xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/si_dma.c (revision 2a598d0b)
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->rptr_cpu_addr;
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], (ring->wptr << 2) & 0x3fffc);
60 }
61 
62 static 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  */
91 static 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 
113 static void si_dma_stop(struct amdgpu_device *adev)
114 {
115 	u32 rb_cntl;
116 	unsigned i;
117 
118 	amdgpu_sdma_unset_buffer_funcs_helper(adev);
119 
120 	for (i = 0; i < adev->sdma.num_instances; i++) {
121 		/* dma0 */
122 		rb_cntl = RREG32(DMA_RB_CNTL + sdma_offsets[i]);
123 		rb_cntl &= ~DMA_RB_ENABLE;
124 		WREG32(DMA_RB_CNTL + sdma_offsets[i], rb_cntl);
125 	}
126 }
127 
128 static int si_dma_start(struct amdgpu_device *adev)
129 {
130 	struct amdgpu_ring *ring;
131 	u32 rb_cntl, dma_cntl, ib_cntl, rb_bufsz;
132 	int i, r;
133 	uint64_t rptr_addr;
134 
135 	for (i = 0; i < adev->sdma.num_instances; i++) {
136 		ring = &adev->sdma.instance[i].ring;
137 
138 		WREG32(DMA_SEM_INCOMPLETE_TIMER_CNTL + sdma_offsets[i], 0);
139 		WREG32(DMA_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
140 
141 		/* Set ring buffer size in dwords */
142 		rb_bufsz = order_base_2(ring->ring_size / 4);
143 		rb_cntl = rb_bufsz << 1;
144 #ifdef __BIG_ENDIAN
145 		rb_cntl |= DMA_RB_SWAP_ENABLE | DMA_RPTR_WRITEBACK_SWAP_ENABLE;
146 #endif
147 		WREG32(DMA_RB_CNTL + sdma_offsets[i], rb_cntl);
148 
149 		/* Initialize the ring buffer's read and write pointers */
150 		WREG32(DMA_RB_RPTR + sdma_offsets[i], 0);
151 		WREG32(DMA_RB_WPTR + sdma_offsets[i], 0);
152 
153 		rptr_addr = ring->rptr_gpu_addr;
154 
155 		WREG32(DMA_RB_RPTR_ADDR_LO + sdma_offsets[i], lower_32_bits(rptr_addr));
156 		WREG32(DMA_RB_RPTR_ADDR_HI + sdma_offsets[i], upper_32_bits(rptr_addr) & 0xFF);
157 
158 		rb_cntl |= DMA_RPTR_WRITEBACK_ENABLE;
159 
160 		WREG32(DMA_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
161 
162 		/* enable DMA IBs */
163 		ib_cntl = DMA_IB_ENABLE | CMD_VMID_FORCE;
164 #ifdef __BIG_ENDIAN
165 		ib_cntl |= DMA_IB_SWAP_ENABLE;
166 #endif
167 		WREG32(DMA_IB_CNTL + sdma_offsets[i], ib_cntl);
168 
169 		dma_cntl = RREG32(DMA_CNTL + sdma_offsets[i]);
170 		dma_cntl &= ~CTXEMPTY_INT_ENABLE;
171 		WREG32(DMA_CNTL + sdma_offsets[i], dma_cntl);
172 
173 		ring->wptr = 0;
174 		WREG32(DMA_RB_WPTR + sdma_offsets[i], ring->wptr << 2);
175 		WREG32(DMA_RB_CNTL + sdma_offsets[i], rb_cntl | DMA_RB_ENABLE);
176 
177 		r = amdgpu_ring_test_helper(ring);
178 		if (r)
179 			return r;
180 
181 		if (adev->mman.buffer_funcs_ring == ring)
182 			amdgpu_ttm_set_buffer_funcs_status(adev, true);
183 	}
184 
185 	return 0;
186 }
187 
188 /**
189  * si_dma_ring_test_ring - simple async dma engine test
190  *
191  * @ring: amdgpu_ring structure holding ring information
192  *
193  * Test the DMA engine by writing using it to write an
194  * value to memory. (VI).
