ldiskfs/kernel_patches/patches/sles15sp1/ext4-pdirop.patch

   1 Single directory performance is a critical for HPC workloads. In a
   2 typical use case an application creates a separate output file for
   3 each node and task in a job. As nodes and tasks increase, hundreds
   4 of thousands of files may be created in a single directory within
   5 a short window of time.
   6 Today, both filename lookup and file system modifying operations
   7 (such as create and unlink) are protected with a single lock for
   8 an entire ldiskfs directory. PDO project will remove this
   9 bottleneck by introducing a parallel locking mechanism for entire
  10 ldiskfs directories. This work will enable multiple application
  11 threads to simultaneously lookup, create and unlink in parallel.
  12
  13 This patch contains:
  14  - pdirops support for ldiskfs
  15  - integrate with osd-ldiskfs
  16
  17 ---
  18  fs/ext4/Makefile           |    1
  19  fs/ext4/ext4.h             |   78 +++
  20  fs/ext4/namei.c            |  446 ++++++++++++++++++++--
  21  fs/ext4/super.c            |    1
  22  6 files changed, 1567 insertions(+), 37 deletions(-)
  23
  24 --- a/fs/ext4/Makefile
  25 +++ b/fs/ext4/Makefile
  26 @@ -6,6 +6,7 @@ obj-$(CONFIG_EXT4_FS) += ext4.o
  27
  28  ext4-y := balloc.o bitmap.o block_validity.o dir.o ext4_jbd2.o extents.o \
  29                 extents_status.o file.o fsmap.o fsync.o hash.o ialloc.o \
  30 +               htree_lock.o \
  31                 indirect.o inline.o inode.o ioctl.o mballoc.o migrate.o \
  32                 mmp.o move_extent.o namei.o page-io.o readpage.o resize.o \
  33                 super.o symlink.o sysfs.o xattr.o xattr_trusted.o xattr_user.o
  34 --- a/fs/ext4/ext4.h
  35 +++ b/fs/ext4/ext4.h
  36 @@ -28,6 +28,7 @@
  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 @@ -980,6 +981,9 @@ 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 @@ -2162,6 +2166,72 @@ 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(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(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 @@ -2473,11 +2543,19 @@ 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 +#if 0
 138         if (!ext4_has_feature_dir_index(inode->i_sb)) {
 139                 /* ext4_iget() should have caught this... */
 140                 WARN_ON_ONCE(ext4_has_feature_metadata_csum(inode->i_sb));
 141                 ext4_clear_inode_flag(inode, EXT4_INODE_INDEX);
 142         }
 143 +#endif
 144  }
 145  static const unsigned char ext4_filetype_table[] = {
 146         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
 147 --- a/fs/ext4/namei.c
 148 +++ b/fs/ext4/namei.c
 149 @@ -53,6 +53,7 @@ struct buffer_head *ext4_append(handle_t
 150                                         ext4_lblk_t *block)
 151  {
 152         struct buffer_head *bh;
 153 +       struct ext4_inode_info *ei = EXT4_I(inode);
 154         int err;
 155
 156         if (unlikely(EXT4_SB(inode->i_sb)->s_max_dir_size_kb &&
 157 @@ -60,15 +61,22 @@ struct buffer_head *ext4_append(handle_t
 158                       EXT4_SB(inode->i_sb)->s_max_dir_size_kb)))
 159                 return ERR_PTR(-ENOSPC);
 160
 161 +       /* with parallel dir operations all appends
 162 +       * have to be serialized -bzzz */
 163 +       down(&ei->i_append_sem);
 164 +
 165         *block = inode->i_size >> inode->i_sb->s_blocksize_bits;
 166
 167         bh = ext4_bread(handle, inode, *block, EXT4_GET_BLOCKS_CREATE);
 168 -       if (IS_ERR(bh))
 169 +       if (IS_ERR(bh)) {
 170 +               up(&ei->i_append_sem);
 171                 return bh;
 172 +       }
 173         inode->i_size += inode->i_sb->s_blocksize;
 174         EXT4_I(inode)->i_disksize = inode->i_size;
 175         BUFFER_TRACE(bh, "get_write_access");
 176         err = ext4_journal_get_write_access(handle, bh);
 177 +       up(&ei->i_append_sem);
 178         if (err) {
 179                 brelse(bh);
 180                 ext4_std_error(inode->i_sb, err);
 181 @@ -249,7 +257,8 @@ static unsigned dx_node_limit(struct ino
 182  static struct dx_frame *dx_probe(struct ext4_filename *fname,
 183                                  struct inode *dir,
 184                                  struct dx_hash_info *hinfo,
