Re: [PATCH] drm/i915,agp/intel: Do not clear stolen entries

From: Mario Kleiner
Date: Sat Jan 29 2011 - 23:13:51 EST


On Jan 30, 2011, at 1:28 AM, Hugh Dickins wrote:

On Fri, Jan 28, 2011 at 6:59 PM, Mario Kleiner
<mario.kleiner@xxxxxxxxxxxxxxxx> wrote:
On Jan 28, 2011, at 11:00 PM, Hugh Dickins wrote:

Sorry, this is now abount vblank or scanout rather than stolen entries.

On Mon, 24 Jan 2011, Chris Wilson wrote:

On Sun, 23 Jan 2011 23:40:41 -0800 (PST), Hugh Dickins <hughd@xxxxxxxxxx>
wrote:

On this laptop I'm typing from (GM965 with KMS), I've had no trouble
getting X up; but when typing in one of the xterms, typed characters
often stop echoing, until I shift to a different window, whereupon
they appear. This condition cleared (for a while) by switching to
VESA fb console and back; no such problem observed on that console.

Does that sound familiar? I have no evidence whatever that i915 is
to blame here. Several times I tried bisecting last week, but each
attempt ended up in a nonsensical place, because the effect does not
occur to order. So I'd sometimes mark a bisection point as good when
I guess it must actually have been bad. Perhaps it's a matter of
timing or an uninitialized variable. But while I'm here, worth asking
if that behaviour sounds like anything you might be responsible for?

Sounds suspiciously like the batch buffer is not being dispatched and
flushed to the scanout. A very similar bug was recently fixed for
xf86-video-intel 2.14.0 which was causing deferred output.

I made a more patient bisection during the week, on x86_64 which
seemed more consistent than i386, and this time it converged sensibly:
to commit 0af7e4dff50454905092d468e91c1ef92e10e6b4
drm/i915: Add support for precise vblank timestamping (v2)

Which kindly notes in its commit message:
This code has been only tested on a HP-Mini Netbook with
Atom processor and Intel 945GME gpu. The codepath for
(IS_G4X(dev) || IS_GEN5(dev) || IS_GEN6(dev)) gpu's
has not been tested so far due to lack of hardware.
so not surprising that it doesn't work on GM965.

I'm now running with this silly revert:

--- a/drivers/gpu/drm/i915/i915_drv.c 2011-01-18 22:04:29.000000000
-0800
+++ b/drivers/gpu/drm/i915/i915_drv.c 2011-01-24 19:35:51.000000000
-0800
@@ -674,8 +674,8 @@ static struct drm_driver driver = {
.device_is_agp = i915_driver_device_is_agp,
.enable_vblank = i915_enable_vblank,
.disable_vblank = i915_disable_vblank,
- .get_vblank_timestamp = i915_get_vblank_timestamp,
- .get_scanout_position = i915_get_crtc_scanoutpos,
+ .get_vblank_timestamp = NULL /* i915_get_vblank_timestamp */,
+ .get_scanout_position = NULL /* i915_get_crtc_scanoutpos */,
.irq_preinstall = i915_driver_irq_preinstall,
.irq_postinstall = i915_driver_irq_postinstall,
.irq_uninstall = i915_driver_irq_uninstall,

which makes 2.6.38-rc usable; though I do believe that I've seen
the same issue (unflushed text) occur a couple of times since, much
too rare to bisect or get upset by, but indicative of some remaining bug.


Hi,

just skimmed through the archives of this thread. Do i understand correctly
that the problem that gets fixed by your revert is that

<snip>

when typing in one of the xterms, typed characters
often stop echoing, until I shift to a different window, whereupon
they appear. This condition cleared (for a while) by switching to
VESA fb console and back; no such problem observed on that console.

</snip>

Yes, that's the problem that's fixed by the little revert patch I
posted last time.
Sorry, this thread started out with other problems, then I asked Chris
if this might also be an i915 issue.


Is this with desktop composition enabled?

Not that I'm aware of. The see-through business. I'm just using four
xterms in fvwm2 on openSUSE11.2 with own kernel. If desktop
composition might be enabled by the X startup script, expecting me to
use gnome rather than fvwm2, then I suppose it might be enabled; but
it's not something I've chosen to turn on. What should I check to
answer you for sure, if it matters?


