Loading...
1/*
2 * Copyright 2014 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/delay.h>
26#include <linux/firmware.h>
27#include <linux/module.h>
28
29#include "amdgpu.h"
30#include "amdgpu_ucode.h"
31#include "amdgpu_trace.h"
32#include "vi.h"
33#include "vid.h"
34
35#include "oss/oss_3_0_d.h"
36#include "oss/oss_3_0_sh_mask.h"
37
38#include "gmc/gmc_8_1_d.h"
39#include "gmc/gmc_8_1_sh_mask.h"
40
41#include "gca/gfx_8_0_d.h"
42#include "gca/gfx_8_0_enum.h"
43#include "gca/gfx_8_0_sh_mask.h"
44
45#include "bif/bif_5_0_d.h"
46#include "bif/bif_5_0_sh_mask.h"
47
48#include "tonga_sdma_pkt_open.h"
49
50#include "ivsrcid/ivsrcid_vislands30.h"
51
52static void sdma_v3_0_set_ring_funcs(struct amdgpu_device *adev);
53static void sdma_v3_0_set_buffer_funcs(struct amdgpu_device *adev);
54static void sdma_v3_0_set_vm_pte_funcs(struct amdgpu_device *adev);
55static void sdma_v3_0_set_irq_funcs(struct amdgpu_device *adev);
56
57MODULE_FIRMWARE("amdgpu/tonga_sdma.bin");
58MODULE_FIRMWARE("amdgpu/tonga_sdma1.bin");
59MODULE_FIRMWARE("amdgpu/carrizo_sdma.bin");
60MODULE_FIRMWARE("amdgpu/carrizo_sdma1.bin");
61MODULE_FIRMWARE("amdgpu/fiji_sdma.bin");
62MODULE_FIRMWARE("amdgpu/fiji_sdma1.bin");
63MODULE_FIRMWARE("amdgpu/stoney_sdma.bin");
64MODULE_FIRMWARE("amdgpu/polaris10_sdma.bin");
65MODULE_FIRMWARE("amdgpu/polaris10_sdma1.bin");
66MODULE_FIRMWARE("amdgpu/polaris11_sdma.bin");
67MODULE_FIRMWARE("amdgpu/polaris11_sdma1.bin");
68MODULE_FIRMWARE("amdgpu/polaris12_sdma.bin");
69MODULE_FIRMWARE("amdgpu/polaris12_sdma1.bin");
70MODULE_FIRMWARE("amdgpu/vegam_sdma.bin");
71MODULE_FIRMWARE("amdgpu/vegam_sdma1.bin");
72
73
74static const u32 sdma_offsets[SDMA_MAX_INSTANCE] =
75{
76 SDMA0_REGISTER_OFFSET,
77 SDMA1_REGISTER_OFFSET
78};
79
80static const u32 golden_settings_tonga_a11[] =
81{
82 mmSDMA0_CHICKEN_BITS, 0xfc910007, 0x00810007,
83 mmSDMA0_CLK_CTRL, 0xff000fff, 0x00000000,
84 mmSDMA0_GFX_IB_CNTL, 0x800f0111, 0x00000100,
85 mmSDMA0_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
86 mmSDMA0_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
87 mmSDMA1_CHICKEN_BITS, 0xfc910007, 0x00810007,
88 mmSDMA1_CLK_CTRL, 0xff000fff, 0x00000000,
89 mmSDMA1_GFX_IB_CNTL, 0x800f0111, 0x00000100,
90 mmSDMA1_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
91 mmSDMA1_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
92};
93
94static const u32 tonga_mgcg_cgcg_init[] =
95{
96 mmSDMA0_CLK_CTRL, 0xff000ff0, 0x00000100,
97 mmSDMA1_CLK_CTRL, 0xff000ff0, 0x00000100
98};
99
100static const u32 golden_settings_fiji_a10[] =
101{
102 mmSDMA0_CHICKEN_BITS, 0xfc910007, 0x00810007,
103 mmSDMA0_GFX_IB_CNTL, 0x800f0111, 0x00000100,
104 mmSDMA0_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
105 mmSDMA0_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
106 mmSDMA1_CHICKEN_BITS, 0xfc910007, 0x00810007,
107 mmSDMA1_GFX_IB_CNTL, 0x800f0111, 0x00000100,
108 mmSDMA1_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
109 mmSDMA1_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
110};
111
112static const u32 fiji_mgcg_cgcg_init[] =
113{
114 mmSDMA0_CLK_CTRL, 0xff000ff0, 0x00000100,
115 mmSDMA1_CLK_CTRL, 0xff000ff0, 0x00000100
116};
117
118static const u32 golden_settings_polaris11_a11[] =
119{
120 mmSDMA0_CHICKEN_BITS, 0xfc910007, 0x00810007,
121 mmSDMA0_CLK_CTRL, 0xff000fff, 0x00000000,
122 mmSDMA0_GFX_IB_CNTL, 0x800f0111, 0x00000100,
123 mmSDMA0_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
124 mmSDMA0_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
125 mmSDMA1_CHICKEN_BITS, 0xfc910007, 0x00810007,
126 mmSDMA1_CLK_CTRL, 0xff000fff, 0x00000000,
127 mmSDMA1_GFX_IB_CNTL, 0x800f0111, 0x00000100,
128 mmSDMA1_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
129 mmSDMA1_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
130};
131
132static const u32 golden_settings_polaris10_a11[] =
133{
134 mmSDMA0_CHICKEN_BITS, 0xfc910007, 0x00810007,
135 mmSDMA0_CLK_CTRL, 0xff000fff, 0x00000000,
136 mmSDMA0_GFX_IB_CNTL, 0x800f0111, 0x00000100,
137 mmSDMA0_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
138 mmSDMA0_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
139 mmSDMA1_CHICKEN_BITS, 0xfc910007, 0x00810007,
140 mmSDMA1_CLK_CTRL, 0xff000fff, 0x00000000,
141 mmSDMA1_GFX_IB_CNTL, 0x800f0111, 0x00000100,
142 mmSDMA1_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
143 mmSDMA1_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
144};
145
146static const u32 cz_golden_settings_a11[] =
147{
148 mmSDMA0_CHICKEN_BITS, 0xfc910007, 0x00810007,
149 mmSDMA0_CLK_CTRL, 0xff000fff, 0x00000000,
150 mmSDMA0_GFX_IB_CNTL, 0x00000100, 0x00000100,
151 mmSDMA0_POWER_CNTL, 0x00000800, 0x0003c800,
152 mmSDMA0_RLC0_IB_CNTL, 0x00000100, 0x00000100,
153 mmSDMA0_RLC1_IB_CNTL, 0x00000100, 0x00000100,
154 mmSDMA1_CHICKEN_BITS, 0xfc910007, 0x00810007,
155 mmSDMA1_CLK_CTRL, 0xff000fff, 0x00000000,
156 mmSDMA1_GFX_IB_CNTL, 0x00000100, 0x00000100,
157 mmSDMA1_POWER_CNTL, 0x00000800, 0x0003c800,
158 mmSDMA1_RLC0_IB_CNTL, 0x00000100, 0x00000100,
159 mmSDMA1_RLC1_IB_CNTL, 0x00000100, 0x00000100,
160};
161
162static const u32 cz_mgcg_cgcg_init[] =
163{
164 mmSDMA0_CLK_CTRL, 0xff000ff0, 0x00000100,
165 mmSDMA1_CLK_CTRL, 0xff000ff0, 0x00000100
166};
167
168static const u32 stoney_golden_settings_a11[] =
169{
170 mmSDMA0_GFX_IB_CNTL, 0x00000100, 0x00000100,
171 mmSDMA0_POWER_CNTL, 0x00000800, 0x0003c800,
172 mmSDMA0_RLC0_IB_CNTL, 0x00000100, 0x00000100,
173 mmSDMA0_RLC1_IB_CNTL, 0x00000100, 0x00000100,
174};
175
176static const u32 stoney_mgcg_cgcg_init[] =
177{
178 mmSDMA0_CLK_CTRL, 0xffffffff, 0x00000100,
179};
180
181/*
182 * sDMA - System DMA
183 * Starting with CIK, the GPU has new asynchronous
184 * DMA engines. These engines are used for compute
185 * and gfx. There are two DMA engines (SDMA0, SDMA1)
186 * and each one supports 1 ring buffer used for gfx
187 * and 2 queues used for compute.
188 *
189 * The programming model is very similar to the CP
190 * (ring buffer, IBs, etc.), but sDMA has it's own
191 * packet format that is different from the PM4 format
192 * used by the CP. sDMA supports copying data, writing
193 * embedded data, solid fills, and a number of other
194 * things. It also has support for tiling/detiling of
195 * buffers.
196 */
197
198static void sdma_v3_0_init_golden_registers(struct amdgpu_device *adev)
199{
200 switch (adev->asic_type) {
201 case CHIP_FIJI:
202 amdgpu_device_program_register_sequence(adev,
203 fiji_mgcg_cgcg_init,
204 ARRAY_SIZE(fiji_mgcg_cgcg_init));
205 amdgpu_device_program_register_sequence(adev,
206 golden_settings_fiji_a10,
207 ARRAY_SIZE(golden_settings_fiji_a10));
208 break;
209 case CHIP_TONGA:
210 amdgpu_device_program_register_sequence(adev,
211 tonga_mgcg_cgcg_init,
212 ARRAY_SIZE(tonga_mgcg_cgcg_init));
213 amdgpu_device_program_register_sequence(adev,
214 golden_settings_tonga_a11,
215 ARRAY_SIZE(golden_settings_tonga_a11));
216 break;
217 case CHIP_POLARIS11:
218 case CHIP_POLARIS12:
219 case CHIP_VEGAM:
220 amdgpu_device_program_register_sequence(adev,
221 golden_settings_polaris11_a11,
222 ARRAY_SIZE(golden_settings_polaris11_a11));
223 break;
224 case CHIP_POLARIS10:
225 amdgpu_device_program_register_sequence(adev,
226 golden_settings_polaris10_a11,
227 ARRAY_SIZE(golden_settings_polaris10_a11));
228 break;
229 case CHIP_CARRIZO:
230 amdgpu_device_program_register_sequence(adev,
231 cz_mgcg_cgcg_init,
232 ARRAY_SIZE(cz_mgcg_cgcg_init));
233 amdgpu_device_program_register_sequence(adev,
234 cz_golden_settings_a11,
235 ARRAY_SIZE(cz_golden_settings_a11));
236 break;
237 case CHIP_STONEY:
238 amdgpu_device_program_register_sequence(adev,
239 stoney_mgcg_cgcg_init,
240 ARRAY_SIZE(stoney_mgcg_cgcg_init));
241 amdgpu_device_program_register_sequence(adev,
242 stoney_golden_settings_a11,
243 ARRAY_SIZE(stoney_golden_settings_a11));
244 break;
245 default:
246 break;
247 }
248}
249
250static void sdma_v3_0_free_microcode(struct amdgpu_device *adev)
251{
252 int i;
253
254 for (i = 0; i < adev->sdma.num_instances; i++)
255 amdgpu_ucode_release(&adev->sdma.instance[i].fw);
256}
257
258/**
259 * sdma_v3_0_init_microcode - load ucode images from disk
260 *
261 * @adev: amdgpu_device pointer
262 *
263 * Use the firmware interface to load the ucode images into
264 * the driver (not loaded into hw).
265 * Returns 0 on success, error on failure.
266 */
267static int sdma_v3_0_init_microcode(struct amdgpu_device *adev)
268{
269 const char *chip_name;
270 int err = 0, i;
271 struct amdgpu_firmware_info *info = NULL;
272 const struct common_firmware_header *header = NULL;
273 const struct sdma_firmware_header_v1_0 *hdr;
274
275 DRM_DEBUG("\n");
276
277 switch (adev->asic_type) {
278 case CHIP_TONGA:
279 chip_name = "tonga";
280 break;
281 case CHIP_FIJI:
282 chip_name = "fiji";
283 break;
284 case CHIP_POLARIS10:
285 chip_name = "polaris10";
286 break;
287 case CHIP_POLARIS11:
288 chip_name = "polaris11";
289 break;
290 case CHIP_POLARIS12:
291 chip_name = "polaris12";
292 break;
293 case CHIP_VEGAM:
294 chip_name = "vegam";
295 break;
296 case CHIP_CARRIZO:
297 chip_name = "carrizo";
298 break;
299 case CHIP_STONEY:
300 chip_name = "stoney";
301 break;
302 default: BUG();
303 }
304
305 for (i = 0; i < adev->sdma.num_instances; i++) {
306 if (i == 0)
307 err = amdgpu_ucode_request(adev, &adev->sdma.instance[i].fw,
308 "amdgpu/%s_sdma.bin", chip_name);
309 else
310 err = amdgpu_ucode_request(adev, &adev->sdma.instance[i].fw,
311 "amdgpu/%s_sdma1.bin", chip_name);
312 if (err)
313 goto out;
314 hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
315 adev->sdma.instance[i].fw_version = le32_to_cpu(hdr->header.ucode_version);
316 adev->sdma.instance[i].feature_version = le32_to_cpu(hdr->ucode_feature_version);
317 if (adev->sdma.instance[i].feature_version >= 20)
318 adev->sdma.instance[i].burst_nop = true;
319
320 info = &adev->firmware.ucode[AMDGPU_UCODE_ID_SDMA0 + i];
321 info->ucode_id = AMDGPU_UCODE_ID_SDMA0 + i;
322 info->fw = adev->sdma.instance[i].fw;
323 header = (const struct common_firmware_header *)info->fw->data;
324 adev->firmware.fw_size +=
325 ALIGN(le32_to_cpu(header->ucode_size_bytes), PAGE_SIZE);
326
327 }
328out:
329 if (err) {
330 pr_err("sdma_v3_0: Failed to load firmware \"%s_sdma%s.bin\"\n",
331 chip_name, i == 0 ? "" : "1");
332 for (i = 0; i < adev->sdma.num_instances; i++)
333 amdgpu_ucode_release(&adev->sdma.instance[i].fw);
334 }
335 return err;
336}
337
338/**
339 * sdma_v3_0_ring_get_rptr - get the current read pointer
340 *
341 * @ring: amdgpu ring pointer
342 *
343 * Get the current rptr from the hardware (VI+).
344 */
345static uint64_t sdma_v3_0_ring_get_rptr(struct amdgpu_ring *ring)
346{
347 /* XXX check if swapping is necessary on BE */
348 return *ring->rptr_cpu_addr >> 2;
349}
350
351/**
352 * sdma_v3_0_ring_get_wptr - get the current write pointer
353 *
354 * @ring: amdgpu ring pointer
355 *
356 * Get the current wptr from the hardware (VI+).
357 */
358static uint64_t sdma_v3_0_ring_get_wptr(struct amdgpu_ring *ring)
359{
360 struct amdgpu_device *adev = ring->adev;
361 u32 wptr;
362
363 if (ring->use_doorbell || ring->use_pollmem) {
364 /* XXX check if swapping is necessary on BE */
365 wptr = *ring->wptr_cpu_addr >> 2;
366 } else {
367 wptr = RREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[ring->me]) >> 2;
368 }
369
370 return wptr;
371}
372
373/**
374 * sdma_v3_0_ring_set_wptr - commit the write pointer
375 *
376 * @ring: amdgpu ring pointer
377 *
378 * Write the wptr back to the hardware (VI+).
