ldiskfs/kernel_patches/patches/rhel8/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-4.18.0-80.1.2.el8_0/fs/ext4/Makefile
  26 ===================================================================
  27 --- linux-4.18.0-80.1.2.el8_0.orig/fs/ext4/Makefile
  28 +++ linux-4.18.0-80.1.2.el8_0/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-4.18.0-80.1.2.el8_0/fs/ext4/ext4.h
  38 ===================================================================
  39 --- linux-4.18.0-80.1.2.el8_0.orig/fs/ext4/ext4.h
  40 +++ linux-4.18.0-80.1.2.el8_0/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 @@ -934,6 +935,9 @@ 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 @@ -2109,6 +2113,72 @@ 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 @@ -2416,8 +2486,16 @@ 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_clear_inode_flag(inode, EXT4_INODE_INDEX);
 145 +#endif
 146  }
 147  static const unsigned char ext4_filetype_table[] = {
 148         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
 149 Index: linux-4.18.0-80.1.2.el8_0/fs/ext4/namei.c
 150 ===================================================================
 151 --- linux-4.18.0-80.1.2.el8_0.orig/fs/ext4/namei.c
 152 +++ linux-4.18.0-80.1.2.el8_0/fs/ext4/namei.c
 153 @@ -54,6 +54,7 @@ struct buffer_head *ext4_append(handle_t
 154                                         ext4_lblk_t *block)
 155  {
 156         struct buffer_head *bh;
 157 +       struct ext4_inode_info *ei = EXT4_I(inode);
 158         int err;
 159
 160         if (unlikely(EXT4_SB(inode->i_sb)->s_max_dir_size_kb &&
 161 @@ -61,15 +62,22 @@ struct buffer_head *ext4_append(handle_t
 162                       EXT4_SB(inode->i_sb)->s_max_dir_size_kb)))
 163                 return ERR_PTR(-ENOSPC);
 164
 165 +       /* with parallel dir operations all appends
 166 +       * have to be serialized -bzzz */
 167 +       down(&ei->i_append_sem);
 168 +
 169         *block = inode->i_size >> inode->i_sb->s_blocksize_bits;
 170
 171         bh = ext4_bread(handle, inode, *block, EXT4_GET_BLOCKS_CREATE);
 172 -       if (IS_ERR(bh))
 173 +       if (IS_ERR(bh)) {
 174 +               up(&ei->i_append_sem);
 175                 return bh;
 176 +       }
 177         inode->i_size += inode->i_sb->s_blocksize;
 178         EXT4_I(inode)->i_disksize = inode->i_size;
 179         BUFFER_TRACE(bh, "get_write_access");
 180         err = ext4_journal_get_write_access(handle, bh);
 181 +       up(&ei->i_append_sem);
 182         if (err) {
 183                 brelse(bh);
 184                 ext4_std_error(inode->i_sb, err);
 185 @@ -250,7 +258,8 @@ static unsigned dx_node_limit(struct ino
 186  static struct dx_frame *dx_probe(struct ext4_filename *fname,
 187                                  struct inode *dir,
 188                                  struct dx_hash_info *hinfo,
 189 -                                struct dx_frame *frame);
 190 +                                struct dx_frame *frame,
 191 +                                struct htree_lock *lck);
 192  static void dx_release(struct dx_frame *frames);
 193  static int dx_make_map(struct inode *dir, struct ext4_dir_entry_2 *de,
 194                        unsigned blocksize, struct dx_hash_info *hinfo,
 195 @@ -264,12 +273,13 @@ static void dx_insert_block(struct dx_fr
 196  static int ext4_htree_next_block(struct inode *dir, __u32 hash,
 197                                  struct dx_frame *frame,
 198                                  struct dx_frame *frames,
 199 -                                __u32 *start_hash);
 200 +                                __u32 *start_hash, struct htree_lock *lck);
 201  static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
 202                 struct ext4_filename *fname,
 203 -               struct ext4_dir_entry_2 **res_dir);
 204 +               struct ext4_dir_entry_2 **res_dir, struct htree_lock *lck);
 205  static int ext4_dx_add_entry(handle_t *handle, struct ext4_filename *fname,
 206 -                            struct inode *dir, struct inode *inode);
 207 +                            struct inode *dir, struct inode *inode,
 208 +                            struct htree_lock *lck);
 209
 210  /* checksumming functions */
 211  void initialize_dirent_tail(struct ext4_dir_entry_tail *t,
 212 @@ -733,6 +743,227 @@ struct stats dx_show_entries(struct dx_h
 213  }
 214  #endif /* DX_DEBUG */
