Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
---
drivers/gpu/drm/i915/i915_gem.c | 159 ++++++++++++++++++++++++---------------
1 file changed, 100 insertions(+), 59 deletions(-)
diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index a6a2893..9e66e29 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -663,19 +663,17 @@ out:
* needs_clflush_before is set and flushes out any written cachelines after
* writing if needs_clflush is set. */
static int
-shmem_pwrite_fast(struct page *page, int shmem_page_offset, int page_length,
+shmem_pwrite_fast(char *vaddr, int shmem_page_offset, int page_length,
char __user *user_data,
bool page_do_bit17_swizzling,
bool needs_clflush_before,
bool needs_clflush_after)
{
- char *vaddr;
int ret;
if (unlikely(page_do_bit17_swizzling))
return -EINVAL;
- vaddr = kmap_atomic(page);
if (needs_clflush_before)
drm_clflush_virt_range(vaddr + shmem_page_offset,
page_length);
@@ -685,7 +683,6 @@ shmem_pwrite_fast(struct page *page, int shmem_page_offset, int page_length,
if (needs_clflush_after)
drm_clflush_virt_range(vaddr + shmem_page_offset,
page_length);
- kunmap_atomic(vaddr);
return ret ? -EFAULT : 0;
}
@@ -693,16 +690,14 @@ shmem_pwrite_fast(struct page *page, int shmem_page_offset, int page_length,
/* Only difference to the fast-path function is that this can handle bit17
* and uses non-atomic copy and kmap functions. */
static int
-shmem_pwrite_slow(struct page *page, int shmem_page_offset, int page_length,
+shmem_pwrite_slow(char *vaddr, int shmem_page_offset, int page_length,
char __user *user_data,
bool page_do_bit17_swizzling,
bool needs_clflush_before,
bool needs_clflush_after)
{
- char *vaddr;
int ret;
- vaddr = kmap(page);
if (unlikely(needs_clflush_before || page_do_bit17_swizzling))
shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
page_length,
@@ -719,7 +714,6 @@ shmem_pwrite_slow(struct page *page, int shmem_page_offset, int page_length,
shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
page_length,
page_do_bit17_swizzling);
- kunmap(page);
return ret ? -EFAULT : 0;
}
@@ -730,10 +724,11 @@ i915_gem_shmem_pwrite(struct drm_device *dev,
struct drm_i915_gem_pwrite *args,
struct drm_file *file)
{
+ struct drm_i915_private *dev_priv = dev->dev_private;
ssize_t remain;
loff_t offset;
char __user *user_data;
- int shmem_page_offset, page_length, ret = 0;
+ int page_length, ret = 0;
int obj_do_bit17_swizzling, page_do_bit17_swizzling;
int hit_slowpath = 0;
int needs_clflush_after = 0;
@@ -769,74 +764,120 @@ i915_gem_shmem_pwrite(struct drm_device *dev,
if (ret)
return ret;
- i915_gem_object_pin_pages(obj);
-
offset = args->offset;
obj->dirty = 1;
- for_each_sg(obj->pages->sgl, sg, obj->pages->nents, i) {
- struct page *page;
- int partial_cacheline_write;
+ if (obj->stolen) {
+ char *vaddr;
- if (i < offset >> PAGE_SHIFT)
- continue;
+ vaddr = (char *)dev_priv->mm.stolen_base;
+ vaddr += obj->stolen->start + (offset & PAGE_MASK);
- if (remain <= 0)
- break;
+ offset = offset_in_page(offset);
+ while (remain > 0) {
+ int partial_cacheline_write;
- /* Operation in this page
- *
- * shmem_page_offset = offset within page in shmem file
- * page_length = bytes to copy for this page
- */
- shmem_page_offset = offset_in_page(offset);
+ page_length = remain;
+ if ((offset + page_length) > PAGE_SIZE)
+ page_length = PAGE_SIZE - offset;
- page_length = remain;
- if ((shmem_page_offset + page_length) > PAGE_SIZE)
- page_length = PAGE_SIZE - shmem_page_offset;
+ /* If we don't overwrite a cacheline completely we need to be
