xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/si_dma.c (revision e3d786a3)
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 
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_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  */
88 static 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 
110 static 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 
129 static 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  */
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 		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  */
261 static 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 
314 err1:
315 	amdgpu_ib_free(adev, &ib, NULL);
316 	dma_fence_put(f);
317 err0:
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  */
332 static 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  */
357 static 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  */
385 static 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  */
425 static 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  */
438 static 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  */
462 static 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 
476 static 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 
484 static 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 
498 static 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_IRQ_CLIENTID_LEGACY, 224,
506 			      &adev->sdma.trap_irq);
507 	if (r)
508 		return r;
509 
510 	/* DMA1 trap event */
511 	r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 244,
512 			      &adev->sdma.trap_irq);
513 	if (r)
514 		return r;
515 
516 	for (i = 0; i < adev->sdma.num_instances; i++) {
517 		ring = &adev->sdma.instance[i].ring;
518 		ring->ring_obj = NULL;
519 		ring->use_doorbell = false;
520 		sprintf(ring->name, "sdma%d", i);
521 		r = amdgpu_ring_init(adev, ring, 1024,
522 				     &adev->sdma.trap_irq,
523 				     (i == 0) ?
524 				     AMDGPU_SDMA_IRQ_TRAP0 :
525 				     AMDGPU_SDMA_IRQ_TRAP1);
526 		if (r)
527 			return r;
528 	}
529 
530 	return r;
531 }
532 
533 static int si_dma_sw_fini(void *handle)
534 {
535 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
536 	int i;
537 
538 	for (i = 0; i < adev->sdma.num_instances; i++)
539 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
540 
541 	return 0;
542 }
543 
544 static int si_dma_hw_init(void *handle)
545 {
546 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
547 
548 	return si_dma_start(adev);
549 }
550 
551 static int si_dma_hw_fini(void *handle)
552 {
553 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
554 
555 	si_dma_stop(adev);
556 
557 	return 0;
558 }
559 
560 static int si_dma_suspend(void *handle)
561 {
562 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
563 
564 	return si_dma_hw_fini(adev);
565 }
566 
567 static int si_dma_resume(void *handle)
568 {
569 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
570 
571 	return si_dma_hw_init(adev);
572 }
573 
574 static bool si_dma_is_idle(void *handle)
575 {
576 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
577 	u32 tmp = RREG32(SRBM_STATUS2);
578 
579 	if (tmp & (DMA_BUSY_MASK | DMA1_BUSY_MASK))
580 	    return false;
581 
582 	return true;
583 }
584 
585 static int si_dma_wait_for_idle(void *handle)
586 {
587 	unsigned i;
588 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
589 
590 	for (i = 0; i < adev->usec_timeout; i++) {
591 		if (si_dma_is_idle(handle))
592 			return 0;
593 		udelay(1);
594 	}
595 	return -ETIMEDOUT;
596 }
597 
598 static int si_dma_soft_reset(void *handle)
599 {
600 	DRM_INFO("si_dma_soft_reset --- not implemented !!!!!!!\n");
601 	return 0;
602 }
603 
604 static int si_dma_set_trap_irq_state(struct amdgpu_device *adev,
605 					struct amdgpu_irq_src *src,
606 					unsigned type,
607 					enum amdgpu_interrupt_state state)
608 {
609 	u32 sdma_cntl;
610 
611 	switch (type) {
612 	case AMDGPU_SDMA_IRQ_TRAP0:
613 		switch (state) {
614 		case AMDGPU_IRQ_STATE_DISABLE:
615 			sdma_cntl = RREG32(DMA_CNTL + DMA0_REGISTER_OFFSET);
616 			sdma_cntl &= ~TRAP_ENABLE;
617 			WREG32(DMA_CNTL + DMA0_REGISTER_OFFSET, sdma_cntl);
618 			break;
619 		case AMDGPU_IRQ_STATE_ENABLE:
620 			sdma_cntl = RREG32(DMA_CNTL + DMA0_REGISTER_OFFSET);
621 			sdma_cntl |= TRAP_ENABLE;
