On 2021/9/24 15:23, Yunsheng Lin wrote:
On 2021/9/23 20:08, Jesper Dangaard Brouer wrote:
quoted
On 22/09/2021 11.41, Yunsheng Lin wrote:
quoted
Currently when PP_FLAG_PAGE_FRAG is set, the caller is not
expected to call page_pool_alloc_pages() directly because of
the PP_FLAG_PAGE_FRAG checking in __page_pool_put_page().
The patch removes the above checking to enable non-split page
support when PP_FLAG_PAGE_FRAG is set.
Reviewed-by: Alexander Duyck <alexanderduyck@fb.com>
Signed-off-by: Yunsheng Lin <redacted>
---
net/core/page_pool.c | 12 +++++++-----
1 file changed, 7 insertions(+), 5 deletions(-)
diff --git a/net/core/page_pool.c b/net/core/page_pool.c
index a65bd7972e37..f7e71dcb6a2e 100644
--- a/net/core/page_pool.c
+++ b/net/core/page_pool.c
@@ -315,11 +315,14 @@ struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp)
/* Fast-path: Get a page from cache */
page = __page_pool_get_cached(pool);
- if (page)
- return page;
/* Slow-path: cache empty, do real allocation */
- page = __page_pool_alloc_pages_slow(pool, gfp);
+ if (!page)
+ page = __page_pool_alloc_pages_slow(pool, gfp);
+
+ if (likely(page))
+ page_pool_set_frag_count(page, 1);
+
I really don't like that you add one atomic_long_set operation per page alloc call.
This is a fast-path for XDP use-cases, which you are ignoring as you drivers doesn't implement XDP.
As I cannot ask you to run XDP benchmarks, I fortunately have some page_pool specific microbenchmarks you can run instead.
I will ask you to provide before and after results from running these benchmarks [1] and [2].
[1] https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/lib/bench_page_pool_simple.c
[2] https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/lib/bench_page_pool_cross_cpu.c
How to use these module is documented here[3]:
[3] https://prototype-kernel.readthedocs.io/en/latest/prototype-kernel/build-process.html
Will running these benchmarks to see if any performance overhead noticable here,
thanks for the benchmarks.
You are right, there is notiable overhead for bench_page_pool_cross_cpu test
case above, possibly due to the cache bouncing caused by page_pool_set_frag_count().
As memntioned by Ilias, mlx5 use page pool and also do the recycling internally,
so handling the page frag tracking consistently for both PP_FLAG_PAGE_FRAG and
non-PP_FLAG_PAGE_FRAG will break the mlx5 driver.
So I will drop this patch for now.
quoted
quoted
return page;
}
EXPORT_SYMBOL(page_pool_alloc_pages);
@@ -428,8 +431,7 @@ __page_pool_put_page(struct page_pool *pool, struct page *page,
unsigned int dma_sync_size, bool allow_direct)
{
/* It is not the last user for the page frag case */
- if (pool->p.flags & PP_FLAG_PAGE_FRAG &&
- page_pool_atomic_sub_frag_count_return(page, 1))
+ if (page_pool_atomic_sub_frag_count_return(page, 1))
return NULL;
This adds an atomic_long_read, even when PP_FLAG_PAGE_FRAG is not set.
The point here is to have consistent handling for both PP_FLAG_PAGE_FRAG
and non-PP_FLAG_PAGE_FRAG case in the following patch.
As the page->_refcount is accessed in "page_ref_count(page) == 1" checking
in __page_pool_put_page(), and page->pp_frag_count is most likely in the
same cache line as the page->_refcount, So I am not expecting a noticable
overhead here.
Anyway, will use the above benchmarks as an example to verify it.
quoted
quoted
/* This allocator is optimized for the XDP mode that uses
.
_______________________________________________
Linuxarm mailing list -- linuxarm@openeuler.org
To unsubscribe send an email to linuxarm-leave@openeuler.org