ldiskfs/kernel_patches/patches/rhel8.7/ext4-pdirop.patch

   1 LU-50 ldiskfs: pdirops patch for ldiskfs
   2
   3 Single directory performance is a critical for HPC workloads. In a
   4 typical use case an application creates a separate output file for
   5 each node and task in a job. As nodes and tasks increase, hundreds
   6 of thousands of files may be created in a single directory within
   7 a short window of time.
   8 Today, both filename lookup and file system modifying operations
   9 (such as create and unlink) are protected with a single lock for
  10 an entire ldiskfs directory. PDO project will remove this
  11 bottleneck by introducing a parallel locking mechanism for entire
  12 ldiskfs directories. This work will enable multiple application
  13 threads to simultaneously lookup, create and unlink in parallel.
  14
  15 This patch contains:
  16   - pdirops support for ldiskfs
  17   - N-level htree directory
  18   - integrate with osd-ldiskfs
  19
  20 Signed-off-by: Liang Zhen <liang@whamcloud.com>
  21 Change-Id: I269c0e3112e68f3acd79e860dab052a68c7d7aaa
  22 Reviewed-on: http://review.whamcloud.com/375
  23 Reviewed-by: Andreas Dilger <adilger@whamcloud.com>
  24
  25 Index: linux-4.18.0-423.el8/fs/ext4/Makefile
  26 ===================================================================
  27 --- linux-4.18.0-423.el8.orig/fs/ext4/Makefile
  28 +++ linux-4.18.0-423.el8/fs/ext4/Makefile
  29 @@ -7,6 +7,7 @@ obj-$(CONFIG_EXT4_FS) += ext4.o
  30
  31  ext4-y := balloc.o bitmap.o block_validity.o dir.o ext4_jbd2.o extents.o \
  32                 extents_status.o file.o fsmap.o fsync.o hash.o ialloc.o \
  33 +               htree_lock.o \
  34                 indirect.o inline.o inode.o ioctl.o mballoc.o migrate.o \
  35                 mmp.o move_extent.o namei.o page-io.o readpage.o resize.o \
  36                 super.o symlink.o sysfs.o xattr.o xattr_trusted.o xattr_user.o
  37 Index: linux-4.18.0-423.el8/fs/ext4/ext4.h
  38 ===================================================================
  39 --- linux-4.18.0-423.el8.orig/fs/ext4/ext4.h
  40 +++ linux-4.18.0-423.el8/fs/ext4/ext4.h
  41 @@ -29,6 +29,7 @@
  42  #include <linux/timer.h>
  43  #include <linux/version.h>
  44  #include <linux/wait.h>
  45 +#include <linux/htree_lock.h>
  46  #include <linux/sched/signal.h>
  47  #include <linux/blockgroup_lock.h>
  48  #include <linux/percpu_counter.h>
  49 @@ -966,6 +967,9 @@ struct ext4_inode_info {
  50         __u32   i_dtime;
  51         ext4_fsblk_t    i_file_acl;
  52
  53 +       /* following fields for parallel directory operations -bzzz */
  54 +       struct semaphore i_append_sem;
  55 +
  56         /*
  57          * i_block_group is the number of the block group which contains
  58          * this file's inode.  Constant across the lifetime of the inode,
  59 @@ -2217,6 +2221,72 @@ struct dx_hash_info
  60   */
  61  #define HASH_NB_ALWAYS         1
  62
  63 +/* assume name-hash is protected by upper layer */
  64 +#define EXT4_HTREE_LOCK_HASH   0
  65 +
  66 +enum ext4_pdo_lk_types {
  67 +#if EXT4_HTREE_LOCK_HASH
  68 +       EXT4_LK_HASH,
  69 +#endif
  70 +       EXT4_LK_DX,             /* index block */
  71 +       EXT4_LK_DE,             /* directory entry block */
  72 +       EXT4_LK_SPIN,           /* spinlock */
  73 +       EXT4_LK_MAX,
  74 +};
  75 +
  76 +/* read-only bit */
  77 +#define EXT4_LB_RO(b)          (1 << (b))
  78 +/* read + write, high bits for writer */
  79 +#define EXT4_LB_RW(b)          ((1 << (b)) | (1 << (EXT4_LK_MAX + (b))))
  80 +
  81 +enum ext4_pdo_lock_bits {
  82 +       /* DX lock bits */
  83 +       EXT4_LB_DX_RO           = EXT4_LB_RO(EXT4_LK_DX),
  84 +       EXT4_LB_DX              = EXT4_LB_RW(EXT4_LK_DX),
  85 +       /* DE lock bits */
  86 +       EXT4_LB_DE_RO           = EXT4_LB_RO(EXT4_LK_DE),
  87 +       EXT4_LB_DE              = EXT4_LB_RW(EXT4_LK_DE),
  88 +       /* DX spinlock bits */
  89 +       EXT4_LB_SPIN_RO         = EXT4_LB_RO(EXT4_LK_SPIN),
  90 +       EXT4_LB_SPIN            = EXT4_LB_RW(EXT4_LK_SPIN),
  91 +       /* accurate searching */
  92 +       EXT4_LB_EXACT           = EXT4_LB_RO(EXT4_LK_MAX << 1),
