ldiskfs/kernel_patches/patches/ubuntu20/ext4-pdirop.patch

   1 From 1a0f7f0b9c13ef0aa86e125f350b6733bff8db3c Mon Sep 17 00:00:00 2001
   2 From: Liang Zhen <liang.zhen@intel.com>
   3 Date: Wed, 15 Jan 2020 07:35:13 -0600
   4 Subject: [PATCH] LU-50 ldiskfs: parallel directory operations for ext4
   5
   6  In a typical use case an application creates a separate output file for each
   7  node and task in a job. As nodes and tasks increase, hundreds of thousands of
   8  files may be created in a single directory within a short window of time.
   9  Today, both filename lookup and file system modifying operations (such as
  10  create and unlink) are protected with a single lock for an entire ldiskfs
  11  directory. PDO project will remove this bottleneck by introducing a parallel
  12  locking mechanism for entire ldiskfs directories. This work will enable
  13  multiple application threads to simultaneously lookup, create and unlink in
  14  parallel.
  15
  16 This patch contains:
  17  - pdirops support for ldiskfs
  18  - integrate with osd-ldiskfs
  19 Signed-off-by: Liang Zhen <liang.zhen@intel.com>
  20 Change-Id: I269c0e3112e68f3acd79e860dab052a68c7d7aaa
  21 ---
  22  fs/ext4/Makefile           |   1 +
  23  fs/ext4/ext4.h             |  78 ++++
  24  fs/ext4/htree_lock.c       | 891 +++++++++++++++++++++++++++++++++++++
  25  fs/ext4/namei.c            | 454 +++++++++++++++++--
  26  fs/ext4/super.c            |   1 +
  27  include/linux/htree_lock.h | 187 ++++++++
  28  6 files changed, 1572 insertions(+), 40 deletions(-)
  29  create mode 100644 fs/ext4/htree_lock.c
  30  create mode 100644 include/linux/htree_lock.h
  31
  32 diff -wur a/fs/ext4/ext4.h b/fs/ext4/ext4.h
  33 --- a/fs/ext4/ext4.h    2020-08-30 12:06:02.782523259 -0600
  34 +++ b/fs/ext4/ext4.h    2020-08-30 12:09:18.997212399 -0600
  35 @@ -29,6 +29,7 @@
  36  #include <linux/timer.h>
  37  #include <linux/version.h>
  38  #include <linux/wait.h>
  39 +#include <linux/htree_lock.h>
  40  #include <linux/sched/signal.h>
  41  #include <linux/blockgroup_lock.h>
  42  #include <linux/percpu_counter.h>
  43 @@ -961,6 +962,9 @@
  44         __u32   i_dtime;
  45         ext4_fsblk_t    i_file_acl;
  46
  47 +       /* following fields for parallel directory operations -bzzz */
  48 +       struct semaphore i_append_sem;
  49 +
  50         /*
  51          * i_block_group is the number of the block group which contains
  52          * this file's inode.  Constant across the lifetime of the inode,
  53 @@ -2206,6 +2210,72 @@
  54   */
  55  #define HASH_NB_ALWAYS         1
  56
  57 +/* assume name-hash is protected by upper layer */
  58 +#define EXT4_HTREE_LOCK_HASH   0
  59 +
  60 +enum ext4_pdo_lk_types {
  61 +#if EXT4_HTREE_LOCK_HASH
  62 +       EXT4_LK_HASH,
  63 +#endif
  64 +       EXT4_LK_DX,             /* index block */
  65 +       EXT4_LK_DE,             /* directory entry block */
  66 +       EXT4_LK_SPIN,           /* spinlock */
  67 +       EXT4_LK_MAX,
  68 +};
  69 +
  70 +/* read-only bit */
  71 +#define EXT4_LB_RO(b)          (1 << (b))
  72 +/* read + write, high bits for writer */
  73 +#define EXT4_LB_RW(b)          ((1 << (b)) | (1 << (EXT4_LK_MAX + (b))))
  74 +
  75 +enum ext4_pdo_lock_bits {
  76 +       /* DX lock bits */
  77 +       EXT4_LB_DX_RO           = EXT4_LB_RO(EXT4_LK_DX),
  78 +       EXT4_LB_DX              = EXT4_LB_RW(EXT4_LK_DX),
  79 +       /* DE lock bits */
  80 +       EXT4_LB_DE_RO           = EXT4_LB_RO(EXT4_LK_DE),
  81 +       EXT4_LB_DE              = EXT4_LB_RW(EXT4_LK_DE),
  82 +       /* DX spinlock bits */
  83 +       EXT4_LB_SPIN_RO         = EXT4_LB_RO(EXT4_LK_SPIN),
  84 +       EXT4_LB_SPIN            = EXT4_LB_RW(EXT4_LK_SPIN),
  85 +       /* accurate searching */
  86 +       EXT4_LB_EXACT           = EXT4_LB_RO(EXT4_LK_MAX << 1),
  87 +};
  88 +
  89 +enum ext4_pdo_lock_opc {
  90 +       /* external */
  91 +       EXT4_HLOCK_READDIR      = (EXT4_LB_DE_RO | EXT4_LB_DX_RO),
  92 +       EXT4_HLOCK_LOOKUP       = (EXT4_LB_DE_RO | EXT4_LB_SPIN_RO |
  93 +                                  EXT4_LB_EXACT),
  94 +       EXT4_HLOCK_DEL          = (EXT4_LB_DE | EXT4_LB_SPIN_RO |
  95 +                                  EXT4_LB_EXACT),
  96 +       EXT4_HLOCK_ADD          = (EXT4_LB_DE | EXT4_LB_SPIN_RO),
  97 +
  98 +       /* internal */
  99 +       EXT4_HLOCK_LOOKUP_SAFE  = (EXT4_LB_DE_RO | EXT4_LB_DX_RO |
 100 +                                  EXT4_LB_EXACT),
 101 +       EXT4_HLOCK_DEL_SAFE     = (EXT4_LB_DE | EXT4_LB_DX_RO | EXT4_LB_EXACT),
 102 +       EXT4_HLOCK_SPLIT        = (EXT4_LB_DE | EXT4_LB_DX | EXT4_LB_SPIN),
 103 +};
 104 +
 105 +extern struct htree_lock_head *ext4_htree_lock_head_alloc(unsigned hbits);
 106 +#define ext4_htree_lock_head_free(lhead)       htree_lock_head_free(lhead)
 107 +
 108 +extern struct htree_lock *ext4_htree_lock_alloc(void);
 109 +#define ext4_htree_lock_free(lck)              htree_lock_free(lck)
 110 +
 111 +extern void ext4_htree_lock(struct htree_lock *lck,
 112 +                           struct htree_lock_head *lhead,
