ldiskfs/kernel_patches/patches/rhel8.3/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-4.18.0-80.1.2.el8_0/fs/ext4/Makefile
  18 ===================================================================
  19 --- linux-4.18.0-80.1.2.el8_0.orig/fs/ext4/Makefile
  20 +++ linux-4.18.0-80.1.2.el8_0/fs/ext4/Makefile
  21 @@ -7,6 +7,7 @@ obj-$(CONFIG_EXT4_FS) += ext4.o
  22
  23  ext4-y := balloc.o bitmap.o block_validity.o dir.o ext4_jbd2.o extents.o \
  24                 extents_status.o file.o fsmap.o fsync.o hash.o ialloc.o \
  25 +               htree_lock.o \
  26                 indirect.o inline.o inode.o ioctl.o mballoc.o migrate.o \
  27                 mmp.o move_extent.o namei.o page-io.o readpage.o resize.o \
  28                 super.o symlink.o sysfs.o xattr.o xattr_trusted.o xattr_user.o
  29 Index: linux-4.18.0-240.1.1.el8/fs/ext4/ext4.h
  30 ===================================================================
  31 --- linux-4.18.0-240.1.1.el8.orig/fs/ext4/ext4.h
  32 +++ linux-4.18.0-240.1.1.el8/fs/ext4/ext4.h
  33 @@ -29,6 +29,7 @@
  34  #include <linux/timer.h>
  35  #include <linux/version.h>
  36  #include <linux/wait.h>
  37 +#include <linux/htree_lock.h>
  38  #include <linux/sched/signal.h>
  39  #include <linux/blockgroup_lock.h>
  40  #include <linux/percpu_counter.h>
  41 @@ -946,6 +947,9 @@ 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 @@ -2185,6 +2189,72 @@ 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  struct ext4_filename {
 122         const struct qstr *usr_fname;
 123         struct fscrypt_str disk_name;
 124 @@ -2487,11 +2557,19 @@ void ext4_insert_dentry(struct inode *in
 125                         struct ext4_filename *fname, void *data);
 126  static inline void ext4_update_dx_flag(struct inode *inode)
 127  {
 128 +       /* Disable it for ldiskfs, because going from a DX directory to
 129 +        * a non-DX directory while it is in use will completely break
 130 +        * the htree-locking.
 131 +        * If we really want to support this operation in the future,
 132 +        * we need to exclusively lock the directory at here which will
 133 +        * increase complexity of code */
 134 +#if 0
 135         if (!ext4_has_feature_dir_index(inode->i_sb)) {
 136                 /* ext4_iget() should have caught this... */
 137                 WARN_ON_ONCE(ext4_has_feature_metadata_csum(inode->i_sb));
 138                 ext4_clear_inode_flag(inode, EXT4_INODE_INDEX);
 139         }
 140 +#endif
 141  }
 142  static const unsigned char ext4_filetype_table[] = {
 143         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
 144 Index: linux-4.18.0-240.1.1.el8/fs/ext4/namei.c
 145 ===================================================================
 146 --- linux-4.18.0-240.1.1.el8.orig/fs/ext4/namei.c
 147 +++ linux-4.18.0-240.1.1.el8/fs/ext4/namei.c
 148 @@ -54,6 +54,7 @@ struct buffer_head *ext4_append(handle_t
 149                                         ext4_lblk_t *block)
 150  {
 151         struct buffer_head *bh;
 152 +       struct ext4_inode_info *ei = EXT4_I(inode);
 153         int err;
 154
 155         if (unlikely(EXT4_SB(inode->i_sb)->s_max_dir_size_kb &&
 156 @@ -61,15 +62,22 @@ struct buffer_head *ext4_append(handle_t
 157                       EXT4_SB(inode->i_sb)->s_max_dir_size_kb)))
 158                 return ERR_PTR(-ENOSPC);
 159
 160 +       /* with parallel dir operations all appends
 161 +       * have to be serialized -bzzz */
 162 +       down(&ei->i_append_sem);
 163 +
 164         *block = inode->i_size >> inode->i_sb->s_blocksize_bits;
 165
 166         bh = ext4_bread(handle, inode, *block, EXT4_GET_BLOCKS_CREATE);
 167 -       if (IS_ERR(bh))
 168 +       if (IS_ERR(bh)) {
 169 +               up(&ei->i_append_sem);
 170                 return bh;
