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_2_4_d.h"
36#include "oss/oss_2_4_sh_mask.h"
37
38#include "gmc/gmc_7_1_d.h"
39#include "gmc/gmc_7_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 "iceland_sdma_pkt_open.h"
49
50#include "ivsrcid/ivsrcid_vislands30.h"
51
52static void sdma_v2_4_set_ring_funcs(struct amdgpu_device *adev);
53static void sdma_v2_4_set_buffer_funcs(struct amdgpu_device *adev);
54static void sdma_v2_4_set_vm_pte_funcs(struct amdgpu_device *adev);
55static void sdma_v2_4_set_irq_funcs(struct amdgpu_device *adev);
56
57MODULE_FIRMWARE("amdgpu/topaz_sdma.bin");
58MODULE_FIRMWARE("amdgpu/topaz_sdma1.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_iceland_a11[] =
67{
68 mmSDMA0_CHICKEN_BITS, 0xfc910007, 0x00810007,
69 mmSDMA0_CLK_CTRL, 0xff000fff, 0x00000000,
70 mmSDMA1_CHICKEN_BITS, 0xfc910007, 0x00810007,
71 mmSDMA1_CLK_CTRL, 0xff000fff, 0x00000000,
72};
73
74static const u32 iceland_mgcg_cgcg_init[] =
75{
76 mmSDMA0_CLK_CTRL, 0xff000ff0, 0x00000100,
77 mmSDMA1_CLK_CTRL, 0xff000ff0, 0x00000100
78};
79
80/*
81 * sDMA - System DMA
82 * Starting with CIK, the GPU has new asynchronous
83 * DMA engines. These engines are used for compute
84 * and gfx. There are two DMA engines (SDMA0, SDMA1)
85 * and each one supports 1 ring buffer used for gfx
86 * and 2 queues used for compute.
87 *
88 * The programming model is very similar to the CP
89 * (ring buffer, IBs, etc.), but sDMA has it's own
90 * packet format that is different from the PM4 format
91 * used by the CP. sDMA supports copying data, writing
92 * embedded data, solid fills, and a number of other
93 * things. It also has support for tiling/detiling of
94 * buffers.
95 */
96
97static void sdma_v2_4_init_golden_registers(struct amdgpu_device *adev)
98{
99 switch (adev->asic_type) {
100 case CHIP_TOPAZ:
101 amdgpu_device_program_register_sequence(adev,
102 iceland_mgcg_cgcg_init,
103 ARRAY_SIZE(iceland_mgcg_cgcg_init));
104 amdgpu_device_program_register_sequence(adev,
105 golden_settings_iceland_a11,
106 ARRAY_SIZE(golden_settings_iceland_a11));
107 break;
108 default:
109 break;
110 }
111}
112
113static void sdma_v2_4_free_microcode(struct amdgpu_device *adev)
114{
115 int i;
116 for (i = 0; i < adev->sdma.num_instances; i++) {
117 release_firmware(adev->sdma.instance[i].fw);
118 adev->sdma.instance[i].fw = NULL;
119 }
120}
121
122/**
123 * sdma_v2_4_init_microcode - load ucode images from disk
124 *
125 * @adev: amdgpu_device pointer
126 *
127 * Use the firmware interface to load the ucode images into
128 * the driver (not loaded into hw).
129 * Returns 0 on success, error on failure.
130 */
131static int sdma_v2_4_init_microcode(struct amdgpu_device *adev)
132{
133 const char *chip_name;
134 char fw_name[30];
135 int err = 0, i;
136 struct amdgpu_firmware_info *info = NULL;
137 const struct common_firmware_header *header = NULL;
138 const struct sdma_firmware_header_v1_0 *hdr;
139
140 DRM_DEBUG("\n");
141
142 switch (adev->asic_type) {
143 case CHIP_TOPAZ:
144 chip_name = "topaz";
145 break;
146 default: BUG();
147 }
148
149 for (i = 0; i < adev->sdma.num_instances; i++) {
150 if (i == 0)
151 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma.bin", chip_name);
152 else
153 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma1.bin", chip_name);
154 err = request_firmware(&adev->sdma.instance[i].fw, fw_name, adev->dev);
155 if (err)
156 goto out;
157 err = amdgpu_ucode_validate(adev->sdma.instance[i].fw);
158 if (err)
159 goto out;
160 hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
161 adev->sdma.instance[i].fw_version = le32_to_cpu(hdr->header.ucode_version);
162 adev->sdma.instance[i].feature_version = le32_to_cpu(hdr->ucode_feature_version);
163 if (adev->sdma.instance[i].feature_version >= 20)
164 adev->sdma.instance[i].burst_nop = true;
165
166 if (adev->firmware.load_type == AMDGPU_FW_LOAD_SMU) {
167 info = &adev->firmware.ucode[AMDGPU_UCODE_ID_SDMA0 + i];
168 info->ucode_id = AMDGPU_UCODE_ID_SDMA0 + i;
169 info->fw = adev->sdma.instance[i].fw;
170 header = (const struct common_firmware_header *)info->fw->data;
171 adev->firmware.fw_size +=
172 ALIGN(le32_to_cpu(header->ucode_size_bytes), PAGE_SIZE);
173 }
174 }
175
176out:
177 if (err) {
178 pr_err("sdma_v2_4: Failed to load firmware \"%s\"\n", fw_name);
179 for (i = 0; i < adev->sdma.num_instances; i++) {
180 release_firmware(adev->sdma.instance[i].fw);
181 adev->sdma.instance[i].fw = NULL;
182 }
183 }
184 return err;
185}
186
187/**
188 * sdma_v2_4_ring_get_rptr - get the current read pointer
189 *
190 * @ring: amdgpu ring pointer
191 *
192 * Get the current rptr from the hardware (VI+).
193 */
194static uint64_t sdma_v2_4_ring_get_rptr(struct amdgpu_ring *ring)
195{
196 /* XXX check if swapping is necessary on BE */
197 return ring->adev->wb.wb[ring->rptr_offs] >> 2;
198}
199
200/**
201 * sdma_v2_4_ring_get_wptr - get the current write pointer
202 *
203 * @ring: amdgpu ring pointer
204 *
205 * Get the current wptr from the hardware (VI+).
206 */
207static uint64_t sdma_v2_4_ring_get_wptr(struct amdgpu_ring *ring)
208{
209 struct amdgpu_device *adev = ring->adev;
210 u32 wptr = RREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[ring->me]) >> 2;
211
212 return wptr;
213}
214
215/**
216 * sdma_v2_4_ring_set_wptr - commit the write pointer
217 *
218 * @ring: amdgpu ring pointer
219 *
220 * Write the wptr back to the hardware (VI+).
221 */
222static void sdma_v2_4_ring_set_wptr(struct amdgpu_ring *ring)
223{
224 struct amdgpu_device *adev = ring->adev;
225
226 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[ring->me], lower_32_bits(ring->wptr) << 2);
227}
228
229static void sdma_v2_4_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
230{
231 struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
232 int i;
233
234 for (i = 0; i < count; i++)
235 if (sdma && sdma->burst_nop && (i == 0))
236 amdgpu_ring_write(ring, ring->funcs->nop |
237 SDMA_PKT_NOP_HEADER_COUNT(count - 1));
238 else
239 amdgpu_ring_write(ring, ring->funcs->nop);
240}
241
242/**
243 * sdma_v2_4_ring_emit_ib - Schedule an IB on the DMA engine
244 *
245 * @ring: amdgpu ring pointer
246 * @ib: IB object to schedule
247 *
248 * Schedule an IB in the DMA ring (VI).
249 */
250static void sdma_v2_4_ring_emit_ib(struct amdgpu_ring *ring,
251 struct amdgpu_job *job,
252 struct amdgpu_ib *ib,
253 uint32_t flags)
254{
255 unsigned vmid = AMDGPU_JOB_GET_VMID(job);
256
257 /* IB packet must end on a 8 DW boundary */
258 sdma_v2_4_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7);
259
260 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) |
261 SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf));
262 /* base must be 32 byte aligned */
263 amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
264 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
265 amdgpu_ring_write(ring, ib->length_dw);
266 amdgpu_ring_write(ring, 0);
267 amdgpu_ring_write(ring, 0);
268
269}
270
271/**
272 * sdma_v2_4_hdp_flush_ring_emit - emit an hdp flush on the DMA ring
273 *
274 * @ring: amdgpu ring pointer
275 *
276 * Emit an hdp flush packet on the requested DMA ring.
277 */
278static void sdma_v2_4_ring_emit_hdp_flush(struct amdgpu_ring *ring)
279{
280 u32 ref_and_mask = 0;
281
282 if (ring->me == 0)
283 ref_and_mask = REG_SET_FIELD(ref_and_mask, GPU_HDP_FLUSH_DONE, SDMA0, 1);
284 else
285 ref_and_mask = REG_SET_FIELD(ref_and_mask, GPU_HDP_FLUSH_DONE, SDMA1, 1);
286
287 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
288 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
289 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
290 amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_DONE << 2);
291 amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_REQ << 2);
292 amdgpu_ring_write(ring, ref_and_mask); /* reference */
293 amdgpu_ring_write(ring, ref_and_mask); /* mask */
294 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
295 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
296}
297
298/**
299 * sdma_v2_4_ring_emit_fence - emit a fence on the DMA ring
300 *
301 * @ring: amdgpu ring pointer
302 * @fence: amdgpu fence object
303 *
304 * Add a DMA fence packet to the ring to write
305 * the fence seq number and DMA trap packet to generate
306 * an interrupt if needed (VI).
307 */
308static void sdma_v2_4_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
309 unsigned flags)
310{
311 bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
312 /* write the fence */
313 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE));
314 amdgpu_ring_write(ring, lower_32_bits(addr));
315 amdgpu_ring_write(ring, upper_32_bits(addr));
316 amdgpu_ring_write(ring, lower_32_bits(seq));
317
318 /* optionally write high bits as well */
319 if (write64bit) {
320 addr += 4;
321 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE));
322 amdgpu_ring_write(ring, lower_32_bits(addr));
323 amdgpu_ring_write(ring, upper_32_bits(addr));
324 amdgpu_ring_write(ring, upper_32_bits(seq));
325 }
326
327 /* generate an interrupt */
328 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP));
329 amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(0));
330}
331
332/**
333 * sdma_v2_4_gfx_stop - stop the gfx async dma engines
334 *
335 * @adev: amdgpu_device pointer
336 *
337 * Stop the gfx async dma ring buffers (VI).
338 */
339static void sdma_v2_4_gfx_stop(struct amdgpu_device *adev)
340{
341 struct amdgpu_ring *sdma0 = &adev->sdma.instance[0].ring;
342 struct amdgpu_ring *sdma1 = &adev->sdma.instance[1].ring;
343 u32 rb_cntl, ib_cntl;
344 int i;
345
346 if ((adev->mman.buffer_funcs_ring == sdma0) ||
347 (adev->mman.buffer_funcs_ring == sdma1))
348 amdgpu_ttm_set_buffer_funcs_status(adev, false);
349
350 for (i = 0; i < adev->sdma.num_instances; i++) {
351 rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
352 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0);
353 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
354 ib_cntl = RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]);
355 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0);
356 WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
357 }
358}
359
360/**
361 * sdma_v2_4_rlc_stop - stop the compute async dma engines
362 *
363 * @adev: amdgpu_device pointer
364 *
365 * Stop the compute async dma queues (VI).
