ldiskfs/kernel_patches/patches/rhel9.2/ext4-pdirop.patch

   1 Subject: [PATCH] ext4-pdirop
   2
   3  In a typical use case an application creates a separate output file for each
   4  node and task in a job. As nodes and tasks increase, hundreds of thousands of
   5  files may be created in a single directory within a short window of time.
   6  Today, both filename lookup and file system modifying operations (such as
   7  create and unlink) are protected with a single lock for an entire ldiskfs
   8  directory. PDO project will remove this bottleneck by introducing a parallel
   9  locking mechanism for entire ldiskfs directories. This work will enable
  10  multiple application threads to simultaneously lookup, create and unlink in
  11  parallel.
  12
  13 This patch contains:
  14  - pdirops support for ldiskfs
  15  - integrate with osd-ldiskfs
  16 Signed-off-by: Liang Zhen <liang.zhen@intel.com>
  17 Change-Id: I269c0e3112e68f3acd79e860dab052a68c7d7aaa
  18 ---
  19  fs/ext4/Makefile |   1 +
  20  fs/ext4/ext4.h   |  78 ++++++++
  21  fs/ext4/namei.c  | 467 ++++++++++++++++++++++++++++++++++++++++++-----
  22  fs/ext4/super.c  |   1 +
  23  4 files changed, 505 insertions(+), 42 deletions(-)
  24
  25 diff --git a/fs/ext4/Makefile b/fs/ext4/Makefile
  26 index 72206a2..1d15a3a 100644
  27 --- a/fs/ext4/Makefile
  28 +++ b/fs/ext4/Makefile
  29 @@ -7,6 +7,7 @@ obj-$(CONFIG_EXT4_FS) += ext4.o
  30
  31  ext4-y := balloc.o bitmap.o block_validity.o dir.o ext4_jbd2.o extents.o \
  32                 extents_status.o file.o fsmap.o fsync.o hash.o ialloc.o \
  33 +               htree_lock.o \
  34                 indirect.o inline.o inode.o ioctl.o mballoc.o migrate.o \
  35                 mmp.o move_extent.o namei.o page-io.o readpage.o resize.o \
  36                 super.o symlink.o sysfs.o xattr.o xattr_hurd.o xattr_trusted.o \
  37 diff --git a/fs/ext4/ext4.h b/fs/ext4/ext4.h
  38 index 9986f39..cbb58c7 100644
  39 --- a/fs/ext4/ext4.h
  40 +++ b/fs/ext4/ext4.h
  41 @@ -29,6 +29,7 @@
  42  #include <linux/mutex.h>
  43  #include <linux/timer.h>
  44  #include <linux/wait.h>
  45 +#include <linux/htree_lock.h>
  46  #include <linux/sched/signal.h>
  47  #include <linux/blockgroup_lock.h>
  48  #include <linux/percpu_counter.h>
  49 @@ -1030,6 +1031,9 @@ struct ext4_inode_info {
  50         __u32   i_dtime;
  51         ext4_fsblk_t    i_file_acl;
  52
  53 +       /* following fields for parallel directory operations -bzzz */
  54 +       struct semaphore i_append_sem;
  55 +
  56         /*
  57          * i_block_group is the number of the block group which contains
  58          * this file's inode.  Constant across the lifetime of the inode,
  59 @@ -2566,6 +2570,72 @@ struct dx_hash_info
  60   */
  61  #define HASH_NB_ALWAYS         1
  62
  63 +/* assume name-hash is protected by upper layer */
  64 +#define EXT4_HTREE_LOCK_HASH   0
  65 +
  66 +enum ext4_pdo_lk_types {
  67 +#if EXT4_HTREE_LOCK_HASH
  68 +       EXT4_LK_HASH,
  69 +#endif
  70 +       EXT4_LK_DX,             /* index block */
  71 +       EXT4_LK_DE,             /* directory entry block */
  72 +       EXT4_LK_SPIN,           /* spinlock */
  73 +       EXT4_LK_MAX,
  74 +};
  75 +
  76 +/* read-only bit */
  77 +#define EXT4_LB_RO(b)          (1 << (b))
  78 +/* read + write, high bits for writer */
  79 +#define EXT4_LB_RW(b)          ((1 << (b)) | (1 << (EXT4_LK_MAX + (b))))
  80 +
  81 +enum ext4_pdo_lock_bits {
  82 +       /* DX lock bits */
  83 +       EXT4_LB_DX_RO           = EXT4_LB_RO(EXT4_LK_DX),
  84 +       EXT4_LB_DX              = EXT4_LB_RW(EXT4_LK_DX),
  85 +       /* DE lock bits */
  86 +       EXT4_LB_DE_RO           = EXT4_LB_RO(EXT4_LK_DE),
  87 +       EXT4_LB_DE              = EXT4_LB_RW(EXT4_LK_DE),
  88 +       /* DX spinlock bits */
  89 +       EXT4_LB_SPIN_RO         = EXT4_LB_RO(EXT4_LK_SPIN),
  90 +       EXT4_LB_SPIN            = EXT4_LB_RW(EXT4_LK_SPIN),
  91 +       /* accurate searching */
  92 +       EXT4_LB_EXACT           = EXT4_LB_RO(EXT4_LK_MAX << 1),
  93 +};
  94 +
  95 +enum ext4_pdo_lock_opc {
  96 +       /* external */
  97 +       EXT4_HLOCK_READDIR      = (EXT4_LB_DE_RO | EXT4_LB_DX_RO),
  98 +       EXT4_HLOCK_LOOKUP       = (EXT4_LB_DE_RO | EXT4_LB_SPIN_RO |
  99 +                                  EXT4_LB_EXACT),
 100 +       EXT4_HLOCK_DEL          = (EXT4_LB_DE | EXT4_LB_SPIN_RO |
 101 +                                  EXT4_LB_EXACT),
 102 +       EXT4_HLOCK_ADD          = (EXT4_LB_DE | EXT4_LB_SPIN_RO),
 103 +
 104 +       /* internal */
 105 +       EXT4_HLOCK_LOOKUP_SAFE  = (EXT4_LB_DE_RO | EXT4_LB_DX_RO |
 106 +                                  EXT4_LB_EXACT),
 107 +       EXT4_HLOCK_DEL_SAFE     = (EXT4_LB_DE | EXT4_LB_DX_RO | EXT4_LB_EXACT),
 108 +       EXT4_HLOCK_SPLIT        = (EXT4_LB_DE | EXT4_LB_DX | EXT4_LB_SPIN),
 109 +};
 110 +
 111 +extern struct htree_lock_head *ext4_htree_lock_head_alloc(unsigned hbits);
 112 +#define ext4_htree_lock_head_free(lhead)       htree_lock_head_free(lhead)
 113 +
 114 +extern struct htree_lock *ext4_htree_lock_alloc(void);
 115 +#define ext4_htree_lock_free(lck)              htree_lock_free(lck)
 116 +
 117 +extern void ext4_htree_lock(struct htree_lock *lck,
 118 +                           struct htree_lock_head *lhead,
 119 +                           struct inode *dir, unsigned flags);
 120 +#define ext4_htree_unlock(lck)                  htree_unlock(lck)