 215
 216 +/* private data for htree_lock */
 217 +struct ext4_dir_lock_data {
 218 +       unsigned                ld_flags;  /* bits-map for lock types */
 219 +       unsigned                ld_count;  /* # entries of the last DX block */
 220 +       struct dx_entry         ld_at_entry; /* copy of leaf dx_entry */
 221 +       struct dx_entry         *ld_at;    /* position of leaf dx_entry */
 222 +};
 223 +
 224 +#define ext4_htree_lock_data(l)        ((struct ext4_dir_lock_data *)(l)->lk_private)
 225 +#define ext4_find_entry(dir, name, dirent, inline) \
 226 +                       __ext4_find_entry(dir, name, dirent, inline, NULL)
 227 +#define ext4_add_entry(handle, dentry, inode) \
 228 +                       __ext4_add_entry(handle, dentry, inode, NULL)
 229 +
 230 +/* NB: ext4_lblk_t is 32 bits so we use high bits to identify invalid blk */
 231 +#define EXT4_HTREE_NODE_CHANGED        (0xcafeULL << 32)
 232 +
 233 +static void ext4_htree_event_cb(void *target, void *event)
 234 +{
 235 +       u64 *block = (u64 *)target;
 236 +
 237 +       if (*block == dx_get_block((struct dx_entry *)event))
 238 +               *block = EXT4_HTREE_NODE_CHANGED;
 239 +}
 240 +
 241 +struct htree_lock_head *ext4_htree_lock_head_alloc(unsigned hbits)
 242 +{
 243 +       struct htree_lock_head *lhead;
 244 +
 245 +       lhead = htree_lock_head_alloc(EXT4_LK_MAX, hbits, 0);
 246 +       if (lhead != NULL) {
 247 +               htree_lock_event_attach(lhead, EXT4_LK_SPIN, HTREE_EVENT_WR,
 248 +                                       ext4_htree_event_cb);
 249 +       }
 250 +       return lhead;
 251 +}
 252 +EXPORT_SYMBOL(ext4_htree_lock_head_alloc);
 253 +
 254 +struct htree_lock *ext4_htree_lock_alloc(void)
 255 +{
 256 +       return htree_lock_alloc(EXT4_LK_MAX,
 257 +                               sizeof(struct ext4_dir_lock_data));
 258 +}
 259 +EXPORT_SYMBOL(ext4_htree_lock_alloc);
 260 +
 261 +static htree_lock_mode_t ext4_htree_mode(unsigned flags)
 262 +{
 263 +       switch (flags) {
 264 +       default: /* 0 or unknown flags require EX lock */
 265 +               return HTREE_LOCK_EX;
 266 +       case EXT4_HLOCK_READDIR:
 267 +               return HTREE_LOCK_PR;
 268 +       case EXT4_HLOCK_LOOKUP:
 269 +               return HTREE_LOCK_CR;
 270 +       case EXT4_HLOCK_DEL:
 271 +       case EXT4_HLOCK_ADD:
 272 +               return HTREE_LOCK_CW;
 273 +       }
 274 +}
 275 +
 276 +/* return PR for read-only operations, otherwise return EX */
 277 +static inline htree_lock_mode_t ext4_htree_safe_mode(unsigned flags)
 278 +{
 279 +       int writer = (flags & EXT4_LB_DE) == EXT4_LB_DE;
 280 +
 281 +       /* 0 requires EX lock */
 282 +       return (flags == 0 || writer) ? HTREE_LOCK_EX : HTREE_LOCK_PR;
 283 +}
 284 +
 285 +static int ext4_htree_safe_locked(struct htree_lock *lck)
 286 +{
 287 +       int writer;
 288 +
 289 +       if (lck == NULL || lck->lk_mode == HTREE_LOCK_EX)
 290 +               return 1;
 291 +
 292 +       writer = (ext4_htree_lock_data(lck)->ld_flags & EXT4_LB_DE) ==
 293 +                EXT4_LB_DE;
 294 +       if (writer) /* all readers & writers are excluded? */
 295 +               return lck->lk_mode == HTREE_LOCK_EX;
 296 +
 297 +       /* all writers are excluded? */
 298 +       return lck->lk_mode == HTREE_LOCK_PR ||
 299 +              lck->lk_mode == HTREE_LOCK_PW ||
 300 +              lck->lk_mode == HTREE_LOCK_EX;
 301 +}
 302 +
 303 +/* relock htree_lock with EX mode if it's change operation, otherwise
 304 + * relock it with PR mode. It's noop if PDO is disabled. */
 305 +static void ext4_htree_safe_relock(struct htree_lock *lck)
 306 +{
 307 +       if (!ext4_htree_safe_locked(lck)) {
 308 +               unsigned flags = ext4_htree_lock_data(lck)->ld_flags;
 309 +
 310 +               htree_change_lock(lck, ext4_htree_safe_mode(flags));
 311 +       }
 312 +}
 313 +
 314 +void ext4_htree_lock(struct htree_lock *lck, struct htree_lock_head *lhead,
 315 +                    struct inode *dir, unsigned flags)
 316 +{
 317 +       htree_lock_mode_t mode = is_dx(dir) ? ext4_htree_mode(flags) :
 318 +                                             ext4_htree_safe_mode(flags);
 319 +
 320 +       ext4_htree_lock_data(lck)->ld_flags = flags;
 321 +       htree_lock(lck, lhead, mode);
 322 +       if (!is_dx(dir))
 323 +               ext4_htree_safe_relock(lck); /* make sure it's safe locked */
