ldiskfs/kernel_patches/patches/ubuntu18/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-4.15.0/fs/ext4/Makefile
  18 ===================================================================
  19 --- linux-4.15.0.orig/fs/ext4/Makefile
  20 +++ linux-4.15.0/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-4.15.0/fs/ext4/ext4.h
  30 ===================================================================
  31 --- linux-4.15.0.orig/fs/ext4/ext4.h
  32 +++ linux-4.15.0/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 @@ -932,6 +933,9 @@ 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 @@ -2114,6 +2118,72 @@ 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 @@ -2421,8 +2491,16 @@ 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-4.15.0/fs/ext4/htree_lock.c
 142 ===================================================================
 143 --- /dev/null
 144 +++ linux-4.15.0/fs/ext4/htree_lock.c
 145 @@ -0,0 +1,891 @@
 146 +/*
 147 + * fs/ext4/htree_lock.c
 148 + *
 149 + * Copyright (c) 2011, 2012, Intel Corporation.
 150 + *
 151 + * Author: Liang Zhen <liang@whamcloud.com>
 152 + */
 153 +#include <linux/jbd2.h>
 154 +#include <linux/hash.h>
 155 +#include <linux/module.h>
 156 +#include <linux/htree_lock.h>
 157 +
 158 +enum {
 159 +       HTREE_LOCK_BIT_EX       = (1 << HTREE_LOCK_EX),
 160 +       HTREE_LOCK_BIT_PW       = (1 << HTREE_LOCK_PW),
 161 +       HTREE_LOCK_BIT_PR       = (1 << HTREE_LOCK_PR),
 162 +       HTREE_LOCK_BIT_CW       = (1 << HTREE_LOCK_CW),
 163 +       HTREE_LOCK_BIT_CR       = (1 << HTREE_LOCK_CR),
 164 +};
 165 +
 166 +enum {
 167 +       HTREE_LOCK_COMPAT_EX    = 0,
 168 +       HTREE_LOCK_COMPAT_PW    = HTREE_LOCK_COMPAT_EX | HTREE_LOCK_BIT_CR,
 169 +       HTREE_LOCK_COMPAT_PR    = HTREE_LOCK_COMPAT_PW | HTREE_LOCK_BIT_PR,
 170 +       HTREE_LOCK_COMPAT_CW    = HTREE_LOCK_COMPAT_PW | HTREE_LOCK_BIT_CW,
 171 +       HTREE_LOCK_COMPAT_CR    = HTREE_LOCK_COMPAT_CW | HTREE_LOCK_BIT_PR |
 172 +                                 HTREE_LOCK_BIT_PW,
 173 +};
 174 +
 175 +static int htree_lock_compat[] = {
 176 +       [HTREE_LOCK_EX]         HTREE_LOCK_COMPAT_EX,
 177 +       [HTREE_LOCK_PW]         HTREE_LOCK_COMPAT_PW,
 178 +       [HTREE_LOCK_PR]         HTREE_LOCK_COMPAT_PR,
 179 +       [HTREE_LOCK_CW]         HTREE_LOCK_COMPAT_CW,
 180 +       [HTREE_LOCK_CR]         HTREE_LOCK_COMPAT_CR,
 181 +};
 182 +
 183 +/* max allowed htree-lock depth.
 184 + * We only need depth=3 for ext4 although user can have higher value. */
 185 +#define HTREE_LOCK_DEP_MAX     16
 186 +
 187 +#ifdef HTREE_LOCK_DEBUG
 188 +
 189 +static char *hl_name[] = {
 190 +       [HTREE_LOCK_EX]         "EX",
 191 +       [HTREE_LOCK_PW]         "PW",
 192 +       [HTREE_LOCK_PR]         "PR",
 193 +       [HTREE_LOCK_CW]         "CW",
 194 +       [HTREE_LOCK_CR]         "CR",
 195 +};
 196 +
 197 +/* lock stats */
 198 +struct htree_lock_node_stats {
 199 +       unsigned long long      blocked[HTREE_LOCK_MAX];
 200 +       unsigned long long      granted[HTREE_LOCK_MAX];
 201 +       unsigned long long      retried[HTREE_LOCK_MAX];
 202 +       unsigned long long      events;
 203 +};
 204 +
 205 +struct htree_lock_stats {
 206 +       struct htree_lock_node_stats    nodes[HTREE_LOCK_DEP_MAX];
 207 +       unsigned long long      granted[HTREE_LOCK_MAX];
 208 +       unsigned long long      blocked[HTREE_LOCK_MAX];
 209 +};
 210 +
 211 +static struct htree_lock_stats hl_stats;
 212 +
 213 +void htree_lock_stat_reset(void)
 214 +{
 215 +       memset(&hl_stats, 0, sizeof(hl_stats));
 216 +}
 217 +
 218 +void htree_lock_stat_print(int depth)
 219 +{
 220 +       int     i;
 221 +       int     j;
 222 +
 223 +       printk(KERN_DEBUG "HTREE LOCK STATS:\n");
 224 +       for (i = 0; i < HTREE_LOCK_MAX; i++) {
 225 +               printk(KERN_DEBUG "[%s]: G [%10llu], B [%10llu]\n",
 226 +                      hl_name[i], hl_stats.granted[i], hl_stats.blocked[i]);
 227 +       }
 228 +       for (i = 0; i < depth; i++) {
 229 +               printk(KERN_DEBUG "HTREE CHILD [%d] STATS:\n", i);
 230 +               for (j = 0; j < HTREE_LOCK_MAX; j++) {
 231 +                       printk(KERN_DEBUG
 232 +                               "[%s]: G [%10llu], B [%10llu], R [%10llu]\n",
 233 +                               hl_name[j], hl_stats.nodes[i].granted[j],
 234 +                               hl_stats.nodes[i].blocked[j],
 235 +                               hl_stats.nodes[i].retried[j]);
 236 +               }
 237 +       }
 238 +}
 239 +
 240 +#define lk_grant_inc(m)       do { hl_stats.granted[m]++; } while (0)
 241 +#define lk_block_inc(m)       do { hl_stats.blocked[m]++; } while (0)
 242 +#define ln_grant_inc(d, m)    do { hl_stats.nodes[d].granted[m]++; } while (0)
 243 +#define ln_block_inc(d, m)    do { hl_stats.nodes[d].blocked[m]++; } while (0)
 244 +#define ln_retry_inc(d, m)    do { hl_stats.nodes[d].retried[m]++; } while (0)
 245 +#define ln_event_inc(d)       do { hl_stats.nodes[d].events++; } while (0)
 246 +
 247 +#else /* !DEBUG */
 248 +
 249 +void htree_lock_stat_reset(void) {}
 250 +void htree_lock_stat_print(int depth) {}
 251 +
 252 +#define lk_grant_inc(m)              do {} while (0)
 253 +#define lk_block_inc(m)              do {} while (0)
 254 +#define ln_grant_inc(d, m)    do {} while (0)
 255 +#define ln_block_inc(d, m)    do {} while (0)
 256 +#define ln_retry_inc(d, m)    do {} while (0)
 257 +#define ln_event_inc(d)              do {} while (0)
 258 +
 259 +#endif /* DEBUG */
 260 +
 261 +EXPORT_SYMBOL(htree_lock_stat_reset);
 262 +EXPORT_SYMBOL(htree_lock_stat_print);
 263 +
 264 +#define HTREE_DEP_ROOT           (-1)
 265 +
 266 +#define htree_spin_lock(lhead, dep)                            \
 267 +       bit_spin_lock((dep) + 1, &(lhead)->lh_lock)
 268 +#define htree_spin_unlock(lhead, dep)                          \
 269 +       bit_spin_unlock((dep) + 1, &(lhead)->lh_lock)
 270 +
 271 +#define htree_key_event_ignore(child, ln)                      \
 272 +       (!((child)->lc_events & (1 << (ln)->ln_mode)))
 273 +
 274 +static int
 275 +htree_key_list_empty(struct htree_lock_node *ln)
 276 +{
 277 +       return list_empty(&ln->ln_major_list) && list_empty(&ln->ln_minor_list);
 278 +}
 279 +
 280 +static void
 281 +htree_key_list_del_init(struct htree_lock_node *ln)
 282 +{
 283 +       struct htree_lock_node *tmp = NULL;
 284 +
 285 +       if (!list_empty(&ln->ln_minor_list)) {
 286 +               tmp = list_entry(ln->ln_minor_list.next,
 287 +                                struct htree_lock_node, ln_minor_list);
 288 +               list_del_init(&ln->ln_minor_list);
 289 +       }
 290 +
 291 +       if (list_empty(&ln->ln_major_list))
 292 +               return;
 293 +
 294 +       if (tmp == NULL) { /* not on minor key list */
 295 +               list_del_init(&ln->ln_major_list);
 296 +       } else {
 297 +               BUG_ON(!list_empty(&tmp->ln_major_list));
 298 +               list_replace_init(&ln->ln_major_list, &tmp->ln_major_list);
 299 +       }
 300 +}
 301 +
 302 +static void
 303 +htree_key_list_replace_init(struct htree_lock_node *old,
 304 +                           struct htree_lock_node *new)
 305 +{
 306 +       if (!list_empty(&old->ln_major_list))
 307 +               list_replace_init(&old->ln_major_list, &new->ln_major_list);
 308 +
 309 +       if (!list_empty(&old->ln_minor_list))
 310 +               list_replace_init(&old->ln_minor_list, &new->ln_minor_list);
 311 +}
 312 +
 313 +static void
 314 +htree_key_event_enqueue(struct htree_lock_child *child,
 315 +                       struct htree_lock_node *ln, int dep, void *event)
 316 +{
 317 +       struct htree_lock_node *tmp;
 318 +
 319 +       /* NB: ALWAYS called holding lhead::lh_lock(dep) */
 320 +       BUG_ON(ln->ln_mode == HTREE_LOCK_NL);
 321 +       if (event == NULL || htree_key_event_ignore(child, ln))
 322 +               return;
 323 +
 324 +       /* shouldn't be a very long list */
 325 +       list_for_each_entry(tmp, &ln->ln_alive_list, ln_alive_list) {
 326 +               if (tmp->ln_mode == HTREE_LOCK_NL) {
 327 +                       ln_event_inc(dep);
 328 +                       if (child->lc_callback != NULL)
 329 +                               child->lc_callback(tmp->ln_ev_target, event);
 330 +               }
 331 +       }
 332 +}
 333 +
 334 +static int
 335 +htree_node_lock_enqueue(struct htree_lock *newlk, struct htree_lock *curlk,
 336 +                       unsigned dep, int wait, void *event)
 337 +{
 338 +       struct htree_lock_child *child = &newlk->lk_head->lh_children[dep];
 339 +       struct htree_lock_node *newln = &newlk->lk_nodes[dep];
 340 +       struct htree_lock_node *curln = &curlk->lk_nodes[dep];
 341 +
 342 +       /* NB: ALWAYS called holding lhead::lh_lock(dep) */
 343 +       /* NB: we only expect PR/PW lock mode at here, only these two modes are
 344 +        * allowed for htree_node_lock(asserted in htree_node_lock_internal),
 345 +        * NL is only used for listener, user can't directly require NL mode */
 346 +       if ((curln->ln_mode == HTREE_LOCK_NL) ||
 347 +           (curln->ln_mode != HTREE_LOCK_PW &&
 348 +            newln->ln_mode != HTREE_LOCK_PW)) {
 349 +               /* no conflict, attach it on granted list of @curlk */
 350 +               if (curln->ln_mode != HTREE_LOCK_NL) {
 351 +                       list_add(&newln->ln_granted_list,
 352 +                                &curln->ln_granted_list);
 353 +               } else {
 354 +                       /* replace key owner */
 355 +                       htree_key_list_replace_init(curln, newln);
 356 +               }
 357 +
 358 +               list_add(&newln->ln_alive_list, &curln->ln_alive_list);
 359 +               htree_key_event_enqueue(child, newln, dep, event);
 360 +               ln_grant_inc(dep, newln->ln_mode);
 361 +               return 1; /* still hold lh_lock */
 362 +       }
 363 +
 364 +       if (!wait) { /* can't grant and don't want to wait */
 365 +               ln_retry_inc(dep, newln->ln_mode);
 366 +               newln->ln_mode = HTREE_LOCK_INVAL;
 367 +               return -1; /* don't wait and just return -1 */
 368 +       }
 369 +
 370 +       newlk->lk_task = current;
 371 +       set_current_state(TASK_UNINTERRUPTIBLE);
 372 +       /* conflict, attach it on blocked list of curlk */
 373 +       list_add_tail(&newln->ln_blocked_list, &curln->ln_blocked_list);
 374 +       list_add(&newln->ln_alive_list, &curln->ln_alive_list);
 375 +       ln_block_inc(dep, newln->ln_mode);
 376 +
 377 +       htree_spin_unlock(newlk->lk_head, dep);
 378 +       /* wait to be given the lock */
 379 +       if (newlk->lk_task != NULL)
 380 +               schedule();
 381 +       /* granted, no doubt, wake up will set me RUNNING */
 382 +       if (event == NULL || htree_key_event_ignore(child, newln))
 383 +               return 0; /* granted without lh_lock */
 384 +
 385 +       htree_spin_lock(newlk->lk_head, dep);
 386 +       htree_key_event_enqueue(child, newln, dep, event);
 387 +       return 1; /* still hold lh_lock */
 388 +}
 389 +
 390 +/*
 391 + * get PR/PW access to particular tree-node according to @dep and @key,
 392 + * it will return -1 if @wait is false and can't immediately grant this lock.
