ldiskfs/kernel_patches/patches/rhel7.6/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 Index: linux-3.10.0-229.1.2.fc21.x86_64/fs/ext4/Makefile
  18 ===================================================================
  19 --- linux-3.10.0-229.1.2.fc21.x86_64.orig/fs/ext4/Makefile
  20 +++ linux-3.10.0-229.1.2.fc21.x86_64/fs/ext4/Makefile
  21 @@ -6,6 +6,7 @@ obj-$(CONFIG_EXT4_FS) += ext4.o
  22
  23  ext4-y := balloc.o bitmap.o dir.o file.o fsync.o ialloc.o inode.o page-io.o \
  24                 ioctl.o namei.o super.o symlink.o hash.o resize.o extents.o \
  25 +               htree_lock.o \
  26                 ext4_jbd2.o migrate.o mballoc.o block_validity.o move_extent.o \
  27                 mmp.o indirect.o extents_status.o xattr.o xattr_user.o \
  28                 xattr_trusted.o inline.o
  29 Index: linux-3.10.0-229.1.2.fc21.x86_64/fs/ext4/ext4.h
  30 ===================================================================
  31 --- linux-3.10.0-229.1.2.fc21.x86_64.orig/fs/ext4/ext4.h
  32 +++ linux-3.10.0-229.1.2.fc21.x86_64/fs/ext4/ext4.h
  33 @@ -27,6 +27,7 @@
  34  #include <linux/mutex.h>
  35  #include <linux/timer.h>
  36  #include <linux/wait.h>
  37 +#include <linux/htree_lock.h>
  38  #include <linux/blockgroup_lock.h>
  39  #include <linux/percpu_counter.h>
  40  #include <linux/ratelimit.h>
  41 @@ -821,6 +822,9 @@ struct ext4_inode_info {
  42         __u32   i_dtime;
  43         ext4_fsblk_t    i_file_acl;
  44
  45 +       /* following fields for parallel directory operations -bzzz */
  46 +       struct semaphore i_append_sem;
  47 +
  48         /*
  49          * i_block_group is the number of the block group which contains
  50          * this file's inode.  Constant across the lifetime of the inode,
  51 @@ -1846,6 +1850,71 @@ struct dx_hash_info
  52   */
  53  #define HASH_NB_ALWAYS         1
  54
  55 +/* assume name-hash is protected by upper layer */
  56 +#define EXT4_HTREE_LOCK_HASH   0
  57 +
  58 +enum ext4_pdo_lk_types {
  59 +#if EXT4_HTREE_LOCK_HASH
  60 +       EXT4_LK_HASH,
  61 +#endif
  62 +       EXT4_LK_DX,             /* index block */
  63 +       EXT4_LK_DE,             /* directory entry block */
  64 +       EXT4_LK_SPIN,           /* spinlock */
  65 +       EXT4_LK_MAX,
  66 +};
  67 +
  68 +/* read-only bit */
  69 +#define EXT4_LB_RO(b)          (1 << (b))
  70 +/* read + write, high bits for writer */
  71 +#define EXT4_LB_RW(b)          ((1 << (b)) | (1 << (EXT4_LK_MAX + (b))))
  72 +
  73 +enum ext4_pdo_lock_bits {
  74 +       /* DX lock bits */
  75 +       EXT4_LB_DX_RO           = EXT4_LB_RO(EXT4_LK_DX),
  76 +       EXT4_LB_DX              = EXT4_LB_RW(EXT4_LK_DX),
  77 +       /* DE lock bits */
  78 +       EXT4_LB_DE_RO           = EXT4_LB_RO(EXT4_LK_DE),
  79 +       EXT4_LB_DE              = EXT4_LB_RW(EXT4_LK_DE),
  80 +       /* DX spinlock bits */
  81 +       EXT4_LB_SPIN_RO         = EXT4_LB_RO(EXT4_LK_SPIN),
  82 +       EXT4_LB_SPIN            = EXT4_LB_RW(EXT4_LK_SPIN),
  83 +       /* accurate searching */
  84 +       EXT4_LB_EXACT           = EXT4_LB_RO(EXT4_LK_MAX << 1),
  85 +};
  86 +
  87 +enum ext4_pdo_lock_opc {
  88 +       /* external */
  89 +       EXT4_HLOCK_READDIR      = (EXT4_LB_DE_RO | EXT4_LB_DX_RO),
  90 +       EXT4_HLOCK_LOOKUP       = (EXT4_LB_DE_RO | EXT4_LB_SPIN_RO |
  91 +                                  EXT4_LB_EXACT),
  92 +       EXT4_HLOCK_DEL          = (EXT4_LB_DE | EXT4_LB_SPIN_RO |
  93 +                                  EXT4_LB_EXACT),
  94 +       EXT4_HLOCK_ADD          = (EXT4_LB_DE | EXT4_LB_SPIN_RO),
  95 +
  96 +       /* internal */
  97 +       EXT4_HLOCK_LOOKUP_SAFE  = (EXT4_LB_DE_RO | EXT4_LB_DX_RO |
  98 +                                  EXT4_LB_EXACT),
  99 +       EXT4_HLOCK_DEL_SAFE     = (EXT4_LB_DE | EXT4_LB_DX_RO | EXT4_LB_EXACT),
 100 +       EXT4_HLOCK_SPLIT        = (EXT4_LB_DE | EXT4_LB_DX | EXT4_LB_SPIN),
 101 +};
 102 +
 103 +extern struct htree_lock_head *ext4_htree_lock_head_alloc(unsigned hbits);
 104 +#define ext4_htree_lock_head_free(lhead)       htree_lock_head_free(lhead)
 105 +
 106 +extern struct htree_lock *ext4_htree_lock_alloc(void);
 107 +#define ext4_htree_lock_free(lck)              htree_lock_free(lck)
 108 +
 109 +extern void ext4_htree_lock(struct htree_lock *lck,
 110 +                           struct htree_lock_head *lhead,
 111 +                           struct inode *dir, unsigned flags);
 112 +#define ext4_htree_unlock(lck)                  htree_unlock(lck)
 113 +
 114 +extern struct buffer_head *__ext4_find_entry(struct inode *dir,
 115 +                                       const struct qstr *d_name,
 116 +                                       struct ext4_dir_entry_2 **res_dir,
 117 +                                       int *inlined, struct htree_lock *lck);
 118 +extern int __ext4_add_entry(handle_t *handle, struct dentry *dentry,
 119 +                     struct inode *inode, struct htree_lock *lck);
 120
 121  /*
 122   * Describe an inode's exact location on disk and in memory
 123 @@ -2088,9 +2157,17 @@ void ext4_insert_dentry(struct inode *in
 124                         const char *name, int namelen, void *data);
 125  static inline void ext4_update_dx_flag(struct inode *inode)
 126  {
 127 +       /* Disable it for ldiskfs, because going from a DX directory to
 128 +        * a non-DX directory while it is in use will completely break
 129 +        * the htree-locking.