 324 +}
 325 +EXPORT_SYMBOL(ext4_htree_lock);
 326 +
 327 +static int ext4_htree_node_lock(struct htree_lock *lck, struct dx_entry *at,
 328 +                               unsigned lmask, int wait, void *ev)
 329 +{
 330 +       u32     key = (at == NULL) ? 0 : dx_get_block(at);
 331 +       u32     mode;
 332 +
 333 +       /* NOOP if htree is well protected or caller doesn't require the lock */
 334 +       if (ext4_htree_safe_locked(lck) ||
 335 +          !(ext4_htree_lock_data(lck)->ld_flags & lmask))
 336 +               return 1;
 337 +
 338 +       mode = (ext4_htree_lock_data(lck)->ld_flags & lmask) == lmask ?
 339 +               HTREE_LOCK_PW : HTREE_LOCK_PR;
 340 +       while (1) {
 341 +               if (htree_node_lock_try(lck, mode, key, ffz(~lmask), wait, ev))
 342 +                       return 1;
 343 +               if (!(lmask & EXT4_LB_SPIN)) /* not a spinlock */
 344 +                       return 0;
 345 +               cpu_relax(); /* spin until granted */
 346 +       }
 347 +}
 348 +
 349 +static int ext4_htree_node_locked(struct htree_lock *lck, unsigned lmask)
 350 +{
 351 +       return ext4_htree_safe_locked(lck) ||
 352 +              htree_node_is_granted(lck, ffz(~lmask));
 353 +}
 354 +
 355 +static void ext4_htree_node_unlock(struct htree_lock *lck,
 356 +                                  unsigned lmask, void *buf)
 357 +{
 358 +       /* NB: it's safe to call mutiple times or even it's not locked */
 359 +       if (!ext4_htree_safe_locked(lck) &&
 360 +            htree_node_is_granted(lck, ffz(~lmask)))
 361 +               htree_node_unlock(lck, ffz(~lmask), buf);
 362 +}
 363 +
 364 +#define ext4_htree_dx_lock(lck, key)           \
 365 +       ext4_htree_node_lock(lck, key, EXT4_LB_DX, 1, NULL)
 366 +#define ext4_htree_dx_lock_try(lck, key)       \
 367 +       ext4_htree_node_lock(lck, key, EXT4_LB_DX, 0, NULL)
 368 +#define ext4_htree_dx_unlock(lck)              \
 369 +       ext4_htree_node_unlock(lck, EXT4_LB_DX, NULL)
 370 +#define ext4_htree_dx_locked(lck)              \
 371 +       ext4_htree_node_locked(lck, EXT4_LB_DX)
 372 +
 373 +static void ext4_htree_dx_need_lock(struct htree_lock *lck)
 374 +{
 375 +       struct ext4_dir_lock_data *ld;
 376 +
 377 +       if (ext4_htree_safe_locked(lck))
 378 +               return;
 379 +
 380 +       ld = ext4_htree_lock_data(lck);
 381 +       switch (ld->ld_flags) {
 382 +       default:
 383 +               return;
 384 +       case EXT4_HLOCK_LOOKUP:
 385 +               ld->ld_flags = EXT4_HLOCK_LOOKUP_SAFE;
 386 +               return;
 387 +       case EXT4_HLOCK_DEL:
 388 +               ld->ld_flags = EXT4_HLOCK_DEL_SAFE;
 389 +               return;
 390 +       case EXT4_HLOCK_ADD:
 391 +               ld->ld_flags = EXT4_HLOCK_SPLIT;
 392 +               return;
 393 +       }
 394 +}
 395 +
 396 +#define ext4_htree_de_lock(lck, key)           \
 397 +       ext4_htree_node_lock(lck, key, EXT4_LB_DE, 1, NULL)
 398 +#define ext4_htree_de_unlock(lck)              \
 399 +       ext4_htree_node_unlock(lck, EXT4_LB_DE, NULL)
 400 +
 401 +#define ext4_htree_spin_lock(lck, key, event)  \
 402 +       ext4_htree_node_lock(lck, key, EXT4_LB_SPIN, 0, event)
 403 +#define ext4_htree_spin_unlock(lck)            \
 404 +       ext4_htree_node_unlock(lck, EXT4_LB_SPIN, NULL)
 405 +#define ext4_htree_spin_unlock_listen(lck, p)  \
 406 +       ext4_htree_node_unlock(lck, EXT4_LB_SPIN, p)
 407 +
 408 +static void ext4_htree_spin_stop_listen(struct htree_lock *lck)
 409 +{
 410 +       if (!ext4_htree_safe_locked(lck) &&
 411 +           htree_node_is_listening(lck, ffz(~EXT4_LB_SPIN)))
 412 +               htree_node_stop_listen(lck, ffz(~EXT4_LB_SPIN));
 413 +}
 414 +
 415 +enum {
 416 +       DX_HASH_COL_IGNORE,     /* ignore collision while probing frames */
 417 +       DX_HASH_COL_YES,        /* there is collision and it does matter */
 418 +       DX_HASH_COL_NO,         /* there is no collision */
 419 +};
 420 +
 421 +static int dx_probe_hash_collision(struct htree_lock *lck,
 422 +                                  struct dx_entry *entries,
 423 +                                  struct dx_entry *at, u32 hash)
 424 +{
 425 +       if (!(lck && ext4_htree_lock_data(lck)->ld_flags & EXT4_LB_EXACT)) {
 426 +               return DX_HASH_COL_IGNORE; /* don't care about collision */
