2 * Copyright 2013 Advanced Micro Devices, Inc.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
22 * Authors: Alex Deucher
24 #include <linux/firmware.h>
27 #include "radeon_ucode.h"
28 #include "radeon_asic.h"
29 #include "radeon_trace.h"
33 #define CIK_SDMA_UCODE_SIZE 1050
34 #define CIK_SDMA_UCODE_VERSION 64
36 u32 cik_gpu_check_soft_reset(struct radeon_device *rdev);
40 * Starting with CIK, the GPU has new asynchronous
41 * DMA engines. These engines are used for compute
42 * and gfx. There are two DMA engines (SDMA0, SDMA1)
43 * and each one supports 1 ring buffer used for gfx
44 * and 2 queues used for compute.
46 * The programming model is very similar to the CP
47 * (ring buffer, IBs, etc.), but sDMA has it's own
48 * packet format that is different from the PM4 format
49 * used by the CP. sDMA supports copying data, writing
50 * embedded data, solid fills, and a number of other
51 * things. It also has support for tiling/detiling of
56 * cik_sdma_get_rptr - get the current read pointer
58 * @rdev: radeon_device pointer
59 * @ring: radeon ring pointer
61 * Get the current rptr from the hardware (CIK+).
63 uint32_t cik_sdma_get_rptr(struct radeon_device *rdev,
64 struct radeon_ring *ring)
68 if (rdev->wb.enabled) {
69 rptr = rdev->wb.wb[ring->rptr_offs/4];
71 if (ring->idx == R600_RING_TYPE_DMA_INDEX)
72 reg = SDMA0_GFX_RB_RPTR + SDMA0_REGISTER_OFFSET;
74 reg = SDMA0_GFX_RB_RPTR + SDMA1_REGISTER_OFFSET;
79 return (rptr & 0x3fffc) >> 2;
83 * cik_sdma_get_wptr - get the current write pointer
85 * @rdev: radeon_device pointer
86 * @ring: radeon ring pointer
88 * Get the current wptr from the hardware (CIK+).
90 uint32_t cik_sdma_get_wptr(struct radeon_device *rdev,
91 struct radeon_ring *ring)
95 if (ring->idx == R600_RING_TYPE_DMA_INDEX)
96 reg = SDMA0_GFX_RB_WPTR + SDMA0_REGISTER_OFFSET;
98 reg = SDMA0_GFX_RB_WPTR + SDMA1_REGISTER_OFFSET;
100 return (RREG32(reg) & 0x3fffc) >> 2;
104 * cik_sdma_set_wptr - commit the write pointer
106 * @rdev: radeon_device pointer
107 * @ring: radeon ring pointer
109 * Write the wptr back to the hardware (CIK+).
111 void cik_sdma_set_wptr(struct radeon_device *rdev,
112 struct radeon_ring *ring)
116 if (ring->idx == R600_RING_TYPE_DMA_INDEX)
117 reg = SDMA0_GFX_RB_WPTR + SDMA0_REGISTER_OFFSET;
119 reg = SDMA0_GFX_RB_WPTR + SDMA1_REGISTER_OFFSET;
121 WREG32(reg, (ring->wptr << 2) & 0x3fffc);
126 * cik_sdma_ring_ib_execute - Schedule an IB on the DMA engine
128 * @rdev: radeon_device pointer
129 * @ib: IB object to schedule
131 * Schedule an IB in the DMA ring (CIK).