379 */
380static void sdma_v3_0_ring_set_wptr(struct amdgpu_ring *ring)
381{
382 struct amdgpu_device *adev = ring->adev;
383
384 if (ring->use_doorbell) {
385 u32 *wb = (u32 *)ring->wptr_cpu_addr;
386 /* XXX check if swapping is necessary on BE */
387 WRITE_ONCE(*wb, ring->wptr << 2);
388 WDOORBELL32(ring->doorbell_index, ring->wptr << 2);
389 } else if (ring->use_pollmem) {
390 u32 *wb = (u32 *)ring->wptr_cpu_addr;
391
392 WRITE_ONCE(*wb, ring->wptr << 2);
393 } else {
394 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[ring->me], ring->wptr << 2);
395 }
396}
397
398static void sdma_v3_0_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
399{
400 struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
401 int i;
402
403 for (i = 0; i < count; i++)
404 if (sdma && sdma->burst_nop && (i == 0))
405 amdgpu_ring_write(ring, ring->funcs->nop |
406 SDMA_PKT_NOP_HEADER_COUNT(count - 1));
407 else
408 amdgpu_ring_write(ring, ring->funcs->nop);
409}
410
411/**
412 * sdma_v3_0_ring_emit_ib - Schedule an IB on the DMA engine
413 *
414 * @ring: amdgpu ring pointer
415 * @job: job to retrieve vmid from
416 * @ib: IB object to schedule
417 * @flags: unused
418 *
419 * Schedule an IB in the DMA ring (VI).
420 */
421static void sdma_v3_0_ring_emit_ib(struct amdgpu_ring *ring,
422 struct amdgpu_job *job,
423 struct amdgpu_ib *ib,
424 uint32_t flags)
425{
426 unsigned vmid = AMDGPU_JOB_GET_VMID(job);
427
428 /* IB packet must end on a 8 DW boundary */
429 sdma_v3_0_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7);
430
431 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) |
432 SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf));
433 /* base must be 32 byte aligned */
434 amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
435 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
436 amdgpu_ring_write(ring, ib->length_dw);
437 amdgpu_ring_write(ring, 0);
438 amdgpu_ring_write(ring, 0);
439
440}
441
442/**
443 * sdma_v3_0_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
444 *
445 * @ring: amdgpu ring pointer
446 *
447 * Emit an hdp flush packet on the requested DMA ring.
448 */
449static void sdma_v3_0_ring_emit_hdp_flush(struct amdgpu_ring *ring)
450{
451 u32 ref_and_mask = 0;
452
453 if (ring->me == 0)
454 ref_and_mask = REG_SET_FIELD(ref_and_mask, GPU_HDP_FLUSH_DONE, SDMA0, 1);
455 else
456 ref_and_mask = REG_SET_FIELD(ref_and_mask, GPU_HDP_FLUSH_DONE, SDMA1, 1);
457
458 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
459 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
460 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
461 amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_DONE << 2);
462 amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_REQ << 2);
463 amdgpu_ring_write(ring, ref_and_mask); /* reference */
464 amdgpu_ring_write(ring, ref_and_mask); /* mask */
465 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
466 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
467}
468
469/**
470 * sdma_v3_0_ring_emit_fence - emit a fence on the DMA ring
471 *
472 * @ring: amdgpu ring pointer
473 * @addr: address
474 * @seq: sequence number
475 * @flags: fence related flags
476 *
477 * Add a DMA fence packet to the ring to write
478 * the fence seq number and DMA trap packet to generate
479 * an interrupt if needed (VI).
480 */
481static void sdma_v3_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
482 unsigned flags)
483{
484 bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
485 /* write the fence */
486 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE));
487 amdgpu_ring_write(ring, lower_32_bits(addr));
488 amdgpu_ring_write(ring, upper_32_bits(addr));
489 amdgpu_ring_write(ring, lower_32_bits(seq));
490
491 /* optionally write high bits as well */
492 if (write64bit) {
493 addr += 4;
494 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE));
495 amdgpu_ring_write(ring, lower_32_bits(addr));
496 amdgpu_ring_write(ring, upper_32_bits(addr));
497 amdgpu_ring_write(ring, upper_32_bits(seq));
498 }
499
500 /* generate an interrupt */
501 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP));
502 amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(0));
503}
504
505/**
506 * sdma_v3_0_gfx_stop - stop the gfx async dma engines
507 *
508 * @adev: amdgpu_device pointer
509 *
510 * Stop the gfx async dma ring buffers (VI).
511 */
512static void sdma_v3_0_gfx_stop(struct amdgpu_device *adev)
513{
514 u32 rb_cntl, ib_cntl;
515 int i;
516
517 for (i = 0; i < adev->sdma.num_instances; i++) {
518 rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
519 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0);
520 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
521 ib_cntl = RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]);
522 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0);
523 WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
524 }
525}
526
527/**
528 * sdma_v3_0_rlc_stop - stop the compute async dma engines
529 *
530 * @adev: amdgpu_device pointer
531 *
532 * Stop the compute async dma queues (VI).
533 */
534static void sdma_v3_0_rlc_stop(struct amdgpu_device *adev)
535{
536 /* XXX todo */
537}
538
539/**
540 * sdma_v3_0_ctx_switch_enable - stop the async dma engines context switch
541 *
542 * @adev: amdgpu_device pointer
543 * @enable: enable/disable the DMA MEs context switch.
544 *
545 * Halt or unhalt the async dma engines context switch (VI).
546 */
547static void sdma_v3_0_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
548{
549 u32 f32_cntl, phase_quantum = 0;
550 int i;
551
552 if (amdgpu_sdma_phase_quantum) {
553 unsigned value = amdgpu_sdma_phase_quantum;
554 unsigned unit = 0;
555
556 while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
557 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
558 value = (value + 1) >> 1;
559 unit++;
560 }
561 if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
562 SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
563 value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
564 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
565 unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
566 SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
567 WARN_ONCE(1,
568 "clamping sdma_phase_quantum to %uK clock cycles\n",
569 value << unit);
570 }
571 phase_quantum =
572 value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
573 unit << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
574 }
575
576 for (i = 0; i < adev->sdma.num_instances; i++) {
577 f32_cntl = RREG32(mmSDMA0_CNTL + sdma_offsets[i]);
578 if (enable) {
579 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
580 AUTO_CTXSW_ENABLE, 1);
581 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
582 ATC_L1_ENABLE, 1);
583 if (amdgpu_sdma_phase_quantum) {
584 WREG32(mmSDMA0_PHASE0_QUANTUM + sdma_offsets[i],
585 phase_quantum);
586 WREG32(mmSDMA0_PHASE1_QUANTUM + sdma_offsets[i],
587 phase_quantum);
588 }
589 } else {
590 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
591 AUTO_CTXSW_ENABLE, 0);
592 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
593 ATC_L1_ENABLE, 1);
594 }
595
596 WREG32(mmSDMA0_CNTL + sdma_offsets[i], f32_cntl);
597 }
598}
599
600/**
601 * sdma_v3_0_enable - stop the async dma engines
602 *
603 * @adev: amdgpu_device pointer
604 * @enable: enable/disable the DMA MEs.
605 *
606 * Halt or unhalt the async dma engines (VI).
607 */
608static void sdma_v3_0_enable(struct amdgpu_device *adev, bool enable)
609{
610 u32 f32_cntl;
611 int i;
612
613 if (!enable) {
614 sdma_v3_0_gfx_stop(adev);
615 sdma_v3_0_rlc_stop(adev);
616 }
617
618 for (i = 0; i < adev->sdma.num_instances; i++) {
619 f32_cntl = RREG32(mmSDMA0_F32_CNTL + sdma_offsets[i]);
620 if (enable)
621 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 0);
622 else
623 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 1);
624 WREG32(mmSDMA0_F32_CNTL + sdma_offsets[i], f32_cntl);
625 }
626}
627
628/**
629 * sdma_v3_0_gfx_resume - setup and start the async dma engines
630 *
631 * @adev: amdgpu_device pointer
632 *
633 * Set up the gfx DMA ring buffers and enable them (VI).
634 * Returns 0 for success, error for failure.
635 */
636static int sdma_v3_0_gfx_resume(struct amdgpu_device *adev)
637{
638 struct amdgpu_ring *ring;
639 u32 rb_cntl, ib_cntl, wptr_poll_cntl;
640 u32 rb_bufsz;
641 u32 doorbell;
642 u64 wptr_gpu_addr;
643 int i, j, r;
644
645 for (i = 0; i < adev->sdma.num_instances; i++) {
646 ring = &adev->sdma.instance[i].ring;
647 amdgpu_ring_clear_ring(ring);
648
649 mutex_lock(&adev->srbm_mutex);
650 for (j = 0; j < 16; j++) {
651 vi_srbm_select(adev, 0, 0, 0, j);
652 /* SDMA GFX */
653 WREG32(mmSDMA0_GFX_VIRTUAL_ADDR + sdma_offsets[i], 0);
654 WREG32(mmSDMA0_GFX_APE1_CNTL + sdma_offsets[i], 0);
655 }
656 vi_srbm_select(adev, 0, 0, 0, 0);
657 mutex_unlock(&adev->srbm_mutex);
658
659 WREG32(mmSDMA0_TILING_CONFIG + sdma_offsets[i],
660 adev->gfx.config.gb_addr_config & 0x70);
661
662 WREG32(mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
663
664 /* Set ring buffer size in dwords */
665 rb_bufsz = order_base_2(ring->ring_size / 4);
666 rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
667 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz);
668#ifdef __BIG_ENDIAN
669 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1);
670 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
671 RPTR_WRITEBACK_SWAP_ENABLE, 1);
672#endif
673 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
674
675 /* Initialize the ring buffer's read and write pointers */
676 ring->wptr = 0;
677 WREG32(mmSDMA0_GFX_RB_RPTR + sdma_offsets[i], 0);
678 sdma_v3_0_ring_set_wptr(ring);
679 WREG32(mmSDMA0_GFX_IB_RPTR + sdma_offsets[i], 0);
680 WREG32(mmSDMA0_GFX_IB_OFFSET + sdma_offsets[i], 0);
681
682 /* set the wb address whether it's enabled or not */
683 WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_HI + sdma_offsets[i],
684 upper_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFF);
685 WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_LO + sdma_offsets[i],
686 lower_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFC);
687
688 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
689
690 WREG32(mmSDMA0_GFX_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
691 WREG32(mmSDMA0_GFX_RB_BASE_HI + sdma_offsets[i], ring->gpu_addr >> 40);
692
693 doorbell = RREG32(mmSDMA0_GFX_DOORBELL + sdma_offsets[i]);
694
695 if (ring->use_doorbell) {
696 doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL,
697 OFFSET, ring->doorbell_index);
698 doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 1);
699 } else {
700 doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 0);
701 }
702 WREG32(mmSDMA0_GFX_DOORBELL + sdma_offsets[i], doorbell);
703
704 /* setup the wptr shadow polling */
705 wptr_gpu_addr = ring->wptr_gpu_addr;
706
707 WREG32(mmSDMA0_GFX_RB_WPTR_POLL_ADDR_LO + sdma_offsets[i],
708 lower_32_bits(wptr_gpu_addr));
709 WREG32(mmSDMA0_GFX_RB_WPTR_POLL_ADDR_HI + sdma_offsets[i],
710 upper_32_bits(wptr_gpu_addr));
711 wptr_poll_cntl = RREG32(mmSDMA0_GFX_RB_WPTR_POLL_CNTL + sdma_offsets[i]);
712 if (ring->use_pollmem) {
713 /*wptr polling is not enough fast, directly clean the wptr register */
714 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], 0);
715 wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl,
716 SDMA0_GFX_RB_WPTR_POLL_CNTL,
717 ENABLE, 1);
718 } else {
719 wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl,
720 SDMA0_GFX_RB_WPTR_POLL_CNTL,
721 ENABLE, 0);
722 }
723 WREG32(mmSDMA0_GFX_RB_WPTR_POLL_CNTL + sdma_offsets[i], wptr_poll_cntl);
724
725 /* enable DMA RB */
726 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1);
727 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
728
729 ib_cntl = RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]);
730 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1);
731#ifdef __BIG_ENDIAN
732 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1);
733#endif
734 /* enable DMA IBs */
735 WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
736 }
737
738 /* unhalt the MEs */
739 sdma_v3_0_enable(adev, true);
740 /* enable sdma ring preemption */
741 sdma_v3_0_ctx_switch_enable(adev, true);
742
743 for (i = 0; i < adev->sdma.num_instances; i++) {
744 ring = &adev->sdma.instance[i].ring;
745 r = amdgpu_ring_test_helper(ring);
746 if (r)
747 return r;
748 }
749
750 return 0;
751}
752
753/**
754 * sdma_v3_0_rlc_resume - setup and start the async dma engines
755 *
756 * @adev: amdgpu_device pointer
757 *
758 * Set up the compute DMA queues and enable them (VI).
759 * Returns 0 for success, error for failure.
760 */
761static int sdma_v3_0_rlc_resume(struct amdgpu_device *adev)
762{
763 /* XXX todo */
764 return 0;
765}
766
767/**
768 * sdma_v3_0_start - setup and start the async dma engines
769 *
770 * @adev: amdgpu_device pointer
771 *
772 * Set up the DMA engines and enable them (VI).
773 * Returns 0 for success, error for failure.
774 */
775static int sdma_v3_0_start(struct amdgpu_device *adev)
776{
777 int r;
778
779 /* disable sdma engine before programing it */
780 sdma_v3_0_ctx_switch_enable(adev, false);
781 sdma_v3_0_enable(adev, false);
782
783 /* start the gfx rings and rlc compute queues */
784 r = sdma_v3_0_gfx_resume(adev);
785 if (r)
786 return r;
787 r = sdma_v3_0_rlc_resume(adev);
788 if (r)
789 return r;
790
791 return 0;
792}
793
794/**
795 * sdma_v3_0_ring_test_ring - simple async dma engine test
796 *
797 * @ring: amdgpu_ring structure holding ring information
798 *
799 * Test the DMA engine by writing using it to write an
800 * value to memory. (VI).
801 * Returns 0 for success, error for failure.
802 */
803static int sdma_v3_0_ring_test_ring(struct amdgpu_ring *ring)
804{
805 struct amdgpu_device *adev = ring->adev;
806 unsigned i;
807 unsigned index;
808 int r;
809 u32 tmp;
810 u64 gpu_addr;
811
812 r = amdgpu_device_wb_get(adev, &index);
813 if (r)
814 return r;
815
816 gpu_addr = adev->wb.gpu_addr + (index * 4);
817 tmp = 0xCAFEDEAD;
818 adev->wb.wb[index] = cpu_to_le32(tmp);
819
820 r = amdgpu_ring_alloc(ring, 5);
821 if (r)
822 goto error_free_wb;
823
824 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
825 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
826 amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
827 amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
828 amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1));
829 amdgpu_ring_write(ring, 0xDEADBEEF);
830 amdgpu_ring_commit(ring);
831
832 for (i = 0; i < adev->usec_timeout; i++) {
833 tmp = le32_to_cpu(adev->wb.wb[index]);
834 if (tmp == 0xDEADBEEF)
835 break;
836 udelay(1);
837 }
838
839 if (i >= adev->usec_timeout)
840 r = -ETIMEDOUT;
841
842error_free_wb:
843 amdgpu_device_wb_free(adev, index);
844 return r;
845}
846
847/**
848 * sdma_v3_0_ring_test_ib - test an IB on the DMA engine
849 *
850 * @ring: amdgpu_ring structure holding ring information
851 * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
852 *
853 * Test a simple IB in the DMA ring (VI).
854 * Returns 0 on success, error on failure.