 130 +        * If we really want to support this operation in the future,
 131 +        * we need to exclusively lock the directory at here which will
 132 +        * increase complexity of code */
 133 +#if 0
 134         if (!EXT4_HAS_COMPAT_FEATURE(inode->i_sb,
 135                                      EXT4_FEATURE_COMPAT_DIR_INDEX))
 136                 ext4_clear_inode_flag(inode, EXT4_INODE_INDEX);
 137 +#endif
 138  }
 139  static unsigned char ext4_filetype_table[] = {
 140         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
 141 Index: linux-3.10.0-229.1.2.fc21.x86_64/fs/ext4/namei.c
 142 ===================================================================
 143 --- linux-3.10.0-229.1.2.fc21.x86_64.orig/fs/ext4/namei.c
 144 +++ linux-3.10.0-229.1.2.fc21.x86_64/fs/ext4/namei.c
 145 @@ -53,6 +53,7 @@ struct buffer_head *ext4_append(handle_t
 146                                         ext4_lblk_t *block)
 147  {
 148         struct buffer_head *bh;
 149 +       struct ext4_inode_info *ei = EXT4_I(inode);
 150         int err = 0;
 151
 152         if (unlikely(EXT4_SB(inode->i_sb)->s_max_dir_size_kb &&
 153 @@ -60,15 +61,22 @@ struct buffer_head *ext4_append(handle_t
 154                       EXT4_SB(inode->i_sb)->s_max_dir_size_kb)))
 155                 return ERR_PTR(-ENOSPC);
 156
 157 +       /* with parallel dir operations all appends
 158 +       * have to be serialized -bzzz */
 159 +       down(&ei->i_append_sem);
 160 +
 161         *block = inode->i_size >> inode->i_sb->s_blocksize_bits;
 162
 163         bh = ext4_bread(handle, inode, *block, 1, &err);
 164 -       if (!bh)
 165 +       if (!bh) {
 166 +               up(&ei->i_append_sem);
 167                 return ERR_PTR(err);
 168 +       }
 169         inode->i_size += inode->i_sb->s_blocksize;
 170         EXT4_I(inode)->i_disksize = inode->i_size;
 171         BUFFER_TRACE(bh, "get_write_access");
 172         err = ext4_journal_get_write_access(handle, bh);
 173 +       up(&ei->i_append_sem);
 174         if (err) {
 175                 brelse(bh);
 176                 ext4_std_error(inode->i_sb, err);
 177 @@ -246,7 +254,7 @@ static struct dx_frame *dx_probe(const s
 178                                  struct inode *dir,
 179                                  struct dx_hash_info *hinfo,
 180                                  struct dx_frame *frame,
 181 -                                int *err);
 182 +                                struct htree_lock *lck, int *err);
 183  static void dx_release(struct dx_frame *frames);
 184  static int dx_make_map(struct ext4_dir_entry_2 *de, unsigned blocksize,
 185                        struct dx_hash_info *hinfo, struct dx_map_entry map[]);
 186 @@ -259,13 +267,13 @@ static void dx_insert_block(struct dx_fr
 187  static int ext4_htree_next_block(struct inode *dir, __u32 hash,
 188                                  struct dx_frame *frame,
 189                                  struct dx_frame *frames,
 190 -                                __u32 *start_hash);
 191 +                                __u32 *start_hash, struct htree_lock *lck);
 192  static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
 193                 const struct qstr *d_name,
 194                 struct ext4_dir_entry_2 **res_dir,
 195 -               int *err);
 196 +               struct htree_lock *lck, int *err);
 197  static int ext4_dx_add_entry(handle_t *handle, struct dentry *dentry,
 198 -                            struct inode *inode);
 199 +                            struct inode *inode, struct htree_lock *lck);
 200
 201  /* checksumming functions */
 202  void initialize_dirent_tail(struct ext4_dir_entry_tail *t,
 203 @@ -668,6 +676,227 @@ struct stats dx_show_entries(struct dx_h
 204  }
 205  #endif /* DX_DEBUG */
 206
 207 +/* private data for htree_lock */
 208 +struct ext4_dir_lock_data {
 209 +       unsigned                ld_flags;  /* bits-map for lock types */
 210 +       unsigned                ld_count;  /* # entries of the last DX block */
 211 +       struct dx_entry         ld_at_entry; /* copy of leaf dx_entry */
 212 +       struct dx_entry         *ld_at;    /* position of leaf dx_entry */
 213 +};
 214 +
 215 +#define ext4_htree_lock_data(l)        ((struct ext4_dir_lock_data *)(l)->lk_private)
 216 +#define ext4_find_entry(dir, name, dirent, inline) \
 217 +                       __ext4_find_entry(dir, name, dirent, inline, NULL)