366 */
367static void sdma_v2_4_rlc_stop(struct amdgpu_device *adev)
368{
369 /* XXX todo */
370}
371
372/**
373 * sdma_v2_4_enable - stop the async dma engines
374 *
375 * @adev: amdgpu_device pointer
376 * @enable: enable/disable the DMA MEs.
377 *
378 * Halt or unhalt the async dma engines (VI).
379 */
380static void sdma_v2_4_enable(struct amdgpu_device *adev, bool enable)
381{
382 u32 f32_cntl;
383 int i;
384
385 if (!enable) {
386 sdma_v2_4_gfx_stop(adev);
387 sdma_v2_4_rlc_stop(adev);
388 }
389
390 for (i = 0; i < adev->sdma.num_instances; i++) {
391 f32_cntl = RREG32(mmSDMA0_F32_CNTL + sdma_offsets[i]);
392 if (enable)
393 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 0);
394 else
395 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 1);
396 WREG32(mmSDMA0_F32_CNTL + sdma_offsets[i], f32_cntl);
397 }
398}
399
400/**
401 * sdma_v2_4_gfx_resume - setup and start the async dma engines
402 *
403 * @adev: amdgpu_device pointer
404 *
405 * Set up the gfx DMA ring buffers and enable them (VI).
406 * Returns 0 for success, error for failure.
407 */
408static int sdma_v2_4_gfx_resume(struct amdgpu_device *adev)
409{
410 struct amdgpu_ring *ring;
411 u32 rb_cntl, ib_cntl;
412 u32 rb_bufsz;
413 u32 wb_offset;
414 int i, j, r;
415
416 for (i = 0; i < adev->sdma.num_instances; i++) {
417 ring = &adev->sdma.instance[i].ring;
418 wb_offset = (ring->rptr_offs * 4);
419
420 mutex_lock(&adev->srbm_mutex);
421 for (j = 0; j < 16; j++) {
422 vi_srbm_select(adev, 0, 0, 0, j);
423 /* SDMA GFX */
424 WREG32(mmSDMA0_GFX_VIRTUAL_ADDR + sdma_offsets[i], 0);
425 WREG32(mmSDMA0_GFX_APE1_CNTL + sdma_offsets[i], 0);
426 }
427 vi_srbm_select(adev, 0, 0, 0, 0);
428 mutex_unlock(&adev->srbm_mutex);
429
430 WREG32(mmSDMA0_TILING_CONFIG + sdma_offsets[i],
431 adev->gfx.config.gb_addr_config & 0x70);
432
433 WREG32(mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
434
435 /* Set ring buffer size in dwords */
436 rb_bufsz = order_base_2(ring->ring_size / 4);
437 rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
438 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz);
439#ifdef __BIG_ENDIAN
440 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1);
441 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
442 RPTR_WRITEBACK_SWAP_ENABLE, 1);
443#endif
444 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
445
446 /* Initialize the ring buffer's read and write pointers */
447 WREG32(mmSDMA0_GFX_RB_RPTR + sdma_offsets[i], 0);
448 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], 0);
449 WREG32(mmSDMA0_GFX_IB_RPTR + sdma_offsets[i], 0);
450 WREG32(mmSDMA0_GFX_IB_OFFSET + sdma_offsets[i], 0);
451
452 /* set the wb address whether it's enabled or not */
453 WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_HI + sdma_offsets[i],
454 upper_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF);
455 WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_LO + sdma_offsets[i],
456 lower_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC);
457
458 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
459
460 WREG32(mmSDMA0_GFX_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
461 WREG32(mmSDMA0_GFX_RB_BASE_HI + sdma_offsets[i], ring->gpu_addr >> 40);
462
463 ring->wptr = 0;
464 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], lower_32_bits(ring->wptr) << 2);
465
466 /* enable DMA RB */
467 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1);
468 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
469
470 ib_cntl = RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]);
471 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1);
472#ifdef __BIG_ENDIAN
473 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1);
474#endif
475 /* enable DMA IBs */
476 WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
477
478 ring->sched.ready = true;
479 }
480
481 sdma_v2_4_enable(adev, true);
482 for (i = 0; i < adev->sdma.num_instances; i++) {
483 ring = &adev->sdma.instance[i].ring;
484 r = amdgpu_ring_test_helper(ring);
485 if (r)
486 return r;
487
488 if (adev->mman.buffer_funcs_ring == ring)
489 amdgpu_ttm_set_buffer_funcs_status(adev, true);
490 }
491
492 return 0;
493}
494
495/**
496 * sdma_v2_4_rlc_resume - setup and start the async dma engines
497 *
498 * @adev: amdgpu_device pointer
499 *
500 * Set up the compute DMA queues and enable them (VI).
501 * Returns 0 for success, error for failure.
502 */
503static int sdma_v2_4_rlc_resume(struct amdgpu_device *adev)
504{
505 /* XXX todo */
506 return 0;
507}
508
509
510/**
511 * sdma_v2_4_start - setup and start the async dma engines
512 *
513 * @adev: amdgpu_device pointer
514 *
515 * Set up the DMA engines and enable them (VI).
516 * Returns 0 for success, error for failure.
517 */
518static int sdma_v2_4_start(struct amdgpu_device *adev)
519{
520 int r;
521
522 /* halt the engine before programing */
523 sdma_v2_4_enable(adev, false);
524
525 /* start the gfx rings and rlc compute queues */
526 r = sdma_v2_4_gfx_resume(adev);
527 if (r)
528 return r;
529 r = sdma_v2_4_rlc_resume(adev);
530 if (r)
531 return r;
532
533 return 0;
534}
535
536/**
537 * sdma_v2_4_ring_test_ring - simple async dma engine test
538 *
539 * @ring: amdgpu_ring structure holding ring information
540 *
541 * Test the DMA engine by writing using it to write an
542 * value to memory. (VI).
543 * Returns 0 for success, error for failure.
544 */
545static int sdma_v2_4_ring_test_ring(struct amdgpu_ring *ring)
546{
547 struct amdgpu_device *adev = ring->adev;
548 unsigned i;
549 unsigned index;
550 int r;
551 u32 tmp;
552 u64 gpu_addr;
553
554 r = amdgpu_device_wb_get(adev, &index);
555 if (r)
556 return r;
557
558 gpu_addr = adev->wb.gpu_addr + (index * 4);
559 tmp = 0xCAFEDEAD;
560 adev->wb.wb[index] = cpu_to_le32(tmp);
561
562 r = amdgpu_ring_alloc(ring, 5);
563 if (r)
564 goto error_free_wb;
565
566 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
567 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
568 amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
569 amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
570 amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1));
571 amdgpu_ring_write(ring, 0xDEADBEEF);
572 amdgpu_ring_commit(ring);
573
574 for (i = 0; i < adev->usec_timeout; i++) {
575 tmp = le32_to_cpu(adev->wb.wb[index]);
576 if (tmp == 0xDEADBEEF)
577 break;
578 udelay(1);
579 }
580
581 if (i >= adev->usec_timeout)
582 r = -ETIMEDOUT;
583
584error_free_wb:
585 amdgpu_device_wb_free(adev, index);
586 return r;
587}
588
589/**
590 * sdma_v2_4_ring_test_ib - test an IB on the DMA engine
591 *
592 * @ring: amdgpu_ring structure holding ring information
593 *
594 * Test a simple IB in the DMA ring (VI).
595 * Returns 0 on success, error on failure.
596 */
597static int sdma_v2_4_ring_test_ib(struct amdgpu_ring *ring, long timeout)
598{
599 struct amdgpu_device *adev = ring->adev;
600 struct amdgpu_ib ib;
601 struct dma_fence *f = NULL;
602 unsigned index;
603 u32 tmp = 0;
604 u64 gpu_addr;
605 long r;
606
607 r = amdgpu_device_wb_get(adev, &index);
608 if (r)
609 return r;
610
611 gpu_addr = adev->wb.gpu_addr + (index * 4);
612 tmp = 0xCAFEDEAD;
613 adev->wb.wb[index] = cpu_to_le32(tmp);
614 memset(&ib, 0, sizeof(ib));
615 r = amdgpu_ib_get(adev, NULL, 256,
616 AMDGPU_IB_POOL_DIRECT, &ib);
617 if (r)
618 goto err0;
619
620 ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
621 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
622 ib.ptr[1] = lower_32_bits(gpu_addr);
623 ib.ptr[2] = upper_32_bits(gpu_addr);
624 ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1);
625 ib.ptr[4] = 0xDEADBEEF;
626 ib.ptr[5] = SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
627 ib.ptr[6] = SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
628 ib.ptr[7] = SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
629 ib.length_dw = 8;
630
631 r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
632 if (r)
633 goto err1;
634
635 r = dma_fence_wait_timeout(f, false, timeout);
636 if (r == 0) {
637 r = -ETIMEDOUT;
638 goto err1;
639 } else if (r < 0) {
640 goto err1;
641 }
642 tmp = le32_to_cpu(adev->wb.wb[index]);
643 if (tmp == 0xDEADBEEF)
644 r = 0;
645 else
646 r = -EINVAL;
647
648err1:
649 amdgpu_ib_free(adev, &ib, NULL);
650 dma_fence_put(f);
651err0:
652 amdgpu_device_wb_free(adev, index);
653 return r;
654}
655
656/**
657 * sdma_v2_4_vm_copy_pte - update PTEs by copying them from the GART
658 *
659 * @ib: indirect buffer to fill with commands
660 * @pe: addr of the page entry
661 * @src: src addr to copy from
662 * @count: number of page entries to update
663 *
664 * Update PTEs by copying them from the GART using sDMA (CIK).
665 */
666static void sdma_v2_4_vm_copy_pte(struct amdgpu_ib *ib,
667 uint64_t pe, uint64_t src,
668 unsigned count)
669{
670 unsigned bytes = count * 8;
671
672 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
673 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
674 ib->ptr[ib->length_dw++] = bytes;
675 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
676 ib->ptr[ib->length_dw++] = lower_32_bits(src);
677 ib->ptr[ib->length_dw++] = upper_32_bits(src);
678 ib->ptr[ib->length_dw++] = lower_32_bits(pe);
679 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
680}
681
682/**
683 * sdma_v2_4_vm_write_pte - update PTEs by writing them manually
684 *
685 * @ib: indirect buffer to fill with commands
686 * @pe: addr of the page entry
687 * @value: dst addr to write into pe
688 * @count: number of page entries to update
689 * @incr: increase next addr by incr bytes
690 *
691 * Update PTEs by writing them manually using sDMA (CIK).
