Re: Unhandled IRQs on AMD E-450

[Josh Boyer]

On Fri, Apr 27, 2012 at 7:46 AM, Jeroen Van den Keybus
<jeroen.vandenkeybus@xxxxxxxxx> wrote:
>> So why aren't you guys producing a proper patch for people to test?
>
> I have been working on a patch for 3.4-rc2, in which I also
> implemented another mechanism to detect a stuck interrupt (using a
> level mechanism: an unhandled IRQ increases the level by, say, 10, and
> a handled one reduces the level by 1. It is considered stuck when the
> level reaches e.g. 100). It also implements proper dmesg reporting. I
> also looked for x86'isms as suggested by Alan, but I think there's no
> problem.
>
> However, while testing, I noticed that the stuck interrupt detection
> fired also on MSI interrupts in the 3.4 kernel under heavy load,
> suggesting that these would also be suffering from being unhandled.
>
> The patch also includes Clemens' patch to put the disk IRQ on MSI,
> which makes the E450 board actually usable.
>
> I haven't had time to debug the new issue (false stuck irq detection
> in 3.4 - not too harmful since the interrupt is reenabled a second
> later, but wrong nevertheless). If any of you is willing to help and
> have a look into the code, I have attached it below.
>
> (There's also a small patch that accidentally got in there to
> reimplement cpu_possible_map, as it was needed by the fglrx driver I'm
> using. Please disregard. Or use.)
>
> Another problem is the arming of the timer (mod_timer), which may
> cause an incorrect poll interval when a IRQ gets stuck while anotther
> one is being polled (either because it's also stuck and polling or
> because forced polling is on for that particular one).

We've had a few bugs reported with this ASM108x chip in Fedora.  I
took your original patch you submitted quite a while ago, and tweaked
it a little.  We've been carrying it in Fedora and it makes things
somewhat usable, but it's far from perfect.

Essentially, it adds a PCI quirk to detect the buggy bridge and only
changes the irqpolling mechanism if that bridge is detected.  That
means that people without the buggy hardware aren't impacted by this
at all, which was the first problem we hit.  Apparently there are
some pieces of hardware that generate a small number of spurious IRQs
"normally" and lowering the threshold to such a small value caused
those machines to kick into polling mode when they really didn't need
to.

The other two issues we've run into with this patch are:

1) While the quirk helps shield people without the buggy bridge, it
doesn't help the case where people have the bridge, but they have no
devices actually behind it.  That means such setups hit the polling
mode when they don't really need to as described above.

2) People, rightfully, complain that it makes inter-activity on their
desktop pretty laggy.  The mouse pointer jumps around a lot and key
strokes are often missed.  For a server class machine, I doubt it
would matter much but Fedora is essentially a desktop distro so that
tends to be a high priority.

I've attached the minor rework of the original patch below for
reference.

josh

----

It seems that some motherboard designs using the ASM1083 PCI/PCIe
bridge (PCI device ID 1b21:1080, Rev. 01) suffer from stuck IRQ lines
on the PCI bus (causing the kernel to emit 'IRQxx: nobody cared' and
disable the IRQ). The following patch is an attempt to mitigate the
serious impact of permanently disabling an IRQ in that case and
actually make PCI devices better usable on this platform.

It seems that the bridge fails to issue a IRQ deassertion message on
the PCIe bus, when the relevant driver causes the interrupting PCI
device to deassert its IRQ line. To solve this issue, it was tried to
re-issue an IRQ on a PCI device being able to do so (e1000 in this
case), but we suspect that the attempt to re-assert/deassert may have
occurred too soon after the initial IRQ for the ASM1083. Anyway, it
didn't work but if, after some delay, a new IRQ occurred, the related
IRQ deassertion message eventually did clear the IOAPIC IRQ. It would
be useful to re-enable the IRQ here.

Therefore the patch below to poll_spurious_irqs() in spurious.c is
proposed, It does the following:

1. lets the kernel decide that an IRQ is unhandled after only 10
positives (instead of 100,000);
2. briefly (a few seconds or so, currently 1 s) switches to polling
IRQ at a higher rate than usual (100..1,000Hz instead of 10Hz,
currently 100Hz), but not too high to avoid excessive CPU load. Any
device drivers 'see' their interrupts handled with a higher latency
than usual, but they will still operate properly;
3. afterwards, simply reenable the IRQ.