195  * Returns 0 for success, error for failure.
196  */
197 static int si_dma_ring_test_ring(struct amdgpu_ring *ring)
198 {
199 	struct amdgpu_device *adev = ring->adev;
200 	unsigned i;
201 	unsigned index;
202 	int r;
203 	u32 tmp;
204 	u64 gpu_addr;
205 
206 	r = amdgpu_device_wb_get(adev, &index);
207 	if (r)
208 		return r;
209 
210 	gpu_addr = adev->wb.gpu_addr + (index * 4);
211 	tmp = 0xCAFEDEAD;
212 	adev->wb.wb[index] = cpu_to_le32(tmp);
213 
214 	r = amdgpu_ring_alloc(ring, 4);
215 	if (r)
216 		goto error_free_wb;
217 
218 	amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 0, 1));
219 	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
220 	amdgpu_ring_write(ring, upper_32_bits(gpu_addr) & 0xff);
221 	amdgpu_ring_write(ring, 0xDEADBEEF);
222 	amdgpu_ring_commit(ring);
223 
224 	for (i = 0; i < adev->usec_timeout; i++) {
225 		tmp = le32_to_cpu(adev->wb.wb[index]);
226 		if (tmp == 0xDEADBEEF)
227 			break;
228 		udelay(1);
229 	}
230 
231 	if (i >= adev->usec_timeout)
232 		r = -ETIMEDOUT;
233 
234 error_free_wb:
235 	amdgpu_device_wb_free(adev, index);
236 	return r;
237 }
238 
239 /**
240  * si_dma_ring_test_ib - test an IB on the DMA engine
241  *
242  * @ring: amdgpu_ring structure holding ring information
243  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
244  *
245  * Test a simple IB in the DMA ring (VI).
246  * Returns 0 on success, error on failure.
247  */
248 static int si_dma_ring_test_ib(struct amdgpu_ring *ring, long timeout)
249 {
250 	struct amdgpu_device *adev = ring->adev;
251 	struct amdgpu_ib ib;
252 	struct dma_fence *f = NULL;
253 	unsigned index;
254 	u32 tmp = 0;
255 	u64 gpu_addr;
256 	long r;
257 
258 	r = amdgpu_device_wb_get(adev, &index);
259 	if (r)
260 		return r;
261 
262 	gpu_addr = adev->wb.gpu_addr + (index * 4);
263 	tmp = 0xCAFEDEAD;
264 	adev->wb.wb[index] = cpu_to_le32(tmp);
265 	memset(&ib, 0, sizeof(ib));
266 	r = amdgpu_ib_get(adev, NULL, 256,
267 					AMDGPU_IB_POOL_DIRECT, &ib);
268 	if (r)
269 		goto err0;
270 
271 	ib.ptr[0] = DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 0, 1);
272 	ib.ptr[1] = lower_32_bits(gpu_addr);
273 	ib.ptr[2] = upper_32_bits(gpu_addr) & 0xff;
274 	ib.ptr[3] = 0xDEADBEEF;
275 	ib.length_dw = 4;
276 	r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
277 	if (r)
278 		goto err1;
279 
280 	r = dma_fence_wait_timeout(f, false, timeout);
281 	if (r == 0) {
282 		r = -ETIMEDOUT;
283 		goto err1;
284 	} else if (r < 0) {
285 		goto err1;
286 	}
287 	tmp = le32_to_cpu(adev->wb.wb[index]);
288 	if (tmp == 0xDEADBEEF)
289 		r = 0;
290 	else
291 		r = -EINVAL;
292 
293 err1:
294 	amdgpu_ib_free(adev, &ib, NULL);
295 	dma_fence_put(f);
296 err0:
297 	amdgpu_device_wb_free(adev, index);
298 	return r;
299 }
300 
301 /**
302  * si_dma_vm_copy_pte - update PTEs by copying them from the GART
303  *
304  * @ib: indirect buffer to fill with commands
305  * @pe: addr of the page entry
306  * @src: src addr to copy from
307  * @count: number of page entries to update
308  *
309  * Update PTEs by copying them from the GART using DMA (SI).
