xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/si_dma.c (revision 3cf3cdea)
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  * @addr: address
85  * @seq: sequence number
86  * @flags: fence related flags
87  *
88  * Add a DMA fence packet to the ring to write
89  * the fence seq number and DMA trap packet to generate
90  * an interrupt if needed (VI).
91  */
92 static void si_dma_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
93 				      unsigned flags)
94 {
95 
96 	bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
97 	/* write the fence */
98 	amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_FENCE, 0, 0, 0, 0));
99 	amdgpu_ring_write(ring, addr & 0xfffffffc);
100 	amdgpu_ring_write(ring, (upper_32_bits(addr) & 0xff));
101 	amdgpu_ring_write(ring, seq);
102 	/* optionally write high bits as well */
103 	if (write64bit) {
104 		addr += 4;
105 		amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_FENCE, 0, 0, 0, 0));
106 		amdgpu_ring_write(ring, addr & 0xfffffffc);
107 		amdgpu_ring_write(ring, (upper_32_bits(addr) & 0xff));
108 		amdgpu_ring_write(ring, upper_32_bits(seq));
109 	}
110 	/* generate an interrupt */
111 	amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_TRAP, 0, 0, 0, 0));
112 }
113 
114 static void si_dma_stop(struct amdgpu_device *adev)
115 {
116 	struct amdgpu_ring *ring;
117 	u32 rb_cntl;
118 	unsigned i;
119 
120 	for (i = 0; i < adev->sdma.num_instances; i++) {
121 		ring = &adev->sdma.instance[i].ring;
122 		/* dma0 */
123 		rb_cntl = RREG32(DMA_RB_CNTL + sdma_offsets[i]);
124 		rb_cntl &= ~DMA_RB_ENABLE;
125 		WREG32(DMA_RB_CNTL + sdma_offsets[i], rb_cntl);
126 
127 		if (adev->mman.buffer_funcs_ring == ring)
128 			amdgpu_ttm_set_buffer_funcs_status(adev, false);
129 	}
130 }
131 
132 static int si_dma_start(struct amdgpu_device *adev)
133 {
134 	struct amdgpu_ring *ring;
135 	u32 rb_cntl, dma_cntl, ib_cntl, rb_bufsz;
136 	int i, r;
137 	uint64_t rptr_addr;
138 
139 	for (i = 0; i < adev->sdma.num_instances; i++) {
140 		ring = &adev->sdma.instance[i].ring;
141 
142 		WREG32(DMA_SEM_INCOMPLETE_TIMER_CNTL + sdma_offsets[i], 0);
143 		WREG32(DMA_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
144 
145 		/* Set ring buffer size in dwords */
146 		rb_bufsz = order_base_2(ring->ring_size / 4);
147 		rb_cntl = rb_bufsz << 1;
148 #ifdef __BIG_ENDIAN
149 		rb_cntl |= DMA_RB_SWAP_ENABLE | DMA_RPTR_WRITEBACK_SWAP_ENABLE;
150 #endif
151 		WREG32(DMA_RB_CNTL + sdma_offsets[i], rb_cntl);
152 
153 		/* Initialize the ring buffer's read and write pointers */
154 		WREG32(DMA_RB_RPTR + sdma_offsets[i], 0);
155 		WREG32(DMA_RB_WPTR + sdma_offsets[i], 0);
156 
157 		rptr_addr = adev->wb.gpu_addr + (ring->rptr_offs * 4);
158 
159 		WREG32(DMA_RB_RPTR_ADDR_LO + sdma_offsets[i], lower_32_bits(rptr_addr));
160 		WREG32(DMA_RB_RPTR_ADDR_HI + sdma_offsets[i], upper_32_bits(rptr_addr) & 0xFF);
161 
162 		rb_cntl |= DMA_RPTR_WRITEBACK_ENABLE;
163 
164 		WREG32(DMA_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
165 
166 		/* enable DMA IBs */
167 		ib_cntl = DMA_IB_ENABLE | CMD_VMID_FORCE;
168 #ifdef __BIG_ENDIAN
169 		ib_cntl |= DMA_IB_SWAP_ENABLE;
170 #endif
171 		WREG32(DMA_IB_CNTL + sdma_offsets[i], ib_cntl);
172 
173 		dma_cntl = RREG32(DMA_CNTL + sdma_offsets[i]);
174 		dma_cntl &= ~CTXEMPTY_INT_ENABLE;
175 		WREG32(DMA_CNTL + sdma_offsets[i], dma_cntl);
176 
177 		ring->wptr = 0;
178 		WREG32(DMA_RB_WPTR + sdma_offsets[i], lower_32_bits(ring->wptr) << 2);
179 		WREG32(DMA_RB_CNTL + sdma_offsets[i], rb_cntl | DMA_RB_ENABLE);
180 
181 		ring->sched.ready = true;
182 
183 		r = amdgpu_ring_test_helper(ring);
184 		if (r)
185 			return r;
186 
187 		if (adev->mman.buffer_funcs_ring == ring)
188 			amdgpu_ttm_set_buffer_funcs_status(adev, true);
189 	}
190 
191 	return 0;
192 }
193 
194 /**
195  * si_dma_ring_test_ring - simple async dma engine test
196  *
197  * @ring: amdgpu_ring structure holding ring information
198  *
199  * Test the DMA engine by writing using it to write an
200  * value to memory. (VI).
