[igt-dev] [RFC v5] tests/gem_watchdog: Initial set of tests for GPU watchdog
Carlos Santa
carlos.santa at intel.com
Tue Jun 18 01:35:08 UTC 2019
On Wed, 2019-05-29 at 10:03 +0100, Chris Wilson wrote:
> Quoting Carlos Santa (2019-05-28 22:35:59)
> > This test adds basic set of tests to reset the different
> > GPU engines through the gpu watchdog timer.
> >
> > Credits to Antonio for the original codebase this is based on.
> >
> > v2: remove gem_context_get_param() during set (Antonio)
> > remove clearing of the engines_threshold[] in the default case
> > inside context_set_watchdog(). (Antonio)
> > fix indexing when creating low/high priority contexts
> > get rid of 2 threads idea (Antonio)
> > fix context prio bug due to wrong indexing (Antonio)
> > v3: no need to force clear drm_i915_gem_watchdog_timeout struct
> > when setting the watchdog timeouts (Antonio)
> > -use gem_context_set_priority() instead to use error checking
> > -always use a looper to test all engines (Antonio)
> > -modify test gpu_watchodg_hang_long_batch_single_engine()
> > to use a couple of batches and set the timeout in between them
> > (Antonio)
> > -make use of for_each_engine_class_instance() (Antonio)
> > v4: -tests should be able to handle s/w watchdog timeout (Tvrtko)
> > -remove dead code inside batch_buffer factory (Tvrtko)
> > -set_watchdog() should return ENODEV for fail cases (Tvrtko)
> > -enclose name, priority and reset params inside array (Tvrtko)
> > v5: -Rebased. Tests 5,6 and 8 from the test plan are now added
> >
> > Test Plan:
> >
> > Assumptions:
> > 1. Use fence status to figure out which ctx was reset
> > 2. Use spin batches w/ user space timers to control duration
> > and corking to control ordering
> > 3. Use context priorities to force preemption
> >
> > Tests:
> > 1. ctx1, long_batch -> execute
> > 2. ctx1, set_watchdog -> reset
> > 3. ctx2/ctx1 -> execute, reset
> > 4. ctx1/ctx2 -> reset, execute
> > 5. ctx1_just_below_threshold -> execute
> > 6. ctx_over_the_threshold -> reset
> > 7. set watchdog on some engines -> execute
> > 8. submit long_batch and after half of the
> > expected runtime submit higher prio batch
> > 9. submit low prio batch w/o watchdog then
> > higher prio with watchdog
> >
> > Unsresolved items:
> > 1. The check for sync_fence_status(*fence)
> > returns -1 instead of EIO.
> >
> > Cc: Ursulin Tvrtko <tvrtko.ursulin at intel.com>
> > Cc: Antonio Argenziano <antonio.argenziano at intel.com>
> > Cc: Chris Wilson <chris at chris-wilson.co.uk>
> > Signed-off-by: Carlos Santa <carlos.santa at intel.com>
> > ---
> > tests/Makefile.sources | 3 +
> > tests/i915/gem_watchdog.c | 490
> > ++++++++++++++++++++++++++++++++++++++
> > tests/meson.build | 1 +
> > 3 files changed, 494 insertions(+)
> > create mode 100644 tests/i915/gem_watchdog.c
> >
> > diff --git a/tests/Makefile.sources b/tests/Makefile.sources
> > index 2d5c929e32fc..65b868e7d927 100644
> > --- a/tests/Makefile.sources
> > +++ b/tests/Makefile.sources
> > @@ -446,6 +446,9 @@ gem_userptr_blits_SOURCES =
> > i915/gem_userptr_blits.c
> > TESTS_progs += gem_wait
> > gem_wait_SOURCES = i915/gem_wait.c
> >
> > +TESTS_progs += gem_watchdog
> > +gem_watchdog_SOURCES = i915/gem_watchdog.c
> > +
> > TESTS_progs += gem_workarounds
> > gem_workarounds_SOURCES = i915/gem_workarounds.c
> >
> > diff --git a/tests/i915/gem_watchdog.c b/tests/i915/gem_watchdog.c
> > new file mode 100644
> > index 000000000000..e9cdb886d3d6
> > --- /dev/null
> > +++ b/tests/i915/gem_watchdog.c
> > @@ -0,0 +1,490 @@
> > +/*
> > + * Copyright © 2019 Intel Corporation
> > + *
> > + * Permission is hereby granted, free of charge, to any person
> > obtaining a
> > + * copy of this software and associated documentation files (the
> > "Software"),
> > + * to deal in the Software without restriction, including without
> > limitation
> > + * the rights to use, copy, modify, merge, publish, distribute,
> > sublicense,
> > + * and/or sell copies of the Software, and to permit persons to
> > whom the
> > + * Software is furnished to do so, subject to the following
> > conditions:
> > + *
> > + * The above copyright notice and this permission notice
> > (including the next
> > + * paragraph) shall be included in all copies or substantial
> > portions of the
> > + * Software.
