Maciej Żenczykowski wrote:
> On Mon, Apr 15, 2024 at 7:14 PM Lena Wang (王娜) <Lena.Wang@xxxxxxxxxxxx> wrote:
> >
> > On Mon, 2024-04-15 at 16:53 -0400, Willem de Bruijn wrote:
> > >
> > > External email : Please do not click links or open attachments until
> > > you have verified the sender or the content.
> > > shiming.cheng@ wrote:
> > > > From: Shiming Cheng <shiming.cheng@xxxxxxxxxxxx>
> > > >
> > > > A GRO packet without fraglist is crashed and backtrace is as below:
> > > > [ 1100.812205][ C3] CPU: 3 PID: 0 Comm: swapper/3 Tainted:
> > > > G W OE 6.6.17-android15-0-g380371ea9bf1 #1
> > > > [ 1100.812317][ C3] __udp_gso_segment+0x298/0x4d4
> > > > [ 1100.812335][ C3] __skb_gso_segment+0xc4/0x120
> > > > [ 1100.812339][ C3] udp_rcv_segment+0x50/0x134
> > > > [ 1100.812344][ C3] udp_queue_rcv_skb+0x74/0x114
> > > > [ 1100.812348][ C3] udp_unicast_rcv_skb+0x94/0xac
> > > > [ 1100.812358][ C3] udp_rcv+0x20/0x30
> > > >
> > > > The reason that the packet loses its fraglist is that in ingress
> > > bpf
> > > > it makes a test pull with to make sure it can read packet headers
> > > > via direct packet access: In bpf_progs/offload.c
> > > > try_make_writable -> bpf_skb_pull_data -> pskb_may_pull ->
> > > > __pskb_pull_tail This operation pull the data in fraglist into
> > > linear
> > > > and set the fraglist to null.
> > >
> > > What is the right behavior from BPF with regard to SKB_GSO_FRAGLIST
> > > skbs?
> > >
> > > Some, like SCTP, cannot be linearized ever, as the do not have a
> > > single gso_size.
> > >
> > > Should this BPF operation just fail?
> > >
> > In most situation for big gso size packet, it indeed fails but BPF
> > doesn't check the result. It seems the udp GRO packet can't be pulled/
> > trimed/condensed or else it can't be segmented correctly.
> >
> > As the BPF function comments it doesn't matter if the data pull failed
> > or pull less. It just does a blind best effort pull.
> >
> > A patch to modify bpf pull length is upstreamed to Google before and
> > below are part of Google BPF expert maze's reply:
> > maze@xxxxxxxxxx<maze@xxxxxxxxxx> #5Apr 13, 2024 02:30AM
> > I *think* if that patch fixes anything, then it's really proving that
> > there's a bug in the kernel that needs to be fixed instead.
> > It should be legal to call try_make_writable(skb, X) with *any* value
> > of X.
> >
> > I add maze in loop and we could start more discussion here.
>
> Personally, I think bpf_skb_pull_data() should have automatically
> (ie. in kernel code) reduced how much it pulls so that it would pull
> headers only,
That would be a helper that parses headers to discover header length.
Parsing is better left to the BPF program.
> and not packet content.
> (This is assuming the rest of the code isn't ready to deal with a longer pull,
> which I think is the case atm. Pulling too much, and then crashing or forcing
> the stack to drop packets because of them being malformed seems wrong...)
>
> In general it would be nice if there was a way to just say pull all headers...
> (or possibly all L2/L3/L4 headers)
> You in general need to pull stuff *before* you've even looked at the packet,
> so that you can look at the packet,
> so it's relatively hard/annoying to pull the correct length from bpf
> code itself.
>
> > > > BPF needs to modify a proper length to do pull data. However kernel
> > > > should also improve the flow to avoid crash from a bpf function
> > > call.
> > > > As there is no split flow and app may not decode the merged UDP
> > > packet,
> > > > we should drop the packet without fraglist in skb_segment_list
> > > here.
> > > >
> > > > Fixes: 3a1296a38d0c ("net: Support GRO/GSO fraglist chaining.")
> > > > Signed-off-by: Shiming Cheng <shiming.cheng@xxxxxxxxxxxx>
> > > > Signed-off-by: Lena Wang <lena.wang@xxxxxxxxxxxx>
> > > > ---
> > > > net/core/skbuff.c | 3 +++
> > > > 1 file changed, 3 insertions(+)
> > > >
> > > > diff --git a/net/core/skbuff.c b/net/core/skbuff.c
> > > > index b99127712e67..f68f2679b086 100644
> > > > --- a/net/core/skbuff.c
> > > > +++ b/net/core/skbuff.c
> > > > @@ -4504,6 +4504,9 @@ struct sk_buff *skb_segment_list(struct
> > > sk_buff *skb,
> > > > if (err)
> > > > goto err_linearize;
> > > >
> > > > +if (!list_skb)
> > > > +goto err_linearize;
> > > > +
This would catch the case where the entire data frag_list is
linearized, but not a pskb_may_pull that only pulls in part of the
list.
Even with BPF being privileged, the kernel should not crash if BPF
pulls a FRAGLIST GSO skb.
But the check needs to be refined a bit. For a UDP GSO packet, I
think gso_size is still valid, so if the head_skb length does not
match gso_size, it has been messed with and should be dropped.
For a GSO_BY_FRAGS skb, there is no single gso_size, and this pull
may be entirely undetectable as long as frag_list != NULL?
> > > > skb_shinfo(skb)->frag_list = NULL;
> > >
> > > In absense of plugging the issue in BPF, dropping here is the best
> > > we can do indeed, I think.
> > >
>
> --
> Maciej Żenczykowski, Kernel Networking Developer @ Google
