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  47  * are no longer valid.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                 /* avoid calling iput_final() in mds dispatch threads */
 694                 ceph_async_iput(req->r_inode);
 695         }
 696         if (req->r_parent)
 697                 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
 698         ceph_async_iput(req->r_target_inode);
 699         if (req->r_dentry)
 700                 dput(req->r_dentry);
 701         if (req->r_old_dentry)
 702                 dput(req->r_old_dentry);
 703         if (req->r_old_dentry_dir) {
 704                 /*
 705                  * track (and drop pins for) r_old_dentry_dir
 706                  * separately, since r_old_dentry's d_parent may have
 707                  * changed between the dir mutex being dropped and
 708                  * this request being freed.
 709                  */
 710                 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
 711                                   CEPH_CAP_PIN);
 712                 ceph_async_iput(req->r_old_dentry_dir);
 713         }
 714         kfree(req->r_path1);
 715         kfree(req->r_path2);
 716         if (req->r_pagelist)
 717                 ceph_pagelist_release(req->r_pagelist);
 718         put_request_session(req);
 719         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
 720         kfree(req);
 721 }
 722
 723 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
 724
 725 /*
 726  * lookup session, bump ref if found.
 727  *
 728  * called under mdsc->mutex.
 729  */
 730 static struct ceph_mds_request *
 731 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
 732 {
 733         struct ceph_mds_request *req;
 734
 735         req = lookup_request(&mdsc->request_tree, tid);
 736         if (req)
 737                 ceph_mdsc_get_request(req);
 738
 739         return req;
 740 }
 741
 742 /*
 743  * Register an in-flight request, and assign a tid.  Link to directory
 744  * are modifying (if any).
 745  *
 746  * Called under mdsc->mutex.
 747  */
 748 static void __register_request(struct ceph_mds_client *mdsc,
 749                                struct ceph_mds_request *req,
 750                                struct inode *dir)
 751 {
 752         int ret = 0;
 753
 754         req->r_tid = ++mdsc->last_tid;
 755         if (req->r_num_caps) {
 756                 ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
 757                                         req->r_num_caps);
 758                 if (ret < 0) {
 759                         pr_err("__register_request %p "
 760                                "failed to reserve caps: %d\n", req, ret);
 761                         /* set req->r_err to fail early from __do_request */
 762                         req->r_err = ret;
 763                         return;
 764                 }
 765         }
 766         dout("__register_request %p tid %lld\n", req, req->r_tid);
 767         ceph_mdsc_get_request(req);
 768         insert_request(&mdsc->request_tree, req);
 769
 770         req->r_uid = current_fsuid();
 771         req->r_gid = current_fsgid();
 772
 773         if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
 774                 mdsc->oldest_tid = req->r_tid;
 775
 776         if (dir) {
 777                 ihold(dir);
 778                 req->r_unsafe_dir = dir;
 779         }
 780 }
 781
 782 static void __unregister_request(struct ceph_mds_client *mdsc,
 783                                  struct ceph_mds_request *req)
 784 {
 785         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 786
 787         /* Never leave an unregistered request on an unsafe list! */
 788         list_del_init(&req->r_unsafe_item);
 789
 790         if (req->r_tid == mdsc->oldest_tid) {
 791                 struct rb_node *p = rb_next(&req->r_node);
 792                 mdsc->oldest_tid = 0;
 793                 while (p) {
 794                         struct ceph_mds_request *next_req =
 795                                 rb_entry(p, struct ceph_mds_request, r_node);
 796                         if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
 797                                 mdsc->oldest_tid = next_req->r_tid;
 798                                 break;
 799                         }
 800                         p = rb_next(p);
 801                 }
 802         }
 803
 804         erase_request(&mdsc->request_tree, req);
 805
 806         if (req->r_unsafe_dir  &&
 807             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
 808                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 809                 spin_lock(&ci->i_unsafe_lock);
 810                 list_del_init(&req->r_unsafe_dir_item);
 811                 spin_unlock(&ci->i_unsafe_lock);
 812         }
 813         if (req->r_target_inode &&
 814             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
 815                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
 816                 spin_lock(&ci->i_unsafe_lock);
 817                 list_del_init(&req->r_unsafe_target_item);
 818                 spin_unlock(&ci->i_unsafe_lock);
 819         }
 820
 821         if (req->r_unsafe_dir) {
 822                 /* avoid calling iput_final() in mds dispatch threads */
 823                 ceph_async_iput(req->r_unsafe_dir);
 824                 req->r_unsafe_dir = NULL;
 825         }
 826
 827         complete_all(&req->r_safe_completion);
 828
 829         ceph_mdsc_put_request(req);
 830 }
 831
 832 /*
 833  * Walk back up the dentry tree until we hit a dentry representing a
 834  * non-snapshot inode. We do this using the rcu_read_lock (which must be held
 835  * when calling this) to ensure that the objects won't disappear while we're
 836  * working with them. Once we hit a candidate dentry, we attempt to take a
 837  * reference to it, and return that as the result.
 838  */
 839 static struct inode *get_nonsnap_parent(struct dentry *dentry)
 840 {
 841         struct inode *inode = NULL;
 842
 843         while (dentry && !IS_ROOT(dentry)) {
 844                 inode = d_inode_rcu(dentry);
 845                 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
 846                         break;
 847                 dentry = dentry->d_parent;
 848         }
 849         if (inode)
 850                 inode = igrab(inode);
 851         return inode;
 852 }
 853
 854 /*
 855  * Choose mds to send request to next.  If there is a hint set in the
 856  * request (e.g., due to a prior forward hint from the mds), use that.
 857  * Otherwise, consult frag tree and/or caps to identify the
 858  * appropriate mds.  If all else fails, choose randomly.
 859  *
 860  * Called under mdsc->mutex.
 861  */
 862 static int __choose_mds(struct ceph_mds_client *mdsc,
 863                         struct ceph_mds_request *req)
 864 {
 865         struct inode *inode;
 866         struct ceph_inode_info *ci;
 867         struct ceph_cap *cap;
 868         int mode = req->r_direct_mode;
 869         int mds = -1;
 870         u32 hash = req->r_direct_hash;
 871         bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
 872
 873         /*
 874          * is there a specific mds we should try?  ignore hint if we have
 875          * no session and the mds is not up (active or recovering).
 876          */
 877         if (req->r_resend_mds >= 0 &&
 878             (__have_session(mdsc, req->r_resend_mds) ||
 879              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 880                 dout("choose_mds using resend_mds mds%d\n",
 881                      req->r_resend_mds);
 882                 return req->r_resend_mds;
 883         }
 884
 885         if (mode == USE_RANDOM_MDS)
 886                 goto random;
 887
 888         inode = NULL;
 889         if (req->r_inode) {
 890                 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
 891                         inode = req->r_inode;
 892                         ihold(inode);
 893                 } else {
 894                         /* req->r_dentry is non-null for LSSNAP request */
 895                         rcu_read_lock();
 896                         inode = get_nonsnap_parent(req->r_dentry);
 897                         rcu_read_unlock();
 898                         dout("__choose_mds using snapdir's parent %p\n", inode);
 899                 }
 900         } else if (req->r_dentry) {
 901                 /* ignore race with rename; old or new d_parent is okay */
 902                 struct dentry *parent;
 903                 struct inode *dir;
 904
 905                 rcu_read_lock();
 906                 parent = req->r_dentry->d_parent;
 907                 dir = req->r_parent ? : d_inode_rcu(parent);
 908
 909                 if (!dir || dir->i_sb != mdsc->fsc->sb) {
 910                         /*  not this fs or parent went negative */
 911                         inode = d_inode(req->r_dentry);
 912                         if (inode)
 913                                 ihold(inode);
 914                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
 915                         /* direct snapped/virtual snapdir requests
 916                          * based on parent dir inode */
 917                         inode = get_nonsnap_parent(parent);
 918                         dout("__choose_mds using nonsnap parent %p\n", inode);
 919                 } else {
 920                         /* dentry target */
 921                         inode = d_inode(req->r_dentry);
 922                         if (!inode || mode == USE_AUTH_MDS) {
 923                                 /* dir + name */
 924                                 inode = igrab(dir);
 925                                 hash = ceph_dentry_hash(dir, req->r_dentry);
 926                                 is_hash = true;
 927                         } else {
 928                                 ihold(inode);
 929                         }
 930                 }
 931                 rcu_read_unlock();
 932         }
 933
 934         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 935              (int)hash, mode);
 936         if (!inode)
 937                 goto random;
 938         ci = ceph_inode(inode);
 939
 940         if (is_hash && S_ISDIR(inode->i_mode)) {
 941                 struct ceph_inode_frag frag;
 942                 int found;
 943
 944                 ceph_choose_frag(ci, hash, &frag, &found);
 945                 if (found) {
 946                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 947                                 u8 r;
 948
 949                                 /* choose a random replica */
 950                                 get_random_bytes(&r, 1);
 951                                 r %= frag.ndist;
 952                                 mds = frag.dist[r];
 953                                 dout("choose_mds %p %llx.%llx "
 954                                      "frag %u mds%d (%d/%d)\n",
 955                                      inode, ceph_vinop(inode),
 956                                      frag.frag, mds,
 957                                      (int)r, frag.ndist);
 958                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 959                                     CEPH_MDS_STATE_ACTIVE)
 960                                         goto out;
 961                         }
 962
 963                         /* since this file/dir wasn't known to be
 964                          * replicated, then we want to look for the
 965                          * authoritative mds. */
 966                         mode = USE_AUTH_MDS;
 967                         if (frag.mds >= 0) {
 968                                 /* choose auth mds */
 969                                 mds = frag.mds;
 970                                 dout("choose_mds %p %llx.%llx "
 971                                      "frag %u mds%d (auth)\n",
 972                                      inode, ceph_vinop(inode), frag.frag, mds);
 973                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 974                                     CEPH_MDS_STATE_ACTIVE)
 975                                         goto out;
 976                         }
 977                 }
 978         }
 979
 980         spin_lock(&ci->i_ceph_lock);
 981         cap = NULL;
 982         if (mode == USE_AUTH_MDS)
 983                 cap = ci->i_auth_cap;
 984         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
 985                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
 986         if (!cap) {
 987                 spin_unlock(&ci->i_ceph_lock);
 988                 ceph_async_iput(inode);
 989                 goto random;
 990         }
 991         mds = cap->session->s_mds;
 992         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
 993              inode, ceph_vinop(inode), mds,
 994              cap == ci->i_auth_cap ? "auth " : "", cap);
 995         spin_unlock(&ci->i_ceph_lock);
 996 out:
 997         /* avoid calling iput_final() while holding mdsc->mutex or
 998          * in mds dispatch threads */
 999         ceph_async_iput(inode);
1000         return mds;
1001
1002 random:
1003         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
1004         dout("choose_mds chose random mds%d\n", mds);
1005         return mds;
1006 }
1007
1008
1009 /*
1010  * session messages
1011  */
1012 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
1013 {
1014         struct ceph_msg *msg;
1015         struct ceph_mds_session_head *h;
1016
1017         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
1018                            false);
1019         if (!msg) {
1020                 pr_err("create_session_msg ENOMEM creating msg\n");
1021                 return NULL;
1022         }
1023         h = msg->front.iov_base;
1024         h->op = cpu_to_le32(op);
1025         h->seq = cpu_to_le64(seq);
1026
1027         return msg;
1028 }
1029
1030 static void encode_supported_features(void **p, void *end)
1031 {
1032         static const unsigned char bits[] = CEPHFS_FEATURES_CLIENT_SUPPORTED;
1033         static const size_t count = ARRAY_SIZE(bits);
1034
1035         if (count > 0) {
1036                 size_t i;
1037                 size_t size = ((size_t)bits[count - 1] + 64) / 64 * 8;
1038
1039                 BUG_ON(*p + 4 + size > end);
1040                 ceph_encode_32(p, size);
1041                 memset(*p, 0, size);
1042                 for (i = 0; i < count; i++)
1043                         ((unsigned char*)(*p))[i / 8] |= 1 << (bits[i] % 8);
1044                 *p += size;
1045         } else {
1046                 BUG_ON(*p + 4 > end);
1047                 ceph_encode_32(p, 0);
1048         }
1049 }
1050
1051 /*
1052  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
1053  * to include additional client metadata fields.
1054  */
1055 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
1056 {
1057         struct ceph_msg *msg;
1058         struct ceph_mds_session_head *h;
1059         int i = -1;
1060         int extra_bytes = 0;
1061         int metadata_key_count = 0;
1062         struct ceph_options *opt = mdsc->fsc->client->options;
1063         struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
1064         void *p, *end;
1065
1066         const char* metadata[][2] = {
1067                 {"hostname", mdsc->nodename},
1068                 {"kernel_version", init_utsname()->release},
1069                 {"entity_id", opt->name ? : ""},
1070                 {"root", fsopt->server_path ? : "/"},
1071                 {NULL, NULL}
1072         };
1073
1074         /* Calculate serialized length of metadata */
1075         extra_bytes = 4;  /* map length */
1076         for (i = 0; metadata[i][0]; ++i) {
1077                 extra_bytes += 8 + strlen(metadata[i][0]) +
1078                         strlen(metadata[i][1]);
1079                 metadata_key_count++;
1080         }
1081         /* supported feature */
1082         extra_bytes += 4 + 8;
1083
1084         /* Allocate the message */
1085         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + extra_bytes,
1086                            GFP_NOFS, false);
1087         if (!msg) {
1088                 pr_err("create_session_msg ENOMEM creating msg\n");
1089                 return NULL;
1090         }
1091         p = msg->front.iov_base;
1092         end = p + msg->front.iov_len;
1093
1094         h = p;
1095         h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
1096         h->seq = cpu_to_le64(seq);
1097
1098         /*
1099          * Serialize client metadata into waiting buffer space, using
1100          * the format that userspace expects for map<string, string>
1101          *
1102          * ClientSession messages with metadata are v2
1103          */
1104         msg->hdr.version = cpu_to_le16(3);
1105         msg->hdr.compat_version = cpu_to_le16(1);
1106
1107         /* The write pointer, following the session_head structure */
1108         p += sizeof(*h);
1109
1110         /* Number of entries in the map */
1111         ceph_encode_32(&p, metadata_key_count);
1112
1113         /* Two length-prefixed strings for each entry in the map */
1114         for (i = 0; metadata[i][0]; ++i) {
1115                 size_t const key_len = strlen(metadata[i][0]);
1116                 size_t const val_len = strlen(metadata[i][1]);
1117
1118                 ceph_encode_32(&p, key_len);
1119                 memcpy(p, metadata[i][0], key_len);
1120                 p += key_len;
1121                 ceph_encode_32(&p, val_len);
1122                 memcpy(p, metadata[i][1], val_len);
1123                 p += val_len;
1124         }
1125
1126         encode_supported_features(&p, end);
1127         msg->front.iov_len = p - msg->front.iov_base;
1128         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1129
1130         return msg;
1131 }
1132
1133 /*
1134  * send session open request.
1135  *
1136  * called under mdsc->mutex
1137  */
1138 static int __open_session(struct ceph_mds_client *mdsc,
1139                           struct ceph_mds_session *session)
1140 {
1141         struct ceph_msg *msg;
1142         int mstate;
1143         int mds = session->s_mds;
1144
1145         /* wait for mds to go active? */
1146         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
1147         dout("open_session to mds%d (%s)\n", mds,
1148              ceph_mds_state_name(mstate));
1149         session->s_state = CEPH_MDS_SESSION_OPENING;
1150         session->s_renew_requested = jiffies;
1151
1152         /* send connect message */
1153         msg = create_session_open_msg(mdsc, session->s_seq);
1154         if (!msg)
1155                 return -ENOMEM;
1156         ceph_con_send(&session->s_con, msg);
1157         return 0;
1158 }
1159
1160 /*
1161  * open sessions for any export targets for the given mds
1162  *
1163  * called under mdsc->mutex
1164  */
1165 static struct ceph_mds_session *
1166 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
1167 {
1168         struct ceph_mds_session *session;
1169
1170         session = __ceph_lookup_mds_session(mdsc, target);
1171         if (!session) {
1172                 session = register_session(mdsc, target);
1173                 if (IS_ERR(session))
1174                         return session;
1175         }
1176         if (session->s_state == CEPH_MDS_SESSION_NEW ||
1177             session->s_state == CEPH_MDS_SESSION_CLOSING)
1178                 __open_session(mdsc, session);
1179
1180         return session;
1181 }
1182
1183 struct ceph_mds_session *
1184 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1185 {
1186         struct ceph_mds_session *session;
1187
1188         dout("open_export_target_session to mds%d\n", target);
1189
1190         mutex_lock(&mdsc->mutex);
1191         session = __open_export_target_session(mdsc, target);
1192         mutex_unlock(&mdsc->mutex);
1193
1194         return session;
1195 }
1196
1197 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1198                                           struct ceph_mds_session *session)
1199 {
1200         struct ceph_mds_info *mi;
1201         struct ceph_mds_session *ts;
1202         int i, mds = session->s_mds;
1203
1204         if (mds >= mdsc->mdsmap->m_num_mds)
1205                 return;
1206
1207         mi = &mdsc->mdsmap->m_info[mds];
1208         dout("open_export_target_sessions for mds%d (%d targets)\n",
1209              session->s_mds, mi->num_export_targets);
1210
1211         for (i = 0; i < mi->num_export_targets; i++) {
1212                 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1213                 if (!IS_ERR(ts))
1214                         ceph_put_mds_session(ts);
1215         }
1216 }
1217
1218 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1219                                            struct ceph_mds_session *session)
1220 {
1221         mutex_lock(&mdsc->mutex);
1222         __open_export_target_sessions(mdsc, session);
1223         mutex_unlock(&mdsc->mutex);
1224 }
1225
1226 /*
1227  * session caps
1228  */
1229
1230 static void detach_cap_releases(struct ceph_mds_session *session,
1231                                 struct list_head *target)
1232 {
1233         lockdep_assert_held(&session->s_cap_lock);
1234
1235         list_splice_init(&session->s_cap_releases, target);
1236         session->s_num_cap_releases = 0;
1237         dout("dispose_cap_releases mds%d\n", session->s_mds);
1238 }
1239
1240 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1241                                  struct list_head *dispose)
1242 {
1243         while (!list_empty(dispose)) {
1244                 struct ceph_cap *cap;
1245                 /* zero out the in-progress message */
1246                 cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1247                 list_del(&cap->session_caps);
1248                 ceph_put_cap(mdsc, cap);
1249         }
1250 }
1251
1252 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1253                                      struct ceph_mds_session *session)
1254 {
1255         struct ceph_mds_request *req;
1256         struct rb_node *p;
1257
1258         dout("cleanup_session_requests mds%d\n", session->s_mds);
1259         mutex_lock(&mdsc->mutex);
1260         while (!list_empty(&session->s_unsafe)) {
1261                 req = list_first_entry(&session->s_unsafe,
1262                                        struct ceph_mds_request, r_unsafe_item);
1263                 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1264                                     req->r_tid);
1265                 __unregister_request(mdsc, req);
1266         }
1267         /* zero r_attempts, so kick_requests() will re-send requests */
1268         p = rb_first(&mdsc->request_tree);
1269         while (p) {
1270                 req = rb_entry(p, struct ceph_mds_request, r_node);
1271                 p = rb_next(p);
1272                 if (req->r_session &&
1273                     req->r_session->s_mds == session->s_mds)
1274                         req->r_attempts = 0;
1275         }
1276         mutex_unlock(&mdsc->mutex);
1277 }
1278
1279 /*
1280  * Helper to safely iterate over all caps associated with a session, with
1281  * special care taken to handle a racing __ceph_remove_cap().
