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 char fw_name[30];
271 int err = 0, i;
272 struct amdgpu_firmware_info *info = NULL;
273 const struct common_firmware_header *header = NULL;
274 const struct sdma_firmware_header_v1_0 *hdr;
275
276 DRM_DEBUG("\n");
277
278 switch (adev->asic_type) {
279 case CHIP_TONGA:
280 chip_name = "tonga";
281 break;
282 case CHIP_FIJI:
283 chip_name = "fiji";
284 break;
285 case CHIP_POLARIS10:
286 chip_name = "polaris10";
287 break;
288 case CHIP_POLARIS11:
289 chip_name = "polaris11";
290 break;
291 case CHIP_POLARIS12:
292 chip_name = "polaris12";
293 break;
294 case CHIP_VEGAM:
295 chip_name = "vegam";
296 break;
297 case CHIP_CARRIZO:
298 chip_name = "carrizo";
299 break;
300 case CHIP_STONEY:
301 chip_name = "stoney";
302 break;
303 default: BUG();
304 }
305
306 for (i = 0; i < adev->sdma.num_instances; i++) {
307 if (i == 0)
308 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma.bin", chip_name);
309 else
310 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma1.bin", chip_name);
311 err = amdgpu_ucode_request(adev, &adev->sdma.instance[i].fw, fw_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\"\n", fw_name);
331 for (i = 0; i < adev->sdma.num_instances; i++)
332 amdgpu_ucode_release(&adev->sdma.instance[i].fw);
333 }
334 return err;
335}
336
337/**
338 * sdma_v3_0_ring_get_rptr - get the current read pointer
339 *
340 * @ring: amdgpu ring pointer
341 *
342 * Get the current rptr from the hardware (VI+).
343 */
344static uint64_t sdma_v3_0_ring_get_rptr(struct amdgpu_ring *ring)
345{
346 /* XXX check if swapping is necessary on BE */
347 return *ring->rptr_cpu_addr >> 2;
348}
349
350/**
351 * sdma_v3_0_ring_get_wptr - get the current write pointer
352 *
353 * @ring: amdgpu ring pointer
354 *
355 * Get the current wptr from the hardware (VI+).
356 */
357static uint64_t sdma_v3_0_ring_get_wptr(struct amdgpu_ring *ring)
358{
359 struct amdgpu_device *adev = ring->adev;
360 u32 wptr;
361
362 if (ring->use_doorbell || ring->use_pollmem) {
363 /* XXX check if swapping is necessary on BE */
364 wptr = *ring->wptr_cpu_addr >> 2;
365 } else {
366 wptr = RREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[ring->me]) >> 2;
367 }
368
369 return wptr;
370}
371
372/**
373 * sdma_v3_0_ring_set_wptr - commit the write pointer
374 *
375 * @ring: amdgpu ring pointer
376 *
377 * Write the wptr back to the hardware (VI+).
378 */
379static void sdma_v3_0_ring_set_wptr(struct amdgpu_ring *ring)
380{
381 struct amdgpu_device *adev = ring->adev;
382
383 if (ring->use_doorbell) {
384 u32 *wb = (u32 *)ring->wptr_cpu_addr;
385 /* XXX check if swapping is necessary on BE */
386 WRITE_ONCE(*wb, ring->wptr << 2);
387 WDOORBELL32(ring->doorbell_index, ring->wptr << 2);
388 } else if (ring->use_pollmem) {
389 u32 *wb = (u32 *)ring->wptr_cpu_addr;
390
391 WRITE_ONCE(*wb, ring->wptr << 2);
392 } else {
393 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[ring->me], ring->wptr << 2);
394 }
395}
396
397static void sdma_v3_0_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
398{
399 struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
400 int i;
401
402 for (i = 0; i < count; i++)
403 if (sdma && sdma->burst_nop && (i == 0))
404 amdgpu_ring_write(ring, ring->funcs->nop |
405 SDMA_PKT_NOP_HEADER_COUNT(count - 1));
406 else
407 amdgpu_ring_write(ring, ring->funcs->nop);
408}
409
410/**
411 * sdma_v3_0_ring_emit_ib - Schedule an IB on the DMA engine
412 *
413 * @ring: amdgpu ring pointer
414 * @job: job to retrieve vmid from
415 * @ib: IB object to schedule
416 * @flags: unused
417 *
418 * Schedule an IB in the DMA ring (VI).
419 */
420static void sdma_v3_0_ring_emit_ib(struct amdgpu_ring *ring,
421 struct amdgpu_job *job,
422 struct amdgpu_ib *ib,
423 uint32_t flags)
424{
425 unsigned vmid = AMDGPU_JOB_GET_VMID(job);
426
427 /* IB packet must end on a 8 DW boundary */
428 sdma_v3_0_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7);
429
430 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) |
431 SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf));
432 /* base must be 32 byte aligned */
433 amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
434 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
435 amdgpu_ring_write(ring, ib->length_dw);
436 amdgpu_ring_write(ring, 0);
437 amdgpu_ring_write(ring, 0);
438
439}
440
441/**
442 * sdma_v3_0_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
443 *
444 * @ring: amdgpu ring pointer
445 *
446 * Emit an hdp flush packet on the requested DMA ring.
447 */
448static void sdma_v3_0_ring_emit_hdp_flush(struct amdgpu_ring *ring)
449{
450 u32 ref_and_mask = 0;
451
452 if (ring->me == 0)
453 ref_and_mask = REG_SET_FIELD(ref_and_mask, GPU_HDP_FLUSH_DONE, SDMA0, 1);
454 else
455 ref_and_mask = REG_SET_FIELD(ref_and_mask, GPU_HDP_FLUSH_DONE, SDMA1, 1);
456
457 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
458 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
459 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
460 amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_DONE << 2);
461 amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_REQ << 2);
462 amdgpu_ring_write(ring, ref_and_mask); /* reference */
463 amdgpu_ring_write(ring, ref_and_mask); /* mask */
464 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
465 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
466}
467
468/**
469 * sdma_v3_0_ring_emit_fence - emit a fence on the DMA ring
470 *
471 * @ring: amdgpu ring pointer
472 * @addr: address
473 * @seq: sequence number
474 * @flags: fence related flags
475 *
476 * Add a DMA fence packet to the ring to write
477 * the fence seq number and DMA trap packet to generate
478 * an interrupt if needed (VI).
479 */
480static void sdma_v3_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
481 unsigned flags)
482{
483 bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
484 /* write the fence */
485 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE));
486 amdgpu_ring_write(ring, lower_32_bits(addr));
487 amdgpu_ring_write(ring, upper_32_bits(addr));
488 amdgpu_ring_write(ring, lower_32_bits(seq));
489
490 /* optionally write high bits as well */
491 if (write64bit) {
492 addr += 4;
493 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE));
494 amdgpu_ring_write(ring, lower_32_bits(addr));
495 amdgpu_ring_write(ring, upper_32_bits(addr));
496 amdgpu_ring_write(ring, upper_32_bits(seq));
497 }
498
499 /* generate an interrupt */
500 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP));
501 amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(0));
502}
503
504/**
505 * sdma_v3_0_gfx_stop - stop the gfx async dma engines
506 *
507 * @adev: amdgpu_device pointer
508 *
509 * Stop the gfx async dma ring buffers (VI).
510 */
511static void sdma_v3_0_gfx_stop(struct amdgpu_device *adev)
512{
513 u32 rb_cntl, ib_cntl;
514 int i;
515
516 for (i = 0; i < adev->sdma.num_instances; i++) {
517 rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
518 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0);
519 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
520 ib_cntl = RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]);
521 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0);
522 WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
523 }
524}
525
526/**
527 * sdma_v3_0_rlc_stop - stop the compute async dma engines
528 *
529 * @adev: amdgpu_device pointer
530 *
531 * Stop the compute async dma queues (VI).
532 */
533static void sdma_v3_0_rlc_stop(struct amdgpu_device *adev)
534{
535 /* XXX todo */
536}
537
538/**
539 * sdma_v3_0_ctx_switch_enable - stop the async dma engines context switch
540 *
541 * @adev: amdgpu_device pointer
542 * @enable: enable/disable the DMA MEs context switch.
543 *
544 * Halt or unhalt the async dma engines context switch (VI).
545 */
546static void sdma_v3_0_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
547{
548 u32 f32_cntl, phase_quantum = 0;
549 int i;
550
551 if (amdgpu_sdma_phase_quantum) {
552 unsigned value = amdgpu_sdma_phase_quantum;
553 unsigned unit = 0;
554
555 while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
556 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
557 value = (value + 1) >> 1;
558 unit++;
559 }
560 if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
561 SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
562 value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
563 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
564 unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
565 SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
566 WARN_ONCE(1,
567 "clamping sdma_phase_quantum to %uK clock cycles\n",
568 value << unit);
569 }
570 phase_quantum =
571 value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
572 unit << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
573 }
574
575 for (i = 0; i < adev->sdma.num_instances; i++) {
576 f32_cntl = RREG32(mmSDMA0_CNTL + sdma_offsets[i]);
577 if (enable) {
578 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
579 AUTO_CTXSW_ENABLE, 1);
580 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
581 ATC_L1_ENABLE, 1);
582 if (amdgpu_sdma_phase_quantum) {
583 WREG32(mmSDMA0_PHASE0_QUANTUM + sdma_offsets[i],
584 phase_quantum);
585 WREG32(mmSDMA0_PHASE1_QUANTUM + sdma_offsets[i],
586 phase_quantum);
587 }
588 } else {
589 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
590 AUTO_CTXSW_ENABLE, 0);
591 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
592 ATC_L1_ENABLE, 1);
593 }
594
595 WREG32(mmSDMA0_CNTL + sdma_offsets[i], f32_cntl);
596 }
597}
598
599/**
600 * sdma_v3_0_enable - stop the async dma engines
601 *
602 * @adev: amdgpu_device pointer
603 * @enable: enable/disable the DMA MEs.
604 *
605 * Halt or unhalt the async dma engines (VI).
606 */
607static void sdma_v3_0_enable(struct amdgpu_device *adev, bool enable)
608{
609 u32 f32_cntl;
610 int i;
611
612 if (!enable) {
613 sdma_v3_0_gfx_stop(adev);
614 sdma_v3_0_rlc_stop(adev);
615 }
616
617 for (i = 0; i < adev->sdma.num_instances; i++) {
618 f32_cntl = RREG32(mmSDMA0_F32_CNTL + sdma_offsets[i]);
619 if (enable)
620 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 0);
621 else
622 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 1);
623 WREG32(mmSDMA0_F32_CNTL + sdma_offsets[i], f32_cntl);
624 }
625}
626
627/**
628 * sdma_v3_0_gfx_resume - setup and start the async dma engines
629 *
630 * @adev: amdgpu_device pointer
631 *
632 * Set up the gfx DMA ring buffers and enable them (VI).
633 * Returns 0 for success, error for failure.
634 */
635static int sdma_v3_0_gfx_resume(struct amdgpu_device *adev)
636{
637 struct amdgpu_ring *ring;
638 u32 rb_cntl, ib_cntl, wptr_poll_cntl;
639 u32 rb_bufsz;
640 u32 doorbell;
641 u64 wptr_gpu_addr;
642 int i, j, r;
643
644 for (i = 0; i < adev->sdma.num_instances; i++) {
645 ring = &adev->sdma.instance[i].ring;
646 amdgpu_ring_clear_ring(ring);
647
648 mutex_lock(&adev->srbm_mutex);
649 for (j = 0; j < 16; j++) {
650 vi_srbm_select(adev, 0, 0, 0, j);
651 /* SDMA GFX */
652 WREG32(mmSDMA0_GFX_VIRTUAL_ADDR + sdma_offsets[i], 0);
653 WREG32(mmSDMA0_GFX_APE1_CNTL + sdma_offsets[i], 0);
654 }
655 vi_srbm_select(adev, 0, 0, 0, 0);
656 mutex_unlock(&adev->srbm_mutex);
657
658 WREG32(mmSDMA0_TILING_CONFIG + sdma_offsets[i],
659 adev->gfx.config.gb_addr_config & 0x70);
660
661 WREG32(mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
662
663 /* Set ring buffer size in dwords */
664 rb_bufsz = order_base_2(ring->ring_size / 4);
665 rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
666 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz);
667#ifdef __BIG_ENDIAN
668 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1);
669 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
670 RPTR_WRITEBACK_SWAP_ENABLE, 1);
671#endif
672 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
673
674 /* Initialize the ring buffer's read and write pointers */
675 ring->wptr = 0;
676 WREG32(mmSDMA0_GFX_RB_RPTR + sdma_offsets[i], 0);
677 sdma_v3_0_ring_set_wptr(ring);
678 WREG32(mmSDMA0_GFX_IB_RPTR + sdma_offsets[i], 0);
679 WREG32(mmSDMA0_GFX_IB_OFFSET + sdma_offsets[i], 0);
680
681 /* set the wb address whether it's enabled or not */
682 WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_HI + sdma_offsets[i],
683 upper_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFF);
684 WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_LO + sdma_offsets[i],
685 lower_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFC);
686
687 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
688
689 WREG32(mmSDMA0_GFX_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
690 WREG32(mmSDMA0_GFX_RB_BASE_HI + sdma_offsets[i], ring->gpu_addr >> 40);
691
692 doorbell = RREG32(mmSDMA0_GFX_DOORBELL + sdma_offsets[i]);
693
694 if (ring->use_doorbell) {
695 doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL,
696 OFFSET, ring->doorbell_index);
697 doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 1);
698 } else {
699 doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 0);
700 }
701 WREG32(mmSDMA0_GFX_DOORBELL + sdma_offsets[i], doorbell);
702
703 /* setup the wptr shadow polling */
704 wptr_gpu_addr = ring->wptr_gpu_addr;
705
706 WREG32(mmSDMA0_GFX_RB_WPTR_POLL_ADDR_LO + sdma_offsets[i],
707 lower_32_bits(wptr_gpu_addr));
708 WREG32(mmSDMA0_GFX_RB_WPTR_POLL_ADDR_HI + sdma_offsets[i],
709 upper_32_bits(wptr_gpu_addr));
710 wptr_poll_cntl = RREG32(mmSDMA0_GFX_RB_WPTR_POLL_CNTL + sdma_offsets[i]);
711 if (ring->use_pollmem) {
712 /*wptr polling is not enogh fast, directly clean the wptr register */
713 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], 0);
714 wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl,
715 SDMA0_GFX_RB_WPTR_POLL_CNTL,
716 ENABLE, 1);
717 } else {
718 wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl,
719 SDMA0_GFX_RB_WPTR_POLL_CNTL,
720 ENABLE, 0);
721 }
722 WREG32(mmSDMA0_GFX_RB_WPTR_POLL_CNTL + sdma_offsets[i], wptr_poll_cntl);
723
724 /* enable DMA RB */
725 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1);
726 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
727
728 ib_cntl = RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]);
729 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1);
730#ifdef __BIG_ENDIAN
731 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1);
732#endif
733 /* enable DMA IBs */
734 WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
735 }
736
737 /* unhalt the MEs */
738 sdma_v3_0_enable(adev, true);
739 /* enable sdma ring preemption */
740 sdma_v3_0_ctx_switch_enable(adev, true);
741
742 for (i = 0; i < adev->sdma.num_instances; i++) {
743 ring = &adev->sdma.instance[i].ring;
744 r = amdgpu_ring_test_helper(ring);
745 if (r)
746 return r;
747 }
748
749 return 0;
750}
751
752/**
753 * sdma_v3_0_rlc_resume - setup and start the async dma engines
754 *
755 * @adev: amdgpu_device pointer
756 *
757 * Set up the compute DMA queues and enable them (VI).