133 void cik_sdma_ring_ib_execute(struct radeon_device *rdev,
134 struct radeon_ib *ib)
136 struct radeon_ring *ring = &rdev->ring[ib->ring];
137 u32 extra_bits = (ib->vm ? ib->vm->ids[ib->ring].id : 0) & 0xf;
139 if (rdev->wb.enabled) {
140 u32 next_rptr = ring->wptr + 5;
141 while ((next_rptr & 7) != 4)
144 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_WRITE, SDMA_WRITE_SUB_OPCODE_LINEAR, 0));
145 radeon_ring_write(ring, ring->next_rptr_gpu_addr & 0xfffffffc);
146 radeon_ring_write(ring, upper_32_bits(ring->next_rptr_gpu_addr));
147 radeon_ring_write(ring, 1); /* number of DWs to follow */
148 radeon_ring_write(ring, next_rptr);
151 /* IB packet must end on a 8 DW boundary */
152 while ((ring->wptr & 7) != 4)
153 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0));
154 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_INDIRECT_BUFFER, 0, extra_bits));
155 radeon_ring_write(ring, ib->gpu_addr & 0xffffffe0); /* base must be 32 byte aligned */
156 radeon_ring_write(ring, upper_32_bits(ib->gpu_addr));
157 radeon_ring_write(ring, ib->length_dw);
162 * cik_sdma_hdp_flush_ring_emit - emit an hdp flush on the DMA ring
164 * @rdev: radeon_device pointer
165 * @ridx: radeon ring index
167 * Emit an hdp flush packet on the requested DMA ring.
169 static void cik_sdma_hdp_flush_ring_emit(struct radeon_device *rdev,
172 struct radeon_ring *ring = &rdev->ring[ridx];
173 u32 extra_bits = (SDMA_POLL_REG_MEM_EXTRA_OP(1) |
174 SDMA_POLL_REG_MEM_EXTRA_FUNC(3)); /* == */
177 if (ridx == R600_RING_TYPE_DMA_INDEX)
178 ref_and_mask = SDMA0;
180 ref_and_mask = SDMA1;
182 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0, extra_bits));
183 radeon_ring_write(ring, GPU_HDP_FLUSH_DONE);
184 radeon_ring_write(ring, GPU_HDP_FLUSH_REQ);
185 radeon_ring_write(ring, ref_and_mask); /* reference */
186 radeon_ring_write(ring, ref_and_mask); /* mask */
187 radeon_ring_write(ring, (0xfff << 16) | 10); /* retry count, poll interval */
191 * cik_sdma_fence_ring_emit - emit a fence on the DMA ring
193 * @rdev: radeon_device pointer
194 * @fence: radeon fence object
196 * Add a DMA fence packet to the ring to write
197 * the fence seq number and DMA trap packet to generate
198 * an interrupt if needed (CIK).
200 void cik_sdma_fence_ring_emit(struct radeon_device *rdev,
201 struct radeon_fence *fence)
203 struct radeon_ring *ring = &rdev->ring[fence->ring];
204 u64 addr = rdev->fence_drv[fence->ring].gpu_addr;
206 /* write the fence */
207 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_FENCE, 0, 0));
208 radeon_ring_write(ring, lower_32_bits(addr));
209 radeon_ring_write(ring, upper_32_bits(addr));
210 radeon_ring_write(ring, fence->seq);
211 /* generate an interrupt */
212 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_TRAP, 0, 0));
214 cik_sdma_hdp_flush_ring_emit(rdev, fence->ring);
218 * cik_sdma_semaphore_ring_emit - emit a semaphore on the dma ring
220 * @rdev: radeon_device pointer
221 * @ring: radeon_ring structure holding ring information
222 * @semaphore: radeon semaphore object
223 * @emit_wait: wait or signal semaphore
225 * Add a DMA semaphore packet to the ring wait on or signal
228 bool cik_sdma_semaphore_ring_emit(struct radeon_device *rdev,
229 struct radeon_ring *ring,
230 struct radeon_semaphore *semaphore,
233 u64 addr = semaphore->gpu_addr;
234 u32 extra_bits = emit_wait ? 0 : SDMA_SEMAPHORE_EXTRA_S;
236 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SEMAPHORE, 0, extra_bits));
237 radeon_ring_write(ring, addr & 0xfffffff8);
238 radeon_ring_write(ring, upper_32_bits(addr));
244 * cik_sdma_gfx_stop - stop the gfx async dma engines
246 * @rdev: radeon_device pointer
248 * Stop the gfx async dma ring buffers (CIK).