 185 -                                struct dx_frame *frame);
 186 +                                struct dx_frame *frame,
 187 +                                struct htree_lock *lck);
 188  static void dx_release(struct dx_frame *frames);
 189  static int dx_make_map(struct inode *dir, struct ext4_dir_entry_2 *de,
 190                        unsigned blocksize, struct dx_hash_info *hinfo,
 191 @@ -263,12 +272,13 @@ static void dx_insert_block(struct dx_fr
 192  static int ext4_htree_next_block(struct inode *dir, __u32 hash,
 193                                  struct dx_frame *frame,
 194                                  struct dx_frame *frames,
 195 -                                __u32 *start_hash);
 196 +                                __u32 *start_hash, struct htree_lock *lck);
 197  static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
 198                 struct ext4_filename *fname,
 199 -               struct ext4_dir_entry_2 **res_dir);
 200 +               struct ext4_dir_entry_2 **res_dir, struct htree_lock *lck);
 201  static int ext4_dx_add_entry(handle_t *handle, struct ext4_filename *fname,
 202 -                            struct inode *dir, struct inode *inode);
 203 +                            struct inode *dir, struct inode *inode,
 204 +                            struct htree_lock *lck);
 205
 206  /* checksumming functions */
 207  void initialize_dirent_tail(struct ext4_dir_entry_tail *t,
 208 @@ -732,6 +742,227 @@ struct stats dx_show_entries(struct dx_h
 209  }
 210  #endif /* DX_DEBUG */
 211
 212 +/* private data for htree_lock */
 213 +struct ext4_dir_lock_data {
 214 +       unsigned                ld_flags;  /* bits-map for lock types */
 215 +       unsigned                ld_count;  /* # entries of the last DX block */
 216 +       struct dx_entry         ld_at_entry; /* copy of leaf dx_entry */
 217 +       struct dx_entry         *ld_at;    /* position of leaf dx_entry */
 218 +};
 219 +
 220 +#define ext4_htree_lock_data(l)        ((struct ext4_dir_lock_data *)(l)->lk_private)
 221 +#define ext4_find_entry(dir, name, dirent, inline) \
 222 +                       __ext4_find_entry(dir, name, dirent, inline, NULL)
 223 +#define ext4_add_entry(handle, dentry, inode) \
 224 +                       __ext4_add_entry(handle, dentry, inode, NULL)
 225 +
 226 +/* NB: ext4_lblk_t is 32 bits so we use high bits to identify invalid blk */
 227 +#define EXT4_HTREE_NODE_CHANGED        (0xcafeULL << 32)
 228 +
 229 +static void ext4_htree_event_cb(void *target, void *event)
 230 +{
 231 +       u64 *block = (u64 *)target;
 232 +
 233 +       if (*block == dx_get_block((struct dx_entry *)event))
 234 +               *block = EXT4_HTREE_NODE_CHANGED;
 235 +}
 236 +
 237 +struct htree_lock_head *ext4_htree_lock_head_alloc(unsigned hbits)
 238 +{
 239 +       struct htree_lock_head *lhead;
 240 +
 241 +       lhead = htree_lock_head_alloc(EXT4_LK_MAX, hbits, 0);
 242 +       if (lhead != NULL) {
 243 +               htree_lock_event_attach(lhead, EXT4_LK_SPIN, HTREE_EVENT_WR,
 244 +                                       ext4_htree_event_cb);
 245 +       }
 246 +       return lhead;
 247 +}
 248 +EXPORT_SYMBOL(ext4_htree_lock_head_alloc);
 249 +
 250 +struct htree_lock *ext4_htree_lock_alloc(void)
 251 +{
 252 +       return htree_lock_alloc(EXT4_LK_MAX,
 253 +                               sizeof(struct ext4_dir_lock_data));
 254 +}
 255 +EXPORT_SYMBOL(ext4_htree_lock_alloc);
 256 +
 257 +static htree_lock_mode_t ext4_htree_mode(unsigned flags)
 258 +{
 259 +       switch (flags) {
 260 +       default: /* 0 or unknown flags require EX lock */
 261 +               return HTREE_LOCK_EX;
 262 +       case EXT4_HLOCK_READDIR:
 263 +               return HTREE_LOCK_PR;
 264 +       case EXT4_HLOCK_LOOKUP:
 265 +               return HTREE_LOCK_CR;
 266 +       case EXT4_HLOCK_DEL:
 267 +       case EXT4_HLOCK_ADD:
 268 +               return HTREE_LOCK_CW;
 269 +       }
 270 +}
 271 +
 272 +/* return PR for read-only operations, otherwise return EX */
 273 +static inline htree_lock_mode_t ext4_htree_safe_mode(unsigned flags)
 274 +{
 275 +       int writer = (flags & EXT4_LB_DE) == EXT4_LB_DE;
 276 +
 277 +       /* 0 requires EX lock */
 278 +       return (flags == 0 || writer) ? HTREE_LOCK_EX : HTREE_LOCK_PR;
 279 +}