692 */
693static void sdma_v2_4_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
694 uint64_t value, unsigned count,
695 uint32_t incr)
696{
697 unsigned ndw = count * 2;
698
699 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
700 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
701 ib->ptr[ib->length_dw++] = pe;
702 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
703 ib->ptr[ib->length_dw++] = ndw;
704 for (; ndw > 0; ndw -= 2) {
705 ib->ptr[ib->length_dw++] = lower_32_bits(value);
706 ib->ptr[ib->length_dw++] = upper_32_bits(value);
707 value += incr;
708 }
709}
710
711/**
712 * sdma_v2_4_vm_set_pte_pde - update the page tables using sDMA
713 *
714 * @ib: indirect buffer to fill with commands
715 * @pe: addr of the page entry
716 * @addr: dst addr to write into pe
717 * @count: number of page entries to update
718 * @incr: increase next addr by incr bytes
719 * @flags: access flags
720 *
721 * Update the page tables using sDMA (CIK).
722 */
723static void sdma_v2_4_vm_set_pte_pde(struct amdgpu_ib *ib, uint64_t pe,
724 uint64_t addr, unsigned count,
725 uint32_t incr, uint64_t flags)
726{
727 /* for physically contiguous pages (vram) */
728 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_GEN_PTEPDE);
729 ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
730 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
731 ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
732 ib->ptr[ib->length_dw++] = upper_32_bits(flags);
733 ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
734 ib->ptr[ib->length_dw++] = upper_32_bits(addr);
735 ib->ptr[ib->length_dw++] = incr; /* increment size */
736 ib->ptr[ib->length_dw++] = 0;
737 ib->ptr[ib->length_dw++] = count; /* number of entries */
738}
739
740/**
741 * sdma_v2_4_ring_pad_ib - pad the IB to the required number of dw
742 *
743 * @ib: indirect buffer to fill with padding
744 *
745 */
746static void sdma_v2_4_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
747{
748 struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
749 u32 pad_count;
750 int i;
751
752 pad_count = (-ib->length_dw) & 7;
753 for (i = 0; i < pad_count; i++)
754 if (sdma && sdma->burst_nop && (i == 0))
755 ib->ptr[ib->length_dw++] =
756 SDMA_PKT_HEADER_OP(SDMA_OP_NOP) |
757 SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
758 else
759 ib->ptr[ib->length_dw++] =
760 SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
761}
762
763/**
764 * sdma_v2_4_ring_emit_pipeline_sync - sync the pipeline
765 *
766 * @ring: amdgpu_ring pointer
767 *
768 * Make sure all previous operations are completed (CIK).
769 */
770static void sdma_v2_4_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
771{
772 uint32_t seq = ring->fence_drv.sync_seq;
773 uint64_t addr = ring->fence_drv.gpu_addr;
774
775 /* wait for idle */
776 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
777 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
778 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
779 SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
780 amdgpu_ring_write(ring, addr & 0xfffffffc);
781 amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
782 amdgpu_ring_write(ring, seq); /* reference */
783 amdgpu_ring_write(ring, 0xffffffff); /* mask */
784 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
785 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
786}
787
788/**
789 * sdma_v2_4_ring_emit_vm_flush - cik vm flush using sDMA
790 *
791 * @ring: amdgpu_ring pointer
792 * @vm: amdgpu_vm pointer
793 *
794 * Update the page table base and flush the VM TLB
795 * using sDMA (VI).
796 */
797static void sdma_v2_4_ring_emit_vm_flush(struct amdgpu_ring *ring,
798 unsigned vmid, uint64_t pd_addr)
799{
800 amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
801
802 /* wait for flush */
803 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
804 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
805 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(0)); /* always */
806 amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST << 2);
807 amdgpu_ring_write(ring, 0);
808 amdgpu_ring_write(ring, 0); /* reference */
809 amdgpu_ring_write(ring, 0); /* mask */
810 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
811 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
812}
813
814static void sdma_v2_4_ring_emit_wreg(struct amdgpu_ring *ring,
815 uint32_t reg, uint32_t val)
816{
817 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
818 SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
819 amdgpu_ring_write(ring, reg);
820 amdgpu_ring_write(ring, val);
821}
822
823static int sdma_v2_4_early_init(void *handle)
824{
825 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
826
827 adev->sdma.num_instances = SDMA_MAX_INSTANCE;
828
829 sdma_v2_4_set_ring_funcs(adev);
830 sdma_v2_4_set_buffer_funcs(adev);
831 sdma_v2_4_set_vm_pte_funcs(adev);
832 sdma_v2_4_set_irq_funcs(adev);
833
834 return 0;
835}
836
837static int sdma_v2_4_sw_init(void *handle)
838{
839 struct amdgpu_ring *ring;
840 int r, i;
841 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
842
843 /* SDMA trap event */
844 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_SDMA_TRAP,
845 &adev->sdma.trap_irq);
846 if (r)
847 return r;
848
849 /* SDMA Privileged inst */
850 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 241,
851 &adev->sdma.illegal_inst_irq);
852 if (r)
853 return r;
854
855 /* SDMA Privileged inst */
856 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_SDMA_SRBM_WRITE,
857 &adev->sdma.illegal_inst_irq);
858 if (r)
859 return r;
860
861 r = sdma_v2_4_init_microcode(adev);
862 if (r) {
863 DRM_ERROR("Failed to load sdma firmware!\n");
864 return r;
865 }
866
867 for (i = 0; i < adev->sdma.num_instances; i++) {
868 ring = &adev->sdma.instance[i].ring;
869 ring->ring_obj = NULL;
870 ring->use_doorbell = false;
871 sprintf(ring->name, "sdma%d", i);
872 r = amdgpu_ring_init(adev, ring, 1024,
873 &adev->sdma.trap_irq,
874 (i == 0) ?
875 AMDGPU_SDMA_IRQ_INSTANCE0 :
876 AMDGPU_SDMA_IRQ_INSTANCE1,
877 AMDGPU_RING_PRIO_DEFAULT);
878 if (r)
879 return r;
880 }
881
882 return r;
883}
884
885static int sdma_v2_4_sw_fini(void *handle)
886{
887 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
888 int i;
889
890 for (i = 0; i < adev->sdma.num_instances; i++)
891 amdgpu_ring_fini(&adev->sdma.instance[i].ring);
892
893 sdma_v2_4_free_microcode(adev);
894 return 0;
895}
896
897static int sdma_v2_4_hw_init(void *handle)
898{
899 int r;
900 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
901
902 sdma_v2_4_init_golden_registers(adev);
903
904 r = sdma_v2_4_start(adev);
905 if (r)
906 return r;
907
908 return r;
909}
910
911static int sdma_v2_4_hw_fini(void *handle)
912{
913 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
914
915 sdma_v2_4_enable(adev, false);
916
917 return 0;
918}
919
920static int sdma_v2_4_suspend(void *handle)
921{
922 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
923
924 return sdma_v2_4_hw_fini(adev);
925}
926
927static int sdma_v2_4_resume(void *handle)
928{
929 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
930
931 return sdma_v2_4_hw_init(adev);
932}
933
934static bool sdma_v2_4_is_idle(void *handle)
935{
936 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
937 u32 tmp = RREG32(mmSRBM_STATUS2);
938
939 if (tmp & (SRBM_STATUS2__SDMA_BUSY_MASK |
940 SRBM_STATUS2__SDMA1_BUSY_MASK))
941 return false;
942
943 return true;
944}
945
946static int sdma_v2_4_wait_for_idle(void *handle)
947{
948 unsigned i;
949 u32 tmp;
950 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
951
952 for (i = 0; i < adev->usec_timeout; i++) {
953 tmp = RREG32(mmSRBM_STATUS2) & (SRBM_STATUS2__SDMA_BUSY_MASK |
954 SRBM_STATUS2__SDMA1_BUSY_MASK);
955
956 if (!tmp)
957 return 0;
958 udelay(1);
959 }
960 return -ETIMEDOUT;
961}
962
963static int sdma_v2_4_soft_reset(void *handle)
964{
965 u32 srbm_soft_reset = 0;
966 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
967 u32 tmp = RREG32(mmSRBM_STATUS2);
968
969 if (tmp & SRBM_STATUS2__SDMA_BUSY_MASK) {
970 /* sdma0 */
971 tmp = RREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET);
972 tmp = REG_SET_FIELD(tmp, SDMA0_F32_CNTL, HALT, 0);
973 WREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET, tmp);
974 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA_MASK;
975 }
976 if (tmp & SRBM_STATUS2__SDMA1_BUSY_MASK) {
977 /* sdma1 */
978 tmp = RREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET);
979 tmp = REG_SET_FIELD(tmp, SDMA0_F32_CNTL, HALT, 0);
980 WREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET, tmp);
981 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA1_MASK;
982 }
983
984 if (srbm_soft_reset) {
985 tmp = RREG32(mmSRBM_SOFT_RESET);
986 tmp |= srbm_soft_reset;
987 dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
988 WREG32(mmSRBM_SOFT_RESET, tmp);
989 tmp = RREG32(mmSRBM_SOFT_RESET);
990
991 udelay(50);
992
993 tmp &= ~srbm_soft_reset;
994 WREG32(mmSRBM_SOFT_RESET, tmp);
995 tmp = RREG32(mmSRBM_SOFT_RESET);
996
997 /* Wait a little for things to settle down */
998 udelay(50);
999 }
1000
1001 return 0;
1002}
1003
1004static int sdma_v2_4_set_trap_irq_state(struct amdgpu_device *adev,
1005 struct amdgpu_irq_src *src,
1006 unsigned type,
1007 enum amdgpu_interrupt_state state)
1008{
1009 u32 sdma_cntl;
1010
1011 switch (type) {
1012 case AMDGPU_SDMA_IRQ_INSTANCE0:
1013 switch (state) {
1014 case AMDGPU_IRQ_STATE_DISABLE:
1015 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1016 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 0);
1017 WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1018 break;
1019 case AMDGPU_IRQ_STATE_ENABLE:
1020 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1021 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 1);
1022 WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1023 break;
1024 default:
1025 break;
1026 }
1027 break;
1028 case AMDGPU_SDMA_IRQ_INSTANCE1:
1029 switch (state) {
1030 case AMDGPU_IRQ_STATE_DISABLE:
1031 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1032 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 0);
1033 WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1034 break;
1035 case AMDGPU_IRQ_STATE_ENABLE:
1036 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1037 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 1);
1038 WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1039 break;
1040 default:
1041 break;
1042 }
1043 break;
1044 default:
1045 break;
1046 }
1047 return 0;
1048}
1049
1050static int sdma_v2_4_process_trap_irq(struct amdgpu_device *adev,
1051 struct amdgpu_irq_src *source,
1052 struct amdgpu_iv_entry *entry)
1053{
1054 u8 instance_id, queue_id;
1055
1056 instance_id = (entry->ring_id & 0x3) >> 0;
1057 queue_id = (entry->ring_id & 0xc) >> 2;
1058 DRM_DEBUG("IH: SDMA trap\n");
1059 switch (instance_id) {
1060 case 0:
1061 switch (queue_id) {
1062 case 0:
1063 amdgpu_fence_process(&adev->sdma.instance[0].ring);
1064 break;
1065 case 1:
1066 /* XXX compute */
1067 break;
1068 case 2:
1069 /* XXX compute */
1070 break;
1071 }
1072 break;
1073 case 1:
1074 switch (queue_id) {
1075 case 0:
1076 amdgpu_fence_process(&adev->sdma.instance[1].ring);
1077 break;
1078 case 1:
1079 /* XXX compute */
1080 break;
1081 case 2:
1082 /* XXX compute */
1083 break;
1084 }
1085 break;
1086 }
1087 return 0;
1088}
1089
1090static int sdma_v2_4_process_illegal_inst_irq(struct amdgpu_device *adev,
1091 struct amdgpu_irq_src *source,
1092 struct amdgpu_iv_entry *entry)
1093{
1094 u8 instance_id, queue_id;
1095
1096 DRM_ERROR("Illegal instruction in SDMA command stream\n");
1097 instance_id = (entry->ring_id & 0x3) >> 0;
1098 queue_id = (entry->ring_id & 0xc) >> 2;
1099
1100 if (instance_id <= 1 && queue_id == 0)
1101 drm_sched_fault(&adev->sdma.instance[instance_id].ring.sched);
1102 return 0;
1103}
1104
1105static int sdma_v2_4_set_clockgating_state(void *handle,
1106 enum amd_clockgating_state state)
1107{
1108 /* XXX handled via the smc on VI */
1109 return 0;
1110}
1111
1112static int sdma_v2_4_set_powergating_state(void *handle,
1113 enum amd_powergating_state state)
1114{
1115 return 0;
1116}
1117
1118static const struct amd_ip_funcs sdma_v2_4_ip_funcs = {
1119 .name = "sdma_v2_4",
1120 .early_init = sdma_v2_4_early_init,
1121 .late_init = NULL,
1122 .sw_init = sdma_v2_4_sw_init,
1123 .sw_fini = sdma_v2_4_sw_fini,
1124 .hw_init = sdma_v2_4_hw_init,
1125 .hw_fini = sdma_v2_4_hw_fini,
1126 .suspend = sdma_v2_4_suspend,
1127 .resume = sdma_v2_4_resume,
1128 .is_idle = sdma_v2_4_is_idle,
1129 .wait_for_idle = sdma_v2_4_wait_for_idle,
1130 .soft_reset = sdma_v2_4_soft_reset,
1131 .set_clockgating_state = sdma_v2_4_set_clockgating_state,
1132 .set_powergating_state = sdma_v2_4_set_powergating_state,
1133};
1134
1135static const struct amdgpu_ring_funcs sdma_v2_4_ring_funcs = {
1136 .type = AMDGPU_RING_TYPE_SDMA,
1137 .align_mask = 0xf,
1138 .nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
1139 .support_64bit_ptrs = false,
1140 .get_rptr = sdma_v2_4_ring_get_rptr,
1141 .get_wptr = sdma_v2_4_ring_get_wptr,
1142 .set_wptr = sdma_v2_4_ring_set_wptr,
1143 .emit_frame_size =
1144 6 + /* sdma_v2_4_ring_emit_hdp_flush */
1145 3 + /* hdp invalidate */
1146 6 + /* sdma_v2_4_ring_emit_pipeline_sync */
1147 VI_FLUSH_GPU_TLB_NUM_WREG * 3 + 6 + /* sdma_v2_4_ring_emit_vm_flush */
1148 10 + 10 + 10, /* sdma_v2_4_ring_emit_fence x3 for user fence, vm fence */
1149 .emit_ib_size = 7 + 6, /* sdma_v2_4_ring_emit_ib */
1150 .emit_ib = sdma_v2_4_ring_emit_ib,
1151 .emit_fence = sdma_v2_4_ring_emit_fence,
1152 .emit_pipeline_sync = sdma_v2_4_ring_emit_pipeline_sync,
1153 .emit_vm_flush = sdma_v2_4_ring_emit_vm_flush,
1154 .emit_hdp_flush = sdma_v2_4_ring_emit_hdp_flush,
1155 .test_ring = sdma_v2_4_ring_test_ring,
1156 .test_ib = sdma_v2_4_ring_test_ib,
1157 .insert_nop = sdma_v2_4_ring_insert_nop,
1158 .pad_ib = sdma_v2_4_ring_pad_ib,
1159 .emit_wreg = sdma_v2_4_ring_emit_wreg,
1160};
1161
1162static void sdma_v2_4_set_ring_funcs(struct amdgpu_device *adev)
1163{
1164 int i;
1165
1166 for (i = 0; i < adev->sdma.num_instances; i++) {
1167 adev->sdma.instance[i].ring.funcs = &sdma_v2_4_ring_funcs;
1168 adev->sdma.instance[i].ring.me = i;
1169 }
1170}
1171
1172static const struct amdgpu_irq_src_funcs sdma_v2_4_trap_irq_funcs = {
1173 .set = sdma_v2_4_set_trap_irq_state,
1174 .process = sdma_v2_4_process_trap_irq,
1175};
1176
1177static const struct amdgpu_irq_src_funcs sdma_v2_4_illegal_inst_irq_funcs = {
1178 .process = sdma_v2_4_process_illegal_inst_irq,
1179};
1180
1181static void sdma_v2_4_set_irq_funcs(struct amdgpu_device *adev)
1182{
1183 adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
1184 adev->sdma.trap_irq.funcs = &sdma_v2_4_trap_irq_funcs;
1185 adev->sdma.illegal_inst_irq.funcs = &sdma_v2_4_illegal_inst_irq_funcs;
1186}
1187
1188/**
1189 * sdma_v2_4_emit_copy_buffer - copy buffer using the sDMA engine
1190 *
1191 * @ring: amdgpu_ring structure holding ring information
1192 * @src_offset: src GPU address
1193 * @dst_offset: dst GPU address
1194 * @byte_count: number of bytes to xfer
1195 *
1196 * Copy GPU buffers using the DMA engine (VI).
1197 * Used by the amdgpu ttm implementation to move pages if
1198 * registered as the asic copy callback.
1199 */
1200static void sdma_v2_4_emit_copy_buffer(struct amdgpu_ib *ib,
1201 uint64_t src_offset,
1202 uint64_t dst_offset,
1203 uint32_t byte_count,
1204 bool tmz)
1205{
1206 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1207 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1208 ib->ptr[ib->length_dw++] = byte_count;
1209 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1210 ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1211 ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1212 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1213 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1214}
1215
1216/**
1217 * sdma_v2_4_emit_fill_buffer - fill buffer using the sDMA engine
1218 *
1219 * @ring: amdgpu_ring structure holding ring information
1220 * @src_data: value to write to buffer
1221 * @dst_offset: dst GPU address
1222 * @byte_count: number of bytes to xfer
1223 *
1224 * Fill GPU buffers using the DMA engine (VI).
1225 */
1226static void sdma_v2_4_emit_fill_buffer(struct amdgpu_ib *ib,
1227 uint32_t src_data,
1228 uint64_t dst_offset,
1229 uint32_t byte_count)
1230{
1231 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL);
1232 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1233 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1234 ib->ptr[ib->length_dw++] = src_data;
1235 ib->ptr[ib->length_dw++] = byte_count;
1236}
1237
1238static const struct amdgpu_buffer_funcs sdma_v2_4_buffer_funcs = {
1239 .copy_max_bytes = 0x1fffff,
1240 .copy_num_dw = 7,
1241 .emit_copy_buffer = sdma_v2_4_emit_copy_buffer,
1242
1243 .fill_max_bytes = 0x1fffff,
1244 .fill_num_dw = 7,
1245 .emit_fill_buffer = sdma_v2_4_emit_fill_buffer,
1246};
1247
1248static void sdma_v2_4_set_buffer_funcs(struct amdgpu_device *adev)
1249{
1250 adev->mman.buffer_funcs = &sdma_v2_4_buffer_funcs;
1251 adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1252}
1253
1254static const struct amdgpu_vm_pte_funcs sdma_v2_4_vm_pte_funcs = {
1255 .copy_pte_num_dw = 7,
1256 .copy_pte = sdma_v2_4_vm_copy_pte,
1257
1258 .write_pte = sdma_v2_4_vm_write_pte,
1259 .set_pte_pde = sdma_v2_4_vm_set_pte_pde,
1260};
1261
1262static void sdma_v2_4_set_vm_pte_funcs(struct amdgpu_device *adev)
1263{
1264 unsigned i;
1265
1266 adev->vm_manager.vm_pte_funcs = &sdma_v2_4_vm_pte_funcs;
1267 for (i = 0; i < adev->sdma.num_instances; i++) {
1268 adev->vm_manager.vm_pte_scheds[i] =
1269 &adev->sdma.instance[i].ring.sched;
1270 }
1271 adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
1272}
1273
1274const struct amdgpu_ip_block_version sdma_v2_4_ip_block =
1275{
1276 .type = AMD_IP_BLOCK_TYPE_SDMA,
1277 .major = 2,
1278 .minor = 4,
1279 .rev = 0,
1280 .funcs = &sdma_v2_4_ip_funcs,
1281};
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_2_4_d.h"
33#include "oss/oss_2_4_sh_mask.h"
34
35#include "gmc/gmc_7_1_d.h"
36#include "gmc/gmc_7_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 "iceland_sdma_pkt_open.h"
46
47static void sdma_v2_4_set_ring_funcs(struct amdgpu_device *adev);
48static void sdma_v2_4_set_buffer_funcs(struct amdgpu_device *adev);
49static void sdma_v2_4_set_vm_pte_funcs(struct amdgpu_device *adev);
50static void sdma_v2_4_set_irq_funcs(struct amdgpu_device *adev);
51
52MODULE_FIRMWARE("amdgpu/topaz_sdma.bin");
53MODULE_FIRMWARE("amdgpu/topaz_sdma1.bin");
54
55static const u32 sdma_offsets[SDMA_MAX_INSTANCE] =
56{
57 SDMA0_REGISTER_OFFSET,
58 SDMA1_REGISTER_OFFSET
59};
60
61static const u32 golden_settings_iceland_a11[] =
62{
63 mmSDMA0_CHICKEN_BITS, 0xfc910007, 0x00810007,
64 mmSDMA0_CLK_CTRL, 0xff000fff, 0x00000000,
65 mmSDMA1_CHICKEN_BITS, 0xfc910007, 0x00810007,
66 mmSDMA1_CLK_CTRL, 0xff000fff, 0x00000000,
67};
68
69static const u32 iceland_mgcg_cgcg_init[] =
70{
71 mmSDMA0_CLK_CTRL, 0xff000ff0, 0x00000100,
72 mmSDMA1_CLK_CTRL, 0xff000ff0, 0x00000100
73};
74
75/*
76 * sDMA - System DMA
77 * Starting with CIK, the GPU has new asynchronous
78 * DMA engines. These engines are used for compute
79 * and gfx. There are two DMA engines (SDMA0, SDMA1)
80 * and each one supports 1 ring buffer used for gfx
81 * and 2 queues used for compute.