250 static void cik_sdma_gfx_stop(struct radeon_device *rdev)
252 u32 rb_cntl, reg_offset;
255 if ((rdev->asic->copy.copy_ring_index == R600_RING_TYPE_DMA_INDEX) ||
256 (rdev->asic->copy.copy_ring_index == CAYMAN_RING_TYPE_DMA1_INDEX))
257 radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
259 for (i = 0; i < 2; i++) {
261 reg_offset = SDMA0_REGISTER_OFFSET;
263 reg_offset = SDMA1_REGISTER_OFFSET;
264 rb_cntl = RREG32(SDMA0_GFX_RB_CNTL + reg_offset);
265 rb_cntl &= ~SDMA_RB_ENABLE;
266 WREG32(SDMA0_GFX_RB_CNTL + reg_offset, rb_cntl);
267 WREG32(SDMA0_GFX_IB_CNTL + reg_offset, 0);
269 rdev->ring[R600_RING_TYPE_DMA_INDEX].ready = false;
270 rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX].ready = false;
274 * cik_sdma_rlc_stop - stop the compute async dma engines
276 * @rdev: radeon_device pointer
278 * Stop the compute async dma queues (CIK).
280 static void cik_sdma_rlc_stop(struct radeon_device *rdev)
286 * cik_sdma_ctx_switch_enable - enable/disable sdma engine preemption
288 * @rdev: radeon_device pointer
289 * @enable: enable/disable preemption.
291 * Halt or unhalt the async dma engines (CIK).
293 static void cik_sdma_ctx_switch_enable(struct radeon_device *rdev, bool enable)
295 uint32_t reg_offset, value;
298 for (i = 0; i < 2; i++) {
300 reg_offset = SDMA0_REGISTER_OFFSET;
302 reg_offset = SDMA1_REGISTER_OFFSET;
303 value = RREG32(SDMA0_CNTL + reg_offset);
305 value |= AUTO_CTXSW_ENABLE;
307 value &= ~AUTO_CTXSW_ENABLE;
308 WREG32(SDMA0_CNTL + reg_offset, value);
313 * cik_sdma_enable - stop the async dma engines
315 * @rdev: radeon_device pointer
316 * @enable: enable/disable the DMA MEs.
318 * Halt or unhalt the async dma engines (CIK).
320 void cik_sdma_enable(struct radeon_device *rdev, bool enable)
322 u32 me_cntl, reg_offset;
325 if (enable == false) {
326 cik_sdma_gfx_stop(rdev);
327 cik_sdma_rlc_stop(rdev);
330 for (i = 0; i < 2; i++) {
332 reg_offset = SDMA0_REGISTER_OFFSET;
334 reg_offset = SDMA1_REGISTER_OFFSET;
335 me_cntl = RREG32(SDMA0_ME_CNTL + reg_offset);
337 me_cntl &= ~SDMA_HALT;
339 me_cntl |= SDMA_HALT;
340 WREG32(SDMA0_ME_CNTL + reg_offset, me_cntl);
343 cik_sdma_ctx_switch_enable(rdev, enable);
347 * cik_sdma_gfx_resume - setup and start the async dma engines
349 * @rdev: radeon_device pointer
351 * Set up the gfx DMA ring buffers and enable them (CIK).
352 * Returns 0 for success, error for failure.