201  * Returns 0 for success, error for failure.
202  */
203 static int si_dma_ring_test_ring(struct amdgpu_ring *ring)
204 {
205 	struct amdgpu_device *adev = ring->adev;
206 	unsigned i;
207 	unsigned index;
208 	int r;
209 	u32 tmp;
210 	u64 gpu_addr;
211 
212 	r = amdgpu_device_wb_get(adev, &index);
213 	if (r)
214 		return r;
215 
216 	gpu_addr = adev->wb.gpu_addr + (index * 4);
217 	tmp = 0xCAFEDEAD;
218 	adev->wb.wb[index] = cpu_to_le32(tmp);
219 
220 	r = amdgpu_ring_alloc(ring, 4);
221 	if (r)
222 		goto error_free_wb;
223 
224 	amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 0, 1));
225 	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
226 	amdgpu_ring_write(ring, upper_32_bits(gpu_addr) & 0xff);
227 	amdgpu_ring_write(ring, 0xDEADBEEF);
228 	amdgpu_ring_commit(ring);
229 
230 	for (i = 0; i < adev->usec_timeout; i++) {
231 		tmp = le32_to_cpu(adev->wb.wb[index]);
232 		if (tmp == 0xDEADBEEF)
233 			break;
234 		udelay(1);
235 	}
236 
237 	if (i >= adev->usec_timeout)
238 		r = -ETIMEDOUT;
239 
240 error_free_wb:
241 	amdgpu_device_wb_free(adev, index);
242 	return r;
243 }
244 
245 /**
246  * si_dma_ring_test_ib - test an IB on the DMA engine
247  *
248  * @ring: amdgpu_ring structure holding ring information
249  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
250  *
251  * Test a simple IB in the DMA ring (VI).
252  * Returns 0 on success, error on failure.
253  */
254 static int si_dma_ring_test_ib(struct amdgpu_ring *ring, long timeout)
255 {
256 	struct amdgpu_device *adev = ring->adev;
257 	struct amdgpu_ib ib;
258 	struct dma_fence *f = NULL;
259 	unsigned index;
260 	u32 tmp = 0;
261 	u64 gpu_addr;
262 	long r;
263 
264 	r = amdgpu_device_wb_get(adev, &index);
265 	if (r)
266 		return r;
267 
268 	gpu_addr = adev->wb.gpu_addr + (index * 4);
269 	tmp = 0xCAFEDEAD;
270 	adev->wb.wb[index] = cpu_to_le32(tmp);
271 	memset(&ib, 0, sizeof(ib));
272 	r = amdgpu_ib_get(adev, NULL, 256,
273 					AMDGPU_IB_POOL_DIRECT, &ib);
274 	if (r)
275 		goto err0;
276 
277 	ib.ptr[0] = DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 0, 1);
278 	ib.ptr[1] = lower_32_bits(gpu_addr);
279 	ib.ptr[2] = upper_32_bits(gpu_addr) & 0xff;
280 	ib.ptr[3] = 0xDEADBEEF;
281 	ib.length_dw = 4;
282 	r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
283 	if (r)
284 		goto err1;
285 
286 	r = dma_fence_wait_timeout(f, false, timeout);
287 	if (r == 0) {
288 		r = -ETIMEDOUT;
289 		goto err1;
290 	} else if (r < 0) {
291 		goto err1;
292 	}
293 	tmp = le32_to_cpu(adev->wb.wb[index]);
294 	if (tmp == 0xDEADBEEF)
295 		r = 0;
296 	else
297 		r = -EINVAL;
298 
299 err1:
300 	amdgpu_ib_free(adev, &ib, NULL);
301 	dma_fence_put(f);
302 err0:
303 	amdgpu_device_wb_free(adev, index);
304 	return r;
305 }
306 
307 /**
308  * si_dma_vm_copy_pte - update PTEs by copying them from the GART
309  *
310  * @ib: indirect buffer to fill with commands
311  * @pe: addr of the page entry
312  * @src: src addr to copy from
313  * @count: number of page entries to update
314  *
315  * Update PTEs by copying them from the GART using DMA (SI).
