Loading...
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
30const u32 sdma_offsets[SDMA_MAX_INSTANCE] =
31{
32 DMA0_REGISTER_OFFSET,
33 DMA1_REGISTER_OFFSET
34};
35
36static void si_dma_set_ring_funcs(struct amdgpu_device *adev);
37static void si_dma_set_buffer_funcs(struct amdgpu_device *adev);
38static void si_dma_set_vm_pte_funcs(struct amdgpu_device *adev);
39static void si_dma_set_irq_funcs(struct amdgpu_device *adev);
40
41static uint64_t si_dma_ring_get_rptr(struct amdgpu_ring *ring)
42{
43 return *ring->rptr_cpu_addr;
44}
45
46static 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
54static 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
62static 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 */
91static 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
113static void si_dma_stop(struct amdgpu_device *adev)
114{
115 u32 rb_cntl;
116 unsigned i;
117
118 for (i = 0; i < adev->sdma.num_instances; i++) {
119 /* dma0 */
120 rb_cntl = RREG32(DMA_RB_CNTL + sdma_offsets[i]);
121 rb_cntl &= ~DMA_RB_ENABLE;
122 WREG32(DMA_RB_CNTL + sdma_offsets[i], rb_cntl);
123 }
124}
125
126static int si_dma_start(struct amdgpu_device *adev)
127{
128 struct amdgpu_ring *ring;
129 u32 rb_cntl, dma_cntl, ib_cntl, rb_bufsz;
130 int i, r;
131 uint64_t rptr_addr;
132
133 for (i = 0; i < adev->sdma.num_instances; i++) {
134 ring = &adev->sdma.instance[i].ring;
135
136 WREG32(DMA_SEM_INCOMPLETE_TIMER_CNTL + sdma_offsets[i], 0);
137 WREG32(DMA_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
138
139 /* Set ring buffer size in dwords */
140 rb_bufsz = order_base_2(ring->ring_size / 4);
141 rb_cntl = rb_bufsz << 1;
142#ifdef __BIG_ENDIAN
143 rb_cntl |= DMA_RB_SWAP_ENABLE | DMA_RPTR_WRITEBACK_SWAP_ENABLE;
144#endif
145 WREG32(DMA_RB_CNTL + sdma_offsets[i], rb_cntl);
146
147 /* Initialize the ring buffer's read and write pointers */
148 WREG32(DMA_RB_RPTR + sdma_offsets[i], 0);
149 WREG32(DMA_RB_WPTR + sdma_offsets[i], 0);
150
151 rptr_addr = ring->rptr_gpu_addr;
152
153 WREG32(DMA_RB_RPTR_ADDR_LO + sdma_offsets[i], lower_32_bits(rptr_addr));
154 WREG32(DMA_RB_RPTR_ADDR_HI + sdma_offsets[i], upper_32_bits(rptr_addr) & 0xFF);
155
156 rb_cntl |= DMA_RPTR_WRITEBACK_ENABLE;
157
158 WREG32(DMA_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
159
160 /* enable DMA IBs */
161 ib_cntl = DMA_IB_ENABLE | CMD_VMID_FORCE;
162#ifdef __BIG_ENDIAN
163 ib_cntl |= DMA_IB_SWAP_ENABLE;
164#endif
165 WREG32(DMA_IB_CNTL + sdma_offsets[i], ib_cntl);
166
167 dma_cntl = RREG32(DMA_CNTL + sdma_offsets[i]);
168 dma_cntl &= ~CTXEMPTY_INT_ENABLE;
169 WREG32(DMA_CNTL + sdma_offsets[i], dma_cntl);
170
171 ring->wptr = 0;
172 WREG32(DMA_RB_WPTR + sdma_offsets[i], ring->wptr << 2);
173 WREG32(DMA_RB_CNTL + sdma_offsets[i], rb_cntl | DMA_RB_ENABLE);
174
175 r = amdgpu_ring_test_helper(ring);
176 if (r)
177 return r;
178 }
179
180 return 0;
181}
182
183/**
184 * si_dma_ring_test_ring - simple async dma engine test
185 *
186 * @ring: amdgpu_ring structure holding ring information
187 *
188 * Test the DMA engine by writing using it to write an
189 * value to memory. (VI).
190 * Returns 0 for success, error for failure.
191 */
192static int si_dma_ring_test_ring(struct amdgpu_ring *ring)
193{
194 struct amdgpu_device *adev = ring->adev;
195 unsigned i;
196 unsigned index;
197 int r;
198 u32 tmp;
199 u64 gpu_addr;
200
201 r = amdgpu_device_wb_get(adev, &index);
202 if (r)
203 return r;
204
205 gpu_addr = adev->wb.gpu_addr + (index * 4);
206 tmp = 0xCAFEDEAD;
207 adev->wb.wb[index] = cpu_to_le32(tmp);
208
209 r = amdgpu_ring_alloc(ring, 4);
210 if (r)
211 goto error_free_wb;
212
213 amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 0, 1));
214 amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
215 amdgpu_ring_write(ring, upper_32_bits(gpu_addr) & 0xff);
216 amdgpu_ring_write(ring, 0xDEADBEEF);
217 amdgpu_ring_commit(ring);
218
219 for (i = 0; i < adev->usec_timeout; i++) {
220 tmp = le32_to_cpu(adev->wb.wb[index]);
221 if (tmp == 0xDEADBEEF)
222 break;
223 udelay(1);
224 }
225
226 if (i >= adev->usec_timeout)
227 r = -ETIMEDOUT;
228
229error_free_wb:
230 amdgpu_device_wb_free(adev, index);
231 return r;
232}
233
234/**
235 * si_dma_ring_test_ib - test an IB on the DMA engine
236 *
237 * @ring: amdgpu_ring structure holding ring information
238 * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
239 *
240 * Test a simple IB in the DMA ring (VI).
241 * Returns 0 on success, error on failure.
242 */
243static int si_dma_ring_test_ib(struct amdgpu_ring *ring, long timeout)
244{
245 struct amdgpu_device *adev = ring->adev;
246 struct amdgpu_ib ib;
247 struct dma_fence *f = NULL;
248 unsigned index;
249 u32 tmp = 0;
250 u64 gpu_addr;
251 long r;
252
253 r = amdgpu_device_wb_get(adev, &index);
254 if (r)
255 return r;
256
257 gpu_addr = adev->wb.gpu_addr + (index * 4);
258 tmp = 0xCAFEDEAD;
259 adev->wb.wb[index] = cpu_to_le32(tmp);
260 memset(&ib, 0, sizeof(ib));
261 r = amdgpu_ib_get(adev, NULL, 256,
262 AMDGPU_IB_POOL_DIRECT, &ib);
263 if (r)
264 goto err0;
265
266 ib.ptr[0] = DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 0, 1);
267 ib.ptr[1] = lower_32_bits(gpu_addr);
268 ib.ptr[2] = upper_32_bits(gpu_addr) & 0xff;
269 ib.ptr[3] = 0xDEADBEEF;
270 ib.length_dw = 4;
271 r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
272 if (r)
273 goto err1;
274
275 r = dma_fence_wait_timeout(f, false, timeout);
276 if (r == 0) {
277 r = -ETIMEDOUT;
278 goto err1;
279 } else if (r < 0) {
280 goto err1;
281 }
282 tmp = le32_to_cpu(adev->wb.wb[index]);
283 if (tmp == 0xDEADBEEF)
284 r = 0;
285 else
286 r = -EINVAL;
287
288err1:
289 amdgpu_ib_free(adev, &ib, NULL);
290 dma_fence_put(f);
291err0:
292 amdgpu_device_wb_free(adev, index);
293 return r;
294}
295
296/**
297 * si_dma_vm_copy_pte - update PTEs by copying them from the GART
298 *
299 * @ib: indirect buffer to fill with commands
300 * @pe: addr of the page entry
301 * @src: src addr to copy from
302 * @count: number of page entries to update
303 *
304 * Update PTEs by copying them from the GART using DMA (SI).