758 * Returns 0 for success, error for failure.
759 */
760static int sdma_v3_0_rlc_resume(struct amdgpu_device *adev)
761{
762 /* XXX todo */
763 return 0;
764}
765
766/**
767 * sdma_v3_0_start - setup and start the async dma engines
768 *
769 * @adev: amdgpu_device pointer
770 *
771 * Set up the DMA engines and enable them (VI).
772 * Returns 0 for success, error for failure.
773 */
774static int sdma_v3_0_start(struct amdgpu_device *adev)
775{
776 int r;
777
778 /* disable sdma engine before programing it */
779 sdma_v3_0_ctx_switch_enable(adev, false);
780 sdma_v3_0_enable(adev, false);
781
782 /* start the gfx rings and rlc compute queues */
783 r = sdma_v3_0_gfx_resume(adev);
784 if (r)
785 return r;
786 r = sdma_v3_0_rlc_resume(adev);
787 if (r)
788 return r;
789
790 return 0;
791}
792
793/**
794 * sdma_v3_0_ring_test_ring - simple async dma engine test
795 *
796 * @ring: amdgpu_ring structure holding ring information
797 *
798 * Test the DMA engine by writing using it to write an
799 * value to memory. (VI).
800 * Returns 0 for success, error for failure.
801 */
802static int sdma_v3_0_ring_test_ring(struct amdgpu_ring *ring)
803{
804 struct amdgpu_device *adev = ring->adev;
805 unsigned i;
806 unsigned index;
807 int r;
808 u32 tmp;
809 u64 gpu_addr;
810
811 r = amdgpu_device_wb_get(adev, &index);
812 if (r)
813 return r;
814
815 gpu_addr = adev->wb.gpu_addr + (index * 4);
816 tmp = 0xCAFEDEAD;
817 adev->wb.wb[index] = cpu_to_le32(tmp);
818
819 r = amdgpu_ring_alloc(ring, 5);
820 if (r)
821 goto error_free_wb;
822
823 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
824 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
825 amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
826 amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
827 amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1));
828 amdgpu_ring_write(ring, 0xDEADBEEF);
829 amdgpu_ring_commit(ring);
830
831 for (i = 0; i < adev->usec_timeout; i++) {
832 tmp = le32_to_cpu(adev->wb.wb[index]);
833 if (tmp == 0xDEADBEEF)
834 break;
835 udelay(1);
836 }
837
838 if (i >= adev->usec_timeout)
839 r = -ETIMEDOUT;
840
841error_free_wb:
842 amdgpu_device_wb_free(adev, index);
843 return r;
844}
845
846/**
847 * sdma_v3_0_ring_test_ib - test an IB on the DMA engine
848 *
849 * @ring: amdgpu_ring structure holding ring information
850 * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
851 *
852 * Test a simple IB in the DMA ring (VI).
853 * Returns 0 on success, error on failure.
854 */
855static int sdma_v3_0_ring_test_ib(struct amdgpu_ring *ring, long timeout)
856{
857 struct amdgpu_device *adev = ring->adev;
858 struct amdgpu_ib ib;
859 struct dma_fence *f = NULL;
860 unsigned index;
861 u32 tmp = 0;
862 u64 gpu_addr;
863 long r;
864
865 r = amdgpu_device_wb_get(adev, &index);
866 if (r)
867 return r;
868
869 gpu_addr = adev->wb.gpu_addr + (index * 4);
870 tmp = 0xCAFEDEAD;
871 adev->wb.wb[index] = cpu_to_le32(tmp);
872 memset(&ib, 0, sizeof(ib));
873 r = amdgpu_ib_get(adev, NULL, 256,
874 AMDGPU_IB_POOL_DIRECT, &ib);
875 if (r)
876 goto err0;
877
878 ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
879 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
880 ib.ptr[1] = lower_32_bits(gpu_addr);
881 ib.ptr[2] = upper_32_bits(gpu_addr);
882 ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1);
883 ib.ptr[4] = 0xDEADBEEF;
884 ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
885 ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
886 ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
887 ib.length_dw = 8;
888
889 r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
890 if (r)
891 goto err1;
892
893 r = dma_fence_wait_timeout(f, false, timeout);
894 if (r == 0) {
895 r = -ETIMEDOUT;
896 goto err1;
897 } else if (r < 0) {
898 goto err1;
899 }
900 tmp = le32_to_cpu(adev->wb.wb[index]);
901 if (tmp == 0xDEADBEEF)
902 r = 0;
903 else
904 r = -EINVAL;
905err1:
906 amdgpu_ib_free(adev, &ib, NULL);
907 dma_fence_put(f);
908err0:
909 amdgpu_device_wb_free(adev, index);
910 return r;
911}
912
913/**
914 * sdma_v3_0_vm_copy_pte - update PTEs by copying them from the GART
915 *
916 * @ib: indirect buffer to fill with commands
917 * @pe: addr of the page entry
918 * @src: src addr to copy from
919 * @count: number of page entries to update
920 *
921 * Update PTEs by copying them from the GART using sDMA (CIK).
922 */
923static void sdma_v3_0_vm_copy_pte(struct amdgpu_ib *ib,
924 uint64_t pe, uint64_t src,
925 unsigned count)
926{
927 unsigned bytes = count * 8;
928
929 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
930 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
931 ib->ptr[ib->length_dw++] = bytes;
932 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
933 ib->ptr[ib->length_dw++] = lower_32_bits(src);
934 ib->ptr[ib->length_dw++] = upper_32_bits(src);
935 ib->ptr[ib->length_dw++] = lower_32_bits(pe);
936 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
937}
938
939/**
940 * sdma_v3_0_vm_write_pte - update PTEs by writing them manually
941 *
942 * @ib: indirect buffer to fill with commands
943 * @pe: addr of the page entry
944 * @value: dst addr to write into pe
945 * @count: number of page entries to update
946 * @incr: increase next addr by incr bytes
947 *
948 * Update PTEs by writing them manually using sDMA (CIK).
949 */
950static void sdma_v3_0_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
951 uint64_t value, unsigned count,
952 uint32_t incr)
953{
954 unsigned ndw = count * 2;
955
956 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
957 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
958 ib->ptr[ib->length_dw++] = lower_32_bits(pe);
959 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
960 ib->ptr[ib->length_dw++] = ndw;
961 for (; ndw > 0; ndw -= 2) {
962 ib->ptr[ib->length_dw++] = lower_32_bits(value);
963 ib->ptr[ib->length_dw++] = upper_32_bits(value);
964 value += incr;
965 }
966}
967
968/**
969 * sdma_v3_0_vm_set_pte_pde - update the page tables using sDMA
970 *
971 * @ib: indirect buffer to fill with commands
972 * @pe: addr of the page entry
973 * @addr: dst addr to write into pe
974 * @count: number of page entries to update
975 * @incr: increase next addr by incr bytes
976 * @flags: access flags
977 *
978 * Update the page tables using sDMA (CIK).
979 */
980static void sdma_v3_0_vm_set_pte_pde(struct amdgpu_ib *ib, uint64_t pe,
981 uint64_t addr, unsigned count,
982 uint32_t incr, uint64_t flags)
983{
984 /* for physically contiguous pages (vram) */
985 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_GEN_PTEPDE);
986 ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
987 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
988 ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
989 ib->ptr[ib->length_dw++] = upper_32_bits(flags);
990 ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
991 ib->ptr[ib->length_dw++] = upper_32_bits(addr);
992 ib->ptr[ib->length_dw++] = incr; /* increment size */
993 ib->ptr[ib->length_dw++] = 0;
994 ib->ptr[ib->length_dw++] = count; /* number of entries */
995}
996
997/**
998 * sdma_v3_0_ring_pad_ib - pad the IB to the required number of dw
999 *
1000 * @ring: amdgpu_ring structure holding ring information
1001 * @ib: indirect buffer to fill with padding
1002 *
1003 */
1004static void sdma_v3_0_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
1005{
1006 struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
1007 u32 pad_count;
1008 int i;
1009
1010 pad_count = (-ib->length_dw) & 7;
1011 for (i = 0; i < pad_count; i++)
1012 if (sdma && sdma->burst_nop && (i == 0))
1013 ib->ptr[ib->length_dw++] =
1014 SDMA_PKT_HEADER_OP(SDMA_OP_NOP) |
1015 SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
1016 else
1017 ib->ptr[ib->length_dw++] =
1018 SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
1019}
1020
1021/**
1022 * sdma_v3_0_ring_emit_pipeline_sync - sync the pipeline
1023 *
1024 * @ring: amdgpu_ring pointer
1025 *
1026 * Make sure all previous operations are completed (CIK).
1027 */
1028static void sdma_v3_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
1029{
1030 uint32_t seq = ring->fence_drv.sync_seq;
1031 uint64_t addr = ring->fence_drv.gpu_addr;
1032
1033 /* wait for idle */
1034 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1035 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1036 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
1037 SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
1038 amdgpu_ring_write(ring, addr & 0xfffffffc);
1039 amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
1040 amdgpu_ring_write(ring, seq); /* reference */
1041 amdgpu_ring_write(ring, 0xffffffff); /* mask */
1042 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1043 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
1044}
1045
1046/**
1047 * sdma_v3_0_ring_emit_vm_flush - cik vm flush using sDMA
1048 *
1049 * @ring: amdgpu_ring pointer
1050 * @vmid: vmid number to use
1051 * @pd_addr: address
1052 *
1053 * Update the page table base and flush the VM TLB
1054 * using sDMA (VI).
