1 | // SPDX-License-Identifier: MIT |
2 | /* |
3 | * Copyright © 2010 Daniel Vetter |
4 | * Copyright © 2020 Intel Corporation |
5 | */ |
6 | |
7 | #include <linux/slab.h> /* fault-inject.h is not standalone! */ |
8 | |
9 | #include <linux/fault-inject.h> |
10 | #include <linux/log2.h> |
11 | #include <linux/random.h> |
12 | #include <linux/seq_file.h> |
13 | #include <linux/stop_machine.h> |
14 | |
15 | #include <asm/set_memory.h> |
16 | #include <asm/smp.h> |
17 | |
18 | #include "display/intel_frontbuffer.h" |
19 | #include "gt/intel_gt.h" |
20 | #include "gt/intel_gt_requests.h" |
21 | |
22 | #include "i915_drv.h" |
23 | #include "i915_gem_evict.h" |
24 | #include "i915_scatterlist.h" |
25 | #include "i915_trace.h" |
26 | #include "i915_vgpu.h" |
27 | |
28 | int i915_gem_gtt_prepare_pages(struct drm_i915_gem_object *obj, |
29 | struct sg_table *pages) |
30 | { |
31 | do { |
32 | if (dma_map_sg_attrs(dev: obj->base.dev->dev, |
33 | sg: pages->sgl, nents: pages->nents, |
34 | dir: DMA_BIDIRECTIONAL, |
35 | DMA_ATTR_SKIP_CPU_SYNC | |
36 | DMA_ATTR_NO_KERNEL_MAPPING | |
37 | DMA_ATTR_NO_WARN)) |
38 | return 0; |
39 | |
40 | /* |
41 | * If the DMA remap fails, one cause can be that we have |
42 | * too many objects pinned in a small remapping table, |
43 | * such as swiotlb. Incrementally purge all other objects and |
44 | * try again - if there are no more pages to remove from |
45 | * the DMA remapper, i915_gem_shrink will return 0. |
46 | */ |
47 | GEM_BUG_ON(obj->mm.pages == pages); |
48 | } while (i915_gem_shrink(NULL, i915: to_i915(dev: obj->base.dev), |
49 | target: obj->base.size >> PAGE_SHIFT, NULL, |
50 | I915_SHRINK_BOUND | |
51 | I915_SHRINK_UNBOUND)); |
52 | |
53 | return -ENOSPC; |
54 | } |
55 | |
56 | void i915_gem_gtt_finish_pages(struct drm_i915_gem_object *obj, |
57 | struct sg_table *pages) |
58 | { |
59 | struct drm_i915_private *i915 = to_i915(dev: obj->base.dev); |
60 | struct i915_ggtt *ggtt = to_gt(i915)->ggtt; |
61 | |
62 | /* XXX This does not prevent more requests being submitted! */ |
63 | if (unlikely(ggtt->do_idle_maps)) |
64 | /* Wait a bit, in the hope it avoids the hang */ |
65 | usleep_range(min: 100, max: 250); |
66 | |
67 | dma_unmap_sg(i915->drm.dev, pages->sgl, pages->nents, |
68 | DMA_BIDIRECTIONAL); |
69 | } |
70 | |
71 | /** |
72 | * i915_gem_gtt_reserve - reserve a node in an address_space (GTT) |
73 | * @vm: the &struct i915_address_space |
74 | * @ww: An optional struct i915_gem_ww_ctx. |
75 | * @node: the &struct drm_mm_node (typically i915_vma.mode) |
76 | * @size: how much space to allocate inside the GTT, |
77 | * must be #I915_GTT_PAGE_SIZE aligned |
78 | * @offset: where to insert inside the GTT, |
79 | * must be #I915_GTT_MIN_ALIGNMENT aligned, and the node |
80 | * (@offset + @size) must fit within the address space |
81 | * @color: color to apply to node, if this node is not from a VMA, |
82 | * color must be #I915_COLOR_UNEVICTABLE |
83 | * @flags: control search and eviction behaviour |
84 | * |
85 | * i915_gem_gtt_reserve() tries to insert the @node at the exact @offset inside |
86 | * the address space (using @size and @color). If the @node does not fit, it |
87 | * tries to evict any overlapping nodes from the GTT, including any |
88 | * neighbouring nodes if the colors do not match (to ensure guard pages between |
89 | * differing domains). See i915_gem_evict_for_node() for the gory details |
90 | * on the eviction algorithm. #PIN_NONBLOCK may used to prevent waiting on |
91 | * evicting active overlapping objects, and any overlapping node that is pinned |
92 | * or marked as unevictable will also result in failure. |
93 | * |
94 | * Returns: 0 on success, -ENOSPC if no suitable hole is found, -EINTR if |
95 | * asked to wait for eviction and interrupted. |
96 | */ |
97 | int i915_gem_gtt_reserve(struct i915_address_space *vm, |
98 | struct i915_gem_ww_ctx *ww, |
99 | struct drm_mm_node *node, |
100 | u64 size, u64 offset, unsigned long color, |
101 | unsigned int flags) |
