xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/si_dma.c (revision 6726d552)
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 	struct amdgpu_ring *ring;
116 	u32 rb_cntl;
117 	unsigned i;
118 
119 	for (i = 0; i < adev->sdma.num_instances; i++) {
120 		ring = &adev->sdma.instance[i].ring;
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 		if (adev->mman.buffer_funcs_ring == ring)
127 			amdgpu_ttm_set_buffer_funcs_status(adev, false);
128 	}
129 }
130 
131 static int si_dma_start(struct amdgpu_device *adev)
132 {
133 	struct amdgpu_ring *ring;
134 	u32 rb_cntl, dma_cntl, ib_cntl, rb_bufsz;
135 	int i, r;
136 	uint64_t rptr_addr;
137 
138 	for (i = 0; i < adev->sdma.num_instances; i++) {
139 		ring = &adev->sdma.instance[i].ring;
140 
141 		WREG32(DMA_SEM_INCOMPLETE_TIMER_CNTL + sdma_offsets[i], 0);
142 		WREG32(DMA_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
143 
144 		/* Set ring buffer size in dwords */
145 		rb_bufsz = order_base_2(ring->ring_size / 4);
146 		rb_cntl = rb_bufsz << 1;
147 #ifdef __BIG_ENDIAN
148 		rb_cntl |= DMA_RB_SWAP_ENABLE | DMA_RPTR_WRITEBACK_SWAP_ENABLE;
149 #endif
150 		WREG32(DMA_RB_CNTL + sdma_offsets[i], rb_cntl);
151 
152 		/* Initialize the ring buffer's read and write pointers */
153 		WREG32(DMA_RB_RPTR + sdma_offsets[i], 0);
154 		WREG32(DMA_RB_WPTR + sdma_offsets[i], 0);
155 
156 		rptr_addr = ring->rptr_gpu_addr;
157 
158 		WREG32(DMA_RB_RPTR_ADDR_LO + sdma_offsets[i], lower_32_bits(rptr_addr));
159 		WREG32(DMA_RB_RPTR_ADDR_HI + sdma_offsets[i], upper_32_bits(rptr_addr) & 0xFF);
160 
161 		rb_cntl |= DMA_RPTR_WRITEBACK_ENABLE;
162 
163 		WREG32(DMA_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
164 
165 		/* enable DMA IBs */
166 		ib_cntl = DMA_IB_ENABLE | CMD_VMID_FORCE;
167 #ifdef __BIG_ENDIAN
168 		ib_cntl |= DMA_IB_SWAP_ENABLE;
169 #endif
170 		WREG32(DMA_IB_CNTL + sdma_offsets[i], ib_cntl);
171 
172 		dma_cntl = RREG32(DMA_CNTL + sdma_offsets[i]);
173 		dma_cntl &= ~CTXEMPTY_INT_ENABLE;
174 		WREG32(DMA_CNTL + sdma_offsets[i], dma_cntl);
175 
176 		ring->wptr = 0;
177 		WREG32(DMA_RB_WPTR + sdma_offsets[i], ring->wptr << 2);
178 		WREG32(DMA_RB_CNTL + sdma_offsets[i], rb_cntl | DMA_RB_ENABLE);
179 
180 		ring->sched.ready = true;
181 
182 		r = amdgpu_ring_test_helper(ring);
183 		if (r)
184 			return r;
185 
186 		if (adev->mman.buffer_funcs_ring == ring)
187 			amdgpu_ttm_set_buffer_funcs_status(adev, true);
188 	}
189 
190 	return 0;
191 }
192 
193 /**
194  * si_dma_ring_test_ring - simple async dma engine test
195  *
196  * @ring: amdgpu_ring structure holding ring information
197  *
198  * Test the DMA engine by writing using it to write an
199  * value to memory. (VI).
200  * Returns 0 for success, error for failure.