855 */
856static int sdma_v3_0_ring_test_ib(struct amdgpu_ring *ring, long timeout)
857{
858 struct amdgpu_device *adev = ring->adev;
859 struct amdgpu_ib ib;
860 struct dma_fence *f = NULL;
861 unsigned index;
862 u32 tmp = 0;
863 u64 gpu_addr;
864 long r;
865
866 r = amdgpu_device_wb_get(adev, &index);
867 if (r)
868 return r;
869
870 gpu_addr = adev->wb.gpu_addr + (index * 4);
871 tmp = 0xCAFEDEAD;
872 adev->wb.wb[index] = cpu_to_le32(tmp);
873 memset(&ib, 0, sizeof(ib));
874 r = amdgpu_ib_get(adev, NULL, 256,
875 AMDGPU_IB_POOL_DIRECT, &ib);
876 if (r)
877 goto err0;
878
879 ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
880 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
881 ib.ptr[1] = lower_32_bits(gpu_addr);
882 ib.ptr[2] = upper_32_bits(gpu_addr);
883 ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1);
884 ib.ptr[4] = 0xDEADBEEF;
885 ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
886 ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
887 ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
888 ib.length_dw = 8;
889
890 r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
891 if (r)
892 goto err1;
893
894 r = dma_fence_wait_timeout(f, false, timeout);
895 if (r == 0) {
896 r = -ETIMEDOUT;
897 goto err1;
898 } else if (r < 0) {
899 goto err1;
900 }
901 tmp = le32_to_cpu(adev->wb.wb[index]);
902 if (tmp == 0xDEADBEEF)
903 r = 0;
904 else
905 r = -EINVAL;
906err1:
907 amdgpu_ib_free(adev, &ib, NULL);
908 dma_fence_put(f);
909err0:
910 amdgpu_device_wb_free(adev, index);
911 return r;
912}
913
914/**
915 * sdma_v3_0_vm_copy_pte - update PTEs by copying them from the GART
916 *
917 * @ib: indirect buffer to fill with commands
918 * @pe: addr of the page entry
919 * @src: src addr to copy from
920 * @count: number of page entries to update
921 *
922 * Update PTEs by copying them from the GART using sDMA (CIK).
923 */
924static void sdma_v3_0_vm_copy_pte(struct amdgpu_ib *ib,
925 uint64_t pe, uint64_t src,
926 unsigned count)
927{
928 unsigned bytes = count * 8;
929
930 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
931 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
932 ib->ptr[ib->length_dw++] = bytes;
933 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
934 ib->ptr[ib->length_dw++] = lower_32_bits(src);
935 ib->ptr[ib->length_dw++] = upper_32_bits(src);
936 ib->ptr[ib->length_dw++] = lower_32_bits(pe);
937 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
938}
939
940/**
941 * sdma_v3_0_vm_write_pte - update PTEs by writing them manually
942 *
943 * @ib: indirect buffer to fill with commands
944 * @pe: addr of the page entry
945 * @value: dst addr to write into pe
946 * @count: number of page entries to update
947 * @incr: increase next addr by incr bytes
948 *
949 * Update PTEs by writing them manually using sDMA (CIK).
950 */
951static void sdma_v3_0_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
952 uint64_t value, unsigned count,
953 uint32_t incr)
954{
955 unsigned ndw = count * 2;
956
957 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
958 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
959 ib->ptr[ib->length_dw++] = lower_32_bits(pe);
960 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
961 ib->ptr[ib->length_dw++] = ndw;
962 for (; ndw > 0; ndw -= 2) {
963 ib->ptr[ib->length_dw++] = lower_32_bits(value);
964 ib->ptr[ib->length_dw++] = upper_32_bits(value);
965 value += incr;
966 }
967}
968
969/**
970 * sdma_v3_0_vm_set_pte_pde - update the page tables using sDMA
971 *
972 * @ib: indirect buffer to fill with commands
973 * @pe: addr of the page entry
974 * @addr: dst addr to write into pe
975 * @count: number of page entries to update
976 * @incr: increase next addr by incr bytes
977 * @flags: access flags
978 *
979 * Update the page tables using sDMA (CIK).
980 */
981static void sdma_v3_0_vm_set_pte_pde(struct amdgpu_ib *ib, uint64_t pe,
982 uint64_t addr, unsigned count,
983 uint32_t incr, uint64_t flags)
984{
985 /* for physically contiguous pages (vram) */
986 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_GEN_PTEPDE);
987 ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
988 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
989 ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
990 ib->ptr[ib->length_dw++] = upper_32_bits(flags);
991 ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
992 ib->ptr[ib->length_dw++] = upper_32_bits(addr);
993 ib->ptr[ib->length_dw++] = incr; /* increment size */
994 ib->ptr[ib->length_dw++] = 0;
995 ib->ptr[ib->length_dw++] = count; /* number of entries */
996}
997
998/**
999 * sdma_v3_0_ring_pad_ib - pad the IB to the required number of dw
1000 *
1001 * @ring: amdgpu_ring structure holding ring information
1002 * @ib: indirect buffer to fill with padding
1003 *
1004 */
1005static void sdma_v3_0_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
1006{
1007 struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
1008 u32 pad_count;
1009 int i;
1010
1011 pad_count = (-ib->length_dw) & 7;
1012 for (i = 0; i < pad_count; i++)
1013 if (sdma && sdma->burst_nop && (i == 0))
1014 ib->ptr[ib->length_dw++] =
1015 SDMA_PKT_HEADER_OP(SDMA_OP_NOP) |
1016 SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
1017 else
1018 ib->ptr[ib->length_dw++] =
1019 SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
1020}
1021
1022/**
1023 * sdma_v3_0_ring_emit_pipeline_sync - sync the pipeline
1024 *
1025 * @ring: amdgpu_ring pointer
1026 *
1027 * Make sure all previous operations are completed (CIK).
1028 */
1029static void sdma_v3_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
1030{
1031 uint32_t seq = ring->fence_drv.sync_seq;
1032 uint64_t addr = ring->fence_drv.gpu_addr;
1033
1034 /* wait for idle */
1035 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1036 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1037 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
1038 SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
1039 amdgpu_ring_write(ring, addr & 0xfffffffc);
1040 amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
1041 amdgpu_ring_write(ring, seq); /* reference */
1042 amdgpu_ring_write(ring, 0xffffffff); /* mask */
1043 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1044 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
1045}
1046
1047/**
1048 * sdma_v3_0_ring_emit_vm_flush - cik vm flush using sDMA
1049 *
1050 * @ring: amdgpu_ring pointer
1051 * @vmid: vmid number to use
1052 * @pd_addr: address
1053 *
1054 * Update the page table base and flush the VM TLB
1055 * using sDMA (VI).
1056 */
1057static void sdma_v3_0_ring_emit_vm_flush(struct amdgpu_ring *ring,
1058 unsigned vmid, uint64_t pd_addr)
1059{
1060 amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
1061
1062 /* wait for flush */
1063 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1064 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1065 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(0)); /* always */
1066 amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST << 2);
1067 amdgpu_ring_write(ring, 0);
1068 amdgpu_ring_write(ring, 0); /* reference */
1069 amdgpu_ring_write(ring, 0); /* mask */
1070 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1071 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
1072}
1073
1074static void sdma_v3_0_ring_emit_wreg(struct amdgpu_ring *ring,
1075 uint32_t reg, uint32_t val)
1076{
1077 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
1078 SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
1079 amdgpu_ring_write(ring, reg);
1080 amdgpu_ring_write(ring, val);
1081}
1082
1083static int sdma_v3_0_early_init(struct amdgpu_ip_block *ip_block)
1084{
1085 struct amdgpu_device *adev = ip_block->adev;
1086 int r;
1087
1088 switch (adev->asic_type) {
1089 case CHIP_STONEY:
1090 adev->sdma.num_instances = 1;
1091 break;
1092 default:
1093 adev->sdma.num_instances = SDMA_MAX_INSTANCE;
1094 break;
1095 }
1096
1097 r = sdma_v3_0_init_microcode(adev);
1098 if (r)
1099 return r;
1100
1101 sdma_v3_0_set_ring_funcs(adev);
1102 sdma_v3_0_set_buffer_funcs(adev);
1103 sdma_v3_0_set_vm_pte_funcs(adev);
1104 sdma_v3_0_set_irq_funcs(adev);
1105
1106 return 0;
1107}
1108
1109static int sdma_v3_0_sw_init(struct amdgpu_ip_block *ip_block)
1110{
1111 struct amdgpu_ring *ring;
1112 int r, i;
1113 struct amdgpu_device *adev = ip_block->adev;
1114
1115 /* SDMA trap event */
1116 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_SDMA_TRAP,
1117 &adev->sdma.trap_irq);
1118 if (r)
1119 return r;
1120
1121 /* SDMA Privileged inst */
1122 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 241,
1123 &adev->sdma.illegal_inst_irq);
1124 if (r)
1125 return r;
1126
1127 /* SDMA Privileged inst */
1128 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_SDMA_SRBM_WRITE,
1129 &adev->sdma.illegal_inst_irq);
1130 if (r)
1131 return r;
1132
1133 for (i = 0; i < adev->sdma.num_instances; i++) {
1134 ring = &adev->sdma.instance[i].ring;
1135 ring->ring_obj = NULL;
1136 if (!amdgpu_sriov_vf(adev)) {
1137 ring->use_doorbell = true;
1138 ring->doorbell_index = adev->doorbell_index.sdma_engine[i];
1139 } else {
1140 ring->use_pollmem = true;
1141 }
1142
1143 sprintf(ring->name, "sdma%d", i);
1144 r = amdgpu_ring_init(adev, ring, 1024, &adev->sdma.trap_irq,
1145 (i == 0) ? AMDGPU_SDMA_IRQ_INSTANCE0 :
1146 AMDGPU_SDMA_IRQ_INSTANCE1,
1147 AMDGPU_RING_PRIO_DEFAULT, NULL);
1148 if (r)
1149 return r;
1150 }
1151
1152 return r;
1153}
1154
1155static int sdma_v3_0_sw_fini(struct amdgpu_ip_block *ip_block)
1156{
1157 struct amdgpu_device *adev = ip_block->adev;
1158 int i;
1159
1160 for (i = 0; i < adev->sdma.num_instances; i++)
1161 amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1162
1163 sdma_v3_0_free_microcode(adev);
1164 return 0;
1165}
1166
1167static int sdma_v3_0_hw_init(struct amdgpu_ip_block *ip_block)
1168{
1169 int r;
1170 struct amdgpu_device *adev = ip_block->adev;
1171
1172 sdma_v3_0_init_golden_registers(adev);
1173
1174 r = sdma_v3_0_start(adev);
1175 if (r)
1176 return r;
1177
1178 return r;
1179}
1180
1181static int sdma_v3_0_hw_fini(struct amdgpu_ip_block *ip_block)
1182{
1183 struct amdgpu_device *adev = ip_block->adev;
1184
1185 sdma_v3_0_ctx_switch_enable(adev, false);
1186 sdma_v3_0_enable(adev, false);
1187
1188 return 0;
1189}
1190
1191static int sdma_v3_0_suspend(struct amdgpu_ip_block *ip_block)
1192{
1193 return sdma_v3_0_hw_fini(ip_block);
1194}
1195
1196static int sdma_v3_0_resume(struct amdgpu_ip_block *ip_block)
1197{
1198 return sdma_v3_0_hw_init(ip_block);
1199}
1200
1201static bool sdma_v3_0_is_idle(void *handle)
1202{
1203 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1204 u32 tmp = RREG32(mmSRBM_STATUS2);
1205
1206 if (tmp & (SRBM_STATUS2__SDMA_BUSY_MASK |
1207 SRBM_STATUS2__SDMA1_BUSY_MASK))
1208 return false;
1209
1210 return true;
1211}
1212
1213static int sdma_v3_0_wait_for_idle(struct amdgpu_ip_block *ip_block)
1214{
1215 unsigned i;
1216 u32 tmp;
1217 struct amdgpu_device *adev = ip_block->adev;
1218
1219 for (i = 0; i < adev->usec_timeout; i++) {
1220 tmp = RREG32(mmSRBM_STATUS2) & (SRBM_STATUS2__SDMA_BUSY_MASK |
1221 SRBM_STATUS2__SDMA1_BUSY_MASK);
1222
1223 if (!tmp)
1224 return 0;
1225 udelay(1);
1226 }
1227 return -ETIMEDOUT;
1228}
1229
1230static bool sdma_v3_0_check_soft_reset(struct amdgpu_ip_block *ip_block)
1231{
1232 struct amdgpu_device *adev = ip_block->adev;
1233 u32 srbm_soft_reset = 0;
1234 u32 tmp = RREG32(mmSRBM_STATUS2);
1235
1236 if ((tmp & SRBM_STATUS2__SDMA_BUSY_MASK) ||
1237 (tmp & SRBM_STATUS2__SDMA1_BUSY_MASK)) {
1238 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA_MASK;
1239 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA1_MASK;
1240 }
1241
1242 if (srbm_soft_reset) {
1243 adev->sdma.srbm_soft_reset = srbm_soft_reset;
1244 return true;
1245 } else {
1246 adev->sdma.srbm_soft_reset = 0;
1247 return false;
1248 }
1249}
1250
1251static int sdma_v3_0_pre_soft_reset(struct amdgpu_ip_block *ip_block)
1252{
1253 struct amdgpu_device *adev = ip_block->adev;
1254 u32 srbm_soft_reset = 0;
1255
1256 if (!adev->sdma.srbm_soft_reset)
1257 return 0;
1258
1259 srbm_soft_reset = adev->sdma.srbm_soft_reset;
1260
1261 if (REG_GET_FIELD(srbm_soft_reset, SRBM_SOFT_RESET, SOFT_RESET_SDMA) ||