354 static int cik_sdma_gfx_resume(struct radeon_device *rdev)
356 struct radeon_ring *ring;
357 u32 rb_cntl, ib_cntl;
359 u32 reg_offset, wb_offset;
362 for (i = 0; i < 2; i++) {
364 ring = &rdev->ring[R600_RING_TYPE_DMA_INDEX];
365 reg_offset = SDMA0_REGISTER_OFFSET;
366 wb_offset = R600_WB_DMA_RPTR_OFFSET;
368 ring = &rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX];
369 reg_offset = SDMA1_REGISTER_OFFSET;
370 wb_offset = CAYMAN_WB_DMA1_RPTR_OFFSET;
373 WREG32(SDMA0_SEM_INCOMPLETE_TIMER_CNTL + reg_offset, 0);
374 WREG32(SDMA0_SEM_WAIT_FAIL_TIMER_CNTL + reg_offset, 0);
376 /* Set ring buffer size in dwords */
377 rb_bufsz = order_base_2(ring->ring_size / 4);
378 rb_cntl = rb_bufsz << 1;
380 rb_cntl |= SDMA_RB_SWAP_ENABLE | SDMA_RPTR_WRITEBACK_SWAP_ENABLE;
382 WREG32(SDMA0_GFX_RB_CNTL + reg_offset, rb_cntl);
384 /* Initialize the ring buffer's read and write pointers */
385 WREG32(SDMA0_GFX_RB_RPTR + reg_offset, 0);
386 WREG32(SDMA0_GFX_RB_WPTR + reg_offset, 0);
388 /* set the wb address whether it's enabled or not */
389 WREG32(SDMA0_GFX_RB_RPTR_ADDR_HI + reg_offset,
390 upper_32_bits(rdev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF);
391 WREG32(SDMA0_GFX_RB_RPTR_ADDR_LO + reg_offset,
392 ((rdev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC));
394 if (rdev->wb.enabled)
395 rb_cntl |= SDMA_RPTR_WRITEBACK_ENABLE;
397 WREG32(SDMA0_GFX_RB_BASE + reg_offset, ring->gpu_addr >> 8);
398 WREG32(SDMA0_GFX_RB_BASE_HI + reg_offset, ring->gpu_addr >> 40);
401 WREG32(SDMA0_GFX_RB_WPTR + reg_offset, ring->wptr << 2);
404 WREG32(SDMA0_GFX_RB_CNTL + reg_offset, rb_cntl | SDMA_RB_ENABLE);
406 ib_cntl = SDMA_IB_ENABLE;
408 ib_cntl |= SDMA_IB_SWAP_ENABLE;
411 WREG32(SDMA0_GFX_IB_CNTL + reg_offset, ib_cntl);
415 r = radeon_ring_test(rdev, ring->idx, ring);
422 if ((rdev->asic->copy.copy_ring_index == R600_RING_TYPE_DMA_INDEX) ||
423 (rdev->asic->copy.copy_ring_index == CAYMAN_RING_TYPE_DMA1_INDEX))
424 radeon_ttm_set_active_vram_size(rdev, rdev->mc.real_vram_size);
430 * cik_sdma_rlc_resume - setup and start the async dma engines
432 * @rdev: radeon_device pointer
434 * Set up the compute DMA queues and enable them (CIK).
435 * Returns 0 for success, error for failure.
437 static int cik_sdma_rlc_resume(struct radeon_device *rdev)
444 * cik_sdma_load_microcode - load the sDMA ME ucode
446 * @rdev: radeon_device pointer
448 * Loads the sDMA0/1 ucode.
449 * Returns 0 for success, -EINVAL if the ucode is not available.
451 static int cik_sdma_load_microcode(struct radeon_device *rdev)
459 cik_sdma_enable(rdev, false);
462 const struct sdma_firmware_header_v1_0 *hdr =
463 (const struct sdma_firmware_header_v1_0 *)rdev->sdma_fw->data;
464 const __le32 *fw_data;
467 radeon_ucode_print_sdma_hdr(&hdr->header);
470 fw_data = (const __le32 *)
