On 6/4/20 7:13 PM, Dave Chinner wrote:
On Thu, Jun 04, 2020 at 05:01:30PM -0400, Waiman Long wrote:
---
fs/xfs/xfs_log.c | 3 ++-
fs/xfs/xfs_trans.c | 8 +++++++-
2 files changed, 9 insertions(+), 2 deletions(-)
diff --git a/fs/xfs/xfs_log.c b/fs/xfs/xfs_log.c
index 00fda2e8e738..d273d4e74ef8 100644
--- a/fs/xfs/xfs_log.c
+++ b/fs/xfs/xfs_log.c
@@ -433,7 +433,8 @@ xfs_log_reserve(
XFS_STATS_INC(mp, xs_try_logspace);
ASSERT(*ticp == NULL);
- tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent, 0);
+ tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent,
+ mp->m_super->s_writers.frozen ? KM_NOLOCKDEP : 0);
*ticp = tic;
Hi Waiman,
As I originally stated when you posted this the first time 6 months
ago: we are not going to spread this sort of conditional gunk though
the XFS codebase just to shut up lockdep false positives.
I pointed you at the way to conditionally turn of lockdep for
operations where we are doing transactions when the filesystem has
already frozen the transaction subsystem. That is:
xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
diff --git a/fs/xfs/xfs_trans.c b/fs/xfs/xfs_trans.c
index 3c94e5ff4316..3a9f394a0f02 100644
--- a/fs/xfs/xfs_trans.c
+++ b/fs/xfs/xfs_trans.c
@@ -261,8 +261,14 @@ xfs_trans_alloc(
* Allocate the handle before we do our freeze accounting and setting up
* GFP_NOFS allocation context so that we avoid lockdep false positives
* by doing GFP_KERNEL allocations inside sb_start_intwrite().
+ *
+ * To prevent false positive lockdep warning of circular locking
+ * dependency between sb_internal and fs_reclaim, disable the
+ * acquisition of the fs_reclaim pseudo-lock when the superblock
+ * has been frozen or in the process of being frozen.
*/
- tp = kmem_zone_zalloc(xfs_trans_zone, 0);
+ tp = kmem_zone_zalloc(xfs_trans_zone,
+ mp->m_super->s_writers.frozen ? KM_NOLOCKDEP : 0);
if (!(flags & XFS_TRANS_NO_WRITECOUNT))
We only should be setting KM_NOLOCKDEP when XFS_TRANS_NO_WRITECOUNT
is set. That's the flag that transactions set when they run in a
fully frozen context to avoid deadlocking with the freeze in
progress, and that's the only case where we should be turning off
lockdep.
And, as I also mentioned, this should be done via a process flag -
PF_MEMALLOC_NOLOCKDEP - so that it is automatically inherited by
all subsequent memory allocations done in this path. That way we
only need this wrapping code in xfs_trans_alloc():
if (flags & XFS_TRANS_NO_WRITECOUNT)
memalloc_nolockdep_save()
.....
if (flags & XFS_TRANS_NO_WRITECOUNT)
memalloc_nolockdep_restore()
and nothing else needs to change.
Cheers,
Dave.
Thanks for the reminder, I will look into that.
Cheers,
Longman