 393 + * All listeners(HTREE_LOCK_NL) on @dep and with the same @key will get
 394 + * @event if it's not NULL.
 395 + * NB: ALWAYS called holding lhead::lh_lock
 396 + */
 397 +static int
 398 +htree_node_lock_internal(struct htree_lock_head *lhead, struct htree_lock *lck,
 399 +                        htree_lock_mode_t mode, u32 key, unsigned dep,
 400 +                        int wait, void *event)
 401 +{
 402 +       LIST_HEAD(list);
 403 +       struct htree_lock       *tmp;
 404 +       struct htree_lock       *tmp2;
 405 +       u16                     major;
 406 +       u16                     minor;
 407 +       u8                      reverse;
 408 +       u8                      ma_bits;
 409 +       u8                      mi_bits;
 410 +
 411 +       BUG_ON(mode != HTREE_LOCK_PW && mode != HTREE_LOCK_PR);
 412 +       BUG_ON(htree_node_is_granted(lck, dep));
 413 +
 414 +       key = hash_long(key, lhead->lh_hbits);
 415 +
 416 +       mi_bits = lhead->lh_hbits >> 1;
 417 +       ma_bits = lhead->lh_hbits - mi_bits;
 418 +
 419 +       lck->lk_nodes[dep].ln_major_key = major = key & ((1U << ma_bits) - 1);
 420 +       lck->lk_nodes[dep].ln_minor_key = minor = key >> ma_bits;
 421 +       lck->lk_nodes[dep].ln_mode = mode;
 422 +
 423 +       /*
 424 +        * The major key list is an ordered list, so searches are started
 425 +        * at the end of the list that is numerically closer to major_key,
 426 +        * so at most half of the list will be walked (for well-distributed
 427 +        * keys). The list traversal aborts early if the expected key
 428 +        * location is passed.
 429 +        */
 430 +       reverse = (major >= (1 << (ma_bits - 1)));
 431 +
 432 +       if (reverse) {
 433 +               list_for_each_entry_reverse(tmp,
 434 +                                       &lhead->lh_children[dep].lc_list,
 435 +                                       lk_nodes[dep].ln_major_list) {
 436 +                       if (tmp->lk_nodes[dep].ln_major_key == major) {
 437 +                               goto search_minor;
 438 +
 439 +                       } else if (tmp->lk_nodes[dep].ln_major_key < major) {
 440 +                               /* attach _after_ @tmp */
 441 +                               list_add(&lck->lk_nodes[dep].ln_major_list,
 442 +                                        &tmp->lk_nodes[dep].ln_major_list);
 443 +                               goto out_grant_major;
 444 +                       }
 445 +               }
 446 +
 447 +               list_add(&lck->lk_nodes[dep].ln_major_list,
 448 +                        &lhead->lh_children[dep].lc_list);
 449 +               goto out_grant_major;
 450 +
 451 +       } else {
 452 +               list_for_each_entry(tmp, &lhead->lh_children[dep].lc_list,
 453 +                                   lk_nodes[dep].ln_major_list) {
 454 +                       if (tmp->lk_nodes[dep].ln_major_key == major) {
 455 +                               goto search_minor;
 456 +
 457 +                       } else if (tmp->lk_nodes[dep].ln_major_key > major) {
 458 +                               /* insert _before_ @tmp */
 459 +                               list_add_tail(&lck->lk_nodes[dep].ln_major_list,
 460 +                                       &tmp->lk_nodes[dep].ln_major_list);
 461 +                               goto out_grant_major;
 462 +                       }
 463 +               }
 464 +
 465 +               list_add_tail(&lck->lk_nodes[dep].ln_major_list,
 466 +                             &lhead->lh_children[dep].lc_list);
 467 +               goto out_grant_major;
 468 +       }
 469 +
 470 + search_minor:
 471 +       /*
 472 +        * NB: minor_key list doesn't have a "head", @list is just a
 473 +        * temporary stub for helping list searching, make sure it's removed
 474 +        * after searching.
 475 +        * minor_key list is an ordered list too.
 476 +        */
 477 +       list_add_tail(&list, &tmp->lk_nodes[dep].ln_minor_list);
 478 +
 479 +       reverse = (minor >= (1 << (mi_bits - 1)));
 480 +
 481 +       if (reverse) {
 482 +               list_for_each_entry_reverse(tmp2, &list,
 483 +                                           lk_nodes[dep].ln_minor_list) {
 484 +                       if (tmp2->lk_nodes[dep].ln_minor_key == minor) {
 485 +                               goto out_enqueue;
 486 +
 487 +                       } else if (tmp2->lk_nodes[dep].ln_minor_key < minor) {
 488 +                               /* attach _after_ @tmp2 */
 489 +                               list_add(&lck->lk_nodes[dep].ln_minor_list,
 490 +                                        &tmp2->lk_nodes[dep].ln_minor_list);
 491 +                               goto out_grant_minor;
 492 +                       }
 493 +               }
 494 +
 495 +               list_add(&lck->lk_nodes[dep].ln_minor_list, &list);
 496 +
 497 +       } else {
 498 +               list_for_each_entry(tmp2, &list,
 499 +                                   lk_nodes[dep].ln_minor_list) {
 500 +                       if (tmp2->lk_nodes[dep].ln_minor_key == minor) {
 501 +                               goto out_enqueue;
 502 +
 503 +                       } else if (tmp2->lk_nodes[dep].ln_minor_key > minor) {
 504 +                               /* insert _before_ @tmp2 */
 505 +                               list_add_tail(&lck->lk_nodes[dep].ln_minor_list,
 506 +                                       &tmp2->lk_nodes[dep].ln_minor_list);
 507 +                               goto out_grant_minor;
 508 +                       }
 509 +               }
 510 +
 511 +               list_add_tail(&lck->lk_nodes[dep].ln_minor_list, &list);
 512 +       }
 513 +
 514 + out_grant_minor:
 515 +       if (list.next == &lck->lk_nodes[dep].ln_minor_list) {
 516 +               /* new lock @lck is the first one on minor_key list, which
 517 +                * means it has the smallest minor_key and it should
 518 +                * replace @tmp as minor_key owner */
 519 +               list_replace_init(&tmp->lk_nodes[dep].ln_major_list,
 520 +                                 &lck->lk_nodes[dep].ln_major_list);
 521 +       }
 522 +       /* remove the temporary head */
 523 +       list_del(&list);
 524 +
 525 + out_grant_major:
 526 +       ln_grant_inc(dep, lck->lk_nodes[dep].ln_mode);
 527 +       return 1; /* granted with holding lh_lock */
 528 +
 529 + out_enqueue:
 530 +       list_del(&list); /* remove temprary head */
 531 +       return htree_node_lock_enqueue(lck, tmp2, dep, wait, event);
 532 +}
 533 +
 534 +/*
 535 + * release the key of @lck at level @dep, and grant any blocked locks.
 536 + * caller will still listen on @key if @event is not NULL, which means
 537 + * caller can see a event (by event_cb) while granting any lock with
 538 + * the same key at level @dep.
