1 LU-50 ldiskfs: pdirops patch for ldiskfs
3 Single directory performance is a critical for HPC workloads. In a
4 typical use case an application creates a separate output file for
5 each node and task in a job. As nodes and tasks increase, hundreds
6 of thousands of files may be created in a single directory within
7 a short window of time.
8 Today, both filename lookup and file system modifying operations
9 (such as create and unlink) are protected with a single lock for
10 an entire ldiskfs directory. PDO project will remove this
11 bottleneck by introducing a parallel locking mechanism for entire
12 ldiskfs directories. This work will enable multiple application
13 threads to simultaneously lookup, create and unlink in parallel.
16 - pdirops support for ldiskfs
17 - N-level htree directory
18 - integrate with osd-ldiskfs
20 Signed-off-by: Liang Zhen <liang@whamcloud.com>
21 Change-Id: I269c0e3112e68f3acd79e860dab052a68c7d7aaa
22 Reviewed-on: http://review.whamcloud.com/375
23 Reviewed-by: Andreas Dilger <adilger@whamcloud.com>
25 Index: linux-4.18.0-80.1.2.el8_0/fs/ext4/Makefile
26 ===================================================================
27 --- linux-4.18.0-80.1.2.el8_0.orig/fs/ext4/Makefile
28 +++ linux-4.18.0-80.1.2.el8_0/fs/ext4/Makefile
29 @@ -7,6 +7,7 @@ obj-$(CONFIG_EXT4_FS) += ext4.o
31 ext4-y := balloc.o bitmap.o block_validity.o dir.o ext4_jbd2.o extents.o \
32 extents_status.o file.o fsmap.o fsync.o hash.o ialloc.o \
34 indirect.o inline.o inode.o ioctl.o mballoc.o migrate.o \
35 mmp.o move_extent.o namei.o page-io.o readpage.o resize.o \
36 super.o symlink.o sysfs.o xattr.o xattr_trusted.o xattr_user.o
37 Index: linux-4.18.0-305.el8/fs/ext4/ext4.h
38 ===================================================================
39 --- linux-4.18.0-305.el8.orig/fs/ext4/ext4.h
40 +++ linux-4.18.0-305.el8/fs/ext4/ext4.h
42 #include <linux/timer.h>
43 #include <linux/version.h>
44 #include <linux/wait.h>
45 +#include <linux/htree_lock.h>
46 #include <linux/sched/signal.h>
47 #include <linux/blockgroup_lock.h>
48 #include <linux/percpu_counter.h>
49 @@ -956,6 +957,9 @@ struct ext4_inode_info {
51 ext4_fsblk_t i_file_acl;
53 + /* following fields for parallel directory operations -bzzz */
54 + struct semaphore i_append_sem;
57 * i_block_group is the number of the block group which contains
58 * this file's inode. Constant across the lifetime of the inode,
59 @@ -2204,6 +2208,72 @@ struct dx_hash_info
61 #define HASH_NB_ALWAYS 1
63 +/* assume name-hash is protected by upper layer */
64 +#define EXT4_HTREE_LOCK_HASH 0
66 +enum ext4_pdo_lk_types {
67 +#if EXT4_HTREE_LOCK_HASH
70 + EXT4_LK_DX, /* index block */
71 + EXT4_LK_DE, /* directory entry block */
72 + EXT4_LK_SPIN, /* spinlock */
77 +#define EXT4_LB_RO(b) (1 << (b))
78 +/* read + write, high bits for writer */
79 +#define EXT4_LB_RW(b) ((1 << (b)) | (1 << (EXT4_LK_MAX + (b))))
81 +enum ext4_pdo_lock_bits {
83 + EXT4_LB_DX_RO = EXT4_LB_RO(EXT4_LK_DX),
84 + EXT4_LB_DX = EXT4_LB_RW(EXT4_LK_DX),
86 + EXT4_LB_DE_RO = EXT4_LB_RO(EXT4_LK_DE),
87 + EXT4_LB_DE = EXT4_LB_RW(EXT4_LK_DE),
88 + /* DX spinlock bits */
89 + EXT4_LB_SPIN_RO = EXT4_LB_RO(EXT4_LK_SPIN),
90 + EXT4_LB_SPIN = EXT4_LB_RW(EXT4_LK_SPIN),
91 + /* accurate searching */
92 + EXT4_LB_EXACT = EXT4_LB_RO(EXT4_LK_MAX << 1),
95 +enum ext4_pdo_lock_opc {
97 + EXT4_HLOCK_READDIR = (EXT4_LB_DE_RO | EXT4_LB_DX_RO),