  93 +};
  94 +
  95 +enum ext4_pdo_lock_opc {
  96 +       /* external */
  97 +       EXT4_HLOCK_READDIR      = (EXT4_LB_DE_RO | EXT4_LB_DX_RO),
  98 +       EXT4_HLOCK_LOOKUP       = (EXT4_LB_DE_RO | EXT4_LB_SPIN_RO |
  99 +                                  EXT4_LB_EXACT),
 100 +       EXT4_HLOCK_DEL          = (EXT4_LB_DE | EXT4_LB_SPIN_RO |
 101 +                                  EXT4_LB_EXACT),
 102 +       EXT4_HLOCK_ADD          = (EXT4_LB_DE | EXT4_LB_SPIN_RO),
 103 +
 104 +       /* internal */
 105 +       EXT4_HLOCK_LOOKUP_SAFE  = (EXT4_LB_DE_RO | EXT4_LB_DX_RO |
 106 +                                  EXT4_LB_EXACT),
 107 +       EXT4_HLOCK_DEL_SAFE     = (EXT4_LB_DE | EXT4_LB_DX_RO | EXT4_LB_EXACT),
 108 +       EXT4_HLOCK_SPLIT        = (EXT4_LB_DE | EXT4_LB_DX | EXT4_LB_SPIN),
 109 +};
 110 +
 111 +extern struct htree_lock_head *ext4_htree_lock_head_alloc(unsigned hbits);
 112 +#define ext4_htree_lock_head_free(lhead)       htree_lock_head_free(lhead)
 113 +
 114 +extern struct htree_lock *ext4_htree_lock_alloc(void);
 115 +#define ext4_htree_lock_free(lck)              htree_lock_free(lck)
 116 +
 117 +extern void ext4_htree_lock(struct htree_lock *lck,
 118 +                           struct htree_lock_head *lhead,
 119 +                           struct inode *dir, unsigned flags);
 120 +#define ext4_htree_unlock(lck)                  htree_unlock(lck)
 121 +
 122 +extern struct buffer_head *__ext4_find_entry(struct inode *dir,
 123 +                                       const struct qstr *d_name,
 124 +                                       struct ext4_dir_entry_2 **res_dir,
 125 +                                       int *inlined, struct htree_lock *lck);
 126 +extern int __ext4_add_entry(handle_t *handle, struct dentry *dentry,
 127 +                     struct inode *inode, struct htree_lock *lck);
 128 +
 129  struct ext4_filename {
 130         const struct qstr *usr_fname;
 131         struct fscrypt_str disk_name;
 132 @@ -2519,12 +2589,20 @@ void ext4_insert_dentry(struct inode *in
 133                         struct ext4_filename *fname, void *data);
 134  static inline void ext4_update_dx_flag(struct inode *inode)
 135  {
 136 +       /* Disable it for ldiskfs, because going from a DX directory to
 137 +        * a non-DX directory while it is in use will completely break
 138 +        * the htree-locking.
 139 +        * If we really want to support this operation in the future,
 140 +        * we need to exclusively lock the directory at here which will
 141 +        * increase complexity of code */
 142 +#if 0
 143         if (!ext4_has_feature_dir_index(inode->i_sb) &&
 144             ext4_test_inode_flag(inode, EXT4_INODE_INDEX)) {
 145                 /* ext4_iget() should have caught this... */
 146                 WARN_ON_ONCE(ext4_has_feature_metadata_csum(inode->i_sb));
 147                 ext4_clear_inode_flag(inode, EXT4_INODE_INDEX);
 148         }
 149 +#endif
 150  }
 151  static const unsigned char ext4_filetype_table[] = {
 152         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
 153 Index: linux-4.18.0-423.el8/fs/ext4/namei.c
 154 ===================================================================
 155 --- linux-4.18.0-423.el8.orig/fs/ext4/namei.c
 156 +++ linux-4.18.0-423.el8/fs/ext4/namei.c
 157 @@ -55,6 +55,7 @@ struct buffer_head *ext4_append(handle_t
 158  {
 159         struct ext4_map_blocks map;
 160         struct buffer_head *bh;
 161 +       struct ext4_inode_info *ei = EXT4_I(inode);
 162         int err;
 163
 164         if (unlikely(EXT4_SB(inode->i_sb)->s_max_dir_size_kb &&
 165 @@ -62,6 +63,10 @@ struct buffer_head *ext4_append(handle_t
 166                       EXT4_SB(inode->i_sb)->s_max_dir_size_kb)))
 167                 return ERR_PTR(-ENOSPC);
 168
 169 +       /* with parallel dir operations all appends
 170 +       * have to be serialized -bzzz */
 171 +       down(&ei->i_append_sem);
 172 +
 173         *block = inode->i_size >> inode->i_sb->s_blocksize_bits;
 174         map.m_lblk = *block;
 175         map.m_len = 1;
 176 @@ -72,18 +77,24 @@ struct buffer_head *ext4_append(handle_t
 177          * directory.
