Message ID | 20180907084151.10085-1-chris@chris-wilson.co.uk (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | [i-g-t] lib: Cancel all outstanding requests at the end of a test | expand |
On 07/09/18 01:41, Chris Wilson wrote: > Quite often on catastrophic failure the test leaves a long queue of > unterminated batches pending execution. Each runs until hangcheck fires > and skips onto the next, leaving us waiting for a very long time at test > exit. > > On older kernels, this gracefully degrades into the existing mechanism. > > Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> LGTM. Just, if I have to be really picky, the comment for quiescent_gpu() says it is installed as an exit handler. Reviewed-by: Antonio Argenziano <antonio.argenziano@intel.com> > --- > lib/drmtest.c | 25 +++++++++++++++++++------ > lib/igt_debugfs.h | 12 ++++++++++++ > 2 files changed, 31 insertions(+), 6 deletions(-) > > diff --git a/lib/drmtest.c b/lib/drmtest.c > index bfa2e0f0a..fee9d33ad 100644 > --- a/lib/drmtest.c > +++ b/lib/drmtest.c > @@ -302,22 +302,35 @@ static int __drm_open_driver_render(int chipset) > static int at_exit_drm_fd = -1; > static int at_exit_drm_render_fd = -1; > > -static void quiescent_gpu_at_exit(int sig) > +static void __cancel_work_at_exit(int fd) > +{ > + igt_terminate_spin_batches(); /* for older kernels */ > + > + igt_drop_caches_set(fd, > + /* cancel everything */ > + DROP_RESET_ACTIVE | DROP_RESET_SEQNO | > + /* cleanup */ > + DROP_ACTIVE | DROP_RETIRE | DROP_IDLE | DROP_FREED); > +} > + > +static void cancel_work_at_exit(int sig) > { > if (at_exit_drm_fd < 0) > return; > > - gem_quiescent_gpu(at_exit_drm_fd); > + __cancel_work_at_exit(at_exit_drm_fd); > + > close(at_exit_drm_fd); > at_exit_drm_fd = -1; > } > > -static void quiescent_gpu_at_exit_render(int sig) > +static void cancel_work_at_exit_render(int sig) > { > if (at_exit_drm_render_fd < 0) > return; > > - gem_quiescent_gpu(at_exit_drm_render_fd); > + __cancel_work_at_exit(at_exit_drm_render_fd); > + > close(at_exit_drm_render_fd); > at_exit_drm_render_fd = -1; > } > @@ -369,7 +382,7 @@ int drm_open_driver(int chipset) > gem_quiescent_gpu(fd); > > at_exit_drm_fd = __drm_open_driver(chipset); > - igt_install_exit_handler(quiescent_gpu_at_exit); > + igt_install_exit_handler(cancel_work_at_exit); > } > } > > @@ -418,7 +431,7 @@ int drm_open_driver_render(int chipset) > at_exit_drm_render_fd = __drm_open_driver(chipset); > if(chipset & DRIVER_INTEL){ > gem_quiescent_gpu(fd); > - igt_install_exit_handler(quiescent_gpu_at_exit_render); > + igt_install_exit_handler(cancel_work_at_exit_render); > } > > return fd; > diff --git a/lib/igt_debugfs.h b/lib/igt_debugfs.h > index ff8612dc6..9f81be0a2 100644 > --- a/lib/igt_debugfs.h > +++ b/lib/igt_debugfs.h > @@ -190,6 +190,18 @@ void igt_require_hpd_storm_ctl(int fd); > * Flush the driver's idle_worker, releasing internal caches and wakerefs. > */ > #define DROP_IDLE 0x40 > +/** > + * DROP_RESET_ACTIVE: > + * > + * Cancel all outstanding requests by forcing a gpu reset > + */ > +#define DROP_RESET_ACTIVE 0x80 > +/** > + * DROP_RESET_SEQNO: > + * > + * Reset the global request seqno counter back to 0 > + */ > +#define DROP_RESET_SEQNO 0x100 > /** > * DROP_ALL: > * >
