[PATCH] btrfs: Annotate unlocked accesses to transaction state

[PATCH] btrfs: Annotate unlocked accesses to transaction state

[Nikolay Borisov]

btrfs_transaction::state is protected by btrfs_fs_info::trans_lock and
all accesses to this variable should be synchornized by it. However,
there are 2 exceptions, namely checking if currently running transaction
has entered its commit phase in start_transaction and in
btrfs_should_end_transaction.

Annotate those 2, unlocked accesses with READ_ONCE respectively. Also
remove the full memory barrier in start_transaction as it provides no
ordering guarantees due to being unpaired. All other accsess  to
transaction state happen under trans_lock being held.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
---
 fs/btrfs/transaction.c | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/fs/btrfs/transaction.c b/fs/btrfs/transaction.c
index df5687a92798..e5a5c3604a9b 100644
--- a/fs/btrfs/transaction.c
+++ b/fs/btrfs/transaction.c
@@ -695,8 +695,7 @@ start_transaction(struct btrfs_root *root, unsigned int num_items,
 	h->can_flush_pending_bgs = true;
 	INIT_LIST_HEAD(&h->new_bgs);

-	smp_mb();
-	if (cur_trans->state >= TRANS_STATE_COMMIT_START &&
+	if (READ_ONCE(cur_trans->state) >= TRANS_STATE_COMMIT_START &&
 	    may_wait_transaction(fs_info, type)) {
 		current->journal_info = h;
 		btrfs_commit_transaction(h);
@@ -912,7 +911,7 @@ int btrfs_should_end_transaction(struct btrfs_trans_handle *trans)
 {
 	struct btrfs_transaction *cur_trans = trans->transaction;

-	if (cur_trans->state >= TRANS_STATE_COMMIT_START ||
+	if (READ_ONCE(cur_trans->state) >= TRANS_STATE_COMMIT_START ||
 	    test_bit(BTRFS_DELAYED_REFS_FLUSHING,
 		     &cur_trans->delayed_refs.flags))
 		return 1;
--
2.25.1

Re: [PATCH] btrfs: Annotate unlocked accesses to transaction state

[Boris Burkov]

On Tue, Nov 24, 2020 at 04:44:13PM +0200, Nikolay Borisov wrote:
> btrfs_transaction::state is protected by btrfs_fs_info::trans_lock and
> all accesses to this variable should be synchornized by it. However,
> there are 2 exceptions, namely checking if currently running transaction
> has entered its commit phase in start_transaction and in
> btrfs_should_end_transaction.
> 
> Annotate those 2, unlocked accesses with READ_ONCE respectively. Also
> remove the full memory barrier in start_transaction as it provides no
> ordering guarantees due to being unpaired. All other accsess  to
> transaction state happen under trans_lock being held.
> 
> Signed-off-by: Nikolay Borisov <nborisov@suse.com>
> ---
>  fs/btrfs/transaction.c | 5 ++---
>  1 file changed, 2 insertions(+), 3 deletions(-)
> 
> diff --git a/fs/btrfs/transaction.c b/fs/btrfs/transaction.c
> index df5687a92798..e5a5c3604a9b 100644
> --- a/fs/btrfs/transaction.c
> +++ b/fs/btrfs/transaction.c
> @@ -695,8 +695,7 @@ start_transaction(struct btrfs_root *root, unsigned int num_items,
>  	h->can_flush_pending_bgs = true;
>  	INIT_LIST_HEAD(&h->new_bgs);
> 
> -	smp_mb();
> -	if (cur_trans->state >= TRANS_STATE_COMMIT_START &&
> +	if (READ_ONCE(cur_trans->state) >= TRANS_STATE_COMMIT_START &&
>  	    may_wait_transaction(fs_info, type)) {
>  		current->journal_info = h;
>  		btrfs_commit_transaction(h);
> @@ -912,7 +911,7 @@ int btrfs_should_end_transaction(struct btrfs_trans_handle *trans)
>  {
>  	struct btrfs_transaction *cur_trans = trans->transaction;
> 
> -	if (cur_trans->state >= TRANS_STATE_COMMIT_START ||
> +	if (READ_ONCE(cur_trans->state) >= TRANS_STATE_COMMIT_START ||
>  	    test_bit(BTRFS_DELAYED_REFS_FLUSHING,
>  		     &cur_trans->delayed_refs.flags))
>  		return 1;
> --
> 2.25.1
> 

I think some parts of this have already gone in and the conditional has
been lightly refactored, so this patch no longer applies to
kdave/misc-next.

Assuming this is a change for adherence to best practices rather than a
fix for a specific issue (in which case it would be nice to have the
issue described in the message), the patch looks good to me otherwise.