 427 +
 428 +       } else if (at == entries + dx_get_count(entries) - 1) {
 429 +               return DX_HASH_COL_IGNORE; /* not in any leaf of this DX */
 430 +
 431 +       } else { /* hash collision? */
 432 +               return ((dx_get_hash(at + 1) & ~1) == hash) ?
 433 +                       DX_HASH_COL_YES : DX_HASH_COL_NO;
 434 +       }
 435 +}
 436 +
 437  /*
 438   * Probe for a directory leaf block to search.
 439   *
 440 @@ -744,10 +975,11 @@ struct stats dx_show_entries(struct dx_h
 441   */
 442  static struct dx_frame *
 443  dx_probe(struct ext4_filename *fname, struct inode *dir,
 444 -        struct dx_hash_info *hinfo, struct dx_frame *frame_in)
 445 +        struct dx_hash_info *hinfo, struct dx_frame *frame_in,
 446 +        struct htree_lock *lck)
 447  {
 448         unsigned count, indirect;
 449 -       struct dx_entry *at, *entries, *p, *q, *m;
 450 +       struct dx_entry *at, *entries, *p, *q, *m, *dx = NULL;
 451         struct dx_root_info *info;
 452         struct dx_frame *frame = frame_in;
 453         struct dx_frame *ret_err = ERR_PTR(ERR_BAD_DX_DIR);
 454 @@ -809,8 +1041,15 @@ dx_probe(struct ext4_filename *fname, st
 455
 456         dxtrace(printk("Look up %x", hash));
 457         while (1) {
 458 +               if (indirect == 0) { /* the last index level */
 459 +                       /* NB: ext4_htree_dx_lock() could be noop if
 460 +                        * DX-lock flag is not set for current operation */
 461 +                       ext4_htree_dx_lock(lck, dx);
 462 +                       ext4_htree_spin_lock(lck, dx, NULL);
 463 +               }
 464                 count = dx_get_count(entries);
 465 -               if (!count || count > dx_get_limit(entries)) {
 466 +               if (count == 0 || count > dx_get_limit(entries)) {
 467 +                       ext4_htree_spin_unlock(lck); /* release spin */
 468                         ext4_warning_inode(dir,
 469                                            "dx entry: count %u beyond limit %u",
 470                                            count, dx_get_limit(entries));
 471 @@ -849,8 +1088,70 @@ dx_probe(struct ext4_filename *fname, st
 472                                dx_get_block(at)));
 473                 frame->entries = entries;
 474                 frame->at = at;
 475 -               if (!indirect--)
 476 +
 477 +               if (indirect == 0) { /* the last index level */
 478 +                       struct ext4_dir_lock_data *ld;
 479 +                       u64 myblock;
 480 +
 481 +                       /* By default we only lock DE-block, however, we will
 482 +                        * also lock the last level DX-block if:
 483 +                        * a) there is hash collision
 484 +                        *    we will set DX-lock flag (a few lines below)
 485 +                        *    and redo to lock DX-block
 486 +                        *    see detail in dx_probe_hash_collision()
 487 +                        * b) it's a retry from splitting
 488 +                        *    we need to lock the last level DX-block so nobody
 489 +                        *    else can split any leaf blocks under the same
 490 +                        *    DX-block, see detail in ext4_dx_add_entry()
 491 +                        */
 492 +                       if (ext4_htree_dx_locked(lck)) {
 493 +                               /* DX-block is locked, just lock DE-block
 494 +                                * and return */
 495 +                               ext4_htree_spin_unlock(lck);
 496 +                               if (!ext4_htree_safe_locked(lck))
 497 +                                       ext4_htree_de_lock(lck, frame->at);
 498 +                               return frame;
 499 +                       }
 500 +                       /* it's pdirop and no DX lock */
 501 +                       if (dx_probe_hash_collision(lck, entries, at, hash) ==
 502 +                           DX_HASH_COL_YES) {
 503 +                               /* found hash collision, set DX-lock flag
 504 +                                * and retry to abtain DX-lock */
 505 +                               ext4_htree_spin_unlock(lck);
 506 +                               ext4_htree_dx_need_lock(lck);
 507 +                               continue;
 508 +                       }
