| Message ID | 20180228110829.12126-1-tvrtko.ursulin@linux.intel.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
Quoting Tvrtko Ursulin (2018-02-28 11:08:29) > From: Tvrtko Ursulin <tvrtko.ursulin@intel.com> > > Some tests (the ones which call igt_setup_runtime_pm and > igt_pm_enable_audio_runtime_pm) change default system configuration and > never restore it. > > The configured runtime suspend is aggressive and may influence behaviour > of subsequent tests, so it is better to restore to previous values on test > exit. > > This way system behaviour, with regards to a random sequence of executed > tests, will be more consistent from one run to another. Otoh, if behaviour changes in subsequent tests, we likely have a lack of testing :( -Chris
On 28/02/2018 11:12, Chris Wilson wrote: > Quoting Tvrtko Ursulin (2018-02-28 11:08:29) >> From: Tvrtko Ursulin <tvrtko.ursulin@intel.com> >> >> Some tests (the ones which call igt_setup_runtime_pm and >> igt_pm_enable_audio_runtime_pm) change default system configuration and >> never restore it. >> >> The configured runtime suspend is aggressive and may influence behaviour >> of subsequent tests, so it is better to restore to previous values on test >> exit. >> >> This way system behaviour, with regards to a random sequence of executed >> tests, will be more consistent from one run to another. > > Otoh, if behaviour changes in subsequent tests, we likely have a lack of > testing :( Yeah, and I am not saying it does - haven't spotted anything like that, just that it leaves the auto-suspend with zero delay afterwards, compared to otherwise default 10s. Maybe it is good for coverage, even with the downside of randomness considering shard runs, or maybe it needs to be more explicit. Regards, Tvrtko
Quoting Tvrtko Ursulin (2018-02-28 11:38:01) > > On 28/02/2018 11:12, Chris Wilson wrote: > > Quoting Tvrtko Ursulin (2018-02-28 11:08:29) > >> From: Tvrtko Ursulin <tvrtko.ursulin@intel.com> > >> > >> Some tests (the ones which call igt_setup_runtime_pm and > >> igt_pm_enable_audio_runtime_pm) change default system configuration and > >> never restore it. > >> > >> The configured runtime suspend is aggressive and may influence behaviour > >> of subsequent tests, so it is better to restore to previous values on test > >> exit. > >> > >> This way system behaviour, with regards to a random sequence of executed > >> tests, will be more consistent from one run to another. > > > > Otoh, if behaviour changes in subsequent tests, we likely have a lack of > > testing :( > > Yeah, and I am not saying it does - haven't spotted anything like that, > just that it leaves the auto-suspend with zero delay afterwards, > compared to otherwise default 10s. > > Maybe it is good for coverage, even with the downside of randomness > considering shard runs, or maybe it needs to be more explicit. I actually thought (or at least picked up the idea) we enabled 0 autosuspend delay throughout igt. And that it's only because most of time we have a display connected that prevents rpm madness. -Chris
Quoting Tvrtko Ursulin (2018-02-28 11:08:29) > From: Tvrtko Ursulin <tvrtko.ursulin@intel.com> > > Some tests (the ones which call igt_setup_runtime_pm and > igt_pm_enable_audio_runtime_pm) change default system configuration and > never restore it. > > The configured runtime suspend is aggressive and may influence behaviour > of subsequent tests, so it is better to restore to previous values on test > exit. > > This way system behaviour, with regards to a random sequence of executed > tests, will be more consistent from one run to another. > > Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> > Cc: Imre Deak <imre.deak@intel.com> Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> -Chris
On 28/02/2018 12:27, Chris Wilson wrote: > Quoting Tvrtko Ursulin (2018-02-28 11:38:01) >> >> On 28/02/2018 11:12, Chris Wilson wrote: >>> Quoting Tvrtko Ursulin (2018-02-28 11:08:29) >>>> From: Tvrtko Ursulin <tvrtko.ursulin@intel.com> >>>> >>>> Some tests (the ones which call igt_setup_runtime_pm and >>>> igt_pm_enable_audio_runtime_pm) change default system configuration and >>>> never restore it. >>>> >>>> The configured runtime suspend is aggressive and may influence behaviour >>>> of subsequent tests, so it is better to restore to previous values on test >>>> exit. >>>> >>>> This way system behaviour, with regards to a random sequence of executed >>>> tests, will be more consistent from one run to another. >>> >>> Otoh, if behaviour changes in subsequent tests, we likely have a lack of >>> testing :( >> >> Yeah, and I am not saying it does - haven't spotted anything like that, >> just that it leaves the auto-suspend with zero delay afterwards, >> compared to otherwise default 10s. >> >> Maybe it is good for coverage, even with the downside of randomness >> considering shard runs, or maybe it needs to be more explicit. > > I actually thought (or at least picked up the idea) we enabled 0 > autosuspend delay throughout igt. And that it's only because most of time > we have a display connected that prevents rpm madness. Yes fair point, the effect of this patch is only visible on esoteric headless setups like mine. :) Anyway, unless I am missing something, I think it is conceptually correct to restore, as long as the zero autosuspend delay is not a global IGT setup but a helper called by a few tests. Regards, Tvrtko