 539 + * NB: ALWAYS called holding lhead::lh_lock
 540 + * NB: listener will not block anyone because listening mode is HTREE_LOCK_NL
 541 + */
 542 +static void
 543 +htree_node_unlock_internal(struct htree_lock_head *lhead,
 544 +                          struct htree_lock *curlk, unsigned dep, void *event)
 545 +{
 546 +       struct htree_lock_node  *curln = &curlk->lk_nodes[dep];
 547 +       struct htree_lock       *grtlk = NULL;
 548 +       struct htree_lock_node  *grtln;
 549 +       struct htree_lock       *poslk;
 550 +       struct htree_lock       *tmplk;
 551 +
 552 +       if (!htree_node_is_granted(curlk, dep))
 553 +               return;
 554 +
 555 +       if (!list_empty(&curln->ln_granted_list)) {
 556 +               /* there is another granted lock */
 557 +               grtlk = list_entry(curln->ln_granted_list.next,
 558 +                                  struct htree_lock,
 559 +                                  lk_nodes[dep].ln_granted_list);
 560 +               list_del_init(&curln->ln_granted_list);
 561 +       }
 562 +
 563 +       if (grtlk == NULL && !list_empty(&curln->ln_blocked_list)) {
 564 +               /*
 565 +                * @curlk is the only granted lock, so we confirmed:
 566 +                * a) curln is key owner (attached on major/minor_list),
 567 +                *    so if there is any blocked lock, it should be attached
 568 +                *    on curln->ln_blocked_list
 569 +                * b) we always can grant the first blocked lock
 570 +                */
 571 +               grtlk = list_entry(curln->ln_blocked_list.next,
 572 +                                  struct htree_lock,
 573 +                                  lk_nodes[dep].ln_blocked_list);
 574 +               BUG_ON(grtlk->lk_task == NULL);
 575 +               wake_up_process(grtlk->lk_task);
 576 +       }
 577 +
 578 +       if (event != NULL &&
 579 +           lhead->lh_children[dep].lc_events != HTREE_EVENT_DISABLE) {
 580 +               curln->ln_ev_target = event;
 581 +               curln->ln_mode = HTREE_LOCK_NL; /* listen! */
 582 +       } else {
 583 +               curln->ln_mode = HTREE_LOCK_INVAL;
 584 +       }
 585 +
 586 +       if (grtlk == NULL) { /* I must be the only one locking this key */
 587 +               struct htree_lock_node *tmpln;
 588 +
 589 +               BUG_ON(htree_key_list_empty(curln));
 590 +
 591 +               if (curln->ln_mode == HTREE_LOCK_NL) /* listening */
 592 +                       return;
 593 +
 594 +               /* not listening */
 595 +               if (list_empty(&curln->ln_alive_list)) { /* no more listener */
 596 +                       htree_key_list_del_init(curln);
 597 +                       return;
 598 +               }
 599 +
 600 +               tmpln = list_entry(curln->ln_alive_list.next,
 601 +                                  struct htree_lock_node, ln_alive_list);
 602 +
 603 +               BUG_ON(tmpln->ln_mode != HTREE_LOCK_NL);
 604 +
 605 +               htree_key_list_replace_init(curln, tmpln);
 606 +               list_del_init(&curln->ln_alive_list);
 607 +
 608 +               return;
 609 +       }
 610 +
 611 +       /* have a granted lock */
 612 +       grtln = &grtlk->lk_nodes[dep];
 613 +       if (!list_empty(&curln->ln_blocked_list)) {
 614 +               /* only key owner can be on both lists */
 615 +               BUG_ON(htree_key_list_empty(curln));
 616 +
 617 +               if (list_empty(&grtln->ln_blocked_list)) {
 618 +                       list_add(&grtln->ln_blocked_list,
 619 +                                &curln->ln_blocked_list);
 620 +               }
 621 +               list_del_init(&curln->ln_blocked_list);
 622 +       }
 623 +       /*
 624 +        * NB: this is the tricky part:
 625 +        * We have only two modes for child-lock (PR and PW), also,
 626 +        * only owner of the key (attached on major/minor_list) can be on
 627 +        * both blocked_list and granted_list, so @grtlk must be one
 628 +        * of these two cases:
 629 +        *
 630 +        * a) @grtlk is taken from granted_list, which means we've granted
 631 +        *    more than one lock so @grtlk has to be PR, the first blocked
 632 +        *    lock must be PW and we can't grant it at all.
 633 +        *    So even @grtlk is not owner of the key (empty blocked_list),
 634 +        *    we don't care because we can't grant any lock.
 635 +        * b) we just grant a new lock which is taken from head of blocked
 636 +        *    list, and it should be the first granted lock, and it should
 637 +        *    be the first one linked on blocked_list.
 638 +        *
 639 +        * Either way, we can get correct result by iterating blocked_list
 640 +        * of @grtlk, and don't have to bother on how to find out
 641 +        * owner of current key.
 642 +        */
 643 +       list_for_each_entry_safe(poslk, tmplk, &grtln->ln_blocked_list,
 644 +                                lk_nodes[dep].ln_blocked_list) {
 645 +               if (grtlk->lk_nodes[dep].ln_mode == HTREE_LOCK_PW ||
 646 +                   poslk->lk_nodes[dep].ln_mode == HTREE_LOCK_PW)
 647 +                       break;
 648 +               /* grant all readers */
 649 +               list_del_init(&poslk->lk_nodes[dep].ln_blocked_list);
 650 +               list_add(&poslk->lk_nodes[dep].ln_granted_list,
 651 +                        &grtln->ln_granted_list);
 652 +
 653 +               BUG_ON(poslk->lk_task == NULL);
 654 +               wake_up_process(poslk->lk_task);
 655 +       }
 656 +
 657 +       /* if @curln is the owner of this key, replace it with @grtln */
 658 +       if (!htree_key_list_empty(curln))
 659 +               htree_key_list_replace_init(curln, grtln);
 660 +
 661 +       if (curln->ln_mode == HTREE_LOCK_INVAL)
 662 +               list_del_init(&curln->ln_alive_list);
 663 +}
 664 +
 665 +/*
 666 + * it's just wrapper of htree_node_lock_internal, it returns 1 on granted
 667 + * and 0 only if @wait is false and can't grant it immediately
 668 + */
 669 +int
 670 +htree_node_lock_try(struct htree_lock *lck, htree_lock_mode_t mode,
 671 +                   u32 key, unsigned dep, int wait, void *event)
 672 +{
 673 +       struct htree_lock_head *lhead = lck->lk_head;
 674 +       int rc;
 675 +
 676 +       BUG_ON(dep >= lck->lk_depth);
 677 +       BUG_ON(lck->lk_mode == HTREE_LOCK_INVAL);
 678 +
 679 +       htree_spin_lock(lhead, dep);
 680 +       rc = htree_node_lock_internal(lhead, lck, mode, key, dep, wait, event);
 681 +       if (rc != 0)
 682 +               htree_spin_unlock(lhead, dep);
 683 +       return rc >= 0;
 684 +}
 685 +EXPORT_SYMBOL(htree_node_lock_try);
 686 +
 687 +/* it's wrapper of htree_node_unlock_internal */
 688 +void
 689 +htree_node_unlock(struct htree_lock *lck, unsigned dep, void *event)
 690 +{
 691 +       struct htree_lock_head *lhead = lck->lk_head;
 692 +
 693 +       BUG_ON(dep >= lck->lk_depth);
 694 +       BUG_ON(lck->lk_mode == HTREE_LOCK_INVAL);
 695 +
 696 +       htree_spin_lock(lhead, dep);
 697 +       htree_node_unlock_internal(lhead, lck, dep, event);
 698 +       htree_spin_unlock(lhead, dep);
 699 +}
 700 +EXPORT_SYMBOL(htree_node_unlock);
 701 +
 702 +/* stop listening on child-lock level @dep */
 703 +void
 704 +htree_node_stop_listen(struct htree_lock *lck, unsigned dep)
 705 +{
 706 +       struct htree_lock_node *ln = &lck->lk_nodes[dep];
 707 +       struct htree_lock_node *tmp;
 708 +
 709 +       BUG_ON(htree_node_is_granted(lck, dep));
 710 +       BUG_ON(!list_empty(&ln->ln_blocked_list));
 711 +       BUG_ON(!list_empty(&ln->ln_granted_list));
 712 +
 713 +       if (!htree_node_is_listening(lck, dep))
 714 +               return;
 715 +
 716 +       htree_spin_lock(lck->lk_head, dep);
 717 +       ln->ln_mode = HTREE_LOCK_INVAL;
 718 +       ln->ln_ev_target = NULL;
 719 +
 720 +       if (htree_key_list_empty(ln)) { /* not owner */
 721 +               list_del_init(&ln->ln_alive_list);
 722 +               goto out;
 723 +       }
 724 +
 725 +       /* I'm the owner... */
 726 +       if (list_empty(&ln->ln_alive_list)) { /* no more listener */
 727 +               htree_key_list_del_init(ln);
 728 +               goto out;
 729 +       }
 730 +
 731 +       tmp = list_entry(ln->ln_alive_list.next,
 732 +                        struct htree_lock_node, ln_alive_list);
 733 +
 734 +       BUG_ON(tmp->ln_mode != HTREE_LOCK_NL);
 735 +       htree_key_list_replace_init(ln, tmp);
 736 +       list_del_init(&ln->ln_alive_list);
 737 + out:
 738 +       htree_spin_unlock(lck->lk_head, dep);
 739 +}
 740 +EXPORT_SYMBOL(htree_node_stop_listen);
 741 +
 742 +/* release all child-locks if we have any */
 743 +static void
 744 +htree_node_release_all(struct htree_lock *lck)
 745 +{
 746 +       int     i;
 747 +
 748 +       for (i = 0; i < lck->lk_depth; i++) {
 749 +               if (htree_node_is_granted(lck, i))
 750 +                       htree_node_unlock(lck, i, NULL);
 751 +               else if (htree_node_is_listening(lck, i))
 752 +                       htree_node_stop_listen(lck, i);
 753 +       }
 754 +}
 755 +
 756 +/*
 757 + * obtain htree lock, it could be blocked inside if there's conflict
 758 + * with any granted or blocked lock and @wait is true.
 759 + * NB: ALWAYS called holding lhead::lh_lock
 760 + */
 761 +static int
 762 +htree_lock_internal(struct htree_lock *lck, int wait)
 763 +{
 764 +       struct htree_lock_head *lhead = lck->lk_head;
 765 +       int     granted = 0;
 766 +       int     blocked = 0;
 767 +       int     i;
 768 +
 769 +       for (i = 0; i < HTREE_LOCK_MAX; i++) {
 770 +               if (lhead->lh_ngranted[i] != 0)
 771 +                       granted |= 1 << i;
 772 +               if (lhead->lh_nblocked[i] != 0)
 773 +                       blocked |= 1 << i;
 774 +       }
 775 +       if ((htree_lock_compat[lck->lk_mode] & granted) != granted ||
 776 +           (htree_lock_compat[lck->lk_mode] & blocked) != blocked) {
 777 +               /* will block current lock even it just conflicts with any
 778 +                * other blocked lock, so lock like EX wouldn't starve */
 779 +               if (!wait)
 780 +                       return -1;
 781 +               lhead->lh_nblocked[lck->lk_mode]++;
 782 +               lk_block_inc(lck->lk_mode);
 783 +
 784 +               lck->lk_task = current;
 785 +               list_add_tail(&lck->lk_blocked_list, &lhead->lh_blocked_list);
 786 +
 787 +retry:
 788 +               set_current_state(TASK_UNINTERRUPTIBLE);
 789 +               htree_spin_unlock(lhead, HTREE_DEP_ROOT);
 790 +               /* wait to be given the lock */
 791 +               if (lck->lk_task != NULL)
 792 +                       schedule();
 793 +               /* granted, no doubt. wake up will set me RUNNING.