1262 REG_GET_FIELD(srbm_soft_reset, SRBM_SOFT_RESET, SOFT_RESET_SDMA1)) {
1263 sdma_v3_0_ctx_switch_enable(adev, false);
1264 sdma_v3_0_enable(adev, false);
1265 }
1266
1267 return 0;
1268}
1269
1270static int sdma_v3_0_post_soft_reset(struct amdgpu_ip_block *ip_block)
1271{
1272 struct amdgpu_device *adev = ip_block->adev;
1273 u32 srbm_soft_reset = 0;
1274
1275 if (!adev->sdma.srbm_soft_reset)
1276 return 0;
1277
1278 srbm_soft_reset = adev->sdma.srbm_soft_reset;
1279
1280 if (REG_GET_FIELD(srbm_soft_reset, SRBM_SOFT_RESET, SOFT_RESET_SDMA) ||
1281 REG_GET_FIELD(srbm_soft_reset, SRBM_SOFT_RESET, SOFT_RESET_SDMA1)) {
1282 sdma_v3_0_gfx_resume(adev);
1283 sdma_v3_0_rlc_resume(adev);
1284 }
1285
1286 return 0;
1287}
1288
1289static int sdma_v3_0_soft_reset(struct amdgpu_ip_block *ip_block)
1290{
1291 struct amdgpu_device *adev = ip_block->adev;
1292 u32 srbm_soft_reset = 0;
1293 u32 tmp;
1294
1295 if (!adev->sdma.srbm_soft_reset)
1296 return 0;
1297
1298 srbm_soft_reset = adev->sdma.srbm_soft_reset;
1299
1300 if (srbm_soft_reset) {
1301 tmp = RREG32(mmSRBM_SOFT_RESET);
1302 tmp |= srbm_soft_reset;
1303 dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
1304 WREG32(mmSRBM_SOFT_RESET, tmp);
1305 tmp = RREG32(mmSRBM_SOFT_RESET);
1306
1307 udelay(50);
1308
1309 tmp &= ~srbm_soft_reset;
1310 WREG32(mmSRBM_SOFT_RESET, tmp);
1311 tmp = RREG32(mmSRBM_SOFT_RESET);
1312
1313 /* Wait a little for things to settle down */
1314 udelay(50);
1315 }
1316
1317 return 0;
1318}
1319
1320static int sdma_v3_0_set_trap_irq_state(struct amdgpu_device *adev,
1321 struct amdgpu_irq_src *source,
1322 unsigned type,
1323 enum amdgpu_interrupt_state state)
1324{
1325 u32 sdma_cntl;
1326
1327 switch (type) {
1328 case AMDGPU_SDMA_IRQ_INSTANCE0:
1329 switch (state) {
1330 case AMDGPU_IRQ_STATE_DISABLE:
1331 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1332 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 0);
1333 WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1334 break;
1335 case AMDGPU_IRQ_STATE_ENABLE:
1336 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1337 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 1);
1338 WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1339 break;
1340 default:
1341 break;
1342 }
1343 break;
1344 case AMDGPU_SDMA_IRQ_INSTANCE1:
1345 switch (state) {
1346 case AMDGPU_IRQ_STATE_DISABLE:
1347 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1348 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 0);
1349 WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1350 break;
1351 case AMDGPU_IRQ_STATE_ENABLE:
1352 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1353 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 1);
1354 WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1355 break;
1356 default:
1357 break;
1358 }
1359 break;
1360 default:
1361 break;
1362 }
1363 return 0;
1364}
1365
1366static int sdma_v3_0_process_trap_irq(struct amdgpu_device *adev,
1367 struct amdgpu_irq_src *source,
1368 struct amdgpu_iv_entry *entry)
1369{
1370 u8 instance_id, queue_id;
1371
1372 instance_id = (entry->ring_id & 0x3) >> 0;
1373 queue_id = (entry->ring_id & 0xc) >> 2;
1374 DRM_DEBUG("IH: SDMA trap\n");
1375 switch (instance_id) {
1376 case 0:
1377 switch (queue_id) {
1378 case 0:
1379 amdgpu_fence_process(&adev->sdma.instance[0].ring);
1380 break;
1381 case 1:
1382 /* XXX compute */
1383 break;
1384 case 2:
1385 /* XXX compute */
1386 break;
1387 }
1388 break;
1389 case 1:
1390 switch (queue_id) {
1391 case 0:
1392 amdgpu_fence_process(&adev->sdma.instance[1].ring);
1393 break;
1394 case 1:
1395 /* XXX compute */
1396 break;
1397 case 2:
1398 /* XXX compute */
1399 break;
1400 }
1401 break;
1402 }
1403 return 0;
1404}
1405
1406static int sdma_v3_0_process_illegal_inst_irq(struct amdgpu_device *adev,
1407 struct amdgpu_irq_src *source,
1408 struct amdgpu_iv_entry *entry)
1409{
1410 u8 instance_id, queue_id;
1411
1412 DRM_ERROR("Illegal instruction in SDMA command stream\n");
1413 instance_id = (entry->ring_id & 0x3) >> 0;
1414 queue_id = (entry->ring_id & 0xc) >> 2;
1415
1416 if (instance_id <= 1 && queue_id == 0)
1417 drm_sched_fault(&adev->sdma.instance[instance_id].ring.sched);
1418 return 0;
1419}
1420
1421static void sdma_v3_0_update_sdma_medium_grain_clock_gating(
1422 struct amdgpu_device *adev,
1423 bool enable)
1424{
1425 uint32_t temp, data;
1426 int i;
1427
1428 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
1429 for (i = 0; i < adev->sdma.num_instances; i++) {
1430 temp = data = RREG32(mmSDMA0_CLK_CTRL + sdma_offsets[i]);
1431 data &= ~(SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK |
1432 SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK |
1433 SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK |
1434 SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1435 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1436 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1437 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1438 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK);
1439 if (data != temp)
1440 WREG32(mmSDMA0_CLK_CTRL + sdma_offsets[i], data);
1441 }
1442 } else {
1443 for (i = 0; i < adev->sdma.num_instances; i++) {
1444 temp = data = RREG32(mmSDMA0_CLK_CTRL + sdma_offsets[i]);
1445 data |= SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK |
1446 SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK |
1447 SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK |
1448 SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1449 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1450 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1451 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1452 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK;
1453
1454 if (data != temp)
1455 WREG32(mmSDMA0_CLK_CTRL + sdma_offsets[i], data);
1456 }
1457 }
1458}
1459
1460static void sdma_v3_0_update_sdma_medium_grain_light_sleep(
1461 struct amdgpu_device *adev,
1462 bool enable)
1463{
1464 uint32_t temp, data;
1465 int i;
1466
1467 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
1468 for (i = 0; i < adev->sdma.num_instances; i++) {
1469 temp = data = RREG32(mmSDMA0_POWER_CNTL + sdma_offsets[i]);
1470 data |= SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1471
1472 if (temp != data)
1473 WREG32(mmSDMA0_POWER_CNTL + sdma_offsets[i], data);
1474 }
1475 } else {
1476 for (i = 0; i < adev->sdma.num_instances; i++) {
1477 temp = data = RREG32(mmSDMA0_POWER_CNTL + sdma_offsets[i]);
1478 data &= ~SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1479
1480 if (temp != data)
1481 WREG32(mmSDMA0_POWER_CNTL + sdma_offsets[i], data);
1482 }
1483 }
1484}
1485
1486static int sdma_v3_0_set_clockgating_state(void *handle,
1487 enum amd_clockgating_state state)
1488{
1489 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1490
1491 if (amdgpu_sriov_vf(adev))
1492 return 0;
1493
1494 switch (adev->asic_type) {
1495 case CHIP_FIJI:
1496 case CHIP_CARRIZO:
1497 case CHIP_STONEY:
1498 sdma_v3_0_update_sdma_medium_grain_clock_gating(adev,
1499 state == AMD_CG_STATE_GATE);
1500 sdma_v3_0_update_sdma_medium_grain_light_sleep(adev,
1501 state == AMD_CG_STATE_GATE);
1502 break;
1503 default:
1504 break;
1505 }
1506 return 0;
1507}
1508
1509static int sdma_v3_0_set_powergating_state(void *handle,
1510 enum amd_powergating_state state)
1511{
1512 return 0;
1513}
1514
1515static void sdma_v3_0_get_clockgating_state(void *handle, u64 *flags)
1516{
1517 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1518 int data;
1519
1520 if (amdgpu_sriov_vf(adev))
1521 *flags = 0;
1522
1523 /* AMD_CG_SUPPORT_SDMA_MGCG */
1524 data = RREG32(mmSDMA0_CLK_CTRL + sdma_offsets[0]);
1525 if (!(data & SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK))
1526 *flags |= AMD_CG_SUPPORT_SDMA_MGCG;
1527
1528 /* AMD_CG_SUPPORT_SDMA_LS */
1529 data = RREG32(mmSDMA0_POWER_CNTL + sdma_offsets[0]);
1530 if (data & SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK)
1531 *flags |= AMD_CG_SUPPORT_SDMA_LS;
1532}
1533
1534static const struct amd_ip_funcs sdma_v3_0_ip_funcs = {
1535 .name = "sdma_v3_0",
1536 .early_init = sdma_v3_0_early_init,
1537 .sw_init = sdma_v3_0_sw_init,
1538 .sw_fini = sdma_v3_0_sw_fini,
1539 .hw_init = sdma_v3_0_hw_init,
1540 .hw_fini = sdma_v3_0_hw_fini,
1541 .suspend = sdma_v3_0_suspend,
1542 .resume = sdma_v3_0_resume,
1543 .is_idle = sdma_v3_0_is_idle,
1544 .wait_for_idle = sdma_v3_0_wait_for_idle,
1545 .check_soft_reset = sdma_v3_0_check_soft_reset,
1546 .pre_soft_reset = sdma_v3_0_pre_soft_reset,
1547 .post_soft_reset = sdma_v3_0_post_soft_reset,
1548 .soft_reset = sdma_v3_0_soft_reset,
1549 .set_clockgating_state = sdma_v3_0_set_clockgating_state,
1550 .set_powergating_state = sdma_v3_0_set_powergating_state,
1551 .get_clockgating_state = sdma_v3_0_get_clockgating_state,
1552};
1553
1554static const struct amdgpu_ring_funcs sdma_v3_0_ring_funcs = {
1555 .type = AMDGPU_RING_TYPE_SDMA,
1556 .align_mask = 0xf,
1557 .nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
1558 .support_64bit_ptrs = false,
1559 .secure_submission_supported = true,
1560 .get_rptr = sdma_v3_0_ring_get_rptr,
1561 .get_wptr = sdma_v3_0_ring_get_wptr,
1562 .set_wptr = sdma_v3_0_ring_set_wptr,
1563 .emit_frame_size =
1564 6 + /* sdma_v3_0_ring_emit_hdp_flush */
1565 3 + /* hdp invalidate */
1566 6 + /* sdma_v3_0_ring_emit_pipeline_sync */
1567 VI_FLUSH_GPU_TLB_NUM_WREG * 3 + 6 + /* sdma_v3_0_ring_emit_vm_flush */
1568 10 + 10 + 10, /* sdma_v3_0_ring_emit_fence x3 for user fence, vm fence */
1569 .emit_ib_size = 7 + 6, /* sdma_v3_0_ring_emit_ib */
1570 .emit_ib = sdma_v3_0_ring_emit_ib,
1571 .emit_fence = sdma_v3_0_ring_emit_fence,
1572 .emit_pipeline_sync = sdma_v3_0_ring_emit_pipeline_sync,
1573 .emit_vm_flush = sdma_v3_0_ring_emit_vm_flush,
1574 .emit_hdp_flush = sdma_v3_0_ring_emit_hdp_flush,
1575 .test_ring = sdma_v3_0_ring_test_ring,
1576 .test_ib = sdma_v3_0_ring_test_ib,
1577 .insert_nop = sdma_v3_0_ring_insert_nop,
1578 .pad_ib = sdma_v3_0_ring_pad_ib,
1579 .emit_wreg = sdma_v3_0_ring_emit_wreg,
1580};
1581
1582static void sdma_v3_0_set_ring_funcs(struct amdgpu_device *adev)
1583{
1584 int i;
1585
1586 for (i = 0; i < adev->sdma.num_instances; i++) {
1587 adev->sdma.instance[i].ring.funcs = &sdma_v3_0_ring_funcs;
1588 adev->sdma.instance[i].ring.me = i;
1589 }
1590}
1591
1592static const struct amdgpu_irq_src_funcs sdma_v3_0_trap_irq_funcs = {
1593 .set = sdma_v3_0_set_trap_irq_state,
1594 .process = sdma_v3_0_process_trap_irq,
1595};
1596
1597static const struct amdgpu_irq_src_funcs sdma_v3_0_illegal_inst_irq_funcs = {
1598 .process = sdma_v3_0_process_illegal_inst_irq,
1599};
1600
1601static void sdma_v3_0_set_irq_funcs(struct amdgpu_device *adev)
1602{
1603 adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
1604 adev->sdma.trap_irq.funcs = &sdma_v3_0_trap_irq_funcs;
1605 adev->sdma.illegal_inst_irq.funcs = &sdma_v3_0_illegal_inst_irq_funcs;
1606}
1607
1608/**
1609 * sdma_v3_0_emit_copy_buffer - copy buffer using the sDMA engine
1610 *
1611 * @ib: indirect buffer to copy to
1612 * @src_offset: src GPU address
1613 * @dst_offset: dst GPU address
1614 * @byte_count: number of bytes to xfer
1615 * @copy_flags: unused
1616 *
1617 * Copy GPU buffers using the DMA engine (VI).
1618 * Used by the amdgpu ttm implementation to move pages if
1619 * registered as the asic copy callback.
1620 */
1621static void sdma_v3_0_emit_copy_buffer(struct amdgpu_ib *ib,
1622 uint64_t src_offset,
1623 uint64_t dst_offset,
1624 uint32_t byte_count,
1625 uint32_t copy_flags)
1626{
1627 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1628 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1629 ib->ptr[ib->length_dw++] = byte_count;
1630 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1631 ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1632 ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1633 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1634 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1635}
1636
1637/**
1638 * sdma_v3_0_emit_fill_buffer - fill buffer using the sDMA engine
1639 *
1640 * @ib: indirect buffer to copy to
1641 * @src_data: value to write to buffer
1642 * @dst_offset: dst GPU address
1643 * @byte_count: number of bytes to xfer
1644 *
1645 * Fill GPU buffers using the DMA engine (VI).