+ * careful to have up-to-date data by first clflushing. Don't
+ * overcomplicate things and flush the entire patch. */
+ partial_cacheline_write = needs_clflush_before &&
+ ((offset | page_length) & (boot_cpu_data.x86_clflush_size - 1));
- /* If we don't overwrite a cacheline completely we need to be
- * careful to have up-to-date data by first clflushing. Don't
- * overcomplicate things and flush the entire patch. */
- partial_cacheline_write = needs_clflush_before &&
- ((shmem_page_offset | page_length)
- & (boot_cpu_data.x86_clflush_size - 1));
+ page_do_bit17_swizzling = obj_do_bit17_swizzling &&
+ ((uintptr_t)vaddr & (1 << 17)) != 0;
- page = sg_page(sg);
- page_do_bit17_swizzling = obj_do_bit17_swizzling &&
- (page_to_phys(page) & (1 << 17)) != 0;
+ ret = shmem_pwrite_fast(vaddr, offset, page_length,
+ user_data, page_do_bit17_swizzling,
+ partial_cacheline_write,
+ needs_clflush_after);
- ret = shmem_pwrite_fast(page, shmem_page_offset, page_length,
- user_data, page_do_bit17_swizzling,
- partial_cacheline_write,
- needs_clflush_after);
- if (ret == 0)
- goto next_page;
+ if (ret == 0)
+ goto next_stolen;
- hit_slowpath = 1;
- mutex_unlock(&dev->struct_mutex);
- ret = shmem_pwrite_slow(page, shmem_page_offset, page_length,
- user_data, page_do_bit17_swizzling,
- partial_cacheline_write,
- needs_clflush_after);
+ hit_slowpath = 1;
+ mutex_unlock(&dev->struct_mutex);
- mutex_lock(&dev->struct_mutex);
+ ret = shmem_pwrite_slow(vaddr, offset, page_length,
+ user_data, page_do_bit17_swizzling,
+ partial_cacheline_write,
+ needs_clflush_after);
-next_page:
- set_page_dirty(page);
- mark_page_accessed(page);
+ mutex_lock(&dev->struct_mutex);
+ if (ret)
+ goto out;
- if (ret)
- goto out;
+next_stolen:
+ remain -= page_length;
+ user_data += page_length;
+ offset = 0;
+ }
+ } else {
+ i915_gem_object_pin_pages(obj);
- remain -= page_length;
- user_data += page_length;
- offset += page_length;
+ offset = offset_in_page(offset);
+ for_each_sg(obj->pages->sgl, sg, obj->pages->nents, i) {
+ struct page *page;
+ char *vaddr;
+ int partial_cacheline_write;
+
+ if (i < args->offset >> PAGE_SHIFT)
+ continue;
+
+ if (remain <= 0)
+ break;
+
+ page_length = remain;
+ if ((offset + page_length) > PAGE_SIZE)
+ page_length = PAGE_SIZE - offset;
+
+ /* If we don't overwrite a cacheline completely we need to be
+ * careful to have up-to-date data by first clflushing. Don't
+ * overcomplicate things and flush the entire patch. */
+ partial_cacheline_write = needs_clflush_before &&
+ ((offset | page_length) & (boot_cpu_data.x86_clflush_size - 1));
+
+ page = sg_page(sg);
+ page_do_bit17_swizzling = obj_do_bit17_swizzling &&
+ (page_to_phys(page) & (1 << 17)) != 0;
+
+ vaddr = kmap_atomic(page);
+ ret = shmem_pwrite_fast(vaddr, offset, page_length,
+ user_data, page_do_bit17_swizzling,
+ partial_cacheline_write,
+ needs_clflush_after);
+ kunmap_atomic(vaddr);
+
+ if (ret == 0)
+ goto next_page;
+
+ hit_slowpath = 1;
+ mutex_unlock(&dev->struct_mutex);
+
+ vaddr = kmap(page);
+ ret = shmem_pwrite_slow(vaddr, offset, page_length,
+ user_data, page_do_bit17_swizzling,
+ partial_cacheline_write,
+ needs_clflush_after);
+ kunmap(page);
+
+ mutex_lock(&dev->struct_mutex);
+ if (ret)
+ goto out_unpin;
+
+next_page:
+ remain -= page_length;
+ user_data += page_length;
+ offset = 0;
+ }
+out_unpin:
+ i915_gem_object_unpin_pages(obj);
}
out:
- i915_gem_object_unpin_pages(obj);
-
if (hit_slowpath) {
/* Fixup: Kill any reinstated backing storage pages */
if (obj->madv == __I915_MADV_PURGED)
--
1.7.10.4