> > + *
> > + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
> > EXPRESS OR
> > + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
> > MERCHANTABILITY,
> > + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO
> > EVENT SHALL
> > + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM,
> > DAMAGES OR OTHER
> > + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
> > ARISING
> > + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
> > OTHER DEALINGS
> > + * IN THE SOFTWARE.
> > + */
> > +#include "igt.h"
> > +#include "igt_sysfs.h"
> > +#include "sw_sync.h"
> > +
> > +#include <pthread.h>
> > +#include <fcntl.h>
> > +
> > +#include <sys/time.h>
> > +#include <sys/ioctl.h>
> > +#include <sys/poll.h>
> > +#include <sys/signal.h>
> > +#include "i915/gem_ring.h"
> > +
> > +#define LOCAL_I915_EXEC_BSD_SHIFT (13)
> > +#define LOCAL_I915_EXEC_BSD_RING1 (1 <<
> > LOCAL_I915_EXEC_BSD_SHIFT)
> > +#define LOCAL_I915_EXEC_BSD_RING2 (2 <<
> > LOCAL_I915_EXEC_BSD_SHIFT)
> > +
> > +#define MAX_PRIO LOCAL_I915_CONTEXT_MAX_USER_PRIORITY
> > +#define DEFAULT_PRIO LOCAL_I915_CONTEXT_DEFAULT_PRIORITY
> > +#define MIN_PRIO LOCAL_I915_CONTEXT_MIN_USER_PRIORITY
> > +#define HIGH 1
> > +#define LOW 0
> > +#define WATCHDOG_OVER_THRESHOLD (2070000)
> > +#define WATCHDOG_BELOW_THRESHOLD (1200000)
> > +#define WATCHDOG_THRESHOLD (1200)
> > +#define MAX_ENGINES 5
> > +#define RENDER_CLASS 0
> > +#define VIDEO_DECODE_CLASS 1
> > +#define VIDEO_ENHANCEMENT_CLASS 2
> > +#define COPY_ENGINE_CLASS 3
> > +#define LOCAL_I915_CONTEXT_PARAM_WATCHDOG 0x10
> > +
> > +#define GET_RESET_STATS_IOCTL DRM_IOWR(DRM_COMMAND_BASE + 0x32,
> > struct local_drm_i915_reset_stats)
> > +struct local_drm_i915_reset_stats {
> > + __u32 ctx_id;
> > + __u32 flags;
> > + __u32 reset_count;
> > + __u32 batch_active;
> > + __u32 batch_pending;
> > + __u32 pad;
> > +};
> > +
> > +const uint64_t timeout_100ms = 100000000LL;
> > +float timedifference_msec(struct timeval t0, struct timeval t1);
> > +
> > +struct drm_i915_gem_watchdog_timeout {
> > + union {
> > + struct {
> > + /*
> > + * Engine class & instance to be configured
> > or queried.
> > + */
> > + __u16 engine_class;
> > + __u16 engine_instance;
> > + };
> > + /* Index based addressing mode */
> > + __u32 index;
> > + };
> > + /* GPU Engine watchdog resets timeout in us */
> > + __u32 timeout_us;
> > +};
> > +
> > +static void clear_error_state(int fd)
> > +{
> > + int dir;
> > +
> > + dir = igt_sysfs_open(fd);
> > +
> > + if (dir < 0)
> > + return;
> > +
> > + /* Any write to the error state clears it */
> > + igt_sysfs_set(dir, "error", "");
> > + close(dir);
> > +}
> > +
> > +static int context_set_watchdog(int fd, unsigned engine_id,
> > + const char *engine_name,
> > + unsigned ctx_id, unsigned
> > threshold)
> > +{
> > + struct drm_i915_gem_watchdog_timeout
> > engines_threshold[MAX_ENGINES];
> > + struct drm_i915_gem_context_param arg = {
> > + .param = LOCAL_I915_CONTEXT_PARAM_WATCHDOG,
> > + .ctx_id = ctx_id,
> > + .size = sizeof(engines_threshold),
>
> Won't the kernel complain about the invalid values?
