Are you saying your code works on top of 4.11-rc2, but not on top of my
for-linus?
I was actually on Linus 4.11-rc1 before I rebased on top of your
for-linus.
That seems odd. Looking at the oops, you are crashing with
!tags in __blk_mq_tag_idle. The below should work around it, but I'm
puzzled why this is new.
I got it just once (out of a single run :)), but maybe it is
possible that its racy and not really new.
But another example where this can happen:
blk_mq_realloc_hw_ctxs explicitly checks on hctx->tags != NULL
but right after calls blk_mq_exit_hctx() which goes in the
same route, won't this happen there too? Or is it assumed that
hctx->state does not have BLK_MQ_S_TAG_ACTIVE on here?
Is it related to the other path you fixed in this patch:
commit 0067d4b020ea07a58540acb2c5fcd3364bf326e0
Author: Sagi Grimberg <sagi@xxxxxxxxxxx>
Date: Mon Mar 13 16:10:11 2017 +0200
blk-mq: Fix tagset reinit in the presence of cpu hot-unplug
Since that's also handling hctx->tags == NULL.
The above patch prevented a NULL deref earlier when the
tags were reinitialized, now we are all setup and we
happen to remove an old namespace.
diff --git a/block/blk-mq-tag.c b/block/blk-mq-tag.c
index 9d97bfc4d465..1283f74bfdfb 100644
--- a/block/blk-mq-tag.c
+++ b/block/blk-mq-tag.c
@@ -54,9 +54,11 @@ void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
return;
- atomic_dec(&tags->active_queues);
+ if (tags) {
+ atomic_dec(&tags->active_queues);
- blk_mq_tag_wakeup_all(tags, false);
+ blk_mq_tag_wakeup_all(tags, false);
+ }
}
/*
I'll see if I can test it out later this week. thanks.
