xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/si_dma.c (revision 3e30a927)
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 
30 const u32 sdma_offsets[SDMA_MAX_INSTANCE] =
31 {
32 	DMA0_REGISTER_OFFSET,
33 	DMA1_REGISTER_OFFSET
34 };
35 
36 static void si_dma_set_ring_funcs(struct amdgpu_device *adev);
37 static void si_dma_set_buffer_funcs(struct amdgpu_device *adev);
38 static void si_dma_set_vm_pte_funcs(struct amdgpu_device *adev);
39 static void si_dma_set_irq_funcs(struct amdgpu_device *adev);
40 
41 static uint64_t si_dma_ring_get_rptr(struct amdgpu_ring *ring)
42 {
43 	return ring->adev->wb.wb[ring->rptr_offs>>2];
44 }
45 
46 static 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 
54 static 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 
63 static void si_dma_ring_emit_ib(struct amdgpu_ring *ring,
64 				struct amdgpu_job *job,
65 				struct amdgpu_ib *ib,
66 				uint32_t flags)
67 {
68 	unsigned vmid = AMDGPU_JOB_GET_VMID(job);
69 	/* The indirect buffer packet must end on an 8 DW boundary in the DMA ring.
70 	 * Pad as necessary with NOPs.
71 	 */
72 	while ((lower_32_bits(ring->wptr) & 7) != 5)
73 		amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0));
74 	amdgpu_ring_write(ring, DMA_IB_PACKET(DMA_PACKET_INDIRECT_BUFFER, vmid, 0));
75 	amdgpu_ring_write(ring, (ib->gpu_addr & 0xFFFFFFE0));
76 	amdgpu_ring_write(ring, (ib->length_dw << 12) | (upper_32_bits(ib->gpu_addr) & 0xFF));
77 
78 }
79 
80 /**
81  * si_dma_ring_emit_fence - emit a fence on the DMA ring
82  *
83  * @ring: amdgpu ring pointer
84  * @fence: amdgpu fence object
85  *
86  * Add a DMA fence packet to the ring to write
87  * the fence seq number and DMA trap packet to generate
88  * an interrupt if needed (VI).
89  */
90 static void si_dma_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
91 				      unsigned flags)
92 {
93 
94 	bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
95 	/* write the fence */
96 	amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_FENCE, 0, 0, 0, 0));
97 	amdgpu_ring_write(ring, addr & 0xfffffffc);
98 	amdgpu_ring_write(ring, (upper_32_bits(addr) & 0xff));
99 	amdgpu_ring_write(ring, seq);
100 	/* optionally write high bits as well */
101 	if (write64bit) {
102 		addr += 4;
103 		amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_FENCE, 0, 0, 0, 0));
104 		amdgpu_ring_write(ring, addr & 0xfffffffc);
105 		amdgpu_ring_write(ring, (upper_32_bits(addr) & 0xff));
106 		amdgpu_ring_write(ring, upper_32_bits(seq));
107 	}
108 	/* generate an interrupt */
109 	amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_TRAP, 0, 0, 0, 0));
110 }
111 
112 static void si_dma_stop(struct amdgpu_device *adev)
113 {
114 	struct amdgpu_ring *ring;
115 	u32 rb_cntl;
116 	unsigned i;
117 
118 	for (i = 0; i < adev->sdma.num_instances; i++) {
119 		ring = &adev->sdma.instance[i].ring;
120 		/* dma0 */
121 		rb_cntl = RREG32(DMA_RB_CNTL + sdma_offsets[i]);
122 		rb_cntl &= ~DMA_RB_ENABLE;
123 		WREG32(DMA_RB_CNTL + sdma_offsets[i], rb_cntl);
124 
125 		if (adev->mman.buffer_funcs_ring == ring)
126 			amdgpu_ttm_set_buffer_funcs_status(adev, false);
