ldiskfs/kernel_patches/patches/linux-5.8/ext4-pdirop.patch

   1 From 1a0f7f0b9c13ef0aa86e125f350b6733bff8db3c Mon Sep 17 00:00:00 2001
   2 From: Shaun Tancheff <stancheff@cray.com>
   3 Date: Wed, 15 Jan 2020 07:35:13 -0600
   4 Subject: [PATCH] Single directory performance is a critical for HPC workloads.
   5  In a typical use case an application creates a separate output file for each
   6  node and task in a job. As nodes and tasks increase, hundreds of thousands of
   7  files may be created in a single directory within a short window of time.
   8  Today, both filename lookup and file system modifying operations (such as
   9  create and unlink) are protected with a single lock for an entire ldiskfs
  10  directory. PDO project will remove this bottleneck by introducing a parallel
  11  locking mechanism for entire ldiskfs directories. This work will enable
  12  multiple application threads to simultaneously lookup, create and unlink in
  13  parallel.
  14
  15 This patch contains:
  16  - pdirops support for ldiskfs
  17  - integrate with osd-ldiskfs
  18 ---
  19  fs/ext4/Makefile |    1
  20  fs/ext4/ext4.h   |   78 +++++++++
  21  fs/ext4/namei.c  |  454 ++++++++++++++++++++++++++++++++++++++++++++++++++-----
  22  fs/ext4/super.c  |    1
  23  4 files changed, 494 insertions(+), 40 deletions(-)
  24  create mode 100644 fs/ext4/htree_lock.c
  25  create mode 100644 include/linux/htree_lock.h
  26
  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 --- a/fs/ext4/ext4.h
  38 +++ b/fs/ext4/ext4.h
  39 @@ -29,6 +29,7 @@
  40  #include <linux/timer.h>
  41  #include <linux/version.h>
  42  #include <linux/wait.h>
  43 +#include <linux/htree_lock.h>
  44  #include <linux/sched/signal.h>
  45  #include <linux/blockgroup_lock.h>
  46  #include <linux/percpu_counter.h>
  47 @@ -987,6 +988,9 @@ struct ext4_inode_info {
  48         __u32   i_dtime;
  49         ext4_fsblk_t    i_file_acl;
  50
  51 +       /* following fields for parallel directory operations -bzzz */
  52 +       struct semaphore i_append_sem;
  53 +
  54         /*
  55          * i_block_group is the number of the block group which contains
  56          * this file's inode.  Constant across the lifetime of the inode,
  57 @@ -2299,6 +2303,72 @@ struct dx_hash_info
  58   */
  59  #define HASH_NB_ALWAYS         1
  60
  61 +/* assume name-hash is protected by upper layer */
  62 +#define EXT4_HTREE_LOCK_HASH   0
  63 +
  64 +enum ext4_pdo_lk_types {
  65 +#if EXT4_HTREE_LOCK_HASH
  66 +       EXT4_LK_HASH,
  67 +#endif
  68 +       EXT4_LK_DX,             /* index block */
  69 +       EXT4_LK_DE,             /* directory entry block */
  70 +       EXT4_LK_SPIN,           /* spinlock */
  71 +       EXT4_LK_MAX,
  72 +};
  73 +
  74 +/* read-only bit */
  75 +#define EXT4_LB_RO(b)          (1 << (b))
  76 +/* read + write, high bits for writer */
  77 +#define EXT4_LB_RW(b)          ((1 << (b)) | (1 << (EXT4_LK_MAX + (b))))
  78 +
  79 +enum ext4_pdo_lock_bits {
  80 +       /* DX lock bits */
  81 +       EXT4_LB_DX_RO           = EXT4_LB_RO(EXT4_LK_DX),
  82 +       EXT4_LB_DX              = EXT4_LB_RW(EXT4_LK_DX),
  83 +       /* DE lock bits */
  84 +       EXT4_LB_DE_RO           = EXT4_LB_RO(EXT4_LK_DE),
  85 +       EXT4_LB_DE              = EXT4_LB_RW(EXT4_LK_DE),
  86 +       /* DX spinlock bits */
  87 +       EXT4_LB_SPIN_RO         = EXT4_LB_RO(EXT4_LK_SPIN),
  88 +       EXT4_LB_SPIN            = EXT4_LB_RW(EXT4_LK_SPIN),
  89 +       /* accurate searching */
  90 +       EXT4_LB_EXACT           = EXT4_LB_RO(EXT4_LK_MAX << 1),
  91 +};
  92 +
  93 +enum ext4_pdo_lock_opc {
  94 +       /* external */
  95 +       EXT4_HLOCK_READDIR      = (EXT4_LB_DE_RO | EXT4_LB_DX_RO),
  96 +       EXT4_HLOCK_LOOKUP       = (EXT4_LB_DE_RO | EXT4_LB_SPIN_RO |
  97 +                                  EXT4_LB_EXACT),
  98 +       EXT4_HLOCK_DEL          = (EXT4_LB_DE | EXT4_LB_SPIN_RO |
  99 +                                  EXT4_LB_EXACT),
 100 +       EXT4_HLOCK_ADD          = (EXT4_LB_DE | EXT4_LB_SPIN_RO),
 101 +
 102 +       /* internal */
 103 +       EXT4_HLOCK_LOOKUP_SAFE  = (EXT4_LB_DE_RO | EXT4_LB_DX_RO |
 104 +                                  EXT4_LB_EXACT),
 105 +       EXT4_HLOCK_DEL_SAFE     = (EXT4_LB_DE | EXT4_LB_DX_RO | EXT4_LB_EXACT),
 106 +       EXT4_HLOCK_SPLIT        = (EXT4_LB_DE | EXT4_LB_DX | EXT4_LB_SPIN),
 107 +};
 108 +
 109 +extern struct htree_lock_head *ext4_htree_lock_head_alloc(unsigned hbits);
 110 +#define ext4_htree_lock_head_free(lhead)       htree_lock_head_free(lhead)
 111 +
 112 +extern struct htree_lock *ext4_htree_lock_alloc(void);
 113 +#define ext4_htree_lock_free(lck)              htree_lock_free(lck)
 114 +
 115 +extern void ext4_htree_lock(struct htree_lock *lck,
 116 +                           struct htree_lock_head *lhead,
 117 +                           struct inode *dir, unsigned flags);
 118 +#define ext4_htree_unlock(lck)                  htree_unlock(lck)
 119 +
 120 +extern struct buffer_head *ext4_find_entry_locked(struct inode *dir,
 121 +                                       const struct qstr *d_name,
 122 +                                       struct ext4_dir_entry_2 **res_dir,
 123 +                                       int *inlined, struct htree_lock *lck);
 124 +extern int ext4_add_entry_locked(handle_t *handle, struct dentry *dentry,
 125 +                     struct inode *inode, struct htree_lock *lck);
 126 +
 127  struct ext4_filename {
 128         const struct qstr *usr_fname;
 129         struct fscrypt_str disk_name;
 130 @@ -2666,11 +2736,19 @@ void ext4_insert_dentry(struct inode *in
 131                         struct ext4_filename *fname, void *data);
 132  static inline void ext4_update_dx_flag(struct inode *inode)
 133  {
 134 +       /* Disable it for ldiskfs, because going from a DX directory to
 135 +        * a non-DX directory while it is in use will completely break
 136 +        * the htree-locking.