305 */
306static void si_dma_vm_copy_pte(struct amdgpu_ib *ib,
307 uint64_t pe, uint64_t src,
308 unsigned count)
309{
310 unsigned bytes = count * 8;
311
312 ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_COPY,
313 1, 0, 0, bytes);
314 ib->ptr[ib->length_dw++] = lower_32_bits(pe);
315 ib->ptr[ib->length_dw++] = lower_32_bits(src);
316 ib->ptr[ib->length_dw++] = upper_32_bits(pe) & 0xff;
317 ib->ptr[ib->length_dw++] = upper_32_bits(src) & 0xff;
318}
319
320/**
321 * si_dma_vm_write_pte - update PTEs by writing them manually
322 *
323 * @ib: indirect buffer to fill with commands
324 * @pe: addr of the page entry
325 * @value: dst addr to write into pe
326 * @count: number of page entries to update
327 * @incr: increase next addr by incr bytes
328 *
329 * Update PTEs by writing them manually using DMA (SI).
330 */
331static void si_dma_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
332 uint64_t value, unsigned count,
333 uint32_t incr)
334{
335 unsigned ndw = count * 2;
336
337 ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 0, ndw);
338 ib->ptr[ib->length_dw++] = lower_32_bits(pe);
339 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
340 for (; ndw > 0; ndw -= 2) {
341 ib->ptr[ib->length_dw++] = lower_32_bits(value);
342 ib->ptr[ib->length_dw++] = upper_32_bits(value);
343 value += incr;
344 }
345}
346
347/**
348 * si_dma_vm_set_pte_pde - update the page tables using sDMA
349 *
350 * @ib: indirect buffer to fill with commands
351 * @pe: addr of the page entry
352 * @addr: dst addr to write into pe
353 * @count: number of page entries to update
354 * @incr: increase next addr by incr bytes
355 * @flags: access flags
356 *
357 * Update the page tables using sDMA (CIK).
358 */
359static void si_dma_vm_set_pte_pde(struct amdgpu_ib *ib,
360 uint64_t pe,
361 uint64_t addr, unsigned count,
362 uint32_t incr, uint64_t flags)
363{
364 uint64_t value;
365 unsigned ndw;
366
367 while (count) {
368 ndw = count * 2;
369 if (ndw > 0xFFFFE)
370 ndw = 0xFFFFE;
371
372 if (flags & AMDGPU_PTE_VALID)
373 value = addr;
374 else
375 value = 0;
376
377 /* for physically contiguous pages (vram) */
378 ib->ptr[ib->length_dw++] = DMA_PTE_PDE_PACKET(ndw);
379 ib->ptr[ib->length_dw++] = pe; /* dst addr */
380 ib->ptr[ib->length_dw++] = upper_32_bits(pe) & 0xff;
381 ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
382 ib->ptr[ib->length_dw++] = upper_32_bits(flags);
383 ib->ptr[ib->length_dw++] = value; /* value */
384 ib->ptr[ib->length_dw++] = upper_32_bits(value);
385 ib->ptr[ib->length_dw++] = incr; /* increment size */
386 ib->ptr[ib->length_dw++] = 0;
387 pe += ndw * 4;
388 addr += (ndw / 2) * incr;
389 count -= ndw / 2;
390 }
391}
392
393/**
394 * si_dma_ring_pad_ib - pad the IB to the required number of dw
395 *
396 * @ring: amdgpu_ring pointer
397 * @ib: indirect buffer to fill with padding
398 *
399 */
400static void si_dma_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
401{
402 while (ib->length_dw & 0x7)
403 ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0);
404}
405
406/**
407 * si_dma_ring_emit_pipeline_sync - sync the pipeline
408 *
409 * @ring: amdgpu_ring pointer
410 *
411 * Make sure all previous operations are completed (CIK).
412 */
413static void si_dma_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
414{
415 uint32_t seq = ring->fence_drv.sync_seq;
416 uint64_t addr = ring->fence_drv.gpu_addr;
417
418 /* wait for idle */
419 amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_POLL_REG_MEM, 0, 0, 0, 0) |
420 (1 << 27)); /* Poll memory */
421 amdgpu_ring_write(ring, lower_32_bits(addr));
422 amdgpu_ring_write(ring, (0xff << 16) | upper_32_bits(addr)); /* retry, addr_hi */
423 amdgpu_ring_write(ring, 0xffffffff); /* mask */
424 amdgpu_ring_write(ring, seq); /* value */
425 amdgpu_ring_write(ring, (3 << 28) | 0x20); /* func(equal) | poll interval */
426}
427
428/**
429 * si_dma_ring_emit_vm_flush - cik vm flush using sDMA
430 *
431 * @ring: amdgpu_ring pointer
432 * @vmid: vmid number to use
433 * @pd_addr: address
434 *
435 * Update the page table base and flush the VM TLB
436 * using sDMA (VI).
437 */
438static void si_dma_ring_emit_vm_flush(struct amdgpu_ring *ring,
439 unsigned vmid, uint64_t pd_addr)
440{
441 amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
442
443 /* wait for invalidate to complete */
444 amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_POLL_REG_MEM, 0, 0, 0, 0));
445 amdgpu_ring_write(ring, VM_INVALIDATE_REQUEST);
446 amdgpu_ring_write(ring, 0xff << 16); /* retry */
447 amdgpu_ring_write(ring, 1 << vmid); /* mask */
448 amdgpu_ring_write(ring, 0); /* value */
449 amdgpu_ring_write(ring, (0 << 28) | 0x20); /* func(always) | poll interval */
450}
451
452static void si_dma_ring_emit_wreg(struct amdgpu_ring *ring,
453 uint32_t reg, uint32_t val)
454{
455 amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0, 0, 0));
456 amdgpu_ring_write(ring, (0xf << 16) | reg);
457 amdgpu_ring_write(ring, val);
458}
459
460static int si_dma_early_init(void *handle)
461{
462 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
463
464 adev->sdma.num_instances = 2;
465
466 si_dma_set_ring_funcs(adev);
467 si_dma_set_buffer_funcs(adev);
468 si_dma_set_vm_pte_funcs(adev);
469 si_dma_set_irq_funcs(adev);
470
471 return 0;
472}
473
474static int si_dma_sw_init(void *handle)
475{
476 struct amdgpu_ring *ring;
477 int r, i;
478 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
479
480 /* DMA0 trap event */
481 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 224,
482 &adev->sdma.trap_irq);
483 if (r)
484 return r;
485
486 /* DMA1 trap event */
487 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 244,
488 &adev->sdma.trap_irq);
489 if (r)
490 return r;
491
492 for (i = 0; i < adev->sdma.num_instances; i++) {
493 ring = &adev->sdma.instance[i].ring;
494 ring->ring_obj = NULL;
495 ring->use_doorbell = false;
496 sprintf(ring->name, "sdma%d", i);
497 r = amdgpu_ring_init(adev, ring, 1024,
498 &adev->sdma.trap_irq,
499 (i == 0) ? AMDGPU_SDMA_IRQ_INSTANCE0 :
500 AMDGPU_SDMA_IRQ_INSTANCE1,
501 AMDGPU_RING_PRIO_DEFAULT, NULL);
502 if (r)
503 return r;
504 }
505
506 return r;
507}
508
509static int si_dma_sw_fini(void *handle)
510{
511 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
512 int i;
513
514 for (i = 0; i < adev->sdma.num_instances; i++)
515 amdgpu_ring_fini(&adev->sdma.instance[i].ring);
516
517 return 0;
518}
519
520static int si_dma_hw_init(void *handle)
521{
522 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
523
524 return si_dma_start(adev);
525}
526
527static int si_dma_hw_fini(void *handle)