 218 +#define ext4_add_entry(handle, dentry, inode) \
 219 +                       __ext4_add_entry(handle, dentry, inode, NULL)
 220 +
 221 +/* NB: ext4_lblk_t is 32 bits so we use high bits to identify invalid blk */
 222 +#define EXT4_HTREE_NODE_CHANGED        (0xcafeULL << 32)
 223 +
 224 +static void ext4_htree_event_cb(void *target, void *event)
 225 +{
 226 +       u64 *block = (u64 *)target;
 227 +
 228 +       if (*block == dx_get_block((struct dx_entry *)event))
 229 +               *block = EXT4_HTREE_NODE_CHANGED;
 230 +}
 231 +
 232 +struct htree_lock_head *ext4_htree_lock_head_alloc(unsigned hbits)
 233 +{
 234 +       struct htree_lock_head *lhead;
 235 +
 236 +       lhead = htree_lock_head_alloc(EXT4_LK_MAX, hbits, 0);
 237 +       if (lhead != NULL) {
 238 +               htree_lock_event_attach(lhead, EXT4_LK_SPIN, HTREE_EVENT_WR,
 239 +                                       ext4_htree_event_cb);
 240 +       }
 241 +       return lhead;
 242 +}
 243 +EXPORT_SYMBOL(ext4_htree_lock_head_alloc);
 244 +
 245 +struct htree_lock *ext4_htree_lock_alloc(void)
 246 +{
 247 +       return htree_lock_alloc(EXT4_LK_MAX,
 248 +                               sizeof(struct ext4_dir_lock_data));
 249 +}
 250 +EXPORT_SYMBOL(ext4_htree_lock_alloc);
 251 +
 252 +static htree_lock_mode_t ext4_htree_mode(unsigned flags)
 253 +{
 254 +       switch (flags) {
 255 +       default: /* 0 or unknown flags require EX lock */
 256 +               return HTREE_LOCK_EX;
 257 +       case EXT4_HLOCK_READDIR:
 258 +               return HTREE_LOCK_PR;
 259 +       case EXT4_HLOCK_LOOKUP:
 260 +               return HTREE_LOCK_CR;
 261 +       case EXT4_HLOCK_DEL:
 262 +       case EXT4_HLOCK_ADD:
 263 +               return HTREE_LOCK_CW;
 264 +       }
 265 +}
 266 +
 267 +/* return PR for read-only operations, otherwise return EX */
 268 +static inline htree_lock_mode_t ext4_htree_safe_mode(unsigned flags)
 269 +{
 270 +       int writer = (flags & EXT4_LB_DE) == EXT4_LB_DE;
 271 +
 272 +       /* 0 requires EX lock */
 273 +       return (flags == 0 || writer) ? HTREE_LOCK_EX : HTREE_LOCK_PR;
 274 +}
 275 +
 276 +static int ext4_htree_safe_locked(struct htree_lock *lck)
 277 +{
 278 +       int writer;
 279 +
 280 +       if (lck == NULL || lck->lk_mode == HTREE_LOCK_EX)
 281 +               return 1;
 282 +
 283 +       writer = (ext4_htree_lock_data(lck)->ld_flags & EXT4_LB_DE) ==
 284 +                EXT4_LB_DE;
 285 +       if (writer) /* all readers & writers are excluded? */
 286 +               return lck->lk_mode == HTREE_LOCK_EX;
 287 +
 288 +       /* all writers are excluded? */
 289 +       return lck->lk_mode == HTREE_LOCK_PR ||
 290 +              lck->lk_mode == HTREE_LOCK_PW ||
 291 +              lck->lk_mode == HTREE_LOCK_EX;
 292 +}
 293 +
 294 +/* relock htree_lock with EX mode if it's change operation, otherwise
 295 + * relock it with PR mode. It's noop if PDO is disabled. */
 296 +static void ext4_htree_safe_relock(struct htree_lock *lck)
 297 +{
 298 +       if (!ext4_htree_safe_locked(lck)) {
 299 +               unsigned flags = ext4_htree_lock_data(lck)->ld_flags;
 300 +
 301 +               htree_change_lock(lck, ext4_htree_safe_mode(flags));
 302 +       }
 303 +}
 304 +
 305 +void ext4_htree_lock(struct htree_lock *lck, struct htree_lock_head *lhead,
 306 +                    struct inode *dir, unsigned flags)
 307 +{
 308 +       htree_lock_mode_t mode = is_dx(dir) ? ext4_htree_mode(flags) :
 309 +                                             ext4_htree_safe_mode(flags);
 310 +
 311 +       ext4_htree_lock_data(lck)->ld_flags = flags;
 312 +       htree_lock(lck, lhead, mode);
 313 +       if (!is_dx(dir))
 314 +               ext4_htree_safe_relock(lck); /* make sure it's safe locked */
 315 +}
 316 +EXPORT_SYMBOL(ext4_htree_lock);
 317 +
 318 +static int ext4_htree_node_lock(struct htree_lock *lck, struct dx_entry *at,
 319 +                               unsigned lmask, int wait, void *ev)
 320 +{
 321 +       u32     key = (at == NULL) ? 0 : dx_get_block(at);
 322 +       u32     mode;
 323 +
 324 +       /* NOOP if htree is well protected or caller doesn't require the lock */
 325 +       if (ext4_htree_safe_locked(lck) ||
 326 +          !(ext4_htree_lock_data(lck)->ld_flags & lmask))
 327 +               return 1;
 328 +
 329 +       mode = (ext4_htree_lock_data(lck)->ld_flags & lmask) == lmask ?
