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 "gt/intel_gt.h"
19 #include "gt/intel_gt_requests.h"
20 
21 #include "i915_drv.h"
22 #include "i915_gem_evict.h"
23 #include "i915_scatterlist.h"
24 #include "i915_trace.h"
25 #include "i915_vgpu.h"
26 
i915_gem_gtt_prepare_pages(struct drm_i915_gem_object * obj,struct sg_table * pages)27 int i915_gem_gtt_prepare_pages(struct drm_i915_gem_object *obj,
28 			       struct sg_table *pages)
29 {
30 #ifdef __linux__
31 	do {
32 		if (dma_map_sg_attrs(obj->base.dev->dev,
33 				     pages->sgl, pages->nents,
34 				     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, to_i915(obj->base.dev),
49 				 obj->base.size >> PAGE_SHIFT, NULL,
50 				 I915_SHRINK_BOUND |
51 				 I915_SHRINK_UNBOUND));
52 
53 	return -ENOSPC;
54 #else
55 	return 0;
56 #endif
57 }
58 
i915_gem_gtt_finish_pages(struct drm_i915_gem_object * obj,struct sg_table * pages)59 void i915_gem_gtt_finish_pages(struct drm_i915_gem_object *obj,
60 			       struct sg_table *pages)
61 {
62 	struct drm_i915_private *i915 = to_i915(obj->base.dev);
63 	struct i915_ggtt *ggtt = to_gt(i915)->ggtt;
64 
65 	/* XXX This does not prevent more requests being submitted! */
66 	if (unlikely(ggtt->do_idle_maps))
67 		/* Wait a bit, in the hope it avoids the hang */
68 		usleep_range(100, 250);
69 
70 #ifdef notyet
71 	dma_unmap_sg(i915->drm.dev, pages->sgl, pages->nents,
72 		     DMA_BIDIRECTIONAL);
73 #endif
74 }
75 
76 /**
77  * i915_gem_gtt_reserve - reserve a node in an address_space (GTT)
78  * @vm: the &struct i915_address_space
79  * @ww: An optional struct i915_gem_ww_ctx.
80  * @node: the &struct drm_mm_node (typically i915_vma.mode)
81  * @size: how much space to allocate inside the GTT,
82  *        must be #I915_GTT_PAGE_SIZE aligned
83  * @offset: where to insert inside the GTT,
84  *          must be #I915_GTT_MIN_ALIGNMENT aligned, and the node
85  *          (@offset + @size) must fit within the address space
86  * @color: color to apply to node, if this node is not from a VMA,
87  *         color must be #I915_COLOR_UNEVICTABLE
88  * @flags: control search and eviction behaviour
89  *
90  * i915_gem_gtt_reserve() tries to insert the @node at the exact @offset inside
91  * the address space (using @size and @color). If the @node does not fit, it
92  * tries to evict any overlapping nodes from the GTT, including any
93  * neighbouring nodes if the colors do not match (to ensure guard pages between
94  * differing domains). See i915_gem_evict_for_node() for the gory details
95  * on the eviction algorithm. #PIN_NONBLOCK may used to prevent waiting on
96  * evicting active overlapping objects, and any overlapping node that is pinned
97  * or marked as unevictable will also result in failure.
98  *
99  * Returns: 0 on success, -ENOSPC if no suitable hole is found, -EINTR if
100  * asked to wait for eviction and interrupted.
