ldiskfs/kernel_patches/patches/linux-5.10/ext4-pdirop.patch

   1 LU-50 ldiskfs: pdirops patch for ldiskfs
   2
   3 Single directory performance is a critical for HPC workloads. In a
   4 typical use case an application creates a separate output file for
   5 each node and task in a job. As nodes and tasks increase, hundreds
   6 of thousands of files may be created in a single directory within
   7 a short window of time.
   8 Today, both filename lookup and file system modifying operations
   9 (such as create and unlink) are protected with a single lock for
  10 an entire ldiskfs directory. PDO project will remove this
  11 bottleneck by introducing a parallel locking mechanism for entire
  12 ldiskfs directories. This work will enable multiple application
  13 threads to simultaneously lookup, create and unlink in parallel.
  14
  15 This patch contains:
  16   - pdirops support for ldiskfs
  17   - N-level htree directory
  18   - integrate with osd-ldiskfs
  19
  20 Signed-off-by: Liang Zhen <liang@whamcloud.com>
  21 Change-Id: I269c0e3112e68f3acd79e860dab052a68c7d7aaa
  22 Reviewed-on: http://review.whamcloud.com/375
  23 Reviewed-by: Andreas Dilger <adilger@whamcloud.com>
  24 ---
  25  fs/ext4/Makefile |    1
  26  fs/ext4/ext4.h   |   78 +++++++++
  27  fs/ext4/namei.c  |  454 ++++++++++++++++++++++++++++++++++++++++++++++++++-----
  28  fs/ext4/super.c  |    1
  29  4 files changed, 494 insertions(+), 40 deletions(-)
  30  create mode 100644 fs/ext4/htree_lock.c
  31  create mode 100644 include/linux/htree_lock.h
  32
  33 --- a/fs/ext4/Makefile
  34 +++ b/fs/ext4/Makefile
  35 @@ -7,6 +7,7 @@ obj-$(CONFIG_EXT4_FS) += ext4.o
  36
  37  ext4-y := balloc.o bitmap.o block_validity.o dir.o ext4_jbd2.o extents.o \
  38                 extents_status.o file.o fsmap.o fsync.o hash.o ialloc.o \
  39 +               htree_lock.o \
  40                 indirect.o inline.o inode.o ioctl.o mballoc.o migrate.o \
  41                 mmp.o move_extent.o namei.o page-io.o readpage.o resize.o \
  42                 super.o symlink.o sysfs.o xattr.o xattr_hurd.o xattr_trusted.o \
  43 --- a/fs/ext4/ext4.h
  44 +++ b/fs/ext4/ext4.h
  45 @@ -28,6 +28,7 @@
  46  #include <linux/mutex.h>
  47  #include <linux/timer.h>
  48  #include <linux/wait.h>
  49 +#include <linux/htree_lock.h>
  50  #include <linux/sched/signal.h>
  51  #include <linux/blockgroup_lock.h>
  52  #include <linux/percpu_counter.h>
  53 @@ -1003,6 +1004,9 @@ struct ext4_inode_info {
  54         __u32   i_dtime;
  55         ext4_fsblk_t    i_file_acl;
  56
  57 +       /* following fields for parallel directory operations -bzzz */
  58 +       struct semaphore i_append_sem;
  59 +
  60         /*
  61          * i_block_group is the number of the block group which contains
  62          * this file's inode.  Constant across the lifetime of the inode,
  63 @@ -2398,6 +2402,72 @@ struct dx_hash_info
  64   */
  65  #define HASH_NB_ALWAYS         1
  66
  67 +/* assume name-hash is protected by upper layer */
  68 +#define EXT4_HTREE_LOCK_HASH   0
  69 +
  70 +enum ext4_pdo_lk_types {
  71 +#if EXT4_HTREE_LOCK_HASH
  72 +       EXT4_LK_HASH,
  73 +#endif
  74 +       EXT4_LK_DX,             /* index block */
  75 +       EXT4_LK_DE,             /* directory entry block */
  76 +       EXT4_LK_SPIN,           /* spinlock */
  77 +       EXT4_LK_MAX,
  78 +};
  79 +
  80 +/* read-only bit */
  81 +#define EXT4_LB_RO(b)          (1 << (b))
  82 +/* read + write, high bits for writer */
  83 +#define EXT4_LB_RW(b)          ((1 << (b)) | (1 << (EXT4_LK_MAX + (b))))
  84 +
  85 +enum ext4_pdo_lock_bits {
  86 +       /* DX lock bits */
  87 +       EXT4_LB_DX_RO           = EXT4_LB_RO(EXT4_LK_DX),
  88 +       EXT4_LB_DX              = EXT4_LB_RW(EXT4_LK_DX),
  89 +       /* DE lock bits */
  90 +       EXT4_LB_DE_RO           = EXT4_LB_RO(EXT4_LK_DE),
  91 +       EXT4_LB_DE              = EXT4_LB_RW(EXT4_LK_DE),
  92 +       /* DX spinlock bits */
  93 +       EXT4_LB_SPIN_RO         = EXT4_LB_RO(EXT4_LK_SPIN),
  94 +       EXT4_LB_SPIN            = EXT4_LB_RW(EXT4_LK_SPIN),
  95 +       /* accurate searching */
  96 +       EXT4_LB_EXACT           = EXT4_LB_RO(EXT4_LK_MAX << 1),
  97 +};
  98 +
  99 +enum ext4_pdo_lock_opc {
 100 +       /* external */
 101 +       EXT4_HLOCK_READDIR      = (EXT4_LB_DE_RO | EXT4_LB_DX_RO),
 102 +       EXT4_HLOCK_LOOKUP       = (EXT4_LB_DE_RO | EXT4_LB_SPIN_RO |
 103 +                                  EXT4_LB_EXACT),
 104 +       EXT4_HLOCK_DEL          = (EXT4_LB_DE | EXT4_LB_SPIN_RO |
 105 +                                  EXT4_LB_EXACT),
 106 +       EXT4_HLOCK_ADD          = (EXT4_LB_DE | EXT4_LB_SPIN_RO),
 107 +
 108 +       /* internal */
 109 +       EXT4_HLOCK_LOOKUP_SAFE  = (EXT4_LB_DE_RO | EXT4_LB_DX_RO |
 110 +                                  EXT4_LB_EXACT),
 111 +       EXT4_HLOCK_DEL_SAFE     = (EXT4_LB_DE | EXT4_LB_DX_RO | EXT4_LB_EXACT),
 112 +       EXT4_HLOCK_SPLIT        = (EXT4_LB_DE | EXT4_LB_DX | EXT4_LB_SPIN),
 113 +};
 114 +
 115 +extern struct htree_lock_head *ext4_htree_lock_head_alloc(unsigned hbits);
 116 +#define ext4_htree_lock_head_free(lhead)       htree_lock_head_free(lhead)