471 (rdev->sdma_fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
472 fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
473 WREG32(SDMA0_UCODE_ADDR + SDMA0_REGISTER_OFFSET, 0);
474 for (i = 0; i < fw_size; i++)
475 WREG32(SDMA0_UCODE_DATA + SDMA0_REGISTER_OFFSET, le32_to_cpup(fw_data++));
476 WREG32(SDMA0_UCODE_DATA + SDMA0_REGISTER_OFFSET, CIK_SDMA_UCODE_VERSION);
479 fw_data = (const __le32 *)
480 (rdev->sdma_fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
481 fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
482 WREG32(SDMA0_UCODE_ADDR + SDMA1_REGISTER_OFFSET, 0);
483 for (i = 0; i < fw_size; i++)
484 WREG32(SDMA0_UCODE_DATA + SDMA1_REGISTER_OFFSET, le32_to_cpup(fw_data++));
485 WREG32(SDMA0_UCODE_DATA + SDMA1_REGISTER_OFFSET, CIK_SDMA_UCODE_VERSION);
487 const __be32 *fw_data;
490 fw_data = (const __be32 *)rdev->sdma_fw->data;
491 WREG32(SDMA0_UCODE_ADDR + SDMA0_REGISTER_OFFSET, 0);
492 for (i = 0; i < CIK_SDMA_UCODE_SIZE; i++)
493 WREG32(SDMA0_UCODE_DATA + SDMA0_REGISTER_OFFSET, be32_to_cpup(fw_data++));
494 WREG32(SDMA0_UCODE_DATA + SDMA0_REGISTER_OFFSET, CIK_SDMA_UCODE_VERSION);
497 fw_data = (const __be32 *)rdev->sdma_fw->data;
498 WREG32(SDMA0_UCODE_ADDR + SDMA1_REGISTER_OFFSET, 0);
499 for (i = 0; i < CIK_SDMA_UCODE_SIZE; i++)
500 WREG32(SDMA0_UCODE_DATA + SDMA1_REGISTER_OFFSET, be32_to_cpup(fw_data++));
501 WREG32(SDMA0_UCODE_DATA + SDMA1_REGISTER_OFFSET, CIK_SDMA_UCODE_VERSION);
504 WREG32(SDMA0_UCODE_ADDR + SDMA0_REGISTER_OFFSET, 0);
505 WREG32(SDMA0_UCODE_ADDR + SDMA1_REGISTER_OFFSET, 0);
510 * cik_sdma_resume - setup and start the async dma engines
512 * @rdev: radeon_device pointer
514 * Set up the DMA engines and enable them (CIK).
515 * Returns 0 for success, error for failure.
517 int cik_sdma_resume(struct radeon_device *rdev)
521 r = cik_sdma_load_microcode(rdev);
526 cik_sdma_enable(rdev, true);
528 /* start the gfx rings and rlc compute queues */
529 r = cik_sdma_gfx_resume(rdev);
532 r = cik_sdma_rlc_resume(rdev);
540 * cik_sdma_fini - tear down the async dma engines
542 * @rdev: radeon_device pointer
544 * Stop the async dma engines and free the rings (CIK).
546 void cik_sdma_fini(struct radeon_device *rdev)
549 cik_sdma_enable(rdev, false);
550 radeon_ring_fini(rdev, &rdev->ring[R600_RING_TYPE_DMA_INDEX]);
551 radeon_ring_fini(rdev, &rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX]);
552 /* XXX - compute dma queue tear down */
556 * cik_copy_dma - copy pages using the DMA engine
558 * @rdev: radeon_device pointer
559 * @src_offset: src GPU address
560 * @dst_offset: dst GPU address
561 * @num_gpu_pages: number of GPU pages to xfer
562 * @resv: reservation object to sync to
564 * Copy GPU paging using the DMA engine (CIK).
565 * Used by the radeon ttm implementation to move pages if
566 * registered as the asic copy callback.