 794 +                * Since thread would be waken up accidentally,
 795 +                * so we need check lock whether granted or not again. */
 796 +               if (!list_empty(&lck->lk_blocked_list)) {
 797 +                       htree_spin_lock(lhead, HTREE_DEP_ROOT);
 798 +                       if (list_empty(&lck->lk_blocked_list)) {
 799 +                               htree_spin_unlock(lhead, HTREE_DEP_ROOT);
 800 +                               return 0;
 801 +                       }
 802 +                       goto retry;
 803 +               }
 804 +               return 0; /* without lh_lock */
 805 +       }
 806 +       lhead->lh_ngranted[lck->lk_mode]++;
 807 +       lk_grant_inc(lck->lk_mode);
 808 +       return 1;
 809 +}
 810 +
 811 +/* release htree lock. NB: ALWAYS called holding lhead::lh_lock */
 812 +static void
 813 +htree_unlock_internal(struct htree_lock *lck)
 814 +{
 815 +       struct htree_lock_head *lhead = lck->lk_head;
 816 +       struct htree_lock *tmp;
 817 +       struct htree_lock *tmp2;
 818 +       int granted = 0;
 819 +       int i;
 820 +
 821 +       BUG_ON(lhead->lh_ngranted[lck->lk_mode] == 0);
 822 +
 823 +       lhead->lh_ngranted[lck->lk_mode]--;
 824 +       lck->lk_mode = HTREE_LOCK_INVAL;
 825 +
 826 +       for (i = 0; i < HTREE_LOCK_MAX; i++) {
 827 +               if (lhead->lh_ngranted[i] != 0)
 828 +                       granted |= 1 << i;
 829 +       }
 830 +       list_for_each_entry_safe(tmp, tmp2,
 831 +                                &lhead->lh_blocked_list, lk_blocked_list) {
 832 +               /* conflict with any granted lock? */
 833 +               if ((htree_lock_compat[tmp->lk_mode] & granted) != granted)
 834 +                       break;
 835 +
 836 +               list_del_init(&tmp->lk_blocked_list);
 837 +
 838 +               BUG_ON(lhead->lh_nblocked[tmp->lk_mode] == 0);
 839 +
 840 +               lhead->lh_nblocked[tmp->lk_mode]--;
 841 +               lhead->lh_ngranted[tmp->lk_mode]++;
 842 +               granted |= 1 << tmp->lk_mode;
 843 +
 844 +               BUG_ON(tmp->lk_task == NULL);
 845 +               wake_up_process(tmp->lk_task);
 846 +       }
 847 +}
 848 +
 849 +/* it's wrapper of htree_lock_internal and exported interface.
 850 + * It always return 1 with granted lock if @wait is true, it can return 0
 851 + * if @wait is false and locking request can't be granted immediately */
 852 +int
 853 +htree_lock_try(struct htree_lock *lck, struct htree_lock_head *lhead,
 854 +              htree_lock_mode_t mode, int wait)
 855 +{
 856 +       int     rc;
 857 +
 858 +       BUG_ON(lck->lk_depth > lhead->lh_depth);
 859 +       BUG_ON(lck->lk_head != NULL);
 860 +       BUG_ON(lck->lk_task != NULL);
 861 +
 862 +       lck->lk_head = lhead;
 863 +       lck->lk_mode = mode;
 864 +
 865 +       htree_spin_lock(lhead, HTREE_DEP_ROOT);
 866 +       rc = htree_lock_internal(lck, wait);
 867 +       if (rc != 0)
 868 +               htree_spin_unlock(lhead, HTREE_DEP_ROOT);
 869 +       return rc >= 0;
 870 +}
 871 +EXPORT_SYMBOL(htree_lock_try);
 872 +
 873 +/* it's wrapper of htree_unlock_internal and exported interface.
 874 + * It will release all htree_node_locks and htree_lock */
 875 +void
 876 +htree_unlock(struct htree_lock *lck)
 877 +{
 878 +       BUG_ON(lck->lk_head == NULL);
 879 +       BUG_ON(lck->lk_mode == HTREE_LOCK_INVAL);
 880 +
 881 +       htree_node_release_all(lck);
 882 +
 883 +       htree_spin_lock(lck->lk_head, HTREE_DEP_ROOT);
 884 +       htree_unlock_internal(lck);
 885 +       htree_spin_unlock(lck->lk_head, HTREE_DEP_ROOT);
 886 +       lck->lk_head = NULL;
 887 +       lck->lk_task = NULL;
 888 +}
 889 +EXPORT_SYMBOL(htree_unlock);
 890 +
 891 +/* change lock mode */
 892 +void
 893 +htree_change_mode(struct htree_lock *lck, htree_lock_mode_t mode)
 894 +{
 895 +       BUG_ON(lck->lk_mode == HTREE_LOCK_INVAL);
 896 +       lck->lk_mode = mode;
 897 +}
 898 +EXPORT_SYMBOL(htree_change_mode);
 899 +
 900 +/* release htree lock, and lock it again with new mode.
 901 + * This function will first release all htree_node_locks and htree_lock,
 902 + * then try to gain htree_lock with new @mode.
 903 + * It always return 1 with granted lock if @wait is true, it can return 0
 904 + * if @wait is false and locking request can't be granted immediately */
 905 +int
 906 +htree_change_lock_try(struct htree_lock *lck, htree_lock_mode_t mode, int wait)
 907 +{
 908 +       struct htree_lock_head *lhead = lck->lk_head;
 909 +       int rc;
 910 +
 911 +       BUG_ON(lhead == NULL);
 912 +       BUG_ON(lck->lk_mode == mode);
 913 +       BUG_ON(lck->lk_mode == HTREE_LOCK_INVAL || mode == HTREE_LOCK_INVAL);
 914 +
 915 +       htree_node_release_all(lck);
 916 +
 917 +       htree_spin_lock(lhead, HTREE_DEP_ROOT);
 918 +       htree_unlock_internal(lck);
 919 +       lck->lk_mode = mode;
 920 +       rc = htree_lock_internal(lck, wait);
 921 +       if (rc != 0)
 922 +               htree_spin_unlock(lhead, HTREE_DEP_ROOT);
 923 +       return rc >= 0;
 924 +}
 925 +EXPORT_SYMBOL(htree_change_lock_try);
 926 +
 927 +/* create a htree_lock head with @depth levels (number of child-locks),
 928 + * it is a per resoruce structure */
 929 +struct htree_lock_head *
 930 +htree_lock_head_alloc(unsigned depth, unsigned hbits, unsigned priv)
 931 +{
 932 +       struct htree_lock_head *lhead;
 933 +       int  i;
 934 +
 935 +       if (depth > HTREE_LOCK_DEP_MAX) {
 936 +               printk(KERN_ERR "%d is larger than max htree_lock depth %d\n",
 937 +                       depth, HTREE_LOCK_DEP_MAX);
 938 +               return NULL;
 939 +       }
 940 +
 941 +       lhead = kzalloc(offsetof(struct htree_lock_head,
 942 +                                lh_children[depth]) + priv, GFP_NOFS);
 943 +       if (lhead == NULL)
 944 +               return NULL;
 945 +
 946 +       if (hbits < HTREE_HBITS_MIN)
 947 +               lhead->lh_hbits = HTREE_HBITS_MIN;
 948 +       else if (hbits > HTREE_HBITS_MAX)
 949 +               lhead->lh_hbits = HTREE_HBITS_MAX;
 950 +
 951 +       lhead->lh_lock = 0;
 952 +       lhead->lh_depth = depth;
 953 +       INIT_LIST_HEAD(&lhead->lh_blocked_list);
 954 +       if (priv > 0) {
 955 +               lhead->lh_private = (void *)lhead +
 956 +                       offsetof(struct htree_lock_head, lh_children[depth]);
 957 +       }
 958 +
 959 +       for (i = 0; i < depth; i++) {
 960 +               INIT_LIST_HEAD(&lhead->lh_children[i].lc_list);
 961 +               lhead->lh_children[i].lc_events = HTREE_EVENT_DISABLE;
 962 +       }
 963 +       return lhead;
 964 +}
 965 +EXPORT_SYMBOL(htree_lock_head_alloc);
 966 +
 967 +/* free the htree_lock head */
 968 +void
 969 +htree_lock_head_free(struct htree_lock_head *lhead)
 970 +{
 971 +       int     i;
 972 +
 973 +       BUG_ON(!list_empty(&lhead->lh_blocked_list));
 974 +       for (i = 0; i < lhead->lh_depth; i++)
 975 +               BUG_ON(!list_empty(&lhead->lh_children[i].lc_list));
 976 +       kfree(lhead);
 977 +}
 978 +EXPORT_SYMBOL(htree_lock_head_free);
 979 +
 980 +/* register event callback for @events of child-lock at level @dep */
 981 +void
 982 +htree_lock_event_attach(struct htree_lock_head *lhead, unsigned dep,
 983 +                       unsigned events, htree_event_cb_t callback)
 984 +{
 985 +       BUG_ON(lhead->lh_depth <= dep);
 986 +       lhead->lh_children[dep].lc_events = events;
 987 +       lhead->lh_children[dep].lc_callback = callback;
 988 +}
 989 +EXPORT_SYMBOL(htree_lock_event_attach);
 990 +
 991 +/* allocate a htree_lock, which is per-thread structure, @pbytes is some
 992 + * extra-bytes as private data for caller */
 993 +struct htree_lock *
 994 +htree_lock_alloc(unsigned depth, unsigned pbytes)
 995 +{
 996 +       struct htree_lock *lck;
 997 +       int i = offsetof(struct htree_lock, lk_nodes[depth]);
 998 +
 999 +       if (depth > HTREE_LOCK_DEP_MAX) {
1000 +               printk(KERN_ERR "%d is larger than max htree_lock depth %d\n",
1001 +                       depth, HTREE_LOCK_DEP_MAX);
1002 +               return NULL;
1003 +       }
1004 +       lck = kzalloc(i + pbytes, GFP_NOFS);
1005 +       if (lck == NULL)
1006 +               return NULL;
1007 +
1008 +       if (pbytes != 0)
1009 +               lck->lk_private = (void *)lck + i;
1010 +       lck->lk_mode = HTREE_LOCK_INVAL;
1011 +       lck->lk_depth = depth;
1012 +       INIT_LIST_HEAD(&lck->lk_blocked_list);
1013 +
1014 +       for (i = 0; i < depth; i++) {
1015 +               struct htree_lock_node *node = &lck->lk_nodes[i];
1016 +
1017 +               node->ln_mode = HTREE_LOCK_INVAL;
1018 +               INIT_LIST_HEAD(&node->ln_major_list);
1019 +               INIT_LIST_HEAD(&node->ln_minor_list);
1020 +               INIT_LIST_HEAD(&node->ln_alive_list);