528{
529 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
530
531 si_dma_stop(adev);
532
533 return 0;
534}
535
536static int si_dma_suspend(void *handle)
537{
538 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
539
540 return si_dma_hw_fini(adev);
541}
542
543static int si_dma_resume(void *handle)
544{
545 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
546
547 return si_dma_hw_init(adev);
548}
549
550static bool si_dma_is_idle(void *handle)
551{
552 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
553 u32 tmp = RREG32(SRBM_STATUS2);
554
555 if (tmp & (DMA_BUSY_MASK | DMA1_BUSY_MASK))
556 return false;
557
558 return true;
559}
560
561static int si_dma_wait_for_idle(void *handle)
562{
563 unsigned i;
564 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
565
566 for (i = 0; i < adev->usec_timeout; i++) {
567 if (si_dma_is_idle(handle))
568 return 0;
569 udelay(1);
570 }
571 return -ETIMEDOUT;
572}
573
574static int si_dma_soft_reset(void *handle)
575{
576 DRM_INFO("si_dma_soft_reset --- not implemented !!!!!!!\n");
577 return 0;
578}
579
580static int si_dma_set_trap_irq_state(struct amdgpu_device *adev,
581 struct amdgpu_irq_src *src,
582 unsigned type,
583 enum amdgpu_interrupt_state state)
584{
585 u32 sdma_cntl;
586
587 switch (type) {
588 case AMDGPU_SDMA_IRQ_INSTANCE0:
589 switch (state) {
590 case AMDGPU_IRQ_STATE_DISABLE:
591 sdma_cntl = RREG32(DMA_CNTL + DMA0_REGISTER_OFFSET);
592 sdma_cntl &= ~TRAP_ENABLE;
593 WREG32(DMA_CNTL + DMA0_REGISTER_OFFSET, sdma_cntl);
594 break;
595 case AMDGPU_IRQ_STATE_ENABLE:
596 sdma_cntl = RREG32(DMA_CNTL + DMA0_REGISTER_OFFSET);
597 sdma_cntl |= TRAP_ENABLE;
598 WREG32(DMA_CNTL + DMA0_REGISTER_OFFSET, sdma_cntl);
599 break;
600 default:
601 break;
602 }
603 break;
604 case AMDGPU_SDMA_IRQ_INSTANCE1:
605 switch (state) {
606 case AMDGPU_IRQ_STATE_DISABLE:
607 sdma_cntl = RREG32(DMA_CNTL + DMA1_REGISTER_OFFSET);
608 sdma_cntl &= ~TRAP_ENABLE;
609 WREG32(DMA_CNTL + DMA1_REGISTER_OFFSET, sdma_cntl);
610 break;
611 case AMDGPU_IRQ_STATE_ENABLE:
612 sdma_cntl = RREG32(DMA_CNTL + DMA1_REGISTER_OFFSET);
613 sdma_cntl |= TRAP_ENABLE;
614 WREG32(DMA_CNTL + DMA1_REGISTER_OFFSET, sdma_cntl);
615 break;
616 default:
617 break;
618 }
619 break;
620 default:
621 break;
622 }
623 return 0;
624}
625
626static int si_dma_process_trap_irq(struct amdgpu_device *adev,
627 struct amdgpu_irq_src *source,
628 struct amdgpu_iv_entry *entry)
629{
630 if (entry->src_id == 224)
631 amdgpu_fence_process(&adev->sdma.instance[0].ring);
632 else
633 amdgpu_fence_process(&adev->sdma.instance[1].ring);
634 return 0;
635}
636
637static int si_dma_set_clockgating_state(void *handle,
638 enum amd_clockgating_state state)
639{
640 u32 orig, data, offset;
641 int i;
642 bool enable;
643 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
644
645 enable = (state == AMD_CG_STATE_GATE);
646
647 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
648 for (i = 0; i < adev->sdma.num_instances; i++) {
649 if (i == 0)
650 offset = DMA0_REGISTER_OFFSET;
651 else
652 offset = DMA1_REGISTER_OFFSET;
653 orig = data = RREG32(DMA_POWER_CNTL + offset);
654 data &= ~MEM_POWER_OVERRIDE;
655 if (data != orig)
656 WREG32(DMA_POWER_CNTL + offset, data);
657 WREG32(DMA_CLK_CTRL + offset, 0x00000100);
658 }
659 } else {
660 for (i = 0; i < adev->sdma.num_instances; i++) {
661 if (i == 0)
662 offset = DMA0_REGISTER_OFFSET;
663 else
664 offset = DMA1_REGISTER_OFFSET;
665 orig = data = RREG32(DMA_POWER_CNTL + offset);
666 data |= MEM_POWER_OVERRIDE;
667 if (data != orig)
668 WREG32(DMA_POWER_CNTL + offset, data);
669
670 orig = data = RREG32(DMA_CLK_CTRL + offset);
671 data = 0xff000000;
672 if (data != orig)
673 WREG32(DMA_CLK_CTRL + offset, data);
674 }
675 }
676
677 return 0;
678}
679
680static int si_dma_set_powergating_state(void *handle,
681 enum amd_powergating_state state)
682{
683 u32 tmp;
684
685 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
686
687 WREG32(DMA_PGFSM_WRITE, 0x00002000);
688 WREG32(DMA_PGFSM_CONFIG, 0x100010ff);
689
690 for (tmp = 0; tmp < 5; tmp++)
691 WREG32(DMA_PGFSM_WRITE, 0);
692
693 return 0;
694}
695
696static const struct amd_ip_funcs si_dma_ip_funcs = {
697 .name = "si_dma",
698 .early_init = si_dma_early_init,
699 .late_init = NULL,
700 .sw_init = si_dma_sw_init,
701 .sw_fini = si_dma_sw_fini,
702 .hw_init = si_dma_hw_init,
703 .hw_fini = si_dma_hw_fini,
704 .suspend = si_dma_suspend,
705 .resume = si_dma_resume,
706 .is_idle = si_dma_is_idle,
707 .wait_for_idle = si_dma_wait_for_idle,
708 .soft_reset = si_dma_soft_reset,
709 .set_clockgating_state = si_dma_set_clockgating_state,
710 .set_powergating_state = si_dma_set_powergating_state,
711};
712
713static const struct amdgpu_ring_funcs si_dma_ring_funcs = {
714 .type = AMDGPU_RING_TYPE_SDMA,
715 .align_mask = 0xf,
716 .nop = DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0),
717 .support_64bit_ptrs = false,
718 .get_rptr = si_dma_ring_get_rptr,
719 .get_wptr = si_dma_ring_get_wptr,
720 .set_wptr = si_dma_ring_set_wptr,
721 .emit_frame_size =
722 3 + 3 + /* hdp flush / invalidate */
723 6 + /* si_dma_ring_emit_pipeline_sync */
724 SI_FLUSH_GPU_TLB_NUM_WREG * 3 + 6 + /* si_dma_ring_emit_vm_flush */
725 9 + 9 + 9, /* si_dma_ring_emit_fence x3 for user fence, vm fence */
726 .emit_ib_size = 7 + 3, /* si_dma_ring_emit_ib */
727 .emit_ib = si_dma_ring_emit_ib,
728 .emit_fence = si_dma_ring_emit_fence,
729 .emit_pipeline_sync = si_dma_ring_emit_pipeline_sync,
730 .emit_vm_flush = si_dma_ring_emit_vm_flush,
731 .test_ring = si_dma_ring_test_ring,
732 .test_ib = si_dma_ring_test_ib,
733 .insert_nop = amdgpu_ring_insert_nop,
734 .pad_ib = si_dma_ring_pad_ib,
735 .emit_wreg = si_dma_ring_emit_wreg,
736};
737
738static void si_dma_set_ring_funcs(struct amdgpu_device *adev)
739{
740 int i;
741
742 for (i = 0; i < adev->sdma.num_instances; i++)
743 adev->sdma.instance[i].ring.funcs = &si_dma_ring_funcs;
744}
745
746static const struct amdgpu_irq_src_funcs si_dma_trap_irq_funcs = {
747 .set = si_dma_set_trap_irq_state,
748 .process = si_dma_process_trap_irq,
749};
750
751static void si_dma_set_irq_funcs(struct amdgpu_device *adev)
752{
753 adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
754 adev->sdma.trap_irq.funcs = &si_dma_trap_irq_funcs;
755}
756
757/**
758 * si_dma_emit_copy_buffer - copy buffer using the sDMA engine
759 *
760 * @ib: indirect buffer to copy to
761 * @src_offset: src GPU address
762 * @dst_offset: dst GPU address
763 * @byte_count: number of bytes to xfer
764 * @tmz: is this a secure operation
765 *
766 * Copy GPU buffers using the DMA engine (VI).
