 314  * parse readdir results
 315  */
 316 static int parse_reply_info_readdir(void **p, void *end,
 317                                 struct ceph_mds_reply_info_parsed *info,
 318                                 u64 features)
 319 {
 320         u32 num, i = 0;
 321         int err;
 322
 323         err = parse_reply_info_dir(p, end, &info->dir_dir, features);
 324         if (err < 0)
 325                 goto out_bad;
 326
 327         ceph_decode_need(p, end, sizeof(num) + 2, bad);
 328         num = ceph_decode_32(p);
 329         {
 330                 u16 flags = ceph_decode_16(p);
 331                 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
 332                 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
 333                 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
 334                 info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
 335         }
 336         if (num == 0)
 337                 goto done;
 338
 339         BUG_ON(!info->dir_entries);
 340         if ((unsigned long)(info->dir_entries + num) >
 341             (unsigned long)info->dir_entries + info->dir_buf_size) {
 342                 pr_err("dir contents are larger than expected\n");
 343                 WARN_ON(1);
 344                 goto bad;
 345         }
 346
 347         info->dir_nr = num;
 348         while (num) {
 349                 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
 350                 /* dentry */
 351                 ceph_decode_32_safe(p, end, rde->name_len, bad);
 352                 ceph_decode_need(p, end, rde->name_len, bad);
 353                 rde->name = *p;
 354                 *p += rde->name_len;
 355                 dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
 356
 357                 /* dentry lease */
 358                 err = parse_reply_info_lease(p, end, &rde->lease, features);
 359                 if (err)
 360                         goto out_bad;
 361                 /* inode */
 362                 err = parse_reply_info_in(p, end, &rde->inode, features);
 363                 if (err < 0)
 364                         goto out_bad;
 365                 /* ceph_readdir_prepopulate() will update it */
 366                 rde->offset = 0;
 367                 i++;
 368                 num--;
 369         }
 370
 371 done:
 372         if (*p != end)
 373                 goto bad;
 374         return 0;
 375
 376 bad:
 377         err = -EIO;
 378 out_bad:
 379         pr_err("problem parsing dir contents %d\n", err);
 380         return err;
 381 }
 382
 383 /*
 384  * parse fcntl F_GETLK results
 385  */
 386 static int parse_reply_info_filelock(void **p, void *end,
 387                                      struct ceph_mds_reply_info_parsed *info,
 388                                      u64 features)
 389 {
 390         if (*p + sizeof(*info->filelock_reply) > end)
 391                 goto bad;
 392
 393         info->filelock_reply = *p;
 394         *p += sizeof(*info->filelock_reply);
 395
 396         if (unlikely(*p != end))
 397                 goto bad;
 398         return 0;
 399
 400 bad:
 401         return -EIO;
 402 }
 403
 404 /*
 405  * parse create results
 406  */
 407 static int parse_reply_info_create(void **p, void *end,
 408                                   struct ceph_mds_reply_info_parsed *info,
 409                                   u64 features)
 410 {
 411         if (features == (u64)-1 ||
 412             (features & CEPH_FEATURE_REPLY_CREATE_INODE)) {
 413                 if (*p == end) {
 414                         info->has_create_ino = false;
 415                 } else {
 416                         info->has_create_ino = true;
 417                         info->ino = ceph_decode_64(p);
 418                 }
 419         }
 420
 421         if (unlikely(*p != end))
 422                 goto bad;
 423         return 0;
 424
 425 bad:
 426         return -EIO;
 427 }
 428
 429 /*
 430  * parse extra results
 431  */
 432 static int parse_reply_info_extra(void **p, void *end,
 433                                   struct ceph_mds_reply_info_parsed *info,
 434                                   u64 features)
 435 {
 436         u32 op = le32_to_cpu(info->head->op);
 437
 438         if (op == CEPH_MDS_OP_GETFILELOCK)
 439                 return parse_reply_info_filelock(p, end, info, features);
 440         else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
 441                 return parse_reply_info_readdir(p, end, info, features);
 442         else if (op == CEPH_MDS_OP_CREATE)
 443                 return parse_reply_info_create(p, end, info, features);
 444         else
 445                 return -EIO;
 446 }
 447
 448 /*
 449  * parse entire mds reply
 450  */
 451 static int parse_reply_info(struct ceph_msg *msg,
 452                             struct ceph_mds_reply_info_parsed *info,
 453                             u64 features)
 454 {
 455         void *p, *end;
 456         u32 len;
 457         int err;
 458
 459         info->head = msg->front.iov_base;
 460         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 461         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 462
 463         /* trace */
 464         ceph_decode_32_safe(&p, end, len, bad);
 465         if (len > 0) {
 466                 ceph_decode_need(&p, end, len, bad);
 467                 err = parse_reply_info_trace(&p, p+len, info, features);
 468                 if (err < 0)
 469                         goto out_bad;
 470         }
 471
 472         /* extra */
 473         ceph_decode_32_safe(&p, end, len, bad);
 474         if (len > 0) {
 475                 ceph_decode_need(&p, end, len, bad);
 476                 err = parse_reply_info_extra(&p, p+len, info, features);
 477                 if (err < 0)
 478                         goto out_bad;
 479         }
 480
 481         /* snap blob */
 482         ceph_decode_32_safe(&p, end, len, bad);
 483         info->snapblob_len = len;
 484         info->snapblob = p;
 485         p += len;
 486
 487         if (p != end)
 488                 goto bad;
 489         return 0;
 490
 491 bad:
 492         err = -EIO;
 493 out_bad:
 494         pr_err("mds parse_reply err %d\n", err);
 495         return err;
 496 }
 497
 498 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
 499 {
 500         if (!info->dir_entries)
 501                 return;
 502         free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
 503 }
 504
 505
 506 /*
 507  * sessions
 508  */
 509 const char *ceph_session_state_name(int s)
 510 {
 511         switch (s) {
 512         case CEPH_MDS_SESSION_NEW: return "new";
 513         case CEPH_MDS_SESSION_OPENING: return "opening";
 514         case CEPH_MDS_SESSION_OPEN: return "open";
 515         case CEPH_MDS_SESSION_HUNG: return "hung";
 516         case CEPH_MDS_SESSION_CLOSING: return "closing";
 517         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
 518         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
 519         case CEPH_MDS_SESSION_REJECTED: return "rejected";
 520         default: return "???";
 521         }
 522 }
 523
 524 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
 525 {
 526         if (refcount_inc_not_zero(&s->s_ref)) {
 527                 dout("mdsc get_session %p %d -> %d\n", s,
 528                      refcount_read(&s->s_ref)-1, refcount_read(&s->s_ref));
 529                 return s;
 530         } else {
 531                 dout("mdsc get_session %p 0 -- FAIL\n", s);
 532                 return NULL;
 533         }
 534 }
 535
 536 void ceph_put_mds_session(struct ceph_mds_session *s)
 537 {
 538         dout("mdsc put_session %p %d -> %d\n", s,
 539              refcount_read(&s->s_ref), refcount_read(&s->s_ref)-1);
 540         if (refcount_dec_and_test(&s->s_ref)) {
 541                 if (s->s_auth.authorizer)
 542                         ceph_auth_destroy_authorizer(s->s_auth.authorizer);
 543                 kfree(s);
 544         }
 545 }
 546
 547 /*
 548  * called under mdsc->mutex
 549  */
 550 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
 551                                                    int mds)
 552 {
 553         if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
 554                 return NULL;
 555         return get_session(mdsc->sessions[mds]);
 556 }
 557
 558 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
 559 {
 560         if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
 561                 return false;
 562         else
 563                 return true;
 564 }
 565
 566 static int __verify_registered_session(struct ceph_mds_client *mdsc,
 567                                        struct ceph_mds_session *s)
 568 {
 569         if (s->s_mds >= mdsc->max_sessions ||
 570             mdsc->sessions[s->s_mds] != s)
 571                 return -ENOENT;
 572         return 0;
 573 }
 574
 575 /*
 576  * create+register a new session for given mds.
 577  * called under mdsc->mutex.
 578  */
 579 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 580                                                  int mds)
 581 {
 582         struct ceph_mds_session *s;
 583
 584         if (mds >= mdsc->mdsmap->m_num_mds)
 585                 return ERR_PTR(-EINVAL);
 586
 587         s = kzalloc(sizeof(*s), GFP_NOFS);
 588         if (!s)
 589                 return ERR_PTR(-ENOMEM);
 590
 591         if (mds >= mdsc->max_sessions) {
 592                 int newmax = 1 << get_count_order(mds + 1);
 593                 struct ceph_mds_session **sa;
 594
 595                 dout("%s: realloc to %d\n", __func__, newmax);
 596                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 597                 if (!sa)
 598                         goto fail_realloc;
 599                 if (mdsc->sessions) {
 600                         memcpy(sa, mdsc->sessions,
 601                                mdsc->max_sessions * sizeof(void *));
 602                         kfree(mdsc->sessions);
 603                 }
 604                 mdsc->sessions = sa;
 605                 mdsc->max_sessions = newmax;
 606         }
 607
 608         dout("%s: mds%d\n", __func__, mds);
 609         s->s_mdsc = mdsc;
 610         s->s_mds = mds;
 611         s->s_state = CEPH_MDS_SESSION_NEW;
 612         s->s_ttl = 0;
 613         s->s_seq = 0;
 614         mutex_init(&s->s_mutex);
 615
 616         ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
 617
 618         spin_lock_init(&s->s_gen_ttl_lock);
 619         s->s_cap_gen = 1;
 620         s->s_cap_ttl = jiffies - 1;
 621
 622         spin_lock_init(&s->s_cap_lock);
 623         s->s_renew_requested = 0;
 624         s->s_renew_seq = 0;
 625         INIT_LIST_HEAD(&s->s_caps);
 626         s->s_nr_caps = 0;
 627         s->s_trim_caps = 0;
 628         refcount_set(&s->s_ref, 1);
 629         INIT_LIST_HEAD(&s->s_waiting);
 630         INIT_LIST_HEAD(&s->s_unsafe);
 631         s->s_num_cap_releases = 0;
 632         s->s_cap_reconnect = 0;
 633         s->s_cap_iterator = NULL;
 634         INIT_LIST_HEAD(&s->s_cap_releases);
 635         INIT_WORK(&s->s_cap_release_work, ceph_cap_release_work);
 636
 637         INIT_LIST_HEAD(&s->s_cap_flushing);
 638
 639         mdsc->sessions[mds] = s;
 640         atomic_inc(&mdsc->num_sessions);
 641         refcount_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
 642
 643         ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
 644                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
 645
 646         return s;
 647
 648 fail_realloc:
 649         kfree(s);
 650         return ERR_PTR(-ENOMEM);
 651 }
 652
 653 /*
 654  * called under mdsc->mutex
 655  */
 656 static void __unregister_session(struct ceph_mds_client *mdsc,
 657                                struct ceph_mds_session *s)
 658 {
 659         dout("__unregister_session mds%d %p\n", s->s_mds, s);
 660         BUG_ON(mdsc->sessions[s->s_mds] != s);
 661         mdsc->sessions[s->s_mds] = NULL;
 662         s->s_state = 0;
 663         ceph_con_close(&s->s_con);
 664         ceph_put_mds_session(s);
 665         atomic_dec(&mdsc->num_sessions);
 666 }
 667
 668 /*
 669  * drop session refs in request.
 670  *
 671  * should be last request ref, or hold mdsc->mutex
 672  */
 673 static void put_request_session(struct ceph_mds_request *req)
 674 {
 675         if (req->r_session) {
 676                 ceph_put_mds_session(req->r_session);
 677                 req->r_session = NULL;
 678         }
 679 }
 680
 681 void ceph_mdsc_release_request(struct kref *kref)
 682 {
 683         struct ceph_mds_request *req = container_of(kref,
 684                                                     struct ceph_mds_request,
 685                                                     r_kref);
 686         destroy_reply_info(&req->r_reply_info);
 687         if (req->r_request)
 688                 ceph_msg_put(req->r_request);
 689         if (req->r_reply)
 690                 ceph_msg_put(req->r_reply);
 691         if (req->r_inode) {
 692                 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
 693                 iput(req->r_inode);
 694         }
 695         if (req->r_parent)
 696                 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
 697         iput(req->r_target_inode);
 698         if (req->r_dentry)
 699                 dput(req->r_dentry);
 700         if (req->r_old_dentry)
 701                 dput(req->r_old_dentry);
 702         if (req->r_old_dentry_dir) {
 703                 /*
 704                  * track (and drop pins for) r_old_dentry_dir
 705                  * separately, since r_old_dentry's d_parent may have
 706                  * changed between the dir mutex being dropped and
 707                  * this request being freed.
 708                  */
 709                 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
 710                                   CEPH_CAP_PIN);
 711                 iput(req->r_old_dentry_dir);
 712         }
 713         kfree(req->r_path1);
 714         kfree(req->r_path2);
 715         if (req->r_pagelist)
 716                 ceph_pagelist_release(req->r_pagelist);
 717         put_request_session(req);
 718         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
 719         kfree(req);
 720 }
 721
 722 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
 723
 724 /*
 725  * lookup session, bump ref if found.
 726  *
 727  * called under mdsc->mutex.
 728  */
 729 static struct ceph_mds_request *
 730 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
 731 {
 732         struct ceph_mds_request *req;
 733
 734         req = lookup_request(&mdsc->request_tree, tid);
 735         if (req)
 736                 ceph_mdsc_get_request(req);
 737
 738         return req;
 739 }
 740
 741 /*
 742  * Register an in-flight request, and assign a tid.  Link to directory
 743  * are modifying (if any).
 744  *
 745  * Called under mdsc->mutex.
 746  */
 747 static void __register_request(struct ceph_mds_client *mdsc,
 748                                struct ceph_mds_request *req,
 749                                struct inode *dir)
 750 {
 751         int ret = 0;
 752
 753         req->r_tid = ++mdsc->last_tid;
 754         if (req->r_num_caps) {
 755                 ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
 756                                         req->r_num_caps);
 757                 if (ret < 0) {
 758                         pr_err("__register_request %p "
 759                                "failed to reserve caps: %d\n", req, ret);
 760                         /* set req->r_err to fail early from __do_request */
 761                         req->r_err = ret;
 762                         return;
 763                 }
 764         }
 765         dout("__register_request %p tid %lld\n", req, req->r_tid);
 766         ceph_mdsc_get_request(req);
 767         insert_request(&mdsc->request_tree, req);
 768
 769         req->r_uid = current_fsuid();
 770         req->r_gid = current_fsgid();
 771
 772         if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
 773                 mdsc->oldest_tid = req->r_tid;
 774
 775         if (dir) {
 776                 ihold(dir);
 777                 req->r_unsafe_dir = dir;
 778         }
 779 }
 780
 781 static void __unregister_request(struct ceph_mds_client *mdsc,
 782                                  struct ceph_mds_request *req)
 783 {
 784         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 785
 786         /* Never leave an unregistered request on an unsafe list! */
 787         list_del_init(&req->r_unsafe_item);
 788
 789         if (req->r_tid == mdsc->oldest_tid) {
 790                 struct rb_node *p = rb_next(&req->r_node);
 791                 mdsc->oldest_tid = 0;
 792                 while (p) {
 793                         struct ceph_mds_request *next_req =
 794                                 rb_entry(p, struct ceph_mds_request, r_node);
 795                         if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
 796                                 mdsc->oldest_tid = next_req->r_tid;
 797                                 break;
 798                         }
 799                         p = rb_next(p);
 800                 }
 801         }
 802
 803         erase_request(&mdsc->request_tree, req);
 804
 805         if (req->r_unsafe_dir  &&
 806             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
 807                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 808                 spin_lock(&ci->i_unsafe_lock);
 809                 list_del_init(&req->r_unsafe_dir_item);
 810                 spin_unlock(&ci->i_unsafe_lock);
 811         }
 812         if (req->r_target_inode &&
 813             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
 814                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
 815                 spin_lock(&ci->i_unsafe_lock);
 816                 list_del_init(&req->r_unsafe_target_item);
 817                 spin_unlock(&ci->i_unsafe_lock);
 818         }
 819
 820         if (req->r_unsafe_dir) {
 821                 iput(req->r_unsafe_dir);
 822                 req->r_unsafe_dir = NULL;
 823         }
 824
 825         complete_all(&req->r_safe_completion);
 826
 827         ceph_mdsc_put_request(req);
 828 }
 829
 830 /*
 831  * Walk back up the dentry tree until we hit a dentry representing a
 832  * non-snapshot inode. We do this using the rcu_read_lock (which must be held
 833  * when calling this) to ensure that the objects won't disappear while we're
 834  * working with them. Once we hit a candidate dentry, we attempt to take a
 835  * reference to it, and return that as the result.
 836  */
 837 static struct inode *get_nonsnap_parent(struct dentry *dentry)
 838 {
 839         struct inode *inode = NULL;
 840
 841         while (dentry && !IS_ROOT(dentry)) {
 842                 inode = d_inode_rcu(dentry);
 843                 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
 844                         break;
 845                 dentry = dentry->d_parent;
 846         }
 847         if (inode)
 848                 inode = igrab(inode);
 849         return inode;
 850 }
 851
 852 /*
 853  * Choose mds to send request to next.  If there is a hint set in the
 854  * request (e.g., due to a prior forward hint from the mds), use that.
 855  * Otherwise, consult frag tree and/or caps to identify the
 856  * appropriate mds.  If all else fails, choose randomly.
 857  *
 858  * Called under mdsc->mutex.
 859  */
 860 static int __choose_mds(struct ceph_mds_client *mdsc,
 861                         struct ceph_mds_request *req)
 862 {
 863         struct inode *inode;
 864         struct ceph_inode_info *ci;
 865         struct ceph_cap *cap;
 866         int mode = req->r_direct_mode;
 867         int mds = -1;
 868         u32 hash = req->r_direct_hash;
 869         bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
 870
 871         /*
 872          * is there a specific mds we should try?  ignore hint if we have
 873          * no session and the mds is not up (active or recovering).
