Loading...
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#include <drm/drmP.h>
26#include "amdgpu.h"
27#include "amdgpu_ucode.h"
28#include "amdgpu_trace.h"
29#include "cikd.h"
30#include "cik.h"
31
32#include "bif/bif_4_1_d.h"
33#include "bif/bif_4_1_sh_mask.h"
34
35#include "gca/gfx_7_2_d.h"
36#include "gca/gfx_7_2_enum.h"
37#include "gca/gfx_7_2_sh_mask.h"
38
39#include "gmc/gmc_7_1_d.h"
40#include "gmc/gmc_7_1_sh_mask.h"
41
42#include "oss/oss_2_0_d.h"
43#include "oss/oss_2_0_sh_mask.h"
44
45static const u32 sdma_offsets[SDMA_MAX_INSTANCE] =
46{
47 SDMA0_REGISTER_OFFSET,
48 SDMA1_REGISTER_OFFSET
49};
50
51static void cik_sdma_set_ring_funcs(struct amdgpu_device *adev);
52static void cik_sdma_set_irq_funcs(struct amdgpu_device *adev);
53static void cik_sdma_set_buffer_funcs(struct amdgpu_device *adev);
54static void cik_sdma_set_vm_pte_funcs(struct amdgpu_device *adev);
55
56MODULE_FIRMWARE("radeon/bonaire_sdma.bin");
57MODULE_FIRMWARE("radeon/bonaire_sdma1.bin");
58MODULE_FIRMWARE("radeon/hawaii_sdma.bin");
59MODULE_FIRMWARE("radeon/hawaii_sdma1.bin");
60MODULE_FIRMWARE("radeon/kaveri_sdma.bin");
61MODULE_FIRMWARE("radeon/kaveri_sdma1.bin");
62MODULE_FIRMWARE("radeon/kabini_sdma.bin");
63MODULE_FIRMWARE("radeon/kabini_sdma1.bin");
64MODULE_FIRMWARE("radeon/mullins_sdma.bin");
65MODULE_FIRMWARE("radeon/mullins_sdma1.bin");
66
67u32 amdgpu_cik_gpu_check_soft_reset(struct amdgpu_device *adev);
68
69/*
70 * sDMA - System DMA
71 * Starting with CIK, the GPU has new asynchronous
72 * DMA engines. These engines are used for compute
73 * and gfx. There are two DMA engines (SDMA0, SDMA1)
74 * and each one supports 1 ring buffer used for gfx
75 * and 2 queues used for compute.
76 *
77 * The programming model is very similar to the CP
78 * (ring buffer, IBs, etc.), but sDMA has it's own
79 * packet format that is different from the PM4 format
80 * used by the CP. sDMA supports copying data, writing
81 * embedded data, solid fills, and a number of other
82 * things. It also has support for tiling/detiling of
83 * buffers.
84 */
85
86/**
87 * cik_sdma_init_microcode - load ucode images from disk
88 *
89 * @adev: amdgpu_device pointer
90 *
91 * Use the firmware interface to load the ucode images into
92 * the driver (not loaded into hw).
93 * Returns 0 on success, error on failure.
94 */
95static int cik_sdma_init_microcode(struct amdgpu_device *adev)
96{
97 const char *chip_name;
98 char fw_name[30];
99 int err = 0, i;
100
101 DRM_DEBUG("\n");
102
103 switch (adev->asic_type) {
104 case CHIP_BONAIRE:
105 chip_name = "bonaire";
106 break;
107 case CHIP_HAWAII:
108 chip_name = "hawaii";
109 break;
110 case CHIP_KAVERI:
111 chip_name = "kaveri";
112 break;
113 case CHIP_KABINI:
114 chip_name = "kabini";
115 break;
116 case CHIP_MULLINS:
117 chip_name = "mullins";
118 break;
119 default: BUG();
120 }
121
122 for (i = 0; i < adev->sdma.num_instances; i++) {
123 if (i == 0)
124 snprintf(fw_name, sizeof(fw_name), "radeon/%s_sdma.bin", chip_name);
125 else
126 snprintf(fw_name, sizeof(fw_name), "radeon/%s_sdma1.bin", chip_name);
127 err = request_firmware(&adev->sdma.instance[i].fw, fw_name, adev->dev);
128 if (err)
129 goto out;
130 err = amdgpu_ucode_validate(adev->sdma.instance[i].fw);
131 }
132out:
133 if (err) {
134 printk(KERN_ERR
135 "cik_sdma: Failed to load firmware \"%s\"\n",
136 fw_name);
137 for (i = 0; i < adev->sdma.num_instances; i++) {
138 release_firmware(adev->sdma.instance[i].fw);
139 adev->sdma.instance[i].fw = NULL;
140 }
141 }
142 return err;
143}
144
145/**
146 * cik_sdma_ring_get_rptr - get the current read pointer
147 *
148 * @ring: amdgpu ring pointer
149 *
150 * Get the current rptr from the hardware (CIK+).
151 */
152static uint32_t cik_sdma_ring_get_rptr(struct amdgpu_ring *ring)
153{
154 u32 rptr;
155
156 rptr = ring->adev->wb.wb[ring->rptr_offs];
157
158 return (rptr & 0x3fffc) >> 2;
159}
160
161/**
162 * cik_sdma_ring_get_wptr - get the current write pointer
163 *
164 * @ring: amdgpu ring pointer
165 *
166 * Get the current wptr from the hardware (CIK+).
167 */
168static uint32_t cik_sdma_ring_get_wptr(struct amdgpu_ring *ring)
169{
170 struct amdgpu_device *adev = ring->adev;
171 u32 me = (ring == &adev->sdma.instance[0].ring) ? 0 : 1;
172
173 return (RREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[me]) & 0x3fffc) >> 2;
174}
175
176/**
177 * cik_sdma_ring_set_wptr - commit the write pointer
178 *
179 * @ring: amdgpu ring pointer
180 *
181 * Write the wptr back to the hardware (CIK+).
182 */
183static void cik_sdma_ring_set_wptr(struct amdgpu_ring *ring)
184{
185 struct amdgpu_device *adev = ring->adev;
186 u32 me = (ring == &adev->sdma.instance[0].ring) ? 0 : 1;
187
188 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[me], (ring->wptr << 2) & 0x3fffc);
189}
190
191static void cik_sdma_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
192{
193 struct amdgpu_sdma_instance *sdma = amdgpu_get_sdma_instance(ring);
194 int i;
195
196 for (i = 0; i < count; i++)
197 if (sdma && sdma->burst_nop && (i == 0))
198 amdgpu_ring_write(ring, ring->nop |
199 SDMA_NOP_COUNT(count - 1));
200 else
201 amdgpu_ring_write(ring, ring->nop);
202}
203
204/**
205 * cik_sdma_ring_emit_ib - Schedule an IB on the DMA engine
206 *
207 * @ring: amdgpu ring pointer
208 * @ib: IB object to schedule
209 *
210 * Schedule an IB in the DMA ring (CIK).
211 */
212static void cik_sdma_ring_emit_ib(struct amdgpu_ring *ring,
213 struct amdgpu_ib *ib)
214{
215 u32 extra_bits = ib->vm_id & 0xf;
216 u32 next_rptr = ring->wptr + 5;
217
218 while ((next_rptr & 7) != 4)
219 next_rptr++;
220
221 next_rptr += 4;
222 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_WRITE, SDMA_WRITE_SUB_OPCODE_LINEAR, 0));
223 amdgpu_ring_write(ring, ring->next_rptr_gpu_addr & 0xfffffffc);
224 amdgpu_ring_write(ring, upper_32_bits(ring->next_rptr_gpu_addr) & 0xffffffff);
225 amdgpu_ring_write(ring, 1); /* number of DWs to follow */
226 amdgpu_ring_write(ring, next_rptr);
227
228 /* IB packet must end on a 8 DW boundary */
229 cik_sdma_ring_insert_nop(ring, (12 - (ring->wptr & 7)) % 8);
230
231 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_INDIRECT_BUFFER, 0, extra_bits));
232 amdgpu_ring_write(ring, ib->gpu_addr & 0xffffffe0); /* base must be 32 byte aligned */
233 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr) & 0xffffffff);
234 amdgpu_ring_write(ring, ib->length_dw);
235
236}
237
238/**
239 * cik_sdma_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
240 *
241 * @ring: amdgpu ring pointer
242 *
243 * Emit an hdp flush packet on the requested DMA ring.
244 */
245static void cik_sdma_ring_emit_hdp_flush(struct amdgpu_ring *ring)
246{
247 u32 extra_bits = (SDMA_POLL_REG_MEM_EXTRA_OP(1) |
248 SDMA_POLL_REG_MEM_EXTRA_FUNC(3)); /* == */
249 u32 ref_and_mask;
250
251 if (ring == &ring->adev->sdma.instance[0].ring)
252 ref_and_mask = GPU_HDP_FLUSH_DONE__SDMA0_MASK;
253 else
254 ref_and_mask = GPU_HDP_FLUSH_DONE__SDMA1_MASK;
255
256 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0, extra_bits));
257 amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_DONE << 2);
258 amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_REQ << 2);
259 amdgpu_ring_write(ring, ref_and_mask); /* reference */
260 amdgpu_ring_write(ring, ref_and_mask); /* mask */
261 amdgpu_ring_write(ring, (0xfff << 16) | 10); /* retry count, poll interval */
262}
263
264static void cik_sdma_ring_emit_hdp_invalidate(struct amdgpu_ring *ring)
265{
266 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
267 amdgpu_ring_write(ring, mmHDP_DEBUG0);
268 amdgpu_ring_write(ring, 1);
269}
270
271/**
272 * cik_sdma_ring_emit_fence - emit a fence on the DMA ring
273 *
274 * @ring: amdgpu ring pointer
275 * @fence: amdgpu fence object
276 *
277 * Add a DMA fence packet to the ring to write
278 * the fence seq number and DMA trap packet to generate
279 * an interrupt if needed (CIK).
280 */
281static void cik_sdma_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
282 unsigned flags)
283{
284 bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
285 /* write the fence */
286 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_FENCE, 0, 0));
287 amdgpu_ring_write(ring, lower_32_bits(addr));
288 amdgpu_ring_write(ring, upper_32_bits(addr));
289 amdgpu_ring_write(ring, lower_32_bits(seq));
290
291 /* optionally write high bits as well */
292 if (write64bit) {
293 addr += 4;
294 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_FENCE, 0, 0));
295 amdgpu_ring_write(ring, lower_32_bits(addr));
296 amdgpu_ring_write(ring, upper_32_bits(addr));
297 amdgpu_ring_write(ring, upper_32_bits(seq));
298 }
299
300 /* generate an interrupt */
301 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_TRAP, 0, 0));
302}
303
304/**
305 * cik_sdma_gfx_stop - stop the gfx async dma engines
306 *
307 * @adev: amdgpu_device pointer
308 *
309 * Stop the gfx async dma ring buffers (CIK).
310 */
311static void cik_sdma_gfx_stop(struct amdgpu_device *adev)
312{
313 struct amdgpu_ring *sdma0 = &adev->sdma.instance[0].ring;
314 struct amdgpu_ring *sdma1 = &adev->sdma.instance[1].ring;
315 u32 rb_cntl;
316 int i;
317
318 if ((adev->mman.buffer_funcs_ring == sdma0) ||
319 (adev->mman.buffer_funcs_ring == sdma1))
320 amdgpu_ttm_set_active_vram_size(adev, adev->mc.visible_vram_size);
321
322 for (i = 0; i < adev->sdma.num_instances; i++) {
323 rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
324 rb_cntl &= ~SDMA0_GFX_RB_CNTL__RB_ENABLE_MASK;
325 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
326 WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], 0);
327 }
328 sdma0->ready = false;
329 sdma1->ready = false;
330}
331
332/**
333 * cik_sdma_rlc_stop - stop the compute async dma engines
334 *
335 * @adev: amdgpu_device pointer
336 *
337 * Stop the compute async dma queues (CIK).
338 */
339static void cik_sdma_rlc_stop(struct amdgpu_device *adev)
340{
341 /* XXX todo */
342}
343
344/**
345 * cik_sdma_enable - stop the async dma engines
346 *
347 * @adev: amdgpu_device pointer
348 * @enable: enable/disable the DMA MEs.
349 *
350 * Halt or unhalt the async dma engines (CIK).