767 * Used by the amdgpu ttm implementation to move pages if
768 * registered as the asic copy callback.
769 */
770static void si_dma_emit_copy_buffer(struct amdgpu_ib *ib,
771 uint64_t src_offset,
772 uint64_t dst_offset,
773 uint32_t byte_count,
774 bool tmz)
775{
776 ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_COPY,
777 1, 0, 0, byte_count);
778 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
779 ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
780 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset) & 0xff;
781 ib->ptr[ib->length_dw++] = upper_32_bits(src_offset) & 0xff;
782}
783
784/**
785 * si_dma_emit_fill_buffer - fill buffer using the sDMA engine
786 *
787 * @ib: indirect buffer to copy to
788 * @src_data: value to write to buffer
789 * @dst_offset: dst GPU address
790 * @byte_count: number of bytes to xfer
791 *
792 * Fill GPU buffers using the DMA engine (VI).
793 */
794static void si_dma_emit_fill_buffer(struct amdgpu_ib *ib,
795 uint32_t src_data,
796 uint64_t dst_offset,
797 uint32_t byte_count)
798{
799 ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_CONSTANT_FILL,
800 0, 0, 0, byte_count / 4);
801 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
802 ib->ptr[ib->length_dw++] = src_data;
803 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset) << 16;
804}
805
806
807static const struct amdgpu_buffer_funcs si_dma_buffer_funcs = {
808 .copy_max_bytes = 0xffff8,
809 .copy_num_dw = 5,
810 .emit_copy_buffer = si_dma_emit_copy_buffer,
811
812 .fill_max_bytes = 0xffff8,
813 .fill_num_dw = 4,
814 .emit_fill_buffer = si_dma_emit_fill_buffer,
815};
816
817static void si_dma_set_buffer_funcs(struct amdgpu_device *adev)
818{
819 adev->mman.buffer_funcs = &si_dma_buffer_funcs;
820 adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
821}
822
823static const struct amdgpu_vm_pte_funcs si_dma_vm_pte_funcs = {
824 .copy_pte_num_dw = 5,
825 .copy_pte = si_dma_vm_copy_pte,
826
827 .write_pte = si_dma_vm_write_pte,
828 .set_pte_pde = si_dma_vm_set_pte_pde,
829};
830
831static void si_dma_set_vm_pte_funcs(struct amdgpu_device *adev)
832{
833 unsigned i;
834
835 adev->vm_manager.vm_pte_funcs = &si_dma_vm_pte_funcs;
836 for (i = 0; i < adev->sdma.num_instances; i++) {
837 adev->vm_manager.vm_pte_scheds[i] =
838 &adev->sdma.instance[i].ring.sched;
839 }
840 adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
841}
842
843const struct amdgpu_ip_block_version si_dma_ip_block =
844{
845 .type = AMD_IP_BLOCK_TYPE_SDMA,
846 .major = 1,
847 .minor = 0,
848 .rev = 0,
849 .funcs = &si_dma_ip_funcs,
850};
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/sid.h"
28
29const u32 sdma_offsets[SDMA_MAX_INSTANCE] =
30{
31 DMA0_REGISTER_OFFSET,
32 DMA1_REGISTER_OFFSET
33};
34
35static void si_dma_set_ring_funcs(struct amdgpu_device *adev);
36static void si_dma_set_buffer_funcs(struct amdgpu_device *adev);
37static void si_dma_set_vm_pte_funcs(struct amdgpu_device *adev);
38static void si_dma_set_irq_funcs(struct amdgpu_device *adev);
39
40static uint32_t si_dma_ring_get_rptr(struct amdgpu_ring *ring)
41{
42 return ring->adev->wb.wb[ring->rptr_offs>>2];
43}
44
45static uint32_t si_dma_ring_get_wptr(struct amdgpu_ring *ring)
46{
47 struct amdgpu_device *adev = ring->adev;
48 u32 me = (ring == &adev->sdma.instance[0].ring) ? 0 : 1;
49
50 return (RREG32(DMA_RB_WPTR + sdma_offsets[me]) & 0x3fffc) >> 2;
51}
52
53static void si_dma_ring_set_wptr(struct amdgpu_ring *ring)
54{
55 struct amdgpu_device *adev = ring->adev;
56 u32 me = (ring == &adev->sdma.instance[0].ring) ? 0 : 1;
57
58 WREG32(DMA_RB_WPTR + sdma_offsets[me], (ring->wptr << 2) & 0x3fffc);
59}
60
61static void si_dma_ring_emit_ib(struct amdgpu_ring *ring,
62 struct amdgpu_ib *ib,
63 unsigned vm_id, bool ctx_switch)
64{
65 /* The indirect buffer packet must end on an 8 DW boundary in the DMA ring.
66 * Pad as necessary with NOPs.
67 */
68 while ((ring->wptr & 7) != 5)
69 amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0));
70 amdgpu_ring_write(ring, DMA_IB_PACKET(DMA_PACKET_INDIRECT_BUFFER, vm_id, 0));
71 amdgpu_ring_write(ring, (ib->gpu_addr & 0xFFFFFFE0));
72 amdgpu_ring_write(ring, (ib->length_dw << 12) | (upper_32_bits(ib->gpu_addr) & 0xFF));
73
74}
75
76static void si_dma_ring_emit_hdp_flush(struct amdgpu_ring *ring)
77{
78 amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0, 0, 0));
79 amdgpu_ring_write(ring, (0xf << 16) | (HDP_MEM_COHERENCY_FLUSH_CNTL));
80 amdgpu_ring_write(ring, 1);
81}
82
83static void si_dma_ring_emit_hdp_invalidate(struct amdgpu_ring *ring)
84{
85 amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0, 0, 0));
86 amdgpu_ring_write(ring, (0xf << 16) | (HDP_DEBUG0));
87 amdgpu_ring_write(ring, 1);
88}
89
90/**
91 * si_dma_ring_emit_fence - emit a fence on the DMA ring
92 *
93 * @ring: amdgpu ring pointer
94 * @fence: amdgpu fence object
95 *
96 * Add a DMA fence packet to the ring to write
97 * the fence seq number and DMA trap packet to generate
98 * an interrupt if needed (VI).
