ldiskfs/kernel_patches/patches/ubuntu14+16/ext4-pdirop.patch

   1 diff --git a/fs/ext4/Makefile b/fs/ext4/Makefile
   2 index f52cf54..3f16939 100644
   3 --- a/fs/ext4/Makefile
   4 +++ b/fs/ext4/Makefile
   5 @@ -6,6 +6,7 @@ obj-$(CONFIG_EXT4_FS) += ext4.o
   6
   7  ext4-y := balloc.o bitmap.o dir.o file.o fsync.o ialloc.o inode.o page-io.o \
   8                 ioctl.o namei.o super.o symlink.o hash.o resize.o extents.o \
   9 +               htree_lock.o \
  10                 ext4_jbd2.o migrate.o mballoc.o block_validity.o move_extent.o \
  11                 mmp.o indirect.o extents_status.o xattr.o xattr_user.o \
  12                 xattr_trusted.o inline.o readpage.o sysfs.o
  13 diff --git a/fs/ext4/ext4.h b/fs/ext4/ext4.h
  14 index 2d22f1a..005c9b3 100644
  15 --- a/fs/ext4/ext4.h
  16 +++ b/fs/ext4/ext4.h
  17 @@ -28,6 +28,7 @@
  18  #include <linux/timer.h>
  19  #include <linux/version.h>
  20  #include <linux/wait.h>
  21 +#include <linux/htree_lock.h>
  22  #include <linux/blockgroup_lock.h>
  23  #include <linux/percpu_counter.h>
  24  #include <linux/ratelimit.h>
  25 @@ -880,6 +881,9 @@ struct ext4_inode_info {
  26         __u32   i_dtime;
  27         ext4_fsblk_t    i_file_acl;
  28
  29 +       /* following fields for parallel directory operations -bzzz */
  30 +       struct semaphore i_append_sem;
  31 +
  32         /*
  33          * i_block_group is the number of the block group which contains
  34          * this file's inode.  Constant across the lifetime of the inode,
  35 @@ -2086,6 +2090,71 @@ struct dx_hash_info
  36         u32             *seed;
  37  };
  38
  39 +/* assume name-hash is protected by upper layer */
  40 +#define EXT4_HTREE_LOCK_HASH   0
  41 +
  42 +enum ext4_pdo_lk_types {
  43 +#if EXT4_HTREE_LOCK_HASH
  44 +       EXT4_LK_HASH,
  45 +#endif
  46 +       EXT4_LK_DX,             /* index block */
  47 +       EXT4_LK_DE,             /* directory entry block */
  48 +       EXT4_LK_SPIN,           /* spinlock */
  49 +       EXT4_LK_MAX,
  50 +};
  51 +
  52 +/* read-only bit */
  53 +#define EXT4_LB_RO(b)          (1 << (b))
  54 +/* read + write, high bits for writer */
  55 +#define EXT4_LB_RW(b)          ((1 << (b)) | (1 << (EXT4_LK_MAX + (b))))
  56 +
  57 +enum ext4_pdo_lock_bits {
  58 +       /* DX lock bits */
  59 +       EXT4_LB_DX_RO           = EXT4_LB_RO(EXT4_LK_DX),
  60 +       EXT4_LB_DX              = EXT4_LB_RW(EXT4_LK_DX),
  61 +       /* DE lock bits */
  62 +       EXT4_LB_DE_RO           = EXT4_LB_RO(EXT4_LK_DE),
  63 +       EXT4_LB_DE              = EXT4_LB_RW(EXT4_LK_DE),
  64 +       /* DX spinlock bits */
  65 +       EXT4_LB_SPIN_RO         = EXT4_LB_RO(EXT4_LK_SPIN),
  66 +       EXT4_LB_SPIN            = EXT4_LB_RW(EXT4_LK_SPIN),
  67 +       /* accurate searching */
  68 +       EXT4_LB_EXACT           = EXT4_LB_RO(EXT4_LK_MAX << 1),
  69 +};
  70 +
  71 +enum ext4_pdo_lock_opc {
  72 +       /* external */
  73 +       EXT4_HLOCK_READDIR      = (EXT4_LB_DE_RO | EXT4_LB_DX_RO),
  74 +       EXT4_HLOCK_LOOKUP       = (EXT4_LB_DE_RO | EXT4_LB_SPIN_RO |
  75 +                                  EXT4_LB_EXACT),
  76 +       EXT4_HLOCK_DEL          = (EXT4_LB_DE | EXT4_LB_SPIN_RO |
  77 +                                  EXT4_LB_EXACT),
  78 +       EXT4_HLOCK_ADD          = (EXT4_LB_DE | EXT4_LB_SPIN_RO),
  79 +
  80 +       /* internal */
  81 +       EXT4_HLOCK_LOOKUP_SAFE  = (EXT4_LB_DE_RO | EXT4_LB_DX_RO |
  82 +                                  EXT4_LB_EXACT),
  83 +       EXT4_HLOCK_DEL_SAFE     = (EXT4_LB_DE | EXT4_LB_DX_RO | EXT4_LB_EXACT),
  84 +       EXT4_HLOCK_SPLIT        = (EXT4_LB_DE | EXT4_LB_DX | EXT4_LB_SPIN),
  85 +};
  86 +
  87 +extern struct htree_lock_head *ext4_htree_lock_head_alloc(unsigned hbits);
  88 +#define ext4_htree_lock_head_free(lhead)       htree_lock_head_free(lhead)
  89 +
  90 +extern struct htree_lock *ext4_htree_lock_alloc(void);
  91 +#define ext4_htree_lock_free(lck)              htree_lock_free(lck)
  92 +
  93 +extern void ext4_htree_lock(struct htree_lock *lck,
  94 +                           struct htree_lock_head *lhead,
  95 +                           struct inode *dir, unsigned flags);
  96 +#define ext4_htree_unlock(lck)                  htree_unlock(lck)
  97 +
  98 +extern struct buffer_head *__ext4_find_entry(struct inode *dir,
  99 +                                       const struct qstr *d_name,
 100 +                                       struct ext4_dir_entry_2 **res_dir,
 101 +                                       int *inlined, struct htree_lock *lck);
 102 +extern int __ext4_add_entry(handle_t *handle, struct dentry *dentry,
 103 +                     struct inode *inode, struct htree_lock *lck);
 104
 105  /* 32 and 64 bit signed EOF for dx directories */
 106  #define EXT4_HTREE_EOF_32BIT   ((1UL  << (32 - 1)) - 1)
 107 @@ -2475,8 +2544,16 @@ int ext4_insert_dentry(struct inode *dir,
 108                        struct ext4_filename *fname, void *data);
 109  static inline void ext4_update_dx_flag(struct inode *inode)
 110  {
 111 +       /* Disable it for ldiskfs, because going from a DX directory to
 112 +        * a non-DX directory while it is in use will completely break
 113 +        * the htree-locking.
 114 +        * If we really want to support this operation in the future,
 115 +        * we need to exclusively lock the directory at here which will
 116 +        * increase complexity of code */
 117 +#if 0
 118         if (!ext4_has_feature_dir_index(inode->i_sb))
 119                 ext4_clear_inode_flag(inode, EXT4_INODE_INDEX);
 120 +#endif
 121  }
 122  static unsigned char ext4_filetype_table[] = {
 123         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
 124 diff --git a/fs/ext4/htree_lock.c b/fs/ext4/htree_lock.c
 125 new file mode 100644
 126 index 0000000..99e7375
 127 --- /dev/null
 128 +++ b/fs/ext4/htree_lock.c
 129 @@ -0,0 +1,880 @@
 130 +/*
 131 + * fs/ext4/htree_lock.c
 132 + *
 133 + * Copyright (c) 2011, 2012, Intel Corporation.
 134 + *
 135 + * Author: Liang Zhen <liang@whamcloud.com>
 136 + */
 137 +#include <linux/jbd2.h>
 138 +#include <linux/hash.h>
 139 +#include <linux/module.h>
 140 +#include <linux/htree_lock.h>
 141 +
 142 +enum {
 143 +       HTREE_LOCK_BIT_EX       = (1 << HTREE_LOCK_EX),
 144 +       HTREE_LOCK_BIT_PW       = (1 << HTREE_LOCK_PW),
 145 +       HTREE_LOCK_BIT_PR       = (1 << HTREE_LOCK_PR),
 146 +       HTREE_LOCK_BIT_CW       = (1 << HTREE_LOCK_CW),
 147 +       HTREE_LOCK_BIT_CR       = (1 << HTREE_LOCK_CR),
 148 +};
 149 +
 150 +enum {
 151 +       HTREE_LOCK_COMPAT_EX    = 0,
 152 +       HTREE_LOCK_COMPAT_PW    = HTREE_LOCK_COMPAT_EX | HTREE_LOCK_BIT_CR,
 153 +       HTREE_LOCK_COMPAT_PR    = HTREE_LOCK_COMPAT_PW | HTREE_LOCK_BIT_PR,
 154 +       HTREE_LOCK_COMPAT_CW    = HTREE_LOCK_COMPAT_PW | HTREE_LOCK_BIT_CW,
 155 +       HTREE_LOCK_COMPAT_CR    = HTREE_LOCK_COMPAT_CW | HTREE_LOCK_BIT_PR |
 156 +                                 HTREE_LOCK_BIT_PW,
 157 +};
 158 +
 159 +static int htree_lock_compat[] = {
 160 +       [HTREE_LOCK_EX]         HTREE_LOCK_COMPAT_EX,
 161 +       [HTREE_LOCK_PW]         HTREE_LOCK_COMPAT_PW,
 162 +       [HTREE_LOCK_PR]         HTREE_LOCK_COMPAT_PR,
 163 +       [HTREE_LOCK_CW]         HTREE_LOCK_COMPAT_CW,
 164 +       [HTREE_LOCK_CR]         HTREE_LOCK_COMPAT_CR,
 165 +};
 166 +
 167 +/* max allowed htree-lock depth.