 117 +
 118 +extern struct htree_lock *ext4_htree_lock_alloc(void);
 119 +#define ext4_htree_lock_free(lck)              htree_lock_free(lck)
 120 +
 121 +extern void ext4_htree_lock(struct htree_lock *lck,
 122 +                           struct htree_lock_head *lhead,
 123 +                           struct inode *dir, unsigned flags);
 124 +#define ext4_htree_unlock(lck)                  htree_unlock(lck)
 125 +
 126 +extern struct buffer_head *ext4_find_entry_locked(struct inode *dir,
 127 +                                       const struct qstr *d_name,
 128 +                                       struct ext4_dir_entry_2 **res_dir,
 129 +                                       int *inlined, struct htree_lock *lck);
 130 +extern int ext4_add_entry_locked(handle_t *handle, struct dentry *dentry,
 131 +                     struct inode *inode, struct htree_lock *lck);
 132 +
 133  struct ext4_filename {
 134         const struct qstr *usr_fname;
 135         struct fscrypt_str disk_name;
 136 @@ -2772,12 +2842,20 @@ void ext4_insert_dentry(struct inode *in
 137                         struct ext4_filename *fname, void *data);
 138  static inline void ext4_update_dx_flag(struct inode *inode)
 139  {
 140 +       /* Disable it for ldiskfs, because going from a DX directory to
 141 +        * a non-DX directory while it is in use will completely break
 142 +        * the htree-locking.
 143 +        * If we really want to support this operation in the future,
 144 +        * we need to exclusively lock the directory at here which will
 145 +        * increase complexity of code */
 146 +#if 0
 147         if (!ext4_has_feature_dir_index(inode->i_sb) &&
 148             ext4_test_inode_flag(inode, EXT4_INODE_INDEX)) {
 149                 /* ext4_iget() should have caught this... */
 150                 WARN_ON_ONCE(ext4_has_feature_metadata_csum(inode->i_sb));
 151                 ext4_clear_inode_flag(inode, EXT4_INODE_INDEX);
 152         }
 153 +#endif
 154  }
 155  static const unsigned char ext4_filetype_table[] = {
 156         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
 157 --- a/fs/ext4/namei.c
 158 +++ b/fs/ext4/namei.c
 159 @@ -55,6 +55,7 @@ struct buffer_head *ext4_append(handle_t
 160                                         ext4_lblk_t *block)
 161  {
 162         struct buffer_head *bh;
 163 +       struct ext4_inode_info *ei = EXT4_I(inode);
 164         int err;
 165
 166         if (unlikely(EXT4_SB(inode->i_sb)->s_max_dir_size_kb &&
 167 @@ -62,15 +63,22 @@ struct buffer_head *ext4_append(handle_t
 168                       EXT4_SB(inode->i_sb)->s_max_dir_size_kb)))
 169                 return ERR_PTR(-ENOSPC);
 170
 171 +       /* with parallel dir operations all appends
 172 +       * have to be serialized -bzzz */
 173 +       down(&ei->i_append_sem);
 174 +
 175         *block = inode->i_size >> inode->i_sb->s_blocksize_bits;
 176
 177         bh = ext4_bread(handle, inode, *block, EXT4_GET_BLOCKS_CREATE);
 178 -       if (IS_ERR(bh))
 179 +       if (IS_ERR(bh)) {
 180 +               up(&ei->i_append_sem);
 181                 return bh;
 182 +       }
 183         inode->i_size += inode->i_sb->s_blocksize;
 184         EXT4_I(inode)->i_disksize = inode->i_size;
 185         BUFFER_TRACE(bh, "get_write_access");
 186         err = ext4_journal_get_write_access(handle, bh);
 187 +       up(&ei->i_append_sem);
 188         if (err) {
 189                 brelse(bh);
 190                 ext4_std_error(inode->i_sb, err);
 191 @@ -271,7 +279,8 @@ static unsigned dx_node_limit(struct ino
 192  static struct dx_frame *dx_probe(struct ext4_filename *fname,
 193                                  struct inode *dir,
 194                                  struct dx_hash_info *hinfo,
 195 -                                struct dx_frame *frame);
 196 +                                struct dx_frame *frame,
 197 +                                struct htree_lock *lck);
 198  static void dx_release(struct dx_frame *frames);
 199  static int dx_make_map(struct inode *dir, struct ext4_dir_entry_2 *de,
 200                        unsigned blocksize, struct dx_hash_info *hinfo,
 201 @@ -285,12 +294,13 @@ static void dx_insert_block(struct dx_fr
 202  static int ext4_htree_next_block(struct inode *dir, __u32 hash,
 203                                  struct dx_frame *frame,
 204                                  struct dx_frame *frames,
 205 -                                __u32 *start_hash);
 206 +                                __u32 *start_hash, struct htree_lock *lck);
 207  static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
 208                 struct ext4_filename *fname,
 209 -               struct ext4_dir_entry_2 **res_dir);
 210 +               struct ext4_dir_entry_2 **res_dir, struct htree_lock *lck);
 211  static int ext4_dx_add_entry(handle_t *handle, struct ext4_filename *fname,
 212 -                            struct inode *dir, struct inode *inode);
 213 +                            struct inode *dir, struct inode *inode,
 214 +                            struct htree_lock *lck);
 215
 216  /* checksumming functions */
 217  void ext4_initialize_dirent_tail(struct buffer_head *bh,