102 | { |
103 | int err; |
104 | |
105 | GEM_BUG_ON(!size); |
106 | GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE)); |
107 | GEM_BUG_ON(!IS_ALIGNED(offset, I915_GTT_MIN_ALIGNMENT)); |
108 | GEM_BUG_ON(range_overflows(offset, size, vm->total)); |
109 | GEM_BUG_ON(vm == &to_gt(vm->i915)->ggtt->alias->vm); |
110 | GEM_BUG_ON(drm_mm_node_allocated(node)); |
111 | |
112 | node->size = size; |
113 | node->start = offset; |
114 | node->color = color; |
115 | |
116 | err = drm_mm_reserve_node(mm: &vm->mm, node); |
117 | if (err != -ENOSPC) |
118 | return err; |
119 | |
120 | if (flags & PIN_NOEVICT) |
121 | return -ENOSPC; |
122 | |
123 | err = i915_gem_evict_for_node(vm, ww, node, flags); |
124 | if (err == 0) |
125 | err = drm_mm_reserve_node(mm: &vm->mm, node); |
126 | |
127 | return err; |
128 | } |
129 | |
130 | static u64 random_offset(u64 start, u64 end, u64 len, u64 align) |
131 | { |
132 | u64 range, addr; |
133 | |
134 | GEM_BUG_ON(range_overflows(start, len, end)); |
135 | GEM_BUG_ON(round_up(start, align) > round_down(end - len, align)); |
136 | |
137 | range = round_down(end - len, align) - round_up(start, align); |
138 | if (range) { |
139 | if (sizeof(unsigned long) == sizeof(u64)) { |
140 | addr = get_random_u64(); |
141 | } else { |
142 | addr = get_random_u32(); |
143 | if (range > U32_MAX) { |
144 | addr <<= 32; |
145 | addr |= get_random_u32(); |
146 | } |
147 | } |
148 | div64_u64_rem(dividend: addr, divisor: range, remainder: &addr); |
149 | start += addr; |
150 | } |
151 | |
152 | return round_up(start, align); |
153 | } |
154 | |
155 | /** |
156 | * i915_gem_gtt_insert - insert a node into an address_space (GTT) |
157 | * @vm: the &struct i915_address_space |
158 | * @ww: An optional struct i915_gem_ww_ctx. |
159 | * @node: the &struct drm_mm_node (typically i915_vma.node) |
160 | * @size: how much space to allocate inside the GTT, |
161 | * must be #I915_GTT_PAGE_SIZE aligned |
162 | * @alignment: required alignment of starting offset, may be 0 but |
163 | * if specified, this must be a power-of-two and at least |
164 | * #I915_GTT_MIN_ALIGNMENT |
165 | * @color: color to apply to node |
166 | * @start: start of any range restriction inside GTT (0 for all), |
167 | * must be #I915_GTT_PAGE_SIZE aligned |
168 | * @end: end of any range restriction inside GTT (U64_MAX for all), |
169 | * must be #I915_GTT_PAGE_SIZE aligned if not U64_MAX |
170 | * @flags: control search and eviction behaviour |
171 | * |
172 | * i915_gem_gtt_insert() first searches for an available hole into which |
173 | * is can insert the node. The hole address is aligned to @alignment and |
174 | * its @size must then fit entirely within the [@start, @end] bounds. The |
175 | * nodes on either side of the hole must match @color, or else a guard page |
176 | * will be inserted between the two nodes (or the node evicted). If no |
177 | * suitable hole is found, first a victim is randomly selected and tested |
178 | * for eviction, otherwise then the LRU list of objects within the GTT |
179 | * is scanned to find the first set of replacement nodes to create the hole. |
180 | * Those old overlapping nodes are evicted from the GTT (and so must be |
181 | * rebound before any future use). Any node that is currently pinned cannot |
182 | * be evicted (see i915_vma_pin()). Similar if the node's VMA is currently |
183 | * active and #PIN_NONBLOCK is specified, that node is also skipped when |
184 | * searching for an eviction candidate. See i915_gem_evict_something() for |
185 | * the gory details on the eviction algorithm. |
186 | * |
187 | * Returns: 0 on success, -ENOSPC if no suitable hole is found, -EINTR if |
188 | * asked to wait for eviction and interrupted. |
189 | */ |
190 | int i915_gem_gtt_insert(struct i915_address_space *vm, |
191 | struct i915_gem_ww_ctx *ww, |
192 | struct drm_mm_node *node, |
193 | u64 size, u64 alignment, unsigned long color, |
194 | u64 start, u64 end, unsigned int flags) |
195 | { |