351 */
352static void cik_sdma_enable(struct amdgpu_device *adev, bool enable)
353{
354 u32 me_cntl;
355 int i;
356
357 if (enable == false) {
358 cik_sdma_gfx_stop(adev);
359 cik_sdma_rlc_stop(adev);
360 }
361
362 for (i = 0; i < adev->sdma.num_instances; i++) {
363 me_cntl = RREG32(mmSDMA0_F32_CNTL + sdma_offsets[i]);
364 if (enable)
365 me_cntl &= ~SDMA0_F32_CNTL__HALT_MASK;
366 else
367 me_cntl |= SDMA0_F32_CNTL__HALT_MASK;
368 WREG32(mmSDMA0_F32_CNTL + sdma_offsets[i], me_cntl);
369 }
370}
371
372/**
373 * cik_sdma_gfx_resume - setup and start the async dma engines
374 *
375 * @adev: amdgpu_device pointer
376 *
377 * Set up the gfx DMA ring buffers and enable them (CIK).
378 * Returns 0 for success, error for failure.
379 */
380static int cik_sdma_gfx_resume(struct amdgpu_device *adev)
381{
382 struct amdgpu_ring *ring;
383 u32 rb_cntl, ib_cntl;
384 u32 rb_bufsz;
385 u32 wb_offset;
386 int i, j, r;
387
388 for (i = 0; i < adev->sdma.num_instances; i++) {
389 ring = &adev->sdma.instance[i].ring;
390 wb_offset = (ring->rptr_offs * 4);
391
392 mutex_lock(&adev->srbm_mutex);
393 for (j = 0; j < 16; j++) {
394 cik_srbm_select(adev, 0, 0, 0, j);
395 /* SDMA GFX */
396 WREG32(mmSDMA0_GFX_VIRTUAL_ADDR + sdma_offsets[i], 0);
397 WREG32(mmSDMA0_GFX_APE1_CNTL + sdma_offsets[i], 0);
398 /* XXX SDMA RLC - todo */
399 }
400 cik_srbm_select(adev, 0, 0, 0, 0);
401 mutex_unlock(&adev->srbm_mutex);
402
403 WREG32(mmSDMA0_TILING_CONFIG + sdma_offsets[i],
404 adev->gfx.config.gb_addr_config & 0x70);
405
406 WREG32(mmSDMA0_SEM_INCOMPLETE_TIMER_CNTL + sdma_offsets[i], 0);
407 WREG32(mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
408
409 /* Set ring buffer size in dwords */
410 rb_bufsz = order_base_2(ring->ring_size / 4);
411 rb_cntl = rb_bufsz << 1;
412#ifdef __BIG_ENDIAN
413 rb_cntl |= SDMA0_GFX_RB_CNTL__RB_SWAP_ENABLE_MASK |
414 SDMA0_GFX_RB_CNTL__RPTR_WRITEBACK_SWAP_ENABLE_MASK;
415#endif
416 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
417
418 /* Initialize the ring buffer's read and write pointers */
419 WREG32(mmSDMA0_GFX_RB_RPTR + sdma_offsets[i], 0);
420 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], 0);
421
422 /* set the wb address whether it's enabled or not */
423 WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_HI + sdma_offsets[i],
424 upper_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF);
425 WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_LO + sdma_offsets[i],
426 ((adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC));
427
428 rb_cntl |= SDMA0_GFX_RB_CNTL__RPTR_WRITEBACK_ENABLE_MASK;
429
430 WREG32(mmSDMA0_GFX_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
431 WREG32(mmSDMA0_GFX_RB_BASE_HI + sdma_offsets[i], ring->gpu_addr >> 40);
432
433 ring->wptr = 0;
434 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], ring->wptr << 2);
435
436 /* enable DMA RB */
437 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i],
438 rb_cntl | SDMA0_GFX_RB_CNTL__RB_ENABLE_MASK);
439
440 ib_cntl = SDMA0_GFX_IB_CNTL__IB_ENABLE_MASK;
441#ifdef __BIG_ENDIAN
442 ib_cntl |= SDMA0_GFX_IB_CNTL__IB_SWAP_ENABLE_MASK;
443#endif
444 /* enable DMA IBs */
445 WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
446
447 ring->ready = true;
448
449 r = amdgpu_ring_test_ring(ring);
450 if (r) {
451 ring->ready = false;
452 return r;
453 }
454
455 if (adev->mman.buffer_funcs_ring == ring)
456 amdgpu_ttm_set_active_vram_size(adev, adev->mc.real_vram_size);
457 }
458
459 return 0;
460}
461
462/**
463 * cik_sdma_rlc_resume - setup and start the async dma engines
464 *
465 * @adev: amdgpu_device pointer
466 *
467 * Set up the compute DMA queues and enable them (CIK).
468 * Returns 0 for success, error for failure.
469 */
470static int cik_sdma_rlc_resume(struct amdgpu_device *adev)
471{
472 /* XXX todo */
473 return 0;
474}
475
476/**
477 * cik_sdma_load_microcode - load the sDMA ME ucode
478 *
479 * @adev: amdgpu_device pointer
480 *
481 * Loads the sDMA0/1 ucode.
482 * Returns 0 for success, -EINVAL if the ucode is not available.
483 */
484static int cik_sdma_load_microcode(struct amdgpu_device *adev)
485{
486 const struct sdma_firmware_header_v1_0 *hdr;
487 const __le32 *fw_data;
488 u32 fw_size;
489 int i, j;
490
491 /* halt the MEs */
492 cik_sdma_enable(adev, false);
493
494 for (i = 0; i < adev->sdma.num_instances; i++) {
495 if (!adev->sdma.instance[i].fw)
496 return -EINVAL;
497 hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
498 amdgpu_ucode_print_sdma_hdr(&hdr->header);
499 fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
500 adev->sdma.instance[i].fw_version = le32_to_cpu(hdr->header.ucode_version);
501 adev->sdma.instance[i].feature_version = le32_to_cpu(hdr->ucode_feature_version);
502 if (adev->sdma.instance[i].feature_version >= 20)
503 adev->sdma.instance[i].burst_nop = true;
504 fw_data = (const __le32 *)
505 (adev->sdma.instance[i].fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
506 WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], 0);
507 for (j = 0; j < fw_size; j++)
508 WREG32(mmSDMA0_UCODE_DATA + sdma_offsets[i], le32_to_cpup(fw_data++));
509 WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], adev->sdma.instance[i].fw_version);
510 }
511
512 return 0;
513}
514
515/**
516 * cik_sdma_start - setup and start the async dma engines
517 *
518 * @adev: amdgpu_device pointer
519 *
520 * Set up the DMA engines and enable them (CIK).
521 * Returns 0 for success, error for failure.
522 */
523static int cik_sdma_start(struct amdgpu_device *adev)
524{
525 int r;
526
527 r = cik_sdma_load_microcode(adev);
528 if (r)
529 return r;
530
531 /* unhalt the MEs */
532 cik_sdma_enable(adev, true);
533
534 /* start the gfx rings and rlc compute queues */
535 r = cik_sdma_gfx_resume(adev);
536 if (r)
537 return r;
538 r = cik_sdma_rlc_resume(adev);
539 if (r)
540 return r;
541
542 return 0;
543}
544
545/**
546 * cik_sdma_ring_test_ring - simple async dma engine test
547 *
548 * @ring: amdgpu_ring structure holding ring information
549 *
550 * Test the DMA engine by writing using it to write an
551 * value to memory. (CIK).
552 * Returns 0 for success, error for failure.
553 */
554static int cik_sdma_ring_test_ring(struct amdgpu_ring *ring)
555{
556 struct amdgpu_device *adev = ring->adev;
557 unsigned i;
558 unsigned index;
559 int r;
560 u32 tmp;
561 u64 gpu_addr;
562
563 r = amdgpu_wb_get(adev, &index);
564 if (r) {
565 dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
566 return r;
567 }
568
569 gpu_addr = adev->wb.gpu_addr + (index * 4);
570 tmp = 0xCAFEDEAD;
571 adev->wb.wb[index] = cpu_to_le32(tmp);
572
573 r = amdgpu_ring_alloc(ring, 5);
574 if (r) {
575 DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
576 amdgpu_wb_free(adev, index);
577 return r;
578 }
579 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_WRITE, SDMA_WRITE_SUB_OPCODE_LINEAR, 0));
580 amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
581 amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
582 amdgpu_ring_write(ring, 1); /* number of DWs to follow */
583 amdgpu_ring_write(ring, 0xDEADBEEF);
584 amdgpu_ring_commit(ring);
585
586 for (i = 0; i < adev->usec_timeout; i++) {
587 tmp = le32_to_cpu(adev->wb.wb[index]);
588 if (tmp == 0xDEADBEEF)
589 break;
590 DRM_UDELAY(1);
591 }
592
593 if (i < adev->usec_timeout) {
594 DRM_INFO("ring test on %d succeeded in %d usecs\n", ring->idx, i);
595 } else {
596 DRM_ERROR("amdgpu: ring %d test failed (0x%08X)\n",
597 ring->idx, tmp);
598 r = -EINVAL;
599 }
600 amdgpu_wb_free(adev, index);
601
602 return r;
603}
604
605/**
606 * cik_sdma_ring_test_ib - test an IB on the DMA engine
607 *
608 * @ring: amdgpu_ring structure holding ring information
609 *
610 * Test a simple IB in the DMA ring (CIK).
611 * Returns 0 on success, error on failure.
612 */
613static int cik_sdma_ring_test_ib(struct amdgpu_ring *ring)
614{
615 struct amdgpu_device *adev = ring->adev;
616 struct amdgpu_ib ib;
617 struct fence *f = NULL;
618 unsigned i;
619 unsigned index;
620 int r;
621 u32 tmp = 0;
622 u64 gpu_addr;
623
624 r = amdgpu_wb_get(adev, &index);
625 if (r) {
626 dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
627 return r;
628 }
629
630 gpu_addr = adev->wb.gpu_addr + (index * 4);
631 tmp = 0xCAFEDEAD;
632 adev->wb.wb[index] = cpu_to_le32(tmp);
633 memset(&ib, 0, sizeof(ib));
634 r = amdgpu_ib_get(adev, NULL, 256, &ib);
635 if (r) {
636 DRM_ERROR("amdgpu: failed to get ib (%d).\n", r);
637 goto err0;
638 }
639
640 ib.ptr[0] = SDMA_PACKET(SDMA_OPCODE_WRITE, SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
641 ib.ptr[1] = lower_32_bits(gpu_addr);
642 ib.ptr[2] = upper_32_bits(gpu_addr);
643 ib.ptr[3] = 1;
644 ib.ptr[4] = 0xDEADBEEF;
645 ib.length_dw = 5;
646 r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
647 if (r)
648 goto err1;
649
650 r = fence_wait(f, false);
651 if (r) {
652 DRM_ERROR("amdgpu: fence wait failed (%d).\n", r);
653 goto err1;
654 }
655 for (i = 0; i < adev->usec_timeout; i++) {
656 tmp = le32_to_cpu(adev->wb.wb[index]);
657 if (tmp == 0xDEADBEEF)
658 break;
659 DRM_UDELAY(1);
660 }
661 if (i < adev->usec_timeout) {
662 DRM_INFO("ib test on ring %d succeeded in %u usecs\n",
663 ring->idx, i);
664 goto err1;
665 } else {
666 DRM_ERROR("amdgpu: ib test failed (0x%08X)\n", tmp);
667 r = -EINVAL;
668 }
669
670err1:
671 fence_put(f);
672 amdgpu_ib_free(adev, &ib, NULL);
673 fence_put(f);
674err0:
675 amdgpu_wb_free(adev, index);
676 return r;
677}
678
679/**
680 * cik_sdma_vm_copy_pages - update PTEs by copying them from the GART
681 *
682 * @ib: indirect buffer to fill with commands
683 * @pe: addr of the page entry
684 * @src: src addr to copy from
685 * @count: number of page entries to update
686 *
687 * Update PTEs by copying them from the GART using sDMA (CIK).
