1 /*
2 * Copyright 2013 Advanced Micro Devices, Inc.
3 *
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:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
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.
21 *
22 * Authors: Alex Deucher
23 */
24 #include <linux/firmware.h>
25
26 #include "radeon.h"
27 #include "radeon_ucode.h"
28 #include "radeon_asic.h"
29 #include "radeon_trace.h"
30 #include "cik.h"
31 #include "cikd.h"
32
33 /* sdma */
34 #define CIK_SDMA_UCODE_SIZE 1050
35 #define CIK_SDMA_UCODE_VERSION 64
36
37 /*
38 * sDMA - System DMA
39 * Starting with CIK, the GPU has new asynchronous
40 * DMA engines. These engines are used for compute
41 * and gfx. There are two DMA engines (SDMA0, SDMA1)
42 * and each one supports 1 ring buffer used for gfx
43 * and 2 queues used for compute.
44 *
45 * The programming model is very similar to the CP
46 * (ring buffer, IBs, etc.), but sDMA has it's own
47 * packet format that is different from the PM4 format
48 * used by the CP. sDMA supports copying data, writing
49 * embedded data, solid fills, and a number of other
50 * things. It also has support for tiling/detiling of
51 * buffers.
52 */
53
54 /**
55 * cik_sdma_get_rptr - get the current read pointer
56 *
57 * @rdev: radeon_device pointer
58 * @ring: radeon ring pointer
59 *
60 * Get the current rptr from the hardware (CIK+).
61 */
cik_sdma_get_rptr(struct radeon_device * rdev,struct radeon_ring * ring)62 uint32_t cik_sdma_get_rptr(struct radeon_device *rdev,
63 struct radeon_ring *ring)
64 {
65 u32 rptr, reg;
66
67 if (rdev->wb.enabled) {
68 rptr = rdev->wb.wb[ring->rptr_offs/4];
69 } else {
70 if (ring->idx == R600_RING_TYPE_DMA_INDEX)
71 reg = SDMA0_GFX_RB_RPTR + SDMA0_REGISTER_OFFSET;
72 else
73 reg = SDMA0_GFX_RB_RPTR + SDMA1_REGISTER_OFFSET;
74
75 rptr = RREG32(reg);
76 }
77
78 return (rptr & 0x3fffc) >> 2;
79 }
80
81 /**
82 * cik_sdma_get_wptr - get the current write pointer
83 *
84 * @rdev: radeon_device pointer
85 * @ring: radeon ring pointer
86 *
87 * Get the current wptr from the hardware (CIK+).
88 */
cik_sdma_get_wptr(struct radeon_device * rdev,struct radeon_ring * ring)89 uint32_t cik_sdma_get_wptr(struct radeon_device *rdev,
90 struct radeon_ring *ring)
91 {
92 u32 reg;
93
94 if (ring->idx == R600_RING_TYPE_DMA_INDEX)
95 reg = SDMA0_GFX_RB_WPTR + SDMA0_REGISTER_OFFSET;
96 else
97 reg = SDMA0_GFX_RB_WPTR + SDMA1_REGISTER_OFFSET;
98
99 return (RREG32(reg) & 0x3fffc) >> 2;
100 }
101
102 /**
103 * cik_sdma_set_wptr - commit the write pointer
104 *
105 * @rdev: radeon_device pointer
106 * @ring: radeon ring pointer
107 *
108 * Write the wptr back to the hardware (CIK+).
109 */
cik_sdma_set_wptr(struct radeon_device * rdev,struct radeon_ring * ring)110 void cik_sdma_set_wptr(struct radeon_device *rdev,
111 struct radeon_ring *ring)
112 {
113 u32 reg;
114
115 if (ring->idx == R600_RING_TYPE_DMA_INDEX)
116 reg = SDMA0_GFX_RB_WPTR + SDMA0_REGISTER_OFFSET;
117 else
118 reg = SDMA0_GFX_RB_WPTR + SDMA1_REGISTER_OFFSET;
119
120 WREG32(reg, (ring->wptr << 2) & 0x3fffc);
121 (void)RREG32(reg);
122 }
123
124 /**
125 * cik_sdma_ring_ib_execute - Schedule an IB on the DMA engine
126 *
127 * @rdev: radeon_device pointer
128 * @ib: IB object to schedule
129 *
130 * Schedule an IB in the DMA ring (CIK).
