1 // SPDX-License-Identifier: MIT
3 * Copyright © 2010 Daniel Vetter
4 * Copyright © 2020 Intel Corporation
7 #include <linux/slab.h> /* fault-inject.h is not standalone! */
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>
15 #include <asm/set_memory.h>
18 #include "display/intel_frontbuffer.h"
19 #include "gt/intel_gt.h"
20 #include "gt/intel_gt_requests.h"
23 #include "i915_scatterlist.h"
24 #include "i915_trace.h"
25 #include "i915_vgpu.h"
27 int i915_gem_gtt_prepare_pages(struct drm_i915_gem_object *obj,
28 struct sg_table *pages)
31 if (dma_map_sg_attrs(&obj->base.dev->pdev->dev,
32 pages->sgl, pages->nents,
33 PCI_DMA_BIDIRECTIONAL,
38 * If the DMA remap fails, one cause can be that we have
39 * too many objects pinned in a small remapping table,
40 * such as swiotlb. Incrementally purge all other objects and
41 * try again - if there are no more pages to remove from
42 * the DMA remapper, i915_gem_shrink will return 0.
44 GEM_BUG_ON(obj->mm.pages == pages);
45 } while (i915_gem_shrink(to_i915(obj->base.dev),
46 obj->base.size >> PAGE_SHIFT, NULL,
48 I915_SHRINK_UNBOUND));
53 void i915_gem_gtt_finish_pages(struct drm_i915_gem_object *obj,
54 struct sg_table *pages)
56 struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
57 struct device *kdev = &dev_priv->drm.pdev->dev;
58 struct i915_ggtt *ggtt = &dev_priv->ggtt;
60 if (unlikely(ggtt->do_idle_maps)) {
61 /* XXX This does not prevent more requests being submitted! */
62 if (intel_gt_retire_requests_timeout(ggtt->vm.gt,
63 -MAX_SCHEDULE_TIMEOUT)) {
64 drm_err(&dev_priv->drm,
65 "Failed to wait for idle; VT'd may hang.\n");
66 /* Wait a bit, in hopes it avoids the hang */
71 dma_unmap_sg(kdev, pages->sgl, pages->nents, PCI_DMA_BIDIRECTIONAL);
75 * i915_gem_gtt_reserve - reserve a node in an address_space (GTT)
76 * @vm: the &struct i915_address_space
77 * @node: the &struct drm_mm_node (typically i915_vma.mode)
78 * @size: how much space to allocate inside the GTT,
79 * must be #I915_GTT_PAGE_SIZE aligned
80 * @offset: where to insert inside the GTT,
81 * must be #I915_GTT_MIN_ALIGNMENT aligned, and the node
82 * (@offset + @size) must fit within the address space
83 * @color: color to apply to node, if this node is not from a VMA,
84 * color must be #I915_COLOR_UNEVICTABLE
85 * @flags: control search and eviction behaviour
87 * i915_gem_gtt_reserve() tries to insert the @node at the exact @offset inside
88 * the address space (using @size and @color). If the @node does not fit, it
89 * tries to evict any overlapping nodes from the GTT, including any
90 * neighbouring nodes if the colors do not match (to ensure guard pages between
91 * differing domains). See i915_gem_evict_for_node() for the gory details
92 * on the eviction algorithm. #PIN_NONBLOCK may used to prevent waiting on
93 * evicting active overlapping objects, and any overlapping node that is pinned
94 * or marked as unevictable will also result in failure.
96 * Returns: 0 on success, -ENOSPC if no suitable hole is found, -EINTR if
97 * asked to wait for eviction and interrupted.
99 int i915_gem_gtt_reserve(struct i915_address_space *vm,
100 struct drm_mm_node *node,
101 u64 size, u64 offset, unsigned long color,
107 GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
108 GEM_BUG_ON(!IS_ALIGNED(offset, I915_GTT_MIN_ALIGNMENT));
109 GEM_BUG_ON(range_overflows(offset, size, vm->total));
110 GEM_BUG_ON(vm == &vm->i915->ggtt.alias->vm);
111 GEM_BUG_ON(drm_mm_node_allocated(node));
114 node->start = offset;
117 err = drm_mm_reserve_node(&vm->mm, node);
121 if (flags & PIN_NOEVICT)
124 err = i915_gem_evict_for_node(vm, node, flags);
126 err = drm_mm_reserve_node(&vm->mm, node);
131 static u64 random_offset(u64 start, u64 end, u64 len, u64 align)
135 GEM_BUG_ON(range_overflows(start, len, end));
136 GEM_BUG_ON(round_up(start, align) > round_down(end - len, align));
138 range = round_down(end - len, align) - round_up(start, align);
140 if (sizeof(unsigned long) == sizeof(u64)) {
141 addr = get_random_long();
143 addr = get_random_int();
144 if (range > U32_MAX) {
146 addr |= get_random_int();
149 div64_u64_rem(addr, range, &addr);
153 return round_up(start, align);
157 * i915_gem_gtt_insert - insert a node into an address_space (GTT)
158 * @vm: the &struct i915_address_space
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
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.