316  */
317 static void si_dma_vm_copy_pte(struct amdgpu_ib *ib,
318 			       uint64_t pe, uint64_t src,
319 			       unsigned count)
320 {
321 	unsigned bytes = count * 8;
322 
323 	ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_COPY,
324 					      1, 0, 0, bytes);
325 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
326 	ib->ptr[ib->length_dw++] = lower_32_bits(src);
327 	ib->ptr[ib->length_dw++] = upper_32_bits(pe) & 0xff;
328 	ib->ptr[ib->length_dw++] = upper_32_bits(src) & 0xff;
329 }
330 
331 /**
332  * si_dma_vm_write_pte - update PTEs by writing them manually
333  *
334  * @ib: indirect buffer to fill with commands
335  * @pe: addr of the page entry
336  * @value: dst addr to write into pe
337  * @count: number of page entries to update
338  * @incr: increase next addr by incr bytes
339  *
340  * Update PTEs by writing them manually using DMA (SI).
341  */
342 static void si_dma_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
343 				uint64_t value, unsigned count,
344 				uint32_t incr)
345 {
346 	unsigned ndw = count * 2;
347 
348 	ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 0, ndw);
349 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
350 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
351 	for (; ndw > 0; ndw -= 2) {
352 		ib->ptr[ib->length_dw++] = lower_32_bits(value);
353 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
354 		value += incr;
355 	}
356 }
357 
358 /**
359  * si_dma_vm_set_pte_pde - update the page tables using sDMA
360  *
361  * @ib: indirect buffer to fill with commands
362  * @pe: addr of the page entry
363  * @addr: dst addr to write into pe
364  * @count: number of page entries to update
365  * @incr: increase next addr by incr bytes
366  * @flags: access flags
367  *
368  * Update the page tables using sDMA (CIK).
369  */
370 static void si_dma_vm_set_pte_pde(struct amdgpu_ib *ib,
371 				     uint64_t pe,
372 				     uint64_t addr, unsigned count,
373 				     uint32_t incr, uint64_t flags)
374 {
375 	uint64_t value;
376 	unsigned ndw;
377 
378 	while (count) {
379 		ndw = count * 2;
380 		if (ndw > 0xFFFFE)
381 			ndw = 0xFFFFE;
382 
383 		if (flags & AMDGPU_PTE_VALID)
384 			value = addr;
385 		else
386 			value = 0;
387 
388 		/* for physically contiguous pages (vram) */
389 		ib->ptr[ib->length_dw++] = DMA_PTE_PDE_PACKET(ndw);
390 		ib->ptr[ib->length_dw++] = pe; /* dst addr */
391 		ib->ptr[ib->length_dw++] = upper_32_bits(pe) & 0xff;
392 		ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
393 		ib->ptr[ib->length_dw++] = upper_32_bits(flags);
394 		ib->ptr[ib->length_dw++] = value; /* value */
395 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
396 		ib->ptr[ib->length_dw++] = incr; /* increment size */
397 		ib->ptr[ib->length_dw++] = 0;
398 		pe += ndw * 4;
399 		addr += (ndw / 2) * incr;
400 		count -= ndw / 2;
401 	}
402 }
403 
404 /**
405  * si_dma_ring_pad_ib - pad the IB to the required number of dw
406  *
407  * @ring: amdgpu_ring pointer
408  * @ib: indirect buffer to fill with padding
409  *
410  */
411 static void si_dma_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
412 {
413 	while (ib->length_dw & 0x7)
414 		ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0);
415 }
416 
417 /**
418  * si_dma_ring_emit_pipeline_sync - sync the pipeline
419  *
420  * @ring: amdgpu_ring pointer
421  *
422  * Make sure all previous operations are completed (CIK).