 280 +
 281 +static int ext4_htree_safe_locked(struct htree_lock *lck)
 282 +{
 283 +       int writer;
 284 +
 285 +       if (lck == NULL || lck->lk_mode == HTREE_LOCK_EX)
 286 +               return 1;
 287 +
 288 +       writer = (ext4_htree_lock_data(lck)->ld_flags & EXT4_LB_DE) ==
 289 +                EXT4_LB_DE;
 290 +       if (writer) /* all readers & writers are excluded? */
 291 +               return lck->lk_mode == HTREE_LOCK_EX;
 292 +
 293 +       /* all writers are excluded? */
 294 +       return lck->lk_mode == HTREE_LOCK_PR ||
 295 +              lck->lk_mode == HTREE_LOCK_PW ||
 296 +              lck->lk_mode == HTREE_LOCK_EX;
 297 +}
 298 +
 299 +/* relock htree_lock with EX mode if it's change operation, otherwise
 300 + * relock it with PR mode. It's noop if PDO is disabled. */
 301 +static void ext4_htree_safe_relock(struct htree_lock *lck)
 302 +{
 303 +       if (!ext4_htree_safe_locked(lck)) {
 304 +               unsigned flags = ext4_htree_lock_data(lck)->ld_flags;
 305 +
 306 +               htree_change_lock(lck, ext4_htree_safe_mode(flags));
 307 +       }
 308 +}
 309 +
 310 +void ext4_htree_lock(struct htree_lock *lck, struct htree_lock_head *lhead,
 311 +                    struct inode *dir, unsigned flags)
 312 +{
 313 +       htree_lock_mode_t mode = is_dx(dir) ? ext4_htree_mode(flags) :
 314 +                                             ext4_htree_safe_mode(flags);
 315 +
 316 +       ext4_htree_lock_data(lck)->ld_flags = flags;
 317 +       htree_lock(lck, lhead, mode);
 318 +       if (!is_dx(dir))
 319 +               ext4_htree_safe_relock(lck); /* make sure it's safe locked */
 320 +}
 321 +EXPORT_SYMBOL(ext4_htree_lock);
 322 +
 323 +static int ext4_htree_node_lock(struct htree_lock *lck, struct dx_entry *at,
 324 +                               unsigned lmask, int wait, void *ev)
 325 +{
 326 +       u32     key = (at == NULL) ? 0 : dx_get_block(at);
 327 +       u32     mode;
 328 +
 329 +       /* NOOP if htree is well protected or caller doesn't require the lock */
 330 +       if (ext4_htree_safe_locked(lck) ||
 331 +          !(ext4_htree_lock_data(lck)->ld_flags & lmask))
 332 +               return 1;
 333 +
 334 +       mode = (ext4_htree_lock_data(lck)->ld_flags & lmask) == lmask ?
 335 +               HTREE_LOCK_PW : HTREE_LOCK_PR;
 336 +       while (1) {
 337 +               if (htree_node_lock_try(lck, mode, key, ffz(~lmask), wait, ev))
 338 +                       return 1;
 339 +               if (!(lmask & EXT4_LB_SPIN)) /* not a spinlock */
 340 +                       return 0;
 341 +               cpu_relax(); /* spin until granted */
 342 +       }
 343 +}
 344 +
 345 +static int ext4_htree_node_locked(struct htree_lock *lck, unsigned lmask)
 346 +{
 347 +       return ext4_htree_safe_locked(lck) ||
 348 +              htree_node_is_granted(lck, ffz(~lmask));
 349 +}
 350 +
 351 +static void ext4_htree_node_unlock(struct htree_lock *lck,
 352 +                                  unsigned lmask, void *buf)
 353 +{
 354 +       /* NB: it's safe to call mutiple times or even it's not locked */
 355 +       if (!ext4_htree_safe_locked(lck) &&
 356 +            htree_node_is_granted(lck, ffz(~lmask)))
 357 +               htree_node_unlock(lck, ffz(~lmask), buf);
 358 +}
 359 +
 360 +#define ext4_htree_dx_lock(lck, key)           \
 361 +       ext4_htree_node_lock(lck, key, EXT4_LB_DX, 1, NULL)
 362 +#define ext4_htree_dx_lock_try(lck, key)       \
 363 +       ext4_htree_node_lock(lck, key, EXT4_LB_DX, 0, NULL)
 364 +#define ext4_htree_dx_unlock(lck)              \
 365 +       ext4_htree_node_unlock(lck, EXT4_LB_DX, NULL)
 366 +#define ext4_htree_dx_locked(lck)              \
 367 +       ext4_htree_node_locked(lck, EXT4_LB_DX)
 368 +
 369 +static void ext4_htree_dx_need_lock(struct htree_lock *lck)
 370 +{
 371 +       struct ext4_dir_lock_data *ld;
 372 +
 373 +       if (ext4_htree_safe_locked(lck))
 374 +               return;
 375 +
 376 +       ld = ext4_htree_lock_data(lck);
 377 +       switch (ld->ld_flags) {
 378 +       default:
 379 +               return;
 380 +       case EXT4_HLOCK_LOOKUP:
 381 +               ld->ld_flags = EXT4_HLOCK_LOOKUP_SAFE;
 382 +               return;
 383 +       case EXT4_HLOCK_DEL:
 384 +               ld->ld_flags = EXT4_HLOCK_DEL_SAFE;
