Loading...
1/*
2 * Copyright 2015 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 "amdgpu.h"
26#include "amdgpu_trace.h"
27#include "si.h"
28#include "sid.h"
29
30const u32 sdma_offsets[SDMA_MAX_INSTANCE] =
31{
32 DMA0_REGISTER_OFFSET,
33 DMA1_REGISTER_OFFSET
34};
35
36static void si_dma_set_ring_funcs(struct amdgpu_device *adev);
37static void si_dma_set_buffer_funcs(struct amdgpu_device *adev);
38static void si_dma_set_vm_pte_funcs(struct amdgpu_device *adev);
39static void si_dma_set_irq_funcs(struct amdgpu_device *adev);
40
41static uint64_t si_dma_ring_get_rptr(struct amdgpu_ring *ring)
42{
43 return *ring->rptr_cpu_addr;
44}
45
46static uint64_t si_dma_ring_get_wptr(struct amdgpu_ring *ring)
47{
48 struct amdgpu_device *adev = ring->adev;
49 u32 me = (ring == &adev->sdma.instance[0].ring) ? 0 : 1;
50
51 return (RREG32(DMA_RB_WPTR + sdma_offsets[me]) & 0x3fffc) >> 2;
52}
53
54static void si_dma_ring_set_wptr(struct amdgpu_ring *ring)
55{
56 struct amdgpu_device *adev = ring->adev;
57 u32 me = (ring == &adev->sdma.instance[0].ring) ? 0 : 1;
58
59 WREG32(DMA_RB_WPTR + sdma_offsets[me], (ring->wptr << 2) & 0x3fffc);
60}
61
62static void si_dma_ring_emit_ib(struct amdgpu_ring *ring,
63 struct amdgpu_job *job,
64 struct amdgpu_ib *ib,
65 uint32_t flags)
66{
67 unsigned vmid = AMDGPU_JOB_GET_VMID(job);
68 /* The indirect buffer packet must end on an 8 DW boundary in the DMA ring.
69 * Pad as necessary with NOPs.
70 */
71 while ((lower_32_bits(ring->wptr) & 7) != 5)
72 amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0));
73 amdgpu_ring_write(ring, DMA_IB_PACKET(DMA_PACKET_INDIRECT_BUFFER, vmid, 0));
74 amdgpu_ring_write(ring, (ib->gpu_addr & 0xFFFFFFE0));
75 amdgpu_ring_write(ring, (ib->length_dw << 12) | (upper_32_bits(ib->gpu_addr) & 0xFF));
76
77}
78
79/**
80 * si_dma_ring_emit_fence - emit a fence on the DMA ring
81 *
82 * @ring: amdgpu ring pointer
83 * @addr: address
84 * @seq: sequence number
85 * @flags: fence related flags
86 *
87 * Add a DMA fence packet to the ring to write
88 * the fence seq number and DMA trap packet to generate
89 * an interrupt if needed (VI).
90 */
91static void si_dma_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
92 unsigned flags)
93{
94
95 bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
96 /* write the fence */
97 amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_FENCE, 0, 0, 0, 0));
98 amdgpu_ring_write(ring, addr & 0xfffffffc);
99 amdgpu_ring_write(ring, (upper_32_bits(addr) & 0xff));
100 amdgpu_ring_write(ring, seq);
101 /* optionally write high bits as well */
102 if (write64bit) {
103 addr += 4;
104 amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_FENCE, 0, 0, 0, 0));
105 amdgpu_ring_write(ring, addr & 0xfffffffc);
106 amdgpu_ring_write(ring, (upper_32_bits(addr) & 0xff));
107 amdgpu_ring_write(ring, upper_32_bits(seq));
108 }
109 /* generate an interrupt */
110 amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_TRAP, 0, 0, 0, 0));
111}
112
113static void si_dma_stop(struct amdgpu_device *adev)
114{
115 u32 rb_cntl;
116 unsigned i;
117
118 for (i = 0; i < adev->sdma.num_instances; i++) {
119 /* dma0 */
120 rb_cntl = RREG32(DMA_RB_CNTL + sdma_offsets[i]);
121 rb_cntl &= ~DMA_RB_ENABLE;
122 WREG32(DMA_RB_CNTL + sdma_offsets[i], rb_cntl);
123 }
124}
125
126static int si_dma_start(struct amdgpu_device *adev)
127{
128 struct amdgpu_ring *ring;
129 u32 rb_cntl, dma_cntl, ib_cntl, rb_bufsz;
130 int i, r;
131 uint64_t rptr_addr;
132
133 for (i = 0; i < adev->sdma.num_instances; i++) {
134 ring = &adev->sdma.instance[i].ring;
135
136 WREG32(DMA_SEM_INCOMPLETE_TIMER_CNTL + sdma_offsets[i], 0);
137 WREG32(DMA_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
138
139 /* Set ring buffer size in dwords */
140 rb_bufsz = order_base_2(ring->ring_size / 4);
141 rb_cntl = rb_bufsz << 1;
142#ifdef __BIG_ENDIAN
143 rb_cntl |= DMA_RB_SWAP_ENABLE | DMA_RPTR_WRITEBACK_SWAP_ENABLE;
144#endif
145 WREG32(DMA_RB_CNTL + sdma_offsets[i], rb_cntl);
146
147 /* Initialize the ring buffer's read and write pointers */
148 WREG32(DMA_RB_RPTR + sdma_offsets[i], 0);
149 WREG32(DMA_RB_WPTR + sdma_offsets[i], 0);
150
151 rptr_addr = ring->rptr_gpu_addr;
152
153 WREG32(DMA_RB_RPTR_ADDR_LO + sdma_offsets[i], lower_32_bits(rptr_addr));
154 WREG32(DMA_RB_RPTR_ADDR_HI + sdma_offsets[i], upper_32_bits(rptr_addr) & 0xFF);
155
156 rb_cntl |= DMA_RPTR_WRITEBACK_ENABLE;
157
158 WREG32(DMA_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
159
160 /* enable DMA IBs */
161 ib_cntl = DMA_IB_ENABLE | CMD_VMID_FORCE;
162#ifdef __BIG_ENDIAN
163 ib_cntl |= DMA_IB_SWAP_ENABLE;
164#endif
165 WREG32(DMA_IB_CNTL + sdma_offsets[i], ib_cntl);
166
167 dma_cntl = RREG32(DMA_CNTL + sdma_offsets[i]);
168 dma_cntl &= ~CTXEMPTY_INT_ENABLE;
169 WREG32(DMA_CNTL + sdma_offsets[i], dma_cntl);
170
171 ring->wptr = 0;
172 WREG32(DMA_RB_WPTR + sdma_offsets[i], ring->wptr << 2);
173 WREG32(DMA_RB_CNTL + sdma_offsets[i], rb_cntl | DMA_RB_ENABLE);
174
175 r = amdgpu_ring_test_helper(ring);
176 if (r)
177 return r;
178 }
179
180 return 0;
181}
182
183/**
184 * si_dma_ring_test_ring - simple async dma engine test
185 *
186 * @ring: amdgpu_ring structure holding ring information
187 *
188 * Test the DMA engine by writing using it to write an
189 * value to memory. (VI).
190 * Returns 0 for success, error for failure.
