Re: [PATCH i-g-t v11 4/5] tests/intel/xe_exec_system_allocator: Add madvise-swizzle test

[Matthew Brost <matthew.brost@intel.com>]

On Wed, Sep 03, 2025 at 02:13:38PM +0000, nishit.sharma@intel.com wrote:
> From: Nishit Sharma <nishit.sharma@intel.com>
> 
> madvise-swizzle test introduced which is called in combination with other
> tests as well. In this test the buffer object preferred location is
> system memory. MADVISE_SWIZZLE called in main loop to toggle between
> SYSTEM/DEVICE on each pass.
> 
> Signed-off-by: Nishit Sharma <nishit.sharma@intel.com>
> ---
>  tests/intel/xe_exec_system_allocator.c | 32 ++++++++++++++++++++++++++
>  1 file changed, 32 insertions(+)
> 
> diff --git a/tests/intel/xe_exec_system_allocator.c b/tests/intel/xe_exec_system_allocator.c
> index e7f3d423a..4a208f6d8 100644
> --- a/tests/intel/xe_exec_system_allocator.c
> +++ b/tests/intel/xe_exec_system_allocator.c
> @@ -777,6 +777,7 @@ partial(int fd, struct drm_xe_engine_class_instance *eci, unsigned int flags)
>  #define PROCESSES		(0x1 << 24)
>  #define PREFETCH_BENCHMARK	(0x1 << 25)
>  #define PREFETCH_SYS_BENCHMARK	(0x1 << 26)
> +#define MADVISE_SWIZZLE			(0x1 << 27)
>  
>  #define N_MULTI_FAULT		4
>  
> @@ -885,7 +886,9 @@ partial(int fd, struct drm_xe_engine_class_instance *eci, unsigned int flags)
>   * arg[1]:
>   *
>   * @malloc:				malloc single buffer for all execs, issue a command which will trigger multiple faults
> + * @malloc-madvise:			malloc single buffer for all execs, issue a command which will trigger multiple faults, perfoems madvise operation
>   * @malloc-prefetch:			malloc single buffer for all execs, prefetch buffer before each exec
> + * @malloc-prefetch-madvise:		malloc single buffer for all execs, prefetch buffer before each exec, performs madvise operation
>   * @malloc-multi-fault:			malloc single buffer for all execs
>   * @malloc-fork-read:			malloc single buffer for all execs, fork a process to read test output
>   * @malloc-fork-read-after:		malloc single buffer for all execs, fork a process to read test output, check again after fork returns in parent
> @@ -897,6 +900,7 @@ partial(int fd, struct drm_xe_engine_class_instance *eci, unsigned int flags)
>   * @mmap:				mmap single buffer for all execs
>   * @mmap-prefetch:			mmap single buffer for all execs, prefetch buffer before each exec
>   * @mmap-remap:				mmap and mremap a buffer for all execs
> + * @mmap-remap-madvise:			mmap and mremap a buffer for all execs, performs madvise operations
>   * @mmap-remap-dontunmap:		mmap and mremap a buffer with dontunmap flag for all execs
>   * @mmap-remap-ro:			mmap and mremap a read-only buffer for all execs
>   * @mmap-remap-ro-dontunmap:		mmap and mremap a read-only buffer with dontunmap flag for all execs
> @@ -916,8 +920,10 @@ partial(int fd, struct drm_xe_engine_class_instance *eci, unsigned int flags)
>   * @mmap-file-mlock:			mmap and mlock single buffer, with file backing, for all execs
>   * @mmap-race:				mmap single buffer for all execs with race between cpu and gpu access
>   * @free:				malloc and free buffer for each exec
> + * @free-madvise:			malloc and free buffer for each exec, performs madvise operation
>   * @free-race:				malloc and free buffer for each exec with race between cpu and gpu access
>   * @new:				malloc a new buffer for each exec
> + * @new-madvise:			malloc a new buffer for each exec, performs madvise operation

You have mismatches compare to the section names. Building the
IGTs should complain.