 385 +               return;
 386 +       case EXT4_HLOCK_ADD:
 387 +               ld->ld_flags = EXT4_HLOCK_SPLIT;
 388 +               return;
 389 +       }
 390 +}
 391 +
 392 +#define ext4_htree_de_lock(lck, key)           \
 393 +       ext4_htree_node_lock(lck, key, EXT4_LB_DE, 1, NULL)
 394 +#define ext4_htree_de_unlock(lck)              \
 395 +       ext4_htree_node_unlock(lck, EXT4_LB_DE, NULL)
 396 +
 397 +#define ext4_htree_spin_lock(lck, key, event)  \
 398 +       ext4_htree_node_lock(lck, key, EXT4_LB_SPIN, 0, event)
 399 +#define ext4_htree_spin_unlock(lck)            \
 400 +       ext4_htree_node_unlock(lck, EXT4_LB_SPIN, NULL)
 401 +#define ext4_htree_spin_unlock_listen(lck, p)  \
 402 +       ext4_htree_node_unlock(lck, EXT4_LB_SPIN, p)
 403 +
 404 +static void ext4_htree_spin_stop_listen(struct htree_lock *lck)
 405 +{
 406 +       if (!ext4_htree_safe_locked(lck) &&
 407 +           htree_node_is_listening(lck, ffz(~EXT4_LB_SPIN)))
 408 +               htree_node_stop_listen(lck, ffz(~EXT4_LB_SPIN));
 409 +}
 410 +
 411 +enum {
 412 +       DX_HASH_COL_IGNORE,     /* ignore collision while probing frames */
 413 +       DX_HASH_COL_YES,        /* there is collision and it does matter */
 414 +       DX_HASH_COL_NO,         /* there is no collision */
 415 +};
 416 +
 417 +static int dx_probe_hash_collision(struct htree_lock *lck,
 418 +                                  struct dx_entry *entries,
 419 +                                  struct dx_entry *at, u32 hash)
 420 +{
 421 +       if (!(lck && ext4_htree_lock_data(lck)->ld_flags & EXT4_LB_EXACT)) {
 422 +               return DX_HASH_COL_IGNORE; /* don't care about collision */
 423 +
 424 +       } else if (at == entries + dx_get_count(entries) - 1) {
 425 +               return DX_HASH_COL_IGNORE; /* not in any leaf of this DX */
 426 +
 427 +       } else { /* hash collision? */
 428 +               return ((dx_get_hash(at + 1) & ~1) == hash) ?
 429 +                       DX_HASH_COL_YES : DX_HASH_COL_NO;
 430 +       }
 431 +}
 432 +
 433  /*
 434   * Probe for a directory leaf block to search.
 435   *
 436 @@ -743,10 +974,11 @@ struct stats dx_show_entries(struct dx_h
 437   */
 438  static struct dx_frame *
 439  dx_probe(struct ext4_filename *fname, struct inode *dir,
 440 -        struct dx_hash_info *hinfo, struct dx_frame *frame_in)
 441 +        struct dx_hash_info *hinfo, struct dx_frame *frame_in,
 442 +        struct htree_lock *lck)
 443  {
 444         unsigned count, indirect;
 445 -       struct dx_entry *at, *entries, *p, *q, *m;
 446 +       struct dx_entry *at, *entries, *p, *q, *m, *dx = NULL;
 447         struct dx_root_info *info;
 448         struct dx_frame *frame = frame_in;
 449         struct dx_frame *ret_err = ERR_PTR(ERR_BAD_DX_DIR);
 450 @@ -809,8 +1041,15 @@ dx_probe(struct ext4_filename *fname, st
 451
 452         dxtrace(printk("Look up %x", hash));
 453         while (1) {
 454 +               if (indirect == 0) { /* the last index level */
 455 +                       /* NB: ext4_htree_dx_lock() could be noop if
 456 +                        * DX-lock flag is not set for current operation */
 457 +                       ext4_htree_dx_lock(lck, dx);
 458 +                       ext4_htree_spin_lock(lck, dx, NULL);
 459 +               }
 460                 count = dx_get_count(entries);
 461 -               if (!count || count > dx_get_limit(entries)) {
 462 +               if (count == 0 || count > dx_get_limit(entries)) {
 463 +                       ext4_htree_spin_unlock(lck); /* release spin */
 464                         ext4_warning_inode(dir,
 465                                            "dx entry: count %u beyond limit %u",
 466                                            count, dx_get_limit(entries));
 467 @@ -849,8 +1088,70 @@ dx_probe(struct ext4_filename *fname, st
 468                                dx_get_block(at)));
 469                 frame->entries = entries;
 470                 frame->at = at;
 471 -               if (!indirect--)
 472 +
 473 +               if (indirect == 0) { /* the last index level */
 474 +                       struct ext4_dir_lock_data *ld;
 475 +                       u64 myblock;
 476 +