 874          */
 875         if (req->r_resend_mds >= 0 &&
 876             (__have_session(mdsc, req->r_resend_mds) ||
 877              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 878                 dout("choose_mds using resend_mds mds%d\n",
 879                      req->r_resend_mds);
 880                 return req->r_resend_mds;
 881         }
 882
 883         if (mode == USE_RANDOM_MDS)
 884                 goto random;
 885
 886         inode = NULL;
 887         if (req->r_inode) {
 888                 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
 889                         inode = req->r_inode;
 890                         ihold(inode);
 891                 } else {
 892                         /* req->r_dentry is non-null for LSSNAP request */
 893                         rcu_read_lock();
 894                         inode = get_nonsnap_parent(req->r_dentry);
 895                         rcu_read_unlock();
 896                         dout("__choose_mds using snapdir's parent %p\n", inode);
 897                 }
 898         } else if (req->r_dentry) {
 899                 /* ignore race with rename; old or new d_parent is okay */
 900                 struct dentry *parent;
 901                 struct inode *dir;
 902
 903                 rcu_read_lock();
 904                 parent = req->r_dentry->d_parent;
 905                 dir = req->r_parent ? : d_inode_rcu(parent);
 906
 907                 if (!dir || dir->i_sb != mdsc->fsc->sb) {
 908                         /*  not this fs or parent went negative */
 909                         inode = d_inode(req->r_dentry);
 910                         if (inode)
 911                                 ihold(inode);
 912                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
 913                         /* direct snapped/virtual snapdir requests
 914                          * based on parent dir inode */
 915                         inode = get_nonsnap_parent(parent);
 916                         dout("__choose_mds using nonsnap parent %p\n", inode);
 917                 } else {
 918                         /* dentry target */
 919                         inode = d_inode(req->r_dentry);
 920                         if (!inode || mode == USE_AUTH_MDS) {
 921                                 /* dir + name */
 922                                 inode = igrab(dir);
 923                                 hash = ceph_dentry_hash(dir, req->r_dentry);
 924                                 is_hash = true;
 925                         } else {
 926                                 ihold(inode);
 927                         }
 928                 }
 929                 rcu_read_unlock();
 930         }
 931
 932         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 933              (int)hash, mode);
 934         if (!inode)
 935                 goto random;
 936         ci = ceph_inode(inode);
 937
 938         if (is_hash && S_ISDIR(inode->i_mode)) {
 939                 struct ceph_inode_frag frag;
 940                 int found;
 941
 942                 ceph_choose_frag(ci, hash, &frag, &found);
 943                 if (found) {
 944                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 945                                 u8 r;
 946
 947                                 /* choose a random replica */
 948                                 get_random_bytes(&r, 1);
 949                                 r %= frag.ndist;
 950                                 mds = frag.dist[r];
 951                                 dout("choose_mds %p %llx.%llx "
 952                                      "frag %u mds%d (%d/%d)\n",
 953                                      inode, ceph_vinop(inode),
 954                                      frag.frag, mds,
 955                                      (int)r, frag.ndist);
 956                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 957                                     CEPH_MDS_STATE_ACTIVE)
 958                                         goto out;
 959                         }
 960
 961                         /* since this file/dir wasn't known to be
 962                          * replicated, then we want to look for the
 963                          * authoritative mds. */
 964                         mode = USE_AUTH_MDS;
 965                         if (frag.mds >= 0) {
 966                                 /* choose auth mds */
 967                                 mds = frag.mds;
 968                                 dout("choose_mds %p %llx.%llx "
 969                                      "frag %u mds%d (auth)\n",
 970                                      inode, ceph_vinop(inode), frag.frag, mds);
 971                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 972                                     CEPH_MDS_STATE_ACTIVE)
 973                                         goto out;
 974                         }
 975                 }
 976         }
 977
 978         spin_lock(&ci->i_ceph_lock);
 979         cap = NULL;
 980         if (mode == USE_AUTH_MDS)
 981                 cap = ci->i_auth_cap;
 982         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
 983                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
 984         if (!cap) {
 985                 spin_unlock(&ci->i_ceph_lock);
 986                 iput(inode);
 987                 goto random;
 988         }
 989         mds = cap->session->s_mds;
 990         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
 991              inode, ceph_vinop(inode), mds,
 992              cap == ci->i_auth_cap ? "auth " : "", cap);
 993         spin_unlock(&ci->i_ceph_lock);
 994 out:
 995         iput(inode);
 996         return mds;
 997
 998 random:
 999         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
1000         dout("choose_mds chose random mds%d\n", mds);
1001         return mds;
1002 }
1003
1004
1005 /*
1006  * session messages
1007  */
1008 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
1009 {
1010         struct ceph_msg *msg;
1011         struct ceph_mds_session_head *h;
1012
1013         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
1014                            false);
1015         if (!msg) {
1016                 pr_err("create_session_msg ENOMEM creating msg\n");
1017                 return NULL;
1018         }
1019         h = msg->front.iov_base;
1020         h->op = cpu_to_le32(op);
1021         h->seq = cpu_to_le64(seq);
1022
1023         return msg;
1024 }
1025
1026 static void encode_supported_features(void **p, void *end)
1027 {
1028         static const unsigned char bits[] = CEPHFS_FEATURES_CLIENT_SUPPORTED;
1029         static const size_t count = ARRAY_SIZE(bits);
1030
1031         if (count > 0) {
1032                 size_t i;
1033                 size_t size = ((size_t)bits[count - 1] + 64) / 64 * 8;
1034
1035                 BUG_ON(*p + 4 + size > end);
1036                 ceph_encode_32(p, size);
1037                 memset(*p, 0, size);
1038                 for (i = 0; i < count; i++)
1039                         ((unsigned char*)(*p))[i / 8] |= 1 << (bits[i] % 8);
1040                 *p += size;
1041         } else {
1042                 BUG_ON(*p + 4 > end);
1043                 ceph_encode_32(p, 0);
1044         }
1045 }
1046
1047 /*
1048  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
1049  * to include additional client metadata fields.
1050  */
1051 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
1052 {
1053         struct ceph_msg *msg;
1054         struct ceph_mds_session_head *h;
1055         int i = -1;
1056         int extra_bytes = 0;
1057         int metadata_key_count = 0;
1058         struct ceph_options *opt = mdsc->fsc->client->options;
1059         struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
1060         void *p, *end;
1061
1062         const char* metadata[][2] = {
1063                 {"hostname", mdsc->nodename},
1064                 {"kernel_version", init_utsname()->release},
1065                 {"entity_id", opt->name ? : ""},
1066                 {"root", fsopt->server_path ? : "/"},
1067                 {NULL, NULL}
1068         };
1069
1070         /* Calculate serialized length of metadata */
1071         extra_bytes = 4;  /* map length */
1072         for (i = 0; metadata[i][0]; ++i) {
1073                 extra_bytes += 8 + strlen(metadata[i][0]) +
1074                         strlen(metadata[i][1]);
1075                 metadata_key_count++;
1076         }
1077         /* supported feature */
1078         extra_bytes += 4 + 8;
1079
1080         /* Allocate the message */
1081         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + extra_bytes,
1082                            GFP_NOFS, false);
1083         if (!msg) {
1084                 pr_err("create_session_msg ENOMEM creating msg\n");
1085                 return NULL;
1086         }
1087         p = msg->front.iov_base;
1088         end = p + msg->front.iov_len;
1089
1090         h = p;
1091         h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
1092         h->seq = cpu_to_le64(seq);
1093
1094         /*
1095          * Serialize client metadata into waiting buffer space, using
1096          * the format that userspace expects for map<string, string>
1097          *
1098          * ClientSession messages with metadata are v2
1099          */
1100         msg->hdr.version = cpu_to_le16(3);
1101         msg->hdr.compat_version = cpu_to_le16(1);
1102
1103         /* The write pointer, following the session_head structure */
1104         p += sizeof(*h);
1105
1106         /* Number of entries in the map */
1107         ceph_encode_32(&p, metadata_key_count);
1108
1109         /* Two length-prefixed strings for each entry in the map */
1110         for (i = 0; metadata[i][0]; ++i) {
1111                 size_t const key_len = strlen(metadata[i][0]);
1112                 size_t const val_len = strlen(metadata[i][1]);
1113
1114                 ceph_encode_32(&p, key_len);
1115                 memcpy(p, metadata[i][0], key_len);
1116                 p += key_len;
1117                 ceph_encode_32(&p, val_len);
1118                 memcpy(p, metadata[i][1], val_len);
1119                 p += val_len;
1120         }
1121
1122         encode_supported_features(&p, end);
1123         msg->front.iov_len = p - msg->front.iov_base;
1124         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1125
1126         return msg;
1127 }
1128
1129 /*
1130  * send session open request.
1131  *
1132  * called under mdsc->mutex
1133  */
1134 static int __open_session(struct ceph_mds_client *mdsc,
1135                           struct ceph_mds_session *session)
1136 {
1137         struct ceph_msg *msg;
1138         int mstate;
1139         int mds = session->s_mds;
1140
1141         /* wait for mds to go active? */
1142         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
1143         dout("open_session to mds%d (%s)\n", mds,
1144              ceph_mds_state_name(mstate));
1145         session->s_state = CEPH_MDS_SESSION_OPENING;
1146         session->s_renew_requested = jiffies;
1147
1148         /* send connect message */
1149         msg = create_session_open_msg(mdsc, session->s_seq);
1150         if (!msg)
1151                 return -ENOMEM;
1152         ceph_con_send(&session->s_con, msg);
1153         return 0;
1154 }
1155
1156 /*
1157  * open sessions for any export targets for the given mds
1158  *
1159  * called under mdsc->mutex
1160  */
1161 static struct ceph_mds_session *
1162 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
1163 {
1164         struct ceph_mds_session *session;
1165
1166         session = __ceph_lookup_mds_session(mdsc, target);
1167         if (!session) {
1168                 session = register_session(mdsc, target);
1169                 if (IS_ERR(session))
1170                         return session;
1171         }
1172         if (session->s_state == CEPH_MDS_SESSION_NEW ||
1173             session->s_state == CEPH_MDS_SESSION_CLOSING)
1174                 __open_session(mdsc, session);
1175
1176         return session;
1177 }
1178
1179 struct ceph_mds_session *
1180 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1181 {
1182         struct ceph_mds_session *session;
1183
1184         dout("open_export_target_session to mds%d\n", target);
1185
1186         mutex_lock(&mdsc->mutex);
1187         session = __open_export_target_session(mdsc, target);
1188         mutex_unlock(&mdsc->mutex);
1189
1190         return session;
1191 }
1192
1193 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1194                                           struct ceph_mds_session *session)
1195 {
1196         struct ceph_mds_info *mi;
1197         struct ceph_mds_session *ts;
1198         int i, mds = session->s_mds;
1199
1200         if (mds >= mdsc->mdsmap->m_num_mds)
1201                 return;
1202
1203         mi = &mdsc->mdsmap->m_info[mds];
1204         dout("open_export_target_sessions for mds%d (%d targets)\n",
1205              session->s_mds, mi->num_export_targets);
1206
1207         for (i = 0; i < mi->num_export_targets; i++) {
1208                 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1209                 if (!IS_ERR(ts))
1210                         ceph_put_mds_session(ts);
1211         }
1212 }
1213
1214 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1215                                            struct ceph_mds_session *session)
1216 {
1217         mutex_lock(&mdsc->mutex);
1218         __open_export_target_sessions(mdsc, session);
1219         mutex_unlock(&mdsc->mutex);
1220 }
1221
1222 /*
1223  * session caps
1224  */
1225
1226 static void detach_cap_releases(struct ceph_mds_session *session,
1227                                 struct list_head *target)
1228 {
1229         lockdep_assert_held(&session->s_cap_lock);
1230
1231         list_splice_init(&session->s_cap_releases, target);
1232         session->s_num_cap_releases = 0;
1233         dout("dispose_cap_releases mds%d\n", session->s_mds);
1234 }
1235
1236 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1237                                  struct list_head *dispose)
1238 {
1239         while (!list_empty(dispose)) {
1240                 struct ceph_cap *cap;
1241                 /* zero out the in-progress message */
1242                 cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1243                 list_del(&cap->session_caps);
1244                 ceph_put_cap(mdsc, cap);
1245         }
1246 }
1247
1248 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1249                                      struct ceph_mds_session *session)
1250 {
1251         struct ceph_mds_request *req;
1252         struct rb_node *p;
1253
1254         dout("cleanup_session_requests mds%d\n", session->s_mds);
1255         mutex_lock(&mdsc->mutex);
1256         while (!list_empty(&session->s_unsafe)) {
1257                 req = list_first_entry(&session->s_unsafe,
1258                                        struct ceph_mds_request, r_unsafe_item);
1259                 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1260                                     req->r_tid);
1261                 __unregister_request(mdsc, req);
1262         }
1263         /* zero r_attempts, so kick_requests() will re-send requests */
1264         p = rb_first(&mdsc->request_tree);
1265         while (p) {
1266                 req = rb_entry(p, struct ceph_mds_request, r_node);
1267                 p = rb_next(p);
1268                 if (req->r_session &&
1269                     req->r_session->s_mds == session->s_mds)
1270                         req->r_attempts = 0;
1271         }
1272         mutex_unlock(&mdsc->mutex);
1273 }
1274
1275 /*
1276  * Helper to safely iterate over all caps associated with a session, with
1277  * special care taken to handle a racing __ceph_remove_cap().
1278  *
1279  * Caller must hold session s_mutex.
1280  */
1281 int ceph_iterate_session_caps(struct ceph_mds_session *session,
1282                               int (*cb)(struct inode *, struct ceph_cap *,
1283                                         void *), void *arg)
1284 {
1285         struct list_head *p;
1286         struct ceph_cap *cap;
1287         struct inode *inode, *last_inode = NULL;
1288         struct ceph_cap *old_cap = NULL;
1289         int ret;
1290
1291         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1292         spin_lock(&session->s_cap_lock);
1293         p = session->s_caps.next;
1294         while (p != &session->s_caps) {
1295                 cap = list_entry(p, struct ceph_cap, session_caps);
1296                 inode = igrab(&cap->ci->vfs_inode);
1297                 if (!inode) {
1298                         p = p->next;
1299                         continue;
1300                 }
1301                 session->s_cap_iterator = cap;
1302                 spin_unlock(&session->s_cap_lock);
1303
1304                 if (last_inode) {
1305                         iput(last_inode);
1306                         last_inode = NULL;
1307                 }
1308                 if (old_cap) {
1309                         ceph_put_cap(session->s_mdsc, old_cap);
1310                         old_cap = NULL;
1311                 }
1312
1313                 ret = cb(inode, cap, arg);
1314                 last_inode = inode;
1315
1316                 spin_lock(&session->s_cap_lock);
1317                 p = p->next;
1318                 if (!cap->ci) {
1319                         dout("iterate_session_caps  finishing cap %p removal\n",
1320                              cap);
1321                         BUG_ON(cap->session != session);
1322                         cap->session = NULL;
1323                         list_del_init(&cap->session_caps);
1324                         session->s_nr_caps--;
1325                         if (cap->queue_release)
1326                                 __ceph_queue_cap_release(session, cap);
1327                         else
1328                                 old_cap = cap;  /* put_cap it w/o locks held */
1329                 }
1330                 if (ret < 0)
1331                         goto out;
1332         }
1333         ret = 0;
1334 out:
1335         session->s_cap_iterator = NULL;
1336         spin_unlock(&session->s_cap_lock);
1337
1338         iput(last_inode);
1339         if (old_cap)
1340                 ceph_put_cap(session->s_mdsc, old_cap);
1341
1342         return ret;
1343 }
1344
1345 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1346                                   void *arg)
1347 {
1348         struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1349         struct ceph_inode_info *ci = ceph_inode(inode);
1350         LIST_HEAD(to_remove);
1351         bool drop = false;
1352         bool invalidate = false;
1353
1354         dout("removing cap %p, ci is %p, inode is %p\n",
1355              cap, ci, &ci->vfs_inode);
1356         spin_lock(&ci->i_ceph_lock);
1357         if (cap->mds_wanted | cap->issued)
1358                 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1359         __ceph_remove_cap(cap, false);
1360         if (!ci->i_auth_cap) {
1361                 struct ceph_cap_flush *cf;
1362                 struct ceph_mds_client *mdsc = fsc->mdsc;
1363
1364                 if (ci->i_wrbuffer_ref > 0 &&
1365                     READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
1366                         invalidate = true;
1367
1368                 while (!list_empty(&ci->i_cap_flush_list)) {
1369                         cf = list_first_entry(&ci->i_cap_flush_list,
1370                                               struct ceph_cap_flush, i_list);
1371                         list_move(&cf->i_list, &to_remove);
1372                 }
1373
1374                 spin_lock(&mdsc->cap_dirty_lock);
1375
1376                 list_for_each_entry(cf, &to_remove, i_list)
1377                         list_del(&cf->g_list);
1378
1379                 if (!list_empty(&ci->i_dirty_item)) {
1380                         pr_warn_ratelimited(
1381                                 " dropping dirty %s state for %p %lld\n",
1382                                 ceph_cap_string(ci->i_dirty_caps),
1383                                 inode, ceph_ino(inode));
1384                         ci->i_dirty_caps = 0;
1385                         list_del_init(&ci->i_dirty_item);
1386                         drop = true;
1387                 }
1388                 if (!list_empty(&ci->i_flushing_item)) {
1389                         pr_warn_ratelimited(
1390                                 " dropping dirty+flushing %s state for %p %lld\n",
1391                                 ceph_cap_string(ci->i_flushing_caps),
1392                                 inode, ceph_ino(inode));
1393                         ci->i_flushing_caps = 0;
1394                         list_del_init(&ci->i_flushing_item);
1395                         mdsc->num_cap_flushing--;
1396                         drop = true;
1397                 }
1398                 spin_unlock(&mdsc->cap_dirty_lock);
1399
1400                 if (atomic_read(&ci->i_filelock_ref) > 0) {
1401                         /* make further file lock syscall return -EIO */
1402                         ci->i_ceph_flags |= CEPH_I_ERROR_FILELOCK;
1403                         pr_warn_ratelimited(" dropping file locks for %p %lld\n",
1404                                             inode, ceph_ino(inode));
1405                 }
1406
1407                 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1408                         list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1409                         ci->i_prealloc_cap_flush = NULL;
1410                 }
1411
1412                if (drop &&
1413                   ci->i_wrbuffer_ref_head == 0 &&
1414                   ci->i_wr_ref == 0 &&
1415                   ci->i_dirty_caps == 0 &&
1416                   ci->i_flushing_caps == 0) {
1417                       ceph_put_snap_context(ci->i_head_snapc);
1418                       ci->i_head_snapc = NULL;
1419                }
1420         }
1421         spin_unlock(&ci->i_ceph_lock);
1422         while (!list_empty(&to_remove)) {
1423                 struct ceph_cap_flush *cf;
1424                 cf = list_first_entry(&to_remove,
1425                                       struct ceph_cap_flush, i_list);
1426                 list_del(&cf->i_list);
1427                 ceph_free_cap_flush(cf);
1428         }
1429
1430         wake_up_all(&ci->i_cap_wq);
1431         if (invalidate)
1432                 ceph_queue_invalidate(inode);
1433         if (drop)
1434                 iput(inode);
1435         return 0;
1436 }
1437
1438 /*
1439  * caller must hold session s_mutex
1440  */
1441 static void remove_session_caps(struct ceph_mds_session *session)
1442 {
1443         struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1444         struct super_block *sb = fsc->sb;
1445         LIST_HEAD(dispose);
1446
1447         dout("remove_session_caps on %p\n", session);
1448         ceph_iterate_session_caps(session, remove_session_caps_cb, fsc);
1449
1450         wake_up_all(&fsc->mdsc->cap_flushing_wq);
1451
1452         spin_lock(&session->s_cap_lock);
1453         if (session->s_nr_caps > 0) {
1454                 struct inode *inode;
1455                 struct ceph_cap *cap, *prev = NULL;
1456                 struct ceph_vino vino;
1457                 /*
1458                  * iterate_session_caps() skips inodes that are being
1459                  * deleted, we need to wait until deletions are complete.