98 + EXT4_HLOCK_LOOKUP = (EXT4_LB_DE_RO | EXT4_LB_SPIN_RO |
100 + EXT4_HLOCK_DEL = (EXT4_LB_DE | EXT4_LB_SPIN_RO |
102 + EXT4_HLOCK_ADD = (EXT4_LB_DE | EXT4_LB_SPIN_RO),
105 + EXT4_HLOCK_LOOKUP_SAFE = (EXT4_LB_DE_RO | EXT4_LB_DX_RO |
107 + EXT4_HLOCK_DEL_SAFE = (EXT4_LB_DE | EXT4_LB_DX_RO | EXT4_LB_EXACT),
108 + EXT4_HLOCK_SPLIT = (EXT4_LB_DE | EXT4_LB_DX | EXT4_LB_SPIN),
111 +extern struct htree_lock_head *ext4_htree_lock_head_alloc(unsigned hbits);
112 +#define ext4_htree_lock_head_free(lhead) htree_lock_head_free(lhead)
114 +extern struct htree_lock *ext4_htree_lock_alloc(void);
115 +#define ext4_htree_lock_free(lck) htree_lock_free(lck)
117 +extern void ext4_htree_lock(struct htree_lock *lck,
118 + struct htree_lock_head *lhead,
119 + struct inode *dir, unsigned flags);
120 +#define ext4_htree_unlock(lck) htree_unlock(lck)
122 +extern struct buffer_head *__ext4_find_entry(struct inode *dir,
123 + const struct qstr *d_name,
124 + struct ext4_dir_entry_2 **res_dir,
125 + int *inlined, struct htree_lock *lck);
126 +extern int __ext4_add_entry(handle_t *handle, struct dentry *dentry,
127 + struct inode *inode, struct htree_lock *lck);
129 struct ext4_filename {
130 const struct qstr *usr_fname;
131 struct fscrypt_str disk_name;
132 @@ -2506,12 +2576,20 @@ void ext4_insert_dentry(struct inode *in
133 struct ext4_filename *fname, void *data);
134 static inline void ext4_update_dx_flag(struct inode *inode)
136 + /* Disable it for ldiskfs, because going from a DX directory to
137 + * a non-DX directory while it is in use will completely break
138 + * the htree-locking.
139 + * If we really want to support this operation in the future,
140 + * we need to exclusively lock the directory at here which will
141 + * increase complexity of code */
143 if (!ext4_has_feature_dir_index(inode->i_sb) &&
144 ext4_test_inode_flag(inode, EXT4_INODE_INDEX)) {
145 /* ext4_iget() should have caught this... */
146 WARN_ON_ONCE(ext4_has_feature_metadata_csum(inode->i_sb));
147 ext4_clear_inode_flag(inode, EXT4_INODE_INDEX);
151 static const unsigned char ext4_filetype_table[] = {
152 DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
153 Index: linux-4.18.0-240.1.1.el8/fs/ext4/namei.c
154 ===================================================================
155 --- linux-4.18.0-240.1.1.el8.orig/fs/ext4/namei.c
156 +++ linux-4.18.0-240.1.1.el8/fs/ext4/namei.c
157 @@ -54,6 +54,7 @@ struct buffer_head *ext4_append(handle_t
160 struct buffer_head *bh;
161 + struct ext4_inode_info *ei = EXT4_I(inode);
164 if (unlikely(EXT4_SB(inode->i_sb)->s_max_dir_size_kb &&
165 @@ -61,15 +62,22 @@ struct buffer_head *ext4_append(handle_t
166 EXT4_SB(inode->i_sb)->s_max_dir_size_kb)))
167 return ERR_PTR(-ENOSPC);
169 + /* with parallel dir operations all appends
170 + * have to be serialized -bzzz */
171 + down(&ei->i_append_sem);
173 *block = inode->i_size >> inode->i_sb->s_blocksize_bits;
175 bh = ext4_bread(handle, inode, *block, EXT4_GET_BLOCKS_CREATE);
178 + up(&ei->i_append_sem);
181 inode->i_size += inode->i_sb->s_blocksize;
182 EXT4_I(inode)->i_disksize = inode->i_size;
183 BUFFER_TRACE(bh, "get_write_access");
184 err = ext4_journal_get_write_access(handle, bh);
185 + up(&ei->i_append_sem);
188 ext4_std_error(inode->i_sb, err);
189 @@ -252,7 +260,8 @@ static unsigned dx_node_limit(struct ino
190 static struct dx_frame *dx_probe(struct ext4_filename *fname,
