aboutsummaryrefslogtreecommitdiffstats
path: root/fs
diff options
context:
space:
mode:
authorChristian Brauner <brauner@kernel.org>2025-04-22 09:50:32 +0200
committerChristian Brauner <brauner@kernel.org>2025-04-22 18:16:09 +0200
commit53f7eedd88d144d8d1a83cad5fba1fb75b22b19d (patch)
tree1ec506089a27903c16eb743dcd238290518261ea /fs
parent559a0d7bf1a6e5a5d0ad4ab4b0089145042e3109 (diff)
parent2d900efff915fe24c3948d28eef9078953d87fec (diff)
downloadnet-53f7eedd88d144d8d1a83cad5fba1fb75b22b19d.tar.gz
Merge patch series "fs/buffer: split pagecache lookups into atomic or blocking"
Davidlohr Bueso <dave@stgolabs.net> says: This is a respin of the series[0] to address the sleep in atomic scenarios for noref migration with large folios, introduced in: 3c20917120ce61 ("block/bdev: enable large folio support for large logical block sizes") The main difference is that it removes the first patch and moves the fix (reducing the i_private_lock critical region in the migration path) to the final patch, which also introduces the new BH_Migrate flag. It also simplifies the locking scheme in patch 1 to avoid folio trylocking in the atomic lookup cases. So essentially blocking users will take the folio lock and hence wait for migration, and otherwise nonblocking callers will bail the lookup if a noref migration is on-going. Blocking callers will also benefit from potential performance gains by reducing contention on the spinlock for bdev mappings. * patches from https://lore.kernel.org/20250418015921.132400-1-dave@stgolabs.net: mm/migrate: fix sleep in atomic for large folios and buffer heads fs/ext4: use sleeping version of sb_find_get_block() fs/jbd2: use sleeping version of __find_get_block() fs/ocfs2: use sleeping version of __find_get_block() fs/buffer: use sleeping version of __find_get_block() fs/buffer: introduce sleeping flavors for pagecache lookups fs/buffer: split locking for pagecache lookups Link: https://lore.kernel.org/20250418015921.132400-1-dave@stgolabs.net Signed-off-by: Christian Brauner <brauner@kernel.org>
Diffstat (limited to 'fs')
-rw-r--r--fs/buffer.c73
-rw-r--r--fs/ext4/ialloc.c3
-rw-r--r--fs/ext4/mballoc.c3
-rw-r--r--fs/jbd2/revoke.c15
-rw-r--r--fs/ocfs2/journal.c2
5 files changed, 68 insertions, 28 deletions
diff --git a/fs/buffer.c b/fs/buffer.c
index c7abb4a029dc84..7be23ff20b2733 100644
--- a/fs/buffer.c
+++ b/fs/buffer.c
@@ -176,18 +176,8 @@ void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
}
EXPORT_SYMBOL(end_buffer_write_sync);
-/*
- * Various filesystems appear to want __find_get_block to be non-blocking.
- * But it's the page lock which protects the buffers. To get around this,
- * we get exclusion from try_to_free_buffers with the blockdev mapping's
- * i_private_lock.
- *
- * Hack idea: for the blockdev mapping, i_private_lock contention
- * may be quite high. This code could TryLock the page, and if that
- * succeeds, there is no need to take i_private_lock.
- */
static struct buffer_head *
-__find_get_block_slow(struct block_device *bdev, sector_t block)
+__find_get_block_slow(struct block_device *bdev, sector_t block, bool atomic)
{
struct address_space *bd_mapping = bdev->bd_mapping;
const int blkbits = bd_mapping->host->i_blkbits;
@@ -204,10 +194,28 @@ __find_get_block_slow(struct block_device *bdev, sector_t block)
if (IS_ERR(folio))
goto out;
- spin_lock(&bd_mapping->i_private_lock);
+ /*
+ * Folio lock protects the buffers. Callers that cannot block
+ * will fallback to serializing vs try_to_free_buffers() via
+ * the i_private_lock.