622 			WREG32(DMA_CNTL + DMA0_REGISTER_OFFSET, sdma_cntl);
623 			break;
624 		default:
625 			break;
626 		}
627 		break;
628 	case AMDGPU_SDMA_IRQ_TRAP1:
629 		switch (state) {
630 		case AMDGPU_IRQ_STATE_DISABLE:
631 			sdma_cntl = RREG32(DMA_CNTL + DMA1_REGISTER_OFFSET);
632 			sdma_cntl &= ~TRAP_ENABLE;
633 			WREG32(DMA_CNTL + DMA1_REGISTER_OFFSET, sdma_cntl);
634 			break;
635 		case AMDGPU_IRQ_STATE_ENABLE:
636 			sdma_cntl = RREG32(DMA_CNTL + DMA1_REGISTER_OFFSET);
637 			sdma_cntl |= TRAP_ENABLE;
638 			WREG32(DMA_CNTL + DMA1_REGISTER_OFFSET, sdma_cntl);
639 			break;
640 		default:
641 			break;
642 		}
643 		break;
644 	default:
645 		break;
646 	}
647 	return 0;
648 }
649 
650 static int si_dma_process_trap_irq(struct amdgpu_device *adev,
651 				      struct amdgpu_irq_src *source,
652 				      struct amdgpu_iv_entry *entry)
653 {
654 	if (entry->src_id == 224)
655 		amdgpu_fence_process(&adev->sdma.instance[0].ring);
656 	else
657 		amdgpu_fence_process(&adev->sdma.instance[1].ring);
658 	return 0;
659 }
660 
661 static int si_dma_process_illegal_inst_irq(struct amdgpu_device *adev,
662 					      struct amdgpu_irq_src *source,
663 					      struct amdgpu_iv_entry *entry)
664 {
665 	DRM_ERROR("Illegal instruction in SDMA command stream\n");
666 	schedule_work(&adev->reset_work);
667 	return 0;
668 }
669 
670 static int si_dma_set_clockgating_state(void *handle,
671 					  enum amd_clockgating_state state)
672 {
673 	u32 orig, data, offset;
674 	int i;
675 	bool enable;
676 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
677 
678 	enable = (state == AMD_CG_STATE_GATE) ? true : false;
679 
680 	if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
681 		for (i = 0; i < adev->sdma.num_instances; i++) {
682 			if (i == 0)
683 				offset = DMA0_REGISTER_OFFSET;
684 			else
685 				offset = DMA1_REGISTER_OFFSET;
686 			orig = data = RREG32(DMA_POWER_CNTL + offset);
687 			data &= ~MEM_POWER_OVERRIDE;
688 			if (data != orig)
689 				WREG32(DMA_POWER_CNTL + offset, data);
690 			WREG32(DMA_CLK_CTRL + offset, 0x00000100);
691 		}
692 	} else {
693 		for (i = 0; i < adev->sdma.num_instances; i++) {
694 			if (i == 0)
695 				offset = DMA0_REGISTER_OFFSET;
696 			else
697 				offset = DMA1_REGISTER_OFFSET;
698 			orig = data = RREG32(DMA_POWER_CNTL + offset);
699 			data |= MEM_POWER_OVERRIDE;
700 			if (data != orig)
701 				WREG32(DMA_POWER_CNTL + offset, data);
702 
703 			orig = data = RREG32(DMA_CLK_CTRL + offset);
704 			data = 0xff000000;
705 			if (data != orig)
706 				WREG32(DMA_CLK_CTRL + offset, data);
707 		}
708 	}
709 
710 	return 0;
711 }
712 
713 static int si_dma_set_powergating_state(void *handle,
714 					  enum amd_powergating_state state)
715 {
716 	u32 tmp;
717 
718 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
719 
720 	WREG32(DMA_PGFSM_WRITE,  0x00002000);
721 	WREG32(DMA_PGFSM_CONFIG, 0x100010ff);
722 
723 	for (tmp = 0; tmp < 5; tmp++)
724 		WREG32(DMA_PGFSM_WRITE, 0);
725 
726 	return 0;
727 }
728 
729 static const struct amd_ip_funcs si_dma_ip_funcs = {
730 	.name = "si_dma",
731 	.early_init = si_dma_early_init,
732 	.late_init = NULL,
733 	.sw_init = si_dma_sw_init,
734 	.sw_fini = si_dma_sw_fini,
735 	.hw_init = si_dma_hw_init,
736 	.hw_fini = si_dma_hw_fini,
737 	.suspend = si_dma_suspend,
738 	.resume = si_dma_resume,
739 	.is_idle = si_dma_is_idle,
740 	.wait_for_idle = si_dma_wait_for_idle,
741 	.soft_reset = si_dma_soft_reset,
742 	.set_clockgating_state = si_dma_set_clockgating_state,
743 	.set_powergating_state = si_dma_set_powergating_state,
744 };
745 
746 static const struct amdgpu_ring_funcs si_dma_ring_funcs = {
747 	.type = AMDGPU_RING_TYPE_SDMA,
748 	.align_mask = 0xf,
749 	.nop = DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0),
750 	.support_64bit_ptrs = false,
751 	.get_rptr = si_dma_ring_get_rptr,
752 	.get_wptr = si_dma_ring_get_wptr,
753 	.set_wptr = si_dma_ring_set_wptr,
754 	.emit_frame_size =
755 		3 + 3 + /* hdp flush / invalidate */
756 		6 + /* si_dma_ring_emit_pipeline_sync */
757 		SI_FLUSH_GPU_TLB_NUM_WREG * 3 + 6 + /* si_dma_ring_emit_vm_flush */
758 		9 + 9 + 9, /* si_dma_ring_emit_fence x3 for user fence, vm fence */
759 	.emit_ib_size = 7 + 3, /* si_dma_ring_emit_ib */
760 	.emit_ib = si_dma_ring_emit_ib,