 178          */
 179         err = ext4_map_blocks(NULL, inode, &map, 0);
 180 -       if (err < 0)
 181 +       if (err < 0) {
 182 +               up(&ei->i_append_sem);
 183                 return ERR_PTR(err);
 184 +       }
 185         if (err) {
 186 +               up(&ei->i_append_sem);
 187                 EXT4_ERROR_INODE(inode, "Logical block already allocated");
 188                 return ERR_PTR(-EFSCORRUPTED);
 189         }
 190
 191         bh = ext4_bread(handle, inode, *block, EXT4_GET_BLOCKS_CREATE);
 192 -       if (IS_ERR(bh))
 193 +       if (IS_ERR(bh)) {
 194 +               up(&ei->i_append_sem);
 195                 return bh;
 196 +       }
 197         inode->i_size += inode->i_sb->s_blocksize;
 198         EXT4_I(inode)->i_disksize = inode->i_size;
 199 +       up(&ei->i_append_sem);
 200         BUFFER_TRACE(bh, "get_write_access");
 201         err = ext4_journal_get_write_access(handle, bh);
 202         if (err) {
 203 @@ -288,7 +299,8 @@ static unsigned dx_node_limit(struct ino
 204  static struct dx_frame *dx_probe(struct ext4_filename *fname,
 205                                  struct inode *dir,
 206                                  struct dx_hash_info *hinfo,
 207 -                                struct dx_frame *frame);
 208 +                                struct dx_frame *frame,
 209 +                                struct htree_lock *lck);
 210  static void dx_release(struct dx_frame *frames);
 211  static int dx_make_map(struct inode *dir, struct buffer_head *bh,
 212                        struct dx_hash_info *hinfo,
 213 @@ -302,12 +314,13 @@ static void dx_insert_block(struct dx_fr
 214  static int ext4_htree_next_block(struct inode *dir, __u32 hash,
 215                                  struct dx_frame *frame,
 216                                  struct dx_frame *frames,
 217 -                                __u32 *start_hash);
 218 +                                __u32 *start_hash, struct htree_lock *lck);
 219  static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
 220                 struct ext4_filename *fname,
 221 -               struct ext4_dir_entry_2 **res_dir);
 222 +               struct ext4_dir_entry_2 **res_dir, struct htree_lock *lck);
 223  static int ext4_dx_add_entry(handle_t *handle, struct ext4_filename *fname,
 224 -                            struct inode *dir, struct inode *inode);
 225 +                            struct inode *dir, struct inode *inode,
 226 +                            struct htree_lock *lck);
 227
 228  /* checksumming functions */
 229  void initialize_dirent_tail(struct ext4_dir_entry_tail *t,
 230 @@ -771,6 +784,227 @@ struct stats dx_show_entries(struct dx_h
 231  }
 232  #endif /* DX_DEBUG */
 233
 234 +/* private data for htree_lock */
 235 +struct ext4_dir_lock_data {
 236 +       unsigned                ld_flags;  /* bits-map for lock types */
 237 +       unsigned                ld_count;  /* # entries of the last DX block */
 238 +       struct dx_entry         ld_at_entry; /* copy of leaf dx_entry */
 239 +       struct dx_entry         *ld_at;    /* position of leaf dx_entry */
 240 +};
 241 +
 242 +#define ext4_htree_lock_data(l)        ((struct ext4_dir_lock_data *)(l)->lk_private)
 243 +#define ext4_find_entry(dir, name, dirent, inline) \
 244 +                       __ext4_find_entry(dir, name, dirent, inline, NULL)
 245 +#define ext4_add_entry(handle, dentry, inode) \
 246 +                       __ext4_add_entry(handle, dentry, inode, NULL)
 247 +
 248 +/* NB: ext4_lblk_t is 32 bits so we use high bits to identify invalid blk */
 249 +#define EXT4_HTREE_NODE_CHANGED        (0xcafeULL << 32)
 250 +
 251 +static void ext4_htree_event_cb(void *target, void *event)
 252 +{
 253 +       u64 *block = (u64 *)target;
 254 +
 255 +       if (*block == dx_get_block((struct dx_entry *)event))
 256 +               *block = EXT4_HTREE_NODE_CHANGED;
 257 +}
 258 +
 259 +struct htree_lock_head *ext4_htree_lock_head_alloc(unsigned hbits)
 260 +{
 261 +       struct htree_lock_head *lhead;
 262 +
 263 +       lhead = htree_lock_head_alloc(EXT4_LK_MAX, hbits, 0);
 264 +       if (lhead != NULL) {
 265 +               htree_lock_event_attach(lhead, EXT4_LK_SPIN, HTREE_EVENT_WR,
 266 +                                       ext4_htree_event_cb);
 267 +       }
 268 +       return lhead;
 269 +}
 270 +EXPORT_SYMBOL(ext4_htree_lock_head_alloc);
 271 +
 272 +struct htree_lock *ext4_htree_lock_alloc(void)
 273 +{
 274 +       return htree_lock_alloc(EXT4_LK_MAX,
 275 +                               sizeof(struct ext4_dir_lock_data));
 276 +}
 277 +EXPORT_SYMBOL(ext4_htree_lock_alloc);
 278 +
 279 +static htree_lock_mode_t ext4_htree_mode(unsigned flags)
 280 +{
 281 +       switch (flags) {
 282 +       default: /* 0 or unknown flags require EX lock */
 283 +               return HTREE_LOCK_EX;
 284 +       case EXT4_HLOCK_READDIR:
 285 +               return HTREE_LOCK_PR;
 286 +       case EXT4_HLOCK_LOOKUP:
 287 +               return HTREE_LOCK_CR;
 288 +       case EXT4_HLOCK_DEL:
 289 +       case EXT4_HLOCK_ADD:
 290 +               return HTREE_LOCK_CW;
 291 +       }
 292 +}
 293 +
 294 +/* return PR for read-only operations, otherwise return EX */
 295 +static inline htree_lock_mode_t ext4_htree_safe_mode(unsigned flags)
 296 +{
 297 +       int writer = (flags & EXT4_LB_DE) == EXT4_LB_DE;
 298 +
 299 +       /* 0 requires EX lock */
 300 +       return (flags == 0 || writer) ? HTREE_LOCK_EX : HTREE_LOCK_PR;
 301 +}
 302 +
 303 +static int ext4_htree_safe_locked(struct htree_lock *lck)
 304 +{
 305 +       int writer;
 306 +
 307 +       if (lck == NULL || lck->lk_mode == HTREE_LOCK_EX)
 308 +               return 1;
 309 +
 310 +       writer = (ext4_htree_lock_data(lck)->ld_flags & EXT4_LB_DE) ==
 311 +                EXT4_LB_DE;
 312 +       if (writer) /* all readers & writers are excluded? */