1460                  * __wait_on_freeing_inode() is designed for the job,
1461                  * but it is not exported, so use lookup inode function
1462                  * to access it.
1463                  */
1464                 while (!list_empty(&session->s_caps)) {
1465                         cap = list_entry(session->s_caps.next,
1466                                          struct ceph_cap, session_caps);
1467                         if (cap == prev)
1468                                 break;
1469                         prev = cap;
1470                         vino = cap->ci->i_vino;
1471                         spin_unlock(&session->s_cap_lock);
1472
1473                         inode = ceph_find_inode(sb, vino);
1474                         iput(inode);
1475
1476                         spin_lock(&session->s_cap_lock);
1477                 }
1478         }
1479
1480         // drop cap expires and unlock s_cap_lock
1481         detach_cap_releases(session, &dispose);
1482
1483         BUG_ON(session->s_nr_caps > 0);
1484         BUG_ON(!list_empty(&session->s_cap_flushing));
1485         spin_unlock(&session->s_cap_lock);
1486         dispose_cap_releases(session->s_mdsc, &dispose);
1487 }
1488
1489 enum {
1490         RECONNECT,
1491         RENEWCAPS,
1492         FORCE_RO,
1493 };
1494
1495 /*
1496  * wake up any threads waiting on this session's caps.  if the cap is
1497  * old (didn't get renewed on the client reconnect), remove it now.
1498  *
1499  * caller must hold s_mutex.
1500  */
1501 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1502                               void *arg)
1503 {
1504         struct ceph_inode_info *ci = ceph_inode(inode);
1505         unsigned long ev = (unsigned long)arg;
1506
1507         if (ev == RECONNECT) {
1508                 spin_lock(&ci->i_ceph_lock);
1509                 ci->i_wanted_max_size = 0;
1510                 ci->i_requested_max_size = 0;
1511                 spin_unlock(&ci->i_ceph_lock);
1512         } else if (ev == RENEWCAPS) {
1513                 if (cap->cap_gen < cap->session->s_cap_gen) {
1514                         /* mds did not re-issue stale cap */
1515                         spin_lock(&ci->i_ceph_lock);
1516                         cap->issued = cap->implemented = CEPH_CAP_PIN;
1517                         /* make sure mds knows what we want */
1518                         if (__ceph_caps_file_wanted(ci) & ~cap->mds_wanted)
1519                                 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1520                         spin_unlock(&ci->i_ceph_lock);
1521                 }
1522         } else if (ev == FORCE_RO) {
1523         }
1524         wake_up_all(&ci->i_cap_wq);
1525         return 0;
1526 }
1527
1528 static void wake_up_session_caps(struct ceph_mds_session *session, int ev)
1529 {
1530         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1531         ceph_iterate_session_caps(session, wake_up_session_cb,
1532                                   (void *)(unsigned long)ev);
1533 }
1534
1535 /*
1536  * Send periodic message to MDS renewing all currently held caps.  The
1537  * ack will reset the expiration for all caps from this session.
1538  *
1539  * caller holds s_mutex
1540  */
1541 static int send_renew_caps(struct ceph_mds_client *mdsc,
1542                            struct ceph_mds_session *session)
1543 {
1544         struct ceph_msg *msg;
1545         int state;
1546
1547         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1548             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1549                 pr_info("mds%d caps stale\n", session->s_mds);
1550         session->s_renew_requested = jiffies;
1551
1552         /* do not try to renew caps until a recovering mds has reconnected
1553          * with its clients. */
1554         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1555         if (state < CEPH_MDS_STATE_RECONNECT) {
1556                 dout("send_renew_caps ignoring mds%d (%s)\n",
1557                      session->s_mds, ceph_mds_state_name(state));
1558                 return 0;
1559         }
1560
1561         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1562                 ceph_mds_state_name(state));
1563         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1564                                  ++session->s_renew_seq);
1565         if (!msg)
1566                 return -ENOMEM;
1567         ceph_con_send(&session->s_con, msg);
1568         return 0;
1569 }
1570
1571 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1572                              struct ceph_mds_session *session, u64 seq)
1573 {
1574         struct ceph_msg *msg;
1575
1576         dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1577              session->s_mds, ceph_session_state_name(session->s_state), seq);
1578         msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1579         if (!msg)
1580                 return -ENOMEM;
1581         ceph_con_send(&session->s_con, msg);
1582         return 0;
1583 }
1584
1585
1586 /*
1587  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1588  *
1589  * Called under session->s_mutex
1590  */
1591 static void renewed_caps(struct ceph_mds_client *mdsc,
1592                          struct ceph_mds_session *session, int is_renew)
1593 {
1594         int was_stale;
1595         int wake = 0;
1596
1597         spin_lock(&session->s_cap_lock);
1598         was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1599
1600         session->s_cap_ttl = session->s_renew_requested +
1601                 mdsc->mdsmap->m_session_timeout*HZ;
1602
1603         if (was_stale) {
1604                 if (time_before(jiffies, session->s_cap_ttl)) {
1605                         pr_info("mds%d caps renewed\n", session->s_mds);
1606                         wake = 1;
1607                 } else {
1608                         pr_info("mds%d caps still stale\n", session->s_mds);
1609                 }
1610         }
1611         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1612              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1613              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1614         spin_unlock(&session->s_cap_lock);
1615
1616         if (wake)
1617                 wake_up_session_caps(session, RENEWCAPS);
1618 }
1619
1620 /*
1621  * send a session close request
1622  */
1623 static int request_close_session(struct ceph_mds_client *mdsc,
1624                                  struct ceph_mds_session *session)
1625 {
1626         struct ceph_msg *msg;
1627
1628         dout("request_close_session mds%d state %s seq %lld\n",
1629              session->s_mds, ceph_session_state_name(session->s_state),
1630              session->s_seq);
1631         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1632         if (!msg)
1633                 return -ENOMEM;
1634         ceph_con_send(&session->s_con, msg);
1635         return 1;
1636 }
1637
1638 /*
1639  * Called with s_mutex held.
1640  */
1641 static int __close_session(struct ceph_mds_client *mdsc,
1642                          struct ceph_mds_session *session)
1643 {
1644         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1645                 return 0;
1646         session->s_state = CEPH_MDS_SESSION_CLOSING;
1647         return request_close_session(mdsc, session);
1648 }
1649
1650 static bool drop_negative_children(struct dentry *dentry)
1651 {
1652         struct dentry *child;
1653         bool all_negative = true;
1654
1655         if (!d_is_dir(dentry))
1656                 goto out;
1657
1658         spin_lock(&dentry->d_lock);
1659         list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1660                 if (d_really_is_positive(child)) {
1661                         all_negative = false;
1662                         break;
1663                 }
1664         }
1665         spin_unlock(&dentry->d_lock);
1666
1667         if (all_negative)
1668                 shrink_dcache_parent(dentry);
1669 out:
1670         return all_negative;
1671 }
1672
1673 /*
1674  * Trim old(er) caps.
1675  *
1676  * Because we can't cache an inode without one or more caps, we do
1677  * this indirectly: if a cap is unused, we prune its aliases, at which
1678  * point the inode will hopefully get dropped to.
1679  *
1680  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1681  * memory pressure from the MDS, though, so it needn't be perfect.
1682  */
1683 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1684 {
1685         struct ceph_mds_session *session = arg;
1686         struct ceph_inode_info *ci = ceph_inode(inode);
1687         int used, wanted, oissued, mine;
1688
1689         if (session->s_trim_caps <= 0)
1690                 return -1;
1691
1692         spin_lock(&ci->i_ceph_lock);
1693         mine = cap->issued | cap->implemented;
1694         used = __ceph_caps_used(ci);
1695         wanted = __ceph_caps_file_wanted(ci);
1696         oissued = __ceph_caps_issued_other(ci, cap);
1697
1698         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1699              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1700              ceph_cap_string(used), ceph_cap_string(wanted));
1701         if (cap == ci->i_auth_cap) {
1702                 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1703                     !list_empty(&ci->i_cap_snaps))
1704                         goto out;
1705                 if ((used | wanted) & CEPH_CAP_ANY_WR)
1706                         goto out;
1707                 /* Note: it's possible that i_filelock_ref becomes non-zero
1708                  * after dropping auth caps. It doesn't hurt because reply
1709                  * of lock mds request will re-add auth caps. */
1710                 if (atomic_read(&ci->i_filelock_ref) > 0)
1711                         goto out;
1712         }
1713         /* The inode has cached pages, but it's no longer used.
1714          * we can safely drop it */
1715         if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1716             !(oissued & CEPH_CAP_FILE_CACHE)) {
1717           used = 0;
1718           oissued = 0;
1719         }
1720         if ((used | wanted) & ~oissued & mine)
1721                 goto out;   /* we need these caps */
1722
1723         if (oissued) {
1724                 /* we aren't the only cap.. just remove us */
1725                 __ceph_remove_cap(cap, true);
1726                 session->s_trim_caps--;
1727         } else {
1728                 struct dentry *dentry;
1729                 /* try dropping referring dentries */
1730                 spin_unlock(&ci->i_ceph_lock);
1731                 dentry = d_find_any_alias(inode);
1732                 if (dentry && drop_negative_children(dentry)) {
1733                         int count;
1734                         dput(dentry);
1735                         d_prune_aliases(inode);
1736                         count = atomic_read(&inode->i_count);
1737                         if (count == 1)
1738                                 session->s_trim_caps--;
1739                         dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1740                              inode, cap, count);
1741                 } else {
1742                         dput(dentry);
1743                 }
1744                 return 0;
1745         }
1746
1747 out:
1748         spin_unlock(&ci->i_ceph_lock);
1749         return 0;
1750 }
1751
1752 /*
1753  * Trim session cap count down to some max number.
1754  */
1755 int ceph_trim_caps(struct ceph_mds_client *mdsc,
1756                    struct ceph_mds_session *session,
1757                    int max_caps)
1758 {
1759         int trim_caps = session->s_nr_caps - max_caps;
1760
1761         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1762              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1763         if (trim_caps > 0) {
1764                 session->s_trim_caps = trim_caps;
1765                 ceph_iterate_session_caps(session, trim_caps_cb, session);
1766                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1767                      session->s_mds, session->s_nr_caps, max_caps,
1768                         trim_caps - session->s_trim_caps);
1769                 session->s_trim_caps = 0;
1770         }
1771
1772         ceph_flush_cap_releases(mdsc, session);
1773         return 0;
1774 }
1775
1776 static int check_caps_flush(struct ceph_mds_client *mdsc,
1777                             u64 want_flush_tid)
1778 {
1779         int ret = 1;
1780
1781         spin_lock(&mdsc->cap_dirty_lock);
1782         if (!list_empty(&mdsc->cap_flush_list)) {
1783                 struct ceph_cap_flush *cf =
1784                         list_first_entry(&mdsc->cap_flush_list,
1785                                          struct ceph_cap_flush, g_list);
1786                 if (cf->tid <= want_flush_tid) {
1787                         dout("check_caps_flush still flushing tid "
1788                              "%llu <= %llu\n", cf->tid, want_flush_tid);
1789                         ret = 0;
1790                 }
1791         }
1792         spin_unlock(&mdsc->cap_dirty_lock);
1793         return ret;
1794 }
1795
1796 /*
1797  * flush all dirty inode data to disk.
1798  *
1799  * returns true if we've flushed through want_flush_tid
1800  */
1801 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1802                             u64 want_flush_tid)
1803 {
1804         dout("check_caps_flush want %llu\n", want_flush_tid);
1805
1806         wait_event(mdsc->cap_flushing_wq,
1807                    check_caps_flush(mdsc, want_flush_tid));
1808
1809         dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1810 }
1811
1812 /*
1813  * called under s_mutex
1814  */
1815 static void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1816                                    struct ceph_mds_session *session)
1817 {
1818         struct ceph_msg *msg = NULL;
1819         struct ceph_mds_cap_release *head;
1820         struct ceph_mds_cap_item *item;
1821         struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
1822         struct ceph_cap *cap;
1823         LIST_HEAD(tmp_list);
1824         int num_cap_releases;
1825         __le32  barrier, *cap_barrier;
1826
1827         down_read(&osdc->lock);
1828         barrier = cpu_to_le32(osdc->epoch_barrier);
1829         up_read(&osdc->lock);
1830
1831         spin_lock(&session->s_cap_lock);
1832 again:
1833         list_splice_init(&session->s_cap_releases, &tmp_list);
1834         num_cap_releases = session->s_num_cap_releases;
1835         session->s_num_cap_releases = 0;
1836         spin_unlock(&session->s_cap_lock);
1837
1838         while (!list_empty(&tmp_list)) {
1839                 if (!msg) {
1840                         msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1841                                         PAGE_SIZE, GFP_NOFS, false);
1842                         if (!msg)
1843                                 goto out_err;
1844                         head = msg->front.iov_base;
1845                         head->num = cpu_to_le32(0);
1846                         msg->front.iov_len = sizeof(*head);
1847
1848                         msg->hdr.version = cpu_to_le16(2);
1849                         msg->hdr.compat_version = cpu_to_le16(1);
1850                 }
1851
1852                 cap = list_first_entry(&tmp_list, struct ceph_cap,
1853                                         session_caps);
1854                 list_del(&cap->session_caps);
1855                 num_cap_releases--;
1856
1857                 head = msg->front.iov_base;
1858                 put_unaligned_le32(get_unaligned_le32(&head->num) + 1,
1859                                    &head->num);
1860                 item = msg->front.iov_base + msg->front.iov_len;
1861                 item->ino = cpu_to_le64(cap->cap_ino);
1862                 item->cap_id = cpu_to_le64(cap->cap_id);
1863                 item->migrate_seq = cpu_to_le32(cap->mseq);
1864                 item->seq = cpu_to_le32(cap->issue_seq);
1865                 msg->front.iov_len += sizeof(*item);
1866
1867                 ceph_put_cap(mdsc, cap);
1868
1869                 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1870                         // Append cap_barrier field
1871                         cap_barrier = msg->front.iov_base + msg->front.iov_len;
1872                         *cap_barrier = barrier;
1873                         msg->front.iov_len += sizeof(*cap_barrier);
1874
1875                         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1876                         dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1877                         ceph_con_send(&session->s_con, msg);
1878                         msg = NULL;
1879                 }
1880         }
1881
1882         BUG_ON(num_cap_releases != 0);
1883
1884         spin_lock(&session->s_cap_lock);
1885         if (!list_empty(&session->s_cap_releases))
1886                 goto again;
1887         spin_unlock(&session->s_cap_lock);
1888
1889         if (msg) {
1890                 // Append cap_barrier field
1891                 cap_barrier = msg->front.iov_base + msg->front.iov_len;
1892                 *cap_barrier = barrier;
1893                 msg->front.iov_len += sizeof(*cap_barrier);
1894
1895                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1896                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1897                 ceph_con_send(&session->s_con, msg);
1898         }
1899         return;
1900 out_err:
1901         pr_err("send_cap_releases mds%d, failed to allocate message\n",
1902                 session->s_mds);
1903         spin_lock(&session->s_cap_lock);
1904         list_splice(&tmp_list, &session->s_cap_releases);
1905         session->s_num_cap_releases += num_cap_releases;
1906         spin_unlock(&session->s_cap_lock);
1907 }
1908
1909 static void ceph_cap_release_work(struct work_struct *work)
1910 {
1911         struct ceph_mds_session *session =
1912                 container_of(work, struct ceph_mds_session, s_cap_release_work);
1913
1914         mutex_lock(&session->s_mutex);
1915         if (session->s_state == CEPH_MDS_SESSION_OPEN ||
1916             session->s_state == CEPH_MDS_SESSION_HUNG)
1917                 ceph_send_cap_releases(session->s_mdsc, session);
1918         mutex_unlock(&session->s_mutex);
1919         ceph_put_mds_session(session);
1920 }
1921
1922 void ceph_flush_cap_releases(struct ceph_mds_client *mdsc,
1923                              struct ceph_mds_session *session)
1924 {
1925         if (mdsc->stopping)
1926                 return;
1927
1928         get_session(session);
1929         if (queue_work(mdsc->fsc->cap_wq,
1930                        &session->s_cap_release_work)) {
1931                 dout("cap release work queued\n");
1932         } else {
1933                 ceph_put_mds_session(session);
1934                 dout("failed to queue cap release work\n");
1935         }
1936 }
1937
1938 /*
1939  * caller holds session->s_cap_lock
1940  */
1941 void __ceph_queue_cap_release(struct ceph_mds_session *session,
1942                               struct ceph_cap *cap)
1943 {
1944         list_add_tail(&cap->session_caps, &session->s_cap_releases);
1945         session->s_num_cap_releases++;
1946
1947         if (!(session->s_num_cap_releases % CEPH_CAPS_PER_RELEASE))
1948                 ceph_flush_cap_releases(session->s_mdsc, session);
1949 }
1950
1951 static void ceph_cap_reclaim_work(struct work_struct *work)
1952 {
1953         struct ceph_mds_client *mdsc =
1954                 container_of(work, struct ceph_mds_client, cap_reclaim_work);
1955         int ret = ceph_trim_dentries(mdsc);
1956         if (ret == -EAGAIN)
1957                 ceph_queue_cap_reclaim_work(mdsc);
1958 }
1959
1960 void ceph_queue_cap_reclaim_work(struct ceph_mds_client *mdsc)
1961 {
1962         if (mdsc->stopping)
1963                 return;
1964
1965         if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_reclaim_work)) {
1966                 dout("caps reclaim work queued\n");
1967         } else {
1968                 dout("failed to queue caps release work\n");
1969         }
1970 }
1971
1972 void ceph_reclaim_caps_nr(struct ceph_mds_client *mdsc, int nr)
1973 {
1974         int val;
1975         if (!nr)
1976                 return;
1977         val = atomic_add_return(nr, &mdsc->cap_reclaim_pending);
1978         if (!(val % CEPH_CAPS_PER_RELEASE)) {
1979                 atomic_set(&mdsc->cap_reclaim_pending, 0);
1980                 ceph_queue_cap_reclaim_work(mdsc);
1981         }
1982 }
1983
1984 /*
1985  * requests
1986  */
1987
1988 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1989                                     struct inode *dir)
1990 {
1991         struct ceph_inode_info *ci = ceph_inode(dir);
1992         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1993         struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1994         size_t size = sizeof(struct ceph_mds_reply_dir_entry);
1995         int order, num_entries;
1996
1997         spin_lock(&ci->i_ceph_lock);
1998         num_entries = ci->i_files + ci->i_subdirs;
1999         spin_unlock(&ci->i_ceph_lock);
2000         num_entries = max(num_entries, 1);
2001         num_entries = min(num_entries, opt->max_readdir);
2002
2003         order = get_order(size * num_entries);
2004         while (order >= 0) {
2005                 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
2006                                                              __GFP_NOWARN,
2007                                                              order);
2008                 if (rinfo->dir_entries)
2009                         break;
2010                 order--;
2011         }
2012         if (!rinfo->dir_entries)
2013                 return -ENOMEM;
2014
2015         num_entries = (PAGE_SIZE << order) / size;
2016         num_entries = min(num_entries, opt->max_readdir);
2017
2018         rinfo->dir_buf_size = PAGE_SIZE << order;
2019         req->r_num_caps = num_entries + 1;
2020         req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
2021         req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
2022         return 0;
2023 }
2024
2025 /*
2026  * Create an mds request.