688 */
689static void cik_sdma_vm_copy_pte(struct amdgpu_ib *ib,
690 uint64_t pe, uint64_t src,
691 unsigned count)
692{
693 while (count) {
694 unsigned bytes = count * 8;
695 if (bytes > 0x1FFFF8)
696 bytes = 0x1FFFF8;
697
698 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_COPY,
699 SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
700 ib->ptr[ib->length_dw++] = bytes;
701 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
702 ib->ptr[ib->length_dw++] = lower_32_bits(src);
703 ib->ptr[ib->length_dw++] = upper_32_bits(src);
704 ib->ptr[ib->length_dw++] = lower_32_bits(pe);
705 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
706
707 pe += bytes;
708 src += bytes;
709 count -= bytes / 8;
710 }
711}
712
713/**
714 * cik_sdma_vm_write_pages - update PTEs by writing them manually
715 *
716 * @ib: indirect buffer to fill with commands
717 * @pe: addr of the page entry
718 * @addr: dst addr to write into pe
719 * @count: number of page entries to update
720 * @incr: increase next addr by incr bytes
721 * @flags: access flags
722 *
723 * Update PTEs by writing them manually using sDMA (CIK).
724 */
725static void cik_sdma_vm_write_pte(struct amdgpu_ib *ib,
726 const dma_addr_t *pages_addr, uint64_t pe,
727 uint64_t addr, unsigned count,
728 uint32_t incr, uint32_t flags)
729{
730 uint64_t value;
731 unsigned ndw;
732
733 while (count) {
734 ndw = count * 2;
735 if (ndw > 0xFFFFE)
736 ndw = 0xFFFFE;
737
738 /* for non-physically contiguous pages (system) */
739 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_WRITE,
740 SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
741 ib->ptr[ib->length_dw++] = pe;
742 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
743 ib->ptr[ib->length_dw++] = ndw;
744 for (; ndw > 0; ndw -= 2, --count, pe += 8) {
745 value = amdgpu_vm_map_gart(pages_addr, addr);
746 addr += incr;
747 value |= flags;
748 ib->ptr[ib->length_dw++] = value;
749 ib->ptr[ib->length_dw++] = upper_32_bits(value);
750 }
751 }
752}
753
754/**
755 * cik_sdma_vm_set_pages - update the page tables using sDMA
756 *
757 * @ib: indirect buffer to fill with commands
758 * @pe: addr of the page entry
759 * @addr: dst addr to write into pe
760 * @count: number of page entries to update
761 * @incr: increase next addr by incr bytes
762 * @flags: access flags
763 *
764 * Update the page tables using sDMA (CIK).
765 */
766static void cik_sdma_vm_set_pte_pde(struct amdgpu_ib *ib,
767 uint64_t pe,
768 uint64_t addr, unsigned count,
769 uint32_t incr, uint32_t flags)
770{
771 uint64_t value;
772 unsigned ndw;
773
774 while (count) {
775 ndw = count;
776 if (ndw > 0x7FFFF)
777 ndw = 0x7FFFF;
778
779 if (flags & AMDGPU_PTE_VALID)
780 value = addr;
781 else
782 value = 0;
783
784 /* for physically contiguous pages (vram) */
785 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_GENERATE_PTE_PDE, 0, 0);
786 ib->ptr[ib->length_dw++] = pe; /* dst addr */
787 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
788 ib->ptr[ib->length_dw++] = flags; /* mask */
789 ib->ptr[ib->length_dw++] = 0;
790 ib->ptr[ib->length_dw++] = value; /* value */
791 ib->ptr[ib->length_dw++] = upper_32_bits(value);
792 ib->ptr[ib->length_dw++] = incr; /* increment size */
793 ib->ptr[ib->length_dw++] = 0;
794 ib->ptr[ib->length_dw++] = ndw; /* number of entries */
795
796 pe += ndw * 8;
797 addr += ndw * incr;
798 count -= ndw;
799 }
800}
801
802/**
803 * cik_sdma_vm_pad_ib - pad the IB to the required number of dw
804 *
805 * @ib: indirect buffer to fill with padding
806 *
807 */
808static void cik_sdma_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
809{
810 struct amdgpu_sdma_instance *sdma = amdgpu_get_sdma_instance(ring);
811 u32 pad_count;
812 int i;
813
814 pad_count = (8 - (ib->length_dw & 0x7)) % 8;
815 for (i = 0; i < pad_count; i++)
816 if (sdma && sdma->burst_nop && (i == 0))
817 ib->ptr[ib->length_dw++] =
818 SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0) |
819 SDMA_NOP_COUNT(pad_count - 1);
820 else
821 ib->ptr[ib->length_dw++] =
822 SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0);
823}
824
825/**
826 * cik_sdma_ring_emit_pipeline_sync - sync the pipeline
827 *
828 * @ring: amdgpu_ring pointer
829 *
830 * Make sure all previous operations are completed (CIK).
831 */
832static void cik_sdma_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
833{
834 uint32_t seq = ring->fence_drv.sync_seq;
835 uint64_t addr = ring->fence_drv.gpu_addr;
836
837 /* wait for idle */
838 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0,
839 SDMA_POLL_REG_MEM_EXTRA_OP(0) |
840 SDMA_POLL_REG_MEM_EXTRA_FUNC(3) | /* equal */
841 SDMA_POLL_REG_MEM_EXTRA_M));
842 amdgpu_ring_write(ring, addr & 0xfffffffc);
843 amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
844 amdgpu_ring_write(ring, seq); /* reference */
845 amdgpu_ring_write(ring, 0xfffffff); /* mask */
846 amdgpu_ring_write(ring, (0xfff << 16) | 4); /* retry count, poll interval */
847}
848
849/**
850 * cik_sdma_ring_emit_vm_flush - cik vm flush using sDMA
851 *
852 * @ring: amdgpu_ring pointer
853 * @vm: amdgpu_vm pointer
854 *
855 * Update the page table base and flush the VM TLB
856 * using sDMA (CIK).
857 */
858static void cik_sdma_ring_emit_vm_flush(struct amdgpu_ring *ring,
859 unsigned vm_id, uint64_t pd_addr)
860{
861 u32 extra_bits = (SDMA_POLL_REG_MEM_EXTRA_OP(0) |
862 SDMA_POLL_REG_MEM_EXTRA_FUNC(0)); /* always */
863
864 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
865 if (vm_id < 8) {
866 amdgpu_ring_write(ring, (mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + vm_id));
867 } else {
868 amdgpu_ring_write(ring, (mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vm_id - 8));
869 }
870 amdgpu_ring_write(ring, pd_addr >> 12);
871
872 /* flush TLB */
873 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
874 amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST);
875 amdgpu_ring_write(ring, 1 << vm_id);
876
877 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0, extra_bits));
878 amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST << 2);
879 amdgpu_ring_write(ring, 0);
880 amdgpu_ring_write(ring, 0); /* reference */
881 amdgpu_ring_write(ring, 0); /* mask */
882 amdgpu_ring_write(ring, (0xfff << 16) | 10); /* retry count, poll interval */
883}
884
885static void cik_enable_sdma_mgcg(struct amdgpu_device *adev,
886 bool enable)
887{
888 u32 orig, data;
889
890 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
891 WREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET, 0x00000100);
892 WREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET, 0x00000100);
893 } else {
894 orig = data = RREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET);
895 data |= 0xff000000;
896 if (data != orig)
897 WREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET, data);
898
899 orig = data = RREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET);
900 data |= 0xff000000;
901 if (data != orig)
902 WREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET, data);
903 }
904}
905
906static void cik_enable_sdma_mgls(struct amdgpu_device *adev,
907 bool enable)
908{
909 u32 orig, data;
910
911 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
912 orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET);
913 data |= 0x100;
914 if (orig != data)
915 WREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET, data);
916
917 orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET);
918 data |= 0x100;
919 if (orig != data)
920 WREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET, data);
921 } else {
922 orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET);
923 data &= ~0x100;
924 if (orig != data)
925 WREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET, data);
926
927 orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET);
928 data &= ~0x100;
929 if (orig != data)
930 WREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET, data);
931 }
932}
933
934static int cik_sdma_early_init(void *handle)
935{
936 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
937
938 adev->sdma.num_instances = SDMA_MAX_INSTANCE;
939
940 cik_sdma_set_ring_funcs(adev);
941 cik_sdma_set_irq_funcs(adev);
942 cik_sdma_set_buffer_funcs(adev);
943 cik_sdma_set_vm_pte_funcs(adev);
944
945 return 0;
946}
947
948static int cik_sdma_sw_init(void *handle)
949{
950 struct amdgpu_ring *ring;
951 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
952 int r, i;
953
954 r = cik_sdma_init_microcode(adev);
955 if (r) {
956 DRM_ERROR("Failed to load sdma firmware!\n");
957 return r;
958 }
959
960 /* SDMA trap event */
961 r = amdgpu_irq_add_id(adev, 224, &adev->sdma.trap_irq);
962 if (r)
963 return r;
964
965 /* SDMA Privileged inst */
966 r = amdgpu_irq_add_id(adev, 241, &adev->sdma.illegal_inst_irq);
967 if (r)
968 return r;
969
970 /* SDMA Privileged inst */
971 r = amdgpu_irq_add_id(adev, 247, &adev->sdma.illegal_inst_irq);
972 if (r)
973 return r;
974
975 for (i = 0; i < adev->sdma.num_instances; i++) {
976 ring = &adev->sdma.instance[i].ring;
977 ring->ring_obj = NULL;
978 sprintf(ring->name, "sdma%d", i);
979 r = amdgpu_ring_init(adev, ring, 256 * 1024,
980 SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0), 0xf,
981 &adev->sdma.trap_irq,
982 (i == 0) ?