127 		ring->sched.ready = false;
128 	}
129 }
130 
131 static int si_dma_start(struct amdgpu_device *adev)
132 {
133 	struct amdgpu_ring *ring;
134 	u32 rb_cntl, dma_cntl, ib_cntl, rb_bufsz;
135 	int i, r;
136 	uint64_t rptr_addr;
137 
138 	for (i = 0; i < adev->sdma.num_instances; i++) {
139 		ring = &adev->sdma.instance[i].ring;
140 
141 		WREG32(DMA_SEM_INCOMPLETE_TIMER_CNTL + sdma_offsets[i], 0);
142 		WREG32(DMA_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
143 
144 		/* Set ring buffer size in dwords */
145 		rb_bufsz = order_base_2(ring->ring_size / 4);
146 		rb_cntl = rb_bufsz << 1;
147 #ifdef __BIG_ENDIAN
148 		rb_cntl |= DMA_RB_SWAP_ENABLE | DMA_RPTR_WRITEBACK_SWAP_ENABLE;
149 #endif
150 		WREG32(DMA_RB_CNTL + sdma_offsets[i], rb_cntl);
151 
152 		/* Initialize the ring buffer's read and write pointers */
153 		WREG32(DMA_RB_RPTR + sdma_offsets[i], 0);
154 		WREG32(DMA_RB_WPTR + sdma_offsets[i], 0);
155 
156 		rptr_addr = adev->wb.gpu_addr + (ring->rptr_offs * 4);
157 
158 		WREG32(DMA_RB_RPTR_ADDR_LO + sdma_offsets[i], lower_32_bits(rptr_addr));
159 		WREG32(DMA_RB_RPTR_ADDR_HI + sdma_offsets[i], upper_32_bits(rptr_addr) & 0xFF);
160 
161 		rb_cntl |= DMA_RPTR_WRITEBACK_ENABLE;
162 
163 		WREG32(DMA_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
164 
165 		/* enable DMA IBs */
166 		ib_cntl = DMA_IB_ENABLE | CMD_VMID_FORCE;
167 #ifdef __BIG_ENDIAN
168 		ib_cntl |= DMA_IB_SWAP_ENABLE;
169 #endif
170 		WREG32(DMA_IB_CNTL + sdma_offsets[i], ib_cntl);
171 
172 		dma_cntl = RREG32(DMA_CNTL + sdma_offsets[i]);
173 		dma_cntl &= ~CTXEMPTY_INT_ENABLE;
174 		WREG32(DMA_CNTL + sdma_offsets[i], dma_cntl);
175 
176 		ring->wptr = 0;
177 		WREG32(DMA_RB_WPTR + sdma_offsets[i], lower_32_bits(ring->wptr) << 2);
178 		WREG32(DMA_RB_CNTL + sdma_offsets[i], rb_cntl | DMA_RB_ENABLE);
179 
180 		ring->sched.ready = true;
181 
182 		r = amdgpu_ring_test_helper(ring);
183 		if (r)
184 			return r;
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  */
202 static 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 		return r;
214 
215 	gpu_addr = adev->wb.gpu_addr + (index * 4);
216 	tmp = 0xCAFEDEAD;
217 	adev->wb.wb[index] = cpu_to_le32(tmp);
218 
219 	r = amdgpu_ring_alloc(ring, 4);
220 	if (r)
221 		goto error_free_wb;
222 
223 	amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 0, 1));
224 	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
225 	amdgpu_ring_write(ring, upper_32_bits(gpu_addr) & 0xff);
226 	amdgpu_ring_write(ring, 0xDEADBEEF);
227 	amdgpu_ring_commit(ring);
228 
229 	for (i = 0; i < adev->usec_timeout; i++) {
230 		tmp = le32_to_cpu(adev->wb.wb[index]);
231 		if (tmp == 0xDEADBEEF)
232 			break;
233 		udelay(1);
234 	}
235 
236 	if (i >= adev->usec_timeout)
237 		r = -ETIMEDOUT;
238 
239 error_free_wb:
240 	amdgpu_device_wb_free(adev, index);
241 	return r;
242 }
243 
244 /**
245  * si_dma_ring_test_ib - test an IB on the DMA engine
246  *
247  * @ring: amdgpu_ring structure holding ring information
248  *
249  * Test a simple IB in the DMA ring (VI).
250  * Returns 0 on success, error on failure.