 113 +                           struct inode *dir, unsigned flags);
 114 +#define ext4_htree_unlock(lck)                  htree_unlock(lck)
 115 +
 116 +extern struct buffer_head *ext4_find_entry_locked(struct inode *dir,
 117 +                                       const struct qstr *d_name,
 118 +                                       struct ext4_dir_entry_2 **res_dir,
 119 +                                       int *inlined, struct htree_lock *lck);
 120 +extern int ext4_add_entry_locked(handle_t *handle, struct dentry *dentry,
 121 +                     struct inode *inode, struct htree_lock *lck);
 122 +
 123  struct ext4_filename {
 124         const struct qstr *usr_fname;
 125         struct fscrypt_str disk_name;
 126 @@ -2573,11 +2643,20 @@
 127                         struct ext4_filename *fname, void *data);
 128  static inline void ext4_update_dx_flag(struct inode *inode)
 129  {
 130 +       /* Disable it for ldiskfs, because going from a DX directory to
 131 +        * a non-DX directory while it is in use will completely break
 132 +        * the htree-locking.
 133 +        * If we really want to support this operation in the future,
 134 +        * we need to exclusively lock the directory at here which will
 135 +        * increase complexity of code
 136 +        */
 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 diff -wur a/fs/ext4/Makefile b/fs/ext4/Makefile
 148 --- a/fs/ext4/Makefile  2020-08-30 12:06:02.378525933 -0600
 149 +++ b/fs/ext4/Makefile  2020-08-30 12:07:32.337927838 -0600
 150 @@ -7,6 +7,7 @@
 151
 152  ext4-y := balloc.o bitmap.o block_validity.o dir.o ext4_jbd2.o extents.o \
 153                 extents_status.o file.o fsmap.o fsync.o hash.o ialloc.o \
 154 +               htree_lock.o \
 155                 indirect.o inline.o inode.o ioctl.o mballoc.o migrate.o \
 156                 mmp.o move_extent.o namei.o page-io.o readpage.o resize.o \
 157                 super.o symlink.o sysfs.o xattr.o xattr_trusted.o xattr_user.o
 158 diff -wur a/fs/ext4/namei.c b/fs/ext4/namei.c
 159 --- a/fs/ext4/namei.c   2020-08-30 12:06:02.746523498 -0600
 160 +++ b/fs/ext4/namei.c   2020-08-30 12:11:25.136359125 -0600
 161 @@ -55,6 +55,7 @@
 162                                         ext4_lblk_t *block)
 163  {
 164         struct buffer_head *bh;
 165 +       struct ext4_inode_info *ei = EXT4_I(inode);
 166         int err;
 167
 168         if (unlikely(EXT4_SB(inode->i_sb)->s_max_dir_size_kb &&
 169 @@ -62,15 +63,22 @@
 170                       EXT4_SB(inode->i_sb)->s_max_dir_size_kb)))
 171                 return ERR_PTR(-ENOSPC);
 172
 173 +       /* with parallel dir operations all appends
 174 +       * have to be serialized -bzzz */
 175 +       down(&ei->i_append_sem);
 176 +
 177         *block = inode->i_size >> inode->i_sb->s_blocksize_bits;
 178
 179         bh = ext4_bread(handle, inode, *block, EXT4_GET_BLOCKS_CREATE);
 180 -       if (IS_ERR(bh))
 181 +       if (IS_ERR(bh)) {
 182 +               up(&ei->i_append_sem);
 183                 return bh;
 184 +       }
 185         inode->i_size += inode->i_sb->s_blocksize;
 186         EXT4_I(inode)->i_disksize = inode->i_size;
 187         BUFFER_TRACE(bh, "get_write_access");
 188         err = ext4_journal_get_write_access(handle, bh);
 189 +       up(&ei->i_append_sem);
 190         if (err) {
 191                 brelse(bh);
 192                 ext4_std_error(inode->i_sb, err);
 193 @@ -264,7 +272,8 @@
 194  static struct dx_frame *dx_probe(struct ext4_filename *fname,
 195                                  struct inode *dir,
 196                                  struct dx_hash_info *hinfo,
 197 -                                struct dx_frame *frame);
 198 +                                struct dx_frame *frame,
 199 +                                struct htree_lock *lck);
 200  static void dx_release(struct dx_frame *frames);
 201  static int dx_make_map(struct inode *dir, struct ext4_dir_entry_2 *de,
 202                        unsigned blocksize, struct dx_hash_info *hinfo,
 203 @@ -278,12 +287,13 @@
 204  static int ext4_htree_next_block(struct inode *dir, __u32 hash,
 205                                  struct dx_frame *frame,
 206                                  struct dx_frame *frames,
 207 -                                __u32 *start_hash);
 208 +                                __u32 *start_hash, struct htree_lock *lck);
 209  static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
 210                 struct ext4_filename *fname,
 211 -               struct ext4_dir_entry_2 **res_dir);
 212 +               struct ext4_dir_entry_2 **res_dir, struct htree_lock *lck);
 213  static int ext4_dx_add_entry(handle_t *handle, struct ext4_filename *fname,
 214 -                            struct inode *dir, struct inode *inode);
 215 +                            struct inode *dir, struct inode *inode,
 216 +                            struct htree_lock *lck);
 217
 218  /* checksumming functions */
 219  void ext4_initialize_dirent_tail(struct buffer_head *bh,
 220 @@ -748,6 +758,227 @@
