Loading...
1/*
2 * Copyright 2013 Advanced Micro Devices, Inc.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
21 *
22 * Authors: Alex Deucher
23 */
24#include <linux/firmware.h>
25#include <drm/drmP.h>
26#include "amdgpu.h"
27#include "amdgpu_ucode.h"
28#include "amdgpu_trace.h"
29#include "cikd.h"
30#include "cik.h"
31
32#include "bif/bif_4_1_d.h"
33#include "bif/bif_4_1_sh_mask.h"
34
35#include "gca/gfx_7_2_d.h"
36#include "gca/gfx_7_2_enum.h"
37#include "gca/gfx_7_2_sh_mask.h"
38
39#include "gmc/gmc_7_1_d.h"
40#include "gmc/gmc_7_1_sh_mask.h"
41
42#include "oss/oss_2_0_d.h"
43#include "oss/oss_2_0_sh_mask.h"
44
45static const u32 sdma_offsets[SDMA_MAX_INSTANCE] =
46{
47 SDMA0_REGISTER_OFFSET,
48 SDMA1_REGISTER_OFFSET
49};
50
51static void cik_sdma_set_ring_funcs(struct amdgpu_device *adev);
52static void cik_sdma_set_irq_funcs(struct amdgpu_device *adev);
53static void cik_sdma_set_buffer_funcs(struct amdgpu_device *adev);
54static void cik_sdma_set_vm_pte_funcs(struct amdgpu_device *adev);
55static int cik_sdma_soft_reset(void *handle);
56
57MODULE_FIRMWARE("radeon/bonaire_sdma.bin");
58MODULE_FIRMWARE("radeon/bonaire_sdma1.bin");
59MODULE_FIRMWARE("radeon/hawaii_sdma.bin");
60MODULE_FIRMWARE("radeon/hawaii_sdma1.bin");
61MODULE_FIRMWARE("radeon/kaveri_sdma.bin");
62MODULE_FIRMWARE("radeon/kaveri_sdma1.bin");
63MODULE_FIRMWARE("radeon/kabini_sdma.bin");
64MODULE_FIRMWARE("radeon/kabini_sdma1.bin");
65MODULE_FIRMWARE("radeon/mullins_sdma.bin");
66MODULE_FIRMWARE("radeon/mullins_sdma1.bin");
67
68u32 amdgpu_cik_gpu_check_soft_reset(struct amdgpu_device *adev);
69
70
71static void cik_sdma_free_microcode(struct amdgpu_device *adev)
72{
73 int i;
74 for (i = 0; i < adev->sdma.num_instances; i++) {
75 release_firmware(adev->sdma.instance[i].fw);
76 adev->sdma.instance[i].fw = NULL;
77 }
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
97/**
98 * cik_sdma_init_microcode - load ucode images from disk
99 *
100 * @adev: amdgpu_device pointer
101 *
102 * Use the firmware interface to load the ucode images into
103 * the driver (not loaded into hw).
104 * Returns 0 on success, error on failure.
105 */
106static int cik_sdma_init_microcode(struct amdgpu_device *adev)
107{
108 const char *chip_name;
109 char fw_name[30];
110 int err = 0, i;
111
112 DRM_DEBUG("\n");
113
114 switch (adev->asic_type) {
115 case CHIP_BONAIRE:
116 chip_name = "bonaire";
117 break;
118 case CHIP_HAWAII:
119 chip_name = "hawaii";
120 break;
121 case CHIP_KAVERI:
122 chip_name = "kaveri";
123 break;
124 case CHIP_KABINI:
125 chip_name = "kabini";
126 break;
127 case CHIP_MULLINS:
128 chip_name = "mullins";
129 break;
130 default: BUG();
131 }
132
133 for (i = 0; i < adev->sdma.num_instances; i++) {
134 if (i == 0)
135 snprintf(fw_name, sizeof(fw_name), "radeon/%s_sdma.bin", chip_name);
136 else
137 snprintf(fw_name, sizeof(fw_name), "radeon/%s_sdma1.bin", chip_name);
138 err = request_firmware(&adev->sdma.instance[i].fw, fw_name, adev->dev);
139 if (err)
140 goto out;
141 err = amdgpu_ucode_validate(adev->sdma.instance[i].fw);
142 }
143out:
144 if (err) {
145 pr_err("cik_sdma: Failed to load firmware \"%s\"\n", fw_name);
146 for (i = 0; i < adev->sdma.num_instances; i++) {
147 release_firmware(adev->sdma.instance[i].fw);
148 adev->sdma.instance[i].fw = NULL;
149 }
150 }
151 return err;
152}
153
154/**
155 * cik_sdma_ring_get_rptr - get the current read pointer
156 *
157 * @ring: amdgpu ring pointer
158 *
159 * Get the current rptr from the hardware (CIK+).
160 */
161static uint64_t cik_sdma_ring_get_rptr(struct amdgpu_ring *ring)
162{
163 u32 rptr;
164
165 rptr = ring->adev->wb.wb[ring->rptr_offs];
166
167 return (rptr & 0x3fffc) >> 2;
168}
169
170/**
171 * cik_sdma_ring_get_wptr - get the current write pointer
172 *
173 * @ring: amdgpu ring pointer
174 *
175 * Get the current wptr from the hardware (CIK+).
176 */
177static uint64_t cik_sdma_ring_get_wptr(struct amdgpu_ring *ring)
178{
179 struct amdgpu_device *adev = ring->adev;
180 u32 me = (ring == &adev->sdma.instance[0].ring) ? 0 : 1;
181
182 return (RREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[me]) & 0x3fffc) >> 2;
183}
184
185/**
186 * cik_sdma_ring_set_wptr - commit the write pointer
187 *
188 * @ring: amdgpu ring pointer
189 *
190 * Write the wptr back to the hardware (CIK+).
191 */
192static void cik_sdma_ring_set_wptr(struct amdgpu_ring *ring)
193{
194 struct amdgpu_device *adev = ring->adev;
195 u32 me = (ring == &adev->sdma.instance[0].ring) ? 0 : 1;
196
197 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[me],
198 (lower_32_bits(ring->wptr) << 2) & 0x3fffc);
199}
200
201static void cik_sdma_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
202{
203 struct amdgpu_sdma_instance *sdma = amdgpu_get_sdma_instance(ring);
204 int i;
205
206 for (i = 0; i < count; i++)
207 if (sdma && sdma->burst_nop && (i == 0))
208 amdgpu_ring_write(ring, ring->funcs->nop |
209 SDMA_NOP_COUNT(count - 1));
210 else
211 amdgpu_ring_write(ring, ring->funcs->nop);
212}
213
214/**
215 * cik_sdma_ring_emit_ib - Schedule an IB on the DMA engine
216 *
217 * @ring: amdgpu ring pointer
218 * @ib: IB object to schedule
219 *
220 * Schedule an IB in the DMA ring (CIK).
221 */
222static void cik_sdma_ring_emit_ib(struct amdgpu_ring *ring,
223 struct amdgpu_ib *ib,
224 unsigned vmid, bool ctx_switch)
225{
226 u32 extra_bits = vmid & 0xf;
227
228 /* IB packet must end on a 8 DW boundary */
229 cik_sdma_ring_insert_nop(ring, (12 - (lower_32_bits(ring->wptr) & 7)) % 8);
230
231 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_INDIRECT_BUFFER, 0, extra_bits));
232 amdgpu_ring_write(ring, ib->gpu_addr & 0xffffffe0); /* base must be 32 byte aligned */
233 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr) & 0xffffffff);
234 amdgpu_ring_write(ring, ib->length_dw);
235
236}
237
238/**
239 * cik_sdma_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
240 *
241 * @ring: amdgpu ring pointer
242 *
243 * Emit an hdp flush packet on the requested DMA ring.
244 */
245static void cik_sdma_ring_emit_hdp_flush(struct amdgpu_ring *ring)
246{
247 u32 extra_bits = (SDMA_POLL_REG_MEM_EXTRA_OP(1) |
248 SDMA_POLL_REG_MEM_EXTRA_FUNC(3)); /* == */
249 u32 ref_and_mask;
250
251 if (ring == &ring->adev->sdma.instance[0].ring)
252 ref_and_mask = GPU_HDP_FLUSH_DONE__SDMA0_MASK;
253 else
254 ref_and_mask = GPU_HDP_FLUSH_DONE__SDMA1_MASK;
255
256 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0, extra_bits));
257 amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_DONE << 2);
258 amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_REQ << 2);
259 amdgpu_ring_write(ring, ref_and_mask); /* reference */
260 amdgpu_ring_write(ring, ref_and_mask); /* mask */
261 amdgpu_ring_write(ring, (0xfff << 16) | 10); /* retry count, poll interval */
262}
263
264/**
265 * cik_sdma_ring_emit_fence - emit a fence on the DMA ring
266 *
267 * @ring: amdgpu ring pointer
268 * @fence: amdgpu fence object
269 *
270 * Add a DMA fence packet to the ring to write
271 * the fence seq number and DMA trap packet to generate
272 * an interrupt if needed (CIK).
273 */
274static void cik_sdma_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
275 unsigned flags)
276{
277 bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
278 /* write the fence */
279 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_FENCE, 0, 0));
280 amdgpu_ring_write(ring, lower_32_bits(addr));
281 amdgpu_ring_write(ring, upper_32_bits(addr));
282 amdgpu_ring_write(ring, lower_32_bits(seq));
283
284 /* optionally write high bits as well */
285 if (write64bit) {
286 addr += 4;
287 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_FENCE, 0, 0));
288 amdgpu_ring_write(ring, lower_32_bits(addr));
289 amdgpu_ring_write(ring, upper_32_bits(addr));
290 amdgpu_ring_write(ring, upper_32_bits(seq));
291 }
292
293 /* generate an interrupt */
294 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_TRAP, 0, 0));
295}
296
297/**
298 * cik_sdma_gfx_stop - stop the gfx async dma engines
299 *
300 * @adev: amdgpu_device pointer
301 *
302 * Stop the gfx async dma ring buffers (CIK).
303 */
304static void cik_sdma_gfx_stop(struct amdgpu_device *adev)
305{
306 struct amdgpu_ring *sdma0 = &adev->sdma.instance[0].ring;
307 struct amdgpu_ring *sdma1 = &adev->sdma.instance[1].ring;
308 u32 rb_cntl;
309 int i;
310
311 if ((adev->mman.buffer_funcs_ring == sdma0) ||
312 (adev->mman.buffer_funcs_ring == sdma1))
313 amdgpu_ttm_set_buffer_funcs_status(adev, false);
314
315 for (i = 0; i < adev->sdma.num_instances; i++) {
316 rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
317 rb_cntl &= ~SDMA0_GFX_RB_CNTL__RB_ENABLE_MASK;
318 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
319 WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], 0);
320 }
321 sdma0->ready = false;
322 sdma1->ready = false;
323}
324
325/**
326 * cik_sdma_rlc_stop - stop the compute async dma engines
327 *
328 * @adev: amdgpu_device pointer
329 *
330 * Stop the compute async dma queues (CIK).
331 */
332static void cik_sdma_rlc_stop(struct amdgpu_device *adev)
333{
334 /* XXX todo */
335}
336
337/**
338 * cik_ctx_switch_enable - stop the async dma engines context switch
339 *
340 * @adev: amdgpu_device pointer
341 * @enable: enable/disable the DMA MEs context switch.
342 *
343 * Halt or unhalt the async dma engines context switch (VI).