+ */
+ if (atomic)
+ spin_lock(&bd_mapping->i_private_lock);
+ else
+ folio_lock(folio);
+
head = folio_buffers(folio);
if (!head)
goto out_unlock;
+ /*
+ * Upon a noref migration, the folio lock serializes here;
+ * otherwise bail.
+ */
+ if (test_bit_acquire(BH_Migrate, &head->b_state)) {
+ WARN_ON(!atomic);
+ goto out_unlock;
+ }
+
bh = head;
do {
if (!buffer_mapped(bh))
@@ -236,7 +244,10 @@ __find_get_block_slow(struct block_device *bdev, sector_t block)
1 << blkbits);
}
out_unlock:
- spin_unlock(&bd_mapping->i_private_lock);
+ if (atomic)
+ spin_unlock(&bd_mapping->i_private_lock);
+ else
+ folio_unlock(folio);
folio_put(folio);
out:
return ret;
@@ -656,7 +667,9 @@ EXPORT_SYMBOL(generic_buffers_fsync);
void write_boundary_block(struct block_device *bdev,
sector_t bblock, unsigned blocksize)
{
- struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
+ struct buffer_head *bh;
+
+ bh = __find_get_block_nonatomic(bdev, bblock + 1, blocksize);
if (bh) {
if (buffer_dirty(bh))
write_dirty_buffer(bh, 0);
@@ -1386,16 +1399,18 @@ lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
/*
* Perform a pagecache lookup for the matching buffer. If it's there, refresh
* it in the LRU and mark it as accessed. If it is not present then return
- * NULL
+ * NULL. Atomic context callers may also return NULL if the buffer is being
+ * migrated; similarly the page is not marked accessed either.
*/
-struct buffer_head *
-__find_get_block(struct block_device *bdev, sector_t block, unsigned size)
+static struct buffer_head *
+find_get_block_common(struct block_device *bdev, sector_t block,
+ unsigned size, bool atomic)
{
struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
if (bh == NULL) {
/* __find_get_block_slow will mark the page accessed */
- bh = __find_get_block_slow(bdev, block);
+ bh = __find_get_block_slow(bdev, block, atomic);
if (bh)
bh_lru_install(bh);
} else
@@ -1403,8 +1418,23 @@ __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
return bh;
}
+
+struct buffer_head *
+__find_get_block(struct block_device *bdev, sector_t block, unsigned size)
+{
+ return find_get_block_common(bdev, block, size, true);
+}
EXPORT_SYMBOL(__find_get_block);
+/* same as __find_get_block() but allows sleeping contexts */
+struct buffer_head *
+__find_get_block_nonatomic(struct block_device *bdev, sector_t block,
+ unsigned size)
+{
+ return find_get_block_common(bdev, block, size, false);
+}
+EXPORT_SYMBOL(__find_get_block_nonatomic);
+
/**
* bdev_getblk - Get a buffer_head in a block device's buffer cache.
* @bdev: The block device.
@@ -1422,7 +1452,12 @@ EXPORT_SYMBOL(__find_get_block);
struct buffer_head *bdev_getblk(struct block_device *bdev, sector_t block,
unsigned size, gfp_t gfp)
{
- struct buffer_head *bh = __find_get_block(bdev, block, size);
+ struct buffer_head *bh;
+
+ if (gfpflags_allow_blocking(gfp))
+ bh = __find_get_block_nonatomic(bdev, block, size);
+ else
+ bh = __find_get_block(bdev, block, size);
might_alloc(gfp);
if (bh)
diff --git a/fs/ext4/ialloc.c b/fs/ext4/ialloc.c
index 38bc8d74f4cc23..e7ecc7c8a72969 100644
--- a/fs/ext4/ialloc.c
+++ b/fs/ext4/ialloc.c
@@ -691,7 +691,8 @@ static int recently_deleted(struct super_block *sb, ext4_group_t group, int ino)
if (!bh || !buffer_uptodate(bh))
/*
* If the block is not in the buffer cache, then it
- * must have been written out.