If proper operation of the PCIe legacy IRQ line emulation is restored
after 3, the system operates again at normal performance. If the IRQ
is still stuck after this procedure, the sequence repeats.

If a genuinely stuck IRQ is used with this solution, the system would
simply sustain short bursts of 10 unhandled IRQs per second, and use
polling mode indefinitely at a moderate 100Hz rate. It seemed a good
alternative to the default irqpoll behaviour to me, which is why I
left it in poll_spurious_irqs() (instead of creating a new kernel
option). Additionally, if any device happens to share an IRQ with a
faulty one, that device is no longer banned forever.

Debugging output is still present and may be removed. Bad IRQ
reporting is also commented out now.

I have now tried it for about 2 months and I can conclude the following:

1. The patch works and, judging from my Firewire card interrupt on
IRQ16, which repeats every 64 secs, I can confirm that the IRQ usually
gets reset when a new IRQ arrives (polling mode runs for 64 seconds
every time).
2. When testing a SiL-3114 SATA PCI card behind the ASM1083, I could
keep this running at fairly high speeds (50..70MB/s) for an hour or
so, but eventually the SiL driver crashed. In such conditions the PCI
system had to deal with a few hundred IRQs per second / polling mode
kicking in every 5..10 seconds).

I would like to thank Clemens Ladisch for his invaluable help in
finding a solution (and providing a patch to avoid my SATA going down
every time during debugging).


Signed-off-by: Jeroen Van den Keybus <jeroen.vandenkeybus@xxxxxxxxx>

Make it less chatty.  Only kick it in if we detect an ASM1083 PCI bridge.

Josh Boyer <jwboyer@xxxxxxxxxx>
======

--- linux-2.6.orig/kernel/irq/spurious.c
+++ linux-2.6/kernel/irq/spurious.c
@@ -18,6 +18,8 @@

 static int irqfixup __read_mostly;

+int irq_poll_and_retry = 0;
+
 #define POLL_SPURIOUS_IRQ_INTERVAL (HZ/10)
 static void poll_spurious_irqs(unsigned long dummy);
 static DEFINE_TIMER(poll_spurious_irq_timer, poll_spurious_irqs, 0, 0);
@@ -141,12 +143,13 @@ out:
 static void poll_spurious_irqs(unsigned long dummy)
 {
 	struct irq_desc *desc;
-	int i;
+	int i, poll_again;

 	if (atomic_inc_return(&irq_poll_active) != 1)
 		goto out;
 	irq_poll_cpu = smp_processor_id();

+	poll_again = 0; /* Will stay false as long as no polling candidate is found */
 	for_each_irq_desc(i, desc) {
 		unsigned int state;

@@ -159,14 +162,33 @@ static void poll_spurious_irqs(unsigned
 		if (!(state & IRQS_SPURIOUS_DISABLED))
 			continue;

-		local_irq_disable();
-		try_one_irq(i, desc, true);
-		local_irq_enable();
+		/* We end up here with a disabled spurious interrupt.
+		   desc->irqs_unhandled now tracks the number of times
+		   the interrupt has been polled */
+		if (irq_poll_and_retry) {
+			if (desc->irqs_unhandled < 100) { /* 1 second delay with poll
frequency 100 Hz */
+				local_irq_disable();
+				try_one_irq(i, desc, true);
+				local_irq_enable();
+				desc->irqs_unhandled++;
+				poll_again = 1;
+			} else {
+				irq_enable(desc); /* Reenable the interrupt line */
+				desc->depth--;
+				desc->istate &= (~IRQS_SPURIOUS_DISABLED);
+				desc->irqs_unhandled = 0;
+			}
+		} else {
+			local_irq_disable();
+			try_one_irq(i, desc, true);
+			local_irq_enable();
+		}
 	}
+	if (poll_again)
+		mod_timer(&poll_spurious_irq_timer,
+			  jiffies + POLL_SPURIOUS_IRQ_INTERVAL);
 out:
 	atomic_dec(&irq_poll_active);
-	mod_timer(&poll_spurious_irq_timer,
-		  jiffies + POLL_SPURIOUS_IRQ_INTERVAL);
 }

 static inline int bad_action_ret(irqreturn_t action_ret)
@@ -177,11 +199,19 @@ static inline int bad_action_ret(irqretu
 }