 168 + * We only need depth=3 for ext4 although user can have higher value. */
 169 +#define HTREE_LOCK_DEP_MAX     16
 170 +
 171 +#ifdef HTREE_LOCK_DEBUG
 172 +
 173 +static char *hl_name[] = {
 174 +       [HTREE_LOCK_EX]         "EX",
 175 +       [HTREE_LOCK_PW]         "PW",
 176 +       [HTREE_LOCK_PR]         "PR",
 177 +       [HTREE_LOCK_CW]         "CW",
 178 +       [HTREE_LOCK_CR]         "CR",
 179 +};
 180 +
 181 +/* lock stats */
 182 +struct htree_lock_node_stats {
 183 +       unsigned long long      blocked[HTREE_LOCK_MAX];
 184 +       unsigned long long      granted[HTREE_LOCK_MAX];
 185 +       unsigned long long      retried[HTREE_LOCK_MAX];
 186 +       unsigned long long      events;
 187 +};
 188 +
 189 +struct htree_lock_stats {
 190 +       struct htree_lock_node_stats    nodes[HTREE_LOCK_DEP_MAX];
 191 +       unsigned long long      granted[HTREE_LOCK_MAX];
 192 +       unsigned long long      blocked[HTREE_LOCK_MAX];
 193 +};
 194 +
 195 +static struct htree_lock_stats hl_stats;
 196 +
 197 +void htree_lock_stat_reset(void)
 198 +{
 199 +       memset(&hl_stats, 0, sizeof(hl_stats));
 200 +}
 201 +
 202 +void htree_lock_stat_print(int depth)
 203 +{
 204 +       int     i;
 205 +       int     j;
 206 +
 207 +       printk(KERN_DEBUG "HTREE LOCK STATS:\n");
 208 +       for (i = 0; i < HTREE_LOCK_MAX; i++) {
 209 +               printk(KERN_DEBUG "[%s]: G [%10llu], B [%10llu]\n",
 210 +                      hl_name[i], hl_stats.granted[i], hl_stats.blocked[i]);
 211 +       }
 212 +       for (i = 0; i < depth; i++) {
 213 +               printk(KERN_DEBUG "HTREE CHILD [%d] STATS:\n", i);
 214 +               for (j = 0; j < HTREE_LOCK_MAX; j++) {
 215 +                       printk(KERN_DEBUG
 216 +                               "[%s]: G [%10llu], B [%10llu], R [%10llu]\n",
 217 +                               hl_name[j], hl_stats.nodes[i].granted[j],
 218 +                               hl_stats.nodes[i].blocked[j],
 219 +                               hl_stats.nodes[i].retried[j]);
 220 +               }
 221 +       }
 222 +}
 223 +
 224 +#define lk_grant_inc(m)       do { hl_stats.granted[m]++; } while (0)
 225 +#define lk_block_inc(m)       do { hl_stats.blocked[m]++; } while (0)
 226 +#define ln_grant_inc(d, m)    do { hl_stats.nodes[d].granted[m]++; } while (0)
 227 +#define ln_block_inc(d, m)    do { hl_stats.nodes[d].blocked[m]++; } while (0)
 228 +#define ln_retry_inc(d, m)    do { hl_stats.nodes[d].retried[m]++; } while (0)
 229 +#define ln_event_inc(d)       do { hl_stats.nodes[d].events++; } while (0)
 230 +
 231 +#else /* !DEBUG */
 232 +
 233 +void htree_lock_stat_reset(void) {}
 234 +void htree_lock_stat_print(int depth) {}
 235 +
 236 +#define lk_grant_inc(m)              do {} while (0)
 237 +#define lk_block_inc(m)              do {} while (0)
 238 +#define ln_grant_inc(d, m)    do {} while (0)
 239 +#define ln_block_inc(d, m)    do {} while (0)
 240 +#define ln_retry_inc(d, m)    do {} while (0)
 241 +#define ln_event_inc(d)              do {} while (0)
 242 +
 243 +#endif /* DEBUG */
 244 +
 245 +EXPORT_SYMBOL(htree_lock_stat_reset);
 246 +EXPORT_SYMBOL(htree_lock_stat_print);
 247 +
 248 +#define HTREE_DEP_ROOT           (-1)
 249 +
 250 +#define htree_spin_lock(lhead, dep)                            \
 251 +       bit_spin_lock((dep) + 1, &(lhead)->lh_lock)
 252 +#define htree_spin_unlock(lhead, dep)                          \
 253 +       bit_spin_unlock((dep) + 1, &(lhead)->lh_lock)
 254 +
 255 +#define htree_key_event_ignore(child, ln)                      \
 256 +       (!((child)->lc_events & (1 << (ln)->ln_mode)))
 257 +
 258 +static int
 259 +htree_key_list_empty(struct htree_lock_node *ln)
 260 +{
 261 +       return list_empty(&ln->ln_major_list) && list_empty(&ln->ln_minor_list);
 262 +}
 263 +
 264 +static void
 265 +htree_key_list_del_init(struct htree_lock_node *ln)
 266 +{
 267 +       struct htree_lock_node *tmp = NULL;
 268 +
 269 +       if (!list_empty(&ln->ln_minor_list)) {
 270 +               tmp = list_entry(ln->ln_minor_list.next,
 271 +                                struct htree_lock_node, ln_minor_list);
 272 +               list_del_init(&ln->ln_minor_list);
 273 +       }
 274 +
 275 +       if (list_empty(&ln->ln_major_list))
 276 +               return;
 277 +
 278 +       if (tmp == NULL) { /* not on minor key list */
 279 +               list_del_init(&ln->ln_major_list);
 280 +       } else {
 281 +               BUG_ON(!list_empty(&tmp->ln_major_list));
 282 +               list_replace_init(&ln->ln_major_list, &tmp->ln_major_list);
 283 +       }
 284 +}
 285 +
 286 +static void
 287 +htree_key_list_replace_init(struct htree_lock_node *old,
 288 +                           struct htree_lock_node *new)
 289 +{
 290 +       if (!list_empty(&old->ln_major_list))
 291 +               list_replace_init(&old->ln_major_list, &new->ln_major_list);
 292 +
 293 +       if (!list_empty(&old->ln_minor_list))
 294 +               list_replace_init(&old->ln_minor_list, &new->ln_minor_list);
 295 +}
 296 +
 297 +static void
 298 +htree_key_event_enqueue(struct htree_lock_child *child,
 299 +                       struct htree_lock_node *ln, int dep, void *event)
 300 +{
 301 +       struct htree_lock_node *tmp;
 302 +
 303 +       /* NB: ALWAYS called holding lhead::lh_lock(dep) */
 304 +       BUG_ON(ln->ln_mode == HTREE_LOCK_NL);
 305 +       if (event == NULL || htree_key_event_ignore(child, ln))
 306 +               return;
 307 +
 308 +       /* shouldn't be a very long list */
 309 +       list_for_each_entry(tmp, &ln->ln_alive_list, ln_alive_list) {
 310 +               if (tmp->ln_mode == HTREE_LOCK_NL) {
 311 +                       ln_event_inc(dep);
 312 +                       if (child->lc_callback != NULL)
 313 +                               child->lc_callback(tmp->ln_ev_target, event);
 314 +               }
 315 +       }
 316 +}
 317 +
 318 +static int
 319 +htree_node_lock_enqueue(struct htree_lock *newlk, struct htree_lock *curlk,
 320 +                       unsigned dep, int wait, void *event)
 321 +{
 322 +       struct htree_lock_child *child = &newlk->lk_head->lh_children[dep];
 323 +       struct htree_lock_node *newln = &newlk->lk_nodes[dep];
 324 +       struct htree_lock_node *curln = &curlk->lk_nodes[dep];
 325 +
 326 +       /* NB: ALWAYS called holding lhead::lh_lock(dep) */
 327 +       /* NB: we only expect PR/PW lock mode at here, only these two modes are
 328 +        * allowed for htree_node_lock(asserted in htree_node_lock_internal),
 329 +        * NL is only used for listener, user can't directly require NL mode */
 330 +       if ((curln->ln_mode == HTREE_LOCK_NL) ||
 331 +           (curln->ln_mode != HTREE_LOCK_PW &&
 332 +            newln->ln_mode != HTREE_LOCK_PW)) {
 333 +               /* no conflict, attach it on granted list of @curlk */
 334 +               if (curln->ln_mode != HTREE_LOCK_NL) {
 335 +                       list_add(&newln->ln_granted_list,
 336 +                                &curln->ln_granted_list);
 337 +               } else {
 338 +                       /* replace key owner */
 339 +                       htree_key_list_replace_init(curln, newln);
 340 +               }
 341 +
 342 +               list_add(&newln->ln_alive_list, &curln->ln_alive_list);
 343 +               htree_key_event_enqueue(child, newln, dep, event);
 344 +               ln_grant_inc(dep, newln->ln_mode);
 345 +               return 1; /* still hold lh_lock */
 346 +       }
 347 +
 348 +       if (!wait) { /* can't grant and don't want to wait */
 349 +               ln_retry_inc(dep, newln->ln_mode);
 350 +               newln->ln_mode = HTREE_LOCK_INVAL;
 351 +               return -1; /* don't wait and just return -1 */
 352 +       }
 353 +
 354 +       newlk->lk_task = current;
 355 +       set_current_state(TASK_UNINTERRUPTIBLE);
 356 +       /* conflict, attach it on blocked list of curlk */
 357 +       list_add_tail(&newln->ln_blocked_list, &curln->ln_blocked_list);
 358 +       list_add(&newln->ln_alive_list, &curln->ln_alive_list);
 359 +       ln_block_inc(dep, newln->ln_mode);
 360 +
 361 +       htree_spin_unlock(newlk->lk_head, dep);
 362 +       /* wait to be given the lock */
 363 +       if (newlk->lk_task != NULL)
 364 +               schedule();
 365 +       /* granted, no doubt, wake up will set me RUNNING */
 366 +       if (event == NULL || htree_key_event_ignore(child, newln))
 367 +               return 0; /* granted without lh_lock */
 368 +
 369 +       htree_spin_lock(newlk->lk_head, dep);
 370 +       htree_key_event_enqueue(child, newln, dep, event);
 371 +       return 1; /* still hold lh_lock */
 372 +}
 373 +
 374 +/*
 375 + * get PR/PW access to particular tree-node according to @dep and @key,
 376 + * it will return -1 if @wait is false and can't immediately grant this lock.
 377 + * All listeners(HTREE_LOCK_NL) on @dep and with the same @key will get
 378 + * @event if it's not NULL.
 379 + * NB: ALWAYS called holding lhead::lh_lock
 380 + */
 381 +static int
 382 +htree_node_lock_internal(struct htree_lock_head *lhead, struct htree_lock *lck,
 383 +                        htree_lock_mode_t mode, u32 key, unsigned dep,
 384 +                        int wait, void *event)
 385 +{
 386 +       LIST_HEAD(list);
 387 +       struct htree_lock       *tmp;
 388 +       struct htree_lock       *tmp2;
 389 +       u16                     major;
 390 +       u16                     minor;
 391 +       u8                      reverse;
 392 +       u8                      ma_bits;
 393 +       u8                      mi_bits;
 394 +
 395 +       BUG_ON(mode != HTREE_LOCK_PW && mode != HTREE_LOCK_PR);
 396 +       BUG_ON(htree_node_is_granted(lck, dep));
 397 +
 398 +       key = hash_long(key, lhead->lh_hbits);
 399 +
 400 +       mi_bits = lhead->lh_hbits >> 1;
 401 +       ma_bits = lhead->lh_hbits - mi_bits;
 402 +
 403 +       lck->lk_nodes[dep].ln_major_key = major = key & ((1U << ma_bits) - 1);
 404 +       lck->lk_nodes[dep].ln_minor_key = minor = key >> ma_bits;
 405 +       lck->lk_nodes[dep].ln_mode = mode;
 406 +
 407 +       /*
 408 +        * The major key list is an ordered list, so searches are started
 409 +        * at the end of the list that is numerically closer to major_key,
 410 +        * so at most half of the list will be walked (for well-distributed
 411 +        * keys). The list traversal aborts early if the expected key
 412 +        * location is passed.
 413 +        */
 414 +       reverse = (major >= (1 << (ma_bits - 1)));
 415 +
 416 +       if (reverse) {
 417 +               list_for_each_entry_reverse(tmp,
 418 +                                       &lhead->lh_children[dep].lc_list,
 419 +                                       lk_nodes[dep].ln_major_list) {
 420 +                       if (tmp->lk_nodes[dep].ln_major_key == major) {
 421 +                               goto search_minor;
 422 +
 423 +                       } else if (tmp->lk_nodes[dep].ln_major_key < major) {
 424 +                               /* attach _after_ @tmp */
 425 +                               list_add(&lck->lk_nodes[dep].ln_major_list,
 426 +                                        &tmp->lk_nodes[dep].ln_major_list);
 427 +                               goto out_grant_major;
 428 +                       }
 429 +               }
 430 +
 431 +               list_add(&lck->lk_nodes[dep].ln_major_list,
 432 +                        &lhead->lh_children[dep].lc_list);
 433 +               goto out_grant_major;
 434 +
 435 +       } else {
 436 +               list_for_each_entry(tmp, &lhead->lh_children[dep].lc_list,
 437 +                                   lk_nodes[dep].ln_major_list) {
 438 +                       if (tmp->lk_nodes[dep].ln_major_key == major) {
 439 +                               goto search_minor;
 440 +
 441 +                       } else if (tmp->lk_nodes[dep].ln_major_key > major) {
 442 +                               /* insert _before_ @tmp */
 443 +                               list_add_tail(&lck->lk_nodes[dep].ln_major_list,
 444 +                                       &tmp->lk_nodes[dep].ln_major_list);
 445 +                               goto out_grant_major;
 446 +                       }
 447 +               }
 448 +
 449 +               list_add_tail(&lck->lk_nodes[dep].ln_major_list,
 450 +                             &lhead->lh_children[dep].lc_list);
 451 +               goto out_grant_major;
 452 +       }
 453 +
 454 + search_minor:
 455 +       /*
 456 +        * NB: minor_key list doesn't have a "head", @list is just a
 457 +        * temporary stub for helping list searching, make sure it's removed
 458 +        * after searching.