1055 */
1056static void sdma_v3_0_ring_emit_vm_flush(struct amdgpu_ring *ring,
1057 unsigned vmid, uint64_t pd_addr)
1058{
1059 amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
1060
1061 /* wait for flush */
1062 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1063 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1064 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(0)); /* always */
1065 amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST << 2);
1066 amdgpu_ring_write(ring, 0);
1067 amdgpu_ring_write(ring, 0); /* reference */
1068 amdgpu_ring_write(ring, 0); /* mask */
1069 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1070 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
1071}
1072
1073static void sdma_v3_0_ring_emit_wreg(struct amdgpu_ring *ring,
1074 uint32_t reg, uint32_t val)
1075{
1076 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
1077 SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
1078 amdgpu_ring_write(ring, reg);
1079 amdgpu_ring_write(ring, val);
1080}
1081
1082static int sdma_v3_0_early_init(void *handle)
1083{
1084 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1085 int r;
1086
1087 switch (adev->asic_type) {
1088 case CHIP_STONEY:
1089 adev->sdma.num_instances = 1;
1090 break;
1091 default:
1092 adev->sdma.num_instances = SDMA_MAX_INSTANCE;
1093 break;
1094 }
1095
1096 r = sdma_v3_0_init_microcode(adev);
1097 if (r)
1098 return r;
1099
1100 sdma_v3_0_set_ring_funcs(adev);
1101 sdma_v3_0_set_buffer_funcs(adev);
1102 sdma_v3_0_set_vm_pte_funcs(adev);
1103 sdma_v3_0_set_irq_funcs(adev);
1104
1105 return 0;
1106}
1107
1108static int sdma_v3_0_sw_init(void *handle)
1109{
1110 struct amdgpu_ring *ring;
1111 int r, i;
1112 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1113
1114 /* SDMA trap event */
1115 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_SDMA_TRAP,
1116 &adev->sdma.trap_irq);
1117 if (r)
1118 return r;
1119
1120 /* SDMA Privileged inst */
1121 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 241,
1122 &adev->sdma.illegal_inst_irq);
1123 if (r)
1124 return r;
1125
1126 /* SDMA Privileged inst */
1127 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_SDMA_SRBM_WRITE,
1128 &adev->sdma.illegal_inst_irq);
1129 if (r)
1130 return r;
1131
1132 for (i = 0; i < adev->sdma.num_instances; i++) {
1133 ring = &adev->sdma.instance[i].ring;
1134 ring->ring_obj = NULL;
1135 if (!amdgpu_sriov_vf(adev)) {
1136 ring->use_doorbell = true;
1137 ring->doorbell_index = adev->doorbell_index.sdma_engine[i];
1138 } else {
1139 ring->use_pollmem = true;
1140 }
1141
1142 sprintf(ring->name, "sdma%d", i);
1143 r = amdgpu_ring_init(adev, ring, 1024, &adev->sdma.trap_irq,
1144 (i == 0) ? AMDGPU_SDMA_IRQ_INSTANCE0 :
1145 AMDGPU_SDMA_IRQ_INSTANCE1,
1146 AMDGPU_RING_PRIO_DEFAULT, NULL);
1147 if (r)
1148 return r;
1149 }
1150
1151 return r;
1152}
1153
1154static int sdma_v3_0_sw_fini(void *handle)
1155{
1156 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1157 int i;
1158
1159 for (i = 0; i < adev->sdma.num_instances; i++)
1160 amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1161
1162 sdma_v3_0_free_microcode(adev);
1163 return 0;
1164}
1165
1166static int sdma_v3_0_hw_init(void *handle)
1167{
1168 int r;
1169 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1170
1171 sdma_v3_0_init_golden_registers(adev);
1172
1173 r = sdma_v3_0_start(adev);
1174 if (r)
1175 return r;
1176
1177 return r;
1178}
1179
1180static int sdma_v3_0_hw_fini(void *handle)
1181{
1182 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1183
1184 sdma_v3_0_ctx_switch_enable(adev, false);
1185 sdma_v3_0_enable(adev, false);
1186
1187 return 0;
1188}
1189
1190static int sdma_v3_0_suspend(void *handle)
1191{
1192 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1193
1194 return sdma_v3_0_hw_fini(adev);
1195}
1196
1197static int sdma_v3_0_resume(void *handle)
1198{
1199 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1200
1201 return sdma_v3_0_hw_init(adev);
1202}
1203
1204static bool sdma_v3_0_is_idle(void *handle)
1205{
1206 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1207 u32 tmp = RREG32(mmSRBM_STATUS2);
1208
1209 if (tmp & (SRBM_STATUS2__SDMA_BUSY_MASK |
1210 SRBM_STATUS2__SDMA1_BUSY_MASK))
1211 return false;
1212
1213 return true;
1214}
1215
1216static int sdma_v3_0_wait_for_idle(void *handle)
1217{
1218 unsigned i;
1219 u32 tmp;
1220 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1221
1222 for (i = 0; i < adev->usec_timeout; i++) {
1223 tmp = RREG32(mmSRBM_STATUS2) & (SRBM_STATUS2__SDMA_BUSY_MASK |
1224 SRBM_STATUS2__SDMA1_BUSY_MASK);
1225
1226 if (!tmp)
1227 return 0;
1228 udelay(1);
1229 }
1230 return -ETIMEDOUT;
1231}
1232
1233static bool sdma_v3_0_check_soft_reset(void *handle)
1234{
1235 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1236 u32 srbm_soft_reset = 0;
1237 u32 tmp = RREG32(mmSRBM_STATUS2);
1238
1239 if ((tmp & SRBM_STATUS2__SDMA_BUSY_MASK) ||
1240 (tmp & SRBM_STATUS2__SDMA1_BUSY_MASK)) {
1241 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA_MASK;
1242 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA1_MASK;
1243 }
1244
1245 if (srbm_soft_reset) {
1246 adev->sdma.srbm_soft_reset = srbm_soft_reset;
1247 return true;
1248 } else {
1249 adev->sdma.srbm_soft_reset = 0;
1250 return false;
1251 }
1252}
1253
1254static int sdma_v3_0_pre_soft_reset(void *handle)
1255{
1256 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1257 u32 srbm_soft_reset = 0;
1258
1259 if (!adev->sdma.srbm_soft_reset)
1260 return 0;
1261
1262 srbm_soft_reset = adev->sdma.srbm_soft_reset;
1263
1264 if (REG_GET_FIELD(srbm_soft_reset, SRBM_SOFT_RESET, SOFT_RESET_SDMA) ||
1265 REG_GET_FIELD(srbm_soft_reset, SRBM_SOFT_RESET, SOFT_RESET_SDMA1)) {
1266 sdma_v3_0_ctx_switch_enable(adev, false);
1267 sdma_v3_0_enable(adev, false);
1268 }
1269
1270 return 0;
1271}
1272
1273static int sdma_v3_0_post_soft_reset(void *handle)
1274{
1275 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1276 u32 srbm_soft_reset = 0;
1277
1278 if (!adev->sdma.srbm_soft_reset)
1279 return 0;
1280
1281 srbm_soft_reset = adev->sdma.srbm_soft_reset;
1282
1283 if (REG_GET_FIELD(srbm_soft_reset, SRBM_SOFT_RESET, SOFT_RESET_SDMA) ||
1284 REG_GET_FIELD(srbm_soft_reset, SRBM_SOFT_RESET, SOFT_RESET_SDMA1)) {
1285 sdma_v3_0_gfx_resume(adev);
1286 sdma_v3_0_rlc_resume(adev);
1287 }
1288
1289 return 0;
1290}
1291
1292static int sdma_v3_0_soft_reset(void *handle)
1293{
1294 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1295 u32 srbm_soft_reset = 0;
1296 u32 tmp;
1297
1298 if (!adev->sdma.srbm_soft_reset)
1299 return 0;
1300
1301 srbm_soft_reset = adev->sdma.srbm_soft_reset;
1302
1303 if (srbm_soft_reset) {
1304 tmp = RREG32(mmSRBM_SOFT_RESET);
1305 tmp |= srbm_soft_reset;
1306 dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
1307 WREG32(mmSRBM_SOFT_RESET, tmp);
1308 tmp = RREG32(mmSRBM_SOFT_RESET);
1309
1310 udelay(50);
1311
1312 tmp &= ~srbm_soft_reset;
1313 WREG32(mmSRBM_SOFT_RESET, tmp);
1314 tmp = RREG32(mmSRBM_SOFT_RESET);
1315
1316 /* Wait a little for things to settle down */
1317 udelay(50);
1318 }
1319
1320 return 0;
1321}
1322
1323static int sdma_v3_0_set_trap_irq_state(struct amdgpu_device *adev,
1324 struct amdgpu_irq_src *source,
1325 unsigned type,
1326 enum amdgpu_interrupt_state state)
1327{
1328 u32 sdma_cntl;
1329
1330 switch (type) {
1331 case AMDGPU_SDMA_IRQ_INSTANCE0:
1332 switch (state) {
1333 case AMDGPU_IRQ_STATE_DISABLE:
1334 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1335 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 0);
1336 WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1337 break;
1338 case AMDGPU_IRQ_STATE_ENABLE:
1339 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1340 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 1);
1341 WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1342 break;
1343 default:
1344 break;
1345 }
1346 break;
1347 case AMDGPU_SDMA_IRQ_INSTANCE1:
1348 switch (state) {
1349 case AMDGPU_IRQ_STATE_DISABLE:
1350 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1351 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 0);
1352 WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1353 break;
1354 case AMDGPU_IRQ_STATE_ENABLE:
1355 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1356 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 1);
1357 WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1358 break;
1359 default:
1360 break;
1361 }
1362 break;
1363 default:
1364 break;
1365 }
1366 return 0;
1367}
1368
1369static int sdma_v3_0_process_trap_irq(struct amdgpu_device *adev,
1370 struct amdgpu_irq_src *source,
1371 struct amdgpu_iv_entry *entry)
1372{
1373 u8 instance_id, queue_id;
1374
1375 instance_id = (entry->ring_id & 0x3) >> 0;
1376 queue_id = (entry->ring_id & 0xc) >> 2;
1377 DRM_DEBUG("IH: SDMA trap\n");
1378 switch (instance_id) {
1379 case 0:
1380 switch (queue_id) {
1381 case 0:
1382 amdgpu_fence_process(&adev->sdma.instance[0].ring);
1383 break;
1384 case 1:
1385 /* XXX compute */
1386 break;
1387 case 2:
1388 /* XXX compute */
1389 break;
1390 }
1391 break;
1392 case 1:
1393 switch (queue_id) {
1394 case 0:
1395 amdgpu_fence_process(&adev->sdma.instance[1].ring);
1396 break;
1397 case 1:
1398 /* XXX compute */
1399 break;
1400 case 2:
1401 /* XXX compute */
1402 break;
1403 }
1404 break;
1405 }
1406 return 0;
1407}
1408
1409static int sdma_v3_0_process_illegal_inst_irq(struct amdgpu_device *adev,
1410 struct amdgpu_irq_src *source,
1411 struct amdgpu_iv_entry *entry)
1412{
1413 u8 instance_id, queue_id;
1414
1415 DRM_ERROR("Illegal instruction in SDMA command stream\n");
1416 instance_id = (entry->ring_id & 0x3) >> 0;
1417 queue_id = (entry->ring_id & 0xc) >> 2;
1418
1419 if (instance_id <= 1 && queue_id == 0)
1420 drm_sched_fault(&adev->sdma.instance[instance_id].ring.sched);
1421 return 0;
1422}
1423
1424static void sdma_v3_0_update_sdma_medium_grain_clock_gating(
1425 struct amdgpu_device *adev,
1426 bool enable)
1427{
1428 uint32_t temp, data;
1429 int i;
1430
1431 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
1432 for (i = 0; i < adev->sdma.num_instances; i++) {
1433 temp = data = RREG32(mmSDMA0_CLK_CTRL + sdma_offsets[i]);
1434 data &= ~(SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK |
1435 SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK |
1436 SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK |
1437 SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1438 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1439 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1440 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1441 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK);
1442 if (data != temp)
1443 WREG32(mmSDMA0_CLK_CTRL + sdma_offsets[i], data);
1444 }
1445 } else {
1446 for (i = 0; i < adev->sdma.num_instances; i++) {
1447 temp = data = RREG32(mmSDMA0_CLK_CTRL + sdma_offsets[i]);
1448 data |= SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK |
1449 SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK |
1450 SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK |
1451 SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1452 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1453 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1454 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1455 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK;
1456
1457 if (data != temp)
1458 WREG32(mmSDMA0_CLK_CTRL + sdma_offsets[i], data);
1459 }
1460 }
1461}
1462
1463static void sdma_v3_0_update_sdma_medium_grain_light_sleep(
1464 struct amdgpu_device *adev,
1465 bool enable)
1466{
1467 uint32_t temp, data;
1468 int i;
1469
1470 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
1471 for (i = 0; i < adev->sdma.num_instances; i++) {
1472 temp = data = RREG32(mmSDMA0_POWER_CNTL + sdma_offsets[i]);
1473 data |= SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1474
1475 if (temp != data)
1476 WREG32(mmSDMA0_POWER_CNTL + sdma_offsets[i], data);
1477 }
1478 } else {
1479 for (i = 0; i < adev->sdma.num_instances; i++) {
