[igt-dev] [PATCH] tests/xe/xe_noexec_ping_pong: Add a test to expose unnecessary rebinds

[Thomas Hellström]

Thanks, for reviewing, Matthew, some answers below:

BTW there is a patch 7/7 "drm/xe/vm: Defer vm rebind until the next exec 
if nothing to execute" that fixes
the ping-ponging. On the list now.

On 3/15/23 22:15, Matthew Brost wrote:
> On Wed, Mar 15, 2023 at 03:15:24PM +0100, Thomas Hellström wrote:
>> This test creates compute vms, binds a couple of bos and an engine each,
>> thus redying it for execution. However, VRAM memory is over-
>> committed and while there is still nothing to execute, an eviction
>> will trigger the VM's rebind worker to rebind the evicted bo, which
>> will in turn trigger another eviction and so on.
>>
>> Since we don't have eviction stats yet we need to watch "top" for
>> the rebind kworkers using a lot of CPU while the test idles.
>>
>> The correct driver behaviour should be not to rebind anything unless
>> there is work queued on one of the VM's compute engines.
>>
>> Signed-off-by: Thomas Hellström <thomas.hellstrom at linux.intel.com>
>> ---
>>   tests/meson.build              |   1 +
>>   tests/xe/xe_noexec_ping_pong.c | 105 +++++++++++++++++++++++++++++++++
>>   2 files changed, 106 insertions(+)
>>   create mode 100644 tests/xe/xe_noexec_ping_pong.c
>>
>> diff --git a/tests/meson.build b/tests/meson.build
>> index 0863fac0..8bf15f90 100644
>> --- a/tests/meson.build
>> +++ b/tests/meson.build
>> @@ -258,6 +258,7 @@ xe_progs = [
>>   	'xe_live_ktest',
>>   	'xe_mmap',
>>   	'xe_mmio',
>> +	'xe_noexec_ping_pong',
>>   	'xe_pm',
>>   	'xe_prime_self_import',
>>   	'xe_query',
>> diff --git a/tests/xe/xe_noexec_ping_pong.c b/tests/xe/xe_noexec_ping_pong.c
>> new file mode 100644
>> index 00000000..39f9e961
>> --- /dev/null
>> +++ b/tests/xe/xe_noexec_ping_pong.c
>> @@ -0,0 +1,105 @@
>> +// SPDX-License-Identifier: MIT
>> +/*
>> + * Copyright © 2023 Intel Corporation
>> + */
>> +
>> +#include <unistd.h>
>> +
>> +#include "igt.h"
>> +#include "xe_drm.h"
>> +
>> +#include "xe/xe_ioctl.h"
>> +#include "xe/xe_query.h"
>> +
>> +#define NUM_VMS 10
>> +#define NUM_BOS 1
>> +
>> +/**
>> + * TEST: Expose compute VM's unnecessary rebinds
>> + * Category: Software building block
>> + * Sub-category: compute
>> + * Test category: functionality test
>> + */
>> +
>> +/*
>> + * This test creates compute vms, binds a couple of bos and an engine each,
>> + * thus redying it for execution. However, VRAM memory is over-
>> + * committed and while there is still nothing to execute, an eviction
>> + * will trigger the VM's rebind worker to rebind the evicted bo, which
>> + * will in turn trigger another eviction and so on.
>> + *
>> + * Since we don't have eviction stats yet we need to watch "top" for
>> + * the rebind kworkers using a lot of CPU while the test idles.
>> + *
>> + * The correct driver behaviour should be not to rebind anything unless
>> + * there is worked queued on one of the VM's compute engines.
>> + */
>> +static void test_ping_pong(int fd, struct drm_xe_engine_class_instance *eci)
>> +{
>> +	size_t vram_size = xe_vram_size(fd);
>> +	size_t align = xe_get_default_alignment(fd);
>> +	size_t bo_size = vram_size / NUM_VMS / NUM_BOS;
> Does this math work out to over commit? If I'm reading this correctly
> this is going to be less than 100% of the VRAm. Thinking it should be:
>
> (vram_size * 3) / 2) / (NUM_VMS / NUM_BOS);

The bo size above calculates the size to exactly (except for rounding) 
fill VRAM, since we have (NUM_VMS * NUM_BOS) in total. The size is then 
aligned upwards and increased with 10% below, so I believe it is correct.

>
>> +	uint32_t vm[NUM_VMS];
>> +	uint32_t bo[NUM_VMS][NUM_BOS];
>> +	uint32_t engines[NUM_VMS];
>> +	unsigned int i, j;
>> +
>> +	if (!bo_size)
>> +		return;
>> +
>> +	/* Align and make sure we overcommit vram with at least 10% */
>> +	bo_size = (bo_size + bo_size / 10 + align - 1) / align * align;
Here.
>> +
>> +	/*
>> +	 * This should not start ping-ponging memory between system and
>> +	 * VRAM. For now look at top to determine. TODO: Look at eviction
>> +	 * stats.
>> +	 */
>> +	for (i = 0; i < NUM_VMS; ++i) {
>> +		struct drm_xe_ext_engine_set_property ext = {
>> +			.base.next_extension = 0,
>> +			.base.name = XE_ENGINE_EXTENSION_SET_PROPERTY,
>> +			.property = XE_ENGINE_SET_PROPERTY_COMPUTE_MODE,
>> +			.value = 1,
>> +		};
>> +
>> +		vm[i] = xe_vm_create(fd, DRM_XE_VM_CREATE_COMPUTE_MODE, 0);
>> +		for (j = 0; j < NUM_BOS; ++j) {
>> +			igt_debug("Creating bo size %lu for vm %u\n",
>> +				  (unsigned long) bo_size,
>> +				  (unsigned int) vm[i]);
>> +
>> +			bo[i][j] = xe_bo_create_flags(fd, vm[i], bo_size,
>> +						      vram_memory(fd, 0));
>> +			xe_vm_bind(fd, vm[i], bo[i][j], 0, 0x40000 + j*bo_size,
>> +				   bo_size, NULL, 0);
>> +		}
>> +		engines[i] = xe_engine_create(fd, vm[i], eci,
>> +					      to_user_pointer(&ext));
>> +	}
>> +
>> +	igt_info("Now sleeping for 20s.\n");
>> +	igt_info("Watch \"top\" for high-cpu kworkers!\n");
>> +	sleep(20);
> Maybe 5, 20 is kinda a long time.

Sure, I can cut it down to 10 so there are at least two top refreshes 
happening. Until we have eviction stats there is no point running this 
test as part of CI anyway, and with eviction stats a couple of seconds 
should be enough.

/Thomas

>
> Matt
>
>> +
>> +	for (i = 0; i < NUM_VMS; ++i) {
>> +		xe_engine_destroy(fd, engines[i]);
>> +		for (j = 0; j < NUM_BOS; ++j)
>> +			gem_close(fd, bo[i][j]);
>> +		xe_vm_destroy(fd, vm[i]);
>> +	}
>> +}
>> +
>> +static int fd;
>> +
>> +igt_simple_main
>> +{
>> +
>> +	fd = drm_open_driver(DRIVER_XE);
>> +	xe_device_get(fd);
>> +
>> +	test_ping_pong(fd, xe_hw_engine(fd, 0));
>> +
>> +	xe_device_put(fd);
>> +	close(fd);
>> +}
>> -- 
>> 2.39.2
>>