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