1480 temp = data = RREG32(mmSDMA0_POWER_CNTL + sdma_offsets[i]);
1481 data &= ~SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1482
1483 if (temp != data)
1484 WREG32(mmSDMA0_POWER_CNTL + sdma_offsets[i], data);
1485 }
1486 }
1487}
1488
1489static int sdma_v3_0_set_clockgating_state(void *handle,
1490 enum amd_clockgating_state state)
1491{
1492 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1493
1494 if (amdgpu_sriov_vf(adev))
1495 return 0;
1496
1497 switch (adev->asic_type) {
1498 case CHIP_FIJI:
1499 case CHIP_CARRIZO:
1500 case CHIP_STONEY:
1501 sdma_v3_0_update_sdma_medium_grain_clock_gating(adev,
1502 state == AMD_CG_STATE_GATE);
1503 sdma_v3_0_update_sdma_medium_grain_light_sleep(adev,
1504 state == AMD_CG_STATE_GATE);
1505 break;
1506 default:
1507 break;
1508 }
1509 return 0;
1510}
1511
1512static int sdma_v3_0_set_powergating_state(void *handle,
1513 enum amd_powergating_state state)
1514{
1515 return 0;
1516}
1517
1518static void sdma_v3_0_get_clockgating_state(void *handle, u64 *flags)
1519{
1520 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1521 int data;
1522
1523 if (amdgpu_sriov_vf(adev))
1524 *flags = 0;
1525
1526 /* AMD_CG_SUPPORT_SDMA_MGCG */
1527 data = RREG32(mmSDMA0_CLK_CTRL + sdma_offsets[0]);
1528 if (!(data & SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK))
1529 *flags |= AMD_CG_SUPPORT_SDMA_MGCG;
1530
1531 /* AMD_CG_SUPPORT_SDMA_LS */
1532 data = RREG32(mmSDMA0_POWER_CNTL + sdma_offsets[0]);
1533 if (data & SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK)
1534 *flags |= AMD_CG_SUPPORT_SDMA_LS;
1535}
1536
1537static const struct amd_ip_funcs sdma_v3_0_ip_funcs = {
1538 .name = "sdma_v3_0",
1539 .early_init = sdma_v3_0_early_init,
1540 .late_init = NULL,
1541 .sw_init = sdma_v3_0_sw_init,
1542 .sw_fini = sdma_v3_0_sw_fini,
1543 .hw_init = sdma_v3_0_hw_init,
1544 .hw_fini = sdma_v3_0_hw_fini,
1545 .suspend = sdma_v3_0_suspend,
1546 .resume = sdma_v3_0_resume,
1547 .is_idle = sdma_v3_0_is_idle,
1548 .wait_for_idle = sdma_v3_0_wait_for_idle,
1549 .check_soft_reset = sdma_v3_0_check_soft_reset,
1550 .pre_soft_reset = sdma_v3_0_pre_soft_reset,
1551 .post_soft_reset = sdma_v3_0_post_soft_reset,
1552 .soft_reset = sdma_v3_0_soft_reset,
1553 .set_clockgating_state = sdma_v3_0_set_clockgating_state,
1554 .set_powergating_state = sdma_v3_0_set_powergating_state,
1555 .get_clockgating_state = sdma_v3_0_get_clockgating_state,
1556};
1557
1558static const struct amdgpu_ring_funcs sdma_v3_0_ring_funcs = {
1559 .type = AMDGPU_RING_TYPE_SDMA,
1560 .align_mask = 0xf,
1561 .nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
1562 .support_64bit_ptrs = false,
1563 .secure_submission_supported = true,
1564 .get_rptr = sdma_v3_0_ring_get_rptr,
1565 .get_wptr = sdma_v3_0_ring_get_wptr,
1566 .set_wptr = sdma_v3_0_ring_set_wptr,
1567 .emit_frame_size =
1568 6 + /* sdma_v3_0_ring_emit_hdp_flush */
1569 3 + /* hdp invalidate */
1570 6 + /* sdma_v3_0_ring_emit_pipeline_sync */
1571 VI_FLUSH_GPU_TLB_NUM_WREG * 3 + 6 + /* sdma_v3_0_ring_emit_vm_flush */
1572 10 + 10 + 10, /* sdma_v3_0_ring_emit_fence x3 for user fence, vm fence */
1573 .emit_ib_size = 7 + 6, /* sdma_v3_0_ring_emit_ib */
1574 .emit_ib = sdma_v3_0_ring_emit_ib,
1575 .emit_fence = sdma_v3_0_ring_emit_fence,
1576 .emit_pipeline_sync = sdma_v3_0_ring_emit_pipeline_sync,
1577 .emit_vm_flush = sdma_v3_0_ring_emit_vm_flush,
1578 .emit_hdp_flush = sdma_v3_0_ring_emit_hdp_flush,
1579 .test_ring = sdma_v3_0_ring_test_ring,
1580 .test_ib = sdma_v3_0_ring_test_ib,
1581 .insert_nop = sdma_v3_0_ring_insert_nop,
1582 .pad_ib = sdma_v3_0_ring_pad_ib,
1583 .emit_wreg = sdma_v3_0_ring_emit_wreg,
1584};
1585
1586static void sdma_v3_0_set_ring_funcs(struct amdgpu_device *adev)
1587{
1588 int i;
1589
1590 for (i = 0; i < adev->sdma.num_instances; i++) {
1591 adev->sdma.instance[i].ring.funcs = &sdma_v3_0_ring_funcs;
1592 adev->sdma.instance[i].ring.me = i;
1593 }
1594}
1595
1596static const struct amdgpu_irq_src_funcs sdma_v3_0_trap_irq_funcs = {
1597 .set = sdma_v3_0_set_trap_irq_state,
1598 .process = sdma_v3_0_process_trap_irq,
1599};
1600
1601static const struct amdgpu_irq_src_funcs sdma_v3_0_illegal_inst_irq_funcs = {
1602 .process = sdma_v3_0_process_illegal_inst_irq,
1603};
1604
1605static void sdma_v3_0_set_irq_funcs(struct amdgpu_device *adev)
1606{
1607 adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
1608 adev->sdma.trap_irq.funcs = &sdma_v3_0_trap_irq_funcs;
1609 adev->sdma.illegal_inst_irq.funcs = &sdma_v3_0_illegal_inst_irq_funcs;
1610}
1611
1612/**
1613 * sdma_v3_0_emit_copy_buffer - copy buffer using the sDMA engine
1614 *
1615 * @ib: indirect buffer to copy to
1616 * @src_offset: src GPU address
1617 * @dst_offset: dst GPU address
1618 * @byte_count: number of bytes to xfer
1619 * @tmz: unused
1620 *
1621 * Copy GPU buffers using the DMA engine (VI).
1622 * Used by the amdgpu ttm implementation to move pages if
1623 * registered as the asic copy callback.
1624 */
1625static void sdma_v3_0_emit_copy_buffer(struct amdgpu_ib *ib,
1626 uint64_t src_offset,
1627 uint64_t dst_offset,
1628 uint32_t byte_count,
1629 bool tmz)
1630{
1631 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1632 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1633 ib->ptr[ib->length_dw++] = byte_count;
1634 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1635 ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1636 ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1637 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1638 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1639}
1640
1641/**
1642 * sdma_v3_0_emit_fill_buffer - fill buffer using the sDMA engine
1643 *
1644 * @ib: indirect buffer to copy to
1645 * @src_data: value to write to buffer
1646 * @dst_offset: dst GPU address
1647 * @byte_count: number of bytes to xfer
1648 *
1649 * Fill GPU buffers using the DMA engine (VI).
1650 */
1651static void sdma_v3_0_emit_fill_buffer(struct amdgpu_ib *ib,
1652 uint32_t src_data,
1653 uint64_t dst_offset,
1654 uint32_t byte_count)
1655{
1656 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL);
1657 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1658 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1659 ib->ptr[ib->length_dw++] = src_data;
1660 ib->ptr[ib->length_dw++] = byte_count;
1661}
1662
1663static const struct amdgpu_buffer_funcs sdma_v3_0_buffer_funcs = {
1664 .copy_max_bytes = 0x3fffe0, /* not 0x3fffff due to HW limitation */
1665 .copy_num_dw = 7,
1666 .emit_copy_buffer = sdma_v3_0_emit_copy_buffer,
1667
1668 .fill_max_bytes = 0x3fffe0, /* not 0x3fffff due to HW limitation */
1669 .fill_num_dw = 5,
1670 .emit_fill_buffer = sdma_v3_0_emit_fill_buffer,
1671};
1672
1673static void sdma_v3_0_set_buffer_funcs(struct amdgpu_device *adev)
1674{
1675 adev->mman.buffer_funcs = &sdma_v3_0_buffer_funcs;
1676 adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1677}
1678
1679static const struct amdgpu_vm_pte_funcs sdma_v3_0_vm_pte_funcs = {
1680 .copy_pte_num_dw = 7,
1681 .copy_pte = sdma_v3_0_vm_copy_pte,
1682
1683 .write_pte = sdma_v3_0_vm_write_pte,
1684 .set_pte_pde = sdma_v3_0_vm_set_pte_pde,
1685};
1686
1687static void sdma_v3_0_set_vm_pte_funcs(struct amdgpu_device *adev)
1688{
1689 unsigned i;
1690
1691 adev->vm_manager.vm_pte_funcs = &sdma_v3_0_vm_pte_funcs;
1692 for (i = 0; i < adev->sdma.num_instances; i++) {
1693 adev->vm_manager.vm_pte_scheds[i] =
1694 &adev->sdma.instance[i].ring.sched;
1695 }
1696 adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
1697}
1698
1699const struct amdgpu_ip_block_version sdma_v3_0_ip_block =
1700{
1701 .type = AMD_IP_BLOCK_TYPE_SDMA,
1702 .major = 3,
1703 .minor = 0,
1704 .rev = 0,
1705 .funcs = &sdma_v3_0_ip_funcs,
1706};
1707
1708const struct amdgpu_ip_block_version sdma_v3_1_ip_block =
1709{
1710 .type = AMD_IP_BLOCK_TYPE_SDMA,
1711 .major = 3,
1712 .minor = 1,
1713 .rev = 0,
1714 .funcs = &sdma_v3_0_ip_funcs,
1715};
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#include <linux/firmware.h>
25#include <drm/drmP.h>
26#include "amdgpu.h"
27#include "amdgpu_ucode.h"
28#include "amdgpu_trace.h"
29#include "vi.h"
30#include "vid.h"
31
32#include "oss/oss_3_0_d.h"
33#include "oss/oss_3_0_sh_mask.h"
34
35#include "gmc/gmc_8_1_d.h"
36#include "gmc/gmc_8_1_sh_mask.h"
37
38#include "gca/gfx_8_0_d.h"
39#include "gca/gfx_8_0_enum.h"
40#include "gca/gfx_8_0_sh_mask.h"
41
42#include "bif/bif_5_0_d.h"
43#include "bif/bif_5_0_sh_mask.h"
44
45#include "tonga_sdma_pkt_open.h"
46
47static void sdma_v3_0_set_ring_funcs(struct amdgpu_device *adev);
48static void sdma_v3_0_set_buffer_funcs(struct amdgpu_device *adev);
49static void sdma_v3_0_set_vm_pte_funcs(struct amdgpu_device *adev);
50static void sdma_v3_0_set_irq_funcs(struct amdgpu_device *adev);
51
52MODULE_FIRMWARE("amdgpu/tonga_sdma.bin");
53MODULE_FIRMWARE("amdgpu/tonga_sdma1.bin");
54MODULE_FIRMWARE("amdgpu/carrizo_sdma.bin");
55MODULE_FIRMWARE("amdgpu/carrizo_sdma1.bin");
56MODULE_FIRMWARE("amdgpu/fiji_sdma.bin");
57MODULE_FIRMWARE("amdgpu/fiji_sdma1.bin");
58MODULE_FIRMWARE("amdgpu/stoney_sdma.bin");
59
60static const u32 sdma_offsets[SDMA_MAX_INSTANCE] =
61{
62 SDMA0_REGISTER_OFFSET,
63 SDMA1_REGISTER_OFFSET
64};
65
66static const u32 golden_settings_tonga_a11[] =
67{
68 mmSDMA0_CHICKEN_BITS, 0xfc910007, 0x00810007,
69 mmSDMA0_CLK_CTRL, 0xff000fff, 0x00000000,
70 mmSDMA0_GFX_IB_CNTL, 0x800f0111, 0x00000100,
71 mmSDMA0_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
72 mmSDMA0_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
73 mmSDMA1_CHICKEN_BITS, 0xfc910007, 0x00810007,
74 mmSDMA1_CLK_CTRL, 0xff000fff, 0x00000000,
75 mmSDMA1_GFX_IB_CNTL, 0x800f0111, 0x00000100,
76 mmSDMA1_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
77 mmSDMA1_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
78};
79
80static const u32 tonga_mgcg_cgcg_init[] =
81{
82 mmSDMA0_CLK_CTRL, 0xff000ff0, 0x00000100,
83 mmSDMA1_CLK_CTRL, 0xff000ff0, 0x00000100
84};
85
86static const u32 golden_settings_fiji_a10[] =
87{
88 mmSDMA0_CHICKEN_BITS, 0xfc910007, 0x00810007,
89 mmSDMA0_GFX_IB_CNTL, 0x800f0111, 0x00000100,
90 mmSDMA0_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
91 mmSDMA0_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
92 mmSDMA1_CHICKEN_BITS, 0xfc910007, 0x00810007,
93 mmSDMA1_GFX_IB_CNTL, 0x800f0111, 0x00000100,
94 mmSDMA1_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
95 mmSDMA1_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
96};
97
98static const u32 fiji_mgcg_cgcg_init[] =
99{
100 mmSDMA0_CLK_CTRL, 0xff000ff0, 0x00000100,
101 mmSDMA1_CLK_CTRL, 0xff000ff0, 0x00000100
102};
103
104static const u32 cz_golden_settings_a11[] =
105{
106 mmSDMA0_CHICKEN_BITS, 0xfc910007, 0x00810007,
107 mmSDMA0_CLK_CTRL, 0xff000fff, 0x00000000,
108 mmSDMA0_GFX_IB_CNTL, 0x00000100, 0x00000100,
109 mmSDMA0_POWER_CNTL, 0x00000800, 0x0003c800,
110 mmSDMA0_RLC0_IB_CNTL, 0x00000100, 0x00000100,
111 mmSDMA0_RLC1_IB_CNTL, 0x00000100, 0x00000100,
112 mmSDMA1_CHICKEN_BITS, 0xfc910007, 0x00810007,
113 mmSDMA1_CLK_CTRL, 0xff000fff, 0x00000000,
114 mmSDMA1_GFX_IB_CNTL, 0x00000100, 0x00000100,
115 mmSDMA1_POWER_CNTL, 0x00000800, 0x0003c800,
116 mmSDMA1_RLC0_IB_CNTL, 0x00000100, 0x00000100,
117 mmSDMA1_RLC1_IB_CNTL, 0x00000100, 0x00000100,
118};
119
120static const u32 cz_mgcg_cgcg_init[] =
121{
122 mmSDMA0_CLK_CTRL, 0xff000ff0, 0x00000100,
123 mmSDMA1_CLK_CTRL, 0xff000ff0, 0x00000100
124};
125
126static const u32 stoney_golden_settings_a11[] =
127{
128 mmSDMA0_GFX_IB_CNTL, 0x00000100, 0x00000100,
129 mmSDMA0_POWER_CNTL, 0x00000800, 0x0003c800,
130 mmSDMA0_RLC0_IB_CNTL, 0x00000100, 0x00000100,
131 mmSDMA0_RLC1_IB_CNTL, 0x00000100, 0x00000100,
132};
133
134static const u32 stoney_mgcg_cgcg_init[] =
135{
136 mmSDMA0_CLK_CTRL, 0xffffffff, 0x00000100,
137};
138
139/*
140 * sDMA - System DMA
141 * Starting with CIK, the GPU has new asynchronous
142 * DMA engines. These engines are used for compute
143 * and gfx. There are two DMA engines (SDMA0, SDMA1)
144 * and each one supports 1 ring buffer used for gfx
145 * and 2 queues used for compute.