192 struct dx_hash_info *hinfo,
193 - struct dx_frame *frame);
194 + struct dx_frame *frame,
195 + struct htree_lock *lck);
196 static void dx_release(struct dx_frame *frames);
197 static int dx_make_map(struct inode *dir, struct ext4_dir_entry_2 *de,
198 unsigned blocksize, struct dx_hash_info *hinfo,
199 @@ -266,12 +275,13 @@ static void dx_insert_block(struct dx_fr
200 static int ext4_htree_next_block(struct inode *dir, __u32 hash,
201 struct dx_frame *frame,
202 struct dx_frame *frames,
203 - __u32 *start_hash);
204 + __u32 *start_hash, struct htree_lock *lck);
205 static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
206 struct ext4_filename *fname,
207 - struct ext4_dir_entry_2 **res_dir);
208 + struct ext4_dir_entry_2 **res_dir, struct htree_lock *lck);
209 static int ext4_dx_add_entry(handle_t *handle, struct ext4_filename *fname,
210 - struct inode *dir, struct inode *inode);
211 + struct inode *dir, struct inode *inode,
212 + struct htree_lock *lck);
214 /* checksumming functions */
215 void initialize_dirent_tail(struct ext4_dir_entry_tail *t,
216 @@ -735,6 +745,227 @@ struct stats dx_show_entries(struct dx_h
218 #endif /* DX_DEBUG */
220 +/* private data for htree_lock */
221 +struct ext4_dir_lock_data {
222 + unsigned ld_flags; /* bits-map for lock types */
223 + unsigned ld_count; /* # entries of the last DX block */
224 + struct dx_entry ld_at_entry; /* copy of leaf dx_entry */
225 + struct dx_entry *ld_at; /* position of leaf dx_entry */
228 +#define ext4_htree_lock_data(l) ((struct ext4_dir_lock_data *)(l)->lk_private)
229 +#define ext4_find_entry(dir, name, dirent, inline) \
230 + __ext4_find_entry(dir, name, dirent, inline, NULL)
231 +#define ext4_add_entry(handle, dentry, inode) \
232 + __ext4_add_entry(handle, dentry, inode, NULL)
234 +/* NB: ext4_lblk_t is 32 bits so we use high bits to identify invalid blk */
235 +#define EXT4_HTREE_NODE_CHANGED (0xcafeULL << 32)
237 +static void ext4_htree_event_cb(void *target, void *event)
239 + u64 *block = (u64 *)target;
241 + if (*block == dx_get_block((struct dx_entry *)event))
242 + *block = EXT4_HTREE_NODE_CHANGED;
245 +struct htree_lock_head *ext4_htree_lock_head_alloc(unsigned hbits)
247 + struct htree_lock_head *lhead;
249 + lhead = htree_lock_head_alloc(EXT4_LK_MAX, hbits, 0);
250 + if (lhead != NULL) {
251 + htree_lock_event_attach(lhead, EXT4_LK_SPIN, HTREE_EVENT_WR,
252 + ext4_htree_event_cb);
256 +EXPORT_SYMBOL(ext4_htree_lock_head_alloc);
258 +struct htree_lock *ext4_htree_lock_alloc(void)
260 + return htree_lock_alloc(EXT4_LK_MAX,
261 + sizeof(struct ext4_dir_lock_data));
263 +EXPORT_SYMBOL(ext4_htree_lock_alloc);
265 +static htree_lock_mode_t ext4_htree_mode(unsigned flags)
268 + default: /* 0 or unknown flags require EX lock */
269 + return HTREE_LOCK_EX;
270 + case EXT4_HLOCK_READDIR:
271 + return HTREE_LOCK_PR;
272 + case EXT4_HLOCK_LOOKUP:
273 + return HTREE_LOCK_CR;
274 + case EXT4_HLOCK_DEL:
275 + case EXT4_HLOCK_ADD:
276 + return HTREE_LOCK_CW;
280 +/* return PR for read-only operations, otherwise return EX */
281 +static inline htree_lock_mode_t ext4_htree_safe_mode(unsigned flags)
283 + int writer = (flags & EXT4_LB_DE) == EXT4_LB_DE;
285 + /* 0 requires EX lock */
286 + return (flags == 0 || writer) ? HTREE_LOCK_EX : HTREE_LOCK_PR;
289 +static int ext4_htree_safe_locked(struct htree_lock *lck)
293 + if (lck == NULL || lck->lk_mode == HTREE_LOCK_EX)
