1 /*
2 * Copyright 2010 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 "radeon.h"
26 #include "radeon_asic.h"
27 #include "radeon_trace.h"
28 #include "nid.h"
29
30 u32 cayman_gpu_check_soft_reset(struct radeon_device *rdev);
31
32 /*
33 * DMA
34 * Starting with R600, the GPU has an asynchronous
35 * DMA engine. The programming model is very similar
36 * to the 3D engine (ring buffer, IBs, etc.), but the
37 * DMA controller has it's own packet format that is
38 * different form the PM4 format used by the 3D engine.
39 * It supports copying data, writing embedded data,
40 * solid fills, and a number of other things. It also
41 * has support for tiling/detiling of buffers.
42 * Cayman and newer support two asynchronous DMA engines.
43 */
44
45 /**
46 * cayman_dma_get_rptr - get the current read pointer
47 *
48 * @rdev: radeon_device pointer
49 * @ring: radeon ring pointer
50 *
51 * Get the current rptr from the hardware (cayman+).
52 */
cayman_dma_get_rptr(struct radeon_device * rdev,struct radeon_ring * ring)53 uint32_t cayman_dma_get_rptr(struct radeon_device *rdev,
54 struct radeon_ring *ring)
55 {
56 u32 rptr, reg;
57
58 if (rdev->wb.enabled) {
59 rptr = rdev->wb.wb[ring->rptr_offs/4];
60 } else {
61 if (ring->idx == R600_RING_TYPE_DMA_INDEX)
62 reg = DMA_RB_RPTR + DMA0_REGISTER_OFFSET;
63 else
64 reg = DMA_RB_RPTR + DMA1_REGISTER_OFFSET;
65
66 rptr = RREG32(reg);
67 }
68
69 return (rptr & 0x3fffc) >> 2;
70 }
71
72 /**
73 * cayman_dma_get_wptr - get the current write pointer
74 *
75 * @rdev: radeon_device pointer
76 * @ring: radeon ring pointer
77 *
78 * Get the current wptr from the hardware (cayman+).
79 */
cayman_dma_get_wptr(struct radeon_device * rdev,struct radeon_ring * ring)80 uint32_t cayman_dma_get_wptr(struct radeon_device *rdev,
81 struct radeon_ring *ring)
82 {
83 u32 reg;
84
85 if (ring->idx == R600_RING_TYPE_DMA_INDEX)
86 reg = DMA_RB_WPTR + DMA0_REGISTER_OFFSET;
87 else
88 reg = DMA_RB_WPTR + DMA1_REGISTER_OFFSET;
89
90 return (RREG32(reg) & 0x3fffc) >> 2;
91 }
92
93 /**
94 * cayman_dma_set_wptr - commit the write pointer
95 *
96 * @rdev: radeon_device pointer
97 * @ring: radeon ring pointer
98 *
99 * Write the wptr back to the hardware (cayman+).
100 */
cayman_dma_set_wptr(struct radeon_device * rdev,struct radeon_ring * ring)101 void cayman_dma_set_wptr(struct radeon_device *rdev,
102 struct radeon_ring *ring)
103 {
104 u32 reg;
105
106 if (ring->idx == R600_RING_TYPE_DMA_INDEX)
107 reg = DMA_RB_WPTR + DMA0_REGISTER_OFFSET;
108 else
109 reg = DMA_RB_WPTR + DMA1_REGISTER_OFFSET;
110
111 WREG32(reg, (ring->wptr << 2) & 0x3fffc);
112 }
113
114 /**
115 * cayman_dma_ring_ib_execute - Schedule an IB on the DMA engine
116 *
117 * @rdev: radeon_device pointer
118 * @ib: IB object to schedule
119 *
120 * Schedule an IB in the DMA ring (cayman-SI).