 171 +       }
 172         inode->i_size += inode->i_sb->s_blocksize;
 173         EXT4_I(inode)->i_disksize = inode->i_size;
 174         BUFFER_TRACE(bh, "get_write_access");
 175         err = ext4_journal_get_write_access(handle, bh);
 176 +       up(&ei->i_append_sem);
 177         if (err) {
 178                 brelse(bh);
 179                 ext4_std_error(inode->i_sb, err);
 180 @@ -252,7 +260,8 @@ static unsigned dx_node_limit(struct ino
 181  static struct dx_frame *dx_probe(struct ext4_filename *fname,
 182                                  struct inode *dir,
 183                                  struct dx_hash_info *hinfo,
 184 -                                struct dx_frame *frame);
 185 +                                struct dx_frame *frame,
 186 +                                struct htree_lock *lck);
 187  static void dx_release(struct dx_frame *frames);
 188  static int dx_make_map(struct inode *dir, struct ext4_dir_entry_2 *de,
 189                        unsigned blocksize, struct dx_hash_info *hinfo,
 190 @@ -266,12 +275,13 @@ static void dx_insert_block(struct dx_fr
 191  static int ext4_htree_next_block(struct inode *dir, __u32 hash,
 192                                  struct dx_frame *frame,
 193                                  struct dx_frame *frames,
 194 -                                __u32 *start_hash);
 195 +                                __u32 *start_hash, struct htree_lock *lck);
 196  static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
 197                 struct ext4_filename *fname,
 198 -               struct ext4_dir_entry_2 **res_dir);
 199 +               struct ext4_dir_entry_2 **res_dir, struct htree_lock *lck);
 200  static int ext4_dx_add_entry(handle_t *handle, struct ext4_filename *fname,
 201 -                            struct inode *dir, struct inode *inode);
 202 +                            struct inode *dir, struct inode *inode,
 203 +                            struct htree_lock *lck);
 204
 205  /* checksumming functions */
 206  void initialize_dirent_tail(struct ext4_dir_entry_tail *t,
 207 @@ -735,6 +745,227 @@ struct stats dx_show_entries(struct dx_h
 208  }
 209  #endif /* DX_DEBUG */
 210
 211 +/* private data for htree_lock */
 212 +struct ext4_dir_lock_data {
 213 +       unsigned                ld_flags;  /* bits-map for lock types */
 214 +       unsigned                ld_count;  /* # entries of the last DX block */
 215 +       struct dx_entry         ld_at_entry; /* copy of leaf dx_entry */
 216 +       struct dx_entry         *ld_at;    /* position of leaf dx_entry */
 217 +};
 218 +
 219 +#define ext4_htree_lock_data(l)        ((struct ext4_dir_lock_data *)(l)->lk_private)
 220 +#define ext4_find_entry(dir, name, dirent, inline) \
 221 +                       __ext4_find_entry(dir, name, dirent, inline, NULL)
 222 +#define ext4_add_entry(handle, dentry, inode) \
 223 +                       __ext4_add_entry(handle, dentry, inode, NULL)
 224 +
 225 +/* NB: ext4_lblk_t is 32 bits so we use high bits to identify invalid blk */
 226 +#define EXT4_HTREE_NODE_CHANGED        (0xcafeULL << 32)
 227 +
 228 +static void ext4_htree_event_cb(void *target, void *event)
 229 +{
 230 +       u64 *block = (u64 *)target;
 231 +
 232 +       if (*block == dx_get_block((struct dx_entry *)event))
 233 +               *block = EXT4_HTREE_NODE_CHANGED;
 234 +}
 235 +
 236 +struct htree_lock_head *ext4_htree_lock_head_alloc(unsigned hbits)
 237 +{
 238 +       struct htree_lock_head *lhead;
 239 +
 240 +       lhead = htree_lock_head_alloc(EXT4_LK_MAX, hbits, 0);
 241 +       if (lhead != NULL) {
 242 +               htree_lock_event_attach(lhead, EXT4_LK_SPIN, HTREE_EVENT_WR,
 243 +                                       ext4_htree_event_cb);
 244 +       }
