i915_gem_gtt.c revision 1.19
1 /* $NetBSD: i915_gem_gtt.c,v 1.19 2021/12/18 23:45:28 riastradh Exp $ */ 2 3 // SPDX-License-Identifier: MIT 4 /* 5 * Copyright 2010 Daniel Vetter 6 * Copyright 2020 Intel Corporation 7 */ 8 9 #include <sys/cdefs.h> 10 __KERNEL_RCSID(0, "$NetBSD: i915_gem_gtt.c,v 1.19 2021/12/18 23:45:28 riastradh Exp $"); 11 12 #include <linux/slab.h> /* fault-inject.h is not standalone! */ 13 14 #include <linux/fault-inject.h> 15 #include <linux/log2.h> 16 #include <linux/random.h> 17 #include <linux/seq_file.h> 18 #include <linux/stop_machine.h> 19 20 #include <asm/set_memory.h> 21 #include <asm/smp.h> 22 23 #include <drm/i915_drm.h> 24 25 #include "display/intel_frontbuffer.h" 26 #include "gt/intel_gt.h" 27 #include "gt/intel_gt_requests.h" 28 29 #include "i915_drv.h" 30 #include "i915_scatterlist.h" 31 #include "i915_trace.h" 32 #include "i915_vgpu.h" 33 34 #ifdef __NetBSD__ 35 #include <drm/bus_dma_hacks.h> 36 #include <x86/machdep.h> 37 #include <x86/pte.h> 38 #define _PAGE_PRESENT PTE_P /* 0x01 PTE is present */ 39 #define _PAGE_RW PTE_W /* 0x02 read/write */ 40 #define _PAGE_PWT PTE_PWT /* 0x08 page write-through */ 41 #define _PAGE_PCD PTE_PCD /* 0x10 page cache disabled */ 42 #define _PAGE_PAT PTE_PAT /* 0x80 page attribute table on PTE */ 43 #endif 44 45 int i915_gem_gtt_prepare_pages(struct drm_i915_gem_object *obj, 46 struct sg_table *pages) 47 { 48 do { 49 #ifdef __NetBSD__ 50 KASSERT(0 < obj->base.size); 51 /* XXX errno NetBSD->Linux */ 52 return -bus_dmamap_load_pglist(obj->base.dev->dmat, obj->pages, 53 &obj->pageq, obj->base.size, BUS_DMA_NOWAIT); 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 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 struct device *kdev = &dev_priv->drm.pdev->dev; 83 struct i915_ggtt *ggtt = &dev_priv->ggtt; 84 85 if (unlikely(ggtt->do_idle_maps)) { 86 /* XXX This does not prevent more requests being submitted! */ 87 if (intel_gt_retire_requests_timeout(ggtt->vm.gt, 88 -MAX_SCHEDULE_TIMEOUT)) { 89 DRM_ERROR("Failed to wait for idle; VT'd may hang.\n"); 90 /* Wait a bit, in hopes it avoids the hang */ 91 udelay(10); 92 } 93 } 94 95 dma_unmap_sg(kdev, pages->sgl, pages->nents, PCI_DMA_BIDIRECTIONAL); 96 } 97 98 /** 99 * i915_gem_gtt_reserve - reserve a node in an address_space (GTT) 100 * @vm: the &struct i915_address_space 101 * @node: the &struct drm_mm_node (typically i915_vma.mode) 102 * @size: how much space to allocate inside the GTT, 103 * must be #I915_GTT_PAGE_SIZE aligned 104 * @offset: where to insert inside the GTT, 105 * must be #I915_GTT_MIN_ALIGNMENT aligned, and the node 106 * (@offset + @size) must fit within the address space 107 * @color: color to apply to node, if this node is not from a VMA, 108 * color must be #I915_COLOR_UNEVICTABLE 109 * @flags: control search and eviction behaviour 110 * 111 * i915_gem_gtt_reserve() tries to insert the @node at the exact @offset inside 112 * the address space (using @size and @color). If the @node does not fit, it 113 * tries to evict any overlapping nodes from the GTT, including any 114 * neighbouring nodes if the colors do not match (to ensure guard pages between 115 * differing domains). See i915_gem_evict_for_node() for the gory details 116 * on the eviction algorithm. #PIN_NONBLOCK may used to prevent waiting on 117 * evicting active overlapping objects, and any overlapping node that is pinned 118 * or marked as unevictable will also result in failure. 119 * 120 * Returns: 0 on success, -ENOSPC if no suitable hole is found, -EINTR if 121 * asked to wait for eviction and interrupted. 