1282  *
1283  * Caller must hold session s_mutex.
1284  */
1285 int ceph_iterate_session_caps(struct ceph_mds_session *session,
1286                               int (*cb)(struct inode *, struct ceph_cap *,
1287                                         void *), void *arg)
1288 {
1289         struct list_head *p;
1290         struct ceph_cap *cap;
1291         struct inode *inode, *last_inode = NULL;
1292         struct ceph_cap *old_cap = NULL;
1293         int ret;
1294
1295         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1296         spin_lock(&session->s_cap_lock);
1297         p = session->s_caps.next;
1298         while (p != &session->s_caps) {
1299                 cap = list_entry(p, struct ceph_cap, session_caps);
1300                 inode = igrab(&cap->ci->vfs_inode);
1301                 if (!inode) {
1302                         p = p->next;
1303                         continue;
1304                 }
1305                 session->s_cap_iterator = cap;
1306                 spin_unlock(&session->s_cap_lock);
1307
1308                 if (last_inode) {
1309                         /* avoid calling iput_final() while holding
1310                          * s_mutex or in mds dispatch threads */
1311                         ceph_async_iput(last_inode);
1312                         last_inode = NULL;
1313                 }
1314                 if (old_cap) {
1315                         ceph_put_cap(session->s_mdsc, old_cap);
1316                         old_cap = NULL;
1317                 }
1318
1319                 ret = cb(inode, cap, arg);
1320                 last_inode = inode;
1321
1322                 spin_lock(&session->s_cap_lock);
1323                 p = p->next;
1324                 if (!cap->ci) {
1325                         dout("iterate_session_caps  finishing cap %p removal\n",
1326                              cap);
1327                         BUG_ON(cap->session != session);
1328                         cap->session = NULL;
1329                         list_del_init(&cap->session_caps);
1330                         session->s_nr_caps--;
1331                         if (cap->queue_release)
1332                                 __ceph_queue_cap_release(session, cap);
1333                         else
1334                                 old_cap = cap;  /* put_cap it w/o locks held */
1335                 }
1336                 if (ret < 0)
1337                         goto out;
1338         }
1339         ret = 0;
1340 out:
1341         session->s_cap_iterator = NULL;
1342         spin_unlock(&session->s_cap_lock);
1343
1344         ceph_async_iput(last_inode);
1345         if (old_cap)
1346                 ceph_put_cap(session->s_mdsc, old_cap);
1347
1348         return ret;
1349 }
1350
1351 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1352                                   void *arg)
1353 {
1354         struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1355         struct ceph_inode_info *ci = ceph_inode(inode);
1356         LIST_HEAD(to_remove);
1357         bool drop = false;
1358         bool invalidate = false;
1359
1360         dout("removing cap %p, ci is %p, inode is %p\n",
1361              cap, ci, &ci->vfs_inode);
1362         spin_lock(&ci->i_ceph_lock);
1363         if (cap->mds_wanted | cap->issued)
1364                 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1365         __ceph_remove_cap(cap, false);
1366         if (!ci->i_auth_cap) {
1367                 struct ceph_cap_flush *cf;
1368                 struct ceph_mds_client *mdsc = fsc->mdsc;
1369
1370                 if (ci->i_wrbuffer_ref > 0 &&
1371                     READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
1372                         invalidate = true;
1373
1374                 while (!list_empty(&ci->i_cap_flush_list)) {
1375                         cf = list_first_entry(&ci->i_cap_flush_list,
1376                                               struct ceph_cap_flush, i_list);
1377                         list_move(&cf->i_list, &to_remove);
1378                 }
1379
1380                 spin_lock(&mdsc->cap_dirty_lock);
1381
1382                 list_for_each_entry(cf, &to_remove, i_list)
1383                         list_del(&cf->g_list);
1384
1385                 if (!list_empty(&ci->i_dirty_item)) {
1386                         pr_warn_ratelimited(
1387                                 " dropping dirty %s state for %p %lld\n",
1388                                 ceph_cap_string(ci->i_dirty_caps),
1389                                 inode, ceph_ino(inode));
1390                         ci->i_dirty_caps = 0;
1391                         list_del_init(&ci->i_dirty_item);
1392                         drop = true;
1393                 }
1394                 if (!list_empty(&ci->i_flushing_item)) {
1395                         pr_warn_ratelimited(
1396                                 " dropping dirty+flushing %s state for %p %lld\n",
1397                                 ceph_cap_string(ci->i_flushing_caps),
1398                                 inode, ceph_ino(inode));
1399                         ci->i_flushing_caps = 0;
1400                         list_del_init(&ci->i_flushing_item);
1401                         mdsc->num_cap_flushing--;
1402                         drop = true;
1403                 }
1404                 spin_unlock(&mdsc->cap_dirty_lock);
1405
1406                 if (atomic_read(&ci->i_filelock_ref) > 0) {
1407                         /* make further file lock syscall return -EIO */
1408                         ci->i_ceph_flags |= CEPH_I_ERROR_FILELOCK;
1409                         pr_warn_ratelimited(" dropping file locks for %p %lld\n",
1410                                             inode, ceph_ino(inode));
1411                 }
1412
1413                 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1414                         list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1415                         ci->i_prealloc_cap_flush = NULL;
1416                 }
1417
1418                if (drop &&
1419                   ci->i_wrbuffer_ref_head == 0 &&
1420                   ci->i_wr_ref == 0 &&
1421                   ci->i_dirty_caps == 0 &&
1422                   ci->i_flushing_caps == 0) {
1423                       ceph_put_snap_context(ci->i_head_snapc);
1424                       ci->i_head_snapc = NULL;
1425                }
1426         }
1427         spin_unlock(&ci->i_ceph_lock);
1428         while (!list_empty(&to_remove)) {
1429                 struct ceph_cap_flush *cf;
1430                 cf = list_first_entry(&to_remove,
1431                                       struct ceph_cap_flush, i_list);
1432                 list_del(&cf->i_list);
1433                 ceph_free_cap_flush(cf);
1434         }
1435
1436         wake_up_all(&ci->i_cap_wq);
1437         if (invalidate)
1438                 ceph_queue_invalidate(inode);
1439         if (drop)
1440                 iput(inode);
1441         return 0;
1442 }
1443
1444 /*
1445  * caller must hold session s_mutex
1446  */
1447 static void remove_session_caps(struct ceph_mds_session *session)
1448 {
1449         struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1450         struct super_block *sb = fsc->sb;
1451         LIST_HEAD(dispose);
1452
1453         dout("remove_session_caps on %p\n", session);
1454         ceph_iterate_session_caps(session, remove_session_caps_cb, fsc);
1455
1456         wake_up_all(&fsc->mdsc->cap_flushing_wq);
1457
1458         spin_lock(&session->s_cap_lock);
1459         if (session->s_nr_caps > 0) {
1460                 struct inode *inode;
1461                 struct ceph_cap *cap, *prev = NULL;
1462                 struct ceph_vino vino;
1463                 /*
1464                  * iterate_session_caps() skips inodes that are being
1465                  * deleted, we need to wait until deletions are complete.
1466                  * __wait_on_freeing_inode() is designed for the job,
1467                  * but it is not exported, so use lookup inode function
1468                  * to access it.
1469                  */
1470                 while (!list_empty(&session->s_caps)) {
1471                         cap = list_entry(session->s_caps.next,
1472                                          struct ceph_cap, session_caps);
1473                         if (cap == prev)
1474                                 break;
1475                         prev = cap;
1476                         vino = cap->ci->i_vino;
1477                         spin_unlock(&session->s_cap_lock);
1478
1479                         inode = ceph_find_inode(sb, vino);
1480                          /* avoid calling iput_final() while holding s_mutex */
1481                         ceph_async_iput(inode);
1482
1483                         spin_lock(&session->s_cap_lock);
1484                 }
1485         }
1486
1487         // drop cap expires and unlock s_cap_lock
1488         detach_cap_releases(session, &dispose);
1489
1490         BUG_ON(session->s_nr_caps > 0);
1491         BUG_ON(!list_empty(&session->s_cap_flushing));
1492         spin_unlock(&session->s_cap_lock);
1493         dispose_cap_releases(session->s_mdsc, &dispose);
1494 }
1495
1496 enum {
1497         RECONNECT,
1498         RENEWCAPS,
1499         FORCE_RO,
1500 };
1501
1502 /*
1503  * wake up any threads waiting on this session's caps.  if the cap is
1504  * old (didn't get renewed on the client reconnect), remove it now.
1505  *
1506  * caller must hold s_mutex.
1507  */
1508 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1509                               void *arg)
1510 {
1511         struct ceph_inode_info *ci = ceph_inode(inode);
1512         unsigned long ev = (unsigned long)arg;
1513
1514         if (ev == RECONNECT) {
1515                 spin_lock(&ci->i_ceph_lock);
1516                 ci->i_wanted_max_size = 0;
1517                 ci->i_requested_max_size = 0;
1518                 spin_unlock(&ci->i_ceph_lock);
1519         } else if (ev == RENEWCAPS) {
1520                 if (cap->cap_gen < cap->session->s_cap_gen) {
1521                         /* mds did not re-issue stale cap */
1522                         spin_lock(&ci->i_ceph_lock);
1523                         cap->issued = cap->implemented = CEPH_CAP_PIN;
1524                         /* make sure mds knows what we want */
1525                         if (__ceph_caps_file_wanted(ci) & ~cap->mds_wanted)
1526                                 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1527                         spin_unlock(&ci->i_ceph_lock);
1528                 }
1529         } else if (ev == FORCE_RO) {
1530         }
1531         wake_up_all(&ci->i_cap_wq);
1532         return 0;
1533 }
1534
1535 static void wake_up_session_caps(struct ceph_mds_session *session, int ev)
1536 {
1537         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1538         ceph_iterate_session_caps(session, wake_up_session_cb,
1539                                   (void *)(unsigned long)ev);
1540 }
1541
1542 /*
1543  * Send periodic message to MDS renewing all currently held caps.  The
1544  * ack will reset the expiration for all caps from this session.
1545  *
1546  * caller holds s_mutex
1547  */
1548 static int send_renew_caps(struct ceph_mds_client *mdsc,
1549                            struct ceph_mds_session *session)
1550 {
1551         struct ceph_msg *msg;
1552         int state;
1553
1554         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1555             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1556                 pr_info("mds%d caps stale\n", session->s_mds);
1557         session->s_renew_requested = jiffies;
1558
1559         /* do not try to renew caps until a recovering mds has reconnected
1560          * with its clients. */
1561         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1562         if (state < CEPH_MDS_STATE_RECONNECT) {
1563                 dout("send_renew_caps ignoring mds%d (%s)\n",
1564                      session->s_mds, ceph_mds_state_name(state));
1565                 return 0;
1566         }
1567
1568         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1569                 ceph_mds_state_name(state));
1570         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1571                                  ++session->s_renew_seq);
1572         if (!msg)
1573                 return -ENOMEM;
1574         ceph_con_send(&session->s_con, msg);
1575         return 0;
1576 }
1577
1578 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1579                              struct ceph_mds_session *session, u64 seq)
1580 {
1581         struct ceph_msg *msg;
1582
1583         dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1584              session->s_mds, ceph_session_state_name(session->s_state), seq);
1585         msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1586         if (!msg)
1587                 return -ENOMEM;
1588         ceph_con_send(&session->s_con, msg);
1589         return 0;
1590 }
1591
1592
1593 /*
1594  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1595  *
1596  * Called under session->s_mutex
1597  */
1598 static void renewed_caps(struct ceph_mds_client *mdsc,
1599                          struct ceph_mds_session *session, int is_renew)
1600 {
1601         int was_stale;
1602         int wake = 0;
1603
1604         spin_lock(&session->s_cap_lock);
1605         was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1606
1607         session->s_cap_ttl = session->s_renew_requested +
1608                 mdsc->mdsmap->m_session_timeout*HZ;
1609
1610         if (was_stale) {
1611                 if (time_before(jiffies, session->s_cap_ttl)) {
1612                         pr_info("mds%d caps renewed\n", session->s_mds);
1613                         wake = 1;
1614                 } else {
1615                         pr_info("mds%d caps still stale\n", session->s_mds);
1616                 }
1617         }
1618         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1619              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1620              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1621         spin_unlock(&session->s_cap_lock);
1622
1623         if (wake)
1624                 wake_up_session_caps(session, RENEWCAPS);
1625 }
1626
1627 /*
1628  * send a session close request
1629  */
1630 static int request_close_session(struct ceph_mds_client *mdsc,
1631                                  struct ceph_mds_session *session)
1632 {
1633         struct ceph_msg *msg;
1634
1635         dout("request_close_session mds%d state %s seq %lld\n",
1636              session->s_mds, ceph_session_state_name(session->s_state),
1637              session->s_seq);
1638         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1639         if (!msg)
1640                 return -ENOMEM;
1641         ceph_con_send(&session->s_con, msg);
1642         return 1;
1643 }
1644
1645 /*
1646  * Called with s_mutex held.
1647  */
1648 static int __close_session(struct ceph_mds_client *mdsc,
1649                          struct ceph_mds_session *session)
1650 {
1651         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1652                 return 0;
1653         session->s_state = CEPH_MDS_SESSION_CLOSING;
1654         return request_close_session(mdsc, session);
1655 }
1656
1657 static bool drop_negative_children(struct dentry *dentry)
1658 {
1659         struct dentry *child;
1660         bool all_negative = true;
1661
1662         if (!d_is_dir(dentry))
1663                 goto out;
1664
1665         spin_lock(&dentry->d_lock);
1666         list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1667                 if (d_really_is_positive(child)) {
1668                         all_negative = false;
1669                         break;
1670                 }
1671         }
1672         spin_unlock(&dentry->d_lock);
1673
1674         if (all_negative)
1675                 shrink_dcache_parent(dentry);
1676 out:
1677         return all_negative;
1678 }
1679
1680 /*
1681  * Trim old(er) caps.
1682  *
1683  * Because we can't cache an inode without one or more caps, we do
1684  * this indirectly: if a cap is unused, we prune its aliases, at which
1685  * point the inode will hopefully get dropped to.
1686  *
1687  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1688  * memory pressure from the MDS, though, so it needn't be perfect.
1689  */
1690 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1691 {
1692         struct ceph_mds_session *session = arg;
1693         struct ceph_inode_info *ci = ceph_inode(inode);
1694         int used, wanted, oissued, mine;
1695
1696         if (session->s_trim_caps <= 0)
1697                 return -1;
1698
1699         spin_lock(&ci->i_ceph_lock);
1700         mine = cap->issued | cap->implemented;
1701         used = __ceph_caps_used(ci);
1702         wanted = __ceph_caps_file_wanted(ci);
1703         oissued = __ceph_caps_issued_other(ci, cap);
1704
1705         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1706              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1707              ceph_cap_string(used), ceph_cap_string(wanted));
1708         if (cap == ci->i_auth_cap) {
1709                 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1710                     !list_empty(&ci->i_cap_snaps))
1711                         goto out;
1712                 if ((used | wanted) & CEPH_CAP_ANY_WR)
1713                         goto out;
1714                 /* Note: it's possible that i_filelock_ref becomes non-zero
1715                  * after dropping auth caps. It doesn't hurt because reply
1716                  * of lock mds request will re-add auth caps. */
1717                 if (atomic_read(&ci->i_filelock_ref) > 0)
1718                         goto out;
1719         }
1720         /* The inode has cached pages, but it's no longer used.
