ldiskfs/kernel_patches/patches/rhel8/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-4.18.0-80.1.2.el8_0/fs/ext4/ext4.h
  30 ===================================================================
  31 --- linux-4.18.0-80.1.2.el8_0.orig/fs/ext4/ext4.h
  32 +++ linux-4.18.0-80.1.2.el8_0/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 @@ -934,6 +935,9 @@ 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 @@ -2109,6 +2113,72 @@ 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 @@ -2416,8 +2486,16 @@ 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_clear_inode_flag(inode, EXT4_INODE_INDEX);
 137 +#endif
 138  }
 139  static const unsigned char ext4_filetype_table[] = {
 140         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
 141 Index: linux-4.18.0-80.1.2.el8_0/fs/ext4/namei.c
 142 ===================================================================
 143 --- linux-4.18.0-80.1.2.el8_0.orig/fs/ext4/namei.c
 144 +++ linux-4.18.0-80.1.2.el8_0/fs/ext4/namei.c
 145 @@ -54,6 +54,7 @@ struct buffer_head *ext4_append(handle_t
 146                                         ext4_lblk_t *block)
 147  {
 148         struct buffer_head *bh;
 149 +       struct ext4_inode_info *ei = EXT4_I(inode);
 150         int err;
 151
 152         if (unlikely(EXT4_SB(inode->i_sb)->s_max_dir_size_kb &&
 153 @@ -61,15 +62,22 @@ struct buffer_head *ext4_append(handle_t
 154                       EXT4_SB(inode->i_sb)->s_max_dir_size_kb)))
 155                 return ERR_PTR(-ENOSPC);
 156
 157 +       /* with parallel dir operations all appends
 158 +       * have to be serialized -bzzz */
 159 +       down(&ei->i_append_sem);
 160 +
 161         *block = inode->i_size >> inode->i_sb->s_blocksize_bits;
 162
 163         bh = ext4_bread(handle, inode, *block, EXT4_GET_BLOCKS_CREATE);
 164 -       if (IS_ERR(bh))
 165 +       if (IS_ERR(bh)) {
 166 +               up(&ei->i_append_sem);
 167                 return bh;
 168 +       }
 169         inode->i_size += inode->i_sb->s_blocksize;
 170         EXT4_I(inode)->i_disksize = inode->i_size;
 171         BUFFER_TRACE(bh, "get_write_access");
 172         err = ext4_journal_get_write_access(handle, bh);
 173 +       up(&ei->i_append_sem);
 174         if (err) {
 175                 brelse(bh);
 176                 ext4_std_error(inode->i_sb, err);
 177 @@ -250,7 +258,8 @@ static unsigned dx_node_limit(struct ino
 178  static struct dx_frame *dx_probe(struct ext4_filename *fname,
 179                                  struct inode *dir,
 180                                  struct dx_hash_info *hinfo,
 181 -                                struct dx_frame *frame);
 182 +                                struct dx_frame *frame,
 183 +                                struct htree_lock *lck);
 184  static void dx_release(struct dx_frame *frames);
 185  static int dx_make_map(struct inode *dir, struct ext4_dir_entry_2 *de,
 186                        unsigned blocksize, struct dx_hash_info *hinfo,
 187 @@ -264,12 +273,13 @@ static void dx_insert_block(struct dx_fr
 188  static int ext4_htree_next_block(struct inode *dir, __u32 hash,
 189                                  struct dx_frame *frame,
 190                                  struct dx_frame *frames,
 191 -                                __u32 *start_hash);
 192 +                                __u32 *start_hash, struct htree_lock *lck);
 193  static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
 194                 struct ext4_filename *fname,
 195 -               struct ext4_dir_entry_2 **res_dir);
 196 +               struct ext4_dir_entry_2 **res_dir, struct htree_lock *lck);
 197  static int ext4_dx_add_entry(handle_t *handle, struct ext4_filename *fname,
 198 -                            struct inode *dir, struct inode *inode);
 199 +                            struct inode *dir, struct inode *inode,
 200 +                            struct htree_lock *lck);
 201
 202  /* checksumming functions */
 203  void initialize_dirent_tail(struct ext4_dir_entry_tail *t,
 204 @@ -733,6 +743,227 @@ struct stats dx_show_entries(struct dx_h
 205  }
 206  #endif /* DX_DEBUG */
 207
 208 +/* private data for htree_lock */
 209 +struct ext4_dir_lock_data {
 210 +       unsigned                ld_flags;  /* bits-map for lock types */
 211 +       unsigned                ld_count;  /* # entries of the last DX block */
 212 +       struct dx_entry         ld_at_entry; /* copy of leaf dx_entry */
 213 +       struct dx_entry         *ld_at;    /* position of leaf dx_entry */
 214 +};
 215 +