82 *
83 * The programming model is very similar to the CP
84 * (ring buffer, IBs, etc.), but sDMA has it's own
85 * packet format that is different from the PM4 format
86 * used by the CP. sDMA supports copying data, writing
87 * embedded data, solid fills, and a number of other
88 * things. It also has support for tiling/detiling of
89 * buffers.
90 */
91
92static void sdma_v2_4_init_golden_registers(struct amdgpu_device *adev)
93{
94 switch (adev->asic_type) {
95 case CHIP_TOPAZ:
96 amdgpu_program_register_sequence(adev,
97 iceland_mgcg_cgcg_init,
98 (const u32)ARRAY_SIZE(iceland_mgcg_cgcg_init));
99 amdgpu_program_register_sequence(adev,
100 golden_settings_iceland_a11,
101 (const u32)ARRAY_SIZE(golden_settings_iceland_a11));
102 break;
103 default:
104 break;
105 }
106}
107
108static void sdma_v2_4_free_microcode(struct amdgpu_device *adev)
109{
110 int i;
111 for (i = 0; i < adev->sdma.num_instances; i++) {
112 release_firmware(adev->sdma.instance[i].fw);
113 adev->sdma.instance[i].fw = NULL;
114 }
115}
116
117/**
118 * sdma_v2_4_init_microcode - load ucode images from disk
119 *
120 * @adev: amdgpu_device pointer
121 *
122 * Use the firmware interface to load the ucode images into
123 * the driver (not loaded into hw).
124 * Returns 0 on success, error on failure.
125 */
126static int sdma_v2_4_init_microcode(struct amdgpu_device *adev)
127{
128 const char *chip_name;
129 char fw_name[30];
130 int err = 0, i;
131 struct amdgpu_firmware_info *info = NULL;
132 const struct common_firmware_header *header = NULL;
133 const struct sdma_firmware_header_v1_0 *hdr;
134
135 DRM_DEBUG("\n");
136
137 switch (adev->asic_type) {
138 case CHIP_TOPAZ:
139 chip_name = "topaz";
140 break;
141 default: BUG();
142 }
143
144 for (i = 0; i < adev->sdma.num_instances; i++) {
145 if (i == 0)
146 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma.bin", chip_name);
147 else
148 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma1.bin", chip_name);
149 err = request_firmware(&adev->sdma.instance[i].fw, fw_name, adev->dev);
150 if (err)
151 goto out;
152 err = amdgpu_ucode_validate(adev->sdma.instance[i].fw);
153 if (err)
154 goto out;
155 hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
156 adev->sdma.instance[i].fw_version = le32_to_cpu(hdr->header.ucode_version);
157 adev->sdma.instance[i].feature_version = le32_to_cpu(hdr->ucode_feature_version);
158 if (adev->sdma.instance[i].feature_version >= 20)
159 adev->sdma.instance[i].burst_nop = true;
160
161 if (adev->firmware.smu_load) {
162 info = &adev->firmware.ucode[AMDGPU_UCODE_ID_SDMA0 + i];
163 info->ucode_id = AMDGPU_UCODE_ID_SDMA0 + i;
164 info->fw = adev->sdma.instance[i].fw;
165 header = (const struct common_firmware_header *)info->fw->data;
166 adev->firmware.fw_size +=
167 ALIGN(le32_to_cpu(header->ucode_size_bytes), PAGE_SIZE);
168 }
169 }
170
171out:
172 if (err) {
173 printk(KERN_ERR
174 "sdma_v2_4: Failed to load firmware \"%s\"\n",
175 fw_name);
176 for (i = 0; i < adev->sdma.num_instances; i++) {
177 release_firmware(adev->sdma.instance[i].fw);
178 adev->sdma.instance[i].fw = NULL;
179 }
180 }
181 return err;
182}
183
184/**
185 * sdma_v2_4_ring_get_rptr - get the current read pointer
186 *
187 * @ring: amdgpu ring pointer
188 *
189 * Get the current rptr from the hardware (VI+).
190 */
191static uint32_t sdma_v2_4_ring_get_rptr(struct amdgpu_ring *ring)
192{
193 /* XXX check if swapping is necessary on BE */
194 return ring->adev->wb.wb[ring->rptr_offs] >> 2;
195}
196
197/**
198 * sdma_v2_4_ring_get_wptr - get the current write pointer
199 *
200 * @ring: amdgpu ring pointer
201 *
202 * Get the current wptr from the hardware (VI+).
203 */
204static uint32_t sdma_v2_4_ring_get_wptr(struct amdgpu_ring *ring)
205{
206 struct amdgpu_device *adev = ring->adev;
207 int me = (ring == &ring->adev->sdma.instance[0].ring) ? 0 : 1;
208 u32 wptr = RREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[me]) >> 2;
209
210 return wptr;
211}
212
213/**
214 * sdma_v2_4_ring_set_wptr - commit the write pointer
215 *
216 * @ring: amdgpu ring pointer
217 *
218 * Write the wptr back to the hardware (VI+).
219 */
220static void sdma_v2_4_ring_set_wptr(struct amdgpu_ring *ring)
221{
222 struct amdgpu_device *adev = ring->adev;
223 int me = (ring == &ring->adev->sdma.instance[0].ring) ? 0 : 1;
224
225 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[me], ring->wptr << 2);
226}
227
228static void sdma_v2_4_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
229{
230 struct amdgpu_sdma_instance *sdma = amdgpu_get_sdma_instance(ring);
231 int i;
232
233 for (i = 0; i < count; i++)
234 if (sdma && sdma->burst_nop && (i == 0))
235 amdgpu_ring_write(ring, ring->funcs->nop |
236 SDMA_PKT_NOP_HEADER_COUNT(count - 1));
237 else
238 amdgpu_ring_write(ring, ring->funcs->nop);
239}
240
241/**
242 * sdma_v2_4_ring_emit_ib - Schedule an IB on the DMA engine
243 *
244 * @ring: amdgpu ring pointer
245 * @ib: IB object to schedule
246 *
247 * Schedule an IB in the DMA ring (VI).
248 */
249static void sdma_v2_4_ring_emit_ib(struct amdgpu_ring *ring,
250 struct amdgpu_ib *ib,
251 unsigned vm_id, bool ctx_switch)
252{
253 u32 vmid = vm_id & 0xf;
254
255 /* IB packet must end on a 8 DW boundary */
256 sdma_v2_4_ring_insert_nop(ring, (10 - (ring->wptr & 7)) % 8);
257
258 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) |
259 SDMA_PKT_INDIRECT_HEADER_VMID(vmid));
260 /* base must be 32 byte aligned */
261 amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
262 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
263 amdgpu_ring_write(ring, ib->length_dw);
264 amdgpu_ring_write(ring, 0);
265 amdgpu_ring_write(ring, 0);
266
267}
268
269/**
270 * sdma_v2_4_hdp_flush_ring_emit - emit an hdp flush on the DMA ring
271 *
272 * @ring: amdgpu ring pointer
273 *
274 * Emit an hdp flush packet on the requested DMA ring.
275 */
276static void sdma_v2_4_ring_emit_hdp_flush(struct amdgpu_ring *ring)
277{
278 u32 ref_and_mask = 0;
279
280 if (ring == &ring->adev->sdma.instance[0].ring)
281 ref_and_mask = REG_SET_FIELD(ref_and_mask, GPU_HDP_FLUSH_DONE, SDMA0, 1);
282 else
283 ref_and_mask = REG_SET_FIELD(ref_and_mask, GPU_HDP_FLUSH_DONE, SDMA1, 1);
284
285 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
286 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
287 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
288 amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_DONE << 2);
289 amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_REQ << 2);
290 amdgpu_ring_write(ring, ref_and_mask); /* reference */
291 amdgpu_ring_write(ring, ref_and_mask); /* mask */
292 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
293 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
294}
295
296static void sdma_v2_4_ring_emit_hdp_invalidate(struct amdgpu_ring *ring)
297{
298 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
299 SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
300 amdgpu_ring_write(ring, mmHDP_DEBUG0);
301 amdgpu_ring_write(ring, 1);
302}
303/**
304 * sdma_v2_4_ring_emit_fence - emit a fence on the DMA ring
305 *
306 * @ring: amdgpu ring pointer
307 * @fence: amdgpu fence object
308 *
309 * Add a DMA fence packet to the ring to write
310 * the fence seq number and DMA trap packet to generate
311 * an interrupt if needed (VI).
312 */
313static void sdma_v2_4_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
314 unsigned flags)
315{
316 bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
317 /* write the fence */
318 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE));
319 amdgpu_ring_write(ring, lower_32_bits(addr));
320 amdgpu_ring_write(ring, upper_32_bits(addr));
321 amdgpu_ring_write(ring, lower_32_bits(seq));
322
323 /* optionally write high bits as well */
324 if (write64bit) {
325 addr += 4;
326 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE));
327 amdgpu_ring_write(ring, lower_32_bits(addr));
328 amdgpu_ring_write(ring, upper_32_bits(addr));
329 amdgpu_ring_write(ring, upper_32_bits(seq));
330 }
331
332 /* generate an interrupt */
333 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP));
334 amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(0));
335}
336
337/**
338 * sdma_v2_4_gfx_stop - stop the gfx async dma engines
339 *
340 * @adev: amdgpu_device pointer
341 *
342 * Stop the gfx async dma ring buffers (VI).
343 */
344static void sdma_v2_4_gfx_stop(struct amdgpu_device *adev)
345{
346 struct amdgpu_ring *sdma0 = &adev->sdma.instance[0].ring;
347 struct amdgpu_ring *sdma1 = &adev->sdma.instance[1].ring;
348 u32 rb_cntl, ib_cntl;
349 int i;
350
351 if ((adev->mman.buffer_funcs_ring == sdma0) ||
352 (adev->mman.buffer_funcs_ring == sdma1))
353 amdgpu_ttm_set_active_vram_size(adev, adev->mc.visible_vram_size);
354
355 for (i = 0; i < adev->sdma.num_instances; i++) {
356 rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
357 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0);
358 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
359 ib_cntl = RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]);
360 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0);
361 WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
362 }
363 sdma0->ready = false;
364 sdma1->ready = false;
365}
366
367/**
368 * sdma_v2_4_rlc_stop - stop the compute async dma engines
369 *
370 * @adev: amdgpu_device pointer
371 *
372 * Stop the compute async dma queues (VI).
373 */
374static void sdma_v2_4_rlc_stop(struct amdgpu_device *adev)
375{
376 /* XXX todo */
377}
378
379/**
380 * sdma_v2_4_enable - stop the async dma engines
381 *
382 * @adev: amdgpu_device pointer
383 * @enable: enable/disable the DMA MEs.
384 *
385 * Halt or unhalt the async dma engines (VI).