 313 +               return lck->lk_mode == HTREE_LOCK_EX;
 314 +
 315 +       /* all writers are excluded? */
 316 +       return lck->lk_mode == HTREE_LOCK_PR ||
 317 +              lck->lk_mode == HTREE_LOCK_PW ||
 318 +              lck->lk_mode == HTREE_LOCK_EX;
 319 +}
 320 +
 321 +/* relock htree_lock with EX mode if it's change operation, otherwise
 322 + * relock it with PR mode. It's noop if PDO is disabled. */
 323 +static void ext4_htree_safe_relock(struct htree_lock *lck)
 324 +{
 325 +       if (!ext4_htree_safe_locked(lck)) {
 326 +               unsigned flags = ext4_htree_lock_data(lck)->ld_flags;
 327 +
 328 +               htree_change_lock(lck, ext4_htree_safe_mode(flags));
 329 +       }
 330 +}
 331 +
 332 +void ext4_htree_lock(struct htree_lock *lck, struct htree_lock_head *lhead,
 333 +                    struct inode *dir, unsigned flags)
 334 +{
 335 +       htree_lock_mode_t mode = is_dx(dir) ? ext4_htree_mode(flags) :
 336 +                                             ext4_htree_safe_mode(flags);
 337 +
 338 +       ext4_htree_lock_data(lck)->ld_flags = flags;
 339 +       htree_lock(lck, lhead, mode);
 340 +       if (!is_dx(dir))
 341 +               ext4_htree_safe_relock(lck); /* make sure it's safe locked */
 342 +}
 343 +EXPORT_SYMBOL(ext4_htree_lock);
 344 +
 345 +static int ext4_htree_node_lock(struct htree_lock *lck, struct dx_entry *at,
 346 +                               unsigned lmask, int wait, void *ev)
 347 +{
 348 +       u32     key = (at == NULL) ? 0 : dx_get_block(at);
 349 +       u32     mode;
 350 +
 351 +       /* NOOP if htree is well protected or caller doesn't require the lock */
 352 +       if (ext4_htree_safe_locked(lck) ||
 353 +          !(ext4_htree_lock_data(lck)->ld_flags & lmask))
 354 +               return 1;
 355 +
 356 +       mode = (ext4_htree_lock_data(lck)->ld_flags & lmask) == lmask ?
 357 +               HTREE_LOCK_PW : HTREE_LOCK_PR;
 358 +       while (1) {
 359 +               if (htree_node_lock_try(lck, mode, key, ffz(~lmask), wait, ev))
 360 +                       return 1;
 361 +               if (!(lmask & EXT4_LB_SPIN)) /* not a spinlock */
 362 +                       return 0;
 363 +               cpu_relax(); /* spin until granted */
 364 +       }
 365 +}
 366 +
 367 +static int ext4_htree_node_locked(struct htree_lock *lck, unsigned lmask)
 368 +{
 369 +       return ext4_htree_safe_locked(lck) ||
 370 +              htree_node_is_granted(lck, ffz(~lmask));
 371 +}
 372 +
 373 +static void ext4_htree_node_unlock(struct htree_lock *lck,
 374 +                                  unsigned lmask, void *buf)
 375 +{
 376 +       /* NB: it's safe to call mutiple times or even it's not locked */
 377 +       if (!ext4_htree_safe_locked(lck) &&
 378 +            htree_node_is_granted(lck, ffz(~lmask)))
 379 +               htree_node_unlock(lck, ffz(~lmask), buf);
 380 +}
 381 +
 382 +#define ext4_htree_dx_lock(lck, key)           \
 383 +       ext4_htree_node_lock(lck, key, EXT4_LB_DX, 1, NULL)
 384 +#define ext4_htree_dx_lock_try(lck, key)       \
 385 +       ext4_htree_node_lock(lck, key, EXT4_LB_DX, 0, NULL)
 386 +#define ext4_htree_dx_unlock(lck)              \
 387 +       ext4_htree_node_unlock(lck, EXT4_LB_DX, NULL)
 388 +#define ext4_htree_dx_locked(lck)              \
 389 +       ext4_htree_node_locked(lck, EXT4_LB_DX)
 390 +
 391 +static void ext4_htree_dx_need_lock(struct htree_lock *lck)
 392 +{
 393 +       struct ext4_dir_lock_data *ld;
 394 +
 395 +       if (ext4_htree_safe_locked(lck))
 396 +               return;
 397 +
 398 +       ld = ext4_htree_lock_data(lck);
 399 +       switch (ld->ld_flags) {
 400 +       default:
 401 +               return;
 402 +       case EXT4_HLOCK_LOOKUP:
 403 +               ld->ld_flags = EXT4_HLOCK_LOOKUP_SAFE;
 404 +               return;
 405 +       case EXT4_HLOCK_DEL:
 406 +               ld->ld_flags = EXT4_HLOCK_DEL_SAFE;
 407 +               return;
 408 +       case EXT4_HLOCK_ADD:
 409 +               ld->ld_flags = EXT4_HLOCK_SPLIT;
 410 +               return;
 411 +       }
 412 +}
 413 +
 414 +#define ext4_htree_de_lock(lck, key)           \
 415 +       ext4_htree_node_lock(lck, key, EXT4_LB_DE, 1, NULL)
 416 +#define ext4_htree_de_unlock(lck)              \
 417 +       ext4_htree_node_unlock(lck, EXT4_LB_DE, NULL)
 418 +
 419 +#define ext4_htree_spin_lock(lck, key, event)  \
 420 +       ext4_htree_node_lock(lck, key, EXT4_LB_SPIN, 0, event)
 421 +#define ext4_htree_spin_unlock(lck)            \
 422 +       ext4_htree_node_unlock(lck, EXT4_LB_SPIN, NULL)
 423 +#define ext4_htree_spin_unlock_listen(lck, p)  \
 424 +       ext4_htree_node_unlock(lck, EXT4_LB_SPIN, p)
 425 +
 426 +static void ext4_htree_spin_stop_listen(struct htree_lock *lck)
 427 +{
 428 +       if (!ext4_htree_safe_locked(lck) &&
 429 +           htree_node_is_listening(lck, ffz(~EXT4_LB_SPIN)))
 430 +               htree_node_stop_listen(lck, ffz(~EXT4_LB_SPIN));
 431 +}
 432 +
 433 +enum {
 434 +       DX_HASH_COL_IGNORE,     /* ignore collision while probing frames */
 435 +       DX_HASH_COL_YES,        /* there is collision and it does matter */
 436 +       DX_HASH_COL_NO,         /* there is no collision */
 437 +};
 438 +
 439 +static int dx_probe_hash_collision(struct htree_lock *lck,
 440 +                                  struct dx_entry *entries,
 441 +                                  struct dx_entry *at, u32 hash)
 442 +{
 443 +       if (!(lck && ext4_htree_lock_data(lck)->ld_flags & EXT4_LB_EXACT)) {
 444 +               return DX_HASH_COL_IGNORE; /* don't care about collision */
 445 +
 446 +       } else if (at == entries + dx_get_count(entries) - 1) {
 447 +               return DX_HASH_COL_IGNORE; /* not in any leaf of this DX */
 448 +
 449 +       } else { /* hash collision? */
 450 +               return ((dx_get_hash(at + 1) & ~1) == hash) ?