201  */
202 static int si_dma_ring_test_ring(struct amdgpu_ring *ring)
203 {
204 	struct amdgpu_device *adev = ring->adev;
205 	unsigned i;
206 	unsigned index;
207 	int r;
208 	u32 tmp;
209 	u64 gpu_addr;
210 
211 	r = amdgpu_device_wb_get(adev, &index);
212 	if (r)
213 		return r;
214 
215 	gpu_addr = adev->wb.gpu_addr + (index * 4);
216 	tmp = 0xCAFEDEAD;
217 	adev->wb.wb[index] = cpu_to_le32(tmp);
218 
219 	r = amdgpu_ring_alloc(ring, 4);
220 	if (r)
221 		goto error_free_wb;
222 
223 	amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 0, 1));
224 	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
225 	amdgpu_ring_write(ring, upper_32_bits(gpu_addr) & 0xff);
226 	amdgpu_ring_write(ring, 0xDEADBEEF);
227 	amdgpu_ring_commit(ring);
228 
229 	for (i = 0; i < adev->usec_timeout; i++) {
230 		tmp = le32_to_cpu(adev->wb.wb[index]);
231 		if (tmp == 0xDEADBEEF)
232 			break;
233 		udelay(1);
234 	}
235 
236 	if (i >= adev->usec_timeout)
237 		r = -ETIMEDOUT;
238 
239 error_free_wb:
240 	amdgpu_device_wb_free(adev, index);
241 	return r;
242 }
243 
244 /**
245  * si_dma_ring_test_ib - test an IB on the DMA engine
246  *
247  * @ring: amdgpu_ring structure holding ring information
248  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
249  *
250  * Test a simple IB in the DMA ring (VI).
251  * Returns 0 on success, error on failure.
252  */
253 static int si_dma_ring_test_ib(struct amdgpu_ring *ring, long timeout)
254 {
255 	struct amdgpu_device *adev = ring->adev;
256 	struct amdgpu_ib ib;
257 	struct dma_fence *f = NULL;
258 	unsigned index;
259 	u32 tmp = 0;
260 	u64 gpu_addr;
261 	long r;
262 
263 	r = amdgpu_device_wb_get(adev, &index);
264 	if (r)
265 		return r;
266 
267 	gpu_addr = adev->wb.gpu_addr + (index * 4);
268 	tmp = 0xCAFEDEAD;
269 	adev->wb.wb[index] = cpu_to_le32(tmp);
270 	memset(&ib, 0, sizeof(ib));
271 	r = amdgpu_ib_get(adev, NULL, 256,
272 					AMDGPU_IB_POOL_DIRECT, &ib);
273 	if (r)
274 		goto err0;
275 
276 	ib.ptr[0] = DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 0, 1);
277 	ib.ptr[1] = lower_32_bits(gpu_addr);
278 	ib.ptr[2] = upper_32_bits(gpu_addr) & 0xff;
279 	ib.ptr[3] = 0xDEADBEEF;
280 	ib.length_dw = 4;
281 	r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
282 	if (r)
283 		goto err1;
284 
285 	r = dma_fence_wait_timeout(f, false, timeout);
286 	if (r == 0) {
287 		r = -ETIMEDOUT;
288 		goto err1;
289 	} else if (r < 0) {
290 		goto err1;
291 	}
292 	tmp = le32_to_cpu(adev->wb.wb[index]);
293 	if (tmp == 0xDEADBEEF)
294 		r = 0;
295 	else
296 		r = -EINVAL;
297 
298 err1:
299 	amdgpu_ib_free(adev, &ib, NULL);
300 	dma_fence_put(f);
301 err0:
302 	amdgpu_device_wb_free(adev, index);
303 	return r;
304 }
305 
306 /**
307  * si_dma_vm_copy_pte - update PTEs by copying them from the GART
308  *
309  * @ib: indirect buffer to fill with commands
310  * @pe: addr of the page entry
311  * @src: src addr to copy from
312  * @count: number of page entries to update
313  *
314  * Update PTEs by copying them from the GART using DMA (SI).