 121 +
 122 +extern struct buffer_head *ext4_find_entry_locked(struct inode *dir,
 123 +                                       const struct qstr *d_name,
 124 +                                       struct ext4_dir_entry_2 **res_dir,
 125 +                                       int *inlined, struct htree_lock *lck);
 126 +extern int ext4_add_entry_locked(handle_t *handle, struct dentry *dentry,
 127 +                     struct inode *inode, struct htree_lock *lck);
 128 +
 129  struct ext4_filename {
 130         const struct qstr *usr_fname;
 131         struct fscrypt_str disk_name;
 132 @@ -2896,12 +2966,20 @@ void ext4_insert_dentry(struct inode *dir, struct inode *inode,
 133                         void *data);
 134  static inline void ext4_update_dx_flag(struct inode *inode)
 135  {
 136 +       /* Disable it for ldiskfs, because going from a DX directory to
 137 +        * a non-DX directory while it is in use will completely break
 138 +        * the htree-locking.
 139 +        * If we really want to support this operation in the future,
 140 +        * we need to exclusively lock the directory at here which will
 141 +        * increase complexity of code */
 142 +#if 0
 143         if (!ext4_has_feature_dir_index(inode->i_sb) &&
 144             ext4_test_inode_flag(inode, EXT4_INODE_INDEX)) {
 145                 /* ext4_iget() should have caught this... */
 146                 WARN_ON_ONCE(ext4_has_feature_metadata_csum(inode->i_sb));
 147                 ext4_clear_inode_flag(inode, EXT4_INODE_INDEX);
 148         }
 149 +#endif
 150  }
 151  static const unsigned char ext4_filetype_table[] = {
 152         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
 153 diff --git a/fs/ext4/namei.c b/fs/ext4/namei.c
 154 index 54bbe22..9b20bc5 100644
 155 --- a/fs/ext4/namei.c
 156 +++ b/fs/ext4/namei.c
 157 @@ -56,6 +56,7 @@ struct buffer_head *ext4_append(handle_t *handle,
 158  {
 159         struct ext4_map_blocks map;
 160         struct buffer_head *bh;
 161 +       struct ext4_inode_info *ei = EXT4_I(inode);
 162         int err;
 163
 164         if (unlikely(EXT4_SB(inode->i_sb)->s_max_dir_size_kb &&
 165 @@ -63,6 +64,10 @@ struct buffer_head *ext4_append(handle_t *handle,
 166                       EXT4_SB(inode->i_sb)->s_max_dir_size_kb)))
 167                 return ERR_PTR(-ENOSPC);
 168
 169 +       /* with parallel dir operations all appends
 170 +       * have to be serialized -bzzz */
 171 +       down(&ei->i_append_sem);
 172 +
 173         *block = inode->i_size >> inode->i_sb->s_blocksize_bits;
 174         map.m_lblk = *block;
 175         map.m_len = 1;
 176 @@ -74,15 +79,18 @@ struct buffer_head *ext4_append(handle_t *handle,
 177          */
 178         err = ext4_map_blocks(NULL, inode, &map, 0);
 179         if (err < 0)
 180 -               return ERR_PTR(err);
 181 +               goto err_unlock;
 182         if (err) {
 183                 EXT4_ERROR_INODE(inode, "Logical block already allocated");
 184 -               return ERR_PTR(-EFSCORRUPTED);
 185 +               err = -EFSCORRUPTED;
 186 +               goto err_unlock;
 187         }
 188
 189         bh = ext4_bread(handle, inode, *block, EXT4_GET_BLOCKS_CREATE);
 190 -       if (IS_ERR(bh))
 191 +       if (IS_ERR(bh)) {
 192 +               up(&ei->i_append_sem);
 193                 return bh;
 194 +       }
 195         inode->i_size += inode->i_sb->s_blocksize;
 196         EXT4_I(inode)->i_disksize = inode->i_size;
 197         err = ext4_mark_inode_dirty(handle, inode);
 198 @@ -93,11 +101,14 @@ struct buffer_head *ext4_append(handle_t *handle,
 199                                             EXT4_JTR_NONE);
 200         if (err)
 201                 goto out;
 202 +       up(&ei->i_append_sem);
 203         return bh;
 204
 205  out:
 206         brelse(bh);
 207         ext4_std_error(inode->i_sb, err);
 208 +err_unlock:
 209 +       up(&ei->i_append_sem);
 210         return ERR_PTR(err);
 211  }
 212
 213 @@ -296,7 +307,8 @@ static unsigned dx_node_limit(struct inode *dir);
 214  static struct dx_frame *dx_probe(struct ext4_filename *fname,
 215                                  struct inode *dir,
 216                                  struct dx_hash_info *hinfo,
 217 -                                struct dx_frame *frame);
 218 +                                struct dx_frame *frame,
 219 +                                struct htree_lock *lck);
 220  static void dx_release(struct dx_frame *frames, struct inode *dir);
 221  static int dx_make_map(struct inode *dir, struct buffer_head *bh,
 222                        struct dx_hash_info *hinfo,
 223 @@ -312,12 +324,13 @@ static void dx_insert_block(struct dx_frame *frame,
 224  static int ext4_htree_next_block(struct inode *dir, __u32 hash,
 225                                  struct dx_frame *frame,
 226                                  struct dx_frame *frames,
 227 -                                __u32 *start_hash);
 228 +                                __u32 *start_hash, struct htree_lock *lck);
 229  static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
 230                 struct ext4_filename *fname,
 231 -               struct ext4_dir_entry_2 **res_dir);
 232 +               struct ext4_dir_entry_2 **res_dir, struct htree_lock *lck);
 233  static int ext4_dx_add_entry(handle_t *handle, struct ext4_filename *fname,
 234 -                            struct inode *dir, struct inode *inode);