121 */
cayman_dma_ring_ib_execute(struct radeon_device * rdev,struct radeon_ib * ib)122 void cayman_dma_ring_ib_execute(struct radeon_device *rdev,
123 struct radeon_ib *ib)
124 {
125 struct radeon_ring *ring = &rdev->ring[ib->ring];
126 unsigned vm_id = ib->vm ? ib->vm->ids[ib->ring].id : 0;
127
128 if (rdev->wb.enabled) {
129 u32 next_rptr = ring->wptr + 4;
130 while ((next_rptr & 7) != 5)
131 next_rptr++;
132 next_rptr += 3;
133 radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 1));
134 radeon_ring_write(ring, ring->next_rptr_gpu_addr & 0xfffffffc);
135 radeon_ring_write(ring, upper_32_bits(ring->next_rptr_gpu_addr) & 0xff);
136 radeon_ring_write(ring, next_rptr);
137 }
138
139 /* The indirect buffer packet must end on an 8 DW boundary in the DMA ring.
140 * Pad as necessary with NOPs.
141 */
142 while ((ring->wptr & 7) != 5)
143 radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0));
144 radeon_ring_write(ring, DMA_IB_PACKET(DMA_PACKET_INDIRECT_BUFFER, vm_id, 0));
145 radeon_ring_write(ring, (ib->gpu_addr & 0xFFFFFFE0));
146 radeon_ring_write(ring, (ib->length_dw << 12) | (upper_32_bits(ib->gpu_addr) & 0xFF));
147
148 }
149
150 /**
151 * cayman_dma_stop - stop the async dma engines
152 *
153 * @rdev: radeon_device pointer
154 *
155 * Stop the async dma engines (cayman-SI).
156 */
cayman_dma_stop(struct radeon_device * rdev)157 void cayman_dma_stop(struct radeon_device *rdev)
158 {
159 u32 rb_cntl;
160
161 if ((rdev->asic->copy.copy_ring_index == R600_RING_TYPE_DMA_INDEX) ||
162 (rdev->asic->copy.copy_ring_index == CAYMAN_RING_TYPE_DMA1_INDEX))
163 radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
164
165 /* dma0 */
166 rb_cntl = RREG32(DMA_RB_CNTL + DMA0_REGISTER_OFFSET);
167 rb_cntl &= ~DMA_RB_ENABLE;
168 WREG32(DMA_RB_CNTL + DMA0_REGISTER_OFFSET, rb_cntl);
169
170 /* dma1 */
171 rb_cntl = RREG32(DMA_RB_CNTL + DMA1_REGISTER_OFFSET);
172 rb_cntl &= ~DMA_RB_ENABLE;
173 WREG32(DMA_RB_CNTL + DMA1_REGISTER_OFFSET, rb_cntl);
174
175 rdev->ring[R600_RING_TYPE_DMA_INDEX].ready = false;
176 rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX].ready = false;
177 }
178
179 /**
180 * cayman_dma_resume - setup and start the async dma engines
181 *
182 * @rdev: radeon_device pointer
183 *
184 * Set up the DMA ring buffers and enable them. (cayman-SI).
185 * Returns 0 for success, error for failure.
