xref: /openbmc/linux/drivers/gpu/drm/radeon/ni_dma.c (revision 275876e2)
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 #include <drm/drmP.h>
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  */
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  */
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  */
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  */
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 
127 	if (rdev->wb.enabled) {
128 		u32 next_rptr = ring->wptr + 4;
129 		while ((next_rptr & 7) != 5)
130 			next_rptr++;
131 		next_rptr += 3;
132 		radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 1));
133 		radeon_ring_write(ring, ring->next_rptr_gpu_addr & 0xfffffffc);
134 		radeon_ring_write(ring, upper_32_bits(ring->next_rptr_gpu_addr) & 0xff);
135 		radeon_ring_write(ring, next_rptr);
136 	}
137 
138 	/* The indirect buffer packet must end on an 8 DW boundary in the DMA ring.
139 	 * Pad as necessary with NOPs.
140 	 */
141 	while ((ring->wptr & 7) != 5)
142 		radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0));
143 	radeon_ring_write(ring, DMA_IB_PACKET(DMA_PACKET_INDIRECT_BUFFER, ib->vm ? ib->vm->id : 0, 0));
144 	radeon_ring_write(ring, (ib->gpu_addr & 0xFFFFFFE0));
145 	radeon_ring_write(ring, (ib->length_dw << 12) | (upper_32_bits(ib->gpu_addr) & 0xFF));
146 
147 }
148 
149 /**
150  * cayman_dma_stop - stop the async dma engines
151  *
152  * @rdev: radeon_device pointer
153  *
154  * Stop the async dma engines (cayman-SI).
155  */
156 void cayman_dma_stop(struct radeon_device *rdev)
157 {
158 	u32 rb_cntl;
159 
160 	if ((rdev->asic->copy.copy_ring_index == R600_RING_TYPE_DMA_INDEX) ||
161 	    (rdev->asic->copy.copy_ring_index == CAYMAN_RING_TYPE_DMA1_INDEX))
162 		radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
163 
164 	/* dma0 */
165 	rb_cntl = RREG32(DMA_RB_CNTL + DMA0_REGISTER_OFFSET);
166 	rb_cntl &= ~DMA_RB_ENABLE;
167 	WREG32(DMA_RB_CNTL + DMA0_REGISTER_OFFSET, rb_cntl);
168 
169 	/* dma1 */
170 	rb_cntl = RREG32(DMA_RB_CNTL + DMA1_REGISTER_OFFSET);
171 	rb_cntl &= ~DMA_RB_ENABLE;
172 	WREG32(DMA_RB_CNTL + DMA1_REGISTER_OFFSET, rb_cntl);
173 
174 	rdev->ring[R600_RING_TYPE_DMA_INDEX].ready = false;
175 	rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX].ready = false;
176 }
177 
178 /**
179  * cayman_dma_resume - setup and start the async dma engines
180  *
181  * @rdev: radeon_device pointer
182  *
183  * Set up the DMA ring buffers and enable them. (cayman-SI).
184  * Returns 0 for success, error for failure.
185  */
186 int cayman_dma_resume(struct radeon_device *rdev)
187 {
188 	struct radeon_ring *ring;
189 	u32 rb_cntl, dma_cntl, ib_cntl;
190 	u32 rb_bufsz;
191 	u32 reg_offset, wb_offset;
192 	int i, r;
193 
194 	/* Reset dma */
195 	WREG32(SRBM_SOFT_RESET, SOFT_RESET_DMA | SOFT_RESET_DMA1);
196 	RREG32(SRBM_SOFT_RESET);
197 	udelay(50);
198 	WREG32(SRBM_SOFT_RESET, 0);
199 
200 	for (i = 0; i < 2; i++) {
201 		if (i == 0) {
202 			ring = &rdev->ring[R600_RING_TYPE_DMA_INDEX];
203 			reg_offset = DMA0_REGISTER_OFFSET;
204 			wb_offset = R600_WB_DMA_RPTR_OFFSET;
205 		} else {
206 			ring = &rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX];
207 			reg_offset = DMA1_REGISTER_OFFSET;
208 			wb_offset = CAYMAN_WB_DMA1_RPTR_OFFSET;
209 		}
210 
211 		WREG32(DMA_SEM_INCOMPLETE_TIMER_CNTL + reg_offset, 0);
212 		WREG32(DMA_SEM_WAIT_FAIL_TIMER_CNTL + reg_offset, 0);
213 
214 		/* Set ring buffer size in dwords */
215 		rb_bufsz = order_base_2(ring->ring_size / 4);
216 		rb_cntl = rb_bufsz << 1;
217 #ifdef __BIG_ENDIAN
218 		rb_cntl |= DMA_RB_SWAP_ENABLE | DMA_RPTR_WRITEBACK_SWAP_ENABLE;
219 #endif
220 		WREG32(DMA_RB_CNTL + reg_offset, rb_cntl);
221 
222 		/* Initialize the ring buffer's read and write pointers */
223 		WREG32(DMA_RB_RPTR + reg_offset, 0);
224 		WREG32(DMA_RB_WPTR + reg_offset, 0);
225 
226 		/* set the wb address whether it's enabled or not */
227 		WREG32(DMA_RB_RPTR_ADDR_HI + reg_offset,