191 */
192static int si_dma_ring_test_ring(struct amdgpu_ring *ring)
193{
194 struct amdgpu_device *adev = ring->adev;
195 unsigned i;
196 unsigned index;
197 int r;
198 u32 tmp;
199 u64 gpu_addr;
200
201 r = amdgpu_device_wb_get(adev, &index);
202 if (r)
203 return r;
204
205 gpu_addr = adev->wb.gpu_addr + (index * 4);
206 tmp = 0xCAFEDEAD;
207 adev->wb.wb[index] = cpu_to_le32(tmp);
208
209 r = amdgpu_ring_alloc(ring, 4);
210 if (r)
211 goto error_free_wb;
212
213 amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 0, 1));
214 amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
215 amdgpu_ring_write(ring, upper_32_bits(gpu_addr) & 0xff);
216 amdgpu_ring_write(ring, 0xDEADBEEF);
217 amdgpu_ring_commit(ring);
218
219 for (i = 0; i < adev->usec_timeout; i++) {
220 tmp = le32_to_cpu(adev->wb.wb[index]);
221 if (tmp == 0xDEADBEEF)
222 break;
223 udelay(1);
224 }
225
226 if (i >= adev->usec_timeout)
227 r = -ETIMEDOUT;
228
229error_free_wb:
230 amdgpu_device_wb_free(adev, index);
231 return r;
232}
233
234/**
235 * si_dma_ring_test_ib - test an IB on the DMA engine
236 *
237 * @ring: amdgpu_ring structure holding ring information
238 * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
239 *
240 * Test a simple IB in the DMA ring (VI).
241 * Returns 0 on success, error on failure.
242 */
243static int si_dma_ring_test_ib(struct amdgpu_ring *ring, long timeout)
244{
245 struct amdgpu_device *adev = ring->adev;
246 struct amdgpu_ib ib;
247 struct dma_fence *f = NULL;
248 unsigned index;
249 u32 tmp = 0;
250 u64 gpu_addr;
251 long r;
252
253 r = amdgpu_device_wb_get(adev, &index);
254 if (r)
255 return r;
256
257 gpu_addr = adev->wb.gpu_addr + (index * 4);
258 tmp = 0xCAFEDEAD;
259 adev->wb.wb[index] = cpu_to_le32(tmp);
260 memset(&ib, 0, sizeof(ib));
261 r = amdgpu_ib_get(adev, NULL, 256,
262 AMDGPU_IB_POOL_DIRECT, &ib);
263 if (r)
264 goto err0;
265
266 ib.ptr[0] = DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 0, 1);
267 ib.ptr[1] = lower_32_bits(gpu_addr);
268 ib.ptr[2] = upper_32_bits(gpu_addr) & 0xff;
269 ib.ptr[3] = 0xDEADBEEF;
270 ib.length_dw = 4;
271 r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
272 if (r)
273 goto err1;
274
275 r = dma_fence_wait_timeout(f, false, timeout);
276 if (r == 0) {
277 r = -ETIMEDOUT;
278 goto err1;
279 } else if (r < 0) {
280 goto err1;
281 }
282 tmp = le32_to_cpu(adev->wb.wb[index]);
283 if (tmp == 0xDEADBEEF)
284 r = 0;
285 else
286 r = -EINVAL;
287
288err1:
289 amdgpu_ib_free(adev, &ib, NULL);
290 dma_fence_put(f);
291err0:
292 amdgpu_device_wb_free(adev, index);
293 return r;
294}
295
296/**
297 * si_dma_vm_copy_pte - update PTEs by copying them from the GART
298 *
299 * @ib: indirect buffer to fill with commands
300 * @pe: addr of the page entry
301 * @src: src addr to copy from
302 * @count: number of page entries to update
303 *
304 * Update PTEs by copying them from the GART using DMA (SI).
305 */
306static void si_dma_vm_copy_pte(struct amdgpu_ib *ib,
307 uint64_t pe, uint64_t src,
308 unsigned count)
309{
310 unsigned bytes = count * 8;
311
312 ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_COPY,
313 1, 0, 0, bytes);
314 ib->ptr[ib->length_dw++] = lower_32_bits(pe);
315 ib->ptr[ib->length_dw++] = lower_32_bits(src);
316 ib->ptr[ib->length_dw++] = upper_32_bits(pe) & 0xff;
317 ib->ptr[ib->length_dw++] = upper_32_bits(src) & 0xff;
318}
319
320/**
321 * si_dma_vm_write_pte - update PTEs by writing them manually
322 *
323 * @ib: indirect buffer to fill with commands
324 * @pe: addr of the page entry
325 * @value: dst addr to write into pe
326 * @count: number of page entries to update
327 * @incr: increase next addr by incr bytes
328 *
329 * Update PTEs by writing them manually using DMA (SI).
330 */
331static void si_dma_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
332 uint64_t value, unsigned count,
333 uint32_t incr)
334{
335 unsigned ndw = count * 2;
336
337 ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 0, ndw);
338 ib->ptr[ib->length_dw++] = lower_32_bits(pe);
339 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
340 for (; ndw > 0; ndw -= 2) {
341 ib->ptr[ib->length_dw++] = lower_32_bits(value);
342 ib->ptr[ib->length_dw++] = upper_32_bits(value);
343 value += incr;
344 }
345}
346
347/**
348 * si_dma_vm_set_pte_pde - update the page tables using sDMA
349 *
350 * @ib: indirect buffer to fill with commands
351 * @pe: addr of the page entry
352 * @addr: dst addr to write into pe
353 * @count: number of page entries to update
354 * @incr: increase next addr by incr bytes
355 * @flags: access flags
356 *
357 * Update the page tables using sDMA (CIK).
358 */
359static void si_dma_vm_set_pte_pde(struct amdgpu_ib *ib,
360 uint64_t pe,
361 uint64_t addr, unsigned count,
362 uint32_t incr, uint64_t flags)
363{
364 uint64_t value;
365 unsigned ndw;
366
367 while (count) {
368 ndw = count * 2;
369 if (ndw > 0xFFFFE)
370 ndw = 0xFFFFE;
371
372 if (flags & AMDGPU_PTE_VALID)
373 value = addr;
374 else
375 value = 0;
376
377 /* for physically contiguous pages (vram) */
378 ib->ptr[ib->length_dw++] = DMA_PTE_PDE_PACKET(ndw);
379 ib->ptr[ib->length_dw++] = pe; /* dst addr */
380 ib->ptr[ib->length_dw++] = upper_32_bits(pe) & 0xff;
381 ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
382 ib->ptr[ib->length_dw++] = upper_32_bits(flags);
383 ib->ptr[ib->length_dw++] = value; /* value */
384 ib->ptr[ib->length_dw++] = upper_32_bits(value);
385 ib->ptr[ib->length_dw++] = incr; /* increment size */
386 ib->ptr[ib->length_dw++] = 0;
387 pe += ndw * 4;
388 addr += (ndw / 2) * incr;
389 count -= ndw / 2;
390 }
391}
392
393/**
394 * si_dma_ring_pad_ib - pad the IB to the required number of dw
395 *
396 * @ring: amdgpu_ring pointer
397 * @ib: indirect buffer to fill with padding
398 *
399 */
400static void si_dma_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
401{
402 while (ib->length_dw & 0x7)
403 ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0);
404}
405
406/**
407 * si_dma_ring_emit_pipeline_sync - sync the pipeline
408 *
409 * @ring: amdgpu_ring pointer
410 *
411 * Make sure all previous operations are completed (CIK).
412 */
413static void si_dma_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
414{
415 uint32_t seq = ring->fence_drv.sync_seq;
416 uint64_t addr = ring->fence_drv.gpu_addr;
417
418 /* wait for idle */
419 amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_POLL_REG_MEM, 0, 0, 0, 0) |
420 (1 << 27)); /* Poll memory */
421 amdgpu_ring_write(ring, lower_32_bits(addr));
422 amdgpu_ring_write(ring, (0xff << 16) | upper_32_bits(addr)); /* retry, addr_hi */
423 amdgpu_ring_write(ring, 0xffffffff); /* mask */
424 amdgpu_ring_write(ring, seq); /* value */
425 amdgpu_ring_write(ring, (3 << 28) | 0x20); /* func(equal) | poll interval */
426}
427
428/**
429 * si_dma_ring_emit_vm_flush - cik vm flush using sDMA
430 *
431 * @ring: amdgpu_ring pointer
432 * @vmid: vmid number to use
433 * @pd_addr: address
434 *
435 * Update the page table base and flush the VM TLB
436 * using sDMA (VI).