983 AMDGPU_SDMA_IRQ_TRAP0 : AMDGPU_SDMA_IRQ_TRAP1,
984 AMDGPU_RING_TYPE_SDMA);
985 if (r)
986 return r;
987 }
988
989 return r;
990}
991
992static int cik_sdma_sw_fini(void *handle)
993{
994 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
995 int i;
996
997 for (i = 0; i < adev->sdma.num_instances; i++)
998 amdgpu_ring_fini(&adev->sdma.instance[i].ring);
999
1000 return 0;
1001}
1002
1003static int cik_sdma_hw_init(void *handle)
1004{
1005 int r;
1006 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1007
1008 r = cik_sdma_start(adev);
1009 if (r)
1010 return r;
1011
1012 return r;
1013}
1014
1015static int cik_sdma_hw_fini(void *handle)
1016{
1017 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1018
1019 cik_sdma_enable(adev, false);
1020
1021 return 0;
1022}
1023
1024static int cik_sdma_suspend(void *handle)
1025{
1026 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1027
1028 return cik_sdma_hw_fini(adev);
1029}
1030
1031static int cik_sdma_resume(void *handle)
1032{
1033 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1034
1035 return cik_sdma_hw_init(adev);
1036}
1037
1038static bool cik_sdma_is_idle(void *handle)
1039{
1040 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1041 u32 tmp = RREG32(mmSRBM_STATUS2);
1042
1043 if (tmp & (SRBM_STATUS2__SDMA_BUSY_MASK |
1044 SRBM_STATUS2__SDMA1_BUSY_MASK))
1045 return false;
1046
1047 return true;
1048}
1049
1050static int cik_sdma_wait_for_idle(void *handle)
1051{
1052 unsigned i;
1053 u32 tmp;
1054 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1055
1056 for (i = 0; i < adev->usec_timeout; i++) {
1057 tmp = RREG32(mmSRBM_STATUS2) & (SRBM_STATUS2__SDMA_BUSY_MASK |
1058 SRBM_STATUS2__SDMA1_BUSY_MASK);
1059
1060 if (!tmp)
1061 return 0;
1062 udelay(1);
1063 }
1064 return -ETIMEDOUT;
1065}
1066
1067static void cik_sdma_print_status(void *handle)
1068{
1069 int i, j;
1070 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1071
1072 dev_info(adev->dev, "CIK SDMA registers\n");
1073 dev_info(adev->dev, " SRBM_STATUS2=0x%08X\n",
1074 RREG32(mmSRBM_STATUS2));
1075 for (i = 0; i < adev->sdma.num_instances; i++) {
1076 dev_info(adev->dev, " SDMA%d_STATUS_REG=0x%08X\n",
1077 i, RREG32(mmSDMA0_STATUS_REG + sdma_offsets[i]));
1078 dev_info(adev->dev, " SDMA%d_ME_CNTL=0x%08X\n",
1079 i, RREG32(mmSDMA0_F32_CNTL + sdma_offsets[i]));
1080 dev_info(adev->dev, " SDMA%d_CNTL=0x%08X\n",
1081 i, RREG32(mmSDMA0_CNTL + sdma_offsets[i]));
1082 dev_info(adev->dev, " SDMA%d_SEM_INCOMPLETE_TIMER_CNTL=0x%08X\n",
1083 i, RREG32(mmSDMA0_SEM_INCOMPLETE_TIMER_CNTL + sdma_offsets[i]));
1084 dev_info(adev->dev, " SDMA%d_SEM_WAIT_FAIL_TIMER_CNTL=0x%08X\n",
1085 i, RREG32(mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i]));
1086 dev_info(adev->dev, " SDMA%d_GFX_IB_CNTL=0x%08X\n",
1087 i, RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]));
1088 dev_info(adev->dev, " SDMA%d_GFX_RB_CNTL=0x%08X\n",
1089 i, RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]));
1090 dev_info(adev->dev, " SDMA%d_GFX_RB_RPTR=0x%08X\n",
1091 i, RREG32(mmSDMA0_GFX_RB_RPTR + sdma_offsets[i]));
1092 dev_info(adev->dev, " SDMA%d_GFX_RB_WPTR=0x%08X\n",
1093 i, RREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i]));
1094 dev_info(adev->dev, " SDMA%d_GFX_RB_RPTR_ADDR_HI=0x%08X\n",
1095 i, RREG32(mmSDMA0_GFX_RB_RPTR_ADDR_HI + sdma_offsets[i]));
1096 dev_info(adev->dev, " SDMA%d_GFX_RB_RPTR_ADDR_LO=0x%08X\n",
1097 i, RREG32(mmSDMA0_GFX_RB_RPTR_ADDR_LO + sdma_offsets[i]));
1098 dev_info(adev->dev, " SDMA%d_GFX_RB_BASE=0x%08X\n",
1099 i, RREG32(mmSDMA0_GFX_RB_BASE + sdma_offsets[i]));
1100 dev_info(adev->dev, " SDMA%d_GFX_RB_BASE_HI=0x%08X\n",
1101 i, RREG32(mmSDMA0_GFX_RB_BASE_HI + sdma_offsets[i]));
1102 dev_info(adev->dev, " SDMA%d_TILING_CONFIG=0x%08X\n",
1103 i, RREG32(mmSDMA0_TILING_CONFIG + sdma_offsets[i]));
1104 mutex_lock(&adev->srbm_mutex);
1105 for (j = 0; j < 16; j++) {
1106 cik_srbm_select(adev, 0, 0, 0, j);
1107 dev_info(adev->dev, " VM %d:\n", j);
1108 dev_info(adev->dev, " SDMA0_GFX_VIRTUAL_ADDR=0x%08X\n",
1109 RREG32(mmSDMA0_GFX_VIRTUAL_ADDR + sdma_offsets[i]));
1110 dev_info(adev->dev, " SDMA0_GFX_APE1_CNTL=0x%08X\n",
1111 RREG32(mmSDMA0_GFX_APE1_CNTL + sdma_offsets[i]));
1112 }
1113 cik_srbm_select(adev, 0, 0, 0, 0);
1114 mutex_unlock(&adev->srbm_mutex);
1115 }
1116}
1117
1118static int cik_sdma_soft_reset(void *handle)
1119{
1120 u32 srbm_soft_reset = 0;
1121 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1122 u32 tmp = RREG32(mmSRBM_STATUS2);
1123
1124 if (tmp & SRBM_STATUS2__SDMA_BUSY_MASK) {
1125 /* sdma0 */
1126 tmp = RREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET);
1127 tmp |= SDMA0_F32_CNTL__HALT_MASK;
1128 WREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET, tmp);
1129 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA_MASK;
1130 }
1131 if (tmp & SRBM_STATUS2__SDMA1_BUSY_MASK) {
1132 /* sdma1 */
1133 tmp = RREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET);
1134 tmp |= SDMA0_F32_CNTL__HALT_MASK;
1135 WREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET, tmp);
1136 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA1_MASK;
1137 }
1138
1139 if (srbm_soft_reset) {
1140 cik_sdma_print_status((void *)adev);
1141
1142 tmp = RREG32(mmSRBM_SOFT_RESET);
1143 tmp |= srbm_soft_reset;
1144 dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
1145 WREG32(mmSRBM_SOFT_RESET, tmp);
1146 tmp = RREG32(mmSRBM_SOFT_RESET);
1147
1148 udelay(50);
1149
1150 tmp &= ~srbm_soft_reset;
1151 WREG32(mmSRBM_SOFT_RESET, tmp);
1152 tmp = RREG32(mmSRBM_SOFT_RESET);
1153
1154 /* Wait a little for things to settle down */
1155 udelay(50);
1156
1157 cik_sdma_print_status((void *)adev);
1158 }
1159
1160 return 0;
1161}
1162
1163static int cik_sdma_set_trap_irq_state(struct amdgpu_device *adev,
1164 struct amdgpu_irq_src *src,
1165 unsigned type,
1166 enum amdgpu_interrupt_state state)
1167{
1168 u32 sdma_cntl;
1169
1170 switch (type) {
1171 case AMDGPU_SDMA_IRQ_TRAP0:
1172 switch (state) {
1173 case AMDGPU_IRQ_STATE_DISABLE:
1174 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1175 sdma_cntl &= ~SDMA0_CNTL__TRAP_ENABLE_MASK;
1176 WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1177 break;
1178 case AMDGPU_IRQ_STATE_ENABLE:
1179 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1180 sdma_cntl |= SDMA0_CNTL__TRAP_ENABLE_MASK;
1181 WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1182 break;
1183 default:
1184 break;
1185 }
1186 break;
1187 case AMDGPU_SDMA_IRQ_TRAP1:
1188 switch (state) {
1189 case AMDGPU_IRQ_STATE_DISABLE:
1190 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1191 sdma_cntl &= ~SDMA0_CNTL__TRAP_ENABLE_MASK;
1192 WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1193 break;
1194 case AMDGPU_IRQ_STATE_ENABLE:
1195 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1196 sdma_cntl |= SDMA0_CNTL__TRAP_ENABLE_MASK;
1197 WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1198 break;
1199 default:
1200 break;
1201 }
1202 break;
1203 default:
1204 break;
1205 }
1206 return 0;
1207}
1208
1209static int cik_sdma_process_trap_irq(struct amdgpu_device *adev,
1210 struct amdgpu_irq_src *source,
1211 struct amdgpu_iv_entry *entry)
1212{
1213 u8 instance_id, queue_id;
1214
1215 instance_id = (entry->ring_id & 0x3) >> 0;
1216 queue_id = (entry->ring_id & 0xc) >> 2;
1217 DRM_DEBUG("IH: SDMA trap\n");
1218 switch (instance_id) {
1219 case 0:
1220 switch (queue_id) {
1221 case 0:
1222 amdgpu_fence_process(&adev->sdma.instance[0].ring);
1223 break;
1224 case 1:
1225 /* XXX compute */
1226 break;
1227 case 2:
1228 /* XXX compute */
1229 break;
1230 }
1231 break;
1232 case 1:
1233 switch (queue_id) {
1234 case 0:
1235 amdgpu_fence_process(&adev->sdma.instance[1].ring);
1236 break;
1237 case 1:
1238 /* XXX compute */
1239 break;
1240 case 2:
1241 /* XXX compute */
1242 break;
1243 }
1244 break;
1245 }
1246
1247 return 0;
1248}
1249
1250static int cik_sdma_process_illegal_inst_irq(struct amdgpu_device *adev,
1251 struct amdgpu_irq_src *source,
1252 struct amdgpu_iv_entry *entry)
1253{
1254 DRM_ERROR("Illegal instruction in SDMA command stream\n");
1255 schedule_work(&adev->reset_work);
1256 return 0;
1257}
1258
1259static int cik_sdma_set_clockgating_state(void *handle,
1260 enum amd_clockgating_state state)
1261{
1262 bool gate = false;
1263 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1264
1265 if (state == AMD_CG_STATE_GATE)
1266 gate = true;
1267
1268 cik_enable_sdma_mgcg(adev, gate);
1269 cik_enable_sdma_mgls(adev, gate);
1270
1271 return 0;
1272}
1273
1274static int cik_sdma_set_powergating_state(void *handle,
1275 enum amd_powergating_state state)
1276{
1277 return 0;
1278}
1279
1280const struct amd_ip_funcs cik_sdma_ip_funcs = {
1281 .early_init = cik_sdma_early_init,
1282 .late_init = NULL,
1283 .sw_init = cik_sdma_sw_init,
1284 .sw_fini = cik_sdma_sw_fini,
1285 .hw_init = cik_sdma_hw_init,
1286 .hw_fini = cik_sdma_hw_fini,
1287 .suspend = cik_sdma_suspend,
1288 .resume = cik_sdma_resume,
1289 .is_idle = cik_sdma_is_idle,
1290 .wait_for_idle = cik_sdma_wait_for_idle,
1291 .soft_reset = cik_sdma_soft_reset,
1292 .print_status = cik_sdma_print_status,
1293 .set_clockgating_state = cik_sdma_set_clockgating_state,
1294 .set_powergating_state = cik_sdma_set_powergating_state,
1295};
1296
1297static const struct amdgpu_ring_funcs cik_sdma_ring_funcs = {
1298 .get_rptr = cik_sdma_ring_get_rptr,
1299 .get_wptr = cik_sdma_ring_get_wptr,
1300 .set_wptr = cik_sdma_ring_set_wptr,
1301 .parse_cs = NULL,
1302 .emit_ib = cik_sdma_ring_emit_ib,
1303 .emit_fence = cik_sdma_ring_emit_fence,
1304 .emit_pipeline_sync = cik_sdma_ring_emit_pipeline_sync,
1305 .emit_vm_flush = cik_sdma_ring_emit_vm_flush,
1306 .emit_hdp_flush = cik_sdma_ring_emit_hdp_flush,
1307 .emit_hdp_invalidate = cik_sdma_ring_emit_hdp_invalidate,
1308 .test_ring = cik_sdma_ring_test_ring,
1309 .test_ib = cik_sdma_ring_test_ib,
1310 .insert_nop = cik_sdma_ring_insert_nop,
1311 .pad_ib = cik_sdma_ring_pad_ib,
1312};
1313
1314static void cik_sdma_set_ring_funcs(struct amdgpu_device *adev)
1315{
1316 int i;
1317
1318 for (i = 0; i < adev->sdma.num_instances; i++)
1319 adev->sdma.instance[i].ring.funcs = &cik_sdma_ring_funcs;
1320}
1321
1322static const struct amdgpu_irq_src_funcs cik_sdma_trap_irq_funcs = {
1323 .set = cik_sdma_set_trap_irq_state,
1324 .process = cik_sdma_process_trap_irq,
1325};
1326
1327static const struct amdgpu_irq_src_funcs cik_sdma_illegal_inst_irq_funcs = {
1328 .process = cik_sdma_process_illegal_inst_irq,
1329};
1330
1331static void cik_sdma_set_irq_funcs(struct amdgpu_device *adev)
1332{
1333 adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
1334 adev->sdma.trap_irq.funcs = &cik_sdma_trap_irq_funcs;
1335 adev->sdma.illegal_inst_irq.funcs = &cik_sdma_illegal_inst_irq_funcs;
1336}
1337
1338/**
1339 * cik_sdma_emit_copy_buffer - copy buffer using the sDMA engine
1340 *
1341 * @ring: amdgpu_ring structure holding ring information
1342 * @src_offset: src GPU address
1343 * @dst_offset: dst GPU address
1344 * @byte_count: number of bytes to xfer
1345 *
1346 * Copy GPU buffers using the DMA engine (CIK).
1347 * Used by the amdgpu ttm implementation to move pages if
1348 * registered as the asic copy callback.
1349 */
1350static void cik_sdma_emit_copy_buffer(struct amdgpu_ib *ib,
1351 uint64_t src_offset,
1352 uint64_t dst_offset,
1353 uint32_t byte_count)
1354{
1355 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_COPY, SDMA_COPY_SUB_OPCODE_LINEAR, 0);
1356 ib->ptr[ib->length_dw++] = byte_count;
1357 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1358 ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1359 ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1360 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1361 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1362}
1363
1364/**
1365 * cik_sdma_emit_fill_buffer - fill buffer using the sDMA engine
1366 *
1367 * @ring: amdgpu_ring structure holding ring information
1368 * @src_data: value to write to buffer
1369 * @dst_offset: dst GPU address
1370 * @byte_count: number of bytes to xfer
1371 *
1372 * Fill GPU buffers using the DMA engine (CIK).