122 */ 123 int i915_gem_gtt_reserve(struct i915_address_space *vm, 124 struct drm_mm_node *node, 125 u64 size, u64 offset, unsigned long color, 126 unsigned int flags) 127 { 128 int err; 129 130 GEM_BUG_ON(!size); 131 GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE)); 132 GEM_BUG_ON(!IS_ALIGNED(offset, I915_GTT_MIN_ALIGNMENT)); 133 GEM_BUG_ON(range_overflows(offset, size, vm->total)); 134 GEM_BUG_ON(vm == &vm->i915->ggtt.alias->vm); 135 GEM_BUG_ON(drm_mm_node_allocated(node)); 136 137 node->size = size; 138 node->start = offset; 139 node->color = color; 140 141 err = drm_mm_reserve_node(&vm->mm, node); 142 if (err != -ENOSPC) 143 return err; 144 145 if (flags & PIN_NOEVICT) 146 return -ENOSPC; 147 148 err = i915_gem_evict_for_node(vm, node, flags); 149 if (err == 0) 150 err = drm_mm_reserve_node(&vm->mm, node); 151 152 return err; 153 } 154 155 static u64 random_offset(u64 start, u64 end, u64 len, u64 align) 156 { 157 u64 range, addr; 158 159 GEM_BUG_ON(range_overflows(start, len, end)); 160 GEM_BUG_ON(round_up(start, align) > round_down(end - len, align)); 161 162 range = round_down(end - len, align) - round_up(start, align); 163 if (range) { 164 if (sizeof(unsigned long) == sizeof(u64)) { 165 addr = get_random_long(); 166 } else { 167 addr = get_random_int(); 168 if (range > U32_MAX) { 169 addr <<= 32; 170 addr |= get_random_int(); 171 } 172 } 173 div64_u64_rem(addr, range, &addr); 174 start += addr; 175 } 176 177 return round_up(start, align); 178 } 179 180 /** 181 * i915_gem_gtt_insert - insert a node into an address_space (GTT) 182 * @vm: the &struct i915_address_space 183 * @node: the &struct drm_mm_node (typically i915_vma.node) 184 * @size: how much space to allocate inside the GTT, 185 * must be #I915_GTT_PAGE_SIZE aligned 186 * @alignment: required alignment of starting offset, may be 0 but 187 * if specified, this must be a power-of-two and at least 188 * #I915_GTT_MIN_ALIGNMENT 189 * @color: color to apply to node 190 * @start: start of any range restriction inside GTT (0 for all), 191 * must be #I915_GTT_PAGE_SIZE aligned 192 * @end: end of any range restriction inside GTT (U64_MAX for all), 193 * must be #I915_GTT_PAGE_SIZE aligned if not U64_MAX 194 * @flags: control search and eviction behaviour 195 * 196 * i915_gem_gtt_insert() first searches for an available hole into which 197 * is can insert the node. The hole address is aligned to @alignment and 198 * its @size must then fit entirely within the [@start, @end] bounds. The 199 * nodes on either side of the hole must match @color, or else a guard page 200 * will be inserted between the two nodes (or the node evicted). If no 201 * suitable hole is found, first a victim is randomly selected and tested 202 * for eviction, otherwise then the LRU list of objects within the GTT 203 * is scanned to find the first set of replacement nodes to create the hole. 204 * Those old overlapping nodes are evicted from the GTT (and so must be 205 * rebound before any future use). Any node that is currently pinned cannot 206 * be evicted (see i915_vma_pin()). Similar if the node's VMA is currently 207 * active and #PIN_NONBLOCK is specified, that node is also skipped when 208 * searching for an eviction candidate. See i915_gem_evict_something() for 209 * the gory details on the eviction algorithm. 210 * 211 * Returns: 0 on success, -ENOSPC if no suitable hole is found, -EINTR if 212 * asked to wait for eviction and interrupted. 