btrfs: locking: use atomic for DREW lock writers

The DREW lock uses percpu variable to track lock counters and for that
it needs to allocate the structure. In btrfs_read_tree_root() or
btrfs_init_fs_root() this may add another error case or requires the
NOFS scope protection.

One way is to preallocate the structure as was suggested in
https://lore.kernel.org/linux-btrfs/20221214021125.28289-1-robbieko@synology.com/

We may avoid the allocation altogether if we don't use the percpu
variables but an atomic for the writer counter. This should not make any
difference, the DREW lock is used for truncate and NOCOW writes along
with other IO operations.

The percpu counter for writers has been there since the original commit
8257b2dc3c1a1057 "Btrfs: introduce btrfs_{start, end}_nocow_write() for
each subvolume". The reason could be to avoid hammering the same
cacheline from all the readers but then the writers do that anyway.

Signed-off-by: David Sterba <dsterba@suse.com>

diff --git a/fs/btrfs/disk-io.c b/fs/btrfs/disk-io.c
index 991ff26fb8c62af8c7502c03c1abdd260f9168e9..1b1b9e8197d66e60b4161336e8fe969d7e28150d 100644 (file)
--- a/fs/btrfs/disk-io.c
+++ b/fs/btrfs/disk-io.c
@@ -1341,17 +1341,8 @@ struct btrfs_root *btrfs_read_tree_root(struct btrfs_root *tree_root,
 static int btrfs_init_fs_root(struct btrfs_root *root, dev_t anon_dev)
 {
        int ret;
-       unsigned int nofs_flag;
 
-       /*
-        * We might be called under a transaction (e.g. indirect backref
-        * resolution) which could deadlock if it triggers memory reclaim
-        */
-       nofs_flag = memalloc_nofs_save();
-       ret = btrfs_drew_lock_init(&root->snapshot_lock);
-       memalloc_nofs_restore(nofs_flag);
-       if (ret)
-               goto fail;
+       btrfs_drew_lock_init(&root->snapshot_lock);
 
        if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID &&
            !btrfs_is_data_reloc_root(root)) {
@@ -2065,7 +2056,6 @@ void btrfs_put_root(struct btrfs_root *root)
                WARN_ON(test_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state));
                if (root->anon_dev)
                        free_anon_bdev(root->anon_dev);
-               btrfs_drew_lock_destroy(&root->snapshot_lock);
                free_root_extent_buffers(root);
 #ifdef CONFIG_BTRFS_DEBUG
                spin_lock(&root->fs_info->fs_roots_radix_lock);

diff --git a/fs/btrfs/locking.c b/fs/btrfs/locking.c
index 870528d87526572c8cba03631445c9f6fb4b599b..3a496b0d3d2b0327599bafd7bbace9629cffc177 100644 (file)
--- a/fs/btrfs/locking.c
+++ b/fs/btrfs/locking.c
@@ -325,24 +325,12 @@ struct extent_buffer *btrfs_try_read_lock_root_node(struct btrfs_root *root)
  * acquire the lock.
  */
 
-int btrfs_drew_lock_init(struct btrfs_drew_lock *lock)
+void btrfs_drew_lock_init(struct btrfs_drew_lock *lock)
 {
-       int ret;
-
-       ret = percpu_counter_init(&lock->writers, 0, GFP_KERNEL);
-       if (ret)
-               return ret;
-
        atomic_set(&lock->readers, 0);
+       atomic_set(&lock->writers, 0);
        init_waitqueue_head(&lock->pending_readers);
        init_waitqueue_head(&lock->pending_writers);
-
-       return 0;
-}
-
-void btrfs_drew_lock_destroy(struct btrfs_drew_lock *lock)
-{
-       percpu_counter_destroy(&lock->writers);
 }
 
 /* Return true if acquisition is successful, false otherwise */
@@ -351,10 +339,10 @@ bool btrfs_drew_try_write_lock(struct btrfs_drew_lock *lock)
        if (atomic_read(&lock->readers))
                return false;
 
-       percpu_counter_inc(&lock->writers);
+       atomic_inc(&lock->writers);
 
        /* Ensure writers count is updated before we check for pending readers */
-       smp_mb();
+       smp_mb__after_atomic();
        if (atomic_read(&lock->readers)) {
                btrfs_drew_write_unlock(lock);
                return false;
@@ -374,7 +362,7 @@ void btrfs_drew_write_lock(struct btrfs_drew_lock *lock)
 
 void btrfs_drew_write_unlock(struct btrfs_drew_lock *lock)
 {
-       percpu_counter_dec(&lock->writers);
+       atomic_dec(&lock->writers);
        cond_wake_up(&lock->pending_readers);
 }
 
@@ -390,8 +378,7 @@ void btrfs_drew_read_lock(struct btrfs_drew_lock *lock)
         */
        smp_mb__after_atomic();
 
-       wait_event(lock->pending_readers,
-                  percpu_counter_sum(&lock->writers) == 0);
+       wait_event(lock->pending_readers, atomic_read(&lock->writers) == 0);
 }
 
 void btrfs_drew_read_unlock(struct btrfs_drew_lock *lock)

diff --git a/fs/btrfs/locking.h b/fs/btrfs/locking.h
index 11c2269b4b6f6504a3d722ee9dda4a6f94c8c45e..edb9b4a0dba15d7e444da3f2945138d1aaa4881e 100644 (file)
--- a/fs/btrfs/locking.h
+++ b/fs/btrfs/locking.h
@@ -195,13 +195,12 @@ static inline void btrfs_tree_unlock_rw(struct extent_buffer *eb, int rw)
 
 struct btrfs_drew_lock {
        atomic_t readers;
-       struct percpu_counter writers;
+       atomic_t writers;
        wait_queue_head_t pending_writers;
        wait_queue_head_t pending_readers;
 };
 
-int btrfs_drew_lock_init(struct btrfs_drew_lock *lock);
-void btrfs_drew_lock_destroy(struct btrfs_drew_lock *lock);
+void btrfs_drew_lock_init(struct btrfs_drew_lock *lock);
 void btrfs_drew_write_lock(struct btrfs_drew_lock *lock);
 bool btrfs_drew_try_write_lock(struct btrfs_drew_lock *lock);
 void btrfs_drew_write_unlock(struct btrfs_drew_lock *lock);