[igt-dev] [RFC] tests/gem_watchdog: Initial set of tests for GPU watchdog

[Antonio Argenziano]

On 16/04/19 13:21, Chris Wilson wrote:
> Quoting Antonio Argenziano (2019-04-16 21:16:35)
>>
>>
>> On 15/04/19 19:02, Carlos Santa wrote:
>>> On Mon, 2019-04-15 at 14:50 -0700, Antonio Argenziano wrote:
>>>>
>>>> On 15/04/19 11:22, Carlos Santa wrote:
>>>>> +{
>>>>> +    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;
>>>>> +    int fence;
>>>>> +
>>>>> +    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 | LOCAL_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();
>>>>> +
>>>>> +    if (handle) {
>>>>> +        *handle = obj[1].handle;
>>>>> +        return;
>>>>> +    }
>>>>> +    gem_sync(fd, obj[1].handle);
>>>>> +    execbuf.rsvd1 = ctx_id;
>>>>> +    execbuf.rsvd2 = -1;
>>>>> +
>>>>> +    spin = igt_spin_batch_new(fd, .dependency = obj[0].handle);
>>>>> +    igt_spin_batch_set_timeout(spin, timeout);
>>>>> +    igt_assert(gem_bo_busy(fd, obj[0].handle));
>>>>> +
>>>>> +    gem_execbuf_wr(fd, &execbuf);
>>>>> +    igt_spin_batch_free(fd, spin);
>>>>> +
>>>>> +    fence = execbuf.rsvd2 >> 32;
>>>>> +
>>>>> +#if 0
>>>>> +    igt_info("fence:%d, fence status : %d EIO: %d
>>>>> ctx_id:%d\n",fence, sync_fence_status(fence),-EIO, ctx_id);
>>>>> +#endif
>>>>> +
>>>
>>> Any ideas why would the fence wouldn't signal correctly (i.e., EIO)
>>> after resetting?  I need help with this...
>>
>> Not sure, it looks correct on this end, just one small comment would be
>> to use I915_EXEC_FENCE_OUT instead of the local flag.
> 
> Hint: it doesn't run to the timeout.

Oh this is the canary batch, for some reason I was convinced we were 
printing this for the hanging batch as well...

Antonio

> -Chris
>