 459 +        * minor_key list is an ordered list too.
 460 +        */
 461 +       list_add_tail(&list, &tmp->lk_nodes[dep].ln_minor_list);
 462 +
 463 +       reverse = (minor >= (1 << (mi_bits - 1)));
 464 +
 465 +       if (reverse) {
 466 +               list_for_each_entry_reverse(tmp2, &list,
 467 +                                           lk_nodes[dep].ln_minor_list) {
 468 +                       if (tmp2->lk_nodes[dep].ln_minor_key == minor) {
 469 +                               goto out_enqueue;
 470 +
 471 +                       } else if (tmp2->lk_nodes[dep].ln_minor_key < minor) {
 472 +                               /* attach _after_ @tmp2 */
 473 +                               list_add(&lck->lk_nodes[dep].ln_minor_list,
 474 +                                        &tmp2->lk_nodes[dep].ln_minor_list);
 475 +                               goto out_grant_minor;
 476 +                       }
 477 +               }
 478 +
 479 +               list_add(&lck->lk_nodes[dep].ln_minor_list, &list);
 480 +
 481 +       } else {
 482 +               list_for_each_entry(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 +                               /* insert _before_ @tmp2 */
 489 +                               list_add_tail(&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_tail(&lck->lk_nodes[dep].ln_minor_list, &list);
 496 +       }
 497 +
 498 + out_grant_minor:
 499 +       if (list.next == &lck->lk_nodes[dep].ln_minor_list) {
 500 +               /* new lock @lck is the first one on minor_key list, which
 501 +                * means it has the smallest minor_key and it should
 502 +                * replace @tmp as minor_key owner */
 503 +               list_replace_init(&tmp->lk_nodes[dep].ln_major_list,
 504 +                                 &lck->lk_nodes[dep].ln_major_list);
 505 +       }
 506 +       /* remove the temporary head */
 507 +       list_del(&list);
 508 +
 509 + out_grant_major:
 510 +       ln_grant_inc(dep, lck->lk_nodes[dep].ln_mode);
 511 +       return 1; /* granted with holding lh_lock */
 512 +
 513 + out_enqueue:
 514 +       list_del(&list); /* remove temprary head */
 515 +       return htree_node_lock_enqueue(lck, tmp2, dep, wait, event);
 516 +}
 517 +
 518 +/*
 519 + * release the key of @lck at level @dep, and grant any blocked locks.
 520 + * caller will still listen on @key if @event is not NULL, which means
 521 + * caller can see a event (by event_cb) while granting any lock with
 522 + * the same key at level @dep.
 523 + * NB: ALWAYS called holding lhead::lh_lock
 524 + * NB: listener will not block anyone because listening mode is HTREE_LOCK_NL
 525 + */
 526 +static void
 527 +htree_node_unlock_internal(struct htree_lock_head *lhead,
 528 +                          struct htree_lock *curlk, unsigned dep, void *event)
 529 +{
 530 +       struct htree_lock_node  *curln = &curlk->lk_nodes[dep];
 531 +       struct htree_lock       *grtlk = NULL;
 532 +       struct htree_lock_node  *grtln;
 533 +       struct htree_lock       *poslk;
 534 +       struct htree_lock       *tmplk;
 535 +
 536 +       if (!htree_node_is_granted(curlk, dep))
 537 +               return;
 538 +
 539 +       if (!list_empty(&curln->ln_granted_list)) {
 540 +               /* there is another granted lock */
 541 +               grtlk = list_entry(curln->ln_granted_list.next,
 542 +                                  struct htree_lock,
 543 +                                  lk_nodes[dep].ln_granted_list);
 544 +               list_del_init(&curln->ln_granted_list);
 545 +       }
 546 +
 547 +       if (grtlk == NULL && !list_empty(&curln->ln_blocked_list)) {
 548 +               /*
 549 +                * @curlk is the only granted lock, so we confirmed:
 550 +                * a) curln is key owner (attached on major/minor_list),
 551 +                *    so if there is any blocked lock, it should be attached
 552 +                *    on curln->ln_blocked_list
 553 +                * b) we always can grant the first blocked lock
 554 +                */
 555 +               grtlk = list_entry(curln->ln_blocked_list.next,
 556 +                                  struct htree_lock,
 557 +                                  lk_nodes[dep].ln_blocked_list);
 558 +               BUG_ON(grtlk->lk_task == NULL);
 559 +               wake_up_process(grtlk->lk_task);
 560 +       }
 561 +
 562 +       if (event != NULL &&
 563 +           lhead->lh_children[dep].lc_events != HTREE_EVENT_DISABLE) {
 564 +               curln->ln_ev_target = event;
 565 +               curln->ln_mode = HTREE_LOCK_NL; /* listen! */
 566 +       } else {
 567 +               curln->ln_mode = HTREE_LOCK_INVAL;
 568 +       }
 569 +
 570 +       if (grtlk == NULL) { /* I must be the only one locking this key */
 571 +               struct htree_lock_node *tmpln;
 572 +
 573 +               BUG_ON(htree_key_list_empty(curln));
 574 +
 575 +               if (curln->ln_mode == HTREE_LOCK_NL) /* listening */
 576 +                       return;
 577 +
 578 +               /* not listening */
 579 +               if (list_empty(&curln->ln_alive_list)) { /* no more listener */
 580 +                       htree_key_list_del_init(curln);
 581 +                       return;
 582 +               }
 583 +
 584 +               tmpln = list_entry(curln->ln_alive_list.next,
 585 +                                  struct htree_lock_node, ln_alive_list);
 586 +
 587 +               BUG_ON(tmpln->ln_mode != HTREE_LOCK_NL);
 588 +
 589 +               htree_key_list_replace_init(curln, tmpln);
 590 +               list_del_init(&curln->ln_alive_list);
 591 +
 592 +               return;
 593 +       }
 594 +
 595 +       /* have a granted lock */
 596 +       grtln = &grtlk->lk_nodes[dep];
 597 +       if (!list_empty(&curln->ln_blocked_list)) {
 598 +               /* only key owner can be on both lists */
 599 +               BUG_ON(htree_key_list_empty(curln));
 600 +
 601 +               if (list_empty(&grtln->ln_blocked_list)) {
 602 +                       list_add(&grtln->ln_blocked_list,
 603 +                                &curln->ln_blocked_list);
 604 +               }
 605 +               list_del_init(&curln->ln_blocked_list);
 606 +       }
 607 +       /*
 608 +        * NB: this is the tricky part:
 609 +        * We have only two modes for child-lock (PR and PW), also,
 610 +        * only owner of the key (attached on major/minor_list) can be on
 611 +        * both blocked_list and granted_list, so @grtlk must be one
 612 +        * of these two cases:
 613 +        *
 614 +        * a) @grtlk is taken from granted_list, which means we've granted
 615 +        *    more than one lock so @grtlk has to be PR, the first blocked
 616 +        *    lock must be PW and we can't grant it at all.
 617 +        *    So even @grtlk is not owner of the key (empty blocked_list),
 618 +        *    we don't care because we can't grant any lock.
 619 +        * b) we just grant a new lock which is taken from head of blocked
 620 +        *    list, and it should be the first granted lock, and it should
 621 +        *    be the first one linked on blocked_list.
 622 +        *
 623 +        * Either way, we can get correct result by iterating blocked_list
 624 +        * of @grtlk, and don't have to bother on how to find out
 625 +        * owner of current key.
 626 +        */
 627 +       list_for_each_entry_safe(poslk, tmplk, &grtln->ln_blocked_list,
 628 +                                lk_nodes[dep].ln_blocked_list) {
 629 +               if (grtlk->lk_nodes[dep].ln_mode == HTREE_LOCK_PW ||
 630 +                   poslk->lk_nodes[dep].ln_mode == HTREE_LOCK_PW)
 631 +                       break;
 632 +               /* grant all readers */
 633 +               list_del_init(&poslk->lk_nodes[dep].ln_blocked_list);
 634 +               list_add(&poslk->lk_nodes[dep].ln_granted_list,
 635 +                        &grtln->ln_granted_list);
 636 +
 637 +               BUG_ON(poslk->lk_task == NULL);
 638 +               wake_up_process(poslk->lk_task);
 639 +       }
 640 +
 641 +       /* if @curln is the owner of this key, replace it with @grtln */
 642 +       if (!htree_key_list_empty(curln))
 643 +               htree_key_list_replace_init(curln, grtln);
 644 +
 645 +       if (curln->ln_mode == HTREE_LOCK_INVAL)
 646 +               list_del_init(&curln->ln_alive_list);
 647 +}
 648 +
 649 +/*
 650 + * it's just wrapper of htree_node_lock_internal, it returns 1 on granted
 651 + * and 0 only if @wait is false and can't grant it immediately
 652 + */
 653 +int
 654 +htree_node_lock_try(struct htree_lock *lck, htree_lock_mode_t mode,
 655 +                   u32 key, unsigned dep, int wait, void *event)
 656 +{
 657 +       struct htree_lock_head *lhead = lck->lk_head;
 658 +       int rc;
 659 +
 660 +       BUG_ON(dep >= lck->lk_depth);
 661 +       BUG_ON(lck->lk_mode == HTREE_LOCK_INVAL);
 662 +
 663 +       htree_spin_lock(lhead, dep);
 664 +       rc = htree_node_lock_internal(lhead, lck, mode, key, dep, wait, event);
 665 +       if (rc != 0)
 666 +               htree_spin_unlock(lhead, dep);
 667 +       return rc >= 0;
 668 +}
 669 +EXPORT_SYMBOL(htree_node_lock_try);
 670 +
 671 +/* it's wrapper of htree_node_unlock_internal */
 672 +void
 673 +htree_node_unlock(struct htree_lock *lck, unsigned dep, void *event)
 674 +{
 675 +       struct htree_lock_head *lhead = lck->lk_head;
 676 +
 677 +       BUG_ON(dep >= lck->lk_depth);
 678 +       BUG_ON(lck->lk_mode == HTREE_LOCK_INVAL);
 679 +
 680 +       htree_spin_lock(lhead, dep);
 681 +       htree_node_unlock_internal(lhead, lck, dep, event);
 682 +       htree_spin_unlock(lhead, dep);
 683 +}
 684 +EXPORT_SYMBOL(htree_node_unlock);
 685 +
 686 +/* stop listening on child-lock level @dep */
 687 +void
 688 +htree_node_stop_listen(struct htree_lock *lck, unsigned dep)
 689 +{
 690 +       struct htree_lock_node *ln = &lck->lk_nodes[dep];
 691 +       struct htree_lock_node *tmp;
 692 +
 693 +       BUG_ON(htree_node_is_granted(lck, dep));
 694 +       BUG_ON(!list_empty(&ln->ln_blocked_list));
 695 +       BUG_ON(!list_empty(&ln->ln_granted_list));
 696 +
 697 +       if (!htree_node_is_listening(lck, dep))
 698 +               return;
 699 +
 700 +       htree_spin_lock(lck->lk_head, dep);
 701 +       ln->ln_mode = HTREE_LOCK_INVAL;
 702 +       ln->ln_ev_target = NULL;
 703 +
 704 +       if (htree_key_list_empty(ln)) { /* not owner */
 705 +               list_del_init(&ln->ln_alive_list);
 706 +               goto out;
 707 +       }
 708 +
 709 +       /* I'm the owner... */
 710 +       if (list_empty(&ln->ln_alive_list)) { /* no more listener */
 711 +               htree_key_list_del_init(ln);
 712 +               goto out;
 713 +       }
 714 +
 715 +       tmp = list_entry(ln->ln_alive_list.next,
 716 +                        struct htree_lock_node, ln_alive_list);
 717 +
 718 +       BUG_ON(tmp->ln_mode != HTREE_LOCK_NL);
 719 +       htree_key_list_replace_init(ln, tmp);
 720 +       list_del_init(&ln->ln_alive_list);
 721 + out:
 722 +       htree_spin_unlock(lck->lk_head, dep);
 723 +}
 724 +EXPORT_SYMBOL(htree_node_stop_listen);
 725 +
 726 +/* release all child-locks if we have any */
 727 +static void
 728 +htree_node_release_all(struct htree_lock *lck)
 729 +{
 730 +       int     i;
 731 +
 732 +       for (i = 0; i < lck->lk_depth; i++) {
 733 +               if (htree_node_is_granted(lck, i))
 734 +                       htree_node_unlock(lck, i, NULL);
 735 +               else if (htree_node_is_listening(lck, i))
 736 +                       htree_node_stop_listen(lck, i);
 737 +       }
 738 +}
 739 +
 740 +/*
 741 + * obtain htree lock, it could be blocked inside if there's conflict
 742 + * with any granted or blocked lock and @wait is true.