Hm, probably not. You could check if compiz is running or if debug messages from "drm_wait_vblank" show up in syslog while you move and resize windows etc., or if there is visible tearing.

What xserver version is this?

Do things like glxgears in a
window work correctly? If desktop composition is off?

Yes, glxgears appears to work correctly: I type "glxgears" at the
xterm shell prompt, those letters and carriage return are not echoed
back to me, but the glxgears window appears with the gears turning
correctly, then I close that window, type more and again my typing is
not echoed.


For a softer fix to the problem you can revert your revert and disable use
of those functions by the drm core via:

echo 0 > /sys/modules/drm/parameters/timestamp_precision_usec

Thanks for the info. ("module" rather than "modules".)


But can you run it with echo 7 > /sys/modules/drm/parameters/debug

and show me bits of the syslog output when the problem happens? Especially
output from the functions "drm_calc_vbltimestamp_from_scanoutpos" and
"drm_handle_vblank" and maybe for "vblank_disable_fn",
"drm_update_vblank_count", and "drm_vblank_get".

Wow, millions of lines of output (partly because I couldn't see the
typo that had prevented me from turning it off after a few seconds).
I rebuilt the kernel with the DRM_DEBUG at the head of drm_ioctl()
edited out: that generates so many messages (cmd=0x400c645f
mostly, but some cmd=0x6458) that the logging cannot keep up, and
hardly gets a chance to print anything else.

But even with that edited out, nothing from any of the functions that
you suggest: only, and perhaps this is the problem?,
[drm:i915_driver_irq_handler}, pipe a underrun
about 64 times per second.


That doesn't sound healthy. Chris Wilson will know, but afaik it means that the display pipeline doesn't get enough memory bandwidth to fetch pixels from the framebuffer fast/often enough for display, so there could be some misconfiguration of the memory controller or the display watermarks?

At the moment i can't see how the timestamping functions in the i915 kms driver from my patch could cause your problem, except that disabling them apparently changes something to the better.

Could it be some side-effect, e.g., that if a display pipe suffers a fifo underrun and my timestamping code reads some mmio registers related to that pipe after the underrun, the chip somehow doesn't like that and chokes on it a bit?

If you haven't uncommented the kms timestamping functions with your patch and they are enabled, e.g., via echo 1000 > /sys/module/drm/ parameters/timestamp_precision_usec and logging of drm debug output is enabled, you should see some output from those functions in the syslog or dmesg output. There isn't any path through them that wouldn't produce a DRM_DEBUG, DRM_ERROR, DRM_DEBUG_DRIVER or DRM_DEBUG_KMS output at some point. An additional function to check would be "i915_get_vblank_timestamp".

If they exit with an error, you should see debug output and the drm core would fall back to the implementation you get when you apply your revert, or disable them via the echo 0 ...

If they succeed you should see output from drm_calc_vbltimestamp_from_scanoutpos with sensible timestamps and status info.

If they "succeed" but return garbage results due to some lingering bug in the untested codepath for GM965 and there would be an additional bug in the filtering of timestamps, you should see some debug output from drm_handle_vblank (..."Redundant vblirq ignored"...). OpenGL apps like glxgears could show some slight hiccups or jerks in their redraws. A desktop with composition enabled could also show some small jerks and delays. The only possible kind of bug i would expect in the untested codepath would be calculated timestamps that could be off by a few milliseconds, nothing more serious than that.

I'm not sure why a non-composited desktop would be affected at all, unless it makes use of the vblank counter and vblank counter incrementing is broken.

You could add a

DRM_DEBUG("crtc %d: vblcount %d\n", crtc, (int) vblcount);

right before the atomic_inc(&dev->_vblank_count[crtc]);
in drm_handle_vblank() at the bottom of drivers/gpu/drm/drm_irq.c
to print out the vblank counter at each vblank irq to verify that it increments nicely as it should.

You could also turn the abs(diff_ns) in that function into a abs64 (diff_ns). The use of abs() is wrong, as it truncates the s64 diff_ns numbers to 32 bits. I'll prepare a patch to fix that. But this mistake should only cause trouble if the timestamps are already grossly wrong due to a kms driver bug, and even then it should only cause a slight, barely noticeable hiccup for OpenGL rendering due to at most one skipped frame every two seconds worst-case when the 32 bit value wraps around.