 477 +                       /* By default we only lock DE-block, however, we will
 478 +                        * also lock the last level DX-block if:
 479 +                        * a) there is hash collision
 480 +                        *    we will set DX-lock flag (a few lines below)
 481 +                        *    and redo to lock DX-block
 482 +                        *    see detail in dx_probe_hash_collision()
 483 +                        * b) it's a retry from splitting
 484 +                        *    we need to lock the last level DX-block so nobody
 485 +                        *    else can split any leaf blocks under the same
 486 +                        *    DX-block, see detail in ext4_dx_add_entry()
 487 +                        */
 488 +                       if (ext4_htree_dx_locked(lck)) {
 489 +                               /* DX-block is locked, just lock DE-block
 490 +                                * and return */
 491 +                               ext4_htree_spin_unlock(lck);
 492 +                               if (!ext4_htree_safe_locked(lck))
 493 +                                       ext4_htree_de_lock(lck, frame->at);
 494 +                               return frame;
 495 +                       }
 496 +                       /* it's pdirop and no DX lock */
 497 +                       if (dx_probe_hash_collision(lck, entries, at, hash) ==
 498 +                           DX_HASH_COL_YES) {
 499 +                               /* found hash collision, set DX-lock flag
 500 +                                * and retry to abtain DX-lock */
 501 +                               ext4_htree_spin_unlock(lck);
 502 +                               ext4_htree_dx_need_lock(lck);
 503 +                               continue;
 504 +                       }
 505 +                       ld = ext4_htree_lock_data(lck);
 506 +                       /* because I don't lock DX, so @at can't be trusted
 507 +                        * after I release spinlock so I have to save it */
 508 +                       ld->ld_at = at;
 509 +                       ld->ld_at_entry = *at;
 510 +                       ld->ld_count = dx_get_count(entries);
 511 +
 512 +                       frame->at = &ld->ld_at_entry;
 513 +                       myblock = dx_get_block(at);
 514 +
 515 +                       /* NB: ordering locking */
 516 +                       ext4_htree_spin_unlock_listen(lck, &myblock);
 517 +                       /* other thread can split this DE-block because:
 518 +                        * a) I don't have lock for the DE-block yet
 519 +                        * b) I released spinlock on DX-block
 520 +                        * if it happened I can detect it by listening
 521 +                        * splitting event on this DE-block */
 522 +                       ext4_htree_de_lock(lck, frame->at);
 523 +                       ext4_htree_spin_stop_listen(lck);
 524 +
 525 +                       if (myblock == EXT4_HTREE_NODE_CHANGED) {
 526 +                               /* someone split this DE-block before
 527 +                                * I locked it, I need to retry and lock
 528 +                                * valid DE-block */
 529 +                               ext4_htree_de_unlock(lck);
 530 +                               continue;
 531 +                       }
 532                         return frame;
 533 +               }
 534 +               dx = at;
 535 +               indirect--;
 536                 frame++;
 537                 frame->bh = ext4_read_dirblock(dir, dx_get_block(at), INDEX);
 538                 if (IS_ERR(frame->bh)) {
 539 @@ -917,7 +1218,7 @@ static void dx_release(struct dx_frame *
 540  static int ext4_htree_next_block(struct inode *dir, __u32 hash,
 541                                  struct dx_frame *frame,
 542                                  struct dx_frame *frames,
 543 -                                __u32 *start_hash)
 544 +                                __u32 *start_hash, struct htree_lock *lck)
 545  {
 546         struct dx_frame *p;
 547         struct buffer_head *bh;
 548 @@ -932,12 +1233,22 @@ static int ext4_htree_next_block(struct
 549          * this loop, num_frames indicates the number of interior
 550          * nodes need to be read.