186 */
cayman_dma_resume(struct radeon_device * rdev)187 int cayman_dma_resume(struct radeon_device *rdev)
188 {
189 struct radeon_ring *ring;
190 u32 rb_cntl, dma_cntl, ib_cntl;
191 u32 rb_bufsz;
192 u32 reg_offset, wb_offset;
193 int i, r;
194
195 for (i = 0; i < 2; i++) {
196 if (i == 0) {
197 ring = &rdev->ring[R600_RING_TYPE_DMA_INDEX];
198 reg_offset = DMA0_REGISTER_OFFSET;
199 wb_offset = R600_WB_DMA_RPTR_OFFSET;
200 } else {
201 ring = &rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX];
202 reg_offset = DMA1_REGISTER_OFFSET;
203 wb_offset = CAYMAN_WB_DMA1_RPTR_OFFSET;
204 }
205
206 WREG32(DMA_SEM_INCOMPLETE_TIMER_CNTL + reg_offset, 0);
207 WREG32(DMA_SEM_WAIT_FAIL_TIMER_CNTL + reg_offset, 0);
208
209 /* Set ring buffer size in dwords */
210 rb_bufsz = order_base_2(ring->ring_size / 4);
211 rb_cntl = rb_bufsz << 1;
212 #ifdef __BIG_ENDIAN
213 rb_cntl |= DMA_RB_SWAP_ENABLE | DMA_RPTR_WRITEBACK_SWAP_ENABLE;
214 #endif
215 WREG32(DMA_RB_CNTL + reg_offset, rb_cntl);
216
217 /* Initialize the ring buffer's read and write pointers */
218 WREG32(DMA_RB_RPTR + reg_offset, 0);
219 WREG32(DMA_RB_WPTR + reg_offset, 0);
220
221 /* set the wb address whether it's enabled or not */
222 WREG32(DMA_RB_RPTR_ADDR_HI + reg_offset,
223 upper_32_bits(rdev->wb.gpu_addr + wb_offset) & 0xFF);
224 WREG32(DMA_RB_RPTR_ADDR_LO + reg_offset,
225 ((rdev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC));
226
227 if (rdev->wb.enabled)
228 rb_cntl |= DMA_RPTR_WRITEBACK_ENABLE;
229
230 WREG32(DMA_RB_BASE + reg_offset, ring->gpu_addr >> 8);
231
232 /* enable DMA IBs */
233 ib_cntl = DMA_IB_ENABLE | CMD_VMID_FORCE;
234 #ifdef __BIG_ENDIAN
235 ib_cntl |= DMA_IB_SWAP_ENABLE;
236 #endif
237 WREG32(DMA_IB_CNTL + reg_offset, ib_cntl);
238
239 dma_cntl = RREG32(DMA_CNTL + reg_offset);
240 dma_cntl &= ~CTXEMPTY_INT_ENABLE;
241 WREG32(DMA_CNTL + reg_offset, dma_cntl);
242
243 ring->wptr = 0;
244 WREG32(DMA_RB_WPTR + reg_offset, ring->wptr << 2);
245
246 WREG32(DMA_RB_CNTL + reg_offset, rb_cntl | DMA_RB_ENABLE);
247
248 ring->ready = true;
249
250 r = radeon_ring_test(rdev, ring->idx, ring);
251 if (r) {
252 ring->ready = false;
253 return r;
254 }
255 }
256
257 if ((rdev->asic->copy.copy_ring_index == R600_RING_TYPE_DMA_INDEX) ||
258 (rdev->asic->copy.copy_ring_index == CAYMAN_RING_TYPE_DMA1_INDEX))
259 radeon_ttm_set_active_vram_size(rdev, rdev->mc.real_vram_size);
260
261 return 0;
262 }
263
264 /**
265 * cayman_dma_fini - tear down the async dma engines
266 *
267 * @rdev: radeon_device pointer
268 *
269 * Stop the async dma engines and free the rings (cayman-SI).
270 */
cayman_dma_fini(struct radeon_device * rdev)271 void cayman_dma_fini(struct radeon_device *rdev)
272 {
273 cayman_dma_stop(rdev);
274 radeon_ring_fini(rdev, &rdev->ring[R600_RING_TYPE_DMA_INDEX]);
275 radeon_ring_fini(rdev, &rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX]);
276 }
277
278 /**
279 * cayman_dma_is_lockup - Check if the DMA engine is locked up
280 *
281 * @rdev: radeon_device pointer
282 * @ring: radeon_ring structure holding ring information
283 *
284 * Check if the async DMA engine is locked up.
285 * Returns true if the engine appears to be locked up, false if not.