 137 +        * If we really want to support this operation in the future,
 138 +        * we need to exclusively lock the directory at here which will
 139 +        * increase complexity of code */
 140 +#if 0
 141         if (!ext4_has_feature_dir_index(inode->i_sb)) {
 142                 /* ext4_iget() should have caught this... */
 143                 WARN_ON_ONCE(ext4_has_feature_metadata_csum(inode->i_sb));
 144                 ext4_clear_inode_flag(inode, EXT4_INODE_INDEX);
 145         }
 146 +#endif
 147  }
 148  static const unsigned char ext4_filetype_table[] = {
 149         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
 150 --- a/fs/ext4/namei.c
 151 +++ b/fs/ext4/namei.c
 152 @@ -55,6 +55,7 @@ struct buffer_head *ext4_append(handle_t
 153                                         ext4_lblk_t *block)
 154  {
 155         struct buffer_head *bh;
 156 +       struct ext4_inode_info *ei = EXT4_I(inode);
 157         int err;
 158
 159         if (unlikely(EXT4_SB(inode->i_sb)->s_max_dir_size_kb &&
 160 @@ -62,15 +63,22 @@ struct buffer_head *ext4_append(handle_t
 161                       EXT4_SB(inode->i_sb)->s_max_dir_size_kb)))
 162                 return ERR_PTR(-ENOSPC);
 163
 164 +       /* with parallel dir operations all appends
 165 +       * have to be serialized -bzzz */
 166 +       down(&ei->i_append_sem);
 167 +
 168         *block = inode->i_size >> inode->i_sb->s_blocksize_bits;
 169
 170         bh = ext4_bread(handle, inode, *block, EXT4_GET_BLOCKS_CREATE);
 171 -       if (IS_ERR(bh))
 172 +       if (IS_ERR(bh)) {
 173 +               up(&ei->i_append_sem);
 174                 return bh;
 175 +       }
 176         inode->i_size += inode->i_sb->s_blocksize;
 177         EXT4_I(inode)->i_disksize = inode->i_size;
 178         BUFFER_TRACE(bh, "get_write_access");
 179         err = ext4_journal_get_write_access(handle, bh);
 180 +       up(&ei->i_append_sem);
 181         if (err) {
 182                 brelse(bh);
 183                 ext4_std_error(inode->i_sb, err);
 184 @@ -271,7 +279,8 @@ static unsigned dx_node_limit(struct ino
 185  static struct dx_frame *dx_probe(struct ext4_filename *fname,
 186                                  struct inode *dir,
 187                                  struct dx_hash_info *hinfo,
 188 -                                struct dx_frame *frame);
 189 +                                struct dx_frame *frame,
 190 +                                struct htree_lock *lck);
 191  static void dx_release(struct dx_frame *frames);
 192  static int dx_make_map(struct inode *dir, struct ext4_dir_entry_2 *de,
 193                        unsigned blocksize, struct dx_hash_info *hinfo,
 194 @@ -285,12 +294,13 @@ static void dx_insert_block(struct dx_fr
 195  static int ext4_htree_next_block(struct inode *dir, __u32 hash,
 196                                  struct dx_frame *frame,
 197                                  struct dx_frame *frames,
 198 -                                __u32 *start_hash);
 199 +                                __u32 *start_hash, struct htree_lock *lck);
 200  static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
 201                 struct ext4_filename *fname,
 202 -               struct ext4_dir_entry_2 **res_dir);
 203 +               struct ext4_dir_entry_2 **res_dir, struct htree_lock *lck);
 204  static int ext4_dx_add_entry(handle_t *handle, struct ext4_filename *fname,
 205 -                            struct inode *dir, struct inode *inode);
 206 +                            struct inode *dir, struct inode *inode,
 207 +                            struct htree_lock *lck);
 208
 209  /* checksumming functions */
 210  void ext4_initialize_dirent_tail(struct buffer_head *bh,
 211 @@ -755,6 +765,227 @@ struct stats dx_show_entries(struct dx_h
 212  }
 213  #endif /* DX_DEBUG */
 214
 215 +/* private data for htree_lock */
 216 +struct ext4_dir_lock_data {
 217 +       unsigned                ld_flags;  /* bits-map for lock types */
 218 +       unsigned                ld_count;  /* # entries of the last DX block */
 219 +       struct dx_entry         ld_at_entry; /* copy of leaf dx_entry */
 220 +       struct dx_entry         *ld_at;    /* position of leaf dx_entry */
 221 +};
 222 +
 223 +#define ext4_htree_lock_data(l)        ((struct ext4_dir_lock_data *)(l)->lk_private)
 224 +#define ext4_find_entry(dir, name, dirent, inline) \