315  */
316 static void si_dma_vm_copy_pte(struct amdgpu_ib *ib,
317 			       uint64_t pe, uint64_t src,
318 			       unsigned count)
319 {
320 	unsigned bytes = count * 8;
321 
322 	ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_COPY,
323 					      1, 0, 0, bytes);
324 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
325 	ib->ptr[ib->length_dw++] = lower_32_bits(src);
326 	ib->ptr[ib->length_dw++] = upper_32_bits(pe) & 0xff;
327 	ib->ptr[ib->length_dw++] = upper_32_bits(src) & 0xff;
328 }
329 
330 /**
331  * si_dma_vm_write_pte - update PTEs by writing them manually
332  *
333  * @ib: indirect buffer to fill with commands
334  * @pe: addr of the page entry
335  * @value: dst addr to write into pe
336  * @count: number of page entries to update
337  * @incr: increase next addr by incr bytes
338  *
339  * Update PTEs by writing them manually using DMA (SI).
340  */
341 static void si_dma_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
342 				uint64_t value, unsigned count,
343 				uint32_t incr)
344 {
345 	unsigned ndw = count * 2;
346 
347 	ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 0, ndw);
348 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
349 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
350 	for (; ndw > 0; ndw -= 2) {
351 		ib->ptr[ib->length_dw++] = lower_32_bits(value);
352 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
353 		value += incr;
354 	}
355 }
356 
357 /**
358  * si_dma_vm_set_pte_pde - update the page tables using sDMA
359  *
360  * @ib: indirect buffer to fill with commands
361  * @pe: addr of the page entry
362  * @addr: dst addr to write into pe
363  * @count: number of page entries to update
364  * @incr: increase next addr by incr bytes
365  * @flags: access flags
366  *
367  * Update the page tables using sDMA (CIK).
368  */
369 static void si_dma_vm_set_pte_pde(struct amdgpu_ib *ib,
370 				     uint64_t pe,
371 				     uint64_t addr, unsigned count,
372 				     uint32_t incr, uint64_t flags)
373 {
374 	uint64_t value;
375 	unsigned ndw;
376 
377 	while (count) {
378 		ndw = count * 2;
379 		if (ndw > 0xFFFFE)
380 			ndw = 0xFFFFE;
381 
382 		if (flags & AMDGPU_PTE_VALID)
383 			value = addr;
384 		else
385 			value = 0;
386 
387 		/* for physically contiguous pages (vram) */
388 		ib->ptr[ib->length_dw++] = DMA_PTE_PDE_PACKET(ndw);
389 		ib->ptr[ib->length_dw++] = pe; /* dst addr */
390 		ib->ptr[ib->length_dw++] = upper_32_bits(pe) & 0xff;
391 		ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
392 		ib->ptr[ib->length_dw++] = upper_32_bits(flags);
393 		ib->ptr[ib->length_dw++] = value; /* value */
394 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
395 		ib->ptr[ib->length_dw++] = incr; /* increment size */
396 		ib->ptr[ib->length_dw++] = 0;
397 		pe += ndw * 4;
398 		addr += (ndw / 2) * incr;
399 		count -= ndw / 2;
400 	}
401 }
402 
403 /**
404  * si_dma_ring_pad_ib - pad the IB to the required number of dw
405  *
406  * @ring: amdgpu_ring pointer
407  * @ib: indirect buffer to fill with padding
408  *
409  */
410 static void si_dma_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
411 {
412 	while (ib->length_dw & 0x7)
413 		ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0);
414 }
415 
416 /**
417  * si_dma_ring_emit_pipeline_sync - sync the pipeline
418  *
419  * @ring: amdgpu_ring pointer
420  *
421  * Make sure all previous operations are completed (CIK).