1373 */
1374static void cik_sdma_emit_fill_buffer(struct amdgpu_ib *ib,
1375 uint32_t src_data,
1376 uint64_t dst_offset,
1377 uint32_t byte_count)
1378{
1379 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_CONSTANT_FILL, 0, 0);
1380 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1381 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1382 ib->ptr[ib->length_dw++] = src_data;
1383 ib->ptr[ib->length_dw++] = byte_count;
1384}
1385
1386static const struct amdgpu_buffer_funcs cik_sdma_buffer_funcs = {
1387 .copy_max_bytes = 0x1fffff,
1388 .copy_num_dw = 7,
1389 .emit_copy_buffer = cik_sdma_emit_copy_buffer,
1390
1391 .fill_max_bytes = 0x1fffff,
1392 .fill_num_dw = 5,
1393 .emit_fill_buffer = cik_sdma_emit_fill_buffer,
1394};
1395
1396static void cik_sdma_set_buffer_funcs(struct amdgpu_device *adev)
1397{
1398 if (adev->mman.buffer_funcs == NULL) {
1399 adev->mman.buffer_funcs = &cik_sdma_buffer_funcs;
1400 adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1401 }
1402}
1403
1404static const struct amdgpu_vm_pte_funcs cik_sdma_vm_pte_funcs = {
1405 .copy_pte = cik_sdma_vm_copy_pte,
1406 .write_pte = cik_sdma_vm_write_pte,
1407 .set_pte_pde = cik_sdma_vm_set_pte_pde,
1408};
1409
1410static void cik_sdma_set_vm_pte_funcs(struct amdgpu_device *adev)
1411{
1412 unsigned i;
1413
1414 if (adev->vm_manager.vm_pte_funcs == NULL) {
1415 adev->vm_manager.vm_pte_funcs = &cik_sdma_vm_pte_funcs;
1416 for (i = 0; i < adev->sdma.num_instances; i++)
1417 adev->vm_manager.vm_pte_rings[i] =
1418 &adev->sdma.instance[i].ring;
1419
1420 adev->vm_manager.vm_pte_num_rings = adev->sdma.num_instances;
1421 }
1422}
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
25#include <linux/firmware.h>
26#include <linux/module.h>
27
28#include "amdgpu.h"
29#include "amdgpu_ucode.h"
30#include "amdgpu_trace.h"
31#include "cikd.h"
32#include "cik.h"
33
34#include "bif/bif_4_1_d.h"
35#include "bif/bif_4_1_sh_mask.h"
36
37#include "gca/gfx_7_2_d.h"
38#include "gca/gfx_7_2_enum.h"
39#include "gca/gfx_7_2_sh_mask.h"
40
41#include "gmc/gmc_7_1_d.h"
42#include "gmc/gmc_7_1_sh_mask.h"
43
44#include "oss/oss_2_0_d.h"
45#include "oss/oss_2_0_sh_mask.h"
46
47static const u32 sdma_offsets[SDMA_MAX_INSTANCE] =
48{
49 SDMA0_REGISTER_OFFSET,
50 SDMA1_REGISTER_OFFSET
51};
52
53static void cik_sdma_set_ring_funcs(struct amdgpu_device *adev);
54static void cik_sdma_set_irq_funcs(struct amdgpu_device *adev);
55static void cik_sdma_set_buffer_funcs(struct amdgpu_device *adev);
56static void cik_sdma_set_vm_pte_funcs(struct amdgpu_device *adev);
57static int cik_sdma_soft_reset(struct amdgpu_ip_block *ip_block);
58
59MODULE_FIRMWARE("amdgpu/bonaire_sdma.bin");
60MODULE_FIRMWARE("amdgpu/bonaire_sdma1.bin");
61MODULE_FIRMWARE("amdgpu/hawaii_sdma.bin");
62MODULE_FIRMWARE("amdgpu/hawaii_sdma1.bin");
63MODULE_FIRMWARE("amdgpu/kaveri_sdma.bin");
64MODULE_FIRMWARE("amdgpu/kaveri_sdma1.bin");
65MODULE_FIRMWARE("amdgpu/kabini_sdma.bin");
66MODULE_FIRMWARE("amdgpu/kabini_sdma1.bin");
67MODULE_FIRMWARE("amdgpu/mullins_sdma.bin");
68MODULE_FIRMWARE("amdgpu/mullins_sdma1.bin");
69
70u32 amdgpu_cik_gpu_check_soft_reset(struct amdgpu_device *adev);
71
72
73static void cik_sdma_free_microcode(struct amdgpu_device *adev)
74{
75 int i;
76
77 for (i = 0; i < adev->sdma.num_instances; i++)
78 amdgpu_ucode_release(&adev->sdma.instance[i].fw);
79}
80
81/*
82 * sDMA - System DMA
83 * Starting with CIK, the GPU has new asynchronous
84 * DMA engines. These engines are used for compute
85 * and gfx. There are two DMA engines (SDMA0, SDMA1)
86 * and each one supports 1 ring buffer used for gfx
87 * and 2 queues used for compute.
88 *
89 * The programming model is very similar to the CP
90 * (ring buffer, IBs, etc.), but sDMA has it's own
91 * packet format that is different from the PM4 format
92 * used by the CP. sDMA supports copying data, writing
93 * embedded data, solid fills, and a number of other
94 * things. It also has support for tiling/detiling of
95 * buffers.
96 */
97
98/**
99 * cik_sdma_init_microcode - load ucode images from disk
100 *
101 * @adev: amdgpu_device pointer
102 *
103 * Use the firmware interface to load the ucode images into
104 * the driver (not loaded into hw).
105 * Returns 0 on success, error on failure.
106 */
107static int cik_sdma_init_microcode(struct amdgpu_device *adev)
108{
109 const char *chip_name;
110 int err = 0, i;
111
112 DRM_DEBUG("\n");
113
114 switch (adev->asic_type) {
115 case CHIP_BONAIRE:
116 chip_name = "bonaire";
117 break;
118 case CHIP_HAWAII:
119 chip_name = "hawaii";
120 break;
121 case CHIP_KAVERI:
122 chip_name = "kaveri";
123 break;
124 case CHIP_KABINI:
125 chip_name = "kabini";
126 break;
127 case CHIP_MULLINS:
128 chip_name = "mullins";
129 break;
130 default: BUG();
131 }
132
133 for (i = 0; i < adev->sdma.num_instances; i++) {
134 if (i == 0)
135 err = amdgpu_ucode_request(adev, &adev->sdma.instance[i].fw,
136 "amdgpu/%s_sdma.bin", chip_name);
137 else
138 err = amdgpu_ucode_request(adev, &adev->sdma.instance[i].fw,
139 "amdgpu/%s_sdma1.bin", chip_name);
140 if (err)
141 goto out;
142 }
143out:
144 if (err) {
145 pr_err("cik_sdma: Failed to load firmware \"%s_sdma%s.bin\"\n",
146 chip_name, i == 0 ? "" : "1");
147 for (i = 0; i < adev->sdma.num_instances; i++)
148 amdgpu_ucode_release(&adev->sdma.instance[i].fw);
149 }
150 return err;
151}
152
153/**
154 * cik_sdma_ring_get_rptr - get the current read pointer
155 *
156 * @ring: amdgpu ring pointer
157 *
158 * Get the current rptr from the hardware (CIK+).
159 */
160static uint64_t cik_sdma_ring_get_rptr(struct amdgpu_ring *ring)
161{
162 u32 rptr;
163
164 rptr = *ring->rptr_cpu_addr;
165
166 return (rptr & 0x3fffc) >> 2;
167}
168
169/**
170 * cik_sdma_ring_get_wptr - get the current write pointer
171 *
172 * @ring: amdgpu ring pointer
173 *
174 * Get the current wptr from the hardware (CIK+).
175 */
176static uint64_t cik_sdma_ring_get_wptr(struct amdgpu_ring *ring)
177{
178 struct amdgpu_device *adev = ring->adev;
179
180 return (RREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[ring->me]) & 0x3fffc) >> 2;
181}
182
183/**
184 * cik_sdma_ring_set_wptr - commit the write pointer
185 *
186 * @ring: amdgpu ring pointer
187 *
188 * Write the wptr back to the hardware (CIK+).
189 */
190static void cik_sdma_ring_set_wptr(struct amdgpu_ring *ring)
191{
192 struct amdgpu_device *adev = ring->adev;
193
194 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[ring->me],
195 (ring->wptr << 2) & 0x3fffc);
196}
197
198static void cik_sdma_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
199{
200 struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
201 int i;
202
203 for (i = 0; i < count; i++)
204 if (sdma && sdma->burst_nop && (i == 0))
205 amdgpu_ring_write(ring, ring->funcs->nop |
206 SDMA_NOP_COUNT(count - 1));
207 else
208 amdgpu_ring_write(ring, ring->funcs->nop);
209}
210
211/**
212 * cik_sdma_ring_emit_ib - Schedule an IB on the DMA engine
213 *
214 * @ring: amdgpu ring pointer
215 * @job: job to retrive vmid from
216 * @ib: IB object to schedule
217 * @flags: unused
218 *
219 * Schedule an IB in the DMA ring (CIK).
220 */
221static void cik_sdma_ring_emit_ib(struct amdgpu_ring *ring,
222 struct amdgpu_job *job,
223 struct amdgpu_ib *ib,
224 uint32_t flags)
225{
226 unsigned vmid = AMDGPU_JOB_GET_VMID(job);
227 u32 extra_bits = vmid & 0xf;
228
229 /* IB packet must end on a 8 DW boundary */
230 cik_sdma_ring_insert_nop(ring, (4 - lower_32_bits(ring->wptr)) & 7);
231
232 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_INDIRECT_BUFFER, 0, extra_bits));
233 amdgpu_ring_write(ring, ib->gpu_addr & 0xffffffe0); /* base must be 32 byte aligned */
234 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr) & 0xffffffff);
235 amdgpu_ring_write(ring, ib->length_dw);
236
237}
238
239/**
240 * cik_sdma_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
241 *
242 * @ring: amdgpu ring pointer
243 *
244 * Emit an hdp flush packet on the requested DMA ring.
245 */
246static void cik_sdma_ring_emit_hdp_flush(struct amdgpu_ring *ring)
247{
248 u32 extra_bits = (SDMA_POLL_REG_MEM_EXTRA_OP(1) |
249 SDMA_POLL_REG_MEM_EXTRA_FUNC(3)); /* == */
250 u32 ref_and_mask;
251
252 if (ring->me == 0)
253 ref_and_mask = GPU_HDP_FLUSH_DONE__SDMA0_MASK;
254 else
255 ref_and_mask = GPU_HDP_FLUSH_DONE__SDMA1_MASK;
256
257 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0, extra_bits));
258 amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_DONE << 2);
259 amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_REQ << 2);
260 amdgpu_ring_write(ring, ref_and_mask); /* reference */
261 amdgpu_ring_write(ring, ref_and_mask); /* mask */
262 amdgpu_ring_write(ring, (0xfff << 16) | 10); /* retry count, poll interval */
263}
264
265/**
266 * cik_sdma_ring_emit_fence - emit a fence on the DMA ring
267 *
268 * @ring: amdgpu ring pointer
269 * @addr: address
270 * @seq: sequence number
271 * @flags: fence related flags
272 *
273 * Add a DMA fence packet to the ring to write
274 * the fence seq number and DMA trap packet to generate
275 * an interrupt if needed (CIK).
276 */
277static void cik_sdma_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
278 unsigned flags)
279{
280 bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
281 /* write the fence */
282 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_FENCE, 0, 0));
283 amdgpu_ring_write(ring, lower_32_bits(addr));
284 amdgpu_ring_write(ring, upper_32_bits(addr));
285 amdgpu_ring_write(ring, lower_32_bits(seq));
286
287 /* optionally write high bits as well */
288 if (write64bit) {
289 addr += 4;
290 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_FENCE, 0, 0));
291 amdgpu_ring_write(ring, lower_32_bits(addr));
292 amdgpu_ring_write(ring, upper_32_bits(addr));
293 amdgpu_ring_write(ring, upper_32_bits(seq));
294 }
295
296 /* generate an interrupt */
297 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_TRAP, 0, 0));
298}
299
300/**
301 * cik_sdma_gfx_stop - stop the gfx async dma engines
302 *
303 * @adev: amdgpu_device pointer
304 *
305 * Stop the gfx async dma ring buffers (CIK).