146 *
147 * The programming model is very similar to the CP
148 * (ring buffer, IBs, etc.), but sDMA has it's own
149 * packet format that is different from the PM4 format
150 * used by the CP. sDMA supports copying data, writing
151 * embedded data, solid fills, and a number of other
152 * things. It also has support for tiling/detiling of
153 * buffers.
154 */
155
156static void sdma_v3_0_init_golden_registers(struct amdgpu_device *adev)
157{
158 switch (adev->asic_type) {
159 case CHIP_FIJI:
160 amdgpu_program_register_sequence(adev,
161 fiji_mgcg_cgcg_init,
162 (const u32)ARRAY_SIZE(fiji_mgcg_cgcg_init));
163 amdgpu_program_register_sequence(adev,
164 golden_settings_fiji_a10,
165 (const u32)ARRAY_SIZE(golden_settings_fiji_a10));
166 break;
167 case CHIP_TONGA:
168 amdgpu_program_register_sequence(adev,
169 tonga_mgcg_cgcg_init,
170 (const u32)ARRAY_SIZE(tonga_mgcg_cgcg_init));
171 amdgpu_program_register_sequence(adev,
172 golden_settings_tonga_a11,
173 (const u32)ARRAY_SIZE(golden_settings_tonga_a11));
174 break;
175 case CHIP_CARRIZO:
176 amdgpu_program_register_sequence(adev,
177 cz_mgcg_cgcg_init,
178 (const u32)ARRAY_SIZE(cz_mgcg_cgcg_init));
179 amdgpu_program_register_sequence(adev,
180 cz_golden_settings_a11,
181 (const u32)ARRAY_SIZE(cz_golden_settings_a11));
182 break;
183 case CHIP_STONEY:
184 amdgpu_program_register_sequence(adev,
185 stoney_mgcg_cgcg_init,
186 (const u32)ARRAY_SIZE(stoney_mgcg_cgcg_init));
187 amdgpu_program_register_sequence(adev,
188 stoney_golden_settings_a11,
189 (const u32)ARRAY_SIZE(stoney_golden_settings_a11));
190 break;
191 default:
192 break;
193 }
194}
195
196/**
197 * sdma_v3_0_init_microcode - load ucode images from disk
198 *
199 * @adev: amdgpu_device pointer
200 *
201 * Use the firmware interface to load the ucode images into
202 * the driver (not loaded into hw).
203 * Returns 0 on success, error on failure.
204 */
205static int sdma_v3_0_init_microcode(struct amdgpu_device *adev)
206{
207 const char *chip_name;
208 char fw_name[30];
209 int err = 0, i;
210 struct amdgpu_firmware_info *info = NULL;
211 const struct common_firmware_header *header = NULL;
212 const struct sdma_firmware_header_v1_0 *hdr;
213
214 DRM_DEBUG("\n");
215
216 switch (adev->asic_type) {
217 case CHIP_TONGA:
218 chip_name = "tonga";
219 break;
220 case CHIP_FIJI:
221 chip_name = "fiji";
222 break;
223 case CHIP_CARRIZO:
224 chip_name = "carrizo";
225 break;
226 case CHIP_STONEY:
227 chip_name = "stoney";
228 break;
229 default: BUG();
230 }
231
232 for (i = 0; i < adev->sdma.num_instances; i++) {
233 if (i == 0)
234 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma.bin", chip_name);
235 else
236 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma1.bin", chip_name);
237 err = request_firmware(&adev->sdma.instance[i].fw, fw_name, adev->dev);
238 if (err)
239 goto out;
240 err = amdgpu_ucode_validate(adev->sdma.instance[i].fw);
241 if (err)
242 goto out;
243 hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
244 adev->sdma.instance[i].fw_version = le32_to_cpu(hdr->header.ucode_version);
245 adev->sdma.instance[i].feature_version = le32_to_cpu(hdr->ucode_feature_version);
246 if (adev->sdma.instance[i].feature_version >= 20)
247 adev->sdma.instance[i].burst_nop = true;
248
249 if (adev->firmware.smu_load) {
250 info = &adev->firmware.ucode[AMDGPU_UCODE_ID_SDMA0 + i];
251 info->ucode_id = AMDGPU_UCODE_ID_SDMA0 + i;
252 info->fw = adev->sdma.instance[i].fw;
253 header = (const struct common_firmware_header *)info->fw->data;
254 adev->firmware.fw_size +=
255 ALIGN(le32_to_cpu(header->ucode_size_bytes), PAGE_SIZE);
256 }
257 }
258out:
259 if (err) {
260 printk(KERN_ERR
261 "sdma_v3_0: Failed to load firmware \"%s\"\n",
262 fw_name);
263 for (i = 0; i < adev->sdma.num_instances; i++) {
264 release_firmware(adev->sdma.instance[i].fw);
265 adev->sdma.instance[i].fw = NULL;
266 }
267 }
268 return err;
269}
270
271/**
272 * sdma_v3_0_ring_get_rptr - get the current read pointer
273 *
274 * @ring: amdgpu ring pointer
275 *
276 * Get the current rptr from the hardware (VI+).
277 */
278static uint32_t sdma_v3_0_ring_get_rptr(struct amdgpu_ring *ring)
279{
280 u32 rptr;
281
282 /* XXX check if swapping is necessary on BE */
283 rptr = ring->adev->wb.wb[ring->rptr_offs] >> 2;
284
285 return rptr;
286}
287
288/**
289 * sdma_v3_0_ring_get_wptr - get the current write pointer
290 *
291 * @ring: amdgpu ring pointer
292 *
293 * Get the current wptr from the hardware (VI+).
294 */
295static uint32_t sdma_v3_0_ring_get_wptr(struct amdgpu_ring *ring)
296{
297 struct amdgpu_device *adev = ring->adev;
298 u32 wptr;
299
300 if (ring->use_doorbell) {
301 /* XXX check if swapping is necessary on BE */
302 wptr = ring->adev->wb.wb[ring->wptr_offs] >> 2;
303 } else {
304 int me = (ring == &ring->adev->sdma.instance[0].ring) ? 0 : 1;
305
306 wptr = RREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[me]) >> 2;
307 }
308
309 return wptr;
310}
311
312/**
313 * sdma_v3_0_ring_set_wptr - commit the write pointer
314 *
315 * @ring: amdgpu ring pointer
316 *
317 * Write the wptr back to the hardware (VI+).
318 */
319static void sdma_v3_0_ring_set_wptr(struct amdgpu_ring *ring)
320{
321 struct amdgpu_device *adev = ring->adev;
322
323 if (ring->use_doorbell) {
324 /* XXX check if swapping is necessary on BE */
325 adev->wb.wb[ring->wptr_offs] = ring->wptr << 2;
326 WDOORBELL32(ring->doorbell_index, ring->wptr << 2);
327 } else {
328 int me = (ring == &ring->adev->sdma.instance[0].ring) ? 0 : 1;
329
330 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[me], ring->wptr << 2);
331 }
332}
333
334static void sdma_v3_0_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
335{
336 struct amdgpu_sdma_instance *sdma = amdgpu_get_sdma_instance(ring);
337 int i;
338
339 for (i = 0; i < count; i++)
340 if (sdma && sdma->burst_nop && (i == 0))
341 amdgpu_ring_write(ring, ring->nop |
342 SDMA_PKT_NOP_HEADER_COUNT(count - 1));
343 else
344 amdgpu_ring_write(ring, ring->nop);
345}
346
347/**
348 * sdma_v3_0_ring_emit_ib - Schedule an IB on the DMA engine
349 *
350 * @ring: amdgpu ring pointer
351 * @ib: IB object to schedule
352 *
353 * Schedule an IB in the DMA ring (VI).
354 */
355static void sdma_v3_0_ring_emit_ib(struct amdgpu_ring *ring,
356 struct amdgpu_ib *ib)
357{
358 u32 vmid = ib->vm_id & 0xf;
359 u32 next_rptr = ring->wptr + 5;
360
361 while ((next_rptr & 7) != 2)
362 next_rptr++;
363 next_rptr += 6;
364
365 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
366 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
367 amdgpu_ring_write(ring, lower_32_bits(ring->next_rptr_gpu_addr) & 0xfffffffc);
368 amdgpu_ring_write(ring, upper_32_bits(ring->next_rptr_gpu_addr));
369 amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1));
370 amdgpu_ring_write(ring, next_rptr);
371
372 /* IB packet must end on a 8 DW boundary */
373 sdma_v3_0_ring_insert_nop(ring, (10 - (ring->wptr & 7)) % 8);
374
375 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) |
376 SDMA_PKT_INDIRECT_HEADER_VMID(vmid));
377 /* base must be 32 byte aligned */
378 amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
379 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
380 amdgpu_ring_write(ring, ib->length_dw);
381 amdgpu_ring_write(ring, 0);
382 amdgpu_ring_write(ring, 0);
383
384}
385
386/**
387 * sdma_v3_0_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
388 *
389 * @ring: amdgpu ring pointer
390 *
391 * Emit an hdp flush packet on the requested DMA ring.
392 */
393static void sdma_v3_0_ring_emit_hdp_flush(struct amdgpu_ring *ring)
394{
395 u32 ref_and_mask = 0;
396
397 if (ring == &ring->adev->sdma.instance[0].ring)
398 ref_and_mask = REG_SET_FIELD(ref_and_mask, GPU_HDP_FLUSH_DONE, SDMA0, 1);
399 else
400 ref_and_mask = REG_SET_FIELD(ref_and_mask, GPU_HDP_FLUSH_DONE, SDMA1, 1);
401
402 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
403 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
404 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
405 amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_DONE << 2);
406 amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_REQ << 2);
407 amdgpu_ring_write(ring, ref_and_mask); /* reference */
408 amdgpu_ring_write(ring, ref_and_mask); /* mask */
409 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
410 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
411}
412
413static void sdma_v3_0_ring_emit_hdp_invalidate(struct amdgpu_ring *ring)
414{
415 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
416 SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
417 amdgpu_ring_write(ring, mmHDP_DEBUG0);
418 amdgpu_ring_write(ring, 1);
419}
420
421/**
422 * sdma_v3_0_ring_emit_fence - emit a fence on the DMA ring
423 *
424 * @ring: amdgpu ring pointer
425 * @fence: amdgpu fence object
426 *
427 * Add a DMA fence packet to the ring to write
428 * the fence seq number and DMA trap packet to generate
429 * an interrupt if needed (VI).
430 */
431static void sdma_v3_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
432 unsigned flags)
433{
434 bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
435 /* write the fence */
436 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE));
437 amdgpu_ring_write(ring, lower_32_bits(addr));
438 amdgpu_ring_write(ring, upper_32_bits(addr));
439 amdgpu_ring_write(ring, lower_32_bits(seq));
440
441 /* optionally write high bits as well */
442 if (write64bit) {
443 addr += 4;
444 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE));
445 amdgpu_ring_write(ring, lower_32_bits(addr));
446 amdgpu_ring_write(ring, upper_32_bits(addr));
447 amdgpu_ring_write(ring, upper_32_bits(seq));
448 }
449
450 /* generate an interrupt */
451 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP));
452 amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(0));
453}
454
455/**
456 * sdma_v3_0_gfx_stop - stop the gfx async dma engines
457 *
458 * @adev: amdgpu_device pointer
459 *
460 * Stop the gfx async dma ring buffers (VI).
461 */
462static void sdma_v3_0_gfx_stop(struct amdgpu_device *adev)
463{
464 struct amdgpu_ring *sdma0 = &adev->sdma.instance[0].ring;
465 struct amdgpu_ring *sdma1 = &adev->sdma.instance[1].ring;
466 u32 rb_cntl, ib_cntl;
467 int i;
468
469 if ((adev->mman.buffer_funcs_ring == sdma0) ||
470 (adev->mman.buffer_funcs_ring == sdma1))
471 amdgpu_ttm_set_active_vram_size(adev, adev->mc.visible_vram_size);
472
473 for (i = 0; i < adev->sdma.num_instances; i++) {
474 rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
475 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0);
476 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
477 ib_cntl = RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]);
478 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0);
479 WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
480 }
481 sdma0->ready = false;
482 sdma1->ready = false;
483}
484
485/**
486 * sdma_v3_0_rlc_stop - stop the compute async dma engines
487 *
488 * @adev: amdgpu_device pointer
489 *
490 * Stop the compute async dma queues (VI).
491 */
492static void sdma_v3_0_rlc_stop(struct amdgpu_device *adev)
493{
494 /* XXX todo */
495}
496
497/**
498 * sdma_v3_0_ctx_switch_enable - stop the async dma engines context switch
499 *
500 * @adev: amdgpu_device pointer
501 * @enable: enable/disable the DMA MEs context switch.
502 *
503 * Halt or unhalt the async dma engines context switch (VI).
504 */
505static void sdma_v3_0_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
506{
507 u32 f32_cntl;
508 int i;
509
510 for (i = 0; i < adev->sdma.num_instances; i++) {
511 f32_cntl = RREG32(mmSDMA0_CNTL + sdma_offsets[i]);
512 if (enable)
513 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
514 AUTO_CTXSW_ENABLE, 1);
515 else
516 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
517 AUTO_CTXSW_ENABLE, 0);
518 WREG32(mmSDMA0_CNTL + sdma_offsets[i], f32_cntl);
519 }
520}
521
522/**
523 * sdma_v3_0_enable - stop the async dma engines
524 *
525 * @adev: amdgpu_device pointer
526 * @enable: enable/disable the DMA MEs.