1721          * we can safely drop it */
1722         if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1723             !(oissued & CEPH_CAP_FILE_CACHE)) {
1724           used = 0;
1725           oissued = 0;
1726         }
1727         if ((used | wanted) & ~oissued & mine)
1728                 goto out;   /* we need these caps */
1729
1730         if (oissued) {
1731                 /* we aren't the only cap.. just remove us */
1732                 __ceph_remove_cap(cap, true);
1733                 session->s_trim_caps--;
1734         } else {
1735                 struct dentry *dentry;
1736                 /* try dropping referring dentries */
1737                 spin_unlock(&ci->i_ceph_lock);
1738                 dentry = d_find_any_alias(inode);
1739                 if (dentry && drop_negative_children(dentry)) {
1740                         int count;
1741                         dput(dentry);
1742                         d_prune_aliases(inode);
1743                         count = atomic_read(&inode->i_count);
1744                         if (count == 1)
1745                                 session->s_trim_caps--;
1746                         dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1747                              inode, cap, count);
1748                 } else {
1749                         dput(dentry);
1750                 }
1751                 return 0;
1752         }
1753
1754 out:
1755         spin_unlock(&ci->i_ceph_lock);
1756         return 0;
1757 }
1758
1759 /*
1760  * Trim session cap count down to some max number.
1761  */
1762 int ceph_trim_caps(struct ceph_mds_client *mdsc,
1763                    struct ceph_mds_session *session,
1764                    int max_caps)
1765 {
1766         int trim_caps = session->s_nr_caps - max_caps;
1767
1768         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1769              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1770         if (trim_caps > 0) {
1771                 session->s_trim_caps = trim_caps;
1772                 ceph_iterate_session_caps(session, trim_caps_cb, session);
1773                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1774                      session->s_mds, session->s_nr_caps, max_caps,
1775                         trim_caps - session->s_trim_caps);
1776                 session->s_trim_caps = 0;
1777         }
1778
1779         ceph_flush_cap_releases(mdsc, session);
1780         return 0;
1781 }
1782
1783 static int check_caps_flush(struct ceph_mds_client *mdsc,
1784                             u64 want_flush_tid)
1785 {
1786         int ret = 1;
1787
1788         spin_lock(&mdsc->cap_dirty_lock);
1789         if (!list_empty(&mdsc->cap_flush_list)) {
1790                 struct ceph_cap_flush *cf =
1791                         list_first_entry(&mdsc->cap_flush_list,
1792                                          struct ceph_cap_flush, g_list);
1793                 if (cf->tid <= want_flush_tid) {
1794                         dout("check_caps_flush still flushing tid "
1795                              "%llu <= %llu\n", cf->tid, want_flush_tid);
1796                         ret = 0;
1797                 }
1798         }
1799         spin_unlock(&mdsc->cap_dirty_lock);
1800         return ret;
1801 }
1802
1803 /*
1804  * flush all dirty inode data to disk.
1805  *
1806  * returns true if we've flushed through want_flush_tid
1807  */
1808 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1809                             u64 want_flush_tid)
1810 {
1811         dout("check_caps_flush want %llu\n", want_flush_tid);
1812
1813         wait_event(mdsc->cap_flushing_wq,
1814                    check_caps_flush(mdsc, want_flush_tid));
1815
1816         dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1817 }
1818
1819 /*
1820  * called under s_mutex
1821  */
1822 static void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1823                                    struct ceph_mds_session *session)
1824 {
1825         struct ceph_msg *msg = NULL;
1826         struct ceph_mds_cap_release *head;
1827         struct ceph_mds_cap_item *item;
1828         struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
1829         struct ceph_cap *cap;
1830         LIST_HEAD(tmp_list);
1831         int num_cap_releases;
1832         __le32  barrier, *cap_barrier;
1833
1834         down_read(&osdc->lock);
1835         barrier = cpu_to_le32(osdc->epoch_barrier);
1836         up_read(&osdc->lock);
1837
1838         spin_lock(&session->s_cap_lock);
1839 again:
1840         list_splice_init(&session->s_cap_releases, &tmp_list);
1841         num_cap_releases = session->s_num_cap_releases;
1842         session->s_num_cap_releases = 0;
1843         spin_unlock(&session->s_cap_lock);
1844
1845         while (!list_empty(&tmp_list)) {
1846                 if (!msg) {
1847                         msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1848                                         PAGE_SIZE, GFP_NOFS, false);
1849                         if (!msg)
1850                                 goto out_err;
1851                         head = msg->front.iov_base;
1852                         head->num = cpu_to_le32(0);
1853                         msg->front.iov_len = sizeof(*head);
1854
1855                         msg->hdr.version = cpu_to_le16(2);
1856                         msg->hdr.compat_version = cpu_to_le16(1);
1857                 }
1858
1859                 cap = list_first_entry(&tmp_list, struct ceph_cap,
1860                                         session_caps);
1861                 list_del(&cap->session_caps);
1862                 num_cap_releases--;
1863
1864                 head = msg->front.iov_base;
1865                 put_unaligned_le32(get_unaligned_le32(&head->num) + 1,
1866                                    &head->num);
1867                 item = msg->front.iov_base + msg->front.iov_len;
1868                 item->ino = cpu_to_le64(cap->cap_ino);
1869                 item->cap_id = cpu_to_le64(cap->cap_id);
1870                 item->migrate_seq = cpu_to_le32(cap->mseq);
1871                 item->seq = cpu_to_le32(cap->issue_seq);
1872                 msg->front.iov_len += sizeof(*item);
1873
1874                 ceph_put_cap(mdsc, cap);
1875
1876                 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1877                         // Append cap_barrier field
1878                         cap_barrier = msg->front.iov_base + msg->front.iov_len;
1879                         *cap_barrier = barrier;
1880                         msg->front.iov_len += sizeof(*cap_barrier);
1881
1882                         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1883                         dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1884                         ceph_con_send(&session->s_con, msg);
1885                         msg = NULL;
1886                 }
1887         }
1888
1889         BUG_ON(num_cap_releases != 0);
1890
1891         spin_lock(&session->s_cap_lock);
1892         if (!list_empty(&session->s_cap_releases))
1893                 goto again;
1894         spin_unlock(&session->s_cap_lock);
1895
1896         if (msg) {
1897                 // Append cap_barrier field
1898                 cap_barrier = msg->front.iov_base + msg->front.iov_len;
1899                 *cap_barrier = barrier;
1900                 msg->front.iov_len += sizeof(*cap_barrier);
1901
1902                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1903                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1904                 ceph_con_send(&session->s_con, msg);
1905         }
1906         return;
1907 out_err:
1908         pr_err("send_cap_releases mds%d, failed to allocate message\n",
1909                 session->s_mds);
1910         spin_lock(&session->s_cap_lock);
1911         list_splice(&tmp_list, &session->s_cap_releases);
1912         session->s_num_cap_releases += num_cap_releases;
1913         spin_unlock(&session->s_cap_lock);
1914 }
1915
1916 static void ceph_cap_release_work(struct work_struct *work)
1917 {
1918         struct ceph_mds_session *session =
1919                 container_of(work, struct ceph_mds_session, s_cap_release_work);
1920
1921         mutex_lock(&session->s_mutex);
1922         if (session->s_state == CEPH_MDS_SESSION_OPEN ||
1923             session->s_state == CEPH_MDS_SESSION_HUNG)
1924                 ceph_send_cap_releases(session->s_mdsc, session);
1925         mutex_unlock(&session->s_mutex);
1926         ceph_put_mds_session(session);
1927 }
1928
1929 void ceph_flush_cap_releases(struct ceph_mds_client *mdsc,
1930                              struct ceph_mds_session *session)
1931 {
1932         if (mdsc->stopping)
1933                 return;
1934
1935         get_session(session);
1936         if (queue_work(mdsc->fsc->cap_wq,
1937                        &session->s_cap_release_work)) {
1938                 dout("cap release work queued\n");
1939         } else {
1940                 ceph_put_mds_session(session);
1941                 dout("failed to queue cap release work\n");
1942         }
1943 }
1944
1945 /*
1946  * caller holds session->s_cap_lock
1947  */
1948 void __ceph_queue_cap_release(struct ceph_mds_session *session,
1949                               struct ceph_cap *cap)
1950 {
1951         list_add_tail(&cap->session_caps, &session->s_cap_releases);
1952         session->s_num_cap_releases++;
1953
1954         if (!(session->s_num_cap_releases % CEPH_CAPS_PER_RELEASE))
1955                 ceph_flush_cap_releases(session->s_mdsc, session);
1956 }
1957
1958 static void ceph_cap_reclaim_work(struct work_struct *work)
1959 {
1960         struct ceph_mds_client *mdsc =
1961                 container_of(work, struct ceph_mds_client, cap_reclaim_work);
1962         int ret = ceph_trim_dentries(mdsc);
1963         if (ret == -EAGAIN)
1964                 ceph_queue_cap_reclaim_work(mdsc);
1965 }
1966
1967 void ceph_queue_cap_reclaim_work(struct ceph_mds_client *mdsc)
1968 {
1969         if (mdsc->stopping)
1970                 return;
1971
1972         if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_reclaim_work)) {
1973                 dout("caps reclaim work queued\n");
1974         } else {
1975                 dout("failed to queue caps release work\n");
1976         }
1977 }
1978
1979 void ceph_reclaim_caps_nr(struct ceph_mds_client *mdsc, int nr)
1980 {
1981         int val;
1982         if (!nr)
1983                 return;
1984         val = atomic_add_return(nr, &mdsc->cap_reclaim_pending);
1985         if (!(val % CEPH_CAPS_PER_RELEASE)) {
1986                 atomic_set(&mdsc->cap_reclaim_pending, 0);
1987                 ceph_queue_cap_reclaim_work(mdsc);
1988         }
1989 }
1990
1991 /*
1992  * requests
1993  */
1994
1995 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1996                                     struct inode *dir)
1997 {
1998         struct ceph_inode_info *ci = ceph_inode(dir);
1999         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
2000         struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
2001         size_t size = sizeof(struct ceph_mds_reply_dir_entry);
2002         int order, num_entries;
2003
2004         spin_lock(&ci->i_ceph_lock);
2005         num_entries = ci->i_files + ci->i_subdirs;
2006         spin_unlock(&ci->i_ceph_lock);
2007         num_entries = max(num_entries, 1);
2008         num_entries = min(num_entries, opt->max_readdir);
2009
2010         order = get_order(size * num_entries);
2011         while (order >= 0) {
2012                 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
2013                                                              __GFP_NOWARN,
2014                                                              order);
2015                 if (rinfo->dir_entries)
2016                         break;
2017                 order--;
2018         }
2019         if (!rinfo->dir_entries)
2020                 return -ENOMEM;
2021
2022         num_entries = (PAGE_SIZE << order) / size;
2023         num_entries = min(num_entries, opt->max_readdir);
2024
2025         rinfo->dir_buf_size = PAGE_SIZE << order;
2026         req->r_num_caps = num_entries + 1;
2027         req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
2028         req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
2029         return 0;
2030 }
2031
2032 /*
2033  * Create an mds request.
2034  */
2035 struct ceph_mds_request *
2036 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
2037 {
2038         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
2039         struct timespec64 ts;
2040
2041         if (!req)
2042                 return ERR_PTR(-ENOMEM);
2043
2044         mutex_init(&req->r_fill_mutex);
2045         req->r_mdsc = mdsc;
2046         req->r_started = jiffies;
2047         req->r_resend_mds = -1;
2048         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
2049         INIT_LIST_HEAD(&req->r_unsafe_target_item);
2050         req->r_fmode = -1;
2051         kref_init(&req->r_kref);
2052         RB_CLEAR_NODE(&req->r_node);
2053         INIT_LIST_HEAD(&req->r_wait);
2054         init_completion(&req->r_completion);
2055         init_completion(&req->r_safe_completion);
2056         INIT_LIST_HEAD(&req->r_unsafe_item);
2057
2058         ktime_get_coarse_real_ts64(&ts);
2059         req->r_stamp = timespec64_trunc(ts, mdsc->fsc->sb->s_time_gran);
2060
2061         req->r_op = op;
2062         req->r_direct_mode = mode;
2063         return req;
2064 }
2065
2066 /*
2067  * return oldest (lowest) request, tid in request tree, 0 if none.
2068  *
2069  * called under mdsc->mutex.
2070  */
2071 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
2072 {
2073         if (RB_EMPTY_ROOT(&mdsc->request_tree))
2074                 return NULL;
2075         return rb_entry(rb_first(&mdsc->request_tree),
2076                         struct ceph_mds_request, r_node);
2077 }
2078
2079 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
2080 {
2081         return mdsc->oldest_tid;
2082 }
2083
2084 /*
2085  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
2086  * on build_path_from_dentry in fs/cifs/dir.c.
2087  *
2088  * If @stop_on_nosnap, generate path relative to the first non-snapped
2089  * inode.
2090  *
2091  * Encode hidden .snap dirs as a double /, i.e.
2092  *   foo/.snap/bar -> foo//bar
2093  */
2094 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *pbase,
2095                            int stop_on_nosnap)
2096 {
2097         struct dentry *temp;
2098         char *path;
2099         int pos;
2100         unsigned seq;
2101         u64 base;
2102
2103         if (!dentry)
2104                 return ERR_PTR(-EINVAL);
2105
2106         path = __getname();
2107         if (!path)
2108                 return ERR_PTR(-ENOMEM);
2109 retry:
2110         pos = PATH_MAX - 1;
2111         path[pos] = '\0';
2112
2113         seq = read_seqbegin(&rename_lock);
2114         rcu_read_lock();
2115         temp = dentry;
2116         for (;;) {
2117                 struct inode *inode;
2118
2119                 spin_lock(&temp->d_lock);
2120                 inode = d_inode(temp);
2121                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
2122                         dout("build_path path+%d: %p SNAPDIR\n",
2123                              pos, temp);
2124                 } else if (stop_on_nosnap && inode &&
2125                            ceph_snap(inode) == CEPH_NOSNAP) {
2126                         spin_unlock(&temp->d_lock);
2127                         break;
2128                 } else {
2129                         pos -= temp->d_name.len;
2130                         if (pos < 0) {
2131                                 spin_unlock(&temp->d_lock);
2132                                 break;
2133                         }
2134                         memcpy(path + pos, temp->d_name.name, temp->d_name.len);
2135                 }
2136                 spin_unlock(&temp->d_lock);
2137                 temp = temp->d_parent;
2138
2139                 /* Are we at the root? */
2140                 if (IS_ROOT(temp))
2141                         break;
2142
2143                 /* Are we out of buffer? */
2144                 if (--pos < 0)
2145                         break;
2146
2147                 path[pos] = '/';
2148         }
2149         base = ceph_ino(d_inode(temp));
2150         rcu_read_unlock();
2151         if (pos < 0 || read_seqretry(&rename_lock, seq)) {
2152                 pr_err("build_path did not end path lookup where "
2153                        "expected, pos is %d\n", pos);
2154                 /* presumably this is only possible if racing with a
2155                    rename of one of the parent directories (we can not
2156                    lock the dentries above us to prevent this, but
2157                    retrying should be harmless) */
2158                 goto retry;
2159         }
2160
2161         *pbase = base;
2162         *plen = PATH_MAX - 1 - pos;
2163         dout("build_path on %p %d built %llx '%.*s'\n",
2164              dentry, d_count(dentry), base, *plen, path + pos);
2165         return path + pos;
2166 }
2167
2168 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
2169                              const char **ppath, int *ppathlen, u64 *pino,
2170                              bool *pfreepath, bool parent_locked)
2171 {
2172         char *path;
2173
2174         rcu_read_lock();
2175         if (!dir)
2176                 dir = d_inode_rcu(dentry->d_parent);
2177         if (dir && parent_locked && ceph_snap(dir) == CEPH_NOSNAP) {
2178                 *pino = ceph_ino(dir);
2179                 rcu_read_unlock();
2180                 *ppath = dentry->d_name.name;
2181                 *ppathlen = dentry->d_name.len;
2182                 return 0;
2183         }
2184         rcu_read_unlock();
2185         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2186         if (IS_ERR(path))
2187                 return PTR_ERR(path);
2188         *ppath = path;
2189         *pfreepath = true;
2190         return 0;
2191 }
2192
2193 static int build_inode_path(struct inode *inode,
2194                             const char **ppath, int *ppathlen, u64 *pino,
2195                             bool *pfreepath)
2196 {
2197         struct dentry *dentry;
2198         char *path;
2199
2200         if (ceph_snap(inode) == CEPH_NOSNAP) {
2201                 *pino = ceph_ino(inode);
2202                 *ppathlen = 0;
2203                 return 0;
2204         }
2205         dentry = d_find_alias(inode);
2206         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2207         dput(dentry);
2208         if (IS_ERR(path))
2209                 return PTR_ERR(path);
2210         *ppath = path;
2211         *pfreepath = true;
2212         return 0;
2213 }
2214
2215 /*
2216  * request arguments may be specified via an inode *, a dentry *, or
2217  * an explicit ino+path.