344 */
345static void cik_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
346{
347 u32 f32_cntl, phase_quantum = 0;
348 int i;
349
350 if (amdgpu_sdma_phase_quantum) {
351 unsigned value = amdgpu_sdma_phase_quantum;
352 unsigned unit = 0;
353
354 while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
355 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
356 value = (value + 1) >> 1;
357 unit++;
358 }
359 if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
360 SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
361 value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
362 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
363 unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
364 SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
365 WARN_ONCE(1,
366 "clamping sdma_phase_quantum to %uK clock cycles\n",
367 value << unit);
368 }
369 phase_quantum =
370 value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
371 unit << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
372 }
373
374 for (i = 0; i < adev->sdma.num_instances; i++) {
375 f32_cntl = RREG32(mmSDMA0_CNTL + sdma_offsets[i]);
376 if (enable) {
377 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
378 AUTO_CTXSW_ENABLE, 1);
379 if (amdgpu_sdma_phase_quantum) {
380 WREG32(mmSDMA0_PHASE0_QUANTUM + sdma_offsets[i],
381 phase_quantum);
382 WREG32(mmSDMA0_PHASE1_QUANTUM + sdma_offsets[i],
383 phase_quantum);
384 }
385 } else {
386 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
387 AUTO_CTXSW_ENABLE, 0);
388 }
389
390 WREG32(mmSDMA0_CNTL + sdma_offsets[i], f32_cntl);
391 }
392}
393
394/**
395 * cik_sdma_enable - stop the async dma engines
396 *
397 * @adev: amdgpu_device pointer
398 * @enable: enable/disable the DMA MEs.
399 *
400 * Halt or unhalt the async dma engines (CIK).
401 */
402static void cik_sdma_enable(struct amdgpu_device *adev, bool enable)
403{
404 u32 me_cntl;
405 int i;
406
407 if (!enable) {
408 cik_sdma_gfx_stop(adev);
409 cik_sdma_rlc_stop(adev);
410 }
411
412 for (i = 0; i < adev->sdma.num_instances; i++) {
413 me_cntl = RREG32(mmSDMA0_F32_CNTL + sdma_offsets[i]);
414 if (enable)
415 me_cntl &= ~SDMA0_F32_CNTL__HALT_MASK;
416 else
417 me_cntl |= SDMA0_F32_CNTL__HALT_MASK;
418 WREG32(mmSDMA0_F32_CNTL + sdma_offsets[i], me_cntl);
419 }
420}
421
422/**
423 * cik_sdma_gfx_resume - setup and start the async dma engines
424 *
425 * @adev: amdgpu_device pointer
426 *
427 * Set up the gfx DMA ring buffers and enable them (CIK).
428 * Returns 0 for success, error for failure.
429 */
430static int cik_sdma_gfx_resume(struct amdgpu_device *adev)
431{
432 struct amdgpu_ring *ring;
433 u32 rb_cntl, ib_cntl;
434 u32 rb_bufsz;
435 u32 wb_offset;
436 int i, j, r;
437
438 for (i = 0; i < adev->sdma.num_instances; i++) {
439 ring = &adev->sdma.instance[i].ring;
440 wb_offset = (ring->rptr_offs * 4);
441
442 mutex_lock(&adev->srbm_mutex);
443 for (j = 0; j < 16; j++) {
444 cik_srbm_select(adev, 0, 0, 0, j);
445 /* SDMA GFX */
446 WREG32(mmSDMA0_GFX_VIRTUAL_ADDR + sdma_offsets[i], 0);
447 WREG32(mmSDMA0_GFX_APE1_CNTL + sdma_offsets[i], 0);
448 /* XXX SDMA RLC - todo */
449 }
450 cik_srbm_select(adev, 0, 0, 0, 0);
451 mutex_unlock(&adev->srbm_mutex);
452
453 WREG32(mmSDMA0_TILING_CONFIG + sdma_offsets[i],
454 adev->gfx.config.gb_addr_config & 0x70);
455
456 WREG32(mmSDMA0_SEM_INCOMPLETE_TIMER_CNTL + sdma_offsets[i], 0);
457 WREG32(mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
458
459 /* Set ring buffer size in dwords */
460 rb_bufsz = order_base_2(ring->ring_size / 4);
461 rb_cntl = rb_bufsz << 1;
462#ifdef __BIG_ENDIAN
463 rb_cntl |= SDMA0_GFX_RB_CNTL__RB_SWAP_ENABLE_MASK |
464 SDMA0_GFX_RB_CNTL__RPTR_WRITEBACK_SWAP_ENABLE_MASK;
465#endif
466 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
467
468 /* Initialize the ring buffer's read and write pointers */
469 WREG32(mmSDMA0_GFX_RB_RPTR + sdma_offsets[i], 0);
470 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], 0);
471 WREG32(mmSDMA0_GFX_IB_RPTR + sdma_offsets[i], 0);
472 WREG32(mmSDMA0_GFX_IB_OFFSET + sdma_offsets[i], 0);
473
474 /* set the wb address whether it's enabled or not */
475 WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_HI + sdma_offsets[i],
476 upper_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF);
477 WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_LO + sdma_offsets[i],
478 ((adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC));
479
480 rb_cntl |= SDMA0_GFX_RB_CNTL__RPTR_WRITEBACK_ENABLE_MASK;
481
482 WREG32(mmSDMA0_GFX_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
483 WREG32(mmSDMA0_GFX_RB_BASE_HI + sdma_offsets[i], ring->gpu_addr >> 40);
484
485 ring->wptr = 0;
486 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], lower_32_bits(ring->wptr) << 2);
487
488 /* enable DMA RB */
489 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i],
490 rb_cntl | SDMA0_GFX_RB_CNTL__RB_ENABLE_MASK);
491
492 ib_cntl = SDMA0_GFX_IB_CNTL__IB_ENABLE_MASK;
493#ifdef __BIG_ENDIAN
494 ib_cntl |= SDMA0_GFX_IB_CNTL__IB_SWAP_ENABLE_MASK;
495#endif
496 /* enable DMA IBs */
497 WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
498
499 ring->ready = true;
500 }
501
502 cik_sdma_enable(adev, true);
503
504 for (i = 0; i < adev->sdma.num_instances; i++) {
505 ring = &adev->sdma.instance[i].ring;
506 r = amdgpu_ring_test_ring(ring);
507 if (r) {
508 ring->ready = false;
509 return r;
510 }
511
512 if (adev->mman.buffer_funcs_ring == ring)
513 amdgpu_ttm_set_buffer_funcs_status(adev, true);
514 }
515
516 return 0;
517}
518
519/**
520 * cik_sdma_rlc_resume - setup and start the async dma engines
521 *
522 * @adev: amdgpu_device pointer
523 *
524 * Set up the compute DMA queues and enable them (CIK).
525 * Returns 0 for success, error for failure.
526 */
527static int cik_sdma_rlc_resume(struct amdgpu_device *adev)
528{
529 /* XXX todo */
530 return 0;
531}
532
533/**
534 * cik_sdma_load_microcode - load the sDMA ME ucode
535 *
536 * @adev: amdgpu_device pointer
537 *
538 * Loads the sDMA0/1 ucode.
539 * Returns 0 for success, -EINVAL if the ucode is not available.
540 */
541static int cik_sdma_load_microcode(struct amdgpu_device *adev)
542{
543 const struct sdma_firmware_header_v1_0 *hdr;
544 const __le32 *fw_data;
545 u32 fw_size;
546 int i, j;
547
548 /* halt the MEs */
549 cik_sdma_enable(adev, false);
550
551 for (i = 0; i < adev->sdma.num_instances; i++) {
552 if (!adev->sdma.instance[i].fw)
553 return -EINVAL;
554 hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
555 amdgpu_ucode_print_sdma_hdr(&hdr->header);
556 fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
557 adev->sdma.instance[i].fw_version = le32_to_cpu(hdr->header.ucode_version);
558 adev->sdma.instance[i].feature_version = le32_to_cpu(hdr->ucode_feature_version);
559 if (adev->sdma.instance[i].feature_version >= 20)
560 adev->sdma.instance[i].burst_nop = true;
561 fw_data = (const __le32 *)
562 (adev->sdma.instance[i].fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
563 WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], 0);
564 for (j = 0; j < fw_size; j++)
565 WREG32(mmSDMA0_UCODE_DATA + sdma_offsets[i], le32_to_cpup(fw_data++));
566 WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], adev->sdma.instance[i].fw_version);
567 }
568
569 return 0;
570}
571
572/**
573 * cik_sdma_start - setup and start the async dma engines
574 *
575 * @adev: amdgpu_device pointer
576 *
577 * Set up the DMA engines and enable them (CIK).
578 * Returns 0 for success, error for failure.
579 */
580static int cik_sdma_start(struct amdgpu_device *adev)
581{
582 int r;
583
584 r = cik_sdma_load_microcode(adev);
585 if (r)
586 return r;
587
588 /* halt the engine before programing */
589 cik_sdma_enable(adev, false);
590 /* enable sdma ring preemption */
591 cik_ctx_switch_enable(adev, true);
592
593 /* start the gfx rings and rlc compute queues */
594 r = cik_sdma_gfx_resume(adev);
595 if (r)
596 return r;
597 r = cik_sdma_rlc_resume(adev);
598 if (r)
599 return r;
600
601 return 0;
602}
603
604/**
605 * cik_sdma_ring_test_ring - simple async dma engine test
606 *
607 * @ring: amdgpu_ring structure holding ring information
608 *
609 * Test the DMA engine by writing using it to write an
610 * value to memory. (CIK).
611 * Returns 0 for success, error for failure.
612 */
613static int cik_sdma_ring_test_ring(struct amdgpu_ring *ring)
614{
615 struct amdgpu_device *adev = ring->adev;
616 unsigned i;
617 unsigned index;
618 int r;
619 u32 tmp;
620 u64 gpu_addr;
621
622 r = amdgpu_device_wb_get(adev, &index);
623 if (r) {
624 dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
625 return r;
626 }
627
628 gpu_addr = adev->wb.gpu_addr + (index * 4);
629 tmp = 0xCAFEDEAD;
630 adev->wb.wb[index] = cpu_to_le32(tmp);
631
632 r = amdgpu_ring_alloc(ring, 5);
633 if (r) {
634 DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
635 amdgpu_device_wb_free(adev, index);
636 return r;
637 }
638 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_WRITE, SDMA_WRITE_SUB_OPCODE_LINEAR, 0));
639 amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
640 amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
641 amdgpu_ring_write(ring, 1); /* number of DWs to follow */
642 amdgpu_ring_write(ring, 0xDEADBEEF);
643 amdgpu_ring_commit(ring);
644
645 for (i = 0; i < adev->usec_timeout; i++) {
646 tmp = le32_to_cpu(adev->wb.wb[index]);
647 if (tmp == 0xDEADBEEF)
648 break;
649 DRM_UDELAY(1);
650 }
651
652 if (i < adev->usec_timeout) {
653 DRM_DEBUG("ring test on %d succeeded in %d usecs\n", ring->idx, i);
654 } else {
655 DRM_ERROR("amdgpu: ring %d test failed (0x%08X)\n",
656 ring->idx, tmp);
657 r = -EINVAL;
658 }
659 amdgpu_device_wb_free(adev, index);
660
661 return r;
662}
663
664/**
665 * cik_sdma_ring_test_ib - test an IB on the DMA engine
666 *
667 * @ring: amdgpu_ring structure holding ring information
668 *
669 * Test a simple IB in the DMA ring (CIK).
670 * Returns 0 on success, error on failure.