2027  */
2028 struct ceph_mds_request *
2029 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
2030 {
2031         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
2032         struct timespec64 ts;
2033
2034         if (!req)
2035                 return ERR_PTR(-ENOMEM);
2036
2037         mutex_init(&req->r_fill_mutex);
2038         req->r_mdsc = mdsc;
2039         req->r_started = jiffies;
2040         req->r_resend_mds = -1;
2041         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
2042         INIT_LIST_HEAD(&req->r_unsafe_target_item);
2043         req->r_fmode = -1;
2044         kref_init(&req->r_kref);
2045         RB_CLEAR_NODE(&req->r_node);
2046         INIT_LIST_HEAD(&req->r_wait);
2047         init_completion(&req->r_completion);
2048         init_completion(&req->r_safe_completion);
2049         INIT_LIST_HEAD(&req->r_unsafe_item);
2050
2051         ktime_get_coarse_real_ts64(&ts);
2052         req->r_stamp = timespec64_trunc(ts, mdsc->fsc->sb->s_time_gran);
2053
2054         req->r_op = op;
2055         req->r_direct_mode = mode;
2056         return req;
2057 }
2058
2059 /*
2060  * return oldest (lowest) request, tid in request tree, 0 if none.
2061  *
2062  * called under mdsc->mutex.
2063  */
2064 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
2065 {
2066         if (RB_EMPTY_ROOT(&mdsc->request_tree))
2067                 return NULL;
2068         return rb_entry(rb_first(&mdsc->request_tree),
2069                         struct ceph_mds_request, r_node);
2070 }
2071
2072 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
2073 {
2074         return mdsc->oldest_tid;
2075 }
2076
2077 /*
2078  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
2079  * on build_path_from_dentry in fs/cifs/dir.c.
2080  *
2081  * If @stop_on_nosnap, generate path relative to the first non-snapped
2082  * inode.
2083  *
2084  * Encode hidden .snap dirs as a double /, i.e.
2085  *   foo/.snap/bar -> foo//bar
2086  */
2087 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *pbase,
2088                            int stop_on_nosnap)
2089 {
2090         struct dentry *temp;
2091         char *path;
2092         int pos;
2093         unsigned seq;
2094         u64 base;
2095
2096         if (!dentry)
2097                 return ERR_PTR(-EINVAL);
2098
2099         path = __getname();
2100         if (!path)
2101                 return ERR_PTR(-ENOMEM);
2102 retry:
2103         pos = PATH_MAX - 1;
2104         path[pos] = '\0';
2105
2106         seq = read_seqbegin(&rename_lock);
2107         rcu_read_lock();
2108         temp = dentry;
2109         for (;;) {
2110                 struct inode *inode;
2111
2112                 spin_lock(&temp->d_lock);
2113                 inode = d_inode(temp);
2114                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
2115                         dout("build_path path+%d: %p SNAPDIR\n",
2116                              pos, temp);
2117                 } else if (stop_on_nosnap && inode &&
2118                            ceph_snap(inode) == CEPH_NOSNAP) {
2119                         spin_unlock(&temp->d_lock);
2120                         break;
2121                 } else {
2122                         pos -= temp->d_name.len;
2123                         if (pos < 0) {
2124                                 spin_unlock(&temp->d_lock);
2125                                 break;
2126                         }
2127                         memcpy(path + pos, temp->d_name.name, temp->d_name.len);
2128                 }
2129                 spin_unlock(&temp->d_lock);
2130                 temp = temp->d_parent;
2131
2132                 /* Are we at the root? */
2133                 if (IS_ROOT(temp))
2134                         break;
2135
2136                 /* Are we out of buffer? */
2137                 if (--pos < 0)
2138                         break;
2139
2140                 path[pos] = '/';
2141         }
2142         base = ceph_ino(d_inode(temp));
2143         rcu_read_unlock();
2144         if (pos < 0 || read_seqretry(&rename_lock, seq)) {
2145                 pr_err("build_path did not end path lookup where "
2146                        "expected, pos is %d\n", pos);
2147                 /* presumably this is only possible if racing with a
2148                    rename of one of the parent directories (we can not
2149                    lock the dentries above us to prevent this, but
2150                    retrying should be harmless) */
2151                 goto retry;
2152         }
2153
2154         *pbase = base;
2155         *plen = PATH_MAX - 1 - pos;
2156         dout("build_path on %p %d built %llx '%.*s'\n",
2157              dentry, d_count(dentry), base, *plen, path + pos);
2158         return path + pos;
2159 }
2160
2161 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
2162                              const char **ppath, int *ppathlen, u64 *pino,
2163                              bool *pfreepath, bool parent_locked)
2164 {
2165         char *path;
2166
2167         rcu_read_lock();
2168         if (!dir)
2169                 dir = d_inode_rcu(dentry->d_parent);
2170         if (dir && parent_locked && ceph_snap(dir) == CEPH_NOSNAP) {
2171                 *pino = ceph_ino(dir);
2172                 rcu_read_unlock();
2173                 *ppath = dentry->d_name.name;
2174                 *ppathlen = dentry->d_name.len;
2175                 return 0;
2176         }
2177         rcu_read_unlock();
2178         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2179         if (IS_ERR(path))
2180                 return PTR_ERR(path);
2181         *ppath = path;
2182         *pfreepath = true;
2183         return 0;
2184 }
2185
2186 static int build_inode_path(struct inode *inode,
2187                             const char **ppath, int *ppathlen, u64 *pino,
2188                             bool *pfreepath)
2189 {
2190         struct dentry *dentry;
2191         char *path;
2192
2193         if (ceph_snap(inode) == CEPH_NOSNAP) {
2194                 *pino = ceph_ino(inode);
2195                 *ppathlen = 0;
2196                 return 0;
2197         }
2198         dentry = d_find_alias(inode);
2199         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2200         dput(dentry);
2201         if (IS_ERR(path))
2202                 return PTR_ERR(path);
2203         *ppath = path;
2204         *pfreepath = true;
2205         return 0;
2206 }
2207
2208 /*
2209  * request arguments may be specified via an inode *, a dentry *, or
2210  * an explicit ino+path.
2211  */
2212 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
2213                                   struct inode *rdiri, const char *rpath,
2214                                   u64 rino, const char **ppath, int *pathlen,
2215                                   u64 *ino, bool *freepath, bool parent_locked)
2216 {
2217         int r = 0;
2218
2219         if (rinode) {
2220                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2221                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2222                      ceph_snap(rinode));
2223         } else if (rdentry) {
2224                 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
2225                                         freepath, parent_locked);
2226                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
2227                      *ppath);
2228         } else if (rpath || rino) {
2229                 *ino = rino;
2230                 *ppath = rpath;
2231                 *pathlen = rpath ? strlen(rpath) : 0;
2232                 dout(" path %.*s\n", *pathlen, rpath);
2233         }
2234
2235         return r;
2236 }
2237
2238 /*
2239  * called under mdsc->mutex
2240  */
2241 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
2242                                                struct ceph_mds_request *req,
2243                                                int mds, bool drop_cap_releases)
2244 {
2245         struct ceph_msg *msg;
2246         struct ceph_mds_request_head *head;
2247         const char *path1 = NULL;
2248         const char *path2 = NULL;
2249         u64 ino1 = 0, ino2 = 0;
2250         int pathlen1 = 0, pathlen2 = 0;
2251         bool freepath1 = false, freepath2 = false;
2252         int len;
2253         u16 releases;
2254         void *p, *end;
2255         int ret;
2256
2257         ret = set_request_path_attr(req->r_inode, req->r_dentry,
2258                               req->r_parent, req->r_path1, req->r_ino1.ino,
2259                               &path1, &pathlen1, &ino1, &freepath1,
2260                               test_bit(CEPH_MDS_R_PARENT_LOCKED,
2261                                         &req->r_req_flags));
2262         if (ret < 0) {
2263                 msg = ERR_PTR(ret);
2264                 goto out;
2265         }
2266
2267         /* If r_old_dentry is set, then assume that its parent is locked */
2268         ret = set_request_path_attr(NULL, req->r_old_dentry,
2269                               req->r_old_dentry_dir,
2270                               req->r_path2, req->r_ino2.ino,
2271                               &path2, &pathlen2, &ino2, &freepath2, true);
2272         if (ret < 0) {
2273                 msg = ERR_PTR(ret);
2274                 goto out_free1;
2275         }
2276
2277         len = sizeof(*head) +
2278                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2279                 sizeof(struct ceph_timespec);
2280
2281         /* calculate (max) length for cap releases */
2282         len += sizeof(struct ceph_mds_request_release) *
2283                 (!!req->r_inode_drop + !!req->r_dentry_drop +
2284                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2285         if (req->r_dentry_drop)
2286                 len += pathlen1;
2287         if (req->r_old_dentry_drop)
2288                 len += pathlen2;
2289
2290         msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
2291         if (!msg) {
2292                 msg = ERR_PTR(-ENOMEM);
2293                 goto out_free2;
2294         }
2295
2296         msg->hdr.version = cpu_to_le16(2);
2297         msg->hdr.tid = cpu_to_le64(req->r_tid);
2298
2299         head = msg->front.iov_base;
2300         p = msg->front.iov_base + sizeof(*head);
2301         end = msg->front.iov_base + msg->front.iov_len;
2302
2303         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2304         head->op = cpu_to_le32(req->r_op);
2305         head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
2306         head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
2307         head->args = req->r_args;
2308
2309         ceph_encode_filepath(&p, end, ino1, path1);
2310         ceph_encode_filepath(&p, end, ino2, path2);
2311
2312         /* make note of release offset, in case we need to replay */
2313         req->r_request_release_offset = p - msg->front.iov_base;
2314
2315         /* cap releases */
2316         releases = 0;
2317         if (req->r_inode_drop)
2318                 releases += ceph_encode_inode_release(&p,
2319                       req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2320                       mds, req->r_inode_drop, req->r_inode_unless, 0);
2321         if (req->r_dentry_drop)
2322                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
2323                                 req->r_parent, mds, req->r_dentry_drop,
2324                                 req->r_dentry_unless);
2325         if (req->r_old_dentry_drop)
2326                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2327                                 req->r_old_dentry_dir, mds,
2328                                 req->r_old_dentry_drop,
2329                                 req->r_old_dentry_unless);
2330         if (req->r_old_inode_drop)
2331                 releases += ceph_encode_inode_release(&p,
2332                       d_inode(req->r_old_dentry),
2333                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2334
2335         if (drop_cap_releases) {
2336                 releases = 0;
2337                 p = msg->front.iov_base + req->r_request_release_offset;
2338         }
2339
2340         head->num_releases = cpu_to_le16(releases);
2341
2342         /* time stamp */
2343         {
2344                 struct ceph_timespec ts;
2345                 ceph_encode_timespec64(&ts, &req->r_stamp);
2346                 ceph_encode_copy(&p, &ts, sizeof(ts));
2347         }
2348
2349         BUG_ON(p > end);
2350         msg->front.iov_len = p - msg->front.iov_base;
2351         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2352
2353         if (req->r_pagelist) {
2354                 struct ceph_pagelist *pagelist = req->r_pagelist;
2355                 ceph_msg_data_add_pagelist(msg, pagelist);
2356                 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2357         } else {
2358                 msg->hdr.data_len = 0;
2359         }
2360
2361         msg->hdr.data_off = cpu_to_le16(0);
2362
2363 out_free2:
2364         if (freepath2)
2365                 ceph_mdsc_free_path((char *)path2, pathlen2);
2366 out_free1:
2367         if (freepath1)
2368                 ceph_mdsc_free_path((char *)path1, pathlen1);
2369 out:
2370         return msg;
2371 }
2372
2373 /*
2374  * called under mdsc->mutex if error, under no mutex if
2375  * success.
2376  */
2377 static void complete_request(struct ceph_mds_client *mdsc,
2378                              struct ceph_mds_request *req)
2379 {
2380         if (req->r_callback)
2381                 req->r_callback(mdsc, req);
2382         complete_all(&req->r_completion);
2383 }
2384
2385 /*
2386  * called under mdsc->mutex
2387  */
2388 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2389                                   struct ceph_mds_request *req,
2390                                   int mds, bool drop_cap_releases)
2391 {
2392         struct ceph_mds_request_head *rhead;
2393         struct ceph_msg *msg;
2394         int flags = 0;
2395
2396         req->r_attempts++;
2397         if (req->r_inode) {
2398                 struct ceph_cap *cap =
2399                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2400
2401                 if (cap)
2402                         req->r_sent_on_mseq = cap->mseq;
2403                 else
2404                         req->r_sent_on_mseq = -1;
2405         }
2406         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2407              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2408
2409         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2410                 void *p;
2411                 /*
2412                  * Replay.  Do not regenerate message (and rebuild
2413                  * paths, etc.); just use the original message.
2414                  * Rebuilding paths will break for renames because
2415                  * d_move mangles the src name.
2416                  */
2417                 msg = req->r_request;
2418                 rhead = msg->front.iov_base;
2419
2420                 flags = le32_to_cpu(rhead->flags);
2421                 flags |= CEPH_MDS_FLAG_REPLAY;
2422                 rhead->flags = cpu_to_le32(flags);
2423
2424                 if (req->r_target_inode)
2425                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2426
2427                 rhead->num_retry = req->r_attempts - 1;
2428
2429                 /* remove cap/dentry releases from message */
2430                 rhead->num_releases = 0;
2431
2432                 /* time stamp */
2433                 p = msg->front.iov_base + req->r_request_release_offset;
2434                 {
2435                         struct ceph_timespec ts;
2436                         ceph_encode_timespec64(&ts, &req->r_stamp);
2437                         ceph_encode_copy(&p, &ts, sizeof(ts));
2438                 }
2439
2440                 msg->front.iov_len = p - msg->front.iov_base;
2441                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2442                 return 0;
2443         }
2444
2445         if (req->r_request) {
2446                 ceph_msg_put(req->r_request);
2447                 req->r_request = NULL;
2448         }
2449         msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2450         if (IS_ERR(msg)) {
2451                 req->r_err = PTR_ERR(msg);
2452                 return PTR_ERR(msg);
2453         }
2454         req->r_request = msg;
2455
2456         rhead = msg->front.iov_base;
2457         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2458         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2459                 flags |= CEPH_MDS_FLAG_REPLAY;
2460         if (req->r_parent)
2461                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2462         rhead->flags = cpu_to_le32(flags);
2463         rhead->num_fwd = req->r_num_fwd;
2464         rhead->num_retry = req->r_attempts - 1;
2465         rhead->ino = 0;
2466
2467         dout(" r_parent = %p\n", req->r_parent);
2468         return 0;
2469 }
2470
2471 /*
2472  * send request, or put it on the appropriate wait list.
2473  */
2474 static void __do_request(struct ceph_mds_client *mdsc,
2475                         struct ceph_mds_request *req)
2476 {
2477         struct ceph_mds_session *session = NULL;
2478         int mds = -1;
2479         int err = 0;
2480
2481         if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2482                 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2483                         __unregister_request(mdsc, req);
2484                 return;
2485         }
2486
2487         if (req->r_timeout &&
2488             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2489                 dout("do_request timed out\n");
2490                 err = -EIO;
2491                 goto finish;
2492         }
2493         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2494                 dout("do_request forced umount\n");
2495                 err = -EIO;
2496                 goto finish;
2497         }
2498         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2499                 if (mdsc->mdsmap_err) {
2500                         err = mdsc->mdsmap_err;
2501                         dout("do_request mdsmap err %d\n", err);
2502                         goto finish;
2503                 }
2504                 if (mdsc->mdsmap->m_epoch == 0) {
2505                         dout("do_request no mdsmap, waiting for map\n");
2506                         list_add(&req->r_wait, &mdsc->waiting_for_map);
2507                         return;
2508                 }
2509                 if (!(mdsc->fsc->mount_options->flags &
2510                       CEPH_MOUNT_OPT_MOUNTWAIT) &&
2511                     !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2512                         err = -ENOENT;
2513                         pr_info("probably no mds server is up\n");
2514                         goto finish;
2515                 }
2516         }
2517
2518         put_request_session(req);
2519
2520         mds = __choose_mds(mdsc, req);
2521         if (mds < 0 ||
2522             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2523                 dout("do_request no mds or not active, waiting for map\n");
2524                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2525                 return;
2526         }
2527
2528         /* get, open session */
2529         session = __ceph_lookup_mds_session(mdsc, mds);
2530         if (!session) {
2531                 session = register_session(mdsc, mds);
2532                 if (IS_ERR(session)) {
2533                         err = PTR_ERR(session);
2534                         goto finish;
2535                 }
2536         }
2537         req->r_session = get_session(session);
2538
2539         dout("do_request mds%d session %p state %s\n", mds, session,
2540              ceph_session_state_name(session->s_state));
2541         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2542             session->s_state != CEPH_MDS_SESSION_HUNG) {
2543                 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2544                         err = -EACCES;
2545                         goto out_session;
2546                 }
2547                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2548                     session->s_state == CEPH_MDS_SESSION_CLOSING)
2549                         __open_session(mdsc, session);
2550                 list_add(&req->r_wait, &session->s_waiting);
2551                 goto out_session;
2552         }
2553
2554         /* send request */
2555         req->r_resend_mds = -1;   /* forget any previous mds hint */
2556
2557         if (req->r_request_started == 0)   /* note request start time */
2558                 req->r_request_started = jiffies;
2559
2560         err = __prepare_send_request(mdsc, req, mds, false);
2561         if (!err) {
2562                 ceph_msg_get(req->r_request);
2563                 ceph_con_send(&session->s_con, req->r_request);
2564         }
2565
2566 out_session:
2567         ceph_put_mds_session(session);
2568 finish:
2569         if (err) {
2570                 dout("__do_request early error %d\n", err);
2571                 req->r_err = err;
2572                 complete_request(mdsc, req);
2573                 __unregister_request(mdsc, req);
2574         }
2575         return;
2576 }
2577
2578 /*
2579  * called under mdsc->mutex
2580  */
2581 static void __wake_requests(struct ceph_mds_client *mdsc,
2582                             struct list_head *head)
2583 {
2584         struct ceph_mds_request *req;
2585         LIST_HEAD(tmp_list);
2586
2587         list_splice_init(head, &tmp_list);
2588
2589         while (!list_empty(&tmp_list)) {
2590                 req = list_entry(tmp_list.next,
2591                                  struct ceph_mds_request, r_wait);
2592                 list_del_init(&req->r_wait);
2593                 dout(" wake request %p tid %llu\n", req, req->r_tid);
2594                 __do_request(mdsc, req);
2595         }
2596 }
2597
2598 /*
2599  * Wake up threads with requests pending for @mds, so that they can
2600  * resubmit their requests to a possibly different mds.