296 + writer = (ext4_htree_lock_data(lck)->ld_flags & EXT4_LB_DE) ==
298 + if (writer) /* all readers & writers are excluded? */
299 + return lck->lk_mode == HTREE_LOCK_EX;
301 + /* all writers are excluded? */
302 + return lck->lk_mode == HTREE_LOCK_PR ||
303 + lck->lk_mode == HTREE_LOCK_PW ||
304 + lck->lk_mode == HTREE_LOCK_EX;
307 +/* relock htree_lock with EX mode if it's change operation, otherwise
308 + * relock it with PR mode. It's noop if PDO is disabled. */
309 +static void ext4_htree_safe_relock(struct htree_lock *lck)
311 + if (!ext4_htree_safe_locked(lck)) {
312 + unsigned flags = ext4_htree_lock_data(lck)->ld_flags;
314 + htree_change_lock(lck, ext4_htree_safe_mode(flags));
318 +void ext4_htree_lock(struct htree_lock *lck, struct htree_lock_head *lhead,
319 + struct inode *dir, unsigned flags)
321 + htree_lock_mode_t mode = is_dx(dir) ? ext4_htree_mode(flags) :
322 + ext4_htree_safe_mode(flags);
324 + ext4_htree_lock_data(lck)->ld_flags = flags;
325 + htree_lock(lck, lhead, mode);
327 + ext4_htree_safe_relock(lck); /* make sure it's safe locked */
329 +EXPORT_SYMBOL(ext4_htree_lock);
331 +static int ext4_htree_node_lock(struct htree_lock *lck, struct dx_entry *at,
332 + unsigned lmask, int wait, void *ev)
334 + u32 key = (at == NULL) ? 0 : dx_get_block(at);
337 + /* NOOP if htree is well protected or caller doesn't require the lock */
338 + if (ext4_htree_safe_locked(lck) ||
339 + !(ext4_htree_lock_data(lck)->ld_flags & lmask))
342 + mode = (ext4_htree_lock_data(lck)->ld_flags & lmask) == lmask ?
343 + HTREE_LOCK_PW : HTREE_LOCK_PR;
345 + if (htree_node_lock_try(lck, mode, key, ffz(~lmask), wait, ev))
347 + if (!(lmask & EXT4_LB_SPIN)) /* not a spinlock */
349 + cpu_relax(); /* spin until granted */
353 +static int ext4_htree_node_locked(struct htree_lock *lck, unsigned lmask)
355 + return ext4_htree_safe_locked(lck) ||
356 + htree_node_is_granted(lck, ffz(~lmask));
359 +static void ext4_htree_node_unlock(struct htree_lock *lck,
360 + unsigned lmask, void *buf)
362 + /* NB: it's safe to call mutiple times or even it's not locked */
363 + if (!ext4_htree_safe_locked(lck) &&
364 + htree_node_is_granted(lck, ffz(~lmask)))
365 + htree_node_unlock(lck, ffz(~lmask), buf);
368 +#define ext4_htree_dx_lock(lck, key) \
369 + ext4_htree_node_lock(lck, key, EXT4_LB_DX, 1, NULL)
370 +#define ext4_htree_dx_lock_try(lck, key) \
371 + ext4_htree_node_lock(lck, key, EXT4_LB_DX, 0, NULL)
372 +#define ext4_htree_dx_unlock(lck) \
373 + ext4_htree_node_unlock(lck, EXT4_LB_DX, NULL)
374 +#define ext4_htree_dx_locked(lck) \
375 + ext4_htree_node_locked(lck, EXT4_LB_DX)
377 +static void ext4_htree_dx_need_lock(struct htree_lock *lck)
379 + struct ext4_dir_lock_data *ld;
381 + if (ext4_htree_safe_locked(lck))
384 + ld = ext4_htree_lock_data(lck);
385 + switch (ld->ld_flags) {
388 + case EXT4_HLOCK_LOOKUP:
389 + ld->ld_flags = EXT4_HLOCK_LOOKUP_SAFE;
391 + case EXT4_HLOCK_DEL:
392 + ld->ld_flags = EXT4_HLOCK_DEL_SAFE;
394 + case EXT4_HLOCK_ADD:
395 + ld->ld_flags = EXT4_HLOCK_SPLIT;
400 +#define ext4_htree_de_lock(lck, key) \
401 + ext4_htree_node_lock(lck, key, EXT4_LB_DE, 1, NULL)
402 +#define ext4_htree_de_unlock(lck) \
403 + ext4_htree_node_unlock(lck, EXT4_LB_DE, NULL)
405 +#define ext4_htree_spin_lock(lck, key, event) \
406 + ext4_htree_node_lock(lck, key, EXT4_LB_SPIN, 0, event)
407 +#define ext4_htree_spin_unlock(lck) \
