1
0
Fork 0
mirror of synced 2025-03-06 20:59:54 +01:00
linux/drivers/gpu/drm/i915/gem/selftests/huge_gem_object.c
Matthew Auld 8c94951560 drm/i915: use i915_sg_dma_sizes() for all backends
We rely on page_sizes.sg in setup_scratch_page() reporting the correct
value if the underlying sgl is not contiguous, however in
get_pages_internal() we are only looking at the layout of the created
pages when calculating the sg_page_sizes, and not the final sgl, which
could in theory be completely different. In such a situation we might
incorrectly think we have a 64K scratch page, when it is actually only
4K or similar split over multiple non-contiguous entries, which could
lead to broken behaviour when touching the scratch space within the
padding of a 64K GTT page-table. For most of the other backends we
already just call i915_sg_dma_sizes() on the final mapping, so rather
just move that into __i915_gem_object_set_pages() to avoid such issues
coming back to bite us later.

v2: Update missing conversion in gvt

Suggested-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Cc: Stuart Summers <stuart.summers@intel.com>
Cc: Andrzej Hajda <andrzej.hajda@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20221108103238.165447-1-matthew.auld@intel.com
2022-11-09 10:07:22 +00:00

127 lines
2.8 KiB
C

/*
* SPDX-License-Identifier: MIT
*
* Copyright © 2016 Intel Corporation
*/
#include "i915_scatterlist.h"
#include "huge_gem_object.h"
static void huge_free_pages(struct drm_i915_gem_object *obj,
struct sg_table *pages)
{
unsigned long nreal = obj->scratch / PAGE_SIZE;
struct sgt_iter sgt_iter;
struct page *page;
for_each_sgt_page(page, sgt_iter, pages) {
__free_page(page);
if (!--nreal)
break;
}
sg_free_table(pages);
kfree(pages);
}
static int huge_get_pages(struct drm_i915_gem_object *obj)
{
#define GFP (GFP_KERNEL | __GFP_NOWARN | __GFP_RETRY_MAYFAIL)
const unsigned long nreal = obj->scratch / PAGE_SIZE;
const unsigned long npages = obj->base.size / PAGE_SIZE;
struct scatterlist *sg, *src, *end;
struct sg_table *pages;
unsigned long n;
pages = kmalloc(sizeof(*pages), GFP);
if (!pages)
return -ENOMEM;
if (sg_alloc_table(pages, npages, GFP)) {
kfree(pages);
return -ENOMEM;
}
sg = pages->sgl;
for (n = 0; n < nreal; n++) {
struct page *page;
page = alloc_page(GFP | __GFP_HIGHMEM);
if (!page) {
sg_mark_end(sg);
goto err;
}
sg_set_page(sg, page, PAGE_SIZE, 0);
sg = __sg_next(sg);
}
if (nreal < npages) {
for (end = sg, src = pages->sgl; sg; sg = __sg_next(sg)) {
sg_set_page(sg, sg_page(src), PAGE_SIZE, 0);
src = __sg_next(src);
if (src == end)
src = pages->sgl;
}
}
if (i915_gem_gtt_prepare_pages(obj, pages))
goto err;
__i915_gem_object_set_pages(obj, pages);
return 0;
err:
huge_free_pages(obj, pages);
return -ENOMEM;
#undef GFP
}
static void huge_put_pages(struct drm_i915_gem_object *obj,
struct sg_table *pages)
{
i915_gem_gtt_finish_pages(obj, pages);
huge_free_pages(obj, pages);
obj->mm.dirty = false;
}
static const struct drm_i915_gem_object_ops huge_ops = {
.name = "huge-gem",
.get_pages = huge_get_pages,
.put_pages = huge_put_pages,
};
struct drm_i915_gem_object *
huge_gem_object(struct drm_i915_private *i915,
phys_addr_t phys_size,
dma_addr_t dma_size)
{
static struct lock_class_key lock_class;
struct drm_i915_gem_object *obj;
unsigned int cache_level;
GEM_BUG_ON(!phys_size || phys_size > dma_size);
GEM_BUG_ON(!IS_ALIGNED(phys_size, PAGE_SIZE));
GEM_BUG_ON(!IS_ALIGNED(dma_size, I915_GTT_PAGE_SIZE));
if (overflows_type(dma_size, obj->base.size))
return ERR_PTR(-E2BIG);
obj = i915_gem_object_alloc();
if (!obj)
return ERR_PTR(-ENOMEM);
drm_gem_private_object_init(&i915->drm, &obj->base, dma_size);
i915_gem_object_init(obj, &huge_ops, &lock_class, 0);
obj->mem_flags |= I915_BO_FLAG_STRUCT_PAGE;
obj->read_domains = I915_GEM_DOMAIN_CPU;
obj->write_domain = I915_GEM_DOMAIN_CPU;
cache_level = HAS_LLC(i915) ? I915_CACHE_LLC : I915_CACHE_NONE;
i915_gem_object_set_cache_coherency(obj, cache_level);
obj->scratch = phys_size;
return obj;
}