 509 +                       ld = ext4_htree_lock_data(lck);
 510 +                       /* because I don't lock DX, so @at can't be trusted
 511 +                        * after I release spinlock so I have to save it */
 512 +                       ld->ld_at = at;
 513 +                       ld->ld_at_entry = *at;
 514 +                       ld->ld_count = dx_get_count(entries);
 515 +
 516 +                       frame->at = &ld->ld_at_entry;
 517 +                       myblock = dx_get_block(at);
 518 +
 519 +                       /* NB: ordering locking */
 520 +                       ext4_htree_spin_unlock_listen(lck, &myblock);
 521 +                       /* other thread can split this DE-block because:
 522 +                        * a) I don't have lock for the DE-block yet
 523 +                        * b) I released spinlock on DX-block
 524 +                        * if it happened I can detect it by listening
 525 +                        * splitting event on this DE-block */
 526 +                       ext4_htree_de_lock(lck, frame->at);
 527 +                       ext4_htree_spin_stop_listen(lck);
 528 +
 529 +                       if (myblock == EXT4_HTREE_NODE_CHANGED) {
 530 +                               /* someone split this DE-block before
 531 +                                * I locked it, I need to retry and lock
 532 +                                * valid DE-block */
 533 +                               ext4_htree_de_unlock(lck);
 534 +                               continue;
 535 +                       }
 536                         return frame;
 537 +               }
 538 +               dx = at;
 539 +               indirect--;
 540                 frame++;
 541                 frame->bh = ext4_read_dirblock(dir, dx_get_block(at), INDEX);
 542                 if (IS_ERR(frame->bh)) {
 543 @@ -916,7 +1217,7 @@ static void dx_release(struct dx_frame *
 544  static int ext4_htree_next_block(struct inode *dir, __u32 hash,
 545                                  struct dx_frame *frame,
 546                                  struct dx_frame *frames,
 547 -                                __u32 *start_hash)
 548 +                                __u32 *start_hash, struct htree_lock *lck)
 549  {
 550         struct dx_frame *p;
 551         struct buffer_head *bh;
 552 @@ -931,12 +1232,22 @@ static int ext4_htree_next_block(struct
 553          * this loop, num_frames indicates the number of interior
 554          * nodes need to be read.
 555          */
 556 +       ext4_htree_de_unlock(lck);
 557         while (1) {
 558 -               if (++(p->at) < p->entries + dx_get_count(p->entries))
 559 -                       break;
 560 +               if (num_frames > 0 || ext4_htree_dx_locked(lck)) {
 561 +                       /* num_frames > 0 :
 562 +                        *   DX block
 563 +                        * ext4_htree_dx_locked:
 564 +                        *   frame->at is reliable pointer returned by dx_probe,
 565 +                        *   otherwise dx_probe already knew no collision */
 566 +                       if (++(p->at) < p->entries + dx_get_count(p->entries))
 567 +                               break;
 568 +               }
 569                 if (p == frames)
 570                         return 0;
 571                 num_frames++;
 572 +               if (num_frames == 1)
 573 +                       ext4_htree_dx_unlock(lck);
 574                 p--;
 575         }
 576
 577 @@ -959,6 +1270,13 @@ static int ext4_htree_next_block(struct
 578          * block so no check is necessary
 579          */
 580         while (num_frames--) {
 581 +               if (num_frames == 0) {
 582 +                       /* it's not always necessary, we just don't want to
 583 +                        * detect hash collision again */
 584 +                       ext4_htree_dx_need_lock(lck);
 585 +                       ext4_htree_dx_lock(lck, p->at);
 586 +               }
 587 +
 588                 bh = ext4_read_dirblock(dir, dx_get_block(p->at), INDEX);
 589                 if (IS_ERR(bh))
 590                         return PTR_ERR(bh);
 591 @@ -967,6 +1285,7 @@ static int ext4_htree_next_block(struct
 592                 p->bh = bh;
 593                 p->at = p->entries = ((struct dx_node *) bh->b_data)->entries;
 594         }
 595 +       ext4_htree_de_lock(lck, p->at);