+ * must have been written out, or, most unlikely, is
+ * being migrated - false failure should be OK here.
*/
goto out;
diff --git a/fs/ext4/mballoc.c b/fs/ext4/mballoc.c
index f88424c281943f..1e98c5be4e0ad5 100644
--- a/fs/ext4/mballoc.c
+++ b/fs/ext4/mballoc.c
@@ -6642,7 +6642,8 @@ void ext4_free_blocks(handle_t *handle, struct inode *inode,
for (i = 0; i < count; i++) {
cond_resched();
if (is_metadata)
- bh = sb_find_get_block(inode->i_sb, block + i);
+ bh = sb_find_get_block_nonatomic(inode->i_sb,
+ block + i);
ext4_forget(handle, is_metadata, inode, bh, block + i);
}
}
diff --git a/fs/jbd2/revoke.c b/fs/jbd2/revoke.c
index 0cf0fddbee81c8..1467f6790747d8 100644
--- a/fs/jbd2/revoke.c
+++ b/fs/jbd2/revoke.c
@@ -345,7 +345,8 @@ int jbd2_journal_revoke(handle_t *handle, unsigned long long blocknr,
bh = bh_in;
if (!bh) {
- bh = __find_get_block(bdev, blocknr, journal->j_blocksize);
+ bh = __find_get_block_nonatomic(bdev, blocknr,
+ journal->j_blocksize);
if (bh)
BUFFER_TRACE(bh, "found on hash");
}
@@ -355,7 +356,8 @@ int jbd2_journal_revoke(handle_t *handle, unsigned long long blocknr,
/* If there is a different buffer_head lying around in
* memory anywhere... */
- bh2 = __find_get_block(bdev, blocknr, journal->j_blocksize);
+ bh2 = __find_get_block_nonatomic(bdev, blocknr,
+ journal->j_blocksize);
if (bh2) {
/* ... and it has RevokeValid status... */
if (bh2 != bh && buffer_revokevalid(bh2))
@@ -464,7 +466,8 @@ void jbd2_journal_cancel_revoke(handle_t *handle, struct journal_head *jh)
* state machine will get very upset later on. */
if (need_cancel) {
struct buffer_head *bh2;
- bh2 = __find_get_block(bh->b_bdev, bh->b_blocknr, bh->b_size);
+ bh2 = __find_get_block_nonatomic(bh->b_bdev, bh->b_blocknr,
+ bh->b_size);
if (bh2) {
if (bh2 != bh)
clear_buffer_revoked(bh2);
@@ -492,9 +495,9 @@ void jbd2_clear_buffer_revoked_flags(journal_t *journal)
struct jbd2_revoke_record_s *record;
struct buffer_head *bh;
record = (struct jbd2_revoke_record_s *)list_entry;
- bh = __find_get_block(journal->j_fs_dev,
- record->blocknr,
- journal->j_blocksize);
+ bh = __find_get_block_nonatomic(journal->j_fs_dev,
+ record->blocknr,
+ journal->j_blocksize);
if (bh) {
clear_buffer_revoked(bh);
__brelse(bh);
diff --git a/fs/ocfs2/journal.c b/fs/ocfs2/journal.c
index f1b4b3e611cb9b..c7a9729dc9d087 100644
--- a/fs/ocfs2/journal.c
+++ b/fs/ocfs2/journal.c
@@ -1249,7 +1249,7 @@ static int ocfs2_force_read_journal(struct inode *inode)
}
for (i = 0; i < p_blocks; i++, p_blkno++) {
- bh = __find_get_block(osb->sb->s_bdev, p_blkno,
+ bh = __find_get_block_nonatomic(osb->sb->s_bdev, p_blkno,
osb->sb->s_blocksize);
/* block not cached. */
if (!bh)
generated by cgit 1.2.3-korg (git 2.43.0) at 2025-12-03 10:32:04 +0000