 235 +                            struct inode *dir, struct inode *inode,
 236 +                            struct htree_lock *lck);
 237
 238  /* checksumming functions */
 239  void ext4_initialize_dirent_tail(struct buffer_head *bh,
 240 @@ -802,6 +815,227 @@ static inline void htree_rep_invariant_check(struct dx_entry *at,
 241  }
 242  #endif /* DX_DEBUG */
 243
 244 +/* private data for htree_lock */
 245 +struct ext4_dir_lock_data {
 246 +       unsigned                ld_flags;  /* bits-map for lock types */
 247 +       unsigned                ld_count;  /* # entries of the last DX block */
 248 +       struct dx_entry         ld_at_entry; /* copy of leaf dx_entry */
 249 +       struct dx_entry         *ld_at;    /* position of leaf dx_entry */
 250 +};
 251 +
 252 +#define ext4_htree_lock_data(l)        ((struct ext4_dir_lock_data *)(l)->lk_private)
 253 +#define ext4_find_entry(dir, name, dirent, inline) \
 254 +                       ext4_find_entry_locked(dir, name, dirent, inline, NULL)
 255 +#define ext4_add_entry(handle, dentry, inode) \
 256 +                       ext4_add_entry_locked(handle, dentry, inode, NULL)
 257 +
 258 +/* NB: ext4_lblk_t is 32 bits so we use high bits to identify invalid blk */
 259 +#define EXT4_HTREE_NODE_CHANGED        (0xcafeULL << 32)
 260 +
 261 +static void ext4_htree_event_cb(void *target, void *event)
 262 +{
 263 +       u64 *block = (u64 *)target;
 264 +
 265 +       if (*block == dx_get_block((struct dx_entry *)event))
 266 +               *block = EXT4_HTREE_NODE_CHANGED;
 267 +}
 268 +
 269 +struct htree_lock_head *ext4_htree_lock_head_alloc(unsigned hbits)
 270 +{
 271 +       struct htree_lock_head *lhead;
 272 +
 273 +       lhead = htree_lock_head_alloc(EXT4_LK_MAX, hbits, 0);
 274 +       if (lhead != NULL) {
 275 +               htree_lock_event_attach(lhead, EXT4_LK_SPIN, HTREE_EVENT_WR,
 276 +                                       ext4_htree_event_cb);
 277 +       }
 278 +       return lhead;
 279 +}
 280 +EXPORT_SYMBOL(ext4_htree_lock_head_alloc);
 281 +
 282 +struct htree_lock *ext4_htree_lock_alloc(void)
 283 +{
 284 +       return htree_lock_alloc(EXT4_LK_MAX,
 285 +                               sizeof(struct ext4_dir_lock_data));
 286 +}
 287 +EXPORT_SYMBOL(ext4_htree_lock_alloc);
 288 +
 289 +static htree_lock_mode_t ext4_htree_mode(unsigned flags)
 290 +{
 291 +       switch (flags) {
 292 +       default: /* 0 or unknown flags require EX lock */
 293 +               return HTREE_LOCK_EX;
 294 +       case EXT4_HLOCK_READDIR:
 295 +               return HTREE_LOCK_PR;
 296 +       case EXT4_HLOCK_LOOKUP:
 297 +               return HTREE_LOCK_CR;
 298 +       case EXT4_HLOCK_DEL:
 299 +       case EXT4_HLOCK_ADD:
 300 +               return HTREE_LOCK_CW;
 301 +       }
 302 +}
 303 +
 304 +/* return PR for read-only operations, otherwise return EX */
 305 +static inline htree_lock_mode_t ext4_htree_safe_mode(unsigned flags)
 306 +{
 307 +       int writer = (flags & EXT4_LB_DE) == EXT4_LB_DE;
 308 +
 309 +       /* 0 requires EX lock */
 310 +       return (flags == 0 || writer) ? HTREE_LOCK_EX : HTREE_LOCK_PR;
 311 +}
 312 +
 313 +static int ext4_htree_safe_locked(struct htree_lock *lck)
 314 +{
 315 +       int writer;
 316 +
 317 +       if (lck == NULL || lck->lk_mode == HTREE_LOCK_EX)
 318 +               return 1;
 319 +
 320 +       writer = (ext4_htree_lock_data(lck)->ld_flags & EXT4_LB_DE) ==
 321 +                EXT4_LB_DE;
 322 +       if (writer) /* all readers & writers are excluded? */
 323 +               return lck->lk_mode == HTREE_LOCK_EX;
 324 +
 325 +       /* all writers are excluded? */
 326 +       return lck->lk_mode == HTREE_LOCK_PR ||
 327 +              lck->lk_mode == HTREE_LOCK_PW ||
 328 +              lck->lk_mode == HTREE_LOCK_EX;
 329 +}
 330 +
 331 +/* relock htree_lock with EX mode if it's change operation, otherwise
 332 + * relock it with PR mode. It's noop if PDO is disabled. */
 333 +static void ext4_htree_safe_relock(struct htree_lock *lck)
 334 +{
 335 +       if (!ext4_htree_safe_locked(lck)) {
 336 +               unsigned flags = ext4_htree_lock_data(lck)->ld_flags;
 337 +
 338 +               htree_change_lock(lck, ext4_htree_safe_mode(flags));
 339 +       }
 340 +}
 341 +
 342 +void ext4_htree_lock(struct htree_lock *lck, struct htree_lock_head *lhead,
 343 +                    struct inode *dir, unsigned flags)
 344 +{
 345 +       htree_lock_mode_t mode = is_dx(dir) ? ext4_htree_mode(flags) :
 346 +                                             ext4_htree_safe_mode(flags);
 347 +
 348 +       ext4_htree_lock_data(lck)->ld_flags = flags;
 349 +       htree_lock(lck, lhead, mode);
 350 +       if (!is_dx(dir))
 351 +               ext4_htree_safe_relock(lck); /* make sure it's safe locked */
 352 +}
 353 +EXPORT_SYMBOL(ext4_htree_lock);
 354 +
 355 +static int ext4_htree_node_lock(struct htree_lock *lck, struct dx_entry *at,
 356 +                               unsigned lmask, int wait, void *ev)
 357 +{
 358 +       u32     key = (at == NULL) ? 0 : dx_get_block(at);
 359 +       u32     mode;
 360 +
 361 +       /* NOOP if htree is well protected or caller doesn't require the lock */
 362 +       if (ext4_htree_safe_locked(lck) ||
 363 +          !(ext4_htree_lock_data(lck)->ld_flags & lmask))
 364 +               return 1;
 365 +
 366 +       mode = (ext4_htree_lock_data(lck)->ld_flags & lmask) == lmask ?
