Re: WARNING in sk_stream_kill_queues (5)
From: Eric Dumazet
Date: Thu Dec 10 2020 - 12:42:45 EST
On Thu, Dec 10, 2020 at 5:51 PM Marco Elver <elver@xxxxxxxxxx> wrote:
>
> On Wed, Dec 09, 2020 at 01:47PM +0100, Marco Elver wrote:
> > On Tue, Dec 08, 2020 at 08:06PM +0100, Marco Elver wrote:
> > > On Thu, 3 Dec 2020 at 19:01, Eric Dumazet <eric.dumazet@xxxxxxxxx> wrote:
> > > > On 12/3/20 6:41 PM, Marco Elver wrote:
> > > >
> > > > > One more experiment -- simply adding
> > > > >
> > > > > --- a/net/core/skbuff.c
> > > > > +++ b/net/core/skbuff.c
> > > > > @@ -207,7 +207,21 @@ struct sk_buff *__alloc_skb(unsigned int size, gfp_t gfp_mask,
> > > > > */
> > > > > size = SKB_DATA_ALIGN(size);
> > > > > size += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
> > > > > + size = 1 << kmalloc_index(size); /* HACK */
> > > > > data = kmalloc_reserve(size, gfp_mask, node, &pfmemalloc);
> > > > >
> > > > >
> > > > > also got rid of the warnings. Something must be off with some value that
> > > > > is computed in terms of ksize(). If not, I don't have any explanation
> > > > > for why the above hides the problem.
> > > >
> > > > Maybe the implementations of various macros (SKB_DATA_ALIGN and friends)
> > > > hae some kind of assumptions, I will double check this.
> > >
> > > If I force kfence to return 4K sized allocations for everything, the
> > > warnings remain. That might suggest that it's not due to a missed
> > > ALIGN.
> > >
> > > Is it possible that copies or moves are a problem? E.g. we copy
> > > something from kfence -> non-kfence object (or vice-versa), and
> > > ksize() no longer matches, then things go wrong?
> >
> > I was able to narrow it down to allocations of size 640. I also narrowed
> > it down to 5 allocations that go through kfence that start triggering
> > the issue. I have attached the list of those 5 allocations with
> > allocation + free stacks. I'll try to go through them, maybe I get
> > lucky eventually. :-)
>
> [...]
>
> > kfence-#3 [0xffff88843681ac00-0xffff88843681ae7f, size=640, cache=kmalloc-1k] allocated by task 17012:
> > __kmalloc_reserve net/core/skbuff.c:142 [inline]
> > __alloc_skb+0xb8/0x3f0 net/core/skbuff.c:210
> > alloc_skb_fclone include/linux/skbuff.h:1144 [inline]
> > sk_stream_alloc_skb+0xd3/0x650 net/ipv4/tcp.c:888
> > tso_fragment net/ipv4/tcp_output.c:2124 [inline]
> > tcp_write_xmit+0x1366/0x3510 net/ipv4/tcp_output.c:2674
> > __tcp_push_pending_frames+0x68/0x1f0 net/ipv4/tcp_output.c:2866
> > tcp_push_pending_frames include/net/tcp.h:1864 [inline]
> > tcp_data_snd_check net/ipv4/tcp_input.c:5374 [inline]
> > tcp_rcv_established+0x57c/0x10b0 net/ipv4/tcp_input.c:5869
> > tcp_v4_do_rcv+0x361/0x4c0 net/ipv4/tcp_ipv4.c:1668
> > sk_backlog_rcv include/net/sock.h:1010 [inline]
> > __release_sock+0xd7/0x260 net/core/sock.c:2523
> > release_sock+0x40/0x120 net/core/sock.c:3053
> > sk_wait_data+0x127/0x2b0 net/core/sock.c:2565
> > tcp_recvmsg+0x1106/0x1b60 net/ipv4/tcp.c:2181
> > inet_recvmsg+0xb1/0x270 net/ipv4/af_inet.c:848
> > sock_recvmsg_nosec net/socket.c:885 [inline]
> > sock_recvmsg net/socket.c:903 [inline]
> > sock_recvmsg net/socket.c:899 [inline]
> > ____sys_recvmsg+0x2fd/0x3a0 net/socket.c:2563
> > ___sys_recvmsg+0xd9/0x1b0 net/socket.c:2605
> > __sys_recvmsg+0x8b/0x130 net/socket.c:2641
> > __do_sys_recvmsg net/socket.c:2651 [inline]
> > __se_sys_recvmsg net/socket.c:2648 [inline]
> > __x64_sys_recvmsg+0x43/0x50 net/socket.c:2648
> > do_syscall_64+0x34/0x80 arch/x86/entry/common.c:46
> > entry_SYSCALL_64_after_hwframe+0x44/0xa9
>
> So I started putting gdb to work, and whenever I see an allocation
> exactly like the above that goes through tso_fragment() a warning
> immediately follows.