No, but I'll check again since I had some extra debugging added...
>
> > + .value = (uint64_t)&engines_threshold
> > + };
> > +
> > + memset(&engines_threshold, 0, sizeof(engines_threshold));
>
> You set all the unspecified values to be rcs0 with 0 timeout.
I think I forced everything to 0 to avoid garbage data...
>
> > +
> > + switch (engine_id & I915_EXEC_RING_MASK) {
> > + case I915_EXEC_RENDER:
> > + engines_threshold[0].timeout_us = threshold;
>
> I presume this is a u64, so might as well use _ns for consistency.
The uApi is u32 but I'll change it to u64.
>
> > + engines_threshold[0].engine_class = RENDER_CLASS;
> > + engines_threshold[0].engine_instance = 0;
>
> Use struct i915_engine_class_instance.
Ok.
>
> > + break;
> > + case I915_EXEC_BLT:
> > + if (__gem_context_get_param(fd, &arg) == -ENODEV)
> > + return -ENODEV;
>
> And otherwise setting random values into engine_thresholds.
Hmm, Ursulin still wanted to add support for CPU watchdog so this is
how we check whether there's support for watchdog on BLT, it returns
ENODEV for this case...
>
> > + else
> > + engines_threshold[3].timeout_us =
> > threshold;
> > + engines_threshold[3].engine_class =
> > COPY_ENGINE_CLASS;
> > + engines_threshold[3].engine_instance = 0;
> > + break;
> > + case I915_EXEC_BSD:
> > + engines_threshold[1].timeout_us = threshold;
> > + engines_threshold[1].engine_class =
> > VIDEO_DECODE_CLASS;
> > + engines_threshold[1].engine_instance = 0;
> > + break;
> > + case I915_EXEC_VEBOX:
> > + engines_threshold[2].timeout_us = threshold;
> > + engines_threshold[2].engine_class =
> > VIDEO_ENHANCEMENT_CLASS;
> > + engines_threshold[2].engine_instance = 0;
> > + break;
> > + default:
> > + break;
> > + }
>
> Ok, so you only set one engine and the rest to zero. That seems like
> a
> bad example in how to use the uAPI.
The alternative would be to set them all at once?? but then what about
the media decode engine that only cares about the media engines,
wouldn't this case only be setting the threshold to one specific
engine? Also, I am looping through all engines so eventually they all
get populated.
>
> > + gem_context_set_param(fd, &arg);
> > +
> > + return 0;
> > +}
> > +
> > +static void batch_buffer_factory(uint32_t fd, uint32_t ctx_id,
> > unsigned exec_id,
> > + uint32_t target, uint32_t offset,
> > + uint32_t *handle, uint64_t
> > timeout,
> > + int *fence, int fence_index)
> > +{
> > + struct drm_i915_gem_exec_object2 obj[2];
> > + struct drm_i915_gem_relocation_entry reloc;
> > + struct drm_i915_gem_execbuffer2 execbuf;
> > + igt_spin_t *spin = NULL;
> > + const uint32_t bbe = MI_BATCH_BUFFER_END;
> > + int i = 0;
> > +
> > + gem_quiescent_gpu(fd);
> > +
> > + memset(&execbuf, 0, sizeof(execbuf));
> > + memset(&obj, 0, sizeof(obj));
> > + memset(&reloc, 0, sizeof(reloc));
> > +
> > + execbuf.buffers_ptr = to_user_pointer(obj);
> > +
> > + execbuf.buffer_count = 2;
> > + execbuf.flags = exec_id | I915_EXEC_FENCE_OUT ;
> > +
> > + obj[0].handle = target;
> > + obj[1].handle = gem_create(fd, 4096);
> > +
> > + obj[1].relocation_count = 1;
> > + obj[1].relocs_ptr = to_user_pointer(&reloc);
> > +
> > + reloc.target_handle = obj[0].handle;
> > + reloc.read_domains = I915_GEM_DOMAIN_COMMAND;
> > + reloc.write_domain = I915_GEM_DOMAIN_COMMAND;
> > + reloc.delta = offset * sizeof(uint32_t);
> > +
> > + reloc.offset = i * sizeof(uint32_t);
> > + gem_write(fd, obj[1].handle, 0, &bbe, sizeof(bbe));
> > +
> > + __sync_synchronize();
>
> What?