437 */
438static void si_dma_ring_emit_vm_flush(struct amdgpu_ring *ring,
439 unsigned vmid, uint64_t pd_addr)
440{
441 amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
442
443 /* wait for invalidate to complete */
444 amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_POLL_REG_MEM, 0, 0, 0, 0));
445 amdgpu_ring_write(ring, VM_INVALIDATE_REQUEST);
446 amdgpu_ring_write(ring, 0xff << 16); /* retry */
447 amdgpu_ring_write(ring, 1 << vmid); /* mask */
448 amdgpu_ring_write(ring, 0); /* value */
449 amdgpu_ring_write(ring, (0 << 28) | 0x20); /* func(always) | poll interval */
450}
451
452static void si_dma_ring_emit_wreg(struct amdgpu_ring *ring,
453 uint32_t reg, uint32_t val)
454{
455 amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0, 0, 0));
456 amdgpu_ring_write(ring, (0xf << 16) | reg);
457 amdgpu_ring_write(ring, val);
458}
459
460static int si_dma_early_init(void *handle)
461{
462 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
463
464 adev->sdma.num_instances = 2;
465
466 si_dma_set_ring_funcs(adev);
467 si_dma_set_buffer_funcs(adev);
468 si_dma_set_vm_pte_funcs(adev);
469 si_dma_set_irq_funcs(adev);
470
471 return 0;
472}
473
474static int si_dma_sw_init(void *handle)
475{
476 struct amdgpu_ring *ring;
477 int r, i;
478 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
479
480 /* DMA0 trap event */
481 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 224,
482 &adev->sdma.trap_irq);
483 if (r)
484 return r;
485
486 /* DMA1 trap event */
487 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 244,
488 &adev->sdma.trap_irq);
489 if (r)
490 return r;
491
492 for (i = 0; i < adev->sdma.num_instances; i++) {
493 ring = &adev->sdma.instance[i].ring;
494 ring->ring_obj = NULL;
495 ring->use_doorbell = false;
496 sprintf(ring->name, "sdma%d", i);
497 r = amdgpu_ring_init(adev, ring, 1024,
498 &adev->sdma.trap_irq,
499 (i == 0) ? AMDGPU_SDMA_IRQ_INSTANCE0 :
500 AMDGPU_SDMA_IRQ_INSTANCE1,
501 AMDGPU_RING_PRIO_DEFAULT, NULL);
502 if (r)
503 return r;
504 }
505
506 return r;
507}
508
509static int si_dma_sw_fini(void *handle)
510{
511 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
512 int i;
513
514 for (i = 0; i < adev->sdma.num_instances; i++)
515 amdgpu_ring_fini(&adev->sdma.instance[i].ring);
516
517 return 0;
518}
519
520static int si_dma_hw_init(void *handle)
521{
522 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
523
524 return si_dma_start(adev);
525}
526
527static int si_dma_hw_fini(void *handle)
528{
529 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
530
531 si_dma_stop(adev);
532
533 return 0;
534}
535
536static int si_dma_suspend(void *handle)
537{
538 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
539
540 return si_dma_hw_fini(adev);
541}
542
543static int si_dma_resume(void *handle)
544{
545 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
546
547 return si_dma_hw_init(adev);
548}
549
550static bool si_dma_is_idle(void *handle)
551{
552 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
553 u32 tmp = RREG32(SRBM_STATUS2);
554
555 if (tmp & (DMA_BUSY_MASK | DMA1_BUSY_MASK))
556 return false;
557
558 return true;
559}
560
561static int si_dma_wait_for_idle(void *handle)
562{
563 unsigned i;
564 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
565
566 for (i = 0; i < adev->usec_timeout; i++) {
567 if (si_dma_is_idle(handle))
568 return 0;
569 udelay(1);
570 }
571 return -ETIMEDOUT;
572}
573
574static int si_dma_soft_reset(void *handle)
575{
576 DRM_INFO("si_dma_soft_reset --- not implemented !!!!!!!\n");
577 return 0;
578}
579
580static int si_dma_set_trap_irq_state(struct amdgpu_device *adev,
581 struct amdgpu_irq_src *src,
582 unsigned type,
583 enum amdgpu_interrupt_state state)
584{
585 u32 sdma_cntl;
586
587 switch (type) {
588 case AMDGPU_SDMA_IRQ_INSTANCE0:
589 switch (state) {
590 case AMDGPU_IRQ_STATE_DISABLE:
591 sdma_cntl = RREG32(DMA_CNTL + DMA0_REGISTER_OFFSET);
592 sdma_cntl &= ~TRAP_ENABLE;
593 WREG32(DMA_CNTL + DMA0_REGISTER_OFFSET, sdma_cntl);
594 break;
595 case AMDGPU_IRQ_STATE_ENABLE:
596 sdma_cntl = RREG32(DMA_CNTL + DMA0_REGISTER_OFFSET);
597 sdma_cntl |= TRAP_ENABLE;
598 WREG32(DMA_CNTL + DMA0_REGISTER_OFFSET, sdma_cntl);
599 break;
600 default:
601 break;
602 }
603 break;
604 case AMDGPU_SDMA_IRQ_INSTANCE1:
605 switch (state) {
606 case AMDGPU_IRQ_STATE_DISABLE:
607 sdma_cntl = RREG32(DMA_CNTL + DMA1_REGISTER_OFFSET);
608 sdma_cntl &= ~TRAP_ENABLE;
609 WREG32(DMA_CNTL + DMA1_REGISTER_OFFSET, sdma_cntl);
610 break;
611 case AMDGPU_IRQ_STATE_ENABLE:
612 sdma_cntl = RREG32(DMA_CNTL + DMA1_REGISTER_OFFSET);
613 sdma_cntl |= TRAP_ENABLE;
614 WREG32(DMA_CNTL + DMA1_REGISTER_OFFSET, sdma_cntl);
615 break;
616 default:
617 break;
618 }
619 break;
620 default:
621 break;
622 }
623 return 0;
624}
625
626static int si_dma_process_trap_irq(struct amdgpu_device *adev,
627 struct amdgpu_irq_src *source,
628 struct amdgpu_iv_entry *entry)
629{
630 if (entry->src_id == 224)
631 amdgpu_fence_process(&adev->sdma.instance[0].ring);
632 else
633 amdgpu_fence_process(&adev->sdma.instance[1].ring);
634 return 0;
635}
636
637static int si_dma_set_clockgating_state(void *handle,
638 enum amd_clockgating_state state)
639{
640 u32 orig, data, offset;
641 int i;
642 bool enable;
643 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
644
645 enable = (state == AMD_CG_STATE_GATE);
646
647 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
648 for (i = 0; i < adev->sdma.num_instances; i++) {
649 if (i == 0)
650 offset = DMA0_REGISTER_OFFSET;
651 else
652 offset = DMA1_REGISTER_OFFSET;
653 orig = data = RREG32(DMA_POWER_CNTL + offset);
654 data &= ~MEM_POWER_OVERRIDE;
655 if (data != orig)
656 WREG32(DMA_POWER_CNTL + offset, data);
657 WREG32(DMA_CLK_CTRL + offset, 0x00000100);
658 }
659 } else {
660 for (i = 0; i < adev->sdma.num_instances; i++) {
661 if (i == 0)
662 offset = DMA0_REGISTER_OFFSET;
663 else
664 offset = DMA1_REGISTER_OFFSET;
665 orig = data = RREG32(DMA_POWER_CNTL + offset);
666 data |= MEM_POWER_OVERRIDE;
667 if (data != orig)
668 WREG32(DMA_POWER_CNTL + offset, data);
669
670 orig = data = RREG32(DMA_CLK_CTRL + offset);
671 data = 0xff000000;
672 if (data != orig)
673 WREG32(DMA_CLK_CTRL + offset, data);
674 }
675 }
676
677 return 0;
678}
679
680static int si_dma_set_powergating_state(void *handle,
681 enum amd_powergating_state state)
682{
683 u32 tmp;
684
685 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
686
687 WREG32(DMA_PGFSM_WRITE, 0x00002000);
688 WREG32(DMA_PGFSM_CONFIG, 0x100010ff);
689