 596         return 1;
 597  }
 598
 599 @@ -1114,10 +1433,10 @@ int ext4_htree_fill_tree(struct file *di
 600         }
 601         hinfo.hash = start_hash;
 602         hinfo.minor_hash = 0;
 603 -       frame = dx_probe(NULL, dir, &hinfo, frames);
 604 +       /* assume it's PR locked */
 605 +       frame = dx_probe(NULL, dir, &hinfo, frames, NULL);
 606         if (IS_ERR(frame))
 607                 return PTR_ERR(frame);
 608 -
 609         /* Add '.' and '..' from the htree header */
 610         if (!start_hash && !start_minor_hash) {
 611                 de = (struct ext4_dir_entry_2 *) frames[0].bh->b_data;
 612 @@ -1157,7 +1476,7 @@ int ext4_htree_fill_tree(struct file *di
 613                 count += ret;
 614                 hashval = ~0;
 615                 ret = ext4_htree_next_block(dir, HASH_NB_ALWAYS,
 616 -                                           frame, frames, &hashval);
 617 +                                           frame, frames, &hashval, NULL);
 618                 *next_hash = hashval;
 619                 if (ret < 0) {
 620                         err = ret;
 621 @@ -1349,10 +1668,10 @@ static int is_dx_internal_node(struct in
 622   * The returned buffer_head has ->b_count elevated.  The caller is expected
 623   * to brelse() it when appropriate.
 624   */
 625 -static struct buffer_head * ext4_find_entry (struct inode *dir,
 626 +struct buffer_head *__ext4_find_entry(struct inode *dir,
 627                                         const struct qstr *d_name,
 628                                         struct ext4_dir_entry_2 **res_dir,
 629 -                                       int *inlined)
 630 +                                       int *inlined, struct htree_lock *lck)
 631  {
 632         struct super_block *sb;
 633         struct buffer_head *bh_use[NAMEI_RA_SIZE];
 634 @@ -1401,7 +1720,7 @@ static struct buffer_head * ext4_find_en
 635                 goto restart;
 636         }
 637         if (is_dx(dir)) {
 638 -               ret = ext4_dx_find_entry(dir, &fname, res_dir);
 639 +               ret = ext4_dx_find_entry(dir, &fname, res_dir, lck);
 640                 /*
 641                  * On success, or if the error was file not found,
 642                  * return.  Otherwise, fall back to doing a search the
 643 @@ -1411,6 +1730,7 @@ static struct buffer_head * ext4_find_en
 644                         goto cleanup_and_exit;
 645                 dxtrace(printk(KERN_DEBUG "ext4_find_entry: dx failed, "
 646                                "falling back\n"));
 647 +               ext4_htree_safe_relock(lck);
 648                 ret = NULL;
 649         }
 650         nblocks = dir->i_size >> EXT4_BLOCK_SIZE_BITS(sb);
 651 @@ -1499,10 +1819,12 @@ cleanup_and_exit:
 652         ext4_fname_free_filename(&fname);
 653         return ret;
 654  }
 655 +EXPORT_SYMBOL(__ext4_find_entry);
 656
 657  static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
 658                         struct ext4_filename *fname,
 659 -                       struct ext4_dir_entry_2 **res_dir)
 660 +                       struct ext4_dir_entry_2 **res_dir,
 661 +                       struct htree_lock *lck)
 662  {
 663         struct super_block * sb = dir->i_sb;
 664         struct dx_frame frames[EXT4_HTREE_LEVEL], *frame;
 665 @@ -1513,7 +1835,7 @@ static struct buffer_head * ext4_dx_find
 666  #ifdef CONFIG_EXT4_FS_ENCRYPTION
 667         *res_dir = NULL;
 668  #endif
 669 -       frame = dx_probe(fname, dir, NULL, frames);
 670 +       frame = dx_probe(fname, dir, NULL, frames, lck);
 671         if (IS_ERR(frame))
 672                 return (struct buffer_head *) frame;
 673         do {
 674 @@ -1535,7 +1857,7 @@ static struct buffer_head * ext4_dx_find
 675
 676                 /* Check to see if we should continue to search */
 677                 retval = ext4_htree_next_block(dir, fname->hinfo.hash, frame,
 678 -                                              frames, NULL);
 679 +                                              frames, NULL, lck);
 680                 if (retval < 0) {
 681                         ext4_warning_inode(dir,
 682                                 "error %d reading directory index block",
 683 @@ -1710,8 +2032,9 @@ static struct ext4_dir_entry_2* dx_pack_
 684   * Returns pointer to de in block into which the new entry will be inserted.
