1 /* $NetBSD: i915_gem_gtt.c,v 1.24 2022/08/20 23:19:09 riastradh Exp $ */
2
3 // SPDX-License-Identifier: MIT
4 /*
5 * Copyright © 2010 Daniel Vetter
6 * Copyright © 2020 Intel Corporation
7 */
8 #include <sys/cdefs.h>
9 __KERNEL_RCSID(0, "$NetBSD: i915_gem_gtt.c,v 1.24 2022/08/20 23:19:09 riastradh Exp $");
10
11 #include <linux/slab.h> /* fault-inject.h is not standalone! */
12
13 #include <linux/fault-inject.h>
14 #include <linux/log2.h>
15 #include <linux/random.h>
16 #include <linux/seq_file.h>
17 #include <linux/stop_machine.h>
18
19 #include <asm/set_memory.h>
20 #include <asm/smp.h>
21
22 #include <drm/i915_drm.h>
23
24 #include "display/intel_frontbuffer.h"
25 #include "gt/intel_gt.h"
26 #include "gt/intel_gt_requests.h"
27
28 #include "i915_drv.h"
29 #include "i915_scatterlist.h"
30 #include "i915_trace.h"
31 #include "i915_vgpu.h"
32
33 #ifdef __NetBSD__
34 #include <drm/bus_dma_hacks.h>
35 #include <x86/machdep.h>
36 #include <machine/pte.h>
37 #define _PAGE_PRESENT PTE_P /* 0x01 PTE is present */
38 #define _PAGE_RW PTE_W /* 0x02 read/write */
39 #define _PAGE_PWT PTE_PWT /* 0x08 page write-through */
40 #define _PAGE_PCD PTE_PCD /* 0x10 page cache disabled */
41 #define _PAGE_PAT PTE_PAT /* 0x80 page attribute table on PTE */
42 #endif
43
i915_gem_gtt_prepare_pages(struct drm_i915_gem_object * obj,struct sg_table * pages)44 int i915_gem_gtt_prepare_pages(struct drm_i915_gem_object *obj,
45 struct sg_table *pages)
46 {
47 do {
48 #ifdef __NetBSD__
49 if (dma_map_sg_attrs(obj->base.dev->dmat,
50 pages->sgl, pages->nents,
51 PCI_DMA_BIDIRECTIONAL,
52 DMA_ATTR_NO_WARN))
53 return 0;
54 #else
55 if (dma_map_sg_attrs(&obj->base.dev->pdev->dev,
56 pages->sgl, pages->nents,
57 PCI_DMA_BIDIRECTIONAL,
58 DMA_ATTR_NO_WARN))
59 return 0;
60 #endif
61
62 /*
63 * If the DMA remap fails, one cause can be that we have
64 * too many objects pinned in a small remapping table,
65 * such as swiotlb. Incrementally purge all other objects and
66 * try again - if there are no more pages to remove from
67 * the DMA remapper, i915_gem_shrink will return 0.
68 */
69 GEM_BUG_ON(obj->mm.pages == pages);
70 } while (i915_gem_shrink(to_i915(obj->base.dev),
71 obj->base.size >> PAGE_SHIFT, NULL,
72 I915_SHRINK_BOUND |
73 I915_SHRINK_UNBOUND));
74
75 return -ENOSPC;
76 }
77
i915_gem_gtt_finish_pages(struct drm_i915_gem_object * obj,struct sg_table * pages)78 void i915_gem_gtt_finish_pages(struct drm_i915_gem_object *obj,
79 struct sg_table *pages)
80 {
81 struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
82 #ifdef __NetBSD__
83 bus_dma_tag_t kdev = dev_priv->drm.dmat;
84 #else
85 struct device *kdev = &dev_priv->drm.pdev->dev;
86 #endif
87 struct i915_ggtt *ggtt = &dev_priv->ggtt;
88
89 if (unlikely(ggtt->do_idle_maps)) {
90 /* XXX This does not prevent more requests being submitted! */
91 if (intel_gt_retire_requests_timeout(ggtt->vm.gt,
92 -MAX_SCHEDULE_TIMEOUT)) {
93 DRM_ERROR("Failed to wait for idle; VT'd may hang.\n");
94 /* Wait a bit, in hopes it avoids the hang */
95 udelay(10);
96 }
97 }
98
99 dma_unmap_sg(kdev, pages->sgl, pages->nents, PCI_DMA_BIDIRECTIONAL);
100 }
101
102 /**
103 * i915_gem_gtt_reserve - reserve a node in an address_space (GTT)
104 * @vm: the &struct i915_address_space
105 * @node: the &struct drm_mm_node (typically i915_vma.mode)
106 * @size: how much space to allocate inside the GTT,
107 * must be #I915_GTT_PAGE_SIZE aligned
108 * @offset: where to insert inside the GTT,
109 * must be #I915_GTT_MIN_ALIGNMENT aligned, and the node
110 * (@offset + @size) must fit within the address space
111 * @color: color to apply to node, if this node is not from a VMA,
112 * color must be #I915_COLOR_UNEVICTABLE
113 * @flags: control search and eviction behaviour
114 *
115 * i915_gem_gtt_reserve() tries to insert the @node at the exact @offset inside
116 * the address space (using @size and @color). If the @node does not fit, it
117 * tries to evict any overlapping nodes from the GTT, including any
118 * neighbouring nodes if the colors do not match (to ensure guard pages between
119 * differing domains). See i915_gem_evict_for_node() for the gory details
120 * on the eviction algorithm. #PIN_NONBLOCK may used to prevent waiting on
121 * evicting active overlapping objects, and any overlapping node that is pinned
122 * or marked as unevictable will also result in failure.
