Re: x86/mce: suspicious RCU usage in 4.13.4

[Borislav Petkov]

On Tue, Oct 10, 2017 at 03:00:09PM -0400, Jeremy Cline wrote:
> Hello,
> 
> A Fedora user has reported an issue about suspicious RCU usage in
> dev-mcelog. It looks like perhaps the notifier call chain is not
> acquiring the mce_chrdev_read_mutex? The traceback is
> 
> [36915.633804] =============================
> [36915.633805] WARNING: suspicious RCU usage
> [36915.633808] 4.13.4-301.fc27.x86_64+debug #1 Not tainted
> [36915.633809] -----------------------------
> [36915.633811] arch/x86/kernel/cpu/mcheck/dev-mcelog.c:60 suspicious
> mce_log_get_idx_check() usage!
> [36915.633812]
>                other info that might help us debug this:
> 
> [36915.633813]
>                rcu_scheduler_active = 2, debug_locks = 1
> [36915.633815] 3 locks held by kworker/1:2/14637:
> [36915.633816]  #0:  ("events"){.+.+.+}, at: [<ffffffffaa0d2ac0>]
> process_one_work+0x1d0/0x6a0
> [36915.633827]  #1:  ((&mce_work)){+.+...}, at: [<ffffffffaa0d2ac0>]
> process_one_work+0x1d0/0x6a0
> [36915.633833]  #2:  ((x86_mce_decoder_chain).rwsem){++++..}, at:
> [<ffffffffaa0dc92f>] blocking_notifier_call_chain+0x2f/0x70
> [36915.633840]
>                stack backtrace:
> [36915.633843] CPU: 1 PID: 14637 Comm: kworker/1:2 Not tainted
> 4.13.4-301.fc27.x86_64+debug #1
> [36915.633844] Hardware name: Gigabyte Technology Co., Ltd.
> Z87M-D3H/Z87M-D3H, BIOS F11 08/12/2014
> [36915.633847] Workqueue: events mce_gen_pool_process
> [36915.633849] Call Trace:
> [36915.633854]  dump_stack+0x8e/0xd6
> [36915.633858]  lockdep_rcu_suspicious+0xc5/0x100
> [36915.633862]  dev_mce_log+0xf6/0x1e0
> [36915.633865]  notifier_call_chain+0x39/0x90
> [36915.633869]  blocking_notifier_call_chain+0x49/0x70
> [36915.633873]  mce_gen_pool_process+0x41/0x70

Right, so dev_mce_log() is called in process context now and thus can be
greatly simplified by removing all those memory barriers and cmpxchg()
fun which was for atomic context back then. And simply grab the mutex
instead.

IOW, something like this totally untested hunk. Tony?

---
diff --git a/arch/x86/kernel/cpu/mcheck/dev-mcelog.c b/arch/x86/kernel/cpu/mcheck/dev-mcelog.c
index 10cec43aac38..1dacebb6a23b 100644
--- a/arch/x86/kernel/cpu/mcheck/dev-mcelog.c
+++ b/arch/x86/kernel/cpu/mcheck/dev-mcelog.c
@@ -53,9 +53,10 @@ static int dev_mce_log(struct notifier_block *nb, unsigned long val,
 				void *data)
 {
 	struct mce *mce = (struct mce *)data;
-	unsigned int next, entry;
+	unsigned int entry;
+
+	mutex_lock(&mce_chrdev_read_mutex);
 
-	wmb();
 	for (;;) {
 		entry = mce_log_get_idx_check(mcelog.next);
 		for (;;) {
@@ -66,10 +67,10 @@ static int dev_mce_log(struct notifier_block *nb, unsigned long val,
 			 * interesting ones:
 			 */
 			if (entry >= MCE_LOG_LEN) {
-				set_bit(MCE_OVERFLOW,
-					(unsigned long *)&mcelog.flags);
+				set_bit(MCE_OVERFLOW, (unsigned long *)&mcelog.flags);
 				return NOTIFY_OK;
 			}
+
 			/* Old left over entry. Skip: */
 			if (mcelog.entry[entry].finished) {
 				entry++;
@@ -77,15 +78,13 @@ static int dev_mce_log(struct notifier_block *nb, unsigned long val,
 			}
 			break;
 		}
-		smp_rmb();
-		next = entry + 1;
-		if (cmpxchg(&mcelog.next, entry, next) == entry)
-			break;
+		mcelog.next = entry + 1;
 	}
+
 	memcpy(mcelog.entry + entry, mce, sizeof(struct mce));
-	wmb();
 	mcelog.entry[entry].finished = 1;
-	wmb();
+
+	mutex_unlock(&mce_chrdev_read_mutex);
 
 	/* wake processes polling /dev/mcelog */
 	wake_up_interruptible(&mce_chrdev_wait);

-- 
Regards/Gruss,
    Boris.

Good mailing practices for 400: avoid top-posting and trim the reply.