 685   */
 686  static struct ext4_dir_entry_2 *do_split(handle_t *handle, struct inode *dir,
 687 -                       struct buffer_head **bh,struct dx_frame *frame,
 688 -                       struct dx_hash_info *hinfo)
 689 +                       struct buffer_head **bh, struct dx_frame *frames,
 690 +                       struct dx_frame *frame, struct dx_hash_info *hinfo,
 691 +                       struct htree_lock *lck)
 692  {
 693         unsigned blocksize = dir->i_sb->s_blocksize;
 694         unsigned count, continued;
 695 @@ -1773,8 +2096,14 @@ static struct ext4_dir_entry_2 *do_split
 696                                         hash2, split, count-split));
 697
 698         /* Fancy dance to stay within two buffers */
 699 -       de2 = dx_move_dirents(data1, data2, map + split, count - split,
 700 -                             blocksize);
 701 +       if (hinfo->hash < hash2) {
 702 +               de2 = dx_move_dirents(data1, data2, map + split,
 703 +                                     count - split, blocksize);
 704 +       } else {
 705 +               /* make sure we will add entry to the same block which
 706 +                * we have already locked */
 707 +               de2 = dx_move_dirents(data1, data2, map, split, blocksize);
 708 +       }
 709         de = dx_pack_dirents(data1, blocksize);
 710         de->rec_len = ext4_rec_len_to_disk(data1 + (blocksize - csum_size) -
 711                                            (char *) de,
 712 @@ -1795,12 +2124,21 @@ static struct ext4_dir_entry_2 *do_split
 713         dxtrace(dx_show_leaf(dir, hinfo, (struct ext4_dir_entry_2 *) data2,
 714                         blocksize, 1));
 715
 716 -       /* Which block gets the new entry? */
 717 -       if (hinfo->hash >= hash2) {
 718 -               swap(*bh, bh2);
 719 -               de = de2;
 720 +       ext4_htree_spin_lock(lck, frame > frames ? (frame - 1)->at : NULL,
 721 +                            frame->at); /* notify block is being split */
 722 +       if (hinfo->hash < hash2) {
 723 +               dx_insert_block(frame, hash2 + continued, newblock);
 724 +
 725 +       } else {
 726 +               /* switch block number */
 727 +               dx_insert_block(frame, hash2 + continued,
 728 +                               dx_get_block(frame->at));
 729 +               dx_set_block(frame->at, newblock);
 730 +               (frame->at)++;
 731         }
 732 -       dx_insert_block(frame, hash2 + continued, newblock);
 733 +       ext4_htree_spin_unlock(lck);
 734 +       ext4_htree_dx_unlock(lck);
 735 +
 736         err = ext4_handle_dirty_dirent_node(handle, dir, bh2);
 737         if (err)
 738                 goto journal_error;
 739 @@ -2074,7 +2412,7 @@ static int make_indexed_dir(handle_t *ha
 740         if (retval)
 741                 goto out_frames;
 742
 743 -       de = do_split(handle,dir, &bh2, frame, &fname->hinfo);
 744 +       de = do_split(handle, dir, &bh2, frames, frame, &fname->hinfo, NULL);
 745         if (IS_ERR(de)) {
 746                 retval = PTR_ERR(de);
 747                 goto out_frames;
 748 @@ -2184,8 +2522,8 @@ out:
 749   * may not sleep between calling this and putting something into
 750   * the entry, as someone else might have used it while you slept.