101  */
i915_gem_gtt_reserve(struct i915_address_space * vm,struct i915_gem_ww_ctx * ww,struct drm_mm_node * node,u64 size,u64 offset,unsigned long color,unsigned int flags)102 int i915_gem_gtt_reserve(struct i915_address_space *vm,
103 			 struct i915_gem_ww_ctx *ww,
104 			 struct drm_mm_node *node,
105 			 u64 size, u64 offset, unsigned long color,
106 			 unsigned int flags)
107 {
108 	int err;
109 
110 	GEM_BUG_ON(!size);
111 	GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
112 	GEM_BUG_ON(!IS_ALIGNED(offset, I915_GTT_MIN_ALIGNMENT));
113 	GEM_BUG_ON(range_overflows(offset, size, vm->total));
114 	GEM_BUG_ON(vm == &to_gt(vm->i915)->ggtt->alias->vm);
115 	GEM_BUG_ON(drm_mm_node_allocated(node));
116 
117 	node->size = size;
118 	node->start = offset;
119 	node->color = color;
120 
121 	err = drm_mm_reserve_node(&vm->mm, node);
122 	if (err != -ENOSPC)
123 		return err;
124 
125 	if (flags & PIN_NOEVICT)
126 		return -ENOSPC;
127 
128 	err = i915_gem_evict_for_node(vm, ww, node, flags);
129 	if (err == 0)
130 		err = drm_mm_reserve_node(&vm->mm, node);
131 
132 	return err;
133 }
134 
random_offset(u64 start,u64 end,u64 len,u64 align)135 static u64 random_offset(u64 start, u64 end, u64 len, u64 align)
136 {
137 	u64 range, addr;
138 
139 	GEM_BUG_ON(range_overflows(start, len, end));
140 	GEM_BUG_ON(round_up(start, align) > round_down(end - len, align));
141 
142 	range = round_down(end - len, align) - round_up(start, align);
143 	if (range) {
144 		if (sizeof(unsigned long) == sizeof(u64)) {
145 			addr = get_random_u64();
146 		} else {
147 			addr = get_random_u32();
148 			if (range > U32_MAX) {
149 				addr <<= 32;
150 				addr |= get_random_u32();
151 			}
152 		}
153 		div64_u64_rem(addr, range, &addr);
154 		start += addr;
155 	}
156 
157 	return round_up(start, align);
158 }
159 
160 /**
161  * i915_gem_gtt_insert - insert a node into an address_space (GTT)
162  * @vm: the &struct i915_address_space
163  * @ww: An optional struct i915_gem_ww_ctx.
164  * @node: the &struct drm_mm_node (typically i915_vma.node)
165  * @size: how much space to allocate inside the GTT,
166  *        must be #I915_GTT_PAGE_SIZE aligned
167  * @alignment: required alignment of starting offset, may be 0 but
168  *             if specified, this must be a power-of-two and at least
169  *             #I915_GTT_MIN_ALIGNMENT
170  * @color: color to apply to node
171  * @start: start of any range restriction inside GTT (0 for all),
172  *         must be #I915_GTT_PAGE_SIZE aligned
173  * @end: end of any range restriction inside GTT (U64_MAX for all),
174  *       must be #I915_GTT_PAGE_SIZE aligned if not U64_MAX
175  * @flags: control search and eviction behaviour
176  *
177  * i915_gem_gtt_insert() first searches for an available hole into which
178  * is can insert the node. The hole address is aligned to @alignment and
179  * its @size must then fit entirely within the [@start, @end] bounds. The
180  * nodes on either side of the hole must match @color, or else a guard page
181  * will be inserted between the two nodes (or the node evicted). If no
182  * suitable hole is found, first a victim is randomly selected and tested
183  * for eviction, otherwise then the LRU list of objects within the GTT
184  * is scanned to find the first set of replacement nodes to create the hole.
185  * Those old overlapping nodes are evicted from the GTT (and so must be
186  * rebound before any future use). Any node that is currently pinned cannot
187  * be evicted (see i915_vma_pin()). Similar if the node's VMA is currently
188  * active and #PIN_NONBLOCK is specified, that node is also skipped when
189  * searching for an eviction candidate. See i915_gem_evict_something() for
190  * the gory details on the eviction algorithm.
191  *
192  * Returns: 0 on success, -ENOSPC if no suitable hole is found, -EINTR if
193  * asked to wait for eviction and interrupted.
