Re: configurable block error injection v4

[Christoph Hellwig <hch@lst.de>]

Sashiko had two comments on this:

 - the one about the race of removal vs. addition is spot on, and
   trivially fixed by actually removing the code (unlock/lock cycle).
 - the other about zero-sized commands is valid, but more of an
   enhancement.  And one that if implemented right now actually
   make things worse, as flushed actually show up as empty writes
   with the preflush bit and not as REQ_OP_FLUSH.  So if we'd special
   case zero-sized bios, we'd make a flush hit all write rules,
   which would defeat the point.  We really need to do flushes as
   REQ_OP_FLUSH at the bio level :(

Below is what I plan to fold in, and I'm thinking of the empty bio
issue above can be caught in a comment or the documentation nicely.

diff --git a/block/error-injection.c b/block/error-injection.c
index 7f7f0d3327bc..3bc91f199dc7 100644
--- a/block/error-injection.c
+++ b/block/error-injection.c
@@ -119,11 +119,7 @@ static void error_inject_removeall(struct gendisk *disk)
 	while ((inj = list_first_entry_or_null(&disk->error_injection_list,
 			struct blk_error_inject, entry))) {
 		list_del_rcu(&inj->entry);
-		mutex_unlock(&disk->error_injection_lock);
-
 		kfree_rcu_mightsleep(inj);
-
-		mutex_lock(&disk->error_injection_lock);
 	}
 	static_branch_dec(&blk_error_injection_enabled);
 	mutex_unlock(&disk->error_injection_lock);