xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/si_dma.c (revision 0cb4228f)
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 		ring->sched.ready = true;
178 
179 		r = amdgpu_ring_test_helper(ring);
180 		if (r)
181 			return r;
182 
183 		if (adev->mman.buffer_funcs_ring == ring)
184 			amdgpu_ttm_set_buffer_funcs_status(adev, true);
185 	}
186 
187 	return 0;
188 }
189 
190 /**
191  * si_dma_ring_test_ring - simple async dma engine test
192  *
193  * @ring: amdgpu_ring structure holding ring information
194  *
195  * Test the DMA engine by writing using it to write an
196  * value to memory. (VI).
197  * Returns 0 for success, error for failure.
198  */
199 static int si_dma_ring_test_ring(struct amdgpu_ring *ring)
200 {
201 	struct amdgpu_device *adev = ring->adev;
202 	unsigned i;
203 	unsigned index;
204 	int r;
205 	u32 tmp;
206 	u64 gpu_addr;
207 
208 	r = amdgpu_device_wb_get(adev, &index);
209 	if (r)
210 		return r;
211 
212 	gpu_addr = adev->wb.gpu_addr + (index * 4);
213 	tmp = 0xCAFEDEAD;
214 	adev->wb.wb[index] = cpu_to_le32(tmp);
215 
216 	r = amdgpu_ring_alloc(ring, 4);
217 	if (r)
218 		goto error_free_wb;
219 
220 	amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 0, 1));
221 	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
222 	amdgpu_ring_write(ring, upper_32_bits(gpu_addr) & 0xff);
223 	amdgpu_ring_write(ring, 0xDEADBEEF);
224 	amdgpu_ring_commit(ring);
225 
226 	for (i = 0; i < adev->usec_timeout; i++) {
227 		tmp = le32_to_cpu(adev->wb.wb[index]);
228 		if (tmp == 0xDEADBEEF)
229 			break;
230 		udelay(1);
231 	}
232 
233 	if (i >= adev->usec_timeout)
234 		r = -ETIMEDOUT;
235 
236 error_free_wb:
237 	amdgpu_device_wb_free(adev, index);
238 	return r;
239 }
240 
241 /**
242  * si_dma_ring_test_ib - test an IB on the DMA engine
243  *
244  * @ring: amdgpu_ring structure holding ring information
245  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
246  *
247  * Test a simple IB in the DMA ring (VI).
248  * Returns 0 on success, error on failure.
249  */
250 static int si_dma_ring_test_ib(struct amdgpu_ring *ring, long timeout)
251 {
252 	struct amdgpu_device *adev = ring->adev;
253 	struct amdgpu_ib ib;
254 	struct dma_fence *f = NULL;
255 	unsigned index;
256 	u32 tmp = 0;
257 	u64 gpu_addr;
258 	long r;
259 
260 	r = amdgpu_device_wb_get(adev, &index);
261 	if (r)
262 		return r;
263 
264 	gpu_addr = adev->wb.gpu_addr + (index * 4);
265 	tmp = 0xCAFEDEAD;
266 	adev->wb.wb[index] = cpu_to_le32(tmp);
267 	memset(&ib, 0, sizeof(ib));
268 	r = amdgpu_ib_get(adev, NULL, 256,
269 					AMDGPU_IB_POOL_DIRECT, &ib);
270 	if (r)
271 		goto err0;
272 
273 	ib.ptr[0] = DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 0, 1);
274 	ib.ptr[1] = lower_32_bits(gpu_addr);
275 	ib.ptr[2] = upper_32_bits(gpu_addr) & 0xff;
276 	ib.ptr[3] = 0xDEADBEEF;
277 	ib.length_dw = 4;
278 	r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
279 	if (r)
280 		goto err1;
281 
282 	r = dma_fence_wait_timeout(f, false, timeout);
283 	if (r == 0) {
284 		r = -ETIMEDOUT;
285 		goto err1;
286 	} else if (r < 0) {
287 		goto err1;
288 	}
289 	tmp = le32_to_cpu(adev->wb.wb[index]);
290 	if (tmp == 0xDEADBEEF)
291 		r = 0;
292 	else
293 		r = -EINVAL;
294 
295 err1:
296 	amdgpu_ib_free(adev, &ib, NULL);
297 	dma_fence_put(f);
298 err0:
299 	amdgpu_device_wb_free(adev, index);
300 	return r;
301 }
302 
303 /**
304  * si_dma_vm_copy_pte - update PTEs by copying them from the GART
305  *
306  * @ib: indirect buffer to fill with commands
307  * @pe: addr of the page entry
308  * @src: src addr to copy from
309  * @count: number of page entries to update
310  *
311  * Update PTEs by copying them from the GART using DMA (SI).