 245 +       return lhead;
 246 +}
 247 +EXPORT_SYMBOL(ext4_htree_lock_head_alloc);
 248 +
 249 +struct htree_lock *ext4_htree_lock_alloc(void)
 250 +{
 251 +       return htree_lock_alloc(EXT4_LK_MAX,
 252 +                               sizeof(struct ext4_dir_lock_data));
 253 +}
 254 +EXPORT_SYMBOL(ext4_htree_lock_alloc);
 255 +
 256 +static htree_lock_mode_t ext4_htree_mode(unsigned flags)
 257 +{
 258 +       switch (flags) {
 259 +       default: /* 0 or unknown flags require EX lock */
 260 +               return HTREE_LOCK_EX;
 261 +       case EXT4_HLOCK_READDIR:
 262 +               return HTREE_LOCK_PR;
 263 +       case EXT4_HLOCK_LOOKUP:
 264 +               return HTREE_LOCK_CR;
 265 +       case EXT4_HLOCK_DEL:
 266 +       case EXT4_HLOCK_ADD:
 267 +               return HTREE_LOCK_CW;
 268 +       }
 269 +}
 270 +
 271 +/* return PR for read-only operations, otherwise return EX */
 272 +static inline htree_lock_mode_t ext4_htree_safe_mode(unsigned flags)
 273 +{
 274 +       int writer = (flags & EXT4_LB_DE) == EXT4_LB_DE;
 275 +
 276 +       /* 0 requires EX lock */
 277 +       return (flags == 0 || writer) ? HTREE_LOCK_EX : HTREE_LOCK_PR;
 278 +}
 279 +
 280 +static int ext4_htree_safe_locked(struct htree_lock *lck)
 281 +{
 282 +       int writer;
 283 +
 284 +       if (lck == NULL || lck->lk_mode == HTREE_LOCK_EX)
 285 +               return 1;
 286 +
 287 +       writer = (ext4_htree_lock_data(lck)->ld_flags & EXT4_LB_DE) ==
 288 +                EXT4_LB_DE;
 289 +       if (writer) /* all readers & writers are excluded? */
 290 +               return lck->lk_mode == HTREE_LOCK_EX;
 291 +
 292 +       /* all writers are excluded? */
 293 +       return lck->lk_mode == HTREE_LOCK_PR ||
 294 +              lck->lk_mode == HTREE_LOCK_PW ||
 295 +              lck->lk_mode == HTREE_LOCK_EX;
 296 +}
 297 +
 298 +/* relock htree_lock with EX mode if it's change operation, otherwise
 299 + * relock it with PR mode. It's noop if PDO is disabled. */
 300 +static void ext4_htree_safe_relock(struct htree_lock *lck)
 301 +{
 302 +       if (!ext4_htree_safe_locked(lck)) {
 303 +               unsigned flags = ext4_htree_lock_data(lck)->ld_flags;
 304 +
 305 +               htree_change_lock(lck, ext4_htree_safe_mode(flags));
 306 +       }
 307 +}
 308 +
 309 +void ext4_htree_lock(struct htree_lock *lck, struct htree_lock_head *lhead,
 310 +                    struct inode *dir, unsigned flags)
 311 +{
 312 +       htree_lock_mode_t mode = is_dx(dir) ? ext4_htree_mode(flags) :
 313 +                                             ext4_htree_safe_mode(flags);
 314 +
 315 +       ext4_htree_lock_data(lck)->ld_flags = flags;
 316 +       htree_lock(lck, lhead, mode);
 317 +       if (!is_dx(dir))
 318 +               ext4_htree_safe_relock(lck); /* make sure it's safe locked */
 319 +}
 320 +EXPORT_SYMBOL(ext4_htree_lock);
 321 +
 322 +static int ext4_htree_node_lock(struct htree_lock *lck, struct dx_entry *at,
 323 +                               unsigned lmask, int wait, void *ev)
 324 +{
 325 +       u32     key = (at == NULL) ? 0 : dx_get_block(at);
 326 +       u32     mode;
 327 +
 328 +       /* NOOP if htree is well protected or caller doesn't require the lock */
 329 +       if (ext4_htree_safe_locked(lck) ||
 330 +          !(ext4_htree_lock_data(lck)->ld_flags & lmask))
 331 +               return 1;
 332 +
 333 +       mode = (ext4_htree_lock_data(lck)->ld_flags & lmask) == lmask ?
 334 +               HTREE_LOCK_PW : HTREE_LOCK_PR;
 335 +       while (1) {
 336 +               if (htree_node_lock_try(lck, mode, key, ffz(~lmask), wait, ev))
 337 +                       return 1;
 338 +               if (!(lmask & EXT4_LB_SPIN)) /* not a spinlock */
 339 +                       return 0;
 340 +               cpu_relax(); /* spin until granted */