310  */
311 static void si_dma_vm_copy_pte(struct amdgpu_ib *ib,
312 			       uint64_t pe, uint64_t src,
313 			       unsigned count)
314 {
315 	unsigned bytes = count * 8;
316 
317 	ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_COPY,
318 					      1, 0, 0, bytes);
319 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
320 	ib->ptr[ib->length_dw++] = lower_32_bits(src);
321 	ib->ptr[ib->length_dw++] = upper_32_bits(pe) & 0xff;
322 	ib->ptr[ib->length_dw++] = upper_32_bits(src) & 0xff;
323 }
324 
325 /**
326  * si_dma_vm_write_pte - update PTEs by writing them manually
327  *
328  * @ib: indirect buffer to fill with commands
329  * @pe: addr of the page entry
330  * @value: dst addr to write into pe
331  * @count: number of page entries to update
332  * @incr: increase next addr by incr bytes
333  *
334  * Update PTEs by writing them manually using DMA (SI).
335  */
336 static void si_dma_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
337 				uint64_t value, unsigned count,
338 				uint32_t incr)
339 {
340 	unsigned ndw = count * 2;
341 
342 	ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 0, ndw);
343 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
344 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
345 	for (; ndw > 0; ndw -= 2) {
346 		ib->ptr[ib->length_dw++] = lower_32_bits(value);
347 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
348 		value += incr;
349 	}
350 }
351 
352 /**
353  * si_dma_vm_set_pte_pde - update the page tables using sDMA
354  *
355  * @ib: indirect buffer to fill with commands
356  * @pe: addr of the page entry
357  * @addr: dst addr to write into pe
358  * @count: number of page entries to update
359  * @incr: increase next addr by incr bytes
360  * @flags: access flags
361  *
362  * Update the page tables using sDMA (CIK).
363  */
364 static void si_dma_vm_set_pte_pde(struct amdgpu_ib *ib,
365 				     uint64_t pe,
366 				     uint64_t addr, unsigned count,
367 				     uint32_t incr, uint64_t flags)
368 {
369 	uint64_t value;
370 	unsigned ndw;
371 
372 	while (count) {
373 		ndw = count * 2;
374 		if (ndw > 0xFFFFE)
375 			ndw = 0xFFFFE;
376 
377 		if (flags & AMDGPU_PTE_VALID)
378 			value = addr;
379 		else
380 			value = 0;
381 
382 		/* for physically contiguous pages (vram) */
383 		ib->ptr[ib->length_dw++] = DMA_PTE_PDE_PACKET(ndw);
384 		ib->ptr[ib->length_dw++] = pe; /* dst addr */
385 		ib->ptr[ib->length_dw++] = upper_32_bits(pe) & 0xff;
386 		ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
387 		ib->ptr[ib->length_dw++] = upper_32_bits(flags);
388 		ib->ptr[ib->length_dw++] = value; /* value */
389 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
390 		ib->ptr[ib->length_dw++] = incr; /* increment size */
391 		ib->ptr[ib->length_dw++] = 0;
392 		pe += ndw * 4;
393 		addr += (ndw / 2) * incr;
394 		count -= ndw / 2;
395 	}
396 }
397 
398 /**
399  * si_dma_ring_pad_ib - pad the IB to the required number of dw
400  *
401  * @ring: amdgpu_ring pointer
402  * @ib: indirect buffer to fill with padding
403  *
404  */
405 static void si_dma_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
406 {
407 	while (ib->length_dw & 0x7)
408 		ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0);
409 }
410 
411 /**
412  * si_dma_ring_emit_pipeline_sync - sync the pipeline
413  *
414  * @ring: amdgpu_ring pointer
415  *
416  * Make sure all previous operations are completed (CIK).