131 */
cik_sdma_ring_ib_execute(struct radeon_device * rdev,struct radeon_ib * ib)132 void cik_sdma_ring_ib_execute(struct radeon_device *rdev,
133 struct radeon_ib *ib)
134 {
135 struct radeon_ring *ring = &rdev->ring[ib->ring];
136 u32 extra_bits = (ib->vm ? ib->vm->ids[ib->ring].id : 0) & 0xf;
137
138 if (rdev->wb.enabled) {
139 u32 next_rptr = ring->wptr + 5;
140 while ((next_rptr & 7) != 4)
141 next_rptr++;
142 next_rptr += 4;
143 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_WRITE, SDMA_WRITE_SUB_OPCODE_LINEAR, 0));
144 radeon_ring_write(ring, ring->next_rptr_gpu_addr & 0xfffffffc);
145 radeon_ring_write(ring, upper_32_bits(ring->next_rptr_gpu_addr));
146 radeon_ring_write(ring, 1); /* number of DWs to follow */
147 radeon_ring_write(ring, next_rptr);
148 }
149
150 /* IB packet must end on a 8 DW boundary */
151 while ((ring->wptr & 7) != 4)
152 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0));
153 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_INDIRECT_BUFFER, 0, extra_bits));
154 radeon_ring_write(ring, ib->gpu_addr & 0xffffffe0); /* base must be 32 byte aligned */
155 radeon_ring_write(ring, upper_32_bits(ib->gpu_addr));
156 radeon_ring_write(ring, ib->length_dw);
157
158 }
159
160 /**
161 * cik_sdma_hdp_flush_ring_emit - emit an hdp flush on the DMA ring
162 *
163 * @rdev: radeon_device pointer
164 * @ridx: radeon ring index
165 *
166 * Emit an hdp flush packet on the requested DMA ring.
167 */
cik_sdma_hdp_flush_ring_emit(struct radeon_device * rdev,int ridx)168 static void cik_sdma_hdp_flush_ring_emit(struct radeon_device *rdev,
169 int ridx)
170 {
171 struct radeon_ring *ring = &rdev->ring[ridx];
172 u32 extra_bits = (SDMA_POLL_REG_MEM_EXTRA_OP(1) |
173 SDMA_POLL_REG_MEM_EXTRA_FUNC(3)); /* == */
174 u32 ref_and_mask;
175
176 if (ridx == R600_RING_TYPE_DMA_INDEX)
177 ref_and_mask = SDMA0;
178 else
179 ref_and_mask = SDMA1;
180
181 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0, extra_bits));
182 radeon_ring_write(ring, GPU_HDP_FLUSH_DONE);
183 radeon_ring_write(ring, GPU_HDP_FLUSH_REQ);
184 radeon_ring_write(ring, ref_and_mask); /* reference */
185 radeon_ring_write(ring, ref_and_mask); /* mask */
186 radeon_ring_write(ring, (0xfff << 16) | 10); /* retry count, poll interval */
187 }
188
189 /**
190 * cik_sdma_fence_ring_emit - emit a fence on the DMA ring
191 *
192 * @rdev: radeon_device pointer
193 * @fence: radeon fence object
194 *
195 * Add a DMA fence packet to the ring to write
196 * the fence seq number and DMA trap packet to generate
197 * an interrupt if needed (CIK).
198 */
cik_sdma_fence_ring_emit(struct radeon_device * rdev,struct radeon_fence * fence)199 void cik_sdma_fence_ring_emit(struct radeon_device *rdev,
200 struct radeon_fence *fence)
201 {
202 struct radeon_ring *ring = &rdev->ring[fence->ring];
203 u64 addr = rdev->fence_drv[fence->ring].gpu_addr;
204
205 /* write the fence */
206 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_FENCE, 0, 0));
207 radeon_ring_write(ring, lower_32_bits(addr));
208 radeon_ring_write(ring, upper_32_bits(addr));
209 radeon_ring_write(ring, fence->seq);
210 /* generate an interrupt */
211 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_TRAP, 0, 0));
212 /* flush HDP */
213 cik_sdma_hdp_flush_ring_emit(rdev, fence->ring);
214 }
215
216 /**
217 * cik_sdma_semaphore_ring_emit - emit a semaphore on the dma ring
218 *
219 * @rdev: radeon_device pointer
220 * @ring: radeon_ring structure holding ring information
221 * @semaphore: radeon semaphore object
222 * @emit_wait: wait or signal semaphore
223 *
224 * Add a DMA semaphore packet to the ring wait on or signal
225 * other rings (CIK).
226 */
cik_sdma_semaphore_ring_emit(struct radeon_device * rdev,struct radeon_ring * ring,struct radeon_semaphore * semaphore,bool emit_wait)227 bool cik_sdma_semaphore_ring_emit(struct radeon_device *rdev,
228 struct radeon_ring *ring,
229 struct radeon_semaphore *semaphore,
230 bool emit_wait)
231 {
232 u64 addr = semaphore->gpu_addr;
233 u32 extra_bits = emit_wait ? 0 : SDMA_SEMAPHORE_EXTRA_S;
234
235 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SEMAPHORE, 0, extra_bits));
236 radeon_ring_write(ring, addr & 0xfffffff8);
237 radeon_ring_write(ring, upper_32_bits(addr));
238
239 return true;
240 }
241
242 /**
243 * cik_sdma_gfx_stop - stop the gfx async dma engines
244 *
245 * @rdev: radeon_device pointer
246 *
247 * Stop the gfx async dma ring buffers (CIK).