1021 +               INIT_LIST_HEAD(&node->ln_blocked_list);
1022 +               INIT_LIST_HEAD(&node->ln_granted_list);
1023 +       }
1024 +
1025 +       return lck;
1026 +}
1027 +EXPORT_SYMBOL(htree_lock_alloc);
1028 +
1029 +/* free htree_lock node */
1030 +void
1031 +htree_lock_free(struct htree_lock *lck)
1032 +{
1033 +       BUG_ON(lck->lk_mode != HTREE_LOCK_INVAL);
1034 +       kfree(lck);
1035 +}
1036 +EXPORT_SYMBOL(htree_lock_free);
1037 Index: linux-4.15.0/fs/ext4/namei.c
1038 ===================================================================
1039 --- linux-4.15.0.orig/fs/ext4/namei.c
1040 +++ linux-4.15.0/fs/ext4/namei.c
1041 @@ -53,6 +53,7 @@ struct buffer_head *ext4_append(handle_t
1042                                         ext4_lblk_t *block)
1043  {
1044         struct buffer_head *bh;
1045 +       struct ext4_inode_info *ei = EXT4_I(inode);
1046         int err;
1047
1048         if (unlikely(EXT4_SB(inode->i_sb)->s_max_dir_size_kb &&
1049 @@ -60,15 +61,22 @@ struct buffer_head *ext4_append(handle_t
1050                       EXT4_SB(inode->i_sb)->s_max_dir_size_kb)))
1051                 return ERR_PTR(-ENOSPC);
1052
1053 +       /* with parallel dir operations all appends
1054 +       * have to be serialized -bzzz */
1055 +       down(&ei->i_append_sem);
1056 +
1057         *block = inode->i_size >> inode->i_sb->s_blocksize_bits;
1058
1059         bh = ext4_bread(handle, inode, *block, EXT4_GET_BLOCKS_CREATE);
1060 -       if (IS_ERR(bh))
1061 +       if (IS_ERR(bh)) {
1062 +               up(&ei->i_append_sem);
1063                 return bh;
1064 +       }
1065         inode->i_size += inode->i_sb->s_blocksize;
1066         EXT4_I(inode)->i_disksize = inode->i_size;
1067         BUFFER_TRACE(bh, "get_write_access");
1068         err = ext4_journal_get_write_access(handle, bh);
1069 +       up(&ei->i_append_sem);
1070         if (err) {
1071                 brelse(bh);
1072                 ext4_std_error(inode->i_sb, err);
1073 @@ -248,7 +256,8 @@ static unsigned dx_node_limit(struct ino
1074  static struct dx_frame *dx_probe(struct ext4_filename *fname,
1075                                  struct inode *dir,
1076                                  struct dx_hash_info *hinfo,
1077 -                                struct dx_frame *frame);
1078 +                                struct dx_frame *frame,
1079 +                                struct htree_lock *lck);
1080  static void dx_release(struct dx_frame *frames);
1081  static int dx_make_map(struct inode *dir, struct ext4_dir_entry_2 *de,
1082                        unsigned blocksize, struct dx_hash_info *hinfo,
1083 @@ -262,12 +271,13 @@ static void dx_insert_block(struct dx_fr
1084  static int ext4_htree_next_block(struct inode *dir, __u32 hash,
1085                                  struct dx_frame *frame,
1086                                  struct dx_frame *frames,
1087 -                                __u32 *start_hash);
1088 +                                __u32 *start_hash, struct htree_lock *lck);
1089  static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
1090                 struct ext4_filename *fname,
1091 -               struct ext4_dir_entry_2 **res_dir);
1092 +               struct ext4_dir_entry_2 **res_dir, struct htree_lock *lck);
1093  static int ext4_dx_add_entry(handle_t *handle, struct ext4_filename *fname,
1094 -                            struct inode *dir, struct inode *inode);
1095 +                            struct inode *dir, struct inode *inode,
1096 +                            struct htree_lock *lck);
1097
1098  /* checksumming functions */
1099  void initialize_dirent_tail(struct ext4_dir_entry_tail *t,
1100 @@ -730,6 +740,227 @@ struct stats dx_show_entries(struct dx_h
1101  }
1102  #endif /* DX_DEBUG */
1103
1104 +/* private data for htree_lock */
1105 +struct ext4_dir_lock_data {
1106 +       unsigned                ld_flags;  /* bits-map for lock types */
1107 +       unsigned                ld_count;  /* # entries of the last DX block */
1108 +       struct dx_entry         ld_at_entry; /* copy of leaf dx_entry */
1109 +       struct dx_entry         *ld_at;    /* position of leaf dx_entry */
1110 +};
1111 +
1112 +#define ext4_htree_lock_data(l)        ((struct ext4_dir_lock_data *)(l)->lk_private)
1113 +#define ext4_find_entry(dir, name, dirent, inline) \
1114 +                       __ext4_find_entry(dir, name, dirent, inline, NULL)
1115 +#define ext4_add_entry(handle, dentry, inode) \
1116 +                       __ext4_add_entry(handle, dentry, inode, NULL)
1117 +
1118 +/* NB: ext4_lblk_t is 32 bits so we use high bits to identify invalid blk */
1119 +#define EXT4_HTREE_NODE_CHANGED        (0xcafeULL << 32)
1120 +
1121 +static void ext4_htree_event_cb(void *target, void *event)
1122 +{
1123 +       u64 *block = (u64 *)target;
1124 +
1125 +       if (*block == dx_get_block((struct dx_entry *)event))
1126 +               *block = EXT4_HTREE_NODE_CHANGED;
1127 +}
1128 +
1129 +struct htree_lock_head *ext4_htree_lock_head_alloc(unsigned hbits)
1130 +{
1131 +       struct htree_lock_head *lhead;
1132 +
1133 +       lhead = htree_lock_head_alloc(EXT4_LK_MAX, hbits, 0);
1134 +       if (lhead != NULL) {
1135 +               htree_lock_event_attach(lhead, EXT4_LK_SPIN, HTREE_EVENT_WR,
1136 +                                       ext4_htree_event_cb);
1137 +       }
1138 +       return lhead;
1139 +}
1140 +EXPORT_SYMBOL(ext4_htree_lock_head_alloc);
1141 +
1142 +struct htree_lock *ext4_htree_lock_alloc(void)
1143 +{
1144 +       return htree_lock_alloc(EXT4_LK_MAX,
1145 +                               sizeof(struct ext4_dir_lock_data));
1146 +}
1147 +EXPORT_SYMBOL(ext4_htree_lock_alloc);
1148 +
1149 +static htree_lock_mode_t ext4_htree_mode(unsigned flags)
1150 +{
1151 +       switch (flags) {
1152 +       default: /* 0 or unknown flags require EX lock */
1153 +               return HTREE_LOCK_EX;
1154 +       case EXT4_HLOCK_READDIR:
1155 +               return HTREE_LOCK_PR;
1156 +       case EXT4_HLOCK_LOOKUP:
1157 +               return HTREE_LOCK_CR;
1158 +       case EXT4_HLOCK_DEL:
1159 +       case EXT4_HLOCK_ADD:
1160 +               return HTREE_LOCK_CW;
1161 +       }
1162 +}
1163 +
1164 +/* return PR for read-only operations, otherwise return EX */
1165 +static inline htree_lock_mode_t ext4_htree_safe_mode(unsigned flags)
1166 +{
1167 +       int writer = (flags & EXT4_LB_DE) == EXT4_LB_DE;
1168 +
1169 +       /* 0 requires EX lock */
1170 +       return (flags == 0 || writer) ? HTREE_LOCK_EX : HTREE_LOCK_PR;
1171 +}
1172 +
1173 +static int ext4_htree_safe_locked(struct htree_lock *lck)
1174 +{
1175 +       int writer;
1176 +
1177 +       if (lck == NULL || lck->lk_mode == HTREE_LOCK_EX)
1178 +               return 1;
1179 +
1180 +       writer = (ext4_htree_lock_data(lck)->ld_flags & EXT4_LB_DE) ==
1181 +                EXT4_LB_DE;
1182 +       if (writer) /* all readers & writers are excluded? */
1183 +               return lck->lk_mode == HTREE_LOCK_EX;
1184 +
1185 +       /* all writers are excluded? */
1186 +       return lck->lk_mode == HTREE_LOCK_PR ||
1187 +              lck->lk_mode == HTREE_LOCK_PW ||
1188 +              lck->lk_mode == HTREE_LOCK_EX;
1189 +}
1190 +
1191 +/* relock htree_lock with EX mode if it's change operation, otherwise
1192 + * relock it with PR mode. It's noop if PDO is disabled. */
1193 +static void ext4_htree_safe_relock(struct htree_lock *lck)
1194 +{
1195 +       if (!ext4_htree_safe_locked(lck)) {
1196 +               unsigned flags = ext4_htree_lock_data(lck)->ld_flags;
1197 +
1198 +               htree_change_lock(lck, ext4_htree_safe_mode(flags));
1199 +       }
1200 +}
1201 +
1202 +void ext4_htree_lock(struct htree_lock *lck, struct htree_lock_head *lhead,
1203 +                    struct inode *dir, unsigned flags)
1204 +{
1205 +       htree_lock_mode_t mode = is_dx(dir) ? ext4_htree_mode(flags) :
1206 +                                             ext4_htree_safe_mode(flags);
1207 +
1208 +       ext4_htree_lock_data(lck)->ld_flags = flags;
1209 +       htree_lock(lck, lhead, mode);
1210 +       if (!is_dx(dir))
1211 +               ext4_htree_safe_relock(lck); /* make sure it's safe locked */
1212 +}
1213 +EXPORT_SYMBOL(ext4_htree_lock);
1214 +
1215 +static int ext4_htree_node_lock(struct htree_lock *lck, struct dx_entry *at,
1216 +                               unsigned lmask, int wait, void *ev)
1217 +{
1218 +       u32     key = (at == NULL) ? 0 : dx_get_block(at);
1219 +       u32     mode;
1220 +
1221 +       /* NOOP if htree is well protected or caller doesn't require the lock */
1222 +       if (ext4_htree_safe_locked(lck) ||
1223 +          !(ext4_htree_lock_data(lck)->ld_flags & lmask))
1224 +               return 1;
1225 +
1226 +       mode = (ext4_htree_lock_data(lck)->ld_flags & lmask) == lmask ?