 218 @@ -754,6 +764,227 @@ struct stats dx_show_entries(struct dx_h
 219  }
 220  #endif /* DX_DEBUG */
 221
 222 +/* private data for htree_lock */
 223 +struct ext4_dir_lock_data {
 224 +       unsigned                ld_flags;  /* bits-map for lock types */
 225 +       unsigned                ld_count;  /* # entries of the last DX block */
 226 +       struct dx_entry         ld_at_entry; /* copy of leaf dx_entry */
 227 +       struct dx_entry         *ld_at;    /* position of leaf dx_entry */
 228 +};
 229 +
 230 +#define ext4_htree_lock_data(l)        ((struct ext4_dir_lock_data *)(l)->lk_private)
 231 +#define ext4_find_entry(dir, name, dirent, inline) \
 232 +                       ext4_find_entry_locked(dir, name, dirent, inline, NULL)
 233 +#define ext4_add_entry(handle, dentry, inode) \
 234 +                       ext4_add_entry_locked(handle, dentry, inode, NULL)
 235 +
 236 +/* NB: ext4_lblk_t is 32 bits so we use high bits to identify invalid blk */
 237 +#define EXT4_HTREE_NODE_CHANGED        (0xcafeULL << 32)
 238 +
 239 +static void ext4_htree_event_cb(void *target, void *event)
 240 +{
 241 +       u64 *block = (u64 *)target;
 242 +
 243 +       if (*block == dx_get_block((struct dx_entry *)event))
 244 +               *block = EXT4_HTREE_NODE_CHANGED;
 245 +}
 246 +
 247 +struct htree_lock_head *ext4_htree_lock_head_alloc(unsigned hbits)
 248 +{
 249 +       struct htree_lock_head *lhead;
 250 +
 251 +       lhead = htree_lock_head_alloc(EXT4_LK_MAX, hbits, 0);
 252 +       if (lhead != NULL) {
 253 +               htree_lock_event_attach(lhead, EXT4_LK_SPIN, HTREE_EVENT_WR,
 254 +                                       ext4_htree_event_cb);
 255 +       }
 256 +       return lhead;
 257 +}
 258 +EXPORT_SYMBOL(ext4_htree_lock_head_alloc);
 259 +
 260 +struct htree_lock *ext4_htree_lock_alloc(void)
 261 +{
 262 +       return htree_lock_alloc(EXT4_LK_MAX,
 263 +                               sizeof(struct ext4_dir_lock_data));
 264 +}
 265 +EXPORT_SYMBOL(ext4_htree_lock_alloc);
 266 +
 267 +static htree_lock_mode_t ext4_htree_mode(unsigned flags)
 268 +{
 269 +       switch (flags) {
 270 +       default: /* 0 or unknown flags require EX lock */
 271 +               return HTREE_LOCK_EX;
 272 +       case EXT4_HLOCK_READDIR:
 273 +               return HTREE_LOCK_PR;
 274 +       case EXT4_HLOCK_LOOKUP:
 275 +               return HTREE_LOCK_CR;
 276 +       case EXT4_HLOCK_DEL:
 277 +       case EXT4_HLOCK_ADD:
 278 +               return HTREE_LOCK_CW;
 279 +       }
 280 +}
 281 +
 282 +/* return PR for read-only operations, otherwise return EX */
 283 +static inline htree_lock_mode_t ext4_htree_safe_mode(unsigned flags)
 284 +{
 285 +       int writer = (flags & EXT4_LB_DE) == EXT4_LB_DE;
 286 +
 287 +       /* 0 requires EX lock */
 288 +       return (flags == 0 || writer) ? HTREE_LOCK_EX : HTREE_LOCK_PR;
 289 +}
 290 +
 291 +static int ext4_htree_safe_locked(struct htree_lock *lck)
 292 +{
 293 +       int writer;
 294 +
 295 +       if (lck == NULL || lck->lk_mode == HTREE_LOCK_EX)
 296 +               return 1;
 297 +
 298 +       writer = (ext4_htree_lock_data(lck)->ld_flags & EXT4_LB_DE) ==
 299 +                EXT4_LB_DE;
 300 +       if (writer) /* all readers & writers are excluded? */
 301 +               return lck->lk_mode == HTREE_LOCK_EX;
 302 +
 303 +       /* all writers are excluded? */
 304 +       return lck->lk_mode == HTREE_LOCK_PR ||
 305 +              lck->lk_mode == HTREE_LOCK_PW ||
 306 +              lck->lk_mode == HTREE_LOCK_EX;
 307 +}
 308 +
 309 +/* relock htree_lock with EX mode if it's change operation, otherwise
 310 + * relock it with PR mode. It's noop if PDO is disabled. */
 311 +static void ext4_htree_safe_relock(struct htree_lock *lck)
 312 +{
 313 +       if (!ext4_htree_safe_locked(lck)) {
 314 +               unsigned flags = ext4_htree_lock_data(lck)->ld_flags;
 315 +
 316 +               htree_change_lock(lck, ext4_htree_safe_mode(flags));
 317 +       }
 318 +}
 319 +
 320 +void ext4_htree_lock(struct htree_lock *lck, struct htree_lock_head *lhead,
 321 +                    struct inode *dir, unsigned flags)
 322 +{
 323 +       htree_lock_mode_t mode = is_dx(dir) ? ext4_htree_mode(flags) :
 324 +                                             ext4_htree_safe_mode(flags);
 325 +
 326 +       ext4_htree_lock_data(lck)->ld_flags = flags;
 327 +       htree_lock(lck, lhead, mode);
 328 +       if (!is_dx(dir))
 329 +               ext4_htree_safe_relock(lck); /* make sure it's safe locked */
 330 +}
 331 +EXPORT_SYMBOL(ext4_htree_lock);
 332 +
 333 +static int ext4_htree_node_lock(struct htree_lock *lck, struct dx_entry *at,
 334 +                               unsigned lmask, int wait, void *ev)
 335 +{
 336 +       u32     key = (at == NULL) ? 0 : dx_get_block(at);
 337 +       u32     mode;
 338 +
 339 +       /* NOOP if htree is well protected or caller doesn't require the lock */
 340 +       if (ext4_htree_safe_locked(lck) ||
 341 +          !(ext4_htree_lock_data(lck)->ld_flags & lmask))
 342 +               return 1;
 343 +
 344 +       mode = (ext4_htree_lock_data(lck)->ld_flags & lmask) == lmask ?