248 */
cik_sdma_gfx_stop(struct radeon_device * rdev)249 static void cik_sdma_gfx_stop(struct radeon_device *rdev)
250 {
251 u32 rb_cntl, reg_offset;
252 int i;
253
254 if ((rdev->asic->copy.copy_ring_index == R600_RING_TYPE_DMA_INDEX) ||
255 (rdev->asic->copy.copy_ring_index == CAYMAN_RING_TYPE_DMA1_INDEX))
256 radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
257
258 for (i = 0; i < 2; i++) {
259 if (i == 0)
260 reg_offset = SDMA0_REGISTER_OFFSET;
261 else
262 reg_offset = SDMA1_REGISTER_OFFSET;
263 rb_cntl = RREG32(SDMA0_GFX_RB_CNTL + reg_offset);
264 rb_cntl &= ~SDMA_RB_ENABLE;
265 WREG32(SDMA0_GFX_RB_CNTL + reg_offset, rb_cntl);
266 WREG32(SDMA0_GFX_IB_CNTL + reg_offset, 0);
267 }
268 rdev->ring[R600_RING_TYPE_DMA_INDEX].ready = false;
269 rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX].ready = false;
270
271 /* FIXME use something else than big hammer but after few days can not
272 * seem to find good combination so reset SDMA blocks as it seems we
273 * do not shut them down properly. This fix hibernation and does not
274 * affect suspend to ram.
275 */
276 WREG32(SRBM_SOFT_RESET, SOFT_RESET_SDMA | SOFT_RESET_SDMA1);
277 (void)RREG32(SRBM_SOFT_RESET);
278 udelay(50);
279 WREG32(SRBM_SOFT_RESET, 0);
280 (void)RREG32(SRBM_SOFT_RESET);
281 }
282
283 /**
284 * cik_sdma_rlc_stop - stop the compute async dma engines
285 *
286 * @rdev: radeon_device pointer
287 *
288 * Stop the compute async dma queues (CIK).
289 */
cik_sdma_rlc_stop(struct radeon_device * rdev)290 static void cik_sdma_rlc_stop(struct radeon_device *rdev)
291 {
292 /* XXX todo */
293 }
294
295 /**
296 * cik_sdma_ctx_switch_enable - enable/disable sdma engine preemption
297 *
298 * @rdev: radeon_device pointer
299 * @enable: enable/disable preemption.
300 *
301 * Halt or unhalt the async dma engines (CIK).
302 */
cik_sdma_ctx_switch_enable(struct radeon_device * rdev,bool enable)303 static void cik_sdma_ctx_switch_enable(struct radeon_device *rdev, bool enable)
304 {
305 uint32_t reg_offset, value;
306 int i;
307
308 for (i = 0; i < 2; i++) {
309 if (i == 0)
310 reg_offset = SDMA0_REGISTER_OFFSET;
311 else
312 reg_offset = SDMA1_REGISTER_OFFSET;
313 value = RREG32(SDMA0_CNTL + reg_offset);
314 if (enable)
315 value |= AUTO_CTXSW_ENABLE;
316 else
317 value &= ~AUTO_CTXSW_ENABLE;
318 WREG32(SDMA0_CNTL + reg_offset, value);
319 }
320 }
321
322 /**
323 * cik_sdma_enable - stop the async dma engines
324 *
325 * @rdev: radeon_device pointer
326 * @enable: enable/disable the DMA MEs.
327 *
328 * Halt or unhalt the async dma engines (CIK).
329 */
cik_sdma_enable(struct radeon_device * rdev,bool enable)330 void cik_sdma_enable(struct radeon_device *rdev, bool enable)
331 {
332 u32 me_cntl, reg_offset;
333 int i;
334
335 if (!enable) {
336 cik_sdma_gfx_stop(rdev);
337 cik_sdma_rlc_stop(rdev);
338 }
339
340 for (i = 0; i < 2; i++) {
341 if (i == 0)
342 reg_offset = SDMA0_REGISTER_OFFSET;
343 else
344 reg_offset = SDMA1_REGISTER_OFFSET;
345 me_cntl = RREG32(SDMA0_ME_CNTL + reg_offset);
346 if (enable)
347 me_cntl &= ~SDMA_HALT;
348 else
349 me_cntl |= SDMA_HALT;
350 WREG32(SDMA0_ME_CNTL + reg_offset, me_cntl);
351 }
352
353 cik_sdma_ctx_switch_enable(rdev, enable);
354 }
355
356 /**
357 * cik_sdma_gfx_resume - setup and start the async dma engines
358 *
359 * @rdev: radeon_device pointer
360 *
361 * Set up the gfx DMA ring buffers and enable them (CIK).
362 * Returns 0 for success, error for failure.