 330 +               HTREE_LOCK_PW : HTREE_LOCK_PR;
 331 +       while (1) {
 332 +               if (htree_node_lock_try(lck, mode, key, ffz(~lmask), wait, ev))
 333 +                       return 1;
 334 +               if (!(lmask & EXT4_LB_SPIN)) /* not a spinlock */
 335 +                       return 0;
 336 +               cpu_relax(); /* spin until granted */
 337 +       }
 338 +}
 339 +
 340 +static int ext4_htree_node_locked(struct htree_lock *lck, unsigned lmask)
 341 +{
 342 +       return ext4_htree_safe_locked(lck) ||
 343 +              htree_node_is_granted(lck, ffz(~lmask));
 344 +}
 345 +
 346 +static void ext4_htree_node_unlock(struct htree_lock *lck,
 347 +                                  unsigned lmask, void *buf)
 348 +{
 349 +       /* NB: it's safe to call mutiple times or even it's not locked */
 350 +       if (!ext4_htree_safe_locked(lck) &&
 351 +            htree_node_is_granted(lck, ffz(~lmask)))
 352 +               htree_node_unlock(lck, ffz(~lmask), buf);
 353 +}
 354 +
 355 +#define ext4_htree_dx_lock(lck, key)           \
 356 +       ext4_htree_node_lock(lck, key, EXT4_LB_DX, 1, NULL)
 357 +#define ext4_htree_dx_lock_try(lck, key)       \
 358 +       ext4_htree_node_lock(lck, key, EXT4_LB_DX, 0, NULL)
 359 +#define ext4_htree_dx_unlock(lck)              \
 360 +       ext4_htree_node_unlock(lck, EXT4_LB_DX, NULL)
 361 +#define ext4_htree_dx_locked(lck)              \
 362 +       ext4_htree_node_locked(lck, EXT4_LB_DX)
 363 +
 364 +static void ext4_htree_dx_need_lock(struct htree_lock *lck)
 365 +{
 366 +       struct ext4_dir_lock_data *ld;
 367 +
 368 +       if (ext4_htree_safe_locked(lck))
 369 +               return;
 370 +
 371 +       ld = ext4_htree_lock_data(lck);
 372 +       switch (ld->ld_flags) {
 373 +       default:
 374 +               return;
 375 +       case EXT4_HLOCK_LOOKUP:
 376 +               ld->ld_flags = EXT4_HLOCK_LOOKUP_SAFE;
 377 +               return;
 378 +       case EXT4_HLOCK_DEL:
 379 +               ld->ld_flags = EXT4_HLOCK_DEL_SAFE;
 380 +               return;
 381 +       case EXT4_HLOCK_ADD:
 382 +               ld->ld_flags = EXT4_HLOCK_SPLIT;
 383 +               return;
 384 +       }
 385 +}
 386 +
 387 +#define ext4_htree_de_lock(lck, key)           \
 388 +       ext4_htree_node_lock(lck, key, EXT4_LB_DE, 1, NULL)
 389 +#define ext4_htree_de_unlock(lck)              \
 390 +       ext4_htree_node_unlock(lck, EXT4_LB_DE, NULL)
 391 +
 392 +#define ext4_htree_spin_lock(lck, key, event)  \
 393 +       ext4_htree_node_lock(lck, key, EXT4_LB_SPIN, 0, event)
 394 +#define ext4_htree_spin_unlock(lck)            \
 395 +       ext4_htree_node_unlock(lck, EXT4_LB_SPIN, NULL)
 396 +#define ext4_htree_spin_unlock_listen(lck, p)  \
 397 +       ext4_htree_node_unlock(lck, EXT4_LB_SPIN, p)
 398 +
 399 +static void ext4_htree_spin_stop_listen(struct htree_lock *lck)
 400 +{
 401 +       if (!ext4_htree_safe_locked(lck) &&
 402 +           htree_node_is_listening(lck, ffz(~EXT4_LB_SPIN)))
 403 +               htree_node_stop_listen(lck, ffz(~EXT4_LB_SPIN));
 404 +}
 405 +
 406 +enum {
 407 +       DX_HASH_COL_IGNORE,     /* ignore collision while probing frames */
 408 +       DX_HASH_COL_YES,        /* there is collision and it does matter */
 409 +       DX_HASH_COL_NO,         /* there is no collision */
 410 +};
 411 +
 412 +static int dx_probe_hash_collision(struct htree_lock *lck,
 413 +                                  struct dx_entry *entries,
 414 +                                  struct dx_entry *at, u32 hash)
 415 +{
 416 +       if (!(ext4_htree_lock_data(lck)->ld_flags & EXT4_LB_EXACT)) {
 417 +               return DX_HASH_COL_IGNORE; /* don't care about collision */
 418 +
 419 +       } else if (at == entries + dx_get_count(entries) - 1) {
 420 +               return DX_HASH_COL_IGNORE; /* not in any leaf of this DX */
 421 +
 422 +       } else { /* hash collision? */
 423 +               return ((dx_get_hash(at + 1) & ~1) == hash) ?