Not sure but I'll remove it, this is from the older reference code this
is based on.
>
> > +
> > + gem_sync(fd, obj[1].handle);
> > + execbuf.rsvd1 = ctx_id;
> > + execbuf.rsvd2 = -1;
> > +
> > + spin = igt_spin_new(fd, .dependency = obj[0].handle);
> > + igt_spin_set_timeout(spin, timeout);
> > + igt_assert(gem_bo_busy(fd, obj[0].handle));
> > +
> > + gem_execbuf_wr(fd, &execbuf);
> > + igt_spin_free(fd, spin);
>
> This still doesn't do what you think it does. Hint, this never runs
> to
> timeout.
>
> > + fence[fence_index] = execbuf.rsvd2 >> 32;
>
> You can just ask the spinner to return a fence.
> spin = igt_spin_new(fd, .engine =engine, .flags =
> IGT_SPIN_FENCE_OUT);
> then fence = spin->out_fence.
>
Ok, I did the abovebut please review the new implementation when I send
it (hopefully tomorrow) because I did just that but I still must be
missing something as I now get a -9 when I check the status of the
fence.
> > + gem_close(fd, obj[1].handle);
> > + gem_quiescent_gpu(fd);
> > +}
> > +
> > +static void inject_hang(uint32_t fd, unsigned engine_id, uint32_t
> > ctx_id,
> > + unsigned flags)
> > +{
> > + igt_hang_t hang;
> > + hang = igt_hang_ctx(fd, ctx_id, engine_id, flags);
> > + gem_sync(fd, hang.spin->handle);
> > +}
> > +
> > +/* Test#1: create some work and let it run on all engines */
> > +static void long_batch_test1(int fd, int prio1, int prio2, int
> > reset_ctx1,
> > + int reset_ctx2, unsigned threshold)
> > +{
> > + unsigned engine_id = 0;
> > + unsigned nengine = 0;
> > + struct intel_execution_engine2 *e_;
> > + uint32_t ctx[8];
> > + uint32_t scratch[8];
> > + uint32_t ctx2[8];
> > + uint32_t scratch2[8];
>
> A struct might spell it out a little more concisely.
Agreed.
>
> And 8! That's cutting it a little fine. MAX_ENGINES is 64 and we are
> already dreading the day we exceed that.
Ok.
>
> > + int *fence, i = 0;
> > + const uint64_t batch_timeout_ms = timeout_100ms * 4;
> > +
> > + fence = (int *)malloc(sizeof(int)*8);
>
> Please be kind to readers. The cast is implicit, and spaces are
> required
> around '*'.
Sorry about that.
>
> > + igt_assert(fence);
> > +
> > + __for_each_physical_engine(fd, e_) {
> > + engine_id = e_->flags;
> > +
> > + scratch2[nengine] = gem_create(fd, 4096);
> > + ctx2[nengine] = gem_context_create(fd);
> > + gem_context_set_priority(fd, ctx2[nengine], prio2);
> > + batch_buffer_factory(fd, ctx2[nengine], engine_id,
> > + scratch2[nengine], 0, NULL,
> > + batch_timeout_ms, fence, nengine);
> > +
> > + scratch[nengine] = gem_create(fd, 4096);
> > + ctx[nengine] = gem_context_create(fd);
> > + gem_context_set_priority(fd, ctx[nengine], prio1);
> > + batch_buffer_factory(fd, ctx[nengine], engine_id,
> > + scratch[nengine], 0, NULL,
> > + batch_timeout_ms, fence, nengine);
> > + igt_info("Test #1: ctx1/ctx2 --> execute on
> > %s\n",e_->name);
> > + nengine++;
>
> So that doesn't test anything at all.
Hmm, Ursulin has provided a test plan (please see the commit message
above) some time back and that's part of that, we were trying to cover
different cases and that's one of them, it simply create some work and
let it run on all the different engines....
