ldiskfs/kernel_patches/patches/linux-5.4/ext4-pdirop.patch

   1 From 1a0f7f0b9c13ef0aa86e125f350b6733bff8db3c Mon Sep 17 00:00:00 2001
   2 From: Liang Zhen <liang.zhen@intel.com>
   3 Date: Wed, 15 Jan 2020 07:35:13 -0600
   4 Subject: [PATCH] LU-50 ldiskfs: parallel directory operations for ext4
   5
   6  In a typical use case an application creates a separate output file for each
   7  node and task in a job. As nodes and tasks increase, hundreds of thousands of
   8  files may be created in a single directory within a short window of time.
   9  Today, both filename lookup and file system modifying operations (such as
  10  create and unlink) are protected with a single lock for an entire ldiskfs
  11  directory. PDO project will remove this bottleneck by introducing a parallel
  12  locking mechanism for entire ldiskfs directories. This work will enable
  13  multiple application threads to simultaneously lookup, create and unlink in
  14  parallel.
  15
  16 This patch contains:
  17  - pdirops support for ldiskfs
  18  - integrate with osd-ldiskfs
  19 Signed-off-by: Liang Zhen <liang.zhen@intel.com>
  20 Change-Id: I269c0e3112e68f3acd79e860dab052a68c7d7aaa
  21 ---
  22  fs/ext4/Makefile           |   1 +
  23  fs/ext4/ext4.h             |  78 ++++
  24  fs/ext4/htree_lock.c       | 891 +++++++++++++++++++++++++++++++++++++
  25  fs/ext4/namei.c            | 454 +++++++++++++++++--
  26  fs/ext4/super.c            |   1 +
  27  include/linux/htree_lock.h | 187 ++++++++
  28  6 files changed, 1572 insertions(+), 40 deletions(-)
  29  create mode 100644 fs/ext4/htree_lock.c
  30  create mode 100644 include/linux/htree_lock.h
  31
  32 Index: linux-stage/fs/ext4/ext4.h
  33 ===================================================================
  34 --- linux-stage.orig/fs/ext4/ext4.h
  35 +++ linux-stage/fs/ext4/ext4.h
  36 @@ -29,6 +29,7 @@
  37  #include <linux/timer.h>
  38  #include <linux/version.h>
  39  #include <linux/wait.h>
  40 +#include <linux/htree_lock.h>
  41  #include <linux/sched/signal.h>
  42  #include <linux/blockgroup_lock.h>
  43  #include <linux/percpu_counter.h>
  44 @@ -961,6 +962,9 @@ struct ext4_inode_info {
  45         __u32   i_dtime;
  46         ext4_fsblk_t    i_file_acl;
  47
  48 +       /* following fields for parallel directory operations -bzzz */
  49 +       struct semaphore i_append_sem;
  50 +
  51         /*
  52          * i_block_group is the number of the block group which contains
  53          * this file's inode.  Constant across the lifetime of the inode,
  54 @@ -2207,6 +2211,72 @@ struct dx_hash_info
  55   */
  56  #define HASH_NB_ALWAYS         1
  57
  58 +/* assume name-hash is protected by upper layer */
  59 +#define EXT4_HTREE_LOCK_HASH   0
  60 +
  61 +enum ext4_pdo_lk_types {
  62 +#if EXT4_HTREE_LOCK_HASH
  63 +       EXT4_LK_HASH,
  64 +#endif
  65 +       EXT4_LK_DX,             /* index block */
  66 +       EXT4_LK_DE,             /* directory entry block */
  67 +       EXT4_LK_SPIN,           /* spinlock */
  68 +       EXT4_LK_MAX,
  69 +};
  70 +
  71 +/* read-only bit */
  72 +#define EXT4_LB_RO(b)          (1 << (b))
  73 +/* read + write, high bits for writer */
  74 +#define EXT4_LB_RW(b)          ((1 << (b)) | (1 << (EXT4_LK_MAX + (b))))
  75 +
  76 +enum ext4_pdo_lock_bits {
  77 +       /* DX lock bits */
  78 +       EXT4_LB_DX_RO           = EXT4_LB_RO(EXT4_LK_DX),
  79 +       EXT4_LB_DX              = EXT4_LB_RW(EXT4_LK_DX),
  80 +       /* DE lock bits */
  81 +       EXT4_LB_DE_RO           = EXT4_LB_RO(EXT4_LK_DE),
  82 +       EXT4_LB_DE              = EXT4_LB_RW(EXT4_LK_DE),
  83 +       /* DX spinlock bits */
  84 +       EXT4_LB_SPIN_RO         = EXT4_LB_RO(EXT4_LK_SPIN),
  85 +       EXT4_LB_SPIN            = EXT4_LB_RW(EXT4_LK_SPIN),
  86 +       /* accurate searching */
  87 +       EXT4_LB_EXACT           = EXT4_LB_RO(EXT4_LK_MAX << 1),
  88 +};
  89 +
  90 +enum ext4_pdo_lock_opc {
  91 +       /* external */
  92 +       EXT4_HLOCK_READDIR      = (EXT4_LB_DE_RO | EXT4_LB_DX_RO),
  93 +       EXT4_HLOCK_LOOKUP       = (EXT4_LB_DE_RO | EXT4_LB_SPIN_RO |
  94 +                                  EXT4_LB_EXACT),
  95 +       EXT4_HLOCK_DEL          = (EXT4_LB_DE | EXT4_LB_SPIN_RO |
  96 +                                  EXT4_LB_EXACT),
  97 +       EXT4_HLOCK_ADD          = (EXT4_LB_DE | EXT4_LB_SPIN_RO),
  98 +
  99 +       /* internal */
 100 +       EXT4_HLOCK_LOOKUP_SAFE  = (EXT4_LB_DE_RO | EXT4_LB_DX_RO |
 101 +                                  EXT4_LB_EXACT),
 102 +       EXT4_HLOCK_DEL_SAFE     = (EXT4_LB_DE | EXT4_LB_DX_RO | EXT4_LB_EXACT),
 103 +       EXT4_HLOCK_SPLIT        = (EXT4_LB_DE | EXT4_LB_DX | EXT4_LB_SPIN),
 104 +};
 105 +
 106 +extern struct htree_lock_head *ext4_htree_lock_head_alloc(unsigned hbits);
 107 +#define ext4_htree_lock_head_free(lhead)       htree_lock_head_free(lhead)
 108 +
 109 +extern struct htree_lock *ext4_htree_lock_alloc(void);
 110 +#define ext4_htree_lock_free(lck)              htree_lock_free(lck)
 111 +
 112 +extern void ext4_htree_lock(struct htree_lock *lck,
 113 +                           struct htree_lock_head *lhead,
 114 +                           struct inode *dir, unsigned flags);
 115 +#define ext4_htree_unlock(lck)                  htree_unlock(lck)
 116 +
 117 +extern struct buffer_head *ext4_find_entry_locked(struct inode *dir,
