[CRIU] [PATCH] tcp: Try harder to restore recv queue

[Andrew Vagin]

On Thu, Dec 24, 2015 at 10:50:46PM +0300, Pavel Emelyanov wrote:
> On restore we try to put data back into recv queue with quite
> big chunks. However the kernel doesn't try hard to split the
> data into packets in repair mode for this queue and just
> allocates the linear skb of the given size. If the size is
> moderately big, the allocation is subject to fail, slab doesn't
> reliably allocates memory over 4k.
> 
> So, when failing with big chunk on recv queue -- shrink the
> chunk and try again.
>

Acked-by: Andrew Vagin <avagin at virtuozzo.com> 
> Signed-off-by: Pavel Emelyanov <xemul at parallels.com>
> ---
>  sk-tcp.c | 25 ++++++++++++++++++++-----
>  1 file changed, 20 insertions(+), 5 deletions(-)
> 
> diff --git a/sk-tcp.c b/sk-tcp.c
> index 9f3bf0d..b5e66b0 100644
> --- a/sk-tcp.c
> +++ b/sk-tcp.c
> @@ -461,7 +461,7 @@ static int restore_tcp_seqs(int sk, TcpStreamEntry *tse)
>  
>  static int __send_tcp_queue(int sk, int queue, u32 len, struct cr_img *img)
>  {
> -	int ret, err = -1;
> +	int ret, err = -1, max_chunk;
>  	int off;
>  	char *buf;
>  
> @@ -472,17 +472,32 @@ static int __send_tcp_queue(int sk, int queue, u32 len, struct cr_img *img)
>  	if (read_img_buf(img, buf, len) < 0)
>  		goto err;
>  
> +	max_chunk = (queue == TCP_RECV_QUEUE ? kdat.tcp_max_rshare : len);
>  	off = 0;
>  	while (len) {
>  		int chunk = len;
>  
> -		if (queue == TCP_RECV_QUEUE && len > kdat.tcp_max_rshare)
> -			chunk = kdat.tcp_max_rshare;
> +		if (chunk > max_chunk)
> +			chunk = max_chunk;
>  
>  		ret = send(sk, buf + off, chunk, 0);
>  		if (ret <= 0) {
> -			pr_perror("Can't restore %d queue data (%d), want (%d:%d)",
> -				  queue, ret, chunk, len);
> +			if ((queue == TCP_RECV_QUEUE) && (max_chunk > 1024) && (errno == ENOMEM)) {
> +				/*
> +				 * When restoring recv queue in repair mode
> +				 * kernel doesn't try hard and just allocates
> +				 * a linear skb with the size we pass to the
> +				 * system call. Thus, if the size is too big
> +				 * for slab allocator, the send just fails
> +				 * with ENOMEM. Try smaller chunk, hopefully
> +				 * there's still enough memory in the system.
> +				 */
> +				max_chunk >>= 1;
> +				continue;
> +			}
> +
> +			pr_perror("Can't restore %d queue data (%d), want (%d:%d:%d)",
> +				  queue, ret, chunk, len, max_chunk);
>  			goto err;
>  		}
>  		off += ret;
> -- 
> 1.9.3
> 
> _______________________________________________
> CRIU mailing list
> CRIU at openvz.org
> https://lists.openvz.org/mailman/listinfo/criu