 367 +               HTREE_LOCK_PW : HTREE_LOCK_PR;
 368 +       while (1) {
 369 +               if (htree_node_lock_try(lck, mode, key, ffz(~lmask), wait, ev))
 370 +                       return 1;
 371 +               if (!(lmask & EXT4_LB_SPIN)) /* not a spinlock */
 372 +                       return 0;
 373 +               cpu_relax(); /* spin until granted */
 374 +       }
 375 +}
 376 +
 377 +static int ext4_htree_node_locked(struct htree_lock *lck, unsigned lmask)
 378 +{
 379 +       return ext4_htree_safe_locked(lck) ||
 380 +              htree_node_is_granted(lck, ffz(~lmask));
 381 +}
 382 +
 383 +static void ext4_htree_node_unlock(struct htree_lock *lck,
 384 +                                  unsigned lmask, void *buf)
 385 +{
 386 +       /* NB: it's safe to call mutiple times or even it's not locked */
 387 +       if (!ext4_htree_safe_locked(lck) &&
 388 +            htree_node_is_granted(lck, ffz(~lmask)))
 389 +               htree_node_unlock(lck, ffz(~lmask), buf);
 390 +}
 391 +
 392 +#define ext4_htree_dx_lock(lck, key)           \
 393 +       ext4_htree_node_lock(lck, key, EXT4_LB_DX, 1, NULL)
 394 +#define ext4_htree_dx_lock_try(lck, key)       \
 395 +       ext4_htree_node_lock(lck, key, EXT4_LB_DX, 0, NULL)
 396 +#define ext4_htree_dx_unlock(lck)              \
 397 +       ext4_htree_node_unlock(lck, EXT4_LB_DX, NULL)
 398 +#define ext4_htree_dx_locked(lck)              \
 399 +       ext4_htree_node_locked(lck, EXT4_LB_DX)
 400 +
 401 +static void ext4_htree_dx_need_lock(struct htree_lock *lck)
 402 +{
 403 +       struct ext4_dir_lock_data *ld;
 404 +
 405 +       if (ext4_htree_safe_locked(lck))
 406 +               return;
 407 +
 408 +       ld = ext4_htree_lock_data(lck);
 409 +       switch (ld->ld_flags) {
 410 +       default:
 411 +               return;
 412 +       case EXT4_HLOCK_LOOKUP:
 413 +               ld->ld_flags = EXT4_HLOCK_LOOKUP_SAFE;
 414 +               return;
 415 +       case EXT4_HLOCK_DEL:
 416 +               ld->ld_flags = EXT4_HLOCK_DEL_SAFE;
 417 +               return;
 418 +       case EXT4_HLOCK_ADD:
 419 +               ld->ld_flags = EXT4_HLOCK_SPLIT;
 420 +               return;
 421 +       }
 422 +}
 423 +
 424 +#define ext4_htree_de_lock(lck, key)           \
 425 +       ext4_htree_node_lock(lck, key, EXT4_LB_DE, 1, NULL)
 426 +#define ext4_htree_de_unlock(lck)              \
 427 +       ext4_htree_node_unlock(lck, EXT4_LB_DE, NULL)
 428 +
 429 +#define ext4_htree_spin_lock(lck, key, event)  \
 430 +       ext4_htree_node_lock(lck, key, EXT4_LB_SPIN, 0, event)
 431 +#define ext4_htree_spin_unlock(lck)            \
 432 +       ext4_htree_node_unlock(lck, EXT4_LB_SPIN, NULL)
 433 +#define ext4_htree_spin_unlock_listen(lck, p)  \
 434 +       ext4_htree_node_unlock(lck, EXT4_LB_SPIN, p)
 435 +
 436 +static void ext4_htree_spin_stop_listen(struct htree_lock *lck)
 437 +{
 438 +       if (!ext4_htree_safe_locked(lck) &&
 439 +           htree_node_is_listening(lck, ffz(~EXT4_LB_SPIN)))
 440 +               htree_node_stop_listen(lck, ffz(~EXT4_LB_SPIN));
 441 +}
 442 +
 443 +enum {
 444 +       DX_HASH_COL_IGNORE,     /* ignore collision while probing frames */
 445 +       DX_HASH_COL_YES,        /* there is collision and it does matter */
 446 +       DX_HASH_COL_NO,         /* there is no collision */
 447 +};
 448 +
 449 +static int dx_probe_hash_collision(struct htree_lock *lck,
 450 +                                  struct dx_entry *entries,
 451 +                                  struct dx_entry *at, u32 hash)
 452 +{
 453 +       if (!(lck && ext4_htree_lock_data(lck)->ld_flags & EXT4_LB_EXACT)) {
 454 +               return DX_HASH_COL_IGNORE; /* don't care about collision */
 455 +
 456 +       } else if (at == entries + dx_get_count(entries) - 1) {
 457 +               return DX_HASH_COL_IGNORE; /* not in any leaf of this DX */
 458 +
 459 +       } else { /* hash collision? */
 460 +               return ((dx_get_hash(at + 1) & ~1) == hash) ?