 424 +                       DX_HASH_COL_YES : DX_HASH_COL_NO;
 425 +       }
 426 +}
 427 +
 428  /*
 429   * Probe for a directory leaf block to search.
 430   *
 431 @@ -679,10 +908,11 @@ struct stats dx_show_entries(struct dx_h
 432   */
 433  static struct dx_frame *
 434  dx_probe(const struct qstr *d_name, struct inode *dir,
 435 -        struct dx_hash_info *hinfo, struct dx_frame *frame_in, int *err)
 436 +        struct dx_hash_info *hinfo, struct dx_frame *frame_in,
 437 +        struct htree_lock *lck, int *err)
 438  {
 439         unsigned count, indirect;
 440 -       struct dx_entry *at, *entries, *p, *q, *m;
 441 +       struct dx_entry *at, *entries, *p, *q, *m, *dx = NULL;
 442         struct dx_root_info *info;
 443         struct buffer_head *bh;
 444         struct dx_frame *frame = frame_in;
 445 @@ -750,8 +980,15 @@ dx_probe(const struct qstr *d_name, stru
 446         dxtrace(printk("Look up %x", hash));
 447         while (1)
 448         {
 449 +               if (indirect == 0) { /* the last index level */
 450 +                       /* NB: ext4_htree_dx_lock() could be noop if
 451 +                        * DX-lock flag is not set for current operation */
 452 +                       ext4_htree_dx_lock(lck, dx);
 453 +                       ext4_htree_spin_lock(lck, dx, NULL);
 454 +               }
 455                 count = dx_get_count(entries);
 456 -               if (!count || count > dx_get_limit(entries)) {
 457 +               if (count == 0 || count > dx_get_limit(entries)) {
 458 +                       ext4_htree_spin_unlock(lck); /* release spin */
 459                         ext4_warning(dir->i_sb,
 460                                      "dx entry: no count or count > limit");
 461                         brelse(bh);
 462 @@ -792,7 +1029,70 @@ dx_probe(const struct qstr *d_name, stru
 463                 frame->bh = bh;
 464                 frame->entries = entries;
 465                 frame->at = at;
 466 -               if (!indirect--) return frame;
 467 +
 468 +               if (indirect == 0) { /* the last index level */
 469 +                       struct ext4_dir_lock_data *ld;
 470 +                       u64 myblock;
 471 +
 472 +                       /* By default we only lock DE-block, however, we will
 473 +                        * also lock the last level DX-block if:
 474 +                        * a) there is hash collision
 475 +                        *    we will set DX-lock flag (a few lines below)
 476 +                        *    and redo to lock DX-block
 477 +                        *    see detail in dx_probe_hash_collision()
 478 +                        * b) it's a retry from splitting
 479 +                        *    we need to lock the last level DX-block so nobody
 480 +                        *    else can split any leaf blocks under the same
 481 +                        *    DX-block, see detail in ext4_dx_add_entry()
 482 +                        */
 483 +                       if (ext4_htree_dx_locked(lck)) {
 484 +                               /* DX-block is locked, just lock DE-block
 485 +                                * and return */
 486 +                               ext4_htree_spin_unlock(lck);
 487 +                               if (!ext4_htree_safe_locked(lck))
 488 +                                       ext4_htree_de_lock(lck, frame->at);
 489 +                               return frame;
 490 +                       }
 491 +                       /* it's pdirop and no DX lock */
 492 +                       if (dx_probe_hash_collision(lck, entries, at, hash) ==
 493 +                           DX_HASH_COL_YES) {
 494 +                               /* found hash collision, set DX-lock flag
 495 +                                * and retry to abtain DX-lock */
 496 +                               ext4_htree_spin_unlock(lck);
 497 +                               ext4_htree_dx_need_lock(lck);
 498 +                               continue;
 499 +                       }
 500 +                       ld = ext4_htree_lock_data(lck);
 501 +                       /* because I don't lock DX, so @at can't be trusted
 502 +                        * after I release spinlock so I have to save it */
 503 +                       ld->ld_at = at;
 504 +                       ld->ld_at_entry = *at;
 505 +                       ld->ld_count = dx_get_count(entries);
 506 +
 507 +                       frame->at = &ld->ld_at_entry;
 508 +                       myblock = dx_get_block(at);
 509 +
 510 +                       /* NB: ordering locking */
 511 +                       ext4_htree_spin_unlock_listen(lck, &myblock);
 512 +                       /* other thread can split this DE-block because:
 513 +                        * a) I don't have lock for the DE-block yet
 514 +                        * b) I released spinlock on DX-block
 515 +                        * if it happened I can detect it by listening
 516 +                        * splitting event on this DE-block */
 517 +                       ext4_htree_de_lock(lck, frame->at);
 518 +                       ext4_htree_spin_stop_listen(lck);
 519 +
 520 +                       if (myblock == EXT4_HTREE_NODE_CHANGED) {
 521 +                               /* someone split this DE-block before
 522 +                                * I locked it, I need to retry and lock
 523 +                                * valid DE-block */
 524 +                               ext4_htree_de_unlock(lck);
 525 +                               continue;
 526 +                       }
 527 +                       return frame;
 528 +               }
 529 +               dx = at;
 530 +               indirect--;
 531                 bh = ext4_read_dirblock(dir, dx_get_block(at), INDEX);
 532                 if (IS_ERR(bh)) {
 533                         *err = PTR_ERR(bh);
 534 @@ -860,7 +1160,7 @@ static void dx_release (struct dx_frame
 535  static int ext4_htree_next_block(struct inode *dir, __u32 hash,
 536                                  struct dx_frame *frame,
 537                                  struct dx_frame *frames,
 538 -                                __u32 *start_hash)
 539 +                                __u32 *start_hash, struct htree_lock *lck)
 540  {
 541         struct dx_frame *p;
 542         struct buffer_head *bh;
 543 @@ -875,12 +1175,22 @@ static int ext4_htree_next_block(struct
 544          * this loop, num_frames indicates the number of interior
 545          * nodes need to be read.
