[CRIU] [PATCH 2/2] tcp: don't split packets for restoring a send queue

[Pavel Emelyanov]

On 11/21/2014 05:10 PM, Andrey Vagin wrote:
> The kernel can do it better. The problem exists only for recv queues.

But what the problem is? We still can feed recv queue with
small chunks, kernel just doesn't eat _parts_ of it. If the
recv queue is too big, you try to feed it all at once with
this fix, which increases the risk of failure.

> Signed-off-by: Andrey Vagin <avagin at openvz.org>
> ---
>  include/kerndat.h | 1 -
>  kerndat.c         | 7 ++-----
>  sk-tcp.c          | 8 +++++---
>  3 files changed, 7 insertions(+), 9 deletions(-)
> 
> diff --git a/include/kerndat.h b/include/kerndat.h
> index 1f511cf..12517d1 100644
> --- a/include/kerndat.h
> +++ b/include/kerndat.h
> @@ -16,7 +16,6 @@ extern int kerndat_get_dirty_track(void);
>  
>  struct kerndat_s {
>  	dev_t shmem_dev;
> -	int tcp_max_wshare;
>  	int tcp_max_rshare;
>  	int last_cap;
>  	u64 zero_page_pfn;
> diff --git a/kerndat.c b/kerndat.c
> index 68c3e35..7855066 100644
> --- a/kerndat.c
> +++ b/kerndat.c
> @@ -19,7 +19,6 @@
>  #include "util.h"
>  
>  struct kerndat_s kdat = {
> -	.tcp_max_wshare = 2U << 20,
>  	.tcp_max_rshare = 3U << 20,
>  };
>  
> @@ -197,7 +196,6 @@ static int tcp_read_sysctl_limits(void)
>  	int ret;
>  
>  	struct sysctl_req req[] = {
> -		{ "net/ipv4/tcp_wmem", &vect[0], CTL_U32A(ARRAY_SIZE(vect[0])) },
>  		{ "net/ipv4/tcp_rmem", &vect[1], CTL_U32A(ARRAY_SIZE(vect[1])) },
>  		{ },
>  	};
> @@ -212,13 +210,12 @@ static int tcp_read_sysctl_limits(void)
>  		goto out;
>  	}
>  
> -	kdat.tcp_max_wshare = min(kdat.tcp_max_wshare, (int)vect[0][2]);
>  	kdat.tcp_max_rshare = min(kdat.tcp_max_rshare, (int)vect[1][2]);
>  
> -	if (kdat.tcp_max_wshare < 128 || kdat.tcp_max_rshare < 128)
> +	if (kdat.tcp_max_rshare < 128)
>  		pr_warn("The memory limits for TCP queues are suspiciously small\n");
>  out:
> -	pr_debug("TCP queue memory limits are %d:%d\n", kdat.tcp_max_wshare, kdat.tcp_max_rshare);
> +	pr_debug("TCP recv queue memory limit is %d\n", kdat.tcp_max_rshare);
>  	return 0;
>  }
>  
> diff --git a/sk-tcp.c b/sk-tcp.c
> index 3ef7015..e280ee6 100644
> --- a/sk-tcp.c
> +++ b/sk-tcp.c
> @@ -456,7 +456,7 @@ 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 off, max;
> +	int off;
>  	char *buf;
>  
>  	buf = xmalloc(len);
> @@ -466,10 +466,12 @@ 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 = (queue == TCP_SEND_QUEUE) ? kdat.tcp_max_wshare : kdat.tcp_max_rshare;
>  	off = 0;
>  	while (len) {
> -		int chunk = (len > max ? max : len);
> +		int chunk = len;
> +
> +		if (queue == TCP_RECV_QUEUE && len > kdat.tcp_max_rshare)
> +			chunk = kdat.tcp_max_rshare;
>  
>  		ret = send(sk, buf + off, chunk, 0);
>  		if (ret <= 0) {
>