 341 +       }
 342 +}
 343 +
 344 +static int ext4_htree_node_locked(struct htree_lock *lck, unsigned lmask)
 345 +{
 346 +       return ext4_htree_safe_locked(lck) ||
 347 +              htree_node_is_granted(lck, ffz(~lmask));
 348 +}
 349 +
 350 +static void ext4_htree_node_unlock(struct htree_lock *lck,
 351 +                                  unsigned lmask, void *buf)
 352 +{
 353 +       /* NB: it's safe to call mutiple times or even it's not locked */
 354 +       if (!ext4_htree_safe_locked(lck) &&
 355 +            htree_node_is_granted(lck, ffz(~lmask)))
 356 +               htree_node_unlock(lck, ffz(~lmask), buf);
 357 +}
 358 +
 359 +#define ext4_htree_dx_lock(lck, key)           \
 360 +       ext4_htree_node_lock(lck, key, EXT4_LB_DX, 1, NULL)
 361 +#define ext4_htree_dx_lock_try(lck, key)       \
 362 +       ext4_htree_node_lock(lck, key, EXT4_LB_DX, 0, NULL)
 363 +#define ext4_htree_dx_unlock(lck)              \
 364 +       ext4_htree_node_unlock(lck, EXT4_LB_DX, NULL)
 365 +#define ext4_htree_dx_locked(lck)              \
 366 +       ext4_htree_node_locked(lck, EXT4_LB_DX)
 367 +
 368 +static void ext4_htree_dx_need_lock(struct htree_lock *lck)
 369 +{
 370 +       struct ext4_dir_lock_data *ld;
 371 +
 372 +       if (ext4_htree_safe_locked(lck))
 373 +               return;
 374 +
 375 +       ld = ext4_htree_lock_data(lck);
 376 +       switch (ld->ld_flags) {
 377 +       default:
 378 +               return;
 379 +       case EXT4_HLOCK_LOOKUP:
 380 +               ld->ld_flags = EXT4_HLOCK_LOOKUP_SAFE;
 381 +               return;
 382 +       case EXT4_HLOCK_DEL:
 383 +               ld->ld_flags = EXT4_HLOCK_DEL_SAFE;
 384 +               return;
 385 +       case EXT4_HLOCK_ADD:
 386 +               ld->ld_flags = EXT4_HLOCK_SPLIT;
 387 +               return;
 388 +       }
 389 +}
 390 +
 391 +#define ext4_htree_de_lock(lck, key)           \
 392 +       ext4_htree_node_lock(lck, key, EXT4_LB_DE, 1, NULL)
 393 +#define ext4_htree_de_unlock(lck)              \
 394 +       ext4_htree_node_unlock(lck, EXT4_LB_DE, NULL)
 395 +
 396 +#define ext4_htree_spin_lock(lck, key, event)  \
 397 +       ext4_htree_node_lock(lck, key, EXT4_LB_SPIN, 0, event)
 398 +#define ext4_htree_spin_unlock(lck)            \
 399 +       ext4_htree_node_unlock(lck, EXT4_LB_SPIN, NULL)
 400 +#define ext4_htree_spin_unlock_listen(lck, p)  \
 401 +       ext4_htree_node_unlock(lck, EXT4_LB_SPIN, p)
 402 +
 403 +static void ext4_htree_spin_stop_listen(struct htree_lock *lck)
 404 +{
 405 +       if (!ext4_htree_safe_locked(lck) &&
 406 +           htree_node_is_listening(lck, ffz(~EXT4_LB_SPIN)))
 407 +               htree_node_stop_listen(lck, ffz(~EXT4_LB_SPIN));
 408 +}
 409 +
 410 +enum {
 411 +       DX_HASH_COL_IGNORE,     /* ignore collision while probing frames */
 412 +       DX_HASH_COL_YES,        /* there is collision and it does matter */
 413 +       DX_HASH_COL_NO,         /* there is no collision */
 414 +};
 415 +
 416 +static int dx_probe_hash_collision(struct htree_lock *lck,
 417 +                                  struct dx_entry *entries,
 418 +                                  struct dx_entry *at, u32 hash)
 419 +{
 420 +       if (!(lck && ext4_htree_lock_data(lck)->ld_flags & EXT4_LB_EXACT)) {
 421 +               return DX_HASH_COL_IGNORE; /* don't care about collision */
 422 +
 423 +       } else if (at == entries + dx_get_count(entries) - 1) {
 424 +               return DX_HASH_COL_IGNORE; /* not in any leaf of this DX */
 425 +
 426 +       } else { /* hash collision? */
 427 +               return ((dx_get_hash(at + 1) & ~1) == hash) ?
 428 +                       DX_HASH_COL_YES : DX_HASH_COL_NO;
 429 +       }
 430 +}
 431 +
 432  /*
 433   * Probe for a directory leaf block to search.