568 struct radeon_fence *cik_copy_dma(struct radeon_device *rdev,
569 uint64_t src_offset, uint64_t dst_offset,
570 unsigned num_gpu_pages,
571 struct reservation_object *resv)
573 struct radeon_fence *fence;
574 struct radeon_sync sync;
575 int ring_index = rdev->asic->copy.dma_ring_index;
576 struct radeon_ring *ring = &rdev->ring[ring_index];
577 u32 size_in_bytes, cur_size_in_bytes;
581 radeon_sync_create(&sync);
583 size_in_bytes = (num_gpu_pages << RADEON_GPU_PAGE_SHIFT);
584 num_loops = DIV_ROUND_UP(size_in_bytes, 0x1fffff);
585 r = radeon_ring_lock(rdev, ring, num_loops * 7 + 14);
587 DRM_ERROR("radeon: moving bo (%d).\n", r);
588 radeon_sync_free(rdev, &sync, NULL);
592 radeon_sync_resv(rdev, &sync, resv, false);
593 radeon_sync_rings(rdev, &sync, ring->idx);
595 for (i = 0; i < num_loops; i++) {
596 cur_size_in_bytes = size_in_bytes;
597 if (cur_size_in_bytes > 0x1fffff)
598 cur_size_in_bytes = 0x1fffff;
599 size_in_bytes -= cur_size_in_bytes;
600 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_COPY, SDMA_COPY_SUB_OPCODE_LINEAR, 0));
601 radeon_ring_write(ring, cur_size_in_bytes);
602 radeon_ring_write(ring, 0); /* src/dst endian swap */
603 radeon_ring_write(ring, lower_32_bits(src_offset));
604 radeon_ring_write(ring, upper_32_bits(src_offset));
605 radeon_ring_write(ring, lower_32_bits(dst_offset));
606 radeon_ring_write(ring, upper_32_bits(dst_offset));
607 src_offset += cur_size_in_bytes;
608 dst_offset += cur_size_in_bytes;
611 r = radeon_fence_emit(rdev, &fence, ring->idx);
613 radeon_ring_unlock_undo(rdev, ring);
614 radeon_sync_free(rdev, &sync, NULL);
618 radeon_ring_unlock_commit(rdev, ring, false);
619 radeon_sync_free(rdev, &sync, fence);
625 * cik_sdma_ring_test - simple async dma engine test
627 * @rdev: radeon_device pointer
628 * @ring: radeon_ring structure holding ring information
630 * Test the DMA engine by writing using it to write an
631 * value to memory. (CIK).
632 * Returns 0 for success, error for failure.
634 int cik_sdma_ring_test(struct radeon_device *rdev,
635 struct radeon_ring *ring)
643 if (ring->idx == R600_RING_TYPE_DMA_INDEX)
644 index = R600_WB_DMA_RING_TEST_OFFSET;
646 index = CAYMAN_WB_DMA1_RING_TEST_OFFSET;
648 gpu_addr = rdev->wb.gpu_addr + index;
651 rdev->wb.wb[index/4] = cpu_to_le32(tmp);
653 r = radeon_ring_lock(rdev, ring, 5);
655 DRM_ERROR("radeon: dma failed to lock ring %d (%d).\n", ring->idx, r);
658 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_WRITE, SDMA_WRITE_SUB_OPCODE_LINEAR, 0));
659 radeon_ring_write(ring, lower_32_bits(gpu_addr));
660 radeon_ring_write(ring, upper_32_bits(gpu_addr));
661 radeon_ring_write(ring, 1); /* number of DWs to follow */
662 radeon_ring_write(ring, 0xDEADBEEF);
663 radeon_ring_unlock_commit(rdev, ring, false);
665 for (i = 0; i < rdev->usec_timeout; i++) {
666 tmp = le32_to_cpu(rdev->wb.wb[index/4]);
667 if (tmp == 0xDEADBEEF)
672 if (i < rdev->usec_timeout) {
673 DRM_INFO("ring test on %d succeeded in %d usecs\n", ring->idx, i);
675 DRM_ERROR("radeon: ring %d test failed (0x%08X)\n",
683 * cik_sdma_ib_test - test an IB on the DMA engine
685 * @rdev: radeon_device pointer
686 * @ring: radeon_ring structure holding ring information
688 * Test a simple IB in the DMA ring (CIK).
689 * Returns 0 on success, error on failure.