Quoting Antonio Argenziano (2018-09-07 17:37:45) > > > On 07/09/18 01:41, Chris Wilson wrote: > > Quite often on catastrophic failure the test leaves a long queue of > > unterminated batches pending execution. Each runs until hangcheck fires > > and skips onto the next, leaving us waiting for a very long time at test > > exit. > > > > On older kernels, this gracefully degrades into the existing mechanism. > > > > Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> > > LGTM. Just, if I have to be really picky, the comment for > quiescent_gpu() says it is installed as an exit handler. It's still correct in spirit. But no caller of quiescent_gpu() cares because they need it for idling right there and then. -Chris
diff --git a/lib/drmtest.c b/lib/drmtest.c index bfa2e0f0a..fee9d33ad 100644 --- a/lib/drmtest.c +++ b/lib/drmtest.c @@ -302,22 +302,35 @@ static int __drm_open_driver_render(int chipset) static int at_exit_drm_fd = -1; static int at_exit_drm_render_fd = -1; -static void quiescent_gpu_at_exit(int sig) +static void __cancel_work_at_exit(int fd) +{ + igt_terminate_spin_batches(); /* for older kernels */ + + igt_drop_caches_set(fd, + /* cancel everything */ + DROP_RESET_ACTIVE | DROP_RESET_SEQNO | + /* cleanup */ + DROP_ACTIVE | DROP_RETIRE | DROP_IDLE | DROP_FREED); +} + +static void cancel_work_at_exit(int sig) { if (at_exit_drm_fd < 0) return; - gem_quiescent_gpu(at_exit_drm_fd); + __cancel_work_at_exit(at_exit_drm_fd); + close(at_exit_drm_fd); at_exit_drm_fd = -1; } -static void quiescent_gpu_at_exit_render(int sig) +static void cancel_work_at_exit_render(int sig) { if (at_exit_drm_render_fd < 0) return; - gem_quiescent_gpu(at_exit_drm_render_fd); + __cancel_work_at_exit(at_exit_drm_render_fd); + close(at_exit_drm_render_fd); at_exit_drm_render_fd = -1; } @@ -369,7 +382,7 @@ int drm_open_driver(int chipset) gem_quiescent_gpu(fd); at_exit_drm_fd = __drm_open_driver(chipset); - igt_install_exit_handler(quiescent_gpu_at_exit); + igt_install_exit_handler(cancel_work_at_exit); } } @@ -418,7 +431,7 @@ int drm_open_driver_render(int chipset) at_exit_drm_render_fd = __drm_open_driver(chipset); if(chipset & DRIVER_INTEL){ gem_quiescent_gpu(fd); - igt_install_exit_handler(quiescent_gpu_at_exit_render); + igt_install_exit_handler(cancel_work_at_exit_render); } return fd; diff --git a/lib/igt_debugfs.h b/lib/igt_debugfs.h index ff8612dc6..9f81be0a2 100644 --- a/lib/igt_debugfs.h +++ b/lib/igt_debugfs.h @@ -190,6 +190,18 @@ void igt_require_hpd_storm_ctl(int fd); * Flush the driver's idle_worker, releasing internal caches and wakerefs. */ #define DROP_IDLE 0x40 +/** + * DROP_RESET_ACTIVE: + * + * Cancel all outstanding requests by forcing a gpu reset + */ +#define DROP_RESET_ACTIVE 0x80 +/** + * DROP_RESET_SEQNO: + * + * Reset the global request seqno counter back to 0 + */ +#define DROP_RESET_SEQNO 0x100 /** * DROP_ALL: *
Quite often on catastrophic failure the test leaves a long queue of unterminated batches pending execution. Each runs until hangcheck fires and skips onto the next, leaving us waiting for a very long time at test exit. On older kernels, this gracefully degrades into the existing mechanism. Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> --- lib/drmtest.c | 25 +++++++++++++++++++------ lib/igt_debugfs.h | 12 ++++++++++++ 2 files changed, 31 insertions(+), 6 deletions(-)