 434   *
 435 @@ -746,10 +977,11 @@ struct stats dx_show_entries(struct dx_h
 436   */
 437  static struct dx_frame *
 438  dx_probe(struct ext4_filename *fname, struct inode *dir,
 439 -        struct dx_hash_info *hinfo, struct dx_frame *frame_in)
 440 +        struct dx_hash_info *hinfo, struct dx_frame *frame_in,
 441 +        struct htree_lock *lck)
 442  {
 443         unsigned count, indirect;
 444 -       struct dx_entry *at, *entries, *p, *q, *m;
 445 +       struct dx_entry *at, *entries, *p, *q, *m, *dx = NULL;
 446         struct dx_root_info *info;
 447         struct dx_frame *frame = frame_in;
 448         struct dx_frame *ret_err = ERR_PTR(ERR_BAD_DX_DIR);
 449 @@ -811,8 +1043,15 @@ dx_probe(struct ext4_filename *fname, st
 450
 451         dxtrace(printk("Look up %x", hash));
 452         while (1) {
 453 +               if (indirect == 0) { /* the last index level */
 454 +                       /* NB: ext4_htree_dx_lock() could be noop if
 455 +                        * DX-lock flag is not set for current operation */
 456 +                       ext4_htree_dx_lock(lck, dx);
 457 +                       ext4_htree_spin_lock(lck, dx, NULL);
 458 +               }
 459                 count = dx_get_count(entries);
 460 -               if (!count || count > dx_get_limit(entries)) {
 461 +               if (count == 0 || count > dx_get_limit(entries)) {
 462 +                       ext4_htree_spin_unlock(lck); /* release spin */
 463                         ext4_warning_inode(dir,
 464                                            "dx entry: count %u beyond limit %u",
 465                                            count, dx_get_limit(entries));
 466 @@ -851,8 +1090,70 @@ dx_probe(struct ext4_filename *fname, st
 467                                dx_get_block(at)));
 468                 frame->entries = entries;
 469                 frame->at = at;
 470 -               if (!indirect--)
 471 +
 472 +               if (indirect == 0) { /* the last index level */
 473 +                       struct ext4_dir_lock_data *ld;
 474 +                       u64 myblock;
 475 +
 476 +                       /* By default we only lock DE-block, however, we will
 477 +                        * also lock the last level DX-block if:
 478 +                        * a) there is hash collision
 479 +                        *    we will set DX-lock flag (a few lines below)
 480 +                        *    and redo to lock DX-block
 481 +                        *    see detail in dx_probe_hash_collision()
 482 +                        * b) it's a retry from splitting
 483 +                        *    we need to lock the last level DX-block so nobody
 484 +                        *    else can split any leaf blocks under the same
 485 +                        *    DX-block, see detail in ext4_dx_add_entry()
 486 +                        */
 487 +                       if (ext4_htree_dx_locked(lck)) {
 488 +                               /* DX-block is locked, just lock DE-block
 489 +                                * and return */
 490 +                               ext4_htree_spin_unlock(lck);
 491 +                               if (!ext4_htree_safe_locked(lck))
 492 +                                       ext4_htree_de_lock(lck, frame->at);
 493 +                               return frame;
 494 +                       }
 495 +                       /* it's pdirop and no DX lock */
 496 +                       if (dx_probe_hash_collision(lck, entries, at, hash) ==
 497 +                           DX_HASH_COL_YES) {
 498 +                               /* found hash collision, set DX-lock flag
 499 +                                * and retry to abtain DX-lock */
 500 +                               ext4_htree_spin_unlock(lck);
 501 +                               ext4_htree_dx_need_lock(lck);
 502 +                               continue;
 503 +                       }
 504 +                       ld = ext4_htree_lock_data(lck);
 505 +                       /* because I don't lock DX, so @at can't be trusted
 506 +                        * after I release spinlock so I have to save it */
 507 +                       ld->ld_at = at;
 508 +                       ld->ld_at_entry = *at;
 509 +                       ld->ld_count = dx_get_count(entries);
 510 +
 511 +                       frame->at = &ld->ld_at_entry;
 512 +                       myblock = dx_get_block(at);
 513 +
 514 +                       /* NB: ordering locking */
 515 +                       ext4_htree_spin_unlock_listen(lck, &myblock);
 516 +                       /* other thread can split this DE-block because:
 517 +                        * a) I don't have lock for the DE-block yet
 518 +                        * b) I released spinlock on DX-block
 519 +                        * if it happened I can detect it by listening
 520 +                        * splitting event on this DE-block */
 521 +                       ext4_htree_de_lock(lck, frame->at);
 522 +                       ext4_htree_spin_stop_listen(lck);
 523 +
 524 +                       if (myblock == EXT4_HTREE_NODE_CHANGED) {
 525 +                               /* someone split this DE-block before
 526 +                                * I locked it, I need to retry and lock
 527 +                                * valid DE-block */
 528 +                               ext4_htree_de_unlock(lck);
 529 +                               continue;
 530 +                       }
 531                         return frame;
 532 +               }
 533 +               dx = at;
 534 +               indirect--;
 535                 frame++;
 536                 frame->bh = ext4_read_dirblock(dir, dx_get_block(at), INDEX);
 537                 if (IS_ERR(frame->bh)) {
 538 @@ -921,7 +1222,7 @@ static void dx_release(struct dx_frame *
 539  static int ext4_htree_next_block(struct inode *dir, __u32 hash,
 540                                  struct dx_frame *frame,
 541                                  struct dx_frame *frames,
 542 -                                __u32 *start_hash)
 543 +                                __u32 *start_hash, struct htree_lock *lck)
 544  {
 545         struct dx_frame *p;
 546         struct buffer_head *bh;
 547 @@ -936,12 +1237,22 @@ static int ext4_htree_next_block(struct
 548          * this loop, num_frames indicates the number of interior
 549          * nodes need to be read.