408 + ext4_htree_node_unlock(lck, EXT4_LB_SPIN, NULL)
409 +#define ext4_htree_spin_unlock_listen(lck, p) \
410 + ext4_htree_node_unlock(lck, EXT4_LB_SPIN, p)
412 +static void ext4_htree_spin_stop_listen(struct htree_lock *lck)
414 + if (!ext4_htree_safe_locked(lck) &&
415 + htree_node_is_listening(lck, ffz(~EXT4_LB_SPIN)))
416 + htree_node_stop_listen(lck, ffz(~EXT4_LB_SPIN));
420 + DX_HASH_COL_IGNORE, /* ignore collision while probing frames */
421 + DX_HASH_COL_YES, /* there is collision and it does matter */
422 + DX_HASH_COL_NO, /* there is no collision */
425 +static int dx_probe_hash_collision(struct htree_lock *lck,
426 + struct dx_entry *entries,
427 + struct dx_entry *at, u32 hash)
429 + if (!(lck && ext4_htree_lock_data(lck)->ld_flags & EXT4_LB_EXACT)) {
430 + return DX_HASH_COL_IGNORE; /* don't care about collision */
432 + } else if (at == entries + dx_get_count(entries) - 1) {
433 + return DX_HASH_COL_IGNORE; /* not in any leaf of this DX */
435 + } else { /* hash collision? */
436 + return ((dx_get_hash(at + 1) & ~1) == hash) ?
437 + DX_HASH_COL_YES : DX_HASH_COL_NO;
442 * Probe for a directory leaf block to search.
444 @@ -746,10 +977,11 @@ struct stats dx_show_entries(struct dx_h
446 static struct dx_frame *
447 dx_probe(struct ext4_filename *fname, struct inode *dir,
448 - struct dx_hash_info *hinfo, struct dx_frame *frame_in)
449 + struct dx_hash_info *hinfo, struct dx_frame *frame_in,
450 + struct htree_lock *lck)
452 unsigned count, indirect;
453 - struct dx_entry *at, *entries, *p, *q, *m;
454 + struct dx_entry *at, *entries, *p, *q, *m, *dx = NULL;
455 struct dx_root_info *info;
456 struct dx_frame *frame = frame_in;
457 struct dx_frame *ret_err = ERR_PTR(ERR_BAD_DX_DIR);
458 @@ -811,8 +1043,15 @@ dx_probe(struct ext4_filename *fname, st
460 dxtrace(printk("Look up %x", hash));
462 + if (indirect == 0) { /* the last index level */
463 + /* NB: ext4_htree_dx_lock() could be noop if
464 + * DX-lock flag is not set for current operation */
465 + ext4_htree_dx_lock(lck, dx);
466 + ext4_htree_spin_lock(lck, dx, NULL);
468 count = dx_get_count(entries);
469 - if (!count || count > dx_get_limit(entries)) {
470 + if (count == 0 || count > dx_get_limit(entries)) {
471 + ext4_htree_spin_unlock(lck); /* release spin */
472 ext4_warning_inode(dir,
473 "dx entry: count %u beyond limit %u",
474 count, dx_get_limit(entries));
475 @@ -851,8 +1090,70 @@ dx_probe(struct ext4_filename *fname, st
477 frame->entries = entries;
481 + if (indirect == 0) { /* the last index level */
482 + struct ext4_dir_lock_data *ld;
485 + /* By default we only lock DE-block, however, we will
486 + * also lock the last level DX-block if:
487 + * a) there is hash collision
488 + * we will set DX-lock flag (a few lines below)
489 + * and redo to lock DX-block
490 + * see detail in dx_probe_hash_collision()
491 + * b) it's a retry from splitting
492 + * we need to lock the last level DX-block so nobody
493 + * else can split any leaf blocks under the same
494 + * DX-block, see detail in ext4_dx_add_entry()
496 + if (ext4_htree_dx_locked(lck)) {
497 + /* DX-block is locked, just lock DE-block
499 + ext4_htree_spin_unlock(lck);
500 + if (!ext4_htree_safe_locked(lck))
501 + ext4_htree_de_lock(lck, frame->at);
504 + /* it's pdirop and no DX lock */
505 + if (dx_probe_hash_collision(lck, entries, at, hash) ==
507 + /* found hash collision, set DX-lock flag