 345 +               HTREE_LOCK_PW : HTREE_LOCK_PR;
 346 +       while (1) {
 347 +               if (htree_node_lock_try(lck, mode, key, ffz(~lmask), wait, ev))
 348 +                       return 1;
 349 +               if (!(lmask & EXT4_LB_SPIN)) /* not a spinlock */
 350 +                       return 0;
 351 +               cpu_relax(); /* spin until granted */
 352 +       }
 353 +}
 354 +
 355 +static int ext4_htree_node_locked(struct htree_lock *lck, unsigned lmask)
 356 +{
 357 +       return ext4_htree_safe_locked(lck) ||
 358 +              htree_node_is_granted(lck, ffz(~lmask));
 359 +}
 360 +
 361 +static void ext4_htree_node_unlock(struct htree_lock *lck,
 362 +                                  unsigned lmask, void *buf)
 363 +{
 364 +       /* NB: it's safe to call mutiple times or even it's not locked */
 365 +       if (!ext4_htree_safe_locked(lck) &&
 366 +            htree_node_is_granted(lck, ffz(~lmask)))
 367 +               htree_node_unlock(lck, ffz(~lmask), buf);
 368 +}
 369 +
 370 +#define ext4_htree_dx_lock(lck, key)           \
 371 +       ext4_htree_node_lock(lck, key, EXT4_LB_DX, 1, NULL)
 372 +#define ext4_htree_dx_lock_try(lck, key)       \
 373 +       ext4_htree_node_lock(lck, key, EXT4_LB_DX, 0, NULL)
 374 +#define ext4_htree_dx_unlock(lck)              \
 375 +       ext4_htree_node_unlock(lck, EXT4_LB_DX, NULL)
 376 +#define ext4_htree_dx_locked(lck)              \
 377 +       ext4_htree_node_locked(lck, EXT4_LB_DX)
 378 +
 379 +static void ext4_htree_dx_need_lock(struct htree_lock *lck)
 380 +{
 381 +       struct ext4_dir_lock_data *ld;
 382 +
 383 +       if (ext4_htree_safe_locked(lck))
 384 +               return;
 385 +
 386 +       ld = ext4_htree_lock_data(lck);
 387 +       switch (ld->ld_flags) {
 388 +       default:
 389 +               return;
 390 +       case EXT4_HLOCK_LOOKUP:
 391 +               ld->ld_flags = EXT4_HLOCK_LOOKUP_SAFE;
 392 +               return;
 393 +       case EXT4_HLOCK_DEL:
 394 +               ld->ld_flags = EXT4_HLOCK_DEL_SAFE;
 395 +               return;
 396 +       case EXT4_HLOCK_ADD:
 397 +               ld->ld_flags = EXT4_HLOCK_SPLIT;
 398 +               return;
 399 +       }
 400 +}
 401 +
 402 +#define ext4_htree_de_lock(lck, key)           \
 403 +       ext4_htree_node_lock(lck, key, EXT4_LB_DE, 1, NULL)
 404 +#define ext4_htree_de_unlock(lck)              \
 405 +       ext4_htree_node_unlock(lck, EXT4_LB_DE, NULL)
 406 +
 407 +#define ext4_htree_spin_lock(lck, key, event)  \
 408 +       ext4_htree_node_lock(lck, key, EXT4_LB_SPIN, 0, event)
 409 +#define ext4_htree_spin_unlock(lck)            \
 410 +       ext4_htree_node_unlock(lck, EXT4_LB_SPIN, NULL)
 411 +#define ext4_htree_spin_unlock_listen(lck, p)  \
 412 +       ext4_htree_node_unlock(lck, EXT4_LB_SPIN, p)
 413 +
 414 +static void ext4_htree_spin_stop_listen(struct htree_lock *lck)
 415 +{
 416 +       if (!ext4_htree_safe_locked(lck) &&
 417 +           htree_node_is_listening(lck, ffz(~EXT4_LB_SPIN)))
 418 +               htree_node_stop_listen(lck, ffz(~EXT4_LB_SPIN));
 419 +}
 420 +
 421 +enum {
 422 +       DX_HASH_COL_IGNORE,     /* ignore collision while probing frames */
 423 +       DX_HASH_COL_YES,        /* there is collision and it does matter */
 424 +       DX_HASH_COL_NO,         /* there is no collision */
 425 +};
 426 +
 427 +static int dx_probe_hash_collision(struct htree_lock *lck,
 428 +                                  struct dx_entry *entries,
 429 +                                  struct dx_entry *at, u32 hash)
 430 +{
 431 +       if (!(lck && ext4_htree_lock_data(lck)->ld_flags & EXT4_LB_EXACT)) {
 432 +               return DX_HASH_COL_IGNORE; /* don't care about collision */
 433 +
 434 +       } else if (at == entries + dx_get_count(entries) - 1) {
 435 +               return DX_HASH_COL_IGNORE; /* not in any leaf of this DX */
 436 +
 437 +       } else { /* hash collision? */
 438 +               return ((dx_get_hash(at + 1) & ~1) == hash) ?