 221  }
 222  #endif /* DX_DEBUG */
 223
 224 +/* private data for htree_lock */
 225 +struct ext4_dir_lock_data {
 226 +       unsigned                ld_flags;  /* bits-map for lock types */
 227 +       unsigned                ld_count;  /* # entries of the last DX block */
 228 +       struct dx_entry         ld_at_entry; /* copy of leaf dx_entry */
 229 +       struct dx_entry         *ld_at;    /* position of leaf dx_entry */
 230 +};
 231 +
 232 +#define ext4_htree_lock_data(l)        ((struct ext4_dir_lock_data *)(l)->lk_private)
 233 +#define ext4_find_entry(dir, name, dirent, inline) \
 234 +                       ext4_find_entry_locked(dir, name, dirent, inline, NULL)
 235 +#define ext4_add_entry(handle, dentry, inode) \
 236 +                       ext4_add_entry_locked(handle, dentry, inode, NULL)
 237 +
 238 +/* NB: ext4_lblk_t is 32 bits so we use high bits to identify invalid blk */
 239 +#define EXT4_HTREE_NODE_CHANGED        (0xcafeULL << 32)
 240 +
 241 +static void ext4_htree_event_cb(void *target, void *event)
 242 +{
 243 +       u64 *block = (u64 *)target;
 244 +
 245 +       if (*block == dx_get_block((struct dx_entry *)event))
 246 +               *block = EXT4_HTREE_NODE_CHANGED;
 247 +}
 248 +
 249 +struct htree_lock_head *ext4_htree_lock_head_alloc(unsigned hbits)
 250 +{
 251 +       struct htree_lock_head *lhead;
 252 +
 253 +       lhead = htree_lock_head_alloc(EXT4_LK_MAX, hbits, 0);
 254 +       if (lhead != NULL) {
 255 +               htree_lock_event_attach(lhead, EXT4_LK_SPIN, HTREE_EVENT_WR,
 256 +                                       ext4_htree_event_cb);
 257 +       }
 258 +       return lhead;
 259 +}
 260 +EXPORT_SYMBOL(ext4_htree_lock_head_alloc);
 261 +
 262 +struct htree_lock *ext4_htree_lock_alloc(void)
 263 +{
 264 +       return htree_lock_alloc(EXT4_LK_MAX,
 265 +                               sizeof(struct ext4_dir_lock_data));
 266 +}
 267 +EXPORT_SYMBOL(ext4_htree_lock_alloc);
 268 +
 269 +static htree_lock_mode_t ext4_htree_mode(unsigned flags)
 270 +{
 271 +       switch (flags) {
 272 +       default: /* 0 or unknown flags require EX lock */
 273 +               return HTREE_LOCK_EX;
 274 +       case EXT4_HLOCK_READDIR:
 275 +               return HTREE_LOCK_PR;
 276 +       case EXT4_HLOCK_LOOKUP:
 277 +               return HTREE_LOCK_CR;
 278 +       case EXT4_HLOCK_DEL:
 279 +       case EXT4_HLOCK_ADD:
 280 +               return HTREE_LOCK_CW;
 281 +       }
 282 +}
 283 +
 284 +/* return PR for read-only operations, otherwise return EX */
 285 +static inline htree_lock_mode_t ext4_htree_safe_mode(unsigned flags)
 286 +{
 287 +       int writer = (flags & EXT4_LB_DE) == EXT4_LB_DE;
 288 +
 289 +       /* 0 requires EX lock */
 290 +       return (flags == 0 || writer) ? HTREE_LOCK_EX : HTREE_LOCK_PR;
 291 +}
 292 +
 293 +static int ext4_htree_safe_locked(struct htree_lock *lck)
 294 +{
 295 +       int writer;
 296 +
 297 +       if (lck == NULL || lck->lk_mode == HTREE_LOCK_EX)
 298 +               return 1;
 299 +
 300 +       writer = (ext4_htree_lock_data(lck)->ld_flags & EXT4_LB_DE) ==
 301 +                EXT4_LB_DE;
 302 +       if (writer) /* all readers & writers are excluded? */
 303 +               return lck->lk_mode == HTREE_LOCK_EX;
 304 +
 305 +       /* all writers are excluded? */
 306 +       return lck->lk_mode == HTREE_LOCK_PR ||
 307 +              lck->lk_mode == HTREE_LOCK_PW ||
 308 +              lck->lk_mode == HTREE_LOCK_EX;
 309 +}
 310 +
 311 +/* relock htree_lock with EX mode if it's change operation, otherwise
 312 + * relock it with PR mode. It's noop if PDO is disabled. */
 313 +static void ext4_htree_safe_relock(struct htree_lock *lck)
 314 +{
 315 +       if (!ext4_htree_safe_locked(lck)) {
 316 +               unsigned flags = ext4_htree_lock_data(lck)->ld_flags;
 317 +
 318 +               htree_change_lock(lck, ext4_htree_safe_mode(flags));
 319 +       }
 320 +}
 321 +
 322 +void ext4_htree_lock(struct htree_lock *lck, struct htree_lock_head *lhead,
 323 +                    struct inode *dir, unsigned flags)
 324 +{
 325 +       htree_lock_mode_t mode = is_dx(dir) ? ext4_htree_mode(flags) :
 326 +                                             ext4_htree_safe_mode(flags);
 327 +
 328 +       ext4_htree_lock_data(lck)->ld_flags = flags;
 329 +       htree_lock(lck, lhead, mode);
 330 +       if (!is_dx(dir))
 331 +               ext4_htree_safe_relock(lck); /* make sure it's safe locked */
 332 +}
 333 +EXPORT_SYMBOL(ext4_htree_lock);
 334 +
 335 +static int ext4_htree_node_lock(struct htree_lock *lck, struct dx_entry *at,
 336 +                               unsigned lmask, int wait, void *ev)
 337 +{
 338 +       u32     key = (at == NULL) ? 0 : dx_get_block(at);
 339 +       u32     mode;
 340 +
 341 +       /* NOOP if htree is well protected or caller doesn't require the lock */
 342 +       if (ext4_htree_safe_locked(lck) ||
 343 +          !(ext4_htree_lock_data(lck)->ld_flags & lmask))
 344 +               return 1;
 345 +
 346 +       mode = (ext4_htree_lock_data(lck)->ld_flags & lmask) == lmask ?