690 for (tmp = 0; tmp < 5; tmp++)
691 WREG32(DMA_PGFSM_WRITE, 0);
692
693 return 0;
694}
695
696static const struct amd_ip_funcs si_dma_ip_funcs = {
697 .name = "si_dma",
698 .early_init = si_dma_early_init,
699 .late_init = NULL,
700 .sw_init = si_dma_sw_init,
701 .sw_fini = si_dma_sw_fini,
702 .hw_init = si_dma_hw_init,
703 .hw_fini = si_dma_hw_fini,
704 .suspend = si_dma_suspend,
705 .resume = si_dma_resume,
706 .is_idle = si_dma_is_idle,
707 .wait_for_idle = si_dma_wait_for_idle,
708 .soft_reset = si_dma_soft_reset,
709 .set_clockgating_state = si_dma_set_clockgating_state,
710 .set_powergating_state = si_dma_set_powergating_state,
711};
712
713static const struct amdgpu_ring_funcs si_dma_ring_funcs = {
714 .type = AMDGPU_RING_TYPE_SDMA,
715 .align_mask = 0xf,
716 .nop = DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0),
717 .support_64bit_ptrs = false,
718 .get_rptr = si_dma_ring_get_rptr,
719 .get_wptr = si_dma_ring_get_wptr,
720 .set_wptr = si_dma_ring_set_wptr,
721 .emit_frame_size =
722 3 + 3 + /* hdp flush / invalidate */
723 6 + /* si_dma_ring_emit_pipeline_sync */
724 SI_FLUSH_GPU_TLB_NUM_WREG * 3 + 6 + /* si_dma_ring_emit_vm_flush */
725 9 + 9 + 9, /* si_dma_ring_emit_fence x3 for user fence, vm fence */
726 .emit_ib_size = 7 + 3, /* si_dma_ring_emit_ib */
727 .emit_ib = si_dma_ring_emit_ib,
728 .emit_fence = si_dma_ring_emit_fence,
729 .emit_pipeline_sync = si_dma_ring_emit_pipeline_sync,
730 .emit_vm_flush = si_dma_ring_emit_vm_flush,
731 .test_ring = si_dma_ring_test_ring,
732 .test_ib = si_dma_ring_test_ib,
733 .insert_nop = amdgpu_ring_insert_nop,
734 .pad_ib = si_dma_ring_pad_ib,
735 .emit_wreg = si_dma_ring_emit_wreg,
736};
737
738static void si_dma_set_ring_funcs(struct amdgpu_device *adev)
739{
740 int i;
741
742 for (i = 0; i < adev->sdma.num_instances; i++)
743 adev->sdma.instance[i].ring.funcs = &si_dma_ring_funcs;
744}
745
746static const struct amdgpu_irq_src_funcs si_dma_trap_irq_funcs = {
747 .set = si_dma_set_trap_irq_state,
748 .process = si_dma_process_trap_irq,
749};
750
751static void si_dma_set_irq_funcs(struct amdgpu_device *adev)
752{
753 adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
754 adev->sdma.trap_irq.funcs = &si_dma_trap_irq_funcs;
755}
756
757/**
758 * si_dma_emit_copy_buffer - copy buffer using the sDMA engine
759 *
760 * @ib: indirect buffer to copy to
761 * @src_offset: src GPU address
762 * @dst_offset: dst GPU address
763 * @byte_count: number of bytes to xfer
764 * @tmz: is this a secure operation
765 *
766 * Copy GPU buffers using the DMA engine (VI).
767 * Used by the amdgpu ttm implementation to move pages if
768 * registered as the asic copy callback.
769 */
770static void si_dma_emit_copy_buffer(struct amdgpu_ib *ib,
771 uint64_t src_offset,
772 uint64_t dst_offset,
773 uint32_t byte_count,
774 bool tmz)
775{
776 ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_COPY,
777 1, 0, 0, byte_count);
778 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
779 ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
780 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset) & 0xff;
781 ib->ptr[ib->length_dw++] = upper_32_bits(src_offset) & 0xff;
782}
783
784/**
785 * si_dma_emit_fill_buffer - fill buffer using the sDMA engine
786 *
787 * @ib: indirect buffer to copy to
788 * @src_data: value to write to buffer
789 * @dst_offset: dst GPU address
790 * @byte_count: number of bytes to xfer
791 *
792 * Fill GPU buffers using the DMA engine (VI).
793 */
794static void si_dma_emit_fill_buffer(struct amdgpu_ib *ib,
795 uint32_t src_data,
796 uint64_t dst_offset,
797 uint32_t byte_count)
798{
799 ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_CONSTANT_FILL,
800 0, 0, 0, byte_count / 4);
801 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
802 ib->ptr[ib->length_dw++] = src_data;
803 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset) << 16;
804}
805
806
807static const struct amdgpu_buffer_funcs si_dma_buffer_funcs = {
808 .copy_max_bytes = 0xffff8,
809 .copy_num_dw = 5,
810 .emit_copy_buffer = si_dma_emit_copy_buffer,
811
812 .fill_max_bytes = 0xffff8,
813 .fill_num_dw = 4,
814 .emit_fill_buffer = si_dma_emit_fill_buffer,
815};
816
817static void si_dma_set_buffer_funcs(struct amdgpu_device *adev)
818{
819 adev->mman.buffer_funcs = &si_dma_buffer_funcs;
820 adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
821}
822
823static const struct amdgpu_vm_pte_funcs si_dma_vm_pte_funcs = {
824 .copy_pte_num_dw = 5,
825 .copy_pte = si_dma_vm_copy_pte,
826
827 .write_pte = si_dma_vm_write_pte,
828 .set_pte_pde = si_dma_vm_set_pte_pde,
829};
830
831static void si_dma_set_vm_pte_funcs(struct amdgpu_device *adev)
832{
833 unsigned i;
834
835 adev->vm_manager.vm_pte_funcs = &si_dma_vm_pte_funcs;
836 for (i = 0; i < adev->sdma.num_instances; i++) {
837 adev->vm_manager.vm_pte_scheds[i] =
838 &adev->sdma.instance[i].ring.sched;
839 }
840 adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
841}
842
843const struct amdgpu_ip_block_version si_dma_ip_block =
844{
845 .type = AMD_IP_BLOCK_TYPE_SDMA,
846 .major = 1,
847 .minor = 0,
848 .rev = 0,
849 .funcs = &si_dma_ip_funcs,
850};
1/*
2 * Copyright 2015 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 <drm/drmP.h>
25#include "amdgpu.h"
26#include "amdgpu_trace.h"
27#include "si.h"
28#include "sid.h"
29
30const u32 sdma_offsets[SDMA_MAX_INSTANCE] =
31{
32 DMA0_REGISTER_OFFSET,
33 DMA1_REGISTER_OFFSET
34};
35
36static void si_dma_set_ring_funcs(struct amdgpu_device *adev);
37static void si_dma_set_buffer_funcs(struct amdgpu_device *adev);
38static void si_dma_set_vm_pte_funcs(struct amdgpu_device *adev);
39static void si_dma_set_irq_funcs(struct amdgpu_device *adev);
40
41static uint64_t si_dma_ring_get_rptr(struct amdgpu_ring *ring)
42{
43 return ring->adev->wb.wb[ring->rptr_offs>>2];
44}
45
46static uint64_t si_dma_ring_get_wptr(struct amdgpu_ring *ring)
47{
48 struct amdgpu_device *adev = ring->adev;
49 u32 me = (ring == &adev->sdma.instance[0].ring) ? 0 : 1;
50
51 return (RREG32(DMA_RB_WPTR + sdma_offsets[me]) & 0x3fffc) >> 2;
52}
53
54static void si_dma_ring_set_wptr(struct amdgpu_ring *ring)
55{
56 struct amdgpu_device *adev = ring->adev;
57 u32 me = (ring == &adev->sdma.instance[0].ring) ? 0 : 1;
58
59 WREG32(DMA_RB_WPTR + sdma_offsets[me],
60 (lower_32_bits(ring->wptr) << 2) & 0x3fffc);
61}
62
63static void si_dma_ring_emit_ib(struct amdgpu_ring *ring,
64 struct amdgpu_ib *ib,
65 unsigned vmid, bool ctx_switch)
66{
67 /* The indirect buffer packet must end on an 8 DW boundary in the DMA ring.