 439 +                       DX_HASH_COL_YES : DX_HASH_COL_NO;
 440 +       }
 441 +}
 442 +
 443  /*
 444   * Probe for a directory leaf block to search.
 445   *
 446 @@ -765,10 +996,11 @@ struct stats dx_show_entries(struct dx_h
 447   */
 448  static struct dx_frame *
 449  dx_probe(struct ext4_filename *fname, struct inode *dir,
 450 -        struct dx_hash_info *hinfo, struct dx_frame *frame_in)
 451 +        struct dx_hash_info *hinfo, struct dx_frame *frame_in,
 452 +        struct htree_lock *lck)
 453  {
 454         unsigned count, indirect;
 455 -       struct dx_entry *at, *entries, *p, *q, *m;
 456 +       struct dx_entry *at, *entries, *p, *q, *m, *dx = NULL;
 457         struct dx_root_info *info;
 458         struct dx_frame *frame = frame_in;
 459         struct dx_frame *ret_err = ERR_PTR(ERR_BAD_DX_DIR);
 460 @@ -830,8 +1062,15 @@ dx_probe(struct ext4_filename *fname, st
 461
 462         dxtrace(printk("Look up %x", hash));
 463         while (1) {
 464 +               if (indirect == 0) { /* the last index level */
 465 +                       /* NB: ext4_htree_dx_lock() could be noop if
 466 +                        * DX-lock flag is not set for current operation */
 467 +                       ext4_htree_dx_lock(lck, dx);
 468 +                       ext4_htree_spin_lock(lck, dx, NULL);
 469 +               }
 470                 count = dx_get_count(entries);
 471 -               if (!count || count > dx_get_limit(entries)) {
 472 +               if (count == 0 || count > dx_get_limit(entries)) {
 473 +                       ext4_htree_spin_unlock(lck); /* release spin */
 474                         ext4_warning_inode(dir,
 475                                            "dx entry: count %u beyond limit %u",
 476                                            count, dx_get_limit(entries));
 477 @@ -870,8 +1109,70 @@ dx_probe(struct ext4_filename *fname, st
 478                                dx_get_block(at)));
 479                 frame->entries = entries;
 480                 frame->at = at;
 481 -               if (!indirect--)
 482 +
 483 +               if (indirect == 0) { /* the last index level */
 484 +                       struct ext4_dir_lock_data *ld;
 485 +                       u64 myblock;
 486 +
 487 +                       /* By default we only lock DE-block, however, we will
 488 +                        * also lock the last level DX-block if:
 489 +                        * a) there is hash collision
 490 +                        *    we will set DX-lock flag (a few lines below)
 491 +                        *    and redo to lock DX-block
 492 +                        *    see detail in dx_probe_hash_collision()
 493 +                        * b) it's a retry from splitting
 494 +                        *    we need to lock the last level DX-block so nobody
 495 +                        *    else can split any leaf blocks under the same
 496 +                        *    DX-block, see detail in ext4_dx_add_entry()
 497 +                        */
 498 +                       if (ext4_htree_dx_locked(lck)) {
 499 +                               /* DX-block is locked, just lock DE-block
 500 +                                * and return */
 501 +                               ext4_htree_spin_unlock(lck);
 502 +                               if (!ext4_htree_safe_locked(lck))
 503 +                                       ext4_htree_de_lock(lck, frame->at);
 504 +                               return frame;
 505 +                       }
 506 +                       /* it's pdirop and no DX lock */
 507 +                       if (dx_probe_hash_collision(lck, entries, at, hash) ==
 508 +                           DX_HASH_COL_YES) {
 509 +                               /* found hash collision, set DX-lock flag
 510 +                                * and retry to abtain DX-lock */
 511 +                               ext4_htree_spin_unlock(lck);
 512 +                               ext4_htree_dx_need_lock(lck);
 513 +                               continue;
 514 +                       }
 515 +                       ld = ext4_htree_lock_data(lck);
 516 +                       /* because I don't lock DX, so @at can't be trusted
 517 +                        * after I release spinlock so I have to save it */
 518 +                       ld->ld_at = at;
 519 +                       ld->ld_at_entry = *at;
 520 +                       ld->ld_count = dx_get_count(entries);
 521 +
 522 +                       frame->at = &ld->ld_at_entry;
 523 +                       myblock = dx_get_block(at);
 524 +
 525 +                       /* NB: ordering locking */
 526 +                       ext4_htree_spin_unlock_listen(lck, &myblock);
 527 +                       /* other thread can split this DE-block because:
 528 +                        * a) I don't have lock for the DE-block yet
 529 +                        * b) I released spinlock on DX-block
 530 +                        * if it happened I can detect it by listening
 531 +                        * splitting event on this DE-block */
 532 +                       ext4_htree_de_lock(lck, frame->at);
 533 +                       ext4_htree_spin_stop_listen(lck);
 534 +
 535 +                       if (myblock == EXT4_HTREE_NODE_CHANGED) {
 536 +                               /* someone split this DE-block before
 537 +                                * I locked it, I need to retry and lock
 538 +                                * valid DE-block */
 539 +                               ext4_htree_de_unlock(lck);
 540 +                               continue;
 541 +                       }
 542                         return frame;
 543 +               }
 544 +               dx = at;
 545 +               indirect--;
 546                 frame++;
 547                 frame->bh = ext4_read_dirblock(dir, dx_get_block(at), INDEX);
 548                 if (IS_ERR(frame->bh)) {
 549 @@ -940,7 +1241,7 @@ static void dx_release(struct dx_frame *
 550  static int ext4_htree_next_block(struct inode *dir, __u32 hash,
 551                                  struct dx_frame *frame,
 552                                  struct dx_frame *frames,
 553 -                                __u32 *start_hash)
 554 +                                __u32 *start_hash, struct htree_lock *lck)
 555  {
 556         struct dx_frame *p;
 557         struct buffer_head *bh;
 558 @@ -955,12 +1256,22 @@ static int ext4_htree_next_block(struct
 559          * this loop, num_frames indicates the number of interior
 560          * nodes need to be read.