 743 + * NB: ALWAYS called holding lhead::lh_lock
 744 + */
 745 +static int
 746 +htree_lock_internal(struct htree_lock *lck, int wait)
 747 +{
 748 +       struct htree_lock_head *lhead = lck->lk_head;
 749 +       int     granted = 0;
 750 +       int     blocked = 0;
 751 +       int     i;
 752 +
 753 +       for (i = 0; i < HTREE_LOCK_MAX; i++) {
 754 +               if (lhead->lh_ngranted[i] != 0)
 755 +                       granted |= 1 << i;
 756 +               if (lhead->lh_nblocked[i] != 0)
 757 +                       blocked |= 1 << i;
 758 +       }
 759 +       if ((htree_lock_compat[lck->lk_mode] & granted) != granted ||
 760 +           (htree_lock_compat[lck->lk_mode] & blocked) != blocked) {
 761 +               /* will block current lock even it just conflicts with any
 762 +                * other blocked lock, so lock like EX wouldn't starve */
 763 +               if (!wait)
 764 +                       return -1;
 765 +               lhead->lh_nblocked[lck->lk_mode]++;
 766 +               lk_block_inc(lck->lk_mode);
 767 +
 768 +               lck->lk_task = current;
 769 +               list_add_tail(&lck->lk_blocked_list, &lhead->lh_blocked_list);
 770 +
 771 +               set_current_state(TASK_UNINTERRUPTIBLE);
 772 +               htree_spin_unlock(lhead, HTREE_DEP_ROOT);
 773 +               /* wait to be given the lock */
 774 +               if (lck->lk_task != NULL)
 775 +                       schedule();
 776 +               /* granted, no doubt. wake up will set me RUNNING */
 777 +               return 0; /* without lh_lock */
 778 +       }
 779 +       lhead->lh_ngranted[lck->lk_mode]++;
 780 +       lk_grant_inc(lck->lk_mode);
 781 +       return 1;
 782 +}
 783 +
 784 +/* release htree lock. NB: ALWAYS called holding lhead::lh_lock */
 785 +static void
 786 +htree_unlock_internal(struct htree_lock *lck)
 787 +{
 788 +       struct htree_lock_head *lhead = lck->lk_head;
 789 +       struct htree_lock *tmp;
 790 +       struct htree_lock *tmp2;
 791 +       int granted = 0;
 792 +       int i;
 793 +
 794 +       BUG_ON(lhead->lh_ngranted[lck->lk_mode] == 0);
 795 +
 796 +       lhead->lh_ngranted[lck->lk_mode]--;
 797 +       lck->lk_mode = HTREE_LOCK_INVAL;
 798 +
 799 +       for (i = 0; i < HTREE_LOCK_MAX; i++) {
 800 +               if (lhead->lh_ngranted[i] != 0)
 801 +                       granted |= 1 << i;
 802 +       }
 803 +       list_for_each_entry_safe(tmp, tmp2,
 804 +                                &lhead->lh_blocked_list, lk_blocked_list) {
 805 +               /* conflict with any granted lock? */
 806 +               if ((htree_lock_compat[tmp->lk_mode] & granted) != granted)
 807 +                       break;
 808 +
 809 +               list_del_init(&tmp->lk_blocked_list);
 810 +
 811 +               BUG_ON(lhead->lh_nblocked[tmp->lk_mode] == 0);
 812 +
 813 +               lhead->lh_nblocked[tmp->lk_mode]--;
 814 +               lhead->lh_ngranted[tmp->lk_mode]++;
 815 +               granted |= 1 << tmp->lk_mode;
 816 +
 817 +               BUG_ON(tmp->lk_task == NULL);
 818 +               wake_up_process(tmp->lk_task);
 819 +       }
 820 +}
 821 +
 822 +/* it's wrapper of htree_lock_internal and exported interface.
 823 + * It always return 1 with granted lock if @wait is true, it can return 0
 824 + * if @wait is false and locking request can't be granted immediately */
 825 +int
 826 +htree_lock_try(struct htree_lock *lck, struct htree_lock_head *lhead,
 827 +              htree_lock_mode_t mode, int wait)
 828 +{
 829 +       int     rc;
 830 +
 831 +       BUG_ON(lck->lk_depth > lhead->lh_depth);
 832 +       BUG_ON(lck->lk_head != NULL);
 833 +       BUG_ON(lck->lk_task != NULL);
 834 +
 835 +       lck->lk_head = lhead;
 836 +       lck->lk_mode = mode;
 837 +
 838 +       htree_spin_lock(lhead, HTREE_DEP_ROOT);
 839 +       rc = htree_lock_internal(lck, wait);
 840 +       if (rc != 0)
 841 +               htree_spin_unlock(lhead, HTREE_DEP_ROOT);
 842 +       return rc >= 0;
 843 +}
 844 +EXPORT_SYMBOL(htree_lock_try);
 845 +
 846 +/* it's wrapper of htree_unlock_internal and exported interface.
 847 + * It will release all htree_node_locks and htree_lock */
 848 +void
 849 +htree_unlock(struct htree_lock *lck)
 850 +{
 851 +       BUG_ON(lck->lk_head == NULL);
 852 +       BUG_ON(lck->lk_mode == HTREE_LOCK_INVAL);
 853 +
 854 +       htree_node_release_all(lck);
 855 +
 856 +       htree_spin_lock(lck->lk_head, HTREE_DEP_ROOT);
 857 +       htree_unlock_internal(lck);
 858 +       htree_spin_unlock(lck->lk_head, HTREE_DEP_ROOT);
 859 +       lck->lk_head = NULL;
 860 +       lck->lk_task = NULL;
 861 +}
 862 +EXPORT_SYMBOL(htree_unlock);
 863 +
 864 +/* change lock mode */
 865 +void
 866 +htree_change_mode(struct htree_lock *lck, htree_lock_mode_t mode)
 867 +{
 868 +       BUG_ON(lck->lk_mode == HTREE_LOCK_INVAL);
 869 +       lck->lk_mode = mode;
 870 +}
 871 +EXPORT_SYMBOL(htree_change_mode);
 872 +
 873 +/* release htree lock, and lock it again with new mode.
 874 + * This function will first release all htree_node_locks and htree_lock,
 875 + * then try to gain htree_lock with new @mode.
 876 + * It always return 1 with granted lock if @wait is true, it can return 0
 877 + * if @wait is false and locking request can't be granted immediately */
 878 +int
 879 +htree_change_lock_try(struct htree_lock *lck, htree_lock_mode_t mode, int wait)
 880 +{
 881 +       struct htree_lock_head *lhead = lck->lk_head;
 882 +       int rc;
 883 +
 884 +       BUG_ON(lhead == NULL);
 885 +       BUG_ON(lck->lk_mode == mode);
 886 +       BUG_ON(lck->lk_mode == HTREE_LOCK_INVAL || mode == HTREE_LOCK_INVAL);
 887 +
 888 +       htree_node_release_all(lck);
 889 +
 890 +       htree_spin_lock(lhead, HTREE_DEP_ROOT);
 891 +       htree_unlock_internal(lck);
 892 +       lck->lk_mode = mode;
 893 +       rc = htree_lock_internal(lck, wait);
 894 +       if (rc != 0)
 895 +               htree_spin_unlock(lhead, HTREE_DEP_ROOT);
 896 +       return rc >= 0;
 897 +}
 898 +EXPORT_SYMBOL(htree_change_lock_try);
 899 +
 900 +/* create a htree_lock head with @depth levels (number of child-locks),
 901 + * it is a per resoruce structure */
 902 +struct htree_lock_head *
 903 +htree_lock_head_alloc(unsigned depth, unsigned hbits, unsigned priv)
 904 +{
 905 +       struct htree_lock_head *lhead;
 906 +       int  i;
 907 +
 908 +       if (depth > HTREE_LOCK_DEP_MAX) {
 909 +               printk(KERN_ERR "%d is larger than max htree_lock depth %d\n",
 910 +                       depth, HTREE_LOCK_DEP_MAX);
 911 +               return NULL;
 912 +       }
 913 +
 914 +       lhead = kzalloc(offsetof(struct htree_lock_head,
 915 +                                lh_children[depth]) + priv, GFP_NOFS);
 916 +       if (lhead == NULL)
 917 +               return NULL;
 918 +
 919 +       if (hbits < HTREE_HBITS_MIN)
 920 +               lhead->lh_hbits = HTREE_HBITS_MIN;
 921 +       else if (hbits > HTREE_HBITS_MAX)
 922 +               lhead->lh_hbits = HTREE_HBITS_MAX;
 923 +
 924 +       lhead->lh_lock = 0;
 925 +       lhead->lh_depth = depth;
 926 +       INIT_LIST_HEAD(&lhead->lh_blocked_list);
 927 +       if (priv > 0) {
 928 +               lhead->lh_private = (void *)lhead +
 929 +                       offsetof(struct htree_lock_head, lh_children[depth]);
 930 +       }
 931 +
 932 +       for (i = 0; i < depth; i++) {
 933 +               INIT_LIST_HEAD(&lhead->lh_children[i].lc_list);
 934 +               lhead->lh_children[i].lc_events = HTREE_EVENT_DISABLE;
 935 +       }
 936 +       return lhead;
 937 +}
 938 +EXPORT_SYMBOL(htree_lock_head_alloc);
 939 +
 940 +/* free the htree_lock head */
 941 +void
 942 +htree_lock_head_free(struct htree_lock_head *lhead)
 943 +{
 944 +       int     i;
 945 +
 946 +       BUG_ON(!list_empty(&lhead->lh_blocked_list));
 947 +       for (i = 0; i < lhead->lh_depth; i++)
 948 +               BUG_ON(!list_empty(&lhead->lh_children[i].lc_list));
 949 +       kfree(lhead);
 950 +}
 951 +EXPORT_SYMBOL(htree_lock_head_free);
 952 +
 953 +/* register event callback for @events of child-lock at level @dep */
 954 +void
 955 +htree_lock_event_attach(struct htree_lock_head *lhead, unsigned dep,
 956 +                       unsigned events, htree_event_cb_t callback)
 957 +{
 958 +       BUG_ON(lhead->lh_depth <= dep);
 959 +       lhead->lh_children[dep].lc_events = events;
 960 +       lhead->lh_children[dep].lc_callback = callback;
 961 +}
 962 +EXPORT_SYMBOL(htree_lock_event_attach);
 963 +
 964 +/* allocate a htree_lock, which is per-thread structure, @pbytes is some
 965 + * extra-bytes as private data for caller */
 966 +struct htree_lock *
 967 +htree_lock_alloc(unsigned depth, unsigned pbytes)
 968 +{
 969 +       struct htree_lock *lck;
 970 +       int i = offsetof(struct htree_lock, lk_nodes[depth]);
 971 +
 972 +       if (depth > HTREE_LOCK_DEP_MAX) {
 973 +               printk(KERN_ERR "%d is larger than max htree_lock depth %d\n",
 974 +                       depth, HTREE_LOCK_DEP_MAX);
 975 +               return NULL;
 976 +       }
 977 +       lck = kzalloc(i + pbytes, GFP_NOFS);
 978 +       if (lck == NULL)
 979 +               return NULL;
 980 +
 981 +       if (pbytes != 0)
 982 +               lck->lk_private = (void *)lck + i;