 461 +                       DX_HASH_COL_YES : DX_HASH_COL_NO;
 462 +       }
 463 +}
 464 +
 465  /*
 466   * Probe for a directory leaf block to search.
 467   *
 468 @@ -813,10 +1047,11 @@ static inline void htree_rep_invariant_check(struct dx_entry *at,
 469   */
 470  static struct dx_frame *
 471  dx_probe(struct ext4_filename *fname, struct inode *dir,
 472 -        struct dx_hash_info *hinfo, struct dx_frame *frame_in)
 473 +        struct dx_hash_info *hinfo, struct dx_frame *frame_in,
 474 +        struct htree_lock *lck)
 475  {
 476         unsigned count, indirect, level, i;
 477 -       struct dx_entry *at, *entries, *p, *q, *m;
 478 +       struct dx_entry *at, *entries, *p, *q, *m, *dx = NULL;
 479         struct dx_root_info *info;
 480         struct dx_frame *frame = frame_in;
 481         struct dx_frame *ret_err = ERR_PTR(ERR_BAD_DX_DIR);
 482 @@ -900,8 +1135,16 @@ dx_probe(struct ext4_filename *fname, struct inode *dir,
 483         level = 0;
 484         blocks[0] = 0;
 485         while (1) {
 486 +               if (indirect == level) { /* the last index level */
 487 +                       /* NB: ext4_htree_dx_lock() could be noop if
 488 +                        * DX-lock flag is not set for current operation
 489 +                        */
 490 +                       ext4_htree_dx_lock(lck, dx);
 491 +                       ext4_htree_spin_lock(lck, dx, NULL);
 492 +               }
 493                 count = dx_get_count(entries);
 494                 if (!count || count > dx_get_limit(entries)) {
 495 +                       ext4_htree_spin_unlock(lck); /* release spin */
 496                         ext4_warning_inode(dir,
 497                                            "dx entry: count %u beyond limit %u",
 498                                            count, dx_get_limit(entries));
 499 @@ -928,6 +1171,74 @@ dx_probe(struct ext4_filename *fname, struct inode *dir,
 500                 frame->entries = entries;
 501                 frame->at = at;
 502
 503 +               if (indirect == level) { /* the last index level */
 504 +                       struct ext4_dir_lock_data *ld;
 505 +                       u64 myblock;
 506 +
 507 +                       /* By default we only lock DE-block, however, we will
 508 +                        * also lock the last level DX-block if:
 509 +                        * a) there is hash collision
 510 +                        *    we will set DX-lock flag (a few lines below)
 511 +                        *    and redo to lock DX-block
 512 +                        *    see detail in dx_probe_hash_collision()
 513 +                        * b) it's a retry from splitting
 514 +                        *    we need to lock the last level DX-block so nobody
 515 +                        *    else can split any leaf blocks under the same
 516 +                        *    DX-block, see detail in ext4_dx_add_entry()
 517 +                        */
 518 +                       if (ext4_htree_dx_locked(lck)) {
 519 +                               /* DX-block is locked, just lock DE-block
 520 +                                * and return
 521 +                                */
 522 +                               ext4_htree_spin_unlock(lck);
 523 +                               if (!ext4_htree_safe_locked(lck))
 524 +                                       ext4_htree_de_lock(lck, frame->at);
 525 +                               return frame;
 526 +                       }
 527 +                       /* it's pdirop and no DX lock */
 528 +                       if (dx_probe_hash_collision(lck, entries, at, hash) ==
 529 +                           DX_HASH_COL_YES) {
 530 +                               /* found hash collision, set DX-lock flag
 531 +                                * and retry to abtain DX-lock
 532 +                                */
 533 +                               ext4_htree_spin_unlock(lck);
 534 +                               ext4_htree_dx_need_lock(lck);
 535 +                               continue;
 536 +                       }
 537 +                       ld = ext4_htree_lock_data(lck);
 538 +                       /* because I don't lock DX, so @at can't be trusted
 539 +                        * after I release spinlock so I have to save it
 540 +                        */
 541 +                       ld->ld_at = at;
 542 +                       ld->ld_at_entry = *at;
 543 +                       ld->ld_count = dx_get_count(entries);
 544 +
 545 +                       frame->at = &ld->ld_at_entry;
 546 +                       myblock = dx_get_block(at);
 547 +
 548 +                       /* NB: ordering locking */
 549 +                       ext4_htree_spin_unlock_listen(lck, &myblock);
 550 +                       /* other thread can split this DE-block because:
 551 +                        * a) I don't have lock for the DE-block yet
 552 +                        * b) I released spinlock on DX-block
 553 +                        * if it happened I can detect it by listening
 554 +                        * splitting event on this DE-block
 555 +                        */
 556 +                       ext4_htree_de_lock(lck, frame->at);
 557 +                       ext4_htree_spin_stop_listen(lck);
 558 +
 559 +                       if (myblock == EXT4_HTREE_NODE_CHANGED) {
 560 +                               /* someone split this DE-block before
 561 +                                * I locked it, I need to retry and lock
 562 +                                * valid DE-block
 563 +                                */
 564 +                               ext4_htree_de_unlock(lck);
 565 +                               continue;
 566 +                       }
 567 +                       return frame;
 568 +               }
 569 +               dx = at;
 570 +
 571                 block = dx_get_block(at);
 572                 for (i = 0; i <= level; i++) {
 573                         if (blocks[i] == block) {
 574 @@ -937,8 +1248,7 @@ dx_probe(struct ext4_filename *fname, struct inode *dir,
 575                                 goto fail;
 576                         }
 577                 }
 578 -               if (++level > indirect)
 579 -                       return frame;
 580 +               ++level;
 581                 blocks[level] = block;
 582                 frame++;
 583                 frame->bh = ext4_read_dirblock(dir, block, INDEX);
 584 @@ -1009,7 +1319,7 @@ static void dx_release(struct dx_frame *frames, struct inode *dir)
 585  static int ext4_htree_next_block(struct inode *dir, __u32 hash,
 586                                  struct dx_frame *frame,
 587                                  struct dx_frame *frames,
 588 -                                __u32 *start_hash)
 589 +                                __u32 *start_hash, struct htree_lock *lck)
 590  {
 591         struct dx_frame *p;
 592         struct buffer_head *bh;
 593 @@ -1024,12 +1334,22 @@ static int ext4_htree_next_block(struct inode *dir, __u32 hash,
 594          * this loop, num_frames indicates the number of interior
 595          * nodes need to be read.