1227 +               HTREE_LOCK_PW : HTREE_LOCK_PR;
1228 +       while (1) {
1229 +               if (htree_node_lock_try(lck, mode, key, ffz(~lmask), wait, ev))
1230 +                       return 1;
1231 +               if (!(lmask & EXT4_LB_SPIN)) /* not a spinlock */
1232 +                       return 0;
1233 +               cpu_relax(); /* spin until granted */
1234 +       }
1235 +}
1236 +
1237 +static int ext4_htree_node_locked(struct htree_lock *lck, unsigned lmask)
1238 +{
1239 +       return ext4_htree_safe_locked(lck) ||
1240 +              htree_node_is_granted(lck, ffz(~lmask));
1241 +}
1242 +
1243 +static void ext4_htree_node_unlock(struct htree_lock *lck,
1244 +                                  unsigned lmask, void *buf)
1245 +{
1246 +       /* NB: it's safe to call mutiple times or even it's not locked */
1247 +       if (!ext4_htree_safe_locked(lck) &&
1248 +            htree_node_is_granted(lck, ffz(~lmask)))
1249 +               htree_node_unlock(lck, ffz(~lmask), buf);
1250 +}
1251 +
1252 +#define ext4_htree_dx_lock(lck, key)           \
1253 +       ext4_htree_node_lock(lck, key, EXT4_LB_DX, 1, NULL)
1254 +#define ext4_htree_dx_lock_try(lck, key)       \
1255 +       ext4_htree_node_lock(lck, key, EXT4_LB_DX, 0, NULL)
1256 +#define ext4_htree_dx_unlock(lck)              \
1257 +       ext4_htree_node_unlock(lck, EXT4_LB_DX, NULL)
1258 +#define ext4_htree_dx_locked(lck)              \
1259 +       ext4_htree_node_locked(lck, EXT4_LB_DX)
1260 +
1261 +static void ext4_htree_dx_need_lock(struct htree_lock *lck)
1262 +{
1263 +       struct ext4_dir_lock_data *ld;
1264 +
1265 +       if (ext4_htree_safe_locked(lck))
1266 +               return;
1267 +
1268 +       ld = ext4_htree_lock_data(lck);
1269 +       switch (ld->ld_flags) {
1270 +       default:
1271 +               return;
1272 +       case EXT4_HLOCK_LOOKUP:
1273 +               ld->ld_flags = EXT4_HLOCK_LOOKUP_SAFE;
1274 +               return;
1275 +       case EXT4_HLOCK_DEL:
1276 +               ld->ld_flags = EXT4_HLOCK_DEL_SAFE;
1277 +               return;
1278 +       case EXT4_HLOCK_ADD:
1279 +               ld->ld_flags = EXT4_HLOCK_SPLIT;
1280 +               return;
1281 +       }
1282 +}
1283 +
1284 +#define ext4_htree_de_lock(lck, key)           \
1285 +       ext4_htree_node_lock(lck, key, EXT4_LB_DE, 1, NULL)
1286 +#define ext4_htree_de_unlock(lck)              \
1287 +       ext4_htree_node_unlock(lck, EXT4_LB_DE, NULL)
1288 +
1289 +#define ext4_htree_spin_lock(lck, key, event)  \
1290 +       ext4_htree_node_lock(lck, key, EXT4_LB_SPIN, 0, event)
1291 +#define ext4_htree_spin_unlock(lck)            \
1292 +       ext4_htree_node_unlock(lck, EXT4_LB_SPIN, NULL)
1293 +#define ext4_htree_spin_unlock_listen(lck, p)  \
1294 +       ext4_htree_node_unlock(lck, EXT4_LB_SPIN, p)
1295 +
1296 +static void ext4_htree_spin_stop_listen(struct htree_lock *lck)
1297 +{
1298 +       if (!ext4_htree_safe_locked(lck) &&
1299 +           htree_node_is_listening(lck, ffz(~EXT4_LB_SPIN)))
1300 +               htree_node_stop_listen(lck, ffz(~EXT4_LB_SPIN));
1301 +}
1302 +
1303 +enum {
1304 +       DX_HASH_COL_IGNORE,     /* ignore collision while probing frames */
1305 +       DX_HASH_COL_YES,        /* there is collision and it does matter */
1306 +       DX_HASH_COL_NO,         /* there is no collision */
1307 +};
1308 +
1309 +static int dx_probe_hash_collision(struct htree_lock *lck,
1310 +                                  struct dx_entry *entries,
1311 +                                  struct dx_entry *at, u32 hash)
1312 +{
1313 +       if (!(lck && ext4_htree_lock_data(lck)->ld_flags & EXT4_LB_EXACT)) {
1314 +               return DX_HASH_COL_IGNORE; /* don't care about collision */
1315 +
1316 +       } else if (at == entries + dx_get_count(entries) - 1) {
1317 +               return DX_HASH_COL_IGNORE; /* not in any leaf of this DX */
1318 +
1319 +       } else { /* hash collision? */
1320 +               return ((dx_get_hash(at + 1) & ~1) == hash) ?
1321 +                       DX_HASH_COL_YES : DX_HASH_COL_NO;
1322 +       }
1323 +}
1324 +
1325  /*
1326   * Probe for a directory leaf block to search.
1327   *
1328 @@ -741,10 +972,11 @@ struct stats dx_show_entries(struct dx_h
1329   */
1330  static struct dx_frame *
1331  dx_probe(struct ext4_filename *fname, struct inode *dir,
1332 -        struct dx_hash_info *hinfo, struct dx_frame *frame_in)
1333 +        struct dx_hash_info *hinfo, struct dx_frame *frame_in,
1334 +        struct htree_lock *lck)
1335  {
1336         unsigned count, indirect;
1337 -       struct dx_entry *at, *entries, *p, *q, *m;
1338 +       struct dx_entry *at, *entries, *p, *q, *m, *dx = NULL;
1339         struct dx_root_info *info;
1340         struct dx_frame *frame = frame_in;
1341         struct dx_frame *ret_err = ERR_PTR(ERR_BAD_DX_DIR);
1342 @@ -806,8 +1038,15 @@ dx_probe(struct ext4_filename *fname, st
1343
1344         dxtrace(printk("Look up %x", hash));
1345         while (1) {
1346 +               if (indirect == 0) { /* the last index level */
1347 +                       /* NB: ext4_htree_dx_lock() could be noop if
1348 +                        * DX-lock flag is not set for current operation */
1349 +                       ext4_htree_dx_lock(lck, dx);
1350 +                       ext4_htree_spin_lock(lck, dx, NULL);
1351 +               }
1352                 count = dx_get_count(entries);
1353 -               if (!count || count > dx_get_limit(entries)) {
1354 +               if (count == 0 || count > dx_get_limit(entries)) {
1355 +                       ext4_htree_spin_unlock(lck); /* release spin */
1356                         ext4_warning_inode(dir,
1357                                            "dx entry: count %u beyond limit %u",
1358                                            count, dx_get_limit(entries));
1359 @@ -846,8 +1085,70 @@ dx_probe(struct ext4_filename *fname, st
1360                                dx_get_block(at)));
1361                 frame->entries = entries;
1362                 frame->at = at;
1363 -               if (!indirect--)
1364 +
1365 +               if (indirect == 0) { /* the last index level */
1366 +                       struct ext4_dir_lock_data *ld;
1367 +                       u64 myblock;
1368 +
1369 +                       /* By default we only lock DE-block, however, we will
1370 +                        * also lock the last level DX-block if:
1371 +                        * a) there is hash collision
1372 +                        *    we will set DX-lock flag (a few lines below)
1373 +                        *    and redo to lock DX-block
1374 +                        *    see detail in dx_probe_hash_collision()
1375 +                        * b) it's a retry from splitting
1376 +                        *    we need to lock the last level DX-block so nobody
1377 +                        *    else can split any leaf blocks under the same
1378 +                        *    DX-block, see detail in ext4_dx_add_entry()
1379 +                        */
1380 +                       if (ext4_htree_dx_locked(lck)) {
1381 +                               /* DX-block is locked, just lock DE-block
1382 +                                * and return */
1383 +                               ext4_htree_spin_unlock(lck);
1384 +                               if (!ext4_htree_safe_locked(lck))
1385 +                                       ext4_htree_de_lock(lck, frame->at);
1386 +                               return frame;
1387 +                       }
1388 +                       /* it's pdirop and no DX lock */
1389 +                       if (dx_probe_hash_collision(lck, entries, at, hash) ==
1390 +                           DX_HASH_COL_YES) {
1391 +                               /* found hash collision, set DX-lock flag
1392 +                                * and retry to abtain DX-lock */
1393 +                               ext4_htree_spin_unlock(lck);
1394 +                               ext4_htree_dx_need_lock(lck);
1395 +                               continue;
1396 +                       }
1397 +                       ld = ext4_htree_lock_data(lck);
1398 +                       /* because I don't lock DX, so @at can't be trusted
1399 +                        * after I release spinlock so I have to save it */
1400 +                       ld->ld_at = at;
1401 +                       ld->ld_at_entry = *at;
1402 +                       ld->ld_count = dx_get_count(entries);
1403 +
1404 +                       frame->at = &ld->ld_at_entry;
1405 +                       myblock = dx_get_block(at);
1406 +
1407 +                       /* NB: ordering locking */
1408 +                       ext4_htree_spin_unlock_listen(lck, &myblock);
1409 +                       /* other thread can split this DE-block because:
1410 +                        * a) I don't have lock for the DE-block yet
1411 +                        * b) I released spinlock on DX-block
1412 +                        * if it happened I can detect it by listening
1413 +                        * splitting event on this DE-block */
1414 +                       ext4_htree_de_lock(lck, frame->at);
1415 +                       ext4_htree_spin_stop_listen(lck);
1416 +
1417 +                       if (myblock == EXT4_HTREE_NODE_CHANGED) {
1418 +                               /* someone split this DE-block before
1419 +                                * I locked it, I need to retry and lock
1420 +                                * valid DE-block */
1421 +                               ext4_htree_de_unlock(lck);
1422 +                               continue;
1423 +                       }
1424                         return frame;
1425 +               }
1426 +               dx = at;
1427 +               indirect--;
1428                 frame++;
1429                 frame->bh = ext4_read_dirblock(dir, dx_get_block(at), INDEX);
1430                 if (IS_ERR(frame->bh)) {
1431 @@ -913,7 +1214,7 @@ static void dx_release(struct dx_frame *
1432  static int ext4_htree_next_block(struct inode *dir, __u32 hash,
1433                                  struct dx_frame *frame,
1434                                  struct dx_frame *frames,
1435 -                                __u32 *start_hash)
1436 +                                __u32 *start_hash, struct htree_lock *lck)
1437  {
1438         struct dx_frame *p;
1439         struct buffer_head *bh;
1440 @@ -928,12 +1229,22 @@ static int ext4_htree_next_block(struct
1441          * this loop, num_frames indicates the number of interior
1442          * nodes need to be read.
1443          */
1444 +       ext4_htree_de_unlock(lck);
1445         while (1) {
1446 -               if (++(p->at) < p->entries + dx_get_count(p->entries))
1447 -                       break;
1448 +               if (num_frames > 0 || ext4_htree_dx_locked(lck)) {
1449 +                       /* num_frames > 0 :
1450 +                        *   DX block
1451 +                        * ext4_htree_dx_locked:
1452 +                        *   frame->at is reliable pointer returned by dx_probe,
1453 +                        *   otherwise dx_probe already knew no collision */
1454 +                       if (++(p->at) < p->entries + dx_get_count(p->entries))
1455 +                               break;
1456 +               }
1457                 if (p == frames)
1458                         return 0;
1459                 num_frames++;
1460 +               if (num_frames == 1)
1461 +                       ext4_htree_dx_unlock(lck);
1462                 p--;
1463         }
1464
1465 @@ -956,6 +1267,13 @@ static int ext4_htree_next_block(struct
1466          * block so no check is necessary
1467          */
1468         while (num_frames--) {
1469 +               if (num_frames == 0) {
1470 +                       /* it's not always necessary, we just don't want to
1471 +                        * detect hash collision again */
1472 +                       ext4_htree_dx_need_lock(lck);
1473 +                       ext4_htree_dx_lock(lck, p->at);
1474 +               }
1475 +
1476                 bh = ext4_read_dirblock(dir, dx_get_block(p->at), INDEX);
1477                 if (IS_ERR(bh))
1478                         return PTR_ERR(bh);
1479 @@ -964,6 +1282,7 @@ static int ext4_htree_next_block(struct
1480                 p->bh = bh;
1481                 p->at = p->entries = ((struct dx_node *) bh->b_data)->entries;
1482         }
1483 +       ext4_htree_de_lock(lck, p->at);
1484         return 1;
1485  }
1486
1487 @@ -1111,10 +1430,10 @@ int ext4_htree_fill_tree(struct file *di
1488         }
1489         hinfo.hash = start_hash;
1490         hinfo.minor_hash = 0;
1491 -       frame = dx_probe(NULL, dir, &hinfo, frames);
1492 +       /* assume it's PR locked */
1493 +       frame = dx_probe(NULL, dir, &hinfo, frames, NULL);
1494         if (IS_ERR(frame))
1495                 return PTR_ERR(frame);
1496 -
1497         /* Add '.' and '..' from the htree header */
1498         if (!start_hash && !start_minor_hash) {
1499                 de = (struct ext4_dir_entry_2 *) frames[0].bh->b_data;
1500 @@ -1154,7 +1473,7 @@ int ext4_htree_fill_tree(struct file *di
1501                 count += ret;
1502                 hashval = ~0;
1503                 ret = ext4_htree_next_block(dir, HASH_NB_ALWAYS,
1504 -                                           frame, frames, &hashval);
1505 +                                           frame, frames, &hashval, NULL);
1506                 *next_hash = hashval;
1507                 if (ret < 0) {
1508                         err = ret;
1509 @@ -1346,10 +1665,10 @@ static int is_dx_internal_node(struct in
1510   * The returned buffer_head has ->b_count elevated.  The caller is expected
1511   * to brelse() it when appropriate.