423  */
424 static void si_dma_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
425 {
426 	uint32_t seq = ring->fence_drv.sync_seq;
427 	uint64_t addr = ring->fence_drv.gpu_addr;
428 
429 	/* wait for idle */
430 	amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_POLL_REG_MEM, 0, 0, 0, 0) |
431 			  (1 << 27)); /* Poll memory */
432 	amdgpu_ring_write(ring, lower_32_bits(addr));
433 	amdgpu_ring_write(ring, (0xff << 16) | upper_32_bits(addr)); /* retry, addr_hi */
434 	amdgpu_ring_write(ring, 0xffffffff); /* mask */
435 	amdgpu_ring_write(ring, seq); /* value */
436 	amdgpu_ring_write(ring, (3 << 28) | 0x20); /* func(equal) | poll interval */
437 }
438 
439 /**
440  * si_dma_ring_emit_vm_flush - cik vm flush using sDMA
441  *
442  * @ring: amdgpu_ring pointer
443  * @vmid: vmid number to use
444  * @pd_addr: address
445  *
446  * Update the page table base and flush the VM TLB
447  * using sDMA (VI).
448  */
449 static void si_dma_ring_emit_vm_flush(struct amdgpu_ring *ring,
450 				      unsigned vmid, uint64_t pd_addr)
451 {
452 	amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
453 
454 	/* wait for invalidate to complete */
455 	amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_POLL_REG_MEM, 0, 0, 0, 0));
456 	amdgpu_ring_write(ring, VM_INVALIDATE_REQUEST);
457 	amdgpu_ring_write(ring, 0xff << 16); /* retry */
458 	amdgpu_ring_write(ring, 1 << vmid); /* mask */
459 	amdgpu_ring_write(ring, 0); /* value */
460 	amdgpu_ring_write(ring, (0 << 28) | 0x20); /* func(always) | poll interval */
461 }
462 
463 static void si_dma_ring_emit_wreg(struct amdgpu_ring *ring,
464 				  uint32_t reg, uint32_t val)
465 {
466 	amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0, 0, 0));
467 	amdgpu_ring_write(ring, (0xf << 16) | reg);
468 	amdgpu_ring_write(ring, val);
469 }
470 
471 static int si_dma_early_init(void *handle)
472 {
473 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
474 
475 	adev->sdma.num_instances = 2;
476 
477 	si_dma_set_ring_funcs(adev);
478 	si_dma_set_buffer_funcs(adev);
479 	si_dma_set_vm_pte_funcs(adev);
480 	si_dma_set_irq_funcs(adev);
481 
482 	return 0;
483 }
484 
485 static int si_dma_sw_init(void *handle)
486 {
487 	struct amdgpu_ring *ring;
488 	int r, i;
489 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
490 
491 	/* DMA0 trap event */
492 	r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 224,
493 			      &adev->sdma.trap_irq);
494 	if (r)
495 		return r;
496 
497 	/* DMA1 trap event */
498 	r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 244,
499 			      &adev->sdma.trap_irq);
500 	if (r)
501 		return r;
502 
503 	for (i = 0; i < adev->sdma.num_instances; i++) {
504 		ring = &adev->sdma.instance[i].ring;
505 		ring->ring_obj = NULL;
506 		ring->use_doorbell = false;
507 		sprintf(ring->name, "sdma%d", i);
508 		r = amdgpu_ring_init(adev, ring, 1024,
509 				     &adev->sdma.trap_irq,
510 				     (i == 0) ? AMDGPU_SDMA_IRQ_INSTANCE0 :
511 				     AMDGPU_SDMA_IRQ_INSTANCE1,
512 				     AMDGPU_RING_PRIO_DEFAULT, NULL);
513 		if (r)
514 			return r;
515 	}
516 
517 	return r;
518 }
519 
520 static int si_dma_sw_fini(void *handle)
521 {
522 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
523 	int i;
524 
525 	for (i = 0; i < adev->sdma.num_instances; i++)
526 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
527 
528 	return 0;
529 }
530 
531 static int si_dma_hw_init(void *handle)
532 {
