On Fri, Nov 15, 2019 at 11:34:54AM +0800, Liu, Aaron wrote:
> DMA's atomic behavir is unlike GFX,If the comparing data is not
> equal to destination data,
> For GFX, loop again till gfx timeout(system hang).
> For DMA, loop again till timer expired and then send interrupt.
> So testcase can't use interrupt mechanism.
> We take another way to verify. When the comparing data is not
> equal to destination data, overwrite the source data to the destination
> buffer. Otherwise, original destination data unchanged.
> So if the bo_cpu data is overwritten, the result is passed.
>
> Steps:
> 1. use linear write packet to write 0xdeadbeaf to secure buffer,
> 2. use atmoic packet and ATOMIC_CMPSWAP_RTN_32 opcode to compare
> the cmp_data(0xdeadbeaf) to the written data which has been encrypted.
>
> Signed-off-by: Aaron Liu <aaron....@amd.com>
> ---
> tests/amdgpu/basic_tests.c | 55
> ++++++++++++++++++++++++++++++++++++++++++++--
> 1 file changed, 53 insertions(+), 2 deletions(-)
>
> diff --git a/tests/amdgpu/basic_tests.c b/tests/amdgpu/basic_tests.c
> index 907237e..5e687e3 100644
> --- a/tests/amdgpu/basic_tests.c
> +++ b/tests/amdgpu/basic_tests.c
> @@ -109,6 +109,20 @@ CU_TestInfo basic_tests[] = {
> #define SDMA_OPCODE_COPY 1
> # define SDMA_COPY_SUB_OPCODE_LINEAR 0
>
> +#define SDMA_OPCODE_ATOMIC 10
> +# define SDMA_ATOMIC_LOOP(x) ((x) << 0)
> + /* 0 - single_pass_atomic.
> + * 1 - loop_until_compare_satisfied.
> + */
> +# define SDMA_ATOMIC_TMZ(x) ((x) << 2)
> + /* 0 - non-TMZ.
> + * 1 - TMZ.
> + */
> +# define SDMA_ATOMIC_OPCODE(x) ((x) << 9)
> + /* TC_OP_ATOMIC_CMPSWAP_RTN_32 0x00000008
> + * same as Packet 3
> + */
> +
> #define GFX_COMPUTE_NOP 0xffff1000
> #define SDMA_NOP 0x0
>
> @@ -1408,6 +1422,7 @@
> amdgpu_command_submission_write_linear_helper_with_secure(amdgpu_device_handle
> struct amdgpu_cs_request *ibs_request;
> uint64_t bo_mc;
> volatile uint32_t *bo_cpu;
> + uint32_t bo_cpu_origin;
> int i, j, r, loop, ring_id;
> uint64_t gtt_flags[2] = {0, AMDGPU_GEM_CREATE_CPU_GTT_USWC};
> amdgpu_va_handle va_handle;
> @@ -1460,8 +1475,9 @@
> amdgpu_command_submission_write_linear_helper_with_secure(amdgpu_device_handle
>
> sdma_write_length);
> else
> pm4[i++] =
> SDMA_PACKET(SDMA_OPCODE_WRITE,
> -
> SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
> - pm4[i++] = 0xffffffff & bo_mc;
> +
> SDMA_WRITE_SUB_OPCODE_LINEAR,
> + secure ?
> SDMA_ATOMIC_TMZ(1) : 0);
> + pm4[i++] = 0xfffffffc & bo_mc;
> pm4[i++] = (0xffffffff00000000 & bo_mc) >> 32;
> if (family_id >= AMDGPU_FAMILY_AI)
> pm4[i++] = sdma_write_length - 1;
> @@ -1513,6 +1529,41 @@
> amdgpu_command_submission_write_linear_helper_with_secure(amdgpu_device_handle
> ip_type, ring_id, i,
> pm4,
> 1, resources, ib_info,
> ibs_request, true);
> + } else if (ip_type == AMDGPU_HW_IP_DMA) {
> + /* restore the bo_cpu to compare */
> + bo_cpu_origin = bo_cpu[0];
> + memset((void*)pm4, 0, pm4_dw *
> sizeof(uint32_t));
> + /* atomic opcode for 32b w/ RTN and
> ATOMIC_SWAPCMP_RTN
> + * loop, 1-loop_until_compare_satisfied.
> + * single_pass_atomic, 0-lru
> + */
> + pm4[i++] = SDMA_PACKET(SDMA_OPCODE_ATOMIC,
> + 0,
> +
> SDMA_ATOMIC_LOOP(1) |
> +
> SDMA_ATOMIC_TMZ(1) |
> +
> SDMA_ATOMIC_OPCODE(TC_OP_ATOMIC_CMPSWAP_RTN_32));
> + pm4[i++] = 0xfffffffc & bo_mc;
> + pm4[i++] = (0xffffffff00000000 & bo_mc) >> 32;
> + pm4[i++] = 0x12345678;
> + pm4[i++] = 0x0;
> + pm4[i++] = 0xdeadbeaf;
> + pm4[i++] = 0x0;
> + pm4[i++] = 0x100;
> + amdgpu_test_exec_cs_helper_raw(device,
> context_handle,
> + ip_type, ring_id, i,
> pm4,
> + 1, resources, ib_info,
> + ibs_request, true);
> + /* DMA's atomic behavir is unlike GFX
> + * If the comparing data is not equal to
> destination data,
> + * For GFX, loop again till gfx timeout(system
> hang).
> + * For DMA, loop again till timer expired and
> then send interrupt.
> + * So testcase can't use interrupt mechanism.
> + * We take another way to verify. When the
> comparing data is not
> + * equal to destination data, overwrite the
> source data to the destination
> + * buffer. Otherwise, original destination data
> unchanged.
> + * So if the bo_cpu data is overwritten, the
> result is passed.
> + */
> + CU_ASSERT_NOT_EQUAL(bo_cpu[0], bo_cpu_origin);
I think we can add one more test for the case that overwrite the source
data to the destination buffer.
Thanks,
Ray
> }
>
> r = amdgpu_bo_unmap_and_free(bo, va_handle, bo_mc,
> --
> 2.7.4
>
_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