 546          */
 547 +       ext4_htree_de_unlock(lck);
 548         while (1) {
 549 -               if (++(p->at) < p->entries + dx_get_count(p->entries))
 550 -                       break;
 551 +               if (num_frames > 0 || ext4_htree_dx_locked(lck)) {
 552 +                       /* num_frames > 0 :
 553 +                        *   DX block
 554 +                        * ext4_htree_dx_locked:
 555 +                        *   frame->at is reliable pointer returned by dx_probe,
 556 +                        *   otherwise dx_probe already knew no collision */
 557 +                       if (++(p->at) < p->entries + dx_get_count(p->entries))
 558 +                               break;
 559 +               }
 560                 if (p == frames)
 561                         return 0;
 562                 num_frames++;
 563 +               if (num_frames == 1)
 564 +                       ext4_htree_dx_unlock(lck);
 565                 p--;
 566         }
 567
 568 @@ -903,6 +1213,13 @@ static int ext4_htree_next_block(struct
 569          * block so no check is necessary
 570          */
 571         while (num_frames--) {
 572 +               if (num_frames == 0) {
 573 +                       /* it's not always necessary, we just don't want to
 574 +                        * detect hash collision again */
 575 +                       ext4_htree_dx_need_lock(lck);
 576 +                       ext4_htree_dx_lock(lck, p->at);
 577 +               }
 578 +
 579                 bh = ext4_read_dirblock(dir, dx_get_block(p->at), INDEX);
 580                 if (IS_ERR(bh))
 581                         return PTR_ERR(bh);
 582 @@ -911,6 +1228,7 @@ static int ext4_htree_next_block(struct
 583                 p->bh = bh;
 584                 p->at = p->entries = ((struct dx_node *) bh->b_data)->entries;
 585         }
 586 +       ext4_htree_de_lock(lck, p->at);
 587         return 1;
 588  }
 589
 590 @@ -1013,10 +1331,10 @@ int ext4_htree_fill_tree(struct file *di
 591         }
 592         hinfo.hash = start_hash;
 593         hinfo.minor_hash = 0;
 594 -       frame = dx_probe(NULL, dir, &hinfo, frames, &err);
 595 +       /* assume it's PR locked */
 596 +       frame = dx_probe(NULL, dir, &hinfo, frames, NULL, &err);
 597         if (!frame)
 598                 return err;
 599 -
 600         /* Add '.' and '..' from the htree header */
 601         if (!start_hash && !start_minor_hash) {
 602                 de = (struct ext4_dir_entry_2 *) frames[0].bh->b_data;
 603 @@ -1043,7 +1361,7 @@ int ext4_htree_fill_tree(struct file *di
 604                 count += ret;
 605                 hashval = ~0;
 606                 ret = ext4_htree_next_block(dir, HASH_NB_ALWAYS,
 607 -                                           frame, frames, &hashval);
 608 +                                           frame, frames, &hashval, NULL);
 609                 *next_hash = hashval;
 610                 if (ret < 0) {
 611                         err = ret;
 612 @@ -1236,10 +1554,10 @@ static int is_dx_internal_node(struct in
 613   * The returned buffer_head has ->b_count elevated.  The caller is expected
 614   * to brelse() it when appropriate.