386 */
387static void sdma_v2_4_enable(struct amdgpu_device *adev, bool enable)
388{
389 u32 f32_cntl;
390 int i;
391
392 if (!enable) {
393 sdma_v2_4_gfx_stop(adev);
394 sdma_v2_4_rlc_stop(adev);
395 }
396
397 for (i = 0; i < adev->sdma.num_instances; i++) {
398 f32_cntl = RREG32(mmSDMA0_F32_CNTL + sdma_offsets[i]);
399 if (enable)
400 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 0);
401 else
402 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 1);
403 WREG32(mmSDMA0_F32_CNTL + sdma_offsets[i], f32_cntl);
404 }
405}
406
407/**
408 * sdma_v2_4_gfx_resume - setup and start the async dma engines
409 *
410 * @adev: amdgpu_device pointer
411 *
412 * Set up the gfx DMA ring buffers and enable them (VI).
413 * Returns 0 for success, error for failure.
414 */
415static int sdma_v2_4_gfx_resume(struct amdgpu_device *adev)
416{
417 struct amdgpu_ring *ring;
418 u32 rb_cntl, ib_cntl;
419 u32 rb_bufsz;
420 u32 wb_offset;
421 int i, j, r;
422
423 for (i = 0; i < adev->sdma.num_instances; i++) {
424 ring = &adev->sdma.instance[i].ring;
425 wb_offset = (ring->rptr_offs * 4);
426
427 mutex_lock(&adev->srbm_mutex);
428 for (j = 0; j < 16; j++) {
429 vi_srbm_select(adev, 0, 0, 0, j);
430 /* SDMA GFX */
431 WREG32(mmSDMA0_GFX_VIRTUAL_ADDR + sdma_offsets[i], 0);
432 WREG32(mmSDMA0_GFX_APE1_CNTL + sdma_offsets[i], 0);
433 }
434 vi_srbm_select(adev, 0, 0, 0, 0);
435 mutex_unlock(&adev->srbm_mutex);
436
437 WREG32(mmSDMA0_TILING_CONFIG + sdma_offsets[i],
438 adev->gfx.config.gb_addr_config & 0x70);
439
440 WREG32(mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
441
442 /* Set ring buffer size in dwords */
443 rb_bufsz = order_base_2(ring->ring_size / 4);
444 rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
445 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz);
446#ifdef __BIG_ENDIAN
447 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1);
448 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
449 RPTR_WRITEBACK_SWAP_ENABLE, 1);
450#endif
451 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
452
453 /* Initialize the ring buffer's read and write pointers */
454 WREG32(mmSDMA0_GFX_RB_RPTR + sdma_offsets[i], 0);
455 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], 0);
456 WREG32(mmSDMA0_GFX_IB_RPTR + sdma_offsets[i], 0);
457 WREG32(mmSDMA0_GFX_IB_OFFSET + sdma_offsets[i], 0);
458
459 /* set the wb address whether it's enabled or not */
460 WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_HI + sdma_offsets[i],
461 upper_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF);
462 WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_LO + sdma_offsets[i],
463 lower_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC);
464
465 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
466
467 WREG32(mmSDMA0_GFX_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
468 WREG32(mmSDMA0_GFX_RB_BASE_HI + sdma_offsets[i], ring->gpu_addr >> 40);
469
470 ring->wptr = 0;
471 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], ring->wptr << 2);
472
473 /* enable DMA RB */
474 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1);
475 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
476
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, 1);
479#ifdef __BIG_ENDIAN
480 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1);
481#endif
482 /* enable DMA IBs */
483 WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
484
485 ring->ready = true;
486 }
487
488 sdma_v2_4_enable(adev, true);
489 for (i = 0; i < adev->sdma.num_instances; i++) {
490 ring = &adev->sdma.instance[i].ring;
491 r = amdgpu_ring_test_ring(ring);
492 if (r) {
493 ring->ready = false;
494 return r;
495 }
496
497 if (adev->mman.buffer_funcs_ring == ring)
498 amdgpu_ttm_set_active_vram_size(adev, adev->mc.real_vram_size);
499 }
500
501 return 0;
502}
503
504/**
505 * sdma_v2_4_rlc_resume - setup and start the async dma engines
506 *
507 * @adev: amdgpu_device pointer
508 *
509 * Set up the compute DMA queues and enable them (VI).
510 * Returns 0 for success, error for failure.
511 */
512static int sdma_v2_4_rlc_resume(struct amdgpu_device *adev)
513{
514 /* XXX todo */
515 return 0;
516}
517
518/**
519 * sdma_v2_4_load_microcode - load the sDMA ME ucode
520 *
521 * @adev: amdgpu_device pointer
522 *
523 * Loads the sDMA0/1 ucode.
524 * Returns 0 for success, -EINVAL if the ucode is not available.
525 */
526static int sdma_v2_4_load_microcode(struct amdgpu_device *adev)
527{
528 const struct sdma_firmware_header_v1_0 *hdr;
529 const __le32 *fw_data;
530 u32 fw_size;
531 int i, j;
532
533 /* halt the MEs */
534 sdma_v2_4_enable(adev, false);
535
536 for (i = 0; i < adev->sdma.num_instances; i++) {
537 if (!adev->sdma.instance[i].fw)
538 return -EINVAL;
539 hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
540 amdgpu_ucode_print_sdma_hdr(&hdr->header);
541 fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
542 fw_data = (const __le32 *)
543 (adev->sdma.instance[i].fw->data +
544 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
545 WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], 0);
546 for (j = 0; j < fw_size; j++)
547 WREG32(mmSDMA0_UCODE_DATA + sdma_offsets[i], le32_to_cpup(fw_data++));
548 WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], adev->sdma.instance[i].fw_version);
549 }
550
551 return 0;
552}
553
554/**
555 * sdma_v2_4_start - setup and start the async dma engines
556 *
557 * @adev: amdgpu_device pointer
558 *
559 * Set up the DMA engines and enable them (VI).
560 * Returns 0 for success, error for failure.
561 */
562static int sdma_v2_4_start(struct amdgpu_device *adev)
563{
564 int r;
565
566 if (!adev->pp_enabled) {
567 if (!adev->firmware.smu_load) {
568 r = sdma_v2_4_load_microcode(adev);
569 if (r)
570 return r;
571 } else {
572 r = adev->smu.smumgr_funcs->check_fw_load_finish(adev,
573 AMDGPU_UCODE_ID_SDMA0);
574 if (r)
575 return -EINVAL;
576 r = adev->smu.smumgr_funcs->check_fw_load_finish(adev,
577 AMDGPU_UCODE_ID_SDMA1);
578 if (r)
579 return -EINVAL;
580 }
581 }
582
583 /* halt the engine before programing */
584 sdma_v2_4_enable(adev, false);
585
586 /* start the gfx rings and rlc compute queues */
587 r = sdma_v2_4_gfx_resume(adev);
588 if (r)
589 return r;
590 r = sdma_v2_4_rlc_resume(adev);
591 if (r)
592 return r;
593
594 return 0;
595}
596
597/**
598 * sdma_v2_4_ring_test_ring - simple async dma engine test
599 *
600 * @ring: amdgpu_ring structure holding ring information
601 *
602 * Test the DMA engine by writing using it to write an
603 * value to memory. (VI).
604 * Returns 0 for success, error for failure.
605 */
606static int sdma_v2_4_ring_test_ring(struct amdgpu_ring *ring)
607{
608 struct amdgpu_device *adev = ring->adev;
609 unsigned i;
610 unsigned index;
611 int r;
612 u32 tmp;
613 u64 gpu_addr;
614
615 r = amdgpu_wb_get(adev, &index);
616 if (r) {
617 dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
618 return r;
619 }
620
621 gpu_addr = adev->wb.gpu_addr + (index * 4);
622 tmp = 0xCAFEDEAD;
623 adev->wb.wb[index] = cpu_to_le32(tmp);
624
625 r = amdgpu_ring_alloc(ring, 5);
626 if (r) {
627 DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
628 amdgpu_wb_free(adev, index);
629 return r;
630 }
631
632 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
633 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
634 amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
635 amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
636 amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1));
637 amdgpu_ring_write(ring, 0xDEADBEEF);
638 amdgpu_ring_commit(ring);
639
640 for (i = 0; i < adev->usec_timeout; i++) {
641 tmp = le32_to_cpu(adev->wb.wb[index]);
642 if (tmp == 0xDEADBEEF)
643 break;
644 DRM_UDELAY(1);
645 }
646
647 if (i < adev->usec_timeout) {
648 DRM_INFO("ring test on %d succeeded in %d usecs\n", ring->idx, i);
649 } else {
650 DRM_ERROR("amdgpu: ring %d test failed (0x%08X)\n",
651 ring->idx, tmp);
652 r = -EINVAL;
653 }
654 amdgpu_wb_free(adev, index);
655
656 return r;
657}
658
659/**
660 * sdma_v2_4_ring_test_ib - test an IB on the DMA engine
661 *
662 * @ring: amdgpu_ring structure holding ring information
663 *
664 * Test a simple IB in the DMA ring (VI).
665 * Returns 0 on success, error on failure.
666 */
667static int sdma_v2_4_ring_test_ib(struct amdgpu_ring *ring, long timeout)
668{
669 struct amdgpu_device *adev = ring->adev;
670 struct amdgpu_ib ib;
671 struct dma_fence *f = NULL;
672 unsigned index;
673 u32 tmp = 0;
674 u64 gpu_addr;
675 long r;
676
677 r = amdgpu_wb_get(adev, &index);
678 if (r) {
679 dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
680 return r;
681 }
682
683 gpu_addr = adev->wb.gpu_addr + (index * 4);
684 tmp = 0xCAFEDEAD;
685 adev->wb.wb[index] = cpu_to_le32(tmp);
686 memset(&ib, 0, sizeof(ib));
687 r = amdgpu_ib_get(adev, NULL, 256, &ib);
688 if (r) {
689 DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
690 goto err0;
691 }
692
693 ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
694 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
695 ib.ptr[1] = lower_32_bits(gpu_addr);
696 ib.ptr[2] = upper_32_bits(gpu_addr);
697 ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1);
698 ib.ptr[4] = 0xDEADBEEF;
699 ib.ptr[5] = SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
700 ib.ptr[6] = SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
701 ib.ptr[7] = SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
702 ib.length_dw = 8;
703
704 r = amdgpu_ib_schedule(ring, 1, &ib, NULL, NULL, &f);
705 if (r)
706 goto err1;
707
708 r = dma_fence_wait_timeout(f, false, timeout);
709 if (r == 0) {
710 DRM_ERROR("amdgpu: IB test timed out\n");
711 r = -ETIMEDOUT;
712 goto err1;
713 } else if (r < 0) {
714 DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
715 goto err1;
716 }
717 tmp = le32_to_cpu(adev->wb.wb[index]);
718 if (tmp == 0xDEADBEEF) {
719 DRM_INFO("ib test on ring %d succeeded\n", ring->idx);
720 r = 0;
721 } else {
722 DRM_ERROR("amdgpu: ib test failed (0x%08X)\n", tmp);
723 r = -EINVAL;
724 }
725
726err1:
727 amdgpu_ib_free(adev, &ib, NULL);
728 dma_fence_put(f);
729err0:
730 amdgpu_wb_free(adev, index);
731 return r;
732}
733
734/**
735 * sdma_v2_4_vm_copy_pte - update PTEs by copying them from the GART
736 *
737 * @ib: indirect buffer to fill with commands
738 * @pe: addr of the page entry
739 * @src: src addr to copy from
740 * @count: number of page entries to update
741 *
742 * Update PTEs by copying them from the GART using sDMA (CIK).
