Re: [PATCH 4.14] tcp: fix tcp_rtx_queue_tail in case of empty retransmit queue

From: Sasha Levin
Date: Sat Aug 31 2019 - 20:07:52 EST

On Sat, Aug 31, 2019 at 03:14:35PM +0200, Matthieu Baerts wrote:
Hi Sasha,

Thank you for your reply!

On 31/08/2019 14:20, Sasha Levin wrote:
On Sat, Aug 24, 2019 at 08:03:51AM +0200, Tim Froidcoeur wrote:
Commit 8c3088f895a0 ("tcp: be more careful in tcp_fragment()")
triggers following stack trace:

[25244.848046] kernel BUG at ./include/linux/skbuff.h:1406!
[25244.859335] RIP: 0010:skb_queue_prev+0x9/0xc
[25244.888167] Call Trace:
[25244.889182]Â <IRQ>
[25244.890001]Â tcp_fragment+0x9c/0x2cf
[25244.891295]Â tcp_write_xmit+0x68f/0x988
[25244.892732]Â __tcp_push_pending_frames+0x3b/0xa0
[25244.894347]Â tcp_data_snd_check+0x2a/0xc8
[25244.895775]Â tcp_rcv_established+0x2a8/0x30d
[25244.897282]Â tcp_v4_do_rcv+0xb2/0x158
[25244.898666]Â tcp_v4_rcv+0x692/0x956
[25244.899959]Â ip_local_deliver_finish+0xeb/0x169
[25244.901547]Â __netif_receive_skb_core+0x51c/0x582
[25244.903193]Â ? inet_gro_receive+0x239/0x247
[25244.904756]Â netif_receive_skb_internal+0xab/0xc6
[25244.906395]Â napi_gro_receive+0x8a/0xc0
[25244.907760]Â receive_buf+0x9a1/0x9cd
[25244.909160]Â ? load_balance+0x17a/0x7b7
[25244.910536]Â ? vring_unmap_one+0x18/0x61
[25244.911932]Â ? detach_buf+0x60/0xfa
[25244.913234]Â virtnet_poll+0x128/0x1e1
[25244.914607]Â net_rx_action+0x12a/0x2b1
[25244.915953]Â __do_softirq+0x11c/0x26b
[25244.917269]Â ? handle_irq_event+0x44/0x56
[25244.918695]Â irq_exit+0x61/0xa0
[25244.919947]Â do_IRQ+0x9d/0xbb
[25244.921065]Â common_interrupt+0x85/0x85
[25244.922479]Â </IRQ>

tcp_rtx_queue_tail() (called by tcp_fragment()) can call
tcp_write_queue_prev() on the first packet in the queue, which will
the BUG in tcp_write_queue_prev(), because there is no previous packet.

This happens when the retransmit queue is empty, for example in case of a
zero window.

Patch is needed for 4.4, 4.9 and 4.14 stable branches.

There needs to be a better explanation of why it's not needed

Commit 8c3088f895a0 ("tcp: be more careful in tcp_fragment()") was not a
simple cherry-pick of the original one from master (b617158dc096)
because there is a specific TCP rtx queue only since v4.15. For more
details, please see the commit message of b617158dc096 ("tcp: be more
careful in tcp_fragment()").

The BUG() is hit due to the specific code added to versions older than
v4.15. The comment in skb_queue_prev() (include/linux/skbuff.h:1406),
just before the BUG_ON() somehow suggests to add a check before using
it, what Tim did.

In master, this code path causing the issue will not be taken because
the implementation of tcp_rtx_queue_tail() is different:

tcp_fragment() â tcp_rtx_queue_tail() â tcp_write_queue_prev() â
skb_queue_prev() â BUG_ON()

Because this patch is specific to versions older than the two last
stable ones but still linked to the network architecture, who can review
and approve it? :)

Thanks for the explanation. I've changed the commit message to include
this explanation and queued it for 4.4, 4.9 and 4.14.