 550          */
 551 +       ext4_htree_de_unlock(lck);
 552         while (1) {
 553 -               if (++(p->at) < p->entries + dx_get_count(p->entries))
 554 -                       break;
 555 +               if (num_frames > 0 || ext4_htree_dx_locked(lck)) {
 556 +                       /* num_frames > 0 :
 557 +                        *   DX block
 558 +                        * ext4_htree_dx_locked:
 559 +                        *   frame->at is reliable pointer returned by dx_probe,
 560 +                        *   otherwise dx_probe already knew no collision */
 561 +                       if (++(p->at) < p->entries + dx_get_count(p->entries))
 562 +                               break;
 563 +               }
 564                 if (p == frames)
 565                         return 0;
 566                 num_frames++;
 567 +               if (num_frames == 1)
 568 +                       ext4_htree_dx_unlock(lck);
 569                 p--;
 570         }
 571
 572 @@ -964,6 +1275,13 @@ static int ext4_htree_next_block(struct
 573          * block so no check is necessary
 574          */
 575         while (num_frames--) {
 576 +               if (num_frames == 0) {
 577 +                       /* it's not always necessary, we just don't want to
 578 +                        * detect hash collision again */
 579 +                       ext4_htree_dx_need_lock(lck);
 580 +                       ext4_htree_dx_lock(lck, p->at);
 581 +               }
 582 +
 583                 bh = ext4_read_dirblock(dir, dx_get_block(p->at), INDEX);
 584                 if (IS_ERR(bh))
 585                         return PTR_ERR(bh);
 586 @@ -972,6 +1290,7 @@ static int ext4_htree_next_block(struct
 587                 p->bh = bh;
 588                 p->at = p->entries = ((struct dx_node *) bh->b_data)->entries;
 589         }
 590 +       ext4_htree_de_lock(lck, p->at);
 591         return 1;
 592  }
 593
 594 @@ -1119,10 +1438,10 @@ int ext4_htree_fill_tree(struct file *di
 595         }
 596         hinfo.hash = start_hash;
 597         hinfo.minor_hash = 0;
 598 -       frame = dx_probe(NULL, dir, &hinfo, frames);
 599 +       /* assume it's PR locked */
 600 +       frame = dx_probe(NULL, dir, &hinfo, frames, NULL);
 601         if (IS_ERR(frame))
 602                 return PTR_ERR(frame);
 603 -
 604         /* Add '.' and '..' from the htree header */
 605         if (!start_hash && !start_minor_hash) {
 606                 de = (struct ext4_dir_entry_2 *) frames[0].bh->b_data;
 607 @@ -1162,7 +1481,7 @@ int ext4_htree_fill_tree(struct file *di
 608                 count += ret;
 609                 hashval = ~0;
 610                 ret = ext4_htree_next_block(dir, HASH_NB_ALWAYS,
 611 -                                           frame, frames, &hashval);
 612 +                                           frame, frames, &hashval, NULL);
 613                 *next_hash = hashval;
 614                 if (ret < 0) {
 615                         err = ret;
 616 @@ -1354,10 +1673,10 @@ static int is_dx_internal_node(struct in
 617   * The returned buffer_head has ->b_count elevated.  The caller is expected
 618   * to brelse() it when appropriate.
 619   */
 620 -static struct buffer_head * ext4_find_entry (struct inode *dir,
 621 +struct buffer_head *__ext4_find_entry(struct inode *dir,
 622                                         const struct qstr *d_name,
 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 @@ -1406,7 +1725,7 @@ static struct buffer_head * ext4_find_en
 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 @@ -1416,6 +1735,7 @@ static struct buffer_head * ext4_find_en
 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 @@ -1507,10 +1827,12 @@ cleanup_and_exit:
 647         ext4_fname_free_filename(&fname);
 648         return ret;
 649  }
 650 +EXPORT_SYMBOL(__ext4_find_entry);
 651
 652  static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
 653                         struct ext4_filename *fname,
 654 -                       struct ext4_dir_entry_2 **res_dir)
 655 +                       struct ext4_dir_entry_2 **res_dir,
 656 +                       struct htree_lock *lck)
 657  {
 658         struct super_block * sb = dir->i_sb;
 659         struct dx_frame frames[EXT4_HTREE_LEVEL], *frame;
 660 @@ -1521,7 +1843,7 @@ static struct buffer_head * ext4_dx_find
 661  #ifdef CONFIG_EXT4_FS_ENCRYPTION
 662         *res_dir = NULL;
 663  #endif
 664 -       frame = dx_probe(fname, dir, NULL, frames);
 665 +       frame = dx_probe(fname, dir, NULL, frames, lck);
 666         if (IS_ERR(frame))
 667                 return (struct buffer_head *) frame;
 668         do {
 669 @@ -1543,7 +1865,7 @@ static struct buffer_head * ext4_dx_find
 670
 671                 /* Check to see if we should continue to search */
 672                 retval = ext4_htree_next_block(dir, fname->hinfo.hash, frame,
 673 -                                              frames, NULL);
 674 +                                              frames, NULL, lck);
 675                 if (retval < 0) {
 676                         ext4_warning_inode(dir,
 677                                 "error %d reading directory index block",
 678 @@ -1718,8 +2040,9 @@ static struct ext4_dir_entry_2* dx_pack_
 679   * Returns pointer to de in block into which the new entry will be inserted.