diff --git a/lib/igt_pm.c b/lib/igt_pm.c index 5bf5b2e23cdc..0c8b5e8e9257 100644 --- a/lib/igt_pm.c +++ b/lib/igt_pm.c @@ -63,6 +63,46 @@ enum { /* Remember to fix this if adding longer strings */ #define MAX_POLICY_STRLEN strlen(MAX_PERFORMANCE_STR) +static char __igt_pm_audio_runtime_power_save[64]; +static char __igt_pm_audio_runtime_control[64]; + +static void __igt_pm_audio_runtime_exit_handler(int sig) +{ + int fd; + + igt_debug("Restoring audio power management to '%s' and '%s'\n", + __igt_pm_audio_runtime_power_save, + __igt_pm_audio_runtime_control); + + fd = open("/sys/module/snd_hda_intel/parameters/power_save", O_WRONLY); + if (fd < 0) + return; + if (write(fd, __igt_pm_audio_runtime_power_save, + strlen(__igt_pm_audio_runtime_power_save)) != + strlen(__igt_pm_audio_runtime_power_save)) + igt_warn("Failed to restore audio power_save to '%s'\n", + __igt_pm_audio_runtime_power_save); + close(fd); + + fd = open("/sys/bus/pci/devices/0000:00:03.0/power/control", O_WRONLY); + if (fd < 0) + return; + if (write(fd, __igt_pm_audio_runtime_control, + strlen(__igt_pm_audio_runtime_control)) != + strlen(__igt_pm_audio_runtime_control)) + igt_warn("Failed to restore audio control to '%s'\n", + __igt_pm_audio_runtime_control); + close(fd); +} + +static void strchomp(char *str) +{ + int len = strlen(str); + + if (str[len - 1] == '\n') + str[len - 1] = 0; +} + /** * igt_pm_enable_audio_runtime_pm: * @@ -78,16 +118,32 @@ void igt_pm_enable_audio_runtime_pm(void) { int fd; - fd = open("/sys/module/snd_hda_intel/parameters/power_save", O_WRONLY); + /* Check if already enabled. */ + if (__igt_pm_audio_runtime_power_save[0]) + return; + + fd = open("/sys/module/snd_hda_intel/parameters/power_save", O_RDWR); if (fd >= 0) { + igt_assert(read(fd, __igt_pm_audio_runtime_power_save, + sizeof(__igt_pm_audio_runtime_power_save)) > 0); + strchomp(__igt_pm_audio_runtime_power_save); igt_assert_eq(write(fd, "1\n", 2), 2); + igt_install_exit_handler(__igt_pm_audio_runtime_exit_handler); close(fd); } - fd = open("/sys/bus/pci/devices/0000:00:03.0/power/control", O_WRONLY); + fd = open("/sys/bus/pci/devices/0000:00:03.0/power/control", O_RDWR); if (fd >= 0) { + igt_assert(read(fd, __igt_pm_audio_runtime_control, + sizeof(__igt_pm_audio_runtime_control)) > 0); + strchomp(__igt_pm_audio_runtime_control); igt_assert_eq(write(fd, "auto\n", 5), 5); close(fd); } + + igt_debug("Saved audio power management as '%s' and '%s'\n", + __igt_pm_audio_runtime_power_save, + __igt_pm_audio_runtime_control); + /* Give some time for it to react. */ sleep(1); } @@ -238,6 +294,38 @@ void igt_pm_restore_sata_link_power_management(int8_t *pm_data) /* We just leak this on exit ... */ int pm_status_fd = -1; +static char __igt_pm_runtime_autosuspend[64]; +static char __igt_pm_runtime_control[64]; + +static void __igt_pm_runtime_exit_handler(int sig) +{ + int fd; + + igt_debug("Restoring runtime management to '%s' and '%s'\n", + __igt_pm_runtime_autosuspend, + __igt_pm_runtime_control); + + fd = open(POWER_DIR "/autosuspend_delay_ms", O_WRONLY); + if (fd < 0) + return; + if (write(fd, __igt_pm_runtime_autosuspend, + strlen(__igt_pm_runtime_autosuspend)) != + strlen(__igt_pm_runtime_autosuspend)) + igt_warn("Failed to restore runtime pm autosuspend delay to '%s'\n", + __igt_pm_runtime_autosuspend); + close(fd); + + fd = open(POWER_DIR "/control", O_WRONLY); + if (fd < 0) + return; + if (write(fd, __igt_pm_runtime_control, + strlen(__igt_pm_runtime_control)) != + strlen(__igt_pm_runtime_control)) + igt_warn("Failed to restore runtime pm control to '%s'\n", + __igt_pm_runtime_control); + close(fd); +} + /** * igt_setup_runtime_pm: * @@ -261,10 +349,16 @@ bool igt_setup_runtime_pm(void) /* Our implementation uses autosuspend. Try to set it to 0ms so the test * suite goes faster and we have a higher probability of triggering race * conditions. */ - fd = open(POWER_DIR "/autosuspend_delay_ms", O_WRONLY); + fd = open(POWER_DIR "/autosuspend_delay_ms", O_RDWR); igt_assert_f(fd >= 0, "Can't open " POWER_DIR "/autosuspend_delay_ms\n"); + /* Save previous values and install exit handler to restore them. */ + igt_assert(read(fd, __igt_pm_runtime_autosuspend, + sizeof(__igt_pm_runtime_autosuspend)) > 0); + strchomp(__igt_pm_runtime_autosuspend); + igt_install_exit_handler(__igt_pm_runtime_exit_handler); + /* If we fail to write to the file, it means this system doesn't support * runtime PM. */ size = write(fd, "0\n", 2); @@ -278,6 +372,13 @@ bool igt_setup_runtime_pm(void) fd = open(POWER_DIR "/control", O_RDWR); igt_assert_f(fd >= 0, "Can't open " POWER_DIR "/control\n"); + igt_assert(read(fd, __igt_pm_runtime_control, + sizeof(__igt_pm_runtime_control)) > 0); + strchomp(__igt_pm_runtime_control); + + igt_debug("Saved runtime power management as '%s' and '%s'\n", + __igt_pm_runtime_autosuspend, __igt_pm_runtime_control); + size = write(fd, "auto\n", 5); igt_assert(size == 5);