 551          */
 552 +       ext4_htree_de_unlock(lck);
 553         while (1) {
 554 -               if (++(p->at) < p->entries + dx_get_count(p->entries))
 555 -                       break;
 556 +               if (num_frames > 0 || ext4_htree_dx_locked(lck)) {
 557 +                       /* num_frames > 0 :
 558 +                        *   DX block
 559 +                        * ext4_htree_dx_locked:
 560 +                        *   frame->at is reliable pointer returned by dx_probe,
 561 +                        *   otherwise dx_probe already knew no collision */
 562 +                       if (++(p->at) < p->entries + dx_get_count(p->entries))
 563 +                               break;
 564 +               }
 565                 if (p == frames)
 566                         return 0;
 567                 num_frames++;
 568 +               if (num_frames == 1)
 569 +                       ext4_htree_dx_unlock(lck);
 570                 p--;
 571         }
 572
 573 @@ -960,6 +1271,13 @@ static int ext4_htree_next_block(struct
 574          * block so no check is necessary
 575          */
 576         while (num_frames--) {
 577 +               if (num_frames == 0) {
 578 +                       /* it's not always necessary, we just don't want to
 579 +                        * detect hash collision again */
 580 +                       ext4_htree_dx_need_lock(lck);
 581 +                       ext4_htree_dx_lock(lck, p->at);
 582 +               }
 583 +
 584                 bh = ext4_read_dirblock(dir, dx_get_block(p->at), INDEX);
 585                 if (IS_ERR(bh))
 586                         return PTR_ERR(bh);
 587 @@ -968,6 +1286,7 @@ static int ext4_htree_next_block(struct
 588                 p->bh = bh;
 589                 p->at = p->entries = ((struct dx_node *) bh->b_data)->entries;
 590         }
 591 +       ext4_htree_de_lock(lck, p->at);
 592         return 1;
 593  }
 594
 595 @@ -1115,10 +1434,10 @@ int ext4_htree_fill_tree(struct file *di
 596         }
 597         hinfo.hash = start_hash;
 598         hinfo.minor_hash = 0;
 599 -       frame = dx_probe(NULL, dir, &hinfo, frames);
 600 +       /* assume it's PR locked */
 601 +       frame = dx_probe(NULL, dir, &hinfo, frames, NULL);
 602         if (IS_ERR(frame))
 603                 return PTR_ERR(frame);
 604 -
 605         /* Add '.' and '..' from the htree header */
 606         if (!start_hash && !start_minor_hash) {
 607                 de = (struct ext4_dir_entry_2 *) frames[0].bh->b_data;
 608 @@ -1158,7 +1477,7 @@ int ext4_htree_fill_tree(struct file *di
 609                 count += ret;
 610                 hashval = ~0;
 611                 ret = ext4_htree_next_block(dir, HASH_NB_ALWAYS,
 612 -                                           frame, frames, &hashval);
 613 +                                           frame, frames, &hashval, NULL);
 614                 *next_hash = hashval;
 615                 if (ret < 0) {
 616                         err = ret;
 617 @@ -1350,10 +1669,10 @@ static int is_dx_internal_node(struct in
 618   * The returned buffer_head has ->b_count elevated.  The caller is expected
 619   * to brelse() it when appropriate.
 620   */
 621 -static struct buffer_head * ext4_find_entry (struct inode *dir,
 622 +struct buffer_head *__ext4_find_entry(struct inode *dir,
 623                                         const struct qstr *d_name,
 624                                         struct ext4_dir_entry_2 **res_dir,
 625 -                                       int *inlined)
 626 +                                       int *inlined, struct htree_lock *lck)
 627  {
 628         struct super_block *sb;
 629         struct buffer_head *bh_use[NAMEI_RA_SIZE];
 630 @@ -1403,7 +1722,7 @@ static struct buffer_head * ext4_find_en
 631                 goto restart;
 632         }
 633         if (is_dx(dir)) {
 634 -               ret = ext4_dx_find_entry(dir, &fname, res_dir);
 635 +               ret = ext4_dx_find_entry(dir, &fname, res_dir, lck);
 636                 /*
 637                  * On success, or if the error was file not found,
 638                  * return.  Otherwise, fall back to doing a search the
 639 @@ -1413,6 +1732,7 @@ static struct buffer_head * ext4_find_en
 640                         goto cleanup_and_exit;
 641                 dxtrace(printk(KERN_DEBUG "ext4_find_entry: dx failed, "
 642                                "falling back\n"));
 643 +               ext4_htree_safe_relock(lck);
 644                 ret = NULL;
 645         }
 646         nblocks = dir->i_size >> EXT4_BLOCK_SIZE_BITS(sb);
 647 @@ -1514,10 +1834,12 @@ cleanup_and_exit:
 648         ext4_fname_free_filename(&fname);
 649         return ret;
 650  }
 651 +EXPORT_SYMBOL(__ext4_find_entry);
 652
 653  static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
 654                         struct ext4_filename *fname,
 655 -                       struct ext4_dir_entry_2 **res_dir)
 656 +                       struct ext4_dir_entry_2 **res_dir,
 657 +                       struct htree_lock *lck)
 658  {
 659         struct super_block * sb = dir->i_sb;
 660         struct dx_frame frames[EXT4_HTREE_LEVEL], *frame;
 661 @@ -1528,7 +1850,7 @@ static struct buffer_head * ext4_dx_find
 662  #ifdef CONFIG_EXT4_FS_ENCRYPTION
 663         *res_dir = NULL;
 664  #endif
 665 -       frame = dx_probe(fname, dir, NULL, frames);
 666 +       frame = dx_probe(fname, dir, NULL, frames, lck);
 667         if (IS_ERR(frame))
 668                 return (struct buffer_head *) frame;
 669         do {
 670 @@ -1550,7 +1872,7 @@ static struct buffer_head * ext4_dx_find
 671
 672                 /* Check to see if we should continue to search */
 673                 retval = ext4_htree_next_block(dir, fname->hinfo.hash, frame,
 674 -                                              frames, NULL);
 675 +                                              frames, NULL, lck);
 676                 if (retval < 0) {
 677                         ext4_warning_inode(dir,
 678                                 "error %d reading directory index block",
 679 @@ -1735,8 +2057,9 @@ static struct ext4_dir_entry_2* dx_pack_
 680   * Returns pointer to de in block into which the new entry will be inserted.