306 */
307static void cik_sdma_gfx_stop(struct amdgpu_device *adev)
308{
309 u32 rb_cntl;
310 int i;
311
312 for (i = 0; i < adev->sdma.num_instances; i++) {
313 rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
314 rb_cntl &= ~SDMA0_GFX_RB_CNTL__RB_ENABLE_MASK;
315 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
316 WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], 0);
317 }
318}
319
320/**
321 * cik_sdma_rlc_stop - stop the compute async dma engines
322 *
323 * @adev: amdgpu_device pointer
324 *
325 * Stop the compute async dma queues (CIK).
326 */
327static void cik_sdma_rlc_stop(struct amdgpu_device *adev)
328{
329 /* XXX todo */
330}
331
332/**
333 * cik_ctx_switch_enable - stop the async dma engines context switch
334 *
335 * @adev: amdgpu_device pointer
336 * @enable: enable/disable the DMA MEs context switch.
337 *
338 * Halt or unhalt the async dma engines context switch (VI).
339 */
340static void cik_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
341{
342 u32 f32_cntl, phase_quantum = 0;
343 int i;
344
345 if (amdgpu_sdma_phase_quantum) {
346 unsigned value = amdgpu_sdma_phase_quantum;
347 unsigned unit = 0;
348
349 while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
350 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
351 value = (value + 1) >> 1;
352 unit++;
353 }
354 if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
355 SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
356 value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
357 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
358 unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
359 SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
360 WARN_ONCE(1,
361 "clamping sdma_phase_quantum to %uK clock cycles\n",
362 value << unit);
363 }
364 phase_quantum =
365 value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
366 unit << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
367 }
368
369 for (i = 0; i < adev->sdma.num_instances; i++) {
370 f32_cntl = RREG32(mmSDMA0_CNTL + sdma_offsets[i]);
371 if (enable) {
372 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
373 AUTO_CTXSW_ENABLE, 1);
374 if (amdgpu_sdma_phase_quantum) {
375 WREG32(mmSDMA0_PHASE0_QUANTUM + sdma_offsets[i],
376 phase_quantum);
377 WREG32(mmSDMA0_PHASE1_QUANTUM + sdma_offsets[i],
378 phase_quantum);
379 }
380 } else {
381 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
382 AUTO_CTXSW_ENABLE, 0);
383 }
384
385 WREG32(mmSDMA0_CNTL + sdma_offsets[i], f32_cntl);
386 }
387}
388
389/**
390 * cik_sdma_enable - stop the async dma engines
391 *
392 * @adev: amdgpu_device pointer
393 * @enable: enable/disable the DMA MEs.
394 *
395 * Halt or unhalt the async dma engines (CIK).
396 */
397static void cik_sdma_enable(struct amdgpu_device *adev, bool enable)
398{
399 u32 me_cntl;
400 int i;
401
402 if (!enable) {
403 cik_sdma_gfx_stop(adev);
404 cik_sdma_rlc_stop(adev);
405 }
406
407 for (i = 0; i < adev->sdma.num_instances; i++) {
408 me_cntl = RREG32(mmSDMA0_F32_CNTL + sdma_offsets[i]);
409 if (enable)
410 me_cntl &= ~SDMA0_F32_CNTL__HALT_MASK;
411 else
412 me_cntl |= SDMA0_F32_CNTL__HALT_MASK;
413 WREG32(mmSDMA0_F32_CNTL + sdma_offsets[i], me_cntl);
414 }
415}
416
417/**
418 * cik_sdma_gfx_resume - setup and start the async dma engines
419 *
420 * @adev: amdgpu_device pointer
421 *
422 * Set up the gfx DMA ring buffers and enable them (CIK).
423 * Returns 0 for success, error for failure.
424 */
425static int cik_sdma_gfx_resume(struct amdgpu_device *adev)
426{
427 struct amdgpu_ring *ring;
428 u32 rb_cntl, ib_cntl;
429 u32 rb_bufsz;
430 int i, j, r;
431
432 for (i = 0; i < adev->sdma.num_instances; i++) {
433 ring = &adev->sdma.instance[i].ring;
434
435 mutex_lock(&adev->srbm_mutex);
436 for (j = 0; j < 16; j++) {
437 cik_srbm_select(adev, 0, 0, 0, j);
438 /* SDMA GFX */
439 WREG32(mmSDMA0_GFX_VIRTUAL_ADDR + sdma_offsets[i], 0);
440 WREG32(mmSDMA0_GFX_APE1_CNTL + sdma_offsets[i], 0);
441 /* XXX SDMA RLC - todo */
442 }
443 cik_srbm_select(adev, 0, 0, 0, 0);
444 mutex_unlock(&adev->srbm_mutex);
445
446 WREG32(mmSDMA0_TILING_CONFIG + sdma_offsets[i],
447 adev->gfx.config.gb_addr_config & 0x70);
448
449 WREG32(mmSDMA0_SEM_INCOMPLETE_TIMER_CNTL + sdma_offsets[i], 0);
450 WREG32(mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
451
452 /* Set ring buffer size in dwords */
453 rb_bufsz = order_base_2(ring->ring_size / 4);
454 rb_cntl = rb_bufsz << 1;
455#ifdef __BIG_ENDIAN
456 rb_cntl |= SDMA0_GFX_RB_CNTL__RB_SWAP_ENABLE_MASK |
457 SDMA0_GFX_RB_CNTL__RPTR_WRITEBACK_SWAP_ENABLE_MASK;
458#endif
459 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
460
461 /* Initialize the ring buffer's read and write pointers */
462 WREG32(mmSDMA0_GFX_RB_RPTR + sdma_offsets[i], 0);
463 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], 0);
464 WREG32(mmSDMA0_GFX_IB_RPTR + sdma_offsets[i], 0);
465 WREG32(mmSDMA0_GFX_IB_OFFSET + sdma_offsets[i], 0);
466
467 /* set the wb address whether it's enabled or not */
468 WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_HI + sdma_offsets[i],
469 upper_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFF);
470 WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_LO + sdma_offsets[i],
471 ((ring->rptr_gpu_addr) & 0xFFFFFFFC));
472
473 rb_cntl |= SDMA0_GFX_RB_CNTL__RPTR_WRITEBACK_ENABLE_MASK;
474
475 WREG32(mmSDMA0_GFX_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
476 WREG32(mmSDMA0_GFX_RB_BASE_HI + sdma_offsets[i], ring->gpu_addr >> 40);
477
478 ring->wptr = 0;
479 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], ring->wptr << 2);
480
481 /* enable DMA RB */
482 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i],
483 rb_cntl | SDMA0_GFX_RB_CNTL__RB_ENABLE_MASK);
484
485 ib_cntl = SDMA0_GFX_IB_CNTL__IB_ENABLE_MASK;
486#ifdef __BIG_ENDIAN
487 ib_cntl |= SDMA0_GFX_IB_CNTL__IB_SWAP_ENABLE_MASK;
488#endif
489 /* enable DMA IBs */
490 WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
491 }
492
493 cik_sdma_enable(adev, true);
494
495 for (i = 0; i < adev->sdma.num_instances; i++) {
496 ring = &adev->sdma.instance[i].ring;
497 r = amdgpu_ring_test_helper(ring);
498 if (r)
499 return r;
500 }
501
502 return 0;
503}
504
505/**
506 * cik_sdma_rlc_resume - setup and start the async dma engines
507 *
508 * @adev: amdgpu_device pointer
509 *
510 * Set up the compute DMA queues and enable them (CIK).
511 * Returns 0 for success, error for failure.
512 */
513static int cik_sdma_rlc_resume(struct amdgpu_device *adev)
514{
515 /* XXX todo */
516 return 0;
517}
518
519/**
520 * cik_sdma_load_microcode - load the sDMA ME ucode
521 *
522 * @adev: amdgpu_device pointer
523 *
524 * Loads the sDMA0/1 ucode.
525 * Returns 0 for success, -EINVAL if the ucode is not available.
526 */
527static int cik_sdma_load_microcode(struct amdgpu_device *adev)
528{
529 const struct sdma_firmware_header_v1_0 *hdr;
530 const __le32 *fw_data;
531 u32 fw_size;
532 int i, j;
533
534 /* halt the MEs */
535 cik_sdma_enable(adev, false);
536
537 for (i = 0; i < adev->sdma.num_instances; i++) {
538 if (!adev->sdma.instance[i].fw)
539 return -EINVAL;
540 hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
541 amdgpu_ucode_print_sdma_hdr(&hdr->header);
542 fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
543 adev->sdma.instance[i].fw_version = le32_to_cpu(hdr->header.ucode_version);
544 adev->sdma.instance[i].feature_version = le32_to_cpu(hdr->ucode_feature_version);
545 if (adev->sdma.instance[i].feature_version >= 20)
546 adev->sdma.instance[i].burst_nop = true;
547 fw_data = (const __le32 *)
548 (adev->sdma.instance[i].fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
549 WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], 0);
550 for (j = 0; j < fw_size; j++)
551 WREG32(mmSDMA0_UCODE_DATA + sdma_offsets[i], le32_to_cpup(fw_data++));
552 WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], adev->sdma.instance[i].fw_version);
553 }
554
555 return 0;
556}
557
558/**
559 * cik_sdma_start - setup and start the async dma engines
560 *
561 * @adev: amdgpu_device pointer
562 *
563 * Set up the DMA engines and enable them (CIK).
564 * Returns 0 for success, error for failure.
565 */
566static int cik_sdma_start(struct amdgpu_device *adev)
567{
568 int r;
569
570 r = cik_sdma_load_microcode(adev);
571 if (r)
572 return r;
573
574 /* halt the engine before programing */
575 cik_sdma_enable(adev, false);
576 /* enable sdma ring preemption */
577 cik_ctx_switch_enable(adev, true);
578
579 /* start the gfx rings and rlc compute queues */
580 r = cik_sdma_gfx_resume(adev);
581 if (r)
582 return r;
583 r = cik_sdma_rlc_resume(adev);
584 if (r)
585 return r;
586
587 return 0;
588}
589
590/**
591 * cik_sdma_ring_test_ring - simple async dma engine test
592 *
593 * @ring: amdgpu_ring structure holding ring information
594 *
595 * Test the DMA engine by writing using it to write an
596 * value to memory. (CIK).
597 * Returns 0 for success, error for failure.
598 */
599static int cik_sdma_ring_test_ring(struct amdgpu_ring *ring)
600{
601 struct amdgpu_device *adev = ring->adev;
602 unsigned i;
603 unsigned index;
604 int r;
605 u32 tmp;
606 u64 gpu_addr;
607
608 r = amdgpu_device_wb_get(adev, &index);
609 if (r)
610 return r;
611
612 gpu_addr = adev->wb.gpu_addr + (index * 4);
613 tmp = 0xCAFEDEAD;
614 adev->wb.wb[index] = cpu_to_le32(tmp);
615
616 r = amdgpu_ring_alloc(ring, 5);
617 if (r)
618 goto error_free_wb;
619
620 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_WRITE, SDMA_WRITE_SUB_OPCODE_LINEAR, 0));
621 amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
622 amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
623 amdgpu_ring_write(ring, 1); /* number of DWs to follow */
624 amdgpu_ring_write(ring, 0xDEADBEEF);
625 amdgpu_ring_commit(ring);
626
627 for (i = 0; i < adev->usec_timeout; i++) {
628 tmp = le32_to_cpu(adev->wb.wb[index]);
629 if (tmp == 0xDEADBEEF)
630 break;
631 udelay(1);
632 }
633
634 if (i >= adev->usec_timeout)
635 r = -ETIMEDOUT;
636
637error_free_wb:
638 amdgpu_device_wb_free(adev, index);
639 return r;
640}
641
642/**
643 * cik_sdma_ring_test_ib - test an IB on the DMA engine
644 *
645 * @ring: amdgpu_ring structure holding ring information
646 * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
647 *
648 * Test a simple IB in the DMA ring (CIK).
649 * Returns 0 on success, error on failure.