 983 +       lck->lk_mode = HTREE_LOCK_INVAL;
 984 +       lck->lk_depth = depth;
 985 +       INIT_LIST_HEAD(&lck->lk_blocked_list);
 986 +
 987 +       for (i = 0; i < depth; i++) {
 988 +               struct htree_lock_node *node = &lck->lk_nodes[i];
 989 +
 990 +               node->ln_mode = HTREE_LOCK_INVAL;
 991 +               INIT_LIST_HEAD(&node->ln_major_list);
 992 +               INIT_LIST_HEAD(&node->ln_minor_list);
 993 +               INIT_LIST_HEAD(&node->ln_alive_list);
 994 +               INIT_LIST_HEAD(&node->ln_blocked_list);
 995 +               INIT_LIST_HEAD(&node->ln_granted_list);
 996 +       }
 997 +
 998 +       return lck;
 999 +}
1000 +EXPORT_SYMBOL(htree_lock_alloc);
1001 +
1002 +/* free htree_lock node */
1003 +void
1004 +htree_lock_free(struct htree_lock *lck)
1005 +{
1006 +       BUG_ON(lck->lk_mode != HTREE_LOCK_INVAL);
1007 +       kfree(lck);
1008 +}
1009 +EXPORT_SYMBOL(htree_lock_free);
1010 diff --git a/fs/ext4/namei.c b/fs/ext4/namei.c
1011 index 3f70bca..99a8da2 100644
1012 --- a/fs/ext4/namei.c
1013 +++ b/fs/ext4/namei.c
1014 @@ -52,6 +52,7 @@ struct buffer_head *ext4_append(handle_t *handle,
1015                                         ext4_lblk_t *block)
1016  {
1017         struct buffer_head *bh;
1018 +       struct ext4_inode_info *ei = EXT4_I(inode);
1019         int err;
1020
1021         if (unlikely(EXT4_SB(inode->i_sb)->s_max_dir_size_kb &&
1022 @@ -59,15 +60,22 @@ struct buffer_head *ext4_append(handle_t *handle,
1023                       EXT4_SB(inode->i_sb)->s_max_dir_size_kb)))
1024                 return ERR_PTR(-ENOSPC);
1025
1026 +       /* with parallel dir operations all appends
1027 +       * have to be serialized -bzzz */
1028 +       down(&ei->i_append_sem);
1029 +
1030         *block = inode->i_size >> inode->i_sb->s_blocksize_bits;
1031
1032         bh = ext4_bread(handle, inode, *block, EXT4_GET_BLOCKS_CREATE);
1033 -       if (IS_ERR(bh))
1034 +       if (IS_ERR(bh)) {
1035 +               up(&ei->i_append_sem);
1036                 return bh;
1037 +       }
1038         inode->i_size += inode->i_sb->s_blocksize;
1039         EXT4_I(inode)->i_disksize = inode->i_size;
1040         BUFFER_TRACE(bh, "get_write_access");
1041         err = ext4_journal_get_write_access(handle, bh);
1042 +       up(&ei->i_append_sem);
1043         if (err) {
1044                 brelse(bh);
1045                 ext4_std_error(inode->i_sb, err);
1046 @@ -247,7 +255,8 @@ static unsigned dx_node_limit(struct inode *dir);
1047  static struct dx_frame *dx_probe(struct ext4_filename *fname,
1048                                  struct inode *dir,
1049                                  struct dx_hash_info *hinfo,
1050 -                                struct dx_frame *frame);
1051 +                                struct dx_frame *frame,
1052 +                                struct htree_lock *lck);
1053  static void dx_release(struct dx_frame *frames);
1054  static int dx_make_map(struct inode *dir, struct ext4_dir_entry_2 *de,
1055                        unsigned blocksize, struct dx_hash_info *hinfo,
1056 @@ -261,12 +270,13 @@ static void dx_insert_block(struct dx_frame *frame,
1057  static int ext4_htree_next_block(struct inode *dir, __u32 hash,
1058                                  struct dx_frame *frame,
1059                                  struct dx_frame *frames,
1060 -                                __u32 *start_hash);
1061 +                                __u32 *start_hash, struct htree_lock *lck);
1062  static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
1063                 struct ext4_filename *fname,
1064 -               struct ext4_dir_entry_2 **res_dir);
1065 +               struct ext4_dir_entry_2 **res_dir, struct htree_lock *lck);
1066  static int ext4_dx_add_entry(handle_t *handle, struct ext4_filename *fname,
1067 -                            struct dentry *dentry, struct inode *inode);
1068 +                            struct dentry *dentry, struct inode *inode,
1069 +                            struct htree_lock *lck);
1070
1071  /* checksumming functions */
1072  void initialize_dirent_tail(struct ext4_dir_entry_tail *t,
1073 @@ -733,6 +743,227 @@ struct stats dx_show_entries(struct dx_hash_info *hinfo, struct inode *dir,
1074  }
1075  #endif /* DX_DEBUG */
1076
1077 +/* private data for htree_lock */
1078 +struct ext4_dir_lock_data {
1079 +       unsigned                ld_flags;  /* bits-map for lock types */
1080 +       unsigned                ld_count;  /* # entries of the last DX block */
1081 +       struct dx_entry         ld_at_entry; /* copy of leaf dx_entry */
1082 +       struct dx_entry         *ld_at;    /* position of leaf dx_entry */
1083 +};
1084 +
1085 +#define ext4_htree_lock_data(l)        ((struct ext4_dir_lock_data *)(l)->lk_private)
1086 +#define ext4_find_entry(dir, name, dirent, inline) \
1087 +                       __ext4_find_entry(dir, name, dirent, inline, NULL)
1088 +#define ext4_add_entry(handle, dentry, inode) \
1089 +                       __ext4_add_entry(handle, dentry, inode, NULL)
1090 +
1091 +/* NB: ext4_lblk_t is 32 bits so we use high bits to identify invalid blk */
1092 +#define EXT4_HTREE_NODE_CHANGED        (0xcafeULL << 32)
1093 +
1094 +static void ext4_htree_event_cb(void *target, void *event)
1095 +{
1096 +       u64 *block = (u64 *)target;
1097 +
1098 +       if (*block == dx_get_block((struct dx_entry *)event))
1099 +               *block = EXT4_HTREE_NODE_CHANGED;
1100 +}
1101 +
1102 +struct htree_lock_head *ext4_htree_lock_head_alloc(unsigned hbits)
1103 +{
1104 +       struct htree_lock_head *lhead;
1105 +
1106 +       lhead = htree_lock_head_alloc(EXT4_LK_MAX, hbits, 0);
1107 +       if (lhead != NULL) {
1108 +               htree_lock_event_attach(lhead, EXT4_LK_SPIN, HTREE_EVENT_WR,
1109 +                                       ext4_htree_event_cb);
1110 +       }
1111 +       return lhead;
1112 +}
1113 +EXPORT_SYMBOL(ext4_htree_lock_head_alloc);
1114 +
1115 +struct htree_lock *ext4_htree_lock_alloc(void)
1116 +{
1117 +       return htree_lock_alloc(EXT4_LK_MAX,
1118 +                               sizeof(struct ext4_dir_lock_data));
1119 +}
1120 +EXPORT_SYMBOL(ext4_htree_lock_alloc);
1121 +
1122 +static htree_lock_mode_t ext4_htree_mode(unsigned flags)
1123 +{
1124 +       switch (flags) {
1125 +       default: /* 0 or unknown flags require EX lock */
1126 +               return HTREE_LOCK_EX;
1127 +       case EXT4_HLOCK_READDIR:
1128 +               return HTREE_LOCK_PR;
1129 +       case EXT4_HLOCK_LOOKUP:
1130 +               return HTREE_LOCK_CR;
1131 +       case EXT4_HLOCK_DEL:
1132 +       case EXT4_HLOCK_ADD:
1133 +               return HTREE_LOCK_CW;
1134 +       }
1135 +}
1136 +
1137 +/* return PR for read-only operations, otherwise return EX */
1138 +static inline htree_lock_mode_t ext4_htree_safe_mode(unsigned flags)
1139 +{
1140 +       int writer = (flags & EXT4_LB_DE) == EXT4_LB_DE;
1141 +
1142 +       /* 0 requires EX lock */
1143 +       return (flags == 0 || writer) ? HTREE_LOCK_EX : HTREE_LOCK_PR;
1144 +}
1145 +
1146 +static int ext4_htree_safe_locked(struct htree_lock *lck)
1147 +{
1148 +       int writer;
1149 +
1150 +       if (lck == NULL || lck->lk_mode == HTREE_LOCK_EX)
1151 +               return 1;
1152 +
1153 +       writer = (ext4_htree_lock_data(lck)->ld_flags & EXT4_LB_DE) ==
1154 +                EXT4_LB_DE;
1155 +       if (writer) /* all readers & writers are excluded? */
1156 +               return lck->lk_mode == HTREE_LOCK_EX;
1157 +
1158 +       /* all writers are excluded? */
1159 +       return lck->lk_mode == HTREE_LOCK_PR ||
1160 +              lck->lk_mode == HTREE_LOCK_PW ||
1161 +              lck->lk_mode == HTREE_LOCK_EX;
1162 +}
1163 +
1164 +/* relock htree_lock with EX mode if it's change operation, otherwise
1165 + * relock it with PR mode. It's noop if PDO is disabled. */
1166 +static void ext4_htree_safe_relock(struct htree_lock *lck)
1167 +{
1168 +       if (!ext4_htree_safe_locked(lck)) {
1169 +               unsigned flags = ext4_htree_lock_data(lck)->ld_flags;
1170 +
1171 +               htree_change_lock(lck, ext4_htree_safe_mode(flags));
1172 +       }
1173 +}
1174 +
1175 +void ext4_htree_lock(struct htree_lock *lck, struct htree_lock_head *lhead,
1176 +                    struct inode *dir, unsigned flags)
1177 +{
1178 +       htree_lock_mode_t mode = is_dx(dir) ? ext4_htree_mode(flags) :
1179 +                                             ext4_htree_safe_mode(flags);
1180 +
1181 +       ext4_htree_lock_data(lck)->ld_flags = flags;
1182 +       htree_lock(lck, lhead, mode);
1183 +       if (!is_dx(dir))
1184 +               ext4_htree_safe_relock(lck); /* make sure it's safe locked */
1185 +}
1186 +EXPORT_SYMBOL(ext4_htree_lock);
1187 +
1188 +static int ext4_htree_node_lock(struct htree_lock *lck, struct dx_entry *at,
1189 +                               unsigned lmask, int wait, void *ev)
1190 +{
1191 +       u32     key = (at == NULL) ? 0 : dx_get_block(at);
1192 +       u32     mode;
1193 +
1194 +       /* NOOP if htree is well protected or caller doesn't require the lock */
1195 +       if (ext4_htree_safe_locked(lck) ||
1196 +          !(ext4_htree_lock_data(lck)->ld_flags & lmask))
1197 +               return 1;
1198 +
1199 +       mode = (ext4_htree_lock_data(lck)->ld_flags & lmask) == lmask ?
