drivers/gpu/drm/i915/i915_gem_gtt.c

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