[Intel-gfx] [PATCH v3] igt/gem_trtt: Exercise the TRTT hardware
Chris Wilson
chris at chris-wilson.co.uk
Thu Mar 3 10:04:03 UTC 2016
On Thu, Mar 03, 2016 at 10:25:59AM +0530, akash.goel at intel.com wrote:
> +static bool uses_full_ppgtt(int fd, int min)
gem_gtt_type()
> +static bool has_softpin_support(int fd)
gem_has_softpin()
> +static int emit_store_dword(int fd, uint32_t *cmd_buf, uint32_t dw_offset,
> + uint64_t vaddr, uint32_t data)
> +{
> + /* Check that softpin addresses are in the correct form */
> + assert(vaddr == gen8_canonical_addr(vaddr));
> +
> + /* SDI cannot write to unaligned addresses */
> + assert((vaddr & 3) == 0);
> +
> + cmd_buf[dw_offset++] = MI_STORE_DWORD_IMM;
> + cmd_buf[dw_offset++] = vaddr & 0xFFFFFFFC;
> + cmd_buf[dw_offset++] = (vaddr >> 32) & 0xFFFF; /* bits 32:47 */
> + cmd_buf[dw_offset++] = data;
> +
> + return dw_offset;
> +}
> +
> +/* emit_store_qword
> + * populate batch buffer with MI_STORE_DWORD_IMM command
> + * @fd: drm file descriptor
> + * @cmd_buf: batch buffer
> + * @dw_offset: write offset in batch buffer
> + * @vaddr: destination Virtual address
> + * @data: u64 data to be stored at destination
> + */
> +static int emit_store_qword(int fd, uint32_t *cmd_buf, uint32_t dw_offset,
> + uint64_t vaddr, uint64_t data)
> +{
> + /* Check that softpin addresses are in the correct form */
> + assert(vaddr == gen8_canonical_addr(vaddr));
> +
> + /* SDI cannot write to unaligned addresses */
> + assert((vaddr & 3) == 0);
> +
> + cmd_buf[dw_offset++] = MI_STORE_DWORD_IMM | 0x3;
> + cmd_buf[dw_offset++] = vaddr & 0xFFFFFFFC;
You just asserted above that the low bits aren't set.
> + cmd_buf[dw_offset++] = (vaddr >> 32) & 0xFFFF; /* bits 32:47 */
This conflicts with the value that would be set by the kernel (if it had
to the relocation). Who is correct? If not required, please don't raise
the spectre of devastating bugs.
> + cmd_buf[dw_offset++] = data;
> + cmd_buf[dw_offset++] = data >> 32;
> +
> + return dw_offset;
> +}
> +
> +/* emit_bb_end
> + * populate batch buffer with MI_BATCH_BUFFER_END command
> + * @fd: drm file descriptor
> + * @cmd_buf: batch buffer
> + * @dw_offset: write offset in batch buffer
> + */
> +static int emit_bb_end(int fd, uint32_t *cmd_buf, uint32_t dw_offset)
> +{
> + cmd_buf[dw_offset++] = MI_BATCH_BUFFER_END;
> + cmd_buf[dw_offset++] = 0;
Why? execbuf.batch_len must be aligned, but the CS parser doesn't care,
and there is no guarantee that you are aligned coming into
emit_bb_end().
> +
> + return dw_offset;
> +}
> +
> +/* setup_execbuffer
> + * helper for buffer execution
> + * @execbuf - pointer to execbuffer
> + * @exec_object - pointer to exec object2 struct
> + * @ring - ring to be used
> + * @buffer_count - how manu buffers to submit
> + * @batch_length - length of batch buffer
> + */
> +static void setup_execbuffer(struct drm_i915_gem_execbuffer2 *execbuf,
> + struct drm_i915_gem_exec_object2 *exec_object,
> + uint32_t ctx_id, int ring, int buffer_count, int batch_length)
> +{
> + memset(execbuf, 0, sizeof(*execbuf));
> +
> + execbuf->buffers_ptr = (unsigned long)exec_object;
> + execbuf->buffer_count = buffer_count;
> + execbuf->batch_len = batch_length;
> + execbuf->flags = ring;
> + i915_execbuffer2_set_context_id(*execbuf, ctx_id);
> +}
> +
> +#define TABLE_SIZE 0x1000
> +#define TILE_SIZE 0x10000
> +
> +#define TRTT_SEGMENT_SIZE (1ULL << 44)
> +#define PPGTT_SIZE (1ULL << 48)
> +
> +#define NULL_TILE_PATTERN 0xFFFFFFFF
> +#define INVALID_TILE_PATTERN 0xFFFFFFFE
> +
> +struct local_i915_gem_context_trtt_param {
> + uint64_t segment_base_addr;
> + uint64_t l3_table_address;
> + uint32_t invd_tile_val;