650 */
651static int cik_sdma_ring_test_ib(struct amdgpu_ring *ring, long timeout)
652{
653 struct amdgpu_device *adev = ring->adev;
654 struct amdgpu_ib ib;
655 struct dma_fence *f = NULL;
656 unsigned index;
657 u32 tmp = 0;
658 u64 gpu_addr;
659 long r;
660
661 r = amdgpu_device_wb_get(adev, &index);
662 if (r)
663 return r;
664
665 gpu_addr = adev->wb.gpu_addr + (index * 4);
666 tmp = 0xCAFEDEAD;
667 adev->wb.wb[index] = cpu_to_le32(tmp);
668 memset(&ib, 0, sizeof(ib));
669 r = amdgpu_ib_get(adev, NULL, 256,
670 AMDGPU_IB_POOL_DIRECT, &ib);
671 if (r)
672 goto err0;
673
674 ib.ptr[0] = SDMA_PACKET(SDMA_OPCODE_WRITE,
675 SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
676 ib.ptr[1] = lower_32_bits(gpu_addr);
677 ib.ptr[2] = upper_32_bits(gpu_addr);
678 ib.ptr[3] = 1;
679 ib.ptr[4] = 0xDEADBEEF;
680 ib.length_dw = 5;
681 r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
682 if (r)
683 goto err1;
684
685 r = dma_fence_wait_timeout(f, false, timeout);
686 if (r == 0) {
687 r = -ETIMEDOUT;
688 goto err1;
689 } else if (r < 0) {
690 goto err1;
691 }
692 tmp = le32_to_cpu(adev->wb.wb[index]);
693 if (tmp == 0xDEADBEEF)
694 r = 0;
695 else
696 r = -EINVAL;
697
698err1:
699 amdgpu_ib_free(adev, &ib, NULL);
700 dma_fence_put(f);
701err0:
702 amdgpu_device_wb_free(adev, index);
703 return r;
704}
705
706/**
707 * cik_sdma_vm_copy_pte - update PTEs by copying them from the GART
708 *
709 * @ib: indirect buffer to fill with commands
710 * @pe: addr of the page entry
711 * @src: src addr to copy from
712 * @count: number of page entries to update
713 *
714 * Update PTEs by copying them from the GART using sDMA (CIK).
715 */
716static void cik_sdma_vm_copy_pte(struct amdgpu_ib *ib,
717 uint64_t pe, uint64_t src,
718 unsigned count)
719{
720 unsigned bytes = count * 8;
721
722 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_COPY,
723 SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
724 ib->ptr[ib->length_dw++] = bytes;
725 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
726 ib->ptr[ib->length_dw++] = lower_32_bits(src);
727 ib->ptr[ib->length_dw++] = upper_32_bits(src);
728 ib->ptr[ib->length_dw++] = lower_32_bits(pe);
729 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
730}
731
732/**
733 * cik_sdma_vm_write_pte - update PTEs by writing them manually
734 *
735 * @ib: indirect buffer to fill with commands
736 * @pe: addr of the page entry
737 * @value: dst addr to write into pe
738 * @count: number of page entries to update
739 * @incr: increase next addr by incr bytes
740 *
741 * Update PTEs by writing them manually using sDMA (CIK).
742 */
743static void cik_sdma_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
744 uint64_t value, unsigned count,
745 uint32_t incr)
746{
747 unsigned ndw = count * 2;
748
749 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_WRITE,
750 SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
751 ib->ptr[ib->length_dw++] = lower_32_bits(pe);
752 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
753 ib->ptr[ib->length_dw++] = ndw;
754 for (; ndw > 0; ndw -= 2) {
755 ib->ptr[ib->length_dw++] = lower_32_bits(value);
756 ib->ptr[ib->length_dw++] = upper_32_bits(value);
757 value += incr;
758 }
759}
760
761/**
762 * cik_sdma_vm_set_pte_pde - update the page tables using sDMA
763 *
764 * @ib: indirect buffer to fill with commands
765 * @pe: addr of the page entry
766 * @addr: dst addr to write into pe
767 * @count: number of page entries to update
768 * @incr: increase next addr by incr bytes
769 * @flags: access flags
770 *
771 * Update the page tables using sDMA (CIK).
772 */
773static void cik_sdma_vm_set_pte_pde(struct amdgpu_ib *ib, uint64_t pe,
774 uint64_t addr, unsigned count,
775 uint32_t incr, uint64_t flags)
776{
777 /* for physically contiguous pages (vram) */
778 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_GENERATE_PTE_PDE, 0, 0);
779 ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
780 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
781 ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
782 ib->ptr[ib->length_dw++] = upper_32_bits(flags);
783 ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
784 ib->ptr[ib->length_dw++] = upper_32_bits(addr);
785 ib->ptr[ib->length_dw++] = incr; /* increment size */
786 ib->ptr[ib->length_dw++] = 0;
787 ib->ptr[ib->length_dw++] = count; /* number of entries */
788}
789
790/**
791 * cik_sdma_ring_pad_ib - pad the IB to the required number of dw
792 *
793 * @ring: amdgpu_ring structure holding ring information
794 * @ib: indirect buffer to fill with padding
795 *
796 */
797static void cik_sdma_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
798{
799 struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
800 u32 pad_count;
801 int i;
802
803 pad_count = (-ib->length_dw) & 7;
804 for (i = 0; i < pad_count; i++)
805 if (sdma && sdma->burst_nop && (i == 0))
806 ib->ptr[ib->length_dw++] =
807 SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0) |
808 SDMA_NOP_COUNT(pad_count - 1);
809 else
810 ib->ptr[ib->length_dw++] =
811 SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0);
812}
813
814/**
815 * cik_sdma_ring_emit_pipeline_sync - sync the pipeline
816 *
817 * @ring: amdgpu_ring pointer
818 *
819 * Make sure all previous operations are completed (CIK).
820 */
821static void cik_sdma_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
822{
823 uint32_t seq = ring->fence_drv.sync_seq;
824 uint64_t addr = ring->fence_drv.gpu_addr;
825
826 /* wait for idle */
827 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0,
828 SDMA_POLL_REG_MEM_EXTRA_OP(0) |
829 SDMA_POLL_REG_MEM_EXTRA_FUNC(3) | /* equal */
830 SDMA_POLL_REG_MEM_EXTRA_M));
831 amdgpu_ring_write(ring, addr & 0xfffffffc);
832 amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
833 amdgpu_ring_write(ring, seq); /* reference */
834 amdgpu_ring_write(ring, 0xffffffff); /* mask */
835 amdgpu_ring_write(ring, (0xfff << 16) | 4); /* retry count, poll interval */
836}
837
838/**
839 * cik_sdma_ring_emit_vm_flush - cik vm flush using sDMA
840 *
841 * @ring: amdgpu_ring pointer
842 * @vmid: vmid number to use
843 * @pd_addr: address
844 *
845 * Update the page table base and flush the VM TLB
846 * using sDMA (CIK).
847 */
848static void cik_sdma_ring_emit_vm_flush(struct amdgpu_ring *ring,
849 unsigned vmid, uint64_t pd_addr)
850{
851 u32 extra_bits = (SDMA_POLL_REG_MEM_EXTRA_OP(0) |
852 SDMA_POLL_REG_MEM_EXTRA_FUNC(0)); /* always */
853
854 amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
855
856 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0, extra_bits));
857 amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST << 2);
858 amdgpu_ring_write(ring, 0);
859 amdgpu_ring_write(ring, 0); /* reference */
860 amdgpu_ring_write(ring, 0); /* mask */
861 amdgpu_ring_write(ring, (0xfff << 16) | 10); /* retry count, poll interval */
862}
863
864static void cik_sdma_ring_emit_wreg(struct amdgpu_ring *ring,
865 uint32_t reg, uint32_t val)
866{
867 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
868 amdgpu_ring_write(ring, reg);
869 amdgpu_ring_write(ring, val);
870}
871
872static void cik_enable_sdma_mgcg(struct amdgpu_device *adev,
873 bool enable)
874{
875 u32 orig, data;
876
877 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
878 WREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET, 0x00000100);
879 WREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET, 0x00000100);
880 } else {
881 orig = data = RREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET);
882 data |= 0xff000000;
883 if (data != orig)
884 WREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET, data);
885
886 orig = data = RREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET);
887 data |= 0xff000000;
888 if (data != orig)
889 WREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET, data);
890 }
891}
892
893static void cik_enable_sdma_mgls(struct amdgpu_device *adev,
894 bool enable)
895{
896 u32 orig, data;
897
898 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
899 orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET);
900 data |= 0x100;
901 if (orig != data)
902 WREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET, data);
903
904 orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET);
905 data |= 0x100;
906 if (orig != data)
907 WREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET, data);
908 } else {
909 orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET);
910 data &= ~0x100;
911 if (orig != data)
912 WREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET, data);
913
914 orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET);
915 data &= ~0x100;
916 if (orig != data)
917 WREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET, data);
918 }
919}
920
921static int cik_sdma_early_init(struct amdgpu_ip_block *ip_block)
922{
923 struct amdgpu_device *adev = ip_block->adev;
924 int r;
925
926 adev->sdma.num_instances = SDMA_MAX_INSTANCE;
927
928 r = cik_sdma_init_microcode(adev);
929 if (r)
930 return r;
931
932 cik_sdma_set_ring_funcs(adev);
933 cik_sdma_set_irq_funcs(adev);
934 cik_sdma_set_buffer_funcs(adev);
935 cik_sdma_set_vm_pte_funcs(adev);
936
937 return 0;
938}
939
940static int cik_sdma_sw_init(struct amdgpu_ip_block *ip_block)
941{
942 struct amdgpu_ring *ring;
943 struct amdgpu_device *adev = ip_block->adev;
944 int r, i;
945
946 /* SDMA trap event */
947 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 224,
948 &adev->sdma.trap_irq);
949 if (r)
950 return r;
951
952 /* SDMA Privileged inst */
953 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 241,
954 &adev->sdma.illegal_inst_irq);
955 if (r)
956 return r;
957
958 /* SDMA Privileged inst */
959 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 247,
960 &adev->sdma.illegal_inst_irq);
961 if (r)
962 return r;
963
964 for (i = 0; i < adev->sdma.num_instances; i++) {
965 ring = &adev->sdma.instance[i].ring;
966 ring->ring_obj = NULL;
967 sprintf(ring->name, "sdma%d", i);
968 r = amdgpu_ring_init(adev, ring, 1024,
969 &adev->sdma.trap_irq,
970 (i == 0) ? AMDGPU_SDMA_IRQ_INSTANCE0 :
971 AMDGPU_SDMA_IRQ_INSTANCE1,
972 AMDGPU_RING_PRIO_DEFAULT, NULL);
973 if (r)
974 return r;
975 }
976
977 return r;
978}
979
980static int cik_sdma_sw_fini(struct amdgpu_ip_block *ip_block)
981{
982 struct amdgpu_device *adev = ip_block->adev;
983 int i;
984
985 for (i = 0; i < adev->sdma.num_instances; i++)
986 amdgpu_ring_fini(&adev->sdma.instance[i].ring);
987
988 cik_sdma_free_microcode(adev);
989 return 0;
990}
991
992static int cik_sdma_hw_init(struct amdgpu_ip_block *ip_block)
993{
994 int r;
995 struct amdgpu_device *adev = ip_block->adev;
996
997 r = cik_sdma_start(adev);
998 if (r)
999 return r;
1000
1001 return r;
1002}
1003
1004static int cik_sdma_hw_fini(struct amdgpu_ip_block *ip_block)
1005{
1006 struct amdgpu_device *adev = ip_block->adev;
1007
1008 cik_ctx_switch_enable(adev, false);
1009 cik_sdma_enable(adev, false);
1010
1011 return 0;
1012}