422  */
423 static void si_dma_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
424 {
425 	uint32_t seq = ring->fence_drv.sync_seq;
426 	uint64_t addr = ring->fence_drv.gpu_addr;
427 
428 	/* wait for idle */
429 	amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_POLL_REG_MEM, 0, 0, 0, 0) |
430 			  (1 << 27)); /* Poll memory */
431 	amdgpu_ring_write(ring, lower_32_bits(addr));
432 	amdgpu_ring_write(ring, (0xff << 16) | upper_32_bits(addr)); /* retry, addr_hi */
433 	amdgpu_ring_write(ring, 0xffffffff); /* mask */
434 	amdgpu_ring_write(ring, seq); /* value */
435 	amdgpu_ring_write(ring, (3 << 28) | 0x20); /* func(equal) | poll interval */
436 }
437 
438 /**
439  * si_dma_ring_emit_vm_flush - cik vm flush using sDMA
440  *
441  * @ring: amdgpu_ring pointer
442  * @vmid: vmid number to use
443  * @pd_addr: address
444  *
445  * Update the page table base and flush the VM TLB
446  * using sDMA (VI).
447  */
448 static void si_dma_ring_emit_vm_flush(struct amdgpu_ring *ring,
449 				      unsigned vmid, uint64_t pd_addr)
450 {
451 	amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
452 
453 	/* wait for invalidate to complete */
454 	amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_POLL_REG_MEM, 0, 0, 0, 0));
455 	amdgpu_ring_write(ring, VM_INVALIDATE_REQUEST);
456 	amdgpu_ring_write(ring, 0xff << 16); /* retry */
457 	amdgpu_ring_write(ring, 1 << vmid); /* mask */
458 	amdgpu_ring_write(ring, 0); /* value */
459 	amdgpu_ring_write(ring, (0 << 28) | 0x20); /* func(always) | poll interval */
460 }
461 
462 static void si_dma_ring_emit_wreg(struct amdgpu_ring *ring,
463 				  uint32_t reg, uint32_t val)
464 {
465 	amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0, 0, 0));
466 	amdgpu_ring_write(ring, (0xf << 16) | reg);
467 	amdgpu_ring_write(ring, val);
468 }
469 
470 static int si_dma_early_init(void *handle)
471 {
472 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
473 
474 	adev->sdma.num_instances = 2;
475 
476 	si_dma_set_ring_funcs(adev);
477 	si_dma_set_buffer_funcs(adev);
478 	si_dma_set_vm_pte_funcs(adev);
479 	si_dma_set_irq_funcs(adev);
480 
481 	return 0;
482 }
483 
484 static int si_dma_sw_init(void *handle)
485 {
486 	struct amdgpu_ring *ring;
487 	int r, i;
488 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
489 
490 	/* DMA0 trap event */
491 	r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 224,
492 			      &adev->sdma.trap_irq);
493 	if (r)
494 		return r;
495 
496 	/* DMA1 trap event */
497 	r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 244,
498 			      &adev->sdma.trap_irq);
499 	if (r)
500 		return r;
501 
502 	for (i = 0; i < adev->sdma.num_instances; i++) {
503 		ring = &adev->sdma.instance[i].ring;
504 		ring->ring_obj = NULL;
505 		ring->use_doorbell = false;
506 		sprintf(ring->name, "sdma%d", i);
507 		r = amdgpu_ring_init(adev, ring, 1024,
508 				     &adev->sdma.trap_irq,
509 				     (i == 0) ? AMDGPU_SDMA_IRQ_INSTANCE0 :
510 				     AMDGPU_SDMA_IRQ_INSTANCE1,
511 				     AMDGPU_RING_PRIO_DEFAULT, NULL);
512 		if (r)
513 			return r;
514 	}
515 
516 	return r;
517 }
518 
519 static int si_dma_sw_fini(void *handle)
520 {
521 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
522 	int i;
523 
524 	for (i = 0; i < adev->sdma.num_instances; i++)
525 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
526 
527 	return 0;
528 }
529 
530 static int si_dma_hw_init(void *handle)
531 {