 216 +#define ext4_htree_lock_data(l)        ((struct ext4_dir_lock_data *)(l)->lk_private)
 217 +#define ext4_find_entry(dir, name, dirent, inline) \
 218 +                       __ext4_find_entry(dir, name, dirent, inline, NULL)
 219 +#define ext4_add_entry(handle, dentry, inode) \
 220 +                       __ext4_add_entry(handle, dentry, inode, NULL)
 221 +
 222 +/* NB: ext4_lblk_t is 32 bits so we use high bits to identify invalid blk */
 223 +#define EXT4_HTREE_NODE_CHANGED        (0xcafeULL << 32)
 224 +
 225 +static void ext4_htree_event_cb(void *target, void *event)
 226 +{
 227 +       u64 *block = (u64 *)target;
 228 +
 229 +       if (*block == dx_get_block((struct dx_entry *)event))
 230 +               *block = EXT4_HTREE_NODE_CHANGED;
 231 +}
 232 +
 233 +struct htree_lock_head *ext4_htree_lock_head_alloc(unsigned hbits)
 234 +{
 235 +       struct htree_lock_head *lhead;
 236 +
 237 +       lhead = htree_lock_head_alloc(EXT4_LK_MAX, hbits, 0);
 238 +       if (lhead != NULL) {
 239 +               htree_lock_event_attach(lhead, EXT4_LK_SPIN, HTREE_EVENT_WR,
 240 +                                       ext4_htree_event_cb);
 241 +       }
 242 +       return lhead;
 243 +}
 244 +EXPORT_SYMBOL(ext4_htree_lock_head_alloc);
 245 +
 246 +struct htree_lock *ext4_htree_lock_alloc(void)
 247 +{
 248 +       return htree_lock_alloc(EXT4_LK_MAX,
 249 +                               sizeof(struct ext4_dir_lock_data));
 250 +}
 251 +EXPORT_SYMBOL(ext4_htree_lock_alloc);
 252 +
 253 +static htree_lock_mode_t ext4_htree_mode(unsigned flags)
 254 +{
 255 +       switch (flags) {
 256 +       default: /* 0 or unknown flags require EX lock */
 257 +               return HTREE_LOCK_EX;
 258 +       case EXT4_HLOCK_READDIR:
 259 +               return HTREE_LOCK_PR;
 260 +       case EXT4_HLOCK_LOOKUP:
 261 +               return HTREE_LOCK_CR;
 262 +       case EXT4_HLOCK_DEL:
 263 +       case EXT4_HLOCK_ADD:
 264 +               return HTREE_LOCK_CW;
 265 +       }
 266 +}
 267 +
 268 +/* return PR for read-only operations, otherwise return EX */
 269 +static inline htree_lock_mode_t ext4_htree_safe_mode(unsigned flags)
 270 +{
 271 +       int writer = (flags & EXT4_LB_DE) == EXT4_LB_DE;
 272 +
 273 +       /* 0 requires EX lock */
 274 +       return (flags == 0 || writer) ? HTREE_LOCK_EX : HTREE_LOCK_PR;
 275 +}
 276 +
 277 +static int ext4_htree_safe_locked(struct htree_lock *lck)
 278 +{
 279 +       int writer;
 280 +
 281 +       if (lck == NULL || lck->lk_mode == HTREE_LOCK_EX)
 282 +               return 1;
 283 +
 284 +       writer = (ext4_htree_lock_data(lck)->ld_flags & EXT4_LB_DE) ==
 285 +                EXT4_LB_DE;
 286 +       if (writer) /* all readers & writers are excluded? */
 287 +               return lck->lk_mode == HTREE_LOCK_EX;
 288 +
 289 +       /* all writers are excluded? */
 290 +       return lck->lk_mode == HTREE_LOCK_PR ||
 291 +              lck->lk_mode == HTREE_LOCK_PW ||
 292 +              lck->lk_mode == HTREE_LOCK_EX;
 293 +}
 294 +
 295 +/* relock htree_lock with EX mode if it's change operation, otherwise
 296 + * relock it with PR mode. It's noop if PDO is disabled. */
 297 +static void ext4_htree_safe_relock(struct htree_lock *lck)
 298 +{
 299 +       if (!ext4_htree_safe_locked(lck)) {
 300 +               unsigned flags = ext4_htree_lock_data(lck)->ld_flags;
 301 +
 302 +               htree_change_lock(lck, ext4_htree_safe_mode(flags));
 303 +       }
 304 +}
 305 +
 306 +void ext4_htree_lock(struct htree_lock *lck, struct htree_lock_head *lhead,
 307 +                    struct inode *dir, unsigned flags)
 308 +{
 309 +       htree_lock_mode_t mode = is_dx(dir) ? ext4_htree_mode(flags) :
 310 +                                             ext4_htree_safe_mode(flags);
 311 +
 312 +       ext4_htree_lock_data(lck)->ld_flags = flags;
 313 +       htree_lock(lck, lhead, mode);
 314 +       if (!is_dx(dir))
 315 +               ext4_htree_safe_relock(lck); /* make sure it's safe locked */
 316 +}
 317 +EXPORT_SYMBOL(ext4_htree_lock);
 318 +
 319 +static int ext4_htree_node_lock(struct htree_lock *lck, struct dx_entry *at,
 320 +                               unsigned lmask, int wait, void *ev)
 321 +{
 322 +       u32     key = (at == NULL) ? 0 : dx_get_block(at);
 323 +       u32     mode;
 324 +
 325 +       /* NOOP if htree is well protected or caller doesn't require the lock */
 326 +       if (ext4_htree_safe_locked(lck) ||
 327 +          !(ext4_htree_lock_data(lck)->ld_flags & lmask))
 328 +               return 1;
 329 +
 330 +       mode = (ext4_htree_lock_data(lck)->ld_flags & lmask) == lmask ?