2601  */
2602 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2603 {
2604         struct ceph_mds_request *req;
2605         struct rb_node *p = rb_first(&mdsc->request_tree);
2606
2607         dout("kick_requests mds%d\n", mds);
2608         while (p) {
2609                 req = rb_entry(p, struct ceph_mds_request, r_node);
2610                 p = rb_next(p);
2611                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2612                         continue;
2613                 if (req->r_attempts > 0)
2614                         continue; /* only new requests */
2615                 if (req->r_session &&
2616                     req->r_session->s_mds == mds) {
2617                         dout(" kicking tid %llu\n", req->r_tid);
2618                         list_del_init(&req->r_wait);
2619                         __do_request(mdsc, req);
2620                 }
2621         }
2622 }
2623
2624 int ceph_mdsc_submit_request(struct ceph_mds_client *mdsc, struct inode *dir,
2625                               struct ceph_mds_request *req)
2626 {
2627         int err;
2628
2629         /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2630         if (req->r_inode)
2631                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2632         if (req->r_parent)
2633                 ceph_get_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
2634         if (req->r_old_dentry_dir)
2635                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2636                                   CEPH_CAP_PIN);
2637
2638         dout("submit_request on %p for inode %p\n", req, dir);
2639         mutex_lock(&mdsc->mutex);
2640         __register_request(mdsc, req, dir);
2641         __do_request(mdsc, req);
2642         err = req->r_err;
2643         mutex_unlock(&mdsc->mutex);
2644         return err;
2645 }
2646
2647 static int ceph_mdsc_wait_request(struct ceph_mds_client *mdsc,
2648                                   struct ceph_mds_request *req)
2649 {
2650         int err;
2651
2652         /* wait */
2653         dout("do_request waiting\n");
2654         if (!req->r_timeout && req->r_wait_for_completion) {
2655                 err = req->r_wait_for_completion(mdsc, req);
2656         } else {
2657                 long timeleft = wait_for_completion_killable_timeout(
2658                                         &req->r_completion,
2659                                         ceph_timeout_jiffies(req->r_timeout));
2660                 if (timeleft > 0)
2661                         err = 0;
2662                 else if (!timeleft)
2663                         err = -EIO;  /* timed out */
2664                 else
2665                         err = timeleft;  /* killed */
2666         }
2667         dout("do_request waited, got %d\n", err);
2668         mutex_lock(&mdsc->mutex);
2669
2670         /* only abort if we didn't race with a real reply */
2671         if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2672                 err = le32_to_cpu(req->r_reply_info.head->result);
2673         } else if (err < 0) {
2674                 dout("aborted request %lld with %d\n", req->r_tid, err);
2675
2676                 /*
2677                  * ensure we aren't running concurrently with
2678                  * ceph_fill_trace or ceph_readdir_prepopulate, which
2679                  * rely on locks (dir mutex) held by our caller.
2680                  */
2681                 mutex_lock(&req->r_fill_mutex);
2682                 req->r_err = err;
2683                 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2684                 mutex_unlock(&req->r_fill_mutex);
2685
2686                 if (req->r_parent &&
2687                     (req->r_op & CEPH_MDS_OP_WRITE))
2688                         ceph_invalidate_dir_request(req);
2689         } else {
2690                 err = req->r_err;
2691         }
2692
2693         mutex_unlock(&mdsc->mutex);
2694         return err;
2695 }
2696
2697 /*
2698  * Synchrously perform an mds request.  Take care of all of the
2699  * session setup, forwarding, retry details.
2700  */
2701 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2702                          struct inode *dir,
2703                          struct ceph_mds_request *req)
2704 {
2705         int err;
2706
2707         dout("do_request on %p\n", req);
2708
2709         /* issue */
2710         err = ceph_mdsc_submit_request(mdsc, dir, req);
2711         if (!err)
2712                 err = ceph_mdsc_wait_request(mdsc, req);
2713         dout("do_request %p done, result %d\n", req, err);
2714         return err;
2715 }
2716
2717 /*
2718  * Invalidate dir's completeness, dentry lease state on an aborted MDS
2719  * namespace request.
2720  */
2721 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2722 {
2723         struct inode *dir = req->r_parent;
2724         struct inode *old_dir = req->r_old_dentry_dir;
2725
2726         dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
2727
2728         ceph_dir_clear_complete(dir);
2729         if (old_dir)
2730                 ceph_dir_clear_complete(old_dir);
2731         if (req->r_dentry)
2732                 ceph_invalidate_dentry_lease(req->r_dentry);
2733         if (req->r_old_dentry)
2734                 ceph_invalidate_dentry_lease(req->r_old_dentry);
2735 }
2736
2737 /*
2738  * Handle mds reply.
2739  *
2740  * We take the session mutex and parse and process the reply immediately.
2741  * This preserves the logical ordering of replies, capabilities, etc., sent
2742  * by the MDS as they are applied to our local cache.
2743  */
2744 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2745 {
2746         struct ceph_mds_client *mdsc = session->s_mdsc;
2747         struct ceph_mds_request *req;
2748         struct ceph_mds_reply_head *head = msg->front.iov_base;
2749         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2750         struct ceph_snap_realm *realm;
2751         u64 tid;
2752         int err, result;
2753         int mds = session->s_mds;
2754
2755         if (msg->front.iov_len < sizeof(*head)) {
2756                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2757                 ceph_msg_dump(msg);
2758                 return;
2759         }
2760
2761         /* get request, session */
2762         tid = le64_to_cpu(msg->hdr.tid);
2763         mutex_lock(&mdsc->mutex);
2764         req = lookup_get_request(mdsc, tid);
2765         if (!req) {
2766                 dout("handle_reply on unknown tid %llu\n", tid);
2767                 mutex_unlock(&mdsc->mutex);
2768                 return;
2769         }
2770         dout("handle_reply %p\n", req);
2771
2772         /* correct session? */
2773         if (req->r_session != session) {
2774                 pr_err("mdsc_handle_reply got %llu on session mds%d"
2775                        " not mds%d\n", tid, session->s_mds,
2776                        req->r_session ? req->r_session->s_mds : -1);
2777                 mutex_unlock(&mdsc->mutex);
2778                 goto out;
2779         }
2780
2781         /* dup? */
2782         if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
2783             (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
2784                 pr_warn("got a dup %s reply on %llu from mds%d\n",
2785                            head->safe ? "safe" : "unsafe", tid, mds);
2786                 mutex_unlock(&mdsc->mutex);
2787                 goto out;
2788         }
2789         if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
2790                 pr_warn("got unsafe after safe on %llu from mds%d\n",
2791                            tid, mds);
2792                 mutex_unlock(&mdsc->mutex);
2793                 goto out;
2794         }
2795
2796         result = le32_to_cpu(head->result);
2797
2798         /*
2799          * Handle an ESTALE
2800          * if we're not talking to the authority, send to them
2801          * if the authority has changed while we weren't looking,
2802          * send to new authority
2803          * Otherwise we just have to return an ESTALE
2804          */
2805         if (result == -ESTALE) {
2806                 dout("got ESTALE on request %llu\n", req->r_tid);
2807                 req->r_resend_mds = -1;
2808                 if (req->r_direct_mode != USE_AUTH_MDS) {
2809                         dout("not using auth, setting for that now\n");
2810                         req->r_direct_mode = USE_AUTH_MDS;
2811                         __do_request(mdsc, req);
2812                         mutex_unlock(&mdsc->mutex);
2813                         goto out;
2814                 } else  {
2815                         int mds = __choose_mds(mdsc, req);
2816                         if (mds >= 0 && mds != req->r_session->s_mds) {
2817                                 dout("but auth changed, so resending\n");
2818                                 __do_request(mdsc, req);
2819                                 mutex_unlock(&mdsc->mutex);
2820                                 goto out;
2821                         }
2822                 }
2823                 dout("have to return ESTALE on request %llu\n", req->r_tid);
2824         }
2825
2826
2827         if (head->safe) {
2828                 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
2829                 __unregister_request(mdsc, req);
2830
2831                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2832                         /*
2833                          * We already handled the unsafe response, now do the
2834                          * cleanup.  No need to examine the response; the MDS
2835                          * doesn't include any result info in the safe
2836                          * response.  And even if it did, there is nothing
2837                          * useful we could do with a revised return value.
2838                          */
2839                         dout("got safe reply %llu, mds%d\n", tid, mds);
2840
2841                         /* last unsafe request during umount? */
2842                         if (mdsc->stopping && !__get_oldest_req(mdsc))
2843                                 complete_all(&mdsc->safe_umount_waiters);
2844                         mutex_unlock(&mdsc->mutex);
2845                         goto out;
2846                 }
2847         } else {
2848                 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
2849                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2850                 if (req->r_unsafe_dir) {
2851                         struct ceph_inode_info *ci =
2852                                         ceph_inode(req->r_unsafe_dir);
2853                         spin_lock(&ci->i_unsafe_lock);
2854                         list_add_tail(&req->r_unsafe_dir_item,
2855                                       &ci->i_unsafe_dirops);
2856                         spin_unlock(&ci->i_unsafe_lock);
2857                 }
2858         }
2859
2860         dout("handle_reply tid %lld result %d\n", tid, result);
2861         rinfo = &req->r_reply_info;
2862         if (test_bit(CEPHFS_FEATURE_REPLY_ENCODING, &session->s_features))
2863                 err = parse_reply_info(msg, rinfo, (u64)-1);
2864         else
2865                 err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2866         mutex_unlock(&mdsc->mutex);
2867
2868         mutex_lock(&session->s_mutex);
2869         if (err < 0) {
2870                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2871                 ceph_msg_dump(msg);
2872                 goto out_err;
2873         }
2874
2875         /* snap trace */
2876         realm = NULL;
2877         if (rinfo->snapblob_len) {
2878                 down_write(&mdsc->snap_rwsem);
2879                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2880                                 rinfo->snapblob + rinfo->snapblob_len,
2881                                 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2882                                 &realm);
2883                 downgrade_write(&mdsc->snap_rwsem);
2884         } else {
2885                 down_read(&mdsc->snap_rwsem);
2886         }
2887
2888         /* insert trace into our cache */
2889         mutex_lock(&req->r_fill_mutex);
2890         current->journal_info = req;
2891         err = ceph_fill_trace(mdsc->fsc->sb, req);
2892         if (err == 0) {
2893                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2894                                     req->r_op == CEPH_MDS_OP_LSSNAP))
2895                         ceph_readdir_prepopulate(req, req->r_session);
2896         }
2897         current->journal_info = NULL;
2898         mutex_unlock(&req->r_fill_mutex);
2899
2900         up_read(&mdsc->snap_rwsem);
2901         if (realm)
2902                 ceph_put_snap_realm(mdsc, realm);
2903
2904         if (err == 0) {
2905                 if (req->r_target_inode &&
2906                     test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2907                         struct ceph_inode_info *ci =
2908                                 ceph_inode(req->r_target_inode);
2909                         spin_lock(&ci->i_unsafe_lock);
2910                         list_add_tail(&req->r_unsafe_target_item,
2911                                       &ci->i_unsafe_iops);
2912                         spin_unlock(&ci->i_unsafe_lock);
2913                 }
2914
2915                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2916         }
2917 out_err:
2918         mutex_lock(&mdsc->mutex);
2919         if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2920                 if (err) {
2921                         req->r_err = err;
2922                 } else {
2923                         req->r_reply =  ceph_msg_get(msg);
2924                         set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
2925                 }
2926         } else {
2927                 dout("reply arrived after request %lld was aborted\n", tid);
2928         }
2929         mutex_unlock(&mdsc->mutex);
2930
2931         mutex_unlock(&session->s_mutex);
2932
2933         /* kick calling process */
2934         complete_request(mdsc, req);
2935 out:
2936         ceph_mdsc_put_request(req);
2937         return;
2938 }
2939
2940
2941
2942 /*
2943  * handle mds notification that our request has been forwarded.
2944  */
2945 static void handle_forward(struct ceph_mds_client *mdsc,
2946                            struct ceph_mds_session *session,
2947                            struct ceph_msg *msg)
2948 {
2949         struct ceph_mds_request *req;
2950         u64 tid = le64_to_cpu(msg->hdr.tid);
2951         u32 next_mds;
2952         u32 fwd_seq;
2953         int err = -EINVAL;
2954         void *p = msg->front.iov_base;
2955         void *end = p + msg->front.iov_len;
2956
2957         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2958         next_mds = ceph_decode_32(&p);
2959         fwd_seq = ceph_decode_32(&p);
2960
2961         mutex_lock(&mdsc->mutex);
2962         req = lookup_get_request(mdsc, tid);
2963         if (!req) {
2964                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2965                 goto out;  /* dup reply? */
2966         }
2967
2968         if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2969                 dout("forward tid %llu aborted, unregistering\n", tid);
2970                 __unregister_request(mdsc, req);
2971         } else if (fwd_seq <= req->r_num_fwd) {
2972                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2973                      tid, next_mds, req->r_num_fwd, fwd_seq);
2974         } else {
2975                 /* resend. forward race not possible; mds would drop */
2976                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2977                 BUG_ON(req->r_err);
2978                 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
2979                 req->r_attempts = 0;
2980                 req->r_num_fwd = fwd_seq;
2981                 req->r_resend_mds = next_mds;
2982                 put_request_session(req);
2983                 __do_request(mdsc, req);
2984         }
2985         ceph_mdsc_put_request(req);
2986 out:
2987         mutex_unlock(&mdsc->mutex);
2988         return;
2989
2990 bad:
2991         pr_err("mdsc_handle_forward decode error err=%d\n", err);
2992 }
2993
2994 static int __decode_and_drop_session_metadata(void **p, void *end)
2995 {
2996         /* map<string,string> */
2997         u32 n;
2998         ceph_decode_32_safe(p, end, n, bad);
2999         while (n-- > 0) {
3000                 u32 len;
3001                 ceph_decode_32_safe(p, end, len, bad);
3002                 ceph_decode_need(p, end, len, bad);
3003                 *p += len;
3004                 ceph_decode_32_safe(p, end, len, bad);
3005                 ceph_decode_need(p, end, len, bad);
3006                 *p += len;
3007         }
3008         return 0;
3009 bad:
3010         return -1;
3011 }
3012
3013 /*
3014  * handle a mds session control message
3015  */
3016 static void handle_session(struct ceph_mds_session *session,
3017                            struct ceph_msg *msg)
3018 {
3019         struct ceph_mds_client *mdsc = session->s_mdsc;
3020         int mds = session->s_mds;
3021         int msg_version = le16_to_cpu(msg->hdr.version);
3022         void *p = msg->front.iov_base;
3023         void *end = p + msg->front.iov_len;
3024         struct ceph_mds_session_head *h;
3025         u32 op;
3026         u64 seq;
3027         unsigned long features = 0;
3028         int wake = 0;
3029
3030         /* decode */
3031         ceph_decode_need(&p, end, sizeof(*h), bad);
3032         h = p;
3033         p += sizeof(*h);
3034
3035         op = le32_to_cpu(h->op);
3036         seq = le64_to_cpu(h->seq);
3037
3038         if (msg_version >= 3) {
3039                 u32 len;
3040                 /* version >= 2, metadata */
3041                 if (__decode_and_drop_session_metadata(&p, end) < 0)
3042                         goto bad;
3043                 /* version >= 3, feature bits */
3044                 ceph_decode_32_safe(&p, end, len, bad);
3045                 ceph_decode_need(&p, end, len, bad);
3046                 memcpy(&features, p, min_t(size_t, len, sizeof(features)));
3047                 p += len;
3048         }
3049
3050         mutex_lock(&mdsc->mutex);
3051         if (op == CEPH_SESSION_CLOSE) {
3052                 get_session(session);
3053                 __unregister_session(mdsc, session);
3054         }
3055         /* FIXME: this ttl calculation is generous */
3056         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
3057         mutex_unlock(&mdsc->mutex);
3058
3059         mutex_lock(&session->s_mutex);
3060
3061         dout("handle_session mds%d %s %p state %s seq %llu\n",
3062              mds, ceph_session_op_name(op), session,
3063              ceph_session_state_name(session->s_state), seq);
3064
3065         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
3066                 session->s_state = CEPH_MDS_SESSION_OPEN;
3067                 pr_info("mds%d came back\n", session->s_mds);
3068         }
3069
3070         switch (op) {
3071         case CEPH_SESSION_OPEN:
3072                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3073                         pr_info("mds%d reconnect success\n", session->s_mds);
3074                 session->s_state = CEPH_MDS_SESSION_OPEN;
3075                 session->s_features = features;
3076                 renewed_caps(mdsc, session, 0);
3077                 wake = 1;
3078                 if (mdsc->stopping)
3079                         __close_session(mdsc, session);
3080                 break;
3081
3082         case CEPH_SESSION_RENEWCAPS:
3083                 if (session->s_renew_seq == seq)
3084                         renewed_caps(mdsc, session, 1);
3085                 break;
3086
3087         case CEPH_SESSION_CLOSE:
3088                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3089                         pr_info("mds%d reconnect denied\n", session->s_mds);
3090                 cleanup_session_requests(mdsc, session);
3091                 remove_session_caps(session);
3092                 wake = 2; /* for good measure */
3093                 wake_up_all(&mdsc->session_close_wq);
3094                 break;
3095
3096         case CEPH_SESSION_STALE:
3097                 pr_info("mds%d caps went stale, renewing\n",
3098                         session->s_mds);
3099                 spin_lock(&session->s_gen_ttl_lock);
3100                 session->s_cap_gen++;
3101                 session->s_cap_ttl = jiffies - 1;
3102                 spin_unlock(&session->s_gen_ttl_lock);
3103                 send_renew_caps(mdsc, session);
3104                 break;
3105
3106         case CEPH_SESSION_RECALL_STATE:
3107                 ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
3108                 break;
3109
3110         case CEPH_SESSION_FLUSHMSG:
3111                 send_flushmsg_ack(mdsc, session, seq);
3112                 break;
3113
3114         case CEPH_SESSION_FORCE_RO:
3115                 dout("force_session_readonly %p\n", session);
3116                 spin_lock(&session->s_cap_lock);
3117                 session->s_readonly = true;
3118                 spin_unlock(&session->s_cap_lock);
3119                 wake_up_session_caps(session, FORCE_RO);
3120                 break;
3121
3122         case CEPH_SESSION_REJECT:
3123                 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
3124                 pr_info("mds%d rejected session\n", session->s_mds);
3125                 session->s_state = CEPH_MDS_SESSION_REJECTED;
3126                 cleanup_session_requests(mdsc, session);
3127                 remove_session_caps(session);
3128                 wake = 2; /* for good measure */
3129                 break;
3130
3131         default:
3132                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
3133                 WARN_ON(1);
3134         }
3135
3136         mutex_unlock(&session->s_mutex);
3137         if (wake) {
3138                 mutex_lock(&mdsc->mutex);
3139                 __wake_requests(mdsc, &session->s_waiting);
3140                 if (wake == 2)
3141                         kick_requests(mdsc, mds);
3142                 mutex_unlock(&mdsc->mutex);
3143         }
3144         if (op == CEPH_SESSION_CLOSE)
3145                 ceph_put_mds_session(session);
3146         return;
3147
3148 bad:
3149         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
3150                (int)msg->front.iov_len);
3151         ceph_msg_dump(msg);
3152         return;
3153 }
3154
3155
3156 /*
3157  * called under session->mutex.