99 */
100static void si_dma_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
101 unsigned flags)
102{
103
104 bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
105 /* write the fence */
106 amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_FENCE, 0, 0, 0, 0));
107 amdgpu_ring_write(ring, addr & 0xfffffffc);
108 amdgpu_ring_write(ring, (upper_32_bits(addr) & 0xff));
109 amdgpu_ring_write(ring, seq);
110 /* optionally write high bits as well */
111 if (write64bit) {
112 addr += 4;
113 amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_FENCE, 0, 0, 0, 0));
114 amdgpu_ring_write(ring, addr & 0xfffffffc);
115 amdgpu_ring_write(ring, (upper_32_bits(addr) & 0xff));
116 amdgpu_ring_write(ring, upper_32_bits(seq));
117 }
118 /* generate an interrupt */
119 amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_TRAP, 0, 0, 0, 0));
120}
121
122static void si_dma_stop(struct amdgpu_device *adev)
123{
124 struct amdgpu_ring *ring;
125 u32 rb_cntl;
126 unsigned i;
127
128 for (i = 0; i < adev->sdma.num_instances; i++) {
129 ring = &adev->sdma.instance[i].ring;
130 /* dma0 */
131 rb_cntl = RREG32(DMA_RB_CNTL + sdma_offsets[i]);
132 rb_cntl &= ~DMA_RB_ENABLE;
133 WREG32(DMA_RB_CNTL + sdma_offsets[i], rb_cntl);
134
135 if (adev->mman.buffer_funcs_ring == ring)
136 amdgpu_ttm_set_active_vram_size(adev, adev->mc.visible_vram_size);
137 ring->ready = false;
138 }
139}
140
141static int si_dma_start(struct amdgpu_device *adev)
142{
143 struct amdgpu_ring *ring;
144 u32 rb_cntl, dma_cntl, ib_cntl, rb_bufsz;
145 int i, r;
146 uint64_t rptr_addr;
147
148 for (i = 0; i < adev->sdma.num_instances; i++) {
149 ring = &adev->sdma.instance[i].ring;
150
151 WREG32(DMA_SEM_INCOMPLETE_TIMER_CNTL + sdma_offsets[i], 0);
152 WREG32(DMA_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
153
154 /* Set ring buffer size in dwords */
155 rb_bufsz = order_base_2(ring->ring_size / 4);
156 rb_cntl = rb_bufsz << 1;
157#ifdef __BIG_ENDIAN
158 rb_cntl |= DMA_RB_SWAP_ENABLE | DMA_RPTR_WRITEBACK_SWAP_ENABLE;
159#endif
160 WREG32(DMA_RB_CNTL + sdma_offsets[i], rb_cntl);
161
162 /* Initialize the ring buffer's read and write pointers */
163 WREG32(DMA_RB_RPTR + sdma_offsets[i], 0);
164 WREG32(DMA_RB_WPTR + sdma_offsets[i], 0);
165
166 rptr_addr = adev->wb.gpu_addr + (ring->rptr_offs * 4);
167
168 WREG32(DMA_RB_RPTR_ADDR_LO + sdma_offsets[i], lower_32_bits(rptr_addr));
169 WREG32(DMA_RB_RPTR_ADDR_HI + sdma_offsets[i], upper_32_bits(rptr_addr) & 0xFF);
170
171 rb_cntl |= DMA_RPTR_WRITEBACK_ENABLE;
172
173 WREG32(DMA_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
174
175 /* enable DMA IBs */
176 ib_cntl = DMA_IB_ENABLE | CMD_VMID_FORCE;
177#ifdef __BIG_ENDIAN
178 ib_cntl |= DMA_IB_SWAP_ENABLE;
179#endif
180 WREG32(DMA_IB_CNTL + sdma_offsets[i], ib_cntl);
181
182 dma_cntl = RREG32(DMA_CNTL + sdma_offsets[i]);
183 dma_cntl &= ~CTXEMPTY_INT_ENABLE;
184 WREG32(DMA_CNTL + sdma_offsets[i], dma_cntl);
185
186 ring->wptr = 0;
187 WREG32(DMA_RB_WPTR + sdma_offsets[i], ring->wptr << 2);
188 WREG32(DMA_RB_CNTL + sdma_offsets[i], rb_cntl | DMA_RB_ENABLE);
189
190 ring->ready = true;
191
192 r = amdgpu_ring_test_ring(ring);
193 if (r) {
194 ring->ready = false;
195 return r;
196 }
197
198 if (adev->mman.buffer_funcs_ring == ring)
199 amdgpu_ttm_set_active_vram_size(adev, adev->mc.real_vram_size);
200 }
201
202 return 0;
203}
204
205/**
206 * si_dma_ring_test_ring - simple async dma engine test
207 *
208 * @ring: amdgpu_ring structure holding ring information
209 *
210 * Test the DMA engine by writing using it to write an
211 * value to memory. (VI).
212 * Returns 0 for success, error for failure.
213 */
214static int si_dma_ring_test_ring(struct amdgpu_ring *ring)
215{
216 struct amdgpu_device *adev = ring->adev;
217 unsigned i;
218 unsigned index;
219 int r;
220 u32 tmp;
221 u64 gpu_addr;
222
223 r = amdgpu_wb_get(adev, &index);
224 if (r) {
225 dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
226 return r;
227 }
228
229 gpu_addr = adev->wb.gpu_addr + (index * 4);
230 tmp = 0xCAFEDEAD;
231 adev->wb.wb[index] = cpu_to_le32(tmp);
232
233 r = amdgpu_ring_alloc(ring, 4);
234 if (r) {
235 DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
236 amdgpu_wb_free(adev, index);
237 return r;
238 }
239
240 amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 0, 1));
241 amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
242 amdgpu_ring_write(ring, upper_32_bits(gpu_addr) & 0xff);
243 amdgpu_ring_write(ring, 0xDEADBEEF);
244 amdgpu_ring_commit(ring);
245
246 for (i = 0; i < adev->usec_timeout; i++) {
247 tmp = le32_to_cpu(adev->wb.wb[index]);
248 if (tmp == 0xDEADBEEF)
249 break;
250 DRM_UDELAY(1);
251 }
252
253 if (i < adev->usec_timeout) {
254 DRM_INFO("ring test on %d succeeded in %d usecs\n", ring->idx, i);
255 } else {
256 DRM_ERROR("amdgpu: ring %d test failed (0x%08X)\n",
257 ring->idx, tmp);
258 r = -EINVAL;
259 }
260 amdgpu_wb_free(adev, index);
261
262 return r;
263}
264
265/**
266 * si_dma_ring_test_ib - test an IB on the DMA engine
267 *
268 * @ring: amdgpu_ring structure holding ring information
269 *
270 * Test a simple IB in the DMA ring (VI).
271 * Returns 0 on success, error on failure.