1512   */
1513 -static struct buffer_head * ext4_find_entry (struct inode *dir,
1514 +struct buffer_head *__ext4_find_entry(struct inode *dir,
1515                                         const struct qstr *d_name,
1516                                         struct ext4_dir_entry_2 **res_dir,
1517 -                                       int *inlined)
1518 +                                       int *inlined, struct htree_lock *lck)
1519  {
1520         struct super_block *sb;
1521         struct buffer_head *bh_use[NAMEI_RA_SIZE];
1522 @@ -1398,7 +1717,7 @@ static struct buffer_head * ext4_find_en
1523                 goto restart;
1524         }
1525         if (is_dx(dir)) {
1526 -               ret = ext4_dx_find_entry(dir, &fname, res_dir);
1527 +               ret = ext4_dx_find_entry(dir, &fname, res_dir, lck);
1528                 /*
1529                  * On success, or if the error was file not found,
1530                  * return.  Otherwise, fall back to doing a search the
1531 @@ -1408,6 +1727,7 @@ static struct buffer_head * ext4_find_en
1532                         goto cleanup_and_exit;
1533                 dxtrace(printk(KERN_DEBUG "ext4_find_entry: dx failed, "
1534                                "falling back\n"));
1535 +               ext4_htree_safe_relock(lck);
1536         }
1537         nblocks = dir->i_size >> EXT4_BLOCK_SIZE_BITS(sb);
1538         if (!nblocks) {
1539 @@ -1495,10 +1815,12 @@ cleanup_and_exit:
1540         ext4_fname_free_filename(&fname);
1541         return ret;
1542  }
1543 +EXPORT_SYMBOL(__ext4_find_entry);
1544
1545  static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
1546                         struct ext4_filename *fname,
1547 -                       struct ext4_dir_entry_2 **res_dir)
1548 +                       struct ext4_dir_entry_2 **res_dir,
1549 +                       struct htree_lock *lck)
1550  {
1551         struct super_block * sb = dir->i_sb;
1552         struct dx_frame frames[EXT4_HTREE_LEVEL], *frame;
1553 @@ -1509,7 +1831,7 @@ static struct buffer_head * ext4_dx_find
1554  #ifdef CONFIG_EXT4_FS_ENCRYPTION
1555         *res_dir = NULL;
1556  #endif
1557 -       frame = dx_probe(fname, dir, NULL, frames);
1558 +       frame = dx_probe(fname, dir, NULL, frames, lck);
1559         if (IS_ERR(frame))
1560                 return (struct buffer_head *) frame;
1561         do {
1562 @@ -1531,7 +1853,7 @@ static struct buffer_head * ext4_dx_find
1563
1564                 /* Check to see if we should continue to search */
1565                 retval = ext4_htree_next_block(dir, fname->hinfo.hash, frame,
1566 -                                              frames, NULL);
1567 +                                              frames, NULL, lck);
1568                 if (retval < 0) {
1569                         ext4_warning_inode(dir,
1570                                 "error %d reading directory index block",
1571 @@ -1706,8 +2028,9 @@ static struct ext4_dir_entry_2* dx_pack_
1572   * Returns pointer to de in block into which the new entry will be inserted.
1573   */
1574  static struct ext4_dir_entry_2 *do_split(handle_t *handle, struct inode *dir,
1575 -                       struct buffer_head **bh,struct dx_frame *frame,
1576 -                       struct dx_hash_info *hinfo)
1577 +                       struct buffer_head **bh, struct dx_frame *frames,
1578 +                       struct dx_frame *frame, struct dx_hash_info *hinfo,
1579 +                       struct htree_lock *lck)
1580  {
1581         unsigned blocksize = dir->i_sb->s_blocksize;
1582         unsigned count, continued;
1583 @@ -1769,8 +2092,14 @@ static struct ext4_dir_entry_2 *do_split
1584                                         hash2, split, count-split));
1585
1586         /* Fancy dance to stay within two buffers */
1587 -       de2 = dx_move_dirents(data1, data2, map + split, count - split,
1588 -                             blocksize);
1589 +       if (hinfo->hash < hash2) {
1590 +               de2 = dx_move_dirents(data1, data2, map + split,
1591 +                                     count - split, blocksize);
1592 +       } else {
1593 +               /* make sure we will add entry to the same block which
1594 +                * we have already locked */
1595 +               de2 = dx_move_dirents(data1, data2, map, split, blocksize);
1596 +       }
1597         de = dx_pack_dirents(data1, blocksize);
1598         de->rec_len = ext4_rec_len_to_disk(data1 + (blocksize - csum_size) -
1599                                            (char *) de,
1600 @@ -1791,12 +2120,21 @@ static struct ext4_dir_entry_2 *do_split
1601         dxtrace(dx_show_leaf(dir, hinfo, (struct ext4_dir_entry_2 *) data2,
1602                         blocksize, 1));
1603
1604 -       /* Which block gets the new entry? */
1605 -       if (hinfo->hash >= hash2) {
1606 -               swap(*bh, bh2);
1607 -               de = de2;
1608 +       ext4_htree_spin_lock(lck, frame > frames ? (frame - 1)->at : NULL,
1609 +                            frame->at); /* notify block is being split */
1610 +       if (hinfo->hash < hash2) {
1611 +               dx_insert_block(frame, hash2 + continued, newblock);
1612 +
1613 +       } else {
1614 +               /* switch block number */
1615 +               dx_insert_block(frame, hash2 + continued,
1616 +                               dx_get_block(frame->at));
1617 +               dx_set_block(frame->at, newblock);
1618 +               (frame->at)++;
1619         }
1620 -       dx_insert_block(frame, hash2 + continued, newblock);
1621 +       ext4_htree_spin_unlock(lck);
1622 +       ext4_htree_dx_unlock(lck);
1623 +
1624         err = ext4_handle_dirty_dirent_node(handle, dir, bh2);
1625         if (err)
1626                 goto journal_error;
1627 @@ -2070,7 +2408,7 @@ static int make_indexed_dir(handle_t *ha
1628         if (retval)
1629                 goto out_frames;
1630
1631 -       de = do_split(handle,dir, &bh2, frame, &fname->hinfo);
1632 +       de = do_split(handle, dir, &bh2, frames, frame, &fname->hinfo, NULL);
1633         if (IS_ERR(de)) {
1634                 retval = PTR_ERR(de);
1635                 goto out_frames;
1636 @@ -2180,8 +2518,8 @@ out:
1637   * may not sleep between calling this and putting something into
1638   * the entry, as someone else might have used it while you slept.
1639   */
1640 -static int ext4_add_entry(handle_t *handle, struct dentry *dentry,
1641 -                         struct inode *inode)
1642 +int __ext4_add_entry(handle_t *handle, struct dentry *dentry,
1643 +                     struct inode *inode, struct htree_lock *lck)
1644  {
1645         struct inode *dir = d_inode(dentry->d_parent);
1646         struct buffer_head *bh = NULL;
1647 @@ -2222,9 +2560,10 @@ static int ext4_add_entry(handle_t *hand
1648                 if (dentry->d_name.len == 2 &&
1649                     memcmp(dentry->d_name.name, "..", 2) == 0)
1650                         return ext4_update_dotdot(handle, dentry, inode);
1651 -               retval = ext4_dx_add_entry(handle, &fname, dir, inode);
1652 +               retval = ext4_dx_add_entry(handle, &fname, dir, inode, lck);
1653                 if (!retval || (retval != ERR_BAD_DX_DIR))
1654                         goto out;
1655 +               ext4_htree_safe_relock(lck);
1656                 ext4_clear_inode_flag(dir, EXT4_INODE_INDEX);
1657                 dx_fallback++;
1658                 ext4_mark_inode_dirty(handle, dir);
1659 @@ -2274,12 +2613,14 @@ out:
1660                 ext4_set_inode_state(inode, EXT4_STATE_NEWENTRY);
1661         return retval;
1662  }
1663 +EXPORT_SYMBOL(__ext4_add_entry);
1664
1665  /*
1666   * Returns 0 for success, or a negative error value
1667   */
1668  static int ext4_dx_add_entry(handle_t *handle, struct ext4_filename *fname,
1669 -                            struct inode *dir, struct inode *inode)
1670 +                            struct inode *dir, struct inode *inode,
1671 +                            struct htree_lock *lck)
1672  {
1673         struct dx_frame frames[EXT4_HTREE_LEVEL], *frame;
1674         struct dx_entry *entries, *at;
1675 @@ -2291,7 +2632,7 @@ static int ext4_dx_add_entry(handle_t *h
1676
1677  again:
1678         restart = 0;
1679 -       frame = dx_probe(fname, dir, NULL, frames);
1680 +       frame = dx_probe(fname, dir, NULL, frames, lck);
1681         if (IS_ERR(frame))
1682                 return PTR_ERR(frame);
1683         entries = frame->entries;
1684 @@ -2326,6 +2667,11 @@ again:
1685                 struct dx_node *node2;
1686                 struct buffer_head *bh2;
1687
1688 +               if (!ext4_htree_safe_locked(lck)) { /* retry with EX lock */
1689 +                       ext4_htree_safe_relock(lck);
1690 +                       restart = 1;
1691 +                       goto cleanup;
1692 +               }
1693                 while (frame > frames) {
1694                         if (dx_get_count((frame - 1)->entries) <
1695                             dx_get_limit((frame - 1)->entries)) {
1696 @@ -2428,8 +2774,32 @@ again:
1697                         restart = 1;
1698                         goto journal_error;
1699                 }
1700 +       } else if (!ext4_htree_dx_locked(lck)) {
1701 +               struct ext4_dir_lock_data *ld = ext4_htree_lock_data(lck);
1702 +
1703 +               /* not well protected, require DX lock */
1704 +               ext4_htree_dx_need_lock(lck);
1705 +               at = frame > frames ? (frame - 1)->at : NULL;
1706 +
1707 +               /* NB: no risk of deadlock because it's just a try.
1708 +                *
1709 +                * NB: we check ld_count for twice, the first time before
1710 +                * having DX lock, the second time after holding DX lock.
