On Sat, Feb 13, 2021 at 6:10 AM Alexander Lobakin <aloba...@pm.me> wrote:
>
> Currently, all sorts of skb allocation always do allocate
> skbuff_heads one by one via kmem_cache_alloc().
> On the other hand, we have percpu napi_alloc_cache to store
> skbuff_heads queued up for freeing and flush them by bulks.
>
> We can use this cache not only for bulk-wiping, but also to obtain
> heads for new skbs and avoid unconditional allocations, as well as
> for bulk-allocating (like XDP's cpumap code and veth driver already
> do).
>
> As this might affect latencies, cache pressure and lots of hardware
> and driver-dependent stuff, this new feature is mostly optional and
> can be issued via:
> - a new napi_build_skb() function (as a replacement for build_skb());
> - existing {,__}napi_alloc_skb() and napi_get_frags() functions;
> - __alloc_skb() with passing SKB_ALLOC_NAPI in flags.
>
> iperf3 showed 35-70 Mbps bumps for both TCP and UDP while performing
> VLAN NAT on 1.2 GHz MIPS board. The boost is likely to be bigger
> on more powerful hosts and NICs with tens of Mpps.
>
> Note on skbuff_heads from distant slabs or pfmemalloc'ed slabs:
> - kmalloc()/kmem_cache_alloc() itself allows by default allocating
> memory from the remote nodes to defragment their slabs. This is
> controlled by sysctl, but according to this, skbuff_head from a
> remote node is an OK case;
> - The easiest way to check if the slab of skbuff_head is remote or
> pfmemalloc'ed is:
>
> if (!dev_page_is_reusable(virt_to_head_page(skb)))
> /* drop it */;
>
> ...*but*, regarding that most slabs are built of compound pages,
> virt_to_head_page() will hit unlikely-branch every single call.
> This check costed at least 20 Mbps in test scenarios and seems
> like it'd be better to _not_ do this.
<snip>
> Alexander Lobakin (11):
> skbuff: move __alloc_skb() next to the other skb allocation functions
> skbuff: simplify kmalloc_reserve()
> skbuff: make __build_skb_around() return void
> skbuff: simplify __alloc_skb() a bit
> skbuff: use __build_skb_around() in __alloc_skb()
> skbuff: remove __kfree_skb_flush()
> skbuff: move NAPI cache declarations upper in the file
> skbuff: introduce {,__}napi_build_skb() which reuses NAPI cache heads
> skbuff: allow to optionally use NAPI cache from __alloc_skb()
> skbuff: allow to use NAPI cache from __napi_alloc_skb()
> skbuff: queue NAPI_MERGED_FREE skbs into NAPI cache instead of freeing
>
> include/linux/skbuff.h | 4 +-
> net/core/dev.c | 16 +-
> net/core/skbuff.c | 428 +++++++++++++++++++++++------------------
> 3 files changed, 242 insertions(+), 206 deletions(-)
>
With the last few changes and testing to verify the need to drop the
cache clearing this patch set looks good to me.
Reviewed-by: Alexander Duyck <alexanderdu...@fb.com>