 225 +                       ext4_find_entry_locked(dir, name, dirent, inline, NULL)
 226 +#define ext4_add_entry(handle, dentry, inode) \
 227 +                       ext4_add_entry_locked(handle, dentry, inode, NULL)
 228 +
 229 +/* NB: ext4_lblk_t is 32 bits so we use high bits to identify invalid blk */
 230 +#define EXT4_HTREE_NODE_CHANGED        (0xcafeULL << 32)
 231 +
 232 +static void ext4_htree_event_cb(void *target, void *event)
 233 +{
 234 +       u64 *block = (u64 *)target;
 235 +
 236 +       if (*block == dx_get_block((struct dx_entry *)event))
 237 +               *block = EXT4_HTREE_NODE_CHANGED;
 238 +}
 239 +
 240 +struct htree_lock_head *ext4_htree_lock_head_alloc(unsigned hbits)
 241 +{
 242 +       struct htree_lock_head *lhead;
 243 +
 244 +       lhead = htree_lock_head_alloc(EXT4_LK_MAX, hbits, 0);
 245 +       if (lhead != NULL) {
 246 +               htree_lock_event_attach(lhead, EXT4_LK_SPIN, HTREE_EVENT_WR,
 247 +                                       ext4_htree_event_cb);
 248 +       }
 249 +       return lhead;
 250 +}
 251 +EXPORT_SYMBOL(ext4_htree_lock_head_alloc);
 252 +
 253 +struct htree_lock *ext4_htree_lock_alloc(void)
 254 +{
 255 +       return htree_lock_alloc(EXT4_LK_MAX,
 256 +                               sizeof(struct ext4_dir_lock_data));
 257 +}
 258 +EXPORT_SYMBOL(ext4_htree_lock_alloc);
 259 +
 260 +static htree_lock_mode_t ext4_htree_mode(unsigned flags)
 261 +{
 262 +       switch (flags) {
 263 +       default: /* 0 or unknown flags require EX lock */
 264 +               return HTREE_LOCK_EX;
 265 +       case EXT4_HLOCK_READDIR:
 266 +               return HTREE_LOCK_PR;
 267 +       case EXT4_HLOCK_LOOKUP:
 268 +               return HTREE_LOCK_CR;
 269 +       case EXT4_HLOCK_DEL:
 270 +       case EXT4_HLOCK_ADD:
 271 +               return HTREE_LOCK_CW;
 272 +       }
 273 +}
 274 +
 275 +/* return PR for read-only operations, otherwise return EX */
 276 +static inline htree_lock_mode_t ext4_htree_safe_mode(unsigned flags)
 277 +{
 278 +       int writer = (flags & EXT4_LB_DE) == EXT4_LB_DE;
 279 +
 280 +       /* 0 requires EX lock */
 281 +       return (flags == 0 || writer) ? HTREE_LOCK_EX : HTREE_LOCK_PR;
 282 +}
 283 +
 284 +static int ext4_htree_safe_locked(struct htree_lock *lck)
 285 +{
 286 +       int writer;
 287 +
 288 +       if (lck == NULL || lck->lk_mode == HTREE_LOCK_EX)
 289 +               return 1;
 290 +
 291 +       writer = (ext4_htree_lock_data(lck)->ld_flags & EXT4_LB_DE) ==
 292 +                EXT4_LB_DE;
 293 +       if (writer) /* all readers & writers are excluded? */
 294 +               return lck->lk_mode == HTREE_LOCK_EX;
 295 +
 296 +       /* all writers are excluded? */
 297 +       return lck->lk_mode == HTREE_LOCK_PR ||
 298 +              lck->lk_mode == HTREE_LOCK_PW ||
 299 +              lck->lk_mode == HTREE_LOCK_EX;
 300 +}
 301 +
 302 +/* relock htree_lock with EX mode if it's change operation, otherwise
 303 + * relock it with PR mode. It's noop if PDO is disabled. */
 304 +static void ext4_htree_safe_relock(struct htree_lock *lck)
 305 +{
 306 +       if (!ext4_htree_safe_locked(lck)) {
 307 +               unsigned flags = ext4_htree_lock_data(lck)->ld_flags;
 308 +
 309 +               htree_change_lock(lck, ext4_htree_safe_mode(flags));
 310 +       }
 311 +}
 312 +
 313 +void ext4_htree_lock(struct htree_lock *lck, struct htree_lock_head *lhead,
 314 +                    struct inode *dir, unsigned flags)
 315 +{
 316 +       htree_lock_mode_t mode = is_dx(dir) ? ext4_htree_mode(flags) :
 317 +                                             ext4_htree_safe_mode(flags);
 318 +
 319 +       ext4_htree_lock_data(lck)->ld_flags = flags;
 320 +       htree_lock(lck, lhead, mode);
 321 +       if (!is_dx(dir))
 322 +               ext4_htree_safe_relock(lck); /* make sure it's safe locked */
 323 +}
 324 +EXPORT_SYMBOL(ext4_htree_lock);
 325 +
 326 +static int ext4_htree_node_lock(struct htree_lock *lck, struct dx_entry *at,
 327 +                               unsigned lmask, int wait, void *ev)
 328 +{
 329 +       u32     key = (at == NULL) ? 0 : dx_get_block(at);
 330 +       u32     mode;
 331 +
 332 +       /* NOOP if htree is well protected or caller doesn't require the lock */
 333 +       if (ext4_htree_safe_locked(lck) ||
 334 +          !(ext4_htree_lock_data(lck)->ld_flags & lmask))
 335 +               return 1;
 336 +
 337 +       mode = (ext4_htree_lock_data(lck)->ld_flags & lmask) == lmask ?