251  */
252 static int si_dma_ring_test_ib(struct amdgpu_ring *ring, long timeout)
253 {
254 	struct amdgpu_device *adev = ring->adev;
255 	struct amdgpu_ib ib;
256 	struct dma_fence *f = NULL;
257 	unsigned index;
258 	u32 tmp = 0;
259 	u64 gpu_addr;
260 	long r;
261 
262 	r = amdgpu_device_wb_get(adev, &index);
263 	if (r)
264 		return r;
265 
266 	gpu_addr = adev->wb.gpu_addr + (index * 4);
267 	tmp = 0xCAFEDEAD;
268 	adev->wb.wb[index] = cpu_to_le32(tmp);
269 	memset(&ib, 0, sizeof(ib));
270 	r = amdgpu_ib_get(adev, NULL, 256, &ib);
271 	if (r)
272 		goto err0;
273 
274 	ib.ptr[0] = DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 0, 1);
275 	ib.ptr[1] = lower_32_bits(gpu_addr);
276 	ib.ptr[2] = upper_32_bits(gpu_addr) & 0xff;
277 	ib.ptr[3] = 0xDEADBEEF;
278 	ib.length_dw = 4;
279 	r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
280 	if (r)
281 		goto err1;
282 
283 	r = dma_fence_wait_timeout(f, false, timeout);
284 	if (r == 0) {
285 		r = -ETIMEDOUT;
286 		goto err1;
287 	} else if (r < 0) {
288 		goto err1;
289 	}
290 	tmp = le32_to_cpu(adev->wb.wb[index]);
291 	if (tmp == 0xDEADBEEF)
292 		r = 0;
293 	else
294 		r = -EINVAL;
295 
296 err1:
297 	amdgpu_ib_free(adev, &ib, NULL);
298 	dma_fence_put(f);
299 err0:
300 	amdgpu_device_wb_free(adev, index);
301 	return r;
302 }
303 
304 /**
305  * cik_dma_vm_copy_pte - update PTEs by copying them from the GART
306  *
307  * @ib: indirect buffer to fill with commands
308  * @pe: addr of the page entry
309  * @src: src addr to copy from
310  * @count: number of page entries to update
311  *
312  * Update PTEs by copying them from the GART using DMA (SI).
313  */
314 static void si_dma_vm_copy_pte(struct amdgpu_ib *ib,
315 			       uint64_t pe, uint64_t src,
316 			       unsigned count)
317 {
318 	unsigned bytes = count * 8;
319 
320 	ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_COPY,
321 					      1, 0, 0, bytes);
322 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
323 	ib->ptr[ib->length_dw++] = lower_32_bits(src);
324 	ib->ptr[ib->length_dw++] = upper_32_bits(pe) & 0xff;
325 	ib->ptr[ib->length_dw++] = upper_32_bits(src) & 0xff;
326 }
327 
328 /**
329  * si_dma_vm_write_pte - update PTEs by writing them manually
330  *
331  * @ib: indirect buffer to fill with commands
332  * @pe: addr of the page entry
333  * @value: dst addr to write into pe
334  * @count: number of page entries to update
335  * @incr: increase next addr by incr bytes
336  *
337  * Update PTEs by writing them manually using DMA (SI).
338  */
339 static void si_dma_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
340 				uint64_t value, unsigned count,
341 				uint32_t incr)
342 {
343 	unsigned ndw = count * 2;
344 
345 	ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 0, ndw);
346 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
347 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
348 	for (; ndw > 0; ndw -= 2) {
349 		ib->ptr[ib->length_dw++] = lower_32_bits(value);
350 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
351 		value += incr;
352 	}
353 }
354 
355 /**
356  * si_dma_vm_set_pte_pde - update the page tables using sDMA
357  *
358  * @ib: indirect buffer to fill with commands
359  * @pe: addr of the page entry
360  * @addr: dst addr to write into pe
361  * @count: number of page entries to update
362  * @incr: increase next addr by incr bytes
363  * @flags: access flags
364  *
365  * Update the page tables using sDMA (CIK).