272 */
273static int si_dma_ring_test_ib(struct amdgpu_ring *ring, long timeout)
274{
275 struct amdgpu_device *adev = ring->adev;
276 struct amdgpu_ib ib;
277 struct dma_fence *f = NULL;
278 unsigned index;
279 u32 tmp = 0;
280 u64 gpu_addr;
281 long r;
282
283 r = amdgpu_wb_get(adev, &index);
284 if (r) {
285 dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
286 return r;
287 }
288
289 gpu_addr = adev->wb.gpu_addr + (index * 4);
290 tmp = 0xCAFEDEAD;
291 adev->wb.wb[index] = cpu_to_le32(tmp);
292 memset(&ib, 0, sizeof(ib));
293 r = amdgpu_ib_get(adev, NULL, 256, &ib);
294 if (r) {
295 DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
296 goto err0;
297 }
298
299 ib.ptr[0] = DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 0, 1);
300 ib.ptr[1] = lower_32_bits(gpu_addr);
301 ib.ptr[2] = upper_32_bits(gpu_addr) & 0xff;
302 ib.ptr[3] = 0xDEADBEEF;
303 ib.length_dw = 4;
304 r = amdgpu_ib_schedule(ring, 1, &ib, NULL, NULL, &f);
305 if (r)
306 goto err1;
307
308 r = dma_fence_wait_timeout(f, false, timeout);
309 if (r == 0) {
310 DRM_ERROR("amdgpu: IB test timed out\n");
311 r = -ETIMEDOUT;
312 goto err1;
313 } else if (r < 0) {
314 DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
315 goto err1;
316 }
317 tmp = le32_to_cpu(adev->wb.wb[index]);
318 if (tmp == 0xDEADBEEF) {
319 DRM_INFO("ib test on ring %d succeeded\n", ring->idx);
320 r = 0;
321 } else {
322 DRM_ERROR("amdgpu: ib test failed (0x%08X)\n", tmp);
323 r = -EINVAL;
324 }
325
326err1:
327 amdgpu_ib_free(adev, &ib, NULL);
328 dma_fence_put(f);
329err0:
330 amdgpu_wb_free(adev, index);
331 return r;
332}
333
334/**
335 * cik_dma_vm_copy_pte - update PTEs by copying them from the GART
336 *
337 * @ib: indirect buffer to fill with commands
338 * @pe: addr of the page entry
339 * @src: src addr to copy from
340 * @count: number of page entries to update
341 *
342 * Update PTEs by copying them from the GART using DMA (SI).
343 */
344static void si_dma_vm_copy_pte(struct amdgpu_ib *ib,
345 uint64_t pe, uint64_t src,
346 unsigned count)
347{
348 unsigned bytes = count * 8;
349
350 ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_COPY,
351 1, 0, 0, bytes);
352 ib->ptr[ib->length_dw++] = lower_32_bits(pe);
353 ib->ptr[ib->length_dw++] = lower_32_bits(src);
354 ib->ptr[ib->length_dw++] = upper_32_bits(pe) & 0xff;
355 ib->ptr[ib->length_dw++] = upper_32_bits(src) & 0xff;
356}
357
358/**
359 * si_dma_vm_write_pte - update PTEs by writing them manually
360 *
361 * @ib: indirect buffer to fill with commands
362 * @pe: addr of the page entry
363 * @value: dst addr to write into pe
364 * @count: number of page entries to update
365 * @incr: increase next addr by incr bytes
366 *
367 * Update PTEs by writing them manually using DMA (SI).
368 */
369static void si_dma_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
370 uint64_t value, unsigned count,
371 uint32_t incr)
372{
373 unsigned ndw = count * 2;
374
375 ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 0, ndw);
376 ib->ptr[ib->length_dw++] = lower_32_bits(pe);
377 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
378 for (; ndw > 0; ndw -= 2) {
379 ib->ptr[ib->length_dw++] = lower_32_bits(value);
380 ib->ptr[ib->length_dw++] = upper_32_bits(value);
381 value += incr;
382 }
383}
384
385/**
386 * si_dma_vm_set_pte_pde - update the page tables using sDMA
387 *
388 * @ib: indirect buffer to fill with commands
389 * @pe: addr of the page entry
390 * @addr: dst addr to write into pe
391 * @count: number of page entries to update
392 * @incr: increase next addr by incr bytes
393 * @flags: access flags
394 *
395 * Update the page tables using sDMA (CIK).
396 */
397static void si_dma_vm_set_pte_pde(struct amdgpu_ib *ib,
398 uint64_t pe,
399 uint64_t addr, unsigned count,
400 uint32_t incr, uint32_t flags)
401{
402 uint64_t value;
403 unsigned ndw;
404
405 while (count) {
406 ndw = count * 2;
407 if (ndw > 0xFFFFE)
408 ndw = 0xFFFFE;
409
410 if (flags & AMDGPU_PTE_VALID)
411 value = addr;
412 else
413 value = 0;
414
415 /* for physically contiguous pages (vram) */
416 ib->ptr[ib->length_dw++] = DMA_PTE_PDE_PACKET(ndw);
417 ib->ptr[ib->length_dw++] = pe; /* dst addr */
418 ib->ptr[ib->length_dw++] = upper_32_bits(pe) & 0xff;
419 ib->ptr[ib->length_dw++] = flags; /* mask */
420 ib->ptr[ib->length_dw++] = 0;
421 ib->ptr[ib->length_dw++] = value; /* value */
422 ib->ptr[ib->length_dw++] = upper_32_bits(value);
423 ib->ptr[ib->length_dw++] = incr; /* increment size */
424 ib->ptr[ib->length_dw++] = 0;
425 pe += ndw * 4;
426 addr += (ndw / 2) * incr;
427 count -= ndw / 2;
428 }
429}
430
431/**
432 * si_dma_pad_ib - pad the IB to the required number of dw
433 *
434 * @ib: indirect buffer to fill with padding
435 *
436 */
437static void si_dma_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
438{
439 while (ib->length_dw & 0x7)
440 ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0);
441}
442
443/**
444 * cik_sdma_ring_emit_pipeline_sync - sync the pipeline
445 *
446 * @ring: amdgpu_ring pointer
447 *
448 * Make sure all previous operations are completed (CIK).
449 */
450static void si_dma_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
451{
452 uint32_t seq = ring->fence_drv.sync_seq;
453 uint64_t addr = ring->fence_drv.gpu_addr;
454
455 /* wait for idle */
456 amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_POLL_REG_MEM, 0, 0, 0, 0) |
457 (1 << 27)); /* Poll memory */
458 amdgpu_ring_write(ring, lower_32_bits(addr));
459 amdgpu_ring_write(ring, (0xff << 16) | upper_32_bits(addr)); /* retry, addr_hi */
460 amdgpu_ring_write(ring, 0xffffffff); /* mask */
461 amdgpu_ring_write(ring, seq); /* value */
462 amdgpu_ring_write(ring, (3 << 28) | 0x20); /* func(equal) | poll interval */
463}
464
465/**
466 * si_dma_ring_emit_vm_flush - cik vm flush using sDMA
467 *
468 * @ring: amdgpu_ring pointer
469 * @vm: amdgpu_vm pointer
470 *
471 * Update the page table base and flush the VM TLB
472 * using sDMA (VI).