 596          */
 597 +       ext4_htree_de_unlock(lck);
 598         while (1) {
 599 -               if (++(p->at) < p->entries + dx_get_count(p->entries))
 600 -                       break;
 601 +               if (num_frames > 0 || ext4_htree_dx_locked(lck)) {
 602 +                       /* num_frames > 0 :
 603 +                        *   DX block
 604 +                        * ext4_htree_dx_locked:
 605 +                        *   frame->at is reliable pointer returned by dx_probe,
 606 +                        *   otherwise dx_probe already knew no collision */
 607 +                       if (++(p->at) < p->entries + dx_get_count(p->entries))
 608 +                               break;
 609 +               }
 610                 if (p == frames)
 611                         return 0;
 612                 num_frames++;
 613 +               if (num_frames == 1)
 614 +                       ext4_htree_dx_unlock(lck);
 615                 p--;
 616         }
 617
 618 @@ -1052,6 +1372,13 @@ static int ext4_htree_next_block(struct inode *dir, __u32 hash,
 619          * block so no check is necessary
 620          */
 621         while (num_frames--) {
 622 +               if (num_frames == 0) {
 623 +                       /* it's not always necessary, we just don't want to
 624 +                        * detect hash collision again */
 625 +                       ext4_htree_dx_need_lock(lck);
 626 +                       ext4_htree_dx_lock(lck, p->at);
 627 +               }
 628 +
 629                 bh = ext4_read_dirblock(dir, dx_get_block(p->at), INDEX);
 630                 if (IS_ERR(bh))
 631                         return PTR_ERR(bh);
 632 @@ -1060,6 +1387,7 @@ static int ext4_htree_next_block(struct inode *dir, __u32 hash,
 633                 p->bh = bh;
 634                 p->at = p->entries = ((struct dx_node *) bh->b_data)->entries;
 635         }
 636 +       ext4_htree_de_lock(lck, p->at);
 637         return 1;
 638  }
 639
 640 @@ -1221,10 +1549,10 @@ int ext4_htree_fill_tree(struct file *dir_file, __u32 start_hash,
 641         }
 642         hinfo.hash = start_hash;
 643         hinfo.minor_hash = 0;
 644 -       frame = dx_probe(NULL, dir, &hinfo, frames);
 645 +       /* assume it's PR locked */
 646 +       frame = dx_probe(NULL, dir, &hinfo, frames, NULL);
 647         if (IS_ERR(frame))
 648                 return PTR_ERR(frame);
 649 -
 650         /* Add '.' and '..' from the htree header */
 651         if (!start_hash && !start_minor_hash) {
 652                 de = (struct ext4_dir_entry_2 *) frames[0].bh->b_data;
 653 @@ -1264,7 +1592,7 @@ int ext4_htree_fill_tree(struct file *dir_file, __u32 start_hash,
 654                 count += ret;
 655                 hashval = ~0;
 656                 ret = ext4_htree_next_block(dir, HASH_NB_ALWAYS,
 657 -                                           frame, frames, &hashval);
 658 +                                           frame, frames, &hashval, NULL);
 659                 *next_hash = hashval;
 660                 if (ret < 0) {
 661                         err = ret;
 662 @@ -1584,7 +1912,7 @@ static int is_dx_internal_node(struct inode *dir, ext4_lblk_t block,
 663  static struct buffer_head *__ext4_find_entry(struct inode *dir,
 664                                              struct ext4_filename *fname,
 665                                              struct ext4_dir_entry_2 **res_dir,
 666 -                                            int *inlined)
 667 +                                            int *inlined, struct htree_lock *lck)
 668  {
 669         struct super_block *sb;
 670         struct buffer_head *bh_use[NAMEI_RA_SIZE];
 671 @@ -1626,7 +1954,7 @@ static struct buffer_head *__ext4_find_entry(struct inode *dir,
 672                 goto restart;
 673         }
 674         if (is_dx(dir)) {
 675 -               ret = ext4_dx_find_entry(dir, fname, res_dir);
 676 +               ret = ext4_dx_find_entry(dir, fname, res_dir, lck);
 677                 /*
 678                  * On success, or if the error was file not found,
 679                  * return.  Otherwise, fall back to doing a search the
 680 @@ -1636,6 +1964,7 @@ static struct buffer_head *__ext4_find_entry(struct inode *dir,
 681                         goto cleanup_and_exit;
 682                 dxtrace(printk(KERN_DEBUG "ext4_find_entry: dx failed, "
 683                                "falling back\n"));
 684 +               ext4_htree_safe_relock(lck);
 685                 ret = NULL;
 686         }
 687         nblocks = dir->i_size >> EXT4_BLOCK_SIZE_BITS(sb);
 688 @@ -1726,10 +2055,10 @@ cleanup_and_exit:
 689         return ret;
 690  }
 691
 692 -static struct buffer_head *ext4_find_entry(struct inode *dir,
 693 +struct buffer_head *ext4_find_entry_locked(struct inode *dir,
 694                                            const struct qstr *d_name,
 695                                            struct ext4_dir_entry_2 **res_dir,
 696 -                                          int *inlined)
 697 +                                          int *inlined, struct htree_lock *lck)
 698  {
 699         int err;
 700         struct ext4_filename fname;
 701 @@ -1741,12 +2070,14 @@ static struct buffer_head *ext4_find_entry(struct inode *dir,
 702         if (err)
 703                 return ERR_PTR(err);
 704
 705 -       bh = __ext4_find_entry(dir, &fname, res_dir, inlined);
 706 +       bh = __ext4_find_entry(dir, &fname, res_dir, inlined, lck);
 707
 708         ext4_fname_free_filename(&fname);
 709         return bh;
 710  }
 711
 712 +EXPORT_SYMBOL(ext4_find_entry_locked);
 713 +