366  */
367 static void si_dma_vm_set_pte_pde(struct amdgpu_ib *ib,
368 				     uint64_t pe,
369 				     uint64_t addr, unsigned count,
370 				     uint32_t incr, uint64_t flags)
371 {
372 	uint64_t value;
373 	unsigned ndw;
374 
375 	while (count) {
376 		ndw = count * 2;
377 		if (ndw > 0xFFFFE)
378 			ndw = 0xFFFFE;
379 
380 		if (flags & AMDGPU_PTE_VALID)
381 			value = addr;
382 		else
383 			value = 0;
384 
385 		/* for physically contiguous pages (vram) */
386 		ib->ptr[ib->length_dw++] = DMA_PTE_PDE_PACKET(ndw);
387 		ib->ptr[ib->length_dw++] = pe; /* dst addr */
388 		ib->ptr[ib->length_dw++] = upper_32_bits(pe) & 0xff;
389 		ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
390 		ib->ptr[ib->length_dw++] = upper_32_bits(flags);
391 		ib->ptr[ib->length_dw++] = value; /* value */
392 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
393 		ib->ptr[ib->length_dw++] = incr; /* increment size */
394 		ib->ptr[ib->length_dw++] = 0;
395 		pe += ndw * 4;
396 		addr += (ndw / 2) * incr;
397 		count -= ndw / 2;
398 	}
399 }
400 
401 /**
402  * si_dma_pad_ib - pad the IB to the required number of dw
403  *
404  * @ib: indirect buffer to fill with padding
405  *
406  */
407 static void si_dma_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
408 {
409 	while (ib->length_dw & 0x7)
410 		ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0);
411 }
412 
413 /**
414  * cik_sdma_ring_emit_pipeline_sync - sync the pipeline
415  *
416  * @ring: amdgpu_ring pointer
417  *
418  * Make sure all previous operations are completed (CIK).
419  */
420 static void si_dma_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
421 {
422 	uint32_t seq = ring->fence_drv.sync_seq;
423 	uint64_t addr = ring->fence_drv.gpu_addr;
424 
425 	/* wait for idle */
426 	amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_POLL_REG_MEM, 0, 0, 0, 0) |
427 			  (1 << 27)); /* Poll memory */
428 	amdgpu_ring_write(ring, lower_32_bits(addr));
429 	amdgpu_ring_write(ring, (0xff << 16) | upper_32_bits(addr)); /* retry, addr_hi */
430 	amdgpu_ring_write(ring, 0xffffffff); /* mask */
431 	amdgpu_ring_write(ring, seq); /* value */
432 	amdgpu_ring_write(ring, (3 << 28) | 0x20); /* func(equal) | poll interval */
433 }
434 
435 /**
436  * si_dma_ring_emit_vm_flush - cik vm flush using sDMA
437  *
438  * @ring: amdgpu_ring pointer
439  * @vm: amdgpu_vm pointer
440  *
441  * Update the page table base and flush the VM TLB
442  * using sDMA (VI).
443  */
444 static void si_dma_ring_emit_vm_flush(struct amdgpu_ring *ring,
445 				      unsigned vmid, uint64_t pd_addr)
446 {
447 	amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
448 
449 	/* wait for invalidate to complete */
450 	amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_POLL_REG_MEM, 0, 0, 0, 0));
451 	amdgpu_ring_write(ring, VM_INVALIDATE_REQUEST);
452 	amdgpu_ring_write(ring, 0xff << 16); /* retry */
453 	amdgpu_ring_write(ring, 1 << vmid); /* mask */
454 	amdgpu_ring_write(ring, 0); /* value */
455 	amdgpu_ring_write(ring, (0 << 28) | 0x20); /* func(always) | poll interval */
456 }
457 
458 static void si_dma_ring_emit_wreg(struct amdgpu_ring *ring,
459 				  uint32_t reg, uint32_t val)
460 {
461 	amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0, 0, 0));
462 	amdgpu_ring_write(ring, (0xf << 16) | reg);
463 	amdgpu_ring_write(ring, val);
464 }
465 
466 static int si_dma_early_init(void *handle)
467 {
468 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
469 
470 	adev->sdma.num_instances = 2;
471 
472 	si_dma_set_ring_funcs(adev);
473 	si_dma_set_buffer_funcs(adev);
474 	si_dma_set_vm_pte_funcs(adev);
475 	si_dma_set_irq_funcs(adev);
476 
477 	return 0;
478 }
479 
480 static int si_dma_sw_init(void *handle)
481 {
482 	struct amdgpu_ring *ring;
483 	int r, i;
484 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
485 
486 	/* DMA0 trap event */
487 	r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 224,
488 			      &adev->sdma.trap_irq);
489 	if (r)
490 		return r;
491 
492 	/* DMA1 trap event */
493 	r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 244,
494 			      &adev->sdma.trap_irq);
495 	if (r)
496 		return r;
497 
498 	for (i = 0; i < adev->sdma.num_instances; i++) {
499 		ring = &adev->sdma.instance[i].ring;
500 		ring->ring_obj = NULL;
501 		ring->use_doorbell = false;
502 		sprintf(ring->name, "sdma%d", i);
503 		r = amdgpu_ring_init(adev, ring, 1024,
504 				     &adev->sdma.trap_irq,
505 				     (i == 0) ?