417  */
418 static void si_dma_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
419 {
420 	uint32_t seq = ring->fence_drv.sync_seq;
421 	uint64_t addr = ring->fence_drv.gpu_addr;
422 
423 	/* wait for idle */
424 	amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_POLL_REG_MEM, 0, 0, 0, 0) |
425 			  (1 << 27)); /* Poll memory */
426 	amdgpu_ring_write(ring, lower_32_bits(addr));
427 	amdgpu_ring_write(ring, (0xff << 16) | upper_32_bits(addr)); /* retry, addr_hi */
428 	amdgpu_ring_write(ring, 0xffffffff); /* mask */
429 	amdgpu_ring_write(ring, seq); /* value */
430 	amdgpu_ring_write(ring, (3 << 28) | 0x20); /* func(equal) | poll interval */
431 }
432 
433 /**
434  * si_dma_ring_emit_vm_flush - cik vm flush using sDMA
435  *
436  * @ring: amdgpu_ring pointer
437  * @vmid: vmid number to use
438  * @pd_addr: address
439  *
440  * Update the page table base and flush the VM TLB
441  * using sDMA (VI).
442  */
443 static void si_dma_ring_emit_vm_flush(struct amdgpu_ring *ring,
444 				      unsigned vmid, uint64_t pd_addr)
445 {
446 	amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
447 
448 	/* wait for invalidate to complete */
449 	amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_POLL_REG_MEM, 0, 0, 0, 0));
450 	amdgpu_ring_write(ring, VM_INVALIDATE_REQUEST);
451 	amdgpu_ring_write(ring, 0xff << 16); /* retry */
452 	amdgpu_ring_write(ring, 1 << vmid); /* mask */
453 	amdgpu_ring_write(ring, 0); /* value */
454 	amdgpu_ring_write(ring, (0 << 28) | 0x20); /* func(always) | poll interval */
455 }
456 
457 static void si_dma_ring_emit_wreg(struct amdgpu_ring *ring,
458 				  uint32_t reg, uint32_t val)
459 {
460 	amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0, 0, 0));
461 	amdgpu_ring_write(ring, (0xf << 16) | reg);
462 	amdgpu_ring_write(ring, val);
463 }
464 
465 static int si_dma_early_init(void *handle)
466 {
467 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
468 
469 	adev->sdma.num_instances = 2;
470 
471 	si_dma_set_ring_funcs(adev);
472 	si_dma_set_buffer_funcs(adev);
473 	si_dma_set_vm_pte_funcs(adev);
474 	si_dma_set_irq_funcs(adev);
475 
476 	return 0;
477 }
478 
479 static int si_dma_sw_init(void *handle)
480 {
481 	struct amdgpu_ring *ring;
482 	int r, i;
483 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
484 
485 	/* DMA0 trap event */
486 	r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 224,
487 			      &adev->sdma.trap_irq);
488 	if (r)
489 		return r;
490 
491 	/* DMA1 trap event */
492 	r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 244,
493 			      &adev->sdma.trap_irq);
494 	if (r)
495 		return r;
496 
497 	for (i = 0; i < adev->sdma.num_instances; i++) {
498 		ring = &adev->sdma.instance[i].ring;
499 		ring->ring_obj = NULL;
500 		ring->use_doorbell = false;
501 		sprintf(ring->name, "sdma%d", i);
502 		r = amdgpu_ring_init(adev, ring, 1024,
503 				     &adev->sdma.trap_irq,
504 				     (i == 0) ? AMDGPU_SDMA_IRQ_INSTANCE0 :
505 				     AMDGPU_SDMA_IRQ_INSTANCE1,
506 				     AMDGPU_RING_PRIO_DEFAULT, NULL);
507 		if (r)
508 			return r;
509 	}
510 
511 	return r;
512 }
513 
514 static int si_dma_sw_fini(void *handle)
515 {
516 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