1200 +               HTREE_LOCK_PW : HTREE_LOCK_PR;
1201 +       while (1) {
1202 +               if (htree_node_lock_try(lck, mode, key, ffz(~lmask), wait, ev))
1203 +                       return 1;
1204 +               if (!(lmask & EXT4_LB_SPIN)) /* not a spinlock */
1205 +                       return 0;
1206 +               cpu_relax(); /* spin until granted */
1207 +       }
1208 +}
1209 +
1210 +static int ext4_htree_node_locked(struct htree_lock *lck, unsigned lmask)
1211 +{
1212 +       return ext4_htree_safe_locked(lck) ||
1213 +              htree_node_is_granted(lck, ffz(~lmask));
1214 +}
1215 +
1216 +static void ext4_htree_node_unlock(struct htree_lock *lck,
1217 +                                  unsigned lmask, void *buf)
1218 +{
1219 +       /* NB: it's safe to call mutiple times or even it's not locked */
1220 +       if (!ext4_htree_safe_locked(lck) &&
1221 +            htree_node_is_granted(lck, ffz(~lmask)))
1222 +               htree_node_unlock(lck, ffz(~lmask), buf);
1223 +}
1224 +
1225 +#define ext4_htree_dx_lock(lck, key)           \
1226 +       ext4_htree_node_lock(lck, key, EXT4_LB_DX, 1, NULL)
1227 +#define ext4_htree_dx_lock_try(lck, key)       \
1228 +       ext4_htree_node_lock(lck, key, EXT4_LB_DX, 0, NULL)
1229 +#define ext4_htree_dx_unlock(lck)              \
1230 +       ext4_htree_node_unlock(lck, EXT4_LB_DX, NULL)
1231 +#define ext4_htree_dx_locked(lck)              \
1232 +       ext4_htree_node_locked(lck, EXT4_LB_DX)
1233 +
1234 +static void ext4_htree_dx_need_lock(struct htree_lock *lck)
1235 +{
1236 +       struct ext4_dir_lock_data *ld;
1237 +
1238 +       if (ext4_htree_safe_locked(lck))
1239 +               return;
1240 +
1241 +       ld = ext4_htree_lock_data(lck);
1242 +       switch (ld->ld_flags) {
1243 +       default:
1244 +               return;
1245 +       case EXT4_HLOCK_LOOKUP:
1246 +               ld->ld_flags = EXT4_HLOCK_LOOKUP_SAFE;
1247 +               return;
1248 +       case EXT4_HLOCK_DEL:
1249 +               ld->ld_flags = EXT4_HLOCK_DEL_SAFE;
1250 +               return;
1251 +       case EXT4_HLOCK_ADD:
1252 +               ld->ld_flags = EXT4_HLOCK_SPLIT;
1253 +               return;
1254 +       }
1255 +}
1256 +
1257 +#define ext4_htree_de_lock(lck, key)           \
1258 +       ext4_htree_node_lock(lck, key, EXT4_LB_DE, 1, NULL)
1259 +#define ext4_htree_de_unlock(lck)              \
1260 +       ext4_htree_node_unlock(lck, EXT4_LB_DE, NULL)
1261 +
1262 +#define ext4_htree_spin_lock(lck, key, event)  \
1263 +       ext4_htree_node_lock(lck, key, EXT4_LB_SPIN, 0, event)
1264 +#define ext4_htree_spin_unlock(lck)            \
1265 +       ext4_htree_node_unlock(lck, EXT4_LB_SPIN, NULL)
1266 +#define ext4_htree_spin_unlock_listen(lck, p)  \
1267 +       ext4_htree_node_unlock(lck, EXT4_LB_SPIN, p)
1268 +
1269 +static void ext4_htree_spin_stop_listen(struct htree_lock *lck)
1270 +{
1271 +       if (!ext4_htree_safe_locked(lck) &&
1272 +           htree_node_is_listening(lck, ffz(~EXT4_LB_SPIN)))
1273 +               htree_node_stop_listen(lck, ffz(~EXT4_LB_SPIN));
1274 +}
1275 +
1276 +enum {
1277 +       DX_HASH_COL_IGNORE,     /* ignore collision while probing frames */
1278 +       DX_HASH_COL_YES,        /* there is collision and it does matter */
1279 +       DX_HASH_COL_NO,         /* there is no collision */
1280 +};
1281 +
1282 +static int dx_probe_hash_collision(struct htree_lock *lck,
1283 +                                  struct dx_entry *entries,
1284 +                                  struct dx_entry *at, u32 hash)
1285 +{
1286 +       if (!(lck && ext4_htree_lock_data(lck)->ld_flags & EXT4_LB_EXACT)) {
1287 +               return DX_HASH_COL_IGNORE; /* don't care about collision */
1288 +
1289 +       } else if (at == entries + dx_get_count(entries) - 1) {
1290 +               return DX_HASH_COL_IGNORE; /* not in any leaf of this DX */
1291 +
1292 +       } else { /* hash collision? */
1293 +               return ((dx_get_hash(at + 1) & ~1) == hash) ?
1294 +                       DX_HASH_COL_YES : DX_HASH_COL_NO;
1295 +       }
1296 +}
1297 +
1298  /*
1299   * Probe for a directory leaf block to search.
1300   *
1301 @@ -744,10 +975,11 @@ struct stats dx_show_entries(struct dx_hash_info *hinfo, struct inode *dir,
1302   */
1303  static struct dx_frame *
1304  dx_probe(struct ext4_filename *fname, struct inode *dir,
1305 -        struct dx_hash_info *hinfo, struct dx_frame *frame_in)
1306 +        struct dx_hash_info *hinfo, struct dx_frame *frame_in,
1307 +        struct htree_lock *lck)
1308  {
1309         unsigned count, indirect;
1310 -       struct dx_entry *at, *entries, *p, *q, *m;
1311 +       struct dx_entry *at, *entries, *p, *q, *m, *dx = NULL;
1312         struct dx_root_info *info;
1313         struct dx_frame *frame = frame_in;
1314         struct dx_frame *ret_err = ERR_PTR(ERR_BAD_DX_DIR);
1315 @@ -808,8 +1040,15 @@ dx_probe(struct ext4_filename *fname, struct inode *dir,
1316
1317         dxtrace(printk("Look up %x", hash));
1318         while (1) {
1319 +               if (indirect == 0) { /* the last index level */
1320 +                       /* NB: ext4_htree_dx_lock() could be noop if
1321 +                        * DX-lock flag is not set for current operation */
1322 +                       ext4_htree_dx_lock(lck, dx);
1323 +                       ext4_htree_spin_lock(lck, dx, NULL);
1324 +               }
1325                 count = dx_get_count(entries);
1326 -               if (!count || count > dx_get_limit(entries)) {
1327 +               if (count == 0 || count > dx_get_limit(entries)) {
1328 +                       ext4_htree_spin_unlock(lck); /* release spin */
1329                         ext4_warning_inode(dir,
1330                                            "dx entry: count %u beyond limit %u",
1331                                            count, dx_get_limit(entries));
1332 @@ -847,8 +1086,70 @@ dx_probe(struct ext4_filename *fname, struct inode *dir,
1333                                dx_get_block(at)));
1334                 frame->entries = entries;
1335                 frame->at = at;
1336 -               if (!indirect--)
1337 +
1338 +               if (indirect == 0) { /* the last index level */
1339 +                       struct ext4_dir_lock_data *ld;
1340 +                       u64 myblock;
1341 +
1342 +                       /* By default we only lock DE-block, however, we will
1343 +                        * also lock the last level DX-block if:
1344 +                        * a) there is hash collision
1345 +                        *    we will set DX-lock flag (a few lines below)
1346 +                        *    and redo to lock DX-block
1347 +                        *    see detail in dx_probe_hash_collision()
1348 +                        * b) it's a retry from splitting
1349 +                        *    we need to lock the last level DX-block so nobody
1350 +                        *    else can split any leaf blocks under the same
1351 +                        *    DX-block, see detail in ext4_dx_add_entry()
1352 +                        */
1353 +                       if (ext4_htree_dx_locked(lck)) {
1354 +                               /* DX-block is locked, just lock DE-block
1355 +                                * and return */
1356 +                               ext4_htree_spin_unlock(lck);
1357 +                               if (!ext4_htree_safe_locked(lck))
1358 +                                       ext4_htree_de_lock(lck, frame->at);
1359 +                               return frame;
1360 +                       }
1361 +                       /* it's pdirop and no DX lock */
1362 +                       if (dx_probe_hash_collision(lck, entries, at, hash) ==
1363 +                           DX_HASH_COL_YES) {
1364 +                               /* found hash collision, set DX-lock flag
1365 +                                * and retry to abtain DX-lock */
1366 +                               ext4_htree_spin_unlock(lck);
1367 +                               ext4_htree_dx_need_lock(lck);
1368 +                               continue;
1369 +                       }
1370 +                       ld = ext4_htree_lock_data(lck);
1371 +                       /* because I don't lock DX, so @at can't be trusted
1372 +                        * after I release spinlock so I have to save it */
1373 +                       ld->ld_at = at;
1374 +                       ld->ld_at_entry = *at;
1375 +                       ld->ld_count = dx_get_count(entries);
1376 +
1377 +                       frame->at = &ld->ld_at_entry;
1378 +                       myblock = dx_get_block(at);
1379 +
1380 +                       /* NB: ordering locking */
1381 +                       ext4_htree_spin_unlock_listen(lck, &myblock);
1382 +                       /* other thread can split this DE-block because:
1383 +                        * a) I don't have lock for the DE-block yet
1384 +                        * b) I released spinlock on DX-block
1385 +                        * if it happened I can detect it by listening
1386 +                        * splitting event on this DE-block */
1387 +                       ext4_htree_de_lock(lck, frame->at);
1388 +                       ext4_htree_spin_stop_listen(lck);
1389 +
1390 +                       if (myblock == EXT4_HTREE_NODE_CHANGED) {
1391 +                               /* someone split this DE-block before
1392 +                                * I locked it, I need to retry and lock
1393 +                                * valid DE-block */
1394 +                               ext4_htree_de_unlock(lck);
1395 +                               continue;
1396 +                       }
1397                         return frame;
1398 +               }
1399 +               dx = at;
1400 +               indirect--;
1401                 frame++;
1402                 frame->bh = ext4_read_dirblock(dir, dx_get_block(at), INDEX);
1403                 if (IS_ERR(frame->bh)) {
1404 @@ -915,7 +1216,7 @@ static void dx_release(struct dx_frame *frames)
1405  static int ext4_htree_next_block(struct inode *dir, __u32 hash,
1406                                  struct dx_frame *frame,
1407                                  struct dx_frame *frames,
1408 -                                __u32 *start_hash)
1409 +                                __u32 *start_hash, struct htree_lock *lck)
1410  {
1411         struct dx_frame *p;
1412         struct buffer_head *bh;
1413 @@ -930,12 +1231,22 @@ static int ext4_htree_next_block(struct inode *dir, __u32 hash,
1414          * this loop, num_frames indicates the number of interior
1415          * nodes need to be read.