312  */
313 static void si_dma_vm_copy_pte(struct amdgpu_ib *ib,
314 			       uint64_t pe, uint64_t src,
315 			       unsigned count)
316 {
317 	unsigned bytes = count * 8;
318 
319 	ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_COPY,
320 					      1, 0, 0, bytes);
321 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
322 	ib->ptr[ib->length_dw++] = lower_32_bits(src);
323 	ib->ptr[ib->length_dw++] = upper_32_bits(pe) & 0xff;
324 	ib->ptr[ib->length_dw++] = upper_32_bits(src) & 0xff;
325 }
326 
327 /**
328  * si_dma_vm_write_pte - update PTEs by writing them manually
329  *
330  * @ib: indirect buffer to fill with commands
331  * @pe: addr of the page entry
332  * @value: dst addr to write into pe
333  * @count: number of page entries to update
334  * @incr: increase next addr by incr bytes
335  *
336  * Update PTEs by writing them manually using DMA (SI).
337  */
338 static void si_dma_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
339 				uint64_t value, unsigned count,
340 				uint32_t incr)
341 {
342 	unsigned ndw = count * 2;
343 
344 	ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 0, ndw);
345 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
346 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
347 	for (; ndw > 0; ndw -= 2) {
348 		ib->ptr[ib->length_dw++] = lower_32_bits(value);
349 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
350 		value += incr;
351 	}
352 }
353 
354 /**
355  * si_dma_vm_set_pte_pde - update the page tables using sDMA
356  *
357  * @ib: indirect buffer to fill with commands
358  * @pe: addr of the page entry
359  * @addr: dst addr to write into pe
360  * @count: number of page entries to update
361  * @incr: increase next addr by incr bytes
362  * @flags: access flags
363  *
364  * Update the page tables using sDMA (CIK).
365  */
366 static void si_dma_vm_set_pte_pde(struct amdgpu_ib *ib,
367 				     uint64_t pe,
368 				     uint64_t addr, unsigned count,
369 				     uint32_t incr, uint64_t flags)
370 {
371 	uint64_t value;
372 	unsigned ndw;
373 
374 	while (count) {
375 		ndw = count * 2;
376 		if (ndw > 0xFFFFE)
377 			ndw = 0xFFFFE;
378 
379 		if (flags & AMDGPU_PTE_VALID)
380 			value = addr;
381 		else
382 			value = 0;
383 
384 		/* for physically contiguous pages (vram) */
385 		ib->ptr[ib->length_dw++] = DMA_PTE_PDE_PACKET(ndw);
386 		ib->ptr[ib->length_dw++] = pe; /* dst addr */
387 		ib->ptr[ib->length_dw++] = upper_32_bits(pe) & 0xff;
388 		ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
389 		ib->ptr[ib->length_dw++] = upper_32_bits(flags);
390 		ib->ptr[ib->length_dw++] = value; /* value */
391 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
392 		ib->ptr[ib->length_dw++] = incr; /* increment size */
393 		ib->ptr[ib->length_dw++] = 0;
394 		pe += ndw * 4;
395 		addr += (ndw / 2) * incr;
396 		count -= ndw / 2;
397 	}
398 }
399 
400 /**
401  * si_dma_ring_pad_ib - pad the IB to the required number of dw
402  *
403  * @ring: amdgpu_ring pointer
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  * si_dma_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  * @vmid: vmid number to use
440  * @pd_addr: address
441  *
442  * Update the page table base and flush the VM TLB
443  * using sDMA (VI).