508 + * and retry to abtain DX-lock */
509 + ext4_htree_spin_unlock(lck);
510 + ext4_htree_dx_need_lock(lck);
513 + ld = ext4_htree_lock_data(lck);
514 + /* because I don't lock DX, so @at can't be trusted
515 + * after I release spinlock so I have to save it */
517 + ld->ld_at_entry = *at;
518 + ld->ld_count = dx_get_count(entries);
520 + frame->at = &ld->ld_at_entry;
521 + myblock = dx_get_block(at);
523 + /* NB: ordering locking */
524 + ext4_htree_spin_unlock_listen(lck, &myblock);
525 + /* other thread can split this DE-block because:
526 + * a) I don't have lock for the DE-block yet
527 + * b) I released spinlock on DX-block
528 + * if it happened I can detect it by listening
529 + * splitting event on this DE-block */
530 + ext4_htree_de_lock(lck, frame->at);
531 + ext4_htree_spin_stop_listen(lck);
533 + if (myblock == EXT4_HTREE_NODE_CHANGED) {
534 + /* someone split this DE-block before
535 + * I locked it, I need to retry and lock
536 + * valid DE-block */
537 + ext4_htree_de_unlock(lck);
545 frame->bh = ext4_read_dirblock(dir, dx_get_block(at), INDEX);
546 if (IS_ERR(frame->bh)) {
547 @@ -921,7 +1222,7 @@ static void dx_release(struct dx_frame *
548 static int ext4_htree_next_block(struct inode *dir, __u32 hash,
549 struct dx_frame *frame,
550 struct dx_frame *frames,
552 + __u32 *start_hash, struct htree_lock *lck)
555 struct buffer_head *bh;
556 @@ -936,12 +1237,22 @@ static int ext4_htree_next_block(struct
557 * this loop, num_frames indicates the number of interior
558 * nodes need to be read.
560 + ext4_htree_de_unlock(lck);
562 - if (++(p->at) < p->entries + dx_get_count(p->entries))
564 + if (num_frames > 0 || ext4_htree_dx_locked(lck)) {
565 + /* num_frames > 0 :
567 + * ext4_htree_dx_locked:
568 + * frame->at is reliable pointer returned by dx_probe,
569 + * otherwise dx_probe already knew no collision */
570 + if (++(p->at) < p->entries + dx_get_count(p->entries))
576 + if (num_frames == 1)
577 + ext4_htree_dx_unlock(lck);
581 @@ -964,6 +1275,13 @@ static int ext4_htree_next_block(struct
582 * block so no check is necessary
584 while (num_frames--) {
585 + if (num_frames == 0) {
586 + /* it's not always necessary, we just don't want to
587 + * detect hash collision again */
588 + ext4_htree_dx_need_lock(lck);
589 + ext4_htree_dx_lock(lck, p->at);
592 bh = ext4_read_dirblock(dir, dx_get_block(p->at), INDEX);
595 @@ -972,6 +1290,7 @@ static int ext4_htree_next_block(struct
597 p->at = p->entries = ((struct dx_node *) bh->b_data)->entries;
599 + ext4_htree_de_lock(lck, p->at);
603 @@ -1119,10 +1438,10 @@ int ext4_htree_fill_tree(struct file *di
605 hinfo.hash = start_hash;
606 hinfo.minor_hash = 0;
607 - frame = dx_probe(NULL, dir, &hinfo, frames);
608 + /* assume it's PR locked */
609 + frame = dx_probe(NULL, dir, &hinfo, frames, NULL);
611 return PTR_ERR(frame);
613 /* Add '.' and '..' from the htree header */
614 if (!start_hash && !start_minor_hash) {
615 de = (struct ext4_dir_entry_2 *) frames[0].bh->b_data;
616 @@ -1162,7 +1481,7 @@ int ext4_htree_fill_tree(struct file *di
619 ret = ext4_htree_next_block(dir, HASH_NB_ALWAYS,
620 - frame, frames, &hashval);
621 + frame, frames, &hashval, NULL);
622 *next_hash = hashval;
625 @@ -1354,10 +1673,10 @@ static int is_dx_internal_node(struct in
626 * The returned buffer_head has ->b_count elevated. The caller is expected
627 * to brelse() it when appropriate.