 118 +                                       const struct qstr *d_name,
 119 +                                       struct ext4_dir_entry_2 **res_dir,
 120 +                                       int *inlined, struct htree_lock *lck);
 121 +extern int ext4_add_entry_locked(handle_t *handle, struct dentry *dentry,
 122 +                     struct inode *inode, struct htree_lock *lck);
 123 +
 124  struct ext4_filename {
 125         const struct qstr *usr_fname;
 126         struct fscrypt_str disk_name;
 127 @@ -2574,12 +2644,21 @@ void ext4_insert_dentry(struct inode *in
 128                         struct ext4_filename *fname, void *data);
 129  static inline void ext4_update_dx_flag(struct inode *inode)
 130  {
 131 +       /* Disable it for ldiskfs, because going from a DX directory to
 132 +        * a non-DX directory while it is in use will completely break
 133 +        * the htree-locking.
 134 +        * If we really want to support this operation in the future,
 135 +        * we need to exclusively lock the directory at here which will
 136 +        * increase complexity of code
 137 +        */
 138 +#if 0
 139         if (!ext4_has_feature_dir_index(inode->i_sb) &&
 140             ext4_test_inode_flag(inode, EXT4_INODE_INDEX)) {
 141                 /* ext4_iget() should have caught this... */
 142                 WARN_ON_ONCE(ext4_has_feature_metadata_csum(inode->i_sb));
 143                 ext4_clear_inode_flag(inode, EXT4_INODE_INDEX);
 144         }
 145 +#endif
 146  }
 147  static const unsigned char ext4_filetype_table[] = {
 148         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
 149 Index: linux-stage/fs/ext4/Makefile
 150 ===================================================================
 151 --- linux-stage.orig/fs/ext4/Makefile
 152 +++ linux-stage/fs/ext4/Makefile
 153 @@ -7,6 +7,7 @@ obj-$(CONFIG_EXT4_FS) += ext4.o
 154
 155  ext4-y := balloc.o bitmap.o block_validity.o dir.o ext4_jbd2.o extents.o \
 156                 extents_status.o file.o fsmap.o fsync.o hash.o ialloc.o \
 157 +               htree_lock.o \
 158                 indirect.o inline.o inode.o ioctl.o mballoc.o migrate.o \
 159                 mmp.o move_extent.o namei.o page-io.o readpage.o resize.o \
 160                 super.o symlink.o sysfs.o xattr.o xattr_trusted.o xattr_user.o
 161 Index: linux-stage/fs/ext4/namei.c
 162 ===================================================================
 163 --- linux-stage.orig/fs/ext4/namei.c
 164 +++ linux-stage/fs/ext4/namei.c
 165 @@ -55,6 +55,7 @@ struct buffer_head *ext4_append(handle_t
 166                                         ext4_lblk_t *block)
 167  {
 168         struct buffer_head *bh;
 169 +       struct ext4_inode_info *ei = EXT4_I(inode);
 170         int err;
 171
 172         if (unlikely(EXT4_SB(inode->i_sb)->s_max_dir_size_kb &&
 173 @@ -62,15 +63,22 @@ struct buffer_head *ext4_append(handle_t
 174                       EXT4_SB(inode->i_sb)->s_max_dir_size_kb)))
 175                 return ERR_PTR(-ENOSPC);
 176
 177 +       /* with parallel dir operations all appends
 178 +       * have to be serialized -bzzz */
 179 +       down(&ei->i_append_sem);
 180 +
 181         *block = inode->i_size >> inode->i_sb->s_blocksize_bits;
 182
 183         bh = ext4_bread(handle, inode, *block, EXT4_GET_BLOCKS_CREATE);
 184 -       if (IS_ERR(bh))
 185 +       if (IS_ERR(bh)) {
 186 +               up(&ei->i_append_sem);
 187                 return bh;
 188 +       }
 189         inode->i_size += inode->i_sb->s_blocksize;
 190         EXT4_I(inode)->i_disksize = inode->i_size;
 191         BUFFER_TRACE(bh, "get_write_access");
 192         err = ext4_journal_get_write_access(handle, bh);
 193 +       up(&ei->i_append_sem);
 194         if (err) {
 195                 brelse(bh);
 196                 ext4_std_error(inode->i_sb, err);
 197 @@ -264,7 +272,8 @@ static unsigned dx_node_limit(struct ino
 198  static struct dx_frame *dx_probe(struct ext4_filename *fname,
 199                                  struct inode *dir,
 200                                  struct dx_hash_info *hinfo,
 201 -                                struct dx_frame *frame);
 202 +                                struct dx_frame *frame,
 203 +                                struct htree_lock *lck);
 204  static void dx_release(struct dx_frame *frames);
 205  static int dx_make_map(struct inode *dir, struct ext4_dir_entry_2 *de,
 206                        unsigned blocksize, struct dx_hash_info *hinfo,
 207 @@ -278,12 +287,13 @@ static void dx_insert_block(struct dx_fr
 208  static int ext4_htree_next_block(struct inode *dir, __u32 hash,
 209                                  struct dx_frame *frame,
 210                                  struct dx_frame *frames,
 211 -                                __u32 *start_hash);
 212 +                                __u32 *start_hash, struct htree_lock *lck);
 213  static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
 214                 struct ext4_filename *fname,
 215 -               struct ext4_dir_entry_2 **res_dir);
 216 +               struct ext4_dir_entry_2 **res_dir, struct htree_lock *lck);
 217  static int ext4_dx_add_entry(handle_t *handle, struct ext4_filename *fname,
 218 -                            struct inode *dir, struct inode *inode);