 331 +               HTREE_LOCK_PW : HTREE_LOCK_PR;
 332 +       while (1) {
 333 +               if (htree_node_lock_try(lck, mode, key, ffz(~lmask), wait, ev))
 334 +                       return 1;
 335 +               if (!(lmask & EXT4_LB_SPIN)) /* not a spinlock */
 336 +                       return 0;
 337 +               cpu_relax(); /* spin until granted */
 338 +       }
 339 +}
 340 +
 341 +static int ext4_htree_node_locked(struct htree_lock *lck, unsigned lmask)
 342 +{
 343 +       return ext4_htree_safe_locked(lck) ||
 344 +              htree_node_is_granted(lck, ffz(~lmask));
 345 +}
 346 +
 347 +static void ext4_htree_node_unlock(struct htree_lock *lck,
 348 +                                  unsigned lmask, void *buf)
 349 +{
 350 +       /* NB: it's safe to call mutiple times or even it's not locked */
 351 +       if (!ext4_htree_safe_locked(lck) &&
 352 +            htree_node_is_granted(lck, ffz(~lmask)))
 353 +               htree_node_unlock(lck, ffz(~lmask), buf);
 354 +}
 355 +
 356 +#define ext4_htree_dx_lock(lck, key)           \
 357 +       ext4_htree_node_lock(lck, key, EXT4_LB_DX, 1, NULL)
 358 +#define ext4_htree_dx_lock_try(lck, key)       \
 359 +       ext4_htree_node_lock(lck, key, EXT4_LB_DX, 0, NULL)
 360 +#define ext4_htree_dx_unlock(lck)              \
 361 +       ext4_htree_node_unlock(lck, EXT4_LB_DX, NULL)
 362 +#define ext4_htree_dx_locked(lck)              \
 363 +       ext4_htree_node_locked(lck, EXT4_LB_DX)
 364 +
 365 +static void ext4_htree_dx_need_lock(struct htree_lock *lck)
 366 +{
 367 +       struct ext4_dir_lock_data *ld;
 368 +
 369 +       if (ext4_htree_safe_locked(lck))
 370 +               return;
 371 +
 372 +       ld = ext4_htree_lock_data(lck);
 373 +       switch (ld->ld_flags) {
 374 +       default:
 375 +               return;
 376 +       case EXT4_HLOCK_LOOKUP:
 377 +               ld->ld_flags = EXT4_HLOCK_LOOKUP_SAFE;
 378 +               return;
 379 +       case EXT4_HLOCK_DEL:
 380 +               ld->ld_flags = EXT4_HLOCK_DEL_SAFE;
 381 +               return;
 382 +       case EXT4_HLOCK_ADD:
 383 +               ld->ld_flags = EXT4_HLOCK_SPLIT;
 384 +               return;
 385 +       }
 386 +}
 387 +
 388 +#define ext4_htree_de_lock(lck, key)           \
 389 +       ext4_htree_node_lock(lck, key, EXT4_LB_DE, 1, NULL)
 390 +#define ext4_htree_de_unlock(lck)              \
 391 +       ext4_htree_node_unlock(lck, EXT4_LB_DE, NULL)
 392 +
 393 +#define ext4_htree_spin_lock(lck, key, event)  \
 394 +       ext4_htree_node_lock(lck, key, EXT4_LB_SPIN, 0, event)
 395 +#define ext4_htree_spin_unlock(lck)            \
 396 +       ext4_htree_node_unlock(lck, EXT4_LB_SPIN, NULL)
 397 +#define ext4_htree_spin_unlock_listen(lck, p)  \
 398 +       ext4_htree_node_unlock(lck, EXT4_LB_SPIN, p)
 399 +
 400 +static void ext4_htree_spin_stop_listen(struct htree_lock *lck)
 401 +{
 402 +       if (!ext4_htree_safe_locked(lck) &&
 403 +           htree_node_is_listening(lck, ffz(~EXT4_LB_SPIN)))
 404 +               htree_node_stop_listen(lck, ffz(~EXT4_LB_SPIN));
 405 +}
 406 +
 407 +enum {
 408 +       DX_HASH_COL_IGNORE,     /* ignore collision while probing frames */
 409 +       DX_HASH_COL_YES,        /* there is collision and it does matter */
 410 +       DX_HASH_COL_NO,         /* there is no collision */
 411 +};
 412 +
 413 +static int dx_probe_hash_collision(struct htree_lock *lck,
 414 +                                  struct dx_entry *entries,
 415 +                                  struct dx_entry *at, u32 hash)
 416 +{
 417 +       if (!(lck && ext4_htree_lock_data(lck)->ld_flags & EXT4_LB_EXACT)) {
 418 +               return DX_HASH_COL_IGNORE; /* don't care about collision */
 419 +
 420 +       } else if (at == entries + dx_get_count(entries) - 1) {
 421 +               return DX_HASH_COL_IGNORE; /* not in any leaf of this DX */
 422 +
 423 +       } else { /* hash collision? */
 424 +               return ((dx_get_hash(at + 1) & ~1) == hash) ?
 425 +                       DX_HASH_COL_YES : DX_HASH_COL_NO;
 426 +       }
 427 +}
 428 +
 429  /*
 430   * Probe for a directory leaf block to search.
 431   *
 432 @@ -744,10 +975,11 @@ struct stats dx_show_entries(struct dx_h