363 */
cik_sdma_gfx_resume(struct radeon_device * rdev)364 static int cik_sdma_gfx_resume(struct radeon_device *rdev)
365 {
366 struct radeon_ring *ring;
367 u32 rb_cntl, ib_cntl;
368 u32 rb_bufsz;
369 u32 reg_offset, wb_offset;
370 int i, r;
371
372 for (i = 0; i < 2; i++) {
373 if (i == 0) {
374 ring = &rdev->ring[R600_RING_TYPE_DMA_INDEX];
375 reg_offset = SDMA0_REGISTER_OFFSET;
376 wb_offset = R600_WB_DMA_RPTR_OFFSET;
377 } else {
378 ring = &rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX];
379 reg_offset = SDMA1_REGISTER_OFFSET;
380 wb_offset = CAYMAN_WB_DMA1_RPTR_OFFSET;
381 }
382
383 WREG32(SDMA0_SEM_INCOMPLETE_TIMER_CNTL + reg_offset, 0);
384 WREG32(SDMA0_SEM_WAIT_FAIL_TIMER_CNTL + reg_offset, 0);
385
386 /* Set ring buffer size in dwords */
387 rb_bufsz = order_base_2(ring->ring_size / 4);
388 rb_cntl = rb_bufsz << 1;
389 #ifdef __BIG_ENDIAN
390 rb_cntl |= SDMA_RB_SWAP_ENABLE | SDMA_RPTR_WRITEBACK_SWAP_ENABLE;
391 #endif
392 WREG32(SDMA0_GFX_RB_CNTL + reg_offset, rb_cntl);
393
394 /* Initialize the ring buffer's read and write pointers */
395 WREG32(SDMA0_GFX_RB_RPTR + reg_offset, 0);
396 WREG32(SDMA0_GFX_RB_WPTR + reg_offset, 0);
397
398 /* set the wb address whether it's enabled or not */
399 WREG32(SDMA0_GFX_RB_RPTR_ADDR_HI + reg_offset,
400 upper_32_bits(rdev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF);
401 WREG32(SDMA0_GFX_RB_RPTR_ADDR_LO + reg_offset,
402 ((rdev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC));
403
404 if (rdev->wb.enabled)
405 rb_cntl |= SDMA_RPTR_WRITEBACK_ENABLE;
406
407 WREG32(SDMA0_GFX_RB_BASE + reg_offset, ring->gpu_addr >> 8);
408 WREG32(SDMA0_GFX_RB_BASE_HI + reg_offset, ring->gpu_addr >> 40);
409
410 ring->wptr = 0;
411 WREG32(SDMA0_GFX_RB_WPTR + reg_offset, ring->wptr << 2);
412
413 /* enable DMA RB */
414 WREG32(SDMA0_GFX_RB_CNTL + reg_offset, rb_cntl | SDMA_RB_ENABLE);
415
416 ib_cntl = SDMA_IB_ENABLE;
417 #ifdef __BIG_ENDIAN
418 ib_cntl |= SDMA_IB_SWAP_ENABLE;
419 #endif
420 /* enable DMA IBs */
421 WREG32(SDMA0_GFX_IB_CNTL + reg_offset, ib_cntl);
422
423 ring->ready = true;
424
425 r = radeon_ring_test(rdev, ring->idx, ring);
426 if (r) {
427 ring->ready = false;
428 return r;
429 }
430 }
431
432 if ((rdev->asic->copy.copy_ring_index == R600_RING_TYPE_DMA_INDEX) ||
433 (rdev->asic->copy.copy_ring_index == CAYMAN_RING_TYPE_DMA1_INDEX))
434 radeon_ttm_set_active_vram_size(rdev, rdev->mc.real_vram_size);
435
436 return 0;
437 }
438
439 /**
440 * cik_sdma_rlc_resume - setup and start the async dma engines
441 *
442 * @rdev: radeon_device pointer
443 *
444 * Set up the compute DMA queues and enable them (CIK).
445 * Returns 0 for success, error for failure.
446 */
cik_sdma_rlc_resume(struct radeon_device * rdev)447 static int cik_sdma_rlc_resume(struct radeon_device *rdev)
448 {
449 /* XXX todo */
450 return 0;
451 }
452
453 /**
454 * cik_sdma_load_microcode - load the sDMA ME ucode
455 *
456 * @rdev: radeon_device pointer
457 *
458 * Loads the sDMA0/1 ucode.
459 * Returns 0 for success, -EINVAL if the ucode is not available.
460 */
cik_sdma_load_microcode(struct radeon_device * rdev)461 static int cik_sdma_load_microcode(struct radeon_device *rdev)
462 {
463 int i;
464
465 if (!rdev->sdma_fw)
466 return -EINVAL;
467
468 /* halt the MEs */
469 cik_sdma_enable(rdev, false);
470
471 if (rdev->new_fw) {
472 const struct sdma_firmware_header_v1_0 *hdr =
473 (const struct sdma_firmware_header_v1_0 *)rdev->sdma_fw->data;
474 const __le32 *fw_data;
475 u32 fw_size;
476
477 radeon_ucode_print_sdma_hdr(&hdr->header);
478
479 /* sdma0 */
480 fw_data = (const __le32 *)
481 (rdev->sdma_fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
482 fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
483 WREG32(SDMA0_UCODE_ADDR + SDMA0_REGISTER_OFFSET, 0);
484 for (i = 0; i < fw_size; i++)
485 WREG32(SDMA0_UCODE_DATA + SDMA0_REGISTER_OFFSET, le32_to_cpup(fw_data++));
486 WREG32(SDMA0_UCODE_DATA + SDMA0_REGISTER_OFFSET, CIK_SDMA_UCODE_VERSION);
487
488 /* sdma1 */
489 fw_data = (const __le32 *)
490 (rdev->sdma_fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
491 fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
492 WREG32(SDMA0_UCODE_ADDR + SDMA1_REGISTER_OFFSET, 0);
493 for (i = 0; i < fw_size; i++)
494 WREG32(SDMA0_UCODE_DATA + SDMA1_REGISTER_OFFSET, le32_to_cpup(fw_data++));
495 WREG32(SDMA0_UCODE_DATA + SDMA1_REGISTER_OFFSET, CIK_SDMA_UCODE_VERSION);
496 } else {
497 const __be32 *fw_data;
498
499 /* sdma0 */
500 fw_data = (const __be32 *)rdev->sdma_fw->data;
501 WREG32(SDMA0_UCODE_ADDR + SDMA0_REGISTER_OFFSET, 0);