 338 +               HTREE_LOCK_PW : HTREE_LOCK_PR;
 339 +       while (1) {
 340 +               if (htree_node_lock_try(lck, mode, key, ffz(~lmask), wait, ev))
 341 +                       return 1;
 342 +               if (!(lmask & EXT4_LB_SPIN)) /* not a spinlock */
 343 +                       return 0;
 344 +               cpu_relax(); /* spin until granted */
 345 +       }
 346 +}
 347 +
 348 +static int ext4_htree_node_locked(struct htree_lock *lck, unsigned lmask)
 349 +{
 350 +       return ext4_htree_safe_locked(lck) ||
 351 +              htree_node_is_granted(lck, ffz(~lmask));
 352 +}
 353 +
 354 +static void ext4_htree_node_unlock(struct htree_lock *lck,
 355 +                                  unsigned lmask, void *buf)
 356 +{
 357 +       /* NB: it's safe to call mutiple times or even it's not locked */
 358 +       if (!ext4_htree_safe_locked(lck) &&
 359 +            htree_node_is_granted(lck, ffz(~lmask)))
 360 +               htree_node_unlock(lck, ffz(~lmask), buf);
 361 +}
 362 +
 363 +#define ext4_htree_dx_lock(lck, key)           \
 364 +       ext4_htree_node_lock(lck, key, EXT4_LB_DX, 1, NULL)
 365 +#define ext4_htree_dx_lock_try(lck, key)       \
 366 +       ext4_htree_node_lock(lck, key, EXT4_LB_DX, 0, NULL)
 367 +#define ext4_htree_dx_unlock(lck)              \
 368 +       ext4_htree_node_unlock(lck, EXT4_LB_DX, NULL)
 369 +#define ext4_htree_dx_locked(lck)              \
 370 +       ext4_htree_node_locked(lck, EXT4_LB_DX)
 371 +
 372 +static void ext4_htree_dx_need_lock(struct htree_lock *lck)
 373 +{
 374 +       struct ext4_dir_lock_data *ld;
 375 +
 376 +       if (ext4_htree_safe_locked(lck))
 377 +               return;
 378 +
 379 +       ld = ext4_htree_lock_data(lck);
 380 +       switch (ld->ld_flags) {
 381 +       default:
 382 +               return;
 383 +       case EXT4_HLOCK_LOOKUP:
 384 +               ld->ld_flags = EXT4_HLOCK_LOOKUP_SAFE;
 385 +               return;
 386 +       case EXT4_HLOCK_DEL:
 387 +               ld->ld_flags = EXT4_HLOCK_DEL_SAFE;
 388 +               return;
 389 +       case EXT4_HLOCK_ADD:
 390 +               ld->ld_flags = EXT4_HLOCK_SPLIT;
 391 +               return;
 392 +       }
 393 +}
 394 +
 395 +#define ext4_htree_de_lock(lck, key)           \
 396 +       ext4_htree_node_lock(lck, key, EXT4_LB_DE, 1, NULL)
 397 +#define ext4_htree_de_unlock(lck)              \
 398 +       ext4_htree_node_unlock(lck, EXT4_LB_DE, NULL)
 399 +
 400 +#define ext4_htree_spin_lock(lck, key, event)  \
 401 +       ext4_htree_node_lock(lck, key, EXT4_LB_SPIN, 0, event)
 402 +#define ext4_htree_spin_unlock(lck)            \
 403 +       ext4_htree_node_unlock(lck, EXT4_LB_SPIN, NULL)
 404 +#define ext4_htree_spin_unlock_listen(lck, p)  \
 405 +       ext4_htree_node_unlock(lck, EXT4_LB_SPIN, p)
 406 +
 407 +static void ext4_htree_spin_stop_listen(struct htree_lock *lck)
 408 +{
 409 +       if (!ext4_htree_safe_locked(lck) &&
 410 +           htree_node_is_listening(lck, ffz(~EXT4_LB_SPIN)))
 411 +               htree_node_stop_listen(lck, ffz(~EXT4_LB_SPIN));
 412 +}
 413 +
 414 +enum {
 415 +       DX_HASH_COL_IGNORE,     /* ignore collision while probing frames */
 416 +       DX_HASH_COL_YES,        /* there is collision and it does matter */
 417 +       DX_HASH_COL_NO,         /* there is no collision */
 418 +};
 419 +
 420 +static int dx_probe_hash_collision(struct htree_lock *lck,
 421 +                                  struct dx_entry *entries,
 422 +                                  struct dx_entry *at, u32 hash)
 423 +{
 424 +       if (!(lck && ext4_htree_lock_data(lck)->ld_flags & EXT4_LB_EXACT)) {
 425 +               return DX_HASH_COL_IGNORE; /* don't care about collision */
 426 +
 427 +       } else if (at == entries + dx_get_count(entries) - 1) {
 428 +               return DX_HASH_COL_IGNORE; /* not in any leaf of this DX */
 429 +
 430 +       } else { /* hash collision? */
 431 +               return ((dx_get_hash(at + 1) & ~1) == hash) ?