You could also try ftrace to check the execution path of the code and its timing:

cd /sys/kernel/debug/tracing/
echo drm_handle_vblank > set_graph_function

or maybe one level deeper: echo i915_driver_irq_handler > set_graph_function if this function is traceable.

echo function_graph > current_tracer
cat trace_pipe > /tmp/trace.out &
sleep 5 ; echo 1 > tracing_enabled ; sleep 10 ; echo 0 > tracing_enabled
more /tmp/trace.out


I just tried setting the debug to 7 for a few seconds on 2.6.37, where
I see no problem: I appear to get the "pipe a underrun" messages with
that too; and the drm_ioctl messages, but much much fewer of them.
Though I've been veering between i386 and x86_64 in these tests, so
keep that in mind if what I'm saying makes no sense: the huge number
of drm_ioctls was with 2.6.36-rc2 (plus some of Chris's fixes) on
i386; the 64 underruns per second was with 2.6.36-rc2 (plus some of

I assume you meant 2.6.38-rc2?

thanks,
-mario


Chris's fixes, minus the drm_ioctl DRM_DEBUG) on x86_64; the underruns
and reasonable number of drm_ioctls was with 2.6.37 on x86_64.

On this laptop I'm working with
# CONFIG_NO_HZ is not set
# CONFIG_HIGH_RES_TIMERS is not set
CONFIG_HZ_250=y
CONFIG_HZ=250
if those have any relevance.

Thanks,
Hugh


Those functions (are supposed to) compute exact timestamps of start of
scanout after each vblank. If they get disabled via the "echo 0 ..." then a
do_gettimeofday() is called for a crude approximation of start of scanout.
The computed timestamps are returned to clients which want them
(oml_sync_control extension). I doubt that many apps use that extension or
its timestamps already, especially not desktop compositors etc., so i
wouldn't expect trouble from such wrong timestamps.

However, the timestamps are also used in drm_handle_vblank() in
drivers/gpu/drm/drm_irq.c at each vblank irq to detect and filter out
redundant vblank irq's to avoid miscounting of vblanks (observed on some
Radeon's). If the kms driver would deliver a grossly wrong timestamp and
something would be wrong in the implementation of that filtering, it could
happen that the vblank counter doesn't get incremented -> delivery of a
vblank event to the x-server gets delayed -> a swapbuffer operation on a
composited desktop gets delayed -> content of a redirected window updates
only with a delay.

The relevant check which could prevent vblank counter increments and delay
vblank event delivery to the x-server in drm_handle_vblank() would be:

if (abs(diff_ns) > DRM_REDUNDANT_VBLIRQ_THRESH_NS) {

The condition should be satisfied if everything works correctly, but also if
timestamps would be grossly wrong, thereby leading to a larger than 1 msec
positive or negative diff_ns. s64 diff_ns is a signed 64 bit integer. Could
abs(diff_ns) somehow miscompute for large 64 bit numbers?

All guesswork, the syslog output should tell us more if the timestamping is
really involved in the problem.

thanks,
-mario

*********************************************************************
Mario Kleiner
Max Planck Institute for Biological Cybernetics
Spemannstr. 38
72076 Tuebingen
Germany

e-mail: mario.kleiner@xxxxxxxxxxxxxxxx
office: +49 (0)7071/601-1623
fax: +49 (0)7071/601-616
www: http://www.kyb.tuebingen.mpg.de/~kleinerm
*********************************************************************
"For a successful technology, reality must take precedence
over public relations, for Nature cannot be fooled."
(Richard Feynman)



*********************************************************************
Mario Kleiner
Max Planck Institute for Biological Cybernetics
Spemannstr. 38
72076 Tuebingen
Germany

e-mail: mario.kleiner@xxxxxxxxxxxxxxxx
office: +49 (0)7071/601-1623
fax: +49 (0)7071/601-616
www: http://www.kyb.tuebingen.mpg.de/~kleinerm
*********************************************************************
"For a successful technology, reality must take precedence
over public relations, for Nature cannot be fooled."
(Richard Feynman)

--
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/