502 for (i = 0; i < CIK_SDMA_UCODE_SIZE; i++)
503 WREG32(SDMA0_UCODE_DATA + SDMA0_REGISTER_OFFSET, be32_to_cpup(fw_data++));
504 WREG32(SDMA0_UCODE_DATA + SDMA0_REGISTER_OFFSET, CIK_SDMA_UCODE_VERSION);
505
506 /* sdma1 */
507 fw_data = (const __be32 *)rdev->sdma_fw->data;
508 WREG32(SDMA0_UCODE_ADDR + SDMA1_REGISTER_OFFSET, 0);
509 for (i = 0; i < CIK_SDMA_UCODE_SIZE; i++)
510 WREG32(SDMA0_UCODE_DATA + SDMA1_REGISTER_OFFSET, be32_to_cpup(fw_data++));
511 WREG32(SDMA0_UCODE_DATA + SDMA1_REGISTER_OFFSET, CIK_SDMA_UCODE_VERSION);
512 }
513
514 WREG32(SDMA0_UCODE_ADDR + SDMA0_REGISTER_OFFSET, 0);
515 WREG32(SDMA0_UCODE_ADDR + SDMA1_REGISTER_OFFSET, 0);
516 return 0;
517 }
518
519 /**
520 * cik_sdma_resume - setup and start the async dma engines
521 *
522 * @rdev: radeon_device pointer
523 *
524 * Set up the DMA engines and enable them (CIK).
525 * Returns 0 for success, error for failure.
526 */
cik_sdma_resume(struct radeon_device * rdev)527 int cik_sdma_resume(struct radeon_device *rdev)
528 {
529 int r;
530
531 r = cik_sdma_load_microcode(rdev);
532 if (r)
533 return r;
534
535 /* unhalt the MEs */
536 cik_sdma_enable(rdev, true);
537
538 /* start the gfx rings and rlc compute queues */
539 r = cik_sdma_gfx_resume(rdev);
540 if (r)
541 return r;
542 r = cik_sdma_rlc_resume(rdev);
543 if (r)
544 return r;
545
546 return 0;
547 }
548
549 /**
550 * cik_sdma_fini - tear down the async dma engines
551 *
552 * @rdev: radeon_device pointer
553 *
554 * Stop the async dma engines and free the rings (CIK).
555 */
cik_sdma_fini(struct radeon_device * rdev)556 void cik_sdma_fini(struct radeon_device *rdev)
557 {
558 /* halt the MEs */
559 cik_sdma_enable(rdev, false);
560 radeon_ring_fini(rdev, &rdev->ring[R600_RING_TYPE_DMA_INDEX]);
561 radeon_ring_fini(rdev, &rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX]);
562 /* XXX - compute dma queue tear down */
563 }
564
565 /**
566 * cik_copy_dma - copy pages using the DMA engine
567 *
568 * @rdev: radeon_device pointer
569 * @src_offset: src GPU address
570 * @dst_offset: dst GPU address
571 * @num_gpu_pages: number of GPU pages to xfer
572 * @resv: reservation object to sync to
573 *
574 * Copy GPU paging using the DMA engine (CIK).
575 * Used by the radeon ttm implementation to move pages if
576 * registered as the asic copy callback.
577 */
cik_copy_dma(struct radeon_device * rdev,uint64_t src_offset,uint64_t dst_offset,unsigned num_gpu_pages,struct dma_resv * resv)578 struct radeon_fence *cik_copy_dma(struct radeon_device *rdev,
579 uint64_t src_offset, uint64_t dst_offset,
580 unsigned num_gpu_pages,
581 struct dma_resv *resv)
582 {
583 struct radeon_fence *fence;
584 struct radeon_sync sync;
585 int ring_index = rdev->asic->copy.dma_ring_index;
586 struct radeon_ring *ring = &rdev->ring[ring_index];
587 u32 size_in_bytes, cur_size_in_bytes;
588 int i, num_loops;
589 int r = 0;
590
591 radeon_sync_create(&sync);
592
593 size_in_bytes = (num_gpu_pages << RADEON_GPU_PAGE_SHIFT);
594 num_loops = DIV_ROUND_UP(size_in_bytes, 0x1fffff);
595 r = radeon_ring_lock(rdev, ring, num_loops * 7 + 14);
596 if (r) {
597 DRM_ERROR("radeon: moving bo (%d).\n", r);
598 radeon_sync_free(rdev, &sync, NULL);
599 return ERR_PTR(r);
600 }
601
602 radeon_sync_resv(rdev, &sync, resv, false);
603 radeon_sync_rings(rdev, &sync, ring->idx);
604
605 for (i = 0; i < num_loops; i++) {
606 cur_size_in_bytes = size_in_bytes;
607 if (cur_size_in_bytes > 0x1fffff)
608 cur_size_in_bytes = 0x1fffff;
609 size_in_bytes -= cur_size_in_bytes;
610 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_COPY, SDMA_COPY_SUB_OPCODE_LINEAR, 0));
611 radeon_ring_write(ring, cur_size_in_bytes);
612 radeon_ring_write(ring, 0); /* src/dst endian swap */
613 radeon_ring_write(ring, lower_32_bits(src_offset));
614 radeon_ring_write(ring, upper_32_bits(src_offset));
615 radeon_ring_write(ring, lower_32_bits(dst_offset));
616 radeon_ring_write(ring, upper_32_bits(dst_offset));
617 src_offset += cur_size_in_bytes;
618 dst_offset += cur_size_in_bytes;
619 }
620
621 r = radeon_fence_emit(rdev, &fence, ring->idx);
622 if (r) {
623 radeon_ring_unlock_undo(rdev, ring);
624 radeon_sync_free(rdev, &sync, NULL);
625 return ERR_PTR(r);
626 }
627
628 radeon_ring_unlock_commit(rdev, ring, false);
629 radeon_sync_free(rdev, &sync, fence);
630
631 return fence;
632 }
633
634 /**
635 * cik_sdma_ring_test - simple async dma engine test
636 *
637 * @rdev: radeon_device pointer
638 * @ring: radeon_ring structure holding ring information
639 *
640 * Test the DMA engine by writing using it to write an
641 * value to memory. (CIK).