>   * @new-prefetch:			malloc a new buffer and prefetch for each exec
>   * @new-race:				malloc a new buffer for each exec with race between cpu and gpu access
>   * @new-bo-map:				malloc a new buffer or map BO for each exec
> @@ -999,6 +1005,23 @@ static void igt_require_hugepages(void)
>  		      "No huge pages available!\n");
>  }
>  
> +static void
> +madvise_swizzle_op_exec(int fd, uint32_t vm, struct test_exec_data *data,
> +		size_t bo_size, uint64_t addr, int index)
> +{
> +	int preferred_loc;
> +
> +	if (index % 2 == 0)
> +		preferred_loc = DRM_XE_PREFERRED_LOC_DEFAULT_SYSTEM;
> +	else
> +		preferred_loc = DRM_XE_PREFERRED_LOC_DEFAULT_DEVICE;
> +
> +	xe_vm_madvise(fd, vm, to_user_pointer(data), bo_size, 0,
> +			DRM_XE_MEM_RANGE_ATTR_PREFERRED_LOC,
> +			preferred_loc,
> +			0);
> +}
> +
>  static void
>  test_exec(int fd, struct drm_xe_engine_class_instance *eci,
>  	  int n_exec_queues, int n_execs, size_t bo_size,
> @@ -1210,6 +1233,9 @@ test_exec(int fd, struct drm_xe_engine_class_instance *eci,
>  		if (barrier)
>  			pthread_barrier_wait(barrier);
>  
> +		if (flags & MADVISE_SWIZZLE)
> +			madvise_swizzle_op_exec(fd, vm, data, bo_size, addr, i);
> +
>  		if (flags & MULTI_FAULT) {
>  			b = 0;
>  			for (j = 0; j < N_MULTI_FAULT - 1; ++j)
> @@ -1790,7 +1816,9 @@ igt_main
>  	struct drm_xe_engine_class_instance *hwe;
>  	const struct section sections[] = {
>  		{ "malloc", 0 },
> +		{ "malloc-madvise", MADVISE_SWIZZLE },
>  		{ "malloc-prefetch", PREFETCH },
> +		{ "malloc-prefetch-madvise", PREFETCH | MADVISE_SWIZZLE },
>  		{ "malloc-multi-fault", MULTI_FAULT },
>  		{ "malloc-fork-read", FORK_READ },
>  		{ "malloc-fork-read-after", FORK_READ | FORK_READ_AFTER },
> @@ -1802,6 +1830,7 @@ igt_main
>  		{ "mmap", MMAP },
>  		{ "mmap-prefetch", MMAP | PREFETCH },
>  		{ "mmap-remap", MMAP | MREMAP },
> +		{ "mmap-remap-madvise", MMAP | MREMAP | MADVISE_SWIZZLE },
>  		{ "mmap-remap-dontunmap", MMAP | MREMAP | DONTUNMAP },
>  		{ "mmap-remap-ro", MMAP | MREMAP | READ_ONLY_REMAP },
>  		{ "mmap-remap-ro-dontunmap", MMAP | MREMAP | DONTUNMAP |
> @@ -1828,13 +1857,16 @@ igt_main
>  		{ "mmap-file-mlock", MMAP | LOCK | FILE_BACKED },
>  		{ "mmap-race", MMAP | RACE },
>  		{ "free", NEW | FREE },
> +		{ "free-madvise", NEW | FREE | MADVISE_SWIZZLE },
>  		{ "free-race", NEW | FREE | RACE },
>  		{ "new", NEW },
> +		{ "new-madvise", NEW | MADVISE_SWIZZLE },
>  		{ "new-prefetch", NEW | PREFETCH },
>  		{ "new-race", NEW | RACE },
>  		{ "new-bo-map", NEW | BO_MAP },
>  		{ "new-busy", NEW | BUSY },
>  		{ "mmap-free", MMAP | NEW | FREE },
> +		{ "mmap-free", MMAP | NEW | FREE | MADVISE_SWIZZLE },

mmap-free-madvise

>  		{ "mmap-free-huge", MMAP | NEW | FREE | HUGE_PAGE },
>  		{ "mmap-free-race", MMAP | NEW | FREE | RACE },
>  		{ "mmap-new", MMAP | NEW },

Also I thought about this a bit more, 'mmap-new-madvise' would
be helpful too as malloc / free patterns are lazy (i.e.,
sometimes free doesn't call into the kernel) whereas mmap /
munmap are not lazy.

Otherwise LGTM.

Matt

> -- 
> 2.43.0
>