1646 */
1647static void sdma_v3_0_emit_fill_buffer(struct amdgpu_ib *ib,
1648 uint32_t src_data,
1649 uint64_t dst_offset,
1650 uint32_t byte_count)
1651{
1652 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL);
1653 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1654 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1655 ib->ptr[ib->length_dw++] = src_data;
1656 ib->ptr[ib->length_dw++] = byte_count;
1657}
1658
1659static const struct amdgpu_buffer_funcs sdma_v3_0_buffer_funcs = {
1660 .copy_max_bytes = 0x3fffe0, /* not 0x3fffff due to HW limitation */
1661 .copy_num_dw = 7,
1662 .emit_copy_buffer = sdma_v3_0_emit_copy_buffer,
1663
1664 .fill_max_bytes = 0x3fffe0, /* not 0x3fffff due to HW limitation */
1665 .fill_num_dw = 5,
1666 .emit_fill_buffer = sdma_v3_0_emit_fill_buffer,
1667};
1668
1669static void sdma_v3_0_set_buffer_funcs(struct amdgpu_device *adev)
1670{
1671 adev->mman.buffer_funcs = &sdma_v3_0_buffer_funcs;
1672 adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1673}
1674
1675static const struct amdgpu_vm_pte_funcs sdma_v3_0_vm_pte_funcs = {
1676 .copy_pte_num_dw = 7,
1677 .copy_pte = sdma_v3_0_vm_copy_pte,
1678
1679 .write_pte = sdma_v3_0_vm_write_pte,
1680 .set_pte_pde = sdma_v3_0_vm_set_pte_pde,
1681};
1682
1683static void sdma_v3_0_set_vm_pte_funcs(struct amdgpu_device *adev)
1684{
1685 unsigned i;
1686
1687 adev->vm_manager.vm_pte_funcs = &sdma_v3_0_vm_pte_funcs;
1688 for (i = 0; i < adev->sdma.num_instances; i++) {
1689 adev->vm_manager.vm_pte_scheds[i] =
1690 &adev->sdma.instance[i].ring.sched;
1691 }
1692 adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
1693}
1694
1695const struct amdgpu_ip_block_version sdma_v3_0_ip_block =
1696{
1697 .type = AMD_IP_BLOCK_TYPE_SDMA,
1698 .major = 3,
1699 .minor = 0,
1700 .rev = 0,
1701 .funcs = &sdma_v3_0_ip_funcs,
1702};
1703
1704const struct amdgpu_ip_block_version sdma_v3_1_ip_block =
1705{
1706 .type = AMD_IP_BLOCK_TYPE_SDMA,
1707 .major = 3,
1708 .minor = 1,
1709 .rev = 0,
1710 .funcs = &sdma_v3_0_ip_funcs,
1711};
1/*
2 * Copyright 2014 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/delay.h>
26#include <linux/firmware.h>
27#include <linux/module.h>
28
29#include "amdgpu.h"
30#include "amdgpu_ucode.h"
31#include "amdgpu_trace.h"
32#include "vi.h"
33#include "vid.h"
34
35#include "oss/oss_3_0_d.h"
36#include "oss/oss_3_0_sh_mask.h"
37
38#include "gmc/gmc_8_1_d.h"
39#include "gmc/gmc_8_1_sh_mask.h"
40
41#include "gca/gfx_8_0_d.h"
42#include "gca/gfx_8_0_enum.h"
43#include "gca/gfx_8_0_sh_mask.h"
44
45#include "bif/bif_5_0_d.h"
46#include "bif/bif_5_0_sh_mask.h"
47
48#include "tonga_sdma_pkt_open.h"
49
50#include "ivsrcid/ivsrcid_vislands30.h"
51
52static void sdma_v3_0_set_ring_funcs(struct amdgpu_device *adev);
53static void sdma_v3_0_set_buffer_funcs(struct amdgpu_device *adev);
54static void sdma_v3_0_set_vm_pte_funcs(struct amdgpu_device *adev);
55static void sdma_v3_0_set_irq_funcs(struct amdgpu_device *adev);
56
57MODULE_FIRMWARE("amdgpu/tonga_sdma.bin");
58MODULE_FIRMWARE("amdgpu/tonga_sdma1.bin");
59MODULE_FIRMWARE("amdgpu/carrizo_sdma.bin");
60MODULE_FIRMWARE("amdgpu/carrizo_sdma1.bin");
61MODULE_FIRMWARE("amdgpu/fiji_sdma.bin");
62MODULE_FIRMWARE("amdgpu/fiji_sdma1.bin");
63MODULE_FIRMWARE("amdgpu/stoney_sdma.bin");
64MODULE_FIRMWARE("amdgpu/polaris10_sdma.bin");
65MODULE_FIRMWARE("amdgpu/polaris10_sdma1.bin");
66MODULE_FIRMWARE("amdgpu/polaris11_sdma.bin");
67MODULE_FIRMWARE("amdgpu/polaris11_sdma1.bin");
68MODULE_FIRMWARE("amdgpu/polaris12_sdma.bin");
69MODULE_FIRMWARE("amdgpu/polaris12_sdma1.bin");
70MODULE_FIRMWARE("amdgpu/vegam_sdma.bin");
71MODULE_FIRMWARE("amdgpu/vegam_sdma1.bin");
72
73
74static const u32 sdma_offsets[SDMA_MAX_INSTANCE] =
75{
76 SDMA0_REGISTER_OFFSET,
77 SDMA1_REGISTER_OFFSET
78};
79
80static const u32 golden_settings_tonga_a11[] =
81{
82 mmSDMA0_CHICKEN_BITS, 0xfc910007, 0x00810007,
83 mmSDMA0_CLK_CTRL, 0xff000fff, 0x00000000,
84 mmSDMA0_GFX_IB_CNTL, 0x800f0111, 0x00000100,
85 mmSDMA0_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
86 mmSDMA0_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
87 mmSDMA1_CHICKEN_BITS, 0xfc910007, 0x00810007,
88 mmSDMA1_CLK_CTRL, 0xff000fff, 0x00000000,
89 mmSDMA1_GFX_IB_CNTL, 0x800f0111, 0x00000100,
90 mmSDMA1_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
91 mmSDMA1_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
92};
93
94static const u32 tonga_mgcg_cgcg_init[] =
95{
96 mmSDMA0_CLK_CTRL, 0xff000ff0, 0x00000100,
97 mmSDMA1_CLK_CTRL, 0xff000ff0, 0x00000100
98};
99
100static const u32 golden_settings_fiji_a10[] =
101{
102 mmSDMA0_CHICKEN_BITS, 0xfc910007, 0x00810007,
103 mmSDMA0_GFX_IB_CNTL, 0x800f0111, 0x00000100,
104 mmSDMA0_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
105 mmSDMA0_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
106 mmSDMA1_CHICKEN_BITS, 0xfc910007, 0x00810007,
107 mmSDMA1_GFX_IB_CNTL, 0x800f0111, 0x00000100,
108 mmSDMA1_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
109 mmSDMA1_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
110};
111
112static const u32 fiji_mgcg_cgcg_init[] =
113{
114 mmSDMA0_CLK_CTRL, 0xff000ff0, 0x00000100,
115 mmSDMA1_CLK_CTRL, 0xff000ff0, 0x00000100
116};
117
118static const u32 golden_settings_polaris11_a11[] =
119{
120 mmSDMA0_CHICKEN_BITS, 0xfc910007, 0x00810007,
121 mmSDMA0_CLK_CTRL, 0xff000fff, 0x00000000,
122 mmSDMA0_GFX_IB_CNTL, 0x800f0111, 0x00000100,
123 mmSDMA0_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
124 mmSDMA0_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
125 mmSDMA1_CHICKEN_BITS, 0xfc910007, 0x00810007,
126 mmSDMA1_CLK_CTRL, 0xff000fff, 0x00000000,
127 mmSDMA1_GFX_IB_CNTL, 0x800f0111, 0x00000100,
128 mmSDMA1_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
129 mmSDMA1_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
130};
131
132static const u32 golden_settings_polaris10_a11[] =
133{
134 mmSDMA0_CHICKEN_BITS, 0xfc910007, 0x00810007,
135 mmSDMA0_CLK_CTRL, 0xff000fff, 0x00000000,
136 mmSDMA0_GFX_IB_CNTL, 0x800f0111, 0x00000100,
137 mmSDMA0_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
138 mmSDMA0_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
139 mmSDMA1_CHICKEN_BITS, 0xfc910007, 0x00810007,
140 mmSDMA1_CLK_CTRL, 0xff000fff, 0x00000000,
141 mmSDMA1_GFX_IB_CNTL, 0x800f0111, 0x00000100,
142 mmSDMA1_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
143 mmSDMA1_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
144};
145
146static const u32 cz_golden_settings_a11[] =
147{
148 mmSDMA0_CHICKEN_BITS, 0xfc910007, 0x00810007,
149 mmSDMA0_CLK_CTRL, 0xff000fff, 0x00000000,
150 mmSDMA0_GFX_IB_CNTL, 0x00000100, 0x00000100,
151 mmSDMA0_POWER_CNTL, 0x00000800, 0x0003c800,
152 mmSDMA0_RLC0_IB_CNTL, 0x00000100, 0x00000100,
153 mmSDMA0_RLC1_IB_CNTL, 0x00000100, 0x00000100,
154 mmSDMA1_CHICKEN_BITS, 0xfc910007, 0x00810007,
155 mmSDMA1_CLK_CTRL, 0xff000fff, 0x00000000,
156 mmSDMA1_GFX_IB_CNTL, 0x00000100, 0x00000100,
157 mmSDMA1_POWER_CNTL, 0x00000800, 0x0003c800,
158 mmSDMA1_RLC0_IB_CNTL, 0x00000100, 0x00000100,
159 mmSDMA1_RLC1_IB_CNTL, 0x00000100, 0x00000100,
160};
161
162static const u32 cz_mgcg_cgcg_init[] =
163{
164 mmSDMA0_CLK_CTRL, 0xff000ff0, 0x00000100,
165 mmSDMA1_CLK_CTRL, 0xff000ff0, 0x00000100
166};
167
168static const u32 stoney_golden_settings_a11[] =
169{
170 mmSDMA0_GFX_IB_CNTL, 0x00000100, 0x00000100,
171 mmSDMA0_POWER_CNTL, 0x00000800, 0x0003c800,
172 mmSDMA0_RLC0_IB_CNTL, 0x00000100, 0x00000100,
173 mmSDMA0_RLC1_IB_CNTL, 0x00000100, 0x00000100,
174};
175
176static const u32 stoney_mgcg_cgcg_init[] =
177{
178 mmSDMA0_CLK_CTRL, 0xffffffff, 0x00000100,
179};
180
181/*
182 * sDMA - System DMA
183 * Starting with CIK, the GPU has new asynchronous
184 * DMA engines. These engines are used for compute
185 * and gfx. There are two DMA engines (SDMA0, SDMA1)
186 * and each one supports 1 ring buffer used for gfx
187 * and 2 queues used for compute.
188 *
189 * The programming model is very similar to the CP
190 * (ring buffer, IBs, etc.), but sDMA has it's own
191 * packet format that is different from the PM4 format
192 * used by the CP. sDMA supports copying data, writing
193 * embedded data, solid fills, and a number of other
194 * things. It also has support for tiling/detiling of
195 * buffers.
196 */
197
198static void sdma_v3_0_init_golden_registers(struct amdgpu_device *adev)
199{
200 switch (adev->asic_type) {
201 case CHIP_FIJI:
202 amdgpu_device_program_register_sequence(adev,
203 fiji_mgcg_cgcg_init,
204 ARRAY_SIZE(fiji_mgcg_cgcg_init));
205 amdgpu_device_program_register_sequence(adev,
206 golden_settings_fiji_a10,
207 ARRAY_SIZE(golden_settings_fiji_a10));
208 break;
209 case CHIP_TONGA:
210 amdgpu_device_program_register_sequence(adev,
211 tonga_mgcg_cgcg_init,
212 ARRAY_SIZE(tonga_mgcg_cgcg_init));
213 amdgpu_device_program_register_sequence(adev,
214 golden_settings_tonga_a11,
215 ARRAY_SIZE(golden_settings_tonga_a11));
216 break;
217 case CHIP_POLARIS11:
218 case CHIP_POLARIS12:
219 case CHIP_VEGAM:
220 amdgpu_device_program_register_sequence(adev,
221 golden_settings_polaris11_a11,
222 ARRAY_SIZE(golden_settings_polaris11_a11));
223 break;
224 case CHIP_POLARIS10:
225 amdgpu_device_program_register_sequence(adev,
226 golden_settings_polaris10_a11,
227 ARRAY_SIZE(golden_settings_polaris10_a11));
228 break;
229 case CHIP_CARRIZO:
230 amdgpu_device_program_register_sequence(adev,
231 cz_mgcg_cgcg_init,
232 ARRAY_SIZE(cz_mgcg_cgcg_init));
233 amdgpu_device_program_register_sequence(adev,
234 cz_golden_settings_a11,
235 ARRAY_SIZE(cz_golden_settings_a11));
236 break;
237 case CHIP_STONEY:
238 amdgpu_device_program_register_sequence(adev,
239 stoney_mgcg_cgcg_init,
240 ARRAY_SIZE(stoney_mgcg_cgcg_init));
241 amdgpu_device_program_register_sequence(adev,
242 stoney_golden_settings_a11,
243 ARRAY_SIZE(stoney_golden_settings_a11));
244 break;
245 default:
246 break;
247 }
248}
249
250static void sdma_v3_0_free_microcode(struct amdgpu_device *adev)
251{
252 int i;
253 for (i = 0; i < adev->sdma.num_instances; i++) {
254 release_firmware(adev->sdma.instance[i].fw);
255 adev->sdma.instance[i].fw = NULL;
256 }
257}
258
259/**
260 * sdma_v3_0_init_microcode - load ucode images from disk
261 *
262 * @adev: amdgpu_device pointer
263 *
264 * Use the firmware interface to load the ucode images into
265 * the driver (not loaded into hw).
266 * Returns 0 on success, error on failure.
267 */
268static int sdma_v3_0_init_microcode(struct amdgpu_device *adev)
269{
270 const char *chip_name;
271 char fw_name[30];
272 int err = 0, i;
273 struct amdgpu_firmware_info *info = NULL;
274 const struct common_firmware_header *header = NULL;
275 const struct sdma_firmware_header_v1_0 *hdr;
276
277 DRM_DEBUG("\n");
278
279 switch (adev->asic_type) {
280 case CHIP_TONGA:
281 chip_name = "tonga";
282 break;
283 case CHIP_FIJI:
284 chip_name = "fiji";
285 break;
286 case CHIP_POLARIS10:
287 chip_name = "polaris10";
288 break;
289 case CHIP_POLARIS11:
290 chip_name = "polaris11";
291 break;
292 case CHIP_POLARIS12:
293 chip_name = "polaris12";
294 break;
295 case CHIP_VEGAM:
296 chip_name = "vegam";
297 break;
298 case CHIP_CARRIZO:
299 chip_name = "carrizo";
300 break;
301 case CHIP_STONEY:
302 chip_name = "stoney";
303 break;
304 default: BUG();
305 }
306
307 for (i = 0; i < adev->sdma.num_instances; i++) {
308 if (i == 0)
309 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma.bin", chip_name);
310 else
311 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma1.bin", chip_name);
312 err = request_firmware(&adev->sdma.instance[i].fw, fw_name, adev->dev);
313 if (err)
314 goto out;
315 err = amdgpu_ucode_validate(adev->sdma.instance[i].fw);
316 if (err)
317 goto out;
318 hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
319 adev->sdma.instance[i].fw_version = le32_to_cpu(hdr->header.ucode_version);
320 adev->sdma.instance[i].feature_version = le32_to_cpu(hdr->ucode_feature_version);
321 if (adev->sdma.instance[i].feature_version >= 20)
322 adev->sdma.instance[i].burst_nop = true;
323
324 info = &adev->firmware.ucode[AMDGPU_UCODE_ID_SDMA0 + i];
325 info->ucode_id = AMDGPU_UCODE_ID_SDMA0 + i;
326 info->fw = adev->sdma.instance[i].fw;
327 header = (const struct common_firmware_header *)info->fw->data;
328 adev->firmware.fw_size +=
329 ALIGN(le32_to_cpu(header->ucode_size_bytes), PAGE_SIZE);
330
331 }
332out:
333 if (err) {
334 pr_err("sdma_v3_0: Failed to load firmware \"%s\"\n", fw_name);
335 for (i = 0; i < adev->sdma.num_instances; i++) {
336 release_firmware(adev->sdma.instance[i].fw);
337 adev->sdma.instance[i].fw = NULL;
338 }
339 }
340 return err;
341}
342
343/**
344 * sdma_v3_0_ring_get_rptr - get the current read pointer
345 *
346 * @ring: amdgpu ring pointer
347 *
348 * Get the current rptr from the hardware (VI+).
349 */
350static uint64_t sdma_v3_0_ring_get_rptr(struct amdgpu_ring *ring)
351{
352 /* XXX check if swapping is necessary on BE */
353 return ring->adev->wb.wb[ring->rptr_offs] >> 2;
354}
355
356/**
357 * sdma_v3_0_ring_get_wptr - get the current write pointer
358 *
359 * @ring: amdgpu ring pointer
360 *
361 * Get the current wptr from the hardware (VI+).
362 */
363static uint64_t sdma_v3_0_ring_get_wptr(struct amdgpu_ring *ring)
364{
365 struct amdgpu_device *adev = ring->adev;
366 u32 wptr;
367
368 if (ring->use_doorbell || ring->use_pollmem) {
369 /* XXX check if swapping is necessary on BE */
370 wptr = ring->adev->wb.wb[ring->wptr_offs] >> 2;
371 } else {
372 wptr = RREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[ring->me]) >> 2;
373 }
374
375 return wptr;
376}
377
378/**
379 * sdma_v3_0_ring_set_wptr - commit the write pointer
380 *
381 * @ring: amdgpu ring pointer
382 *
383 * Write the wptr back to the hardware (VI+).