68 * Pad as necessary with NOPs.
69 */
70 while ((lower_32_bits(ring->wptr) & 7) != 5)
71 amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0));
72 amdgpu_ring_write(ring, DMA_IB_PACKET(DMA_PACKET_INDIRECT_BUFFER, vmid, 0));
73 amdgpu_ring_write(ring, (ib->gpu_addr & 0xFFFFFFE0));
74 amdgpu_ring_write(ring, (ib->length_dw << 12) | (upper_32_bits(ib->gpu_addr) & 0xFF));
75
76}
77
78/**
79 * si_dma_ring_emit_fence - emit a fence on the DMA ring
80 *
81 * @ring: amdgpu ring pointer
82 * @fence: amdgpu fence object
83 *
84 * Add a DMA fence packet to the ring to write
85 * the fence seq number and DMA trap packet to generate
86 * an interrupt if needed (VI).
87 */
88static void si_dma_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
89 unsigned flags)
90{
91
92 bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
93 /* write the fence */
94 amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_FENCE, 0, 0, 0, 0));
95 amdgpu_ring_write(ring, addr & 0xfffffffc);
96 amdgpu_ring_write(ring, (upper_32_bits(addr) & 0xff));
97 amdgpu_ring_write(ring, seq);
98 /* optionally write high bits as well */
99 if (write64bit) {
100 addr += 4;
101 amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_FENCE, 0, 0, 0, 0));
102 amdgpu_ring_write(ring, addr & 0xfffffffc);
103 amdgpu_ring_write(ring, (upper_32_bits(addr) & 0xff));
104 amdgpu_ring_write(ring, upper_32_bits(seq));
105 }
106 /* generate an interrupt */
107 amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_TRAP, 0, 0, 0, 0));
108}
109
110static void si_dma_stop(struct amdgpu_device *adev)
111{
112 struct amdgpu_ring *ring;
113 u32 rb_cntl;
114 unsigned i;
115
116 for (i = 0; i < adev->sdma.num_instances; i++) {
117 ring = &adev->sdma.instance[i].ring;
118 /* dma0 */
119 rb_cntl = RREG32(DMA_RB_CNTL + sdma_offsets[i]);
120 rb_cntl &= ~DMA_RB_ENABLE;
121 WREG32(DMA_RB_CNTL + sdma_offsets[i], rb_cntl);
122
123 if (adev->mman.buffer_funcs_ring == ring)
124 amdgpu_ttm_set_buffer_funcs_status(adev, false);
125 ring->ready = false;
126 }
127}
128
129static int si_dma_start(struct amdgpu_device *adev)
130{
131 struct amdgpu_ring *ring;
132 u32 rb_cntl, dma_cntl, ib_cntl, rb_bufsz;
133 int i, r;
134 uint64_t rptr_addr;
135
136 for (i = 0; i < adev->sdma.num_instances; i++) {
137 ring = &adev->sdma.instance[i].ring;
138
139 WREG32(DMA_SEM_INCOMPLETE_TIMER_CNTL + sdma_offsets[i], 0);
140 WREG32(DMA_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
141
142 /* Set ring buffer size in dwords */
143 rb_bufsz = order_base_2(ring->ring_size / 4);
144 rb_cntl = rb_bufsz << 1;
145#ifdef __BIG_ENDIAN
146 rb_cntl |= DMA_RB_SWAP_ENABLE | DMA_RPTR_WRITEBACK_SWAP_ENABLE;
147#endif
148 WREG32(DMA_RB_CNTL + sdma_offsets[i], rb_cntl);
149
150 /* Initialize the ring buffer's read and write pointers */
151 WREG32(DMA_RB_RPTR + sdma_offsets[i], 0);
152 WREG32(DMA_RB_WPTR + sdma_offsets[i], 0);
153
154 rptr_addr = adev->wb.gpu_addr + (ring->rptr_offs * 4);
155
156 WREG32(DMA_RB_RPTR_ADDR_LO + sdma_offsets[i], lower_32_bits(rptr_addr));
157 WREG32(DMA_RB_RPTR_ADDR_HI + sdma_offsets[i], upper_32_bits(rptr_addr) & 0xFF);
158
159 rb_cntl |= DMA_RPTR_WRITEBACK_ENABLE;
160
161 WREG32(DMA_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
162
163 /* enable DMA IBs */
164 ib_cntl = DMA_IB_ENABLE | CMD_VMID_FORCE;
165#ifdef __BIG_ENDIAN
166 ib_cntl |= DMA_IB_SWAP_ENABLE;
167#endif
168 WREG32(DMA_IB_CNTL + sdma_offsets[i], ib_cntl);
169
170 dma_cntl = RREG32(DMA_CNTL + sdma_offsets[i]);
171 dma_cntl &= ~CTXEMPTY_INT_ENABLE;
172 WREG32(DMA_CNTL + sdma_offsets[i], dma_cntl);
173
174 ring->wptr = 0;
175 WREG32(DMA_RB_WPTR + sdma_offsets[i], lower_32_bits(ring->wptr) << 2);
176 WREG32(DMA_RB_CNTL + sdma_offsets[i], rb_cntl | DMA_RB_ENABLE);
177
178 ring->ready = true;
179
180 r = amdgpu_ring_test_ring(ring);
181 if (r) {
182 ring->ready = false;
183 return r;
184 }
185
186 if (adev->mman.buffer_funcs_ring == ring)
187 amdgpu_ttm_set_buffer_funcs_status(adev, true);
188 }
189
190 return 0;
191}
192
193/**
194 * si_dma_ring_test_ring - simple async dma engine test
195 *
196 * @ring: amdgpu_ring structure holding ring information
197 *
198 * Test the DMA engine by writing using it to write an
199 * value to memory. (VI).
200 * Returns 0 for success, error for failure.
201 */
202static int si_dma_ring_test_ring(struct amdgpu_ring *ring)
203{
204 struct amdgpu_device *adev = ring->adev;
205 unsigned i;
206 unsigned index;
207 int r;
208 u32 tmp;
209 u64 gpu_addr;
210
211 r = amdgpu_device_wb_get(adev, &index);
212 if (r) {
213 dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
214 return r;
215 }
216
217 gpu_addr = adev->wb.gpu_addr + (index * 4);
218 tmp = 0xCAFEDEAD;
219 adev->wb.wb[index] = cpu_to_le32(tmp);
220
221 r = amdgpu_ring_alloc(ring, 4);
222 if (r) {
223 DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
224 amdgpu_device_wb_free(adev, index);
225 return r;
226 }
227
228 amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 0, 1));
229 amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
230 amdgpu_ring_write(ring, upper_32_bits(gpu_addr) & 0xff);
231 amdgpu_ring_write(ring, 0xDEADBEEF);
232 amdgpu_ring_commit(ring);
233
234 for (i = 0; i < adev->usec_timeout; i++) {
235 tmp = le32_to_cpu(adev->wb.wb[index]);
236 if (tmp == 0xDEADBEEF)
237 break;
238 DRM_UDELAY(1);
239 }
240
241 if (i < adev->usec_timeout) {
242 DRM_DEBUG("ring test on %d succeeded in %d usecs\n", ring->idx, i);
243 } else {
244 DRM_ERROR("amdgpu: ring %d test failed (0x%08X)\n",
245 ring->idx, tmp);
246 r = -EINVAL;
247 }
248 amdgpu_device_wb_free(adev, index);
249
250 return r;
251}
252
253/**
254 * si_dma_ring_test_ib - test an IB on the DMA engine
255 *
256 * @ring: amdgpu_ring structure holding ring information
257 *
258 * Test a simple IB in the DMA ring (VI).