761 	.emit_fence = si_dma_ring_emit_fence,
762 	.emit_pipeline_sync = si_dma_ring_emit_pipeline_sync,
763 	.emit_vm_flush = si_dma_ring_emit_vm_flush,
764 	.test_ring = si_dma_ring_test_ring,
765 	.test_ib = si_dma_ring_test_ib,
766 	.insert_nop = amdgpu_ring_insert_nop,
767 	.pad_ib = si_dma_ring_pad_ib,
768 	.emit_wreg = si_dma_ring_emit_wreg,
769 };
770 
771 static void si_dma_set_ring_funcs(struct amdgpu_device *adev)
772 {
773 	int i;
774 
775 	for (i = 0; i < adev->sdma.num_instances; i++)
776 		adev->sdma.instance[i].ring.funcs = &si_dma_ring_funcs;
777 }
778 
779 static const struct amdgpu_irq_src_funcs si_dma_trap_irq_funcs = {
780 	.set = si_dma_set_trap_irq_state,
781 	.process = si_dma_process_trap_irq,
782 };
783 
784 static const struct amdgpu_irq_src_funcs si_dma_illegal_inst_irq_funcs = {
785 	.process = si_dma_process_illegal_inst_irq,
786 };
787 
788 static void si_dma_set_irq_funcs(struct amdgpu_device *adev)
789 {
790 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
791 	adev->sdma.trap_irq.funcs = &si_dma_trap_irq_funcs;
792 	adev->sdma.illegal_inst_irq.funcs = &si_dma_illegal_inst_irq_funcs;
793 }
794 
795 /**
796  * si_dma_emit_copy_buffer - copy buffer using the sDMA engine
797  *
798  * @ring: amdgpu_ring structure holding ring information
799  * @src_offset: src GPU address
800  * @dst_offset: dst GPU address
801  * @byte_count: number of bytes to xfer
802  *
803  * Copy GPU buffers using the DMA engine (VI).
804  * Used by the amdgpu ttm implementation to move pages if
805  * registered as the asic copy callback.
806  */
807 static void si_dma_emit_copy_buffer(struct amdgpu_ib *ib,
808 				       uint64_t src_offset,
809 				       uint64_t dst_offset,
810 				       uint32_t byte_count)
811 {
812 	ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_COPY,
813 					      1, 0, 0, byte_count);
814 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
815 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
816 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset) & 0xff;
817 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset) & 0xff;
818 }
819 
820 /**
821  * si_dma_emit_fill_buffer - fill buffer using the sDMA engine
822  *
823  * @ring: amdgpu_ring structure holding ring information
824  * @src_data: value to write to buffer
825  * @dst_offset: dst GPU address
826  * @byte_count: number of bytes to xfer
827  *
828  * Fill GPU buffers using the DMA engine (VI).
829  */
830 static void si_dma_emit_fill_buffer(struct amdgpu_ib *ib,
831 				       uint32_t src_data,
832 				       uint64_t dst_offset,
833 				       uint32_t byte_count)
834 {
835 	ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_CONSTANT_FILL,
836 					      0, 0, 0, byte_count / 4);
837 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
838 	ib->ptr[ib->length_dw++] = src_data;
839 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset) << 16;
840 }
841 
842 
843 static const struct amdgpu_buffer_funcs si_dma_buffer_funcs = {
844 	.copy_max_bytes = 0xffff8,
845 	.copy_num_dw = 5,
846 	.emit_copy_buffer = si_dma_emit_copy_buffer,
847 
848 	.fill_max_bytes = 0xffff8,
849 	.fill_num_dw = 4,
850 	.emit_fill_buffer = si_dma_emit_fill_buffer,
851 };
852 
853 static void si_dma_set_buffer_funcs(struct amdgpu_device *adev)
854 {
855 	adev->mman.buffer_funcs = &si_dma_buffer_funcs;
856 	adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
857 }
858 
859 static const struct amdgpu_vm_pte_funcs si_dma_vm_pte_funcs = {
860 	.copy_pte_num_dw = 5,
861 	.copy_pte = si_dma_vm_copy_pte,
862 
863 	.write_pte = si_dma_vm_write_pte,
864 	.set_pte_pde = si_dma_vm_set_pte_pde,
865 };
866 
867 static void si_dma_set_vm_pte_funcs(struct amdgpu_device *adev)
868 {
869 	struct drm_gpu_scheduler *sched;
870 	unsigned i;
871 
872 	adev->vm_manager.vm_pte_funcs = &si_dma_vm_pte_funcs;
873 	for (i = 0; i < adev->sdma.num_instances; i++) {
874 		sched = &adev->sdma.instance[i].ring.sched;
875 		adev->vm_manager.vm_pte_rqs[i] =
876 			&sched->sched_rq[DRM_SCHED_PRIORITY_KERNEL];
877 	}
878 	adev->vm_manager.vm_pte_num_rqs = adev->sdma.num_instances;
879 }
880 
881 const struct amdgpu_ip_block_version si_dma_ip_block =
882 {
883 	.type = AMD_IP_BLOCK_TYPE_SDMA,
884 	.major = 1,
885 	.minor = 0,
886 	.rev = 0,
887 	.funcs = &si_dma_ip_funcs,
888 };
889