xref: /openbmc/linux/drivers/gpu/drm/radeon/si_dma.c (revision 1c2f87c2)
1 /*
2  * Copyright 2013 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 "sid.h"
29 
30 u32 si_gpu_check_soft_reset(struct radeon_device *rdev);
31 
32 /**
33  * si_dma_is_lockup - Check if the DMA engine is locked up
34  *
35  * @rdev: radeon_device pointer
36  * @ring: radeon_ring structure holding ring information
37  *
38  * Check if the async DMA engine is locked up.
39  * Returns true if the engine appears to be locked up, false if not.
40  */
41 bool si_dma_is_lockup(struct radeon_device *rdev, struct radeon_ring *ring)
42 {
43 	u32 reset_mask = si_gpu_check_soft_reset(rdev);
44 	u32 mask;
45 
46 	if (ring->idx == R600_RING_TYPE_DMA_INDEX)
47 		mask = RADEON_RESET_DMA;
48 	else
49 		mask = RADEON_RESET_DMA1;
50 
51 	if (!(reset_mask & mask)) {
52 		radeon_ring_lockup_update(rdev, ring);
53 		return false;
54 	}
55 	return radeon_ring_test_lockup(rdev, ring);
56 }
57 
58 /**
59  * si_dma_vm_set_page - update the page tables using the DMA
60  *
61  * @rdev: radeon_device pointer
62  * @ib: indirect buffer to fill with commands
63  * @pe: addr of the page entry
64  * @addr: dst addr to write into pe
65  * @count: number of page entries to update
66  * @incr: increase next addr by incr bytes
67  * @flags: access flags
68  *
69  * Update the page tables using the DMA (SI).
70  */
71 void si_dma_vm_set_page(struct radeon_device *rdev,
72 			struct radeon_ib *ib,
73 			uint64_t pe,
74 			uint64_t addr, unsigned count,
75 			uint32_t incr, uint32_t flags)
76 {
77 	uint64_t value;
78 	unsigned ndw;
79 
80 	trace_radeon_vm_set_page(pe, addr, count, incr, flags);
81 
82 	if (flags & R600_PTE_SYSTEM) {
83 		while (count) {
84 			ndw = count * 2;
85 			if (ndw > 0xFFFFE)
86 				ndw = 0xFFFFE;
87 
88 			/* for non-physically contiguous pages (system) */
89 			ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 0, ndw);
90 			ib->ptr[ib->length_dw++] = pe;
91 			ib->ptr[ib->length_dw++] = upper_32_bits(pe) & 0xff;
92 			for (; ndw > 0; ndw -= 2, --count, pe += 8) {
93 				value = radeon_vm_map_gart(rdev, addr);
94 				value &= 0xFFFFFFFFFFFFF000ULL;
95 				addr += incr;
96 				value |= flags;
97 				ib->ptr[ib->length_dw++] = value;
98 				ib->ptr[ib->length_dw++] = upper_32_bits(value);
99 			}
100 		}
101 	} else {
102 		while (count) {
103 			ndw = count * 2;
104 			if (ndw > 0xFFFFE)
105 				ndw = 0xFFFFE;
106 
107 			if (flags & R600_PTE_VALID)
108 				value = addr;
109 			else
110 				value = 0;
111 			/* for physically contiguous pages (vram) */
112 			ib->ptr[ib->length_dw++] = DMA_PTE_PDE_PACKET(ndw);
113 			ib->ptr[ib->length_dw++] = pe; /* dst addr */
114 			ib->ptr[ib->length_dw++] = upper_32_bits(pe) & 0xff;
115 			ib->ptr[ib->length_dw++] = flags; /* mask */
116 			ib->ptr[ib->length_dw++] = 0;
117 			ib->ptr[ib->length_dw++] = value; /* value */
118 			ib->ptr[ib->length_dw++] = upper_32_bits(value);
119 			ib->ptr[ib->length_dw++] = incr; /* increment size */
120 			ib->ptr[ib->length_dw++] = 0;
121 			pe += ndw * 4;
122 			addr += (ndw / 2) * incr;
123 			count -= ndw / 2;
124 		}
125 	}
126 	while (ib->length_dw & 0x7)
127 		ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0);
128 }
129 
130 void si_dma_vm_flush(struct radeon_device *rdev, int ridx, struct radeon_vm *vm)