533 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
534 
535 	return si_dma_start(adev);
536 }
537 
538 static int si_dma_hw_fini(void *handle)
539 {
540 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
541 
542 	si_dma_stop(adev);
543 
544 	return 0;
545 }
546 
547 static int si_dma_suspend(void *handle)
548 {
549 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
550 
551 	return si_dma_hw_fini(adev);
552 }
553 
554 static int si_dma_resume(void *handle)
555 {
556 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
557 
558 	return si_dma_hw_init(adev);
559 }
560 
561 static bool si_dma_is_idle(void *handle)
562 {
563 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
564 	u32 tmp = RREG32(SRBM_STATUS2);
565 
566 	if (tmp & (DMA_BUSY_MASK | DMA1_BUSY_MASK))
567 	    return false;
568 
569 	return true;
570 }
571 
572 static int si_dma_wait_for_idle(void *handle)
573 {
574 	unsigned i;
575 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
576 
577 	for (i = 0; i < adev->usec_timeout; i++) {
578 		if (si_dma_is_idle(handle))
579 			return 0;
580 		udelay(1);
581 	}
582 	return -ETIMEDOUT;
583 }
584 
585 static int si_dma_soft_reset(void *handle)
586 {
587 	DRM_INFO("si_dma_soft_reset --- not implemented !!!!!!!\n");
588 	return 0;
589 }
590 
591 static int si_dma_set_trap_irq_state(struct amdgpu_device *adev,
592 					struct amdgpu_irq_src *src,
593 					unsigned type,
594 					enum amdgpu_interrupt_state state)
595 {
596 	u32 sdma_cntl;
597 
598 	switch (type) {
599 	case AMDGPU_SDMA_IRQ_INSTANCE0:
600 		switch (state) {
601 		case AMDGPU_IRQ_STATE_DISABLE:
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 		case AMDGPU_IRQ_STATE_ENABLE:
607 			sdma_cntl = RREG32(DMA_CNTL + DMA0_REGISTER_OFFSET);
608 			sdma_cntl |= TRAP_ENABLE;
609 			WREG32(DMA_CNTL + DMA0_REGISTER_OFFSET, sdma_cntl);
610 			break;
611 		default:
612 			break;
613 		}
614 		break;
615 	case AMDGPU_SDMA_IRQ_INSTANCE1:
616 		switch (state) {
617 		case AMDGPU_IRQ_STATE_DISABLE:
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 		case AMDGPU_IRQ_STATE_ENABLE:
623 			sdma_cntl = RREG32(DMA_CNTL + DMA1_REGISTER_OFFSET);
624 			sdma_cntl |= TRAP_ENABLE;
625 			WREG32(DMA_CNTL + DMA1_REGISTER_OFFSET, sdma_cntl);
626 			break;
627 		default:
628 			break;
629 		}
630 		break;
631 	default:
632 		break;
633 	}
634 	return 0;
635 }
636 
637 static int si_dma_process_trap_irq(struct amdgpu_device *adev,
638 				      struct amdgpu_irq_src *source,
639 				      struct amdgpu_iv_entry *entry)
640 {
641 	if (entry->src_id == 224)
642 		amdgpu_fence_process(&adev->sdma.instance[0].ring);
643 	else
644 		amdgpu_fence_process(&adev->sdma.instance[1].ring);
645 	return 0;
646 }
647 
648 static int si_dma_set_clockgating_state(void *handle,
649 					  enum amd_clockgating_state state)
650 {
651 	u32 orig, data, offset;
652 	int i;
653 	bool enable;
654 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
655 
656 	enable = (state == AMD_CG_STATE_GATE);
657 
658 	if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
659 		for (i = 0; i < adev->sdma.num_instances; i++) {
660 			if (i == 0)
661 				offset = DMA0_REGISTER_OFFSET;
662 			else
663 				offset = DMA1_REGISTER_OFFSET;
664 			orig = data = RREG32(DMA_POWER_CNTL + offset);
665 			data &= ~MEM_POWER_OVERRIDE;
666 			if (data != orig)
667 				WREG32(DMA_POWER_CNTL + offset, data);
668 			WREG32(DMA_CLK_CTRL + offset, 0x00000100);
669 		}
670 	} else {