 219 +                            struct inode *dir, struct inode *inode,
 220 +                            struct htree_lock *lck);
 221
 222  /* checksumming functions */
 223  void ext4_initialize_dirent_tail(struct buffer_head *bh,
 224 @@ -748,6 +758,227 @@ struct stats dx_show_entries(struct dx_h
 225  }
 226  #endif /* DX_DEBUG */
 227
 228 +/* private data for htree_lock */
 229 +struct ext4_dir_lock_data {
 230 +       unsigned                ld_flags;  /* bits-map for lock types */
 231 +       unsigned                ld_count;  /* # entries of the last DX block */
 232 +       struct dx_entry         ld_at_entry; /* copy of leaf dx_entry */
 233 +       struct dx_entry         *ld_at;    /* position of leaf dx_entry */
 234 +};
 235 +
 236 +#define ext4_htree_lock_data(l)        ((struct ext4_dir_lock_data *)(l)->lk_private)
 237 +#define ext4_find_entry(dir, name, dirent, inline) \
 238 +                       ext4_find_entry_locked(dir, name, dirent, inline, NULL)
 239 +#define ext4_add_entry(handle, dentry, inode) \
 240 +                       ext4_add_entry_locked(handle, dentry, inode, NULL)
 241 +
 242 +/* NB: ext4_lblk_t is 32 bits so we use high bits to identify invalid blk */
 243 +#define EXT4_HTREE_NODE_CHANGED        (0xcafeULL << 32)
 244 +
 245 +static void ext4_htree_event_cb(void *target, void *event)
 246 +{
 247 +       u64 *block = (u64 *)target;
 248 +
 249 +       if (*block == dx_get_block((struct dx_entry *)event))
 250 +               *block = EXT4_HTREE_NODE_CHANGED;
 251 +}
 252 +
 253 +struct htree_lock_head *ext4_htree_lock_head_alloc(unsigned hbits)
 254 +{
 255 +       struct htree_lock_head *lhead;
 256 +
 257 +       lhead = htree_lock_head_alloc(EXT4_LK_MAX, hbits, 0);
 258 +       if (lhead != NULL) {
 259 +               htree_lock_event_attach(lhead, EXT4_LK_SPIN, HTREE_EVENT_WR,
 260 +                                       ext4_htree_event_cb);
 261 +       }
 262 +       return lhead;
 263 +}
 264 +EXPORT_SYMBOL(ext4_htree_lock_head_alloc);
 265 +
 266 +struct htree_lock *ext4_htree_lock_alloc(void)
 267 +{
 268 +       return htree_lock_alloc(EXT4_LK_MAX,
 269 +                               sizeof(struct ext4_dir_lock_data));
 270 +}
 271 +EXPORT_SYMBOL(ext4_htree_lock_alloc);
 272 +
 273 +static htree_lock_mode_t ext4_htree_mode(unsigned flags)
 274 +{
 275 +       switch (flags) {
 276 +       default: /* 0 or unknown flags require EX lock */
 277 +               return HTREE_LOCK_EX;
 278 +       case EXT4_HLOCK_READDIR:
 279 +               return HTREE_LOCK_PR;
 280 +       case EXT4_HLOCK_LOOKUP:
 281 +               return HTREE_LOCK_CR;
 282 +       case EXT4_HLOCK_DEL:
 283 +       case EXT4_HLOCK_ADD:
 284 +               return HTREE_LOCK_CW;
 285 +       }
 286 +}
 287 +
 288 +/* return PR for read-only operations, otherwise return EX */
 289 +static inline htree_lock_mode_t ext4_htree_safe_mode(unsigned flags)
 290 +{
 291 +       int writer = (flags & EXT4_LB_DE) == EXT4_LB_DE;
 292 +
 293 +       /* 0 requires EX lock */
 294 +       return (flags == 0 || writer) ? HTREE_LOCK_EX : HTREE_LOCK_PR;
 295 +}
 296 +
 297 +static int ext4_htree_safe_locked(struct htree_lock *lck)
 298 +{
 299 +       int writer;
 300 +
 301 +       if (lck == NULL || lck->lk_mode == HTREE_LOCK_EX)
 302 +               return 1;
 303 +
 304 +       writer = (ext4_htree_lock_data(lck)->ld_flags & EXT4_LB_DE) ==
 305 +                EXT4_LB_DE;
 306 +       if (writer) /* all readers & writers are excluded? */
 307 +               return lck->lk_mode == HTREE_LOCK_EX;
 308 +
 309 +       /* all writers are excluded? */
 310 +       return lck->lk_mode == HTREE_LOCK_PR ||
 311 +              lck->lk_mode == HTREE_LOCK_PW ||
 312 +              lck->lk_mode == HTREE_LOCK_EX;
 313 +}
 314 +
 315 +/* relock htree_lock with EX mode if it's change operation, otherwise
 316 + * relock it with PR mode. It's noop if PDO is disabled. */
 317 +static void ext4_htree_safe_relock(struct htree_lock *lck)
 318 +{
 319 +       if (!ext4_htree_safe_locked(lck)) {
 320 +               unsigned flags = ext4_htree_lock_data(lck)->ld_flags;
 321 +
 322 +               htree_change_lock(lck, ext4_htree_safe_mode(flags));
 323 +       }
 324 +}
 325 +
 326 +void ext4_htree_lock(struct htree_lock *lck, struct htree_lock_head *lhead,
 327 +                    struct inode *dir, unsigned flags)
 328 +{
 329 +       htree_lock_mode_t mode = is_dx(dir) ? ext4_htree_mode(flags) :
 330 +                                             ext4_htree_safe_mode(flags);
 331 +
 332 +       ext4_htree_lock_data(lck)->ld_flags = flags;
 333 +       htree_lock(lck, lhead, mode);
 334 +       if (!is_dx(dir))
 335 +               ext4_htree_safe_relock(lck); /* make sure it's safe locked */
 336 +}
 337 +EXPORT_SYMBOL(ext4_htree_lock);
 338 +
 339 +static int ext4_htree_node_lock(struct htree_lock *lck, struct dx_entry *at,
 340 +                               unsigned lmask, int wait, void *ev)
 341 +{
 342 +       u32     key = (at == NULL) ? 0 : dx_get_block(at);
 343 +       u32     mode;
 344 +
 345 +       /* NOOP if htree is well protected or caller doesn't require the lock */
 346 +       if (ext4_htree_safe_locked(lck) ||
 347 +          !(ext4_htree_lock_data(lck)->ld_flags & lmask))
 348 +               return 1;
 349 +
 350 +       mode = (ext4_htree_lock_data(lck)->ld_flags & lmask) == lmask ?