444  */
445 static void si_dma_ring_emit_vm_flush(struct amdgpu_ring *ring,
446 				      unsigned vmid, uint64_t pd_addr)
447 {
448 	amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
449 
450 	/* wait for invalidate to complete */
451 	amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_POLL_REG_MEM, 0, 0, 0, 0));
452 	amdgpu_ring_write(ring, VM_INVALIDATE_REQUEST);
453 	amdgpu_ring_write(ring, 0xff << 16); /* retry */
454 	amdgpu_ring_write(ring, 1 << vmid); /* mask */
455 	amdgpu_ring_write(ring, 0); /* value */
456 	amdgpu_ring_write(ring, (0 << 28) | 0x20); /* func(always) | poll interval */
457 }
458 
459 static void si_dma_ring_emit_wreg(struct amdgpu_ring *ring,
460 				  uint32_t reg, uint32_t val)
461 {
462 	amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0, 0, 0));
463 	amdgpu_ring_write(ring, (0xf << 16) | reg);
464 	amdgpu_ring_write(ring, val);
465 }
466 
467 static int si_dma_early_init(void *handle)
468 {
469 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
470 
471 	adev->sdma.num_instances = 2;
472 
473 	si_dma_set_ring_funcs(adev);
474 	si_dma_set_buffer_funcs(adev);
475 	si_dma_set_vm_pte_funcs(adev);
476 	si_dma_set_irq_funcs(adev);
477 
478 	return 0;
479 }
480 
481 static int si_dma_sw_init(void *handle)
482 {
483 	struct amdgpu_ring *ring;
484 	int r, i;
485 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
486 
487 	/* DMA0 trap event */
488 	r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 224,
489 			      &adev->sdma.trap_irq);
490 	if (r)
491 		return r;
492 
493 	/* DMA1 trap event */
494 	r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 244,
495 			      &adev->sdma.trap_irq);
496 	if (r)
497 		return r;
498 
499 	for (i = 0; i < adev->sdma.num_instances; i++) {
500 		ring = &adev->sdma.instance[i].ring;
501 		ring->ring_obj = NULL;
502 		ring->use_doorbell = false;
503 		sprintf(ring->name, "sdma%d", i);
504 		r = amdgpu_ring_init(adev, ring, 1024,
505 				     &adev->sdma.trap_irq,
506 				     (i == 0) ? AMDGPU_SDMA_IRQ_INSTANCE0 :
507 				     AMDGPU_SDMA_IRQ_INSTANCE1,
508 				     AMDGPU_RING_PRIO_DEFAULT, NULL);
509 		if (r)
510 			return r;
511 	}
512 
513 	return r;
514 }
515 
516 static int si_dma_sw_fini(void *handle)
517 {
518 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
519 	int i;
520 
521 	for (i = 0; i < adev->sdma.num_instances; i++)
522 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
523 
524 	return 0;
525 }
526 
527 static int si_dma_hw_init(void *handle)
528 {
529 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
530 
531 	return si_dma_start(adev);
532 }
533 
534 static int si_dma_hw_fini(void *handle)
535 {
536 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
537 
538 	si_dma_stop(adev);
539 
540 	return 0;
541 }
542 
543 static int si_dma_suspend(void *handle)
544 {
545 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
546 
547 	return si_dma_hw_fini(adev);
548 }
549 
550 static int si_dma_resume(void *handle)
551 {
552 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
553 
554 	return si_dma_hw_init(adev);
555 }
556 
557 static bool si_dma_is_idle(void *handle)
558 {
559 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