213 */ 214 int i915_gem_gtt_insert(struct i915_address_space *vm, 215 struct drm_mm_node *node, 216 u64 size, u64 alignment, unsigned long color, 217 u64 start, u64 end, unsigned int flags) 218 { 219 enum drm_mm_insert_mode mode; 220 u64 offset; 221 int err; 222 223 lockdep_assert_held(&vm->mutex); 224 225 GEM_BUG_ON(!size); 226 GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE)); 227 GEM_BUG_ON(alignment && !is_power_of_2(alignment)); 228 GEM_BUG_ON(alignment && !IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT)); 229 GEM_BUG_ON(start >= end); 230 GEM_BUG_ON(start > 0 && !IS_ALIGNED(start, I915_GTT_PAGE_SIZE)); 231 GEM_BUG_ON(end < U64_MAX && !IS_ALIGNED(end, I915_GTT_PAGE_SIZE)); 232 GEM_BUG_ON(vm == &vm->i915->ggtt.alias->vm); 233 GEM_BUG_ON(drm_mm_node_allocated(node)); 234 235 if (unlikely(range_overflows(start, size, end))) 236 return -ENOSPC; 237 238 if (unlikely(round_up(start, alignment) > round_down(end - size, alignment))) 239 return -ENOSPC; 240 241 mode = DRM_MM_INSERT_BEST; 242 if (flags & PIN_HIGH) 243 mode = DRM_MM_INSERT_HIGHEST; 244 if (flags & PIN_MAPPABLE) 245 mode = DRM_MM_INSERT_LOW; 246 247 /* We only allocate in PAGE_SIZE/GTT_PAGE_SIZE (4096) chunks, 248 * so we know that we always have a minimum alignment of 4096. 249 * The drm_mm range manager is optimised to return results 250 * with zero alignment, so where possible use the optimal 251 * path. 252 */ 253 BUILD_BUG_ON(I915_GTT_MIN_ALIGNMENT > I915_GTT_PAGE_SIZE); 254 if (alignment <= I915_GTT_MIN_ALIGNMENT) 255 alignment = 0; 256 257 err = drm_mm_insert_node_in_range(&vm->mm, node, 258 size, alignment, color, 259 start, end, mode); 260 if (err != -ENOSPC) 261 return err; 262 263 if (mode & DRM_MM_INSERT_ONCE) { 264 err = drm_mm_insert_node_in_range(&vm->mm, node, 265 size, alignment, color, 266 start, end, 267 DRM_MM_INSERT_BEST); 268 if (err != -ENOSPC) 269 return err; 270 } 271 272 if (flags & PIN_NOEVICT) 273 return -ENOSPC; 274 275 /* 276 * No free space, pick a slot at random. 277 * 278 * There is a pathological case here using a GTT shared between 279 * mmap and GPU (i.e. ggtt/aliasing_ppgtt but not full-ppgtt): 280 * 281 * |<-- 256 MiB aperture -->||<-- 1792 MiB unmappable -->| 282 * (64k objects) (448k objects) 283 * 284 * Now imagine that the eviction LRU is ordered top-down (just because 285 * pathology meets real life), and that we need to evict an object to 286 * make room inside the aperture. The eviction scan then has to walk 287 * the 448k list before it finds one within range. And now imagine that 288 * it has to search for a new hole between every byte inside the memcpy, 289 * for several simultaneous clients. 290 * 291 * On a full-ppgtt system, if we have run out of available space, there 292 * will be lots and lots of objects in the eviction list! Again, 293 * searching that LRU list may be slow if we are also applying any 294 * range restrictions (e.g. restriction to low 4GiB) and so, for 295 * simplicity and similarilty between different GTT, try the single 296 * random replacement first. 297 */ 298 offset = random_offset(start, end, 299 size, alignment ?: I915_GTT_MIN_ALIGNMENT); 300 err = i915_gem_gtt_reserve(vm, node, size, offset, color, flags); 301 if (err != -ENOSPC) 302 return err; 303 304 if (flags & PIN_NOSEARCH) 305 return -ENOSPC; 306 307 /* Randomly selected placement is pinned, do a search */ 308 err = i915_gem_evict_something(vm, size, alignment, color, 309 start, end, flags); 310 if (err) 311 return err; 312 313 return drm_mm_insert_node_in_range(&vm->mm, node, 314 size, alignment, color, 315 start, end, DRM_MM_INSERT_EVICT); 316 } 317 318 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) 319 #include "selftests/i915_gem_gtt.c" 320 #endif 321
Indexes created Thu Oct 23 22:10:10 GMT 2025