 351 +               HTREE_LOCK_PW : HTREE_LOCK_PR;
 352 +       while (1) {
 353 +               if (htree_node_lock_try(lck, mode, key, ffz(~lmask), wait, ev))
 354 +                       return 1;
 355 +               if (!(lmask & EXT4_LB_SPIN)) /* not a spinlock */
 356 +                       return 0;
 357 +               cpu_relax(); /* spin until granted */
 358 +       }
 359 +}
 360 +
 361 +static int ext4_htree_node_locked(struct htree_lock *lck, unsigned lmask)
 362 +{
 363 +       return ext4_htree_safe_locked(lck) ||
 364 +              htree_node_is_granted(lck, ffz(~lmask));
 365 +}
 366 +
 367 +static void ext4_htree_node_unlock(struct htree_lock *lck,
 368 +                                  unsigned lmask, void *buf)
 369 +{
 370 +       /* NB: it's safe to call mutiple times or even it's not locked */
 371 +       if (!ext4_htree_safe_locked(lck) &&
 372 +            htree_node_is_granted(lck, ffz(~lmask)))
 373 +               htree_node_unlock(lck, ffz(~lmask), buf);
 374 +}
 375 +
 376 +#define ext4_htree_dx_lock(lck, key)           \
 377 +       ext4_htree_node_lock(lck, key, EXT4_LB_DX, 1, NULL)
 378 +#define ext4_htree_dx_lock_try(lck, key)       \
 379 +       ext4_htree_node_lock(lck, key, EXT4_LB_DX, 0, NULL)
 380 +#define ext4_htree_dx_unlock(lck)              \
 381 +       ext4_htree_node_unlock(lck, EXT4_LB_DX, NULL)
 382 +#define ext4_htree_dx_locked(lck)              \
 383 +       ext4_htree_node_locked(lck, EXT4_LB_DX)
 384 +
 385 +static void ext4_htree_dx_need_lock(struct htree_lock *lck)
 386 +{
 387 +       struct ext4_dir_lock_data *ld;
 388 +
 389 +       if (ext4_htree_safe_locked(lck))
 390 +               return;
 391 +
 392 +       ld = ext4_htree_lock_data(lck);
 393 +       switch (ld->ld_flags) {
 394 +       default:
 395 +               return;
 396 +       case EXT4_HLOCK_LOOKUP:
 397 +               ld->ld_flags = EXT4_HLOCK_LOOKUP_SAFE;
 398 +               return;
 399 +       case EXT4_HLOCK_DEL:
 400 +               ld->ld_flags = EXT4_HLOCK_DEL_SAFE;
 401 +               return;
 402 +       case EXT4_HLOCK_ADD:
 403 +               ld->ld_flags = EXT4_HLOCK_SPLIT;
 404 +               return;
 405 +       }
 406 +}
 407 +
 408 +#define ext4_htree_de_lock(lck, key)           \
 409 +       ext4_htree_node_lock(lck, key, EXT4_LB_DE, 1, NULL)
 410 +#define ext4_htree_de_unlock(lck)              \
 411 +       ext4_htree_node_unlock(lck, EXT4_LB_DE, NULL)
 412 +
 413 +#define ext4_htree_spin_lock(lck, key, event)  \
 414 +       ext4_htree_node_lock(lck, key, EXT4_LB_SPIN, 0, event)
 415 +#define ext4_htree_spin_unlock(lck)            \
 416 +       ext4_htree_node_unlock(lck, EXT4_LB_SPIN, NULL)
 417 +#define ext4_htree_spin_unlock_listen(lck, p)  \
 418 +       ext4_htree_node_unlock(lck, EXT4_LB_SPIN, p)
 419 +
 420 +static void ext4_htree_spin_stop_listen(struct htree_lock *lck)
 421 +{
 422 +       if (!ext4_htree_safe_locked(lck) &&
 423 +           htree_node_is_listening(lck, ffz(~EXT4_LB_SPIN)))
 424 +               htree_node_stop_listen(lck, ffz(~EXT4_LB_SPIN));
 425 +}
 426 +
 427 +enum {
 428 +       DX_HASH_COL_IGNORE,     /* ignore collision while probing frames */
 429 +       DX_HASH_COL_YES,        /* there is collision and it does matter */
 430 +       DX_HASH_COL_NO,         /* there is no collision */
 431 +};
 432 +
 433 +static int dx_probe_hash_collision(struct htree_lock *lck,
 434 +                                  struct dx_entry *entries,
 435 +                                  struct dx_entry *at, u32 hash)
 436 +{
 437 +       if (!(lck && ext4_htree_lock_data(lck)->ld_flags & EXT4_LB_EXACT)) {
 438 +               return DX_HASH_COL_IGNORE; /* don't care about collision */
 439 +
 440 +       } else if (at == entries + dx_get_count(entries) - 1) {
 441 +               return DX_HASH_COL_IGNORE; /* not in any leaf of this DX */
 442 +
 443 +       } else { /* hash collision? */
 444 +               return ((dx_get_hash(at + 1) & ~1) == hash) ?