 347 +               HTREE_LOCK_PW : HTREE_LOCK_PR;
 348 +       while (1) {
 349 +               if (htree_node_lock_try(lck, mode, key, ffz(~lmask), wait, ev))
 350 +                       return 1;
 351 +               if (!(lmask & EXT4_LB_SPIN)) /* not a spinlock */
 352 +                       return 0;
 353 +               cpu_relax(); /* spin until granted */
 354 +       }
 355 +}
 356 +
 357 +static int ext4_htree_node_locked(struct htree_lock *lck, unsigned lmask)
 358 +{
 359 +       return ext4_htree_safe_locked(lck) ||
 360 +              htree_node_is_granted(lck, ffz(~lmask));
 361 +}
 362 +
 363 +static void ext4_htree_node_unlock(struct htree_lock *lck,
 364 +                                  unsigned lmask, void *buf)
 365 +{
 366 +       /* NB: it's safe to call mutiple times or even it's not locked */
 367 +       if (!ext4_htree_safe_locked(lck) &&
 368 +            htree_node_is_granted(lck, ffz(~lmask)))
 369 +               htree_node_unlock(lck, ffz(~lmask), buf);
 370 +}
 371 +
 372 +#define ext4_htree_dx_lock(lck, key)           \
 373 +       ext4_htree_node_lock(lck, key, EXT4_LB_DX, 1, NULL)
 374 +#define ext4_htree_dx_lock_try(lck, key)       \
 375 +       ext4_htree_node_lock(lck, key, EXT4_LB_DX, 0, NULL)
 376 +#define ext4_htree_dx_unlock(lck)              \
 377 +       ext4_htree_node_unlock(lck, EXT4_LB_DX, NULL)
 378 +#define ext4_htree_dx_locked(lck)              \
 379 +       ext4_htree_node_locked(lck, EXT4_LB_DX)
 380 +
 381 +static void ext4_htree_dx_need_lock(struct htree_lock *lck)
 382 +{
 383 +       struct ext4_dir_lock_data *ld;
 384 +
 385 +       if (ext4_htree_safe_locked(lck))
 386 +               return;
 387 +
 388 +       ld = ext4_htree_lock_data(lck);
 389 +       switch (ld->ld_flags) {
 390 +       default:
 391 +               return;
 392 +       case EXT4_HLOCK_LOOKUP:
 393 +               ld->ld_flags = EXT4_HLOCK_LOOKUP_SAFE;
 394 +               return;
 395 +       case EXT4_HLOCK_DEL:
 396 +               ld->ld_flags = EXT4_HLOCK_DEL_SAFE;
 397 +               return;
 398 +       case EXT4_HLOCK_ADD:
 399 +               ld->ld_flags = EXT4_HLOCK_SPLIT;
 400 +               return;
 401 +       }
 402 +}
 403 +
 404 +#define ext4_htree_de_lock(lck, key)           \
 405 +       ext4_htree_node_lock(lck, key, EXT4_LB_DE, 1, NULL)
 406 +#define ext4_htree_de_unlock(lck)              \
 407 +       ext4_htree_node_unlock(lck, EXT4_LB_DE, NULL)
 408 +
 409 +#define ext4_htree_spin_lock(lck, key, event)  \
 410 +       ext4_htree_node_lock(lck, key, EXT4_LB_SPIN, 0, event)
 411 +#define ext4_htree_spin_unlock(lck)            \
 412 +       ext4_htree_node_unlock(lck, EXT4_LB_SPIN, NULL)
 413 +#define ext4_htree_spin_unlock_listen(lck, p)  \
 414 +       ext4_htree_node_unlock(lck, EXT4_LB_SPIN, p)
 415 +
 416 +static void ext4_htree_spin_stop_listen(struct htree_lock *lck)
 417 +{
 418 +       if (!ext4_htree_safe_locked(lck) &&
 419 +           htree_node_is_listening(lck, ffz(~EXT4_LB_SPIN)))
 420 +               htree_node_stop_listen(lck, ffz(~EXT4_LB_SPIN));
 421 +}
 422 +
 423 +enum {
 424 +       DX_HASH_COL_IGNORE,     /* ignore collision while probing frames */
 425 +       DX_HASH_COL_YES,        /* there is collision and it does matter */
 426 +       DX_HASH_COL_NO,         /* there is no collision */
 427 +};
 428 +
 429 +static int dx_probe_hash_collision(struct htree_lock *lck,
 430 +                                  struct dx_entry *entries,
 431 +                                  struct dx_entry *at, u32 hash)
 432 +{
 433 +       if (!(lck && ext4_htree_lock_data(lck)->ld_flags & EXT4_LB_EXACT)) {
 434 +               return DX_HASH_COL_IGNORE; /* don't care about collision */
 435 +
 436 +       } else if (at == entries + dx_get_count(entries) - 1) {
 437 +               return DX_HASH_COL_IGNORE; /* not in any leaf of this DX */
 438 +
 439 +       } else { /* hash collision? */
 440 +               return ((dx_get_hash(at + 1) & ~1) == hash) ?