560 	u32 tmp = RREG32(SRBM_STATUS2);
561 
562 	if (tmp & (DMA_BUSY_MASK | DMA1_BUSY_MASK))
563 	    return false;
564 
565 	return true;
566 }
567 
568 static int si_dma_wait_for_idle(void *handle)
569 {
570 	unsigned i;
571 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
572 
573 	for (i = 0; i < adev->usec_timeout; i++) {
574 		if (si_dma_is_idle(handle))
575 			return 0;
576 		udelay(1);
577 	}
578 	return -ETIMEDOUT;
579 }
580 
581 static int si_dma_soft_reset(void *handle)
582 {
583 	DRM_INFO("si_dma_soft_reset --- not implemented !!!!!!!\n");
584 	return 0;
585 }
586 
587 static int si_dma_set_trap_irq_state(struct amdgpu_device *adev,
588 					struct amdgpu_irq_src *src,
589 					unsigned type,
590 					enum amdgpu_interrupt_state state)
591 {
592 	u32 sdma_cntl;
593 
594 	switch (type) {
595 	case AMDGPU_SDMA_IRQ_INSTANCE0:
596 		switch (state) {
597 		case AMDGPU_IRQ_STATE_DISABLE:
598 			sdma_cntl = RREG32(DMA_CNTL + DMA0_REGISTER_OFFSET);
599 			sdma_cntl &= ~TRAP_ENABLE;
600 			WREG32(DMA_CNTL + DMA0_REGISTER_OFFSET, sdma_cntl);
601 			break;
602 		case AMDGPU_IRQ_STATE_ENABLE:
603 			sdma_cntl = RREG32(DMA_CNTL + DMA0_REGISTER_OFFSET);
604 			sdma_cntl |= TRAP_ENABLE;
605 			WREG32(DMA_CNTL + DMA0_REGISTER_OFFSET, sdma_cntl);
606 			break;
607 		default:
608 			break;
609 		}
610 		break;
611 	case AMDGPU_SDMA_IRQ_INSTANCE1:
612 		switch (state) {
613 		case AMDGPU_IRQ_STATE_DISABLE:
614 			sdma_cntl = RREG32(DMA_CNTL + DMA1_REGISTER_OFFSET);
615 			sdma_cntl &= ~TRAP_ENABLE;
616 			WREG32(DMA_CNTL + DMA1_REGISTER_OFFSET, sdma_cntl);
617 			break;
618 		case AMDGPU_IRQ_STATE_ENABLE:
619 			sdma_cntl = RREG32(DMA_CNTL + DMA1_REGISTER_OFFSET);
620 			sdma_cntl |= TRAP_ENABLE;
621 			WREG32(DMA_CNTL + DMA1_REGISTER_OFFSET, sdma_cntl);
622 			break;
623 		default:
624 			break;
625 		}
626 		break;
627 	default:
628 		break;
629 	}
630 	return 0;
631 }
632 
633 static int si_dma_process_trap_irq(struct amdgpu_device *adev,
634 				      struct amdgpu_irq_src *source,
635 				      struct amdgpu_iv_entry *entry)
636 {
637 	if (entry->src_id == 224)
638 		amdgpu_fence_process(&adev->sdma.instance[0].ring);
639 	else
640 		amdgpu_fence_process(&adev->sdma.instance[1].ring);
641 	return 0;
642 }
643 
644 static int si_dma_set_clockgating_state(void *handle,
645 					  enum amd_clockgating_state state)
646 {
647 	u32 orig, data, offset;
648 	int i;
649 	bool enable;
650 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
651 
652 	enable = (state == AMD_CG_STATE_GATE);
653 
654 	if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
655 		for (i = 0; i < adev->sdma.num_instances; i++) {
656 			if (i == 0)
657 				offset = DMA0_REGISTER_OFFSET;
658 			else
659 				offset = DMA1_REGISTER_OFFSET;
660 			orig = data = RREG32(DMA_POWER_CNTL + offset);
661 			data &= ~MEM_POWER_OVERRIDE;
662 			if (data != orig)
663 				WREG32(DMA_POWER_CNTL + offset, data);
664 			WREG32(DMA_CLK_CTRL + offset, 0x00000100);
665 		}
666 	} else {
667 		for (i = 0; i < adev->sdma.num_instances; i++) {
668 			if (i == 0)
669 				offset = DMA0_REGISTER_OFFSET;
670 			else
671 				offset = DMA1_REGISTER_OFFSET;
672 			orig = data = RREG32(DMA_POWER_CNTL + offset);