 /*
- * If 99,900 of the previous 100,000 interrupts have not been handled
+ * If 9 of the previous 10 interrupts have not been handled
  * then assume that the IRQ is stuck in some manner. Drop a diagnostic
  * and try to turn the IRQ off.
  *
- * (The other 100-of-100,000 interrupts may have been a correctly
+ * Although this may cause early deactivation of a sporadically
+ * malfunctioning IRQ line, the poll system will:
+ * a) Poll it for 100 cycles at a 100 Hz rate
+ * b) Reenable it afterwards
+ *
+ * In worst case, with current settings,  this will cause short bursts
+ * of 10 interrupts every second.
+ *
+ * (The other single interrupt may have been a correctly
  *  functioning device sharing an IRQ with the failing one)
  */
 static void
@@ -269,6 +299,8 @@ try_misrouted_irq(unsigned int irq, stru
 void note_interrupt(unsigned int irq, struct irq_desc *desc,
 		    irqreturn_t action_ret)
 {
+	int unhandled_thresh = 999000;
+
 	if (desc->istate & IRQS_POLL_INPROGRESS)
 		return;

@@ -302,19 +334,31 @@ void note_interrupt(unsigned int irq, st
 	}

 	desc->irq_count++;
-	if (likely(desc->irq_count < 100000))
-		return;
+	if (!irq_poll_and_retry)
+		if (likely(desc->irq_count < 100000))
+			return;
+	else
+		if (likely(desc->irq_count < 10))
+			return;

 	desc->irq_count = 0;
-	if (unlikely(desc->irqs_unhandled > 99900)) {
+	if (irq_poll_and_retry)
+		unhandled_thresh = 9;
+
+	if (unlikely(desc->irqs_unhandled >= unhandled_thresh)) {
 		/*
-		 * The interrupt is stuck
+		 * The interrupt might be stuck
 		 */
-		__report_bad_irq(irq, desc, action_ret);
+		if (!irq_poll_and_retry) {
+			__report_bad_irq(irq, desc, action_ret);
+			printk(KERN_EMERG "Disabling IRQ %d\n", irq);
+		} else {
+			printk(KERN_INFO "IRQ %d might be stuck.  Polling\n",
+				irq);
+		}
 		/*
 		 * Now kill the IRQ
 		 */
-		printk(KERN_EMERG "Disabling IRQ #%d\n", irq);
 		desc->istate |= IRQS_SPURIOUS_DISABLED;
 		desc->depth++;
 		irq_disable(desc);
--- linux-2.6.orig/drivers/pci/quirks.c
+++ linux-2.6/drivers/pci/quirks.c
@@ -1677,6 +1677,22 @@ DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_IN
 DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL,	0x260a, quirk_intel_pcie_pm);
 DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL,	0x260b, quirk_intel_pcie_pm);

+/* ASM108x transparent PCI bridges apparently have broken IRQ deassert
+ * handling.  This causes interrupts to get "stuck" and eventually disabled.
+ * However, the interrupts are often shared and disabling them is fairly bad.
+ * It's been somewhat successful to switch to polling mode and retry after
+ * a bit, so let's do that.
+ */
+extern int irq_poll_and_retry;
+static void quirk_asm108x_poll_interrupts(struct pci_dev *dev)
+{
+	dev_info(&dev->dev, "Buggy bridge found [%04x:%04x]\n",
+		dev->vendor, dev->device);
+	dev_info(&dev->dev, "Stuck interrupts will be polled and retried\n");
+	irq_poll_and_retry = 1;
+}
+DECLARE_PCI_FIXUP_EARLY(PCI_VENDOR_ID_ASMEDIA,	0x1080,	quirk_asm108x_poll_interrupts);
+
 #ifdef CONFIG_X86_IO_APIC
 /*
  * Boot interrupts on some chipsets cannot be turned off. For these chipsets,
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/