532 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
533 
534 	return si_dma_start(adev);
535 }
536 
537 static int si_dma_hw_fini(void *handle)
538 {
539 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
540 
541 	si_dma_stop(adev);
542 
543 	return 0;
544 }
545 
546 static int si_dma_suspend(void *handle)
547 {
548 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
549 
550 	return si_dma_hw_fini(adev);
551 }
552 
553 static int si_dma_resume(void *handle)
554 {
555 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
556 
557 	return si_dma_hw_init(adev);
558 }
559 
560 static bool si_dma_is_idle(void *handle)
561 {
562 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
563 	u32 tmp = RREG32(SRBM_STATUS2);
564 
565 	if (tmp & (DMA_BUSY_MASK | DMA1_BUSY_MASK))
566 	    return false;
567 
568 	return true;
569 }
570 
571 static int si_dma_wait_for_idle(void *handle)
572 {
573 	unsigned i;
574 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
575 
576 	for (i = 0; i < adev->usec_timeout; i++) {
577 		if (si_dma_is_idle(handle))
578 			return 0;
579 		udelay(1);
580 	}
581 	return -ETIMEDOUT;
582 }
583 
584 static int si_dma_soft_reset(void *handle)
585 {
586 	DRM_INFO("si_dma_soft_reset --- not implemented !!!!!!!\n");
587 	return 0;
588 }
589 
590 static int si_dma_set_trap_irq_state(struct amdgpu_device *adev,
591 					struct amdgpu_irq_src *src,
592 					unsigned type,
593 					enum amdgpu_interrupt_state state)
594 {
595 	u32 sdma_cntl;
596 
597 	switch (type) {
598 	case AMDGPU_SDMA_IRQ_INSTANCE0:
599 		switch (state) {
600 		case AMDGPU_IRQ_STATE_DISABLE:
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 		case AMDGPU_IRQ_STATE_ENABLE:
606 			sdma_cntl = RREG32(DMA_CNTL + DMA0_REGISTER_OFFSET);
607 			sdma_cntl |= TRAP_ENABLE;
608 			WREG32(DMA_CNTL + DMA0_REGISTER_OFFSET, sdma_cntl);
609 			break;
610 		default:
611 			break;
612 		}
613 		break;
614 	case AMDGPU_SDMA_IRQ_INSTANCE1:
615 		switch (state) {
616 		case AMDGPU_IRQ_STATE_DISABLE:
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 		case AMDGPU_IRQ_STATE_ENABLE:
622 			sdma_cntl = RREG32(DMA_CNTL + DMA1_REGISTER_OFFSET);
623 			sdma_cntl |= TRAP_ENABLE;
624 			WREG32(DMA_CNTL + DMA1_REGISTER_OFFSET, sdma_cntl);
625 			break;
626 		default:
627 			break;
628 		}
629 		break;
630 	default:
631 		break;
632 	}
633 	return 0;
634 }
635 
636 static int si_dma_process_trap_irq(struct amdgpu_device *adev,
637 				      struct amdgpu_irq_src *source,
638 				      struct amdgpu_iv_entry *entry)
639 {
640 	if (entry->src_id == 224)
641 		amdgpu_fence_process(&adev->sdma.instance[0].ring);
642 	else
643 		amdgpu_fence_process(&adev->sdma.instance[1].ring);
644 	return 0;
645 }
646 
647 static int si_dma_set_clockgating_state(void *handle,
648 					  enum amd_clockgating_state state)
649 {
650 	u32 orig, data, offset;
651 	int i;
652 	bool enable;
653 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
654 
655 	enable = (state == AMD_CG_STATE_GATE);
656 
657 	if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
658 		for (i = 0; i < adev->sdma.num_instances; i++) {
659 			if (i == 0)
660 				offset = DMA0_REGISTER_OFFSET;
661 			else
662 				offset = DMA1_REGISTER_OFFSET;
663 			orig = data = RREG32(DMA_POWER_CNTL + offset);
664 			data &= ~MEM_POWER_OVERRIDE;
665 			if (data != orig)
666 				WREG32(DMA_POWER_CNTL + offset, data);
667 			WREG32(DMA_CLK_CTRL + offset, 0x00000100);
668 		}
669 	} else {