2218  */
2219 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
2220                                   struct inode *rdiri, const char *rpath,
2221                                   u64 rino, const char **ppath, int *pathlen,
2222                                   u64 *ino, bool *freepath, bool parent_locked)
2223 {
2224         int r = 0;
2225
2226         if (rinode) {
2227                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2228                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2229                      ceph_snap(rinode));
2230         } else if (rdentry) {
2231                 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
2232                                         freepath, parent_locked);
2233                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
2234                      *ppath);
2235         } else if (rpath || rino) {
2236                 *ino = rino;
2237                 *ppath = rpath;
2238                 *pathlen = rpath ? strlen(rpath) : 0;
2239                 dout(" path %.*s\n", *pathlen, rpath);
2240         }
2241
2242         return r;
2243 }
2244
2245 /*
2246  * called under mdsc->mutex
2247  */
2248 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
2249                                                struct ceph_mds_request *req,
2250                                                int mds, bool drop_cap_releases)
2251 {
2252         struct ceph_msg *msg;
2253         struct ceph_mds_request_head *head;
2254         const char *path1 = NULL;
2255         const char *path2 = NULL;
2256         u64 ino1 = 0, ino2 = 0;
2257         int pathlen1 = 0, pathlen2 = 0;
2258         bool freepath1 = false, freepath2 = false;
2259         int len;
2260         u16 releases;
2261         void *p, *end;
2262         int ret;
2263
2264         ret = set_request_path_attr(req->r_inode, req->r_dentry,
2265                               req->r_parent, req->r_path1, req->r_ino1.ino,
2266                               &path1, &pathlen1, &ino1, &freepath1,
2267                               test_bit(CEPH_MDS_R_PARENT_LOCKED,
2268                                         &req->r_req_flags));
2269         if (ret < 0) {
2270                 msg = ERR_PTR(ret);
2271                 goto out;
2272         }
2273
2274         /* If r_old_dentry is set, then assume that its parent is locked */
2275         ret = set_request_path_attr(NULL, req->r_old_dentry,
2276                               req->r_old_dentry_dir,
2277                               req->r_path2, req->r_ino2.ino,
2278                               &path2, &pathlen2, &ino2, &freepath2, true);
2279         if (ret < 0) {
2280                 msg = ERR_PTR(ret);
2281                 goto out_free1;
2282         }
2283
2284         len = sizeof(*head) +
2285                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2286                 sizeof(struct ceph_timespec);
2287
2288         /* calculate (max) length for cap releases */
2289         len += sizeof(struct ceph_mds_request_release) *
2290                 (!!req->r_inode_drop + !!req->r_dentry_drop +
2291                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2292         if (req->r_dentry_drop)
2293                 len += pathlen1;
2294         if (req->r_old_dentry_drop)
2295                 len += pathlen2;
2296
2297         msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
2298         if (!msg) {
2299                 msg = ERR_PTR(-ENOMEM);
2300                 goto out_free2;
2301         }
2302
2303         msg->hdr.version = cpu_to_le16(2);
2304         msg->hdr.tid = cpu_to_le64(req->r_tid);
2305
2306         head = msg->front.iov_base;
2307         p = msg->front.iov_base + sizeof(*head);
2308         end = msg->front.iov_base + msg->front.iov_len;
2309
2310         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2311         head->op = cpu_to_le32(req->r_op);
2312         head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
2313         head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
2314         head->args = req->r_args;
2315
2316         ceph_encode_filepath(&p, end, ino1, path1);
2317         ceph_encode_filepath(&p, end, ino2, path2);
2318
2319         /* make note of release offset, in case we need to replay */
2320         req->r_request_release_offset = p - msg->front.iov_base;
2321
2322         /* cap releases */
2323         releases = 0;
2324         if (req->r_inode_drop)
2325                 releases += ceph_encode_inode_release(&p,
2326                       req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2327                       mds, req->r_inode_drop, req->r_inode_unless, 0);
2328         if (req->r_dentry_drop)
2329                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
2330                                 req->r_parent, mds, req->r_dentry_drop,
2331                                 req->r_dentry_unless);
2332         if (req->r_old_dentry_drop)
2333                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2334                                 req->r_old_dentry_dir, mds,
2335                                 req->r_old_dentry_drop,
2336                                 req->r_old_dentry_unless);
2337         if (req->r_old_inode_drop)
2338                 releases += ceph_encode_inode_release(&p,
2339                       d_inode(req->r_old_dentry),
2340                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2341
2342         if (drop_cap_releases) {
2343                 releases = 0;
2344                 p = msg->front.iov_base + req->r_request_release_offset;
2345         }
2346
2347         head->num_releases = cpu_to_le16(releases);
2348
2349         /* time stamp */
2350         {
2351                 struct ceph_timespec ts;
2352                 ceph_encode_timespec64(&ts, &req->r_stamp);
2353                 ceph_encode_copy(&p, &ts, sizeof(ts));
2354         }
2355
2356         BUG_ON(p > end);
2357         msg->front.iov_len = p - msg->front.iov_base;
2358         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2359
2360         if (req->r_pagelist) {
2361                 struct ceph_pagelist *pagelist = req->r_pagelist;
2362                 ceph_msg_data_add_pagelist(msg, pagelist);
2363                 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2364         } else {
2365                 msg->hdr.data_len = 0;
2366         }
2367
2368         msg->hdr.data_off = cpu_to_le16(0);
2369
2370 out_free2:
2371         if (freepath2)
2372                 ceph_mdsc_free_path((char *)path2, pathlen2);
2373 out_free1:
2374         if (freepath1)
2375                 ceph_mdsc_free_path((char *)path1, pathlen1);
2376 out:
2377         return msg;
2378 }
2379
2380 /*
2381  * called under mdsc->mutex if error, under no mutex if
2382  * success.
2383  */
2384 static void complete_request(struct ceph_mds_client *mdsc,
2385                              struct ceph_mds_request *req)
2386 {
2387         if (req->r_callback)
2388                 req->r_callback(mdsc, req);
2389         complete_all(&req->r_completion);
2390 }
2391
2392 /*
2393  * called under mdsc->mutex
2394  */
2395 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2396                                   struct ceph_mds_request *req,
2397                                   int mds, bool drop_cap_releases)
2398 {
2399         struct ceph_mds_request_head *rhead;
2400         struct ceph_msg *msg;
2401         int flags = 0;
2402
2403         req->r_attempts++;
2404         if (req->r_inode) {
2405                 struct ceph_cap *cap =
2406                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2407
2408                 if (cap)
2409                         req->r_sent_on_mseq = cap->mseq;
2410                 else
2411                         req->r_sent_on_mseq = -1;
2412         }
2413         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2414              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2415
2416         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2417                 void *p;
2418                 /*
2419                  * Replay.  Do not regenerate message (and rebuild
2420                  * paths, etc.); just use the original message.
2421                  * Rebuilding paths will break for renames because
2422                  * d_move mangles the src name.
2423                  */
2424                 msg = req->r_request;
2425                 rhead = msg->front.iov_base;
2426
2427                 flags = le32_to_cpu(rhead->flags);
2428                 flags |= CEPH_MDS_FLAG_REPLAY;
2429                 rhead->flags = cpu_to_le32(flags);
2430
2431                 if (req->r_target_inode)
2432                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2433
2434                 rhead->num_retry = req->r_attempts - 1;
2435
2436                 /* remove cap/dentry releases from message */
2437                 rhead->num_releases = 0;
2438
2439                 /* time stamp */
2440                 p = msg->front.iov_base + req->r_request_release_offset;
2441                 {
2442                         struct ceph_timespec ts;
2443                         ceph_encode_timespec64(&ts, &req->r_stamp);
2444                         ceph_encode_copy(&p, &ts, sizeof(ts));
2445                 }
2446
2447                 msg->front.iov_len = p - msg->front.iov_base;
2448                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2449                 return 0;
2450         }
2451
2452         if (req->r_request) {
2453                 ceph_msg_put(req->r_request);
2454                 req->r_request = NULL;
2455         }
2456         msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2457         if (IS_ERR(msg)) {
2458                 req->r_err = PTR_ERR(msg);
2459                 return PTR_ERR(msg);
2460         }
2461         req->r_request = msg;
2462
2463         rhead = msg->front.iov_base;
2464         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2465         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2466                 flags |= CEPH_MDS_FLAG_REPLAY;
2467         if (req->r_parent)
2468                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2469         rhead->flags = cpu_to_le32(flags);
2470         rhead->num_fwd = req->r_num_fwd;
2471         rhead->num_retry = req->r_attempts - 1;
2472         rhead->ino = 0;
2473
2474         dout(" r_parent = %p\n", req->r_parent);
2475         return 0;
2476 }
2477
2478 /*
2479  * send request, or put it on the appropriate wait list.
2480  */
2481 static void __do_request(struct ceph_mds_client *mdsc,
2482                         struct ceph_mds_request *req)
2483 {
2484         struct ceph_mds_session *session = NULL;
2485         int mds = -1;
2486         int err = 0;
2487
2488         if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2489                 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2490                         __unregister_request(mdsc, req);
2491                 return;
2492         }
2493
2494         if (req->r_timeout &&
2495             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2496                 dout("do_request timed out\n");
2497                 err = -EIO;
2498                 goto finish;
2499         }
2500         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2501                 dout("do_request forced umount\n");
2502                 err = -EIO;
2503                 goto finish;
2504         }
2505         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2506                 if (mdsc->mdsmap_err) {
2507                         err = mdsc->mdsmap_err;
2508                         dout("do_request mdsmap err %d\n", err);
2509                         goto finish;
2510                 }
2511                 if (mdsc->mdsmap->m_epoch == 0) {
2512                         dout("do_request no mdsmap, waiting for map\n");
2513                         list_add(&req->r_wait, &mdsc->waiting_for_map);
2514                         return;
2515                 }
2516                 if (!(mdsc->fsc->mount_options->flags &
2517                       CEPH_MOUNT_OPT_MOUNTWAIT) &&
2518                     !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2519                         err = -ENOENT;
2520                         pr_info("probably no mds server is up\n");
2521                         goto finish;
2522                 }
2523         }
2524
2525         put_request_session(req);
2526
2527         mds = __choose_mds(mdsc, req);
2528         if (mds < 0 ||
2529             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2530                 dout("do_request no mds or not active, waiting for map\n");
2531                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2532                 return;
2533         }
2534
2535         /* get, open session */
2536         session = __ceph_lookup_mds_session(mdsc, mds);
2537         if (!session) {
2538                 session = register_session(mdsc, mds);
2539                 if (IS_ERR(session)) {
2540                         err = PTR_ERR(session);
2541                         goto finish;
2542                 }
2543         }
2544         req->r_session = get_session(session);
2545
2546         dout("do_request mds%d session %p state %s\n", mds, session,
2547              ceph_session_state_name(session->s_state));
2548         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2549             session->s_state != CEPH_MDS_SESSION_HUNG) {
2550                 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2551                         err = -EACCES;
2552                         goto out_session;
2553                 }
2554                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2555                     session->s_state == CEPH_MDS_SESSION_CLOSING)
2556                         __open_session(mdsc, session);
2557                 list_add(&req->r_wait, &session->s_waiting);
2558                 goto out_session;
2559         }
2560
2561         /* send request */
2562         req->r_resend_mds = -1;   /* forget any previous mds hint */
2563
2564         if (req->r_request_started == 0)   /* note request start time */
2565                 req->r_request_started = jiffies;
2566
2567         err = __prepare_send_request(mdsc, req, mds, false);
2568         if (!err) {
2569                 ceph_msg_get(req->r_request);
2570                 ceph_con_send(&session->s_con, req->r_request);
2571         }
2572
2573 out_session:
2574         ceph_put_mds_session(session);
2575 finish:
2576         if (err) {
2577                 dout("__do_request early error %d\n", err);
2578                 req->r_err = err;
2579                 complete_request(mdsc, req);
2580                 __unregister_request(mdsc, req);
2581         }
2582         return;
2583 }
2584
2585 /*
2586  * called under mdsc->mutex
2587  */
2588 static void __wake_requests(struct ceph_mds_client *mdsc,
2589                             struct list_head *head)
2590 {
2591         struct ceph_mds_request *req;
2592         LIST_HEAD(tmp_list);
2593
2594         list_splice_init(head, &tmp_list);
2595
2596         while (!list_empty(&tmp_list)) {
2597                 req = list_entry(tmp_list.next,
2598                                  struct ceph_mds_request, r_wait);
2599                 list_del_init(&req->r_wait);
2600                 dout(" wake request %p tid %llu\n", req, req->r_tid);
2601                 __do_request(mdsc, req);
2602         }
2603 }
2604
2605 /*
2606  * Wake up threads with requests pending for @mds, so that they can
2607  * resubmit their requests to a possibly different mds.
2608  */
2609 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2610 {
2611         struct ceph_mds_request *req;
2612         struct rb_node *p = rb_first(&mdsc->request_tree);
2613
2614         dout("kick_requests mds%d\n", mds);
2615         while (p) {
2616                 req = rb_entry(p, struct ceph_mds_request, r_node);
2617                 p = rb_next(p);
2618                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2619                         continue;
2620                 if (req->r_attempts > 0)
2621                         continue; /* only new requests */
2622                 if (req->r_session &&
2623                     req->r_session->s_mds == mds) {
2624                         dout(" kicking tid %llu\n", req->r_tid);
2625                         list_del_init(&req->r_wait);
2626                         __do_request(mdsc, req);
2627                 }
2628         }
2629 }
2630
2631 int ceph_mdsc_submit_request(struct ceph_mds_client *mdsc, struct inode *dir,
2632                               struct ceph_mds_request *req)
2633 {
2634         int err;
2635
2636         /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2637         if (req->r_inode)
2638                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2639         if (req->r_parent)
2640                 ceph_get_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
2641         if (req->r_old_dentry_dir)
2642                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2643                                   CEPH_CAP_PIN);
2644
2645         dout("submit_request on %p for inode %p\n", req, dir);
2646         mutex_lock(&mdsc->mutex);
2647         __register_request(mdsc, req, dir);
2648         __do_request(mdsc, req);
2649         err = req->r_err;
2650         mutex_unlock(&mdsc->mutex);
2651         return err;
2652 }
2653
2654 static int ceph_mdsc_wait_request(struct ceph_mds_client *mdsc,
2655                                   struct ceph_mds_request *req)
2656 {
2657         int err;
2658
2659         /* wait */
2660         dout("do_request waiting\n");
2661         if (!req->r_timeout && req->r_wait_for_completion) {
2662                 err = req->r_wait_for_completion(mdsc, req);
2663         } else {
2664                 long timeleft = wait_for_completion_killable_timeout(
2665                                         &req->r_completion,
2666                                         ceph_timeout_jiffies(req->r_timeout));
2667                 if (timeleft > 0)
2668                         err = 0;
2669                 else if (!timeleft)
2670                         err = -EIO;  /* timed out */
2671                 else
2672                         err = timeleft;  /* killed */
2673         }
2674         dout("do_request waited, got %d\n", err);
2675         mutex_lock(&mdsc->mutex);
2676
2677         /* only abort if we didn't race with a real reply */
2678         if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2679                 err = le32_to_cpu(req->r_reply_info.head->result);
2680         } else if (err < 0) {
2681                 dout("aborted request %lld with %d\n", req->r_tid, err);
2682
2683                 /*
2684                  * ensure we aren't running concurrently with
2685                  * ceph_fill_trace or ceph_readdir_prepopulate, which
2686                  * rely on locks (dir mutex) held by our caller.
2687                  */
2688                 mutex_lock(&req->r_fill_mutex);
2689                 req->r_err = err;
2690                 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2691                 mutex_unlock(&req->r_fill_mutex);
2692
2693                 if (req->r_parent &&
2694                     (req->r_op & CEPH_MDS_OP_WRITE))
2695                         ceph_invalidate_dir_request(req);
2696         } else {
2697                 err = req->r_err;
2698         }
2699
2700         mutex_unlock(&mdsc->mutex);
2701         return err;
2702 }
2703
2704 /*
2705  * Synchrously perform an mds request.  Take care of all of the
2706  * session setup, forwarding, retry details.
2707  */
2708 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2709                          struct inode *dir,
2710                          struct ceph_mds_request *req)
2711 {
2712         int err;
2713
2714         dout("do_request on %p\n", req);
2715
2716         /* issue */
2717         err = ceph_mdsc_submit_request(mdsc, dir, req);
2718         if (!err)
2719                 err = ceph_mdsc_wait_request(mdsc, req);
2720         dout("do_request %p done, result %d\n", req, err);
2721         return err;
2722 }
2723
2724 /*
2725  * Invalidate dir's completeness, dentry lease state on an aborted MDS
2726  * namespace request.
2727  */
2728 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2729 {
2730         struct inode *dir = req->r_parent;
2731         struct inode *old_dir = req->r_old_dentry_dir;
2732
2733         dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
2734
2735         ceph_dir_clear_complete(dir);
2736         if (old_dir)
2737                 ceph_dir_clear_complete(old_dir);
2738         if (req->r_dentry)
2739                 ceph_invalidate_dentry_lease(req->r_dentry);
2740         if (req->r_old_dentry)
2741                 ceph_invalidate_dentry_lease(req->r_old_dentry);
2742 }
2743
2744 /*
2745  * Handle mds reply.
2746  *
2747  * We take the session mutex and parse and process the reply immediately.
2748  * This preserves the logical ordering of replies, capabilities, etc., sent
2749  * by the MDS as they are applied to our local cache.