 561          */
 562 +       ext4_htree_de_unlock(lck);
 563         while (1) {
 564 -               if (++(p->at) < p->entries + dx_get_count(p->entries))
 565 -                       break;
 566 +               if (num_frames > 0 || ext4_htree_dx_locked(lck)) {
 567 +                       /* num_frames > 0 :
 568 +                        *   DX block
 569 +                        * ext4_htree_dx_locked:
 570 +                        *   frame->at is reliable pointer returned by dx_probe,
 571 +                        *   otherwise dx_probe already knew no collision */
 572 +                       if (++(p->at) < p->entries + dx_get_count(p->entries))
 573 +                               break;
 574 +               }
 575                 if (p == frames)
 576                         return 0;
 577                 num_frames++;
 578 +               if (num_frames == 1)
 579 +                       ext4_htree_dx_unlock(lck);
 580                 p--;
 581         }
 582
 583 @@ -983,6 +1294,13 @@ static int ext4_htree_next_block(struct
 584          * block so no check is necessary
 585          */
 586         while (num_frames--) {
 587 +               if (num_frames == 0) {
 588 +                       /* it's not always necessary, we just don't want to
 589 +                        * detect hash collision again */
 590 +                       ext4_htree_dx_need_lock(lck);
 591 +                       ext4_htree_dx_lock(lck, p->at);
 592 +               }
 593 +
 594                 bh = ext4_read_dirblock(dir, dx_get_block(p->at), INDEX);
 595                 if (IS_ERR(bh))
 596                         return PTR_ERR(bh);
 597 @@ -991,6 +1309,7 @@ static int ext4_htree_next_block(struct
 598                 p->bh = bh;
 599                 p->at = p->entries = ((struct dx_node *) bh->b_data)->entries;
 600         }
 601 +       ext4_htree_de_lock(lck, p->at);
 602         return 1;
 603  }
 604
 605 @@ -1135,10 +1454,10 @@ int ext4_htree_fill_tree(struct file *di
 606         }
 607         hinfo.hash = start_hash;
 608         hinfo.minor_hash = 0;
 609 -       frame = dx_probe(NULL, dir, &hinfo, frames);
 610 +       /* assume it's PR locked */
 611 +       frame = dx_probe(NULL, dir, &hinfo, frames, NULL);
 612         if (IS_ERR(frame))
 613                 return PTR_ERR(frame);
 614 -
 615         /* Add '.' and '..' from the htree header */
 616         if (!start_hash && !start_minor_hash) {
 617                 de = (struct ext4_dir_entry_2 *) frames[0].bh->b_data;
 618 @@ -1178,7 +1497,7 @@ int ext4_htree_fill_tree(struct file *di
 619                 count += ret;
 620                 hashval = ~0;
 621                 ret = ext4_htree_next_block(dir, HASH_NB_ALWAYS,
 622 -                                           frame, frames, &hashval);
 623 +                                           frame, frames, &hashval, NULL);
 624                 *next_hash = hashval;
 625                 if (ret < 0) {
 626                         err = ret;
 627 @@ -1454,7 +1773,7 @@ static int is_dx_internal_node(struct in
 628  static struct buffer_head *__ext4_find_entry(struct inode *dir,
 629                                              struct ext4_filename *fname,
 630                                              struct ext4_dir_entry_2 **res_dir,
 631 -                                            int *inlined)
 632 +                                            int *inlined, struct htree_lock *lck)
 633  {
 634         struct super_block *sb;
 635         struct buffer_head *bh_use[NAMEI_RA_SIZE];
 636 @@ -1496,7 +1815,7 @@ static struct buffer_head *__ext4_find_e
 637                 goto restart;
 638         }
 639         if (is_dx(dir)) {
 640 -               ret = ext4_dx_find_entry(dir, fname, res_dir);
 641 +               ret = ext4_dx_find_entry(dir, fname, res_dir, lck);
 642                 /*
 643                  * On success, or if the error was file not found,
 644                  * return.  Otherwise, fall back to doing a search the
 645 @@ -1506,6 +1825,7 @@ static struct buffer_head *__ext4_find_e
 646                         goto cleanup_and_exit;
 647                 dxtrace(printk(KERN_DEBUG "ext4_find_entry: dx failed, "
 648                                "falling back\n"));
 649 +               ext4_htree_safe_relock(lck);
 650                 ret = NULL;
 651         }
 652         nblocks = dir->i_size >> EXT4_BLOCK_SIZE_BITS(sb);
 653 @@ -1596,10 +1916,10 @@ cleanup_and_exit:
 654         return ret;
 655  }
 656
 657 -static struct buffer_head *ext4_find_entry(struct inode *dir,
 658 +struct buffer_head *ext4_find_entry_locked(struct inode *dir,
 659                                            const struct qstr *d_name,
 660                                            struct ext4_dir_entry_2 **res_dir,
 661 -                                          int *inlined)
 662 +                                          int *inlined, struct htree_lock *lck)
 663  {
 664         int err;
 665         struct ext4_filename fname;
 666 @@ -1611,12 +1931,14 @@ static struct buffer_head *ext4_find_ent
 667         if (err)
 668                 return ERR_PTR(err);
 669
 670 -       bh = __ext4_find_entry(dir, &fname, res_dir, inlined);
 671 +       bh = __ext4_find_entry(dir, &fname, res_dir, inlined, lck);
 672
 673         ext4_fname_free_filename(&fname);
 674         return bh;
 675  }
 676
 677 +EXPORT_SYMBOL(ext4_find_entry_locked);