671 */
672static int cik_sdma_ring_test_ib(struct amdgpu_ring *ring, long timeout)
673{
674 struct amdgpu_device *adev = ring->adev;
675 struct amdgpu_ib ib;
676 struct dma_fence *f = NULL;
677 unsigned index;
678 u32 tmp = 0;
679 u64 gpu_addr;
680 long r;
681
682 r = amdgpu_device_wb_get(adev, &index);
683 if (r) {
684 dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
685 return r;
686 }
687
688 gpu_addr = adev->wb.gpu_addr + (index * 4);
689 tmp = 0xCAFEDEAD;
690 adev->wb.wb[index] = cpu_to_le32(tmp);
691 memset(&ib, 0, sizeof(ib));
692 r = amdgpu_ib_get(adev, NULL, 256, &ib);
693 if (r) {
694 DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
695 goto err0;
696 }
697
698 ib.ptr[0] = SDMA_PACKET(SDMA_OPCODE_WRITE,
699 SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
700 ib.ptr[1] = lower_32_bits(gpu_addr);
701 ib.ptr[2] = upper_32_bits(gpu_addr);
702 ib.ptr[3] = 1;
703 ib.ptr[4] = 0xDEADBEEF;
704 ib.length_dw = 5;
705 r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
706 if (r)
707 goto err1;
708
709 r = dma_fence_wait_timeout(f, false, timeout);
710 if (r == 0) {
711 DRM_ERROR("amdgpu: IB test timed out\n");
712 r = -ETIMEDOUT;
713 goto err1;
714 } else if (r < 0) {
715 DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
716 goto err1;
717 }
718 tmp = le32_to_cpu(adev->wb.wb[index]);
719 if (tmp == 0xDEADBEEF) {
720 DRM_DEBUG("ib test on ring %d succeeded\n", ring->idx);
721 r = 0;
722 } else {
723 DRM_ERROR("amdgpu: ib test failed (0x%08X)\n", tmp);
724 r = -EINVAL;
725 }
726
727err1:
728 amdgpu_ib_free(adev, &ib, NULL);
729 dma_fence_put(f);
730err0:
731 amdgpu_device_wb_free(adev, index);
732 return r;
733}
734
735/**
736 * cik_sdma_vm_copy_pages - update PTEs by copying them from the GART
737 *
738 * @ib: indirect buffer to fill with commands
739 * @pe: addr of the page entry
740 * @src: src addr to copy from
741 * @count: number of page entries to update
742 *
743 * Update PTEs by copying them from the GART using sDMA (CIK).
744 */
745static void cik_sdma_vm_copy_pte(struct amdgpu_ib *ib,
746 uint64_t pe, uint64_t src,
747 unsigned count)
748{
749 unsigned bytes = count * 8;
750
751 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_COPY,
752 SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
753 ib->ptr[ib->length_dw++] = bytes;
754 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
755 ib->ptr[ib->length_dw++] = lower_32_bits(src);
756 ib->ptr[ib->length_dw++] = upper_32_bits(src);
757 ib->ptr[ib->length_dw++] = lower_32_bits(pe);
758 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
759}
760
761/**
762 * cik_sdma_vm_write_pages - update PTEs by writing them manually
763 *
764 * @ib: indirect buffer to fill with commands
765 * @pe: addr of the page entry
766 * @value: dst addr to write into pe
767 * @count: number of page entries to update
768 * @incr: increase next addr by incr bytes
769 *
770 * Update PTEs by writing them manually using sDMA (CIK).
771 */
772static void cik_sdma_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
773 uint64_t value, unsigned count,
774 uint32_t incr)
775{
776 unsigned ndw = count * 2;
777
778 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_WRITE,
779 SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
780 ib->ptr[ib->length_dw++] = lower_32_bits(pe);
781 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
782 ib->ptr[ib->length_dw++] = ndw;
783 for (; ndw > 0; ndw -= 2) {
784 ib->ptr[ib->length_dw++] = lower_32_bits(value);
785 ib->ptr[ib->length_dw++] = upper_32_bits(value);
786 value += incr;
787 }
788}
789
790/**
791 * cik_sdma_vm_set_pages - update the page tables using sDMA
792 *
793 * @ib: indirect buffer to fill with commands
794 * @pe: addr of the page entry
795 * @addr: dst addr to write into pe
796 * @count: number of page entries to update
797 * @incr: increase next addr by incr bytes
798 * @flags: access flags
799 *
800 * Update the page tables using sDMA (CIK).
801 */
802static void cik_sdma_vm_set_pte_pde(struct amdgpu_ib *ib, uint64_t pe,
803 uint64_t addr, unsigned count,
804 uint32_t incr, uint64_t flags)
805{
806 /* for physically contiguous pages (vram) */
807 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_GENERATE_PTE_PDE, 0, 0);
808 ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
809 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
810 ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
811 ib->ptr[ib->length_dw++] = upper_32_bits(flags);
812 ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
813 ib->ptr[ib->length_dw++] = upper_32_bits(addr);
814 ib->ptr[ib->length_dw++] = incr; /* increment size */
815 ib->ptr[ib->length_dw++] = 0;
816 ib->ptr[ib->length_dw++] = count; /* number of entries */
817}
818
819/**
820 * cik_sdma_vm_pad_ib - pad the IB to the required number of dw
821 *
822 * @ib: indirect buffer to fill with padding
823 *
824 */
825static void cik_sdma_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
826{
827 struct amdgpu_sdma_instance *sdma = amdgpu_get_sdma_instance(ring);
828 u32 pad_count;
829 int i;
830
831 pad_count = (8 - (ib->length_dw & 0x7)) % 8;
832 for (i = 0; i < pad_count; i++)
833 if (sdma && sdma->burst_nop && (i == 0))
834 ib->ptr[ib->length_dw++] =
835 SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0) |
836 SDMA_NOP_COUNT(pad_count - 1);
837 else
838 ib->ptr[ib->length_dw++] =
839 SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0);
840}
841
842/**
843 * cik_sdma_ring_emit_pipeline_sync - sync the pipeline
844 *
845 * @ring: amdgpu_ring pointer
846 *
847 * Make sure all previous operations are completed (CIK).
848 */
849static void cik_sdma_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
850{
851 uint32_t seq = ring->fence_drv.sync_seq;
852 uint64_t addr = ring->fence_drv.gpu_addr;
853
854 /* wait for idle */
855 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0,
856 SDMA_POLL_REG_MEM_EXTRA_OP(0) |
857 SDMA_POLL_REG_MEM_EXTRA_FUNC(3) | /* equal */
858 SDMA_POLL_REG_MEM_EXTRA_M));
859 amdgpu_ring_write(ring, addr & 0xfffffffc);
860 amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
861 amdgpu_ring_write(ring, seq); /* reference */
862 amdgpu_ring_write(ring, 0xffffffff); /* mask */
863 amdgpu_ring_write(ring, (0xfff << 16) | 4); /* retry count, poll interval */
864}
865
866/**
867 * cik_sdma_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 (CIK).
874 */
875static void cik_sdma_ring_emit_vm_flush(struct amdgpu_ring *ring,
876 unsigned vmid, uint64_t pd_addr)
877{
878 u32 extra_bits = (SDMA_POLL_REG_MEM_EXTRA_OP(0) |
879 SDMA_POLL_REG_MEM_EXTRA_FUNC(0)); /* always */
880
881 amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
882
883 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0, extra_bits));
884 amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST << 2);
885 amdgpu_ring_write(ring, 0);
886 amdgpu_ring_write(ring, 0); /* reference */
887 amdgpu_ring_write(ring, 0); /* mask */
888 amdgpu_ring_write(ring, (0xfff << 16) | 10); /* retry count, poll interval */
889}
890
891static void cik_sdma_ring_emit_wreg(struct amdgpu_ring *ring,
892 uint32_t reg, uint32_t val)
893{
894 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
895 amdgpu_ring_write(ring, reg);
896 amdgpu_ring_write(ring, val);
897}
898
899static void cik_enable_sdma_mgcg(struct amdgpu_device *adev,
900 bool enable)
901{
902 u32 orig, data;
903
904 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
905 WREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET, 0x00000100);
906 WREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET, 0x00000100);
907 } else {
908 orig = data = RREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET);
909 data |= 0xff000000;
910 if (data != orig)
911 WREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET, data);
912
913 orig = data = RREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET);
914 data |= 0xff000000;
915 if (data != orig)
916 WREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET, data);
917 }
918}
919
920static void cik_enable_sdma_mgls(struct amdgpu_device *adev,
921 bool enable)
922{
923 u32 orig, data;
924
925 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
926 orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET);
927 data |= 0x100;
928 if (orig != data)
929 WREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET, data);
930
931 orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET);
932 data |= 0x100;
933 if (orig != data)
934 WREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET, data);
935 } else {
936 orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET);
937 data &= ~0x100;
938 if (orig != data)
939 WREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET, data);
940
941 orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET);
942 data &= ~0x100;
943 if (orig != data)
944 WREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET, data);
945 }
946}
947
948static int cik_sdma_early_init(void *handle)
949{
950 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
951
952 adev->sdma.num_instances = SDMA_MAX_INSTANCE;
953
954 cik_sdma_set_ring_funcs(adev);
955 cik_sdma_set_irq_funcs(adev);
956 cik_sdma_set_buffer_funcs(adev);
957 cik_sdma_set_vm_pte_funcs(adev);
958
959 return 0;
960}
961
962static int cik_sdma_sw_init(void *handle)
963{
964 struct amdgpu_ring *ring;
965 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
966 int r, i;
967
968 r = cik_sdma_init_microcode(adev);
969 if (r) {
970 DRM_ERROR("Failed to load sdma firmware!\n");
971 return r;
972 }
973
974 /* SDMA trap event */
975 r = amdgpu_irq_add_id(adev, AMDGPU_IH_CLIENTID_LEGACY, 224,
976 &adev->sdma.trap_irq);
977 if (r)
978 return r;
979
980 /* SDMA Privileged inst */
981 r = amdgpu_irq_add_id(adev, AMDGPU_IH_CLIENTID_LEGACY, 241,
982 &adev->sdma.illegal_inst_irq);
983 if (r)
984 return r;
985
986 /* SDMA Privileged inst */
987 r = amdgpu_irq_add_id(adev, AMDGPU_IH_CLIENTID_LEGACY, 247,
988 &adev->sdma.illegal_inst_irq);
989 if (r)
990 return r;
991
992 for (i = 0; i < adev->sdma.num_instances; i++) {
993 ring = &adev->sdma.instance[i].ring;
994 ring->ring_obj = NULL;
995 sprintf(ring->name, "sdma%d", i);
996 r = amdgpu_ring_init(adev, ring, 1024,
997 &adev->sdma.trap_irq,
998 (i == 0) ?