384 */
385static void sdma_v3_0_ring_set_wptr(struct amdgpu_ring *ring)
386{
387 struct amdgpu_device *adev = ring->adev;
388
389 if (ring->use_doorbell) {
390 u32 *wb = (u32 *)&adev->wb.wb[ring->wptr_offs];
391 /* XXX check if swapping is necessary on BE */
392 WRITE_ONCE(*wb, (lower_32_bits(ring->wptr) << 2));
393 WDOORBELL32(ring->doorbell_index, lower_32_bits(ring->wptr) << 2);
394 } else if (ring->use_pollmem) {
395 u32 *wb = (u32 *)&adev->wb.wb[ring->wptr_offs];
396
397 WRITE_ONCE(*wb, (lower_32_bits(ring->wptr) << 2));
398 } else {
399 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[ring->me], lower_32_bits(ring->wptr) << 2);
400 }
401}
402
403static void sdma_v3_0_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
404{
405 struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
406 int i;
407
408 for (i = 0; i < count; i++)
409 if (sdma && sdma->burst_nop && (i == 0))
410 amdgpu_ring_write(ring, ring->funcs->nop |
411 SDMA_PKT_NOP_HEADER_COUNT(count - 1));
412 else
413 amdgpu_ring_write(ring, ring->funcs->nop);
414}
415
416/**
417 * sdma_v3_0_ring_emit_ib - Schedule an IB on the DMA engine
418 *
419 * @ring: amdgpu ring pointer
420 * @ib: IB object to schedule
421 *
422 * Schedule an IB in the DMA ring (VI).
423 */
424static void sdma_v3_0_ring_emit_ib(struct amdgpu_ring *ring,
425 struct amdgpu_job *job,
426 struct amdgpu_ib *ib,
427 uint32_t flags)
428{
429 unsigned vmid = AMDGPU_JOB_GET_VMID(job);
430
431 /* IB packet must end on a 8 DW boundary */
432 sdma_v3_0_ring_insert_nop(ring, (10 - (lower_32_bits(ring->wptr) & 7)) % 8);
433
434 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) |
435 SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf));
436 /* base must be 32 byte aligned */
437 amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
438 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
439 amdgpu_ring_write(ring, ib->length_dw);
440 amdgpu_ring_write(ring, 0);
441 amdgpu_ring_write(ring, 0);
442
443}
444
445/**
446 * sdma_v3_0_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
447 *
448 * @ring: amdgpu ring pointer
449 *
450 * Emit an hdp flush packet on the requested DMA ring.
451 */
452static void sdma_v3_0_ring_emit_hdp_flush(struct amdgpu_ring *ring)
453{
454 u32 ref_and_mask = 0;
455
456 if (ring->me == 0)
457 ref_and_mask = REG_SET_FIELD(ref_and_mask, GPU_HDP_FLUSH_DONE, SDMA0, 1);
458 else
459 ref_and_mask = REG_SET_FIELD(ref_and_mask, GPU_HDP_FLUSH_DONE, SDMA1, 1);
460
461 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
462 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
463 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
464 amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_DONE << 2);
465 amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_REQ << 2);
466 amdgpu_ring_write(ring, ref_and_mask); /* reference */
467 amdgpu_ring_write(ring, ref_and_mask); /* mask */
468 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
469 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
470}
471
472/**
473 * sdma_v3_0_ring_emit_fence - emit a fence on the DMA ring
474 *
475 * @ring: amdgpu ring pointer
476 * @fence: amdgpu fence object
477 *
478 * Add a DMA fence packet to the ring to write
479 * the fence seq number and DMA trap packet to generate
480 * an interrupt if needed (VI).
481 */
482static void sdma_v3_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
483 unsigned flags)
484{
485 bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
486 /* write the fence */
487 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE));
488 amdgpu_ring_write(ring, lower_32_bits(addr));
489 amdgpu_ring_write(ring, upper_32_bits(addr));
490 amdgpu_ring_write(ring, lower_32_bits(seq));
491
492 /* optionally write high bits as well */
493 if (write64bit) {
494 addr += 4;
495 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE));
496 amdgpu_ring_write(ring, lower_32_bits(addr));
497 amdgpu_ring_write(ring, upper_32_bits(addr));
498 amdgpu_ring_write(ring, upper_32_bits(seq));
499 }
500
501 /* generate an interrupt */
502 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP));
503 amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(0));
504}
505
506/**
507 * sdma_v3_0_gfx_stop - stop the gfx async dma engines
508 *
509 * @adev: amdgpu_device pointer
510 *
511 * Stop the gfx async dma ring buffers (VI).
512 */
513static void sdma_v3_0_gfx_stop(struct amdgpu_device *adev)
514{
515 struct amdgpu_ring *sdma0 = &adev->sdma.instance[0].ring;
516 struct amdgpu_ring *sdma1 = &adev->sdma.instance[1].ring;
517 u32 rb_cntl, ib_cntl;
518 int i;
519
520 if ((adev->mman.buffer_funcs_ring == sdma0) ||
521 (adev->mman.buffer_funcs_ring == sdma1))
522 amdgpu_ttm_set_buffer_funcs_status(adev, false);
523
524 for (i = 0; i < adev->sdma.num_instances; i++) {
525 rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
526 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0);
527 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
528 ib_cntl = RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]);
529 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0);
530 WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
531 }
532 sdma0->sched.ready = false;
533 sdma1->sched.ready = false;
534}
535
536/**
537 * sdma_v3_0_rlc_stop - stop the compute async dma engines
538 *
539 * @adev: amdgpu_device pointer
540 *
541 * Stop the compute async dma queues (VI).
542 */
543static void sdma_v3_0_rlc_stop(struct amdgpu_device *adev)
544{
545 /* XXX todo */
546}
547
548/**
549 * sdma_v3_0_ctx_switch_enable - stop the async dma engines context switch
550 *
551 * @adev: amdgpu_device pointer
552 * @enable: enable/disable the DMA MEs context switch.
553 *
554 * Halt or unhalt the async dma engines context switch (VI).
555 */
556static void sdma_v3_0_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
557{
558 u32 f32_cntl, phase_quantum = 0;
559 int i;
560
561 if (amdgpu_sdma_phase_quantum) {
562 unsigned value = amdgpu_sdma_phase_quantum;
563 unsigned unit = 0;
564
565 while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
566 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
567 value = (value + 1) >> 1;
568 unit++;
569 }
570 if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
571 SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
572 value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
573 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
574 unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
575 SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
576 WARN_ONCE(1,
577 "clamping sdma_phase_quantum to %uK clock cycles\n",
578 value << unit);
579 }
580 phase_quantum =
581 value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
582 unit << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
583 }
584
585 for (i = 0; i < adev->sdma.num_instances; i++) {
586 f32_cntl = RREG32(mmSDMA0_CNTL + sdma_offsets[i]);
587 if (enable) {
588 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
589 AUTO_CTXSW_ENABLE, 1);
590 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
591 ATC_L1_ENABLE, 1);
592 if (amdgpu_sdma_phase_quantum) {
593 WREG32(mmSDMA0_PHASE0_QUANTUM + sdma_offsets[i],
594 phase_quantum);
595 WREG32(mmSDMA0_PHASE1_QUANTUM + sdma_offsets[i],
596 phase_quantum);
597 }
598 } else {
599 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
600 AUTO_CTXSW_ENABLE, 0);
601 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
602 ATC_L1_ENABLE, 1);
603 }
604
605 WREG32(mmSDMA0_CNTL + sdma_offsets[i], f32_cntl);
606 }
607}
608
609/**
610 * sdma_v3_0_enable - stop the async dma engines
611 *
612 * @adev: amdgpu_device pointer
613 * @enable: enable/disable the DMA MEs.
614 *
615 * Halt or unhalt the async dma engines (VI).
616 */
617static void sdma_v3_0_enable(struct amdgpu_device *adev, bool enable)
618{
619 u32 f32_cntl;
620 int i;
621
622 if (!enable) {
623 sdma_v3_0_gfx_stop(adev);
624 sdma_v3_0_rlc_stop(adev);
625 }
626
627 for (i = 0; i < adev->sdma.num_instances; i++) {
628 f32_cntl = RREG32(mmSDMA0_F32_CNTL + sdma_offsets[i]);
629 if (enable)
630 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 0);
631 else
632 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 1);
633 WREG32(mmSDMA0_F32_CNTL + sdma_offsets[i], f32_cntl);
634 }
635}
636
637/**
638 * sdma_v3_0_gfx_resume - setup and start the async dma engines
639 *
640 * @adev: amdgpu_device pointer
641 *
642 * Set up the gfx DMA ring buffers and enable them (VI).
643 * Returns 0 for success, error for failure.
644 */
645static int sdma_v3_0_gfx_resume(struct amdgpu_device *adev)
646{
647 struct amdgpu_ring *ring;
648 u32 rb_cntl, ib_cntl, wptr_poll_cntl;
649 u32 rb_bufsz;
650 u32 wb_offset;
651 u32 doorbell;
652 u64 wptr_gpu_addr;
653 int i, j, r;
654
655 for (i = 0; i < adev->sdma.num_instances; i++) {
656 ring = &adev->sdma.instance[i].ring;
657 amdgpu_ring_clear_ring(ring);
658 wb_offset = (ring->rptr_offs * 4);
659
660 mutex_lock(&adev->srbm_mutex);
661 for (j = 0; j < 16; j++) {
662 vi_srbm_select(adev, 0, 0, 0, j);
663 /* SDMA GFX */
664 WREG32(mmSDMA0_GFX_VIRTUAL_ADDR + sdma_offsets[i], 0);
665 WREG32(mmSDMA0_GFX_APE1_CNTL + sdma_offsets[i], 0);
666 }
667 vi_srbm_select(adev, 0, 0, 0, 0);
668 mutex_unlock(&adev->srbm_mutex);
669
670 WREG32(mmSDMA0_TILING_CONFIG + sdma_offsets[i],
671 adev->gfx.config.gb_addr_config & 0x70);
672
673 WREG32(mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
674
675 /* Set ring buffer size in dwords */
676 rb_bufsz = order_base_2(ring->ring_size / 4);
677 rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
678 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz);
679#ifdef __BIG_ENDIAN
680 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1);
681 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
682 RPTR_WRITEBACK_SWAP_ENABLE, 1);
683#endif
684 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
685
686 /* Initialize the ring buffer's read and write pointers */
687 ring->wptr = 0;
688 WREG32(mmSDMA0_GFX_RB_RPTR + sdma_offsets[i], 0);
689 sdma_v3_0_ring_set_wptr(ring);
690 WREG32(mmSDMA0_GFX_IB_RPTR + sdma_offsets[i], 0);
691 WREG32(mmSDMA0_GFX_IB_OFFSET + sdma_offsets[i], 0);
692
693 /* set the wb address whether it's enabled or not */
694 WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_HI + sdma_offsets[i],
695 upper_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF);
696 WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_LO + sdma_offsets[i],
697 lower_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC);
698
699 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
700
701 WREG32(mmSDMA0_GFX_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
702 WREG32(mmSDMA0_GFX_RB_BASE_HI + sdma_offsets[i], ring->gpu_addr >> 40);
703
704 doorbell = RREG32(mmSDMA0_GFX_DOORBELL + sdma_offsets[i]);
705
706 if (ring->use_doorbell) {
707 doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL,
708 OFFSET, ring->doorbell_index);
709 doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 1);
710 } else {
711 doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 0);
712 }
713 WREG32(mmSDMA0_GFX_DOORBELL + sdma_offsets[i], doorbell);
714
715 /* setup the wptr shadow polling */
716 wptr_gpu_addr = adev->wb.gpu_addr + (ring->wptr_offs * 4);
717
718 WREG32(mmSDMA0_GFX_RB_WPTR_POLL_ADDR_LO + sdma_offsets[i],
719 lower_32_bits(wptr_gpu_addr));
720 WREG32(mmSDMA0_GFX_RB_WPTR_POLL_ADDR_HI + sdma_offsets[i],
721 upper_32_bits(wptr_gpu_addr));
722 wptr_poll_cntl = RREG32(mmSDMA0_GFX_RB_WPTR_POLL_CNTL + sdma_offsets[i]);
723 if (ring->use_pollmem) {
724 /*wptr polling is not enogh fast, directly clean the wptr register */
725 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], 0);
726 wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl,
727 SDMA0_GFX_RB_WPTR_POLL_CNTL,
728 ENABLE, 1);
729 } else {
730 wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl,
731 SDMA0_GFX_RB_WPTR_POLL_CNTL,
732 ENABLE, 0);
733 }
734 WREG32(mmSDMA0_GFX_RB_WPTR_POLL_CNTL + sdma_offsets[i], wptr_poll_cntl);
735
736 /* enable DMA RB */
737 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1);
738 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
739
740 ib_cntl = RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]);
741 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1);
742#ifdef __BIG_ENDIAN
743 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1);
744#endif
745 /* enable DMA IBs */
746 WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
747
748 ring->sched.ready = true;
749 }
750
751 /* unhalt the MEs */
752 sdma_v3_0_enable(adev, true);
753 /* enable sdma ring preemption */
754 sdma_v3_0_ctx_switch_enable(adev, true);
755
756 for (i = 0; i < adev->sdma.num_instances; i++) {
757 ring = &adev->sdma.instance[i].ring;
758 r = amdgpu_ring_test_helper(ring);
759 if (r)
760 return r;
761
762 if (adev->mman.buffer_funcs_ring == ring)
763 amdgpu_ttm_set_buffer_funcs_status(adev, true);
764 }
765
766 return 0;
767}
768
769/**
770 * sdma_v3_0_rlc_resume - setup and start the async dma engines
771 *
772 * @adev: amdgpu_device pointer
773 *
774 * Set up the compute DMA queues and enable them (VI).
775 * Returns 0 for success, error for failure.
776 */
777static int sdma_v3_0_rlc_resume(struct amdgpu_device *adev)
778{
779 /* XXX todo */
780 return 0;
781}
782
783/**
784 * sdma_v3_0_start - setup and start the async dma engines
785 *
786 * @adev: amdgpu_device pointer
787 *
788 * Set up the DMA engines and enable them (VI).
789 * Returns 0 for success, error for failure.
790 */
791static int sdma_v3_0_start(struct amdgpu_device *adev)
792{
793 int r;
794
795 /* disable sdma engine before programing it */
796 sdma_v3_0_ctx_switch_enable(adev, false);
797 sdma_v3_0_enable(adev, false);
798
799 /* start the gfx rings and rlc compute queues */
800 r = sdma_v3_0_gfx_resume(adev);
801 if (r)
802 return r;
803 r = sdma_v3_0_rlc_resume(adev);
804 if (r)
805 return r;
806
807 return 0;
808}
809
810/**
811 * sdma_v3_0_ring_test_ring - simple async dma engine test
812 *
813 * @ring: amdgpu_ring structure holding ring information
814 *
815 * Test the DMA engine by writing using it to write an
816 * value to memory. (VI).
817 * Returns 0 for success, error for failure.