 681   */
 682  static struct ext4_dir_entry_2 *do_split(handle_t *handle, struct inode *dir,
 683 -                       struct buffer_head **bh,struct dx_frame *frame,
 684 -                       struct dx_hash_info *hinfo)
 685 +                       struct buffer_head **bh, struct dx_frame *frames,
 686 +                       struct dx_frame *frame, struct dx_hash_info *hinfo,
 687 +                       struct htree_lock *lck)
 688  {
 689         unsigned blocksize = dir->i_sb->s_blocksize;
 690         unsigned count, continued;
 691 @@ -1798,8 +2121,14 @@ static struct ext4_dir_entry_2 *do_split
 692                                         hash2, split, count-split));
 693
 694         /* Fancy dance to stay within two buffers */
 695 -       de2 = dx_move_dirents(data1, data2, map + split, count - split,
 696 -                             blocksize);
 697 +       if (hinfo->hash < hash2) {
 698 +               de2 = dx_move_dirents(data1, data2, map + split,
 699 +                                     count - split, blocksize);
 700 +       } else {
 701 +               /* make sure we will add entry to the same block which
 702 +                * we have already locked */
 703 +               de2 = dx_move_dirents(data1, data2, map, split, blocksize);
 704 +       }
 705         de = dx_pack_dirents(data1, blocksize);
 706         de->rec_len = ext4_rec_len_to_disk(data1 + (blocksize - csum_size) -
 707                                            (char *) de,
 708 @@ -1820,12 +2149,21 @@ static struct ext4_dir_entry_2 *do_split
 709         dxtrace(dx_show_leaf(dir, hinfo, (struct ext4_dir_entry_2 *) data2,
 710                         blocksize, 1));
 711
 712 -       /* Which block gets the new entry? */
 713 -       if (hinfo->hash >= hash2) {
 714 -               swap(*bh, bh2);
 715 -               de = de2;
 716 +       ext4_htree_spin_lock(lck, frame > frames ? (frame - 1)->at : NULL,
 717 +                            frame->at); /* notify block is being split */
 718 +       if (hinfo->hash < hash2) {
 719 +               dx_insert_block(frame, hash2 + continued, newblock);
 720 +
 721 +       } else {
 722 +               /* switch block number */
 723 +               dx_insert_block(frame, hash2 + continued,
 724 +                               dx_get_block(frame->at));
 725 +               dx_set_block(frame->at, newblock);
 726 +               (frame->at)++;
 727         }
 728 -       dx_insert_block(frame, hash2 + continued, newblock);
 729 +       ext4_htree_spin_unlock(lck);
 730 +       ext4_htree_dx_unlock(lck);
 731 +
 732         err = ext4_handle_dirty_dirent_node(handle, dir, bh2);
 733         if (err)
 734                 goto journal_error;
 735 @@ -2099,7 +2437,7 @@ static int make_indexed_dir(handle_t *ha
 736         if (retval)
 737                 goto out_frames;
 738
 739 -       de = do_split(handle,dir, &bh2, frame, &fname->hinfo);
 740 +       de = do_split(handle, dir, &bh2, frames, frame, &fname->hinfo, NULL);
 741         if (IS_ERR(de)) {
 742                 retval = PTR_ERR(de);
 743                 goto out_frames;
 744 @@ -2209,8 +2547,8 @@ out:
 745   * may not sleep between calling this and putting something into
 746   * the entry, as someone else might have used it while you slept.