 680   */
 681  static struct ext4_dir_entry_2 *do_split(handle_t *handle, struct inode *dir,
 682 -                       struct buffer_head **bh,struct dx_frame *frame,
 683 -                       struct dx_hash_info *hinfo)
 684 +                       struct buffer_head **bh, struct dx_frame *frames,
 685 +                       struct dx_frame *frame, struct dx_hash_info *hinfo,
 686 +                       struct htree_lock *lck)
 687  {
 688         unsigned blocksize = dir->i_sb->s_blocksize;
 689         unsigned count, continued;
 690 @@ -1781,8 +2104,14 @@ static struct ext4_dir_entry_2 *do_split
 691                                         hash2, split, count-split));
 692
 693         /* Fancy dance to stay within two buffers */
 694 -       de2 = dx_move_dirents(data1, data2, map + split, count - split,
 695 -                             blocksize);
 696 +       if (hinfo->hash < hash2) {
 697 +               de2 = dx_move_dirents(data1, data2, map + split,
 698 +                                     count - split, blocksize);
 699 +       } else {
 700 +               /* make sure we will add entry to the same block which
 701 +                * we have already locked */
 702 +               de2 = dx_move_dirents(data1, data2, map, split, blocksize);
 703 +       }
 704         de = dx_pack_dirents(data1, blocksize);
 705         de->rec_len = ext4_rec_len_to_disk(data1 + (blocksize - csum_size) -
 706                                            (char *) de,
 707 @@ -1803,12 +2132,21 @@ static struct ext4_dir_entry_2 *do_split
 708         dxtrace(dx_show_leaf(dir, hinfo, (struct ext4_dir_entry_2 *) data2,
 709                         blocksize, 1));
 710
 711 -       /* Which block gets the new entry? */
 712 -       if (hinfo->hash >= hash2) {
 713 -               swap(*bh, bh2);
 714 -               de = de2;
 715 +       ext4_htree_spin_lock(lck, frame > frames ? (frame - 1)->at : NULL,
 716 +                            frame->at); /* notify block is being split */
 717 +       if (hinfo->hash < hash2) {
 718 +               dx_insert_block(frame, hash2 + continued, newblock);
 719 +
 720 +       } else {
 721 +               /* switch block number */
 722 +               dx_insert_block(frame, hash2 + continued,
 723 +                               dx_get_block(frame->at));
 724 +               dx_set_block(frame->at, newblock);
 725 +               (frame->at)++;
 726         }
 727 -       dx_insert_block(frame, hash2 + continued, newblock);
 728 +       ext4_htree_spin_unlock(lck);
 729 +       ext4_htree_dx_unlock(lck);
 730 +
 731         err = ext4_handle_dirty_dirent_node(handle, dir, bh2);
 732         if (err)
 733                 goto journal_error;
 734 @@ -2082,7 +2420,7 @@ static int make_indexed_dir(handle_t *ha
 735         if (retval)
 736                 goto out_frames;
 737
 738 -       de = do_split(handle,dir, &bh2, frame, &fname->hinfo);
 739 +       de = do_split(handle, dir, &bh2, frames, frame, &fname->hinfo, NULL);
 740         if (IS_ERR(de)) {
 741                 retval = PTR_ERR(de);
 742                 goto out_frames;
 743 @@ -2192,8 +2530,8 @@ out:
 744   * may not sleep between calling this and putting something into
 745   * the entry, as someone else might have used it while you slept.
