Created attachment 128525 [details]
basic-wait-all dmesg log

Software:
Fedora 24
kernel revision: 5415795 (drm-tip: 2016y-12m-16d-21h-26m-45s UTC integration manifest)
kernel config: https://intel-gfx-ci.01.org/CI/CI_DRM_1968/kernel.config.bz2

Hardware:
Intel i7-3770

Looks like the timeout value isn't being computed properly (possibly just that too much time is unaccounted?). Can you apply:

diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index 1b2f18d0ca2b..dfe751b350ad 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -3019,6 +3019,7 @@ i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
 {
        struct drm_i915_gem_wait *args = data;
        struct drm_i915_gem_object *obj;
+       u64 prev_timeout = args->timeout_ns;
        ktime_t start;
        long ret;
 
@@ -3042,6 +3043,12 @@ i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
                        args->timeout_ns = 0;
        }
 
+       pr_info("%s: timeout %d.%09d -> %d.%09d\n", __func__,
+               prev_timeout < 0 ? -1 : ns_to_timespec(prev_timeout).tv_sec,
+               prev_timeout < 0 ? 0 : ns_to_timespec(prev_timeout).tv_nsec,
+               args->timeout_ns < 0 ? -1 : ns_to_timespec(args->timeout_ns).tv_sec,
+               args->timeout_ns < 0 ? 0 : ns_to_timespec(args->timeout_ns).tv_nsec);
+
        i915_gem_object_put(obj);
        return ret;
 }

and get the output for a failure?

Created attachment 128544 [details]
Dmesg log with applied patch from comment #3

The same config option triggers failures in a similar way in:

igt/gem_sync/basic-each
igt/gem_sync/basic-store-each

Created attachment 128545 [details]
igt/gem_sync/basic-each log

Created attachment 128546 [details]
igt/gem_sync/basic-each - dmesg (truncated beginning)

Created attachment 128547 [details]
igt/gem_sync/basic-store-each - log

Created attachment 128548 [details]
igt/gem_sync/basic-store-each - dmesg (truncated beginning)

[  154.074065] [drm:drm_ioctl] pid=2323, dev=0xe200, auth=1, I915_GEM_WAIT
[  154.074748] i915_gem_wait_ioctl: timeout 0.500000000 -> 0.499993351
[  154.075390] [drm:drm_ioctl] ret = -512
...
[  155.072528] [drm:drm_ioctl] pid=2323, dev=0xe200, auth=1, I915_GEM_WAIT
[  155.073234] i915_gem_wait_ioctl: timeout 0.488539596 -> 0.488535373
[  155.073916] [drm:drm_ioctl] ret = -512
[  155.074605] [drm:drm_ioctl] pid=2323, dev=0xe200, auth=1, I915_GEM_WAIT
[  155.075307] i915_gem_wait_ioctl: timeout 0.488535373 -> 0.488532529

Walltime of 1s, but actual wait time of ~12ms.

Each ioctl is taking ~1.4ms (inside drmIoctl) with a further interval of ~0.7ms between ioctls. And we account for 3us.

Created attachment 128606 [details]
ftrace call graph log for i915_gem_wait_ioctl

The call log capturing i915_gem_wait_ioctl invokation and overhead.
The time of ioctl itself (20ns) is dwarfed by a slowdown somewhere in console_unlock() related to printk() calls (650ns)

cat /proc/consoles?

Also trying with CONFIG_DRM_FBDEV_EMULATION unset.

# cat /proc/consoles
tty0                 -WU (EC p  )    4:1

Upon further digging, the slow stack *seems* like this (may be wrong, the tracing tool is less useful with huge numbers of calls):

console_unlock()
 call_console_drivers.isra.0.constprop.15()
  vt_console_print()
   lf()
    scrup()
     fbcon_scroll()
      fbcon_cursor()
       fbcon_redraw.isra.17()

Inside of this, the delay is split up among many shorter calls, most time occupied by fbcon_putcs -> bit_putcs -> cfb_imageblit.

That's us! Woohoo \o/

What does cat /sys/kernel/debug/dri/0/i915_gem_framebuffer report? Mainly interested in knowing if you have a tiled + fenced fbcon.

Here it is:

$ sudo cat /sys/kernel/debug/dri/0/i915_gem_framebuffer

fbcon size: 1024 x 768, depth 24, 32 bpp, modifier 0x0, refcount 1, obj ffff880221116200:  p        3072KiB 41 00  uncached (pinned x 1) (display) (ggtt offset: 00080000, size: 00300000, type: 0)

Try:

diff --git a/lib/igt_core.c b/lib/igt_core.c
index a18a728c..403b9423 100644
--- a/lib/igt_core.c
+++ b/lib/igt_core.c
@@ -67,6 +67,7 @@
 
 #include "igt_core.h"
 #include "igt_aux.h"
+#include "igt_sysfs.h"
 
 #ifdef HAVE_LIBGEN_H
 #include <libgen.h>   /* for basename() on Solaris */
@@ -512,6 +513,7 @@ static void common_exit_handler(int sig)
 {
        if (!igt_only_list_subtests()) {
                low_mem_killer_disable(false);
+               kick_fbcon(true);
        }
 
        /* When not killed by a signal check that igt_exit() has been properly
@@ -731,6 +733,7 @@ out:
                exit(ret == -1 ? 0 : IGT_EXIT_INVALID);
 
        if (!list_subtests) {
+               kick_fbcon(false);
                kmsg(KERN_INFO "[IGT] %s: executing\n", command_str);
                print_version();

After adding the kick_fbcon() calls, the problem goes away, even without rebooting.

Should this be turned into a patch for igt or should I try to identify the underlying issue causing the slowness?

I just realized the patch was merged. Should this bug be considered fixed now?

I kept the bug around in case we wanted to dig deeper into what the slowdown actually was. If it's below the level of care, lets close and move on.

I haven't seen the underlying slowness causing problems anywhere else yet. Perhaps we can close this and reopen if it reappears.

Closing. Please reopen if needed. Thanks.