 747   */
 748 -static int ext4_add_entry(handle_t *handle, struct dentry *dentry,
 749 -                         struct inode *inode)
 750 +int __ext4_add_entry(handle_t *handle, struct dentry *dentry,
 751 +                     struct inode *inode, struct htree_lock *lck)
 752  {
 753         struct inode *dir = d_inode(dentry->d_parent);
 754         struct buffer_head *bh = NULL;
 755 @@ -2251,7 +2589,7 @@ static int ext4_add_entry(handle_t *hand
 756                 if (dentry->d_name.len == 2 &&
 757                     memcmp(dentry->d_name.name, "..", 2) == 0)
 758                         return ext4_update_dotdot(handle, dentry, inode);
 759 -               retval = ext4_dx_add_entry(handle, &fname, dir, inode);
 760 +               retval = ext4_dx_add_entry(handle, &fname, dir, inode, lck);
 761                 if (!retval || (retval != ERR_BAD_DX_DIR))
 762                         goto out;
 763                 /* Can we just ignore htree data? */
 764 @@ -2261,6 +2599,7 @@ static int ext4_add_entry(handle_t *hand
 765                         retval = -EFSCORRUPTED;
 766                         goto out;
 767                 }
 768 +               ext4_htree_safe_relock(lck);
 769                 ext4_clear_inode_flag(dir, EXT4_INODE_INDEX);
 770                 dx_fallback++;
 771                 ext4_mark_inode_dirty(handle, dir);
 772 @@ -2310,12 +2649,14 @@ out:
 773                 ext4_set_inode_state(inode, EXT4_STATE_NEWENTRY);
 774         return retval;
 775  }
 776 +EXPORT_SYMBOL(__ext4_add_entry);
 777
 778  /*
 779   * Returns 0 for success, or a negative error value
 780   */
 781  static int ext4_dx_add_entry(handle_t *handle, struct ext4_filename *fname,
 782 -                            struct inode *dir, struct inode *inode)
 783 +                            struct inode *dir, struct inode *inode,
 784 +                            struct htree_lock *lck)
 785  {
 786         struct dx_frame frames[EXT4_HTREE_LEVEL], *frame;
 787         struct dx_entry *entries, *at;
 788 @@ -2327,7 +2668,7 @@ static int ext4_dx_add_entry(handle_t *h
 789
 790  again:
 791         restart = 0;
 792 -       frame = dx_probe(fname, dir, NULL, frames);
 793 +       frame = dx_probe(fname, dir, NULL, frames, lck);
 794         if (IS_ERR(frame))
 795                 return PTR_ERR(frame);
 796         entries = frame->entries;
 797 @@ -2362,6 +2703,11 @@ again:
 798                 struct dx_node *node2;
 799                 struct buffer_head *bh2;
 800
 801 +               if (!ext4_htree_safe_locked(lck)) { /* retry with EX lock */
 802 +                       ext4_htree_safe_relock(lck);
 803 +                       restart = 1;
 804 +                       goto cleanup;
 805 +               }
 806                 while (frame > frames) {
 807                         if (dx_get_count((frame - 1)->entries) <
 808                             dx_get_limit((frame - 1)->entries)) {
 809 @@ -2465,8 +2811,32 @@ again:
 810                         restart = 1;
 811                         goto journal_error;
 812                 }
 813 +       } else if (!ext4_htree_dx_locked(lck)) {
 814 +               struct ext4_dir_lock_data *ld = ext4_htree_lock_data(lck);
 815 +
 816 +               /* not well protected, require DX lock */
 817 +               ext4_htree_dx_need_lock(lck);
 818 +               at = frame > frames ? (frame - 1)->at : NULL;
 819 +
 820 +               /* NB: no risk of deadlock because it's just a try.
 821 +                *
 822 +                * NB: we check ld_count for twice, the first time before
 823 +                * having DX lock, the second time after holding DX lock.
 824 +                *
 825 +                * NB: We never free blocks for directory so far, which
 826 +                * means value returned by dx_get_count() should equal to
 827 +                * ld->ld_count if nobody split any DE-block under @at,
 828 +                * and ld->ld_at still points to valid dx_entry. */
 829 +               if ((ld->ld_count != dx_get_count(entries)) ||
 830 +                   !ext4_htree_dx_lock_try(lck, at) ||
 831 +                   (ld->ld_count != dx_get_count(entries))) {
 832 +                       restart = 1;
 833 +                       goto cleanup;
 834 +               }
 835 +               /* OK, I've got DX lock and nothing changed */
 836 +               frame->at = ld->ld_at;
 837         }
 838 -       de = do_split(handle, dir, &bh, frame, &fname->hinfo);
 839 +       de = do_split(handle, dir, &bh, frames, frame, &fname->hinfo, lck);
 840         if (IS_ERR(de)) {
 841                 err = PTR_ERR(de);
 842                 goto cleanup;
 843 @@ -2477,6 +2847,8 @@ again:
 844  journal_error:
 845         ext4_std_error(dir->i_sb, err); /* this is a no-op if err == 0 */
 846  cleanup:
 847 +       ext4_htree_dx_unlock(lck);
 848 +       ext4_htree_de_unlock(lck);
 849         brelse(bh);
 850         dx_release(frames);
 851         /* @restart is true means htree-path has been changed, we need to
 852 --- a/fs/ext4/super.c
 853 +++ b/fs/ext4/super.c
 854 @@ -989,6 +989,7 @@ static struct inode *ext4_alloc_inode(st
 855
 856         inode_set_iversion(&ei->vfs_inode, 1);
 857         spin_lock_init(&ei->i_raw_lock);
 858 +       sema_init(&ei->i_append_sem, 1);
 859         INIT_LIST_HEAD(&ei->i_prealloc_list);
 860         spin_lock_init(&ei->i_prealloc_lock);
 861         ext4_es_init_tree(&ei->i_es_tree);