691 int cik_sdma_ib_test(struct radeon_device *rdev, struct radeon_ring *ring)
700 if (ring->idx == R600_RING_TYPE_DMA_INDEX)
701 index = R600_WB_DMA_RING_TEST_OFFSET;
703 index = CAYMAN_WB_DMA1_RING_TEST_OFFSET;
705 gpu_addr = rdev->wb.gpu_addr + index;
708 rdev->wb.wb[index/4] = cpu_to_le32(tmp);
710 r = radeon_ib_get(rdev, ring->idx, &ib, NULL, 256);
712 DRM_ERROR("radeon: failed to get ib (%d).\n", r);
716 ib.ptr[0] = SDMA_PACKET(SDMA_OPCODE_WRITE, SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
717 ib.ptr[1] = lower_32_bits(gpu_addr);
718 ib.ptr[2] = upper_32_bits(gpu_addr);
720 ib.ptr[4] = 0xDEADBEEF;
723 r = radeon_ib_schedule(rdev, &ib, NULL, false);
725 radeon_ib_free(rdev, &ib);
726 DRM_ERROR("radeon: failed to schedule ib (%d).\n", r);
729 r = radeon_fence_wait(ib.fence, false);
731 DRM_ERROR("radeon: fence wait failed (%d).\n", r);
734 for (i = 0; i < rdev->usec_timeout; i++) {
735 tmp = le32_to_cpu(rdev->wb.wb[index/4]);
736 if (tmp == 0xDEADBEEF)
740 if (i < rdev->usec_timeout) {
741 DRM_INFO("ib test on ring %d succeeded in %u usecs\n", ib.fence->ring, i);
743 DRM_ERROR("radeon: ib test failed (0x%08X)\n", tmp);
746 radeon_ib_free(rdev, &ib);
751 * cik_sdma_is_lockup - Check if the DMA engine is locked up
753 * @rdev: radeon_device pointer
754 * @ring: radeon_ring structure holding ring information
756 * Check if the async DMA engine is locked up (CIK).
757 * Returns true if the engine appears to be locked up, false if not.
759 bool cik_sdma_is_lockup(struct radeon_device *rdev, struct radeon_ring *ring)
761 u32 reset_mask = cik_gpu_check_soft_reset(rdev);
764 if (ring->idx == R600_RING_TYPE_DMA_INDEX)
765 mask = RADEON_RESET_DMA;
767 mask = RADEON_RESET_DMA1;
769 if (!(reset_mask & mask)) {
770 radeon_ring_lockup_update(rdev, ring);
773 return radeon_ring_test_lockup(rdev, ring);
777 * cik_sdma_vm_copy_pages - update PTEs by copying them from the GART
779 * @rdev: radeon_device pointer
780 * @ib: indirect buffer to fill with commands
781 * @pe: addr of the page entry
782 * @src: src addr to copy from
783 * @count: number of page entries to update
785 * Update PTEs by copying them from the GART using sDMA (CIK).
787 void cik_sdma_vm_copy_pages(struct radeon_device *rdev,
788 struct radeon_ib *ib,
789 uint64_t pe, uint64_t src,
793 unsigned bytes = count * 8;
794 if (bytes > 0x1FFFF8)
797 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_COPY,
798 SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
799 ib->ptr[ib->length_dw++] = bytes;
800 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
801 ib->ptr[ib->length_dw++] = lower_32_bits(src);
802 ib->ptr[ib->length_dw++] = upper_32_bits(src);
803 ib->ptr[ib->length_dw++] = lower_32_bits(pe);
804 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
813 * cik_sdma_vm_write_pages - update PTEs by writing them manually
815 * @rdev: radeon_device pointer
816 * @ib: indirect buffer to fill with commands
817 * @pe: addr of the page entry
818 * @addr: dst addr to write into pe
819 * @count: number of page entries to update
820 * @incr: increase next addr by incr bytes
821 * @flags: access flags
823 * Update PTEs by writing them manually using sDMA (CIK).
825 void cik_sdma_vm_write_pages(struct radeon_device *rdev,
826 struct radeon_ib *ib,
828 uint64_t addr, unsigned count,
829 uint32_t incr, uint32_t flags)
839 /* for non-physically contiguous pages (system) */
840 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_WRITE,
841 SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
842 ib->ptr[ib->length_dw++] = pe;
843 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
844 ib->ptr[ib->length_dw++] = ndw;
845 for (; ndw > 0; ndw -= 2, --count, pe += 8) {
846 if (flags & R600_PTE_SYSTEM) {
847 value = radeon_vm_map_gart(rdev, addr);
848 } else if (flags & R600_PTE_VALID) {
855 ib->ptr[ib->length_dw++] = value;
856 ib->ptr[ib->length_dw++] = upper_32_bits(value);
862 * cik_sdma_vm_set_pages - update the page tables using sDMA
864 * @rdev: radeon_device pointer
865 * @ib: indirect buffer to fill with commands
866 * @pe: addr of the page entry
867 * @addr: dst addr to write into pe
868 * @count: number of page entries to update
869 * @incr: increase next addr by incr bytes
870 * @flags: access flags
872 * Update the page tables using sDMA (CIK).