818 */
819static int sdma_v3_0_ring_test_ring(struct amdgpu_ring *ring)
820{
821 struct amdgpu_device *adev = ring->adev;
822 unsigned i;
823 unsigned index;
824 int r;
825 u32 tmp;
826 u64 gpu_addr;
827
828 r = amdgpu_device_wb_get(adev, &index);
829 if (r)
830 return r;
831
832 gpu_addr = adev->wb.gpu_addr + (index * 4);
833 tmp = 0xCAFEDEAD;
834 adev->wb.wb[index] = cpu_to_le32(tmp);
835
836 r = amdgpu_ring_alloc(ring, 5);
837 if (r)
838 goto error_free_wb;
839
840 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
841 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
842 amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
843 amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
844 amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1));
845 amdgpu_ring_write(ring, 0xDEADBEEF);
846 amdgpu_ring_commit(ring);
847
848 for (i = 0; i < adev->usec_timeout; i++) {
849 tmp = le32_to_cpu(adev->wb.wb[index]);
850 if (tmp == 0xDEADBEEF)
851 break;
852 udelay(1);
853 }
854
855 if (i >= adev->usec_timeout)
856 r = -ETIMEDOUT;
857
858error_free_wb:
859 amdgpu_device_wb_free(adev, index);
860 return r;
861}
862
863/**
864 * sdma_v3_0_ring_test_ib - test an IB on the DMA engine
865 *
866 * @ring: amdgpu_ring structure holding ring information
867 *
868 * Test a simple IB in the DMA ring (VI).
869 * Returns 0 on success, error on failure.
870 */
871static int sdma_v3_0_ring_test_ib(struct amdgpu_ring *ring, long timeout)
872{
873 struct amdgpu_device *adev = ring->adev;
874 struct amdgpu_ib ib;
875 struct dma_fence *f = NULL;
876 unsigned index;
877 u32 tmp = 0;
878 u64 gpu_addr;
879 long r;
880
881 r = amdgpu_device_wb_get(adev, &index);
882 if (r)
883 return r;
884
885 gpu_addr = adev->wb.gpu_addr + (index * 4);
886 tmp = 0xCAFEDEAD;
887 adev->wb.wb[index] = cpu_to_le32(tmp);
888 memset(&ib, 0, sizeof(ib));
889 r = amdgpu_ib_get(adev, NULL, 256, &ib);
890 if (r)
891 goto err0;
892
893 ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
894 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
895 ib.ptr[1] = lower_32_bits(gpu_addr);
896 ib.ptr[2] = upper_32_bits(gpu_addr);
897 ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1);
898 ib.ptr[4] = 0xDEADBEEF;
899 ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
900 ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
901 ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
902 ib.length_dw = 8;
903
904 r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
905 if (r)
906 goto err1;
907
908 r = dma_fence_wait_timeout(f, false, timeout);
909 if (r == 0) {
910 r = -ETIMEDOUT;
911 goto err1;
912 } else if (r < 0) {
913 goto err1;
914 }
915 tmp = le32_to_cpu(adev->wb.wb[index]);
916 if (tmp == 0xDEADBEEF)
917 r = 0;
918 else
919 r = -EINVAL;
920err1:
921 amdgpu_ib_free(adev, &ib, NULL);
922 dma_fence_put(f);
923err0:
924 amdgpu_device_wb_free(adev, index);
925 return r;
926}
927
928/**
929 * sdma_v3_0_vm_copy_pte - update PTEs by copying them from the GART
930 *
931 * @ib: indirect buffer to fill with commands
932 * @pe: addr of the page entry
933 * @src: src addr to copy from
934 * @count: number of page entries to update
935 *
936 * Update PTEs by copying them from the GART using sDMA (CIK).
937 */
938static void sdma_v3_0_vm_copy_pte(struct amdgpu_ib *ib,
939 uint64_t pe, uint64_t src,
940 unsigned count)
941{
942 unsigned bytes = count * 8;
943
944 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
945 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
946 ib->ptr[ib->length_dw++] = bytes;
947 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
948 ib->ptr[ib->length_dw++] = lower_32_bits(src);
949 ib->ptr[ib->length_dw++] = upper_32_bits(src);
950 ib->ptr[ib->length_dw++] = lower_32_bits(pe);
951 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
952}
953
954/**
955 * sdma_v3_0_vm_write_pte - update PTEs by writing them manually
956 *
957 * @ib: indirect buffer to fill with commands
958 * @pe: addr of the page entry
959 * @value: dst addr to write into pe
960 * @count: number of page entries to update
961 * @incr: increase next addr by incr bytes
962 *
963 * Update PTEs by writing them manually using sDMA (CIK).
964 */
965static void sdma_v3_0_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
966 uint64_t value, unsigned count,
967 uint32_t incr)
968{
969 unsigned ndw = count * 2;
970
971 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
972 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
973 ib->ptr[ib->length_dw++] = lower_32_bits(pe);
974 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
975 ib->ptr[ib->length_dw++] = ndw;
976 for (; ndw > 0; ndw -= 2) {
977 ib->ptr[ib->length_dw++] = lower_32_bits(value);
978 ib->ptr[ib->length_dw++] = upper_32_bits(value);
979 value += incr;
980 }
981}
982
983/**
984 * sdma_v3_0_vm_set_pte_pde - update the page tables using sDMA
985 *
986 * @ib: indirect buffer to fill with commands
987 * @pe: addr of the page entry
988 * @addr: dst addr to write into pe
989 * @count: number of page entries to update
990 * @incr: increase next addr by incr bytes
991 * @flags: access flags
992 *
993 * Update the page tables using sDMA (CIK).
994 */
995static void sdma_v3_0_vm_set_pte_pde(struct amdgpu_ib *ib, uint64_t pe,
996 uint64_t addr, unsigned count,
997 uint32_t incr, uint64_t flags)
998{
999 /* for physically contiguous pages (vram) */
1000 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_GEN_PTEPDE);
1001 ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
1002 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1003 ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
1004 ib->ptr[ib->length_dw++] = upper_32_bits(flags);
1005 ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
1006 ib->ptr[ib->length_dw++] = upper_32_bits(addr);
1007 ib->ptr[ib->length_dw++] = incr; /* increment size */
1008 ib->ptr[ib->length_dw++] = 0;
1009 ib->ptr[ib->length_dw++] = count; /* number of entries */
1010}
1011
1012/**
1013 * sdma_v3_0_ring_pad_ib - pad the IB to the required number of dw
1014 *
1015 * @ib: indirect buffer to fill with padding
1016 *
1017 */
1018static void sdma_v3_0_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
1019{
1020 struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
1021 u32 pad_count;
1022 int i;
1023
1024 pad_count = (8 - (ib->length_dw & 0x7)) % 8;
1025 for (i = 0; i < pad_count; i++)
1026 if (sdma && sdma->burst_nop && (i == 0))
1027 ib->ptr[ib->length_dw++] =
1028 SDMA_PKT_HEADER_OP(SDMA_OP_NOP) |
1029 SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
1030 else
1031 ib->ptr[ib->length_dw++] =
1032 SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
1033}
1034
1035/**
1036 * sdma_v3_0_ring_emit_pipeline_sync - sync the pipeline
1037 *
1038 * @ring: amdgpu_ring pointer
1039 *
1040 * Make sure all previous operations are completed (CIK).
1041 */
1042static void sdma_v3_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
1043{
1044 uint32_t seq = ring->fence_drv.sync_seq;
1045 uint64_t addr = ring->fence_drv.gpu_addr;
1046
1047 /* wait for idle */
1048 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1049 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1050 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
1051 SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
1052 amdgpu_ring_write(ring, addr & 0xfffffffc);
1053 amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
1054 amdgpu_ring_write(ring, seq); /* reference */
1055 amdgpu_ring_write(ring, 0xffffffff); /* mask */
1056 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1057 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
1058}
1059
1060/**
1061 * sdma_v3_0_ring_emit_vm_flush - cik vm flush using sDMA
1062 *
1063 * @ring: amdgpu_ring pointer
1064 * @vm: amdgpu_vm pointer
1065 *
1066 * Update the page table base and flush the VM TLB
1067 * using sDMA (VI).
1068 */
1069static void sdma_v3_0_ring_emit_vm_flush(struct amdgpu_ring *ring,
1070 unsigned vmid, uint64_t pd_addr)
1071{
1072 amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
1073
1074 /* wait for flush */
1075 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1076 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1077 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(0)); /* always */
1078 amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST << 2);
1079 amdgpu_ring_write(ring, 0);
1080 amdgpu_ring_write(ring, 0); /* reference */
1081 amdgpu_ring_write(ring, 0); /* mask */
1082 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1083 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
1084}
1085
1086static void sdma_v3_0_ring_emit_wreg(struct amdgpu_ring *ring,
1087 uint32_t reg, uint32_t val)
1088{
1089 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
1090 SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
1091 amdgpu_ring_write(ring, reg);
1092 amdgpu_ring_write(ring, val);
1093}
1094
1095static int sdma_v3_0_early_init(void *handle)
1096{
1097 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1098
1099 switch (adev->asic_type) {
1100 case CHIP_STONEY:
1101 adev->sdma.num_instances = 1;
1102 break;
1103 default:
1104 adev->sdma.num_instances = SDMA_MAX_INSTANCE;
1105 break;
1106 }
1107
1108 sdma_v3_0_set_ring_funcs(adev);
1109 sdma_v3_0_set_buffer_funcs(adev);
1110 sdma_v3_0_set_vm_pte_funcs(adev);
1111 sdma_v3_0_set_irq_funcs(adev);
1112
1113 return 0;
1114}
1115
1116static int sdma_v3_0_sw_init(void *handle)
1117{
1118 struct amdgpu_ring *ring;
1119 int r, i;
1120 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1121
1122 /* SDMA trap event */
1123 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_SDMA_TRAP,
1124 &adev->sdma.trap_irq);
1125 if (r)
1126 return r;
1127
1128 /* SDMA Privileged inst */
1129 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 241,
1130 &adev->sdma.illegal_inst_irq);
1131 if (r)
1132 return r;
1133
1134 /* SDMA Privileged inst */
1135 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_SDMA_SRBM_WRITE,
1136 &adev->sdma.illegal_inst_irq);
1137 if (r)
1138 return r;
1139
1140 r = sdma_v3_0_init_microcode(adev);
1141 if (r) {
1142 DRM_ERROR("Failed to load sdma firmware!\n");
1143 return r;
1144 }
1145
1146 for (i = 0; i < adev->sdma.num_instances; i++) {
1147 ring = &adev->sdma.instance[i].ring;
1148 ring->ring_obj = NULL;
1149 if (!amdgpu_sriov_vf(adev)) {
1150 ring->use_doorbell = true;
1151 ring->doorbell_index = adev->doorbell_index.sdma_engine[i];
1152 } else {
1153 ring->use_pollmem = true;
1154 }
1155
1156 sprintf(ring->name, "sdma%d", i);
1157 r = amdgpu_ring_init(adev, ring, 1024,
1158 &adev->sdma.trap_irq,
1159 (i == 0) ?
1160 AMDGPU_SDMA_IRQ_INSTANCE0 :
1161 AMDGPU_SDMA_IRQ_INSTANCE1);
1162 if (r)
1163 return r;
1164 }
1165
1166 return r;
1167}
1168
1169static int sdma_v3_0_sw_fini(void *handle)
1170{
1171 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1172 int i;
1173
1174 for (i = 0; i < adev->sdma.num_instances; i++)
1175 amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1176
1177 sdma_v3_0_free_microcode(adev);
1178 return 0;
1179}
1180
1181static int sdma_v3_0_hw_init(void *handle)
1182{
1183 int r;
1184 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1185
1186 sdma_v3_0_init_golden_registers(adev);
1187
1188 r = sdma_v3_0_start(adev);
1189 if (r)
1190 return r;
1191
1192 return r;
1193}
1194
1195static int sdma_v3_0_hw_fini(void *handle)
1196{
1197 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1198
1199 sdma_v3_0_ctx_switch_enable(adev, false);
1200 sdma_v3_0_enable(adev, false);
1201
1202 return 0;
1203}
1204
1205static int sdma_v3_0_suspend(void *handle)
1206{
1207 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1208
1209 return sdma_v3_0_hw_fini(adev);
1210}
1211
1212static int sdma_v3_0_resume(void *handle)
1213{
1214 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1215
1216 return sdma_v3_0_hw_init(adev);
1217}
1218
1219static bool sdma_v3_0_is_idle(void *handle)
1220{
1221 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1222 u32 tmp = RREG32(mmSRBM_STATUS2);
1223
1224 if (tmp & (SRBM_STATUS2__SDMA_BUSY_MASK |
1225 SRBM_STATUS2__SDMA1_BUSY_MASK))
1226 return false;
1227
1228 return true;
1229}
1230
1231static int sdma_v3_0_wait_for_idle(void *handle)
1232{
1233 unsigned i;
1234 u32 tmp;
1235 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1236
1237 for (i = 0; i < adev->usec_timeout; i++) {
1238 tmp = RREG32(mmSRBM_STATUS2) & (SRBM_STATUS2__SDMA_BUSY_MASK |
1239 SRBM_STATUS2__SDMA1_BUSY_MASK);
1240
1241 if (!tmp)
1242 return 0;
1243 udelay(1);
1244 }
1245 return -ETIMEDOUT;
1246}
1247
1248static bool sdma_v3_0_check_soft_reset(void *handle)
1249{
1250 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1251 u32 srbm_soft_reset = 0;
1252 u32 tmp = RREG32(mmSRBM_STATUS2);
1253
1254 if ((tmp & SRBM_STATUS2__SDMA_BUSY_MASK) ||
1255 (tmp & SRBM_STATUS2__SDMA1_BUSY_MASK)) {
1256 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA_MASK;
1257 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA1_MASK;
1258 }
1259
1260 if (srbm_soft_reset) {
1261 adev->sdma.srbm_soft_reset = srbm_soft_reset;
1262 return true;
1263 } else {
1264 adev->sdma.srbm_soft_reset = 0;
1265 return false;
1266 }
1267}
1268
1269static int sdma_v3_0_pre_soft_reset(void *handle)
1270{
1271 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1272 u32 srbm_soft_reset = 0;
1273
1274 if (!adev->sdma.srbm_soft_reset)
1275 return 0;
1276
1277 srbm_soft_reset = adev->sdma.srbm_soft_reset;
1278
1279 if (REG_GET_FIELD(srbm_soft_reset, SRBM_SOFT_RESET, SOFT_RESET_SDMA) ||
1280 REG_GET_FIELD(srbm_soft_reset, SRBM_SOFT_RESET, SOFT_RESET_SDMA1)) {
1281 sdma_v3_0_ctx_switch_enable(adev, false);
1282 sdma_v3_0_enable(adev, false);
1283 }
1284
1285 return 0;
1286}
1287
1288static int sdma_v3_0_post_soft_reset(void *handle)
1289{
1290 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1291 u32 srbm_soft_reset = 0;
1292
1293 if (!adev->sdma.srbm_soft_reset)
1294 return 0;
1295
1296 srbm_soft_reset = adev->sdma.srbm_soft_reset;