259 * Returns 0 on success, error on failure.
260 */
261static int si_dma_ring_test_ib(struct amdgpu_ring *ring, long timeout)
262{
263 struct amdgpu_device *adev = ring->adev;
264 struct amdgpu_ib ib;
265 struct dma_fence *f = NULL;
266 unsigned index;
267 u32 tmp = 0;
268 u64 gpu_addr;
269 long r;
270
271 r = amdgpu_device_wb_get(adev, &index);
272 if (r) {
273 dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
274 return r;
275 }
276
277 gpu_addr = adev->wb.gpu_addr + (index * 4);
278 tmp = 0xCAFEDEAD;
279 adev->wb.wb[index] = cpu_to_le32(tmp);
280 memset(&ib, 0, sizeof(ib));
281 r = amdgpu_ib_get(adev, NULL, 256, &ib);
282 if (r) {
283 DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
284 goto err0;
285 }
286
287 ib.ptr[0] = DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 0, 1);
288 ib.ptr[1] = lower_32_bits(gpu_addr);
289 ib.ptr[2] = upper_32_bits(gpu_addr) & 0xff;
290 ib.ptr[3] = 0xDEADBEEF;
291 ib.length_dw = 4;
292 r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
293 if (r)
294 goto err1;
295
296 r = dma_fence_wait_timeout(f, false, timeout);
297 if (r == 0) {
298 DRM_ERROR("amdgpu: IB test timed out\n");
299 r = -ETIMEDOUT;
300 goto err1;
301 } else if (r < 0) {
302 DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
303 goto err1;
304 }
305 tmp = le32_to_cpu(adev->wb.wb[index]);
306 if (tmp == 0xDEADBEEF) {
307 DRM_DEBUG("ib test on ring %d succeeded\n", ring->idx);
308 r = 0;
309 } else {
310 DRM_ERROR("amdgpu: ib test failed (0x%08X)\n", tmp);
311 r = -EINVAL;
312 }
313
314err1:
315 amdgpu_ib_free(adev, &ib, NULL);
316 dma_fence_put(f);
317err0:
318 amdgpu_device_wb_free(adev, index);
319 return r;
320}
321
322/**
323 * cik_dma_vm_copy_pte - update PTEs by copying them from the GART
324 *
325 * @ib: indirect buffer to fill with commands
326 * @pe: addr of the page entry
327 * @src: src addr to copy from
328 * @count: number of page entries to update
329 *
330 * Update PTEs by copying them from the GART using DMA (SI).
331 */
332static void si_dma_vm_copy_pte(struct amdgpu_ib *ib,
333 uint64_t pe, uint64_t src,
334 unsigned count)
335{
336 unsigned bytes = count * 8;
337
338 ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_COPY,
339 1, 0, 0, bytes);
340 ib->ptr[ib->length_dw++] = lower_32_bits(pe);
341 ib->ptr[ib->length_dw++] = lower_32_bits(src);
342 ib->ptr[ib->length_dw++] = upper_32_bits(pe) & 0xff;
343 ib->ptr[ib->length_dw++] = upper_32_bits(src) & 0xff;
344}
345
346/**
347 * si_dma_vm_write_pte - update PTEs by writing them manually
348 *
349 * @ib: indirect buffer to fill with commands
350 * @pe: addr of the page entry
351 * @value: dst addr to write into pe
352 * @count: number of page entries to update
353 * @incr: increase next addr by incr bytes
354 *
355 * Update PTEs by writing them manually using DMA (SI).
356 */
357static void si_dma_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
358 uint64_t value, unsigned count,
359 uint32_t incr)
360{
361 unsigned ndw = count * 2;
362
363 ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 0, ndw);
364 ib->ptr[ib->length_dw++] = lower_32_bits(pe);
365 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
366 for (; ndw > 0; ndw -= 2) {
367 ib->ptr[ib->length_dw++] = lower_32_bits(value);
368 ib->ptr[ib->length_dw++] = upper_32_bits(value);
369 value += incr;
370 }
371}
372
373/**
374 * si_dma_vm_set_pte_pde - update the page tables using sDMA
375 *
376 * @ib: indirect buffer to fill with commands
377 * @pe: addr of the page entry
378 * @addr: dst addr to write into pe
379 * @count: number of page entries to update
380 * @incr: increase next addr by incr bytes
381 * @flags: access flags
382 *
383 * Update the page tables using sDMA (CIK).
384 */
385static void si_dma_vm_set_pte_pde(struct amdgpu_ib *ib,
386 uint64_t pe,
387 uint64_t addr, unsigned count,
388 uint32_t incr, uint64_t flags)
389{
390 uint64_t value;
391 unsigned ndw;
392
393 while (count) {
394 ndw = count * 2;
395 if (ndw > 0xFFFFE)
396 ndw = 0xFFFFE;
397
398 if (flags & AMDGPU_PTE_VALID)
399 value = addr;
400 else
401 value = 0;
402
403 /* for physically contiguous pages (vram) */
404 ib->ptr[ib->length_dw++] = DMA_PTE_PDE_PACKET(ndw);
405 ib->ptr[ib->length_dw++] = pe; /* dst addr */
406 ib->ptr[ib->length_dw++] = upper_32_bits(pe) & 0xff;
407 ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
408 ib->ptr[ib->length_dw++] = upper_32_bits(flags);
409 ib->ptr[ib->length_dw++] = value; /* value */
410 ib->ptr[ib->length_dw++] = upper_32_bits(value);
411 ib->ptr[ib->length_dw++] = incr; /* increment size */
412 ib->ptr[ib->length_dw++] = 0;
413 pe += ndw * 4;
414 addr += (ndw / 2) * incr;
415 count -= ndw / 2;
416 }
417}
418
419/**
420 * si_dma_pad_ib - pad the IB to the required number of dw
421 *
422 * @ib: indirect buffer to fill with padding
423 *
424 */
425static void si_dma_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
426{
427 while (ib->length_dw & 0x7)
428 ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0);
429}
430
431/**
432 * cik_sdma_ring_emit_pipeline_sync - sync the pipeline
433 *
434 * @ring: amdgpu_ring pointer
435 *
436 * Make sure all previous operations are completed (CIK).
437 */
438static void si_dma_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
439{
440 uint32_t seq = ring->fence_drv.sync_seq;
441 uint64_t addr = ring->fence_drv.gpu_addr;
442
443 /* wait for idle */
444 amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_POLL_REG_MEM, 0, 0, 0, 0) |
445 (1 << 27)); /* Poll memory */
446 amdgpu_ring_write(ring, lower_32_bits(addr));
447 amdgpu_ring_write(ring, (0xff << 16) | upper_32_bits(addr)); /* retry, addr_hi */
448 amdgpu_ring_write(ring, 0xffffffff); /* mask */
449 amdgpu_ring_write(ring, seq); /* value */
450 amdgpu_ring_write(ring, (3 << 28) | 0x20); /* func(equal) | poll interval */
451}
452
453/**
454 * si_dma_ring_emit_vm_flush - cik vm flush using sDMA
455 *
456 * @ring: amdgpu_ring pointer
457 * @vm: amdgpu_vm pointer
458 *
459 * Update the page table base and flush the VM TLB
460 * using sDMA (VI).