131 {
132 	struct radeon_ring *ring = &rdev->ring[ridx];
133 
134 	if (vm == NULL)
135 		return;
136 
137 	radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0, 0, 0));
138 	if (vm->id < 8) {
139 		radeon_ring_write(ring, (0xf << 16) | ((VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (vm->id << 2)) >> 2));
140 	} else {
141 		radeon_ring_write(ring, (0xf << 16) | ((VM_CONTEXT8_PAGE_TABLE_BASE_ADDR + ((vm->id - 8) << 2)) >> 2));
142 	}
143 	radeon_ring_write(ring, vm->pd_gpu_addr >> 12);
144 
145 	/* flush hdp cache */
146 	radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0, 0, 0));
147 	radeon_ring_write(ring, (0xf << 16) | (HDP_MEM_COHERENCY_FLUSH_CNTL >> 2));
148 	radeon_ring_write(ring, 1);
149 
150 	/* bits 0-7 are the VM contexts0-7 */
151 	radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0, 0, 0));
152 	radeon_ring_write(ring, (0xf << 16) | (VM_INVALIDATE_REQUEST >> 2));
153 	radeon_ring_write(ring, 1 << vm->id);
154 }
155 
156 /**
157  * si_copy_dma - copy pages using the DMA engine
158  *
159  * @rdev: radeon_device pointer
160  * @src_offset: src GPU address
161  * @dst_offset: dst GPU address
162  * @num_gpu_pages: number of GPU pages to xfer
163  * @fence: radeon fence object
164  *
165  * Copy GPU paging using the DMA engine (SI).
166  * Used by the radeon ttm implementation to move pages if
167  * registered as the asic copy callback.
168  */
169 int si_copy_dma(struct radeon_device *rdev,
170 		uint64_t src_offset, uint64_t dst_offset,
171 		unsigned num_gpu_pages,
172 		struct radeon_fence **fence)
173 {
174 	struct radeon_semaphore *sem = NULL;
175 	int ring_index = rdev->asic->copy.dma_ring_index;
176 	struct radeon_ring *ring = &rdev->ring[ring_index];
177 	u32 size_in_bytes, cur_size_in_bytes;
178 	int i, num_loops;
179 	int r = 0;
180 
181 	r = radeon_semaphore_create(rdev, &sem);
182 	if (r) {
183 		DRM_ERROR("radeon: moving bo (%d).\n", r);
184 		return r;
185 	}
186 
187 	size_in_bytes = (num_gpu_pages << RADEON_GPU_PAGE_SHIFT);
188 	num_loops = DIV_ROUND_UP(size_in_bytes, 0xfffff);
189 	r = radeon_ring_lock(rdev, ring, num_loops * 5 + 11);
190 	if (r) {
191 		DRM_ERROR("radeon: moving bo (%d).\n", r);
192 		radeon_semaphore_free(rdev, &sem, NULL);
193 		return r;
194 	}
195 
196 	radeon_semaphore_sync_to(sem, *fence);
197 	radeon_semaphore_sync_rings(rdev, sem, ring->idx);
198 
199 	for (i = 0; i < num_loops; i++) {
200 		cur_size_in_bytes = size_in_bytes;
201 		if (cur_size_in_bytes > 0xFFFFF)
202 			cur_size_in_bytes = 0xFFFFF;
203 		size_in_bytes -= cur_size_in_bytes;
204 		radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_COPY, 1, 0, 0, cur_size_in_bytes));
205 		radeon_ring_write(ring, dst_offset & 0xffffffff);
206 		radeon_ring_write(ring, src_offset & 0xffffffff);
207 		radeon_ring_write(ring, upper_32_bits(dst_offset) & 0xff);
208 		radeon_ring_write(ring, upper_32_bits(src_offset) & 0xff);
209 		src_offset += cur_size_in_bytes;
210 		dst_offset += cur_size_in_bytes;
211 	}
212 
213 	r = radeon_fence_emit(rdev, fence, ring->idx);
214 	if (r) {
215 		radeon_ring_unlock_undo(rdev, ring);
216 		return r;
217 	}
218 
219 	radeon_ring_unlock_commit(rdev, ring);
220 	radeon_semaphore_free(rdev, &sem, *fence);
221 
222 	return r;
223 }
224 
225