1297
1298 if (REG_GET_FIELD(srbm_soft_reset, SRBM_SOFT_RESET, SOFT_RESET_SDMA) ||
1299 REG_GET_FIELD(srbm_soft_reset, SRBM_SOFT_RESET, SOFT_RESET_SDMA1)) {
1300 sdma_v3_0_gfx_resume(adev);
1301 sdma_v3_0_rlc_resume(adev);
1302 }
1303
1304 return 0;
1305}
1306
1307static int sdma_v3_0_soft_reset(void *handle)
1308{
1309 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1310 u32 srbm_soft_reset = 0;
1311 u32 tmp;
1312
1313 if (!adev->sdma.srbm_soft_reset)
1314 return 0;
1315
1316 srbm_soft_reset = adev->sdma.srbm_soft_reset;
1317
1318 if (srbm_soft_reset) {
1319 tmp = RREG32(mmSRBM_SOFT_RESET);
1320 tmp |= srbm_soft_reset;
1321 dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
1322 WREG32(mmSRBM_SOFT_RESET, tmp);
1323 tmp = RREG32(mmSRBM_SOFT_RESET);
1324
1325 udelay(50);
1326
1327 tmp &= ~srbm_soft_reset;
1328 WREG32(mmSRBM_SOFT_RESET, tmp);
1329 tmp = RREG32(mmSRBM_SOFT_RESET);
1330
1331 /* Wait a little for things to settle down */
1332 udelay(50);
1333 }
1334
1335 return 0;
1336}
1337
1338static int sdma_v3_0_set_trap_irq_state(struct amdgpu_device *adev,
1339 struct amdgpu_irq_src *source,
1340 unsigned type,
1341 enum amdgpu_interrupt_state state)
1342{
1343 u32 sdma_cntl;
1344
1345 switch (type) {
1346 case AMDGPU_SDMA_IRQ_INSTANCE0:
1347 switch (state) {
1348 case AMDGPU_IRQ_STATE_DISABLE:
1349 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1350 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 0);
1351 WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1352 break;
1353 case AMDGPU_IRQ_STATE_ENABLE:
1354 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1355 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 1);
1356 WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1357 break;
1358 default:
1359 break;
1360 }
1361 break;
1362 case AMDGPU_SDMA_IRQ_INSTANCE1:
1363 switch (state) {
1364 case AMDGPU_IRQ_STATE_DISABLE:
1365 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1366 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 0);
1367 WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1368 break;
1369 case AMDGPU_IRQ_STATE_ENABLE:
1370 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1371 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 1);
1372 WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1373 break;
1374 default:
1375 break;
1376 }
1377 break;
1378 default:
1379 break;
1380 }
1381 return 0;
1382}
1383
1384static int sdma_v3_0_process_trap_irq(struct amdgpu_device *adev,
1385 struct amdgpu_irq_src *source,
1386 struct amdgpu_iv_entry *entry)
1387{
1388 u8 instance_id, queue_id;
1389
1390 instance_id = (entry->ring_id & 0x3) >> 0;
1391 queue_id = (entry->ring_id & 0xc) >> 2;
1392 DRM_DEBUG("IH: SDMA trap\n");
1393 switch (instance_id) {
1394 case 0:
1395 switch (queue_id) {
1396 case 0:
1397 amdgpu_fence_process(&adev->sdma.instance[0].ring);
1398 break;
1399 case 1:
1400 /* XXX compute */
1401 break;
1402 case 2:
1403 /* XXX compute */
1404 break;
1405 }
1406 break;
1407 case 1:
1408 switch (queue_id) {
1409 case 0:
1410 amdgpu_fence_process(&adev->sdma.instance[1].ring);
1411 break;
1412 case 1:
1413 /* XXX compute */
1414 break;
1415 case 2:
1416 /* XXX compute */
1417 break;
1418 }
1419 break;
1420 }
1421 return 0;
1422}
1423
1424static int sdma_v3_0_process_illegal_inst_irq(struct amdgpu_device *adev,
1425 struct amdgpu_irq_src *source,
1426 struct amdgpu_iv_entry *entry)
1427{
1428 u8 instance_id, queue_id;
1429
1430 DRM_ERROR("Illegal instruction in SDMA command stream\n");
1431 instance_id = (entry->ring_id & 0x3) >> 0;
1432 queue_id = (entry->ring_id & 0xc) >> 2;
1433
1434 if (instance_id <= 1 && queue_id == 0)
1435 drm_sched_fault(&adev->sdma.instance[instance_id].ring.sched);
1436 return 0;
1437}
1438
1439static void sdma_v3_0_update_sdma_medium_grain_clock_gating(
1440 struct amdgpu_device *adev,
1441 bool enable)
1442{
1443 uint32_t temp, data;
1444 int i;
1445
1446 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
1447 for (i = 0; i < adev->sdma.num_instances; i++) {
1448 temp = data = RREG32(mmSDMA0_CLK_CTRL + sdma_offsets[i]);
1449 data &= ~(SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK |
1450 SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK |
1451 SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK |
1452 SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1453 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1454 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1455 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1456 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK);
1457 if (data != temp)
1458 WREG32(mmSDMA0_CLK_CTRL + sdma_offsets[i], data);
1459 }
1460 } else {
1461 for (i = 0; i < adev->sdma.num_instances; i++) {
1462 temp = data = RREG32(mmSDMA0_CLK_CTRL + sdma_offsets[i]);
1463 data |= SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK |
1464 SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK |
1465 SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK |
1466 SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1467 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1468 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1469 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1470 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK;
1471
1472 if (data != temp)
1473 WREG32(mmSDMA0_CLK_CTRL + sdma_offsets[i], data);
1474 }
1475 }
1476}
1477
1478static void sdma_v3_0_update_sdma_medium_grain_light_sleep(
1479 struct amdgpu_device *adev,
1480 bool enable)
1481{
1482 uint32_t temp, data;
1483 int i;
1484
1485 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
1486 for (i = 0; i < adev->sdma.num_instances; i++) {
1487 temp = data = RREG32(mmSDMA0_POWER_CNTL + sdma_offsets[i]);
1488 data |= SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1489
1490 if (temp != data)
1491 WREG32(mmSDMA0_POWER_CNTL + sdma_offsets[i], data);
1492 }
1493 } else {
1494 for (i = 0; i < adev->sdma.num_instances; i++) {
1495 temp = data = RREG32(mmSDMA0_POWER_CNTL + sdma_offsets[i]);
1496 data &= ~SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1497
1498 if (temp != data)
1499 WREG32(mmSDMA0_POWER_CNTL + sdma_offsets[i], data);
1500 }
1501 }
1502}
1503
1504static int sdma_v3_0_set_clockgating_state(void *handle,
1505 enum amd_clockgating_state state)
1506{
1507 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1508
1509 if (amdgpu_sriov_vf(adev))
1510 return 0;
1511
1512 switch (adev->asic_type) {
1513 case CHIP_FIJI:
1514 case CHIP_CARRIZO:
1515 case CHIP_STONEY:
1516 sdma_v3_0_update_sdma_medium_grain_clock_gating(adev,
1517 state == AMD_CG_STATE_GATE);
1518 sdma_v3_0_update_sdma_medium_grain_light_sleep(adev,
1519 state == AMD_CG_STATE_GATE);
1520 break;
1521 default:
1522 break;
1523 }
1524 return 0;
1525}
1526
1527static int sdma_v3_0_set_powergating_state(void *handle,
1528 enum amd_powergating_state state)
1529{
1530 return 0;
1531}
1532
1533static void sdma_v3_0_get_clockgating_state(void *handle, u32 *flags)
1534{
1535 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1536 int data;
1537
1538 if (amdgpu_sriov_vf(adev))
1539 *flags = 0;
1540
1541 /* AMD_CG_SUPPORT_SDMA_MGCG */
1542 data = RREG32(mmSDMA0_CLK_CTRL + sdma_offsets[0]);
1543 if (!(data & SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK))
1544 *flags |= AMD_CG_SUPPORT_SDMA_MGCG;
1545
1546 /* AMD_CG_SUPPORT_SDMA_LS */
1547 data = RREG32(mmSDMA0_POWER_CNTL + sdma_offsets[0]);
1548 if (data & SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK)
1549 *flags |= AMD_CG_SUPPORT_SDMA_LS;
1550}
1551
1552static const struct amd_ip_funcs sdma_v3_0_ip_funcs = {
1553 .name = "sdma_v3_0",
1554 .early_init = sdma_v3_0_early_init,
1555 .late_init = NULL,
1556 .sw_init = sdma_v3_0_sw_init,
1557 .sw_fini = sdma_v3_0_sw_fini,
1558 .hw_init = sdma_v3_0_hw_init,
1559 .hw_fini = sdma_v3_0_hw_fini,
1560 .suspend = sdma_v3_0_suspend,
1561 .resume = sdma_v3_0_resume,
1562 .is_idle = sdma_v3_0_is_idle,
1563 .wait_for_idle = sdma_v3_0_wait_for_idle,
1564 .check_soft_reset = sdma_v3_0_check_soft_reset,
1565 .pre_soft_reset = sdma_v3_0_pre_soft_reset,
1566 .post_soft_reset = sdma_v3_0_post_soft_reset,
1567 .soft_reset = sdma_v3_0_soft_reset,
1568 .set_clockgating_state = sdma_v3_0_set_clockgating_state,
1569 .set_powergating_state = sdma_v3_0_set_powergating_state,
1570 .get_clockgating_state = sdma_v3_0_get_clockgating_state,
1571};
1572
1573static const struct amdgpu_ring_funcs sdma_v3_0_ring_funcs = {
1574 .type = AMDGPU_RING_TYPE_SDMA,
1575 .align_mask = 0xf,
1576 .nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
1577 .support_64bit_ptrs = false,
1578 .get_rptr = sdma_v3_0_ring_get_rptr,
1579 .get_wptr = sdma_v3_0_ring_get_wptr,
1580 .set_wptr = sdma_v3_0_ring_set_wptr,
1581 .emit_frame_size =
1582 6 + /* sdma_v3_0_ring_emit_hdp_flush */
1583 3 + /* hdp invalidate */
1584 6 + /* sdma_v3_0_ring_emit_pipeline_sync */
1585 VI_FLUSH_GPU_TLB_NUM_WREG * 3 + 6 + /* sdma_v3_0_ring_emit_vm_flush */
1586 10 + 10 + 10, /* sdma_v3_0_ring_emit_fence x3 for user fence, vm fence */
1587 .emit_ib_size = 7 + 6, /* sdma_v3_0_ring_emit_ib */
1588 .emit_ib = sdma_v3_0_ring_emit_ib,
1589 .emit_fence = sdma_v3_0_ring_emit_fence,
1590 .emit_pipeline_sync = sdma_v3_0_ring_emit_pipeline_sync,
1591 .emit_vm_flush = sdma_v3_0_ring_emit_vm_flush,
1592 .emit_hdp_flush = sdma_v3_0_ring_emit_hdp_flush,
1593 .test_ring = sdma_v3_0_ring_test_ring,
1594 .test_ib = sdma_v3_0_ring_test_ib,
1595 .insert_nop = sdma_v3_0_ring_insert_nop,
1596 .pad_ib = sdma_v3_0_ring_pad_ib,
1597 .emit_wreg = sdma_v3_0_ring_emit_wreg,
1598};
1599
1600static void sdma_v3_0_set_ring_funcs(struct amdgpu_device *adev)
1601{
1602 int i;
1603
1604 for (i = 0; i < adev->sdma.num_instances; i++) {
1605 adev->sdma.instance[i].ring.funcs = &sdma_v3_0_ring_funcs;
1606 adev->sdma.instance[i].ring.me = i;
1607 }
1608}
1609
1610static const struct amdgpu_irq_src_funcs sdma_v3_0_trap_irq_funcs = {
1611 .set = sdma_v3_0_set_trap_irq_state,
1612 .process = sdma_v3_0_process_trap_irq,
1613};
1614
1615static const struct amdgpu_irq_src_funcs sdma_v3_0_illegal_inst_irq_funcs = {
1616 .process = sdma_v3_0_process_illegal_inst_irq,
1617};
1618
1619static void sdma_v3_0_set_irq_funcs(struct amdgpu_device *adev)
1620{
1621 adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
1622 adev->sdma.trap_irq.funcs = &sdma_v3_0_trap_irq_funcs;
1623 adev->sdma.illegal_inst_irq.funcs = &sdma_v3_0_illegal_inst_irq_funcs;
1624}
1625
1626/**
1627 * sdma_v3_0_emit_copy_buffer - copy buffer using the sDMA engine
1628 *
1629 * @ring: amdgpu_ring structure holding ring information
1630 * @src_offset: src GPU address
1631 * @dst_offset: dst GPU address
1632 * @byte_count: number of bytes to xfer
1633 *
1634 * Copy GPU buffers using the DMA engine (VI).
1635 * Used by the amdgpu ttm implementation to move pages if
1636 * registered as the asic copy callback.
1637 */
1638static void sdma_v3_0_emit_copy_buffer(struct amdgpu_ib *ib,
1639 uint64_t src_offset,
1640 uint64_t dst_offset,
1641 uint32_t byte_count)
1642{
1643 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1644 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1645 ib->ptr[ib->length_dw++] = byte_count;
1646 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1647 ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1648 ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1649 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1650 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1651}
1652
1653/**
1654 * sdma_v3_0_emit_fill_buffer - fill buffer using the sDMA engine
1655 *
1656 * @ring: amdgpu_ring structure holding ring information
1657 * @src_data: value to write to buffer
1658 * @dst_offset: dst GPU address
1659 * @byte_count: number of bytes to xfer
1660 *
1661 * Fill GPU buffers using the DMA engine (VI).
1662 */
1663static void sdma_v3_0_emit_fill_buffer(struct amdgpu_ib *ib,
1664 uint32_t src_data,
1665 uint64_t dst_offset,
1666 uint32_t byte_count)
1667{
1668 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL);
1669 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1670 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1671 ib->ptr[ib->length_dw++] = src_data;
1672 ib->ptr[ib->length_dw++] = byte_count;
1673}
1674
1675static const struct amdgpu_buffer_funcs sdma_v3_0_buffer_funcs = {
1676 .copy_max_bytes = 0x3fffe0, /* not 0x3fffff due to HW limitation */
1677 .copy_num_dw = 7,
1678 .emit_copy_buffer = sdma_v3_0_emit_copy_buffer,
1679
1680 .fill_max_bytes = 0x3fffe0, /* not 0x3fffff due to HW limitation */
1681 .fill_num_dw = 5,
1682 .emit_fill_buffer = sdma_v3_0_emit_fill_buffer,
1683};
1684
1685static void sdma_v3_0_set_buffer_funcs(struct amdgpu_device *adev)
1686{
1687 adev->mman.buffer_funcs = &sdma_v3_0_buffer_funcs;
1688 adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1689}
1690
1691static const struct amdgpu_vm_pte_funcs sdma_v3_0_vm_pte_funcs = {
1692 .copy_pte_num_dw = 7,
1693 .copy_pte = sdma_v3_0_vm_copy_pte,
1694
1695 .write_pte = sdma_v3_0_vm_write_pte,
1696 .set_pte_pde = sdma_v3_0_vm_set_pte_pde,
1697};
1698
1699static void sdma_v3_0_set_vm_pte_funcs(struct amdgpu_device *adev)
1700{
1701 struct drm_gpu_scheduler *sched;
1702 unsigned i;
1703
1704 adev->vm_manager.vm_pte_funcs = &sdma_v3_0_vm_pte_funcs;
1705 for (i = 0; i < adev->sdma.num_instances; i++) {
1706 sched = &adev->sdma.instance[i].ring.sched;
1707 adev->vm_manager.vm_pte_rqs[i] =
1708 &sched->sched_rq[DRM_SCHED_PRIORITY_KERNEL];
1709 }
1710 adev->vm_manager.vm_pte_num_rqs = adev->sdma.num_instances;
1711}
1712
1713const struct amdgpu_ip_block_version sdma_v3_0_ip_block =
1714{
1715 .type = AMD_IP_BLOCK_TYPE_SDMA,
1716 .major = 3,
1717 .minor = 0,
1718 .rev = 0,
1719 .funcs = &sdma_v3_0_ip_funcs,
1720};
1721
1722const struct amdgpu_ip_block_version sdma_v3_1_ip_block =
1723{
1724 .type = AMD_IP_BLOCK_TYPE_SDMA,
1725 .major = 3,
1726 .minor = 1,
1727 .rev = 0,
1728 .funcs = &sdma_v3_0_ip_funcs,
1729};