I noticed this problem. This patch alone is not enough to fix it. I will send my patch
for this soon.
Thanks,
Song
quoted hunk
On Dec 4, 2016, at 12:44 AM, Zhengyuan Liu [off-list ref] wrote:
Currently, we choice write-through mode as the default journal mode.
If there is data-only stripes, we'd rewrite it and add it into raid5d
release list. However, raid5d thread wouldn't put those stripes into
cache(full/partial) list but inactive list instead since the journal
mode is write-through. More futher, later read request would get data
from raid disk directly instead of cache stripe, that's not we want too.
Signed-off-by: Zhengyuan Liu <redacted>
---
drivers/md/raid5-cache.c | 6 +++++-
1 file changed, 5 insertions(+), 1 deletion(-)
diff --git a/drivers/md/raid5-cache.c b/drivers/md/raid5-cache.c
index 07bce0e..0473b33 100644
--- a/drivers/md/raid5-cache.c
+++ b/drivers/md/raid5-cache.c
@@ -2606,7 +2606,6 @@ int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev)
INIT_WORK(&log->deferred_io_work, r5l_submit_io_async);
- log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
INIT_LIST_HEAD(&log->stripe_in_journal_list);
spin_lock_init(&log->stripe_in_journal_lock);
atomic_set(&log->stripe_in_journal_count, 0);
@@ -2614,6 +2613,11 @@ int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev)
if (r5l_load_log(log))
goto error;
+ if (log->last_checkpoint == log->next_checkpoint)
+ log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
+ else
+ log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_BACK;
+
rcu_assign_pointer(conf->log, log);
set_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
return 0;
--
2.7.4