999 AMDGPU_SDMA_IRQ_TRAP0 :
1000 AMDGPU_SDMA_IRQ_TRAP1);
1001 if (r)
1002 return r;
1003 }
1004
1005 return r;
1006}
1007
1008static int cik_sdma_sw_fini(void *handle)
1009{
1010 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1011 int i;
1012
1013 for (i = 0; i < adev->sdma.num_instances; i++)
1014 amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1015
1016 cik_sdma_free_microcode(adev);
1017 return 0;
1018}
1019
1020static int cik_sdma_hw_init(void *handle)
1021{
1022 int r;
1023 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1024
1025 r = cik_sdma_start(adev);
1026 if (r)
1027 return r;
1028
1029 return r;
1030}
1031
1032static int cik_sdma_hw_fini(void *handle)
1033{
1034 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1035
1036 cik_ctx_switch_enable(adev, false);
1037 cik_sdma_enable(adev, false);
1038
1039 return 0;
1040}
1041
1042static int cik_sdma_suspend(void *handle)
1043{
1044 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1045
1046 return cik_sdma_hw_fini(adev);
1047}
1048
1049static int cik_sdma_resume(void *handle)
1050{
1051 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1052
1053 cik_sdma_soft_reset(handle);
1054
1055 return cik_sdma_hw_init(adev);
1056}
1057
1058static bool cik_sdma_is_idle(void *handle)
1059{
1060 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1061 u32 tmp = RREG32(mmSRBM_STATUS2);
1062
1063 if (tmp & (SRBM_STATUS2__SDMA_BUSY_MASK |
1064 SRBM_STATUS2__SDMA1_BUSY_MASK))
1065 return false;
1066
1067 return true;
1068}
1069
1070static int cik_sdma_wait_for_idle(void *handle)
1071{
1072 unsigned i;
1073 u32 tmp;
1074 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1075
1076 for (i = 0; i < adev->usec_timeout; i++) {
1077 tmp = RREG32(mmSRBM_STATUS2) & (SRBM_STATUS2__SDMA_BUSY_MASK |
1078 SRBM_STATUS2__SDMA1_BUSY_MASK);
1079
1080 if (!tmp)
1081 return 0;
1082 udelay(1);
1083 }
1084 return -ETIMEDOUT;
1085}
1086
1087static int cik_sdma_soft_reset(void *handle)
1088{
1089 u32 srbm_soft_reset = 0;
1090 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1091 u32 tmp = RREG32(mmSRBM_STATUS2);
1092
1093 if (tmp & SRBM_STATUS2__SDMA_BUSY_MASK) {
1094 /* sdma0 */
1095 tmp = RREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET);
1096 tmp |= SDMA0_F32_CNTL__HALT_MASK;
1097 WREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET, tmp);
1098 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA_MASK;
1099 }
1100 if (tmp & SRBM_STATUS2__SDMA1_BUSY_MASK) {
1101 /* sdma1 */
1102 tmp = RREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET);
1103 tmp |= SDMA0_F32_CNTL__HALT_MASK;
1104 WREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET, tmp);
1105 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA1_MASK;
1106 }
1107
1108 if (srbm_soft_reset) {
1109 tmp = RREG32(mmSRBM_SOFT_RESET);
1110 tmp |= srbm_soft_reset;
1111 dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
1112 WREG32(mmSRBM_SOFT_RESET, tmp);
1113 tmp = RREG32(mmSRBM_SOFT_RESET);
1114
1115 udelay(50);
1116
1117 tmp &= ~srbm_soft_reset;
1118 WREG32(mmSRBM_SOFT_RESET, tmp);
1119 tmp = RREG32(mmSRBM_SOFT_RESET);
1120
1121 /* Wait a little for things to settle down */
1122 udelay(50);
1123 }
1124
1125 return 0;
1126}
1127
1128static int cik_sdma_set_trap_irq_state(struct amdgpu_device *adev,
1129 struct amdgpu_irq_src *src,
1130 unsigned type,
1131 enum amdgpu_interrupt_state state)
1132{
1133 u32 sdma_cntl;
1134
1135 switch (type) {
1136 case AMDGPU_SDMA_IRQ_TRAP0:
1137 switch (state) {
1138 case AMDGPU_IRQ_STATE_DISABLE:
1139 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1140 sdma_cntl &= ~SDMA0_CNTL__TRAP_ENABLE_MASK;
1141 WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1142 break;
1143 case AMDGPU_IRQ_STATE_ENABLE:
1144 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1145 sdma_cntl |= SDMA0_CNTL__TRAP_ENABLE_MASK;
1146 WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1147 break;
1148 default:
1149 break;
1150 }
1151 break;
1152 case AMDGPU_SDMA_IRQ_TRAP1:
1153 switch (state) {
1154 case AMDGPU_IRQ_STATE_DISABLE:
1155 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1156 sdma_cntl &= ~SDMA0_CNTL__TRAP_ENABLE_MASK;
1157 WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1158 break;
1159 case AMDGPU_IRQ_STATE_ENABLE:
1160 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1161 sdma_cntl |= SDMA0_CNTL__TRAP_ENABLE_MASK;
1162 WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1163 break;
1164 default:
1165 break;
1166 }
1167 break;
1168 default:
1169 break;
1170 }
1171 return 0;
1172}
1173
1174static int cik_sdma_process_trap_irq(struct amdgpu_device *adev,
1175 struct amdgpu_irq_src *source,
1176 struct amdgpu_iv_entry *entry)
1177{
1178 u8 instance_id, queue_id;
1179
1180 instance_id = (entry->ring_id & 0x3) >> 0;
1181 queue_id = (entry->ring_id & 0xc) >> 2;
1182 DRM_DEBUG("IH: SDMA trap\n");
1183 switch (instance_id) {
1184 case 0:
1185 switch (queue_id) {
1186 case 0:
1187 amdgpu_fence_process(&adev->sdma.instance[0].ring);
1188 break;
1189 case 1:
1190 /* XXX compute */
1191 break;
1192 case 2:
1193 /* XXX compute */
1194 break;
1195 }
1196 break;
1197 case 1:
1198 switch (queue_id) {
1199 case 0:
1200 amdgpu_fence_process(&adev->sdma.instance[1].ring);
1201 break;
1202 case 1:
1203 /* XXX compute */
1204 break;
1205 case 2:
1206 /* XXX compute */
1207 break;
1208 }
1209 break;
1210 }
1211
1212 return 0;
1213}
1214
1215static int cik_sdma_process_illegal_inst_irq(struct amdgpu_device *adev,
1216 struct amdgpu_irq_src *source,
1217 struct amdgpu_iv_entry *entry)
1218{
1219 DRM_ERROR("Illegal instruction in SDMA command stream\n");
1220 schedule_work(&adev->reset_work);
1221 return 0;
1222}
1223
1224static int cik_sdma_set_clockgating_state(void *handle,
1225 enum amd_clockgating_state state)
1226{
1227 bool gate = false;
1228 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1229
1230 if (state == AMD_CG_STATE_GATE)
1231 gate = true;
1232
1233 cik_enable_sdma_mgcg(adev, gate);
1234 cik_enable_sdma_mgls(adev, gate);
1235
1236 return 0;
1237}
1238
1239static int cik_sdma_set_powergating_state(void *handle,
1240 enum amd_powergating_state state)
1241{
1242 return 0;
1243}
1244
1245static const struct amd_ip_funcs cik_sdma_ip_funcs = {
1246 .name = "cik_sdma",
1247 .early_init = cik_sdma_early_init,
1248 .late_init = NULL,
1249 .sw_init = cik_sdma_sw_init,
1250 .sw_fini = cik_sdma_sw_fini,
1251 .hw_init = cik_sdma_hw_init,
1252 .hw_fini = cik_sdma_hw_fini,
1253 .suspend = cik_sdma_suspend,
1254 .resume = cik_sdma_resume,
1255 .is_idle = cik_sdma_is_idle,
1256 .wait_for_idle = cik_sdma_wait_for_idle,
1257 .soft_reset = cik_sdma_soft_reset,
1258 .set_clockgating_state = cik_sdma_set_clockgating_state,
1259 .set_powergating_state = cik_sdma_set_powergating_state,
1260};
1261
1262static const struct amdgpu_ring_funcs cik_sdma_ring_funcs = {
1263 .type = AMDGPU_RING_TYPE_SDMA,
1264 .align_mask = 0xf,
1265 .nop = SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0),
1266 .support_64bit_ptrs = false,
1267 .get_rptr = cik_sdma_ring_get_rptr,
1268 .get_wptr = cik_sdma_ring_get_wptr,
1269 .set_wptr = cik_sdma_ring_set_wptr,
1270 .emit_frame_size =
1271 6 + /* cik_sdma_ring_emit_hdp_flush */
1272 3 + /* hdp invalidate */
1273 6 + /* cik_sdma_ring_emit_pipeline_sync */
1274 CIK_FLUSH_GPU_TLB_NUM_WREG * 3 + 6 + /* cik_sdma_ring_emit_vm_flush */
1275 9 + 9 + 9, /* cik_sdma_ring_emit_fence x3 for user fence, vm fence */
1276 .emit_ib_size = 7 + 4, /* cik_sdma_ring_emit_ib */
1277 .emit_ib = cik_sdma_ring_emit_ib,
1278 .emit_fence = cik_sdma_ring_emit_fence,
1279 .emit_pipeline_sync = cik_sdma_ring_emit_pipeline_sync,
1280 .emit_vm_flush = cik_sdma_ring_emit_vm_flush,
1281 .emit_hdp_flush = cik_sdma_ring_emit_hdp_flush,
1282 .test_ring = cik_sdma_ring_test_ring,
1283 .test_ib = cik_sdma_ring_test_ib,
1284 .insert_nop = cik_sdma_ring_insert_nop,
1285 .pad_ib = cik_sdma_ring_pad_ib,
1286 .emit_wreg = cik_sdma_ring_emit_wreg,
1287};
1288
1289static void cik_sdma_set_ring_funcs(struct amdgpu_device *adev)
1290{
1291 int i;
1292
1293 for (i = 0; i < adev->sdma.num_instances; i++)
1294 adev->sdma.instance[i].ring.funcs = &cik_sdma_ring_funcs;
1295}
1296
1297static const struct amdgpu_irq_src_funcs cik_sdma_trap_irq_funcs = {
1298 .set = cik_sdma_set_trap_irq_state,
1299 .process = cik_sdma_process_trap_irq,
1300};
1301
1302static const struct amdgpu_irq_src_funcs cik_sdma_illegal_inst_irq_funcs = {
1303 .process = cik_sdma_process_illegal_inst_irq,
1304};
1305
1306static void cik_sdma_set_irq_funcs(struct amdgpu_device *adev)
1307{
1308 adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
1309 adev->sdma.trap_irq.funcs = &cik_sdma_trap_irq_funcs;
1310 adev->sdma.illegal_inst_irq.funcs = &cik_sdma_illegal_inst_irq_funcs;
1311}
1312
1313/**
1314 * cik_sdma_emit_copy_buffer - copy buffer using the sDMA engine
1315 *
1316 * @ring: amdgpu_ring structure holding ring information
1317 * @src_offset: src GPU address
1318 * @dst_offset: dst GPU address
1319 * @byte_count: number of bytes to xfer
1320 *
1321 * Copy GPU buffers using the DMA engine (CIK).
1322 * Used by the amdgpu ttm implementation to move pages if
1323 * registered as the asic copy callback.
1324 */
1325static void cik_sdma_emit_copy_buffer(struct amdgpu_ib *ib,
1326 uint64_t src_offset,
1327 uint64_t dst_offset,
1328 uint32_t byte_count)
1329{
1330 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_COPY, SDMA_COPY_SUB_OPCODE_LINEAR, 0);
1331 ib->ptr[ib->length_dw++] = byte_count;
1332 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1333 ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1334 ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1335 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1336 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1337}
1338
1339/**
1340 * cik_sdma_emit_fill_buffer - fill buffer using the sDMA engine
1341 *
1342 * @ring: amdgpu_ring structure holding ring information
1343 * @src_data: value to write to buffer
1344 * @dst_offset: dst GPU address
1345 * @byte_count: number of bytes to xfer
1346 *
1347 * Fill GPU buffers using the DMA engine (CIK).