>
> Long story short, I somehow synthesized this patch that appears to fix
> things, but I can't explain why exactly:
>
> | --- a/net/core/skbuff.c
> | +++ b/net/core/skbuff.c
> | @@ -1679,13 +1679,6 @@ int pskb_expand_head(struct sk_buff *skb, int nhead, int ntail,
> |
> | skb_metadata_clear(skb);
> |
> | - /* It is not generally safe to change skb->truesize.
> | - * For the moment, we really care of rx path, or
> | - * when skb is orphaned (not attached to a socket).
> | - */
> | - if (!skb->sk || skb->destructor == sock_edemux)
> | - skb->truesize += size - osize;
> | -
> | return 0;
> |
> | nofrags:
>
> Now, here are the breadcrumbs I followed:
>
>
> 1. Breakpoint on kfence_ksize() -- first allocation that matches the above:
>
> | #0 __kfence_ksize (s=18446612700164612096) at mm/kfence/core.c:726
> | #1 0xffffffff816fbf30 in kfence_ksize (addr=0xffff888436856000) at mm/kfence/core.c:737
> | #2 0xffffffff816217cf in ksize (objp=0xffff888436856000) at mm/slab_common.c:1178
> | #3 0xffffffff84896911 in __alloc_skb (size=914710528, gfp_mask=2592, flags=0, node=-1) at net/core/skbuff.c:217
> | #4 0xffffffff84d0ba73 in alloc_skb_fclone (priority=<optimized out>, size=<optimized out>) at ./include/linux/skbuff.h:1144
> | #5 sk_stream_alloc_skb (sk=0xffff8881176cc000, size=0, gfp=2592, force_schedule=232) at net/ipv4/tcp.c:888
> | #6 0xffffffff84d41c36 in tso_fragment (gfp=<optimized out>, mss_now=<optimized out>, len=<optimized out>,
> | skb=<optimized out>, sk=<optimized out>) at net/ipv4/tcp_output.c:2124
> | #7 tcp_write_xmit (sk=0xffff8881176cc000, mss_now=21950, nonagle=3096, push_one=-1996874776, gfp=0)
> | at net/ipv4/tcp_output.c:2674
> | #8 0xffffffff84d43e48 in __tcp_push_pending_frames (sk=0xffff8881176cc000, cur_mss=337, nonagle=0)
> | at ./include/net/sock.h:918
> | #9 0xffffffff84d3259c in tcp_push_pending_frames (sk=<optimized out>) at ./include/net/tcp.h:1864
> | #10 tcp_data_snd_check (sk=<optimized out>) at net/ipv4/tcp_input.c:5374
> | #11 tcp_rcv_established (sk=0xffff8881176cc000, skb=0x0 <fixed_percpu_data>) at net/ipv4/tcp_input.c:5869
> | #12 0xffffffff84d56731 in tcp_v4_do_rcv (sk=0xffff8881176cc000, skb=0xffff888117f52ea0) at net/ipv4/tcp_ipv4.c:1668
> | [...]