 445 +                       DX_HASH_COL_YES : DX_HASH_COL_NO;
 446 +       }
 447 +}
 448 +
 449  /*
 450   * Probe for a directory leaf block to search.
 451   *
 452 @@ -759,10 +990,11 @@ struct stats dx_show_entries(struct dx_h
 453   */
 454  static struct dx_frame *
 455  dx_probe(struct ext4_filename *fname, struct inode *dir,
 456 -        struct dx_hash_info *hinfo, struct dx_frame *frame_in)
 457 +        struct dx_hash_info *hinfo, struct dx_frame *frame_in,
 458 +        struct htree_lock *lck)
 459  {
 460         unsigned count, indirect;
 461 -       struct dx_entry *at, *entries, *p, *q, *m;
 462 +       struct dx_entry *at, *entries, *p, *q, *m, *dx = NULL;
 463         struct dx_root_info *info;
 464         struct dx_frame *frame = frame_in;
 465         struct dx_frame *ret_err = ERR_PTR(ERR_BAD_DX_DIR);
 466 @@ -824,8 +1056,15 @@ dx_probe(struct ext4_filename *fname, st
 467
 468         dxtrace(printk("Look up %x", hash));
 469         while (1) {
 470 +               if (indirect == 0) { /* the last index level */
 471 +                       /* NB: ext4_htree_dx_lock() could be noop if
 472 +                        * DX-lock flag is not set for current operation */
 473 +                       ext4_htree_dx_lock(lck, dx);
 474 +                       ext4_htree_spin_lock(lck, dx, NULL);
 475 +               }
 476                 count = dx_get_count(entries);
 477 -               if (!count || count > dx_get_limit(entries)) {
 478 +               if (count == 0 || count > dx_get_limit(entries)) {
 479 +                       ext4_htree_spin_unlock(lck); /* release spin */
 480                         ext4_warning_inode(dir,
 481                                            "dx entry: count %u beyond limit %u",
 482                                            count, dx_get_limit(entries));
 483 @@ -864,8 +1103,70 @@ dx_probe(struct ext4_filename *fname, st
 484                                dx_get_block(at)));
 485                 frame->entries = entries;
 486                 frame->at = at;
 487 -               if (!indirect--)
 488 +
 489 +               if (indirect == 0) { /* the last index level */
 490 +                       struct ext4_dir_lock_data *ld;
 491 +                       u64 myblock;
 492 +
 493 +                       /* By default we only lock DE-block, however, we will
 494 +                        * also lock the last level DX-block if:
 495 +                        * a) there is hash collision
 496 +                        *    we will set DX-lock flag (a few lines below)
 497 +                        *    and redo to lock DX-block
 498 +                        *    see detail in dx_probe_hash_collision()
 499 +                        * b) it's a retry from splitting
 500 +                        *    we need to lock the last level DX-block so nobody
 501 +                        *    else can split any leaf blocks under the same
 502 +                        *    DX-block, see detail in ext4_dx_add_entry()
 503 +                        */
 504 +                       if (ext4_htree_dx_locked(lck)) {
 505 +                               /* DX-block is locked, just lock DE-block
 506 +                                * and return */
 507 +                               ext4_htree_spin_unlock(lck);
 508 +                               if (!ext4_htree_safe_locked(lck))
 509 +                                       ext4_htree_de_lock(lck, frame->at);
 510 +                               return frame;
 511 +                       }
 512 +                       /* it's pdirop and no DX lock */
 513 +                       if (dx_probe_hash_collision(lck, entries, at, hash) ==
 514 +                           DX_HASH_COL_YES) {
 515 +                               /* found hash collision, set DX-lock flag
 516 +                                * and retry to abtain DX-lock */
 517 +                               ext4_htree_spin_unlock(lck);
 518 +                               ext4_htree_dx_need_lock(lck);
 519 +                               continue;
 520 +                       }
 521 +                       ld = ext4_htree_lock_data(lck);
 522 +                       /* because I don't lock DX, so @at can't be trusted
 523 +                        * after I release spinlock so I have to save it */
 524 +                       ld->ld_at = at;
 525 +                       ld->ld_at_entry = *at;
 526 +                       ld->ld_count = dx_get_count(entries);
 527 +
 528 +                       frame->at = &ld->ld_at_entry;
 529 +                       myblock = dx_get_block(at);
 530 +
 531 +                       /* NB: ordering locking */
 532 +                       ext4_htree_spin_unlock_listen(lck, &myblock);
 533 +                       /* other thread can split this DE-block because:
 534 +                        * a) I don't have lock for the DE-block yet
 535 +                        * b) I released spinlock on DX-block
 536 +                        * if it happened I can detect it by listening
 537 +                        * splitting event on this DE-block */
 538 +                       ext4_htree_de_lock(lck, frame->at);
 539 +                       ext4_htree_spin_stop_listen(lck);
 540 +
 541 +                       if (myblock == EXT4_HTREE_NODE_CHANGED) {
 542 +                               /* someone split this DE-block before
 543 +                                * I locked it, I need to retry and lock
 544 +                                * valid DE-block */
 545 +                               ext4_htree_de_unlock(lck);
 546 +                               continue;
 547 +                       }
 548                         return frame;
 549 +               }
 550 +               dx = at;
 551 +               indirect--;
 552                 frame++;
 553                 frame->bh = ext4_read_dirblock(dir, dx_get_block(at), INDEX);
 554                 if (IS_ERR(frame->bh)) {
 555 @@ -934,7 +1235,7 @@ static void dx_release(struct dx_frame *
 556  static int ext4_htree_next_block(struct inode *dir, __u32 hash,
 557                                  struct dx_frame *frame,
 558                                  struct dx_frame *frames,
 559 -                                __u32 *start_hash)
 560 +                                __u32 *start_hash, struct htree_lock *lck)
 561  {
 562         struct dx_frame *p;
 563         struct buffer_head *bh;
 564 @@ -949,12 +1250,22 @@ static int ext4_htree_next_block(struct
 565          * this loop, num_frames indicates the number of interior
 566          * nodes need to be read.