527 *
528 * Halt or unhalt the async dma engines (VI).
529 */
530static void sdma_v3_0_enable(struct amdgpu_device *adev, bool enable)
531{
532 u32 f32_cntl;
533 int i;
534
535 if (enable == false) {
536 sdma_v3_0_gfx_stop(adev);
537 sdma_v3_0_rlc_stop(adev);
538 }
539
540 for (i = 0; i < adev->sdma.num_instances; i++) {
541 f32_cntl = RREG32(mmSDMA0_F32_CNTL + sdma_offsets[i]);
542 if (enable)
543 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 0);
544 else
545 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 1);
546 WREG32(mmSDMA0_F32_CNTL + sdma_offsets[i], f32_cntl);
547 }
548}
549
550/**
551 * sdma_v3_0_gfx_resume - setup and start the async dma engines
552 *
553 * @adev: amdgpu_device pointer
554 *
555 * Set up the gfx DMA ring buffers and enable them (VI).
556 * Returns 0 for success, error for failure.
557 */
558static int sdma_v3_0_gfx_resume(struct amdgpu_device *adev)
559{
560 struct amdgpu_ring *ring;
561 u32 rb_cntl, ib_cntl;
562 u32 rb_bufsz;
563 u32 wb_offset;
564 u32 doorbell;
565 int i, j, r;
566
567 for (i = 0; i < adev->sdma.num_instances; i++) {
568 ring = &adev->sdma.instance[i].ring;
569 wb_offset = (ring->rptr_offs * 4);
570
571 mutex_lock(&adev->srbm_mutex);
572 for (j = 0; j < 16; j++) {
573 vi_srbm_select(adev, 0, 0, 0, j);
574 /* SDMA GFX */
575 WREG32(mmSDMA0_GFX_VIRTUAL_ADDR + sdma_offsets[i], 0);
576 WREG32(mmSDMA0_GFX_APE1_CNTL + sdma_offsets[i], 0);
577 }
578 vi_srbm_select(adev, 0, 0, 0, 0);
579 mutex_unlock(&adev->srbm_mutex);
580
581 WREG32(mmSDMA0_TILING_CONFIG + sdma_offsets[i],
582 adev->gfx.config.gb_addr_config & 0x70);
583
584 WREG32(mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
585
586 /* Set ring buffer size in dwords */
587 rb_bufsz = order_base_2(ring->ring_size / 4);
588 rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
589 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz);
590#ifdef __BIG_ENDIAN
591 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1);
592 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
593 RPTR_WRITEBACK_SWAP_ENABLE, 1);
594#endif
595 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
596
597 /* Initialize the ring buffer's read and write pointers */
598 WREG32(mmSDMA0_GFX_RB_RPTR + sdma_offsets[i], 0);
599 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], 0);
600
601 /* set the wb address whether it's enabled or not */
602 WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_HI + sdma_offsets[i],
603 upper_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF);
604 WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_LO + sdma_offsets[i],
605 lower_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC);
606
607 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
608
609 WREG32(mmSDMA0_GFX_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
610 WREG32(mmSDMA0_GFX_RB_BASE_HI + sdma_offsets[i], ring->gpu_addr >> 40);
611
612 ring->wptr = 0;
613 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], ring->wptr << 2);
614
615 doorbell = RREG32(mmSDMA0_GFX_DOORBELL + sdma_offsets[i]);
616
617 if (ring->use_doorbell) {
618 doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL,
619 OFFSET, ring->doorbell_index);
620 doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 1);
621 } else {
622 doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 0);
623 }
624 WREG32(mmSDMA0_GFX_DOORBELL + sdma_offsets[i], doorbell);
625
626 /* enable DMA RB */
627 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1);
628 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
629
630 ib_cntl = RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]);
631 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1);
632#ifdef __BIG_ENDIAN
633 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1);
634#endif
635 /* enable DMA IBs */
636 WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
637
638 ring->ready = true;
639
640 r = amdgpu_ring_test_ring(ring);
641 if (r) {
642 ring->ready = false;
643 return r;
644 }
645
646 if (adev->mman.buffer_funcs_ring == ring)
647 amdgpu_ttm_set_active_vram_size(adev, adev->mc.real_vram_size);
648 }
649
650 return 0;
651}
652
653/**
654 * sdma_v3_0_rlc_resume - setup and start the async dma engines
655 *
656 * @adev: amdgpu_device pointer
657 *
658 * Set up the compute DMA queues and enable them (VI).
659 * Returns 0 for success, error for failure.
660 */
661static int sdma_v3_0_rlc_resume(struct amdgpu_device *adev)
662{
663 /* XXX todo */
664 return 0;
665}
666
667/**
668 * sdma_v3_0_load_microcode - load the sDMA ME ucode
669 *
670 * @adev: amdgpu_device pointer
671 *
672 * Loads the sDMA0/1 ucode.
673 * Returns 0 for success, -EINVAL if the ucode is not available.
674 */
675static int sdma_v3_0_load_microcode(struct amdgpu_device *adev)
676{
677 const struct sdma_firmware_header_v1_0 *hdr;
678 const __le32 *fw_data;
679 u32 fw_size;
680 int i, j;
681
682 /* halt the MEs */
683 sdma_v3_0_enable(adev, false);
684
685 for (i = 0; i < adev->sdma.num_instances; i++) {
686 if (!adev->sdma.instance[i].fw)
687 return -EINVAL;
688 hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
689 amdgpu_ucode_print_sdma_hdr(&hdr->header);
690 fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
691 fw_data = (const __le32 *)
692 (adev->sdma.instance[i].fw->data +
693 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
694 WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], 0);
695 for (j = 0; j < fw_size; j++)
696 WREG32(mmSDMA0_UCODE_DATA + sdma_offsets[i], le32_to_cpup(fw_data++));
697 WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], adev->sdma.instance[i].fw_version);
698 }
699
700 return 0;
701}
702
703/**
704 * sdma_v3_0_start - setup and start the async dma engines
705 *
706 * @adev: amdgpu_device pointer
707 *
708 * Set up the DMA engines and enable them (VI).
709 * Returns 0 for success, error for failure.
710 */
711static int sdma_v3_0_start(struct amdgpu_device *adev)
712{
713 int r, i;
714
715 if (!adev->pp_enabled) {
716 if (!adev->firmware.smu_load) {
717 r = sdma_v3_0_load_microcode(adev);
718 if (r)
719 return r;
720 } else {
721 for (i = 0; i < adev->sdma.num_instances; i++) {
722 r = adev->smu.smumgr_funcs->check_fw_load_finish(adev,
723 (i == 0) ?
724 AMDGPU_UCODE_ID_SDMA0 :
725 AMDGPU_UCODE_ID_SDMA1);
726 if (r)
727 return -EINVAL;
728 }
729 }
730 }
731
732 /* unhalt the MEs */
733 sdma_v3_0_enable(adev, true);
734 /* enable sdma ring preemption */
735 sdma_v3_0_ctx_switch_enable(adev, true);
736
737 /* start the gfx rings and rlc compute queues */
738 r = sdma_v3_0_gfx_resume(adev);
739 if (r)
740 return r;
741 r = sdma_v3_0_rlc_resume(adev);
742 if (r)
743 return r;
744
745 return 0;
746}
747
748/**
749 * sdma_v3_0_ring_test_ring - simple async dma engine test
750 *
751 * @ring: amdgpu_ring structure holding ring information
752 *
753 * Test the DMA engine by writing using it to write an
754 * value to memory. (VI).
755 * Returns 0 for success, error for failure.
756 */
757static int sdma_v3_0_ring_test_ring(struct amdgpu_ring *ring)
758{
759 struct amdgpu_device *adev = ring->adev;
760 unsigned i;
761 unsigned index;
762 int r;
763 u32 tmp;
764 u64 gpu_addr;
765
766 r = amdgpu_wb_get(adev, &index);
767 if (r) {
768 dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
769 return r;
770 }
771
772 gpu_addr = adev->wb.gpu_addr + (index * 4);
773 tmp = 0xCAFEDEAD;
774 adev->wb.wb[index] = cpu_to_le32(tmp);
775
776 r = amdgpu_ring_alloc(ring, 5);
777 if (r) {
778 DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
779 amdgpu_wb_free(adev, index);
780 return r;
781 }
782
783 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
784 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
785 amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
786 amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
787 amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1));
788 amdgpu_ring_write(ring, 0xDEADBEEF);
789 amdgpu_ring_commit(ring);
790
791 for (i = 0; i < adev->usec_timeout; i++) {
792 tmp = le32_to_cpu(adev->wb.wb[index]);
793 if (tmp == 0xDEADBEEF)
794 break;
795 DRM_UDELAY(1);
796 }
797
798 if (i < adev->usec_timeout) {
799 DRM_INFO("ring test on %d succeeded in %d usecs\n", ring->idx, i);
800 } else {
801 DRM_ERROR("amdgpu: ring %d test failed (0x%08X)\n",
802 ring->idx, tmp);
803 r = -EINVAL;
804 }
805 amdgpu_wb_free(adev, index);
806
807 return r;
808}
809
810/**
811 * sdma_v3_0_ring_test_ib - test an IB on the DMA engine
812 *
813 * @ring: amdgpu_ring structure holding ring information
814 *
815 * Test a simple IB in the DMA ring (VI).
816 * Returns 0 on success, error on failure.
817 */
818static int sdma_v3_0_ring_test_ib(struct amdgpu_ring *ring)
819{
820 struct amdgpu_device *adev = ring->adev;
821 struct amdgpu_ib ib;
822 struct fence *f = NULL;
823 unsigned i;
824 unsigned index;
825 int r;
826 u32 tmp = 0;
827 u64 gpu_addr;
828
829 r = amdgpu_wb_get(adev, &index);
830 if (r) {
831 dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
832 return r;
833 }
834
835 gpu_addr = adev->wb.gpu_addr + (index * 4);
836 tmp = 0xCAFEDEAD;
837 adev->wb.wb[index] = cpu_to_le32(tmp);
838 memset(&ib, 0, sizeof(ib));
839 r = amdgpu_ib_get(adev, NULL, 256, &ib);
840 if (r) {
841 DRM_ERROR("amdgpu: failed to get ib (%d).\n", r);
842 goto err0;
843 }
844
845 ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
846 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
847 ib.ptr[1] = lower_32_bits(gpu_addr);
848 ib.ptr[2] = upper_32_bits(gpu_addr);
849 ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1);
850 ib.ptr[4] = 0xDEADBEEF;
851 ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
852 ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
853 ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
854 ib.length_dw = 8;
855
856 r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
857 if (r)
858 goto err1;
859
860 r = fence_wait(f, false);
861 if (r) {
862 DRM_ERROR("amdgpu: fence wait failed (%d).\n", r);
863 goto err1;
864 }
865 for (i = 0; i < adev->usec_timeout; i++) {
866 tmp = le32_to_cpu(adev->wb.wb[index]);
867 if (tmp == 0xDEADBEEF)
868 break;
869 DRM_UDELAY(1);
870 }
871 if (i < adev->usec_timeout) {
872 DRM_INFO("ib test on ring %d succeeded in %u usecs\n",
873 ring->idx, i);
874 goto err1;
875 } else {
876 DRM_ERROR("amdgpu: ib test failed (0x%08X)\n", tmp);
877 r = -EINVAL;
878 }
879err1:
880 fence_put(f);
881 amdgpu_ib_free(adev, &ib, NULL);
882 fence_put(f);
883err0:
884 amdgpu_wb_free(adev, index);
885 return r;
886}
887
888/**
889 * sdma_v3_0_vm_copy_pte - update PTEs by copying them from the GART
890 *
891 * @ib: indirect buffer to fill with commands
892 * @pe: addr of the page entry
893 * @src: src addr to copy from
894 * @count: number of page entries to update
895 *
896 * Update PTEs by copying them from the GART using sDMA (CIK).
897 */
898static void sdma_v3_0_vm_copy_pte(struct amdgpu_ib *ib,
899 uint64_t pe, uint64_t src,
900 unsigned count)
901{
902 while (count) {
903 unsigned bytes = count * 8;
904 if (bytes > 0x1FFFF8)
905 bytes = 0x1FFFF8;
906
907 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
908 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
909 ib->ptr[ib->length_dw++] = bytes;
910 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
911 ib->ptr[ib->length_dw++] = lower_32_bits(src);
912 ib->ptr[ib->length_dw++] = upper_32_bits(src);
913 ib->ptr[ib->length_dw++] = lower_32_bits(pe);
914 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
915
916 pe += bytes;
917 src += bytes;
918 count -= bytes / 8;
919 }
920}
921
922/**
923 * sdma_v3_0_vm_write_pte - update PTEs by writing them manually
924 *
925 * @ib: indirect buffer to fill with commands
926 * @pe: addr of the page entry
927 * @addr: dst addr to write into pe
928 * @count: number of page entries to update
929 * @incr: increase next addr by incr bytes
930 * @flags: access flags
931 *
932 * Update PTEs by writing them manually using sDMA (CIK).