 432 +                       DX_HASH_COL_YES : DX_HASH_COL_NO;
 433 +       }
 434 +}
 435 +
 436  /*
 437   * Probe for a directory leaf block to search.
 438   *
 439 @@ -766,10 +997,11 @@ struct stats dx_show_entries(struct dx_h
 440   */
 441  static struct dx_frame *
 442  dx_probe(struct ext4_filename *fname, struct inode *dir,
 443 -        struct dx_hash_info *hinfo, struct dx_frame *frame_in)
 444 +        struct dx_hash_info *hinfo, struct dx_frame *frame_in,
 445 +        struct htree_lock *lck)
 446  {
 447         unsigned count, indirect;
 448 -       struct dx_entry *at, *entries, *p, *q, *m;
 449 +       struct dx_entry *at, *entries, *p, *q, *m, *dx = NULL;
 450         struct dx_root_info *info;
 451         struct dx_frame *frame = frame_in;
 452         struct dx_frame *ret_err = ERR_PTR(ERR_BAD_DX_DIR);
 453 @@ -831,8 +1063,15 @@ dx_probe(struct ext4_filename *fname, st
 454
 455         dxtrace(printk("Look up %x", hash));
 456         while (1) {
 457 +               if (indirect == 0) { /* the last index level */
 458 +                       /* NB: ext4_htree_dx_lock() could be noop if
 459 +                        * DX-lock flag is not set for current operation */
 460 +                       ext4_htree_dx_lock(lck, dx);
 461 +                       ext4_htree_spin_lock(lck, dx, NULL);
 462 +               }
 463                 count = dx_get_count(entries);
 464 -               if (!count || count > dx_get_limit(entries)) {
 465 +               if (count == 0 || count > dx_get_limit(entries)) {
 466 +                       ext4_htree_spin_unlock(lck); /* release spin */
 467                         ext4_warning_inode(dir,
 468                                            "dx entry: count %u beyond limit %u",
 469                                            count, dx_get_limit(entries));
 470 @@ -871,8 +1110,70 @@ dx_probe(struct ext4_filename *fname, st
 471                                dx_get_block(at)));
 472                 frame->entries = entries;
 473                 frame->at = at;
 474 -               if (!indirect--)
 475 +
 476 +               if (indirect == 0) { /* the last index level */
 477 +                       struct ext4_dir_lock_data *ld;
 478 +                       u64 myblock;
 479 +
 480 +                       /* By default we only lock DE-block, however, we will
 481 +                        * also lock the last level DX-block if:
 482 +                        * a) there is hash collision
 483 +                        *    we will set DX-lock flag (a few lines below)
 484 +                        *    and redo to lock DX-block
 485 +                        *    see detail in dx_probe_hash_collision()
 486 +                        * b) it's a retry from splitting
 487 +                        *    we need to lock the last level DX-block so nobody
 488 +                        *    else can split any leaf blocks under the same
 489 +                        *    DX-block, see detail in ext4_dx_add_entry()
 490 +                        */
 491 +                       if (ext4_htree_dx_locked(lck)) {
 492 +                               /* DX-block is locked, just lock DE-block
 493 +                                * and return */
 494 +                               ext4_htree_spin_unlock(lck);
 495 +                               if (!ext4_htree_safe_locked(lck))
 496 +                                       ext4_htree_de_lock(lck, frame->at);
 497 +                               return frame;
 498 +                       }
 499 +                       /* it's pdirop and no DX lock */
 500 +                       if (dx_probe_hash_collision(lck, entries, at, hash) ==
 501 +                           DX_HASH_COL_YES) {
 502 +                               /* found hash collision, set DX-lock flag
 503 +                                * and retry to abtain DX-lock */
 504 +                               ext4_htree_spin_unlock(lck);
 505 +                               ext4_htree_dx_need_lock(lck);
 506 +                               continue;
 507 +                       }
 508 +                       ld = ext4_htree_lock_data(lck);
 509 +                       /* because I don't lock DX, so @at can't be trusted
 510 +                        * after I release spinlock so I have to save it */
 511 +                       ld->ld_at = at;
 512 +                       ld->ld_at_entry = *at;
 513 +                       ld->ld_count = dx_get_count(entries);
 514 +
 515 +                       frame->at = &ld->ld_at_entry;
 516 +                       myblock = dx_get_block(at);
 517 +
 518 +                       /* NB: ordering locking */
 519 +                       ext4_htree_spin_unlock_listen(lck, &myblock);
 520 +                       /* other thread can split this DE-block because:
 521 +                        * a) I don't have lock for the DE-block yet
 522 +                        * b) I released spinlock on DX-block
 523 +                        * if it happened I can detect it by listening
 524 +                        * splitting event on this DE-block */
 525 +                       ext4_htree_de_lock(lck, frame->at);
 526 +                       ext4_htree_spin_stop_listen(lck);
 527 +
 528 +                       if (myblock == EXT4_HTREE_NODE_CHANGED) {
 529 +                               /* someone split this DE-block before
 530 +                                * I locked it, I need to retry and lock
 531 +                                * valid DE-block */
 532 +                               ext4_htree_de_unlock(lck);
 533 +                               continue;
 534 +                       }
 535                         return frame;
 536 +               }
 537 +               dx = at;
 538 +               indirect--;
 539                 frame++;
 540                 frame->bh = ext4_read_dirblock(dir, dx_get_block(at), INDEX);
 541                 if (IS_ERR(frame->bh)) {
 542 @@ -941,7 +1242,7 @@ static void dx_release(struct dx_frame *
 543  static int ext4_htree_next_block(struct inode *dir, __u32 hash,
 544                                  struct dx_frame *frame,
 545                                  struct dx_frame *frames,
 546 -                                __u32 *start_hash)
 547 +                                __u32 *start_hash, struct htree_lock *lck)
 548  {
 549         struct dx_frame *p;
 550         struct buffer_head *bh;
 551 @@ -956,12 +1257,22 @@ static int ext4_htree_next_block(struct
 552          * this loop, num_frames indicates the number of interior
 553          * nodes need to be read.