743 */
744static void sdma_v2_4_vm_copy_pte(struct amdgpu_ib *ib,
745 uint64_t pe, uint64_t src,
746 unsigned count)
747{
748 unsigned bytes = count * 8;
749
750 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
751 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
752 ib->ptr[ib->length_dw++] = bytes;
753 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
754 ib->ptr[ib->length_dw++] = lower_32_bits(src);
755 ib->ptr[ib->length_dw++] = upper_32_bits(src);
756 ib->ptr[ib->length_dw++] = lower_32_bits(pe);
757 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
758}
759
760/**
761 * sdma_v2_4_vm_write_pte - update PTEs by writing them manually
762 *
763 * @ib: indirect buffer to fill with commands
764 * @pe: addr of the page entry
765 * @value: dst addr to write into pe
766 * @count: number of page entries to update
767 * @incr: increase next addr by incr bytes
768 *
769 * Update PTEs by writing them manually using sDMA (CIK).
770 */
771static void sdma_v2_4_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
772 uint64_t value, unsigned count,
773 uint32_t incr)
774{
775 unsigned ndw = count * 2;
776
777 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
778 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
779 ib->ptr[ib->length_dw++] = pe;
780 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
781 ib->ptr[ib->length_dw++] = ndw;
782 for (; ndw > 0; ndw -= 2) {
783 ib->ptr[ib->length_dw++] = lower_32_bits(value);
784 ib->ptr[ib->length_dw++] = upper_32_bits(value);
785 value += incr;
786 }
787}
788
789/**
790 * sdma_v2_4_vm_set_pte_pde - update the page tables using sDMA
791 *
792 * @ib: indirect buffer to fill with commands
793 * @pe: addr of the page entry
794 * @addr: dst addr to write into pe
795 * @count: number of page entries to update
796 * @incr: increase next addr by incr bytes
797 * @flags: access flags
798 *
799 * Update the page tables using sDMA (CIK).
800 */
801static void sdma_v2_4_vm_set_pte_pde(struct amdgpu_ib *ib, uint64_t pe,
802 uint64_t addr, unsigned count,
803 uint32_t incr, uint32_t flags)
804{
805 /* for physically contiguous pages (vram) */
806 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_GEN_PTEPDE);
807 ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
808 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
809 ib->ptr[ib->length_dw++] = flags; /* mask */
810 ib->ptr[ib->length_dw++] = 0;
811 ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
812 ib->ptr[ib->length_dw++] = upper_32_bits(addr);
813 ib->ptr[ib->length_dw++] = incr; /* increment size */
814 ib->ptr[ib->length_dw++] = 0;
815 ib->ptr[ib->length_dw++] = count; /* number of entries */
816}
817
818/**
819 * sdma_v2_4_ring_pad_ib - pad the IB to the required number of dw
820 *
821 * @ib: indirect buffer to fill with padding
822 *
823 */
824static void sdma_v2_4_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
825{
826 struct amdgpu_sdma_instance *sdma = amdgpu_get_sdma_instance(ring);
827 u32 pad_count;
828 int i;
829
830 pad_count = (8 - (ib->length_dw & 0x7)) % 8;
831 for (i = 0; i < pad_count; i++)
832 if (sdma && sdma->burst_nop && (i == 0))
833 ib->ptr[ib->length_dw++] =
834 SDMA_PKT_HEADER_OP(SDMA_OP_NOP) |
835 SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
836 else
837 ib->ptr[ib->length_dw++] =
838 SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
839}
840
841/**
842 * sdma_v2_4_ring_emit_pipeline_sync - sync the pipeline
843 *
844 * @ring: amdgpu_ring pointer
845 *
846 * Make sure all previous operations are completed (CIK).
847 */
848static void sdma_v2_4_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
849{
850 uint32_t seq = ring->fence_drv.sync_seq;
851 uint64_t addr = ring->fence_drv.gpu_addr;
852
853 /* wait for idle */
854 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
855 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
856 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
857 SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
858 amdgpu_ring_write(ring, addr & 0xfffffffc);
859 amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
860 amdgpu_ring_write(ring, seq); /* reference */
861 amdgpu_ring_write(ring, 0xfffffff); /* mask */
862 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
863 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
864}
865
866/**
867 * sdma_v2_4_ring_emit_vm_flush - cik vm flush using sDMA
868 *
869 * @ring: amdgpu_ring pointer
870 * @vm: amdgpu_vm pointer
871 *
872 * Update the page table base and flush the VM TLB
873 * using sDMA (VI).
874 */
875static void sdma_v2_4_ring_emit_vm_flush(struct amdgpu_ring *ring,
876 unsigned vm_id, uint64_t pd_addr)
877{
878 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
879 SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
880 if (vm_id < 8) {
881 amdgpu_ring_write(ring, (mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + vm_id));
882 } else {
883 amdgpu_ring_write(ring, (mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vm_id - 8));
884 }
885 amdgpu_ring_write(ring, pd_addr >> 12);
886
887 /* flush TLB */
888 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
889 SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
890 amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST);
891 amdgpu_ring_write(ring, 1 << vm_id);
892
893 /* wait for flush */
894 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
895 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
896 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(0)); /* always */
897 amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST << 2);
898 amdgpu_ring_write(ring, 0);
899 amdgpu_ring_write(ring, 0); /* reference */
900 amdgpu_ring_write(ring, 0); /* mask */
901 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
902 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
903}
904
905static int sdma_v2_4_early_init(void *handle)
906{
907 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
908
909 adev->sdma.num_instances = SDMA_MAX_INSTANCE;
910
911 sdma_v2_4_set_ring_funcs(adev);
912 sdma_v2_4_set_buffer_funcs(adev);
913 sdma_v2_4_set_vm_pte_funcs(adev);
914 sdma_v2_4_set_irq_funcs(adev);
915
916 return 0;
917}
918
919static int sdma_v2_4_sw_init(void *handle)
920{
921 struct amdgpu_ring *ring;
922 int r, i;
923 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
924
925 /* SDMA trap event */
926 r = amdgpu_irq_add_id(adev, 224, &adev->sdma.trap_irq);
927 if (r)
928 return r;
929
930 /* SDMA Privileged inst */
931 r = amdgpu_irq_add_id(adev, 241, &adev->sdma.illegal_inst_irq);
932 if (r)
933 return r;
934
935 /* SDMA Privileged inst */
936 r = amdgpu_irq_add_id(adev, 247, &adev->sdma.illegal_inst_irq);
937 if (r)
938 return r;
939
940 r = sdma_v2_4_init_microcode(adev);
941 if (r) {
942 DRM_ERROR("Failed to load sdma firmware!\n");
943 return r;
944 }
945
946 for (i = 0; i < adev->sdma.num_instances; i++) {
947 ring = &adev->sdma.instance[i].ring;
948 ring->ring_obj = NULL;
949 ring->use_doorbell = false;
950 sprintf(ring->name, "sdma%d", i);
951 r = amdgpu_ring_init(adev, ring, 1024,
952 &adev->sdma.trap_irq,
953 (i == 0) ?
954 AMDGPU_SDMA_IRQ_TRAP0 :
955 AMDGPU_SDMA_IRQ_TRAP1);
956 if (r)
957 return r;
958 }
959
960 return r;
961}
962
963static int sdma_v2_4_sw_fini(void *handle)
964{
965 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
966 int i;
967
968 for (i = 0; i < adev->sdma.num_instances; i++)
969 amdgpu_ring_fini(&adev->sdma.instance[i].ring);
970
971 sdma_v2_4_free_microcode(adev);
972 return 0;
973}
974
975static int sdma_v2_4_hw_init(void *handle)
976{
977 int r;
978 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
979
980 sdma_v2_4_init_golden_registers(adev);
981
982 r = sdma_v2_4_start(adev);
983 if (r)
984 return r;
985
986 return r;
987}
988
989static int sdma_v2_4_hw_fini(void *handle)
990{
991 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
992
993 sdma_v2_4_enable(adev, false);
994
995 return 0;
996}
997
998static int sdma_v2_4_suspend(void *handle)
999{
1000 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1001
1002 return sdma_v2_4_hw_fini(adev);
1003}
1004
1005static int sdma_v2_4_resume(void *handle)
1006{
1007 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1008
1009 return sdma_v2_4_hw_init(adev);
1010}
1011
1012static bool sdma_v2_4_is_idle(void *handle)
1013{
1014 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1015 u32 tmp = RREG32(mmSRBM_STATUS2);
1016
1017 if (tmp & (SRBM_STATUS2__SDMA_BUSY_MASK |
1018 SRBM_STATUS2__SDMA1_BUSY_MASK))
1019 return false;
1020
1021 return true;
1022}
1023
1024static int sdma_v2_4_wait_for_idle(void *handle)
1025{
1026 unsigned i;
1027 u32 tmp;
1028 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1029
1030 for (i = 0; i < adev->usec_timeout; i++) {
1031 tmp = RREG32(mmSRBM_STATUS2) & (SRBM_STATUS2__SDMA_BUSY_MASK |
1032 SRBM_STATUS2__SDMA1_BUSY_MASK);
1033
1034 if (!tmp)
1035 return 0;
1036 udelay(1);
1037 }
1038 return -ETIMEDOUT;
1039}
1040
1041static int sdma_v2_4_soft_reset(void *handle)
1042{
1043 u32 srbm_soft_reset = 0;
1044 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1045 u32 tmp = RREG32(mmSRBM_STATUS2);
1046
1047 if (tmp & SRBM_STATUS2__SDMA_BUSY_MASK) {
1048 /* sdma0 */
1049 tmp = RREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET);
1050 tmp = REG_SET_FIELD(tmp, SDMA0_F32_CNTL, HALT, 0);
1051 WREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET, tmp);
1052 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA_MASK;