 751   */
 752 -static int ext4_add_entry(handle_t *handle, struct dentry *dentry,
 753 -                         struct inode *inode)
 754 +int __ext4_add_entry(handle_t *handle, struct dentry *dentry,
 755 +                     struct inode *inode, struct htree_lock *lck)
 756  {
 757         struct inode *dir = d_inode(dentry->d_parent);
 758         struct buffer_head *bh = NULL;
 759 @@ -2226,9 +2564,10 @@ static int ext4_add_entry(handle_t *hand
 760                 if (dentry->d_name.len == 2 &&
 761                     memcmp(dentry->d_name.name, "..", 2) == 0)
 762                         return ext4_update_dotdot(handle, dentry, inode);
 763 -               retval = ext4_dx_add_entry(handle, &fname, dir, inode);
 764 +               retval = ext4_dx_add_entry(handle, &fname, dir, inode, lck);
 765                 if (!retval || (retval != ERR_BAD_DX_DIR))
 766                         goto out;
 767 +               ext4_htree_safe_relock(lck);
 768                 ext4_clear_inode_flag(dir, EXT4_INODE_INDEX);
 769                 dx_fallback++;
 770                 ext4_mark_inode_dirty(handle, dir);
 771 @@ -2278,12 +2617,14 @@ out:
 772                 ext4_set_inode_state(inode, EXT4_STATE_NEWENTRY);
 773         return retval;
 774  }
 775 +EXPORT_SYMBOL(__ext4_add_entry);
 776
 777  /*
 778   * Returns 0 for success, or a negative error value
 779   */
 780  static int ext4_dx_add_entry(handle_t *handle, struct ext4_filename *fname,
 781 -                            struct inode *dir, struct inode *inode)
 782 +                            struct inode *dir, struct inode *inode,
 783 +                            struct htree_lock *lck)
 784  {
 785         struct dx_frame frames[EXT4_HTREE_LEVEL], *frame;
 786         struct dx_entry *entries, *at;
 787 @@ -2295,7 +2636,7 @@ static int ext4_dx_add_entry(handle_t *h
 788
 789  again:
 790         restart = 0;
 791 -       frame = dx_probe(fname, dir, NULL, frames);
 792 +       frame = dx_probe(fname, dir, NULL, frames, lck);
 793         if (IS_ERR(frame))
 794                 return PTR_ERR(frame);
 795         entries = frame->entries;
 796 @@ -2330,6 +2671,11 @@ again:
 797                 struct dx_node *node2;
 798                 struct buffer_head *bh2;
 799
 800 +               if (!ext4_htree_safe_locked(lck)) { /* retry with EX lock */
 801 +                       ext4_htree_safe_relock(lck);
 802 +                       restart = 1;
 803 +                       goto cleanup;
 804 +               }
 805                 while (frame > frames) {
 806                         if (dx_get_count((frame - 1)->entries) <
 807                             dx_get_limit((frame - 1)->entries)) {
 808 @@ -2432,8 +2778,32 @@ again:
 809                         restart = 1;
 810                         goto journal_error;
 811                 }
 812 +       } else if (!ext4_htree_dx_locked(lck)) {
 813 +               struct ext4_dir_lock_data *ld = ext4_htree_lock_data(lck);
 814 +
 815 +               /* not well protected, require DX lock */
 816 +               ext4_htree_dx_need_lock(lck);
 817 +               at = frame > frames ? (frame - 1)->at : NULL;
 818 +
 819 +               /* NB: no risk of deadlock because it's just a try.
 820 +                *
 821 +                * NB: we check ld_count for twice, the first time before
 822 +                * having DX lock, the second time after holding DX lock.
 823 +                *
 824 +                * NB: We never free blocks for directory so far, which
 825 +                * means value returned by dx_get_count() should equal to
 826 +                * ld->ld_count if nobody split any DE-block under @at,
 827 +                * and ld->ld_at still points to valid dx_entry. */
 828 +               if ((ld->ld_count != dx_get_count(entries)) ||
 829 +                   !ext4_htree_dx_lock_try(lck, at) ||
 830 +                   (ld->ld_count != dx_get_count(entries))) {
 831 +                       restart = 1;
 832 +                       goto cleanup;
 833 +               }
 834 +               /* OK, I've got DX lock and nothing changed */
 835 +               frame->at = ld->ld_at;
 836         }
 837 -       de = do_split(handle, dir, &bh, frame, &fname->hinfo);
 838 +       de = do_split(handle, dir, &bh, frames, frame, &fname->hinfo, lck);
 839         if (IS_ERR(de)) {
 840                 err = PTR_ERR(de);
 841                 goto cleanup;
 842 @@ -2444,6 +2814,8 @@ again:
 843  journal_error:
 844         ext4_std_error(dir->i_sb, err); /* this is a no-op if err == 0 */
 845  cleanup:
 846 +       ext4_htree_dx_unlock(lck);
 847 +       ext4_htree_de_unlock(lck);
 848         brelse(bh);
 849         dx_release(frames);
 850         /* @restart is true means htree-path has been changed, we need to
 851 Index: linux-4.18.0-80.1.2.el8_0/fs/ext4/super.c
 852 ===================================================================
 853 --- linux-4.18.0-80.1.2.el8_0.orig/fs/ext4/super.c
 854 +++ linux-4.18.0-80.1.2.el8_0/fs/ext4/super.c
 855 @@ -1009,6 +1009,7 @@ static struct inode *ext4_alloc_inode(st
 856
 857         inode_set_iversion(&ei->vfs_inode, 1);
 858         spin_lock_init(&ei->i_raw_lock);
 859 +       sema_init(&ei->i_append_sem, 1);
 860         INIT_LIST_HEAD(&ei->i_prealloc_list);
 861         spin_lock_init(&ei->i_prealloc_lock);
 862         ext4_es_init_tree(&ei->i_es_tree);