 714  static struct buffer_head *ext4_lookup_entry(struct inode *dir,
 715                                              struct dentry *dentry,
 716                                              struct ext4_dir_entry_2 **res_dir)
 717 @@ -1762,7 +2093,7 @@ static struct buffer_head *ext4_lookup_entry(struct inode *dir,
 718         if (err)
 719                 return ERR_PTR(err);
 720
 721 -       bh = __ext4_find_entry(dir, &fname, res_dir, NULL);
 722 +       bh = __ext4_find_entry(dir, &fname, res_dir, NULL, NULL);
 723
 724         ext4_fname_free_filename(&fname);
 725         return bh;
 726 @@ -1770,7 +2101,8 @@ static struct buffer_head *ext4_lookup_entry(struct inode *dir,
 727
 728  static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
 729                         struct ext4_filename *fname,
 730 -                       struct ext4_dir_entry_2 **res_dir)
 731 +                       struct ext4_dir_entry_2 **res_dir,
 732 +                       struct htree_lock *lck)
 733  {
 734         struct super_block * sb = dir->i_sb;
 735         struct dx_frame frames[EXT4_HTREE_LEVEL], *frame;
 736 @@ -1781,7 +2113,7 @@ static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
 737  #ifdef CONFIG_FS_ENCRYPTION
 738         *res_dir = NULL;
 739  #endif
 740 -       frame = dx_probe(fname, dir, NULL, frames);
 741 +       frame = dx_probe(fname, dir, NULL, frames, lck);
 742         if (IS_ERR(frame))
 743                 return (struct buffer_head *) frame;
 744         do {
 745 @@ -1803,7 +2135,7 @@ static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
 746
 747                 /* Check to see if we should continue to search */
 748                 retval = ext4_htree_next_block(dir, fname->hinfo.hash, frame,
 749 -                                              frames, NULL);
 750 +                                              frames, NULL, lck);
 751                 if (retval < 0) {
 752                         ext4_warning_inode(dir,
 753                                 "error %d reading directory index block",
 754 @@ -1992,8 +2324,9 @@ static struct ext4_dir_entry_2 *dx_pack_dirents(struct inode *dir, char *base,
 755   * Returns pointer to de in block into which the new entry will be inserted.
 756   */
 757  static struct ext4_dir_entry_2 *do_split(handle_t *handle, struct inode *dir,
 758 -                       struct buffer_head **bh,struct dx_frame *frame,
 759 -                       struct dx_hash_info *hinfo)
 760 +                       struct buffer_head **bh, struct dx_frame *frames,
 761 +                       struct dx_frame *frame, struct dx_hash_info *hinfo,
 762 +                       struct htree_lock *lck)
 763  {
 764         unsigned blocksize = dir->i_sb->s_blocksize;
 765         unsigned continued;
 766 @@ -2070,8 +2403,14 @@ static struct ext4_dir_entry_2 *do_split(handle_t *handle, struct inode *dir,
 767                                         hash2, split, count-split));
 768
 769         /* Fancy dance to stay within two buffers */
 770 -       de2 = dx_move_dirents(dir, data1, data2, map + split, count - split,
 771 -                             blocksize);
 772 +       if (hinfo->hash < hash2) {
 773 +               de2 = dx_move_dirents(dir, data1, data2, map + split,
 774 +                                     count - split, blocksize);
 775 +       } else {
 776 +               /* make sure we will add entry to the same block which
 777 +                * we have already locked */
 778 +               de2 = dx_move_dirents(dir, data1, data2, map, split, blocksize);
 779 +       }
 780         de = dx_pack_dirents(dir, data1, blocksize);
 781         de->rec_len = ext4_rec_len_to_disk(data1 + (blocksize - csum_size) -
 782                                            (char *) de,
 783 @@ -2089,12 +2428,21 @@ static struct ext4_dir_entry_2 *do_split(handle_t *handle, struct inode *dir,
 784         dxtrace(dx_show_leaf(dir, hinfo, (struct ext4_dir_entry_2 *) data2,
 785                         blocksize, 1));
 786
 787 -       /* Which block gets the new entry? */
 788 -       if (hinfo->hash >= hash2) {
 789 -               swap(*bh, bh2);
 790 -               de = de2;
 791 +       ext4_htree_spin_lock(lck, frame > frames ? (frame - 1)->at : NULL,
 792 +                            frame->at); /* notify block is being split */
 793 +       if (hinfo->hash < hash2) {
 794 +               dx_insert_block(frame, hash2 + continued, newblock);
 795 +
 796 +       } else {
 797 +               /* switch block number */
 798 +               dx_insert_block(frame, hash2 + continued,
 799 +                               dx_get_block(frame->at));
 800 +               dx_set_block(frame->at, newblock);
 801 +               (frame->at)++;
 802         }
 803 -       dx_insert_block(frame, hash2 + continued, newblock);
 804 +       ext4_htree_spin_unlock(lck);
 805 +       ext4_htree_dx_unlock(lck);
 806 +
 807         err = ext4_handle_dirty_dirblock(handle, dir, bh2);
 808         if (err)
 809                 goto journal_error;
 810 @@ -2402,7 +2750,7 @@ static int make_indexed_dir(handle_t *handle, struct ext4_filename *fname,
 811         if (retval)
 812                 goto out_frames;
 813
 814 -       de = do_split(handle,dir, &bh2, frame, &fname->hinfo);
 815 +       de = do_split(handle, dir, &bh2, frames, frame, &fname->hinfo, NULL);
 816         if (IS_ERR(de)) {
 817                 retval = PTR_ERR(de);
 818                 goto out_frames;
 819 @@ -2512,8 +2860,8 @@ out:
 820   * may not sleep between calling this and putting something into
 821   * the entry, as someone else might have used it while you slept.