506 				     AMDGPU_SDMA_IRQ_INSTANCE0 :
507 				     AMDGPU_SDMA_IRQ_INSTANCE1);
508 		if (r)
509 			return r;
510 	}
511 
512 	return r;
513 }
514 
515 static int si_dma_sw_fini(void *handle)
516 {
517 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
518 	int i;
519 
520 	for (i = 0; i < adev->sdma.num_instances; i++)
521 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
522 
523 	return 0;
524 }
525 
526 static int si_dma_hw_init(void *handle)
527 {
528 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
529 
530 	return si_dma_start(adev);
531 }
532 
533 static int si_dma_hw_fini(void *handle)
534 {
535 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
536 
537 	si_dma_stop(adev);
538 
539 	return 0;
540 }
541 
542 static int si_dma_suspend(void *handle)
543 {
544 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
545 
546 	return si_dma_hw_fini(adev);
547 }
548 
549 static int si_dma_resume(void *handle)
550 {
551 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
552 
553 	return si_dma_hw_init(adev);
554 }
555 
556 static bool si_dma_is_idle(void *handle)
557 {
558 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
559 	u32 tmp = RREG32(SRBM_STATUS2);
560 
561 	if (tmp & (DMA_BUSY_MASK | DMA1_BUSY_MASK))
562 	    return false;
563 
564 	return true;
565 }
566 
567 static int si_dma_wait_for_idle(void *handle)
568 {
569 	unsigned i;
570 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
571 
572 	for (i = 0; i < adev->usec_timeout; i++) {
573 		if (si_dma_is_idle(handle))
574 			return 0;
575 		udelay(1);
576 	}
577 	return -ETIMEDOUT;
578 }
579 
580 static int si_dma_soft_reset(void *handle)
581 {
582 	DRM_INFO("si_dma_soft_reset --- not implemented !!!!!!!\n");
583 	return 0;
584 }
585 
586 static int si_dma_set_trap_irq_state(struct amdgpu_device *adev,
587 					struct amdgpu_irq_src *src,
588 					unsigned type,
589 					enum amdgpu_interrupt_state state)
590 {
591 	u32 sdma_cntl;
592 
593 	switch (type) {
594 	case AMDGPU_SDMA_IRQ_INSTANCE0:
595 		switch (state) {
596 		case AMDGPU_IRQ_STATE_DISABLE:
597 			sdma_cntl = RREG32(DMA_CNTL + DMA0_REGISTER_OFFSET);
598 			sdma_cntl &= ~TRAP_ENABLE;
599 			WREG32(DMA_CNTL + DMA0_REGISTER_OFFSET, sdma_cntl);
600 			break;
601 		case AMDGPU_IRQ_STATE_ENABLE:
602 			sdma_cntl = RREG32(DMA_CNTL + DMA0_REGISTER_OFFSET);
603 			sdma_cntl |= TRAP_ENABLE;
604 			WREG32(DMA_CNTL + DMA0_REGISTER_OFFSET, sdma_cntl);
605 			break;
606 		default:
607 			break;
608 		}
609 		break;
610 	case AMDGPU_SDMA_IRQ_INSTANCE1:
611 		switch (state) {
612 		case AMDGPU_IRQ_STATE_DISABLE:
613 			sdma_cntl = RREG32(DMA_CNTL + DMA1_REGISTER_OFFSET);
614 			sdma_cntl &= ~TRAP_ENABLE;
615 			WREG32(DMA_CNTL + DMA1_REGISTER_OFFSET, sdma_cntl);
616 			break;
617 		case AMDGPU_IRQ_STATE_ENABLE:
618 			sdma_cntl = RREG32(DMA_CNTL + DMA1_REGISTER_OFFSET);
619 			sdma_cntl |= TRAP_ENABLE;
620 			WREG32(DMA_CNTL + DMA1_REGISTER_OFFSET, sdma_cntl);
621 			break;
622 		default:
623 			break;
624 		}
625 		break;
626 	default:
627 		break;
628 	}
629 	return 0;
630 }
631 
632 static int si_dma_process_trap_irq(struct amdgpu_device *adev,
633 				      struct amdgpu_irq_src *source,
634 				      struct amdgpu_iv_entry *entry)
635 {
636 	if (entry->src_id == 224)
637 		amdgpu_fence_process(&adev->sdma.instance[0].ring);
638 	else
639 		amdgpu_fence_process(&adev->sdma.instance[1].ring);
640 	return 0;
641 }
642 
643 static int si_dma_set_clockgating_state(void *handle,
644 					  enum amd_clockgating_state state)
645 {
646 	u32 orig, data, offset;
647 	int i;
648 	bool enable;
649 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
650 
651 	enable = (state == AMD_CG_STATE_GATE);
652 
653 	if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
654 		for (i = 0; i < adev->sdma.num_instances; i++) {