 746   */
 747 -static int ext4_add_entry(handle_t *handle, struct dentry *dentry,
 748 -                         struct inode *inode)
 749 +int __ext4_add_entry(handle_t *handle, struct dentry *dentry,
 750 +                     struct inode *inode, struct htree_lock *lck)
 751  {
 752         struct inode *dir = d_inode(dentry->d_parent);
 753         struct buffer_head *bh = NULL;
 754 @@ -2234,9 +2572,10 @@ static int ext4_add_entry(handle_t *hand
 755                 if (dentry->d_name.len == 2 &&
 756                     memcmp(dentry->d_name.name, "..", 2) == 0)
 757                         return ext4_update_dotdot(handle, dentry, inode);
 758 -               retval = ext4_dx_add_entry(handle, &fname, dir, inode);
 759 +               retval = ext4_dx_add_entry(handle, &fname, dir, inode, lck);
 760                 if (!retval || (retval != ERR_BAD_DX_DIR))
 761                         goto out;
 762 +               ext4_htree_safe_relock(lck);
 763                 /* Can we just ignore htree data? */
 764                 if (ext4_has_metadata_csum(sb)) {
 765                         EXT4_ERROR_INODE(dir,
 766 @@ -2293,12 +2632,14 @@ out:
 767                 ext4_set_inode_state(inode, EXT4_STATE_NEWENTRY);
 768         return retval;
 769  }
 770 +EXPORT_SYMBOL(__ext4_add_entry);
 771
 772  /*
 773   * Returns 0 for success, or a negative error value
 774   */
 775  static int ext4_dx_add_entry(handle_t *handle, struct ext4_filename *fname,
 776 -                            struct inode *dir, struct inode *inode)
 777 +                            struct inode *dir, struct inode *inode,
 778 +                            struct htree_lock *lck)
 779  {
 780         struct dx_frame frames[EXT4_HTREE_LEVEL], *frame;
 781         struct dx_entry *entries, *at;
 782 @@ -2310,7 +2651,7 @@ static int ext4_dx_add_entry(handle_t *h
 783
 784  again:
 785         restart = 0;
 786 -       frame = dx_probe(fname, dir, NULL, frames);
 787 +       frame = dx_probe(fname, dir, NULL, frames, lck);
 788         if (IS_ERR(frame))
 789                 return PTR_ERR(frame);
 790         entries = frame->entries;
 791 @@ -2345,6 +2686,11 @@ again:
 792                 struct dx_node *node2;
 793                 struct buffer_head *bh2;
 794
 795 +               if (!ext4_htree_safe_locked(lck)) { /* retry with EX lock */
 796 +                       ext4_htree_safe_relock(lck);
 797 +                       restart = 1;
 798 +                       goto cleanup;
 799 +               }
 800                 while (frame > frames) {
 801                         if (dx_get_count((frame - 1)->entries) <
 802                             dx_get_limit((frame - 1)->entries)) {
 803 @@ -2447,8 +2793,32 @@ again:
 804                         restart = 1;
 805                         goto journal_error;
 806                 }
 807 +       } else if (!ext4_htree_dx_locked(lck)) {
 808 +               struct ext4_dir_lock_data *ld = ext4_htree_lock_data(lck);
 809 +
 810 +               /* not well protected, require DX lock */
 811 +               ext4_htree_dx_need_lock(lck);
 812 +               at = frame > frames ? (frame - 1)->at : NULL;
 813 +
 814 +               /* NB: no risk of deadlock because it's just a try.
 815 +                *
 816 +                * NB: we check ld_count for twice, the first time before
 817 +                * having DX lock, the second time after holding DX lock.
 818 +                *
 819 +                * NB: We never free blocks for directory so far, which
 820 +                * means value returned by dx_get_count() should equal to
 821 +                * ld->ld_count if nobody split any DE-block under @at,
 822 +                * and ld->ld_at still points to valid dx_entry. */
 823 +               if ((ld->ld_count != dx_get_count(entries)) ||
 824 +                   !ext4_htree_dx_lock_try(lck, at) ||
 825 +                   (ld->ld_count != dx_get_count(entries))) {
 826 +                       restart = 1;
 827 +                       goto cleanup;
 828 +               }
 829 +               /* OK, I've got DX lock and nothing changed */
 830 +               frame->at = ld->ld_at;
 831         }
 832 -       de = do_split(handle, dir, &bh, frame, &fname->hinfo);
 833 +       de = do_split(handle, dir, &bh, frames, frame, &fname->hinfo, lck);
 834         if (IS_ERR(de)) {
 835                 err = PTR_ERR(de);
 836                 goto cleanup;
 837 @@ -2459,6 +2829,8 @@ again:
 838  journal_error:
 839         ext4_std_error(dir->i_sb, err); /* this is a no-op if err == 0 */
 840  cleanup:
 841 +       ext4_htree_dx_unlock(lck);
 842 +       ext4_htree_de_unlock(lck);
 843         brelse(bh);
 844         dx_release(frames);
 845         /* @restart is true means htree-path has been changed, we need to
 846 Index: linux-4.18.0-80.1.2.el8_0/fs/ext4/super.c
 847 ===================================================================
 848 --- linux-4.18.0-80.1.2.el8_0.orig/fs/ext4/super.c
 849 +++ linux-4.18.0-80.1.2.el8_0/fs/ext4/super.c
 850 @@ -1009,6 +1009,7 @@ static struct inode *ext4_alloc_inode(st
 851
 852         inode_set_iversion(&ei->vfs_inode, 1);
 853         spin_lock_init(&ei->i_raw_lock);
 854 +       sema_init(&ei->i_append_sem, 1);
 855         INIT_LIST_HEAD(&ei->i_prealloc_list);
 856         spin_lock_init(&ei->i_prealloc_lock);
 857         ext4_es_init_tree(&ei->i_es_tree);