 433   */
 434  static struct dx_frame *
 435  dx_probe(struct ext4_filename *fname, struct inode *dir,
 436 -        struct dx_hash_info *hinfo, struct dx_frame *frame_in)
 437 +        struct dx_hash_info *hinfo, struct dx_frame *frame_in,
 438 +        struct htree_lock *lck)
 439  {
 440         unsigned count, indirect;
 441 -       struct dx_entry *at, *entries, *p, *q, *m;
 442 +       struct dx_entry *at, *entries, *p, *q, *m, *dx = NULL;
 443         struct dx_root_info *info;
 444         struct dx_frame *frame = frame_in;
 445         struct dx_frame *ret_err = ERR_PTR(ERR_BAD_DX_DIR);
 446 @@ -809,8 +1041,15 @@ dx_probe(struct ext4_filename *fname, st
 447
 448         dxtrace(printk("Look up %x", hash));
 449         while (1) {
 450 +               if (indirect == 0) { /* the last index level */
 451 +                       /* NB: ext4_htree_dx_lock() could be noop if
 452 +                        * DX-lock flag is not set for current operation */
 453 +                       ext4_htree_dx_lock(lck, dx);
 454 +                       ext4_htree_spin_lock(lck, dx, NULL);
 455 +               }
 456                 count = dx_get_count(entries);
 457 -               if (!count || count > dx_get_limit(entries)) {
 458 +               if (count == 0 || count > dx_get_limit(entries)) {
 459 +                       ext4_htree_spin_unlock(lck); /* release spin */
 460                         ext4_warning_inode(dir,
 461                                            "dx entry: count %u beyond limit %u",
 462                                            count, dx_get_limit(entries));
 463 @@ -849,8 +1088,70 @@ dx_probe(struct ext4_filename *fname, st
 464                                dx_get_block(at)));
 465                 frame->entries = entries;
 466                 frame->at = at;
 467 -               if (!indirect--)
 468 +
 469 +               if (indirect == 0) { /* the last index level */
 470 +                       struct ext4_dir_lock_data *ld;
 471 +                       u64 myblock;
 472 +
 473 +                       /* By default we only lock DE-block, however, we will
 474 +                        * also lock the last level DX-block if:
 475 +                        * a) there is hash collision
 476 +                        *    we will set DX-lock flag (a few lines below)
 477 +                        *    and redo to lock DX-block
 478 +                        *    see detail in dx_probe_hash_collision()
 479 +                        * b) it's a retry from splitting
 480 +                        *    we need to lock the last level DX-block so nobody
 481 +                        *    else can split any leaf blocks under the same
 482 +                        *    DX-block, see detail in ext4_dx_add_entry()
 483 +                        */
 484 +                       if (ext4_htree_dx_locked(lck)) {
 485 +                               /* DX-block is locked, just lock DE-block
 486 +                                * and return */
 487 +                               ext4_htree_spin_unlock(lck);
 488 +                               if (!ext4_htree_safe_locked(lck))
 489 +                                       ext4_htree_de_lock(lck, frame->at);
 490 +                               return frame;
 491 +                       }
 492 +                       /* it's pdirop and no DX lock */
 493 +                       if (dx_probe_hash_collision(lck, entries, at, hash) ==
 494 +                           DX_HASH_COL_YES) {
 495 +                               /* found hash collision, set DX-lock flag
 496 +                                * and retry to abtain DX-lock */
 497 +                               ext4_htree_spin_unlock(lck);
 498 +                               ext4_htree_dx_need_lock(lck);
 499 +                               continue;
 500 +                       }
 501 +                       ld = ext4_htree_lock_data(lck);
 502 +                       /* because I don't lock DX, so @at can't be trusted
 503 +                        * after I release spinlock so I have to save it */
 504 +                       ld->ld_at = at;
 505 +                       ld->ld_at_entry = *at;