 567          */
 568 +       ext4_htree_de_unlock(lck);
 569         while (1) {
 570 +               if (num_frames > 0 || ext4_htree_dx_locked(lck)) {
 571 +                       /* num_frames > 0 :
 572 +                        *   DX block
 573 +                        * ext4_htree_dx_locked:
 574 +                        *   frame->at is reliable pointer returned by dx_probe,
 575 +                        *   otherwise dx_probe already knew no collision */
 576                 if (++(p->at) < p->entries + dx_get_count(p->entries))
 577                         break;
 578 +               }
 579                 if (p == frames)
 580                         return 0;
 581                 num_frames++;
 582 +               if (num_frames == 1)
 583 +                       ext4_htree_dx_unlock(lck);
 584                 p--;
 585         }
 586
 587 @@ -977,6 +1288,13 @@ static int ext4_htree_next_block(struct
 588          * block so no check is necessary
 589          */
 590         while (num_frames--) {
 591 +               if (num_frames == 0) {
 592 +                       /* it's not always necessary, we just don't want to
 593 +                        * detect hash collision again */
 594 +                       ext4_htree_dx_need_lock(lck);
 595 +                       ext4_htree_dx_lock(lck, p->at);
 596 +               }
 597 +
 598                 bh = ext4_read_dirblock(dir, dx_get_block(p->at), INDEX);
 599                 if (IS_ERR(bh))
 600                         return PTR_ERR(bh);
 601 @@ -985,6 +1303,7 @@ static int ext4_htree_next_block(struct
 602                 p->bh = bh;
 603                 p->at = p->entries = ((struct dx_node *) bh->b_data)->entries;
 604         }
 605 +       ext4_htree_de_lock(lck, p->at);
 606         return 1;
 607  }
 608
 609 @@ -1132,10 +1451,10 @@ int ext4_htree_fill_tree(struct file *di
 610         }
 611         hinfo.hash = start_hash;
 612         hinfo.minor_hash = 0;
 613 -       frame = dx_probe(NULL, dir, &hinfo, frames);
 614 +       /* assume it's PR locked */
 615 +       frame = dx_probe(NULL, dir, &hinfo, frames, NULL);
 616         if (IS_ERR(frame))
 617                 return PTR_ERR(frame);
 618 -
 619         /* Add '.' and '..' from the htree header */
 620         if (!start_hash && !start_minor_hash) {
 621                 de = (struct ext4_dir_entry_2 *) frames[0].bh->b_data;
 622 @@ -1175,7 +1494,7 @@ int ext4_htree_fill_tree(struct file *di
 623                 count += ret;
 624                 hashval = ~0;
 625                 ret = ext4_htree_next_block(dir, HASH_NB_ALWAYS,
 626 -                                           frame, frames, &hashval);
 627 +                                           frame, frames, &hashval, NULL);
 628                 *next_hash = hashval;
 629                 if (ret < 0) {
 630                         err = ret;
 631 @@ -1451,7 +1770,7 @@ static int is_dx_internal_node(struct in
 632  static struct buffer_head *__ext4_find_entry(struct inode *dir,
 633                                              struct ext4_filename *fname,
 634                                              struct ext4_dir_entry_2 **res_dir,
 635 -                                            int *inlined)
 636 +                                            int *inlined, struct htree_lock *lck)
 637  {
 638         struct super_block *sb;
 639         struct buffer_head *bh_use[NAMEI_RA_SIZE];
 640 @@ -1493,7 +1812,7 @@ static struct buffer_head *__ext4_find_e
 641                 goto restart;
 642         }
 643         if (is_dx(dir)) {
 644 -               ret = ext4_dx_find_entry(dir, fname, res_dir);
 645 +               ret = ext4_dx_find_entry(dir, fname, res_dir, lck);
 646                 /*
 647                  * On success, or if the error was file not found,
 648                  * return.  Otherwise, fall back to doing a search the
 649 @@ -1503,6 +1822,7 @@ static struct buffer_head *__ext4_find_e
 650                         goto cleanup_and_exit;
 651                 dxtrace(printk(KERN_DEBUG "ext4_find_entry: dx failed, "
 652                                "falling back\n"));
 653 +               ext4_htree_safe_relock(lck);
 654                 ret = NULL;
 655         }
 656         nblocks = dir->i_size >> EXT4_BLOCK_SIZE_BITS(sb);
 657 @@ -1591,10 +1911,10 @@ cleanup_and_exit:
 658         return ret;
 659  }
 660
 661 -static struct buffer_head *ext4_find_entry(struct inode *dir,
 662 +struct buffer_head *ext4_find_entry_locked(struct inode *dir,
 663                                            const struct qstr *d_name,
 664                                            struct ext4_dir_entry_2 **res_dir,
 665 -                                          int *inlined)
 666 +                                          int *inlined, struct htree_lock *lck)
 667  {
 668         int err;
 669         struct ext4_filename fname;
 670 @@ -1606,12 +1926,14 @@ static struct buffer_head *ext4_find_ent
 671         if (err)
 672                 return ERR_PTR(err);
 673
 674 -       bh = __ext4_find_entry(dir, &fname, res_dir, inlined);
 675 +       bh = __ext4_find_entry(dir, &fname, res_dir, inlined, lck);
 676
 677         ext4_fname_free_filename(&fname);
 678         return bh;
 679  }
 680
 681 +EXPORT_SYMBOL(ext4_find_entry_locked);
 682 +