286 */
cayman_dma_is_lockup(struct radeon_device * rdev,struct radeon_ring * ring)287 bool cayman_dma_is_lockup(struct radeon_device *rdev, struct radeon_ring *ring)
288 {
289 u32 reset_mask = cayman_gpu_check_soft_reset(rdev);
290 u32 mask;
291
292 if (ring->idx == R600_RING_TYPE_DMA_INDEX)
293 mask = RADEON_RESET_DMA;
294 else
295 mask = RADEON_RESET_DMA1;
296
297 if (!(reset_mask & mask)) {
298 radeon_ring_lockup_update(rdev, ring);
299 return false;
300 }
301 return radeon_ring_test_lockup(rdev, ring);
302 }
303
304 /**
305 * cayman_dma_vm_copy_pages - update PTEs by copying them from the GART
306 *
307 * @rdev: radeon_device pointer
308 * @ib: indirect buffer to fill with commands
309 * @pe: addr of the page entry
310 * @src: src addr where to copy from
311 * @count: number of page entries to update
312 *
313 * Update PTEs by copying them from the GART using the DMA (cayman/TN).
314 */
cayman_dma_vm_copy_pages(struct radeon_device * rdev,struct radeon_ib * ib,uint64_t pe,uint64_t src,unsigned count)315 void cayman_dma_vm_copy_pages(struct radeon_device *rdev,
316 struct radeon_ib *ib,
317 uint64_t pe, uint64_t src,
318 unsigned count)
319 {
320 unsigned ndw;
321
322 while (count) {
323 ndw = count * 2;
324 if (ndw > 0xFFFFE)
325 ndw = 0xFFFFE;
326
327 ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_COPY,
328 0, 0, ndw);
329 ib->ptr[ib->length_dw++] = lower_32_bits(pe);
330 ib->ptr[ib->length_dw++] = lower_32_bits(src);
331 ib->ptr[ib->length_dw++] = upper_32_bits(pe) & 0xff;
332 ib->ptr[ib->length_dw++] = upper_32_bits(src) & 0xff;
333
334 pe += ndw * 4;
335 src += ndw * 4;
336 count -= ndw / 2;
337 }
338 }
339
340 /**
341 * cayman_dma_vm_write_pages - update PTEs by writing them manually
342 *
343 * @rdev: radeon_device pointer
344 * @ib: indirect buffer to fill with commands
345 * @pe: addr of the page entry
346 * @addr: dst addr to write into pe
347 * @count: number of page entries to update
348 * @incr: increase next addr by incr bytes
349 * @flags: hw access flags
350 *
351 * Update PTEs by writing them manually using the DMA (cayman/TN).
352 */
cayman_dma_vm_write_pages(struct radeon_device * rdev,struct radeon_ib * ib,uint64_t pe,uint64_t addr,unsigned count,uint32_t incr,uint32_t flags)353 void cayman_dma_vm_write_pages(struct radeon_device *rdev,
354 struct radeon_ib *ib,
355 uint64_t pe,
356 uint64_t addr, unsigned count,
357 uint32_t incr, uint32_t flags)
358 {
359 uint64_t value;
360 unsigned ndw;
361
362 while (count) {
363 ndw = count * 2;
364 if (ndw > 0xFFFFE)
365 ndw = 0xFFFFE;
366
367 /* for non-physically contiguous pages (system) */
368 ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_WRITE,
369 0, 0, ndw);
370 ib->ptr[ib->length_dw++] = pe;
371 ib->ptr[ib->length_dw++] = upper_32_bits(pe) & 0xff;
372 for (; ndw > 0; ndw -= 2, --count, pe += 8) {
373 if (flags & R600_PTE_SYSTEM) {
374 value = radeon_vm_map_gart(rdev, addr);
375 } else if (flags & R600_PTE_VALID) {
376 value = addr;
377 } else {
378 value = 0;
379 }
380 addr += incr;
381 value |= flags;
382 ib->ptr[ib->length_dw++] = value;
383 ib->ptr[ib->length_dw++] = upper_32_bits(value);
384 }
385 }
386 }
387
388 /**
389 * cayman_dma_vm_set_pages - update the page tables using the DMA
390 *
391 * @rdev: radeon_device pointer
392 * @ib: indirect buffer to fill with commands
393 * @pe: addr of the page entry
394 * @addr: dst addr to write into pe
395 * @count: number of page entries to update
396 * @incr: increase next addr by incr bytes
397 * @flags: hw access flags
398 *
399 * Update the page tables using the DMA (cayman/TN).