673 			data |= MEM_POWER_OVERRIDE;
674 			if (data != orig)
675 				WREG32(DMA_POWER_CNTL + offset, data);
676 
677 			orig = data = RREG32(DMA_CLK_CTRL + offset);
678 			data = 0xff000000;
679 			if (data != orig)
680 				WREG32(DMA_CLK_CTRL + offset, data);
681 		}
682 	}
683 
684 	return 0;
685 }
686 
687 static int si_dma_set_powergating_state(void *handle,
688 					  enum amd_powergating_state state)
689 {
690 	u32 tmp;
691 
692 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
693 
694 	WREG32(DMA_PGFSM_WRITE,  0x00002000);
695 	WREG32(DMA_PGFSM_CONFIG, 0x100010ff);
696 
697 	for (tmp = 0; tmp < 5; tmp++)
698 		WREG32(DMA_PGFSM_WRITE, 0);
699 
700 	return 0;
701 }
702 
703 static const struct amd_ip_funcs si_dma_ip_funcs = {
704 	.name = "si_dma",
705 	.early_init = si_dma_early_init,
706 	.late_init = NULL,
707 	.sw_init = si_dma_sw_init,
708 	.sw_fini = si_dma_sw_fini,
709 	.hw_init = si_dma_hw_init,
710 	.hw_fini = si_dma_hw_fini,
711 	.suspend = si_dma_suspend,
712 	.resume = si_dma_resume,
713 	.is_idle = si_dma_is_idle,
714 	.wait_for_idle = si_dma_wait_for_idle,
715 	.soft_reset = si_dma_soft_reset,
716 	.set_clockgating_state = si_dma_set_clockgating_state,
717 	.set_powergating_state = si_dma_set_powergating_state,
718 };
719 
720 static const struct amdgpu_ring_funcs si_dma_ring_funcs = {
721 	.type = AMDGPU_RING_TYPE_SDMA,
722 	.align_mask = 0xf,
723 	.nop = DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0),
724 	.support_64bit_ptrs = false,
725 	.get_rptr = si_dma_ring_get_rptr,
726 	.get_wptr = si_dma_ring_get_wptr,
727 	.set_wptr = si_dma_ring_set_wptr,
728 	.emit_frame_size =
729 		3 + 3 + /* hdp flush / invalidate */
730 		6 + /* si_dma_ring_emit_pipeline_sync */
731 		SI_FLUSH_GPU_TLB_NUM_WREG * 3 + 6 + /* si_dma_ring_emit_vm_flush */
732 		9 + 9 + 9, /* si_dma_ring_emit_fence x3 for user fence, vm fence */
733 	.emit_ib_size = 7 + 3, /* si_dma_ring_emit_ib */
734 	.emit_ib = si_dma_ring_emit_ib,
735 	.emit_fence = si_dma_ring_emit_fence,
736 	.emit_pipeline_sync = si_dma_ring_emit_pipeline_sync,
737 	.emit_vm_flush = si_dma_ring_emit_vm_flush,
738 	.test_ring = si_dma_ring_test_ring,
739 	.test_ib = si_dma_ring_test_ib,
740 	.insert_nop = amdgpu_ring_insert_nop,
741 	.pad_ib = si_dma_ring_pad_ib,
742 	.emit_wreg = si_dma_ring_emit_wreg,
743 };
744 
745 static void si_dma_set_ring_funcs(struct amdgpu_device *adev)
746 {
747 	int i;
748 
749 	for (i = 0; i < adev->sdma.num_instances; i++)
750 		adev->sdma.instance[i].ring.funcs = &si_dma_ring_funcs;
751 }
752 
753 static const struct amdgpu_irq_src_funcs si_dma_trap_irq_funcs = {
754 	.set = si_dma_set_trap_irq_state,
755 	.process = si_dma_process_trap_irq,
756 };
757 
758 static void si_dma_set_irq_funcs(struct amdgpu_device *adev)
759 {
760 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
761 	adev->sdma.trap_irq.funcs = &si_dma_trap_irq_funcs;
762 }
763 
764 /**
765  * si_dma_emit_copy_buffer - copy buffer using the sDMA engine
766  *
767  * @ib: indirect buffer to copy to
768  * @src_offset: src GPU address
769  * @dst_offset: dst GPU address
770  * @byte_count: number of bytes to xfer
771  * @tmz: is this a secure operation
772  *
773  * Copy GPU buffers using the DMA engine (VI).