517 	int i;
518 
519 	for (i = 0; i < adev->sdma.num_instances; i++)
520 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
521 
522 	return 0;
523 }
524 
525 static int si_dma_hw_init(void *handle)
526 {
527 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
528 
529 	return si_dma_start(adev);
530 }
531 
532 static int si_dma_hw_fini(void *handle)
533 {
534 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
535 
536 	si_dma_stop(adev);
537 
538 	return 0;
539 }
540 
541 static int si_dma_suspend(void *handle)
542 {
543 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
544 
545 	return si_dma_hw_fini(adev);
546 }
547 
548 static int si_dma_resume(void *handle)
549 {
550 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
551 
552 	return si_dma_hw_init(adev);
553 }
554 
555 static bool si_dma_is_idle(void *handle)
556 {
557 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
558 	u32 tmp = RREG32(SRBM_STATUS2);
559 
560 	if (tmp & (DMA_BUSY_MASK | DMA1_BUSY_MASK))
561 	    return false;
562 
563 	return true;
564 }
565 
566 static int si_dma_wait_for_idle(void *handle)
567 {
568 	unsigned i;
569 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
570 
571 	for (i = 0; i < adev->usec_timeout; i++) {
572 		if (si_dma_is_idle(handle))
573 			return 0;
574 		udelay(1);
575 	}
576 	return -ETIMEDOUT;
577 }
578 
579 static int si_dma_soft_reset(void *handle)
580 {
581 	DRM_INFO("si_dma_soft_reset --- not implemented !!!!!!!\n");
582 	return 0;
583 }
584 
585 static int si_dma_set_trap_irq_state(struct amdgpu_device *adev,
586 					struct amdgpu_irq_src *src,
587 					unsigned type,
588 					enum amdgpu_interrupt_state state)
589 {
590 	u32 sdma_cntl;
591 
592 	switch (type) {
593 	case AMDGPU_SDMA_IRQ_INSTANCE0:
594 		switch (state) {
595 		case AMDGPU_IRQ_STATE_DISABLE:
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 		case AMDGPU_IRQ_STATE_ENABLE:
601 			sdma_cntl = RREG32(DMA_CNTL + DMA0_REGISTER_OFFSET);
602 			sdma_cntl |= TRAP_ENABLE;
603 			WREG32(DMA_CNTL + DMA0_REGISTER_OFFSET, sdma_cntl);
604 			break;
605 		default:
606 			break;
607 		}
608 		break;
609 	case AMDGPU_SDMA_IRQ_INSTANCE1:
610 		switch (state) {
611 		case AMDGPU_IRQ_STATE_DISABLE:
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 		case AMDGPU_IRQ_STATE_ENABLE:
617 			sdma_cntl = RREG32(DMA_CNTL + DMA1_REGISTER_OFFSET);
618 			sdma_cntl |= TRAP_ENABLE;
619 			WREG32(DMA_CNTL + DMA1_REGISTER_OFFSET, sdma_cntl);
620 			break;
621 		default:
622 			break;
623 		}
624 		break;
625 	default:
626 		break;
627 	}
628 	return 0;
629 }
630 
631 static int si_dma_process_trap_irq(struct amdgpu_device *adev,
632 				      struct amdgpu_irq_src *source,
633 				      struct amdgpu_iv_entry *entry)
634 {
635 	if (entry->src_id == 224)
636 		amdgpu_fence_process(&adev->sdma.instance[0].ring);
637 	else
638 		amdgpu_fence_process(&adev->sdma.instance[1].ring);
639 	return 0;
640 }
641 
642 static int si_dma_set_clockgating_state(void *handle,
643 					  enum amd_clockgating_state state)
644 {
645 	u32 orig, data, offset;
646 	int i;
647 	bool enable;
648 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
649 
650 	enable = (state == AMD_CG_STATE_GATE);