123 *
124 * Returns: 0 on success, -ENOSPC if no suitable hole is found, -EINTR if
125 * asked to wait for eviction and interrupted.
126 */
i915_gem_gtt_reserve(struct i915_address_space * vm,struct drm_mm_node * node,u64 size,u64 offset,unsigned long color,unsigned int flags)127 int i915_gem_gtt_reserve(struct i915_address_space *vm,
128 struct drm_mm_node *node,
129 u64 size, u64 offset, unsigned long color,
130 unsigned int flags)
131 {
132 int err;
133
134 GEM_BUG_ON(!size);
135 GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
136 GEM_BUG_ON(!IS_ALIGNED(offset, I915_GTT_MIN_ALIGNMENT));
137 GEM_BUG_ON(range_overflows(offset, size, vm->total));
138 GEM_BUG_ON(vm == &vm->i915->ggtt.alias->vm);
139 GEM_BUG_ON(drm_mm_node_allocated(node));
140
141 node->size = size;
142 node->start = offset;
143 node->color = color;
144
145 err = drm_mm_reserve_node(&vm->mm, node);
146 if (err != -ENOSPC)
147 return err;
148
149 if (flags & PIN_NOEVICT)
150 return -ENOSPC;
151
152 err = i915_gem_evict_for_node(vm, node, flags);
153 if (err == 0)
154 err = drm_mm_reserve_node(&vm->mm, node);
155
156 return err;
157 }
158
random_offset(u64 start,u64 end,u64 len,u64 align)159 static u64 random_offset(u64 start, u64 end, u64 len, u64 align)
160 {
161 u64 range, addr;
162
163 GEM_BUG_ON(range_overflows(start, len, end));
164 GEM_BUG_ON(round_up(start, align) > round_down(end - len, align));
165
166 range = round_down(end - len, align) - round_up(start, align);
167 if (range) {
168 if (sizeof(unsigned long) == sizeof(u64)) {
169 addr = get_random_long();
170 } else {
171 addr = get_random_int();
172 if (range > U32_MAX) {
173 addr <<= 32;
174 addr |= get_random_int();
175 }
176 }
177 div64_u64_rem(addr, range, &addr);
178 start += addr;
179 }
180
181 return round_up(start, align);
182 }
183
184 /**
185 * i915_gem_gtt_insert - insert a node into an address_space (GTT)
186 * @vm: the &struct i915_address_space
187 * @node: the &struct drm_mm_node (typically i915_vma.node)
188 * @size: how much space to allocate inside the GTT,
189 * must be #I915_GTT_PAGE_SIZE aligned
190 * @alignment: required alignment of starting offset, may be 0 but
191 * if specified, this must be a power-of-two and at least
192 * #I915_GTT_MIN_ALIGNMENT
193 * @color: color to apply to node
194 * @start: start of any range restriction inside GTT (0 for all),
195 * must be #I915_GTT_PAGE_SIZE aligned
196 * @end: end of any range restriction inside GTT (U64_MAX for all),
197 * must be #I915_GTT_PAGE_SIZE aligned if not U64_MAX
198 * @flags: control search and eviction behaviour
199 *
200 * i915_gem_gtt_insert() first searches for an available hole into which
201 * is can insert the node. The hole address is aligned to @alignment and
202 * its @size must then fit entirely within the [@start, @end] bounds. The
203 * nodes on either side of the hole must match @color, or else a guard page
204 * will be inserted between the two nodes (or the node evicted). If no
205 * suitable hole is found, first a victim is randomly selected and tested
206 * for eviction, otherwise then the LRU list of objects within the GTT
207 * is scanned to find the first set of replacement nodes to create the hole.
208 * Those old overlapping nodes are evicted from the GTT (and so must be
209 * rebound before any future use). Any node that is currently pinned cannot
210 * be evicted (see i915_vma_pin()). Similar if the node's VMA is currently
211 * active and #PIN_NONBLOCK is specified, that node is also skipped when
212 * searching for an eviction candidate. See i915_gem_evict_something() for
213 * the gory details on the eviction algorithm.