187 * Returns: 0 on success, -ENOSPC if no suitable hole is found, -EINTR if
188 * asked to wait for eviction and interrupted.
190 int i915_gem_gtt_insert(struct i915_address_space *vm,
191 struct drm_mm_node *node,
192 u64 size, u64 alignment, unsigned long color,
193 u64 start, u64 end, unsigned int flags)
195 enum drm_mm_insert_mode mode;
199 lockdep_assert_held(&vm->mutex);
202 GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
203 GEM_BUG_ON(alignment && !is_power_of_2(alignment));
204 GEM_BUG_ON(alignment && !IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
205 GEM_BUG_ON(start >= end);
206 GEM_BUG_ON(start > 0 && !IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
207 GEM_BUG_ON(end < U64_MAX && !IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
208 GEM_BUG_ON(vm == &vm->i915->ggtt.alias->vm);
209 GEM_BUG_ON(drm_mm_node_allocated(node));
211 if (unlikely(range_overflows(start, size, end)))
214 if (unlikely(round_up(start, alignment) > round_down(end - size, alignment)))
217 mode = DRM_MM_INSERT_BEST;
218 if (flags & PIN_HIGH)
219 mode = DRM_MM_INSERT_HIGHEST;
220 if (flags & PIN_MAPPABLE)
221 mode = DRM_MM_INSERT_LOW;
223 /* We only allocate in PAGE_SIZE/GTT_PAGE_SIZE (4096) chunks,
224 * so we know that we always have a minimum alignment of 4096.
225 * The drm_mm range manager is optimised to return results
226 * with zero alignment, so where possible use the optimal
229 BUILD_BUG_ON(I915_GTT_MIN_ALIGNMENT > I915_GTT_PAGE_SIZE);
230 if (alignment <= I915_GTT_MIN_ALIGNMENT)
233 err = drm_mm_insert_node_in_range(&vm->mm, node,
234 size, alignment, color,
239 if (mode & DRM_MM_INSERT_ONCE) {
240 err = drm_mm_insert_node_in_range(&vm->mm, node,
241 size, alignment, color,
248 if (flags & PIN_NOEVICT)
252 * No free space, pick a slot at random.
254 * There is a pathological case here using a GTT shared between
255 * mmap and GPU (i.e. ggtt/aliasing_ppgtt but not full-ppgtt):
257 * |<-- 256 MiB aperture -->||<-- 1792 MiB unmappable -->|
258 * (64k objects) (448k objects)
260 * Now imagine that the eviction LRU is ordered top-down (just because
261 * pathology meets real life), and that we need to evict an object to
262 * make room inside the aperture. The eviction scan then has to walk
263 * the 448k list before it finds one within range. And now imagine that
264 * it has to search for a new hole between every byte inside the memcpy,
265 * for several simultaneous clients.
267 * On a full-ppgtt system, if we have run out of available space, there
268 * will be lots and lots of objects in the eviction list! Again,
269 * searching that LRU list may be slow if we are also applying any
270 * range restrictions (e.g. restriction to low 4GiB) and so, for
271 * simplicity and similarilty between different GTT, try the single
272 * random replacement first.
274 offset = random_offset(start, end,
275 size, alignment ?: I915_GTT_MIN_ALIGNMENT);
276 err = i915_gem_gtt_reserve(vm, node, size, offset, color, flags);
280 if (flags & PIN_NOSEARCH)
283 /* Randomly selected placement is pinned, do a search */
284 err = i915_gem_evict_something(vm, size, alignment, color,
289 return drm_mm_insert_node_in_range(&vm->mm, node,
290 size, alignment, color,
291 start, end, DRM_MM_INSERT_EVICT);
294 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
295 #include "selftests/i915_gem_gtt.c"