400 */
cayman_dma_vm_set_pages(struct radeon_device * rdev,struct radeon_ib * ib,uint64_t pe,uint64_t addr,unsigned count,uint32_t incr,uint32_t flags)401 void cayman_dma_vm_set_pages(struct radeon_device *rdev,
402 struct radeon_ib *ib,
403 uint64_t pe,
404 uint64_t addr, unsigned count,
405 uint32_t incr, uint32_t flags)
406 {
407 uint64_t value;
408 unsigned ndw;
409
410 while (count) {
411 ndw = count * 2;
412 if (ndw > 0xFFFFE)
413 ndw = 0xFFFFE;
414
415 if (flags & R600_PTE_VALID)
416 value = addr;
417 else
418 value = 0;
419
420 /* for physically contiguous pages (vram) */
421 ib->ptr[ib->length_dw++] = DMA_PTE_PDE_PACKET(ndw);
422 ib->ptr[ib->length_dw++] = pe; /* dst addr */
423 ib->ptr[ib->length_dw++] = upper_32_bits(pe) & 0xff;
424 ib->ptr[ib->length_dw++] = flags; /* mask */
425 ib->ptr[ib->length_dw++] = 0;
426 ib->ptr[ib->length_dw++] = value; /* value */
427 ib->ptr[ib->length_dw++] = upper_32_bits(value);
428 ib->ptr[ib->length_dw++] = incr; /* increment size */
429 ib->ptr[ib->length_dw++] = 0;
430
431 pe += ndw * 4;
432 addr += (ndw / 2) * incr;
433 count -= ndw / 2;
434 }
435 }
436
437 /**
438 * cayman_dma_vm_pad_ib - pad the IB to the required number of dw
439 *
440 * @ib: indirect buffer to fill with padding
441 *
442 */
cayman_dma_vm_pad_ib(struct radeon_ib * ib)443 void cayman_dma_vm_pad_ib(struct radeon_ib *ib)
444 {
445 while (ib->length_dw & 0x7)
446 ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0);
447 }
448
cayman_dma_vm_flush(struct radeon_device * rdev,struct radeon_ring * ring,unsigned vm_id,uint64_t pd_addr)449 void cayman_dma_vm_flush(struct radeon_device *rdev, struct radeon_ring *ring,
450 unsigned vm_id, uint64_t pd_addr)
451 {
452 radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0, 0));
453 radeon_ring_write(ring, (0xf << 16) | ((VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (vm_id << 2)) >> 2));
454 radeon_ring_write(ring, pd_addr >> 12);
455
456 /* flush hdp cache */
457 radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0, 0));
458 radeon_ring_write(ring, (0xf << 16) | (HDP_MEM_COHERENCY_FLUSH_CNTL >> 2));
459 radeon_ring_write(ring, 1);
460
461 /* bits 0-7 are the VM contexts0-7 */
462 radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0, 0));
463 radeon_ring_write(ring, (0xf << 16) | (VM_INVALIDATE_REQUEST >> 2));
464 radeon_ring_write(ring, 1 << vm_id);
465
466 /* wait for invalidate to complete */
467 radeon_ring_write(ring, DMA_SRBM_READ_PACKET);
468 radeon_ring_write(ring, (0xff << 20) | (VM_INVALIDATE_REQUEST >> 2));
469 radeon_ring_write(ring, 0); /* mask */
470 radeon_ring_write(ring, 0); /* value */
471 }
472
473