642 * Returns 0 for success, error for failure.
643 */
cik_sdma_ring_test(struct radeon_device * rdev,struct radeon_ring * ring)644 int cik_sdma_ring_test(struct radeon_device *rdev,
645 struct radeon_ring *ring)
646 {
647 unsigned i;
648 int r;
649 unsigned index;
650 u32 tmp;
651 u64 gpu_addr;
652
653 if (ring->idx == R600_RING_TYPE_DMA_INDEX)
654 index = R600_WB_DMA_RING_TEST_OFFSET;
655 else
656 index = CAYMAN_WB_DMA1_RING_TEST_OFFSET;
657
658 gpu_addr = rdev->wb.gpu_addr + index;
659
660 tmp = 0xCAFEDEAD;
661 rdev->wb.wb[index/4] = cpu_to_le32(tmp);
662
663 r = radeon_ring_lock(rdev, ring, 5);
664 if (r) {
665 DRM_ERROR("radeon: dma failed to lock ring %d (%d).\n", ring->idx, r);
666 return r;
667 }
668 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_WRITE, SDMA_WRITE_SUB_OPCODE_LINEAR, 0));
669 radeon_ring_write(ring, lower_32_bits(gpu_addr));
670 radeon_ring_write(ring, upper_32_bits(gpu_addr));
671 radeon_ring_write(ring, 1); /* number of DWs to follow */
672 radeon_ring_write(ring, 0xDEADBEEF);
673 radeon_ring_unlock_commit(rdev, ring, false);
674
675 for (i = 0; i < rdev->usec_timeout; i++) {
676 tmp = le32_to_cpu(rdev->wb.wb[index/4]);
677 if (tmp == 0xDEADBEEF)
678 break;
679 udelay(1);
680 }
681
682 if (i < rdev->usec_timeout) {
683 DRM_INFO("ring test on %d succeeded in %d usecs\n", ring->idx, i);
684 } else {
685 DRM_ERROR("radeon: ring %d test failed (0x%08X)\n",
686 ring->idx, tmp);
687 r = -EINVAL;
688 }
689 return r;
690 }
691
692 /**
693 * cik_sdma_ib_test - test an IB on the DMA engine
694 *
695 * @rdev: radeon_device pointer
696 * @ring: radeon_ring structure holding ring information
697 *
698 * Test a simple IB in the DMA ring (CIK).
699 * Returns 0 on success, error on failure.
700 */
cik_sdma_ib_test(struct radeon_device * rdev,struct radeon_ring * ring)701 int cik_sdma_ib_test(struct radeon_device *rdev, struct radeon_ring *ring)
702 {
703 struct radeon_ib ib;
704 unsigned i;
705 unsigned index;
706 int r;
707 u32 tmp = 0;
708 u64 gpu_addr;
709
710 if (ring->idx == R600_RING_TYPE_DMA_INDEX)
711 index = R600_WB_DMA_RING_TEST_OFFSET;
712 else
713 index = CAYMAN_WB_DMA1_RING_TEST_OFFSET;
714
715 gpu_addr = rdev->wb.gpu_addr + index;
716
717 tmp = 0xCAFEDEAD;
718 rdev->wb.wb[index/4] = cpu_to_le32(tmp);
719
720 r = radeon_ib_get(rdev, ring->idx, &ib, NULL, 256);
721 if (r) {
722 DRM_ERROR("radeon: failed to get ib (%d).\n", r);
723 return r;
724 }
725
726 ib.ptr[0] = SDMA_PACKET(SDMA_OPCODE_WRITE, SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
727 ib.ptr[1] = lower_32_bits(gpu_addr);
728 ib.ptr[2] = upper_32_bits(gpu_addr);
729 ib.ptr[3] = 1;
730 ib.ptr[4] = 0xDEADBEEF;
731 ib.length_dw = 5;
732
733 r = radeon_ib_schedule(rdev, &ib, NULL, false);
734 if (r) {
735 radeon_ib_free(rdev, &ib);
736 DRM_ERROR("radeon: failed to schedule ib (%d).\n", r);
737 return r;
738 }
739 r = radeon_fence_wait_timeout(ib.fence, false, usecs_to_jiffies(
740 RADEON_USEC_IB_TEST_TIMEOUT));
741 if (r < 0) {
742 DRM_ERROR("radeon: fence wait failed (%d).\n", r);
743 return r;
744 } else if (r == 0) {
745 DRM_ERROR("radeon: fence wait timed out.\n");
746 return -ETIMEDOUT;
747 }
748 r = 0;
749 for (i = 0; i < rdev->usec_timeout; i++) {
750 tmp = le32_to_cpu(rdev->wb.wb[index/4]);
751 if (tmp == 0xDEADBEEF)
752 break;
753 udelay(1);
754 }
755 if (i < rdev->usec_timeout) {
756 DRM_INFO("ib test on ring %d succeeded in %u usecs\n", ib.fence->ring, i);
757 } else {
758 DRM_ERROR("radeon: ib test failed (0x%08X)\n", tmp);
759 r = -EINVAL;
760 }
761 radeon_ib_free(rdev, &ib);
762 return r;
763 }
764
765 /**
766 * cik_sdma_is_lockup - Check if the DMA engine is locked up
767 *
768 * @rdev: radeon_device pointer
769 * @ring: radeon_ring structure holding ring information
770 *
771 * Check if the async DMA engine is locked up (CIK).