 615   */
 616 -static struct buffer_head * ext4_find_entry (struct inode *dir,
 617 +struct buffer_head *__ext4_find_entry(struct inode *dir,
 618                                         const struct qstr *d_name,
 619                                         struct ext4_dir_entry_2 **res_dir,
 620 -                                       int *inlined)
 621 +                                       int *inlined, struct htree_lock *lck)
 622  {
 623         struct super_block *sb;
 624         struct buffer_head *bh_use[NAMEI_RA_SIZE];
 625 @@ -1283,7 +1601,7 @@ static struct buffer_head * ext4_find_en
 626                 goto restart;
 627         }
 628         if (is_dx(dir)) {
 629 -               bh = ext4_dx_find_entry(dir, d_name, res_dir, &err);
 630 +               bh = ext4_dx_find_entry(dir, d_name, res_dir, lck, &err);
 631                 /*
 632                  * On success, or if the error was file not found,
 633                  * return.  Otherwise, fall back to doing a search the
 634 @@ -1297,6 +1615,7 @@ static struct buffer_head * ext4_find_en
 635                         return bh;
 636                 dxtrace(printk(KERN_DEBUG "ext4_find_entry: dx failed, "
 637                                "falling back\n"));
 638 +               ext4_htree_safe_relock(lck);
 639         }
 640         nblocks = dir->i_size >> EXT4_BLOCK_SIZE_BITS(sb);
 641         start = EXT4_I(dir)->i_dir_start_lookup;
 642 @@ -1389,9 +1708,12 @@ cleanup_and_exit:
 643                 brelse(bh_use[ra_ptr]);
 644         return ret;
 645  }
 646 +EXPORT_SYMBOL(__ext4_find_entry);
 647
 648 -static struct buffer_head * ext4_dx_find_entry(struct inode *dir, const struct qstr *d_name,
 649 -                      struct ext4_dir_entry_2 **res_dir, int *err)
 650 +static struct buffer_head *ext4_dx_find_entry(struct inode *dir,
 651 +                               const struct qstr *d_name,
 652 +                               struct ext4_dir_entry_2 **res_dir,
 653 +                               struct htree_lock *lck, int *err)
 654  {
 655         struct super_block * sb = dir->i_sb;
 656         struct dx_hash_info     hinfo;
 657 @@ -1400,7 +1722,7 @@ static struct buffer_head * ext4_dx_find
 658         ext4_lblk_t block;
 659         int retval;
 660
 661 -       if (!(frame = dx_probe(d_name, dir, &hinfo, frames, err)))
 662 +       if (!(frame = dx_probe(d_name, dir, &hinfo, frames, lck, err)))
 663                 return NULL;
 664         do {
 665                 block = dx_get_block(frame->at);
 666 @@ -1424,7 +1746,7 @@ static struct buffer_head * ext4_dx_find
 667
 668                 /* Check to see if we should continue to search */
 669                 retval = ext4_htree_next_block(dir, hinfo.hash, frame,
 670 -                                              frames, NULL);
 671 +                                              frames, NULL, lck);
 672                 if (retval < 0) {
 673                         ext4_warning(sb,
 674                              "error reading index page in directory #%lu",
 675 @@ -1583,8 +1905,9 @@ static struct ext4_dir_entry_2* dx_pack_
 676   * Returns pointer to de in block into which the new entry will be inserted.
 677   */
 678  static struct ext4_dir_entry_2 *do_split(handle_t *handle, struct inode *dir,
 679 -                       struct buffer_head **bh,struct dx_frame *frame,
 680 -                       struct dx_hash_info *hinfo, int *error)
 681 +                       struct buffer_head **bh, struct dx_frame *frames,
 682 +                       struct dx_frame *frame, struct dx_hash_info *hinfo,
 683 +                       struct htree_lock *lck, int *error)
 684  {
 685         unsigned blocksize = dir->i_sb->s_blocksize;
 686         unsigned count, continued;
 687 @@ -1647,7 +1970,14 @@ static struct ext4_dir_entry_2 *do_split
 688                                         hash2, split, count-split));
 689
 690         /* Fancy dance to stay within two buffers */
 691 -       de2 = dx_move_dirents(data1, data2, map + split, count - split, blocksize);
 692 +       if (hinfo->hash < hash2) {
 693 +               de2 = dx_move_dirents(data1, data2, map + split,
 694 +                                     count - split, blocksize);
 695 +       } else {
 696 +               /* make sure we will add entry to the same block which
 697 +                * we have already locked */
 698 +               de2 = dx_move_dirents(data1, data2, map, split, blocksize);
 699 +       }
 700         de = dx_pack_dirents(data1, blocksize);
 701         de->rec_len = ext4_rec_len_to_disk(data1 + (blocksize - csum_size) -
 702                                            (char *) de,
 703 @@ -1666,13 +1996,21 @@ static struct ext4_dir_entry_2 *do_split
 704         dxtrace(dx_show_leaf (hinfo, (struct ext4_dir_entry_2 *) data1, blocksize, 1));
 705         dxtrace(dx_show_leaf (hinfo, (struct ext4_dir_entry_2 *) data2, blocksize, 1));
 706
 707 -       /* Which block gets the new entry? */
 708 -       if (hinfo->hash >= hash2)
 709 -       {
 710 -               swap(*bh, bh2);
 711 -               de = de2;
 712 +       ext4_htree_spin_lock(lck, frame > frames ? (frame - 1)->at : NULL,
 713 +                            frame->at); /* notify block is being split */
 714 +       if (hinfo->hash < hash2) {
 715 +               dx_insert_block(frame, hash2 + continued, newblock);
 716 +
 717 +       } else {
 718 +               /* switch block number */
 719 +               dx_insert_block(frame, hash2 + continued,
 720 +                               dx_get_block(frame->at));
 721 +               dx_set_block(frame->at, newblock);
 722 +               (frame->at)++;
 723         }
 724 -       dx_insert_block(frame, hash2 + continued, newblock);
 725 +       ext4_htree_spin_unlock(lck);
 726 +       ext4_htree_dx_unlock(lck);
 727 +
 728         err = ext4_handle_dirty_dirent_node(handle, dir, bh2);
 729         if (err)
 730                 goto journal_error;
 731 @@ -1945,7 +2283,7 @@ static int make_indexed_dir(handle_t *ha
 732         if (retval)
 733                 goto out_frames;
 734
 735 -       de = do_split(handle,dir, &bh2, frame, &hinfo, &retval);
 736 +       de = do_split(handle, dir, &bh2, frames, frame, &hinfo, NULL, &retval);
 737         if (!de) {
 738                 goto out_frames;
 739         }
 740 @@ -2051,8 +2389,8 @@ out:
 741   * may not sleep between calling this and putting something into
 742   * the entry, as someone else might have used it while you slept.