651 
652 	if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
653 		for (i = 0; i < adev->sdma.num_instances; i++) {
654 			if (i == 0)
655 				offset = DMA0_REGISTER_OFFSET;
656 			else
657 				offset = DMA1_REGISTER_OFFSET;
658 			orig = data = RREG32(DMA_POWER_CNTL + offset);
659 			data &= ~MEM_POWER_OVERRIDE;
660 			if (data != orig)
661 				WREG32(DMA_POWER_CNTL + offset, data);
662 			WREG32(DMA_CLK_CTRL + offset, 0x00000100);
663 		}
664 	} else {
665 		for (i = 0; i < adev->sdma.num_instances; i++) {
666 			if (i == 0)
667 				offset = DMA0_REGISTER_OFFSET;
668 			else
669 				offset = DMA1_REGISTER_OFFSET;
670 			orig = data = RREG32(DMA_POWER_CNTL + offset);
671 			data |= MEM_POWER_OVERRIDE;
672 			if (data != orig)
673 				WREG32(DMA_POWER_CNTL + offset, data);
674 
675 			orig = data = RREG32(DMA_CLK_CTRL + offset);
676 			data = 0xff000000;
677 			if (data != orig)
678 				WREG32(DMA_CLK_CTRL + offset, data);
679 		}
680 	}
681 
682 	return 0;
683 }
684 
685 static int si_dma_set_powergating_state(void *handle,
686 					  enum amd_powergating_state state)
687 {
688 	u32 tmp;
689 
690 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
691 
692 	WREG32(DMA_PGFSM_WRITE,  0x00002000);
693 	WREG32(DMA_PGFSM_CONFIG, 0x100010ff);
694 
695 	for (tmp = 0; tmp < 5; tmp++)
696 		WREG32(DMA_PGFSM_WRITE, 0);
697 
698 	return 0;
699 }
700 
701 static const struct amd_ip_funcs si_dma_ip_funcs = {
702 	.name = "si_dma",
703 	.early_init = si_dma_early_init,
704 	.late_init = NULL,
705 	.sw_init = si_dma_sw_init,
706 	.sw_fini = si_dma_sw_fini,
707 	.hw_init = si_dma_hw_init,
708 	.hw_fini = si_dma_hw_fini,
709 	.suspend = si_dma_suspend,
710 	.resume = si_dma_resume,
711 	.is_idle = si_dma_is_idle,
712 	.wait_for_idle = si_dma_wait_for_idle,
713 	.soft_reset = si_dma_soft_reset,
714 	.set_clockgating_state = si_dma_set_clockgating_state,
715 	.set_powergating_state = si_dma_set_powergating_state,
716 };
717 
718 static const struct amdgpu_ring_funcs si_dma_ring_funcs = {
719 	.type = AMDGPU_RING_TYPE_SDMA,
720 	.align_mask = 0xf,
721 	.nop = DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0),
722 	.support_64bit_ptrs = false,
723 	.get_rptr = si_dma_ring_get_rptr,
724 	.get_wptr = si_dma_ring_get_wptr,
725 	.set_wptr = si_dma_ring_set_wptr,
726 	.emit_frame_size =
727 		3 + 3 + /* hdp flush / invalidate */
728 		6 + /* si_dma_ring_emit_pipeline_sync */
729 		SI_FLUSH_GPU_TLB_NUM_WREG * 3 + 6 + /* si_dma_ring_emit_vm_flush */
730 		9 + 9 + 9, /* si_dma_ring_emit_fence x3 for user fence, vm fence */
731 	.emit_ib_size = 7 + 3, /* si_dma_ring_emit_ib */
732 	.emit_ib = si_dma_ring_emit_ib,
733 	.emit_fence = si_dma_ring_emit_fence,
734 	.emit_pipeline_sync = si_dma_ring_emit_pipeline_sync,
735 	.emit_vm_flush = si_dma_ring_emit_vm_flush,
736 	.test_ring = si_dma_ring_test_ring,
737 	.test_ib = si_dma_ring_test_ib,
738 	.insert_nop = amdgpu_ring_insert_nop,
739 	.pad_ib = si_dma_ring_pad_ib,
740 	.emit_wreg = si_dma_ring_emit_wreg,
741 };
742 
743 static void si_dma_set_ring_funcs(struct amdgpu_device *adev)
744 {
745 	int i;
746 
747 	for (i = 0; i < adev->sdma.num_instances; i++)
748 		adev->sdma.instance[i].ring.funcs = &si_dma_ring_funcs;