1416          */
1417 +       ext4_htree_de_unlock(lck);
1418         while (1) {
1419 -               if (++(p->at) < p->entries + dx_get_count(p->entries))
1420 -                       break;
1421 +               if (num_frames > 0 || ext4_htree_dx_locked(lck)) {
1422 +                       /* num_frames > 0 :
1423 +                        *   DX block
1424 +                        * ext4_htree_dx_locked:
1425 +                        *   frame->at is reliable pointer returned by dx_probe,
1426 +                        *   otherwise dx_probe already knew no collision */
1427 +                       if (++(p->at) < p->entries + dx_get_count(p->entries))
1428 +                               break;
1429 +               }
1430                 if (p == frames)
1431                         return 0;
1432                 num_frames++;
1433 +               if (num_frames == 1)
1434 +                       ext4_htree_dx_unlock(lck);
1435                 p--;
1436         }
1437
1438 @@ -958,6 +1269,13 @@ static int ext4_htree_next_block(struct inode *dir, __u32 hash,
1439          * block so no check is necessary
1440          */
1441         while (num_frames--) {
1442 +               if (num_frames == 0) {
1443 +                       /* it's not always necessary, we just don't want to
1444 +                        * detect hash collision again */
1445 +                       ext4_htree_dx_need_lock(lck);
1446 +                       ext4_htree_dx_lock(lck, p->at);
1447 +               }
1448 +
1449                 bh = ext4_read_dirblock(dir, dx_get_block(p->at), INDEX);
1450                 if (IS_ERR(bh))
1451                         return PTR_ERR(bh);
1452 @@ -966,6 +1284,7 @@ static int ext4_htree_next_block(struct inode *dir, __u32 hash,
1453                 p->bh = bh;
1454                 p->at = p->entries = ((struct dx_node *) bh->b_data)->entries;
1455         }
1456 +       ext4_htree_de_lock(lck, p->at);
1457         return 1;
1458  }
1459
1460 @@ -1110,10 +1429,10 @@ int ext4_htree_fill_tree(struct file *dir_file, __u32 start_hash,
1461         }
1462         hinfo.hash = start_hash;
1463         hinfo.minor_hash = 0;
1464 -       frame = dx_probe(NULL, dir, &hinfo, frames);
1465 +       /* assume it's PR locked */
1466 +       frame = dx_probe(NULL, dir, &hinfo, frames, NULL);
1467         if (IS_ERR(frame))
1468                 return PTR_ERR(frame);
1469 -
1470         /* Add '.' and '..' from the htree header */
1471         if (!start_hash && !start_minor_hash) {
1472                 de = (struct ext4_dir_entry_2 *) frames[0].bh->b_data;
1473 @@ -1148,7 +1467,7 @@ int ext4_htree_fill_tree(struct file *dir_file, __u32 start_hash,
1474                 count += ret;
1475                 hashval = ~0;
1476                 ret = ext4_htree_next_block(dir, HASH_NB_ALWAYS,
1477 -                                           frame, frames, &hashval);
1478 +                                           frame, frames, &hashval, NULL);
1479                 *next_hash = hashval;
1480                 if (ret < 0) {
1481                         err = ret;
1482 @@ -1372,10 +1691,10 @@ static int is_dx_internal_node(struct inode *dir, ext4_lblk_t block,
1483   * The returned buffer_head has ->b_count elevated.  The caller is expected
1484   * to brelse() it when appropriate.
1485   */
1486 -static struct buffer_head * ext4_find_entry (struct inode *dir,
1487 +struct buffer_head *__ext4_find_entry(struct inode *dir,
1488                                         const struct qstr *d_name,
1489                                         struct ext4_dir_entry_2 **res_dir,
1490 -                                       int *inlined)
1491 +                                       int *inlined, struct htree_lock *lck)
1492  {
1493         struct super_block *sb;
1494         struct buffer_head *bh_use[NAMEI_RA_SIZE];
1495 @@ -1423,7 +1742,7 @@ static struct buffer_head * ext4_find_entry (struct inode *dir,
1496                 goto restart;
1497         }
1498         if (is_dx(dir)) {
1499 -               ret = ext4_dx_find_entry(dir, &fname, res_dir);
1500 +               ret = ext4_dx_find_entry(dir, &fname, res_dir, lck);
1501                 /*
1502                  * On success, or if the error was file not found,
1503                  * return.  Otherwise, fall back to doing a search the
1504 @@ -1433,6 +1752,7 @@ static struct buffer_head * ext4_find_entry (struct inode *dir,
1505                         goto cleanup_and_exit;
1506                 dxtrace(printk(KERN_DEBUG "ext4_find_entry: dx failed, "
1507                                "falling back\n"));
1508 +               ext4_htree_safe_relock(lck);
1509         }
1510         nblocks = dir->i_size >> EXT4_BLOCK_SIZE_BITS(sb);
1511         start = EXT4_I(dir)->i_dir_start_lookup;
1512 @@ -1528,10 +1848,12 @@ cleanup_and_exit:
1513         ext4_fname_free_filename(&fname);
1514         return ret;
1515  }
1516 +EXPORT_SYMBOL(__ext4_find_entry);
1517
1518  static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
1519                         struct ext4_filename *fname,
1520 -                       struct ext4_dir_entry_2 **res_dir)
1521 +                       struct ext4_dir_entry_2 **res_dir,
1522 +                       struct htree_lock *lck)
1523  {
1524         struct super_block * sb = dir->i_sb;
1525         struct dx_frame frames[EXT4_HTREE_LEVEL], *frame;
1526 @@ -1543,7 +1865,7 @@ static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
1527  #ifdef CONFIG_EXT4_FS_ENCRYPTION
1528         *res_dir = NULL;
1529  #endif
1530 -       frame = dx_probe(fname, dir, NULL, frames);
1531 +       frame = dx_probe(fname, dir, NULL, frames, lck);
1532         if (IS_ERR(frame))
1533                 return (struct buffer_head *) frame;
1534         do {
1535 @@ -1565,7 +1887,7 @@ static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
1536
1537                 /* Check to see if we should continue to search */
1538                 retval = ext4_htree_next_block(dir, fname->hinfo.hash, frame,
1539 -                                              frames, NULL);
1540 +                                              frames, NULL, lck);
1541                 if (retval < 0) {
1542                         ext4_warning_inode(dir,
1543                                 "error %d reading directory index block",
1544 @@ -1738,8 +2060,9 @@ static struct ext4_dir_entry_2* dx_pack_dirents(char *base, unsigned blocksize)
1545   * Returns pointer to de in block into which the new entry will be inserted.
1546   */
1547  static struct ext4_dir_entry_2 *do_split(handle_t *handle, struct inode *dir,
1548 -                       struct buffer_head **bh,struct dx_frame *frame,
1549 -                       struct dx_hash_info *hinfo)
1550 +                       struct buffer_head **bh, struct dx_frame *frames,
1551 +                       struct dx_frame *frame, struct dx_hash_info *hinfo,
1552 +                       struct htree_lock *lck)
1553  {
1554         unsigned blocksize = dir->i_sb->s_blocksize;
1555         unsigned count, continued;
1556 @@ -1801,8 +2124,14 @@ static struct ext4_dir_entry_2 *do_split(handle_t *handle, struct inode *dir,
1557                                         hash2, split, count-split));
1558
1559         /* Fancy dance to stay within two buffers */
1560 -       de2 = dx_move_dirents(data1, data2, map + split, count - split,
1561 -                             blocksize);
1562 +       if (hinfo->hash < hash2) {
1563 +               de2 = dx_move_dirents(data1, data2, map + split,
1564 +                                     count - split, blocksize);
1565 +       } else {
1566 +               /* make sure we will add entry to the same block which
1567 +                * we have already locked */
1568 +               de2 = dx_move_dirents(data1, data2, map, split, blocksize);
1569 +       }
1570         de = dx_pack_dirents(data1, blocksize);
1571         de->rec_len = ext4_rec_len_to_disk(data1 + (blocksize - csum_size) -
1572                                            (char *) de,
1573 @@ -1823,12 +2152,21 @@ static struct ext4_dir_entry_2 *do_split(handle_t *handle, struct inode *dir,
1574         dxtrace(dx_show_leaf(dir, hinfo, (struct ext4_dir_entry_2 *) data2,
1575                         blocksize, 1));
1576
1577 -       /* Which block gets the new entry? */
1578 -       if (hinfo->hash >= hash2) {
1579 -               swap(*bh, bh2);
1580 -               de = de2;
1581 +       ext4_htree_spin_lock(lck, frame > frames ? (frame - 1)->at : NULL,
1582 +                            frame->at); /* notify block is being split */
1583 +       if (hinfo->hash < hash2) {
1584 +               dx_insert_block(frame, hash2 + continued, newblock);
1585 +
1586 +       } else {
1587 +               /* switch block number */
1588 +               dx_insert_block(frame, hash2 + continued,
1589 +                               dx_get_block(frame->at));
1590 +               dx_set_block(frame->at, newblock);
1591 +               (frame->at)++;
1592         }
1593 -       dx_insert_block(frame, hash2 + continued, newblock);
1594 +       ext4_htree_spin_unlock(lck);
1595 +       ext4_htree_dx_unlock(lck);
1596 +
1597         err = ext4_handle_dirty_dirent_node(handle, dir, bh2);
1598         if (err)
1599                 goto journal_error;
1600 @@ -2122,7 +2460,7 @@ static int make_indexed_dir(handle_t *handle, struct ext4_filename *fname,
1601         if (retval)
1602                 goto out_frames;
1603
1604 -       de = do_split(handle,dir, &bh, frame, &fname->hinfo);
1605 +       de = do_split(handle,dir, &bh, frames, frame, &fname->hinfo, NULL);
1606         if (IS_ERR(de)) {
1607                 retval = PTR_ERR(de);
1608                 goto out_frames;
1609 @@ -2233,8 +2571,8 @@ out:
1610   * may not sleep between calling this and putting something into
1611   * the entry, as someone else might have used it while you slept.
1612   */
1613 -static int ext4_add_entry(handle_t *handle, struct dentry *dentry,
1614 -                         struct inode *inode)
1615 +int __ext4_add_entry(handle_t *handle, struct dentry *dentry,
1616 +                     struct inode *inode, struct htree_lock *lck)
1617  {
1618         struct inode *dir = d_inode(dentry->d_parent);
1619         struct buffer_head *bh = NULL;
1620 @@ -2275,9 +2613,10 @@ static int ext4_add_entry(handle_t *handle, struct dentry *dentry,
1621                 if (dentry->d_name.len == 2 &&
1622                     memcmp(dentry->d_name.name, "..", 2) == 0)
1623                         return ext4_update_dotdot(handle, dentry, inode);
1624 -               retval = ext4_dx_add_entry(handle, &fname, dentry, inode);
1625 +               retval = ext4_dx_add_entry(handle, &fname, dentry, inode, lck);
1626                 if (!retval || (retval != ERR_BAD_DX_DIR))
1627                         goto out;
1628 +               ext4_htree_safe_relock(lck);
1629                 ext4_clear_inode_flag(dir, EXT4_INODE_INDEX);
1630                 dx_fallback++;
1631                 ext4_mark_inode_dirty(handle, dir);
1632 @@ -2327,12 +2666,14 @@ out:
1633                 ext4_set_inode_state(inode, EXT4_STATE_NEWENTRY);
1634         return retval;
1635  }
1636 +EXPORT_SYMBOL(__ext4_add_entry);
1637
1638  /*
1639   * Returns 0 for success, or a negative error value
1640   */
1641  static int ext4_dx_add_entry(handle_t *handle, struct ext4_filename *fname,
1642 -                            struct dentry *dentry, struct inode *inode)
1643 +                            struct dentry *dentry, struct inode *inode,
1644 +                            struct htree_lock *lck)
1645  {
1646         struct dx_frame frames[EXT4_HTREE_LEVEL], *frame;
1647         struct dx_entry *entries, *at;
1648 @@ -2345,7 +2686,7 @@ static int ext4_dx_add_entry(handle_t *handle, struct ext4_filename *fname,
1649
1650  again:
1651         restart = 0;
1652 -       frame = dx_probe(fname, dir, NULL, frames);
1653 +       frame = dx_probe(fname, dir, NULL, frames, lck);
1654         if (IS_ERR(frame))
1655                 return PTR_ERR(frame);
1656         entries = frame->entries;
1657 @@ -2375,6 +2716,11 @@ again:
1658                 struct dx_node *node2;
1659                 struct buffer_head *bh2;
1660
1661 +               if (!ext4_htree_safe_locked(lck)) { /* retry with EX lock */
1662 +                       ext4_htree_safe_relock(lck);
1663 +                       restart = 1;
1664 +                       goto cleanup;
1665 +               }
1666                 while (frame > frames) {
1667                         if (dx_get_count((frame - 1)->entries) <
1668                             dx_get_limit((frame - 1)->entries)) {
1669 @@ -2474,8 +2820,32 @@ again:
1670                         restart = 1;
1671                         goto cleanup;
1672                 }
1673 +       } else if (!ext4_htree_dx_locked(lck)) {
1674 +               struct ext4_dir_lock_data *ld = ext4_htree_lock_data(lck);
1675 +
1676 +               /* not well protected, require DX lock */
1677 +               ext4_htree_dx_need_lock(lck);
1678 +               at = frame > frames ? (frame - 1)->at : NULL;
1679 +
1680 +               /* NB: no risk of deadlock because it's just a try.