461 */
462static void si_dma_ring_emit_vm_flush(struct amdgpu_ring *ring,
463 unsigned vmid, uint64_t pd_addr)
464{
465 amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
466
467 /* wait for invalidate to complete */
468 amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_POLL_REG_MEM, 0, 0, 0, 0));
469 amdgpu_ring_write(ring, VM_INVALIDATE_REQUEST);
470 amdgpu_ring_write(ring, 0xff << 16); /* retry */
471 amdgpu_ring_write(ring, 1 << vmid); /* mask */
472 amdgpu_ring_write(ring, 0); /* value */
473 amdgpu_ring_write(ring, (0 << 28) | 0x20); /* func(always) | poll interval */
474}
475
476static void si_dma_ring_emit_wreg(struct amdgpu_ring *ring,
477 uint32_t reg, uint32_t val)
478{
479 amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0, 0, 0));
480 amdgpu_ring_write(ring, (0xf << 16) | reg);
481 amdgpu_ring_write(ring, val);
482}
483
484static int si_dma_early_init(void *handle)
485{
486 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
487
488 adev->sdma.num_instances = 2;
489
490 si_dma_set_ring_funcs(adev);
491 si_dma_set_buffer_funcs(adev);
492 si_dma_set_vm_pte_funcs(adev);
493 si_dma_set_irq_funcs(adev);
494
495 return 0;
496}
497
498static int si_dma_sw_init(void *handle)
499{
500 struct amdgpu_ring *ring;
501 int r, i;
502 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
503
504 /* DMA0 trap event */
505 r = amdgpu_irq_add_id(adev, AMDGPU_IH_CLIENTID_LEGACY, 224, &adev->sdma.trap_irq);
506 if (r)
507 return r;
508
509 /* DMA1 trap event */
510 r = amdgpu_irq_add_id(adev, AMDGPU_IH_CLIENTID_LEGACY, 244, &adev->sdma.trap_irq_1);
511 if (r)
512 return r;
513
514 for (i = 0; i < adev->sdma.num_instances; i++) {
515 ring = &adev->sdma.instance[i].ring;
516 ring->ring_obj = NULL;
517 ring->use_doorbell = false;
518 sprintf(ring->name, "sdma%d", i);
519 r = amdgpu_ring_init(adev, ring, 1024,
520 &adev->sdma.trap_irq,
521 (i == 0) ?
522 AMDGPU_SDMA_IRQ_TRAP0 :
523 AMDGPU_SDMA_IRQ_TRAP1);
524 if (r)
525 return r;
526 }
527
528 return r;
529}
530
531static int si_dma_sw_fini(void *handle)
532{
533 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
534 int i;
535
536 for (i = 0; i < adev->sdma.num_instances; i++)
537 amdgpu_ring_fini(&adev->sdma.instance[i].ring);
538
539 return 0;
540}
541
542static int si_dma_hw_init(void *handle)
543{
544 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
545
546 return si_dma_start(adev);
547}
548
549static int si_dma_hw_fini(void *handle)
550{
551 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
552
553 si_dma_stop(adev);
554
555 return 0;
556}
557
558static int si_dma_suspend(void *handle)
559{
560 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
561
562 return si_dma_hw_fini(adev);
563}
564
565static int si_dma_resume(void *handle)
566{
567 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
568
569 return si_dma_hw_init(adev);
570}
571
572static bool si_dma_is_idle(void *handle)
573{
574 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
575 u32 tmp = RREG32(SRBM_STATUS2);
576
577 if (tmp & (DMA_BUSY_MASK | DMA1_BUSY_MASK))
578 return false;
579
580 return true;
581}
582
583static int si_dma_wait_for_idle(void *handle)
584{
585 unsigned i;
586 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
587
588 for (i = 0; i < adev->usec_timeout; i++) {
589 if (si_dma_is_idle(handle))
590 return 0;
591 udelay(1);
592 }
593 return -ETIMEDOUT;
594}
595
596static int si_dma_soft_reset(void *handle)
597{
598 DRM_INFO("si_dma_soft_reset --- not implemented !!!!!!!\n");
599 return 0;
600}
601
602static int si_dma_set_trap_irq_state(struct amdgpu_device *adev,
603 struct amdgpu_irq_src *src,
604 unsigned type,
605 enum amdgpu_interrupt_state state)
606{
607 u32 sdma_cntl;
608
609 switch (type) {
610 case AMDGPU_SDMA_IRQ_TRAP0:
611 switch (state) {
612 case AMDGPU_IRQ_STATE_DISABLE:
613 sdma_cntl = RREG32(DMA_CNTL + DMA0_REGISTER_OFFSET);
614 sdma_cntl &= ~TRAP_ENABLE;
615 WREG32(DMA_CNTL + DMA0_REGISTER_OFFSET, sdma_cntl);
616 break;
617 case AMDGPU_IRQ_STATE_ENABLE:
618 sdma_cntl = RREG32(DMA_CNTL + DMA0_REGISTER_OFFSET);
619 sdma_cntl |= TRAP_ENABLE;
620 WREG32(DMA_CNTL + DMA0_REGISTER_OFFSET, sdma_cntl);
621 break;
622 default:
623 break;
624 }
625 break;
626 case AMDGPU_SDMA_IRQ_TRAP1:
627 switch (state) {
628 case AMDGPU_IRQ_STATE_DISABLE:
629 sdma_cntl = RREG32(DMA_CNTL + DMA1_REGISTER_OFFSET);
630 sdma_cntl &= ~TRAP_ENABLE;
631 WREG32(DMA_CNTL + DMA1_REGISTER_OFFSET, sdma_cntl);
632 break;
633 case AMDGPU_IRQ_STATE_ENABLE:
634 sdma_cntl = RREG32(DMA_CNTL + DMA1_REGISTER_OFFSET);
635 sdma_cntl |= TRAP_ENABLE;
636 WREG32(DMA_CNTL + DMA1_REGISTER_OFFSET, sdma_cntl);
637 break;
638 default:
639 break;
640 }
641 break;
642 default:
643 break;
644 }
645 return 0;
646}
647
648static int si_dma_process_trap_irq(struct amdgpu_device *adev,
649 struct amdgpu_irq_src *source,
650 struct amdgpu_iv_entry *entry)
651{
652 amdgpu_fence_process(&adev->sdma.instance[0].ring);
653
654 return 0;
655}
656
657static int si_dma_process_trap_irq_1(struct amdgpu_device *adev,
658 struct amdgpu_irq_src *source,
659 struct amdgpu_iv_entry *entry)
660{
661 amdgpu_fence_process(&adev->sdma.instance[1].ring);
662
663 return 0;
664}
665
666static int si_dma_process_illegal_inst_irq(struct amdgpu_device *adev,
667 struct amdgpu_irq_src *source,
668 struct amdgpu_iv_entry *entry)
669{
670 DRM_ERROR("Illegal instruction in SDMA command stream\n");
671 schedule_work(&adev->reset_work);
672 return 0;
673}
674
675static int si_dma_set_clockgating_state(void *handle,
676 enum amd_clockgating_state state)
677{
678 u32 orig, data, offset;
679 int i;
680 bool enable;
681 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
682
683 enable = (state == AMD_CG_STATE_GATE) ? true : false;
684
685 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
686 for (i = 0; i < adev->sdma.num_instances; i++) {
687 if (i == 0)
688 offset = DMA0_REGISTER_OFFSET;
689 else
690 offset = DMA1_REGISTER_OFFSET;
691 orig = data = RREG32(DMA_POWER_CNTL + offset);
692 data &= ~MEM_POWER_OVERRIDE;
693 if (data != orig)
694 WREG32(DMA_POWER_CNTL + offset, data);
695 WREG32(DMA_CLK_CTRL + offset, 0x00000100);
696 }
697 } else {
698 for (i = 0; i < adev->sdma.num_instances; i++) {
699 if (i == 0)
700 offset = DMA0_REGISTER_OFFSET;
701 else