214 *
215 * Returns: 0 on success, -ENOSPC if no suitable hole is found, -EINTR if
216 * asked to wait for eviction and interrupted.
217 */
i915_gem_gtt_insert(struct i915_address_space * vm,struct drm_mm_node * node,u64 size,u64 alignment,unsigned long color,u64 start,u64 end,unsigned int flags)218 int i915_gem_gtt_insert(struct i915_address_space *vm,
219 struct drm_mm_node *node,
220 u64 size, u64 alignment, unsigned long color,
221 u64 start, u64 end, unsigned int flags)
222 {
223 enum drm_mm_insert_mode mode;
224 u64 offset;
225 int err;
226
227 lockdep_assert_held(&vm->mutex);
228
229 GEM_BUG_ON(!size);
230 GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
231 GEM_BUG_ON(alignment && !is_power_of_2(alignment));
232 GEM_BUG_ON(alignment && !IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
233 GEM_BUG_ON(start >= end);
234 GEM_BUG_ON(start > 0 && !IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
235 GEM_BUG_ON(end < U64_MAX && !IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
236 GEM_BUG_ON(vm == &vm->i915->ggtt.alias->vm);
237 GEM_BUG_ON(drm_mm_node_allocated(node));
238
239 if (unlikely(range_overflows(start, size, end)))
240 return -ENOSPC;
241
242 if (unlikely(round_up(start, alignment) > round_down(end - size, alignment)))
243 return -ENOSPC;
244
245 mode = DRM_MM_INSERT_BEST;
246 if (flags & PIN_HIGH)
247 mode = DRM_MM_INSERT_HIGHEST;
248 if (flags & PIN_MAPPABLE)
249 mode = DRM_MM_INSERT_LOW;
250
251 /* We only allocate in PAGE_SIZE/GTT_PAGE_SIZE (4096) chunks,
252 * so we know that we always have a minimum alignment of 4096.
253 * The drm_mm range manager is optimised to return results
254 * with zero alignment, so where possible use the optimal
255 * path.
256 */
257 BUILD_BUG_ON(I915_GTT_MIN_ALIGNMENT > I915_GTT_PAGE_SIZE);
258 if (alignment <= I915_GTT_MIN_ALIGNMENT)
259 alignment = 0;
260
261 err = drm_mm_insert_node_in_range(&vm->mm, node,
262 size, alignment, color,
263 start, end, mode);
264 if (err != -ENOSPC)
265 return err;
266
267 if (mode & DRM_MM_INSERT_ONCE) {
268 err = drm_mm_insert_node_in_range(&vm->mm, node,
269 size, alignment, color,
270 start, end,
271 DRM_MM_INSERT_BEST);
272 if (err != -ENOSPC)
273 return err;
274 }
275
276 if (flags & PIN_NOEVICT)
277 return -ENOSPC;
278
279 /*
280 * No free space, pick a slot at random.
281 *
282 * There is a pathological case here using a GTT shared between
283 * mmap and GPU (i.e. ggtt/aliasing_ppgtt but not full-ppgtt):
284 *
285 * |<-- 256 MiB aperture -->||<-- 1792 MiB unmappable -->|
286 * (64k objects) (448k objects)
287 *
288 * Now imagine that the eviction LRU is ordered top-down (just because
289 * pathology meets real life), and that we need to evict an object to
290 * make room inside the aperture. The eviction scan then has to walk
291 * the 448k list before it finds one within range. And now imagine that
292 * it has to search for a new hole between every byte inside the memcpy,
293 * for several simultaneous clients.
294 *
295 * On a full-ppgtt system, if we have run out of available space, there
296 * will be lots and lots of objects in the eviction list! Again,
297 * searching that LRU list may be slow if we are also applying any
298 * range restrictions (e.g. restriction to low 4GiB) and so, for
299 * simplicity and similarilty between different GTT, try the single
300 * random replacement first.
301 */
302 offset = random_offset(start, end,
303 size, alignment ?: I915_GTT_MIN_ALIGNMENT);
304 err = i915_gem_gtt_reserve(vm, node, size, offset, color, flags);
305 if (err != -ENOSPC)
306 return err;
307
308 if (flags & PIN_NOSEARCH)
309 return -ENOSPC;
310
311 /* Randomly selected placement is pinned, do a search */
312 err = i915_gem_evict_something(vm, size, alignment, color,
313 start, end, flags);
314 if (err)
315 return err;
316
317 return drm_mm_insert_node_in_range(&vm->mm, node,
318 size, alignment, color,
319 start, end, DRM_MM_INSERT_EVICT);
320 }
321
322 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
323 #include "selftests/i915_gem_gtt.c"
324 #endif
325