> + uint32_t null_tile_val;
> +};
> +
> +/* setup_trtt
> + * Helper function to request KMD to enable TRTT
> + * @fd - drm fd
> + * @ctx_id - id of the context for which TRTT is to be enabled
> + * @l3_table_address - GFX address of the L3 table
> + * @segment_base_addr - offset of the TRTT segment in PPGTT space
> + */
> +static void
> +setup_trtt(int fd, uint32_t ctx_id, uint64_t l3_table_address,
> + uint64_t segment_base_addr)
> +{
> + struct local_i915_gem_context_param ctx_param;
> + struct local_i915_gem_context_trtt_param trtt_param;
> +
> + memset(&ctx_param, 0, sizeof(ctx_param));
> +
> + trtt_param.null_tile_val = NULL_TILE_PATTERN;
> + trtt_param.invd_tile_val = INVALID_TILE_PATTERN;
> + trtt_param.l3_table_address = l3_table_address;
> + trtt_param.segment_base_addr = segment_base_addr;
> +
> + ctx_param.context = ctx_id;
> + ctx_param.size = sizeof(trtt_param);
> + ctx_param.param = LOCAL_CONTEXT_PARAM_TRTT;
> + ctx_param.value = (uint64_t)&trtt_param;
> +
> + gem_context_set_param(fd, &ctx_param);
> +}
> +
> +/* bo_alloc_setup
> + * allocate bo and populate exec object
> + * @exec_object2 - pointer to exec object
> + * @bo_sizee - buffer size
> + * @flags - exec flags
> + * @bo_offset - pointer to the current PPGTT offset
> + */
> +static void bo_alloc_setup(int fd, struct drm_i915_gem_exec_object2 *exec_object2,
> + uint64_t bo_size, uint64_t flags, uint64_t *bo_offset)
> +{
> + memset(exec_object2, 0, sizeof(*exec_object2));
> + exec_object2->handle = gem_create(fd, bo_size);
> + exec_object2->flags = flags;
> +
> + if (bo_offset)
> + {
> + exec_object2->offset = *bo_offset;
> + *bo_offset += bo_size;
> + }
> +}
> +
> +/* submit_trtt_context
> + * This helper function will create a new context if the TR-TT segment
> + * base address is not zero, allocate a L3 table page, 2 pages apiece
> + * for L2/L1 tables and couple of data buffers of 64KB in size, matching the
> + * Tile size. The 2 data buffers will be mapped to the 2 ends of TRTT virtual
> + * space. Series of MI_STORE_DWORD_IMM commands will be added in the batch
> + * buffer to first update the TR-TT table entries and then to update the data
> + * buffers using their TR-TT VA, exercising the table programming done
> + * previously.
> + * Invoke CONTEXT_SETPARAM ioctl to request KMD to enable TRTT.
> + * Invoke execbuffer to submit the batch buffer.
> + * Verify value of first DWORD in the 2 data buffer matches the data asked
> + * to be written by the GPU.
> + */
> +static void submit_trtt_context(int fd, uint64_t segment_base_addr)
> +{
> + enum {
> + L3_TBL,
> + L2_TBL1,
> + L2_TBL2,
> + L1_TBL1,
> + L1_TBL2,
> + DATA1,
> + DATA2,
> + BATCH,
> + NUM_BUFFERS,
> + };
> +
> + int ring, len = 0;
> + uint32_t *ptr;
> + struct drm_i915_gem_execbuffer2 execbuf;
> + struct drm_i915_gem_exec_object2 exec_object2[NUM_BUFFERS];
> + uint32_t batch_buffer[BO_SIZE];
> + uint32_t ctx_id, data, last_entry_offset;
> + uint64_t cur_ppgtt_off, exec_flags;
> + uint64_t first_tile_addr, last_tile_addr;
> +
> + first_tile_addr = segment_base_addr;
> + last_tile_addr = first_tile_addr + TRTT_SEGMENT_SIZE - TILE_SIZE;
> +
> + if (segment_base_addr == 0) {
> + /* Use the default context for first iteration */
> + ctx_id = 0;
Seems like a variable the callee wants to control. (Otherwise you are
missing a significant test for non-default contexts). It also implies
that we don't test what happens when we call set-trtt-context twice on a
context.
> + /* To avoid conflict with the TR-TT segment */
> + cur_ppgtt_off = TRTT_SEGMENT_SIZE;
> + } else {
> + /* Create a new context to have different TRTT settings
> + * on every iteration.