 683  static struct buffer_head *ext4_lookup_entry(struct inode *dir,
 684                                              struct dentry *dentry,
 685                                              struct ext4_dir_entry_2 **res_dir)
 686 @@ -1626,7 +1948,7 @@ static struct buffer_head *ext4_lookup_e
 687         if (err)
 688                 return ERR_PTR(err);
 689
 690 -       bh = __ext4_find_entry(dir, &fname, res_dir, NULL);
 691 +       bh = __ext4_find_entry(dir, &fname, res_dir, NULL, NULL);
 692
 693         ext4_fname_free_filename(&fname);
 694         return bh;
 695 @@ -1634,7 +1956,8 @@ static struct buffer_head *ext4_lookup_e
 696
 697  static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
 698                         struct ext4_filename *fname,
 699 -                       struct ext4_dir_entry_2 **res_dir)
 700 +                       struct ext4_dir_entry_2 **res_dir,
 701 +                       struct htree_lock *lck)
 702  {
 703         struct super_block * sb = dir->i_sb;
 704         struct dx_frame frames[EXT4_HTREE_LEVEL], *frame;
 705 @@ -1645,7 +1968,7 @@ static struct buffer_head * ext4_dx_find
 706  #ifdef CONFIG_FS_ENCRYPTION
 707         *res_dir = NULL;
 708  #endif
 709 -       frame = dx_probe(fname, dir, NULL, frames);
 710 +       frame = dx_probe(fname, dir, NULL, frames, lck);
 711         if (IS_ERR(frame))
 712                 return (struct buffer_head *) frame;
 713         do {
 714 @@ -1667,7 +1990,7 @@ static struct buffer_head * ext4_dx_find
 715
 716                 /* Check to see if we should continue to search */
 717                 retval = ext4_htree_next_block(dir, fname->hinfo.hash, frame,
 718 -                                              frames, NULL);
 719 +                                              frames, NULL, lck);
 720                 if (retval < 0) {
 721                         ext4_warning_inode(dir,
 722                                 "error %d reading directory index block",
 723 @@ -1847,8 +2170,9 @@ static struct ext4_dir_entry_2* dx_pack_
 724   * Returns pointer to de in block into which the new entry will be inserted.
 725   */
 726  static struct ext4_dir_entry_2 *do_split(handle_t *handle, struct inode *dir,
 727 -                       struct buffer_head **bh,struct dx_frame *frame,
 728 -                       struct dx_hash_info *hinfo)
 729 +                       struct buffer_head **bh, struct dx_frame *frames,
 730 +                       struct dx_frame *frame, struct dx_hash_info *hinfo,
 731 +                       struct htree_lock *lck)
 732  {
 733         unsigned blocksize = dir->i_sb->s_blocksize;
 734         unsigned count, continued;
 735 @@ -1919,8 +2243,14 @@ static struct ext4_dir_entry_2 *do_split
 736                                         hash2, split, count-split));
 737
 738         /* Fancy dance to stay within two buffers */
 739 -       de2 = dx_move_dirents(data1, data2, map + split, count - split,
 740 -                             blocksize);
 741 +       if (hinfo->hash < hash2) {
 742 +               de2 = dx_move_dirents(data1, data2, map + split,
 743 +                                     count - split, blocksize);
 744 +       } else {
 745 +               /* make sure we will add entry to the same block which
 746 +                * we have already locked */
 747 +               de2 = dx_move_dirents(data1, data2, map, split, blocksize);
 748 +       }
 749         de = dx_pack_dirents(data1, blocksize);
 750         de->rec_len = ext4_rec_len_to_disk(data1 + (blocksize - csum_size) -
 751                                            (char *) de,
 752 @@ -1938,12 +2268,21 @@ static struct ext4_dir_entry_2 *do_split
 753         dxtrace(dx_show_leaf(dir, hinfo, (struct ext4_dir_entry_2 *) data2,
 754                         blocksize, 1));
 755
 756 -       /* Which block gets the new entry? */
 757 -       if (hinfo->hash >= hash2) {
 758 -               swap(*bh, bh2);
 759 -               de = de2;
 760 -       }
 761 +       ext4_htree_spin_lock(lck, frame > frames ? (frame - 1)->at : NULL,
 762 +                            frame->at); /* notify block is being split */
 763 +       if (hinfo->hash < hash2) {
 764         dx_insert_block(frame, hash2 + continued, newblock);
 765 +
 766 +       } else {
 767 +               /* switch block number */
 768 +               dx_insert_block(frame, hash2 + continued,
 769 +                               dx_get_block(frame->at));
 770 +               dx_set_block(frame->at, newblock);
 771 +               (frame->at)++;
 772 +       }
 773 +       ext4_htree_spin_unlock(lck);
 774 +       ext4_htree_dx_unlock(lck);
 775 +
 776         err = ext4_handle_dirty_dirblock(handle, dir, bh2);
 777         if (err)
 778                 goto journal_error;
 779 @@ -2213,7 +2552,7 @@ static int make_indexed_dir(handle_t *ha
 780         if (retval)
 781                 goto out_frames;
 782
 783 -       de = do_split(handle,dir, &bh2, frame, &fname->hinfo);
 784 +       de = do_split(handle, dir, &bh2, frames, frame, &fname->hinfo, NULL);
 785         if (IS_ERR(de)) {
 786                 retval = PTR_ERR(de);
 787                 goto out_frames;
 788 @@ -2323,8 +2662,8 @@ out:
 789   * may not sleep between calling this and putting something into
 790   * the entry, as someone else might have used it while you slept.