228 		       upper_32_bits(rdev->wb.gpu_addr + wb_offset) & 0xFF);
229 		WREG32(DMA_RB_RPTR_ADDR_LO + reg_offset,
230 		       ((rdev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC));
231 
232 		if (rdev->wb.enabled)
233 			rb_cntl |= DMA_RPTR_WRITEBACK_ENABLE;
234 
235 		WREG32(DMA_RB_BASE + reg_offset, ring->gpu_addr >> 8);
236 
237 		/* enable DMA IBs */
238 		ib_cntl = DMA_IB_ENABLE | CMD_VMID_FORCE;
239 #ifdef __BIG_ENDIAN
240 		ib_cntl |= DMA_IB_SWAP_ENABLE;
241 #endif
242 		WREG32(DMA_IB_CNTL + reg_offset, ib_cntl);
243 
244 		dma_cntl = RREG32(DMA_CNTL + reg_offset);
245 		dma_cntl &= ~CTXEMPTY_INT_ENABLE;
246 		WREG32(DMA_CNTL + reg_offset, dma_cntl);
247 
248 		ring->wptr = 0;
249 		WREG32(DMA_RB_WPTR + reg_offset, ring->wptr << 2);
250 
251 		WREG32(DMA_RB_CNTL + reg_offset, rb_cntl | DMA_RB_ENABLE);
252 
253 		ring->ready = true;
254 
255 		r = radeon_ring_test(rdev, ring->idx, ring);
256 		if (r) {
257 			ring->ready = false;
258 			return r;
259 		}
260 	}
261 
262 	if ((rdev->asic->copy.copy_ring_index == R600_RING_TYPE_DMA_INDEX) ||
263 	    (rdev->asic->copy.copy_ring_index == CAYMAN_RING_TYPE_DMA1_INDEX))
264 		radeon_ttm_set_active_vram_size(rdev, rdev->mc.real_vram_size);
265 
266 	return 0;
267 }
268 
269 /**
270  * cayman_dma_fini - tear down the async dma engines
271  *
272  * @rdev: radeon_device pointer
273  *
274  * Stop the async dma engines and free the rings (cayman-SI).
275  */
276 void cayman_dma_fini(struct radeon_device *rdev)
277 {
278 	cayman_dma_stop(rdev);
279 	radeon_ring_fini(rdev, &rdev->ring[R600_RING_TYPE_DMA_INDEX]);
280 	radeon_ring_fini(rdev, &rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX]);
281 }
282 
283 /**
284  * cayman_dma_is_lockup - Check if the DMA engine is locked up
285  *
286  * @rdev: radeon_device pointer
287  * @ring: radeon_ring structure holding ring information
288  *
289  * Check if the async DMA engine is locked up.
290  * Returns true if the engine appears to be locked up, false if not.
291  */
292 bool cayman_dma_is_lockup(struct radeon_device *rdev, struct radeon_ring *ring)
293 {
294 	u32 reset_mask = cayman_gpu_check_soft_reset(rdev);
295 	u32 mask;
296 
297 	if (ring->idx == R600_RING_TYPE_DMA_INDEX)
298 		mask = RADEON_RESET_DMA;
299 	else
300 		mask = RADEON_RESET_DMA1;
301 
302 	if (!(reset_mask & mask)) {
303 		radeon_ring_lockup_update(rdev, ring);
304 		return false;
305 	}
306 	return radeon_ring_test_lockup(rdev, ring);
307 }
308 
309 /**
310  * cayman_dma_vm_copy_pages - update PTEs by copying them from the GART
311  *
312  * @rdev: radeon_device pointer
313  * @ib: indirect buffer to fill with commands
314  * @pe: addr of the page entry
315  * @src: src addr where to copy from
316  * @count: number of page entries to update
317  *
318  * Update PTEs by copying them from the GART using the DMA (cayman/TN).
319  */
320 void cayman_dma_vm_copy_pages(struct radeon_device *rdev,
321 			      struct radeon_ib *ib,
322 			      uint64_t pe, uint64_t src,
323 			      unsigned count)
324 {
325 	unsigned ndw;
326 
327 	while (count) {
328 		ndw = count * 2;
329 		if (ndw > 0xFFFFE)
330 			ndw = 0xFFFFE;
331 
332 		ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_COPY,
333 						      0, 0, ndw);
334 		ib->ptr[ib->length_dw++] = lower_32_bits(pe);
335 		ib->ptr[ib->length_dw++] = lower_32_bits(src);
336 		ib->ptr[ib->length_dw++] = upper_32_bits(pe) & 0xff;
337 		ib->ptr[ib->length_dw++] = upper_32_bits(src) & 0xff;
338 
339 		pe += ndw * 4;
340 		src += ndw * 4;
341 		count -= ndw / 2;
342 	}
343 }
344 
345 /**
346  * cayman_dma_vm_write_pages - update PTEs by writing them manually
347  *
348  * @rdev: radeon_device pointer
349  * @ib: indirect buffer to fill with commands
350  * @pe: addr of the page entry
351  * @addr: dst addr to write into pe
352  * @count: number of page entries to update
353  * @incr: increase next addr by incr bytes
354  * @flags: hw access flags
355  *
356  * Update PTEs by writing them manually using the DMA (cayman/TN).