749 }
750 
751 static const struct amdgpu_irq_src_funcs si_dma_trap_irq_funcs = {
752 	.set = si_dma_set_trap_irq_state,
753 	.process = si_dma_process_trap_irq,
754 };
755 
756 static void si_dma_set_irq_funcs(struct amdgpu_device *adev)
757 {
758 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
759 	adev->sdma.trap_irq.funcs = &si_dma_trap_irq_funcs;
760 }
761 
762 /**
763  * si_dma_emit_copy_buffer - copy buffer using the sDMA engine
764  *
765  * @ib: indirect buffer to copy to
766  * @src_offset: src GPU address
767  * @dst_offset: dst GPU address
768  * @byte_count: number of bytes to xfer
769  * @tmz: is this a secure operation
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 				       bool tmz)
780 {
781 	ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_COPY,
782 					      1, 0, 0, byte_count);
783 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
784 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
785 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset) & 0xff;
786 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset) & 0xff;
787 }
788 
789 /**
790  * si_dma_emit_fill_buffer - fill buffer using the sDMA engine
791  *
792  * @ib: indirect buffer to copy to
793  * @src_data: value to write to buffer
794  * @dst_offset: dst GPU address
795  * @byte_count: number of bytes to xfer
796  *
797  * Fill GPU buffers using the DMA engine (VI).
798  */
799 static void si_dma_emit_fill_buffer(struct amdgpu_ib *ib,
800 				       uint32_t src_data,
801 				       uint64_t dst_offset,
802 				       uint32_t byte_count)
803 {
804 	ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_CONSTANT_FILL,
805 					      0, 0, 0, byte_count / 4);
806 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
807 	ib->ptr[ib->length_dw++] = src_data;
808 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset) << 16;
809 }
810 
811 
812 static const struct amdgpu_buffer_funcs si_dma_buffer_funcs = {
813 	.copy_max_bytes = 0xffff8,
814 	.copy_num_dw = 5,
815 	.emit_copy_buffer = si_dma_emit_copy_buffer,
816 
817 	.fill_max_bytes = 0xffff8,
818 	.fill_num_dw = 4,
819 	.emit_fill_buffer = si_dma_emit_fill_buffer,
820 };
821 
822 static void si_dma_set_buffer_funcs(struct amdgpu_device *adev)
823 {
824 	adev->mman.buffer_funcs = &si_dma_buffer_funcs;
825 	adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
826 }
827 
828 static const struct amdgpu_vm_pte_funcs si_dma_vm_pte_funcs = {
829 	.copy_pte_num_dw = 5,
830 	.copy_pte = si_dma_vm_copy_pte,
831 
832 	.write_pte = si_dma_vm_write_pte,
833 	.set_pte_pde = si_dma_vm_set_pte_pde,
834 };
835 
836 static void si_dma_set_vm_pte_funcs(struct amdgpu_device *adev)
837 {
838 	unsigned i;
839 
840 	adev->vm_manager.vm_pte_funcs = &si_dma_vm_pte_funcs;
841 	for (i = 0; i < adev->sdma.num_instances; i++) {
842 		adev->vm_manager.vm_pte_scheds[i] =
843 			&adev->sdma.instance[i].ring.sched;
844 	}
845 	adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
846 }
847 
848 const struct amdgpu_ip_block_version si_dma_ip_block =
849 {
850 	.type = AMD_IP_BLOCK_TYPE_SDMA,
851 	.major = 1,
852 	.minor = 0,
853 	.rev = 0,
854 	.funcs = &si_dma_ip_funcs,
855 };
856