774  * Used by the amdgpu ttm implementation to move pages if
775  * registered as the asic copy callback.
776  */
777 static void si_dma_emit_copy_buffer(struct amdgpu_ib *ib,
778 				       uint64_t src_offset,
779 				       uint64_t dst_offset,
780 				       uint32_t byte_count,
781 				       bool tmz)
782 {
783 	ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_COPY,
784 					      1, 0, 0, byte_count);
785 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
786 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
787 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset) & 0xff;
788 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset) & 0xff;
789 }
790 
791 /**
792  * si_dma_emit_fill_buffer - fill buffer using the sDMA engine
793  *
794  * @ib: indirect buffer to copy to
795  * @src_data: value to write to buffer
796  * @dst_offset: dst GPU address
797  * @byte_count: number of bytes to xfer
798  *
799  * Fill GPU buffers using the DMA engine (VI).
800  */
801 static void si_dma_emit_fill_buffer(struct amdgpu_ib *ib,
802 				       uint32_t src_data,
803 				       uint64_t dst_offset,
804 				       uint32_t byte_count)
805 {
806 	ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_CONSTANT_FILL,
807 					      0, 0, 0, byte_count / 4);
808 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
809 	ib->ptr[ib->length_dw++] = src_data;
810 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset) << 16;
811 }
812 
813 
814 static const struct amdgpu_buffer_funcs si_dma_buffer_funcs = {
815 	.copy_max_bytes = 0xffff8,
816 	.copy_num_dw = 5,
817 	.emit_copy_buffer = si_dma_emit_copy_buffer,
818 
819 	.fill_max_bytes = 0xffff8,
820 	.fill_num_dw = 4,
821 	.emit_fill_buffer = si_dma_emit_fill_buffer,
822 };
823 
824 static void si_dma_set_buffer_funcs(struct amdgpu_device *adev)
825 {
826 	adev->mman.buffer_funcs = &si_dma_buffer_funcs;
827 	adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
828 }
829 
830 static const struct amdgpu_vm_pte_funcs si_dma_vm_pte_funcs = {
831 	.copy_pte_num_dw = 5,
832 	.copy_pte = si_dma_vm_copy_pte,
833 
834 	.write_pte = si_dma_vm_write_pte,
835 	.set_pte_pde = si_dma_vm_set_pte_pde,
836 };
837 
838 static void si_dma_set_vm_pte_funcs(struct amdgpu_device *adev)
839 {
840 	unsigned i;
841 
842 	adev->vm_manager.vm_pte_funcs = &si_dma_vm_pte_funcs;
843 	for (i = 0; i < adev->sdma.num_instances; i++) {
844 		adev->vm_manager.vm_pte_scheds[i] =
845 			&adev->sdma.instance[i].ring.sched;
846 	}
847 	adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
848 }
849 
850 const struct amdgpu_ip_block_version si_dma_ip_block =
851 {
852 	.type = AMD_IP_BLOCK_TYPE_SDMA,
853 	.major = 1,
854 	.minor = 0,
855 	.rev = 0,
856 	.funcs = &si_dma_ip_funcs,
857 };
858