 441 +                       DX_HASH_COL_YES : DX_HASH_COL_NO;
 442 +       }
 443 +}
 444 +
 445  /*
 446   * Probe for a directory leaf block to search.
 447   *
 448 @@ -759,10 +990,11 @@
 449   */
 450  static struct dx_frame *
 451  dx_probe(struct ext4_filename *fname, struct inode *dir,
 452 -        struct dx_hash_info *hinfo, struct dx_frame *frame_in)
 453 +        struct dx_hash_info *hinfo, struct dx_frame *frame_in,
 454 +        struct htree_lock *lck)
 455  {
 456         unsigned count, indirect;
 457 -       struct dx_entry *at, *entries, *p, *q, *m;
 458 +       struct dx_entry *at, *entries, *p, *q, *m, *dx = NULL;
 459         struct dx_root_info *info;
 460         struct dx_frame *frame = frame_in;
 461         struct dx_frame *ret_err = ERR_PTR(ERR_BAD_DX_DIR);
 462 @@ -824,8 +1056,15 @@
 463
 464         dxtrace(printk("Look up %x", hash));
 465         while (1) {
 466 +               if (indirect == 0) { /* the last index level */
 467 +                       /* NB: ext4_htree_dx_lock() could be noop if
 468 +                        * DX-lock flag is not set for current operation */
 469 +                       ext4_htree_dx_lock(lck, dx);
 470 +                       ext4_htree_spin_lock(lck, dx, NULL);
 471 +               }
 472                 count = dx_get_count(entries);
 473 -               if (!count || count > dx_get_limit(entries)) {
 474 +               if (count == 0 || count > dx_get_limit(entries)) {
 475 +                       ext4_htree_spin_unlock(lck); /* release spin */
 476                         ext4_warning_inode(dir,
 477                                            "dx entry: count %u beyond limit %u",
 478                                            count, dx_get_limit(entries));
 479 @@ -864,8 +1103,70 @@
 480                                dx_get_block(at)));
 481                 frame->entries = entries;
 482                 frame->at = at;
 483 -               if (!indirect--)
 484 +
 485 +               if (indirect == 0) { /* the last index level */
 486 +                       struct ext4_dir_lock_data *ld;
 487 +                       u64 myblock;
 488 +
 489 +                       /* By default we only lock DE-block, however, we will
 490 +                        * also lock the last level DX-block if:
 491 +                        * a) there is hash collision
 492 +                        *    we will set DX-lock flag (a few lines below)
 493 +                        *    and redo to lock DX-block
 494 +                        *    see detail in dx_probe_hash_collision()
 495 +                        * b) it's a retry from splitting
 496 +                        *    we need to lock the last level DX-block so nobody
 497 +                        *    else can split any leaf blocks under the same
 498 +                        *    DX-block, see detail in ext4_dx_add_entry()
 499 +                        */
 500 +                       if (ext4_htree_dx_locked(lck)) {
 501 +                               /* DX-block is locked, just lock DE-block
 502 +                                * and return */
 503 +                               ext4_htree_spin_unlock(lck);
 504 +                               if (!ext4_htree_safe_locked(lck))
 505 +                                       ext4_htree_de_lock(lck, frame->at);
 506 +                               return frame;
 507 +                       }
 508 +                       /* it's pdirop and no DX lock */
 509 +                       if (dx_probe_hash_collision(lck, entries, at, hash) ==
 510 +                           DX_HASH_COL_YES) {
 511 +                               /* found hash collision, set DX-lock flag
 512 +                                * and retry to abtain DX-lock */
 513 +                               ext4_htree_spin_unlock(lck);
 514 +                               ext4_htree_dx_need_lock(lck);
 515 +                               continue;
 516 +                       }
 517 +                       ld = ext4_htree_lock_data(lck);
 518 +                       /* because I don't lock DX, so @at can't be trusted
 519 +                        * after I release spinlock so I have to save it */
 520 +                       ld->ld_at = at;
 521 +                       ld->ld_at_entry = *at;
 522 +                       ld->ld_count = dx_get_count(entries);
 523 +
 524 +                       frame->at = &ld->ld_at_entry;
 525 +                       myblock = dx_get_block(at);
 526 +
 527 +                       /* NB: ordering locking */
 528 +                       ext4_htree_spin_unlock_listen(lck, &myblock);
 529 +                       /* other thread can split this DE-block because:
 530 +                        * a) I don't have lock for the DE-block yet
 531 +                        * b) I released spinlock on DX-block
 532 +                        * if it happened I can detect it by listening
 533 +                        * splitting event on this DE-block */
 534 +                       ext4_htree_de_lock(lck, frame->at);
 535 +                       ext4_htree_spin_stop_listen(lck);
 536 +
 537 +                       if (myblock == EXT4_HTREE_NODE_CHANGED) {
 538 +                               /* someone split this DE-block before
 539 +                                * I locked it, I need to retry and lock
 540 +                                * valid DE-block */
 541 +                               ext4_htree_de_unlock(lck);
 542 +                               continue;
 543 +                       }
 544                         return frame;
 545 +               }
 546 +               dx = at;
 547 +               indirect--;
 548                 frame++;
 549                 frame->bh = ext4_read_dirblock(dir, dx_get_block(at), INDEX);
 550                 if (IS_ERR(frame->bh)) {
 551 @@ -934,7 +1235,7 @@
 552  static int ext4_htree_next_block(struct inode *dir, __u32 hash,
 553                                  struct dx_frame *frame,
 554                                  struct dx_frame *frames,
 555 -                                __u32 *start_hash)
 556 +                                __u32 *start_hash, struct htree_lock *lck)
 557  {
 558         struct dx_frame *p;
 559         struct buffer_head *bh;
 560 @@ -949,12 +1250,22 @@
 561          * this loop, num_frames indicates the number of interior
 562          * nodes need to be read.