1348 */
1349static void cik_sdma_emit_fill_buffer(struct amdgpu_ib *ib,
1350 uint32_t src_data,
1351 uint64_t dst_offset,
1352 uint32_t byte_count)
1353{
1354 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_CONSTANT_FILL, 0, 0);
1355 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1356 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1357 ib->ptr[ib->length_dw++] = src_data;
1358 ib->ptr[ib->length_dw++] = byte_count;
1359}
1360
1361static const struct amdgpu_buffer_funcs cik_sdma_buffer_funcs = {
1362 .copy_max_bytes = 0x1fffff,
1363 .copy_num_dw = 7,
1364 .emit_copy_buffer = cik_sdma_emit_copy_buffer,
1365
1366 .fill_max_bytes = 0x1fffff,
1367 .fill_num_dw = 5,
1368 .emit_fill_buffer = cik_sdma_emit_fill_buffer,
1369};
1370
1371static void cik_sdma_set_buffer_funcs(struct amdgpu_device *adev)
1372{
1373 if (adev->mman.buffer_funcs == NULL) {
1374 adev->mman.buffer_funcs = &cik_sdma_buffer_funcs;
1375 adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1376 }
1377}
1378
1379static const struct amdgpu_vm_pte_funcs cik_sdma_vm_pte_funcs = {
1380 .copy_pte_num_dw = 7,
1381 .copy_pte = cik_sdma_vm_copy_pte,
1382
1383 .write_pte = cik_sdma_vm_write_pte,
1384 .set_pte_pde = cik_sdma_vm_set_pte_pde,
1385};
1386
1387static void cik_sdma_set_vm_pte_funcs(struct amdgpu_device *adev)
1388{
1389 unsigned i;
1390
1391 if (adev->vm_manager.vm_pte_funcs == NULL) {
1392 adev->vm_manager.vm_pte_funcs = &cik_sdma_vm_pte_funcs;
1393 for (i = 0; i < adev->sdma.num_instances; i++)
1394 adev->vm_manager.vm_pte_rings[i] =
1395 &adev->sdma.instance[i].ring;
1396
1397 adev->vm_manager.vm_pte_num_rings = adev->sdma.num_instances;
1398 }
1399}
1400
1401const struct amdgpu_ip_block_version cik_sdma_ip_block =
1402{
1403 .type = AMD_IP_BLOCK_TYPE_SDMA,
1404 .major = 2,
1405 .minor = 0,
1406 .rev = 0,
1407 .funcs = &cik_sdma_ip_funcs,
1408};
1/*
2 * Copyright 2013 Advanced Micro Devices, Inc.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
21 *
22 * Authors: Alex Deucher
23 */
24
25#include <linux/firmware.h>
26#include <linux/module.h>
27
28#include "amdgpu.h"
29#include "amdgpu_ucode.h"
30#include "amdgpu_trace.h"
31#include "cikd.h"
32#include "cik.h"
33
34#include "bif/bif_4_1_d.h"
35#include "bif/bif_4_1_sh_mask.h"
36
37#include "gca/gfx_7_2_d.h"
38#include "gca/gfx_7_2_enum.h"
39#include "gca/gfx_7_2_sh_mask.h"
40
41#include "gmc/gmc_7_1_d.h"
42#include "gmc/gmc_7_1_sh_mask.h"
43
44#include "oss/oss_2_0_d.h"
45#include "oss/oss_2_0_sh_mask.h"
46
47static const u32 sdma_offsets[SDMA_MAX_INSTANCE] =
48{
49 SDMA0_REGISTER_OFFSET,
50 SDMA1_REGISTER_OFFSET
51};
52
53static void cik_sdma_set_ring_funcs(struct amdgpu_device *adev);
54static void cik_sdma_set_irq_funcs(struct amdgpu_device *adev);
55static void cik_sdma_set_buffer_funcs(struct amdgpu_device *adev);
56static void cik_sdma_set_vm_pte_funcs(struct amdgpu_device *adev);
57static int cik_sdma_soft_reset(void *handle);
58
59MODULE_FIRMWARE("amdgpu/bonaire_sdma.bin");
60MODULE_FIRMWARE("amdgpu/bonaire_sdma1.bin");
61MODULE_FIRMWARE("amdgpu/hawaii_sdma.bin");
62MODULE_FIRMWARE("amdgpu/hawaii_sdma1.bin");
63MODULE_FIRMWARE("amdgpu/kaveri_sdma.bin");
64MODULE_FIRMWARE("amdgpu/kaveri_sdma1.bin");
65MODULE_FIRMWARE("amdgpu/kabini_sdma.bin");
66MODULE_FIRMWARE("amdgpu/kabini_sdma1.bin");
67MODULE_FIRMWARE("amdgpu/mullins_sdma.bin");
68MODULE_FIRMWARE("amdgpu/mullins_sdma1.bin");
69
70u32 amdgpu_cik_gpu_check_soft_reset(struct amdgpu_device *adev);
71
72
73static void cik_sdma_free_microcode(struct amdgpu_device *adev)
74{
75 int i;
76 for (i = 0; i < adev->sdma.num_instances; i++) {
77 release_firmware(adev->sdma.instance[i].fw);
78 adev->sdma.instance[i].fw = NULL;
79 }
80}
81
82/*
83 * sDMA - System DMA
84 * Starting with CIK, the GPU has new asynchronous
85 * DMA engines. These engines are used for compute
86 * and gfx. There are two DMA engines (SDMA0, SDMA1)
87 * and each one supports 1 ring buffer used for gfx
88 * and 2 queues used for compute.
89 *
90 * The programming model is very similar to the CP
91 * (ring buffer, IBs, etc.), but sDMA has it's own
92 * packet format that is different from the PM4 format
93 * used by the CP. sDMA supports copying data, writing
94 * embedded data, solid fills, and a number of other
95 * things. It also has support for tiling/detiling of
96 * buffers.
97 */
98
99/**
100 * cik_sdma_init_microcode - load ucode images from disk
101 *
102 * @adev: amdgpu_device pointer
103 *
104 * Use the firmware interface to load the ucode images into
105 * the driver (not loaded into hw).
106 * Returns 0 on success, error on failure.
107 */
108static int cik_sdma_init_microcode(struct amdgpu_device *adev)
109{
110 const char *chip_name;
111 char fw_name[30];
112 int err = 0, i;
113
114 DRM_DEBUG("\n");
115
116 switch (adev->asic_type) {
117 case CHIP_BONAIRE:
118 chip_name = "bonaire";
119 break;
120 case CHIP_HAWAII:
121 chip_name = "hawaii";
122 break;
123 case CHIP_KAVERI:
124 chip_name = "kaveri";
125 break;
126 case CHIP_KABINI:
127 chip_name = "kabini";
128 break;
129 case CHIP_MULLINS:
130 chip_name = "mullins";
131 break;
132 default: BUG();
133 }
134
135 for (i = 0; i < adev->sdma.num_instances; i++) {
136 if (i == 0)
137 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma.bin", chip_name);
138 else
139 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma1.bin", chip_name);
140 err = request_firmware(&adev->sdma.instance[i].fw, fw_name, adev->dev);
141 if (err)
142 goto out;
143 err = amdgpu_ucode_validate(adev->sdma.instance[i].fw);
144 }
145out:
146 if (err) {
147 pr_err("cik_sdma: Failed to load firmware \"%s\"\n", fw_name);
148 for (i = 0; i < adev->sdma.num_instances; i++) {
149 release_firmware(adev->sdma.instance[i].fw);
150 adev->sdma.instance[i].fw = NULL;
151 }
152 }
153 return err;
154}
155
156/**
157 * cik_sdma_ring_get_rptr - get the current read pointer
158 *
159 * @ring: amdgpu ring pointer
160 *
161 * Get the current rptr from the hardware (CIK+).
162 */
163static uint64_t cik_sdma_ring_get_rptr(struct amdgpu_ring *ring)
164{
165 u32 rptr;
166
167 rptr = ring->adev->wb.wb[ring->rptr_offs];
168
169 return (rptr & 0x3fffc) >> 2;
170}
171
172/**
173 * cik_sdma_ring_get_wptr - get the current write pointer
174 *
175 * @ring: amdgpu ring pointer
176 *
177 * Get the current wptr from the hardware (CIK+).
178 */
179static uint64_t cik_sdma_ring_get_wptr(struct amdgpu_ring *ring)
180{
181 struct amdgpu_device *adev = ring->adev;
182
183 return (RREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[ring->me]) & 0x3fffc) >> 2;
184}
185
186/**
187 * cik_sdma_ring_set_wptr - commit the write pointer
188 *
189 * @ring: amdgpu ring pointer
190 *
191 * Write the wptr back to the hardware (CIK+).
192 */
193static void cik_sdma_ring_set_wptr(struct amdgpu_ring *ring)
194{
195 struct amdgpu_device *adev = ring->adev;
196
197 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[ring->me],
198 (lower_32_bits(ring->wptr) << 2) & 0x3fffc);
199}
200
201static void cik_sdma_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
202{
203 struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
204 int i;
205
206 for (i = 0; i < count; i++)
207 if (sdma && sdma->burst_nop && (i == 0))
208 amdgpu_ring_write(ring, ring->funcs->nop |
209 SDMA_NOP_COUNT(count - 1));
210 else
211 amdgpu_ring_write(ring, ring->funcs->nop);
212}
213
214/**
215 * cik_sdma_ring_emit_ib - Schedule an IB on the DMA engine
216 *
217 * @ring: amdgpu ring pointer
218 * @job: job to retrive vmid from
219 * @ib: IB object to schedule
220 * @flags: unused
221 *
222 * Schedule an IB in the DMA ring (CIK).
223 */
224static void cik_sdma_ring_emit_ib(struct amdgpu_ring *ring,
225 struct amdgpu_job *job,
226 struct amdgpu_ib *ib,
227 uint32_t flags)
228{
229 unsigned vmid = AMDGPU_JOB_GET_VMID(job);
230 u32 extra_bits = vmid & 0xf;
231
232 /* IB packet must end on a 8 DW boundary */
233 cik_sdma_ring_insert_nop(ring, (4 - lower_32_bits(ring->wptr)) & 7);
234
235 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_INDIRECT_BUFFER, 0, extra_bits));
236 amdgpu_ring_write(ring, ib->gpu_addr & 0xffffffe0); /* base must be 32 byte aligned */
237 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr) & 0xffffffff);
238 amdgpu_ring_write(ring, ib->length_dw);
239
240}
241
242/**
243 * cik_sdma_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
244 *
245 * @ring: amdgpu ring pointer
246 *
247 * Emit an hdp flush packet on the requested DMA ring.
248 */
249static void cik_sdma_ring_emit_hdp_flush(struct amdgpu_ring *ring)
250{
251 u32 extra_bits = (SDMA_POLL_REG_MEM_EXTRA_OP(1) |
252 SDMA_POLL_REG_MEM_EXTRA_FUNC(3)); /* == */
253 u32 ref_and_mask;
254
255 if (ring->me == 0)
256 ref_and_mask = GPU_HDP_FLUSH_DONE__SDMA0_MASK;
257 else
258 ref_and_mask = GPU_HDP_FLUSH_DONE__SDMA1_MASK;
259
260 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0, extra_bits));
261 amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_DONE << 2);
262 amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_REQ << 2);
263 amdgpu_ring_write(ring, ref_and_mask); /* reference */
264 amdgpu_ring_write(ring, ref_and_mask); /* mask */
265 amdgpu_ring_write(ring, (0xfff << 16) | 10); /* retry count, poll interval */
266}
267
268/**
269 * cik_sdma_ring_emit_fence - emit a fence on the DMA ring
270 *
271 * @ring: amdgpu ring pointer
272 * @addr: address
273 * @seq: sequence number
274 * @flags: fence related flags
275 *
276 * Add a DMA fence packet to the ring to write
277 * the fence seq number and DMA trap packet to generate
278 * an interrupt if needed (CIK).
279 */
280static void cik_sdma_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
281 unsigned flags)
282{
283 bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
284 /* write the fence */
285 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_FENCE, 0, 0));
286 amdgpu_ring_write(ring, lower_32_bits(addr));
287 amdgpu_ring_write(ring, upper_32_bits(addr));
288 amdgpu_ring_write(ring, lower_32_bits(seq));
289
290 /* optionally write high bits as well */
291 if (write64bit) {
292 addr += 4;
293 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_FENCE, 0, 0));
294 amdgpu_ring_write(ring, lower_32_bits(addr));
295 amdgpu_ring_write(ring, upper_32_bits(addr));
296 amdgpu_ring_write(ring, upper_32_bits(seq));
297 }
298
299 /* generate an interrupt */
300 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_TRAP, 0, 0));
301}
302
303/**
304 * cik_sdma_gfx_stop - stop the gfx async dma engines
305 *
306 * @adev: amdgpu_device pointer
307 *
308 * Stop the gfx async dma ring buffers (CIK).
309 */
310static void cik_sdma_gfx_stop(struct amdgpu_device *adev)
311{
312 struct amdgpu_ring *sdma0 = &adev->sdma.instance[0].ring;
313 struct amdgpu_ring *sdma1 = &adev->sdma.instance[1].ring;
314 u32 rb_cntl;
315 int i;
316
317 if ((adev->mman.buffer_funcs_ring == sdma0) ||
318 (adev->mman.buffer_funcs_ring == sdma1))
319 amdgpu_ttm_set_buffer_funcs_status(adev, false);
320
321 for (i = 0; i < adev->sdma.num_instances; i++) {
322 rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
323 rb_cntl &= ~SDMA0_GFX_RB_CNTL__RB_ENABLE_MASK;
324 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
325 WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], 0);
326 }
327}
328
329/**
330 * cik_sdma_rlc_stop - stop the compute async dma engines
331 *
332 * @adev: amdgpu_device pointer
333 *
334 * Stop the compute async dma queues (CIK).
335 */
336static void cik_sdma_rlc_stop(struct amdgpu_device *adev)
337{
338 /* XXX todo */
339}
340
341/**
342 * cik_ctx_switch_enable - stop the async dma engines context switch
343 *
344 * @adev: amdgpu_device pointer
345 * @enable: enable/disable the DMA MEs context switch.