 506 +                       ld->ld_count = dx_get_count(entries);
 507 +
 508 +                       frame->at = &ld->ld_at_entry;
 509 +                       myblock = dx_get_block(at);
 510 +
 511 +                       /* NB: ordering locking */
 512 +                       ext4_htree_spin_unlock_listen(lck, &myblock);
 513 +                       /* other thread can split this DE-block because:
 514 +                        * a) I don't have lock for the DE-block yet
 515 +                        * b) I released spinlock on DX-block
 516 +                        * if it happened I can detect it by listening
 517 +                        * splitting event on this DE-block */
 518 +                       ext4_htree_de_lock(lck, frame->at);
 519 +                       ext4_htree_spin_stop_listen(lck);
 520 +
 521 +                       if (myblock == EXT4_HTREE_NODE_CHANGED) {
 522 +                               /* someone split this DE-block before
 523 +                                * I locked it, I need to retry and lock
 524 +                                * valid DE-block */
 525 +                               ext4_htree_de_unlock(lck);
 526 +                               continue;
 527 +                       }
 528                         return frame;
 529 +               }
 530 +               dx = at;
 531 +               indirect--;
 532                 frame++;
 533                 frame->bh = ext4_read_dirblock(dir, dx_get_block(at), INDEX);
 534                 if (IS_ERR(frame->bh)) {
 535 @@ -916,7 +1217,7 @@ static void dx_release(struct dx_frame *
 536  static int ext4_htree_next_block(struct inode *dir, __u32 hash,
 537                                  struct dx_frame *frame,
 538                                  struct dx_frame *frames,
 539 -                                __u32 *start_hash)
 540 +                                __u32 *start_hash, struct htree_lock *lck)
 541  {
 542         struct dx_frame *p;
 543         struct buffer_head *bh;
 544 @@ -931,12 +1232,22 @@ static int ext4_htree_next_block(struct
 545          * this loop, num_frames indicates the number of interior
 546          * nodes need to be read.
 547          */
 548 +       ext4_htree_de_unlock(lck);
 549         while (1) {
 550 -               if (++(p->at) < p->entries + dx_get_count(p->entries))
 551 -                       break;
 552 +               if (num_frames > 0 || ext4_htree_dx_locked(lck)) {
 553 +                       /* num_frames > 0 :
 554 +                        *   DX block
 555 +                        * ext4_htree_dx_locked:
 556 +                        *   frame->at is reliable pointer returned by dx_probe,
 557 +                        *   otherwise dx_probe already knew no collision */
 558 +                       if (++(p->at) < p->entries + dx_get_count(p->entries))
 559 +                               break;
 560 +               }
 561                 if (p == frames)
 562                         return 0;
 563                 num_frames++;
 564 +               if (num_frames == 1)
 565 +                       ext4_htree_dx_unlock(lck);
 566                 p--;
 567         }
 568
 569 @@ -959,6 +1270,13 @@ static int ext4_htree_next_block(struct
 570          * block so no check is necessary
 571          */
 572         while (num_frames--) {
 573 +               if (num_frames == 0) {
 574 +                       /* it's not always necessary, we just don't want to
 575 +                        * detect hash collision again */
 576 +                       ext4_htree_dx_need_lock(lck);
 577 +                       ext4_htree_dx_lock(lck, p->at);
 578 +               }
 579 +
 580                 bh = ext4_read_dirblock(dir, dx_get_block(p->at), INDEX);
 581                 if (IS_ERR(bh))
 582                         return PTR_ERR(bh);
 583 @@ -967,6 +1285,7 @@ static int ext4_htree_next_block(struct
 584                 p->bh = bh;
 585                 p->at = p->entries = ((struct dx_node *) bh->b_data)->entries;
 586         }
 587 +       ext4_htree_de_lock(lck, p->at);
 588         return 1;
 589  }
 590
 591 @@ -1114,10 +1433,10 @@ int ext4_htree_fill_tree(struct file *di
 592         }
 593         hinfo.hash = start_hash;