 563          */
 564 +       ext4_htree_de_unlock(lck);
 565         while (1) {
 566 +               if (num_frames > 0 || ext4_htree_dx_locked(lck)) {
 567 +                       /* num_frames > 0 :
 568 +                        *   DX block
 569 +                        * ext4_htree_dx_locked:
 570 +                        *   frame->at is reliable pointer returned by dx_probe,
 571 +                        *   otherwise dx_probe already knew no collision */
 572                 if (++(p->at) < p->entries + dx_get_count(p->entries))
 573                         break;
 574 +               }
 575                 if (p == frames)
 576                         return 0;
 577                 num_frames++;
 578 +               if (num_frames == 1)
 579 +                       ext4_htree_dx_unlock(lck);
 580                 p--;
 581         }
 582
 583 @@ -977,6 +1288,13 @@
 584          * block so no check is necessary
 585          */
 586         while (num_frames--) {
 587 +               if (num_frames == 0) {
 588 +                       /* it's not always necessary, we just don't want to
 589 +                        * detect hash collision again */
 590 +                       ext4_htree_dx_need_lock(lck);
 591 +                       ext4_htree_dx_lock(lck, p->at);
 592 +               }
 593 +
 594                 bh = ext4_read_dirblock(dir, dx_get_block(p->at), INDEX);
 595                 if (IS_ERR(bh))
 596                         return PTR_ERR(bh);
 597 @@ -985,6 +1303,7 @@
 598                 p->bh = bh;
 599                 p->at = p->entries = ((struct dx_node *) bh->b_data)->entries;
 600         }
 601 +       ext4_htree_de_lock(lck, p->at);
 602         return 1;
 603  }
 604
 605 @@ -1132,10 +1451,10 @@
 606         }
 607         hinfo.hash = start_hash;
 608         hinfo.minor_hash = 0;
 609 -       frame = dx_probe(NULL, dir, &hinfo, frames);
 610 +       /* assume it's PR locked */
 611 +       frame = dx_probe(NULL, dir, &hinfo, frames, NULL);
 612         if (IS_ERR(frame))
 613                 return PTR_ERR(frame);
 614 -
 615         /* Add '.' and '..' from the htree header */
 616         if (!start_hash && !start_minor_hash) {
 617                 de = (struct ext4_dir_entry_2 *) frames[0].bh->b_data;
 618 @@ -1175,7 +1494,7 @@
 619                 count += ret;
 620                 hashval = ~0;
 621                 ret = ext4_htree_next_block(dir, HASH_NB_ALWAYS,
 622 -                                           frame, frames, &hashval);
 623 +                                           frame, frames, &hashval, NULL);
 624                 *next_hash = hashval;
 625                 if (ret < 0) {
 626                         err = ret;
 627 @@ -1451,7 +1770,7 @@
 628  static struct buffer_head *__ext4_find_entry(struct inode *dir,
 629                                              struct ext4_filename *fname,
 630                                              struct ext4_dir_entry_2 **res_dir,
 631 -                                            int *inlined)
 632 +                                            int *inlined, struct htree_lock *lck)
 633  {
 634         struct super_block *sb;
 635         struct buffer_head *bh_use[NAMEI_RA_SIZE];
 636 @@ -1493,7 +1812,7 @@
 637                 goto restart;
 638         }
 639         if (is_dx(dir)) {
 640 -               ret = ext4_dx_find_entry(dir, fname, res_dir);
 641 +               ret = ext4_dx_find_entry(dir, fname, res_dir, lck);
 642                 /*
 643                  * On success, or if the error was file not found,
 644                  * return.  Otherwise, fall back to doing a search the
 645 @@ -1503,6 +1822,7 @@
 646                         goto cleanup_and_exit;
 647                 dxtrace(printk(KERN_DEBUG "ext4_find_entry: dx failed, "
 648                                "falling back\n"));
 649 +               ext4_htree_safe_relock(lck);
 650                 ret = NULL;
 651         }
 652         nblocks = dir->i_size >> EXT4_BLOCK_SIZE_BITS(sb);
 653 @@ -1591,10 +1911,10 @@
 654         return ret;
 655  }
 656
 657 -static struct buffer_head *ext4_find_entry(struct inode *dir,
 658 +struct buffer_head *ext4_find_entry_locked(struct inode *dir,
 659                                            const struct qstr *d_name,
 660                                            struct ext4_dir_entry_2 **res_dir,
 661 -                                          int *inlined)
 662 +                                          int *inlined, struct htree_lock *lck)
 663  {
 664         int err;
 665         struct ext4_filename fname;
 666 @@ -1606,12 +1926,14 @@
 667         if (err)
 668                 return ERR_PTR(err);
 669
 670 -       bh = __ext4_find_entry(dir, &fname, res_dir, inlined);
 671 +       bh = __ext4_find_entry(dir, &fname, res_dir, inlined, lck);
 672
 673         ext4_fname_free_filename(&fname);
 674         return bh;
 675  }
 676
 677 +EXPORT_SYMBOL(ext4_find_entry_locked);
 678 +
