alloc_one_pg_vec_page() first asks the page allocator for a physically contiguous block and falls back to vmalloc if that fails. The first attempt uses __GFP_NORETRY, but that still allows one round of direct compaction and reclaim before giving up. Under fragmentation every block of a ring pays for that round, so setting up a ring can stall the caller for seconds even though vmalloc would have worked right away. On hosts that run tcpdump periodically we saw ring setup (8 order-6 blocks) stall for up to 1.4 s per block and 3.5 s per ring. The contiguous block is only an optimization. User space maps the ring page by page either way. So make the first attempt without direct reclaim, as kvmalloc() already does since commit 46459154f997 ("mm: kvmalloc: make kmalloc fast path real fast path"). kswapd is still woken. The last resort attempt after a failed vmalloc keeps direct reclaim. Order-0 blocks are left alone, the vmalloc fallback would reclaim for them anyway. Signed-off-by: Daniil Tatianin --- net/packet/af_packet.c | 10 +++++++++- 1 file changed, 9 insertions(+), 1 deletion(-) diff --git a/net/packet/af_packet.c b/net/packet/af_packet.c index 8250e6a791f6..38f35a4f9d49 100644 --- a/net/packet/af_packet.c +++ b/net/packet/af_packet.c @@ -4493,8 +4493,16 @@ static char *alloc_one_pg_vec_page(unsigned long order) char *buffer; gfp_t gfp_flags = GFP_KERNEL_ACCOUNT | __GFP_COMP | __GFP_ZERO | __GFP_NOWARN | __GFP_NORETRY; + gfp_t first_gfp_flags = gfp_flags; - buffer = (char *) __get_free_pages(gfp_flags, order); + /* + * A contiguous block is only an optimization because vmalloc works + * too. Do not stall in reclaim or compaction to get one. + */ + if (order) + first_gfp_flags &= ~__GFP_DIRECT_RECLAIM; + + buffer = (char *) __get_free_pages(first_gfp_flags, order); if (buffer) return buffer; base-commit: 073d54dc3ccc8c8109c56e4d984d1ad616554463