 554          */
 555 +       ext4_htree_de_unlock(lck);
 556         while (1) {
 557 -               if (++(p->at) < p->entries + dx_get_count(p->entries))
 558 -                       break;
 559 +               if (num_frames > 0 || ext4_htree_dx_locked(lck)) {
 560 +                       /* num_frames > 0 :
 561 +                        *   DX block
 562 +                        * ext4_htree_dx_locked:
 563 +                        *   frame->at is reliable pointer returned by dx_probe,
 564 +                        *   otherwise dx_probe already knew no collision */
 565 +                       if (++(p->at) < p->entries + dx_get_count(p->entries))
 566 +                               break;
 567 +               }
 568                 if (p == frames)
 569                         return 0;
 570                 num_frames++;
 571 +               if (num_frames == 1)
 572 +                       ext4_htree_dx_unlock(lck);
 573                 p--;
 574         }
 575
 576 @@ -984,6 +1295,13 @@ static int ext4_htree_next_block(struct
 577          * block so no check is necessary
 578          */
 579         while (num_frames--) {
 580 +               if (num_frames == 0) {
 581 +                       /* it's not always necessary, we just don't want to
 582 +                        * detect hash collision again */
 583 +                       ext4_htree_dx_need_lock(lck);
 584 +                       ext4_htree_dx_lock(lck, p->at);
 585 +               }
 586 +
 587                 bh = ext4_read_dirblock(dir, dx_get_block(p->at), INDEX);
 588                 if (IS_ERR(bh))
 589                         return PTR_ERR(bh);
 590 @@ -992,6 +1310,7 @@ static int ext4_htree_next_block(struct
 591                 p->bh = bh;
 592                 p->at = p->entries = ((struct dx_node *) bh->b_data)->entries;
 593         }
 594 +       ext4_htree_de_lock(lck, p->at);
 595         return 1;
 596  }
 597
 598 @@ -1136,10 +1455,10 @@ int ext4_htree_fill_tree(struct file *di
 599         }
 600         hinfo.hash = start_hash;
 601         hinfo.minor_hash = 0;
 602 -       frame = dx_probe(NULL, dir, &hinfo, frames);
 603 +       /* assume it's PR locked */
 604 +       frame = dx_probe(NULL, dir, &hinfo, frames, NULL);
 605         if (IS_ERR(frame))
 606                 return PTR_ERR(frame);
 607 -
 608         /* Add '.' and '..' from the htree header */
 609         if (!start_hash && !start_minor_hash) {
 610                 de = (struct ext4_dir_entry_2 *) frames[0].bh->b_data;
 611 @@ -1179,7 +1498,7 @@ int ext4_htree_fill_tree(struct file *di
 612                 count += ret;
 613                 hashval = ~0;
 614                 ret = ext4_htree_next_block(dir, HASH_NB_ALWAYS,
 615 -                                           frame, frames, &hashval);
 616 +                                           frame, frames, &hashval, NULL);
 617                 *next_hash = hashval;
 618                 if (ret < 0) {
 619                         err = ret;
 620 @@ -1455,7 +1774,7 @@ static int is_dx_internal_node(struct in
 621  static struct buffer_head *__ext4_find_entry(struct inode *dir,
 622                                              struct ext4_filename *fname,
 623                                              struct ext4_dir_entry_2 **res_dir,
 624 -                                            int *inlined)
 625 +                                            int *inlined, struct htree_lock *lck)
 626  {
 627         struct super_block *sb;
 628         struct buffer_head *bh_use[NAMEI_RA_SIZE];
 629 @@ -1497,7 +1816,7 @@ static struct buffer_head *__ext4_find_e
 630                 goto restart;
 631         }
 632         if (is_dx(dir)) {
 633 -               ret = ext4_dx_find_entry(dir, fname, res_dir);
 634 +               ret = ext4_dx_find_entry(dir, fname, res_dir, lck);
 635                 /*
 636                  * On success, or if the error was file not found,
 637                  * return.  Otherwise, fall back to doing a search the
 638 @@ -1507,6 +1826,7 @@ static struct buffer_head *__ext4_find_e
 639                         goto cleanup_and_exit;
 640                 dxtrace(printk(KERN_DEBUG "ext4_find_entry: dx failed, "
 641                                "falling back\n"));
 642 +               ext4_htree_safe_relock(lck);
 643                 ret = NULL;
 644         }
 645         nblocks = dir->i_size >> EXT4_BLOCK_SIZE_BITS(sb);
 646 @@ -1597,10 +1917,10 @@ cleanup_and_exit:
 647         return ret;
 648  }
 649
 650 -static struct buffer_head *ext4_find_entry(struct inode *dir,
 651 +struct buffer_head *ext4_find_entry_locked(struct inode *dir,
 652                                            const struct qstr *d_name,
 653                                            struct ext4_dir_entry_2 **res_dir,
 654 -                                          int *inlined)
 655 +                                          int *inlined, struct htree_lock *lck)
 656  {
 657         int err;
 658         struct ext4_filename fname;
 659 @@ -1612,12 +1932,14 @@ static struct buffer_head *ext4_find_ent
 660         if (err)
 661                 return ERR_PTR(err);
 662
 663 -       bh = __ext4_find_entry(dir, &fname, res_dir, inlined);
 664 +       bh = __ext4_find_entry(dir, &fname, res_dir, inlined, lck);
 665
 666         ext4_fname_free_filename(&fname);
 667         return bh;
 668  }
 669
 670 +EXPORT_SYMBOL(ext4_find_entry_locked);