 679  static struct buffer_head *ext4_lookup_entry(struct inode *dir,
 680                                              struct dentry *dentry,
 681                                              struct ext4_dir_entry_2 **res_dir)
 682 @@ -1626,7 +1948,7 @@
 683         if (err)
 684                 return ERR_PTR(err);
 685
 686 -       bh = __ext4_find_entry(dir, &fname, res_dir, NULL);
 687 +       bh = __ext4_find_entry(dir, &fname, res_dir, NULL, NULL);
 688
 689         ext4_fname_free_filename(&fname);
 690         return bh;
 691 @@ -1634,7 +1956,8 @@
 692
 693  static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
 694                         struct ext4_filename *fname,
 695 -                       struct ext4_dir_entry_2 **res_dir)
 696 +                       struct ext4_dir_entry_2 **res_dir,
 697 +                       struct htree_lock *lck)
 698  {
 699         struct super_block * sb = dir->i_sb;
 700         struct dx_frame frames[EXT4_HTREE_LEVEL], *frame;
 701 @@ -1645,7 +1968,7 @@
 702  #ifdef CONFIG_FS_ENCRYPTION
 703         *res_dir = NULL;
 704  #endif
 705 -       frame = dx_probe(fname, dir, NULL, frames);
 706 +       frame = dx_probe(fname, dir, NULL, frames, lck);
 707         if (IS_ERR(frame))
 708                 return (struct buffer_head *) frame;
 709         do {
 710 @@ -1667,7 +1990,7 @@
 711
 712                 /* Check to see if we should continue to search */
 713                 retval = ext4_htree_next_block(dir, fname->hinfo.hash, frame,
 714 -                                              frames, NULL);
 715 +                                              frames, NULL, lck);
 716                 if (retval < 0) {
 717                         ext4_warning_inode(dir,
 718                                 "error %d reading directory index block",
 719 @@ -1847,8 +2170,9 @@
 720   * Returns pointer to de in block into which the new entry will be inserted.
 721   */
 722  static struct ext4_dir_entry_2 *do_split(handle_t *handle, struct inode *dir,
 723 -                       struct buffer_head **bh,struct dx_frame *frame,
 724 -                       struct dx_hash_info *hinfo)
 725 +                       struct buffer_head **bh, struct dx_frame *frames,
 726 +                       struct dx_frame *frame, struct dx_hash_info *hinfo,
 727 +                       struct htree_lock *lck)
 728  {
 729         unsigned blocksize = dir->i_sb->s_blocksize;
 730         unsigned count, continued;
 731 @@ -1909,8 +2233,14 @@
 732                                         hash2, split, count-split));
 733
 734         /* Fancy dance to stay within two buffers */
 735 -       de2 = dx_move_dirents(data1, data2, map + split, count - split,
 736 -                             blocksize);
 737 +       if (hinfo->hash < hash2) {
 738 +               de2 = dx_move_dirents(data1, data2, map + split,
 739 +                                     count - split, blocksize);
 740 +       } else {
 741 +               /* make sure we will add entry to the same block which
 742 +                * we have already locked */
 743 +               de2 = dx_move_dirents(data1, data2, map, split, blocksize);
 744 +       }
 745         de = dx_pack_dirents(data1, blocksize);
 746         de->rec_len = ext4_rec_len_to_disk(data1 + (blocksize - csum_size) -
 747                                            (char *) de,
 748 @@ -1928,12 +2258,21 @@
 749         dxtrace(dx_show_leaf(dir, hinfo, (struct ext4_dir_entry_2 *) data2,
 750                         blocksize, 1));
 751
 752 -       /* Which block gets the new entry? */
 753 -       if (hinfo->hash >= hash2) {
 754 -               swap(*bh, bh2);
 755 -               de = de2;
 756 -       }
 757 +       ext4_htree_spin_lock(lck, frame > frames ? (frame - 1)->at : NULL,
 758 +                            frame->at); /* notify block is being split */
 759 +       if (hinfo->hash < hash2) {
 760         dx_insert_block(frame, hash2 + continued, newblock);
 761 +
 762 +       } else {
 763 +               /* switch block number */
 764 +               dx_insert_block(frame, hash2 + continued,
 765 +                               dx_get_block(frame->at));
 766 +               dx_set_block(frame->at, newblock);
 767 +               (frame->at)++;
 768 +       }
 769 +       ext4_htree_spin_unlock(lck);
 770 +       ext4_htree_dx_unlock(lck);
 771 +
 772         err = ext4_handle_dirty_dirblock(handle, dir, bh2);
 773         if (err)
 774                 goto journal_error;
 775 @@ -2203,7 +2542,7 @@
 776         if (retval)
 777                 goto out_frames;
 778
 779 -       de = do_split(handle,dir, &bh2, frame, &fname->hinfo);
 780 +       de = do_split(handle, dir, &bh2, frames, frame, &fname->hinfo, NULL);
 781         if (IS_ERR(de)) {
 782                 retval = PTR_ERR(de);
 783                 goto out_frames;
 784 @@ -2313,8 +2652,8 @@
 785   * may not sleep between calling this and putting something into
 786   * the entry, as someone else might have used it while you slept.