 822   */
 823 -static int ext4_add_entry(handle_t *handle, struct dentry *dentry,
 824 -                         struct inode *inode)
 825 +int ext4_add_entry_locked(handle_t *handle, struct dentry *dentry,
 826 +                         struct inode *inode, struct htree_lock *lck)
 827  {
 828         struct inode *dir = d_inode(dentry->d_parent);
 829         struct buffer_head *bh = NULL;
 830 @@ -2562,9 +2910,10 @@ static int ext4_add_entry(handle_t *handle, struct dentry *dentry,
 831                 if (dentry->d_name.len == 2 &&
 832                     memcmp(dentry->d_name.name, "..", 2) == 0)
 833                         return ext4_update_dotdot(handle, dentry, inode);
 834 -               retval = ext4_dx_add_entry(handle, &fname, dir, inode);
 835 +               retval = ext4_dx_add_entry(handle, &fname, dir, inode, lck);
 836                 if (!retval || (retval != ERR_BAD_DX_DIR))
 837                         goto out;
 838 +               ext4_htree_safe_relock(lck);
 839                 /* Can we just ignore htree data? */
 840                 if (ext4_has_metadata_csum(sb)) {
 841                         EXT4_ERROR_INODE(dir,
 842 @@ -2627,12 +2976,14 @@ out:
 843                 ext4_set_inode_state(inode, EXT4_STATE_NEWENTRY);
 844         return retval;
 845  }
 846 +EXPORT_SYMBOL(ext4_add_entry_locked);
 847
 848  /*
 849   * Returns 0 for success, or a negative error value
 850   */
 851  static int ext4_dx_add_entry(handle_t *handle, struct ext4_filename *fname,
 852 -                            struct inode *dir, struct inode *inode)
 853 +                            struct inode *dir, struct inode *inode,
 854 +                            struct htree_lock *lck)
 855  {
 856         struct dx_frame frames[EXT4_HTREE_LEVEL], *frame;
 857         struct dx_entry *entries, *at;
 858 @@ -2644,7 +2995,7 @@ static int ext4_dx_add_entry(handle_t *handle, struct ext4_filename *fname,
 859
 860  again:
 861         restart = 0;
 862 -       frame = dx_probe(fname, dir, NULL, frames);
 863 +       frame = dx_probe(fname, dir, NULL, frames, lck);
 864         if (IS_ERR(frame))
 865                 return PTR_ERR(frame);
 866         entries = frame->entries;
 867 @@ -2679,6 +3030,12 @@ again:
 868                 struct dx_node *node2;
 869                 struct buffer_head *bh2;
 870
 871 +               if (!ext4_htree_safe_locked(lck)) { /* retry with EX lock */
 872 +                       ext4_htree_safe_relock(lck);
 873 +                       restart = 1;
 874 +                       goto cleanup;
 875 +               }
 876 +
 877                 while (frame > frames) {
 878                         if (dx_get_count((frame - 1)->entries) <
 879                             dx_get_limit((frame - 1)->entries)) {
 880 @@ -2782,8 +3139,32 @@ again:
 881                         restart = 1;
 882                         goto journal_error;
 883                 }
 884 +       } else if (!ext4_htree_dx_locked(lck)) {
 885 +               struct ext4_dir_lock_data *ld = ext4_htree_lock_data(lck);
 886 +
 887 +               /* not well protected, require DX lock */
 888 +               ext4_htree_dx_need_lock(lck);
 889 +               at = frame > frames ? (frame - 1)->at : NULL;
 890 +
 891 +               /* NB: no risk of deadlock because it's just a try.
 892 +                *
 893 +                * NB: we check ld_count for twice, the first time before
 894 +                * having DX lock, the second time after holding DX lock.
 895 +                *
 896 +                * NB: We never free blocks for directory so far, which
 897 +                * means value returned by dx_get_count() should equal to
 898 +                * ld->ld_count if nobody split any DE-block under @at,
 899 +                * and ld->ld_at still points to valid dx_entry. */
 900 +               if ((ld->ld_count != dx_get_count(entries)) ||
 901 +                   !ext4_htree_dx_lock_try(lck, at) ||
 902 +                   (ld->ld_count != dx_get_count(entries))) {
 903 +                       restart = 1;
 904 +                       goto cleanup;
 905 +               }
 906 +               /* OK, I've got DX lock and nothing changed */
 907 +               frame->at = ld->ld_at;
 908         }
 909 -       de = do_split(handle, dir, &bh, frame, &fname->hinfo);
 910 +       de = do_split(handle, dir, &bh, frames, frame, &fname->hinfo, lck);
 911         if (IS_ERR(de)) {
 912                 err = PTR_ERR(de);
 913                 goto cleanup;
 914 @@ -2794,6 +3175,8 @@ again:
 915  journal_error:
 916         ext4_std_error(dir->i_sb, err); /* this is a no-op if err == 0 */
 917  cleanup:
 918 +       ext4_htree_dx_unlock(lck);
 919 +       ext4_htree_de_unlock(lck);
 920         brelse(bh);
 921         dx_release(frames, dir);
 922         /* @restart is true means htree-path has been changed, we need to
 923 diff --git a/fs/ext4/super.c b/fs/ext4/super.c
 924 index 5a4b03a..8083662 100644
 925 --- a/fs/ext4/super.c
 926 +++ b/fs/ext4/super.c
 927 @@ -1324,6 +1324,7 @@ static struct inode *ext4_alloc_inode(struct super_block *sb)
 928
 929         inode_set_iversion(&ei->vfs_inode, 1);
 930         spin_lock_init(&ei->i_raw_lock);
 931 +       sema_init(&ei->i_append_sem, 1);
 932         INIT_LIST_HEAD(&ei->i_prealloc_list);
 933         atomic_set(&ei->i_prealloc_active, 0);
 934         spin_lock_init(&ei->i_prealloc_lock);
 935 --
 936 2.34.1
 937