 671 +
 672  static struct buffer_head *ext4_lookup_entry(struct inode *dir,
 673                                              struct dentry *dentry,
 674                                              struct ext4_dir_entry_2 **res_dir)
 675 @@ -1632,7 +1954,7 @@ static struct buffer_head *ext4_lookup_e
 676         if (err)
 677                 return ERR_PTR(err);
 678
 679 -       bh = __ext4_find_entry(dir, &fname, res_dir, NULL);
 680 +       bh = __ext4_find_entry(dir, &fname, res_dir, NULL, NULL);
 681
 682         ext4_fname_free_filename(&fname);
 683         return bh;
 684 @@ -1640,7 +1962,8 @@ static struct buffer_head *ext4_lookup_e
 685
 686  static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
 687                         struct ext4_filename *fname,
 688 -                       struct ext4_dir_entry_2 **res_dir)
 689 +                       struct ext4_dir_entry_2 **res_dir,
 690 +                       struct htree_lock *lck)
 691  {
 692         struct super_block * sb = dir->i_sb;
 693         struct dx_frame frames[EXT4_HTREE_LEVEL], *frame;
 694 @@ -1651,7 +1974,7 @@ static struct buffer_head * ext4_dx_find
 695  #ifdef CONFIG_FS_ENCRYPTION
 696         *res_dir = NULL;
 697  #endif
 698 -       frame = dx_probe(fname, dir, NULL, frames);
 699 +       frame = dx_probe(fname, dir, NULL, frames, lck);
 700         if (IS_ERR(frame))
 701                 return (struct buffer_head *) frame;
 702         do {
 703 @@ -1673,7 +1996,7 @@ static struct buffer_head * ext4_dx_find
 704
 705                 /* Check to see if we should continue to search */
 706                 retval = ext4_htree_next_block(dir, fname->hinfo.hash, frame,
 707 -                                              frames, NULL);
 708 +                                              frames, NULL, lck);
 709                 if (retval < 0) {
 710                         ext4_warning_inode(dir,
 711                                 "error %d reading directory index block",
 712 @@ -1853,8 +2176,9 @@ static struct ext4_dir_entry_2* dx_pack_
 713   * Returns pointer to de in block into which the new entry will be inserted.
 714   */
 715  static struct ext4_dir_entry_2 *do_split(handle_t *handle, struct inode *dir,
 716 -                       struct buffer_head **bh,struct dx_frame *frame,
 717 -                       struct dx_hash_info *hinfo)
 718 +                       struct buffer_head **bh, struct dx_frame *frames,
 719 +                       struct dx_frame *frame, struct dx_hash_info *hinfo,
 720 +                       struct htree_lock *lck)
 721  {
 722         unsigned blocksize = dir->i_sb->s_blocksize;
 723         unsigned count, continued;
 724 @@ -1915,8 +2239,14 @@ static struct ext4_dir_entry_2 *do_split
 725                                         hash2, split, count-split));
 726
 727         /* Fancy dance to stay within two buffers */
 728 -       de2 = dx_move_dirents(data1, data2, map + split, count - split,
 729 -                             blocksize);
 730 +       if (hinfo->hash < hash2) {
 731 +               de2 = dx_move_dirents(data1, data2, map + split,
 732 +                                     count - split, blocksize);
 733 +       } else {
 734 +               /* make sure we will add entry to the same block which
 735 +                * we have already locked */
 736 +               de2 = dx_move_dirents(data1, data2, map, split, blocksize);
 737 +       }
 738         de = dx_pack_dirents(data1, blocksize);
 739         de->rec_len = ext4_rec_len_to_disk(data1 + (blocksize - csum_size) -
 740                                            (char *) de,
 741 @@ -1934,12 +2264,21 @@ static struct ext4_dir_entry_2 *do_split
 742         dxtrace(dx_show_leaf(dir, hinfo, (struct ext4_dir_entry_2 *) data2,
 743                         blocksize, 1));
 744
 745 -       /* Which block gets the new entry? */
 746 -       if (hinfo->hash >= hash2) {
 747 -               swap(*bh, bh2);
 748 -               de = de2;
 749 +       ext4_htree_spin_lock(lck, frame > frames ? (frame - 1)->at : NULL,
 750 +                            frame->at); /* notify block is being split */
 751 +       if (hinfo->hash < hash2) {
 752 +               dx_insert_block(frame, hash2 + continued, newblock);
 753 +
 754 +       } else {
 755 +               /* switch block number */
 756 +               dx_insert_block(frame, hash2 + continued,
 757 +                               dx_get_block(frame->at));
 758 +               dx_set_block(frame->at, newblock);
 759 +               (frame->at)++;
 760         }
 761 -       dx_insert_block(frame, hash2 + continued, newblock);
 762 +       ext4_htree_spin_unlock(lck);
 763 +       ext4_htree_dx_unlock(lck);
 764 +
 765         err = ext4_handle_dirty_dirblock(handle, dir, bh2);
 766         if (err)
 767                 goto journal_error;
 768 @@ -2209,7 +2548,7 @@ static int make_indexed_dir(handle_t *ha
 769         if (retval)
 770                 goto out_frames;
 771
 772 -       de = do_split(handle,dir, &bh2, frame, &fname->hinfo);
 773 +       de = do_split(handle, dir, &bh2, frames, frame, &fname->hinfo, NULL);
 774         if (IS_ERR(de)) {
 775                 retval = PTR_ERR(de);
 776                 goto out_frames;
 777 @@ -2319,8 +2658,8 @@ out:
 778   * may not sleep between calling this and putting something into
 779   * the entry, as someone else might have used it while you slept.