> + */
> + ctx_id = gem_context_create(fd);
> + cur_ppgtt_off = 0;
> + }
> +
> + exec_flags = EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
> +
> + /* first allocate Batch buffer BO */
> + bo_alloc_setup(fd, &exec_object2[BATCH], BO_SIZE, exec_flags, NULL);
> +
> + /* table BOs and data buffer BOs are written by GPU and are soft pinned */
> + exec_flags |= (EXEC_OBJECT_WRITE | EXEC_OBJECT_PINNED);
> +
> + /* Allocate a L3 table BO */
> + bo_alloc_setup(fd, &exec_object2[L3_TBL], TABLE_SIZE, exec_flags, &cur_ppgtt_off);
> +
> + /* Allocate two L2 table BOs */
> + bo_alloc_setup(fd, &exec_object2[L2_TBL1], TABLE_SIZE, exec_flags, &cur_ppgtt_off);
> + bo_alloc_setup(fd, &exec_object2[L2_TBL2], TABLE_SIZE, exec_flags, &cur_ppgtt_off);
> +
> + /* Allocate two L1 table BOs */
> + bo_alloc_setup(fd, &exec_object2[L1_TBL1], TABLE_SIZE, exec_flags, &cur_ppgtt_off);
> + bo_alloc_setup(fd, &exec_object2[L1_TBL2], TABLE_SIZE, exec_flags, &cur_ppgtt_off);
> +
> + /* Align the PPGTT offsets for the 2 data buffers to next 64 KB boundary */
> + cur_ppgtt_off = ALIGN(cur_ppgtt_off, TILE_SIZE);
> +
> + /* Allocate two Data buffer BOs */
> + bo_alloc_setup(fd, &exec_object2[DATA1], TILE_SIZE, exec_flags, &cur_ppgtt_off);
> + bo_alloc_setup(fd, &exec_object2[DATA2], TILE_SIZE, exec_flags, &cur_ppgtt_off);
> +
> + /* Add commands to update the two L3 table entries to point them to the L2 tables*/
> + last_entry_offset = 511*sizeof(uint64_t);
> +
> + len = emit_store_qword(fd, batch_buffer, len,
> + exec_object2[L3_TBL].offset,
> + exec_object2[L2_TBL1].offset);
> +
> + len = emit_store_qword(fd, batch_buffer, len,
> + exec_object2[L3_TBL].offset + last_entry_offset,
> + exec_object2[L2_TBL2].offset);
> +
> + /* Add commands to update an entry of 2 L2 tables to point them to the L1 tables*/
> + len = emit_store_qword(fd, batch_buffer, len,
> + exec_object2[L2_TBL1].offset,
> + exec_object2[L1_TBL1].offset);
> +
> + len = emit_store_qword(fd, batch_buffer, len,
> + exec_object2[L2_TBL2].offset + last_entry_offset,
> + exec_object2[L1_TBL2].offset);
> +
> + /* Add commands to update an entry of 2 L1 tables to point them to the data buffers*/
> + last_entry_offset = 1023*sizeof(uint32_t);
> +
> + len = emit_store_dword(fd, batch_buffer, len,
> + exec_object2[L1_TBL1].offset,
> + exec_object2[DATA1].offset >> 16);
> +
> + len = emit_store_dword(fd, batch_buffer, len,
> + exec_object2[L1_TBL2].offset + last_entry_offset,
> + exec_object2[DATA2].offset >> 16);
> +
> + /* Add commands to update the 2 data buffers, using their TRTT VA */
> + data = 0x12345678;
> + len = emit_store_dword(fd, batch_buffer, len, first_tile_addr, data);
> + len = emit_store_dword(fd, batch_buffer, len, last_tile_addr, data);
> +
> + len = emit_bb_end(fd, batch_buffer, len);
> + gem_write(fd, exec_object2[BATCH].handle, 0, batch_buffer, len*4);
> +
> + /* Request KMD to setup the TR-TT */
> + setup_trtt(fd, ctx_id, exec_object2[L3_TBL].offset, first_tile_addr);
> +
> + ring = I915_EXEC_RENDER;
> + setup_execbuffer(&execbuf, exec_object2, ctx_id, ring, NUM_BUFFERS, len*4);
> +
> + /* submit command buffer */
> + gem_execbuf(fd, &execbuf);
> +
> + /* read the 2 data buffers to check for the value written by the GPU */
> + ptr = mmap_bo(fd, exec_object2[DATA1].handle, TILE_SIZE);
> + igt_fail_on_f(ptr[0] != data,
> + "\nCPU read does not match GPU write,\
> + expected: 0x%x, got: 0x%x\n",
> + data, ptr[0]);
Just igt_assert_eq_u32(ptr[0], expected);
> +
> + ptr = mmap_bo(fd, exec_object2[DATA2].handle, TILE_SIZE);
> + igt_fail_on_f(ptr[0] != data,
> + "\nCPU read does not match GPU write,\
> + expected: 0x%x, got: 0x%x\n",
> + data, ptr[0]);
> +
> + gem_close(fd, exec_object2[L3_TBL].handle);
> + gem_close(fd, exec_object2[L2_TBL1].handle);
> + gem_close(fd, exec_object2[L2_TBL2].handle);
> + gem_close(fd, exec_object2[L1_TBL1].handle);
> + gem_close(fd, exec_object2[L1_TBL2].handle);
> + gem_close(fd, exec_object2[DATA1].handle);
> + gem_close(fd, exec_object2[DATA2].handle);
> + gem_close(fd, exec_object2[BATCH].handle);
Before we destroy the context (or exit), how about a query_trtt().
We should also query after create to ensure that the defaults are set.
Just thinking that is better doing the query after several steps (i.e.
the execbuf) rather than immediately after the set in order to give time
for something to go wrong. We should also ensure that everything remains
set after a GPU hang and suspend/resume.
-Chris
--
Chris Wilson, Intel Open Source Technology Centre
More information about the Intel-gfx
mailing list