473 */
474static void si_dma_ring_emit_vm_flush(struct amdgpu_ring *ring,
475 unsigned vm_id, uint64_t pd_addr)
476{
477 amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0, 0, 0));
478 if (vm_id < 8)
479 amdgpu_ring_write(ring, (0xf << 16) | (VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + vm_id));
480 else
481 amdgpu_ring_write(ring, (0xf << 16) | (VM_CONTEXT8_PAGE_TABLE_BASE_ADDR + (vm_id - 8)));
482 amdgpu_ring_write(ring, pd_addr >> 12);
483
484 /* bits 0-7 are the VM contexts0-7 */
485 amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0, 0, 0));
486 amdgpu_ring_write(ring, (0xf << 16) | (VM_INVALIDATE_REQUEST));
487 amdgpu_ring_write(ring, 1 << vm_id);
488
489 /* wait for invalidate to complete */
490 amdgpu_ring_write(ring, DMA_PACKET(DMA_PACKET_POLL_REG_MEM, 0, 0, 0, 0));
491 amdgpu_ring_write(ring, VM_INVALIDATE_REQUEST);
492 amdgpu_ring_write(ring, 0xff << 16); /* retry */
493 amdgpu_ring_write(ring, 1 << vm_id); /* mask */
494 amdgpu_ring_write(ring, 0); /* value */
495 amdgpu_ring_write(ring, (0 << 28) | 0x20); /* func(always) | poll interval */
496}
497
498static int si_dma_early_init(void *handle)
499{
500 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
501
502 adev->sdma.num_instances = 2;
503
504 si_dma_set_ring_funcs(adev);
505 si_dma_set_buffer_funcs(adev);
506 si_dma_set_vm_pte_funcs(adev);
507 si_dma_set_irq_funcs(adev);
508
509 return 0;
510}
511
512static int si_dma_sw_init(void *handle)
513{
514 struct amdgpu_ring *ring;
515 int r, i;
516 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
517
518 /* DMA0 trap event */
519 r = amdgpu_irq_add_id(adev, 224, &adev->sdma.trap_irq);
520 if (r)
521 return r;
522
523 /* DMA1 trap event */
524 r = amdgpu_irq_add_id(adev, 244, &adev->sdma.trap_irq_1);
525 if (r)
526 return r;
527
528 for (i = 0; i < adev->sdma.num_instances; i++) {
529 ring = &adev->sdma.instance[i].ring;
530 ring->ring_obj = NULL;
531 ring->use_doorbell = false;
532 sprintf(ring->name, "sdma%d", i);
533 r = amdgpu_ring_init(adev, ring, 1024,
534 &adev->sdma.trap_irq,
535 (i == 0) ?
536 AMDGPU_SDMA_IRQ_TRAP0 :
537 AMDGPU_SDMA_IRQ_TRAP1);
538 if (r)
539 return r;
540 }
541
542 return r;
543}
544
545static int si_dma_sw_fini(void *handle)
546{
547 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
548 int i;
549
550 for (i = 0; i < adev->sdma.num_instances; i++)
551 amdgpu_ring_fini(&adev->sdma.instance[i].ring);
552
553 return 0;
554}
555
556static int si_dma_hw_init(void *handle)
557{
558 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
559
560 return si_dma_start(adev);
561}
562
563static int si_dma_hw_fini(void *handle)
564{
565 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
566
567 si_dma_stop(adev);
568
569 return 0;
570}
571
572static int si_dma_suspend(void *handle)
573{
574 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
575
576 return si_dma_hw_fini(adev);
577}
578
579static int si_dma_resume(void *handle)
580{
581 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
582
583 return si_dma_hw_init(adev);
584}
585
586static bool si_dma_is_idle(void *handle)
587{
588 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
589 u32 tmp = RREG32(SRBM_STATUS2);
590
591 if (tmp & (DMA_BUSY_MASK | DMA1_BUSY_MASK))
592 return false;
593
594 return true;
595}
596
597static int si_dma_wait_for_idle(void *handle)
598{
599 unsigned i;
600 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
601
602 for (i = 0; i < adev->usec_timeout; i++) {
603 if (si_dma_is_idle(handle))
604 return 0;
605 udelay(1);
606 }
607 return -ETIMEDOUT;
608}
609
610static int si_dma_soft_reset(void *handle)
611{
612 DRM_INFO("si_dma_soft_reset --- not implemented !!!!!!!\n");
613 return 0;
614}
615
616static int si_dma_set_trap_irq_state(struct amdgpu_device *adev,
617 struct amdgpu_irq_src *src,
618 unsigned type,
619 enum amdgpu_interrupt_state state)
620{
621 u32 sdma_cntl;
622
623 switch (type) {
624 case AMDGPU_SDMA_IRQ_TRAP0:
625 switch (state) {
626 case AMDGPU_IRQ_STATE_DISABLE:
627 sdma_cntl = RREG32(DMA_CNTL + DMA0_REGISTER_OFFSET);
628 sdma_cntl &= ~TRAP_ENABLE;
629 WREG32(DMA_CNTL + DMA0_REGISTER_OFFSET, sdma_cntl);
630 break;
631 case AMDGPU_IRQ_STATE_ENABLE:
632 sdma_cntl = RREG32(DMA_CNTL + DMA0_REGISTER_OFFSET);
633 sdma_cntl |= TRAP_ENABLE;
634 WREG32(DMA_CNTL + DMA0_REGISTER_OFFSET, sdma_cntl);
635 break;
636 default:
637 break;
638 }
639 break;
640 case AMDGPU_SDMA_IRQ_TRAP1:
641 switch (state) {
642 case AMDGPU_IRQ_STATE_DISABLE:
643 sdma_cntl = RREG32(DMA_CNTL + DMA1_REGISTER_OFFSET);
644 sdma_cntl &= ~TRAP_ENABLE;
645 WREG32(DMA_CNTL + DMA1_REGISTER_OFFSET, sdma_cntl);
646 break;
647 case AMDGPU_IRQ_STATE_ENABLE:
648 sdma_cntl = RREG32(DMA_CNTL + DMA1_REGISTER_OFFSET);
649 sdma_cntl |= TRAP_ENABLE;
650 WREG32(DMA_CNTL + DMA1_REGISTER_OFFSET, sdma_cntl);
651 break;
652 default:
653 break;
654 }
655 break;
656 default:
657 break;
658 }
659 return 0;
660}
661
662static int si_dma_process_trap_irq(struct amdgpu_device *adev,
663 struct amdgpu_irq_src *source,
664 struct amdgpu_iv_entry *entry)
665{
666 amdgpu_fence_process(&adev->sdma.instance[0].ring);
667
668 return 0;
669}
670
671static int si_dma_process_trap_irq_1(struct amdgpu_device *adev,
672 struct amdgpu_irq_src *source,
673 struct amdgpu_iv_entry *entry)
674{
675 amdgpu_fence_process(&adev->sdma.instance[1].ring);
676
677 return 0;
678}
679
680static int si_dma_process_illegal_inst_irq(struct amdgpu_device *adev,
681 struct amdgpu_irq_src *source,
682 struct amdgpu_iv_entry *entry)
683{
684 DRM_ERROR("Illegal instruction in SDMA command stream\n");
685 schedule_work(&adev->reset_work);
686 return 0;
687}
688
689static int si_dma_set_clockgating_state(void *handle,
690 enum amd_clockgating_state state)
691{
692 u32 orig, data, offset;
693 int i;
694 bool enable;
695 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
696
697 enable = (state == AMD_CG_STATE_GATE) ? true : false;
698
699 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
700 for (i = 0; i < adev->sdma.num_instances; i++) {
701 if (i == 0)
702 offset = DMA0_REGISTER_OFFSET;
703 else
704 offset = DMA1_REGISTER_OFFSET;
705 orig = data = RREG32(DMA_POWER_CNTL + offset);
706 data &= ~MEM_POWER_OVERRIDE;
707 if (data != orig)
708 WREG32(DMA_POWER_CNTL + offset, data);
709 WREG32(DMA_CLK_CTRL + offset, 0x00000100);
710 }
711 } else {
712 for (i = 0; i < adev->sdma.num_instances; i++) {