655 			if (i == 0)
656 				offset = DMA0_REGISTER_OFFSET;
657 			else
658 				offset = DMA1_REGISTER_OFFSET;
659 			orig = data = RREG32(DMA_POWER_CNTL + offset);
660 			data &= ~MEM_POWER_OVERRIDE;
661 			if (data != orig)
662 				WREG32(DMA_POWER_CNTL + offset, data);
663 			WREG32(DMA_CLK_CTRL + offset, 0x00000100);
664 		}
665 	} else {
666 		for (i = 0; i < adev->sdma.num_instances; i++) {
667 			if (i == 0)
668 				offset = DMA0_REGISTER_OFFSET;
669 			else
670 				offset = DMA1_REGISTER_OFFSET;
671 			orig = data = RREG32(DMA_POWER_CNTL + offset);
672 			data |= MEM_POWER_OVERRIDE;
673 			if (data != orig)
674 				WREG32(DMA_POWER_CNTL + offset, data);
675 
676 			orig = data = RREG32(DMA_CLK_CTRL + offset);
677 			data = 0xff000000;
678 			if (data != orig)
679 				WREG32(DMA_CLK_CTRL + offset, data);
680 		}
681 	}
682 
683 	return 0;
684 }
685 
686 static int si_dma_set_powergating_state(void *handle,
687 					  enum amd_powergating_state state)
688 {
689 	u32 tmp;
690 
691 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
692 
693 	WREG32(DMA_PGFSM_WRITE,  0x00002000);
694 	WREG32(DMA_PGFSM_CONFIG, 0x100010ff);
695 
696 	for (tmp = 0; tmp < 5; tmp++)
697 		WREG32(DMA_PGFSM_WRITE, 0);
698 
699 	return 0;
700 }
701 
702 static const struct amd_ip_funcs si_dma_ip_funcs = {
703 	.name = "si_dma",
704 	.early_init = si_dma_early_init,
705 	.late_init = NULL,
706 	.sw_init = si_dma_sw_init,
707 	.sw_fini = si_dma_sw_fini,
708 	.hw_init = si_dma_hw_init,
709 	.hw_fini = si_dma_hw_fini,
710 	.suspend = si_dma_suspend,
711 	.resume = si_dma_resume,
712 	.is_idle = si_dma_is_idle,
713 	.wait_for_idle = si_dma_wait_for_idle,
714 	.soft_reset = si_dma_soft_reset,
715 	.set_clockgating_state = si_dma_set_clockgating_state,
716 	.set_powergating_state = si_dma_set_powergating_state,
717 };
718 
719 static const struct amdgpu_ring_funcs si_dma_ring_funcs = {
720 	.type = AMDGPU_RING_TYPE_SDMA,
721 	.align_mask = 0xf,
722 	.nop = DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0),
723 	.support_64bit_ptrs = false,
724 	.get_rptr = si_dma_ring_get_rptr,
725 	.get_wptr = si_dma_ring_get_wptr,
726 	.set_wptr = si_dma_ring_set_wptr,
727 	.emit_frame_size =
728 		3 + 3 + /* hdp flush / invalidate */
729 		6 + /* si_dma_ring_emit_pipeline_sync */
730 		SI_FLUSH_GPU_TLB_NUM_WREG * 3 + 6 + /* si_dma_ring_emit_vm_flush */
731 		9 + 9 + 9, /* si_dma_ring_emit_fence x3 for user fence, vm fence */
732 	.emit_ib_size = 7 + 3, /* si_dma_ring_emit_ib */
733 	.emit_ib = si_dma_ring_emit_ib,
734 	.emit_fence = si_dma_ring_emit_fence,
735 	.emit_pipeline_sync = si_dma_ring_emit_pipeline_sync,
736 	.emit_vm_flush = si_dma_ring_emit_vm_flush,
737 	.test_ring = si_dma_ring_test_ring,
738 	.test_ib = si_dma_ring_test_ib,
739 	.insert_nop = amdgpu_ring_insert_nop,
740 	.pad_ib = si_dma_ring_pad_ib,
741 	.emit_wreg = si_dma_ring_emit_wreg,
742 };
743 
744 static void si_dma_set_ring_funcs(struct amdgpu_device *adev)
745 {
746 	int i;
747 
748 	for (i = 0; i < adev->sdma.num_instances; i++)
749 		adev->sdma.instance[i].ring.funcs = &si_dma_ring_funcs;
750 }
751 
752 static const struct amdgpu_irq_src_funcs si_dma_trap_irq_funcs = {
753 	.set = si_dma_set_trap_irq_state,
754 	.process = si_dma_process_trap_irq,
755 };
756 
757 static void si_dma_set_irq_funcs(struct amdgpu_device *adev)
758 {
759 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
760 	adev->sdma.trap_irq.funcs = &si_dma_trap_irq_funcs;
761 }
762 
763 /**
764  * si_dma_emit_copy_buffer - copy buffer using the sDMA engine
765  *
766  * @ring: amdgpu_ring structure holding ring information
767  * @src_offset: src GPU address
768  * @dst_offset: dst GPU address
769  * @byte_count: number of bytes to xfer
770  *
771  * Copy GPU buffers using the DMA engine (VI).