3158  */
3159 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
3160                                    struct ceph_mds_session *session)
3161 {
3162         struct ceph_mds_request *req, *nreq;
3163         struct rb_node *p;
3164         int err;
3165
3166         dout("replay_unsafe_requests mds%d\n", session->s_mds);
3167
3168         mutex_lock(&mdsc->mutex);
3169         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
3170                 err = __prepare_send_request(mdsc, req, session->s_mds, true);
3171                 if (!err) {
3172                         ceph_msg_get(req->r_request);
3173                         ceph_con_send(&session->s_con, req->r_request);
3174                 }
3175         }
3176
3177         /*
3178          * also re-send old requests when MDS enters reconnect stage. So that MDS
3179          * can process completed request in clientreplay stage.
3180          */
3181         p = rb_first(&mdsc->request_tree);
3182         while (p) {
3183                 req = rb_entry(p, struct ceph_mds_request, r_node);
3184                 p = rb_next(p);
3185                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3186                         continue;
3187                 if (req->r_attempts == 0)
3188                         continue; /* only old requests */
3189                 if (req->r_session &&
3190                     req->r_session->s_mds == session->s_mds) {
3191                         err = __prepare_send_request(mdsc, req,
3192                                                      session->s_mds, true);
3193                         if (!err) {
3194                                 ceph_msg_get(req->r_request);
3195                                 ceph_con_send(&session->s_con, req->r_request);
3196                         }
3197                 }
3198         }
3199         mutex_unlock(&mdsc->mutex);
3200 }
3201
3202 static int send_reconnect_partial(struct ceph_reconnect_state *recon_state)
3203 {
3204         struct ceph_msg *reply;
3205         struct ceph_pagelist *_pagelist;
3206         struct page *page;
3207         __le32 *addr;
3208         int err = -ENOMEM;
3209
3210         if (!recon_state->allow_multi)
3211                 return -ENOSPC;
3212
3213         /* can't handle message that contains both caps and realm */
3214         BUG_ON(!recon_state->nr_caps == !recon_state->nr_realms);
3215
3216         /* pre-allocate new pagelist */
3217         _pagelist = ceph_pagelist_alloc(GFP_NOFS);
3218         if (!_pagelist)
3219                 return -ENOMEM;
3220
3221         reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3222         if (!reply)
3223                 goto fail_msg;
3224
3225         /* placeholder for nr_caps */
3226         err = ceph_pagelist_encode_32(_pagelist, 0);
3227         if (err < 0)
3228                 goto fail;
3229
3230         if (recon_state->nr_caps) {
3231                 /* currently encoding caps */
3232                 err = ceph_pagelist_encode_32(recon_state->pagelist, 0);
3233                 if (err)
3234                         goto fail;
3235         } else {
3236                 /* placeholder for nr_realms (currently encoding relams) */
3237                 err = ceph_pagelist_encode_32(_pagelist, 0);
3238                 if (err < 0)
3239                         goto fail;
3240         }
3241
3242         err = ceph_pagelist_encode_8(recon_state->pagelist, 1);
3243         if (err)
3244                 goto fail;
3245
3246         page = list_first_entry(&recon_state->pagelist->head, struct page, lru);
3247         addr = kmap_atomic(page);
3248         if (recon_state->nr_caps) {
3249                 /* currently encoding caps */
3250                 *addr = cpu_to_le32(recon_state->nr_caps);
3251         } else {
3252                 /* currently encoding relams */
3253                 *(addr + 1) = cpu_to_le32(recon_state->nr_realms);
3254         }
3255         kunmap_atomic(addr);
3256
3257         reply->hdr.version = cpu_to_le16(5);
3258         reply->hdr.compat_version = cpu_to_le16(4);
3259
3260         reply->hdr.data_len = cpu_to_le32(recon_state->pagelist->length);
3261         ceph_msg_data_add_pagelist(reply, recon_state->pagelist);
3262
3263         ceph_con_send(&recon_state->session->s_con, reply);
3264         ceph_pagelist_release(recon_state->pagelist);
3265
3266         recon_state->pagelist = _pagelist;
3267         recon_state->nr_caps = 0;
3268         recon_state->nr_realms = 0;
3269         recon_state->msg_version = 5;
3270         return 0;
3271 fail:
3272         ceph_msg_put(reply);
3273 fail_msg:
3274         ceph_pagelist_release(_pagelist);
3275         return err;
3276 }
3277
3278 /*
3279  * Encode information about a cap for a reconnect with the MDS.
3280  */
3281 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
3282                           void *arg)
3283 {
3284         union {
3285                 struct ceph_mds_cap_reconnect v2;
3286                 struct ceph_mds_cap_reconnect_v1 v1;
3287         } rec;
3288         struct ceph_inode_info *ci = cap->ci;
3289         struct ceph_reconnect_state *recon_state = arg;
3290         struct ceph_pagelist *pagelist = recon_state->pagelist;
3291         int err;
3292         u64 snap_follows;
3293
3294         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
3295              inode, ceph_vinop(inode), cap, cap->cap_id,
3296              ceph_cap_string(cap->issued));
3297
3298         spin_lock(&ci->i_ceph_lock);
3299         cap->seq = 0;        /* reset cap seq */
3300         cap->issue_seq = 0;  /* and issue_seq */
3301         cap->mseq = 0;       /* and migrate_seq */
3302         cap->cap_gen = cap->session->s_cap_gen;
3303
3304         if (recon_state->msg_version >= 2) {
3305                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
3306                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3307                 rec.v2.issued = cpu_to_le32(cap->issued);
3308                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3309                 rec.v2.pathbase = 0;
3310                 rec.v2.flock_len = (__force __le32)
3311                         ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
3312         } else {
3313                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
3314                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3315                 rec.v1.issued = cpu_to_le32(cap->issued);
3316                 rec.v1.size = cpu_to_le64(inode->i_size);
3317                 ceph_encode_timespec64(&rec.v1.mtime, &inode->i_mtime);
3318                 ceph_encode_timespec64(&rec.v1.atime, &inode->i_atime);
3319                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3320                 rec.v1.pathbase = 0;
3321         }
3322
3323         if (list_empty(&ci->i_cap_snaps)) {
3324                 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
3325         } else {
3326                 struct ceph_cap_snap *capsnap =
3327                         list_first_entry(&ci->i_cap_snaps,
3328                                          struct ceph_cap_snap, ci_item);
3329                 snap_follows = capsnap->follows;
3330         }
3331         spin_unlock(&ci->i_ceph_lock);
3332
3333         if (recon_state->msg_version >= 2) {
3334                 int num_fcntl_locks, num_flock_locks;
3335                 struct ceph_filelock *flocks = NULL;
3336                 size_t struct_len, total_len = sizeof(u64);
3337                 u8 struct_v = 0;
3338
3339 encode_again:
3340                 if (rec.v2.flock_len) {
3341                         ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
3342                 } else {
3343                         num_fcntl_locks = 0;
3344                         num_flock_locks = 0;
3345                 }
3346                 if (num_fcntl_locks + num_flock_locks > 0) {
3347                         flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
3348                                                sizeof(struct ceph_filelock),
3349                                                GFP_NOFS);
3350                         if (!flocks) {
3351                                 err = -ENOMEM;
3352                                 goto out_err;
3353                         }
3354                         err = ceph_encode_locks_to_buffer(inode, flocks,
3355                                                           num_fcntl_locks,
3356                                                           num_flock_locks);
3357                         if (err) {
3358                                 kfree(flocks);
3359                                 flocks = NULL;
3360                                 if (err == -ENOSPC)
3361                                         goto encode_again;
3362                                 goto out_err;
3363                         }
3364                 } else {
3365                         kfree(flocks);
3366                         flocks = NULL;
3367                 }
3368
3369                 if (recon_state->msg_version >= 3) {
3370                         /* version, compat_version and struct_len */
3371                         total_len += 2 * sizeof(u8) + sizeof(u32);
3372                         struct_v = 2;
3373                 }
3374                 /*
3375                  * number of encoded locks is stable, so copy to pagelist
3376                  */
3377                 struct_len = 2 * sizeof(u32) +
3378                             (num_fcntl_locks + num_flock_locks) *
3379                             sizeof(struct ceph_filelock);
3380                 rec.v2.flock_len = cpu_to_le32(struct_len);
3381
3382                 struct_len += sizeof(u32) + sizeof(rec.v2);
3383
3384                 if (struct_v >= 2)
3385                         struct_len += sizeof(u64); /* snap_follows */
3386
3387                 total_len += struct_len;
3388
3389                 if (pagelist->length + total_len > RECONNECT_MAX_SIZE) {
3390                         err = send_reconnect_partial(recon_state);
3391                         if (err)
3392                                 goto out_freeflocks;
3393                         pagelist = recon_state->pagelist;
3394                 }
3395
3396                 err = ceph_pagelist_reserve(pagelist, total_len);
3397                 if (err)
3398                         goto out_freeflocks;
3399
3400                 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3401                 if (recon_state->msg_version >= 3) {
3402                         ceph_pagelist_encode_8(pagelist, struct_v);
3403                         ceph_pagelist_encode_8(pagelist, 1);
3404                         ceph_pagelist_encode_32(pagelist, struct_len);
3405                 }
3406                 ceph_pagelist_encode_string(pagelist, NULL, 0);
3407                 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
3408                 ceph_locks_to_pagelist(flocks, pagelist,
3409                                        num_fcntl_locks, num_flock_locks);
3410                 if (struct_v >= 2)
3411                         ceph_pagelist_encode_64(pagelist, snap_follows);
3412 out_freeflocks:
3413                 kfree(flocks);
3414         } else {
3415                 u64 pathbase = 0;
3416                 int pathlen = 0;
3417                 char *path = NULL;
3418                 struct dentry *dentry;
3419
3420                 dentry = d_find_alias(inode);
3421                 if (dentry) {
3422                         path = ceph_mdsc_build_path(dentry,
3423                                                 &pathlen, &pathbase, 0);
3424                         dput(dentry);
3425                         if (IS_ERR(path)) {
3426                                 err = PTR_ERR(path);
3427                                 goto out_err;
3428                         }
3429                         rec.v1.pathbase = cpu_to_le64(pathbase);
3430                 }
3431
3432                 err = ceph_pagelist_reserve(pagelist,
3433                                             sizeof(u64) + sizeof(u32) +
3434                                             pathlen + sizeof(rec.v1));
3435                 if (err) {
3436                         goto out_freepath;
3437                 }
3438
3439                 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3440                 ceph_pagelist_encode_string(pagelist, path, pathlen);
3441                 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
3442 out_freepath:
3443                 ceph_mdsc_free_path(path, pathlen);
3444         }
3445
3446 out_err:
3447         if (err >= 0)
3448                 recon_state->nr_caps++;
3449         return err;
3450 }
3451
3452 static int encode_snap_realms(struct ceph_mds_client *mdsc,
3453                               struct ceph_reconnect_state *recon_state)
3454 {
3455         struct rb_node *p;
3456         struct ceph_pagelist *pagelist = recon_state->pagelist;
3457         int err = 0;
3458
3459         if (recon_state->msg_version >= 4) {
3460                 err = ceph_pagelist_encode_32(pagelist, mdsc->num_snap_realms);
3461                 if (err < 0)
3462                         goto fail;
3463         }
3464
3465         /*
3466          * snaprealms.  we provide mds with the ino, seq (version), and
3467          * parent for all of our realms.  If the mds has any newer info,
3468          * it will tell us.
3469          */
3470         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3471                 struct ceph_snap_realm *realm =
3472                        rb_entry(p, struct ceph_snap_realm, node);
3473                 struct ceph_mds_snaprealm_reconnect sr_rec;
3474
3475                 if (recon_state->msg_version >= 4) {
3476                         size_t need = sizeof(u8) * 2 + sizeof(u32) +
3477                                       sizeof(sr_rec);
3478
3479                         if (pagelist->length + need > RECONNECT_MAX_SIZE) {
3480                                 err = send_reconnect_partial(recon_state);
3481                                 if (err)
3482                                         goto fail;
3483                                 pagelist = recon_state->pagelist;
3484                         }
3485
3486                         err = ceph_pagelist_reserve(pagelist, need);
3487                         if (err)
3488                                 goto fail;
3489
3490                         ceph_pagelist_encode_8(pagelist, 1);
3491                         ceph_pagelist_encode_8(pagelist, 1);
3492                         ceph_pagelist_encode_32(pagelist, sizeof(sr_rec));
3493                 }
3494
3495                 dout(" adding snap realm %llx seq %lld parent %llx\n",
3496                      realm->ino, realm->seq, realm->parent_ino);
3497                 sr_rec.ino = cpu_to_le64(realm->ino);
3498                 sr_rec.seq = cpu_to_le64(realm->seq);
3499                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3500
3501                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3502                 if (err)
3503                         goto fail;
3504
3505                 recon_state->nr_realms++;
3506         }
3507 fail:
3508         return err;
3509 }
3510
3511
3512 /*
3513  * If an MDS fails and recovers, clients need to reconnect in order to
3514  * reestablish shared state.  This includes all caps issued through
3515  * this session _and_ the snap_realm hierarchy.  Because it's not
3516  * clear which snap realms the mds cares about, we send everything we
3517  * know about.. that ensures we'll then get any new info the
3518  * recovering MDS might have.
3519  *
3520  * This is a relatively heavyweight operation, but it's rare.
3521  *
3522  * called with mdsc->mutex held.
3523  */
3524 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3525                                struct ceph_mds_session *session)
3526 {
3527         struct ceph_msg *reply;
3528         int mds = session->s_mds;
3529         int err = -ENOMEM;
3530         struct ceph_reconnect_state recon_state = {
3531                 .session = session,
3532         };
3533         LIST_HEAD(dispose);
3534
3535         pr_info("mds%d reconnect start\n", mds);
3536
3537         recon_state.pagelist = ceph_pagelist_alloc(GFP_NOFS);
3538         if (!recon_state.pagelist)
3539                 goto fail_nopagelist;
3540
3541         reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3542         if (!reply)
3543                 goto fail_nomsg;
3544
3545         mutex_lock(&session->s_mutex);
3546         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3547         session->s_seq = 0;
3548
3549         dout("session %p state %s\n", session,
3550              ceph_session_state_name(session->s_state));
3551
3552         spin_lock(&session->s_gen_ttl_lock);
3553         session->s_cap_gen++;
3554         spin_unlock(&session->s_gen_ttl_lock);
3555
3556         spin_lock(&session->s_cap_lock);
3557         /* don't know if session is readonly */
3558         session->s_readonly = 0;
3559         /*
3560          * notify __ceph_remove_cap() that we are composing cap reconnect.
3561          * If a cap get released before being added to the cap reconnect,
3562          * __ceph_remove_cap() should skip queuing cap release.
3563          */
3564         session->s_cap_reconnect = 1;
3565         /* drop old cap expires; we're about to reestablish that state */
3566         detach_cap_releases(session, &dispose);
3567         spin_unlock(&session->s_cap_lock);
3568         dispose_cap_releases(mdsc, &dispose);
3569
3570         /* trim unused caps to reduce MDS's cache rejoin time */
3571         if (mdsc->fsc->sb->s_root)
3572                 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3573
3574         ceph_con_close(&session->s_con);
3575         ceph_con_open(&session->s_con,
3576                       CEPH_ENTITY_TYPE_MDS, mds,
3577                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3578
3579         /* replay unsafe requests */
3580         replay_unsafe_requests(mdsc, session);
3581
3582         ceph_early_kick_flushing_caps(mdsc, session);
3583
3584         down_read(&mdsc->snap_rwsem);
3585
3586         /* placeholder for nr_caps */
3587         err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
3588         if (err)
3589                 goto fail;
3590
3591         if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
3592                 recon_state.msg_version = 3;
3593                 recon_state.allow_multi = true;
3594         } else if (session->s_con.peer_features & CEPH_FEATURE_MDSENC) {
3595                 recon_state.msg_version = 3;
3596         } else {
3597                 recon_state.msg_version = 2;
3598         }
3599         /* trsaverse this session's caps */
3600         err = ceph_iterate_session_caps(session, encode_caps_cb, &recon_state);
3601
3602         spin_lock(&session->s_cap_lock);
3603         session->s_cap_reconnect = 0;
3604         spin_unlock(&session->s_cap_lock);
3605
3606         if (err < 0)
3607                 goto fail;
3608
3609         /* check if all realms can be encoded into current message */
3610         if (mdsc->num_snap_realms) {
3611                 size_t total_len =
3612                         recon_state.pagelist->length +
3613                         mdsc->num_snap_realms *
3614                         sizeof(struct ceph_mds_snaprealm_reconnect);
3615                 if (recon_state.msg_version >= 4) {
3616                         /* number of realms */
3617                         total_len += sizeof(u32);
3618                         /* version, compat_version and struct_len */
3619                         total_len += mdsc->num_snap_realms *
3620                                      (2 * sizeof(u8) + sizeof(u32));
3621                 }
3622                 if (total_len > RECONNECT_MAX_SIZE) {
3623                         if (!recon_state.allow_multi) {
3624                                 err = -ENOSPC;
3625                                 goto fail;
3626                         }
3627                         if (recon_state.nr_caps) {
3628                                 err = send_reconnect_partial(&recon_state);
3629                                 if (err)
3630                                         goto fail;
3631                         }
3632                         recon_state.msg_version = 5;
3633                 }
3634         }
3635
3636         err = encode_snap_realms(mdsc, &recon_state);
3637         if (err < 0)
3638                 goto fail;
3639
3640         if (recon_state.msg_version >= 5) {
3641                 err = ceph_pagelist_encode_8(recon_state.pagelist, 0);
3642                 if (err < 0)
3643                         goto fail;
3644         }
3645
3646         if (recon_state.nr_caps || recon_state.nr_realms) {
3647                 struct page *page =
3648                         list_first_entry(&recon_state.pagelist->head,
3649                                         struct page, lru);
3650                 __le32 *addr = kmap_atomic(page);
3651                 if (recon_state.nr_caps) {
3652                         WARN_ON(recon_state.nr_realms != mdsc->num_snap_realms);
3653                         *addr = cpu_to_le32(recon_state.nr_caps);
3654                 } else if (recon_state.msg_version >= 4) {
3655                         *(addr + 1) = cpu_to_le32(recon_state.nr_realms);
3656                 }
3657                 kunmap_atomic(addr);
3658         }
3659
3660         reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3661         if (recon_state.msg_version >= 4)
3662                 reply->hdr.compat_version = cpu_to_le16(4);
3663
3664         reply->hdr.data_len = cpu_to_le32(recon_state.pagelist->length);
3665         ceph_msg_data_add_pagelist(reply, recon_state.pagelist);
3666
3667         ceph_con_send(&session->s_con, reply);
3668
3669         mutex_unlock(&session->s_mutex);
3670
3671         mutex_lock(&mdsc->mutex);
3672         __wake_requests(mdsc, &session->s_waiting);
3673         mutex_unlock(&mdsc->mutex);
3674
3675         up_read(&mdsc->snap_rwsem);
3676         ceph_pagelist_release(recon_state.pagelist);
3677         return;
3678
3679 fail:
3680         ceph_msg_put(reply);
3681         up_read(&mdsc->snap_rwsem);
3682         mutex_unlock(&session->s_mutex);
3683 fail_nomsg:
3684         ceph_pagelist_release(recon_state.pagelist);
3685 fail_nopagelist:
3686         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3687         return;
3688 }
3689
3690
3691 /*
3692  * compare old and new mdsmaps, kicking requests
3693  * and closing out old connections as necessary
3694  *
3695  * called under mdsc->mutex.