1711 +                *
1712 +                * NB: We never free blocks for directory so far, which
1713 +                * means value returned by dx_get_count() should equal to
1714 +                * ld->ld_count if nobody split any DE-block under @at,
1715 +                * and ld->ld_at still points to valid dx_entry. */
1716 +               if ((ld->ld_count != dx_get_count(entries)) ||
1717 +                   !ext4_htree_dx_lock_try(lck, at) ||
1718 +                   (ld->ld_count != dx_get_count(entries))) {
1719 +                       restart = 1;
1720 +                       goto cleanup;
1721 +               }
1722 +               /* OK, I've got DX lock and nothing changed */
1723 +               frame->at = ld->ld_at;
1724         }
1725 -       de = do_split(handle, dir, &bh, frame, &fname->hinfo);
1726 +       de = do_split(handle, dir, &bh, frames, frame, &fname->hinfo, lck);
1727         if (IS_ERR(de)) {
1728                 err = PTR_ERR(de);
1729                 goto cleanup;
1730 @@ -2440,6 +2810,8 @@ again:
1731  journal_error:
1732         ext4_std_error(dir->i_sb, err); /* this is a no-op if err == 0 */
1733  cleanup:
1734 +       ext4_htree_dx_unlock(lck);
1735 +       ext4_htree_de_unlock(lck);
1736         brelse(bh);
1737         dx_release(frames);
1738         /* @restart is true means htree-path has been changed, we need to
1739 Index: linux-4.15.0/fs/ext4/super.c
1740 ===================================================================
1741 --- linux-4.15.0.orig/fs/ext4/super.c
1742 +++ linux-4.15.0/fs/ext4/super.c
1743 @@ -975,6 +975,7 @@ static struct inode *ext4_alloc_inode(st
1744
1745         ei->vfs_inode.i_version = 1;
1746         spin_lock_init(&ei->i_raw_lock);
1747 +       sema_init(&ei->i_append_sem, 1);
1748         INIT_LIST_HEAD(&ei->i_prealloc_list);
1749         spin_lock_init(&ei->i_prealloc_lock);
1750         ext4_es_init_tree(&ei->i_es_tree);
1751 Index: linux-4.15.0/include/linux/htree_lock.h
1752 ===================================================================
1753 --- /dev/null
1754 +++ linux-4.15.0/include/linux/htree_lock.h
1755 @@ -0,0 +1,187 @@
1756 +/*
1757 + * include/linux/htree_lock.h
1758 + *
1759 + * Copyright (c) 2011, 2012, Intel Corporation.
1760 + *
1761 + * Author: Liang Zhen <liang@whamcloud.com>
1762 + */
1763 +
1764 +/*
1765 + * htree lock
1766 + *
1767 + * htree_lock is an advanced lock, it can support five lock modes (concept is
1768 + * taken from DLM) and it's a sleeping lock.
1769 + *
1770 + * most common use case is:
1771 + * - create a htree_lock_head for data
1772 + * - each thread (contender) creates it's own htree_lock
1773 + * - contender needs to call htree_lock(lock_node, mode) to protect data and
1774 + *   call htree_unlock to release lock
1775 + *
1776 + * Also, there is advanced use-case which is more complex, user can have
1777 + * PW/PR lock on particular key, it's mostly used while user holding shared
1778 + * lock on the htree (CW, CR)
1779 + *
1780 + * htree_lock(lock_node, HTREE_LOCK_CR); lock the htree with CR
1781 + * htree_node_lock(lock_node, HTREE_LOCK_PR, key...); lock @key with PR
1782 + * ...
1783 + * htree_node_unlock(lock_node);; unlock the key
1784 + *
1785 + * Another tip is, we can have N-levels of this kind of keys, all we need to
1786 + * do is specifying N-levels while creating htree_lock_head, then we can
1787 + * lock/unlock a specific level by:
1788 + * htree_node_lock(lock_node, mode1, key1, level1...);
1789 + * do something;
1790 + * htree_node_lock(lock_node, mode1, key2, level2...);
1791 + * do something;
1792 + * htree_node_unlock(lock_node, level2);
1793 + * htree_node_unlock(lock_node, level1);
1794 + *
1795 + * NB: for multi-level, should be careful about locking order to avoid deadlock
1796 + */
1797 +
1798 +#ifndef _LINUX_HTREE_LOCK_H
1799 +#define _LINUX_HTREE_LOCK_H
1800 +
1801 +#include <linux/list.h>
1802 +#include <linux/spinlock.h>
1803 +#include <linux/sched.h>
1804 +
1805 +/*
1806 + * Lock Modes
1807 + * more details can be found here:
1808 + * http://en.wikipedia.org/wiki/Distributed_lock_manager
1809 + */
1810 +typedef enum {
1811 +       HTREE_LOCK_EX   = 0, /* exclusive lock: incompatible with all others */
1812 +       HTREE_LOCK_PW,       /* protected write: allows only CR users */
1813 +       HTREE_LOCK_PR,       /* protected read: allow PR, CR users */
1814 +       HTREE_LOCK_CW,       /* concurrent write: allow CR, CW users */
1815 +       HTREE_LOCK_CR,       /* concurrent read: allow all but EX users */
1816 +       HTREE_LOCK_MAX,      /* number of lock modes */
1817 +} htree_lock_mode_t;
1818 +
1819 +#define HTREE_LOCK_NL          HTREE_LOCK_MAX
1820 +#define HTREE_LOCK_INVAL       0xdead10c
1821 +
1822 +enum {
1823 +       HTREE_HBITS_MIN         = 2,
1824 +       HTREE_HBITS_DEF         = 14,
1825 +       HTREE_HBITS_MAX         = 32,
1826 +};
1827 +
1828 +enum {
1829 +       HTREE_EVENT_DISABLE     = (0),
1830 +       HTREE_EVENT_RD          = (1 << HTREE_LOCK_PR),
1831 +       HTREE_EVENT_WR          = (1 << HTREE_LOCK_PW),
1832 +       HTREE_EVENT_RDWR        = (HTREE_EVENT_RD | HTREE_EVENT_WR),
1833 +};
1834 +
1835 +struct htree_lock;
1836 +
1837 +typedef void (*htree_event_cb_t)(void *target, void *event);
1838 +
1839 +struct htree_lock_child {
1840 +       struct list_head        lc_list;        /* granted list */
1841 +       htree_event_cb_t        lc_callback;    /* event callback */
1842 +       unsigned                lc_events;      /* event types */
1843 +};
1844 +
1845 +struct htree_lock_head {
1846 +       unsigned long           lh_lock;        /* bits lock */
1847 +       /* blocked lock list (htree_lock) */
1848 +       struct list_head        lh_blocked_list;
1849 +       /* # key levels */
1850 +       u16                     lh_depth;
1851 +       /* hash bits for key and limit number of locks */
1852 +       u16                     lh_hbits;
1853 +       /* counters for blocked locks */
1854 +       u16                     lh_nblocked[HTREE_LOCK_MAX];
1855 +       /* counters for granted locks */
1856 +       u16                     lh_ngranted[HTREE_LOCK_MAX];
1857 +       /* private data */
1858 +       void                    *lh_private;
1859 +       /* array of children locks */
1860 +       struct htree_lock_child lh_children[0];
1861 +};
1862 +
1863 +/* htree_lock_node_t is child-lock for a specific key (ln_value) */
1864 +struct htree_lock_node {
1865 +       htree_lock_mode_t       ln_mode;
1866 +       /* major hash key */
1867 +       u16                     ln_major_key;
1868 +       /* minor hash key */
1869 +       u16                     ln_minor_key;
1870 +       struct list_head        ln_major_list;
1871 +       struct list_head        ln_minor_list;
1872 +       /* alive list, all locks (granted, blocked, listening) are on it */
1873 +       struct list_head        ln_alive_list;
1874 +       /* blocked list */
1875 +       struct list_head        ln_blocked_list;
1876 +       /* granted list */
1877 +       struct list_head        ln_granted_list;
1878 +       void                    *ln_ev_target;
1879 +};
1880 +
1881 +struct htree_lock {
1882 +       struct task_struct      *lk_task;
1883 +       struct htree_lock_head  *lk_head;
1884 +       void                    *lk_private;
1885 +       unsigned                lk_depth;
1886 +       htree_lock_mode_t       lk_mode;
1887 +       struct list_head        lk_blocked_list;
1888 +       struct htree_lock_node  lk_nodes[0];
1889 +};
1890 +
1891 +/* create a lock head, which stands for a resource */
1892 +struct htree_lock_head *htree_lock_head_alloc(unsigned depth,
1893 +                                             unsigned hbits, unsigned priv);
1894 +/* free a lock head */
1895 +void htree_lock_head_free(struct htree_lock_head *lhead);
1896 +/* register event callback for child lock at level @depth */
1897 +void htree_lock_event_attach(struct htree_lock_head *lhead, unsigned depth,
1898 +                            unsigned events, htree_event_cb_t callback);
1899 +/* create a lock handle, which stands for a thread */
1900 +struct htree_lock *htree_lock_alloc(unsigned depth, unsigned pbytes);
1901 +/* free a lock handle */
1902 +void htree_lock_free(struct htree_lock *lck);
1903 +/* lock htree, when @wait is true, 0 is returned if the lock can't
1904 + * be granted immediately */
1905 +int htree_lock_try(struct htree_lock *lck, struct htree_lock_head *lhead,
1906 +                  htree_lock_mode_t mode, int wait);
1907 +/* unlock htree */
1908 +void htree_unlock(struct htree_lock *lck);
1909 +/* unlock and relock htree with @new_mode */
1910 +int htree_change_lock_try(struct htree_lock *lck,
1911 +                         htree_lock_mode_t new_mode, int wait);
1912 +void htree_change_mode(struct htree_lock *lck, htree_lock_mode_t mode);
1913 +/* require child lock (key) of htree at level @dep, @event will be sent to all
1914 + * listeners on this @key while lock being granted */
1915 +int htree_node_lock_try(struct htree_lock *lck, htree_lock_mode_t mode,
1916 +                       u32 key, unsigned dep, int wait, void *event);
1917 +/* release child lock at level @dep, this lock will listen on it's key
1918 + * if @event isn't NULL, event_cb will be called against @lck while granting
1919 + * any other lock at level @dep with the same key */
1920 +void htree_node_unlock(struct htree_lock *lck, unsigned dep, void *event);
1921 +/* stop listening on child lock at level @dep */
1922 +void htree_node_stop_listen(struct htree_lock *lck, unsigned dep);
1923 +/* for debug */
1924 +void htree_lock_stat_print(int depth);
1925 +void htree_lock_stat_reset(void);
1926 +
1927 +#define htree_lock(lck, lh, mode)      htree_lock_try(lck, lh, mode, 1)
1928 +#define htree_change_lock(lck, mode)   htree_change_lock_try(lck, mode, 1)
1929 +
1930 +#define htree_lock_mode(lck)           ((lck)->lk_mode)
1931 +
1932 +#define htree_node_lock(lck, mode, key, dep)   \
1933 +       htree_node_lock_try(lck, mode, key, dep, 1, NULL)
1934 +/* this is only safe in thread context of lock owner */
1935 +#define htree_node_is_granted(lck, dep)                \
1936 +       ((lck)->lk_nodes[dep].ln_mode != HTREE_LOCK_INVAL && \
1937 +        (lck)->lk_nodes[dep].ln_mode != HTREE_LOCK_NL)
1938 +/* this is only safe in thread context of lock owner */
1939 +#define htree_node_is_listening(lck, dep)      \
1940 +       ((lck)->lk_nodes[dep].ln_mode == HTREE_LOCK_NL)
1941 +
1942 +#endif