 787   */
 788 -static int ext4_add_entry(handle_t *handle, struct dentry *dentry,
 789 -                         struct inode *inode)
 790 +int ext4_add_entry_locked(handle_t *handle, struct dentry *dentry,
 791 +                         struct inode *inode, struct htree_lock *lck)
 792  {
 793         struct inode *dir = d_inode(dentry->d_parent);
 794         struct buffer_head *bh = NULL;
 795 @@ -2362,9 +2701,10 @@
 796                 if (dentry->d_name.len == 2 &&
 797                     memcmp(dentry->d_name.name, "..", 2) == 0)
 798                         return ext4_update_dotdot(handle, dentry, inode);
 799 -               retval = ext4_dx_add_entry(handle, &fname, dir, inode);
 800 +               retval = ext4_dx_add_entry(handle, &fname, dir, inode, lck);
 801                 if (!retval || (retval != ERR_BAD_DX_DIR))
 802                         goto out;
 803 +               ext4_htree_safe_relock(lck);
 804                 /* Can we just ignore htree data? */
 805                 if (ext4_has_metadata_csum(sb)) {
 806                         EXT4_ERROR_INODE(dir,
 807 @@ -2425,12 +2765,14 @@
 808                 ext4_set_inode_state(inode, EXT4_STATE_NEWENTRY);
 809         return retval;
 810  }
 811 +EXPORT_SYMBOL(ext4_add_entry_locked);
 812
 813  /*
 814   * Returns 0 for success, or a negative error value
 815   */
 816  static int ext4_dx_add_entry(handle_t *handle, struct ext4_filename *fname,
 817 -                            struct inode *dir, struct inode *inode)
 818 +                            struct inode *dir, struct inode *inode,
 819 +                            struct htree_lock *lck)
 820  {
 821         struct dx_frame frames[EXT4_HTREE_LEVEL], *frame;
 822         struct dx_entry *entries, *at;
 823 @@ -2442,7 +2784,7 @@
 824
 825  again:
 826         restart = 0;
 827 -       frame = dx_probe(fname, dir, NULL, frames);
 828 +       frame = dx_probe(fname, dir, NULL, frames, lck);
 829         if (IS_ERR(frame))
 830                 return PTR_ERR(frame);
 831         entries = frame->entries;
 832 @@ -2477,6 +2819,12 @@
 833                 struct dx_node *node2;
 834                 struct buffer_head *bh2;
 835
 836 +               if (!ext4_htree_safe_locked(lck)) { /* retry with EX lock */
 837 +                       ext4_htree_safe_relock(lck);
 838 +                       restart = 1;
 839 +                       goto cleanup;
 840 +               }
 841 +
 842                 while (frame > frames) {
 843                         if (dx_get_count((frame - 1)->entries) <
 844                             dx_get_limit((frame - 1)->entries)) {
 845 @@ -2579,8 +2927,32 @@
 846                         restart = 1;
 847                         goto journal_error;
 848                 }
 849 +       } else if (!ext4_htree_dx_locked(lck)) {
 850 +               struct ext4_dir_lock_data *ld = ext4_htree_lock_data(lck);
 851 +
 852 +               /* not well protected, require DX lock */
 853 +               ext4_htree_dx_need_lock(lck);
 854 +               at = frame > frames ? (frame - 1)->at : NULL;
 855 +
 856 +               /* NB: no risk of deadlock because it's just a try.
 857 +                *
 858 +                * NB: we check ld_count for twice, the first time before
 859 +                * having DX lock, the second time after holding DX lock.
 860 +                *
 861 +                * NB: We never free blocks for directory so far, which
 862 +                * means value returned by dx_get_count() should equal to
 863 +                * ld->ld_count if nobody split any DE-block under @at,
 864 +                * and ld->ld_at still points to valid dx_entry. */
 865 +               if ((ld->ld_count != dx_get_count(entries)) ||
 866 +                   !ext4_htree_dx_lock_try(lck, at) ||
 867 +                   (ld->ld_count != dx_get_count(entries))) {
 868 +                       restart = 1;
 869 +                       goto cleanup;
 870 +               }
 871 +               /* OK, I've got DX lock and nothing changed */
 872 +               frame->at = ld->ld_at;
 873         }
 874 -       de = do_split(handle, dir, &bh, frame, &fname->hinfo);
 875 +       de = do_split(handle, dir, &bh, frames, frame, &fname->hinfo, lck);
 876         if (IS_ERR(de)) {
 877                 err = PTR_ERR(de);
 878                 goto cleanup;
 879 @@ -2591,6 +2963,8 @@
 880  journal_error:
 881         ext4_std_error(dir->i_sb, err); /* this is a no-op if err == 0 */
 882  cleanup:
 883 +       ext4_htree_dx_unlock(lck);
 884 +       ext4_htree_de_unlock(lck);
 885         brelse(bh);
 886         dx_release(frames);
 887         /* @restart is true means htree-path has been changed, we need to
 888 diff -wur a/fs/ext4/super.c b/fs/ext4/super.c
 889 --- a/fs/ext4/super.c   2020-08-30 12:06:02.746523498 -0600
 890 +++ b/fs/ext4/super.c   2020-08-30 12:07:32.345927785 -0600
 891 @@ -1087,6 +1087,7 @@
 892
 893         inode_set_iversion(&ei->vfs_inode, 1);
 894         spin_lock_init(&ei->i_raw_lock);
 895 +       sema_init(&ei->i_append_sem, 1);
 896         INIT_LIST_HEAD(&ei->i_prealloc_list);
 897         spin_lock_init(&ei->i_prealloc_lock);
 898         ext4_es_init_tree(&ei->i_es_tree);