702 offset = DMA1_REGISTER_OFFSET;
703 orig = data = RREG32(DMA_POWER_CNTL + offset);
704 data |= MEM_POWER_OVERRIDE;
705 if (data != orig)
706 WREG32(DMA_POWER_CNTL + offset, data);
707
708 orig = data = RREG32(DMA_CLK_CTRL + offset);
709 data = 0xff000000;
710 if (data != orig)
711 WREG32(DMA_CLK_CTRL + offset, data);
712 }
713 }
714
715 return 0;
716}
717
718static int si_dma_set_powergating_state(void *handle,
719 enum amd_powergating_state state)
720{
721 u32 tmp;
722
723 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
724
725 WREG32(DMA_PGFSM_WRITE, 0x00002000);
726 WREG32(DMA_PGFSM_CONFIG, 0x100010ff);
727
728 for (tmp = 0; tmp < 5; tmp++)
729 WREG32(DMA_PGFSM_WRITE, 0);
730
731 return 0;
732}
733
734static const struct amd_ip_funcs si_dma_ip_funcs = {
735 .name = "si_dma",
736 .early_init = si_dma_early_init,
737 .late_init = NULL,
738 .sw_init = si_dma_sw_init,
739 .sw_fini = si_dma_sw_fini,
740 .hw_init = si_dma_hw_init,
741 .hw_fini = si_dma_hw_fini,
742 .suspend = si_dma_suspend,
743 .resume = si_dma_resume,
744 .is_idle = si_dma_is_idle,
745 .wait_for_idle = si_dma_wait_for_idle,
746 .soft_reset = si_dma_soft_reset,
747 .set_clockgating_state = si_dma_set_clockgating_state,
748 .set_powergating_state = si_dma_set_powergating_state,
749};
750
751static const struct amdgpu_ring_funcs si_dma_ring_funcs = {
752 .type = AMDGPU_RING_TYPE_SDMA,
753 .align_mask = 0xf,
754 .nop = DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0),
755 .support_64bit_ptrs = false,
756 .get_rptr = si_dma_ring_get_rptr,
757 .get_wptr = si_dma_ring_get_wptr,
758 .set_wptr = si_dma_ring_set_wptr,
759 .emit_frame_size =
760 3 + 3 + /* hdp flush / invalidate */
761 6 + /* si_dma_ring_emit_pipeline_sync */
762 SI_FLUSH_GPU_TLB_NUM_WREG * 3 + 6 + /* si_dma_ring_emit_vm_flush */
763 9 + 9 + 9, /* si_dma_ring_emit_fence x3 for user fence, vm fence */
764 .emit_ib_size = 7 + 3, /* si_dma_ring_emit_ib */
765 .emit_ib = si_dma_ring_emit_ib,
766 .emit_fence = si_dma_ring_emit_fence,
767 .emit_pipeline_sync = si_dma_ring_emit_pipeline_sync,
768 .emit_vm_flush = si_dma_ring_emit_vm_flush,
769 .test_ring = si_dma_ring_test_ring,
770 .test_ib = si_dma_ring_test_ib,
771 .insert_nop = amdgpu_ring_insert_nop,
772 .pad_ib = si_dma_ring_pad_ib,
773 .emit_wreg = si_dma_ring_emit_wreg,
774};
775
776static void si_dma_set_ring_funcs(struct amdgpu_device *adev)
777{
778 int i;
779
780 for (i = 0; i < adev->sdma.num_instances; i++)
781 adev->sdma.instance[i].ring.funcs = &si_dma_ring_funcs;
782}
783
784static const struct amdgpu_irq_src_funcs si_dma_trap_irq_funcs = {
785 .set = si_dma_set_trap_irq_state,
786 .process = si_dma_process_trap_irq,
787};
788
789static const struct amdgpu_irq_src_funcs si_dma_trap_irq_funcs_1 = {
790 .set = si_dma_set_trap_irq_state,
791 .process = si_dma_process_trap_irq_1,
792};
793
794static const struct amdgpu_irq_src_funcs si_dma_illegal_inst_irq_funcs = {
795 .process = si_dma_process_illegal_inst_irq,
796};
797
798static void si_dma_set_irq_funcs(struct amdgpu_device *adev)
799{
800 adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
801 adev->sdma.trap_irq.funcs = &si_dma_trap_irq_funcs;
802 adev->sdma.trap_irq_1.funcs = &si_dma_trap_irq_funcs_1;
803 adev->sdma.illegal_inst_irq.funcs = &si_dma_illegal_inst_irq_funcs;
804}
805
806/**
807 * si_dma_emit_copy_buffer - copy buffer using the sDMA engine
808 *
809 * @ring: amdgpu_ring structure holding ring information
810 * @src_offset: src GPU address
811 * @dst_offset: dst GPU address
812 * @byte_count: number of bytes to xfer
813 *
814 * Copy GPU buffers using the DMA engine (VI).
815 * Used by the amdgpu ttm implementation to move pages if
816 * registered as the asic copy callback.
817 */
818static void si_dma_emit_copy_buffer(struct amdgpu_ib *ib,
819 uint64_t src_offset,
820 uint64_t dst_offset,
821 uint32_t byte_count)
822{
823 ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_COPY,
824 1, 0, 0, byte_count);
825 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
826 ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
827 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset) & 0xff;
828 ib->ptr[ib->length_dw++] = upper_32_bits(src_offset) & 0xff;
829}
830
831/**
832 * si_dma_emit_fill_buffer - fill buffer using the sDMA engine
833 *
834 * @ring: amdgpu_ring structure holding ring information
835 * @src_data: value to write to buffer
836 * @dst_offset: dst GPU address
837 * @byte_count: number of bytes to xfer
838 *
839 * Fill GPU buffers using the DMA engine (VI).
840 */
841static void si_dma_emit_fill_buffer(struct amdgpu_ib *ib,
842 uint32_t src_data,
843 uint64_t dst_offset,
844 uint32_t byte_count)
845{
846 ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_CONSTANT_FILL,
847 0, 0, 0, byte_count / 4);
848 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
849 ib->ptr[ib->length_dw++] = src_data;
850 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset) << 16;
851}
852
853
854static const struct amdgpu_buffer_funcs si_dma_buffer_funcs = {
855 .copy_max_bytes = 0xffff8,
856 .copy_num_dw = 5,
857 .emit_copy_buffer = si_dma_emit_copy_buffer,
858
859 .fill_max_bytes = 0xffff8,
860 .fill_num_dw = 4,
861 .emit_fill_buffer = si_dma_emit_fill_buffer,
862};
863
864static void si_dma_set_buffer_funcs(struct amdgpu_device *adev)
865{
866 if (adev->mman.buffer_funcs == NULL) {
867 adev->mman.buffer_funcs = &si_dma_buffer_funcs;
868 adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
869 }
870}
871
872static const struct amdgpu_vm_pte_funcs si_dma_vm_pte_funcs = {
873 .copy_pte_num_dw = 5,
874 .copy_pte = si_dma_vm_copy_pte,
875
876 .write_pte = si_dma_vm_write_pte,
877 .set_pte_pde = si_dma_vm_set_pte_pde,
878};
879
880static void si_dma_set_vm_pte_funcs(struct amdgpu_device *adev)
881{
882 unsigned i;
883
884 if (adev->vm_manager.vm_pte_funcs == NULL) {
885 adev->vm_manager.vm_pte_funcs = &si_dma_vm_pte_funcs;
886 for (i = 0; i < adev->sdma.num_instances; i++)
887 adev->vm_manager.vm_pte_rings[i] =
888 &adev->sdma.instance[i].ring;
889
890 adev->vm_manager.vm_pte_num_rings = adev->sdma.num_instances;
891 }
892}
893
894const struct amdgpu_ip_block_version si_dma_ip_block =
895{
896 .type = AMD_IP_BLOCK_TYPE_SDMA,
897 .major = 1,
898 .minor = 0,
899 .rev = 0,
900 .funcs = &si_dma_ip_funcs,
901};