>
> > + }
> > +
> > + for (i = 0; i < nengine; i++) {
> > + close(fence[i]);
> > + gem_context_destroy(fd, ctx[i]);
> > + gem_context_destroy(fd, ctx2[i]);
> > + gem_close(fd, scratch[i]);
> > + gem_close(fd, scratch2[i]);
> > + }
> > +}
> > +
> > +float timedifference_msec(struct timeval t0, struct timeval t1)
> > +{
> > + return (t1.tv_sec - t0.tv_sec) * 1000.0f + (t1.tv_usec -
> > t0.tv_usec) / 1000.0f;
> > +}
> > +
> > +/* Test#2: submit a long batch and after half of the expected
> > runtime
> > + submit a higher priority batch and then try to cancel the
> > execution*/
> > +static void long_batch_test2(int fd, int prio1, int prio2, int
> > reset_ctx1,
> > + int reset_ctx2, unsigned threshold)
> > +{
> > + unsigned engine_id = 0;
> > + unsigned nengine = 0;
> > + struct intel_execution_engine2 *e_;
> > + uint32_t ctx[8];
> > + uint32_t scratch[8];
> > + uint32_t ctx2[8];
> > + uint32_t scratch2[8];
> > + int *fence, i = 0;
> > + unsigned flags = HANG_ALLOW_CAPTURE;
> > + const uint64_t batch_timeout_ms = timeout_100ms * 4;
> > + struct timeval t0;
> > + struct timeval t1;
> > + float elapsed;
> > +
> > + fence = (int *)malloc(sizeof(int)*8);
> > + igt_assert(fence);
> > +
> > + nengine = 0;
> > +
> > + __for_each_physical_engine(fd, e_) {
> > + engine_id = e_->flags;
> > +
> > + scratch2[nengine] = gem_create(fd, 4096);
> > + ctx2[nengine] = gem_context_create(fd);
> > + gem_context_set_priority(fd, ctx2[nengine], prio2);
> > + batch_buffer_factory(fd, ctx2[nengine], engine_id,
> > + scratch2[nengine], 0, NULL,
> > + batch_timeout_ms, fence, nengine);
> > +
> > + gettimeofday(&t0, 0);
> > + while(elapsed < 200) {
> > + gettimeofday(&t1, 0);
> > + elapsed += timedifference_msec(t0, t1);
> > + }
>
> See usleep() or nanosleep()
Ok.
>
> > + scratch[nengine] = gem_create(fd, 4096);
> > + ctx[nengine] = gem_context_create(fd);
> > + gem_context_set_priority(fd, ctx[nengine], prio1);
> > + batch_buffer_factory(fd, ctx[nengine], engine_id,
> > + scratch[nengine], 0, NULL,
> > + batch_timeout_ms, fence, nengine);
> > +
> > + if (context_set_watchdog(fd, engine_id, e_->name,
> > + ctx[nengine],
> > + threshold) == -ENODEV) {
> > + igt_info("No support for gpu h/w watchdog
> > on %s\n",
> > + e_->name);
> > + goto skip_case;
> > + }
> > + clear_error_state(fd);
> > + inject_hang(fd, engine_id, ctx[nengine], flags);
> > + /* Now check the engine was reset successfully */
> > + //igt_assert_eq(sync_fence_status(*fence), -EIO);
>
> Yeah, that fence has nothing to do with anything.
>
> > + igt_info("Test #2 ctx1/ctx2 --> set watchdog and"
> > + " cancel ctx2 at half expected run
> > with higher"
> > + " priority engine: %s, fence
> > status: 0x%x \n",
> > + e_->name,
> > sync_fence_status(*fence));
>
> You've set a watchdog, but don't validate that it had any effect.
>
> May I suggest, you using a timing batch. Write the initial timestamp
> to
> scratch, and write a second timestamp and recurse back to writing the
> second timestamp. After that batch is forcibly reset, you can compare
> the two timestamps to see how long it executed for (without having to
> dig into the context image to extract the ppHWSP runtimes). The
> contract
> given by the watchdog uAPI would seem to be that the batch is hung
> after
> that runtime exceeds the timeout threshold (though how strictly we
> fail
> for exceeding the timeout???) and most definitely not be reset before
> we
> reach the threshold (give or take a few microseconds required for the
> recursion and context switching/execution that will be charged to the
> watchdog and not recorded by the batch).
> -Chris
I synced with Antonio and we think we have what you're describing
above, however the calculated delta seems to be way to big for the work
being submitted. Please review when I send the new version.
Regards,
Carlos
More information about the igt-dev
mailing list