 791   */
 792 -static int ext4_add_entry(handle_t *handle, struct dentry *dentry,
 793 -                         struct inode *inode)
 794 +int ext4_add_entry_locked(handle_t *handle, struct dentry *dentry,
 795 +                         struct inode *inode, struct htree_lock *lck)
 796  {
 797         struct inode *dir = d_inode(dentry->d_parent);
 798         struct buffer_head *bh = NULL;
 799 @@ -2375,9 +2714,10 @@ static int ext4_add_entry(handle_t *hand
 800                 if (dentry->d_name.len == 2 &&
 801                     memcmp(dentry->d_name.name, "..", 2) == 0)
 802                         return ext4_update_dotdot(handle, dentry, inode);
 803 -               retval = ext4_dx_add_entry(handle, &fname, dir, inode);
 804 +               retval = ext4_dx_add_entry(handle, &fname, dir, inode, lck);
 805                 if (!retval || (retval != ERR_BAD_DX_DIR))
 806                         goto out;
 807 +               ext4_htree_safe_relock(lck);
 808                 /* Can we just ignore htree data? */
 809                 if (ext4_has_metadata_csum(sb)) {
 810                         EXT4_ERROR_INODE(dir,
 811 @@ -2438,12 +2778,14 @@ out:
 812                 ext4_set_inode_state(inode, EXT4_STATE_NEWENTRY);
 813         return retval;
 814  }
 815 +EXPORT_SYMBOL(ext4_add_entry_locked);
 816
 817  /*
 818   * Returns 0 for success, or a negative error value
 819   */
 820  static int ext4_dx_add_entry(handle_t *handle, struct ext4_filename *fname,
 821 -                            struct inode *dir, struct inode *inode)
 822 +                            struct inode *dir, struct inode *inode,
 823 +                            struct htree_lock *lck)
 824  {
 825         struct dx_frame frames[EXT4_HTREE_LEVEL], *frame;
 826         struct dx_entry *entries, *at;
 827 @@ -2455,7 +2797,7 @@ static int ext4_dx_add_entry(handle_t *h
 828
 829  again:
 830         restart = 0;
 831 -       frame = dx_probe(fname, dir, NULL, frames);
 832 +       frame = dx_probe(fname, dir, NULL, frames, lck);
 833         if (IS_ERR(frame))
 834                 return PTR_ERR(frame);
 835         entries = frame->entries;
 836 @@ -2490,6 +2832,12 @@ again:
 837                 struct dx_node *node2;
 838                 struct buffer_head *bh2;
 839
 840 +               if (!ext4_htree_safe_locked(lck)) { /* retry with EX lock */
 841 +                       ext4_htree_safe_relock(lck);
 842 +                       restart = 1;
 843 +                       goto cleanup;
 844 +               }
 845 +
 846                 while (frame > frames) {
 847                         if (dx_get_count((frame - 1)->entries) <
 848                             dx_get_limit((frame - 1)->entries)) {
 849 @@ -2591,8 +2939,32 @@ again:
 850                         restart = 1;
 851                         goto journal_error;
 852                 }
 853 +       } else if (!ext4_htree_dx_locked(lck)) {
 854 +               struct ext4_dir_lock_data *ld = ext4_htree_lock_data(lck);
 855 +
 856 +               /* not well protected, require DX lock */
 857 +               ext4_htree_dx_need_lock(lck);
 858 +               at = frame > frames ? (frame - 1)->at : NULL;
 859 +
 860 +               /* NB: no risk of deadlock because it's just a try.
 861 +                *
 862 +                * NB: we check ld_count for twice, the first time before
 863 +                * having DX lock, the second time after holding DX lock.
 864 +                *
 865 +                * NB: We never free blocks for directory so far, which
 866 +                * means value returned by dx_get_count() should equal to
 867 +                * ld->ld_count if nobody split any DE-block under @at,
 868 +                * and ld->ld_at still points to valid dx_entry. */
 869 +               if ((ld->ld_count != dx_get_count(entries)) ||
 870 +                   !ext4_htree_dx_lock_try(lck, at) ||
 871 +                   (ld->ld_count != dx_get_count(entries))) {
 872 +                       restart = 1;
 873 +                       goto cleanup;
 874 +               }
 875 +               /* OK, I've got DX lock and nothing changed */
 876 +               frame->at = ld->ld_at;
 877         }
 878 -       de = do_split(handle, dir, &bh, frame, &fname->hinfo);
 879 +       de = do_split(handle, dir, &bh, frames, frame, &fname->hinfo, lck);
 880         if (IS_ERR(de)) {
 881                 err = PTR_ERR(de);
 882                 goto cleanup;
 883 @@ -2603,6 +2975,8 @@ again:
 884  journal_error:
 885         ext4_std_error(dir->i_sb, err); /* this is a no-op if err == 0 */
 886  cleanup:
 887 +       ext4_htree_dx_unlock(lck);
 888 +       ext4_htree_de_unlock(lck);
 889         brelse(bh);
 890         dx_release(frames);
 891         /* @restart is true means htree-path has been changed, we need to
 892 Index: linux-stage/fs/ext4/super.c
 893 ===================================================================
 894 --- linux-stage.orig/fs/ext4/super.c
 895 +++ linux-stage/fs/ext4/super.c
 896 @@ -1086,6 +1086,7 @@ static struct inode *ext4_alloc_inode(st
 897
 898         inode_set_iversion(&ei->vfs_inode, 1);
 899         spin_lock_init(&ei->i_raw_lock);
 900 +       sema_init(&ei->i_append_sem, 1);
 901         INIT_LIST_HEAD(&ei->i_prealloc_list);
 902         spin_lock_init(&ei->i_prealloc_lock);
 903         ext4_es_init_tree(&ei->i_es_tree);