772 * Returns true if the engine appears to be locked up, false if not.
773 */
cik_sdma_is_lockup(struct radeon_device * rdev,struct radeon_ring * ring)774 bool cik_sdma_is_lockup(struct radeon_device *rdev, struct radeon_ring *ring)
775 {
776 u32 reset_mask = cik_gpu_check_soft_reset(rdev);
777 u32 mask;
778
779 if (ring->idx == R600_RING_TYPE_DMA_INDEX)
780 mask = RADEON_RESET_DMA;
781 else
782 mask = RADEON_RESET_DMA1;
783
784 if (!(reset_mask & mask)) {
785 radeon_ring_lockup_update(rdev, ring);
786 return false;
787 }
788 return radeon_ring_test_lockup(rdev, ring);
789 }
790
791 /**
792 * cik_sdma_vm_copy_pages - update PTEs by copying them from the GART
793 *
794 * @rdev: radeon_device pointer
795 * @ib: indirect buffer to fill with commands
796 * @pe: addr of the page entry
797 * @src: src addr to copy from
798 * @count: number of page entries to update
799 *
800 * Update PTEs by copying them from the GART using sDMA (CIK).
801 */
cik_sdma_vm_copy_pages(struct radeon_device * rdev,struct radeon_ib * ib,uint64_t pe,uint64_t src,unsigned count)802 void cik_sdma_vm_copy_pages(struct radeon_device *rdev,
803 struct radeon_ib *ib,
804 uint64_t pe, uint64_t src,
805 unsigned count)
806 {
807 while (count) {
808 unsigned bytes = count * 8;
809 if (bytes > 0x1FFFF8)
810 bytes = 0x1FFFF8;
811
812 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_COPY,
813 SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
814 ib->ptr[ib->length_dw++] = bytes;
815 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
816 ib->ptr[ib->length_dw++] = lower_32_bits(src);
817 ib->ptr[ib->length_dw++] = upper_32_bits(src);
818 ib->ptr[ib->length_dw++] = lower_32_bits(pe);
819 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
820
821 pe += bytes;
822 src += bytes;
823 count -= bytes / 8;
824 }
825 }
826
827 /**
828 * cik_sdma_vm_write_pages - update PTEs by writing them manually
829 *
830 * @rdev: radeon_device pointer
831 * @ib: indirect buffer to fill with commands
832 * @pe: addr of the page entry
833 * @addr: dst addr to write into pe
834 * @count: number of page entries to update
835 * @incr: increase next addr by incr bytes
836 * @flags: access flags
837 *
838 * Update PTEs by writing them manually using sDMA (CIK).
839 */
cik_sdma_vm_write_pages(struct radeon_device * rdev,struct radeon_ib * ib,uint64_t pe,uint64_t addr,unsigned count,uint32_t incr,uint32_t flags)840 void cik_sdma_vm_write_pages(struct radeon_device *rdev,
841 struct radeon_ib *ib,
842 uint64_t pe,
843 uint64_t addr, unsigned count,
844 uint32_t incr, uint32_t flags)
845 {
846 uint64_t value;
847 unsigned ndw;
848
849 while (count) {
850 ndw = count * 2;
851 if (ndw > 0xFFFFE)
852 ndw = 0xFFFFE;
853
854 /* for non-physically contiguous pages (system) */
855 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_WRITE,
856 SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
857 ib->ptr[ib->length_dw++] = pe;
858 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
859 ib->ptr[ib->length_dw++] = ndw;
860 for (; ndw > 0; ndw -= 2, --count, pe += 8) {
861 if (flags & R600_PTE_SYSTEM) {
862 value = radeon_vm_map_gart(rdev, addr);
863 } else if (flags & R600_PTE_VALID) {
864 value = addr;
865 } else {
866 value = 0;
867 }
868 addr += incr;
869 value |= flags;
870 ib->ptr[ib->length_dw++] = value;
871 ib->ptr[ib->length_dw++] = upper_32_bits(value);
872 }
873 }
874 }
875
876 /**
877 * cik_sdma_vm_set_pages - update the page tables using sDMA
878 *
879 * @rdev: radeon_device pointer
880 * @ib: indirect buffer to fill with commands
881 * @pe: addr of the page entry
882 * @addr: dst addr to write into pe
883 * @count: number of page entries to update
884 * @incr: increase next addr by incr bytes
885 * @flags: access flags
886 *
887 * Update the page tables using sDMA (CIK).
