xref: /linux/drivers/gpu/drm/i915/i915_gem_gtt.c (revision de73b5a97bba1538f065e1e90d8eeac399db7510)
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_scatterlist.h"
24 #include "i915_trace.h"
25 #include "i915_vgpu.h"
26 
27 int i915_gem_gtt_prepare_pages(struct drm_i915_gem_object *obj,
28 			       struct sg_table *pages)
29 {
30 	do {
31 		if (dma_map_sg_attrs(&obj->base.dev->pdev->dev,
32 				     pages->sgl, pages->nents,
33 				     PCI_DMA_BIDIRECTIONAL,
34 				     DMA_ATTR_SKIP_CPU_SYNC |
35 				     DMA_ATTR_NO_KERNEL_MAPPING |
36 				     DMA_ATTR_NO_WARN))
37 			return 0;
38 
39 		/*
40 		 * If the DMA remap fails, one cause can be that we have
41 		 * too many objects pinned in a small remapping table,
42 		 * such as swiotlb. Incrementally purge all other objects and
43 		 * try again - if there are no more pages to remove from
44 		 * the DMA remapper, i915_gem_shrink will return 0.
45 		 */
46 		GEM_BUG_ON(obj->mm.pages == pages);
47 	} while (i915_gem_shrink(to_i915(obj->base.dev),
48 				 obj->base.size >> PAGE_SHIFT, NULL,
49 				 I915_SHRINK_BOUND |
50 				 I915_SHRINK_UNBOUND));
51 
52 	return -ENOSPC;
53 }
54 
55 void i915_gem_gtt_finish_pages(struct drm_i915_gem_object *obj,
56 			       struct sg_table *pages)
57 {
58 	struct drm_i915_private *i915 = to_i915(obj->base.dev);
59 	struct i915_ggtt *ggtt = &i915->ggtt;
60 
61 	/* XXX This does not prevent more requests being submitted! */
62 	if (unlikely(ggtt->do_idle_maps))
63 		/* Wait a bit, in the hope it avoids the hang */
64 		usleep_range(100, 250);
65 
66 	dma_unmap_sg(&i915->drm.pdev->dev,
67 		     pages->sgl, pages->nents,
68 		     PCI_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  * @node: the &struct drm_mm_node (typically i915_vma.mode)
75  * @size: how much space to allocate inside the GTT,
76  *        must be #I915_GTT_PAGE_SIZE aligned
77  * @offset: where to insert inside the GTT,
78  *          must be #I915_GTT_MIN_ALIGNMENT aligned, and the node
79  *          (@offset + @size) must fit within the address space
80  * @color: color to apply to node, if this node is not from a VMA,
81  *         color must be #I915_COLOR_UNEVICTABLE
82  * @flags: control search and eviction behaviour
83  *
84  * i915_gem_gtt_reserve() tries to insert the @node at the exact @offset inside
85  * the address space (using @size and @color). If the @node does not fit, it
86  * tries to evict any overlapping nodes from the GTT, including any
87  * neighbouring nodes if the colors do not match (to ensure guard pages between
88  * differing domains). See i915_gem_evict_for_node() for the gory details
89  * on the eviction algorithm. #PIN_NONBLOCK may used to prevent waiting on
90  * evicting active overlapping objects, and any overlapping node that is pinned
91  * or marked as unevictable will also result in failure.
92  *
93  * Returns: 0 on success, -ENOSPC if no suitable hole is found, -EINTR if
94  * asked to wait for eviction and interrupted.