 743   */
 744 -static int ext4_add_entry(handle_t *handle, struct dentry *dentry,
 745 -                         struct inode *inode)
 746 +int __ext4_add_entry(handle_t *handle, struct dentry *dentry,
 747 +                     struct inode *inode, struct htree_lock *lck)
 748  {
 749         struct inode *dir = dentry->d_parent->d_inode;
 750         struct buffer_head *bh = NULL;
 751 @@ -2087,9 +2425,10 @@ static int ext4_add_entry(handle_t *hand
 752                 if (dentry->d_name.len == 2 &&
 753                     memcmp(dentry->d_name.name, "..", 2) == 0)
 754                         return ext4_update_dotdot(handle, dentry, inode);
 755 -               retval = ext4_dx_add_entry(handle, dentry, inode);
 756 +               retval = ext4_dx_add_entry(handle, dentry, inode, lck);
 757                 if (!retval || (retval != ERR_BAD_DX_DIR))
 758                         goto out;
 759 +               ext4_htree_safe_relock(lck);
 760                 ext4_clear_inode_flag(dir, EXT4_INODE_INDEX);
 761                 dx_fallback++;
 762                 ext4_mark_inode_dirty(handle, dir);
 763 @@ -2129,12 +2468,13 @@ static int ext4_add_entry(handle_t *hand
 764                 ext4_set_inode_state(inode, EXT4_STATE_NEWENTRY);
 765         return retval;
 766  }
 767 +EXPORT_SYMBOL(__ext4_add_entry);
 768
 769  /*
 770   * Returns 0 for success, or a negative error value
 771   */
 772  static int ext4_dx_add_entry(handle_t *handle, struct dentry *dentry,
 773 -                            struct inode *inode)
 774 +                            struct inode *inode, struct htree_lock *lck)
 775  {
 776         struct dx_frame frames[EXT4_HTREE_LEVEL], *frame;
 777         struct dx_entry *entries, *at;
 778 @@ -2148,7 +2488,7 @@ static int ext4_dx_add_entry(handle_t *h
 779
 780  again:
 781         restart = 0;
 782 -       frame = dx_probe(&dentry->d_name, dir, &hinfo, frames, &err);
 783 +       frame = dx_probe(&dentry->d_name, dir, &hinfo, frames, lck, &err);
 784         if (!frame)
 785                 return err;
 786         entries = frame->entries;
 787 @@ -2178,6 +2518,11 @@ again:
 788                 struct dx_node *node2;
 789                 struct buffer_head *bh2;
 790
 791 +               if (!ext4_htree_safe_locked(lck)) { /* retry with EX lock */
 792 +                       ext4_htree_safe_relock(lck);
 793 +                       restart = 1;
 794 +                       goto cleanup;
 795 +               }
 796                 while (frame > frames) {
 797                         if (dx_get_count((frame - 1)->entries) <
 798                             dx_get_limit((frame - 1)->entries)) {
 799 @@ -2277,16 +2622,43 @@ again:
 800                         restart = 1;
 801                         goto journal_error;
 802                 }
 803 +       } else if (!ext4_htree_dx_locked(lck)) {
 804 +               struct ext4_dir_lock_data *ld = ext4_htree_lock_data(lck);
 805 +
 806 +               /* not well protected, require DX lock */
 807 +               ext4_htree_dx_need_lock(lck);
 808 +               at = frame > frames ? (frame - 1)->at : NULL;
 809 +
 810 +               /* NB: no risk of deadlock because it's just a try.
 811 +                *
 812 +                * NB: we check ld_count for twice, the first time before
 813 +                * having DX lock, the second time after holding DX lock.
 814 +                *
 815 +                * NB: We never free blocks for directory so far, which
 816 +                * means value returned by dx_get_count() should equal to
 817 +                * ld->ld_count if nobody split any DE-block under @at,
 818 +                * and ld->ld_at still points to valid dx_entry. */
 819 +               if ((ld->ld_count != dx_get_count(entries)) ||
 820 +                   !ext4_htree_dx_lock_try(lck, at) ||
 821 +                   (ld->ld_count != dx_get_count(entries))) {
 822 +                       restart = 1;
 823 +                       goto cleanup;
 824 +               }
 825 +               /* OK, I've got DX lock and nothing changed */
 826 +               frame->at = ld->ld_at;
 827         }
 828 -       de = do_split(handle, dir, &bh, frame, &hinfo, &err);
 829 +       de = do_split(handle, dir, &bh, frames, frame, &hinfo, lck, &err);
 830         if (!de)
 831                 goto cleanup;
 832 +
 833         err = add_dirent_to_buf(handle, dentry, inode, de, bh);
 834         goto cleanup;
 835
 836  journal_error:
 837         ext4_std_error(dir->i_sb, err); /* this is a no-op if err == 0 */
 838  cleanup:
 839 +       ext4_htree_dx_unlock(lck);
 840 +       ext4_htree_de_unlock(lck);
 841         brelse(bh);
 842         dx_release(frames);
 843         /* @restart is true means htree-path has been changed, we need to
 844 Index: linux-3.10.0-229.1.2.fc21.x86_64/fs/ext4/super.c
 845 ===================================================================
 846 --- linux-3.10.0-229.1.2.fc21.x86_64.orig/fs/ext4/super.c
 847 +++ linux-3.10.0-229.1.2.fc21.x86_64/fs/ext4/super.c
 848 @@ -875,6 +875,7 @@ static struct inode *ext4_alloc_inode(st
 849
 850         ei->vfs_inode.i_version = 1;
 851         spin_lock_init(&ei->i_raw_lock);
 852 +       sema_init(&ei->i_append_sem, 1);
 853         INIT_LIST_HEAD(&ei->i_prealloc_list);
 854         spin_lock_init(&ei->i_prealloc_lock);
 855         ext4_es_init_tree(&ei->i_es_tree);