2750  */
2751 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2752 {
2753         struct ceph_mds_client *mdsc = session->s_mdsc;
2754         struct ceph_mds_request *req;
2755         struct ceph_mds_reply_head *head = msg->front.iov_base;
2756         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2757         struct ceph_snap_realm *realm;
2758         u64 tid;
2759         int err, result;
2760         int mds = session->s_mds;
2761
2762         if (msg->front.iov_len < sizeof(*head)) {
2763                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2764                 ceph_msg_dump(msg);
2765                 return;
2766         }
2767
2768         /* get request, session */
2769         tid = le64_to_cpu(msg->hdr.tid);
2770         mutex_lock(&mdsc->mutex);
2771         req = lookup_get_request(mdsc, tid);
2772         if (!req) {
2773                 dout("handle_reply on unknown tid %llu\n", tid);
2774                 mutex_unlock(&mdsc->mutex);
2775                 return;
2776         }
2777         dout("handle_reply %p\n", req);
2778
2779         /* correct session? */
2780         if (req->r_session != session) {
2781                 pr_err("mdsc_handle_reply got %llu on session mds%d"
2782                        " not mds%d\n", tid, session->s_mds,
2783                        req->r_session ? req->r_session->s_mds : -1);
2784                 mutex_unlock(&mdsc->mutex);
2785                 goto out;
2786         }
2787
2788         /* dup? */
2789         if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
2790             (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
2791                 pr_warn("got a dup %s reply on %llu from mds%d\n",
2792                            head->safe ? "safe" : "unsafe", tid, mds);
2793                 mutex_unlock(&mdsc->mutex);
2794                 goto out;
2795         }
2796         if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
2797                 pr_warn("got unsafe after safe on %llu from mds%d\n",
2798                            tid, mds);
2799                 mutex_unlock(&mdsc->mutex);
2800                 goto out;
2801         }
2802
2803         result = le32_to_cpu(head->result);
2804
2805         /*
2806          * Handle an ESTALE
2807          * if we're not talking to the authority, send to them
2808          * if the authority has changed while we weren't looking,
2809          * send to new authority
2810          * Otherwise we just have to return an ESTALE
2811          */
2812         if (result == -ESTALE) {
2813                 dout("got ESTALE on request %llu\n", req->r_tid);
2814                 req->r_resend_mds = -1;
2815                 if (req->r_direct_mode != USE_AUTH_MDS) {
2816                         dout("not using auth, setting for that now\n");
2817                         req->r_direct_mode = USE_AUTH_MDS;
2818                         __do_request(mdsc, req);
2819                         mutex_unlock(&mdsc->mutex);
2820                         goto out;
2821                 } else  {
2822                         int mds = __choose_mds(mdsc, req);
2823                         if (mds >= 0 && mds != req->r_session->s_mds) {
2824                                 dout("but auth changed, so resending\n");
2825                                 __do_request(mdsc, req);
2826                                 mutex_unlock(&mdsc->mutex);
2827                                 goto out;
2828                         }
2829                 }
2830                 dout("have to return ESTALE on request %llu\n", req->r_tid);
2831         }
2832
2833
2834         if (head->safe) {
2835                 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
2836                 __unregister_request(mdsc, req);
2837
2838                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2839                         /*
2840                          * We already handled the unsafe response, now do the
2841                          * cleanup.  No need to examine the response; the MDS
2842                          * doesn't include any result info in the safe
2843                          * response.  And even if it did, there is nothing
2844                          * useful we could do with a revised return value.
2845                          */
2846                         dout("got safe reply %llu, mds%d\n", tid, mds);
2847
2848                         /* last unsafe request during umount? */
2849                         if (mdsc->stopping && !__get_oldest_req(mdsc))
2850                                 complete_all(&mdsc->safe_umount_waiters);
2851                         mutex_unlock(&mdsc->mutex);
2852                         goto out;
2853                 }
2854         } else {
2855                 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
2856                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2857                 if (req->r_unsafe_dir) {
2858                         struct ceph_inode_info *ci =
2859                                         ceph_inode(req->r_unsafe_dir);
2860                         spin_lock(&ci->i_unsafe_lock);
2861                         list_add_tail(&req->r_unsafe_dir_item,
2862                                       &ci->i_unsafe_dirops);
2863                         spin_unlock(&ci->i_unsafe_lock);
2864                 }
2865         }
2866
2867         dout("handle_reply tid %lld result %d\n", tid, result);
2868         rinfo = &req->r_reply_info;
2869         if (test_bit(CEPHFS_FEATURE_REPLY_ENCODING, &session->s_features))
2870                 err = parse_reply_info(msg, rinfo, (u64)-1);
2871         else
2872                 err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2873         mutex_unlock(&mdsc->mutex);
2874
2875         mutex_lock(&session->s_mutex);
2876         if (err < 0) {
2877                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2878                 ceph_msg_dump(msg);
2879                 goto out_err;
2880         }
2881
2882         /* snap trace */
2883         realm = NULL;
2884         if (rinfo->snapblob_len) {
2885                 down_write(&mdsc->snap_rwsem);
2886                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2887                                 rinfo->snapblob + rinfo->snapblob_len,
2888                                 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2889                                 &realm);
2890                 downgrade_write(&mdsc->snap_rwsem);
2891         } else {
2892                 down_read(&mdsc->snap_rwsem);
2893         }
2894
2895         /* insert trace into our cache */
2896         mutex_lock(&req->r_fill_mutex);
2897         current->journal_info = req;
2898         err = ceph_fill_trace(mdsc->fsc->sb, req);
2899         if (err == 0) {
2900                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2901                                     req->r_op == CEPH_MDS_OP_LSSNAP))
2902                         ceph_readdir_prepopulate(req, req->r_session);
2903         }
2904         current->journal_info = NULL;
2905         mutex_unlock(&req->r_fill_mutex);
2906
2907         up_read(&mdsc->snap_rwsem);
2908         if (realm)
2909                 ceph_put_snap_realm(mdsc, realm);
2910
2911         if (err == 0) {
2912                 if (req->r_target_inode &&
2913                     test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2914                         struct ceph_inode_info *ci =
2915                                 ceph_inode(req->r_target_inode);
2916                         spin_lock(&ci->i_unsafe_lock);
2917                         list_add_tail(&req->r_unsafe_target_item,
2918                                       &ci->i_unsafe_iops);
2919                         spin_unlock(&ci->i_unsafe_lock);
2920                 }
2921
2922                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2923         }
2924 out_err:
2925         mutex_lock(&mdsc->mutex);
2926         if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2927                 if (err) {
2928                         req->r_err = err;
2929                 } else {
2930                         req->r_reply =  ceph_msg_get(msg);
2931                         set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
2932                 }
2933         } else {
2934                 dout("reply arrived after request %lld was aborted\n", tid);
2935         }
2936         mutex_unlock(&mdsc->mutex);
2937
2938         mutex_unlock(&session->s_mutex);
2939
2940         /* kick calling process */
2941         complete_request(mdsc, req);
2942 out:
2943         ceph_mdsc_put_request(req);
2944         return;
2945 }
2946
2947
2948
2949 /*
2950  * handle mds notification that our request has been forwarded.
2951  */
2952 static void handle_forward(struct ceph_mds_client *mdsc,
2953                            struct ceph_mds_session *session,
2954                            struct ceph_msg *msg)
2955 {
2956         struct ceph_mds_request *req;
2957         u64 tid = le64_to_cpu(msg->hdr.tid);
2958         u32 next_mds;
2959         u32 fwd_seq;
2960         int err = -EINVAL;
2961         void *p = msg->front.iov_base;
2962         void *end = p + msg->front.iov_len;
2963
2964         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2965         next_mds = ceph_decode_32(&p);
2966         fwd_seq = ceph_decode_32(&p);
2967
2968         mutex_lock(&mdsc->mutex);
2969         req = lookup_get_request(mdsc, tid);
2970         if (!req) {
2971                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2972                 goto out;  /* dup reply? */
2973         }
2974
2975         if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2976                 dout("forward tid %llu aborted, unregistering\n", tid);
2977                 __unregister_request(mdsc, req);
2978         } else if (fwd_seq <= req->r_num_fwd) {
2979                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2980                      tid, next_mds, req->r_num_fwd, fwd_seq);
2981         } else {
2982                 /* resend. forward race not possible; mds would drop */
2983                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2984                 BUG_ON(req->r_err);
2985                 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
2986                 req->r_attempts = 0;
2987                 req->r_num_fwd = fwd_seq;
2988                 req->r_resend_mds = next_mds;
2989                 put_request_session(req);
2990                 __do_request(mdsc, req);
2991         }
2992         ceph_mdsc_put_request(req);
2993 out:
2994         mutex_unlock(&mdsc->mutex);
2995         return;
2996
2997 bad:
2998         pr_err("mdsc_handle_forward decode error err=%d\n", err);
2999 }
3000
3001 static int __decode_and_drop_session_metadata(void **p, void *end)
3002 {
3003         /* map<string,string> */
3004         u32 n;
3005         ceph_decode_32_safe(p, end, n, bad);
3006         while (n-- > 0) {
3007                 u32 len;
3008                 ceph_decode_32_safe(p, end, len, bad);
3009                 ceph_decode_need(p, end, len, bad);
3010                 *p += len;
3011                 ceph_decode_32_safe(p, end, len, bad);
3012                 ceph_decode_need(p, end, len, bad);
3013                 *p += len;
3014         }
3015         return 0;
3016 bad:
3017         return -1;
3018 }
3019
3020 /*
3021  * handle a mds session control message
3022  */
3023 static void handle_session(struct ceph_mds_session *session,
3024                            struct ceph_msg *msg)
3025 {
3026         struct ceph_mds_client *mdsc = session->s_mdsc;
3027         int mds = session->s_mds;
3028         int msg_version = le16_to_cpu(msg->hdr.version);
3029         void *p = msg->front.iov_base;
3030         void *end = p + msg->front.iov_len;
3031         struct ceph_mds_session_head *h;
3032         u32 op;
3033         u64 seq;
3034         unsigned long features = 0;
3035         int wake = 0;
3036
3037         /* decode */
3038         ceph_decode_need(&p, end, sizeof(*h), bad);
3039         h = p;
3040         p += sizeof(*h);
3041
3042         op = le32_to_cpu(h->op);
3043         seq = le64_to_cpu(h->seq);
3044
3045         if (msg_version >= 3) {
3046                 u32 len;
3047                 /* version >= 2, metadata */
3048                 if (__decode_and_drop_session_metadata(&p, end) < 0)
3049                         goto bad;
3050                 /* version >= 3, feature bits */
3051                 ceph_decode_32_safe(&p, end, len, bad);
3052                 ceph_decode_need(&p, end, len, bad);
3053                 memcpy(&features, p, min_t(size_t, len, sizeof(features)));
3054                 p += len;
3055         }
3056
3057         mutex_lock(&mdsc->mutex);
3058         if (op == CEPH_SESSION_CLOSE) {
3059                 get_session(session);
3060                 __unregister_session(mdsc, session);
3061         }
3062         /* FIXME: this ttl calculation is generous */
3063         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
3064         mutex_unlock(&mdsc->mutex);
3065
3066         mutex_lock(&session->s_mutex);
3067
3068         dout("handle_session mds%d %s %p state %s seq %llu\n",
3069              mds, ceph_session_op_name(op), session,
3070              ceph_session_state_name(session->s_state), seq);
3071
3072         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
3073                 session->s_state = CEPH_MDS_SESSION_OPEN;
3074                 pr_info("mds%d came back\n", session->s_mds);
3075         }
3076
3077         switch (op) {
3078         case CEPH_SESSION_OPEN:
3079                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3080                         pr_info("mds%d reconnect success\n", session->s_mds);
3081                 session->s_state = CEPH_MDS_SESSION_OPEN;
3082                 session->s_features = features;
3083                 renewed_caps(mdsc, session, 0);
3084                 wake = 1;
3085                 if (mdsc->stopping)
3086                         __close_session(mdsc, session);
3087                 break;
3088
3089         case CEPH_SESSION_RENEWCAPS:
3090                 if (session->s_renew_seq == seq)
3091                         renewed_caps(mdsc, session, 1);
3092                 break;
3093
3094         case CEPH_SESSION_CLOSE:
3095                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3096                         pr_info("mds%d reconnect denied\n", session->s_mds);
3097                 cleanup_session_requests(mdsc, session);
3098                 remove_session_caps(session);
3099                 wake = 2; /* for good measure */
3100                 wake_up_all(&mdsc->session_close_wq);
3101                 break;
3102
3103         case CEPH_SESSION_STALE:
3104                 pr_info("mds%d caps went stale, renewing\n",
3105                         session->s_mds);
3106                 spin_lock(&session->s_gen_ttl_lock);
3107                 session->s_cap_gen++;
3108                 session->s_cap_ttl = jiffies - 1;
3109                 spin_unlock(&session->s_gen_ttl_lock);
3110                 send_renew_caps(mdsc, session);
3111                 break;
3112
3113         case CEPH_SESSION_RECALL_STATE:
3114                 ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
3115                 break;
3116
3117         case CEPH_SESSION_FLUSHMSG:
3118                 send_flushmsg_ack(mdsc, session, seq);
3119                 break;
3120
3121         case CEPH_SESSION_FORCE_RO:
3122                 dout("force_session_readonly %p\n", session);
3123                 spin_lock(&session->s_cap_lock);
3124                 session->s_readonly = true;
3125                 spin_unlock(&session->s_cap_lock);
3126                 wake_up_session_caps(session, FORCE_RO);
3127                 break;
3128
3129         case CEPH_SESSION_REJECT:
3130                 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
3131                 pr_info("mds%d rejected session\n", session->s_mds);
3132                 session->s_state = CEPH_MDS_SESSION_REJECTED;
3133                 cleanup_session_requests(mdsc, session);
3134                 remove_session_caps(session);
3135                 wake = 2; /* for good measure */
3136                 break;
3137
3138         default:
3139                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
3140                 WARN_ON(1);
3141         }
3142
3143         mutex_unlock(&session->s_mutex);
3144         if (wake) {
3145                 mutex_lock(&mdsc->mutex);
3146                 __wake_requests(mdsc, &session->s_waiting);
3147                 if (wake == 2)
3148                         kick_requests(mdsc, mds);
3149                 mutex_unlock(&mdsc->mutex);
3150         }
3151         if (op == CEPH_SESSION_CLOSE)
3152                 ceph_put_mds_session(session);
3153         return;
3154
3155 bad:
3156         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
3157                (int)msg->front.iov_len);
3158         ceph_msg_dump(msg);
3159         return;
3160 }
3161
3162
3163 /*
3164  * called under session->mutex.
3165  */
3166 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
3167                                    struct ceph_mds_session *session)
3168 {
3169         struct ceph_mds_request *req, *nreq;
3170         struct rb_node *p;
3171         int err;
3172
3173         dout("replay_unsafe_requests mds%d\n", session->s_mds);
3174
3175         mutex_lock(&mdsc->mutex);
3176         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
3177                 err = __prepare_send_request(mdsc, req, session->s_mds, true);
3178                 if (!err) {
3179                         ceph_msg_get(req->r_request);
3180                         ceph_con_send(&session->s_con, req->r_request);
3181                 }
3182         }
3183
3184         /*
3185          * also re-send old requests when MDS enters reconnect stage. So that MDS
3186          * can process completed request in clientreplay stage.
3187          */
3188         p = rb_first(&mdsc->request_tree);
3189         while (p) {
3190                 req = rb_entry(p, struct ceph_mds_request, r_node);
3191                 p = rb_next(p);
3192                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3193                         continue;
3194                 if (req->r_attempts == 0)
3195                         continue; /* only old requests */
3196                 if (req->r_session &&
3197                     req->r_session->s_mds == session->s_mds) {
3198                         err = __prepare_send_request(mdsc, req,
3199                                                      session->s_mds, true);
3200                         if (!err) {
3201                                 ceph_msg_get(req->r_request);
3202                                 ceph_con_send(&session->s_con, req->r_request);
3203                         }
3204                 }
3205         }
3206         mutex_unlock(&mdsc->mutex);
3207 }
3208
3209 static int send_reconnect_partial(struct ceph_reconnect_state *recon_state)
3210 {
3211         struct ceph_msg *reply;
3212         struct ceph_pagelist *_pagelist;
3213         struct page *page;
3214         __le32 *addr;
3215         int err = -ENOMEM;
3216
3217         if (!recon_state->allow_multi)
3218                 return -ENOSPC;
3219
3220         /* can't handle message that contains both caps and realm */
3221         BUG_ON(!recon_state->nr_caps == !recon_state->nr_realms);
3222
3223         /* pre-allocate new pagelist */
3224         _pagelist = ceph_pagelist_alloc(GFP_NOFS);
3225         if (!_pagelist)
3226                 return -ENOMEM;
3227
3228         reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3229         if (!reply)
3230                 goto fail_msg;
3231
3232         /* placeholder for nr_caps */
3233         err = ceph_pagelist_encode_32(_pagelist, 0);
3234         if (err < 0)
3235                 goto fail;
3236
3237         if (recon_state->nr_caps) {
3238                 /* currently encoding caps */
3239                 err = ceph_pagelist_encode_32(recon_state->pagelist, 0);
3240                 if (err)
3241                         goto fail;
3242         } else {
3243                 /* placeholder for nr_realms (currently encoding relams) */
3244                 err = ceph_pagelist_encode_32(_pagelist, 0);
3245                 if (err < 0)
3246                         goto fail;
3247         }
3248
3249         err = ceph_pagelist_encode_8(recon_state->pagelist, 1);
3250         if (err)
3251                 goto fail;
3252
3253         page = list_first_entry(&recon_state->pagelist->head, struct page, lru);
3254         addr = kmap_atomic(page);
3255         if (recon_state->nr_caps) {
3256                 /* currently encoding caps */
3257                 *addr = cpu_to_le32(recon_state->nr_caps);
3258         } else {
3259                 /* currently encoding relams */
3260                 *(addr + 1) = cpu_to_le32(recon_state->nr_realms);
3261         }
3262         kunmap_atomic(addr);
3263
3264         reply->hdr.version = cpu_to_le16(5);
3265         reply->hdr.compat_version = cpu_to_le16(4);
3266
3267         reply->hdr.data_len = cpu_to_le32(recon_state->pagelist->length);
3268         ceph_msg_data_add_pagelist(reply, recon_state->pagelist);
3269
3270         ceph_con_send(&recon_state->session->s_con, reply);
3271         ceph_pagelist_release(recon_state->pagelist);
3272
3273         recon_state->pagelist = _pagelist;
3274         recon_state->nr_caps = 0;
3275         recon_state->nr_realms = 0;
3276         recon_state->msg_version = 5;
3277         return 0;
3278 fail:
3279         ceph_msg_put(reply);
3280 fail_msg:
3281         ceph_pagelist_release(_pagelist);
3282         return err;
3283 }
3284
3285 /*
3286  * Encode information about a cap for a reconnect with the MDS.