772  * Used by the amdgpu ttm implementation to move pages if
773  * registered as the asic copy callback.
774  */
775 static void si_dma_emit_copy_buffer(struct amdgpu_ib *ib,
776 				       uint64_t src_offset,
777 				       uint64_t dst_offset,
778 				       uint32_t byte_count)
779 {
780 	ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_COPY,
781 					      1, 0, 0, byte_count);
782 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
783 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
784 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset) & 0xff;
785 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset) & 0xff;
786 }
787 
788 /**
789  * si_dma_emit_fill_buffer - fill buffer using the sDMA engine
790  *
791  * @ring: amdgpu_ring structure holding ring information
792  * @src_data: value to write to buffer
793  * @dst_offset: dst GPU address
794  * @byte_count: number of bytes to xfer
795  *
796  * Fill GPU buffers using the DMA engine (VI).
797  */
798 static void si_dma_emit_fill_buffer(struct amdgpu_ib *ib,
799 				       uint32_t src_data,
800 				       uint64_t dst_offset,
801 				       uint32_t byte_count)
802 {
803 	ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_CONSTANT_FILL,
804 					      0, 0, 0, byte_count / 4);
805 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
806 	ib->ptr[ib->length_dw++] = src_data;
807 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset) << 16;
808 }
809 
810 
811 static const struct amdgpu_buffer_funcs si_dma_buffer_funcs = {
812 	.copy_max_bytes = 0xffff8,
813 	.copy_num_dw = 5,
814 	.emit_copy_buffer = si_dma_emit_copy_buffer,
815 
816 	.fill_max_bytes = 0xffff8,
817 	.fill_num_dw = 4,
818 	.emit_fill_buffer = si_dma_emit_fill_buffer,
819 };
820 
821 static void si_dma_set_buffer_funcs(struct amdgpu_device *adev)
822 {
823 	adev->mman.buffer_funcs = &si_dma_buffer_funcs;
824 	adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
825 }
826 
827 static const struct amdgpu_vm_pte_funcs si_dma_vm_pte_funcs = {
828 	.copy_pte_num_dw = 5,
829 	.copy_pte = si_dma_vm_copy_pte,
830 
831 	.write_pte = si_dma_vm_write_pte,
832 	.set_pte_pde = si_dma_vm_set_pte_pde,
833 };
834 
835 static void si_dma_set_vm_pte_funcs(struct amdgpu_device *adev)
836 {
837 	unsigned i;
838 
839 	adev->vm_manager.vm_pte_funcs = &si_dma_vm_pte_funcs;
840 	for (i = 0; i < adev->sdma.num_instances; i++) {
841 		adev->vm_manager.vm_pte_scheds[i] =
842 			&adev->sdma.instance[i].ring.sched;
843 	}
844 	adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
845 }
846 
847 const struct amdgpu_ip_block_version si_dma_ip_block =
848 {
849 	.type = AMD_IP_BLOCK_TYPE_SDMA,
850 	.major = 1,
851 	.minor = 0,
852 	.rev = 0,
853 	.funcs = &si_dma_ip_funcs,
854 };
855