3696  */
3697 static void check_new_map(struct ceph_mds_client *mdsc,
3698                           struct ceph_mdsmap *newmap,
3699                           struct ceph_mdsmap *oldmap)
3700 {
3701         int i;
3702         int oldstate, newstate;
3703         struct ceph_mds_session *s;
3704
3705         dout("check_new_map new %u old %u\n",
3706              newmap->m_epoch, oldmap->m_epoch);
3707
3708         for (i = 0; i < oldmap->m_num_mds && i < mdsc->max_sessions; i++) {
3709                 if (!mdsc->sessions[i])
3710                         continue;
3711                 s = mdsc->sessions[i];
3712                 oldstate = ceph_mdsmap_get_state(oldmap, i);
3713                 newstate = ceph_mdsmap_get_state(newmap, i);
3714
3715                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3716                      i, ceph_mds_state_name(oldstate),
3717                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3718                      ceph_mds_state_name(newstate),
3719                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3720                      ceph_session_state_name(s->s_state));
3721
3722                 if (i >= newmap->m_num_mds ||
3723                     memcmp(ceph_mdsmap_get_addr(oldmap, i),
3724                            ceph_mdsmap_get_addr(newmap, i),
3725                            sizeof(struct ceph_entity_addr))) {
3726                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3727                                 /* the session never opened, just close it
3728                                  * out now */
3729                                 get_session(s);
3730                                 __unregister_session(mdsc, s);
3731                                 __wake_requests(mdsc, &s->s_waiting);
3732                                 ceph_put_mds_session(s);
3733                         } else if (i >= newmap->m_num_mds) {
3734                                 /* force close session for stopped mds */
3735                                 get_session(s);
3736                                 __unregister_session(mdsc, s);
3737                                 __wake_requests(mdsc, &s->s_waiting);
3738                                 kick_requests(mdsc, i);
3739                                 mutex_unlock(&mdsc->mutex);
3740
3741                                 mutex_lock(&s->s_mutex);
3742                                 cleanup_session_requests(mdsc, s);
3743                                 remove_session_caps(s);
3744                                 mutex_unlock(&s->s_mutex);
3745
3746                                 ceph_put_mds_session(s);
3747
3748                                 mutex_lock(&mdsc->mutex);
3749                         } else {
3750                                 /* just close it */
3751                                 mutex_unlock(&mdsc->mutex);
3752                                 mutex_lock(&s->s_mutex);
3753                                 mutex_lock(&mdsc->mutex);
3754                                 ceph_con_close(&s->s_con);
3755                                 mutex_unlock(&s->s_mutex);
3756                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3757                         }
3758                 } else if (oldstate == newstate) {
3759                         continue;  /* nothing new with this mds */
3760                 }
3761
3762                 /*
3763                  * send reconnect?
3764                  */
3765                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3766                     newstate >= CEPH_MDS_STATE_RECONNECT) {
3767                         mutex_unlock(&mdsc->mutex);
3768                         send_mds_reconnect(mdsc, s);
3769                         mutex_lock(&mdsc->mutex);
3770                 }
3771
3772                 /*
3773                  * kick request on any mds that has gone active.
3774                  */
3775                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3776                     newstate >= CEPH_MDS_STATE_ACTIVE) {
3777                         if (oldstate != CEPH_MDS_STATE_CREATING &&
3778                             oldstate != CEPH_MDS_STATE_STARTING)
3779                                 pr_info("mds%d recovery completed\n", s->s_mds);
3780                         kick_requests(mdsc, i);
3781                         ceph_kick_flushing_caps(mdsc, s);
3782                         wake_up_session_caps(s, RECONNECT);
3783                 }
3784         }
3785
3786         for (i = 0; i < newmap->m_num_mds && i < mdsc->max_sessions; i++) {
3787                 s = mdsc->sessions[i];
3788                 if (!s)
3789                         continue;
3790                 if (!ceph_mdsmap_is_laggy(newmap, i))
3791                         continue;
3792                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3793                     s->s_state == CEPH_MDS_SESSION_HUNG ||
3794                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
3795                         dout(" connecting to export targets of laggy mds%d\n",
3796                              i);
3797                         __open_export_target_sessions(mdsc, s);
3798                 }
3799         }
3800 }
3801
3802
3803
3804 /*
3805  * leases
3806  */
3807
3808 /*
3809  * caller must hold session s_mutex, dentry->d_lock
3810  */
3811 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3812 {
3813         struct ceph_dentry_info *di = ceph_dentry(dentry);
3814
3815         ceph_put_mds_session(di->lease_session);
3816         di->lease_session = NULL;
3817 }
3818
3819 static void handle_lease(struct ceph_mds_client *mdsc,
3820                          struct ceph_mds_session *session,
3821                          struct ceph_msg *msg)
3822 {
3823         struct super_block *sb = mdsc->fsc->sb;
3824         struct inode *inode;
3825         struct dentry *parent, *dentry;
3826         struct ceph_dentry_info *di;
3827         int mds = session->s_mds;
3828         struct ceph_mds_lease *h = msg->front.iov_base;
3829         u32 seq;
3830         struct ceph_vino vino;
3831         struct qstr dname;
3832         int release = 0;
3833
3834         dout("handle_lease from mds%d\n", mds);
3835
3836         /* decode */
3837         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3838                 goto bad;
3839         vino.ino = le64_to_cpu(h->ino);
3840         vino.snap = CEPH_NOSNAP;
3841         seq = le32_to_cpu(h->seq);
3842         dname.len = get_unaligned_le32(h + 1);
3843         if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
3844                 goto bad;
3845         dname.name = (void *)(h + 1) + sizeof(u32);
3846
3847         /* lookup inode */
3848         inode = ceph_find_inode(sb, vino);
3849         dout("handle_lease %s, ino %llx %p %.*s\n",
3850              ceph_lease_op_name(h->action), vino.ino, inode,
3851              dname.len, dname.name);
3852
3853         mutex_lock(&session->s_mutex);
3854         session->s_seq++;
3855
3856         if (!inode) {
3857                 dout("handle_lease no inode %llx\n", vino.ino);
3858                 goto release;
3859         }
3860
3861         /* dentry */
3862         parent = d_find_alias(inode);
3863         if (!parent) {
3864                 dout("no parent dentry on inode %p\n", inode);
3865                 WARN_ON(1);
3866                 goto release;  /* hrm... */
3867         }
3868         dname.hash = full_name_hash(parent, dname.name, dname.len);
3869         dentry = d_lookup(parent, &dname);
3870         dput(parent);
3871         if (!dentry)
3872                 goto release;
3873
3874         spin_lock(&dentry->d_lock);
3875         di = ceph_dentry(dentry);
3876         switch (h->action) {
3877         case CEPH_MDS_LEASE_REVOKE:
3878                 if (di->lease_session == session) {
3879                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3880                                 h->seq = cpu_to_le32(di->lease_seq);
3881                         __ceph_mdsc_drop_dentry_lease(dentry);
3882                 }
3883                 release = 1;
3884                 break;
3885
3886         case CEPH_MDS_LEASE_RENEW:
3887                 if (di->lease_session == session &&
3888                     di->lease_gen == session->s_cap_gen &&
3889                     di->lease_renew_from &&
3890                     di->lease_renew_after == 0) {
3891                         unsigned long duration =
3892                                 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3893
3894                         di->lease_seq = seq;
3895                         di->time = di->lease_renew_from + duration;
3896                         di->lease_renew_after = di->lease_renew_from +
3897                                 (duration >> 1);
3898                         di->lease_renew_from = 0;
3899                 }
3900                 break;
3901         }
3902         spin_unlock(&dentry->d_lock);
3903         dput(dentry);
3904
3905         if (!release)
3906                 goto out;
3907
3908 release:
3909         /* let's just reuse the same message */
3910         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3911         ceph_msg_get(msg);
3912         ceph_con_send(&session->s_con, msg);
3913
3914 out:
3915         iput(inode);
3916         mutex_unlock(&session->s_mutex);
3917         return;
3918
3919 bad:
3920         pr_err("corrupt lease message\n");
3921         ceph_msg_dump(msg);
3922 }
3923
3924 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3925                               struct inode *inode,
3926                               struct dentry *dentry, char action,
3927                               u32 seq)
3928 {
3929         struct ceph_msg *msg;
3930         struct ceph_mds_lease *lease;
3931         int len = sizeof(*lease) + sizeof(u32);
3932         int dnamelen = 0;
3933
3934         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3935              inode, dentry, ceph_lease_op_name(action), session->s_mds);
3936         dnamelen = dentry->d_name.len;
3937         len += dnamelen;
3938
3939         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3940         if (!msg)
3941                 return;
3942         lease = msg->front.iov_base;
3943         lease->action = action;
3944         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3945         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3946         lease->seq = cpu_to_le32(seq);
3947         put_unaligned_le32(dnamelen, lease + 1);
3948         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3949
3950         /*
3951          * if this is a preemptive lease RELEASE, no need to
3952          * flush request stream, since the actual request will
3953          * soon follow.
3954          */
3955         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3956
3957         ceph_con_send(&session->s_con, msg);
3958 }
3959
3960 /*
3961  * lock unlock sessions, to wait ongoing session activities
3962  */
3963 static void lock_unlock_sessions(struct ceph_mds_client *mdsc)
3964 {
3965         int i;
3966
3967         mutex_lock(&mdsc->mutex);
3968         for (i = 0; i < mdsc->max_sessions; i++) {
3969                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3970                 if (!s)
3971                         continue;
3972                 mutex_unlock(&mdsc->mutex);
3973                 mutex_lock(&s->s_mutex);
3974                 mutex_unlock(&s->s_mutex);
3975                 ceph_put_mds_session(s);
3976                 mutex_lock(&mdsc->mutex);
3977         }
3978         mutex_unlock(&mdsc->mutex);
3979 }
3980
3981
3982
3983 /*
3984  * delayed work -- periodically trim expired leases, renew caps with mds
3985  */
3986 static void schedule_delayed(struct ceph_mds_client *mdsc)
3987 {
3988         int delay = 5;
3989         unsigned hz = round_jiffies_relative(HZ * delay);
3990         schedule_delayed_work(&mdsc->delayed_work, hz);
3991 }
3992
3993 static void delayed_work(struct work_struct *work)
3994 {
3995         int i;
3996         struct ceph_mds_client *mdsc =
3997                 container_of(work, struct ceph_mds_client, delayed_work.work);
3998         int renew_interval;
3999         int renew_caps;
4000
4001         dout("mdsc delayed_work\n");
4002
4003         mutex_lock(&mdsc->mutex);
4004         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
4005         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
4006                                    mdsc->last_renew_caps);
4007         if (renew_caps)
4008                 mdsc->last_renew_caps = jiffies;
4009
4010         for (i = 0; i < mdsc->max_sessions; i++) {
4011                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4012                 if (!s)
4013                         continue;
4014                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
4015                         dout("resending session close request for mds%d\n",
4016                              s->s_mds);
4017                         request_close_session(mdsc, s);
4018                         ceph_put_mds_session(s);
4019                         continue;
4020                 }
4021                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
4022                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
4023                                 s->s_state = CEPH_MDS_SESSION_HUNG;
4024                                 pr_info("mds%d hung\n", s->s_mds);
4025                         }
4026                 }
4027                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
4028                         /* this mds is failed or recovering, just wait */
4029                         ceph_put_mds_session(s);
4030                         continue;
4031                 }
4032                 mutex_unlock(&mdsc->mutex);
4033
4034                 mutex_lock(&s->s_mutex);
4035                 if (renew_caps)
4036                         send_renew_caps(mdsc, s);
4037                 else
4038                         ceph_con_keepalive(&s->s_con);
4039                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4040                     s->s_state == CEPH_MDS_SESSION_HUNG)
4041                         ceph_send_cap_releases(mdsc, s);
4042                 mutex_unlock(&s->s_mutex);
4043                 ceph_put_mds_session(s);
4044
4045                 mutex_lock(&mdsc->mutex);
4046         }
4047         mutex_unlock(&mdsc->mutex);
4048
4049         ceph_check_delayed_caps(mdsc);
4050
4051         ceph_queue_cap_reclaim_work(mdsc);
4052
4053         ceph_trim_snapid_map(mdsc);
4054
4055         schedule_delayed(mdsc);
4056 }
4057
4058 int ceph_mdsc_init(struct ceph_fs_client *fsc)
4059
4060 {
4061         struct ceph_mds_client *mdsc;
4062
4063         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
4064         if (!mdsc)
4065                 return -ENOMEM;
4066         mdsc->fsc = fsc;
4067         mutex_init(&mdsc->mutex);
4068         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
4069         if (!mdsc->mdsmap) {
4070                 kfree(mdsc);
4071                 return -ENOMEM;
4072         }
4073
4074         fsc->mdsc = mdsc;
4075         init_completion(&mdsc->safe_umount_waiters);
4076         init_waitqueue_head(&mdsc->session_close_wq);
4077         INIT_LIST_HEAD(&mdsc->waiting_for_map);
4078         mdsc->sessions = NULL;
4079         atomic_set(&mdsc->num_sessions, 0);
4080         mdsc->max_sessions = 0;
4081         mdsc->stopping = 0;
4082         atomic64_set(&mdsc->quotarealms_count, 0);
4083         mdsc->quotarealms_inodes = RB_ROOT;
4084         mutex_init(&mdsc->quotarealms_inodes_mutex);
4085         mdsc->last_snap_seq = 0;
4086         init_rwsem(&mdsc->snap_rwsem);
4087         mdsc->snap_realms = RB_ROOT;
4088         INIT_LIST_HEAD(&mdsc->snap_empty);
4089         mdsc->num_snap_realms = 0;
4090         spin_lock_init(&mdsc->snap_empty_lock);
4091         mdsc->last_tid = 0;
4092         mdsc->oldest_tid = 0;
4093         mdsc->request_tree = RB_ROOT;
4094         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
4095         mdsc->last_renew_caps = jiffies;
4096         INIT_LIST_HEAD(&mdsc->cap_delay_list);
4097         spin_lock_init(&mdsc->cap_delay_lock);
4098         INIT_LIST_HEAD(&mdsc->snap_flush_list);
4099         spin_lock_init(&mdsc->snap_flush_lock);
4100         mdsc->last_cap_flush_tid = 1;
4101         INIT_LIST_HEAD(&mdsc->cap_flush_list);
4102         INIT_LIST_HEAD(&mdsc->cap_dirty);
4103         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
4104         mdsc->num_cap_flushing = 0;
4105         spin_lock_init(&mdsc->cap_dirty_lock);
4106         init_waitqueue_head(&mdsc->cap_flushing_wq);
4107         INIT_WORK(&mdsc->cap_reclaim_work, ceph_cap_reclaim_work);
4108         atomic_set(&mdsc->cap_reclaim_pending, 0);
4109
4110         spin_lock_init(&mdsc->dentry_list_lock);
4111         INIT_LIST_HEAD(&mdsc->dentry_leases);
4112         INIT_LIST_HEAD(&mdsc->dentry_dir_leases);
4113
4114         ceph_caps_init(mdsc);
4115         ceph_adjust_caps_max_min(mdsc, fsc->mount_options);
4116
4117         spin_lock_init(&mdsc->snapid_map_lock);
4118         mdsc->snapid_map_tree = RB_ROOT;
4119         INIT_LIST_HEAD(&mdsc->snapid_map_lru);
4120
4121         init_rwsem(&mdsc->pool_perm_rwsem);
4122         mdsc->pool_perm_tree = RB_ROOT;
4123
4124         strscpy(mdsc->nodename, utsname()->nodename,
4125                 sizeof(mdsc->nodename));
4126         return 0;
4127 }
4128
4129 /*
4130  * Wait for safe replies on open mds requests.  If we time out, drop
4131  * all requests from the tree to avoid dangling dentry refs.
4132  */
4133 static void wait_requests(struct ceph_mds_client *mdsc)
4134 {
4135         struct ceph_options *opts = mdsc->fsc->client->options;
4136         struct ceph_mds_request *req;
4137
4138         mutex_lock(&mdsc->mutex);
4139         if (__get_oldest_req(mdsc)) {
4140                 mutex_unlock(&mdsc->mutex);
4141
4142                 dout("wait_requests waiting for requests\n");
4143                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
4144                                     ceph_timeout_jiffies(opts->mount_timeout));
4145
4146                 /* tear down remaining requests */
4147                 mutex_lock(&mdsc->mutex);
4148                 while ((req = __get_oldest_req(mdsc))) {
4149                         dout("wait_requests timed out on tid %llu\n",
4150                              req->r_tid);
4151                         __unregister_request(mdsc, req);
4152                 }
4153         }
4154         mutex_unlock(&mdsc->mutex);
4155         dout("wait_requests done\n");
4156 }
4157
4158 /*
4159  * called before mount is ro, and before dentries are torn down.
4160  * (hmm, does this still race with new lookups?)