196 | enum drm_mm_insert_mode mode; |
197 | u64 offset; |
198 | int err; |
199 | |
200 | lockdep_assert_held(&vm->mutex); |
201 | |
202 | GEM_BUG_ON(!size); |
203 | GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE)); |
204 | GEM_BUG_ON(alignment && !is_power_of_2(alignment)); |
205 | GEM_BUG_ON(alignment && !IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT)); |
206 | GEM_BUG_ON(start >= end); |
207 | GEM_BUG_ON(start > 0 && !IS_ALIGNED(start, I915_GTT_PAGE_SIZE)); |
208 | GEM_BUG_ON(end < U64_MAX && !IS_ALIGNED(end, I915_GTT_PAGE_SIZE)); |
209 | GEM_BUG_ON(vm == &to_gt(vm->i915)->ggtt->alias->vm); |
210 | GEM_BUG_ON(drm_mm_node_allocated(node)); |
211 | |
212 | if (unlikely(range_overflows(start, size, end))) |
213 | return -ENOSPC; |
214 | |
215 | if (unlikely(round_up(start, alignment) > round_down(end - size, alignment))) |
216 | return -ENOSPC; |
217 | |
218 | mode = DRM_MM_INSERT_BEST; |
219 | if (flags & PIN_HIGH) |
220 | mode = DRM_MM_INSERT_HIGHEST; |
221 | if (flags & PIN_MAPPABLE) |
222 | mode = DRM_MM_INSERT_LOW; |
223 | |
224 | /* We only allocate in PAGE_SIZE/GTT_PAGE_SIZE (4096) chunks, |
225 | * so we know that we always have a minimum alignment of 4096. |
226 | * The drm_mm range manager is optimised to return results |
227 | * with zero alignment, so where possible use the optimal |
228 | * path. |
229 | */ |
230 | BUILD_BUG_ON(I915_GTT_MIN_ALIGNMENT > I915_GTT_PAGE_SIZE); |
231 | if (alignment <= I915_GTT_MIN_ALIGNMENT) |
232 | alignment = 0; |
233 | |
234 | err = drm_mm_insert_node_in_range(mm: &vm->mm, node, |
235 | size, alignment, color, |
236 | start, end, mode); |
237 | if (err != -ENOSPC) |
238 | return err; |
239 | |
240 | if (mode & DRM_MM_INSERT_ONCE) { |
241 | err = drm_mm_insert_node_in_range(mm: &vm->mm, node, |
242 | size, alignment, color, |
243 | start, end, |
244 | mode: DRM_MM_INSERT_BEST); |
245 | if (err != -ENOSPC) |
246 | return err; |
247 | } |
248 | |
249 | if (flags & PIN_NOEVICT) |
250 | return -ENOSPC; |
251 | |
252 | /* |
253 | * No free space, pick a slot at random. |
254 | * |
255 | * There is a pathological case here using a GTT shared between |
256 | * mmap and GPU (i.e. ggtt/aliasing_ppgtt but not full-ppgtt): |
257 | * |
258 | * |<-- 256 MiB aperture -->||<-- 1792 MiB unmappable -->| |
259 | * (64k objects) (448k objects) |
260 | * |
261 | * Now imagine that the eviction LRU is ordered top-down (just because |
262 | * pathology meets real life), and that we need to evict an object to |
263 | * make room inside the aperture. The eviction scan then has to walk |
264 | * the 448k list before it finds one within range. And now imagine that |
265 | * it has to search for a new hole between every byte inside the memcpy, |
266 | * for several simultaneous clients. |
267 | * |
268 | * On a full-ppgtt system, if we have run out of available space, there |
269 | * will be lots and lots of objects in the eviction list! Again, |
270 | * searching that LRU list may be slow if we are also applying any |
271 | * range restrictions (e.g. restriction to low 4GiB) and so, for |
272 | * simplicity and similarilty between different GTT, try the single |
273 | * random replacement first. |
274 | */ |
275 | offset = random_offset(start, end, |
276 | len: size, align: alignment ?: I915_GTT_MIN_ALIGNMENT); |
277 | err = i915_gem_gtt_reserve(vm, ww, node, size, offset, color, flags); |
278 | if (err != -ENOSPC) |
279 | return err; |
280 | |
281 | if (flags & PIN_NOSEARCH) |
282 | return -ENOSPC; |
283 | |
284 | /* Randomly selected placement is pinned, do a search */ |
285 | err = i915_gem_evict_something(vm, ww, min_size: size, alignment, color, |
286 | start, end, flags); |
287 | if (err) |
288 | return err; |
289 | |
290 | return drm_mm_insert_node_in_range(mm: &vm->mm, node, |
291 | size, alignment, color, |
292 | start, end, mode: DRM_MM_INSERT_EVICT); |
293 | } |
294 | |
295 | #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) |
296 | #include "selftests/i915_gem_gtt.c" |
297 | #endif |
298 | |