346 *
347 * Halt or unhalt the async dma engines context switch (VI).
348 */
349static void cik_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
350{
351 u32 f32_cntl, phase_quantum = 0;
352 int i;
353
354 if (amdgpu_sdma_phase_quantum) {
355 unsigned value = amdgpu_sdma_phase_quantum;
356 unsigned unit = 0;
357
358 while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
359 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
360 value = (value + 1) >> 1;
361 unit++;
362 }
363 if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
364 SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
365 value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
366 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
367 unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
368 SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
369 WARN_ONCE(1,
370 "clamping sdma_phase_quantum to %uK clock cycles\n",
371 value << unit);
372 }
373 phase_quantum =
374 value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
375 unit << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
376 }
377
378 for (i = 0; i < adev->sdma.num_instances; i++) {
379 f32_cntl = RREG32(mmSDMA0_CNTL + sdma_offsets[i]);
380 if (enable) {
381 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
382 AUTO_CTXSW_ENABLE, 1);
383 if (amdgpu_sdma_phase_quantum) {
384 WREG32(mmSDMA0_PHASE0_QUANTUM + sdma_offsets[i],
385 phase_quantum);
386 WREG32(mmSDMA0_PHASE1_QUANTUM + sdma_offsets[i],
387 phase_quantum);
388 }
389 } else {
390 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
391 AUTO_CTXSW_ENABLE, 0);
392 }
393
394 WREG32(mmSDMA0_CNTL + sdma_offsets[i], f32_cntl);
395 }
396}
397
398/**
399 * cik_sdma_enable - stop the async dma engines
400 *
401 * @adev: amdgpu_device pointer
402 * @enable: enable/disable the DMA MEs.
403 *
404 * Halt or unhalt the async dma engines (CIK).
405 */
406static void cik_sdma_enable(struct amdgpu_device *adev, bool enable)
407{
408 u32 me_cntl;
409 int i;
410
411 if (!enable) {
412 cik_sdma_gfx_stop(adev);
413 cik_sdma_rlc_stop(adev);
414 }
415
416 for (i = 0; i < adev->sdma.num_instances; i++) {
417 me_cntl = RREG32(mmSDMA0_F32_CNTL + sdma_offsets[i]);
418 if (enable)
419 me_cntl &= ~SDMA0_F32_CNTL__HALT_MASK;
420 else
421 me_cntl |= SDMA0_F32_CNTL__HALT_MASK;
422 WREG32(mmSDMA0_F32_CNTL + sdma_offsets[i], me_cntl);
423 }
424}
425
426/**
427 * cik_sdma_gfx_resume - setup and start the async dma engines
428 *
429 * @adev: amdgpu_device pointer
430 *
431 * Set up the gfx DMA ring buffers and enable them (CIK).
432 * Returns 0 for success, error for failure.
433 */
434static int cik_sdma_gfx_resume(struct amdgpu_device *adev)
435{
436 struct amdgpu_ring *ring;
437 u32 rb_cntl, ib_cntl;
438 u32 rb_bufsz;
439 u32 wb_offset;
440 int i, j, r;
441
442 for (i = 0; i < adev->sdma.num_instances; i++) {
443 ring = &adev->sdma.instance[i].ring;
444 wb_offset = (ring->rptr_offs * 4);
445
446 mutex_lock(&adev->srbm_mutex);
447 for (j = 0; j < 16; j++) {
448 cik_srbm_select(adev, 0, 0, 0, j);
449 /* SDMA GFX */
450 WREG32(mmSDMA0_GFX_VIRTUAL_ADDR + sdma_offsets[i], 0);
451 WREG32(mmSDMA0_GFX_APE1_CNTL + sdma_offsets[i], 0);
452 /* XXX SDMA RLC - todo */
453 }
454 cik_srbm_select(adev, 0, 0, 0, 0);
455 mutex_unlock(&adev->srbm_mutex);
456
457 WREG32(mmSDMA0_TILING_CONFIG + sdma_offsets[i],
458 adev->gfx.config.gb_addr_config & 0x70);
459
460 WREG32(mmSDMA0_SEM_INCOMPLETE_TIMER_CNTL + sdma_offsets[i], 0);
461 WREG32(mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
462
463 /* Set ring buffer size in dwords */
464 rb_bufsz = order_base_2(ring->ring_size / 4);
465 rb_cntl = rb_bufsz << 1;
466#ifdef __BIG_ENDIAN
467 rb_cntl |= SDMA0_GFX_RB_CNTL__RB_SWAP_ENABLE_MASK |
468 SDMA0_GFX_RB_CNTL__RPTR_WRITEBACK_SWAP_ENABLE_MASK;
469#endif
470 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
471
472 /* Initialize the ring buffer's read and write pointers */
473 WREG32(mmSDMA0_GFX_RB_RPTR + sdma_offsets[i], 0);
474 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], 0);
475 WREG32(mmSDMA0_GFX_IB_RPTR + sdma_offsets[i], 0);
476 WREG32(mmSDMA0_GFX_IB_OFFSET + sdma_offsets[i], 0);
477
478 /* set the wb address whether it's enabled or not */
479 WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_HI + sdma_offsets[i],
480 upper_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF);
481 WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_LO + sdma_offsets[i],
482 ((adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC));
483
484 rb_cntl |= SDMA0_GFX_RB_CNTL__RPTR_WRITEBACK_ENABLE_MASK;
485
486 WREG32(mmSDMA0_GFX_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
487 WREG32(mmSDMA0_GFX_RB_BASE_HI + sdma_offsets[i], ring->gpu_addr >> 40);
488
489 ring->wptr = 0;
490 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], lower_32_bits(ring->wptr) << 2);
491
492 /* enable DMA RB */
493 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i],
494 rb_cntl | SDMA0_GFX_RB_CNTL__RB_ENABLE_MASK);
495
496 ib_cntl = SDMA0_GFX_IB_CNTL__IB_ENABLE_MASK;
497#ifdef __BIG_ENDIAN
498 ib_cntl |= SDMA0_GFX_IB_CNTL__IB_SWAP_ENABLE_MASK;
499#endif
500 /* enable DMA IBs */
501 WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
502
503 ring->sched.ready = true;
504 }
505
506 cik_sdma_enable(adev, true);
507
508 for (i = 0; i < adev->sdma.num_instances; i++) {
509 ring = &adev->sdma.instance[i].ring;
510 r = amdgpu_ring_test_helper(ring);
511 if (r)
512 return r;
513
514 if (adev->mman.buffer_funcs_ring == ring)
515 amdgpu_ttm_set_buffer_funcs_status(adev, true);
516 }
517
518 return 0;
519}
520
521/**
522 * cik_sdma_rlc_resume - setup and start the async dma engines
523 *
524 * @adev: amdgpu_device pointer
525 *
526 * Set up the compute DMA queues and enable them (CIK).
527 * Returns 0 for success, error for failure.
528 */
529static int cik_sdma_rlc_resume(struct amdgpu_device *adev)
530{
531 /* XXX todo */
532 return 0;
533}
534
535/**
536 * cik_sdma_load_microcode - load the sDMA ME ucode
537 *
538 * @adev: amdgpu_device pointer
539 *
540 * Loads the sDMA0/1 ucode.
541 * Returns 0 for success, -EINVAL if the ucode is not available.
542 */
543static int cik_sdma_load_microcode(struct amdgpu_device *adev)
544{
545 const struct sdma_firmware_header_v1_0 *hdr;
546 const __le32 *fw_data;
547 u32 fw_size;
548 int i, j;
549
550 /* halt the MEs */
551 cik_sdma_enable(adev, false);
552
553 for (i = 0; i < adev->sdma.num_instances; i++) {
554 if (!adev->sdma.instance[i].fw)
555 return -EINVAL;
556 hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
557 amdgpu_ucode_print_sdma_hdr(&hdr->header);
558 fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
559 adev->sdma.instance[i].fw_version = le32_to_cpu(hdr->header.ucode_version);
560 adev->sdma.instance[i].feature_version = le32_to_cpu(hdr->ucode_feature_version);
561 if (adev->sdma.instance[i].feature_version >= 20)
562 adev->sdma.instance[i].burst_nop = true;
563 fw_data = (const __le32 *)
564 (adev->sdma.instance[i].fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
565 WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], 0);
566 for (j = 0; j < fw_size; j++)
567 WREG32(mmSDMA0_UCODE_DATA + sdma_offsets[i], le32_to_cpup(fw_data++));
568 WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], adev->sdma.instance[i].fw_version);
569 }
570
571 return 0;
572}
573
574/**
575 * cik_sdma_start - setup and start the async dma engines
576 *
577 * @adev: amdgpu_device pointer
578 *
579 * Set up the DMA engines and enable them (CIK).
580 * Returns 0 for success, error for failure.
581 */
582static int cik_sdma_start(struct amdgpu_device *adev)
583{
584 int r;
585
586 r = cik_sdma_load_microcode(adev);
587 if (r)
588 return r;
589
590 /* halt the engine before programing */
591 cik_sdma_enable(adev, false);
592 /* enable sdma ring preemption */
593 cik_ctx_switch_enable(adev, true);
594
595 /* start the gfx rings and rlc compute queues */
596 r = cik_sdma_gfx_resume(adev);
597 if (r)
598 return r;
599 r = cik_sdma_rlc_resume(adev);
600 if (r)
601 return r;
602
603 return 0;
604}
605
606/**
607 * cik_sdma_ring_test_ring - simple async dma engine test
608 *
609 * @ring: amdgpu_ring structure holding ring information
610 *
611 * Test the DMA engine by writing using it to write an
612 * value to memory. (CIK).
613 * Returns 0 for success, error for failure.
614 */
615static int cik_sdma_ring_test_ring(struct amdgpu_ring *ring)
616{
617 struct amdgpu_device *adev = ring->adev;
618 unsigned i;
619 unsigned index;
620 int r;
621 u32 tmp;
622 u64 gpu_addr;
623
624 r = amdgpu_device_wb_get(adev, &index);
625 if (r)
626 return r;
627
628 gpu_addr = adev->wb.gpu_addr + (index * 4);
629 tmp = 0xCAFEDEAD;
630 adev->wb.wb[index] = cpu_to_le32(tmp);
631
632 r = amdgpu_ring_alloc(ring, 5);
633 if (r)
634 goto error_free_wb;
635
636 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_WRITE, SDMA_WRITE_SUB_OPCODE_LINEAR, 0));
637 amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
638 amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
639 amdgpu_ring_write(ring, 1); /* number of DWs to follow */
640 amdgpu_ring_write(ring, 0xDEADBEEF);
641 amdgpu_ring_commit(ring);
642
643 for (i = 0; i < adev->usec_timeout; i++) {
644 tmp = le32_to_cpu(adev->wb.wb[index]);
645 if (tmp == 0xDEADBEEF)
646 break;
647 udelay(1);
648 }
649
650 if (i >= adev->usec_timeout)
651 r = -ETIMEDOUT;
652
653error_free_wb:
654 amdgpu_device_wb_free(adev, index);
655 return r;
656}
657
658/**
659 * cik_sdma_ring_test_ib - test an IB on the DMA engine
660 *
661 * @ring: amdgpu_ring structure holding ring information
662 * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
663 *
664 * Test a simple IB in the DMA ring (CIK).
665 * Returns 0 on success, error on failure.