357  */
358 void cayman_dma_vm_write_pages(struct radeon_device *rdev,
359 			       struct radeon_ib *ib,
360 			       uint64_t pe,
361 			       uint64_t addr, unsigned count,
362 			       uint32_t incr, uint32_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 		/* for non-physically contiguous pages (system) */
373 		ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_WRITE,
374 						      0, 0, ndw);
375 		ib->ptr[ib->length_dw++] = pe;
376 		ib->ptr[ib->length_dw++] = upper_32_bits(pe) & 0xff;
377 		for (; ndw > 0; ndw -= 2, --count, pe += 8) {
378 			if (flags & R600_PTE_SYSTEM) {
379 				value = radeon_vm_map_gart(rdev, addr);
380 				value &= 0xFFFFFFFFFFFFF000ULL;
381 			} else if (flags & R600_PTE_VALID) {
382 				value = addr;
383 			} else {
384 				value = 0;
385 			}
386 			addr += incr;
387 			value |= flags;
388 			ib->ptr[ib->length_dw++] = value;
389 			ib->ptr[ib->length_dw++] = upper_32_bits(value);
390 		}
391 	}
392 }
393 
394 /**
395  * cayman_dma_vm_set_pages - update the page tables using the DMA
396  *
397  * @rdev: radeon_device pointer
398  * @ib: indirect buffer to fill with commands
399  * @pe: addr of the page entry
400  * @addr: dst addr to write into pe
401  * @count: number of page entries to update
402  * @incr: increase next addr by incr bytes
403  * @flags: hw access flags
404  *
405  * Update the page tables using the DMA (cayman/TN).
406  */
407 void cayman_dma_vm_set_pages(struct radeon_device *rdev,
408 			     struct radeon_ib *ib,
409 			     uint64_t pe,
410 			     uint64_t addr, unsigned count,
411 			     uint32_t incr, uint32_t flags)
412 {
413 	uint64_t value;
414 	unsigned ndw;
415 
416 	while (count) {
417 		ndw = count * 2;
418 		if (ndw > 0xFFFFE)
419 			ndw = 0xFFFFE;
420 
421 		if (flags & R600_PTE_VALID)
422 			value = addr;
423 		else
424 			value = 0;
425 
426 		/* for physically contiguous pages (vram) */
427 		ib->ptr[ib->length_dw++] = DMA_PTE_PDE_PACKET(ndw);
428 		ib->ptr[ib->length_dw++] = pe; /* dst addr */
429 		ib->ptr[ib->length_dw++] = upper_32_bits(pe) & 0xff;
430 		ib->ptr[ib->length_dw++] = flags; /* mask */
431 		ib->ptr[ib->length_dw++] = 0;
432 		ib->ptr[ib->length_dw++] = value; /* value */
433 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
434 		ib->ptr[ib->length_dw++] = incr; /* increment size */
435 		ib->ptr[ib->length_dw++] = 0;
436 
437 		pe += ndw * 4;
438 		addr += (ndw / 2) * incr;
439 		count -= ndw / 2;
440 	}
441 }
442 
443 /**
444  * cayman_dma_vm_pad_ib - pad the IB to the required number of dw
445  *
446  * @ib: indirect buffer to fill with padding
447  *
448  */
449 void cayman_dma_vm_pad_ib(struct radeon_ib *ib)
450 {
451 	while (ib->length_dw & 0x7)
452 		ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0);
453 }
454 
455 void cayman_dma_vm_flush(struct radeon_device *rdev, int ridx, struct radeon_vm *vm)
456 {
457 	struct radeon_ring *ring = &rdev->ring[ridx];
458 
459 	if (vm == NULL)
460 		return;
461 
462 	radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0, 0));
463 	radeon_ring_write(ring, (0xf << 16) | ((VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (vm->id << 2)) >> 2));
464 	radeon_ring_write(ring, vm->pd_gpu_addr >> 12);
465 
466 	/* flush hdp cache */
467 	radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0, 0));
468 	radeon_ring_write(ring, (0xf << 16) | (HDP_MEM_COHERENCY_FLUSH_CNTL >> 2));
469 	radeon_ring_write(ring, 1);
470 
471 	/* bits 0-7 are the VM contexts0-7 */
472 	radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0, 0));
473 	radeon_ring_write(ring, (0xf << 16) | (VM_INVALIDATE_REQUEST >> 2));
474 	radeon_ring_write(ring, 1 << vm->id);
475 }
476 
477