933 */
934static void sdma_v3_0_vm_write_pte(struct amdgpu_ib *ib,
935 const dma_addr_t *pages_addr, uint64_t pe,
936 uint64_t addr, unsigned count,
937 uint32_t incr, uint32_t flags)
938{
939 uint64_t value;
940 unsigned ndw;
941
942 while (count) {
943 ndw = count * 2;
944 if (ndw > 0xFFFFE)
945 ndw = 0xFFFFE;
946
947 /* for non-physically contiguous pages (system) */
948 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
949 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
950 ib->ptr[ib->length_dw++] = pe;
951 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
952 ib->ptr[ib->length_dw++] = ndw;
953 for (; ndw > 0; ndw -= 2, --count, pe += 8) {
954 value = amdgpu_vm_map_gart(pages_addr, addr);
955 addr += incr;
956 value |= flags;
957 ib->ptr[ib->length_dw++] = value;
958 ib->ptr[ib->length_dw++] = upper_32_bits(value);
959 }
960 }
961}
962
963/**
964 * sdma_v3_0_vm_set_pte_pde - update the page tables using sDMA
965 *
966 * @ib: indirect buffer to fill with commands
967 * @pe: addr of the page entry
968 * @addr: dst addr to write into pe
969 * @count: number of page entries to update
970 * @incr: increase next addr by incr bytes
971 * @flags: access flags
972 *
973 * Update the page tables using sDMA (CIK).
974 */
975static void sdma_v3_0_vm_set_pte_pde(struct amdgpu_ib *ib,
976 uint64_t pe,
977 uint64_t addr, unsigned count,
978 uint32_t incr, uint32_t flags)
979{
980 uint64_t value;
981 unsigned ndw;
982
983 while (count) {
984 ndw = count;
985 if (ndw > 0x7FFFF)
986 ndw = 0x7FFFF;
987
988 if (flags & AMDGPU_PTE_VALID)
989 value = addr;
990 else
991 value = 0;
992
993 /* for physically contiguous pages (vram) */
994 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_GEN_PTEPDE);
995 ib->ptr[ib->length_dw++] = pe; /* dst addr */
996 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
997 ib->ptr[ib->length_dw++] = flags; /* mask */
998 ib->ptr[ib->length_dw++] = 0;
999 ib->ptr[ib->length_dw++] = value; /* value */
1000 ib->ptr[ib->length_dw++] = upper_32_bits(value);
1001 ib->ptr[ib->length_dw++] = incr; /* increment size */
1002 ib->ptr[ib->length_dw++] = 0;
1003 ib->ptr[ib->length_dw++] = ndw; /* number of entries */
1004
1005 pe += ndw * 8;
1006 addr += ndw * incr;
1007 count -= ndw;
1008 }
1009}
1010
1011/**
1012 * sdma_v3_0_ring_pad_ib - pad the IB to the required number of dw
1013 *
1014 * @ib: indirect buffer to fill with padding
1015 *
1016 */
1017static void sdma_v3_0_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
1018{
1019 struct amdgpu_sdma_instance *sdma = amdgpu_get_sdma_instance(ring);
1020 u32 pad_count;
1021 int i;
1022
1023 pad_count = (8 - (ib->length_dw & 0x7)) % 8;
1024 for (i = 0; i < pad_count; i++)
1025 if (sdma && sdma->burst_nop && (i == 0))
1026 ib->ptr[ib->length_dw++] =
1027 SDMA_PKT_HEADER_OP(SDMA_OP_NOP) |
1028 SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
1029 else
1030 ib->ptr[ib->length_dw++] =
1031 SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
1032}
1033
1034/**
1035 * sdma_v3_0_ring_emit_pipeline_sync - sync the pipeline
1036 *
1037 * @ring: amdgpu_ring pointer
1038 *
1039 * Make sure all previous operations are completed (CIK).
1040 */
1041static void sdma_v3_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
1042{
1043 uint32_t seq = ring->fence_drv.sync_seq;
1044 uint64_t addr = ring->fence_drv.gpu_addr;
1045
1046 /* wait for idle */
1047 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1048 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1049 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
1050 SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
1051 amdgpu_ring_write(ring, addr & 0xfffffffc);
1052 amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
1053 amdgpu_ring_write(ring, seq); /* reference */
1054 amdgpu_ring_write(ring, 0xfffffff); /* mask */
1055 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1056 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
1057}
1058
1059/**
1060 * sdma_v3_0_ring_emit_vm_flush - cik vm flush using sDMA
1061 *
1062 * @ring: amdgpu_ring pointer
1063 * @vm: amdgpu_vm pointer
1064 *
1065 * Update the page table base and flush the VM TLB
1066 * using sDMA (VI).
1067 */
1068static void sdma_v3_0_ring_emit_vm_flush(struct amdgpu_ring *ring,
1069 unsigned vm_id, uint64_t pd_addr)
1070{
1071 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
1072 SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
1073 if (vm_id < 8) {
1074 amdgpu_ring_write(ring, (mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + vm_id));
1075 } else {
1076 amdgpu_ring_write(ring, (mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vm_id - 8));
1077 }
1078 amdgpu_ring_write(ring, pd_addr >> 12);
1079
1080 /* flush TLB */
1081 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
1082 SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
1083 amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST);
1084 amdgpu_ring_write(ring, 1 << vm_id);
1085
1086 /* wait for flush */
1087 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1088 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1089 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(0)); /* always */
1090 amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST << 2);
1091 amdgpu_ring_write(ring, 0);
1092 amdgpu_ring_write(ring, 0); /* reference */
1093 amdgpu_ring_write(ring, 0); /* mask */
1094 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1095 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
1096}
1097
1098static int sdma_v3_0_early_init(void *handle)
1099{
1100 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1101
1102 switch (adev->asic_type) {
1103 case CHIP_STONEY:
1104 adev->sdma.num_instances = 1;
1105 break;
1106 default:
1107 adev->sdma.num_instances = SDMA_MAX_INSTANCE;
1108 break;
1109 }
1110
1111 sdma_v3_0_set_ring_funcs(adev);
1112 sdma_v3_0_set_buffer_funcs(adev);
1113 sdma_v3_0_set_vm_pte_funcs(adev);
1114 sdma_v3_0_set_irq_funcs(adev);
1115
1116 return 0;
1117}
1118
1119static int sdma_v3_0_sw_init(void *handle)
1120{
1121 struct amdgpu_ring *ring;
1122 int r, i;
1123 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1124
1125 /* SDMA trap event */
1126 r = amdgpu_irq_add_id(adev, 224, &adev->sdma.trap_irq);
1127 if (r)
1128 return r;
1129
1130 /* SDMA Privileged inst */
1131 r = amdgpu_irq_add_id(adev, 241, &adev->sdma.illegal_inst_irq);
1132 if (r)
1133 return r;
1134
1135 /* SDMA Privileged inst */
1136 r = amdgpu_irq_add_id(adev, 247, &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 ring->use_doorbell = true;
1150 ring->doorbell_index = (i == 0) ?
1151 AMDGPU_DOORBELL_sDMA_ENGINE0 : AMDGPU_DOORBELL_sDMA_ENGINE1;
1152
1153 sprintf(ring->name, "sdma%d", i);
1154 r = amdgpu_ring_init(adev, ring, 256 * 1024,
1155 SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP), 0xf,
1156 &adev->sdma.trap_irq,
1157 (i == 0) ?
1158 AMDGPU_SDMA_IRQ_TRAP0 : AMDGPU_SDMA_IRQ_TRAP1,
1159 AMDGPU_RING_TYPE_SDMA);
1160 if (r)
1161 return r;
1162 }
1163
1164 return r;
1165}
1166
1167static int sdma_v3_0_sw_fini(void *handle)
1168{
1169 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1170 int i;
1171
1172 for (i = 0; i < adev->sdma.num_instances; i++)
1173 amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1174
1175 return 0;
1176}
1177
1178static int sdma_v3_0_hw_init(void *handle)
1179{
1180 int r;
1181 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1182
1183 sdma_v3_0_init_golden_registers(adev);
1184
1185 r = sdma_v3_0_start(adev);
1186 if (r)
1187 return r;
1188
1189 return r;
1190}
1191
1192static int sdma_v3_0_hw_fini(void *handle)
1193{
1194 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1195
1196 sdma_v3_0_ctx_switch_enable(adev, false);
1197 sdma_v3_0_enable(adev, false);
1198
1199 return 0;
1200}
1201
1202static int sdma_v3_0_suspend(void *handle)
1203{
1204 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1205
1206 return sdma_v3_0_hw_fini(adev);
1207}
1208
1209static int sdma_v3_0_resume(void *handle)
1210{
1211 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1212
1213 return sdma_v3_0_hw_init(adev);
1214}
1215
1216static bool sdma_v3_0_is_idle(void *handle)
1217{
1218 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1219 u32 tmp = RREG32(mmSRBM_STATUS2);
1220
1221 if (tmp & (SRBM_STATUS2__SDMA_BUSY_MASK |
1222 SRBM_STATUS2__SDMA1_BUSY_MASK))
1223 return false;
1224
1225 return true;
1226}
1227
1228static int sdma_v3_0_wait_for_idle(void *handle)
1229{
1230 unsigned i;
1231 u32 tmp;
1232 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1233
1234 for (i = 0; i < adev->usec_timeout; i++) {
1235 tmp = RREG32(mmSRBM_STATUS2) & (SRBM_STATUS2__SDMA_BUSY_MASK |
1236 SRBM_STATUS2__SDMA1_BUSY_MASK);
1237
1238 if (!tmp)
1239 return 0;
1240 udelay(1);
1241 }
1242 return -ETIMEDOUT;
1243}
1244
1245static void sdma_v3_0_print_status(void *handle)
1246{
1247 int i, j;
1248 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1249
1250 dev_info(adev->dev, "VI SDMA registers\n");
1251 dev_info(adev->dev, " SRBM_STATUS2=0x%08X\n",
1252 RREG32(mmSRBM_STATUS2));
1253 for (i = 0; i < adev->sdma.num_instances; i++) {
1254 dev_info(adev->dev, " SDMA%d_STATUS_REG=0x%08X\n",
1255 i, RREG32(mmSDMA0_STATUS_REG + sdma_offsets[i]));
1256 dev_info(adev->dev, " SDMA%d_F32_CNTL=0x%08X\n",
1257 i, RREG32(mmSDMA0_F32_CNTL + sdma_offsets[i]));
1258 dev_info(adev->dev, " SDMA%d_CNTL=0x%08X\n",
1259 i, RREG32(mmSDMA0_CNTL + sdma_offsets[i]));
1260 dev_info(adev->dev, " SDMA%d_SEM_WAIT_FAIL_TIMER_CNTL=0x%08X\n",
1261 i, RREG32(mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i]));
1262 dev_info(adev->dev, " SDMA%d_GFX_IB_CNTL=0x%08X\n",
1263 i, RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]));
1264 dev_info(adev->dev, " SDMA%d_GFX_RB_CNTL=0x%08X\n",
1265 i, RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]));
1266 dev_info(adev->dev, " SDMA%d_GFX_RB_RPTR=0x%08X\n",
1267 i, RREG32(mmSDMA0_GFX_RB_RPTR + sdma_offsets[i]));
1268 dev_info(adev->dev, " SDMA%d_GFX_RB_WPTR=0x%08X\n",
1269 i, RREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i]));
1270 dev_info(adev->dev, " SDMA%d_GFX_RB_RPTR_ADDR_HI=0x%08X\n",
1271 i, RREG32(mmSDMA0_GFX_RB_RPTR_ADDR_HI + sdma_offsets[i]));
1272 dev_info(adev->dev, " SDMA%d_GFX_RB_RPTR_ADDR_LO=0x%08X\n",
1273 i, RREG32(mmSDMA0_GFX_RB_RPTR_ADDR_LO + sdma_offsets[i]));
1274 dev_info(adev->dev, " SDMA%d_GFX_RB_BASE=0x%08X\n",
1275 i, RREG32(mmSDMA0_GFX_RB_BASE + sdma_offsets[i]));
1276 dev_info(adev->dev, " SDMA%d_GFX_RB_BASE_HI=0x%08X\n",
1277 i, RREG32(mmSDMA0_GFX_RB_BASE_HI + sdma_offsets[i]));
1278 dev_info(adev->dev, " SDMA%d_GFX_DOORBELL=0x%08X\n",
1279 i, RREG32(mmSDMA0_GFX_DOORBELL + sdma_offsets[i]));
1280 dev_info(adev->dev, " SDMA%d_TILING_CONFIG=0x%08X\n",
1281 i, RREG32(mmSDMA0_TILING_CONFIG + sdma_offsets[i]));
1282 mutex_lock(&adev->srbm_mutex);
1283 for (j = 0; j < 16; j++) {
1284 vi_srbm_select(adev, 0, 0, 0, j);
1285 dev_info(adev->dev, " VM %d:\n", j);
1286 dev_info(adev->dev, " SDMA%d_GFX_VIRTUAL_ADDR=0x%08X\n",
1287 i, RREG32(mmSDMA0_GFX_VIRTUAL_ADDR + sdma_offsets[i]));
1288 dev_info(adev->dev, " SDMA%d_GFX_APE1_CNTL=0x%08X\n",
1289 i, RREG32(mmSDMA0_GFX_APE1_CNTL + sdma_offsets[i]));
1290 }
1291 vi_srbm_select(adev, 0, 0, 0, 0);
1292 mutex_unlock(&adev->srbm_mutex);
1293 }
1294}
1295
1296static int sdma_v3_0_soft_reset(void *handle)
1297{
1298 u32 srbm_soft_reset = 0;
1299 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1300 u32 tmp = RREG32(mmSRBM_STATUS2);
1301
1302 if (tmp & SRBM_STATUS2__SDMA_BUSY_MASK) {
1303 /* sdma0 */
1304 tmp = RREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET);
1305 tmp = REG_SET_FIELD(tmp, SDMA0_F32_CNTL, HALT, 0);
1306 WREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET, tmp);
1307 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA_MASK;
1308 }
1309 if (tmp & SRBM_STATUS2__SDMA1_BUSY_MASK) {
1310 /* sdma1 */
1311 tmp = RREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET);
1312 tmp = REG_SET_FIELD(tmp, SDMA0_F32_CNTL, HALT, 0);
1313 WREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET, tmp);