888 */
cik_sdma_vm_set_pages(struct radeon_device * rdev,struct radeon_ib * ib,uint64_t pe,uint64_t addr,unsigned count,uint32_t incr,uint32_t flags)889 void cik_sdma_vm_set_pages(struct radeon_device *rdev,
890 struct radeon_ib *ib,
891 uint64_t pe,
892 uint64_t addr, unsigned count,
893 uint32_t incr, uint32_t flags)
894 {
895 uint64_t value;
896 unsigned ndw;
897
898 while (count) {
899 ndw = count;
900 if (ndw > 0x7FFFF)
901 ndw = 0x7FFFF;
902
903 if (flags & R600_PTE_VALID)
904 value = addr;
905 else
906 value = 0;
907
908 /* for physically contiguous pages (vram) */
909 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_GENERATE_PTE_PDE, 0, 0);
910 ib->ptr[ib->length_dw++] = pe; /* dst addr */
911 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
912 ib->ptr[ib->length_dw++] = flags; /* mask */
913 ib->ptr[ib->length_dw++] = 0;
914 ib->ptr[ib->length_dw++] = value; /* value */
915 ib->ptr[ib->length_dw++] = upper_32_bits(value);
916 ib->ptr[ib->length_dw++] = incr; /* increment size */
917 ib->ptr[ib->length_dw++] = 0;
918 ib->ptr[ib->length_dw++] = ndw; /* number of entries */
919
920 pe += ndw * 8;
921 addr += ndw * incr;
922 count -= ndw;
923 }
924 }
925
926 /**
927 * cik_sdma_vm_pad_ib - pad the IB to the required number of dw
928 *
929 * @ib: indirect buffer to fill with padding
930 *
931 */
cik_sdma_vm_pad_ib(struct radeon_ib * ib)932 void cik_sdma_vm_pad_ib(struct radeon_ib *ib)
933 {
934 while (ib->length_dw & 0x7)
935 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0);
936 }
937
938 /*
939 * cik_dma_vm_flush - cik vm flush using sDMA
940 *
941 * Update the page table base and flush the VM TLB
942 * using sDMA (CIK).
943 */
cik_dma_vm_flush(struct radeon_device * rdev,struct radeon_ring * ring,unsigned vm_id,uint64_t pd_addr)944 void cik_dma_vm_flush(struct radeon_device *rdev, struct radeon_ring *ring,
945 unsigned vm_id, uint64_t pd_addr)
946 {
947 u32 extra_bits = (SDMA_POLL_REG_MEM_EXTRA_OP(0) |
948 SDMA_POLL_REG_MEM_EXTRA_FUNC(0)); /* always */
949
950 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
951 if (vm_id < 8) {
952 radeon_ring_write(ring, (VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (vm_id << 2)) >> 2);
953 } else {
954 radeon_ring_write(ring, (VM_CONTEXT8_PAGE_TABLE_BASE_ADDR + ((vm_id - 8) << 2)) >> 2);
955 }
956 radeon_ring_write(ring, pd_addr >> 12);
957
958 /* update SH_MEM_* regs */
959 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
960 radeon_ring_write(ring, SRBM_GFX_CNTL >> 2);
961 radeon_ring_write(ring, VMID(vm_id));
962
963 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
964 radeon_ring_write(ring, SH_MEM_BASES >> 2);
965 radeon_ring_write(ring, 0);
966
967 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
968 radeon_ring_write(ring, SH_MEM_CONFIG >> 2);
969 radeon_ring_write(ring, 0);
970
971 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
972 radeon_ring_write(ring, SH_MEM_APE1_BASE >> 2);
973 radeon_ring_write(ring, 1);
974
975 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
976 radeon_ring_write(ring, SH_MEM_APE1_LIMIT >> 2);
977 radeon_ring_write(ring, 0);
978
979 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
980 radeon_ring_write(ring, SRBM_GFX_CNTL >> 2);
981 radeon_ring_write(ring, VMID(0));
982
983 /* flush HDP */
984 cik_sdma_hdp_flush_ring_emit(rdev, ring->idx);
985
986 /* flush TLB */
987 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
988 radeon_ring_write(ring, VM_INVALIDATE_REQUEST >> 2);
989 radeon_ring_write(ring, 1 << vm_id);
990
991 radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0, extra_bits));
992 radeon_ring_write(ring, VM_INVALIDATE_REQUEST >> 2);
993 radeon_ring_write(ring, 0);
994 radeon_ring_write(ring, 0); /* reference */
995 radeon_ring_write(ring, 0); /* mask */
996 radeon_ring_write(ring, (0xfff << 16) | 10); /* retry count, poll interval */
997 }
998
999