713 if (i == 0)
714 offset = DMA0_REGISTER_OFFSET;
715 else
716 offset = DMA1_REGISTER_OFFSET;
717 orig = data = RREG32(DMA_POWER_CNTL + offset);
718 data |= MEM_POWER_OVERRIDE;
719 if (data != orig)
720 WREG32(DMA_POWER_CNTL + offset, data);
721
722 orig = data = RREG32(DMA_CLK_CTRL + offset);
723 data = 0xff000000;
724 if (data != orig)
725 WREG32(DMA_CLK_CTRL + offset, data);
726 }
727 }
728
729 return 0;
730}
731
732static int si_dma_set_powergating_state(void *handle,
733 enum amd_powergating_state state)
734{
735 u32 tmp;
736
737 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
738
739 WREG32(DMA_PGFSM_WRITE, 0x00002000);
740 WREG32(DMA_PGFSM_CONFIG, 0x100010ff);
741
742 for (tmp = 0; tmp < 5; tmp++)
743 WREG32(DMA_PGFSM_WRITE, 0);
744
745 return 0;
746}
747
748static const struct amd_ip_funcs si_dma_ip_funcs = {
749 .name = "si_dma",
750 .early_init = si_dma_early_init,
751 .late_init = NULL,
752 .sw_init = si_dma_sw_init,
753 .sw_fini = si_dma_sw_fini,
754 .hw_init = si_dma_hw_init,
755 .hw_fini = si_dma_hw_fini,
756 .suspend = si_dma_suspend,
757 .resume = si_dma_resume,
758 .is_idle = si_dma_is_idle,
759 .wait_for_idle = si_dma_wait_for_idle,
760 .soft_reset = si_dma_soft_reset,
761 .set_clockgating_state = si_dma_set_clockgating_state,
762 .set_powergating_state = si_dma_set_powergating_state,
763};
764
765static const struct amdgpu_ring_funcs si_dma_ring_funcs = {
766 .type = AMDGPU_RING_TYPE_SDMA,
767 .align_mask = 0xf,
768 .nop = DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0),
769 .get_rptr = si_dma_ring_get_rptr,
770 .get_wptr = si_dma_ring_get_wptr,
771 .set_wptr = si_dma_ring_set_wptr,
772 .emit_frame_size =
773 3 + /* si_dma_ring_emit_hdp_flush */
774 3 + /* si_dma_ring_emit_hdp_invalidate */
775 6 + /* si_dma_ring_emit_pipeline_sync */
776 12 + /* si_dma_ring_emit_vm_flush */
777 9 + 9 + 9, /* si_dma_ring_emit_fence x3 for user fence, vm fence */
778 .emit_ib_size = 7 + 3, /* si_dma_ring_emit_ib */
779 .emit_ib = si_dma_ring_emit_ib,
780 .emit_fence = si_dma_ring_emit_fence,
781 .emit_pipeline_sync = si_dma_ring_emit_pipeline_sync,
782 .emit_vm_flush = si_dma_ring_emit_vm_flush,
783 .emit_hdp_flush = si_dma_ring_emit_hdp_flush,
784 .emit_hdp_invalidate = si_dma_ring_emit_hdp_invalidate,
785 .test_ring = si_dma_ring_test_ring,
786 .test_ib = si_dma_ring_test_ib,
787 .insert_nop = amdgpu_ring_insert_nop,
788 .pad_ib = si_dma_ring_pad_ib,
789};
790
791static void si_dma_set_ring_funcs(struct amdgpu_device *adev)
792{
793 int i;
794
795 for (i = 0; i < adev->sdma.num_instances; i++)
796 adev->sdma.instance[i].ring.funcs = &si_dma_ring_funcs;
797}
798
799static const struct amdgpu_irq_src_funcs si_dma_trap_irq_funcs = {
800 .set = si_dma_set_trap_irq_state,
801 .process = si_dma_process_trap_irq,
802};
803
804static const struct amdgpu_irq_src_funcs si_dma_trap_irq_funcs_1 = {
805 .set = si_dma_set_trap_irq_state,
806 .process = si_dma_process_trap_irq_1,
807};
808
809static const struct amdgpu_irq_src_funcs si_dma_illegal_inst_irq_funcs = {
810 .process = si_dma_process_illegal_inst_irq,
811};
812
813static void si_dma_set_irq_funcs(struct amdgpu_device *adev)
814{
815 adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
816 adev->sdma.trap_irq.funcs = &si_dma_trap_irq_funcs;
817 adev->sdma.trap_irq_1.funcs = &si_dma_trap_irq_funcs_1;
818 adev->sdma.illegal_inst_irq.funcs = &si_dma_illegal_inst_irq_funcs;
819}
820
821/**
822 * si_dma_emit_copy_buffer - copy buffer using the sDMA engine
823 *
824 * @ring: amdgpu_ring structure holding ring information
825 * @src_offset: src GPU address
826 * @dst_offset: dst GPU address
827 * @byte_count: number of bytes to xfer
828 *
829 * Copy GPU buffers using the DMA engine (VI).
830 * Used by the amdgpu ttm implementation to move pages if
831 * registered as the asic copy callback.
832 */
833static void si_dma_emit_copy_buffer(struct amdgpu_ib *ib,
834 uint64_t src_offset,
835 uint64_t dst_offset,
836 uint32_t byte_count)
837{
838 ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_COPY,
839 1, 0, 0, byte_count);
840 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
841 ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
842 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset) & 0xff;
843 ib->ptr[ib->length_dw++] = upper_32_bits(src_offset) & 0xff;
844}
845
846/**
847 * si_dma_emit_fill_buffer - fill buffer using the sDMA engine
848 *
849 * @ring: amdgpu_ring structure holding ring information
850 * @src_data: value to write to buffer
851 * @dst_offset: dst GPU address
852 * @byte_count: number of bytes to xfer
853 *
854 * Fill GPU buffers using the DMA engine (VI).
855 */
856static void si_dma_emit_fill_buffer(struct amdgpu_ib *ib,
857 uint32_t src_data,
858 uint64_t dst_offset,
859 uint32_t byte_count)
860{
861 ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_CONSTANT_FILL,
862 0, 0, 0, byte_count / 4);
863 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
864 ib->ptr[ib->length_dw++] = src_data;
865 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset) << 16;
866}
867
868
869static const struct amdgpu_buffer_funcs si_dma_buffer_funcs = {
870 .copy_max_bytes = 0xffff8,
871 .copy_num_dw = 5,
872 .emit_copy_buffer = si_dma_emit_copy_buffer,
873
874 .fill_max_bytes = 0xffff8,
875 .fill_num_dw = 4,
876 .emit_fill_buffer = si_dma_emit_fill_buffer,
877};
878
879static void si_dma_set_buffer_funcs(struct amdgpu_device *adev)
880{
881 if (adev->mman.buffer_funcs == NULL) {
882 adev->mman.buffer_funcs = &si_dma_buffer_funcs;
883 adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
884 }
885}
886
887static const struct amdgpu_vm_pte_funcs si_dma_vm_pte_funcs = {
888 .copy_pte = si_dma_vm_copy_pte,
889 .write_pte = si_dma_vm_write_pte,
890 .set_pte_pde = si_dma_vm_set_pte_pde,
891};
892
893static void si_dma_set_vm_pte_funcs(struct amdgpu_device *adev)
894{
895 unsigned i;
896
897 if (adev->vm_manager.vm_pte_funcs == NULL) {
898 adev->vm_manager.vm_pte_funcs = &si_dma_vm_pte_funcs;
899 for (i = 0; i < adev->sdma.num_instances; i++)
900 adev->vm_manager.vm_pte_rings[i] =
901 &adev->sdma.instance[i].ring;
902
903 adev->vm_manager.vm_pte_num_rings = adev->sdma.num_instances;
904 }
905}
906
907const struct amdgpu_ip_block_version si_dma_ip_block =
908{
909 .type = AMD_IP_BLOCK_TYPE_SDMA,
910 .major = 1,
911 .minor = 0,
912 .rev = 0,
913 .funcs = &si_dma_ip_funcs,
914};