 594         hinfo.minor_hash = 0;
 595 -       frame = dx_probe(NULL, dir, &hinfo, frames);
 596 +       /* assume it's PR locked */
 597 +       frame = dx_probe(NULL, dir, &hinfo, frames, NULL);
 598         if (IS_ERR(frame))
 599                 return PTR_ERR(frame);
 600 -
 601         /* Add '.' and '..' from the htree header */
 602         if (!start_hash && !start_minor_hash) {
 603                 de = (struct ext4_dir_entry_2 *) frames[0].bh->b_data;
 604 @@ -1157,7 +1476,7 @@ int ext4_htree_fill_tree(struct file *di
 605                 count += ret;
 606                 hashval = ~0;
 607                 ret = ext4_htree_next_block(dir, HASH_NB_ALWAYS,
 608 -                                           frame, frames, &hashval);
 609 +                                           frame, frames, &hashval, NULL);
 610                 *next_hash = hashval;
 611                 if (ret < 0) {
 612                         err = ret;
 613 @@ -1349,10 +1668,10 @@ static int is_dx_internal_node(struct in
 614   * The returned buffer_head has ->b_count elevated.  The caller is expected
 615   * to brelse() it when appropriate.
 616   */
 617 -static struct buffer_head * ext4_find_entry (struct inode *dir,
 618 +struct buffer_head *__ext4_find_entry(struct inode *dir,
 619                                         const struct qstr *d_name,
 620                                         struct ext4_dir_entry_2 **res_dir,
 621 -                                       int *inlined)
 622 +                                       int *inlined, struct htree_lock *lck)
 623  {
 624         struct super_block *sb;
 625         struct buffer_head *bh_use[NAMEI_RA_SIZE];
 626 @@ -1401,7 +1720,7 @@ static struct buffer_head * ext4_find_en
 627                 goto restart;
 628         }
 629         if (is_dx(dir)) {
 630 -               ret = ext4_dx_find_entry(dir, &fname, res_dir);
 631 +               ret = ext4_dx_find_entry(dir, &fname, res_dir, lck);
 632                 /*
 633                  * On success, or if the error was file not found,
 634                  * return.  Otherwise, fall back to doing a search the
 635 @@ -1411,6 +1730,7 @@ static struct buffer_head * ext4_find_en
 636                         goto cleanup_and_exit;
 637                 dxtrace(printk(KERN_DEBUG "ext4_find_entry: dx failed, "
 638                                "falling back\n"));
 639 +               ext4_htree_safe_relock(lck);
 640                 ret = NULL;
 641         }
 642         nblocks = dir->i_size >> EXT4_BLOCK_SIZE_BITS(sb);
 643 @@ -1499,10 +1819,12 @@ cleanup_and_exit:
 644         ext4_fname_free_filename(&fname);
 645         return ret;
 646  }
 647 +EXPORT_SYMBOL(__ext4_find_entry);
 648
 649  static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
 650                         struct ext4_filename *fname,
 651 -                       struct ext4_dir_entry_2 **res_dir)
 652 +                       struct ext4_dir_entry_2 **res_dir,
 653 +                       struct htree_lock *lck)
 654  {
 655         struct super_block * sb = dir->i_sb;
 656         struct dx_frame frames[EXT4_HTREE_LEVEL], *frame;
 657 @@ -1513,7 +1835,7 @@ static struct buffer_head * ext4_dx_find
 658  #ifdef CONFIG_EXT4_FS_ENCRYPTION
 659         *res_dir = NULL;
 660  #endif
 661 -       frame = dx_probe(fname, dir, NULL, frames);
 662 +       frame = dx_probe(fname, dir, NULL, frames, lck);
 663         if (IS_ERR(frame))
 664                 return (struct buffer_head *) frame;
 665         do {
 666 @@ -1535,7 +1857,7 @@ static struct buffer_head * ext4_dx_find
 667
 668                 /* Check to see if we should continue to search */
 669                 retval = ext4_htree_next_block(dir, fname->hinfo.hash, frame,
 670 -                                              frames, NULL);
 671 +                                              frames, NULL, lck);
 672                 if (retval < 0) {
 673                         ext4_warning_inode(dir,
 674                                 "error %d reading directory index block",
 675 @@ -1710,8 +2032,9 @@ static struct ext4_dir_entry_2* dx_pack_
 676   * Returns pointer to de in block into which the new entry will be inserted.
