diff mbox series

[1/2] btrfs: fix data races when accessing the reserved amount of block reserves

Message ID 5ff1a68f4289d5bb870a499b248d329893d417ae.1708429856.git.fdmanana@suse.com (mailing list archive)
State New, archived
Headers show
Series btrfs: fix a couple KCSAN warnings | expand

Commit Message

Filipe Manana Feb. 20, 2024, 12:24 p.m. UTC 
From: Filipe Manana <fdmanana@suse.com>

At space_info.c we have several places where we access the ->reserved
field of a block reserve without taking the block reserve's spinlock
first, which makes KCSAN warn about a data race since that field is
always updated while holding the spinlock.

The reports from KCSAN are like the following:

  [  117.193526] BUG: KCSAN: data-race in btrfs_block_rsv_release [btrfs] / need_preemptive_reclaim [btrfs]

  [  117.195148] read to 0x000000017f587190 of 8 bytes by task 6303 on cpu 3:
  [  117.195172]  need_preemptive_reclaim+0x222/0x2f0 [btrfs]
  [  117.195992]  __reserve_bytes+0xbb0/0xdc8 [btrfs]
  [  117.196807]  btrfs_reserve_metadata_bytes+0x4c/0x120 [btrfs]
  [  117.197620]  btrfs_block_rsv_add+0x78/0xa8 [btrfs]
  [  117.198434]  btrfs_delayed_update_inode+0x154/0x368 [btrfs]
  [  117.199300]  btrfs_update_inode+0x108/0x1c8 [btrfs]
  [  117.200122]  btrfs_dirty_inode+0xb4/0x140 [btrfs]
  [  117.200937]  btrfs_update_time+0x8c/0xb0 [btrfs]
  [  117.201754]  touch_atime+0x16c/0x1e0
  [  117.201789]  filemap_read+0x674/0x728
  [  117.201823]  btrfs_file_read_iter+0xf8/0x410 [btrfs]
  [  117.202653]  vfs_read+0x2b6/0x498
  [  117.203454]  ksys_read+0xa2/0x150
  [  117.203473]  __s390x_sys_read+0x68/0x88
  [  117.203495]  do_syscall+0x1c6/0x210
  [  117.203517]  __do_syscall+0xc8/0xf0
  [  117.203539]  system_call+0x70/0x98

  [  117.203579] write to 0x000000017f587190 of 8 bytes by task 11 on cpu 0:
  [  117.203604]  btrfs_block_rsv_release+0x2e8/0x578 [btrfs]
  [  117.204432]  btrfs_delayed_inode_release_metadata+0x7c/0x1d0 [btrfs]
  [  117.205259]  __btrfs_update_delayed_inode+0x37c/0x5e0 [btrfs]
  [  117.206093]  btrfs_async_run_delayed_root+0x356/0x498 [btrfs]
  [  117.206917]  btrfs_work_helper+0x160/0x7a0 [btrfs]
  [  117.207738]  process_one_work+0x3b6/0x838
  [  117.207768]  worker_thread+0x75e/0xb10
  [  117.207797]  kthread+0x21a/0x230
  [  117.207830]  __ret_from_fork+0x6c/0xb8
  [  117.207861]  ret_from_fork+0xa/0x30

So add a helper to get the reserved amount of a block reserve while
holding the lock. The value may be not be up to date anymore when used by
need_preemptive_reclaim() and btrfs_preempt_reclaim_metadata_space(), but
that's ok since the worst it can do is cause more reclaim work do be done
sooner rather than later. Reading the field while holding the lock instead
of using the data_race() annotation is used in order to prevent load
tearing.

Signed-off-by: Filipe Manana <fdmanana@suse.com>
---
 fs/btrfs/block-rsv.h  | 16 ++++++++++++++++
 fs/btrfs/space-info.c | 26 +++++++++++++-------------
 2 files changed, 29 insertions(+), 13 deletions(-)

Comments

Sweet Tea Dorminy Feb. 26, 2024, 5:39 p.m. UTC | #1 
> So add a helper to get the reserved amount of a block reserve while
> holding the lock. The value may be not be up to date anymore when used by
> need_preemptive_reclaim() and btrfs_preempt_reclaim_metadata_space(), but
> that's ok since the worst it can do is cause more reclaim work do be done
> sooner rather than later. Reading the field while holding the lock instead
> of using the data_race() annotation is used in order to prevent load
> tearing.

Would READ_ONCE() be enough? My understanding is that it should be 
enough against load tearing if it's aligned, and I think both size and 
reserved are aligned? But I'm always learning more about how to do 
things right.
Filipe Manana Feb. 26, 2024, 5:55 p.m. UTC | #2 
On Mon, Feb 26, 2024 at 5:39 PM Sweet Tea Dorminy
<sweettea-kernel@dorminy.me> wrote:
>
>
> > So add a helper to get the reserved amount of a block reserve while
> > holding the lock. The value may be not be up to date anymore when used by
> > need_preemptive_reclaim() and btrfs_preempt_reclaim_metadata_space(), but
> > that's ok since the worst it can do is cause more reclaim work do be done
> > sooner rather than later. Reading the field while holding the lock instead
> > of using the data_race() annotation is used in order to prevent load
> > tearing.
>
> Would READ_ONCE() be enough? My understanding is that it should be
> enough against load tearing if it's aligned, and I think both size and
> reserved are aligned? But I'm always learning more about how to do
> things right.