3287  */
3288 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
3289                           void *arg)
3290 {
3291         union {
3292                 struct ceph_mds_cap_reconnect v2;
3293                 struct ceph_mds_cap_reconnect_v1 v1;
3294         } rec;
3295         struct ceph_inode_info *ci = cap->ci;
3296         struct ceph_reconnect_state *recon_state = arg;
3297         struct ceph_pagelist *pagelist = recon_state->pagelist;
3298         int err;
3299         u64 snap_follows;
3300
3301         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
3302              inode, ceph_vinop(inode), cap, cap->cap_id,
3303              ceph_cap_string(cap->issued));
3304
3305         spin_lock(&ci->i_ceph_lock);
3306         cap->seq = 0;        /* reset cap seq */
3307         cap->issue_seq = 0;  /* and issue_seq */
3308         cap->mseq = 0;       /* and migrate_seq */
3309         cap->cap_gen = cap->session->s_cap_gen;
3310
3311         if (recon_state->msg_version >= 2) {
3312                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
3313                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3314                 rec.v2.issued = cpu_to_le32(cap->issued);
3315                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3316                 rec.v2.pathbase = 0;
3317                 rec.v2.flock_len = (__force __le32)
3318                         ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
3319         } else {
3320                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
3321                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3322                 rec.v1.issued = cpu_to_le32(cap->issued);
3323                 rec.v1.size = cpu_to_le64(inode->i_size);
3324                 ceph_encode_timespec64(&rec.v1.mtime, &inode->i_mtime);
3325                 ceph_encode_timespec64(&rec.v1.atime, &inode->i_atime);
3326                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3327                 rec.v1.pathbase = 0;
3328         }
3329
3330         if (list_empty(&ci->i_cap_snaps)) {
3331                 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
3332         } else {
3333                 struct ceph_cap_snap *capsnap =
3334                         list_first_entry(&ci->i_cap_snaps,
3335                                          struct ceph_cap_snap, ci_item);
3336                 snap_follows = capsnap->follows;
3337         }
3338         spin_unlock(&ci->i_ceph_lock);
3339
3340         if (recon_state->msg_version >= 2) {
3341                 int num_fcntl_locks, num_flock_locks;
3342                 struct ceph_filelock *flocks = NULL;
3343                 size_t struct_len, total_len = sizeof(u64);
3344                 u8 struct_v = 0;
3345
3346 encode_again:
3347                 if (rec.v2.flock_len) {
3348                         ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
3349                 } else {
3350                         num_fcntl_locks = 0;
3351                         num_flock_locks = 0;
3352                 }
3353                 if (num_fcntl_locks + num_flock_locks > 0) {
3354                         flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
3355                                                sizeof(struct ceph_filelock),
3356                                                GFP_NOFS);
3357                         if (!flocks) {
3358                                 err = -ENOMEM;
3359                                 goto out_err;
3360                         }
3361                         err = ceph_encode_locks_to_buffer(inode, flocks,
3362                                                           num_fcntl_locks,
3363                                                           num_flock_locks);
3364                         if (err) {
3365                                 kfree(flocks);
3366                                 flocks = NULL;
3367                                 if (err == -ENOSPC)
3368                                         goto encode_again;
3369                                 goto out_err;
3370                         }
3371                 } else {
3372                         kfree(flocks);
3373                         flocks = NULL;
3374                 }
3375
3376                 if (recon_state->msg_version >= 3) {
3377                         /* version, compat_version and struct_len */
3378                         total_len += 2 * sizeof(u8) + sizeof(u32);
3379                         struct_v = 2;
3380                 }
3381                 /*
3382                  * number of encoded locks is stable, so copy to pagelist
3383                  */
3384                 struct_len = 2 * sizeof(u32) +
3385                             (num_fcntl_locks + num_flock_locks) *
3386                             sizeof(struct ceph_filelock);
3387                 rec.v2.flock_len = cpu_to_le32(struct_len);
3388
3389                 struct_len += sizeof(u32) + sizeof(rec.v2);
3390
3391                 if (struct_v >= 2)
3392                         struct_len += sizeof(u64); /* snap_follows */
3393
3394                 total_len += struct_len;
3395
3396                 if (pagelist->length + total_len > RECONNECT_MAX_SIZE) {
3397                         err = send_reconnect_partial(recon_state);
3398                         if (err)
3399                                 goto out_freeflocks;
3400                         pagelist = recon_state->pagelist;
3401                 }
3402
3403                 err = ceph_pagelist_reserve(pagelist, total_len);
3404                 if (err)
3405                         goto out_freeflocks;
3406
3407                 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3408                 if (recon_state->msg_version >= 3) {
3409                         ceph_pagelist_encode_8(pagelist, struct_v);
3410                         ceph_pagelist_encode_8(pagelist, 1);
3411                         ceph_pagelist_encode_32(pagelist, struct_len);
3412                 }
3413                 ceph_pagelist_encode_string(pagelist, NULL, 0);
3414                 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
3415                 ceph_locks_to_pagelist(flocks, pagelist,
3416                                        num_fcntl_locks, num_flock_locks);
3417                 if (struct_v >= 2)
3418                         ceph_pagelist_encode_64(pagelist, snap_follows);
3419 out_freeflocks:
3420                 kfree(flocks);
3421         } else {
3422                 u64 pathbase = 0;
3423                 int pathlen = 0;
3424                 char *path = NULL;
3425                 struct dentry *dentry;
3426
3427                 dentry = d_find_alias(inode);
3428                 if (dentry) {
3429                         path = ceph_mdsc_build_path(dentry,
3430                                                 &pathlen, &pathbase, 0);
3431                         dput(dentry);
3432                         if (IS_ERR(path)) {
3433                                 err = PTR_ERR(path);
3434                                 goto out_err;
3435                         }
3436                         rec.v1.pathbase = cpu_to_le64(pathbase);
3437                 }
3438
3439                 err = ceph_pagelist_reserve(pagelist,
3440                                             sizeof(u64) + sizeof(u32) +
3441                                             pathlen + sizeof(rec.v1));
3442                 if (err) {
3443                         goto out_freepath;
3444                 }
3445
3446                 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3447                 ceph_pagelist_encode_string(pagelist, path, pathlen);
3448                 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
3449 out_freepath:
3450                 ceph_mdsc_free_path(path, pathlen);
3451         }
3452
3453 out_err:
3454         if (err >= 0)
3455                 recon_state->nr_caps++;
3456         return err;
3457 }
3458
3459 static int encode_snap_realms(struct ceph_mds_client *mdsc,
3460                               struct ceph_reconnect_state *recon_state)
3461 {
3462         struct rb_node *p;
3463         struct ceph_pagelist *pagelist = recon_state->pagelist;
3464         int err = 0;
3465
3466         if (recon_state->msg_version >= 4) {
3467                 err = ceph_pagelist_encode_32(pagelist, mdsc->num_snap_realms);
3468                 if (err < 0)
3469                         goto fail;
3470         }
3471
3472         /*
3473          * snaprealms.  we provide mds with the ino, seq (version), and
3474          * parent for all of our realms.  If the mds has any newer info,
3475          * it will tell us.
3476          */
3477         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3478                 struct ceph_snap_realm *realm =
3479                        rb_entry(p, struct ceph_snap_realm, node);
3480                 struct ceph_mds_snaprealm_reconnect sr_rec;
3481
3482                 if (recon_state->msg_version >= 4) {
3483                         size_t need = sizeof(u8) * 2 + sizeof(u32) +
3484                                       sizeof(sr_rec);
3485
3486                         if (pagelist->length + need > RECONNECT_MAX_SIZE) {
3487                                 err = send_reconnect_partial(recon_state);
3488                                 if (err)
3489                                         goto fail;
3490                                 pagelist = recon_state->pagelist;
3491                         }
3492
3493                         err = ceph_pagelist_reserve(pagelist, need);
3494                         if (err)
3495                                 goto fail;
3496
3497                         ceph_pagelist_encode_8(pagelist, 1);
3498                         ceph_pagelist_encode_8(pagelist, 1);
3499                         ceph_pagelist_encode_32(pagelist, sizeof(sr_rec));
3500                 }
3501
3502                 dout(" adding snap realm %llx seq %lld parent %llx\n",
3503                      realm->ino, realm->seq, realm->parent_ino);
3504                 sr_rec.ino = cpu_to_le64(realm->ino);
3505                 sr_rec.seq = cpu_to_le64(realm->seq);
3506                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3507
3508                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3509                 if (err)
3510                         goto fail;
3511
3512                 recon_state->nr_realms++;
3513         }
3514 fail:
3515         return err;
3516 }
3517
3518
3519 /*
3520  * If an MDS fails and recovers, clients need to reconnect in order to
3521  * reestablish shared state.  This includes all caps issued through
3522  * this session _and_ the snap_realm hierarchy.  Because it's not
3523  * clear which snap realms the mds cares about, we send everything we
3524  * know about.. that ensures we'll then get any new info the
3525  * recovering MDS might have.
3526  *
3527  * This is a relatively heavyweight operation, but it's rare.
3528  *
3529  * called with mdsc->mutex held.
3530  */
3531 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3532                                struct ceph_mds_session *session)
3533 {
3534         struct ceph_msg *reply;
3535         int mds = session->s_mds;
3536         int err = -ENOMEM;
3537         struct ceph_reconnect_state recon_state = {
3538                 .session = session,
3539         };
3540         LIST_HEAD(dispose);
3541
3542         pr_info("mds%d reconnect start\n", mds);
3543
3544         recon_state.pagelist = ceph_pagelist_alloc(GFP_NOFS);
3545         if (!recon_state.pagelist)
3546                 goto fail_nopagelist;
3547
3548         reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3549         if (!reply)
3550                 goto fail_nomsg;
3551
3552         mutex_lock(&session->s_mutex);
3553         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3554         session->s_seq = 0;
3555
3556         dout("session %p state %s\n", session,
3557              ceph_session_state_name(session->s_state));
3558
3559         spin_lock(&session->s_gen_ttl_lock);
3560         session->s_cap_gen++;
3561         spin_unlock(&session->s_gen_ttl_lock);
3562
3563         spin_lock(&session->s_cap_lock);
3564         /* don't know if session is readonly */
3565         session->s_readonly = 0;
3566         /*
3567          * notify __ceph_remove_cap() that we are composing cap reconnect.
3568          * If a cap get released before being added to the cap reconnect,
3569          * __ceph_remove_cap() should skip queuing cap release.
3570          */
3571         session->s_cap_reconnect = 1;
3572         /* drop old cap expires; we're about to reestablish that state */
3573         detach_cap_releases(session, &dispose);
3574         spin_unlock(&session->s_cap_lock);
3575         dispose_cap_releases(mdsc, &dispose);
3576
3577         /* trim unused caps to reduce MDS's cache rejoin time */
3578         if (mdsc->fsc->sb->s_root)
3579                 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3580
3581         ceph_con_close(&session->s_con);
3582         ceph_con_open(&session->s_con,
3583                       CEPH_ENTITY_TYPE_MDS, mds,
3584                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3585
3586         /* replay unsafe requests */
3587         replay_unsafe_requests(mdsc, session);
3588
3589         ceph_early_kick_flushing_caps(mdsc, session);
3590
3591         down_read(&mdsc->snap_rwsem);
3592
3593         /* placeholder for nr_caps */
3594         err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
3595         if (err)
3596                 goto fail;
3597
3598         if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
3599                 recon_state.msg_version = 3;
3600                 recon_state.allow_multi = true;
3601         } else if (session->s_con.peer_features & CEPH_FEATURE_MDSENC) {
3602                 recon_state.msg_version = 3;
3603         } else {
3604                 recon_state.msg_version = 2;
3605         }
3606         /* trsaverse this session's caps */
3607         err = ceph_iterate_session_caps(session, encode_caps_cb, &recon_state);
3608
3609         spin_lock(&session->s_cap_lock);
3610         session->s_cap_reconnect = 0;
3611         spin_unlock(&session->s_cap_lock);
3612
3613         if (err < 0)
3614                 goto fail;
3615
3616         /* check if all realms can be encoded into current message */
3617         if (mdsc->num_snap_realms) {
3618                 size_t total_len =
3619                         recon_state.pagelist->length +
3620                         mdsc->num_snap_realms *
3621                         sizeof(struct ceph_mds_snaprealm_reconnect);
3622                 if (recon_state.msg_version >= 4) {
3623                         /* number of realms */
3624                         total_len += sizeof(u32);
3625                         /* version, compat_version and struct_len */
3626                         total_len += mdsc->num_snap_realms *
3627                                      (2 * sizeof(u8) + sizeof(u32));
3628                 }
3629                 if (total_len > RECONNECT_MAX_SIZE) {
3630                         if (!recon_state.allow_multi) {
3631                                 err = -ENOSPC;
3632                                 goto fail;
3633                         }
3634                         if (recon_state.nr_caps) {
3635                                 err = send_reconnect_partial(&recon_state);
3636                                 if (err)
3637                                         goto fail;
3638                         }
3639                         recon_state.msg_version = 5;
3640                 }
3641         }
3642
3643         err = encode_snap_realms(mdsc, &recon_state);
3644         if (err < 0)
3645                 goto fail;
3646
3647         if (recon_state.msg_version >= 5) {
3648                 err = ceph_pagelist_encode_8(recon_state.pagelist, 0);
3649                 if (err < 0)
3650                         goto fail;
3651         }
3652
3653         if (recon_state.nr_caps || recon_state.nr_realms) {
3654                 struct page *page =
3655                         list_first_entry(&recon_state.pagelist->head,
3656                                         struct page, lru);
3657                 __le32 *addr = kmap_atomic(page);
3658                 if (recon_state.nr_caps) {
3659                         WARN_ON(recon_state.nr_realms != mdsc->num_snap_realms);
3660                         *addr = cpu_to_le32(recon_state.nr_caps);
3661                 } else if (recon_state.msg_version >= 4) {
3662                         *(addr + 1) = cpu_to_le32(recon_state.nr_realms);
3663                 }
3664                 kunmap_atomic(addr);
3665         }
3666
3667         reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3668         if (recon_state.msg_version >= 4)
3669                 reply->hdr.compat_version = cpu_to_le16(4);
3670
3671         reply->hdr.data_len = cpu_to_le32(recon_state.pagelist->length);
3672         ceph_msg_data_add_pagelist(reply, recon_state.pagelist);
3673
3674         ceph_con_send(&session->s_con, reply);
3675
3676         mutex_unlock(&session->s_mutex);
3677
3678         mutex_lock(&mdsc->mutex);
3679         __wake_requests(mdsc, &session->s_waiting);
3680         mutex_unlock(&mdsc->mutex);
3681
3682         up_read(&mdsc->snap_rwsem);
3683         ceph_pagelist_release(recon_state.pagelist);
3684         return;
3685
3686 fail:
3687         ceph_msg_put(reply);
3688         up_read(&mdsc->snap_rwsem);
3689         mutex_unlock(&session->s_mutex);
3690 fail_nomsg:
3691         ceph_pagelist_release(recon_state.pagelist);
3692 fail_nopagelist:
3693         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3694         return;
3695 }
3696
3697
3698 /*
3699  * compare old and new mdsmaps, kicking requests
3700  * and closing out old connections as necessary
3701  *
3702  * called under mdsc->mutex.