629 -static struct buffer_head * ext4_find_entry (struct inode *dir,
630 +struct buffer_head *__ext4_find_entry(struct inode *dir,
631 const struct qstr *d_name,
632 struct ext4_dir_entry_2 **res_dir,
634 + int *inlined, struct htree_lock *lck)
636 struct super_block *sb;
637 struct buffer_head *bh_use[NAMEI_RA_SIZE];
638 @@ -1406,7 +1725,7 @@ static struct buffer_head * ext4_find_en
642 - ret = ext4_dx_find_entry(dir, &fname, res_dir);
643 + ret = ext4_dx_find_entry(dir, &fname, res_dir, lck);
645 * On success, or if the error was file not found,
646 * return. Otherwise, fall back to doing a search the
647 @@ -1416,6 +1735,7 @@ static struct buffer_head * ext4_find_en
648 goto cleanup_and_exit;
649 dxtrace(printk(KERN_DEBUG "ext4_find_entry: dx failed, "
651 + ext4_htree_safe_relock(lck);
654 nblocks = dir->i_size >> EXT4_BLOCK_SIZE_BITS(sb);
655 @@ -1507,10 +1827,12 @@ cleanup_and_exit:
656 ext4_fname_free_filename(&fname);
659 +EXPORT_SYMBOL(__ext4_find_entry);
661 static struct buffer_head * ext4_dx_find_entry(struct inode *dir,
662 struct ext4_filename *fname,
663 - struct ext4_dir_entry_2 **res_dir)
664 + struct ext4_dir_entry_2 **res_dir,
665 + struct htree_lock *lck)
667 struct super_block * sb = dir->i_sb;
668 struct dx_frame frames[EXT4_HTREE_LEVEL], *frame;
669 @@ -1521,7 +1843,7 @@ static struct buffer_head * ext4_dx_find
670 #ifdef CONFIG_EXT4_FS_ENCRYPTION
673 - frame = dx_probe(fname, dir, NULL, frames);
674 + frame = dx_probe(fname, dir, NULL, frames, lck);
676 return (struct buffer_head *) frame;
678 @@ -1543,7 +1865,7 @@ static struct buffer_head * ext4_dx_find
680 /* Check to see if we should continue to search */
681 retval = ext4_htree_next_block(dir, fname->hinfo.hash, frame,
683 + frames, NULL, lck);
685 ext4_warning_inode(dir,
686 "error %d reading directory index block",
687 @@ -1718,8 +2040,9 @@ static struct ext4_dir_entry_2* dx_pack_
688 * Returns pointer to de in block into which the new entry will be inserted.
690 static struct ext4_dir_entry_2 *do_split(handle_t *handle, struct inode *dir,
691 - struct buffer_head **bh,struct dx_frame *frame,
692 - struct dx_hash_info *hinfo)
693 + struct buffer_head **bh, struct dx_frame *frames,
694 + struct dx_frame *frame, struct dx_hash_info *hinfo,
695 + struct htree_lock *lck)
697 unsigned blocksize = dir->i_sb->s_blocksize;
698 unsigned count, continued;
699 @@ -1781,8 +2104,14 @@ static struct ext4_dir_entry_2 *do_split
700 hash2, split, count-split));
702 /* Fancy dance to stay within two buffers */
703 - de2 = dx_move_dirents(data1, data2, map + split, count - split,
705 + if (hinfo->hash < hash2) {
706 + de2 = dx_move_dirents(data1, data2, map + split,
707 + count - split, blocksize);
709 + /* make sure we will add entry to the same block which
710 + * we have already locked */
711 + de2 = dx_move_dirents(data1, data2, map, split, blocksize);
713 de = dx_pack_dirents(data1, blocksize);
714 de->rec_len = ext4_rec_len_to_disk(data1 + (blocksize - csum_size) -
716 @@ -1803,12 +2132,21 @@ static struct ext4_dir_entry_2 *do_split
717 dxtrace(dx_show_leaf(dir, hinfo, (struct ext4_dir_entry_2 *) data2,
720 - /* Which block gets the new entry? */
721 - if (hinfo->hash >= hash2) {
724 + ext4_htree_spin_lock(lck, frame > frames ? (frame - 1)->at : NULL,
725 + frame->at); /* notify block is being split */
726 + if (hinfo->hash < hash2) {
727 + dx_insert_block(frame, hash2 + continued, newblock);
730 + /* switch block number */
731 + dx_insert_block(frame, hash2 + continued,
732 + dx_get_block(frame->at));
733 + dx_set_block(frame->at, newblock);
736 - dx_insert_block(frame, hash2 + continued, newblock);
737 + ext4_htree_spin_unlock(lck);
738 + ext4_htree_dx_unlock(lck);
740 err = ext4_handle_dirty_dirent_node(handle, dir, bh2);
743 @@ -2082,7 +2420,7 @@ static int make_indexed_dir(handle_t *ha
747 - de = do_split(handle,dir, &bh2, frame, &fname->hinfo);
748 + de = do_split(handle, dir, &bh2, frames, frame, &fname->hinfo, NULL);
750 retval = PTR_ERR(de);
752 @@ -2192,8 +2530,8 @@ out:
753 * may not sleep between calling this and putting something into
754 * the entry, as someone else might have used it while you slept.