671 		for (i = 0; i < adev->sdma.num_instances; i++) {
672 			if (i == 0)
673 				offset = DMA0_REGISTER_OFFSET;
674 			else
675 				offset = DMA1_REGISTER_OFFSET;
676 			orig = data = RREG32(DMA_POWER_CNTL + offset);
677 			data |= MEM_POWER_OVERRIDE;
678 			if (data != orig)
679 				WREG32(DMA_POWER_CNTL + offset, data);
680 
681 			orig = data = RREG32(DMA_CLK_CTRL + offset);
682 			data = 0xff000000;
683 			if (data != orig)
684 				WREG32(DMA_CLK_CTRL + offset, data);
685 		}
686 	}
687 
688 	return 0;
689 }
690 
691 static int si_dma_set_powergating_state(void *handle,
692 					  enum amd_powergating_state state)
693 {
694 	u32 tmp;
695 
696 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
697 
698 	WREG32(DMA_PGFSM_WRITE,  0x00002000);
699 	WREG32(DMA_PGFSM_CONFIG, 0x100010ff);
700 
701 	for (tmp = 0; tmp < 5; tmp++)
702 		WREG32(DMA_PGFSM_WRITE, 0);
703 
704 	return 0;
705 }
706 
707 static const struct amd_ip_funcs si_dma_ip_funcs = {
708 	.name = "si_dma",
709 	.early_init = si_dma_early_init,
710 	.late_init = NULL,
711 	.sw_init = si_dma_sw_init,
712 	.sw_fini = si_dma_sw_fini,
713 	.hw_init = si_dma_hw_init,
714 	.hw_fini = si_dma_hw_fini,
715 	.suspend = si_dma_suspend,
716 	.resume = si_dma_resume,
717 	.is_idle = si_dma_is_idle,
718 	.wait_for_idle = si_dma_wait_for_idle,
719 	.soft_reset = si_dma_soft_reset,
720 	.set_clockgating_state = si_dma_set_clockgating_state,
721 	.set_powergating_state = si_dma_set_powergating_state,
722 };
723 
724 static const struct amdgpu_ring_funcs si_dma_ring_funcs = {
725 	.type = AMDGPU_RING_TYPE_SDMA,
726 	.align_mask = 0xf,
727 	.nop = DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0),
728 	.support_64bit_ptrs = false,
729 	.get_rptr = si_dma_ring_get_rptr,
730 	.get_wptr = si_dma_ring_get_wptr,
731 	.set_wptr = si_dma_ring_set_wptr,
732 	.emit_frame_size =
733 		3 + 3 + /* hdp flush / invalidate */
734 		6 + /* si_dma_ring_emit_pipeline_sync */
735 		SI_FLUSH_GPU_TLB_NUM_WREG * 3 + 6 + /* si_dma_ring_emit_vm_flush */
736 		9 + 9 + 9, /* si_dma_ring_emit_fence x3 for user fence, vm fence */
737 	.emit_ib_size = 7 + 3, /* si_dma_ring_emit_ib */
738 	.emit_ib = si_dma_ring_emit_ib,
739 	.emit_fence = si_dma_ring_emit_fence,
740 	.emit_pipeline_sync = si_dma_ring_emit_pipeline_sync,
741 	.emit_vm_flush = si_dma_ring_emit_vm_flush,
742 	.test_ring = si_dma_ring_test_ring,
743 	.test_ib = si_dma_ring_test_ib,
744 	.insert_nop = amdgpu_ring_insert_nop,
745 	.pad_ib = si_dma_ring_pad_ib,
746 	.emit_wreg = si_dma_ring_emit_wreg,
747 };
748 
749 static void si_dma_set_ring_funcs(struct amdgpu_device *adev)
750 {
751 	int i;
752 
753 	for (i = 0; i < adev->sdma.num_instances; i++)
754 		adev->sdma.instance[i].ring.funcs = &si_dma_ring_funcs;
755 }
756 
757 static const struct amdgpu_irq_src_funcs si_dma_trap_irq_funcs = {
758 	.set = si_dma_set_trap_irq_state,
759 	.process = si_dma_process_trap_irq,
760 };
761 
762 static void si_dma_set_irq_funcs(struct amdgpu_device *adev)
763 {
764 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
765 	adev->sdma.trap_irq.funcs = &si_dma_trap_irq_funcs;
766 }
767 
768 /**
769  * si_dma_emit_copy_buffer - copy buffer using the sDMA engine
770  *
771  * @ib: indirect buffer to copy to
772  * @src_offset: src GPU address
773  * @dst_offset: dst GPU address
774  * @byte_count: number of bytes to xfer
775  * @tmz: is this a secure operation
776  *
777  * Copy GPU buffers using the DMA engine (VI).