 451 +                       DX_HASH_COL_YES : DX_HASH_COL_NO;
 452 +       }
 453 +}
 454 +
 455  /*
 456   * Probe for a directory leaf block to search.
 457   *
 458 @@ -782,10 +1016,11 @@ struct stats dx_show_entries(struct dx_h
 459   */
 460  static struct dx_frame *
 461  dx_probe(struct ext4_filename *fname, struct inode *dir,
 462 -        struct dx_hash_info *hinfo, struct dx_frame *frame_in)
 463 +        struct dx_hash_info *hinfo, struct dx_frame *frame_in,
 464 +        struct htree_lock *lck)
 465  {
 466         unsigned count, indirect, level, i;
 467 -       struct dx_entry *at, *entries, *p, *q, *m;
 468 +       struct dx_entry *at, *entries, *p, *q, *m, *dx = NULL;
 469         struct dx_root_info *info;
 470         struct dx_frame *frame = frame_in;
 471         struct dx_frame *ret_err = ERR_PTR(ERR_BAD_DX_DIR);
 472 @@ -851,8 +1086,15 @@ dx_probe(struct ext4_filename *fname, st
 473         level = 0;
 474         blocks[0] = 0;
 475         while (1) {
 476 +               if (indirect == 0) { /* the last index level */
 477 +                       /* NB: ext4_htree_dx_lock() could be noop if
 478 +                        * DX-lock flag is not set for current operation */
 479 +                       ext4_htree_dx_lock(lck, dx);
 480 +                       ext4_htree_spin_lock(lck, dx, NULL);
 481 +               }
 482                 count = dx_get_count(entries);
 483 -               if (!count || count > dx_get_limit(entries)) {
 484 +               if (count == 0 || count > dx_get_limit(entries)) {
 485 +                       ext4_htree_spin_unlock(lck); /* release spin */
 486                         ext4_warning_inode(dir,
 487                                            "dx entry: count %u beyond limit %u",
 488                                            count, dx_get_limit(entries));
 489 @@ -901,8 +1143,70 @@ dx_probe(struct ext4_filename *fname, st
 490                                 goto fail;
 491                         }
 492                 }
 493 -               if (++level > indirect)
 494 +
 495 +               if (indirect == 0) { /* the last index level */
 496 +                       struct ext4_dir_lock_data *ld;
 497 +                       u64 myblock;
 498 +
 499 +                       /* By default we only lock DE-block, however, we will
 500 +                        * also lock the last level DX-block if:
 501 +                        * a) there is hash collision
 502 +                        *    we will set DX-lock flag (a few lines below)
 503 +                        *    and redo to lock DX-block
 504 +                        *    see detail in dx_probe_hash_collision()
 505 +                        * b) it's a retry from splitting
 506 +                        *    we need to lock the last level DX-block so nobody
 507 +                        *    else can split any leaf blocks under the same
 508 +                        *    DX-block, see detail in ext4_dx_add_entry()
 509 +                        */
 510 +                       if (ext4_htree_dx_locked(lck)) {
 511 +                               /* DX-block is locked, just lock DE-block
 512 +                                * and return */
 513 +                               ext4_htree_spin_unlock(lck);
 514 +                               if (!ext4_htree_safe_locked(lck))
 515 +                                       ext4_htree_de_lock(lck, frame->at);
 516 +                               return frame;
 517 +                       }
 518 +                       /* it's pdirop and no DX lock */
 519 +                       if (dx_probe_hash_collision(lck, entries, at, hash) ==
 520 +                           DX_HASH_COL_YES) {
 521 +                               /* found hash collision, set DX-lock flag
 522 +                                * and retry to abtain DX-lock */
 523 +                               ext4_htree_spin_unlock(lck);
 524 +                               ext4_htree_dx_need_lock(lck);
 525 +                               continue;
 526 +                       }
 527 +                       ld = ext4_htree_lock_data(lck);
 528 +                       /* because I don't lock DX, so @at can't be trusted
 529 +                        * after I release spinlock so I have to save it */
 530 +                       ld->ld_at = at;
 531 +                       ld->ld_at_entry = *at;
 532 +                       ld->ld_count = dx_get_count(entries);
 533 +
 534 +                       frame->at = &ld->ld_at_entry;
 535 +                       myblock = dx_get_block(at);
 536 +
 537 +                       /* NB: ordering locking */
 538 +                       ext4_htree_spin_unlock_listen(lck, &myblock);
 539 +                       /* other thread can split this DE-block because:
 540 +                        * a) I don't have lock for the DE-block yet
 541 +                        * b) I released spinlock on DX-block
 542 +                        * if it happened I can detect it by listening
 543 +                        * splitting event on this DE-block */
 544 +                       ext4_htree_de_lock(lck, frame->at);
 545 +                       ext4_htree_spin_stop_listen(lck);
 546 +
 547 +                       if (myblock == EXT4_HTREE_NODE_CHANGED) {
 548 +                               /* someone split this DE-block before
 549 +                                * I locked it, I need to retry and lock
 550 +                                * valid DE-block */
 551 +                               ext4_htree_de_unlock(lck);
 552 +                               continue;
 553 +                       }
 554                         return frame;
 555 +               }
 556 +               dx = at;
 557 +               indirect--;
 558                 blocks[level] = block;
 559                 frame++;
 560                 frame->bh = ext4_read_dirblock(dir, block, INDEX);
 561 @@ -973,7 +1277,7 @@ static void dx_release(struct dx_frame *
 562  static int ext4_htree_next_block(struct inode *dir, __u32 hash,
 563                                  struct dx_frame *frame,
 564                                  struct dx_frame *frames,
 565 -                                __u32 *start_hash)
 566 +                                __u32 *start_hash, struct htree_lock *lck)
 567  {
 568         struct dx_frame *p;
 569         struct buffer_head *bh;
 570 @@ -988,12 +1292,22 @@ static int ext4_htree_next_block(struct
 571          * this loop, num_frames indicates the number of interior
 572          * nodes need to be read.