1053 }
1054 if (tmp & SRBM_STATUS2__SDMA1_BUSY_MASK) {
1055 /* sdma1 */
1056 tmp = RREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET);
1057 tmp = REG_SET_FIELD(tmp, SDMA0_F32_CNTL, HALT, 0);
1058 WREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET, tmp);
1059 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA1_MASK;
1060 }
1061
1062 if (srbm_soft_reset) {
1063 tmp = RREG32(mmSRBM_SOFT_RESET);
1064 tmp |= srbm_soft_reset;
1065 dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
1066 WREG32(mmSRBM_SOFT_RESET, tmp);
1067 tmp = RREG32(mmSRBM_SOFT_RESET);
1068
1069 udelay(50);
1070
1071 tmp &= ~srbm_soft_reset;
1072 WREG32(mmSRBM_SOFT_RESET, tmp);
1073 tmp = RREG32(mmSRBM_SOFT_RESET);
1074
1075 /* Wait a little for things to settle down */
1076 udelay(50);
1077 }
1078
1079 return 0;
1080}
1081
1082static int sdma_v2_4_set_trap_irq_state(struct amdgpu_device *adev,
1083 struct amdgpu_irq_src *src,
1084 unsigned type,
1085 enum amdgpu_interrupt_state state)
1086{
1087 u32 sdma_cntl;
1088
1089 switch (type) {
1090 case AMDGPU_SDMA_IRQ_TRAP0:
1091 switch (state) {
1092 case AMDGPU_IRQ_STATE_DISABLE:
1093 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1094 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 0);
1095 WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1096 break;
1097 case AMDGPU_IRQ_STATE_ENABLE:
1098 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1099 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 1);
1100 WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1101 break;
1102 default:
1103 break;
1104 }
1105 break;
1106 case AMDGPU_SDMA_IRQ_TRAP1:
1107 switch (state) {
1108 case AMDGPU_IRQ_STATE_DISABLE:
1109 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1110 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 0);
1111 WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1112 break;
1113 case AMDGPU_IRQ_STATE_ENABLE:
1114 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1115 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 1);
1116 WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1117 break;
1118 default:
1119 break;
1120 }
1121 break;
1122 default:
1123 break;
1124 }
1125 return 0;
1126}
1127
1128static int sdma_v2_4_process_trap_irq(struct amdgpu_device *adev,
1129 struct amdgpu_irq_src *source,
1130 struct amdgpu_iv_entry *entry)
1131{
1132 u8 instance_id, queue_id;
1133
1134 instance_id = (entry->ring_id & 0x3) >> 0;
1135 queue_id = (entry->ring_id & 0xc) >> 2;
1136 DRM_DEBUG("IH: SDMA trap\n");
1137 switch (instance_id) {
1138 case 0:
1139 switch (queue_id) {
1140 case 0:
1141 amdgpu_fence_process(&adev->sdma.instance[0].ring);
1142 break;
1143 case 1:
1144 /* XXX compute */
1145 break;
1146 case 2:
1147 /* XXX compute */
1148 break;
1149 }
1150 break;
1151 case 1:
1152 switch (queue_id) {
1153 case 0:
1154 amdgpu_fence_process(&adev->sdma.instance[1].ring);
1155 break;
1156 case 1:
1157 /* XXX compute */
1158 break;
1159 case 2:
1160 /* XXX compute */
1161 break;
1162 }
1163 break;
1164 }
1165 return 0;
1166}
1167
1168static int sdma_v2_4_process_illegal_inst_irq(struct amdgpu_device *adev,
1169 struct amdgpu_irq_src *source,
1170 struct amdgpu_iv_entry *entry)
1171{
1172 DRM_ERROR("Illegal instruction in SDMA command stream\n");
1173 schedule_work(&adev->reset_work);
1174 return 0;
1175}
1176
1177static int sdma_v2_4_set_clockgating_state(void *handle,
1178 enum amd_clockgating_state state)
1179{
1180 /* XXX handled via the smc on VI */
1181 return 0;
1182}
1183
1184static int sdma_v2_4_set_powergating_state(void *handle,
1185 enum amd_powergating_state state)
1186{
1187 return 0;
1188}
1189
1190static const struct amd_ip_funcs sdma_v2_4_ip_funcs = {
1191 .name = "sdma_v2_4",
1192 .early_init = sdma_v2_4_early_init,
1193 .late_init = NULL,
1194 .sw_init = sdma_v2_4_sw_init,
1195 .sw_fini = sdma_v2_4_sw_fini,
1196 .hw_init = sdma_v2_4_hw_init,
1197 .hw_fini = sdma_v2_4_hw_fini,
1198 .suspend = sdma_v2_4_suspend,
1199 .resume = sdma_v2_4_resume,
1200 .is_idle = sdma_v2_4_is_idle,
1201 .wait_for_idle = sdma_v2_4_wait_for_idle,
1202 .soft_reset = sdma_v2_4_soft_reset,
1203 .set_clockgating_state = sdma_v2_4_set_clockgating_state,
1204 .set_powergating_state = sdma_v2_4_set_powergating_state,
1205};
1206
1207static const struct amdgpu_ring_funcs sdma_v2_4_ring_funcs = {
1208 .type = AMDGPU_RING_TYPE_SDMA,
1209 .align_mask = 0xf,
1210 .nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
1211 .get_rptr = sdma_v2_4_ring_get_rptr,
1212 .get_wptr = sdma_v2_4_ring_get_wptr,
1213 .set_wptr = sdma_v2_4_ring_set_wptr,
1214 .emit_frame_size =
1215 6 + /* sdma_v2_4_ring_emit_hdp_flush */
1216 3 + /* sdma_v2_4_ring_emit_hdp_invalidate */
1217 6 + /* sdma_v2_4_ring_emit_pipeline_sync */
1218 12 + /* sdma_v2_4_ring_emit_vm_flush */
1219 10 + 10 + 10, /* sdma_v2_4_ring_emit_fence x3 for user fence, vm fence */
1220 .emit_ib_size = 7 + 6, /* sdma_v2_4_ring_emit_ib */
1221 .emit_ib = sdma_v2_4_ring_emit_ib,
1222 .emit_fence = sdma_v2_4_ring_emit_fence,
1223 .emit_pipeline_sync = sdma_v2_4_ring_emit_pipeline_sync,
1224 .emit_vm_flush = sdma_v2_4_ring_emit_vm_flush,
1225 .emit_hdp_flush = sdma_v2_4_ring_emit_hdp_flush,
1226 .emit_hdp_invalidate = sdma_v2_4_ring_emit_hdp_invalidate,
1227 .test_ring = sdma_v2_4_ring_test_ring,
1228 .test_ib = sdma_v2_4_ring_test_ib,
1229 .insert_nop = sdma_v2_4_ring_insert_nop,
1230 .pad_ib = sdma_v2_4_ring_pad_ib,
1231};
1232
1233static void sdma_v2_4_set_ring_funcs(struct amdgpu_device *adev)
1234{
1235 int i;
1236
1237 for (i = 0; i < adev->sdma.num_instances; i++)
1238 adev->sdma.instance[i].ring.funcs = &sdma_v2_4_ring_funcs;
1239}
1240
1241static const struct amdgpu_irq_src_funcs sdma_v2_4_trap_irq_funcs = {
1242 .set = sdma_v2_4_set_trap_irq_state,
1243 .process = sdma_v2_4_process_trap_irq,
1244};
1245
1246static const struct amdgpu_irq_src_funcs sdma_v2_4_illegal_inst_irq_funcs = {
1247 .process = sdma_v2_4_process_illegal_inst_irq,
1248};
1249
1250static void sdma_v2_4_set_irq_funcs(struct amdgpu_device *adev)
1251{
1252 adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
1253 adev->sdma.trap_irq.funcs = &sdma_v2_4_trap_irq_funcs;
1254 adev->sdma.illegal_inst_irq.funcs = &sdma_v2_4_illegal_inst_irq_funcs;
1255}
1256
1257/**
1258 * sdma_v2_4_emit_copy_buffer - copy buffer using the sDMA engine
1259 *
1260 * @ring: amdgpu_ring structure holding ring information
1261 * @src_offset: src GPU address
1262 * @dst_offset: dst GPU address
1263 * @byte_count: number of bytes to xfer
1264 *
1265 * Copy GPU buffers using the DMA engine (VI).
1266 * Used by the amdgpu ttm implementation to move pages if
1267 * registered as the asic copy callback.
1268 */
1269static void sdma_v2_4_emit_copy_buffer(struct amdgpu_ib *ib,
1270 uint64_t src_offset,
1271 uint64_t dst_offset,
1272 uint32_t byte_count)
1273{
1274 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1275 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1276 ib->ptr[ib->length_dw++] = byte_count;
1277 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1278 ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1279 ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1280 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1281 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1282}
1283
1284/**
1285 * sdma_v2_4_emit_fill_buffer - fill buffer using the sDMA engine
1286 *
1287 * @ring: amdgpu_ring structure holding ring information
1288 * @src_data: value to write to buffer
1289 * @dst_offset: dst GPU address
1290 * @byte_count: number of bytes to xfer
1291 *
1292 * Fill GPU buffers using the DMA engine (VI).
1293 */
1294static void sdma_v2_4_emit_fill_buffer(struct amdgpu_ib *ib,
1295 uint32_t src_data,
1296 uint64_t dst_offset,
1297 uint32_t byte_count)
1298{
1299 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL);
1300 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1301 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1302 ib->ptr[ib->length_dw++] = src_data;
1303 ib->ptr[ib->length_dw++] = byte_count;
1304}
1305
1306static const struct amdgpu_buffer_funcs sdma_v2_4_buffer_funcs = {
1307 .copy_max_bytes = 0x1fffff,
1308 .copy_num_dw = 7,
1309 .emit_copy_buffer = sdma_v2_4_emit_copy_buffer,
1310
1311 .fill_max_bytes = 0x1fffff,
1312 .fill_num_dw = 7,
1313 .emit_fill_buffer = sdma_v2_4_emit_fill_buffer,
1314};
1315
1316static void sdma_v2_4_set_buffer_funcs(struct amdgpu_device *adev)
1317{
1318 if (adev->mman.buffer_funcs == NULL) {
1319 adev->mman.buffer_funcs = &sdma_v2_4_buffer_funcs;
1320 adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1321 }
1322}
1323
1324static const struct amdgpu_vm_pte_funcs sdma_v2_4_vm_pte_funcs = {
1325 .copy_pte = sdma_v2_4_vm_copy_pte,
1326 .write_pte = sdma_v2_4_vm_write_pte,
1327 .set_pte_pde = sdma_v2_4_vm_set_pte_pde,
1328};
1329
1330static void sdma_v2_4_set_vm_pte_funcs(struct amdgpu_device *adev)
1331{
1332 unsigned i;
1333
1334 if (adev->vm_manager.vm_pte_funcs == NULL) {
1335 adev->vm_manager.vm_pte_funcs = &sdma_v2_4_vm_pte_funcs;
1336 for (i = 0; i < adev->sdma.num_instances; i++)
1337 adev->vm_manager.vm_pte_rings[i] =
1338 &adev->sdma.instance[i].ring;
1339
1340 adev->vm_manager.vm_pte_num_rings = adev->sdma.num_instances;
1341 }
1342}
1343
1344const struct amdgpu_ip_block_version sdma_v2_4_ip_block =
1345{
1346 .type = AMD_IP_BLOCK_TYPE_SDMA,
1347 .major = 2,
1348 .minor = 4,
1349 .rev = 0,
1350 .funcs = &sdma_v2_4_ip_funcs,
1351};