670 		for (i = 0; i < adev->sdma.num_instances; i++) {
671 			if (i == 0)
672 				offset = DMA0_REGISTER_OFFSET;
673 			else
674 				offset = DMA1_REGISTER_OFFSET;
675 			orig = data = RREG32(DMA_POWER_CNTL + offset);
676 			data |= MEM_POWER_OVERRIDE;
677 			if (data != orig)
678 				WREG32(DMA_POWER_CNTL + offset, data);
679 
680 			orig = data = RREG32(DMA_CLK_CTRL + offset);
681 			data = 0xff000000;
682 			if (data != orig)
683 				WREG32(DMA_CLK_CTRL + offset, data);
684 		}
685 	}
686 
687 	return 0;
688 }
689 
690 static int si_dma_set_powergating_state(void *handle,
691 					  enum amd_powergating_state state)
692 {
693 	u32 tmp;
694 
695 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
696 
697 	WREG32(DMA_PGFSM_WRITE,  0x00002000);
698 	WREG32(DMA_PGFSM_CONFIG, 0x100010ff);
699 
700 	for (tmp = 0; tmp < 5; tmp++)
701 		WREG32(DMA_PGFSM_WRITE, 0);
702 
703 	return 0;
704 }
705 
706 static const struct amd_ip_funcs si_dma_ip_funcs = {
707 	.name = "si_dma",
708 	.early_init = si_dma_early_init,
709 	.late_init = NULL,
710 	.sw_init = si_dma_sw_init,
711 	.sw_fini = si_dma_sw_fini,
712 	.hw_init = si_dma_hw_init,
713 	.hw_fini = si_dma_hw_fini,
714 	.suspend = si_dma_suspend,
715 	.resume = si_dma_resume,
716 	.is_idle = si_dma_is_idle,
717 	.wait_for_idle = si_dma_wait_for_idle,
718 	.soft_reset = si_dma_soft_reset,
719 	.set_clockgating_state = si_dma_set_clockgating_state,
720 	.set_powergating_state = si_dma_set_powergating_state,
721 };
722 
723 static const struct amdgpu_ring_funcs si_dma_ring_funcs = {
724 	.type = AMDGPU_RING_TYPE_SDMA,
725 	.align_mask = 0xf,
726 	.nop = DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0),
727 	.support_64bit_ptrs = false,
728 	.get_rptr = si_dma_ring_get_rptr,
729 	.get_wptr = si_dma_ring_get_wptr,
730 	.set_wptr = si_dma_ring_set_wptr,
731 	.emit_frame_size =
732 		3 + 3 + /* hdp flush / invalidate */
733 		6 + /* si_dma_ring_emit_pipeline_sync */
734 		SI_FLUSH_GPU_TLB_NUM_WREG * 3 + 6 + /* si_dma_ring_emit_vm_flush */
735 		9 + 9 + 9, /* si_dma_ring_emit_fence x3 for user fence, vm fence */
736 	.emit_ib_size = 7 + 3, /* si_dma_ring_emit_ib */
737 	.emit_ib = si_dma_ring_emit_ib,
738 	.emit_fence = si_dma_ring_emit_fence,
739 	.emit_pipeline_sync = si_dma_ring_emit_pipeline_sync,
740 	.emit_vm_flush = si_dma_ring_emit_vm_flush,
741 	.test_ring = si_dma_ring_test_ring,
742 	.test_ib = si_dma_ring_test_ib,
743 	.insert_nop = amdgpu_ring_insert_nop,
744 	.pad_ib = si_dma_ring_pad_ib,
745 	.emit_wreg = si_dma_ring_emit_wreg,
746 };
747 
748 static void si_dma_set_ring_funcs(struct amdgpu_device *adev)
749 {
750 	int i;
751 
752 	for (i = 0; i < adev->sdma.num_instances; i++)
753 		adev->sdma.instance[i].ring.funcs = &si_dma_ring_funcs;
754 }
755 
756 static const struct amdgpu_irq_src_funcs si_dma_trap_irq_funcs = {
757 	.set = si_dma_set_trap_irq_state,
758 	.process = si_dma_process_trap_irq,
759 };
760 
761 static void si_dma_set_irq_funcs(struct amdgpu_device *adev)
762 {
763 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
764 	adev->sdma.trap_irq.funcs = &si_dma_trap_irq_funcs;
765 }
766 
767 /**
768  * si_dma_emit_copy_buffer - copy buffer using the sDMA engine
769  *
770  * @ib: indirect buffer to copy to
771  * @src_offset: src GPU address
772  * @dst_offset: dst GPU address
773  * @byte_count: number of bytes to xfer
774  * @tmz: is this a secure operation
775  *
776  * Copy GPU buffers using the DMA engine (VI).