 573          */
 574 +       ext4_htree_de_unlock(lck);
 575         while (1) {
 576 -               if (++(p->at) < p->entries + dx_get_count(p->entries))
 577 -                       break;
 578 +               if (num_frames > 0 || ext4_htree_dx_locked(lck)) {
 579 +                       /* num_frames > 0 :
 580 +                        *   DX block
 581 +                        * ext4_htree_dx_locked:
 582 +                        *   frame->at is reliable pointer returned by dx_probe,
 583 +                        *   otherwise dx_probe already knew no collision */
 584 +                       if (++(p->at) < p->entries + dx_get_count(p->entries))
 585 +                               break;
 586 +               }
 587                 if (p == frames)
 588                         return 0;
 589                 num_frames++;
 590 +               if (num_frames == 1)
 591 +                       ext4_htree_dx_unlock(lck);
 592                 p--;
 593         }
 594
 595 @@ -1016,6 +1330,13 @@ static int ext4_htree_next_block(struct
 596          * block so no check is necessary
 597          */
 598         while (num_frames--) {
 599 +               if (num_frames == 0) {
 600 +                       /* it's not always necessary, we just don't want to
 601 +                        * detect hash collision again */
 602 +                       ext4_htree_dx_need_lock(lck);
 603 +                       ext4_htree_dx_lock(lck, p->at);
 604 +               }
 605 +
 606                 bh = ext4_read_dirblock(dir, dx_get_block(p->at), INDEX);
 607                 if (IS_ERR(bh))
 608                         return PTR_ERR(bh);
 609 @@ -1024,6 +1345,7 @@ static int ext4_htree_next_block(struct
 610                 p->bh = bh;
 611                 p->at = p->entries = ((struct dx_node *) bh->b_data)->entries;
 612         }
 613 +       ext4_htree_de_lock(lck, p->at);
 614         return 1;
 615  }
 616
 617 @@ -1171,10 +1493,10 @@ int ext4_htree_fill_tree(struct file *di
 618         }
 619         hinfo.hash = start_hash;
 620         hinfo.minor_hash = 0;
 621 -       frame = dx_probe(NULL, dir, &hinfo, frames);
 622 +       /* assume it's PR locked */
 623 +       frame = dx_probe(NULL, dir, &hinfo, frames, NULL);
 624         if (IS_ERR(frame))
 625                 return PTR_ERR(frame);
 626 -
 627         /* Add '.' and '..' from the htree header */
 628         if (!start_hash && !start_minor_hash) {
 629                 de = (struct ext4_dir_entry_2 *) frames[0].bh->b_data;
 630 @@ -1214,7 +1536,7 @@ int ext4_htree_fill_tree(struct file *di
 631                 count += ret;
 632                 hashval = ~0;
 633                 ret = ext4_htree_next_block(dir, HASH_NB_ALWAYS,
 634 -                                           frame, frames, &hashval);
 635 +                                           frame, frames, &hashval, NULL);
 636                 *next_hash = hashval;
 637                 if (ret < 0) {
 638                         err = ret;
 639 @@ -1413,10 +1735,10 @@ static int is_dx_internal_node(struct in
 640   * The returned buffer_head has ->b_count elevated.  The caller is expected
 641   * to brelse() it when appropriate.