1013
1014static int cik_sdma_suspend(struct amdgpu_ip_block *ip_block)
1015{
1016 return cik_sdma_hw_fini(ip_block);
1017}
1018
1019static int cik_sdma_resume(struct amdgpu_ip_block *ip_block)
1020{
1021 cik_sdma_soft_reset(ip_block);
1022
1023 return cik_sdma_hw_init(ip_block);
1024}
1025
1026static bool cik_sdma_is_idle(void *handle)
1027{
1028 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1029 u32 tmp = RREG32(mmSRBM_STATUS2);
1030
1031 if (tmp & (SRBM_STATUS2__SDMA_BUSY_MASK |
1032 SRBM_STATUS2__SDMA1_BUSY_MASK))
1033 return false;
1034
1035 return true;
1036}
1037
1038static int cik_sdma_wait_for_idle(struct amdgpu_ip_block *ip_block)
1039{
1040 unsigned i;
1041 u32 tmp;
1042 struct amdgpu_device *adev = ip_block->adev;
1043
1044 for (i = 0; i < adev->usec_timeout; i++) {
1045 tmp = RREG32(mmSRBM_STATUS2) & (SRBM_STATUS2__SDMA_BUSY_MASK |
1046 SRBM_STATUS2__SDMA1_BUSY_MASK);
1047
1048 if (!tmp)
1049 return 0;
1050 udelay(1);
1051 }
1052 return -ETIMEDOUT;
1053}
1054
1055static int cik_sdma_soft_reset(struct amdgpu_ip_block *ip_block)
1056{
1057 u32 srbm_soft_reset = 0;
1058 struct amdgpu_device *adev = ip_block->adev;
1059 u32 tmp;
1060
1061 /* sdma0 */
1062 tmp = RREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET);
1063 tmp |= SDMA0_F32_CNTL__HALT_MASK;
1064 WREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET, tmp);
1065 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA_MASK;
1066
1067 /* sdma1 */
1068 tmp = RREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET);
1069 tmp |= SDMA0_F32_CNTL__HALT_MASK;
1070 WREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET, tmp);
1071 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA1_MASK;
1072
1073 if (srbm_soft_reset) {
1074 tmp = RREG32(mmSRBM_SOFT_RESET);
1075 tmp |= srbm_soft_reset;
1076 dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
1077 WREG32(mmSRBM_SOFT_RESET, tmp);
1078 tmp = RREG32(mmSRBM_SOFT_RESET);
1079
1080 udelay(50);
1081
1082 tmp &= ~srbm_soft_reset;
1083 WREG32(mmSRBM_SOFT_RESET, tmp);
1084 tmp = RREG32(mmSRBM_SOFT_RESET);
1085
1086 /* Wait a little for things to settle down */
1087 udelay(50);
1088 }
1089
1090 return 0;
1091}
1092
1093static int cik_sdma_set_trap_irq_state(struct amdgpu_device *adev,
1094 struct amdgpu_irq_src *src,
1095 unsigned type,
1096 enum amdgpu_interrupt_state state)
1097{
1098 u32 sdma_cntl;
1099
1100 switch (type) {
1101 case AMDGPU_SDMA_IRQ_INSTANCE0:
1102 switch (state) {
1103 case AMDGPU_IRQ_STATE_DISABLE:
1104 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1105 sdma_cntl &= ~SDMA0_CNTL__TRAP_ENABLE_MASK;
1106 WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1107 break;
1108 case AMDGPU_IRQ_STATE_ENABLE:
1109 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1110 sdma_cntl |= SDMA0_CNTL__TRAP_ENABLE_MASK;
1111 WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1112 break;
1113 default:
1114 break;
1115 }
1116 break;
1117 case AMDGPU_SDMA_IRQ_INSTANCE1:
1118 switch (state) {
1119 case AMDGPU_IRQ_STATE_DISABLE:
1120 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1121 sdma_cntl &= ~SDMA0_CNTL__TRAP_ENABLE_MASK;
1122 WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1123 break;
1124 case AMDGPU_IRQ_STATE_ENABLE:
1125 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1126 sdma_cntl |= SDMA0_CNTL__TRAP_ENABLE_MASK;
1127 WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1128 break;
1129 default:
1130 break;
1131 }
1132 break;
1133 default:
1134 break;
1135 }
1136 return 0;
1137}
1138
1139static int cik_sdma_process_trap_irq(struct amdgpu_device *adev,
1140 struct amdgpu_irq_src *source,
1141 struct amdgpu_iv_entry *entry)
1142{
1143 u8 instance_id, queue_id;
1144
1145 instance_id = (entry->ring_id & 0x3) >> 0;
1146 queue_id = (entry->ring_id & 0xc) >> 2;
1147 DRM_DEBUG("IH: SDMA trap\n");
1148 switch (instance_id) {
1149 case 0:
1150 switch (queue_id) {
1151 case 0:
1152 amdgpu_fence_process(&adev->sdma.instance[0].ring);
1153 break;
1154 case 1:
1155 /* XXX compute */
1156 break;
1157 case 2:
1158 /* XXX compute */
1159 break;
1160 }
1161 break;
1162 case 1:
1163 switch (queue_id) {
1164 case 0:
1165 amdgpu_fence_process(&adev->sdma.instance[1].ring);
1166 break;
1167 case 1:
1168 /* XXX compute */
1169 break;
1170 case 2:
1171 /* XXX compute */
1172 break;
1173 }
1174 break;
1175 }
1176
1177 return 0;
1178}
1179
1180static int cik_sdma_process_illegal_inst_irq(struct amdgpu_device *adev,
1181 struct amdgpu_irq_src *source,
1182 struct amdgpu_iv_entry *entry)
1183{
1184 u8 instance_id;
1185
1186 DRM_ERROR("Illegal instruction in SDMA command stream\n");
1187 instance_id = (entry->ring_id & 0x3) >> 0;
1188 drm_sched_fault(&adev->sdma.instance[instance_id].ring.sched);
1189 return 0;
1190}
1191
1192static int cik_sdma_set_clockgating_state(void *handle,
1193 enum amd_clockgating_state state)
1194{
1195 bool gate = false;
1196 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1197
1198 if (state == AMD_CG_STATE_GATE)
1199 gate = true;
1200
1201 cik_enable_sdma_mgcg(adev, gate);
1202 cik_enable_sdma_mgls(adev, gate);
1203
1204 return 0;
1205}
1206
1207static int cik_sdma_set_powergating_state(void *handle,
1208 enum amd_powergating_state state)
1209{
1210 return 0;
1211}
1212
1213static const struct amd_ip_funcs cik_sdma_ip_funcs = {
1214 .name = "cik_sdma",
1215 .early_init = cik_sdma_early_init,
1216 .sw_init = cik_sdma_sw_init,
1217 .sw_fini = cik_sdma_sw_fini,
1218 .hw_init = cik_sdma_hw_init,
1219 .hw_fini = cik_sdma_hw_fini,
1220 .suspend = cik_sdma_suspend,
1221 .resume = cik_sdma_resume,
1222 .is_idle = cik_sdma_is_idle,
1223 .wait_for_idle = cik_sdma_wait_for_idle,
1224 .soft_reset = cik_sdma_soft_reset,
1225 .set_clockgating_state = cik_sdma_set_clockgating_state,
1226 .set_powergating_state = cik_sdma_set_powergating_state,
1227};
1228
1229static const struct amdgpu_ring_funcs cik_sdma_ring_funcs = {
1230 .type = AMDGPU_RING_TYPE_SDMA,
1231 .align_mask = 0xf,
1232 .nop = SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0),
1233 .support_64bit_ptrs = false,
1234 .get_rptr = cik_sdma_ring_get_rptr,
1235 .get_wptr = cik_sdma_ring_get_wptr,
1236 .set_wptr = cik_sdma_ring_set_wptr,
1237 .emit_frame_size =
1238 6 + /* cik_sdma_ring_emit_hdp_flush */
1239 3 + /* hdp invalidate */
1240 6 + /* cik_sdma_ring_emit_pipeline_sync */
1241 CIK_FLUSH_GPU_TLB_NUM_WREG * 3 + 6 + /* cik_sdma_ring_emit_vm_flush */
1242 9 + 9 + 9, /* cik_sdma_ring_emit_fence x3 for user fence, vm fence */
1243 .emit_ib_size = 7 + 4, /* cik_sdma_ring_emit_ib */
1244 .emit_ib = cik_sdma_ring_emit_ib,
1245 .emit_fence = cik_sdma_ring_emit_fence,
1246 .emit_pipeline_sync = cik_sdma_ring_emit_pipeline_sync,
1247 .emit_vm_flush = cik_sdma_ring_emit_vm_flush,
1248 .emit_hdp_flush = cik_sdma_ring_emit_hdp_flush,
1249 .test_ring = cik_sdma_ring_test_ring,
1250 .test_ib = cik_sdma_ring_test_ib,
1251 .insert_nop = cik_sdma_ring_insert_nop,
1252 .pad_ib = cik_sdma_ring_pad_ib,
1253 .emit_wreg = cik_sdma_ring_emit_wreg,
1254};
1255
1256static void cik_sdma_set_ring_funcs(struct amdgpu_device *adev)
1257{
1258 int i;
1259
1260 for (i = 0; i < adev->sdma.num_instances; i++) {
1261 adev->sdma.instance[i].ring.funcs = &cik_sdma_ring_funcs;
1262 adev->sdma.instance[i].ring.me = i;
1263 }
1264}
1265
1266static const struct amdgpu_irq_src_funcs cik_sdma_trap_irq_funcs = {
1267 .set = cik_sdma_set_trap_irq_state,
1268 .process = cik_sdma_process_trap_irq,
1269};
1270
1271static const struct amdgpu_irq_src_funcs cik_sdma_illegal_inst_irq_funcs = {
1272 .process = cik_sdma_process_illegal_inst_irq,
1273};
1274
1275static void cik_sdma_set_irq_funcs(struct amdgpu_device *adev)
1276{
1277 adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
1278 adev->sdma.trap_irq.funcs = &cik_sdma_trap_irq_funcs;
1279 adev->sdma.illegal_inst_irq.funcs = &cik_sdma_illegal_inst_irq_funcs;
1280}
1281
1282/**
1283 * cik_sdma_emit_copy_buffer - copy buffer using the sDMA engine
1284 *
1285 * @ib: indirect buffer to copy to
1286 * @src_offset: src GPU address
1287 * @dst_offset: dst GPU address
1288 * @byte_count: number of bytes to xfer
1289 * @copy_flags: unused
1290 *
1291 * Copy GPU buffers using the DMA engine (CIK).
1292 * Used by the amdgpu ttm implementation to move pages if
1293 * registered as the asic copy callback.
1294 */
1295static void cik_sdma_emit_copy_buffer(struct amdgpu_ib *ib,
1296 uint64_t src_offset,
1297 uint64_t dst_offset,
1298 uint32_t byte_count,
1299 uint32_t copy_flags)
1300{
1301 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_COPY, SDMA_COPY_SUB_OPCODE_LINEAR, 0);
1302 ib->ptr[ib->length_dw++] = byte_count;
1303 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1304 ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1305 ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1306 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1307 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1308}
1309
1310/**
1311 * cik_sdma_emit_fill_buffer - fill buffer using the sDMA engine
1312 *
1313 * @ib: indirect buffer to fill
1314 * @src_data: value to write to buffer
1315 * @dst_offset: dst GPU address
1316 * @byte_count: number of bytes to xfer
1317 *
1318 * Fill GPU buffers using the DMA engine (CIK).
1319 */
1320static void cik_sdma_emit_fill_buffer(struct amdgpu_ib *ib,
1321 uint32_t src_data,
1322 uint64_t dst_offset,
1323 uint32_t byte_count)
1324{
1325 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_CONSTANT_FILL, 0, 0);
1326 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1327 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1328 ib->ptr[ib->length_dw++] = src_data;
1329 ib->ptr[ib->length_dw++] = byte_count;
1330}
1331
1332static const struct amdgpu_buffer_funcs cik_sdma_buffer_funcs = {
1333 .copy_max_bytes = 0x1fffff,
1334 .copy_num_dw = 7,
1335 .emit_copy_buffer = cik_sdma_emit_copy_buffer,
1336
1337 .fill_max_bytes = 0x1fffff,
1338 .fill_num_dw = 5,
1339 .emit_fill_buffer = cik_sdma_emit_fill_buffer,
1340};
1341
1342static void cik_sdma_set_buffer_funcs(struct amdgpu_device *adev)
1343{
1344 adev->mman.buffer_funcs = &cik_sdma_buffer_funcs;
1345 adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1346}
1347
1348static const struct amdgpu_vm_pte_funcs cik_sdma_vm_pte_funcs = {
1349 .copy_pte_num_dw = 7,
1350 .copy_pte = cik_sdma_vm_copy_pte,
1351
1352 .write_pte = cik_sdma_vm_write_pte,
1353 .set_pte_pde = cik_sdma_vm_set_pte_pde,
1354};
1355
1356static void cik_sdma_set_vm_pte_funcs(struct amdgpu_device *adev)
1357{
1358 unsigned i;
1359
1360 adev->vm_manager.vm_pte_funcs = &cik_sdma_vm_pte_funcs;
1361 for (i = 0; i < adev->sdma.num_instances; i++) {
1362 adev->vm_manager.vm_pte_scheds[i] =
1363 &adev->sdma.instance[i].ring.sched;
1364 }
1365 adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
1366}
1367
1368const struct amdgpu_ip_block_version cik_sdma_ip_block =
1369{
1370 .type = AMD_IP_BLOCK_TYPE_SDMA,
1371 .major = 2,
1372 .minor = 0,
1373 .rev = 0,
1374 .funcs = &cik_sdma_ip_funcs,
1375};