READ_ONCE() works if it's paired with WRITE_ONCE() in every single
place where we update the value, and there are many such places.
Besides the quantity of places, expressions like:

rsv->reserved += X;

would be become "ugly", like this:

WRITE_ONCE(rsv->reserved, rsv->reserved + X);
diff mbox series

Patch

diff --git a/fs/btrfs/block-rsv.h b/fs/btrfs/block-rsv.h
index b621199b0130..89c99f4e9f16 100644
--- a/fs/btrfs/block-rsv.h
+++ b/fs/btrfs/block-rsv.h
@@ -108,4 +108,20 @@  static inline bool btrfs_block_rsv_full(const struct btrfs_block_rsv *rsv)
 	return data_race(rsv->full);
 }
 
+/*
+ * Get the reserved mount of a block reserve in a context where getting a stale
+ * value is acceptable, instead of accessing it directly and trigger data race
+ * warning from KCSAN.
+ */
+static inline u64 btrfs_block_rsv_reserved(struct btrfs_block_rsv *rsv)
+{
+	u64 ret;
+
+	spin_lock(&rsv->lock);
+	ret = rsv->reserved;
+	spin_unlock(&rsv->lock);
+
+	return ret;
+}
+
 #endif /* BTRFS_BLOCK_RSV_H */
diff --git a/fs/btrfs/space-info.c b/fs/btrfs/space-info.c
index a5b652c1650a..d620323d08ea 100644
--- a/fs/btrfs/space-info.c
+++ b/fs/btrfs/space-info.c
@@ -855,7 +855,7 @@  btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
 static bool need_preemptive_reclaim(struct btrfs_fs_info *fs_info,
 				    struct btrfs_space_info *space_info)
 {
-	u64 global_rsv_size = fs_info->global_block_rsv.reserved;
+	const u64 global_rsv_size = btrfs_block_rsv_reserved(&fs_info->global_block_rsv);
 	u64 ordered, delalloc;
 	u64 thresh;
 	u64 used;
@@ -955,8 +955,8 @@  static bool need_preemptive_reclaim(struct btrfs_fs_info *fs_info,
 	ordered = percpu_counter_read_positive(&fs_info->ordered_bytes) >> 1;
 	delalloc = percpu_counter_read_positive(&fs_info->delalloc_bytes);
 	if (ordered >= delalloc)
-		used += fs_info->delayed_refs_rsv.reserved +
-			fs_info->delayed_block_rsv.reserved;
+		used += btrfs_block_rsv_reserved(&fs_info->delayed_refs_rsv) +
+			btrfs_block_rsv_reserved(&fs_info->delayed_block_rsv);
 	else
 		used += space_info->bytes_may_use - global_rsv_size;
 
@@ -1172,7 +1172,7 @@  static void btrfs_preempt_reclaim_metadata_space(struct work_struct *work)
 		enum btrfs_flush_state flush;
 		u64 delalloc_size = 0;
 		u64 to_reclaim, block_rsv_size;
-		u64 global_rsv_size = global_rsv->reserved;
+		const u64 global_rsv_size = btrfs_block_rsv_reserved(global_rsv);
 
 		loops++;
 
@@ -1184,9 +1184,9 @@  static void btrfs_preempt_reclaim_metadata_space(struct work_struct *work)
 		 * assume it's tied up in delalloc reservations.
 		 */
 		block_rsv_size = global_rsv_size +
-			delayed_block_rsv->reserved +
-			delayed_refs_rsv->reserved +
-			trans_rsv->reserved;
+			btrfs_block_rsv_reserved(delayed_block_rsv) +
+			btrfs_block_rsv_reserved(delayed_refs_rsv) +
+			btrfs_block_rsv_reserved(trans_rsv);
 		if (block_rsv_size < space_info->bytes_may_use)
 			delalloc_size = space_info->bytes_may_use - block_rsv_size;
 
@@ -1206,16 +1206,16 @@  static void btrfs_preempt_reclaim_metadata_space(struct work_struct *work)
 			to_reclaim = delalloc_size;
 			flush = FLUSH_DELALLOC;
 		} else if (space_info->bytes_pinned >
-			   (delayed_block_rsv->reserved +
-			    delayed_refs_rsv->reserved)) {
+			   (btrfs_block_rsv_reserved(delayed_block_rsv) +
+			    btrfs_block_rsv_reserved(delayed_refs_rsv))) {
 			to_reclaim = space_info->bytes_pinned;
 			flush = COMMIT_TRANS;
-		} else if (delayed_block_rsv->reserved >
-			   delayed_refs_rsv->reserved) {
-			to_reclaim = delayed_block_rsv->reserved;
+		} else if (btrfs_block_rsv_reserved(delayed_block_rsv) >
+			   btrfs_block_rsv_reserved(delayed_refs_rsv)) {
+			to_reclaim = btrfs_block_rsv_reserved(delayed_block_rsv);
 			flush = FLUSH_DELAYED_ITEMS_NR;
 		} else {
-			to_reclaim = delayed_refs_rsv->reserved;
+			to_reclaim = btrfs_block_rsv_reserved(delayed_refs_rsv);
 			flush = FLUSH_DELAYED_REFS_NR;
 		}