 642   */
 643 -static struct buffer_head * ext4_find_entry (struct inode *dir,
 644 +struct buffer_head *__ext4_find_entry(struct inode *dir,
 645                                         const struct qstr *d_name,
 646                                         struct ext4_dir_entry_2 **res_dir,
 647 -                                       int *inlined)
 648 +                                       int *inlined, struct htree_lock *lck)
 649  {
 650         struct super_block *sb;
 651         struct buffer_head *bh_use[NAMEI_RA_SIZE];
 652 @@ -1465,7 +1787,7 @@ static struct buffer_head * ext4_find_en
 653                 goto restart;
 654         }
 655         if (is_dx(dir)) {
 656 -               ret = ext4_dx_find_entry(dir, &fname, res_dir);
 657 +               ret = ext4_dx_find_entry(dir, &fname, res_dir, lck);
 658                 /*
 659                  * On success, or if the error was file not found,
 660                  * return.  Otherwise, fall back to doing a search the
 661 @@ -1475,6 +1797,7 @@ static struct buffer_head * ext4_find_en
 662                         goto cleanup_and_exit;
 663                 dxtrace(printk(KERN_DEBUG "ext4_find_entry: dx failed, "
 664                                "falling back\n"));
 665 +               ext4_htree_safe_relock(lck);
 666                 ret = NULL;
 667         }
 668         nblocks = dir->i_size >> EXT4_BLOCK_SIZE_BITS(sb);
 669 @@ -1566,10 +1889,12 @@ cleanup_and_exit:
 670         ext4_fname_free_filename(&fname);
 671         return ret;
 672  }
 673 +EXPORT_SYMBOL(__ext4_find_entry);
 674
 675  static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
 676                         struct ext4_filename *fname,
 677 -                       struct ext4_dir_entry_2 **res_dir)
 678 +                       struct ext4_dir_entry_2 **res_dir,
 679 +                       struct htree_lock *lck)
 680  {
 681         struct super_block * sb = dir->i_sb;
 682         struct dx_frame frames[EXT4_HTREE_LEVEL], *frame;
 683 @@ -1580,7 +1905,7 @@ static struct buffer_head * ext4_dx_find
 684  #ifdef CONFIG_EXT4_FS_ENCRYPTION
 685         *res_dir = NULL;
 686  #endif
 687 -       frame = dx_probe(fname, dir, NULL, frames);
 688 +       frame = dx_probe(fname, dir, NULL, frames, lck);
 689         if (IS_ERR(frame))
 690                 return (struct buffer_head *) frame;
 691         do {
 692 @@ -1602,7 +1927,7 @@ static struct buffer_head * ext4_dx_find
 693
 694                 /* Check to see if we should continue to search */
 695                 retval = ext4_htree_next_block(dir, fname->hinfo.hash, frame,
 696 -                                              frames, NULL);
 697 +                                              frames, NULL, lck);
 698                 if (retval < 0) {
 699                         ext4_warning_inode(dir,
 700                                 "error %d reading directory index block",
 701 @@ -1777,8 +2102,9 @@ static struct ext4_dir_entry_2* dx_pack_
 702   * Returns pointer to de in block into which the new entry will be inserted.
 703   */
 704  static struct ext4_dir_entry_2 *do_split(handle_t *handle, struct inode *dir,
 705 -                       struct buffer_head **bh,struct dx_frame *frame,
 706 -                       struct dx_hash_info *hinfo)
 707 +                       struct buffer_head **bh, struct dx_frame *frames,
 708 +                       struct dx_frame *frame, struct dx_hash_info *hinfo,
 709 +                       struct htree_lock *lck)
 710  {
 711         unsigned blocksize = dir->i_sb->s_blocksize;
 712         unsigned continued;
 713 @@ -1854,8 +2180,14 @@ static struct ext4_dir_entry_2 *do_split
 714                                         hash2, split, count-split));
 715
 716         /* Fancy dance to stay within two buffers */
 717 -       de2 = dx_move_dirents(data1, data2, map + split, count - split,
 718 -                             blocksize);
 719 +       if (hinfo->hash < hash2) {
 720 +               de2 = dx_move_dirents(data1, data2, map + split,
 721 +                                     count - split, blocksize);
 722 +       } else {
 723 +               /* make sure we will add entry to the same block which
 724 +                * we have already locked */
 725 +               de2 = dx_move_dirents(data1, data2, map, split, blocksize);
 726 +       }
 727         de = dx_pack_dirents(data1, blocksize);
 728         de->rec_len = ext4_rec_len_to_disk(data1 + (blocksize - csum_size) -
 729                                            (char *) de,
 730 @@ -1876,12 +2208,21 @@ static struct ext4_dir_entry_2 *do_split
 731         dxtrace(dx_show_leaf(dir, hinfo, (struct ext4_dir_entry_2 *) data2,
 732                         blocksize, 1));
 733
 734 -       /* Which block gets the new entry? */
 735 -       if (hinfo->hash >= hash2) {
 736 -               swap(*bh, bh2);
 737 -               de = de2;
 738 +       ext4_htree_spin_lock(lck, frame > frames ? (frame - 1)->at : NULL,
 739 +                            frame->at); /* notify block is being split */
 740 +       if (hinfo->hash < hash2) {
 741 +               dx_insert_block(frame, hash2 + continued, newblock);
 742 +
 743 +       } else {
 744 +               /* switch block number */
 745 +               dx_insert_block(frame, hash2 + continued,
 746 +                               dx_get_block(frame->at));
 747 +               dx_set_block(frame->at, newblock);
 748 +               (frame->at)++;
 749         }
 750 -       dx_insert_block(frame, hash2 + continued, newblock);
 751 +       ext4_htree_spin_unlock(lck);
 752 +       ext4_htree_dx_unlock(lck);
 753 +
 754         err = ext4_handle_dirty_dirent_node(handle, dir, bh2);
 755         if (err)
 756                 goto journal_error;
 757 @@ -2155,7 +2496,7 @@ static int make_indexed_dir(handle_t *ha
 758         if (retval)
 759                 goto out_frames;
 760
 761 -       de = do_split(handle,dir, &bh2, frame, &fname->hinfo);
 762 +       de = do_split(handle, dir, &bh2, frames, frame, &fname->hinfo, NULL);
 763         if (IS_ERR(de)) {
 764                 retval = PTR_ERR(de);
 765                 goto out_frames;
 766 @@ -2265,8 +2606,8 @@ out:
 767   * may not sleep between calling this and putting something into
 768   * the entry, as someone else might have used it while you slept.