874 void cik_sdma_vm_set_pages(struct radeon_device *rdev,
875 struct radeon_ib *ib,
877 uint64_t addr, unsigned count,
878 uint32_t incr, uint32_t flags)
888 if (flags & R600_PTE_VALID)
893 /* for physically contiguous pages (vram) */
894 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_GENERATE_PTE_PDE, 0, 0);
895 ib->ptr[ib->length_dw++] = pe; /* dst addr */
896 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
897 ib->ptr[ib->length_dw++] = flags; /* mask */
898 ib->ptr[ib->length_dw++] = 0;
899 ib->ptr[ib->length_dw++] = value; /* value */
900 ib->ptr[ib->length_dw++] = upper_32_bits(value);
901 ib->ptr[ib->length_dw++] = incr; /* increment size */
902 ib->ptr[ib->length_dw++] = 0;
903 ib->ptr[ib->length_dw++] = ndw; /* number of entries */
912 * cik_sdma_vm_pad_ib - pad the IB to the required number of dw
914 * @ib: indirect buffer to fill with padding
917 void cik_sdma_vm_pad_ib(struct radeon_ib *ib)
919 while (ib->length_dw & 0x7)
920 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0);
924 * cik_dma_vm_flush - cik vm flush using sDMA
926 * @rdev: radeon_device pointer
928 * Update the page table base and flush the VM TLB
931 void cik_dma_vm_flush(struct radeon_device *rdev, struct radeon_ring *ring,
932 unsigned vm_id, uint64_t pd_addr)
934 u32 extra_bits = (SDMA_POLL_REG_MEM_EXTRA_OP(0) |
935 SDMA_POLL_REG_MEM_EXTRA_FUNC(0)); /* always */
937 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
939 radeon_ring_write(ring, (VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (vm_id << 2)) >> 2);
941 radeon_ring_write(ring, (VM_CONTEXT8_PAGE_TABLE_BASE_ADDR + ((vm_id - 8) << 2)) >> 2);
943 radeon_ring_write(ring, pd_addr >> 12);
945 /* update SH_MEM_* regs */
946 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
947 radeon_ring_write(ring, SRBM_GFX_CNTL >> 2);
948 radeon_ring_write(ring, VMID(vm_id));
950 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
951 radeon_ring_write(ring, SH_MEM_BASES >> 2);
952 radeon_ring_write(ring, 0);
954 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
955 radeon_ring_write(ring, SH_MEM_CONFIG >> 2);
956 radeon_ring_write(ring, 0);
958 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
959 radeon_ring_write(ring, SH_MEM_APE1_BASE >> 2);
960 radeon_ring_write(ring, 1);
962 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
963 radeon_ring_write(ring, SH_MEM_APE1_LIMIT >> 2);
964 radeon_ring_write(ring, 0);
966 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
967 radeon_ring_write(ring, SRBM_GFX_CNTL >> 2);
968 radeon_ring_write(ring, VMID(0));
971 cik_sdma_hdp_flush_ring_emit(rdev, ring->idx);
974 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
975 radeon_ring_write(ring, VM_INVALIDATE_REQUEST >> 2);
976 radeon_ring_write(ring, 1 << vm_id);
978 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0, extra_bits));
979 radeon_ring_write(ring, VM_INVALIDATE_REQUEST >> 2);
980 radeon_ring_write(ring, 0);
981 radeon_ring_write(ring, 0); /* reference */
982 radeon_ring_write(ring, 0); /* mask */
983 radeon_ring_write(ring, (0xfff << 16) | 10); /* retry count, poll interval */