666 */
667static int cik_sdma_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_device_wb_get(adev, &index);
678 if (r)
679 return r;
680
681 gpu_addr = adev->wb.gpu_addr + (index * 4);
682 tmp = 0xCAFEDEAD;
683 adev->wb.wb[index] = cpu_to_le32(tmp);
684 memset(&ib, 0, sizeof(ib));
685 r = amdgpu_ib_get(adev, NULL, 256,
686 AMDGPU_IB_POOL_DIRECT, &ib);
687 if (r)
688 goto err0;
689
690 ib.ptr[0] = SDMA_PACKET(SDMA_OPCODE_WRITE,
691 SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
692 ib.ptr[1] = lower_32_bits(gpu_addr);
693 ib.ptr[2] = upper_32_bits(gpu_addr);
694 ib.ptr[3] = 1;
695 ib.ptr[4] = 0xDEADBEEF;
696 ib.length_dw = 5;
697 r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
698 if (r)
699 goto err1;
700
701 r = dma_fence_wait_timeout(f, false, timeout);
702 if (r == 0) {
703 r = -ETIMEDOUT;
704 goto err1;
705 } else if (r < 0) {
706 goto err1;
707 }
708 tmp = le32_to_cpu(adev->wb.wb[index]);
709 if (tmp == 0xDEADBEEF)
710 r = 0;
711 else
712 r = -EINVAL;
713
714err1:
715 amdgpu_ib_free(adev, &ib, NULL);
716 dma_fence_put(f);
717err0:
718 amdgpu_device_wb_free(adev, index);
719 return r;
720}
721
722/**
723 * cik_sdma_vm_copy_pte - update PTEs by copying them from the GART
724 *
725 * @ib: indirect buffer to fill with commands
726 * @pe: addr of the page entry
727 * @src: src addr to copy from
728 * @count: number of page entries to update
729 *
730 * Update PTEs by copying them from the GART using sDMA (CIK).
731 */
732static void cik_sdma_vm_copy_pte(struct amdgpu_ib *ib,
733 uint64_t pe, uint64_t src,
734 unsigned count)
735{
736 unsigned bytes = count * 8;
737
738 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_COPY,
739 SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
740 ib->ptr[ib->length_dw++] = bytes;
741 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
742 ib->ptr[ib->length_dw++] = lower_32_bits(src);
743 ib->ptr[ib->length_dw++] = upper_32_bits(src);
744 ib->ptr[ib->length_dw++] = lower_32_bits(pe);
745 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
746}
747
748/**
749 * cik_sdma_vm_write_pte - update PTEs by writing them manually
750 *
751 * @ib: indirect buffer to fill with commands
752 * @pe: addr of the page entry
753 * @value: dst addr to write into pe
754 * @count: number of page entries to update
755 * @incr: increase next addr by incr bytes
756 *
757 * Update PTEs by writing them manually using sDMA (CIK).
758 */
759static void cik_sdma_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
760 uint64_t value, unsigned count,
761 uint32_t incr)
762{
763 unsigned ndw = count * 2;
764
765 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_WRITE,
766 SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
767 ib->ptr[ib->length_dw++] = lower_32_bits(pe);
768 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
769 ib->ptr[ib->length_dw++] = ndw;
770 for (; ndw > 0; ndw -= 2) {
771 ib->ptr[ib->length_dw++] = lower_32_bits(value);
772 ib->ptr[ib->length_dw++] = upper_32_bits(value);
773 value += incr;
774 }
775}
776
777/**
778 * cik_sdma_vm_set_pte_pde - update the page tables using sDMA
779 *
780 * @ib: indirect buffer to fill with commands
781 * @pe: addr of the page entry
782 * @addr: dst addr to write into pe
783 * @count: number of page entries to update
784 * @incr: increase next addr by incr bytes
785 * @flags: access flags
786 *
787 * Update the page tables using sDMA (CIK).
788 */
789static void cik_sdma_vm_set_pte_pde(struct amdgpu_ib *ib, uint64_t pe,
790 uint64_t addr, unsigned count,
791 uint32_t incr, uint64_t flags)
792{
793 /* for physically contiguous pages (vram) */
794 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_GENERATE_PTE_PDE, 0, 0);
795 ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
796 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
797 ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
798 ib->ptr[ib->length_dw++] = upper_32_bits(flags);
799 ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
800 ib->ptr[ib->length_dw++] = upper_32_bits(addr);
801 ib->ptr[ib->length_dw++] = incr; /* increment size */
802 ib->ptr[ib->length_dw++] = 0;
803 ib->ptr[ib->length_dw++] = count; /* number of entries */
804}
805
806/**
807 * cik_sdma_ring_pad_ib - pad the IB to the required number of dw
808 *
809 * @ring: amdgpu_ring structure holding ring information
810 * @ib: indirect buffer to fill with padding
811 *
812 */
813static void cik_sdma_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
814{
815 struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
816 u32 pad_count;
817 int i;
818
819 pad_count = (-ib->length_dw) & 7;
820 for (i = 0; i < pad_count; i++)
821 if (sdma && sdma->burst_nop && (i == 0))
822 ib->ptr[ib->length_dw++] =
823 SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0) |
824 SDMA_NOP_COUNT(pad_count - 1);
825 else
826 ib->ptr[ib->length_dw++] =
827 SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0);
828}
829
830/**
831 * cik_sdma_ring_emit_pipeline_sync - sync the pipeline
832 *
833 * @ring: amdgpu_ring pointer
834 *
835 * Make sure all previous operations are completed (CIK).
836 */
837static void cik_sdma_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
838{
839 uint32_t seq = ring->fence_drv.sync_seq;
840 uint64_t addr = ring->fence_drv.gpu_addr;
841
842 /* wait for idle */
843 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0,
844 SDMA_POLL_REG_MEM_EXTRA_OP(0) |
845 SDMA_POLL_REG_MEM_EXTRA_FUNC(3) | /* equal */
846 SDMA_POLL_REG_MEM_EXTRA_M));
847 amdgpu_ring_write(ring, addr & 0xfffffffc);
848 amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
849 amdgpu_ring_write(ring, seq); /* reference */
850 amdgpu_ring_write(ring, 0xffffffff); /* mask */
851 amdgpu_ring_write(ring, (0xfff << 16) | 4); /* retry count, poll interval */
852}
853
854/**
855 * cik_sdma_ring_emit_vm_flush - cik vm flush using sDMA
856 *
857 * @ring: amdgpu_ring pointer
858 * @vmid: vmid number to use
859 * @pd_addr: address
860 *
861 * Update the page table base and flush the VM TLB
862 * using sDMA (CIK).
863 */
864static void cik_sdma_ring_emit_vm_flush(struct amdgpu_ring *ring,
865 unsigned vmid, uint64_t pd_addr)
866{
867 u32 extra_bits = (SDMA_POLL_REG_MEM_EXTRA_OP(0) |
868 SDMA_POLL_REG_MEM_EXTRA_FUNC(0)); /* always */
869
870 amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
871
872 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0, extra_bits));
873 amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST << 2);
874 amdgpu_ring_write(ring, 0);
875 amdgpu_ring_write(ring, 0); /* reference */
876 amdgpu_ring_write(ring, 0); /* mask */
877 amdgpu_ring_write(ring, (0xfff << 16) | 10); /* retry count, poll interval */
878}
879
880static void cik_sdma_ring_emit_wreg(struct amdgpu_ring *ring,
881 uint32_t reg, uint32_t val)
882{
883 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
884 amdgpu_ring_write(ring, reg);
885 amdgpu_ring_write(ring, val);
886}
887
888static void cik_enable_sdma_mgcg(struct amdgpu_device *adev,
889 bool enable)
890{
891 u32 orig, data;
892
893 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
894 WREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET, 0x00000100);
895 WREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET, 0x00000100);
896 } else {
897 orig = data = RREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET);
898 data |= 0xff000000;
899 if (data != orig)
900 WREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET, data);
901
902 orig = data = RREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET);
903 data |= 0xff000000;
904 if (data != orig)
905 WREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET, data);
906 }
907}
908
909static void cik_enable_sdma_mgls(struct amdgpu_device *adev,
910 bool enable)
911{
912 u32 orig, data;
913
914 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
915 orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET);
916 data |= 0x100;
917 if (orig != data)
918 WREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET, data);
919
920 orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET);
921 data |= 0x100;
922 if (orig != data)
923 WREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET, data);
924 } else {
925 orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET);
926 data &= ~0x100;
927 if (orig != data)
928 WREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET, data);
929
930 orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET);
931 data &= ~0x100;
932 if (orig != data)
933 WREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET, data);
934 }
935}
936
937static int cik_sdma_early_init(void *handle)
938{
939 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
940
941 adev->sdma.num_instances = SDMA_MAX_INSTANCE;
942
943 cik_sdma_set_ring_funcs(adev);
944 cik_sdma_set_irq_funcs(adev);
945 cik_sdma_set_buffer_funcs(adev);
946 cik_sdma_set_vm_pte_funcs(adev);
947
948 return 0;
949}
950
951static int cik_sdma_sw_init(void *handle)
952{
953 struct amdgpu_ring *ring;
954 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
955 int r, i;
956
957 r = cik_sdma_init_microcode(adev);
958 if (r) {
959 DRM_ERROR("Failed to load sdma firmware!\n");
960 return r;
961 }
962
963 /* SDMA trap event */
964 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 224,
965 &adev->sdma.trap_irq);
966 if (r)
967 return r;
968
969 /* SDMA Privileged inst */
970 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 241,
971 &adev->sdma.illegal_inst_irq);
972 if (r)
973 return r;
974
975 /* SDMA Privileged inst */
976 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 247,
977 &adev->sdma.illegal_inst_irq);
978 if (r)
979 return r;
980
981 for (i = 0; i < adev->sdma.num_instances; i++) {
982 ring = &adev->sdma.instance[i].ring;
983 ring->ring_obj = NULL;
984 sprintf(ring->name, "sdma%d", i);
985 r = amdgpu_ring_init(adev, ring, 1024,
986 &adev->sdma.trap_irq,
987 (i == 0) ? AMDGPU_SDMA_IRQ_INSTANCE0 :
988 AMDGPU_SDMA_IRQ_INSTANCE1,
989 AMDGPU_RING_PRIO_DEFAULT, NULL);
990 if (r)
991 return r;
992 }
993
994 return r;
995}
996
997static int cik_sdma_sw_fini(void *handle)
998{
999 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1000 int i;
1001
1002 for (i = 0; i < adev->sdma.num_instances; i++)
1003 amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1004
1005 cik_sdma_free_microcode(adev);
1006 return 0;
1007}
1008
1009static int cik_sdma_hw_init(void *handle)
1010{
1011 int r;
1012 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1013
1014 r = cik_sdma_start(adev);
1015 if (r)
1016 return r;
1017
1018 return r;
1019}
1020
1021static int cik_sdma_hw_fini(void *handle)
1022{
1023 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1024
1025 cik_ctx_switch_enable(adev, false);
1026 cik_sdma_enable(adev, false);
1027
1028 return 0;
1029}
1030
1031static int cik_sdma_suspend(void *handle)
1032{
1033 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1034