 677   */
 678  static struct ext4_dir_entry_2 *do_split(handle_t *handle, struct inode *dir,
 679 -                       struct buffer_head **bh,struct dx_frame *frame,
 680 -                       struct dx_hash_info *hinfo)
 681 +                       struct buffer_head **bh, struct dx_frame *frames,
 682 +                       struct dx_frame *frame, struct dx_hash_info *hinfo,
 683 +                       struct htree_lock *lck)
 684  {
 685         unsigned blocksize = dir->i_sb->s_blocksize;
 686         unsigned count, continued;
 687 @@ -1773,8 +2096,14 @@ static struct ext4_dir_entry_2 *do_split
 688                                         hash2, split, count-split));
 689
 690         /* Fancy dance to stay within two buffers */
 691 -       de2 = dx_move_dirents(data1, data2, map + split, count - split,
 692 -                             blocksize);
 693 +       if (hinfo->hash < hash2) {
 694 +               de2 = dx_move_dirents(data1, data2, map + split,
 695 +                                     count - split, blocksize);
 696 +       } else {
 697 +               /* make sure we will add entry to the same block which
 698 +                * we have already locked */
 699 +               de2 = dx_move_dirents(data1, data2, map, split, blocksize);
 700 +       }
 701         de = dx_pack_dirents(data1, blocksize);
 702         de->rec_len = ext4_rec_len_to_disk(data1 + (blocksize - csum_size) -
 703                                            (char *) de,
 704 @@ -1795,12 +2124,21 @@ static struct ext4_dir_entry_2 *do_split
 705         dxtrace(dx_show_leaf(dir, hinfo, (struct ext4_dir_entry_2 *) data2,
 706                         blocksize, 1));
 707
 708 -       /* Which block gets the new entry? */
 709 -       if (hinfo->hash >= hash2) {
 710 -               swap(*bh, bh2);
 711 -               de = de2;
 712 +       ext4_htree_spin_lock(lck, frame > frames ? (frame - 1)->at : NULL,
 713 +                            frame->at); /* notify block is being split */
 714 +       if (hinfo->hash < hash2) {
 715 +               dx_insert_block(frame, hash2 + continued, newblock);
 716 +
 717 +       } else {
 718 +               /* switch block number */
 719 +               dx_insert_block(frame, hash2 + continued,
 720 +                               dx_get_block(frame->at));
 721 +               dx_set_block(frame->at, newblock);
 722 +               (frame->at)++;
 723         }
 724 -       dx_insert_block(frame, hash2 + continued, newblock);
 725 +       ext4_htree_spin_unlock(lck);
 726 +       ext4_htree_dx_unlock(lck);
 727 +
 728         err = ext4_handle_dirty_dirent_node(handle, dir, bh2);
 729         if (err)
 730                 goto journal_error;
 731 @@ -2074,7 +2412,7 @@ static int make_indexed_dir(handle_t *ha
 732         if (retval)
 733                 goto out_frames;
 734
 735 -       de = do_split(handle,dir, &bh2, frame, &fname->hinfo);
 736 +       de = do_split(handle, dir, &bh2, frames, frame, &fname->hinfo, NULL);
 737         if (IS_ERR(de)) {
 738                 retval = PTR_ERR(de);
 739                 goto out_frames;
 740 @@ -2184,8 +2522,8 @@ out:
 741   * may not sleep between calling this and putting something into
 742   * the entry, as someone else might have used it while you slept.