3703  */
3704 static void check_new_map(struct ceph_mds_client *mdsc,
3705                           struct ceph_mdsmap *newmap,
3706                           struct ceph_mdsmap *oldmap)
3707 {
3708         int i;
3709         int oldstate, newstate;
3710         struct ceph_mds_session *s;
3711
3712         dout("check_new_map new %u old %u\n",
3713              newmap->m_epoch, oldmap->m_epoch);
3714
3715         for (i = 0; i < oldmap->m_num_mds && i < mdsc->max_sessions; i++) {
3716                 if (!mdsc->sessions[i])
3717                         continue;
3718                 s = mdsc->sessions[i];
3719                 oldstate = ceph_mdsmap_get_state(oldmap, i);
3720                 newstate = ceph_mdsmap_get_state(newmap, i);
3721
3722                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3723                      i, ceph_mds_state_name(oldstate),
3724                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3725                      ceph_mds_state_name(newstate),
3726                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3727                      ceph_session_state_name(s->s_state));
3728
3729                 if (i >= newmap->m_num_mds ||
3730                     memcmp(ceph_mdsmap_get_addr(oldmap, i),
3731                            ceph_mdsmap_get_addr(newmap, i),
3732                            sizeof(struct ceph_entity_addr))) {
3733                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3734                                 /* the session never opened, just close it
3735                                  * out now */
3736                                 get_session(s);
3737                                 __unregister_session(mdsc, s);
3738                                 __wake_requests(mdsc, &s->s_waiting);
3739                                 ceph_put_mds_session(s);
3740                         } else if (i >= newmap->m_num_mds) {
3741                                 /* force close session for stopped mds */
3742                                 get_session(s);
3743                                 __unregister_session(mdsc, s);
3744                                 __wake_requests(mdsc, &s->s_waiting);
3745                                 kick_requests(mdsc, i);
3746                                 mutex_unlock(&mdsc->mutex);
3747
3748                                 mutex_lock(&s->s_mutex);
3749                                 cleanup_session_requests(mdsc, s);
3750                                 remove_session_caps(s);
3751                                 mutex_unlock(&s->s_mutex);
3752
3753                                 ceph_put_mds_session(s);
3754
3755                                 mutex_lock(&mdsc->mutex);
3756                         } else {
3757                                 /* just close it */
3758                                 mutex_unlock(&mdsc->mutex);
3759                                 mutex_lock(&s->s_mutex);
3760                                 mutex_lock(&mdsc->mutex);
3761                                 ceph_con_close(&s->s_con);
3762                                 mutex_unlock(&s->s_mutex);
3763                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3764                         }
3765                 } else if (oldstate == newstate) {
3766                         continue;  /* nothing new with this mds */
3767                 }
3768
3769                 /*
3770                  * send reconnect?
3771                  */
3772                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3773                     newstate >= CEPH_MDS_STATE_RECONNECT) {
3774                         mutex_unlock(&mdsc->mutex);
3775                         send_mds_reconnect(mdsc, s);
3776                         mutex_lock(&mdsc->mutex);
3777                 }
3778
3779                 /*
3780                  * kick request on any mds that has gone active.
3781                  */
3782                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3783                     newstate >= CEPH_MDS_STATE_ACTIVE) {
3784                         if (oldstate != CEPH_MDS_STATE_CREATING &&
3785                             oldstate != CEPH_MDS_STATE_STARTING)
3786                                 pr_info("mds%d recovery completed\n", s->s_mds);
3787                         kick_requests(mdsc, i);
3788                         ceph_kick_flushing_caps(mdsc, s);
3789                         wake_up_session_caps(s, RECONNECT);
3790                 }
3791         }
3792
3793         for (i = 0; i < newmap->m_num_mds && i < mdsc->max_sessions; i++) {
3794                 s = mdsc->sessions[i];
3795                 if (!s)
3796                         continue;
3797                 if (!ceph_mdsmap_is_laggy(newmap, i))
3798                         continue;
3799                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3800                     s->s_state == CEPH_MDS_SESSION_HUNG ||
3801                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
3802                         dout(" connecting to export targets of laggy mds%d\n",
3803                              i);
3804                         __open_export_target_sessions(mdsc, s);
3805                 }
3806         }
3807 }
3808
3809
3810
3811 /*
3812  * leases
3813  */
3814
3815 /*
3816  * caller must hold session s_mutex, dentry->d_lock
3817  */
3818 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3819 {
3820         struct ceph_dentry_info *di = ceph_dentry(dentry);
3821
3822         ceph_put_mds_session(di->lease_session);
3823         di->lease_session = NULL;
3824 }
3825
3826 static void handle_lease(struct ceph_mds_client *mdsc,
3827                          struct ceph_mds_session *session,
3828                          struct ceph_msg *msg)
3829 {
3830         struct super_block *sb = mdsc->fsc->sb;
3831         struct inode *inode;
3832         struct dentry *parent, *dentry;
3833         struct ceph_dentry_info *di;
3834         int mds = session->s_mds;
3835         struct ceph_mds_lease *h = msg->front.iov_base;
3836         u32 seq;
3837         struct ceph_vino vino;
3838         struct qstr dname;
3839         int release = 0;
3840
3841         dout("handle_lease from mds%d\n", mds);
3842
3843         /* decode */
3844         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3845                 goto bad;
3846         vino.ino = le64_to_cpu(h->ino);
3847         vino.snap = CEPH_NOSNAP;
3848         seq = le32_to_cpu(h->seq);
3849         dname.len = get_unaligned_le32(h + 1);
3850         if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
3851                 goto bad;
3852         dname.name = (void *)(h + 1) + sizeof(u32);
3853
3854         /* lookup inode */
3855         inode = ceph_find_inode(sb, vino);
3856         dout("handle_lease %s, ino %llx %p %.*s\n",
3857              ceph_lease_op_name(h->action), vino.ino, inode,
3858              dname.len, dname.name);
3859
3860         mutex_lock(&session->s_mutex);
3861         session->s_seq++;
3862
3863         if (!inode) {
3864                 dout("handle_lease no inode %llx\n", vino.ino);
3865                 goto release;
3866         }
3867
3868         /* dentry */
3869         parent = d_find_alias(inode);
3870         if (!parent) {
3871                 dout("no parent dentry on inode %p\n", inode);
3872                 WARN_ON(1);
3873                 goto release;  /* hrm... */
3874         }
3875         dname.hash = full_name_hash(parent, dname.name, dname.len);
3876         dentry = d_lookup(parent, &dname);
3877         dput(parent);
3878         if (!dentry)
3879                 goto release;
3880
3881         spin_lock(&dentry->d_lock);
3882         di = ceph_dentry(dentry);
3883         switch (h->action) {
3884         case CEPH_MDS_LEASE_REVOKE:
3885                 if (di->lease_session == session) {
3886                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3887                                 h->seq = cpu_to_le32(di->lease_seq);
3888                         __ceph_mdsc_drop_dentry_lease(dentry);
3889                 }
3890                 release = 1;
3891                 break;
3892
3893         case CEPH_MDS_LEASE_RENEW:
3894                 if (di->lease_session == session &&
3895                     di->lease_gen == session->s_cap_gen &&
3896                     di->lease_renew_from &&
3897                     di->lease_renew_after == 0) {
3898                         unsigned long duration =
3899                                 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3900
3901                         di->lease_seq = seq;
3902                         di->time = di->lease_renew_from + duration;
3903                         di->lease_renew_after = di->lease_renew_from +
3904                                 (duration >> 1);
3905                         di->lease_renew_from = 0;
3906                 }
3907                 break;
3908         }
3909         spin_unlock(&dentry->d_lock);
3910         dput(dentry);
3911
3912         if (!release)
3913                 goto out;
3914
3915 release:
3916         /* let's just reuse the same message */
3917         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3918         ceph_msg_get(msg);
3919         ceph_con_send(&session->s_con, msg);
3920
3921 out:
3922         mutex_unlock(&session->s_mutex);
3923         /* avoid calling iput_final() in mds dispatch threads */
3924         ceph_async_iput(inode);
3925         return;
3926
3927 bad:
3928         pr_err("corrupt lease message\n");
3929         ceph_msg_dump(msg);
3930 }
3931
3932 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3933                               struct inode *inode,
3934                               struct dentry *dentry, char action,
3935                               u32 seq)
3936 {
3937         struct ceph_msg *msg;
3938         struct ceph_mds_lease *lease;
3939         int len = sizeof(*lease) + sizeof(u32);
3940         int dnamelen = 0;
3941
3942         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3943              inode, dentry, ceph_lease_op_name(action), session->s_mds);
3944         dnamelen = dentry->d_name.len;
3945         len += dnamelen;
3946
3947         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3948         if (!msg)
3949                 return;
3950         lease = msg->front.iov_base;
3951         lease->action = action;
3952         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3953         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3954         lease->seq = cpu_to_le32(seq);
3955         put_unaligned_le32(dnamelen, lease + 1);
3956         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3957
3958         /*
3959          * if this is a preemptive lease RELEASE, no need to
3960          * flush request stream, since the actual request will
3961          * soon follow.
3962          */
3963         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3964
3965         ceph_con_send(&session->s_con, msg);
3966 }
3967
3968 /*
3969  * lock unlock sessions, to wait ongoing session activities
3970  */
3971 static void lock_unlock_sessions(struct ceph_mds_client *mdsc)
3972 {
3973         int i;
3974
3975         mutex_lock(&mdsc->mutex);
3976         for (i = 0; i < mdsc->max_sessions; i++) {
3977                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3978                 if (!s)
3979                         continue;
3980                 mutex_unlock(&mdsc->mutex);
3981                 mutex_lock(&s->s_mutex);
3982                 mutex_unlock(&s->s_mutex);
3983                 ceph_put_mds_session(s);
3984                 mutex_lock(&mdsc->mutex);
3985         }
3986         mutex_unlock(&mdsc->mutex);
3987 }
3988
3989
3990
3991 /*
3992  * delayed work -- periodically trim expired leases, renew caps with mds
3993  */
3994 static void schedule_delayed(struct ceph_mds_client *mdsc)
3995 {
3996         int delay = 5;
3997         unsigned hz = round_jiffies_relative(HZ * delay);
3998         schedule_delayed_work(&mdsc->delayed_work, hz);
3999 }
4000
4001 static void delayed_work(struct work_struct *work)
4002 {
4003         int i;
4004         struct ceph_mds_client *mdsc =
4005                 container_of(work, struct ceph_mds_client, delayed_work.work);
4006         int renew_interval;
4007         int renew_caps;
4008
4009         dout("mdsc delayed_work\n");
4010
4011         mutex_lock(&mdsc->mutex);
4012         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
4013         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
4014                                    mdsc->last_renew_caps);
4015         if (renew_caps)
4016                 mdsc->last_renew_caps = jiffies;
4017
4018         for (i = 0; i < mdsc->max_sessions; i++) {
4019                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4020                 if (!s)
4021                         continue;
4022                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
4023                         dout("resending session close request for mds%d\n",
4024                              s->s_mds);
4025                         request_close_session(mdsc, s);
4026                         ceph_put_mds_session(s);
4027                         continue;
4028                 }
4029                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
4030                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
4031                                 s->s_state = CEPH_MDS_SESSION_HUNG;
4032                                 pr_info("mds%d hung\n", s->s_mds);
4033                         }
4034                 }
4035                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
4036                         /* this mds is failed or recovering, just wait */
4037                         ceph_put_mds_session(s);
4038                         continue;
4039                 }
4040                 mutex_unlock(&mdsc->mutex);
4041
4042                 mutex_lock(&s->s_mutex);
4043                 if (renew_caps)
4044                         send_renew_caps(mdsc, s);
4045                 else
4046                         ceph_con_keepalive(&s->s_con);
4047                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4048                     s->s_state == CEPH_MDS_SESSION_HUNG)
4049                         ceph_send_cap_releases(mdsc, s);
4050                 mutex_unlock(&s->s_mutex);
4051                 ceph_put_mds_session(s);
4052
4053                 mutex_lock(&mdsc->mutex);
4054         }
4055         mutex_unlock(&mdsc->mutex);
4056
4057         ceph_check_delayed_caps(mdsc);
4058
4059         ceph_queue_cap_reclaim_work(mdsc);
4060
4061         ceph_trim_snapid_map(mdsc);
4062
4063         schedule_delayed(mdsc);
4064 }
4065
4066 int ceph_mdsc_init(struct ceph_fs_client *fsc)
4067
4068 {
4069         struct ceph_mds_client *mdsc;
4070
4071         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
4072         if (!mdsc)
4073                 return -ENOMEM;
4074         mdsc->fsc = fsc;
4075         mutex_init(&mdsc->mutex);
4076         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
4077         if (!mdsc->mdsmap) {
4078                 kfree(mdsc);
4079                 return -ENOMEM;
4080         }
4081
4082         fsc->mdsc = mdsc;
4083         init_completion(&mdsc->safe_umount_waiters);
4084         init_waitqueue_head(&mdsc->session_close_wq);
4085         INIT_LIST_HEAD(&mdsc->waiting_for_map);
4086         mdsc->sessions = NULL;
4087         atomic_set(&mdsc->num_sessions, 0);
4088         mdsc->max_sessions = 0;
4089         mdsc->stopping = 0;
4090         atomic64_set(&mdsc->quotarealms_count, 0);
4091         mdsc->quotarealms_inodes = RB_ROOT;
4092         mutex_init(&mdsc->quotarealms_inodes_mutex);
4093         mdsc->last_snap_seq = 0;
4094         init_rwsem(&mdsc->snap_rwsem);
4095         mdsc->snap_realms = RB_ROOT;
4096         INIT_LIST_HEAD(&mdsc->snap_empty);
4097         mdsc->num_snap_realms = 0;
4098         spin_lock_init(&mdsc->snap_empty_lock);
4099         mdsc->last_tid = 0;
4100         mdsc->oldest_tid = 0;
4101         mdsc->request_tree = RB_ROOT;
4102         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
4103         mdsc->last_renew_caps = jiffies;
4104         INIT_LIST_HEAD(&mdsc->cap_delay_list);
4105         spin_lock_init(&mdsc->cap_delay_lock);
4106         INIT_LIST_HEAD(&mdsc->snap_flush_list);
4107         spin_lock_init(&mdsc->snap_flush_lock);
4108         mdsc->last_cap_flush_tid = 1;
4109         INIT_LIST_HEAD(&mdsc->cap_flush_list);
4110         INIT_LIST_HEAD(&mdsc->cap_dirty);
4111         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
4112         mdsc->num_cap_flushing = 0;
4113         spin_lock_init(&mdsc->cap_dirty_lock);
4114         init_waitqueue_head(&mdsc->cap_flushing_wq);
4115         INIT_WORK(&mdsc->cap_reclaim_work, ceph_cap_reclaim_work);
4116         atomic_set(&mdsc->cap_reclaim_pending, 0);
4117
4118         spin_lock_init(&mdsc->dentry_list_lock);
4119         INIT_LIST_HEAD(&mdsc->dentry_leases);
4120         INIT_LIST_HEAD(&mdsc->dentry_dir_leases);
4121
4122         ceph_caps_init(mdsc);
4123         ceph_adjust_caps_max_min(mdsc, fsc->mount_options);
4124
4125         spin_lock_init(&mdsc->snapid_map_lock);
4126         mdsc->snapid_map_tree = RB_ROOT;
4127         INIT_LIST_HEAD(&mdsc->snapid_map_lru);
4128
4129         init_rwsem(&mdsc->pool_perm_rwsem);
4130         mdsc->pool_perm_tree = RB_ROOT;
4131
4132         strscpy(mdsc->nodename, utsname()->nodename,
4133                 sizeof(mdsc->nodename));
4134         return 0;
4135 }
4136
4137 /*
4138  * Wait for safe replies on open mds requests.  If we time out, drop
4139  * all requests from the tree to avoid dangling dentry refs.
4140  */
4141 static void wait_requests(struct ceph_mds_client *mdsc)
4142 {
4143         struct ceph_options *opts = mdsc->fsc->client->options;
4144         struct ceph_mds_request *req;
4145
4146         mutex_lock(&mdsc->mutex);
4147         if (__get_oldest_req(mdsc)) {
4148                 mutex_unlock(&mdsc->mutex);
4149
4150                 dout("wait_requests waiting for requests\n");
4151                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
4152                                     ceph_timeout_jiffies(opts->mount_timeout));
4153
4154                 /* tear down remaining requests */
4155                 mutex_lock(&mdsc->mutex);
4156                 while ((req = __get_oldest_req(mdsc))) {
4157                         dout("wait_requests timed out on tid %llu\n",
4158                              req->r_tid);
4159                         __unregister_request(mdsc, req);
4160                 }
4161         }
4162         mutex_unlock(&mdsc->mutex);
4163         dout("wait_requests done\n");
4164 }
4165
4166 /*
4167  * called before mount is ro, and before dentries are torn down.
4168  * (hmm, does this still race with new lookups?)
4169  */
4170 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
4171 {
4172         dout("pre_umount\n");
4173         mdsc->stopping = 1;
4174
4175         lock_unlock_sessions(mdsc);
4176         ceph_flush_dirty_caps(mdsc);
4177         wait_requests(mdsc);
4178
4179         /*
4180          * wait for reply handlers to drop their request refs and
4181          * their inode/dcache refs
4182          */
4183         ceph_msgr_flush();
4184
4185         ceph_cleanup_quotarealms_inodes(mdsc);
4186 }
4187
4188 /*
4189  * wait for all write mds requests to flush.