756 -static int ext4_add_entry(handle_t *handle, struct dentry *dentry,
757 - struct inode *inode)
758 +int __ext4_add_entry(handle_t *handle, struct dentry *dentry,
759 + struct inode *inode, struct htree_lock *lck)
761 struct inode *dir = d_inode(dentry->d_parent);
762 struct buffer_head *bh = NULL;
763 @@ -2234,9 +2572,10 @@ static int ext4_add_entry(handle_t *hand
764 return ext4_update_dotdot(handle, dentry, inode);
767 - retval = ext4_dx_add_entry(handle, &fname, dir, inode);
768 + retval = ext4_dx_add_entry(handle, &fname, dir, inode, lck);
769 if (!retval || (retval != ERR_BAD_DX_DIR))
771 + ext4_htree_safe_relock(lck);
772 /* Can we just ignore htree data? */
773 if (ext4_has_metadata_csum(sb)) {
774 EXT4_ERROR_INODE(dir,
775 @@ -2293,12 +2632,14 @@ out:
776 ext4_set_inode_state(inode, EXT4_STATE_NEWENTRY);
779 +EXPORT_SYMBOL(__ext4_add_entry);
782 * Returns 0 for success, or a negative error value
784 static int ext4_dx_add_entry(handle_t *handle, struct ext4_filename *fname,
785 - struct inode *dir, struct inode *inode)
786 + struct inode *dir, struct inode *inode,
787 + struct htree_lock *lck)
789 struct dx_frame frames[EXT4_HTREE_LEVEL], *frame;
790 struct dx_entry *entries, *at;
791 @@ -2310,7 +2651,7 @@ static int ext4_dx_add_entry(handle_t *h
795 - frame = dx_probe(fname, dir, NULL, frames);
796 + frame = dx_probe(fname, dir, NULL, frames, lck);
798 return PTR_ERR(frame);
799 entries = frame->entries;
800 @@ -2345,6 +2686,11 @@ again:
801 struct dx_node *node2;
802 struct buffer_head *bh2;
804 + if (!ext4_htree_safe_locked(lck)) { /* retry with EX lock */
805 + ext4_htree_safe_relock(lck);
809 while (frame > frames) {
810 if (dx_get_count((frame - 1)->entries) <
811 dx_get_limit((frame - 1)->entries)) {
812 @@ -2447,8 +2793,32 @@ again:
816 + } else if (!ext4_htree_dx_locked(lck)) {
817 + struct ext4_dir_lock_data *ld = ext4_htree_lock_data(lck);
819 + /* not well protected, require DX lock */
820 + ext4_htree_dx_need_lock(lck);
821 + at = frame > frames ? (frame - 1)->at : NULL;
823 + /* NB: no risk of deadlock because it's just a try.
825 + * NB: we check ld_count for twice, the first time before
826 + * having DX lock, the second time after holding DX lock.
828 + * NB: We never free blocks for directory so far, which
829 + * means value returned by dx_get_count() should equal to
830 + * ld->ld_count if nobody split any DE-block under @at,
831 + * and ld->ld_at still points to valid dx_entry. */
832 + if ((ld->ld_count != dx_get_count(entries)) ||
833 + !ext4_htree_dx_lock_try(lck, at) ||
834 + (ld->ld_count != dx_get_count(entries))) {
838 + /* OK, I've got DX lock and nothing changed */
839 + frame->at = ld->ld_at;
841 - de = do_split(handle, dir, &bh, frame, &fname->hinfo);
842 + de = do_split(handle, dir, &bh, frames, frame, &fname->hinfo, lck);
846 @@ -2459,6 +2829,8 @@ again:
848 ext4_std_error(dir->i_sb, err); /* this is a no-op if err == 0 */
850 + ext4_htree_dx_unlock(lck);
851 + ext4_htree_de_unlock(lck);
854 /* @restart is true means htree-path has been changed, we need to
855 Index: linux-4.18.0-80.1.2.el8_0/fs/ext4/super.c
856 ===================================================================
857 --- linux-4.18.0-80.1.2.el8_0.orig/fs/ext4/super.c
858 +++ linux-4.18.0-80.1.2.el8_0/fs/ext4/super.c
859 @@ -1009,6 +1009,7 @@ static struct inode *ext4_alloc_inode(st
861 inode_set_iversion(&ei->vfs_inode, 1);
862 spin_lock_init(&ei->i_raw_lock);
863 + sema_init(&ei->i_append_sem, 1);
864 INIT_LIST_HEAD(&ei->i_prealloc_list);
865 spin_lock_init(&ei->i_prealloc_lock);
866 ext4_es_init_tree(&ei->i_es_tree);