[PATCH] drm/ttm: Merge hugepage attr changes in ttm_dma_page_put.
Zhang, Jerry (Junwei)
Jerry.Zhang at amd.com
Thu Jul 26 05:52:31 UTC 2018
On 07/26/2018 04:29 AM, Bas Nieuwenhuizen wrote:
> Every set_pages_array_wb call resulted in cross-core
> interrupts and TLB flushes. Merge more of them for
> less overhead.
>
> This reduces the time needed to free a 1.6 GiB GTT WC
> buffer as part of Vulkan CTS from ~2 sec to < 0.25 sec.
> (Allocation still takes more than 2 sec though)
>
> Signed-off-by: Bas Nieuwenhuizen <basni at chromium.org>
> ---
> drivers/gpu/drm/ttm/ttm_page_alloc_dma.c | 31 ++++++++++++++++++------
> 1 file changed, 24 insertions(+), 7 deletions(-)
>
> diff --git a/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c b/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c
> index 4c659405a008a..9440ba0a55116 100644
> --- a/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c
> +++ b/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c
> @@ -299,6 +299,25 @@ static int set_pages_array_uc(struct page **pages, int addrinarray)
> #endif
> return 0;
> }
> +
> +static int ttm_set_page_range_wb(struct page *p, unsigned long numpages)
> +{
> +#if IS_ENABLED(CONFIG_AGP)
> + unsigned long i;
> +
> + for (i = 0; i < numpages; i++)
> + unmap_page_from_agp(p + i);
> +#endif
> + return 0;
> +}
> +
> +#else /* for !CONFIG_X86 */
> +
> +static int ttm_set_page_range_wb(struct page *p, unsigned long numpages)
> +{
> + return set_memory_wb((unsigned long)page_address(p), numpages);
> +}
> +
> #endif /* for !CONFIG_X86 */
>
> static int ttm_set_pages_caching(struct dma_pool *pool,
> @@ -387,18 +406,16 @@ static void ttm_pool_update_free_locked(struct dma_pool *pool,
> static void ttm_dma_page_put(struct dma_pool *pool, struct dma_page *d_page)
> {
> struct page *page = d_page->p;
> - unsigned i, num_pages;
> + unsigned num_pages;
> int ret;
>
> /* Don't set WB on WB page pool. */
> if (!(pool->type & IS_CACHED)) {
> num_pages = pool->size / PAGE_SIZE;
> - for (i = 0; i < num_pages; ++i, ++page) {
> - ret = set_pages_array_wb(&page, 1);
> - if (ret) {
> - pr_err("%s: Failed to set %d pages to wb!\n",
> - pool->dev_name, 1);
> - }
> + ret = ttm_set_page_range_wb(page, num_pages);
For AGP enabled, set_pages_array_wc() could works like that by passing "num_pages" instead of "1"
In X86 case, we may use set_pages_array_wb() in arch/x86/mm/pageattr.c.
so, does it work as below?
ret = set_pages_array_wb(page, num_pages);
Jerry
> + if (ret) {
> + pr_err("%s: Failed to set %d pages to wb!\n",
> + pool->dev_name, num_pages);
> }
> }
>
>
More information about the amd-gfx
mailing list