[PATCH net] skbuff: back tiny skbs with kmalloc() in __netdev_alloc_skb() too

[PATCH net] skbuff: back tiny skbs with kmalloc() in __netdev_alloc_skb() too

[Alexander Lobakin]

Commit 3226b158e67c ("net: avoid 32 x truesize under-estimation for
tiny skbs") ensured that skbs with data size lower than 1025 bytes
will be kmalloc'ed to avoid excessive page cache fragmentation and
memory consumption.
However, the same issue can still be achieved manually via
__netdev_alloc_skb(), where the check for size hasn't been changed.
Mirror the condition from __napi_alloc_skb() to prevent from that.

Fixes: 3226b158e67c ("net: avoid 32 x truesize under-estimation for tiny skbs")
Signed-off-by: Alexander Lobakin <alobakin@pm.me>
---
 net/core/skbuff.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/net/core/skbuff.c b/net/core/skbuff.c
index c1a6f262636a..785daff48030 100644
--- a/net/core/skbuff.c
+++ b/net/core/skbuff.c
@@ -437,7 +437,11 @@ struct sk_buff *__netdev_alloc_skb(struct net_device *dev, unsigned int len,
 
 	len += NET_SKB_PAD;
 
-	if ((len > SKB_WITH_OVERHEAD(PAGE_SIZE)) ||
+	/* If requested length is either too small or too big,
+	 * we use kmalloc() for skb->head allocation.
+	 */
+	if (len <= SKB_WITH_OVERHEAD(1024) ||
+	    len > SKB_WITH_OVERHEAD(PAGE_SIZE) ||
 	    (gfp_mask & (__GFP_DIRECT_RECLAIM | GFP_DMA))) {
 		skb = __alloc_skb(len, gfp_mask, SKB_ALLOC_RX, NUMA_NO_NODE);
 		if (!skb)
-- 
2.30.0

Re: [PATCH net] skbuff: back tiny skbs with kmalloc() in __netdev_alloc_skb() too

[Eric Dumazet]

On Fri, Jan 15, 2021 at 12:55 AM Alexander Lobakin <alobakin@pm.me> wrote:
>
> Commit 3226b158e67c ("net: avoid 32 x truesize under-estimation for
> tiny skbs") ensured that skbs with data size lower than 1025 bytes
> will be kmalloc'ed to avoid excessive page cache fragmentation and
> memory consumption.
> However, the same issue can still be achieved manually via
> __netdev_alloc_skb(), where the check for size hasn't been changed.
> Mirror the condition from __napi_alloc_skb() to prevent from that.
>
> Fixes: 3226b158e67c ("net: avoid 32 x truesize under-estimation for tiny skbs")

No, this tag is wrong, if you fix a bug, bug is much older than linux-5.11

My fix was about GRO head and virtio_net heads, both using pre-sized
small buffers.

You want to fix something else, and this is fine, because some drivers
are unfortunately
doing copy break ( at the cost of additional copy, even for packets
that might be consumed right away)

Re: [PATCH net] skbuff: back tiny skbs with kmalloc() in __netdev_alloc_skb() too

[Alexander Lobakin]

From: Eric Dumazet <edumazet@google.com>
Date: Fri, 15 Jan 2021 15:28:37 +0100

> On Fri, Jan 15, 2021 at 12:55 AM Alexander Lobakin <alobakin@pm.me> wrote:
>>
>> Commit 3226b158e67c ("net: avoid 32 x truesize under-estimation for
>> tiny skbs") ensured that skbs with data size lower than 1025 bytes
>> will be kmalloc'ed to avoid excessive page cache fragmentation and
>> memory consumption.
>> However, the same issue can still be achieved manually via
>> __netdev_alloc_skb(), where the check for size hasn't been changed.
>> Mirror the condition from __napi_alloc_skb() to prevent from that.
>>
>> Fixes: 3226b158e67c ("net: avoid 32 x truesize under-estimation for tiny skbs")
>
> No, this tag is wrong, if you fix a bug, bug is much older than linux-5.11
>
> My fix was about GRO head and virtio_net heads, both using pre-sized
> small buffers.
>
> You want to fix something else, and this is fine, because some drivers
> are unfortunately
> doing copy break ( at the cost of additional copy, even for packets
> that might be consumed right away)

You're right, it's about copybreak. I thought about wrong "Fixes"
right after sending, but... Sorry.
Will send v2 soon.

Thanks,
Al