 678 +
 679  static struct buffer_head *ext4_lookup_entry(struct inode *dir,
 680                                              struct dentry *dentry,
 681                                              struct ext4_dir_entry_2 **res_dir)
 682 @@ -1631,7 +1953,7 @@ static struct buffer_head *ext4_lookup_e
 683         if (err)
 684                 return ERR_PTR(err);
 685
 686 -       bh = __ext4_find_entry(dir, &fname, res_dir, NULL);
 687 +       bh = __ext4_find_entry(dir, &fname, res_dir, NULL, NULL);
 688
 689         ext4_fname_free_filename(&fname);
 690         return bh;
 691 @@ -1639,7 +1961,8 @@ static struct buffer_head *ext4_lookup_e
 692
 693  static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
 694                         struct ext4_filename *fname,
 695 -                       struct ext4_dir_entry_2 **res_dir)
 696 +                       struct ext4_dir_entry_2 **res_dir,
 697 +                       struct htree_lock *lck)
 698  {
 699         struct super_block * sb = dir->i_sb;
 700         struct dx_frame frames[EXT4_HTREE_LEVEL], *frame;
 701 @@ -1650,7 +1973,7 @@ static struct buffer_head * ext4_dx_find
 702  #ifdef CONFIG_FS_ENCRYPTION
 703         *res_dir = NULL;
 704  #endif
 705 -       frame = dx_probe(fname, dir, NULL, frames);
 706 +       frame = dx_probe(fname, dir, NULL, frames, lck);
 707         if (IS_ERR(frame))
 708                 return (struct buffer_head *) frame;
 709         do {
 710 @@ -1672,7 +1995,7 @@ static struct buffer_head * ext4_dx_find
 711
 712                 /* Check to see if we should continue to search */
 713                 retval = ext4_htree_next_block(dir, fname->hinfo.hash, frame,
 714 -                                              frames, NULL);
 715 +                                              frames, NULL, lck);
 716                 if (retval < 0) {
 717                         ext4_warning_inode(dir,
 718                                 "error %d reading directory index block",
 719 @@ -1852,8 +2175,9 @@ static struct ext4_dir_entry_2* dx_pack_
 720   * Returns pointer to de in block into which the new entry will be inserted.
 721   */
 722  static struct ext4_dir_entry_2 *do_split(handle_t *handle, struct inode *dir,
 723 -                       struct buffer_head **bh,struct dx_frame *frame,
 724 -                       struct dx_hash_info *hinfo)
 725 +                       struct buffer_head **bh, struct dx_frame *frames,
 726 +                       struct dx_frame *frame, struct dx_hash_info *hinfo,
 727 +                       struct htree_lock *lck)
 728  {
 729         unsigned blocksize = dir->i_sb->s_blocksize;
 730         unsigned count, continued;
 731 @@ -1924,8 +2248,14 @@ static struct ext4_dir_entry_2 *do_split
 732                                         hash2, split, count-split));
 733
 734         /* Fancy dance to stay within two buffers */
 735 -       de2 = dx_move_dirents(data1, data2, map + split, count - split,
 736 -                             blocksize);
 737 +       if (hinfo->hash < hash2) {
 738 +               de2 = dx_move_dirents(data1, data2, map + split,
 739 +                                     count - split, blocksize);
 740 +       } else {
 741 +               /* make sure we will add entry to the same block which
 742 +                * we have already locked */
 743 +               de2 = dx_move_dirents(data1, data2, map, split, blocksize);
 744 +       }
 745         de = dx_pack_dirents(data1, blocksize);
 746         de->rec_len = ext4_rec_len_to_disk(data1 + (blocksize - csum_size) -
 747                                            (char *) de,
 748 @@ -1943,12 +2273,21 @@ static struct ext4_dir_entry_2 *do_split
 749         dxtrace(dx_show_leaf(dir, hinfo, (struct ext4_dir_entry_2 *) data2,
 750                         blocksize, 1));
 751
 752 -       /* Which block gets the new entry? */
 753 -       if (hinfo->hash >= hash2) {
 754 -               swap(*bh, bh2);
 755 -               de = de2;
 756 +       ext4_htree_spin_lock(lck, frame > frames ? (frame - 1)->at : NULL,
 757 +                            frame->at); /* notify block is being split */
 758 +       if (hinfo->hash < hash2) {
 759 +               dx_insert_block(frame, hash2 + continued, newblock);
 760 +
 761 +       } else {
 762 +               /* switch block number */
 763 +               dx_insert_block(frame, hash2 + continued,
 764 +                               dx_get_block(frame->at));
 765 +               dx_set_block(frame->at, newblock);
 766 +               (frame->at)++;
 767         }
 768 -       dx_insert_block(frame, hash2 + continued, newblock);
 769 +       ext4_htree_spin_unlock(lck);
 770 +       ext4_htree_dx_unlock(lck);
 771 +
 772         err = ext4_handle_dirty_dirblock(handle, dir, bh2);
 773         if (err)
 774                 goto journal_error;
 775 @@ -2218,7 +2557,7 @@ static int make_indexed_dir(handle_t *ha
 776         if (retval)
 777                 goto out_frames;
 778
 779 -       de = do_split(handle,dir, &bh2, frame, &fname->hinfo);
 780 +       de = do_split(handle, dir, &bh2, frames, frame, &fname->hinfo, NULL);
 781         if (IS_ERR(de)) {
 782                 retval = PTR_ERR(de);
 783                 goto out_frames;
 784 @@ -2328,8 +2667,8 @@ out:
 785   * may not sleep between calling this and putting something into
 786   * the entry, as someone else might have used it while you slept.
