Bug 74100

Can you please repaste the error message after applying:

commit 48ad03ca0c5f078b8d12a64323fd93b3858041af
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date:   Fri Jan 31 16:56:01 2014 +0000

    lib: Capture errno on entry
    
    When printing the errno, it is important that we capture the user errno
    before we make any library calls - as they may alter the value.
    
    References: https://bugs.freedesktop.org/show_bug.cgi?id=74007
    Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

please retest with latest igt and drm-intel-nightly

Created attachment 93589 [details]
new dmesg

It still causes system hang on latest igt and -nightly kernel.
output:
IGT-Version: 1.5-g0269d1d (x86_64) (Linux: 3.13.0_drm-intel-nightly_8f5284_20140207+ x86_64)
Subtest close-pending-fork-render: SUCCESS
Test requirement not met in function gem_require_ring, file ./../lib/drmtest.h:312:
Last errno: 11, Resource temporarily unavailable
Test requirement: (!(gem_has_vebox(fd)))
Subtest reset-stats-vebox: SKIP
Subtest reset-stats-ctx-vebox: SKIP
Subtest ban-vebox: SKIP
Subtest ban-ctx-vebox: SKIP
Subtest reset-count-vebox: SKIP
Subtest reset-count-ctx-vebox: SKIP
Subtest unrelated-ctx-vebox: SKIP
Subtest close-pending-vebox: SKIP
Subtest close-pending-ctx-vebox: SKIP
Subtest close-pending-fork-vebox: SKIP

Ok, I've cleaned up the superflous SKIP behaviour and some other stuff in the testcase. And I can repro the issue here. Looking at netconsole the system seems to go down with a failed gpu reset:

[   63.823960] [drm] stuck on render ring
[   63.824130] [drm] GPU crash dump saved to /sys/class/drm/card0/error
[   63.824847] [drm] GPU HANG [e77fffff]
[   63.824893] [drm] GPU hangs can indicate a bug anywhere in the entire gfx stack, including userspace.
[   63.824979] [drm] Please file a _new_ bug report on bugs.freedesktop.org against DRI -> DRM/Intel
[   63.825057] [drm] drm/i915 developers can then reassign to the right component if it's not a kernel issue.
[   63.825138] [drm] The gpu crash dump is required to analyze gpu hangs, so please always attach it.
[   63.825978] [drm] Simulated gpu hang, resetting stop_rings
[   65.805203] [drm] Enabling RC6 states: RC6 on, RC6p off, RC6pp off
[   71.812051] [drm] stuck on blitter ring
[   72.316616] [drm:i915_reset] *ERROR* Failed to reset chip: -110

That shouldn't really happen ...

The two render ring hangs are a proper hangs injected by test but the
third one in blitter is unexpected. The recovery from this hang fails
as attempting to reset the GPU hard hangs the whole machine.

The SNB I tested with hangs immediately on setting the reset bit.

As the hang recovery works properly with other tests, I will decrease importance.

With this patch here

http://patchwork.freedesktop.org/patch/21673/

I can convert the hard system hang in a failed gpu hang. Can you please test this patch and confirm that it improves the situation? gem_reset_stat should still fail, but the machine will survive at least.

(In reply to comment #6)
> With this patch here
> 
> http://patchwork.freedesktop.org/patch/21673/
> 
> I can convert the hard system hang in a failed gpu hang. Can you please test
> this patch and confirm that it improves the situation? gem_reset_stat should
> still fail, but the machine will survive at least.

Patch fail.

On Mon, Mar 10, 2014 at 7:03 AM,  <bugzilla-daemon@freedesktop.org> wrote:
> Patch fail.


Please provide more details to the nature of the failure, this
information is next to useless.

patching file drivers/gpu/drm/i915/intel_uncore.c
Hunk #1 FAILED at 989.
1 out of 1 hunk FAILED -- saving rejects to file drivers/gpu/drm/i915/intel_uncore.c.rej

drivers/gpu/drm/i915/intel_uncore.c :
        if (fw_engine)
                dev_priv->uncore.funcs.force_wake_get(dev_priv, fw_engine);

        if (IS_GEN6(dev) || IS_GEN7(dev))
                dev_priv->uncore.fifo_count =
                        __raw_i915_read32(dev_priv, GTFIFOCTL) &
                        GT_FIFO_FREE_ENTRIES_MASK;

        spin_unlock_irqrestore(&dev_priv->uncore.lock, irqflags);
        return ret;
}

int intel_gpu_reset(struct drm_device *dev)

patch:
@@ -989,9 +989,11 @@  static int gen6_do_reset(struct drm_device *dev)
 	if (fw_engine)
 		dev_priv->uncore.funcs.force_wake_get(dev_priv, fw_engine);
 
-	if (IS_GEN6(dev) || IS_GEN7(dev))
-		WARN_ON((__raw_i915_read32(dev_priv, GTFIFOCTL) &
-			 GT_FIFO_FREE_ENTRIES_MASK) != 0);
+	if (IS_GEN6(dev) || IS_GEN7(dev)) {
+		if (WARN_ON((__raw_i915_read32(dev_priv, GTFIFOCTL) &
+			     GT_FIFO_FREE_ENTRIES_MASK) != 0))
+		    ret = -EIO;
+	}
 
 	dev_priv->uncore.fifo_count = 0;

My apologies for not updating this bug, I've already ripped out the offending code. The system hang is still there though.

Reping to Mika to just look into disabling this specific subtest on snb.

Created attachment 95821 [details] [review]
hang preventer hack

I think I'm onto something here. Please test the attached patch and check whether the hangs disappear. Note that there might be some additional test failures with this (since it rips out a bit of code), the important part is whether the snb still hangs or not.

Created attachment 95822 [details] [review]
fix up semaphore hangcheck code

Now also a real patch. Please test both this patch and the earlier quick hack, thanks.

Created attachment 95913 [details]
dmesg

(In reply to comment #12)
> Created attachment 95822 [details] [review] [review]
> fix up semaphore hangcheck code
> 
> Now also a real patch. Please test both this patch and the earlier quick
> hack, thanks.

Test this patch, It still occurs.

Mika, can you please create a patch to just skip the offending tests on snb?

Please test the below patch:

http://patchwork.freedesktop.org/patch/25173/

The hang goes away on latest -nightly and -fixes kernel. Close it.
output:
IGT-Version: 1.6-gc1404e0 (x86_64) (Linux: 3.15.0-rc2_drm-intel-fixes_7f1950_20140429+ x86_64)
Subtest close-pending-fork-render: SUCCESS
Test requirement not met in function gem_require_ring, file ioctl_wrappers.c:820:
Last errno: 0, Success
Test requirement: (!(gem_has_vebox(fd)))

Verified.Fixed.

As I failed to trigger the bug without Daniel's patch I proceed to bisect the significant commit. It turned out to be:

commit 5582e8c3c49150c0e7398688b5ed167d6c3d44fd
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date:   Wed Apr 9 09:19:41 2014 +0100

    drm/i915: Preserve ring buffers objects across resume

Closing verified+fixed.

Closing verified+fixed.