1314 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA1_MASK;
1315 }
1316
1317 if (srbm_soft_reset) {
1318 sdma_v3_0_print_status((void *)adev);
1319
1320 tmp = RREG32(mmSRBM_SOFT_RESET);
1321 tmp |= srbm_soft_reset;
1322 dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
1323 WREG32(mmSRBM_SOFT_RESET, tmp);
1324 tmp = RREG32(mmSRBM_SOFT_RESET);
1325
1326 udelay(50);
1327
1328 tmp &= ~srbm_soft_reset;
1329 WREG32(mmSRBM_SOFT_RESET, tmp);
1330 tmp = RREG32(mmSRBM_SOFT_RESET);
1331
1332 /* Wait a little for things to settle down */
1333 udelay(50);
1334
1335 sdma_v3_0_print_status((void *)adev);
1336 }
1337
1338 return 0;
1339}
1340
1341static int sdma_v3_0_set_trap_irq_state(struct amdgpu_device *adev,
1342 struct amdgpu_irq_src *source,
1343 unsigned type,
1344 enum amdgpu_interrupt_state state)
1345{
1346 u32 sdma_cntl;
1347
1348 switch (type) {
1349 case AMDGPU_SDMA_IRQ_TRAP0:
1350 switch (state) {
1351 case AMDGPU_IRQ_STATE_DISABLE:
1352 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1353 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 0);
1354 WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1355 break;
1356 case AMDGPU_IRQ_STATE_ENABLE:
1357 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1358 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 1);
1359 WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1360 break;
1361 default:
1362 break;
1363 }
1364 break;
1365 case AMDGPU_SDMA_IRQ_TRAP1:
1366 switch (state) {
1367 case AMDGPU_IRQ_STATE_DISABLE:
1368 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1369 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 0);
1370 WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1371 break;
1372 case AMDGPU_IRQ_STATE_ENABLE:
1373 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1374 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 1);
1375 WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1376 break;
1377 default:
1378 break;
1379 }
1380 break;
1381 default:
1382 break;
1383 }
1384 return 0;
1385}
1386
1387static int sdma_v3_0_process_trap_irq(struct amdgpu_device *adev,
1388 struct amdgpu_irq_src *source,
1389 struct amdgpu_iv_entry *entry)
1390{
1391 u8 instance_id, queue_id;
1392
1393 instance_id = (entry->ring_id & 0x3) >> 0;
1394 queue_id = (entry->ring_id & 0xc) >> 2;
1395 DRM_DEBUG("IH: SDMA trap\n");
1396 switch (instance_id) {
1397 case 0:
1398 switch (queue_id) {
1399 case 0:
1400 amdgpu_fence_process(&adev->sdma.instance[0].ring);
1401 break;
1402 case 1:
1403 /* XXX compute */
1404 break;
1405 case 2:
1406 /* XXX compute */
1407 break;
1408 }
1409 break;
1410 case 1:
1411 switch (queue_id) {
1412 case 0:
1413 amdgpu_fence_process(&adev->sdma.instance[1].ring);
1414 break;
1415 case 1:
1416 /* XXX compute */
1417 break;
1418 case 2:
1419 /* XXX compute */
1420 break;
1421 }
1422 break;
1423 }
1424 return 0;
1425}
1426
1427static int sdma_v3_0_process_illegal_inst_irq(struct amdgpu_device *adev,
1428 struct amdgpu_irq_src *source,
1429 struct amdgpu_iv_entry *entry)
1430{
1431 DRM_ERROR("Illegal instruction in SDMA command stream\n");
1432 schedule_work(&adev->reset_work);
1433 return 0;
1434}
1435
1436static void fiji_update_sdma_medium_grain_clock_gating(
1437 struct amdgpu_device *adev,
1438 bool enable)
1439{
1440 uint32_t temp, data;
1441
1442 if (enable) {
1443 temp = data = RREG32(mmSDMA0_CLK_CTRL);
1444 data &= ~(SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK |
1445 SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK |
1446 SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK |
1447 SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1448 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1449 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1450 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1451 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK);
1452 if (data != temp)
1453 WREG32(mmSDMA0_CLK_CTRL, data);
1454
1455 temp = data = RREG32(mmSDMA1_CLK_CTRL);
1456 data &= ~(SDMA1_CLK_CTRL__SOFT_OVERRIDE7_MASK |
1457 SDMA1_CLK_CTRL__SOFT_OVERRIDE6_MASK |
1458 SDMA1_CLK_CTRL__SOFT_OVERRIDE5_MASK |
1459 SDMA1_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1460 SDMA1_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1461 SDMA1_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1462 SDMA1_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1463 SDMA1_CLK_CTRL__SOFT_OVERRIDE0_MASK);
1464
1465 if (data != temp)
1466 WREG32(mmSDMA1_CLK_CTRL, data);
1467 } else {
1468 temp = data = RREG32(mmSDMA0_CLK_CTRL);
1469 data |= SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK |
1470 SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK |
1471 SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK |
1472 SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1473 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1474 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1475 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1476 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK;
1477
1478 if (data != temp)
1479 WREG32(mmSDMA0_CLK_CTRL, data);
1480
1481 temp = data = RREG32(mmSDMA1_CLK_CTRL);
1482 data |= SDMA1_CLK_CTRL__SOFT_OVERRIDE7_MASK |
1483 SDMA1_CLK_CTRL__SOFT_OVERRIDE6_MASK |
1484 SDMA1_CLK_CTRL__SOFT_OVERRIDE5_MASK |
1485 SDMA1_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1486 SDMA1_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1487 SDMA1_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1488 SDMA1_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1489 SDMA1_CLK_CTRL__SOFT_OVERRIDE0_MASK;
1490
1491 if (data != temp)
1492 WREG32(mmSDMA1_CLK_CTRL, data);
1493 }
1494}
1495
1496static void fiji_update_sdma_medium_grain_light_sleep(
1497 struct amdgpu_device *adev,
1498 bool enable)
1499{
1500 uint32_t temp, data;
1501
1502 if (enable) {
1503 temp = data = RREG32(mmSDMA0_POWER_CNTL);
1504 data |= SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1505
1506 if (temp != data)
1507 WREG32(mmSDMA0_POWER_CNTL, data);
1508
1509 temp = data = RREG32(mmSDMA1_POWER_CNTL);
1510 data |= SDMA1_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1511
1512 if (temp != data)
1513 WREG32(mmSDMA1_POWER_CNTL, data);
1514 } else {
1515 temp = data = RREG32(mmSDMA0_POWER_CNTL);
1516 data &= ~SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1517
1518 if (temp != data)
1519 WREG32(mmSDMA0_POWER_CNTL, data);
1520
1521 temp = data = RREG32(mmSDMA1_POWER_CNTL);
1522 data &= ~SDMA1_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1523
1524 if (temp != data)
1525 WREG32(mmSDMA1_POWER_CNTL, data);
1526 }
1527}
1528
1529static int sdma_v3_0_set_clockgating_state(void *handle,
1530 enum amd_clockgating_state state)
1531{
1532 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1533
1534 switch (adev->asic_type) {
1535 case CHIP_FIJI:
1536 fiji_update_sdma_medium_grain_clock_gating(adev,
1537 state == AMD_CG_STATE_GATE ? true : false);
1538 fiji_update_sdma_medium_grain_light_sleep(adev,
1539 state == AMD_CG_STATE_GATE ? true : false);
1540 break;
1541 default:
1542 break;
1543 }
1544 return 0;
1545}
1546
1547static int sdma_v3_0_set_powergating_state(void *handle,
1548 enum amd_powergating_state state)
1549{
1550 return 0;
1551}
1552
1553const struct amd_ip_funcs sdma_v3_0_ip_funcs = {
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 .soft_reset = sdma_v3_0_soft_reset,
1565 .print_status = sdma_v3_0_print_status,
1566 .set_clockgating_state = sdma_v3_0_set_clockgating_state,
1567 .set_powergating_state = sdma_v3_0_set_powergating_state,
1568};
1569
1570static const struct amdgpu_ring_funcs sdma_v3_0_ring_funcs = {
1571 .get_rptr = sdma_v3_0_ring_get_rptr,
1572 .get_wptr = sdma_v3_0_ring_get_wptr,
1573 .set_wptr = sdma_v3_0_ring_set_wptr,
1574 .parse_cs = NULL,
1575 .emit_ib = sdma_v3_0_ring_emit_ib,
1576 .emit_fence = sdma_v3_0_ring_emit_fence,
1577 .emit_pipeline_sync = sdma_v3_0_ring_emit_pipeline_sync,
1578 .emit_vm_flush = sdma_v3_0_ring_emit_vm_flush,
1579 .emit_hdp_flush = sdma_v3_0_ring_emit_hdp_flush,
1580 .emit_hdp_invalidate = sdma_v3_0_ring_emit_hdp_invalidate,
1581 .test_ring = sdma_v3_0_ring_test_ring,
1582 .test_ib = sdma_v3_0_ring_test_ib,
1583 .insert_nop = sdma_v3_0_ring_insert_nop,
1584 .pad_ib = sdma_v3_0_ring_pad_ib,
1585};
1586
1587static void sdma_v3_0_set_ring_funcs(struct amdgpu_device *adev)
1588{
1589 int i;
1590
1591 for (i = 0; i < adev->sdma.num_instances; i++)
1592 adev->sdma.instance[i].ring.funcs = &sdma_v3_0_ring_funcs;
1593}
1594
1595static const struct amdgpu_irq_src_funcs sdma_v3_0_trap_irq_funcs = {
1596 .set = sdma_v3_0_set_trap_irq_state,
1597 .process = sdma_v3_0_process_trap_irq,
1598};
1599
1600static const struct amdgpu_irq_src_funcs sdma_v3_0_illegal_inst_irq_funcs = {
1601 .process = sdma_v3_0_process_illegal_inst_irq,
1602};
1603
1604static void sdma_v3_0_set_irq_funcs(struct amdgpu_device *adev)
1605{
1606 adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
1607 adev->sdma.trap_irq.funcs = &sdma_v3_0_trap_irq_funcs;
1608 adev->sdma.illegal_inst_irq.funcs = &sdma_v3_0_illegal_inst_irq_funcs;
1609}
1610
1611/**
1612 * sdma_v3_0_emit_copy_buffer - copy buffer using the sDMA engine
1613 *
1614 * @ring: amdgpu_ring structure holding ring information
1615 * @src_offset: src GPU address
1616 * @dst_offset: dst GPU address
1617 * @byte_count: number of bytes to xfer
1618 *
1619 * Copy GPU buffers using the DMA engine (VI).
1620 * Used by the amdgpu ttm implementation to move pages if
1621 * registered as the asic copy callback.
1622 */
1623static void sdma_v3_0_emit_copy_buffer(struct amdgpu_ib *ib,
1624 uint64_t src_offset,
1625 uint64_t dst_offset,
1626 uint32_t byte_count)
1627{
1628 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1629 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1630 ib->ptr[ib->length_dw++] = byte_count;
1631 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1632 ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1633 ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1634 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1635 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1636}
1637
1638/**
1639 * sdma_v3_0_emit_fill_buffer - fill buffer using the sDMA engine
1640 *
1641 * @ring: amdgpu_ring structure holding ring information
1642 * @src_data: value to write to buffer
1643 * @dst_offset: dst GPU address
1644 * @byte_count: number of bytes to xfer
1645 *
1646 * Fill GPU buffers using the DMA engine (VI).
1647 */
1648static void sdma_v3_0_emit_fill_buffer(struct amdgpu_ib *ib,
1649 uint32_t src_data,
1650 uint64_t dst_offset,
1651 uint32_t byte_count)
1652{
1653 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL);
1654 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1655 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1656 ib->ptr[ib->length_dw++] = src_data;
1657 ib->ptr[ib->length_dw++] = byte_count;
1658}
1659
1660static const struct amdgpu_buffer_funcs sdma_v3_0_buffer_funcs = {
1661 .copy_max_bytes = 0x1fffff,
1662 .copy_num_dw = 7,
1663 .emit_copy_buffer = sdma_v3_0_emit_copy_buffer,
1664
1665 .fill_max_bytes = 0x1fffff,
1666 .fill_num_dw = 5,
1667 .emit_fill_buffer = sdma_v3_0_emit_fill_buffer,
1668};
1669
1670static void sdma_v3_0_set_buffer_funcs(struct amdgpu_device *adev)
1671{
1672 if (adev->mman.buffer_funcs == NULL) {
1673 adev->mman.buffer_funcs = &sdma_v3_0_buffer_funcs;
1674 adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1675 }
1676}
1677
1678static const struct amdgpu_vm_pte_funcs sdma_v3_0_vm_pte_funcs = {
1679 .copy_pte = sdma_v3_0_vm_copy_pte,
1680 .write_pte = sdma_v3_0_vm_write_pte,
1681 .set_pte_pde = sdma_v3_0_vm_set_pte_pde,
1682};
1683
1684static void sdma_v3_0_set_vm_pte_funcs(struct amdgpu_device *adev)
1685{
1686 unsigned i;
1687
1688 if (adev->vm_manager.vm_pte_funcs == NULL) {
1689 adev->vm_manager.vm_pte_funcs = &sdma_v3_0_vm_pte_funcs;
1690 for (i = 0; i < adev->sdma.num_instances; i++)
1691 adev->vm_manager.vm_pte_rings[i] =
1692 &adev->sdma.instance[i].ring;
1693
1694 adev->vm_manager.vm_pte_num_rings = adev->sdma.num_instances;
1695 }
1696}