 743   */
 744 -static int ext4_add_entry(handle_t *handle, struct dentry *dentry,
 745 -                         struct inode *inode)
 746 +int __ext4_add_entry(handle_t *handle, struct dentry *dentry,
 747 +                     struct inode *inode, struct htree_lock *lck)
 748  {
 749         struct inode *dir = d_inode(dentry->d_parent);
 750         struct buffer_head *bh = NULL;
 751 @@ -2226,9 +2564,10 @@ static int ext4_add_entry(handle_t *hand
 752                 if (dentry->d_name.len == 2 &&
 753                     memcmp(dentry->d_name.name, "..", 2) == 0)
 754                         return ext4_update_dotdot(handle, dentry, inode);
 755 -               retval = ext4_dx_add_entry(handle, &fname, dir, inode);
 756 +               retval = ext4_dx_add_entry(handle, &fname, dir, inode, lck);
 757                 if (!retval || (retval != ERR_BAD_DX_DIR))
 758                         goto out;
 759 +               ext4_htree_safe_relock(lck);
 760                 ext4_clear_inode_flag(dir, EXT4_INODE_INDEX);
 761                 dx_fallback++;
 762                 ext4_mark_inode_dirty(handle, dir);
 763 @@ -2278,12 +2617,14 @@ out:
 764                 ext4_set_inode_state(inode, EXT4_STATE_NEWENTRY);
 765         return retval;
 766  }
 767 +EXPORT_SYMBOL(__ext4_add_entry);
 768
 769  /*
 770   * Returns 0 for success, or a negative error value
 771   */
 772  static int ext4_dx_add_entry(handle_t *handle, struct ext4_filename *fname,
 773 -                            struct inode *dir, struct inode *inode)
 774 +                            struct inode *dir, struct inode *inode,
 775 +                            struct htree_lock *lck)
 776  {
 777         struct dx_frame frames[EXT4_HTREE_LEVEL], *frame;
 778         struct dx_entry *entries, *at;
 779 @@ -2295,7 +2636,7 @@ static int ext4_dx_add_entry(handle_t *h
 780
 781  again:
 782         restart = 0;
 783 -       frame = dx_probe(fname, dir, NULL, frames);
 784 +       frame = dx_probe(fname, dir, NULL, frames, lck);
 785         if (IS_ERR(frame))
 786                 return PTR_ERR(frame);
 787         entries = frame->entries;
 788 @@ -2330,6 +2671,11 @@ again:
 789                 struct dx_node *node2;
 790                 struct buffer_head *bh2;
 791
 792 +               if (!ext4_htree_safe_locked(lck)) { /* retry with EX lock */
 793 +                       ext4_htree_safe_relock(lck);
 794 +                       restart = 1;
 795 +                       goto cleanup;
 796 +               }
 797                 while (frame > frames) {
 798                         if (dx_get_count((frame - 1)->entries) <
 799                             dx_get_limit((frame - 1)->entries)) {
 800 @@ -2432,8 +2778,32 @@ again:
 801                         restart = 1;
 802                         goto journal_error;
 803                 }
 804 +       } else if (!ext4_htree_dx_locked(lck)) {
 805 +               struct ext4_dir_lock_data *ld = ext4_htree_lock_data(lck);
 806 +
 807 +               /* not well protected, require DX lock */
 808 +               ext4_htree_dx_need_lock(lck);
 809 +               at = frame > frames ? (frame - 1)->at : NULL;
 810 +
 811 +               /* NB: no risk of deadlock because it's just a try.
 812 +                *
 813 +                * NB: we check ld_count for twice, the first time before
 814 +                * having DX lock, the second time after holding DX lock.
 815 +                *
 816 +                * NB: We never free blocks for directory so far, which
 817 +                * means value returned by dx_get_count() should equal to
 818 +                * ld->ld_count if nobody split any DE-block under @at,
 819 +                * and ld->ld_at still points to valid dx_entry. */
 820 +               if ((ld->ld_count != dx_get_count(entries)) ||
 821 +                   !ext4_htree_dx_lock_try(lck, at) ||
 822 +                   (ld->ld_count != dx_get_count(entries))) {
 823 +                       restart = 1;
 824 +                       goto cleanup;
 825 +               }
 826 +               /* OK, I've got DX lock and nothing changed */
 827 +               frame->at = ld->ld_at;
 828         }
 829 -       de = do_split(handle, dir, &bh, frame, &fname->hinfo);
 830 +       de = do_split(handle, dir, &bh, frames, frame, &fname->hinfo, lck);
 831         if (IS_ERR(de)) {
 832                 err = PTR_ERR(de);
 833                 goto cleanup;
 834 @@ -2444,6 +2814,8 @@ again:
 835  journal_error:
 836         ext4_std_error(dir->i_sb, err); /* this is a no-op if err == 0 */
 837  cleanup:
 838 +       ext4_htree_dx_unlock(lck);
 839 +       ext4_htree_de_unlock(lck);
 840         brelse(bh);
 841         dx_release(frames);
 842         /* @restart is true means htree-path has been changed, we need to
 843 Index: linux-4.18.0-80.1.2.el8_0/fs/ext4/super.c
 844 ===================================================================
 845 --- linux-4.18.0-80.1.2.el8_0.orig/fs/ext4/super.c
 846 +++ linux-4.18.0-80.1.2.el8_0/fs/ext4/super.c
 847 @@ -1009,6 +1009,7 @@ static struct inode *ext4_alloc_inode(st
 848
 849         inode_set_iversion(&ei->vfs_inode, 1);
 850         spin_lock_init(&ei->i_raw_lock);
 851 +       sema_init(&ei->i_append_sem, 1);
 852         INIT_LIST_HEAD(&ei->i_prealloc_list);
 853         spin_lock_init(&ei->i_prealloc_lock);
 854         ext4_es_init_tree(&ei->i_es_tree);