 787   */
 788 -static int ext4_add_entry(handle_t *handle, struct dentry *dentry,
 789 -                         struct inode *inode)
 790 +int ext4_add_entry_locked(handle_t *handle, struct dentry *dentry,
 791 +                         struct inode *inode, struct htree_lock *lck)
 792  {
 793         struct inode *dir = d_inode(dentry->d_parent);
 794         struct buffer_head *bh = NULL;
 795 @@ -2375,9 +2714,10 @@ static int ext4_add_entry(handle_t *hand
 796                 return ext4_update_dotdot(handle, dentry, inode);
 797
 798         if (is_dx(dir)) {
 799 -               retval = ext4_dx_add_entry(handle, &fname, dir, inode);
 800 +               retval = ext4_dx_add_entry(handle, &fname, dir, inode, lck);
 801                 if (!retval || (retval != ERR_BAD_DX_DIR))
 802                         goto out;
 803 +               ext4_htree_safe_relock(lck);
 804                 /* Can we just ignore htree data? */
 805                 if (ext4_has_metadata_csum(sb)) {
 806                         EXT4_ERROR_INODE(dir,
 807 @@ -2440,12 +2780,14 @@ out:
 808                 ext4_set_inode_state(inode, EXT4_STATE_NEWENTRY);
 809         return retval;
 810  }
 811 +EXPORT_SYMBOL(ext4_add_entry_locked);
 812
 813  /*
 814   * Returns 0 for success, or a negative error value
 815   */
 816  static int ext4_dx_add_entry(handle_t *handle, struct ext4_filename *fname,
 817 -                            struct inode *dir, struct inode *inode)
 818 +                            struct inode *dir, struct inode *inode,
 819 +                            struct htree_lock *lck)
 820  {
 821         struct dx_frame frames[EXT4_HTREE_LEVEL], *frame;
 822         struct dx_entry *entries, *at;
 823 @@ -2457,7 +2799,7 @@ static int ext4_dx_add_entry(handle_t *h
 824
 825  again:
 826         restart = 0;
 827 -       frame = dx_probe(fname, dir, NULL, frames);
 828 +       frame = dx_probe(fname, dir, NULL, frames, lck);
 829         if (IS_ERR(frame))
 830                 return PTR_ERR(frame);
 831         entries = frame->entries;
 832 @@ -2492,6 +2834,12 @@ again:
 833                 struct dx_node *node2;
 834                 struct buffer_head *bh2;
 835
 836 +               if (!ext4_htree_safe_locked(lck)) { /* retry with EX lock */
 837 +                       ext4_htree_safe_relock(lck);
 838 +                       restart = 1;
 839 +                       goto cleanup;
 840 +               }
 841 +
 842                 while (frame > frames) {
 843                         if (dx_get_count((frame - 1)->entries) <
 844                             dx_get_limit((frame - 1)->entries)) {
 845 @@ -2594,8 +2942,32 @@ again:
 846                         restart = 1;
 847                         goto journal_error;
 848                 }
 849 +       } else if (!ext4_htree_dx_locked(lck)) {
 850 +               struct ext4_dir_lock_data *ld = ext4_htree_lock_data(lck);
 851 +
 852 +               /* not well protected, require DX lock */
 853 +               ext4_htree_dx_need_lock(lck);
 854 +               at = frame > frames ? (frame - 1)->at : NULL;
 855 +
 856 +               /* NB: no risk of deadlock because it's just a try.
 857 +                *
 858 +                * NB: we check ld_count for twice, the first time before
 859 +                * having DX lock, the second time after holding DX lock.
 860 +                *
 861 +                * NB: We never free blocks for directory so far, which
 862 +                * means value returned by dx_get_count() should equal to
 863 +                * ld->ld_count if nobody split any DE-block under @at,
 864 +                * and ld->ld_at still points to valid dx_entry. */
 865 +               if ((ld->ld_count != dx_get_count(entries)) ||
 866 +                   !ext4_htree_dx_lock_try(lck, at) ||
 867 +                   (ld->ld_count != dx_get_count(entries))) {
 868 +                       restart = 1;
 869 +                       goto cleanup;
 870 +               }
 871 +               /* OK, I've got DX lock and nothing changed */
 872 +               frame->at = ld->ld_at;
 873         }
 874 -       de = do_split(handle, dir, &bh, frame, &fname->hinfo);
 875 +       de = do_split(handle, dir, &bh, frames, frame, &fname->hinfo, lck);
 876         if (IS_ERR(de)) {
 877                 err = PTR_ERR(de);
 878                 goto cleanup;
 879 @@ -2606,6 +2978,8 @@ again:
 880  journal_error:
 881         ext4_std_error(dir->i_sb, err); /* this is a no-op if err == 0 */
 882  cleanup:
 883 +       ext4_htree_dx_unlock(lck);
 884 +       ext4_htree_de_unlock(lck);
 885         brelse(bh);
 886         dx_release(frames);
 887         /* @restart is true means htree-path has been changed, we need to
 888 --- a/fs/ext4/super.c
 889 +++ b/fs/ext4/super.c
 890 @@ -1297,6 +1297,7 @@ static struct inode *ext4_alloc_inode(st
 891
 892         inode_set_iversion(&ei->vfs_inode, 1);
 893         spin_lock_init(&ei->i_raw_lock);
 894 +       sema_init(&ei->i_append_sem, 1);
 895         INIT_LIST_HEAD(&ei->i_prealloc_list);
 896         atomic_set(&ei->i_prealloc_active, 0);
 897         spin_lock_init(&ei->i_prealloc_lock);