4161  */
4162 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
4163 {
4164         dout("pre_umount\n");
4165         mdsc->stopping = 1;
4166
4167         lock_unlock_sessions(mdsc);
4168         ceph_flush_dirty_caps(mdsc);
4169         wait_requests(mdsc);
4170
4171         /*
4172          * wait for reply handlers to drop their request refs and
4173          * their inode/dcache refs
4174          */
4175         ceph_msgr_flush();
4176
4177         ceph_cleanup_quotarealms_inodes(mdsc);
4178 }
4179
4180 /*
4181  * wait for all write mds requests to flush.
4182  */
4183 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
4184 {
4185         struct ceph_mds_request *req = NULL, *nextreq;
4186         struct rb_node *n;
4187
4188         mutex_lock(&mdsc->mutex);
4189         dout("wait_unsafe_requests want %lld\n", want_tid);
4190 restart:
4191         req = __get_oldest_req(mdsc);
4192         while (req && req->r_tid <= want_tid) {
4193                 /* find next request */
4194                 n = rb_next(&req->r_node);
4195                 if (n)
4196                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
4197                 else
4198                         nextreq = NULL;
4199                 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
4200                     (req->r_op & CEPH_MDS_OP_WRITE)) {
4201                         /* write op */
4202                         ceph_mdsc_get_request(req);
4203                         if (nextreq)
4204                                 ceph_mdsc_get_request(nextreq);
4205                         mutex_unlock(&mdsc->mutex);
4206                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
4207                              req->r_tid, want_tid);
4208                         wait_for_completion(&req->r_safe_completion);
4209                         mutex_lock(&mdsc->mutex);
4210                         ceph_mdsc_put_request(req);
4211                         if (!nextreq)
4212                                 break;  /* next dne before, so we're done! */
4213                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
4214                                 /* next request was removed from tree */
4215                                 ceph_mdsc_put_request(nextreq);
4216                                 goto restart;
4217                         }
4218                         ceph_mdsc_put_request(nextreq);  /* won't go away */
4219                 }
4220                 req = nextreq;
4221         }
4222         mutex_unlock(&mdsc->mutex);
4223         dout("wait_unsafe_requests done\n");
4224 }
4225
4226 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
4227 {
4228         u64 want_tid, want_flush;
4229
4230         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4231                 return;
4232
4233         dout("sync\n");
4234         mutex_lock(&mdsc->mutex);
4235         want_tid = mdsc->last_tid;
4236         mutex_unlock(&mdsc->mutex);
4237
4238         ceph_flush_dirty_caps(mdsc);
4239         spin_lock(&mdsc->cap_dirty_lock);
4240         want_flush = mdsc->last_cap_flush_tid;
4241         if (!list_empty(&mdsc->cap_flush_list)) {
4242                 struct ceph_cap_flush *cf =
4243                         list_last_entry(&mdsc->cap_flush_list,
4244                                         struct ceph_cap_flush, g_list);
4245                 cf->wake = true;
4246         }
4247         spin_unlock(&mdsc->cap_dirty_lock);
4248
4249         dout("sync want tid %lld flush_seq %lld\n",
4250              want_tid, want_flush);
4251
4252         wait_unsafe_requests(mdsc, want_tid);
4253         wait_caps_flush(mdsc, want_flush);
4254 }
4255
4256 /*
4257  * true if all sessions are closed, or we force unmount
4258  */
4259 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
4260 {
4261         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4262                 return true;
4263         return atomic_read(&mdsc->num_sessions) <= skipped;
4264 }
4265
4266 /*
4267  * called after sb is ro.
4268  */
4269 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
4270 {
4271         struct ceph_options *opts = mdsc->fsc->client->options;
4272         struct ceph_mds_session *session;
4273         int i;
4274         int skipped = 0;
4275
4276         dout("close_sessions\n");
4277
4278         /* close sessions */
4279         mutex_lock(&mdsc->mutex);
4280         for (i = 0; i < mdsc->max_sessions; i++) {
4281                 session = __ceph_lookup_mds_session(mdsc, i);
4282                 if (!session)
4283                         continue;
4284                 mutex_unlock(&mdsc->mutex);
4285                 mutex_lock(&session->s_mutex);
4286                 if (__close_session(mdsc, session) <= 0)
4287                         skipped++;
4288                 mutex_unlock(&session->s_mutex);
4289                 ceph_put_mds_session(session);
4290                 mutex_lock(&mdsc->mutex);
4291         }
4292         mutex_unlock(&mdsc->mutex);
4293
4294         dout("waiting for sessions to close\n");
4295         wait_event_timeout(mdsc->session_close_wq,
4296                            done_closing_sessions(mdsc, skipped),
4297                            ceph_timeout_jiffies(opts->mount_timeout));
4298
4299         /* tear down remaining sessions */
4300         mutex_lock(&mdsc->mutex);
4301         for (i = 0; i < mdsc->max_sessions; i++) {
4302                 if (mdsc->sessions[i]) {
4303                         session = get_session(mdsc->sessions[i]);
4304                         __unregister_session(mdsc, session);
4305                         mutex_unlock(&mdsc->mutex);
4306                         mutex_lock(&session->s_mutex);
4307                         remove_session_caps(session);
4308                         mutex_unlock(&session->s_mutex);
4309                         ceph_put_mds_session(session);
4310                         mutex_lock(&mdsc->mutex);
4311                 }
4312         }
4313         WARN_ON(!list_empty(&mdsc->cap_delay_list));
4314         mutex_unlock(&mdsc->mutex);
4315
4316         ceph_cleanup_snapid_map(mdsc);
4317         ceph_cleanup_empty_realms(mdsc);
4318
4319         cancel_work_sync(&mdsc->cap_reclaim_work);
4320         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4321
4322         dout("stopped\n");
4323 }
4324
4325 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
4326 {
4327         struct ceph_mds_session *session;
4328         int mds;
4329
4330         dout("force umount\n");
4331
4332         mutex_lock(&mdsc->mutex);
4333         for (mds = 0; mds < mdsc->max_sessions; mds++) {
4334                 session = __ceph_lookup_mds_session(mdsc, mds);
4335                 if (!session)
4336                         continue;
4337                 mutex_unlock(&mdsc->mutex);
4338                 mutex_lock(&session->s_mutex);
4339                 __close_session(mdsc, session);
4340                 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
4341                         cleanup_session_requests(mdsc, session);
4342                         remove_session_caps(session);
4343                 }
4344                 mutex_unlock(&session->s_mutex);
4345                 ceph_put_mds_session(session);
4346                 mutex_lock(&mdsc->mutex);
4347                 kick_requests(mdsc, mds);
4348         }
4349         __wake_requests(mdsc, &mdsc->waiting_for_map);
4350         mutex_unlock(&mdsc->mutex);
4351 }
4352
4353 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
4354 {
4355         dout("stop\n");
4356         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4357         if (mdsc->mdsmap)
4358                 ceph_mdsmap_destroy(mdsc->mdsmap);
4359         kfree(mdsc->sessions);
4360         ceph_caps_finalize(mdsc);
4361         ceph_pool_perm_destroy(mdsc);
4362 }
4363
4364 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
4365 {
4366         struct ceph_mds_client *mdsc = fsc->mdsc;
4367         dout("mdsc_destroy %p\n", mdsc);
4368
4369         if (!mdsc)
4370                 return;
4371
4372         /* flush out any connection work with references to us */
4373         ceph_msgr_flush();
4374
4375         ceph_mdsc_stop(mdsc);
4376
4377         fsc->mdsc = NULL;
4378         kfree(mdsc);
4379         dout("mdsc_destroy %p done\n", mdsc);
4380 }
4381
4382 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4383 {
4384         struct ceph_fs_client *fsc = mdsc->fsc;
4385         const char *mds_namespace = fsc->mount_options->mds_namespace;
4386         void *p = msg->front.iov_base;
4387         void *end = p + msg->front.iov_len;
4388         u32 epoch;
4389         u32 map_len;
4390         u32 num_fs;
4391         u32 mount_fscid = (u32)-1;
4392         u8 struct_v, struct_cv;
4393         int err = -EINVAL;
4394
4395         ceph_decode_need(&p, end, sizeof(u32), bad);
4396         epoch = ceph_decode_32(&p);
4397
4398         dout("handle_fsmap epoch %u\n", epoch);
4399
4400         ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4401         struct_v = ceph_decode_8(&p);
4402         struct_cv = ceph_decode_8(&p);
4403         map_len = ceph_decode_32(&p);
4404
4405         ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
4406         p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
4407
4408         num_fs = ceph_decode_32(&p);
4409         while (num_fs-- > 0) {
4410                 void *info_p, *info_end;
4411                 u32 info_len;
4412                 u8 info_v, info_cv;
4413                 u32 fscid, namelen;
4414
4415                 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4416                 info_v = ceph_decode_8(&p);
4417                 info_cv = ceph_decode_8(&p);
4418                 info_len = ceph_decode_32(&p);
4419                 ceph_decode_need(&p, end, info_len, bad);
4420                 info_p = p;
4421                 info_end = p + info_len;
4422                 p = info_end;
4423
4424                 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
4425                 fscid = ceph_decode_32(&info_p);
4426                 namelen = ceph_decode_32(&info_p);
4427                 ceph_decode_need(&info_p, info_end, namelen, bad);
4428
4429                 if (mds_namespace &&
4430                     strlen(mds_namespace) == namelen &&
4431                     !strncmp(mds_namespace, (char *)info_p, namelen)) {
4432                         mount_fscid = fscid;
4433                         break;
4434                 }
4435         }
4436
4437         ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
4438         if (mount_fscid != (u32)-1) {
4439                 fsc->client->monc.fs_cluster_id = mount_fscid;
4440                 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
4441                                    0, true);
4442                 ceph_monc_renew_subs(&fsc->client->monc);
4443         } else {
4444                 err = -ENOENT;
4445                 goto err_out;
4446         }
4447         return;
4448
4449 bad:
4450         pr_err("error decoding fsmap\n");
4451 err_out:
4452         mutex_lock(&mdsc->mutex);
4453         mdsc->mdsmap_err = err;
4454         __wake_requests(mdsc, &mdsc->waiting_for_map);
4455         mutex_unlock(&mdsc->mutex);
4456 }
4457
4458 /*
4459  * handle mds map update.
4460  */
4461 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4462 {
4463         u32 epoch;
4464         u32 maplen;
4465         void *p = msg->front.iov_base;
4466         void *end = p + msg->front.iov_len;
4467         struct ceph_mdsmap *newmap, *oldmap;
4468         struct ceph_fsid fsid;
4469         int err = -EINVAL;
4470
4471         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
4472         ceph_decode_copy(&p, &fsid, sizeof(fsid));
4473         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
4474                 return;
4475         epoch = ceph_decode_32(&p);
4476         maplen = ceph_decode_32(&p);
4477         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
4478
4479         /* do we need it? */
4480         mutex_lock(&mdsc->mutex);
4481         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
4482                 dout("handle_map epoch %u <= our %u\n",
4483                      epoch, mdsc->mdsmap->m_epoch);
4484                 mutex_unlock(&mdsc->mutex);
4485                 return;
4486         }
4487
4488         newmap = ceph_mdsmap_decode(&p, end);
4489         if (IS_ERR(newmap)) {
4490                 err = PTR_ERR(newmap);
4491                 goto bad_unlock;
4492         }
4493
4494         /* swap into place */
4495         if (mdsc->mdsmap) {
4496                 oldmap = mdsc->mdsmap;
4497                 mdsc->mdsmap = newmap;
4498                 check_new_map(mdsc, newmap, oldmap);
4499                 ceph_mdsmap_destroy(oldmap);
4500         } else {
4501                 mdsc->mdsmap = newmap;  /* first mds map */
4502         }
4503         mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
4504                                         MAX_LFS_FILESIZE);
4505
4506         __wake_requests(mdsc, &mdsc->waiting_for_map);
4507         ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
4508                           mdsc->mdsmap->m_epoch);
4509
4510         mutex_unlock(&mdsc->mutex);
4511         schedule_delayed(mdsc);
4512         return;
4513
4514 bad_unlock:
4515         mutex_unlock(&mdsc->mutex);
4516 bad:
4517         pr_err("error decoding mdsmap %d\n", err);
4518         return;
4519 }
4520
4521 static struct ceph_connection *con_get(struct ceph_connection *con)
4522 {
4523         struct ceph_mds_session *s = con->private;
4524
4525         if (get_session(s)) {
4526                 dout("mdsc con_get %p ok (%d)\n", s, refcount_read(&s->s_ref));
4527                 return con;
4528         }
4529         dout("mdsc con_get %p FAIL\n", s);
4530         return NULL;
4531 }
4532
4533 static void con_put(struct ceph_connection *con)
4534 {
4535         struct ceph_mds_session *s = con->private;
4536
4537         dout("mdsc con_put %p (%d)\n", s, refcount_read(&s->s_ref) - 1);
4538         ceph_put_mds_session(s);
4539 }
4540
4541 /*
4542  * if the client is unresponsive for long enough, the mds will kill
4543  * the session entirely.
4544  */
4545 static void peer_reset(struct ceph_connection *con)
4546 {
4547         struct ceph_mds_session *s = con->private;
4548         struct ceph_mds_client *mdsc = s->s_mdsc;
4549
4550         pr_warn("mds%d closed our session\n", s->s_mds);
4551         send_mds_reconnect(mdsc, s);
4552 }
4553
4554 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
4555 {
4556         struct ceph_mds_session *s = con->private;
4557         struct ceph_mds_client *mdsc = s->s_mdsc;
4558         int type = le16_to_cpu(msg->hdr.type);
4559
4560         mutex_lock(&mdsc->mutex);
4561         if (__verify_registered_session(mdsc, s) < 0) {
4562                 mutex_unlock(&mdsc->mutex);
4563                 goto out;
4564         }
4565         mutex_unlock(&mdsc->mutex);
4566
4567         switch (type) {
4568         case CEPH_MSG_MDS_MAP:
4569                 ceph_mdsc_handle_mdsmap(mdsc, msg);
4570                 break;
4571         case CEPH_MSG_FS_MAP_USER:
4572                 ceph_mdsc_handle_fsmap(mdsc, msg);
4573                 break;
4574         case CEPH_MSG_CLIENT_SESSION:
4575                 handle_session(s, msg);
4576                 break;
4577         case CEPH_MSG_CLIENT_REPLY:
4578                 handle_reply(s, msg);
4579                 break;
4580         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
4581                 handle_forward(mdsc, s, msg);
4582                 break;
4583         case CEPH_MSG_CLIENT_CAPS:
4584                 ceph_handle_caps(s, msg);
4585                 break;
4586         case CEPH_MSG_CLIENT_SNAP:
4587                 ceph_handle_snap(mdsc, s, msg);
4588                 break;
4589         case CEPH_MSG_CLIENT_LEASE:
4590                 handle_lease(mdsc, s, msg);
4591                 break;
4592         case CEPH_MSG_CLIENT_QUOTA:
4593                 ceph_handle_quota(mdsc, s, msg);
4594                 break;
4595
4596         default:
4597                 pr_err("received unknown message type %d %s\n", type,
4598                        ceph_msg_type_name(type));
4599         }
4600 out:
4601         ceph_msg_put(msg);
4602 }
4603
4604 /*
4605  * authentication
4606  */
4607
4608 /*
4609  * Note: returned pointer is the address of a structure that's
4610  * managed separately.  Caller must *not* attempt to free it.
4611  */
4612 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
4613                                         int *proto, int force_new)
4614 {
4615         struct ceph_mds_session *s = con->private;
4616         struct ceph_mds_client *mdsc = s->s_mdsc;
4617         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4618         struct ceph_auth_handshake *auth = &s->s_auth;
4619
4620         if (force_new && auth->authorizer) {
4621                 ceph_auth_destroy_authorizer(auth->authorizer);
4622                 auth->authorizer = NULL;
4623         }
4624         if (!auth->authorizer) {
4625                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4626                                                       auth);
4627                 if (ret)
4628                         return ERR_PTR(ret);
4629         } else {
4630                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4631                                                       auth);
4632                 if (ret)
4633                         return ERR_PTR(ret);
4634         }
4635         *proto = ac->protocol;
4636
4637         return auth;
4638 }
4639
4640 static int add_authorizer_challenge(struct ceph_connection *con,
4641                                     void *challenge_buf, int challenge_buf_len)
4642 {
4643         struct ceph_mds_session *s = con->private;
4644         struct ceph_mds_client *mdsc = s->s_mdsc;
4645         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4646
4647         return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
4648                                             challenge_buf, challenge_buf_len);
4649 }
4650
4651 static int verify_authorizer_reply(struct ceph_connection *con)
4652 {
4653         struct ceph_mds_session *s = con->private;
4654         struct ceph_mds_client *mdsc = s->s_mdsc;
4655         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4656
4657         return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
4658 }
4659
4660 static int invalidate_authorizer(struct ceph_connection *con)
4661 {
4662         struct ceph_mds_session *s = con->private;
4663         struct ceph_mds_client *mdsc = s->s_mdsc;
4664         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4665
4666         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
4667
4668         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
4669 }
4670
4671 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
4672                                 struct ceph_msg_header *hdr, int *skip)
4673 {
4674         struct ceph_msg *msg;
4675         int type = (int) le16_to_cpu(hdr->type);
4676         int front_len = (int) le32_to_cpu(hdr->front_len);
4677
4678         if (con->in_msg)
4679                 return con->in_msg;
4680
4681         *skip = 0;
4682         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
4683         if (!msg) {
4684                 pr_err("unable to allocate msg type %d len %d\n",
4685                        type, front_len);
4686                 return NULL;
4687         }
4688
4689         return msg;
4690 }
4691
4692 static int mds_sign_message(struct ceph_msg *msg)
4693 {
4694        struct ceph_mds_session *s = msg->con->private;
4695        struct ceph_auth_handshake *auth = &s->s_auth;
4696
4697        return ceph_auth_sign_message(auth, msg);
4698 }
4699
4700 static int mds_check_message_signature(struct ceph_msg *msg)
4701 {
4702        struct ceph_mds_session *s = msg->con->private;
4703        struct ceph_auth_handshake *auth = &s->s_auth;
4704
4705        return ceph_auth_check_message_signature(auth, msg);
4706 }
4707
4708 static const struct ceph_connection_operations mds_con_ops = {
4709         .get = con_get,
4710         .put = con_put,
4711         .dispatch = dispatch,
4712         .get_authorizer = get_authorizer,
4713         .add_authorizer_challenge = add_authorizer_challenge,
4714         .verify_authorizer_reply = verify_authorizer_reply,
4715         .invalidate_authorizer = invalidate_authorizer,
4716         .peer_reset = peer_reset,
4717         .alloc_msg = mds_alloc_msg,
4718         .sign_message = mds_sign_message,
4719         .check_message_signature = mds_check_message_signature,
4720 };
4721
4722 /* eof */