95  */
96 int i915_gem_gtt_reserve(struct i915_address_space *vm,
97 			 struct drm_mm_node *node,
98 			 u64 size, u64 offset, unsigned long color,
99 			 unsigned int flags)
100 {
101 	int err;
102 
103 	GEM_BUG_ON(!size);
104 	GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
105 	GEM_BUG_ON(!IS_ALIGNED(offset, I915_GTT_MIN_ALIGNMENT));
106 	GEM_BUG_ON(range_overflows(offset, size, vm->total));
107 	GEM_BUG_ON(vm == &vm->i915->ggtt.alias->vm);
108 	GEM_BUG_ON(drm_mm_node_allocated(node));
109 
110 	node->size = size;
111 	node->start = offset;
112 	node->color = color;
113 
114 	err = drm_mm_reserve_node(&vm->mm, node);
115 	if (err != -ENOSPC)
116 		return err;
117 
118 	if (flags & PIN_NOEVICT)
119 		return -ENOSPC;
120 
121 	err = i915_gem_evict_for_node(vm, node, flags);
122 	if (err == 0)
123 		err = drm_mm_reserve_node(&vm->mm, node);
124 
125 	return err;
126 }
127 
128 static u64 random_offset(u64 start, u64 end, u64 len, u64 align)
129 {
130 	u64 range, addr;
131 
132 	GEM_BUG_ON(range_overflows(start, len, end));
133 	GEM_BUG_ON(round_up(start, align) > round_down(end - len, align));
134 
135 	range = round_down(end - len, align) - round_up(start, align);
136 	if (range) {
137 		if (sizeof(unsigned long) == sizeof(u64)) {
138 			addr = get_random_long();
139 		} else {
140 			addr = get_random_int();
141 			if (range > U32_MAX) {
142 				addr <<= 32;
143 				addr |= get_random_int();
144 			}
145 		}
146 		div64_u64_rem(addr, range, &addr);
147 		start += addr;
148 	}
149 
150 	return round_up(start, align);
151 }
152 
153 /**
154  * i915_gem_gtt_insert - insert a node into an address_space (GTT)
155  * @vm: the &struct i915_address_space
156  * @node: the &struct drm_mm_node (typically i915_vma.node)
157  * @size: how much space to allocate inside the GTT,
158  *        must be #I915_GTT_PAGE_SIZE aligned
159  * @alignment: required alignment of starting offset, may be 0 but
160  *             if specified, this must be a power-of-two and at least
161  *             #I915_GTT_MIN_ALIGNMENT
162  * @color: color to apply to node
163  * @start: start of any range restriction inside GTT (0 for all),
164  *         must be #I915_GTT_PAGE_SIZE aligned
165  * @end: end of any range restriction inside GTT (U64_MAX for all),
166  *       must be #I915_GTT_PAGE_SIZE aligned if not U64_MAX
167  * @flags: control search and eviction behaviour
168  *
169  * i915_gem_gtt_insert() first searches for an available hole into which
170  * is can insert the node. The hole address is aligned to @alignment and
171  * its @size must then fit entirely within the [@start, @end] bounds. The
172  * nodes on either side of the hole must match @color, or else a guard page
173  * will be inserted between the two nodes (or the node evicted). If no
174  * suitable hole is found, first a victim is randomly selected and tested
175  * for eviction, otherwise then the LRU list of objects within the GTT
176  * is scanned to find the first set of replacement nodes to create the hole.
177  * Those old overlapping nodes are evicted from the GTT (and so must be
178  * rebound before any future use). Any node that is currently pinned cannot
179  * be evicted (see i915_vma_pin()). Similar if the node's VMA is currently
180  * active and #PIN_NONBLOCK is specified, that node is also skipped when
181  * searching for an eviction candidate. See i915_gem_evict_something() for
182  * the gory details on the eviction algorithm.
183  *
184  * Returns: 0 on success, -ENOSPC if no suitable hole is found, -EINTR if
185  * asked to wait for eviction and interrupted.