 780   */
 781 -static int ext4_add_entry(handle_t *handle, struct dentry *dentry,
 782 -                         struct inode *inode)
 783 +int ext4_add_entry_locked(handle_t *handle, struct dentry *dentry,
 784 +                         struct inode *inode, struct htree_lock *lck)
 785  {
 786         struct inode *dir = d_inode(dentry->d_parent);
 787         struct buffer_head *bh = NULL;
 788 @@ -2370,9 +2709,10 @@ static int ext4_add_entry(handle_t *hand
 789                 if (dentry->d_name.len == 2 &&
 790                     memcmp(dentry->d_name.name, "..", 2) == 0)
 791                         return ext4_update_dotdot(handle, dentry, inode);
 792 -               retval = ext4_dx_add_entry(handle, &fname, dir, inode);
 793 +               retval = ext4_dx_add_entry(handle, &fname, dir, inode, lck);
 794                 if (!retval || (retval != ERR_BAD_DX_DIR))
 795                         goto out;
 796 +               ext4_htree_safe_relock(lck);
 797                 /* Can we just ignore htree data? */
 798                 if (ext4_has_metadata_csum(sb)) {
 799                         EXT4_ERROR_INODE(dir,
 800 @@ -2435,12 +2775,14 @@ out:
 801                 ext4_set_inode_state(inode, EXT4_STATE_NEWENTRY);
 802         return retval;
 803  }
 804 +EXPORT_SYMBOL(ext4_add_entry_locked);
 805
 806  /*
 807   * Returns 0 for success, or a negative error value
 808   */
 809  static int ext4_dx_add_entry(handle_t *handle, struct ext4_filename *fname,
 810 -                            struct inode *dir, struct inode *inode)
 811 +                            struct inode *dir, struct inode *inode,
 812 +                            struct htree_lock *lck)
 813  {
 814         struct dx_frame frames[EXT4_HTREE_LEVEL], *frame;
 815         struct dx_entry *entries, *at;
 816 @@ -2452,7 +2794,7 @@ static int ext4_dx_add_entry(handle_t *h
 817
 818  again:
 819         restart = 0;
 820 -       frame = dx_probe(fname, dir, NULL, frames);
 821 +       frame = dx_probe(fname, dir, NULL, frames, lck);
 822         if (IS_ERR(frame))
 823                 return PTR_ERR(frame);
 824         entries = frame->entries;
 825 @@ -2487,6 +2829,12 @@ again:
 826                 struct dx_node *node2;
 827                 struct buffer_head *bh2;
 828
 829 +               if (!ext4_htree_safe_locked(lck)) { /* retry with EX lock */
 830 +                       ext4_htree_safe_relock(lck);
 831 +                       restart = 1;
 832 +                       goto cleanup;
 833 +               }
 834 +
 835                 while (frame > frames) {
 836                         if (dx_get_count((frame - 1)->entries) <
 837                             dx_get_limit((frame - 1)->entries)) {
 838 @@ -2589,8 +2937,32 @@ again:
 839                         restart = 1;
 840                         goto journal_error;
 841                 }
 842 +       } else if (!ext4_htree_dx_locked(lck)) {
 843 +               struct ext4_dir_lock_data *ld = ext4_htree_lock_data(lck);
 844 +
 845 +               /* not well protected, require DX lock */
 846 +               ext4_htree_dx_need_lock(lck);
 847 +               at = frame > frames ? (frame - 1)->at : NULL;
 848 +
 849 +               /* NB: no risk of deadlock because it's just a try.
 850 +                *
 851 +                * NB: we check ld_count for twice, the first time before
 852 +                * having DX lock, the second time after holding DX lock.
 853 +                *
 854 +                * NB: We never free blocks for directory so far, which
 855 +                * means value returned by dx_get_count() should equal to
 856 +                * ld->ld_count if nobody split any DE-block under @at,
 857 +                * and ld->ld_at still points to valid dx_entry. */
 858 +               if ((ld->ld_count != dx_get_count(entries)) ||
 859 +                   !ext4_htree_dx_lock_try(lck, at) ||
 860 +                   (ld->ld_count != dx_get_count(entries))) {
 861 +                       restart = 1;
 862 +                       goto cleanup;
 863 +               }
 864 +               /* OK, I've got DX lock and nothing changed */
 865 +               frame->at = ld->ld_at;
 866         }
 867 -       de = do_split(handle, dir, &bh, frame, &fname->hinfo);
 868 +       de = do_split(handle, dir, &bh, frames, frame, &fname->hinfo, lck);
 869         if (IS_ERR(de)) {
 870                 err = PTR_ERR(de);
 871                 goto cleanup;
 872 @@ -2601,6 +2973,8 @@ again:
 873  journal_error:
 874         ext4_std_error(dir->i_sb, err); /* this is a no-op if err == 0 */
 875  cleanup:
 876 +       ext4_htree_dx_unlock(lck);
 877 +       ext4_htree_de_unlock(lck);
 878         brelse(bh);
 879         dx_release(frames);
 880         /* @restart is true means htree-path has been changed, we need to
 881 --- a/fs/ext4/super.c
 882 +++ b/fs/ext4/super.c
 883 @@ -1122,6 +1122,7 @@ static struct inode *ext4_alloc_inode(st
 884
 885         inode_set_iversion(&ei->vfs_inode, 1);
 886         spin_lock_init(&ei->i_raw_lock);
 887 +       sema_init(&ei->i_append_sem, 1);
 888         INIT_LIST_HEAD(&ei->i_prealloc_list);
 889         spin_lock_init(&ei->i_prealloc_lock);
 890         ext4_es_init_tree(&ei->i_es_tree);