Commit f8cbf6bd authored by Eric Dumazet's avatar Eric Dumazet Committed by Jakub Kicinski
Browse files

netlink: use kvmalloc() in netlink_alloc_large_skb() 



This is a followup of commit 234ec0b6 ("netlink: fix potential
sleeping issue in mqueue_flush_file"), because vfree_atomic()
overhead is unfortunate for medium sized allocations.

1) If the allocation is smaller than PAGE_SIZE, do not bother
   with vmalloc() at all. Some arches have 64KB PAGE_SIZE,
   while NLMSG_GOODSIZE is smaller than 8KB.

2) Use kvmalloc(), which might allocate one high order page
   instead of vmalloc if memory is not too fragmented.

Signed-off-by: default avatarEric Dumazet <edumazet@google.com>
Reviewed-by: default avatarZhengchao Shao <shaozhengchao@huawei.com>
Link: https://lore.kernel.org/r/20240224090630.605917-1-edumazet@google.com


Signed-off-by: default avatarJakub Kicinski <kuba@kernel.org>
parent c4b04a80

net/netlink/af_netlink.c

+8 −10
Original line number Diff line number Diff line
@@ -1202,23 +1202,21 @@ struct sock *netlink_getsockbyfilp(struct file *filp) 


struct sk_buff *netlink_alloc_large_skb(unsigned int size, int broadcast)

{

	size_t head_size = SKB_HEAD_ALIGN(size);

	struct sk_buff *skb;

	void *data;



	if (size <= NLMSG_GOODSIZE || broadcast)

	if (head_size <= PAGE_SIZE || broadcast)

		return alloc_skb(size, GFP_KERNEL);



	size = SKB_DATA_ALIGN(size) +

	       SKB_DATA_ALIGN(sizeof(struct skb_shared_info));



	data = vmalloc(size);

	if (data == NULL)

	data = kvmalloc(head_size, GFP_KERNEL);

	if (!data)

		return NULL;



	skb = __build_skb(data, size);

	if (skb == NULL)

		vfree(data);

	else

	skb = __build_skb(data, head_size);

	if (!skb)

		kvfree(data);

	else if (is_vmalloc_addr(data))

		skb->destructor = netlink_skb_destructor;



	return skb;