186  */
187 int i915_gem_gtt_insert(struct i915_address_space *vm,
188 			struct drm_mm_node *node,
189 			u64 size, u64 alignment, unsigned long color,
190 			u64 start, u64 end, unsigned int flags)
191 {
192 	enum drm_mm_insert_mode mode;
193 	u64 offset;
194 	int err;
195 
196 	lockdep_assert_held(&vm->mutex);
197 
198 	GEM_BUG_ON(!size);
199 	GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
200 	GEM_BUG_ON(alignment && !is_power_of_2(alignment));
201 	GEM_BUG_ON(alignment && !IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
202 	GEM_BUG_ON(start >= end);
203 	GEM_BUG_ON(start > 0  && !IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
204 	GEM_BUG_ON(end < U64_MAX && !IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
205 	GEM_BUG_ON(vm == &vm->i915->ggtt.alias->vm);
206 	GEM_BUG_ON(drm_mm_node_allocated(node));
207 
208 	if (unlikely(range_overflows(start, size, end)))
209 		return -ENOSPC;
210 
211 	if (unlikely(round_up(start, alignment) > round_down(end - size, alignment)))
212 		return -ENOSPC;
213 
214 	mode = DRM_MM_INSERT_BEST;
215 	if (flags & PIN_HIGH)
216 		mode = DRM_MM_INSERT_HIGHEST;
217 	if (flags & PIN_MAPPABLE)
218 		mode = DRM_MM_INSERT_LOW;
219 
220 	/* We only allocate in PAGE_SIZE/GTT_PAGE_SIZE (4096) chunks,
221 	 * so we know that we always have a minimum alignment of 4096.
222 	 * The drm_mm range manager is optimised to return results
223 	 * with zero alignment, so where possible use the optimal
224 	 * path.
225 	 */
226 	BUILD_BUG_ON(I915_GTT_MIN_ALIGNMENT > I915_GTT_PAGE_SIZE);
227 	if (alignment <= I915_GTT_MIN_ALIGNMENT)
228 		alignment = 0;
229 
230 	err = drm_mm_insert_node_in_range(&vm->mm, node,
231 					  size, alignment, color,
232 					  start, end, mode);
233 	if (err != -ENOSPC)
234 		return err;
235 
236 	if (mode & DRM_MM_INSERT_ONCE) {
237 		err = drm_mm_insert_node_in_range(&vm->mm, node,
238 						  size, alignment, color,
239 						  start, end,
240 						  DRM_MM_INSERT_BEST);
241 		if (err != -ENOSPC)
242 			return err;
243 	}
244 
245 	if (flags & PIN_NOEVICT)
246 		return -ENOSPC;
247 
248 	/*
249 	 * No free space, pick a slot at random.
250 	 *
251 	 * There is a pathological case here using a GTT shared between
252 	 * mmap and GPU (i.e. ggtt/aliasing_ppgtt but not full-ppgtt):
253 	 *
254 	 *    |<-- 256 MiB aperture -->||<-- 1792 MiB unmappable -->|
255 	 *         (64k objects)             (448k objects)
256 	 *
257 	 * Now imagine that the eviction LRU is ordered top-down (just because
258 	 * pathology meets real life), and that we need to evict an object to
259 	 * make room inside the aperture. The eviction scan then has to walk
260 	 * the 448k list before it finds one within range. And now imagine that
261 	 * it has to search for a new hole between every byte inside the memcpy,
262 	 * for several simultaneous clients.
263 	 *
264 	 * On a full-ppgtt system, if we have run out of available space, there
265 	 * will be lots and lots of objects in the eviction list! Again,
266 	 * searching that LRU list may be slow if we are also applying any
267 	 * range restrictions (e.g. restriction to low 4GiB) and so, for
268 	 * simplicity and similarilty between different GTT, try the single
269 	 * random replacement first.
270 	 */
271 	offset = random_offset(start, end,
272 			       size, alignment ?: I915_GTT_MIN_ALIGNMENT);
273 	err = i915_gem_gtt_reserve(vm, node, size, offset, color, flags);
274 	if (err != -ENOSPC)
275 		return err;
276 
277 	if (flags & PIN_NOSEARCH)
278 		return -ENOSPC;
279 
280 	/* Randomly selected placement is pinned, do a search */
281 	err = i915_gem_evict_something(vm, size, alignment, color,
282 				       start, end, flags);
283 	if (err)
284 		return err;
285 
286 	return drm_mm_insert_node_in_range(&vm->mm, node,
287 					   size, alignment, color,
288 					   start, end, DRM_MM_INSERT_EVICT);
289 }
290 
291 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
292 #include "selftests/i915_gem_gtt.c"
293 #endif
294