4190  */
4191 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
4192 {
4193         struct ceph_mds_request *req = NULL, *nextreq;
4194         struct rb_node *n;
4195
4196         mutex_lock(&mdsc->mutex);
4197         dout("wait_unsafe_requests want %lld\n", want_tid);
4198 restart:
4199         req = __get_oldest_req(mdsc);
4200         while (req && req->r_tid <= want_tid) {
4201                 /* find next request */
4202                 n = rb_next(&req->r_node);
4203                 if (n)
4204                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
4205                 else
4206                         nextreq = NULL;
4207                 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
4208                     (req->r_op & CEPH_MDS_OP_WRITE)) {
4209                         /* write op */
4210                         ceph_mdsc_get_request(req);
4211                         if (nextreq)
4212                                 ceph_mdsc_get_request(nextreq);
4213                         mutex_unlock(&mdsc->mutex);
4214                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
4215                              req->r_tid, want_tid);
4216                         wait_for_completion(&req->r_safe_completion);
4217                         mutex_lock(&mdsc->mutex);
4218                         ceph_mdsc_put_request(req);
4219                         if (!nextreq)
4220                                 break;  /* next dne before, so we're done! */
4221                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
4222                                 /* next request was removed from tree */
4223                                 ceph_mdsc_put_request(nextreq);
4224                                 goto restart;
4225                         }
4226                         ceph_mdsc_put_request(nextreq);  /* won't go away */
4227                 }
4228                 req = nextreq;
4229         }
4230         mutex_unlock(&mdsc->mutex);
4231         dout("wait_unsafe_requests done\n");
4232 }
4233
4234 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
4235 {
4236         u64 want_tid, want_flush;
4237
4238         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4239                 return;
4240
4241         dout("sync\n");
4242         mutex_lock(&mdsc->mutex);
4243         want_tid = mdsc->last_tid;
4244         mutex_unlock(&mdsc->mutex);
4245
4246         ceph_flush_dirty_caps(mdsc);
4247         spin_lock(&mdsc->cap_dirty_lock);
4248         want_flush = mdsc->last_cap_flush_tid;
4249         if (!list_empty(&mdsc->cap_flush_list)) {
4250                 struct ceph_cap_flush *cf =
4251                         list_last_entry(&mdsc->cap_flush_list,
4252                                         struct ceph_cap_flush, g_list);
4253                 cf->wake = true;
4254         }
4255         spin_unlock(&mdsc->cap_dirty_lock);
4256
4257         dout("sync want tid %lld flush_seq %lld\n",
4258              want_tid, want_flush);
4259
4260         wait_unsafe_requests(mdsc, want_tid);
4261         wait_caps_flush(mdsc, want_flush);
4262 }
4263
4264 /*
4265  * true if all sessions are closed, or we force unmount
4266  */
4267 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
4268 {
4269         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4270                 return true;
4271         return atomic_read(&mdsc->num_sessions) <= skipped;
4272 }
4273
4274 /*
4275  * called after sb is ro.
4276  */
4277 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
4278 {
4279         struct ceph_options *opts = mdsc->fsc->client->options;
4280         struct ceph_mds_session *session;
4281         int i;
4282         int skipped = 0;
4283
4284         dout("close_sessions\n");
4285
4286         /* close sessions */
4287         mutex_lock(&mdsc->mutex);
4288         for (i = 0; i < mdsc->max_sessions; i++) {
4289                 session = __ceph_lookup_mds_session(mdsc, i);
4290                 if (!session)
4291                         continue;
4292                 mutex_unlock(&mdsc->mutex);
4293                 mutex_lock(&session->s_mutex);
4294                 if (__close_session(mdsc, session) <= 0)
4295                         skipped++;
4296                 mutex_unlock(&session->s_mutex);
4297                 ceph_put_mds_session(session);
4298                 mutex_lock(&mdsc->mutex);
4299         }
4300         mutex_unlock(&mdsc->mutex);
4301
4302         dout("waiting for sessions to close\n");
4303         wait_event_timeout(mdsc->session_close_wq,
4304                            done_closing_sessions(mdsc, skipped),
4305                            ceph_timeout_jiffies(opts->mount_timeout));
4306
4307         /* tear down remaining sessions */
4308         mutex_lock(&mdsc->mutex);
4309         for (i = 0; i < mdsc->max_sessions; i++) {
4310                 if (mdsc->sessions[i]) {
4311                         session = get_session(mdsc->sessions[i]);
4312                         __unregister_session(mdsc, session);
4313                         mutex_unlock(&mdsc->mutex);
4314                         mutex_lock(&session->s_mutex);
4315                         remove_session_caps(session);
4316                         mutex_unlock(&session->s_mutex);
4317                         ceph_put_mds_session(session);
4318                         mutex_lock(&mdsc->mutex);
4319                 }
4320         }
4321         WARN_ON(!list_empty(&mdsc->cap_delay_list));
4322         mutex_unlock(&mdsc->mutex);
4323
4324         ceph_cleanup_snapid_map(mdsc);
4325         ceph_cleanup_empty_realms(mdsc);
4326
4327         cancel_work_sync(&mdsc->cap_reclaim_work);
4328         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4329
4330         dout("stopped\n");
4331 }
4332
4333 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
4334 {
4335         struct ceph_mds_session *session;
4336         int mds;
4337
4338         dout("force umount\n");
4339
4340         mutex_lock(&mdsc->mutex);
4341         for (mds = 0; mds < mdsc->max_sessions; mds++) {
4342                 session = __ceph_lookup_mds_session(mdsc, mds);
4343                 if (!session)
4344                         continue;
4345                 mutex_unlock(&mdsc->mutex);
4346                 mutex_lock(&session->s_mutex);
4347                 __close_session(mdsc, session);
4348                 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
4349                         cleanup_session_requests(mdsc, session);
4350                         remove_session_caps(session);
4351                 }
4352                 mutex_unlock(&session->s_mutex);
4353                 ceph_put_mds_session(session);
4354                 mutex_lock(&mdsc->mutex);
4355                 kick_requests(mdsc, mds);
4356         }
4357         __wake_requests(mdsc, &mdsc->waiting_for_map);
4358         mutex_unlock(&mdsc->mutex);
4359 }
4360
4361 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
4362 {
4363         dout("stop\n");
4364         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4365         if (mdsc->mdsmap)
4366                 ceph_mdsmap_destroy(mdsc->mdsmap);
4367         kfree(mdsc->sessions);
4368         ceph_caps_finalize(mdsc);
4369         ceph_pool_perm_destroy(mdsc);
4370 }
4371
4372 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
4373 {
4374         struct ceph_mds_client *mdsc = fsc->mdsc;
4375         dout("mdsc_destroy %p\n", mdsc);
4376
4377         if (!mdsc)
4378                 return;
4379
4380         /* flush out any connection work with references to us */
4381         ceph_msgr_flush();
4382
4383         ceph_mdsc_stop(mdsc);
4384
4385         fsc->mdsc = NULL;
4386         kfree(mdsc);
4387         dout("mdsc_destroy %p done\n", mdsc);
4388 }
4389
4390 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4391 {
4392         struct ceph_fs_client *fsc = mdsc->fsc;
4393         const char *mds_namespace = fsc->mount_options->mds_namespace;
4394         void *p = msg->front.iov_base;
4395         void *end = p + msg->front.iov_len;
4396         u32 epoch;
4397         u32 map_len;
4398         u32 num_fs;
4399         u32 mount_fscid = (u32)-1;
4400         u8 struct_v, struct_cv;
4401         int err = -EINVAL;
4402
4403         ceph_decode_need(&p, end, sizeof(u32), bad);
4404         epoch = ceph_decode_32(&p);
4405
4406         dout("handle_fsmap epoch %u\n", epoch);
4407
4408         ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4409         struct_v = ceph_decode_8(&p);
4410         struct_cv = ceph_decode_8(&p);
4411         map_len = ceph_decode_32(&p);
4412
4413         ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
4414         p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
4415
4416         num_fs = ceph_decode_32(&p);
4417         while (num_fs-- > 0) {
4418                 void *info_p, *info_end;
4419                 u32 info_len;
4420                 u8 info_v, info_cv;
4421                 u32 fscid, namelen;
4422
4423                 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4424                 info_v = ceph_decode_8(&p);
4425                 info_cv = ceph_decode_8(&p);
4426                 info_len = ceph_decode_32(&p);
4427                 ceph_decode_need(&p, end, info_len, bad);
4428                 info_p = p;
4429                 info_end = p + info_len;
4430                 p = info_end;
4431
4432                 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
4433                 fscid = ceph_decode_32(&info_p);
4434                 namelen = ceph_decode_32(&info_p);
4435                 ceph_decode_need(&info_p, info_end, namelen, bad);
4436
4437                 if (mds_namespace &&
4438                     strlen(mds_namespace) == namelen &&
4439                     !strncmp(mds_namespace, (char *)info_p, namelen)) {
4440                         mount_fscid = fscid;
4441                         break;
4442                 }
4443         }
4444
4445         ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
4446         if (mount_fscid != (u32)-1) {
4447                 fsc->client->monc.fs_cluster_id = mount_fscid;
4448                 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
4449                                    0, true);
4450                 ceph_monc_renew_subs(&fsc->client->monc);
4451         } else {
4452                 err = -ENOENT;
4453                 goto err_out;
4454         }
4455         return;
4456
4457 bad:
4458         pr_err("error decoding fsmap\n");
4459 err_out:
4460         mutex_lock(&mdsc->mutex);
4461         mdsc->mdsmap_err = err;
4462         __wake_requests(mdsc, &mdsc->waiting_for_map);
4463         mutex_unlock(&mdsc->mutex);
4464 }
4465
4466 /*
4467  * handle mds map update.
4468  */
4469 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4470 {
4471         u32 epoch;
4472         u32 maplen;
4473         void *p = msg->front.iov_base;
4474         void *end = p + msg->front.iov_len;
4475         struct ceph_mdsmap *newmap, *oldmap;
4476         struct ceph_fsid fsid;
4477         int err = -EINVAL;
4478
4479         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
4480         ceph_decode_copy(&p, &fsid, sizeof(fsid));
4481         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
4482                 return;
4483         epoch = ceph_decode_32(&p);
4484         maplen = ceph_decode_32(&p);
4485         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
4486
4487         /* do we need it? */
4488         mutex_lock(&mdsc->mutex);
4489         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
4490                 dout("handle_map epoch %u <= our %u\n",
4491                      epoch, mdsc->mdsmap->m_epoch);
4492                 mutex_unlock(&mdsc->mutex);
4493                 return;
4494         }
4495
4496         newmap = ceph_mdsmap_decode(&p, end);
4497         if (IS_ERR(newmap)) {
4498                 err = PTR_ERR(newmap);
4499                 goto bad_unlock;
4500         }
4501
4502         /* swap into place */
4503         if (mdsc->mdsmap) {
4504                 oldmap = mdsc->mdsmap;
4505                 mdsc->mdsmap = newmap;
4506                 check_new_map(mdsc, newmap, oldmap);
4507                 ceph_mdsmap_destroy(oldmap);
4508         } else {
4509                 mdsc->mdsmap = newmap;  /* first mds map */
4510         }
4511         mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
4512                                         MAX_LFS_FILESIZE);
4513
4514         __wake_requests(mdsc, &mdsc->waiting_for_map);
4515         ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
4516                           mdsc->mdsmap->m_epoch);
4517
4518         mutex_unlock(&mdsc->mutex);
4519         schedule_delayed(mdsc);
4520         return;
4521
4522 bad_unlock:
4523         mutex_unlock(&mdsc->mutex);
4524 bad:
4525         pr_err("error decoding mdsmap %d\n", err);
4526         return;
4527 }
4528
4529 static struct ceph_connection *con_get(struct ceph_connection *con)
4530 {
4531         struct ceph_mds_session *s = con->private;
4532
4533         if (get_session(s)) {
4534                 dout("mdsc con_get %p ok (%d)\n", s, refcount_read(&s->s_ref));
4535                 return con;
4536         }
4537         dout("mdsc con_get %p FAIL\n", s);
4538         return NULL;
4539 }
4540
4541 static void con_put(struct ceph_connection *con)
4542 {
4543         struct ceph_mds_session *s = con->private;
4544
4545         dout("mdsc con_put %p (%d)\n", s, refcount_read(&s->s_ref) - 1);
4546         ceph_put_mds_session(s);
4547 }
4548
4549 /*
4550  * if the client is unresponsive for long enough, the mds will kill
4551  * the session entirely.
4552  */
4553 static void peer_reset(struct ceph_connection *con)
4554 {
4555         struct ceph_mds_session *s = con->private;
4556         struct ceph_mds_client *mdsc = s->s_mdsc;
4557
4558         pr_warn("mds%d closed our session\n", s->s_mds);
4559         send_mds_reconnect(mdsc, s);
4560 }
4561
4562 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
4563 {
4564         struct ceph_mds_session *s = con->private;
4565         struct ceph_mds_client *mdsc = s->s_mdsc;
4566         int type = le16_to_cpu(msg->hdr.type);
4567
4568         mutex_lock(&mdsc->mutex);
4569         if (__verify_registered_session(mdsc, s) < 0) {
4570                 mutex_unlock(&mdsc->mutex);
4571                 goto out;
4572         }
4573         mutex_unlock(&mdsc->mutex);
4574
4575         switch (type) {
4576         case CEPH_MSG_MDS_MAP:
4577                 ceph_mdsc_handle_mdsmap(mdsc, msg);
4578                 break;
4579         case CEPH_MSG_FS_MAP_USER:
4580                 ceph_mdsc_handle_fsmap(mdsc, msg);
4581                 break;
4582         case CEPH_MSG_CLIENT_SESSION:
4583                 handle_session(s, msg);
4584                 break;
4585         case CEPH_MSG_CLIENT_REPLY:
4586                 handle_reply(s, msg);
4587                 break;
4588         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
4589                 handle_forward(mdsc, s, msg);
4590                 break;
4591         case CEPH_MSG_CLIENT_CAPS:
4592                 ceph_handle_caps(s, msg);
4593                 break;
4594         case CEPH_MSG_CLIENT_SNAP:
4595                 ceph_handle_snap(mdsc, s, msg);
4596                 break;
4597         case CEPH_MSG_CLIENT_LEASE:
4598                 handle_lease(mdsc, s, msg);
4599                 break;
4600         case CEPH_MSG_CLIENT_QUOTA:
4601                 ceph_handle_quota(mdsc, s, msg);
4602                 break;
4603
4604         default:
4605                 pr_err("received unknown message type %d %s\n", type,
4606                        ceph_msg_type_name(type));
4607         }
4608 out:
4609         ceph_msg_put(msg);
4610 }
4611
4612 /*
4613  * authentication
4614  */
4615
4616 /*
4617  * Note: returned pointer is the address of a structure that's
4618  * managed separately.  Caller must *not* attempt to free it.
4619  */
4620 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
4621                                         int *proto, int force_new)
4622 {
4623         struct ceph_mds_session *s = con->private;
4624         struct ceph_mds_client *mdsc = s->s_mdsc;
4625         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4626         struct ceph_auth_handshake *auth = &s->s_auth;
4627
4628         if (force_new && auth->authorizer) {
4629                 ceph_auth_destroy_authorizer(auth->authorizer);
4630                 auth->authorizer = NULL;
4631         }
4632         if (!auth->authorizer) {
4633                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4634                                                       auth);
4635                 if (ret)
4636                         return ERR_PTR(ret);
4637         } else {
4638                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4639                                                       auth);
4640                 if (ret)
4641                         return ERR_PTR(ret);
4642         }
4643         *proto = ac->protocol;
4644
4645         return auth;
4646 }
4647
4648 static int add_authorizer_challenge(struct ceph_connection *con,
4649                                     void *challenge_buf, int challenge_buf_len)
4650 {
4651         struct ceph_mds_session *s = con->private;
4652         struct ceph_mds_client *mdsc = s->s_mdsc;
4653         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4654
4655         return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
4656                                             challenge_buf, challenge_buf_len);
4657 }
4658
4659 static int verify_authorizer_reply(struct ceph_connection *con)
4660 {
4661         struct ceph_mds_session *s = con->private;
4662         struct ceph_mds_client *mdsc = s->s_mdsc;
4663         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4664
4665         return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
4666 }
4667
4668 static int invalidate_authorizer(struct ceph_connection *con)
4669 {
4670         struct ceph_mds_session *s = con->private;
4671         struct ceph_mds_client *mdsc = s->s_mdsc;
4672         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4673
4674         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
4675
4676         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
4677 }
4678
4679 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
4680                                 struct ceph_msg_header *hdr, int *skip)
4681 {
4682         struct ceph_msg *msg;
4683         int type = (int) le16_to_cpu(hdr->type);
4684         int front_len = (int) le32_to_cpu(hdr->front_len);
4685
4686         if (con->in_msg)
4687                 return con->in_msg;
4688
4689         *skip = 0;
4690         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
4691         if (!msg) {
4692                 pr_err("unable to allocate msg type %d len %d\n",
4693                        type, front_len);
4694                 return NULL;
4695         }
4696
4697         return msg;
4698 }
4699
4700 static int mds_sign_message(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_sign_message(auth, msg);
4706 }
4707
4708 static int mds_check_message_signature(struct ceph_msg *msg)
4709 {
4710        struct ceph_mds_session *s = msg->con->private;
4711        struct ceph_auth_handshake *auth = &s->s_auth;
4712
4713        return ceph_auth_check_message_signature(auth, msg);
4714 }
4715
4716 static const struct ceph_connection_operations mds_con_ops = {
4717         .get = con_get,
4718         .put = con_put,
4719         .dispatch = dispatch,
4720         .get_authorizer = get_authorizer,
4721         .add_authorizer_challenge = add_authorizer_challenge,
4722         .verify_authorizer_reply = verify_authorizer_reply,
4723         .invalidate_authorizer = invalidate_authorizer,
4724         .peer_reset = peer_reset,
4725         .alloc_msg = mds_alloc_msg,
4726         .sign_message = mds_sign_message,
4727         .check_message_signature = mds_check_message_signature,
4728 };
4729
4730 /* eof */