1035 return cik_sdma_hw_fini(adev);
1036}
1037
1038static int cik_sdma_resume(void *handle)
1039{
1040 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1041
1042 cik_sdma_soft_reset(handle);
1043
1044 return cik_sdma_hw_init(adev);
1045}
1046
1047static bool cik_sdma_is_idle(void *handle)
1048{
1049 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1050 u32 tmp = RREG32(mmSRBM_STATUS2);
1051
1052 if (tmp & (SRBM_STATUS2__SDMA_BUSY_MASK |
1053 SRBM_STATUS2__SDMA1_BUSY_MASK))
1054 return false;
1055
1056 return true;
1057}
1058
1059static int cik_sdma_wait_for_idle(void *handle)
1060{
1061 unsigned i;
1062 u32 tmp;
1063 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1064
1065 for (i = 0; i < adev->usec_timeout; i++) {
1066 tmp = RREG32(mmSRBM_STATUS2) & (SRBM_STATUS2__SDMA_BUSY_MASK |
1067 SRBM_STATUS2__SDMA1_BUSY_MASK);
1068
1069 if (!tmp)
1070 return 0;
1071 udelay(1);
1072 }
1073 return -ETIMEDOUT;
1074}
1075
1076static int cik_sdma_soft_reset(void *handle)
1077{
1078 u32 srbm_soft_reset = 0;
1079 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1080 u32 tmp;
1081
1082 /* sdma0 */
1083 tmp = RREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET);
1084 tmp |= SDMA0_F32_CNTL__HALT_MASK;
1085 WREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET, tmp);
1086 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA_MASK;
1087
1088 /* sdma1 */
1089 tmp = RREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET);
1090 tmp |= SDMA0_F32_CNTL__HALT_MASK;
1091 WREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET, tmp);
1092 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA1_MASK;
1093
1094 if (srbm_soft_reset) {
1095 tmp = RREG32(mmSRBM_SOFT_RESET);
1096 tmp |= srbm_soft_reset;
1097 dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
1098 WREG32(mmSRBM_SOFT_RESET, tmp);
1099 tmp = RREG32(mmSRBM_SOFT_RESET);
1100
1101 udelay(50);
1102
1103 tmp &= ~srbm_soft_reset;
1104 WREG32(mmSRBM_SOFT_RESET, tmp);
1105 tmp = RREG32(mmSRBM_SOFT_RESET);
1106
1107 /* Wait a little for things to settle down */
1108 udelay(50);
1109 }
1110
1111 return 0;
1112}
1113
1114static int cik_sdma_set_trap_irq_state(struct amdgpu_device *adev,
1115 struct amdgpu_irq_src *src,
1116 unsigned type,
1117 enum amdgpu_interrupt_state state)
1118{
1119 u32 sdma_cntl;
1120
1121 switch (type) {
1122 case AMDGPU_SDMA_IRQ_INSTANCE0:
1123 switch (state) {
1124 case AMDGPU_IRQ_STATE_DISABLE:
1125 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1126 sdma_cntl &= ~SDMA0_CNTL__TRAP_ENABLE_MASK;
1127 WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1128 break;
1129 case AMDGPU_IRQ_STATE_ENABLE:
1130 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1131 sdma_cntl |= SDMA0_CNTL__TRAP_ENABLE_MASK;
1132 WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1133 break;
1134 default:
1135 break;
1136 }
1137 break;
1138 case AMDGPU_SDMA_IRQ_INSTANCE1:
1139 switch (state) {
1140 case AMDGPU_IRQ_STATE_DISABLE:
1141 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1142 sdma_cntl &= ~SDMA0_CNTL__TRAP_ENABLE_MASK;
1143 WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1144 break;
1145 case AMDGPU_IRQ_STATE_ENABLE:
1146 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1147 sdma_cntl |= SDMA0_CNTL__TRAP_ENABLE_MASK;
1148 WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1149 break;
1150 default:
1151 break;
1152 }
1153 break;
1154 default:
1155 break;
1156 }
1157 return 0;
1158}
1159
1160static int cik_sdma_process_trap_irq(struct amdgpu_device *adev,
1161 struct amdgpu_irq_src *source,
1162 struct amdgpu_iv_entry *entry)
1163{
1164 u8 instance_id, queue_id;
1165
1166 instance_id = (entry->ring_id & 0x3) >> 0;
1167 queue_id = (entry->ring_id & 0xc) >> 2;
1168 DRM_DEBUG("IH: SDMA trap\n");
1169 switch (instance_id) {
1170 case 0:
1171 switch (queue_id) {
1172 case 0:
1173 amdgpu_fence_process(&adev->sdma.instance[0].ring);
1174 break;
1175 case 1:
1176 /* XXX compute */
1177 break;
1178 case 2:
1179 /* XXX compute */
1180 break;
1181 }
1182 break;
1183 case 1:
1184 switch (queue_id) {
1185 case 0:
1186 amdgpu_fence_process(&adev->sdma.instance[1].ring);
1187 break;
1188 case 1:
1189 /* XXX compute */
1190 break;
1191 case 2:
1192 /* XXX compute */
1193 break;
1194 }
1195 break;
1196 }
1197
1198 return 0;
1199}
1200
1201static int cik_sdma_process_illegal_inst_irq(struct amdgpu_device *adev,
1202 struct amdgpu_irq_src *source,
1203 struct amdgpu_iv_entry *entry)
1204{
1205 u8 instance_id;
1206
1207 DRM_ERROR("Illegal instruction in SDMA command stream\n");
1208 instance_id = (entry->ring_id & 0x3) >> 0;
1209 drm_sched_fault(&adev->sdma.instance[instance_id].ring.sched);
1210 return 0;
1211}
1212
1213static int cik_sdma_set_clockgating_state(void *handle,
1214 enum amd_clockgating_state state)
1215{
1216 bool gate = false;
1217 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1218
1219 if (state == AMD_CG_STATE_GATE)
1220 gate = true;
1221
1222 cik_enable_sdma_mgcg(adev, gate);
1223 cik_enable_sdma_mgls(adev, gate);
1224
1225 return 0;
1226}
1227
1228static int cik_sdma_set_powergating_state(void *handle,
1229 enum amd_powergating_state state)
1230{
1231 return 0;
1232}
1233
1234static const struct amd_ip_funcs cik_sdma_ip_funcs = {
1235 .name = "cik_sdma",
1236 .early_init = cik_sdma_early_init,
1237 .late_init = NULL,
1238 .sw_init = cik_sdma_sw_init,
1239 .sw_fini = cik_sdma_sw_fini,
1240 .hw_init = cik_sdma_hw_init,
1241 .hw_fini = cik_sdma_hw_fini,
1242 .suspend = cik_sdma_suspend,
1243 .resume = cik_sdma_resume,
1244 .is_idle = cik_sdma_is_idle,
1245 .wait_for_idle = cik_sdma_wait_for_idle,
1246 .soft_reset = cik_sdma_soft_reset,
1247 .set_clockgating_state = cik_sdma_set_clockgating_state,
1248 .set_powergating_state = cik_sdma_set_powergating_state,
1249};
1250
1251static const struct amdgpu_ring_funcs cik_sdma_ring_funcs = {
1252 .type = AMDGPU_RING_TYPE_SDMA,
1253 .align_mask = 0xf,
1254 .nop = SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0),
1255 .support_64bit_ptrs = false,
1256 .get_rptr = cik_sdma_ring_get_rptr,
1257 .get_wptr = cik_sdma_ring_get_wptr,
1258 .set_wptr = cik_sdma_ring_set_wptr,
1259 .emit_frame_size =
1260 6 + /* cik_sdma_ring_emit_hdp_flush */
1261 3 + /* hdp invalidate */
1262 6 + /* cik_sdma_ring_emit_pipeline_sync */
1263 CIK_FLUSH_GPU_TLB_NUM_WREG * 3 + 6 + /* cik_sdma_ring_emit_vm_flush */
1264 9 + 9 + 9, /* cik_sdma_ring_emit_fence x3 for user fence, vm fence */
1265 .emit_ib_size = 7 + 4, /* cik_sdma_ring_emit_ib */
1266 .emit_ib = cik_sdma_ring_emit_ib,
1267 .emit_fence = cik_sdma_ring_emit_fence,
1268 .emit_pipeline_sync = cik_sdma_ring_emit_pipeline_sync,
1269 .emit_vm_flush = cik_sdma_ring_emit_vm_flush,
1270 .emit_hdp_flush = cik_sdma_ring_emit_hdp_flush,
1271 .test_ring = cik_sdma_ring_test_ring,
1272 .test_ib = cik_sdma_ring_test_ib,
1273 .insert_nop = cik_sdma_ring_insert_nop,
1274 .pad_ib = cik_sdma_ring_pad_ib,
1275 .emit_wreg = cik_sdma_ring_emit_wreg,
1276};
1277
1278static void cik_sdma_set_ring_funcs(struct amdgpu_device *adev)
1279{
1280 int i;
1281
1282 for (i = 0; i < adev->sdma.num_instances; i++) {
1283 adev->sdma.instance[i].ring.funcs = &cik_sdma_ring_funcs;
1284 adev->sdma.instance[i].ring.me = i;
1285 }
1286}
1287
1288static const struct amdgpu_irq_src_funcs cik_sdma_trap_irq_funcs = {
1289 .set = cik_sdma_set_trap_irq_state,
1290 .process = cik_sdma_process_trap_irq,
1291};
1292
1293static const struct amdgpu_irq_src_funcs cik_sdma_illegal_inst_irq_funcs = {
1294 .process = cik_sdma_process_illegal_inst_irq,
1295};
1296
1297static void cik_sdma_set_irq_funcs(struct amdgpu_device *adev)
1298{
1299 adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
1300 adev->sdma.trap_irq.funcs = &cik_sdma_trap_irq_funcs;
1301 adev->sdma.illegal_inst_irq.funcs = &cik_sdma_illegal_inst_irq_funcs;
1302}
1303
1304/**
1305 * cik_sdma_emit_copy_buffer - copy buffer using the sDMA engine
1306 *
1307 * @ib: indirect buffer to copy to
1308 * @src_offset: src GPU address
1309 * @dst_offset: dst GPU address
1310 * @byte_count: number of bytes to xfer
1311 * @tmz: is this a secure operation
1312 *
1313 * Copy GPU buffers using the DMA engine (CIK).
1314 * Used by the amdgpu ttm implementation to move pages if
1315 * registered as the asic copy callback.
1316 */
1317static void cik_sdma_emit_copy_buffer(struct amdgpu_ib *ib,
1318 uint64_t src_offset,
1319 uint64_t dst_offset,
1320 uint32_t byte_count,
1321 bool tmz)
1322{
1323 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_COPY, SDMA_COPY_SUB_OPCODE_LINEAR, 0);
1324 ib->ptr[ib->length_dw++] = byte_count;
1325 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1326 ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1327 ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1328 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1329 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1330}
1331
1332/**
1333 * cik_sdma_emit_fill_buffer - fill buffer using the sDMA engine
1334 *
1335 * @ib: indirect buffer to fill
1336 * @src_data: value to write to buffer
1337 * @dst_offset: dst GPU address
1338 * @byte_count: number of bytes to xfer
1339 *
1340 * Fill GPU buffers using the DMA engine (CIK).
1341 */
1342static void cik_sdma_emit_fill_buffer(struct amdgpu_ib *ib,
1343 uint32_t src_data,
1344 uint64_t dst_offset,
1345 uint32_t byte_count)
1346{
1347 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_CONSTANT_FILL, 0, 0);
1348 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1349 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1350 ib->ptr[ib->length_dw++] = src_data;
1351 ib->ptr[ib->length_dw++] = byte_count;
1352}
1353
1354static const struct amdgpu_buffer_funcs cik_sdma_buffer_funcs = {
1355 .copy_max_bytes = 0x1fffff,
1356 .copy_num_dw = 7,
1357 .emit_copy_buffer = cik_sdma_emit_copy_buffer,
1358
1359 .fill_max_bytes = 0x1fffff,
1360 .fill_num_dw = 5,
1361 .emit_fill_buffer = cik_sdma_emit_fill_buffer,
1362};
1363
1364static void cik_sdma_set_buffer_funcs(struct amdgpu_device *adev)
1365{
1366 adev->mman.buffer_funcs = &cik_sdma_buffer_funcs;
1367 adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1368}
1369
1370static const struct amdgpu_vm_pte_funcs cik_sdma_vm_pte_funcs = {
1371 .copy_pte_num_dw = 7,
1372 .copy_pte = cik_sdma_vm_copy_pte,
1373
1374 .write_pte = cik_sdma_vm_write_pte,
1375 .set_pte_pde = cik_sdma_vm_set_pte_pde,
1376};
1377
1378static void cik_sdma_set_vm_pte_funcs(struct amdgpu_device *adev)
1379{
1380 unsigned i;
1381
1382 adev->vm_manager.vm_pte_funcs = &cik_sdma_vm_pte_funcs;
1383 for (i = 0; i < adev->sdma.num_instances; i++) {
1384 adev->vm_manager.vm_pte_scheds[i] =
1385 &adev->sdma.instance[i].ring.sched;
1386 }
1387 adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
1388}
1389
1390const struct amdgpu_ip_block_version cik_sdma_ip_block =
1391{
1392 .type = AMD_IP_BLOCK_TYPE_SDMA,
1393 .major = 2,
1394 .minor = 0,
1395 .rev = 0,
1396 .funcs = &cik_sdma_ip_funcs,
1397};