[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <4967C95E.3070602@cosmosbay.com>
Date: Fri, 09 Jan 2009 23:02:06 +0100
From: Eric Dumazet <dada1@...mosbay.com>
To: Willy Tarreau <w@....eu>
CC: David Miller <davem@...emloft.net>, ben@...s.com,
jarkao2@...il.com, mingo@...e.hu, linux-kernel@...r.kernel.org,
netdev@...r.kernel.org, jens.axboe@...cle.com
Subject: Re: [PATCH] tcp: splice as many packets as possible at once
Willy Tarreau a écrit :
> On Fri, Jan 09, 2009 at 09:51:17PM +0100, Eric Dumazet wrote:
> (...)
>>> Also, in your second mail, you're saying that your change
>>> might return more data than requested by the user. I can't
>>> find why, could you please explain to me, as I'm still quite
>>> ignorant in this area ?
>> Well, I just tested various user programs and indeed got this
>> strange result :
>>
>> Here I call splice() with len=1000 (0x3e8), and you can see
>> it gives a result of 1460 at the second call.
>
> huh, not nice indeed!
>
> While looking at the code to see how this could be possible, I
> came across this minor thing (unrelated IMHO) :
>
> if (__skb_splice_bits(skb, &offset, &tlen, &spd))
> goto done;
>>>>>>> else if (!tlen) <<<<<<
> goto done;
>
> /*
> * now see if we have a frag_list to map
> */
> if (skb_shinfo(skb)->frag_list) {
> struct sk_buff *list = skb_shinfo(skb)->frag_list;
>
> for (; list && tlen; list = list->next) {
> if (__skb_splice_bits(list, &offset, &tlen, &spd))
> break;
> }
> }
>
> done:
>
> Above on the enlighted line, we'd better remove the else and leave a plain
> "if (!tlen)". Otherwise, when the first call to __skb_splice_bits() zeroes
> tlen, we still enter the if and evaluate the for condition for nothing. But
> let's leave that for later.
>
>> I suspect a bug in splice code, that my patch just exposed.
>
> I've checked in skb_splice_bits() and below and can't see how we can move
> more than the requested len.
>
> However, with your change, I don't clearly see how we break out of
> the loop in tcp_read_sock(). Maybe we first read 1000 then loop again
> and read remaining data ? I suspect that we should at least exit when
> ((struct tcp_splice_state *)desc->arg.data)->len = 0.
>
> At least that's something easy to add just before or after !desc->count
> for a test.
>
I believe the bug is in tcp_splice_data_recv()
I am going to test a new patch, but here is the thing I found:
diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index bd6ff90..fbbddf4 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -523,7 +523,7 @@ static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
{
struct tcp_splice_state *tss = rd_desc->arg.data;
- return skb_splice_bits(skb, offset, tss->pipe, tss->len, tss->flags);
+ return skb_splice_bits(skb, offset, tss->pipe, len, tss->flags);
}
static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
Powered by blists - more mailing lists