1681 +                *
1682 +                * NB: we check ld_count for twice, the first time before
1683 +                * having DX lock, the second time after holding DX lock.
1684 +                *
1685 +                * NB: We never free blocks for directory so far, which
1686 +                * means value returned by dx_get_count() should equal to
1687 +                * ld->ld_count if nobody split any DE-block under @at,
1688 +                * and ld->ld_at still points to valid dx_entry. */
1689 +               if ((ld->ld_count != dx_get_count(entries)) ||
1690 +                   !ext4_htree_dx_lock_try(lck, at) ||
1691 +                   (ld->ld_count != dx_get_count(entries))) {
1692 +                       restart = 1;
1693 +                       goto cleanup;
1694 +               }
1695 +               /* OK, I've got DX lock and nothing changed */
1696 +               frame->at = ld->ld_at;
1697         }
1698 -       de = do_split(handle, dir, &bh, frame, &fname->hinfo);
1699 +       de = do_split(handle, dir, &bh, frames, frame, &fname->hinfo, lck);
1700         if (IS_ERR(de)) {
1701                 err = PTR_ERR(de);
1702                 goto cleanup;
1703 @@ -2486,6 +2856,8 @@ again:
1704  journal_error:
1705         ext4_std_error(dir->i_sb, err);
1706  cleanup:
1707 +       ext4_htree_dx_unlock(lck);
1708 +       ext4_htree_de_unlock(lck);
1709         brelse(bh);
1710         dx_release(frames);
1711         /* @restart is true means htree-path has been changed, we need to
1712 diff --git a/fs/ext4/super.c b/fs/ext4/super.c
1713 index 02fe65b..be65ad4 100644
1714 --- a/fs/ext4/super.c
1715 +++ b/fs/ext4/super.c
1716 @@ -896,6 +896,7 @@ static struct inode *ext4_alloc_inode(struct super_block *sb)
1717
1718         ei->vfs_inode.i_version = 1;
1719         spin_lock_init(&ei->i_raw_lock);
1720 +       sema_init(&ei->i_append_sem, 1);
1721         INIT_LIST_HEAD(&ei->i_prealloc_list);
1722         spin_lock_init(&ei->i_prealloc_lock);
1723         ext4_es_init_tree(&ei->i_es_tree);
1724 diff --git a/include/linux/htree_lock.h b/include/linux/htree_lock.h
1725 new file mode 100644
1726 index 0000000..9dc7788
1727 --- /dev/null
1728 +++ b/include/linux/htree_lock.h
1729 @@ -0,0 +1,187 @@
1730 +/*
1731 + * include/linux/htree_lock.h
1732 + *
1733 + * Copyright (c) 2011, 2012, Intel Corporation.
1734 + *
1735 + * Author: Liang Zhen <liang@whamcloud.com>
1736 + */
1737 +
1738 +/*
1739 + * htree lock
1740 + *
1741 + * htree_lock is an advanced lock, it can support five lock modes (concept is
1742 + * taken from DLM) and it's a sleeping lock.
1743 + *
1744 + * most common use case is:
1745 + * - create a htree_lock_head for data
1746 + * - each thread (contender) creates it's own htree_lock
1747 + * - contender needs to call htree_lock(lock_node, mode) to protect data and
1748 + *   call htree_unlock to release lock
1749 + *
1750 + * Also, there is advanced use-case which is more complex, user can have
1751 + * PW/PR lock on particular key, it's mostly used while user holding shared
1752 + * lock on the htree (CW, CR)
1753 + *
1754 + * htree_lock(lock_node, HTREE_LOCK_CR); lock the htree with CR
1755 + * htree_node_lock(lock_node, HTREE_LOCK_PR, key...); lock @key with PR
1756 + * ...
1757 + * htree_node_unlock(lock_node);; unlock the key
1758 + *
1759 + * Another tip is, we can have N-levels of this kind of keys, all we need to
1760 + * do is specifying N-levels while creating htree_lock_head, then we can
1761 + * lock/unlock a specific level by:
1762 + * htree_node_lock(lock_node, mode1, key1, level1...);
1763 + * do something;
1764 + * htree_node_lock(lock_node, mode1, key2, level2...);
1765 + * do something;
1766 + * htree_node_unlock(lock_node, level2);
1767 + * htree_node_unlock(lock_node, level1);
1768 + *
1769 + * NB: for multi-level, should be careful about locking order to avoid deadlock
1770 + */
1771 +
1772 +#ifndef _LINUX_HTREE_LOCK_H
1773 +#define _LINUX_HTREE_LOCK_H
1774 +
1775 +#include <linux/list.h>
1776 +#include <linux/spinlock.h>
1777 +#include <linux/sched.h>
1778 +
1779 +/*
1780 + * Lock Modes
1781 + * more details can be found here:
1782 + * http://en.wikipedia.org/wiki/Distributed_lock_manager
1783 + */
1784 +typedef enum {
1785 +       HTREE_LOCK_EX   = 0, /* exclusive lock: incompatible with all others */
1786 +       HTREE_LOCK_PW,       /* protected write: allows only CR users */
1787 +       HTREE_LOCK_PR,       /* protected read: allow PR, CR users */
1788 +       HTREE_LOCK_CW,       /* concurrent write: allow CR, CW users */
1789 +       HTREE_LOCK_CR,       /* concurrent read: allow all but EX users */
1790 +       HTREE_LOCK_MAX,      /* number of lock modes */
1791 +} htree_lock_mode_t;
1792 +
1793 +#define HTREE_LOCK_NL          HTREE_LOCK_MAX
1794 +#define HTREE_LOCK_INVAL       0xdead10c
1795 +
1796 +enum {
1797 +       HTREE_HBITS_MIN         = 2,
1798 +       HTREE_HBITS_DEF         = 14,
1799 +       HTREE_HBITS_MAX         = 32,
1800 +};
1801 +
1802 +enum {
1803 +       HTREE_EVENT_DISABLE     = (0),
1804 +       HTREE_EVENT_RD          = (1 << HTREE_LOCK_PR),
1805 +       HTREE_EVENT_WR          = (1 << HTREE_LOCK_PW),
1806 +       HTREE_EVENT_RDWR        = (HTREE_EVENT_RD | HTREE_EVENT_WR),
1807 +};
1808 +
1809 +struct htree_lock;
1810 +
1811 +typedef void (*htree_event_cb_t)(void *target, void *event);
1812 +
1813 +struct htree_lock_child {
1814 +       struct list_head        lc_list;        /* granted list */
1815 +       htree_event_cb_t        lc_callback;    /* event callback */
1816 +       unsigned                lc_events;      /* event types */
1817 +};
1818 +
1819 +struct htree_lock_head {
1820 +       unsigned long           lh_lock;        /* bits lock */
1821 +       /* blocked lock list (htree_lock) */
1822 +       struct list_head        lh_blocked_list;
1823 +       /* # key levels */
1824 +       u16                     lh_depth;
1825 +       /* hash bits for key and limit number of locks */
1826 +       u16                     lh_hbits;
1827 +       /* counters for blocked locks */
1828 +       u16                     lh_nblocked[HTREE_LOCK_MAX];
1829 +       /* counters for granted locks */
1830 +       u16                     lh_ngranted[HTREE_LOCK_MAX];
1831 +       /* private data */
1832 +       void                    *lh_private;
1833 +       /* array of children locks */
1834 +       struct htree_lock_child lh_children[0];
1835 +};
1836 +
1837 +/* htree_lock_node_t is child-lock for a specific key (ln_value) */
1838 +struct htree_lock_node {
1839 +       htree_lock_mode_t       ln_mode;
1840 +       /* major hash key */
1841 +       u16                     ln_major_key;
1842 +       /* minor hash key */
1843 +       u16                     ln_minor_key;
1844 +       struct list_head        ln_major_list;
1845 +       struct list_head        ln_minor_list;
1846 +       /* alive list, all locks (granted, blocked, listening) are on it */
1847 +       struct list_head        ln_alive_list;
1848 +       /* blocked list */
1849 +       struct list_head        ln_blocked_list;
1850 +       /* granted list */
1851 +       struct list_head        ln_granted_list;
1852 +       void                    *ln_ev_target;
1853 +};
1854 +
1855 +struct htree_lock {
1856 +       struct task_struct      *lk_task;
1857 +       struct htree_lock_head  *lk_head;
1858 +       void                    *lk_private;
1859 +       unsigned                lk_depth;
1860 +       htree_lock_mode_t       lk_mode;
1861 +       struct list_head        lk_blocked_list;
1862 +       struct htree_lock_node  lk_nodes[0];
1863 +};
1864 +
1865 +/* create a lock head, which stands for a resource */
1866 +struct htree_lock_head *htree_lock_head_alloc(unsigned depth,
1867 +                                             unsigned hbits, unsigned priv);
1868 +/* free a lock head */
1869 +void htree_lock_head_free(struct htree_lock_head *lhead);
1870 +/* register event callback for child lock at level @depth */
1871 +void htree_lock_event_attach(struct htree_lock_head *lhead, unsigned depth,
1872 +                            unsigned events, htree_event_cb_t callback);
1873 +/* create a lock handle, which stands for a thread */
1874 +struct htree_lock *htree_lock_alloc(unsigned depth, unsigned pbytes);
1875 +/* free a lock handle */
1876 +void htree_lock_free(struct htree_lock *lck);
1877 +/* lock htree, when @wait is true, 0 is returned if the lock can't
1878 + * be granted immediately */
1879 +int htree_lock_try(struct htree_lock *lck, struct htree_lock_head *lhead,
1880 +                  htree_lock_mode_t mode, int wait);
1881 +/* unlock htree */
1882 +void htree_unlock(struct htree_lock *lck);
1883 +/* unlock and relock htree with @new_mode */
1884 +int htree_change_lock_try(struct htree_lock *lck,
1885 +                         htree_lock_mode_t new_mode, int wait);
1886 +void htree_change_mode(struct htree_lock *lck, htree_lock_mode_t mode);
1887 +/* require child lock (key) of htree at level @dep, @event will be sent to all
1888 + * listeners on this @key while lock being granted */
1889 +int htree_node_lock_try(struct htree_lock *lck, htree_lock_mode_t mode,
1890 +                       u32 key, unsigned dep, int wait, void *event);
1891 +/* release child lock at level @dep, this lock will listen on it's key
1892 + * if @event isn't NULL, event_cb will be called against @lck while granting
1893 + * any other lock at level @dep with the same key */
1894 +void htree_node_unlock(struct htree_lock *lck, unsigned dep, void *event);
1895 +/* stop listening on child lock at level @dep */
1896 +void htree_node_stop_listen(struct htree_lock *lck, unsigned dep);
1897 +/* for debug */
1898 +void htree_lock_stat_print(int depth);
1899 +void htree_lock_stat_reset(void);
1900 +
1901 +#define htree_lock(lck, lh, mode)      htree_lock_try(lck, lh, mode, 1)
1902 +#define htree_change_lock(lck, mode)   htree_change_lock_try(lck, mode, 1)
1903 +
1904 +#define htree_lock_mode(lck)           ((lck)->lk_mode)
1905 +
1906 +#define htree_node_lock(lck, mode, key, dep)   \
1907 +       htree_node_lock_try(lck, mode, key, dep, 1, NULL)
1908 +/* this is only safe in thread context of lock owner */
1909 +#define htree_node_is_granted(lck, dep)                \
1910 +       ((lck)->lk_nodes[dep].ln_mode != HTREE_LOCK_INVAL && \
1911 +        (lck)->lk_nodes[dep].ln_mode != HTREE_LOCK_NL)
1912 +/* this is only safe in thread context of lock owner */
1913 +#define htree_node_is_listening(lck, dep)      \
1914 +       ((lck)->lk_nodes[dep].ln_mode == HTREE_LOCK_NL)
1915 +
1916 +#endif