 769   */
 770 -static int ext4_add_entry(handle_t *handle, struct dentry *dentry,
 771 -                         struct inode *inode)
 772 +int __ext4_add_entry(handle_t *handle, struct dentry *dentry,
 773 +                     struct inode *inode, struct htree_lock *lck)
 774  {
 775         struct inode *dir = d_inode(dentry->d_parent);
 776         struct buffer_head *bh = NULL;
 777 @@ -2307,9 +2648,10 @@ static int ext4_add_entry(handle_t *hand
 778                 if (dentry->d_name.len == 2 &&
 779                     memcmp(dentry->d_name.name, "..", 2) == 0)
 780                         return ext4_update_dotdot(handle, dentry, inode);
 781 -               retval = ext4_dx_add_entry(handle, &fname, dir, inode);
 782 +               retval = ext4_dx_add_entry(handle, &fname, dir, inode, lck);
 783                 if (!retval || (retval != ERR_BAD_DX_DIR))
 784                         goto out;
 785 +               ext4_htree_safe_relock(lck);
 786                 /* Can we just ignore htree data? */
 787                 if (ext4_has_metadata_csum(sb)) {
 788                         EXT4_ERROR_INODE(dir,
 789 @@ -2372,12 +2714,14 @@ out:
 790                 ext4_set_inode_state(inode, EXT4_STATE_NEWENTRY);
 791         return retval;
 792  }
 793 +EXPORT_SYMBOL(__ext4_add_entry);
 794
 795  /*
 796   * Returns 0 for success, or a negative error value
 797   */
 798  static int ext4_dx_add_entry(handle_t *handle, struct ext4_filename *fname,
 799 -                            struct inode *dir, struct inode *inode)
 800 +                            struct inode *dir, struct inode *inode,
 801 +                            struct htree_lock *lck)
 802  {
 803         struct dx_frame frames[EXT4_HTREE_LEVEL], *frame;
 804         struct dx_entry *entries, *at;
 805 @@ -2389,7 +2733,7 @@ static int ext4_dx_add_entry(handle_t *h
 806
 807  again:
 808         restart = 0;
 809 -       frame = dx_probe(fname, dir, NULL, frames);
 810 +       frame = dx_probe(fname, dir, NULL, frames, lck);
 811         if (IS_ERR(frame))
 812                 return PTR_ERR(frame);
 813         entries = frame->entries;
 814 @@ -2424,6 +2768,11 @@ again:
 815                 struct dx_node *node2;
 816                 struct buffer_head *bh2;
 817
 818 +               if (!ext4_htree_safe_locked(lck)) { /* retry with EX lock */
 819 +                       ext4_htree_safe_relock(lck);
 820 +                       restart = 1;
 821 +                       goto cleanup;
 822 +               }
 823                 while (frame > frames) {
 824                         if (dx_get_count((frame - 1)->entries) <
 825                             dx_get_limit((frame - 1)->entries)) {
 826 @@ -2525,8 +2874,32 @@ again:
 827                         restart = 1;
 828                         goto journal_error;
 829                 }
 830 +       } else if (!ext4_htree_dx_locked(lck)) {
 831 +               struct ext4_dir_lock_data *ld = ext4_htree_lock_data(lck);
 832 +
 833 +               /* not well protected, require DX lock */
 834 +               ext4_htree_dx_need_lock(lck);
 835 +               at = frame > frames ? (frame - 1)->at : NULL;
 836 +
 837 +               /* NB: no risk of deadlock because it's just a try.
 838 +                *
 839 +                * NB: we check ld_count for twice, the first time before
 840 +                * having DX lock, the second time after holding DX lock.
 841 +                *
 842 +                * NB: We never free blocks for directory so far, which
 843 +                * means value returned by dx_get_count() should equal to
 844 +                * ld->ld_count if nobody split any DE-block under @at,
 845 +                * and ld->ld_at still points to valid dx_entry. */
 846 +               if ((ld->ld_count != dx_get_count(entries)) ||
 847 +                   !ext4_htree_dx_lock_try(lck, at) ||
 848 +                   (ld->ld_count != dx_get_count(entries))) {
 849 +                       restart = 1;
 850 +                       goto cleanup;
 851 +               }
 852 +               /* OK, I've got DX lock and nothing changed */
 853 +               frame->at = ld->ld_at;
 854         }
 855 -       de = do_split(handle, dir, &bh, frame, &fname->hinfo);
 856 +       de = do_split(handle, dir, &bh, frames, frame, &fname->hinfo, lck);
 857         if (IS_ERR(de)) {
 858                 err = PTR_ERR(de);
 859                 goto cleanup;
 860 @@ -2537,6 +2910,8 @@ again:
 861  journal_error:
 862         ext4_std_error(dir->i_sb, err); /* this is a no-op if err == 0 */
 863  cleanup:
 864 +       ext4_htree_dx_unlock(lck);
 865 +       ext4_htree_de_unlock(lck);
 866         brelse(bh);
 867         dx_release(frames);
 868         /* @restart is true means htree-path has been changed, we need to
 869 Index: linux-4.18.0-423.el8/fs/ext4/super.c
 870 ===================================================================
 871 --- linux-4.18.0-423.el8.orig/fs/ext4/super.c
 872 +++ linux-4.18.0-423.el8/fs/ext4/super.c
 873 @@ -1136,6 +1136,7 @@ static struct inode *ext4_alloc_inode(st
 874
 875         inode_set_iversion(&ei->vfs_inode, 1);
 876         spin_lock_init(&ei->i_raw_lock);
 877 +       sema_init(&ei->i_append_sem, 1);
 878         INIT_LIST_HEAD(&ei->i_prealloc_list);
 879         spin_lock_init(&ei->i_prealloc_lock);
 880         ext4_es_init_tree(&ei->i_es_tree);