778  * Used by the amdgpu ttm implementation to move pages if
779  * registered as the asic copy callback.
780  */
781 static void si_dma_emit_copy_buffer(struct amdgpu_ib *ib,
782 				       uint64_t src_offset,
783 				       uint64_t dst_offset,
784 				       uint32_t byte_count,
785 				       bool tmz)
786 {
787 	ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_COPY,
788 					      1, 0, 0, byte_count);
789 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
790 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
791 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset) & 0xff;
792 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset) & 0xff;
793 }
794 
795 /**
796  * si_dma_emit_fill_buffer - fill buffer using the sDMA engine
797  *
798  * @ib: indirect buffer to copy to
799  * @src_data: value to write to buffer
800  * @dst_offset: dst GPU address
801  * @byte_count: number of bytes to xfer
802  *
803  * Fill GPU buffers using the DMA engine (VI).
804  */
805 static void si_dma_emit_fill_buffer(struct amdgpu_ib *ib,
806 				       uint32_t src_data,
807 				       uint64_t dst_offset,
808 				       uint32_t byte_count)
809 {
810 	ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_CONSTANT_FILL,
811 					      0, 0, 0, byte_count / 4);
812 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
813 	ib->ptr[ib->length_dw++] = src_data;
814 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset) << 16;
815 }
816 
817 
818 static const struct amdgpu_buffer_funcs si_dma_buffer_funcs = {
819 	.copy_max_bytes = 0xffff8,
820 	.copy_num_dw = 5,
821 	.emit_copy_buffer = si_dma_emit_copy_buffer,
822 
823 	.fill_max_bytes = 0xffff8,
824 	.fill_num_dw = 4,
825 	.emit_fill_buffer = si_dma_emit_fill_buffer,
826 };
827 
828 static void si_dma_set_buffer_funcs(struct amdgpu_device *adev)
829 {
830 	adev->mman.buffer_funcs = &si_dma_buffer_funcs;
831 	adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
832 }
833 
834 static const struct amdgpu_vm_pte_funcs si_dma_vm_pte_funcs = {
835 	.copy_pte_num_dw = 5,
836 	.copy_pte = si_dma_vm_copy_pte,
837 
838 	.write_pte = si_dma_vm_write_pte,
839 	.set_pte_pde = si_dma_vm_set_pte_pde,
840 };
841 
842 static void si_dma_set_vm_pte_funcs(struct amdgpu_device *adev)
843 {
844 	unsigned i;
845 
846 	adev->vm_manager.vm_pte_funcs = &si_dma_vm_pte_funcs;
847 	for (i = 0; i < adev->sdma.num_instances; i++) {
848 		adev->vm_manager.vm_pte_scheds[i] =
849 			&adev->sdma.instance[i].ring.sched;
850 	}
851 	adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
852 }
853 
854 const struct amdgpu_ip_block_version si_dma_ip_block =
855 {
856 	.type = AMD_IP_BLOCK_TYPE_SDMA,
857 	.major = 1,
858 	.minor = 0,
859 	.rev = 0,
860 	.funcs = &si_dma_ip_funcs,
861 };
862