777  * Used by the amdgpu ttm implementation to move pages if
778  * registered as the asic copy callback.
779  */
780 static void si_dma_emit_copy_buffer(struct amdgpu_ib *ib,
781 				       uint64_t src_offset,
782 				       uint64_t dst_offset,
783 				       uint32_t byte_count,
784 				       bool tmz)
785 {
786 	ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_COPY,
787 					      1, 0, 0, byte_count);
788 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
789 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
790 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset) & 0xff;
791 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset) & 0xff;
792 }
793 
794 /**
795  * si_dma_emit_fill_buffer - fill buffer using the sDMA engine
796  *
797  * @ib: indirect buffer to copy to
798  * @src_data: value to write to buffer
799  * @dst_offset: dst GPU address
800  * @byte_count: number of bytes to xfer
801  *
802  * Fill GPU buffers using the DMA engine (VI).
803  */
804 static void si_dma_emit_fill_buffer(struct amdgpu_ib *ib,
805 				       uint32_t src_data,
806 				       uint64_t dst_offset,
807 				       uint32_t byte_count)
808 {
809 	ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_CONSTANT_FILL,
810 					      0, 0, 0, byte_count / 4);
811 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
812 	ib->ptr[ib->length_dw++] = src_data;
813 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset) << 16;
814 }
815 
816 
817 static const struct amdgpu_buffer_funcs si_dma_buffer_funcs = {
818 	.copy_max_bytes = 0xffff8,
819 	.copy_num_dw = 5,
820 	.emit_copy_buffer = si_dma_emit_copy_buffer,
821 
822 	.fill_max_bytes = 0xffff8,
823 	.fill_num_dw = 4,
824 	.emit_fill_buffer = si_dma_emit_fill_buffer,
825 };
826 
827 static void si_dma_set_buffer_funcs(struct amdgpu_device *adev)
828 {
829 	adev->mman.buffer_funcs = &si_dma_buffer_funcs;
830 	adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
831 }
832 
833 static const struct amdgpu_vm_pte_funcs si_dma_vm_pte_funcs = {
834 	.copy_pte_num_dw = 5,
835 	.copy_pte = si_dma_vm_copy_pte,
836 
837 	.write_pte = si_dma_vm_write_pte,
838 	.set_pte_pde = si_dma_vm_set_pte_pde,
839 };
840 
841 static void si_dma_set_vm_pte_funcs(struct amdgpu_device *adev)
842 {
843 	unsigned i;
844 
845 	adev->vm_manager.vm_pte_funcs = &si_dma_vm_pte_funcs;
846 	for (i = 0; i < adev->sdma.num_instances; i++) {
847 		adev->vm_manager.vm_pte_scheds[i] =
848 			&adev->sdma.instance[i].ring.sched;
849 	}
850 	adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
851 }
852 
853 const struct amdgpu_ip_block_version si_dma_ip_block =
854 {
855 	.type = AMD_IP_BLOCK_TYPE_SDMA,
856 	.major = 1,
857 	.minor = 0,
858 	.rev = 0,
859 	.funcs = &si_dma_ip_funcs,
860 };
861