194  */
i915_gem_gtt_insert(struct i915_address_space * vm,struct i915_gem_ww_ctx * ww,struct drm_mm_node * node,u64 size,u64 alignment,unsigned long color,u64 start,u64 end,unsigned int flags)195 int i915_gem_gtt_insert(struct i915_address_space *vm,
196 			struct i915_gem_ww_ctx *ww,
197 			struct drm_mm_node *node,
198 			u64 size, u64 alignment, unsigned long color,
199 			u64 start, u64 end, unsigned int flags)
200 {
201 	enum drm_mm_insert_mode mode;
202 	u64 offset;
203 	int err;
204 
205 	lockdep_assert_held(&vm->mutex);
206 
207 	GEM_BUG_ON(!size);
208 	GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
209 	GEM_BUG_ON(alignment && !is_power_of_2(alignment));
210 	GEM_BUG_ON(alignment && !IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
211 	GEM_BUG_ON(start >= end);
212 	GEM_BUG_ON(start > 0  && !IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
213 	GEM_BUG_ON(end < U64_MAX && !IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
214 	GEM_BUG_ON(vm == &to_gt(vm->i915)->ggtt->alias->vm);
215 	GEM_BUG_ON(drm_mm_node_allocated(node));
216 
217 	if (unlikely(range_overflows(start, size, end)))
218 		return -ENOSPC;
219 
220 	if (unlikely(round_up(start, alignment) > round_down(end - size, alignment)))
221 		return -ENOSPC;
222 
223 	mode = DRM_MM_INSERT_BEST;
224 	if (flags & PIN_HIGH)
225 		mode = DRM_MM_INSERT_HIGHEST;
226 	if (flags & PIN_MAPPABLE)
227 		mode = DRM_MM_INSERT_LOW;
228 
229 	/* We only allocate in PAGE_SIZE/GTT_PAGE_SIZE (4096) chunks,
230 	 * so we know that we always have a minimum alignment of 4096.
231 	 * The drm_mm range manager is optimised to return results
232 	 * with zero alignment, so where possible use the optimal
233 	 * path.
234 	 */
235 	BUILD_BUG_ON(I915_GTT_MIN_ALIGNMENT > I915_GTT_PAGE_SIZE);
236 	if (alignment <= I915_GTT_MIN_ALIGNMENT)
237 		alignment = 0;
238 
239 	err = drm_mm_insert_node_in_range(&vm->mm, node,
240 					  size, alignment, color,
241 					  start, end, mode);
242 	if (err != -ENOSPC)
243 		return err;
244 
245 	if (mode & DRM_MM_INSERT_ONCE) {
246 		err = drm_mm_insert_node_in_range(&vm->mm, node,
247 						  size, alignment, color,
248 						  start, end,
249 						  DRM_MM_INSERT_BEST);
250 		if (err != -ENOSPC)
251 			return err;
252 	}
253 
254 	if (flags & PIN_NOEVICT)
255 		return -ENOSPC;
256 
257 	/*
258 	 * No free space, pick a slot at random.
259 	 *
260 	 * There is a pathological case here using a GTT shared between
261 	 * mmap and GPU (i.e. ggtt/aliasing_ppgtt but not full-ppgtt):
262 	 *
263 	 *    |<-- 256 MiB aperture -->||<-- 1792 MiB unmappable -->|
264 	 *         (64k objects)             (448k objects)
265 	 *
266 	 * Now imagine that the eviction LRU is ordered top-down (just because
267 	 * pathology meets real life), and that we need to evict an object to
268 	 * make room inside the aperture. The eviction scan then has to walk
269 	 * the 448k list before it finds one within range. And now imagine that
270 	 * it has to search for a new hole between every byte inside the memcpy,
271 	 * for several simultaneous clients.
272 	 *
273 	 * On a full-ppgtt system, if we have run out of available space, there
274 	 * will be lots and lots of objects in the eviction list! Again,
275 	 * searching that LRU list may be slow if we are also applying any
276 	 * range restrictions (e.g. restriction to low 4GiB) and so, for
277 	 * simplicity and similarilty between different GTT, try the single
278 	 * random replacement first.
279 	 */
280 	offset = random_offset(start, end,
281 			       size, alignment ?: I915_GTT_MIN_ALIGNMENT);
282 	err = i915_gem_gtt_reserve(vm, ww, node, size, offset, color, flags);
283 	if (err != -ENOSPC)
284 		return err;
285 
286 	if (flags & PIN_NOSEARCH)
287 		return -ENOSPC;
288 
289 	/* Randomly selected placement is pinned, do a search */
290 	err = i915_gem_evict_something(vm, ww, size, alignment, color,
291 				       start, end, flags);
292 	if (err)
293 		return err;
294 
295 	return drm_mm_insert_node_in_range(&vm->mm, node,
296 					   size, alignment, color,
297 					   start, end, DRM_MM_INSERT_EVICT);
298 }
299 
300 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
301 #include "selftests/i915_gem_gtt.c"
302 #endif
303