>
> Set watchpoint on skb->truesize:
>
> | (gdb) frame 3
> | #3 0xffffffff84896911 in __alloc_skb (size=914710528, gfp_mask=2592, flags=0, node=-1) at net/core/skbuff.c:217
> | 217 size = SKB_WITH_OVERHEAD(ksize(data));
> | (gdb) p &skb->truesize
> | $5 = (unsigned int *) 0xffff888117f55f90
> | (gdb) awatch *0xffff888117f55f90
> | Hardware access (read/write) watchpoint 6: *0xffff888117f55f90
>
> 2. Some time later, we see that the skb with kfence-allocated data
> is cloned:
>
> | Thread 7 hit Hardware access (read/write) watchpoint 6: *0xffff888117f55f90
> |
> | Value = 1570
> | 0xffffffff84886947 in __skb_clone (n=0xffff888117f55fa0, skb=0xffff888117f55ec0) at net/core/skbuff.c:1002
> | 1002 C(truesize);
> | (gdb) bt
> | #0 0xffffffff84886947 in __skb_clone (n=0xffff888117f55fa0, skb=0xffff888117f55ec0) at net/core/skbuff.c:1002
> | #1 0xffffffff8488bfb9 in skb_clone (skb=0xffff888117f55ec0, gfp_mask=2592) at net/core/skbuff.c:1454
> | #2 0xffffffff84d3cd1c in __tcp_transmit_skb (sk=0xffff8881176cc000, skb=0xffff888117f55ec0, clone_it=0, gfp_mask=2592,
> | rcv_nxt=0) at net/ipv4/tcp_output.c:1267
> | #3 0xffffffff84d4125b in tcp_transmit_skb (gfp_mask=<optimized out>, clone_it=<optimized out>, skb=<optimized out>,
> | sk=<optimized out>) at ./include/linux/tcp.h:439
> | #4 tcp_write_xmit (sk=0xffff8881176cc000, mss_now=392485600, nonagle=1326, push_one=-1996875104, gfp=0)
> | at net/ipv4/tcp_output.c:2688
> | #5 0xffffffff84d43e48 in __tcp_push_pending_frames (sk=0xffff8881176cc000, cur_mss=337, nonagle=0)
> | at ./include/net/sock.h:918
> | #6 0xffffffff84d3259c in tcp_push_pending_frames (sk=<optimized out>) at ./include/net/tcp.h:1864
> | #7 tcp_data_snd_check (sk=<optimized out>) at net/ipv4/tcp_input.c:5374
> | #8 tcp_rcv_established (sk=0xffff8881176cc000, skb=0x0 <fixed_percpu_data>) at net/ipv4/tcp_input.c:5869
> | #9 0xffffffff84d56731 in tcp_v4_do_rcv (sk=0xffff8881176cc000, skb=0xffff888117f57820) at net/ipv4/tcp_ipv4.c:1668
> | #10 0xffffffff8487bf67 in sk_backlog_rcv (skb=<optimized out>, sk=<optimized out>) at ./include/net/sock.h:1010
> [...]
>
>
> 3. The original skb (that was cloned) has its truesize adjusted
> after a pskb_expand_head():
>
> | Thread 2 hit Hardware access (read/write) watchpoint 6: *0xffff888117f55f90
> |
> | Old value = 1570
> | New value = 1954
>
> ^^ the difference between the old and the new value is exactly
> 384, which is also the final underflow of the sk_wmem_queued
> that triggers the warning. Presumably if the original allocation
> had been through kmalloc-1k and not KFENCE, the difference here
> would have been 0, since ksize() of the original allocation in
> step (1) would have been 1024, and not 640 (difference of 384).
>
> | 0xffffffff8488d84b in pskb_expand_head (skb=0xffff888117f55ec0, nhead=401956752, ntail=1954, gfp_mask=2298092192)
> | at net/core/skbuff.c:1687
> | 1687 skb->truesize += size - osize;
> | (gdb) bt
> | #0 0xffffffff8488d84b in pskb_expand_head (skb=0xffff888117f55ec0, nhead=401956752, ntail=1954, gfp_mask=2298092192)
> | at net/core/skbuff.c:1687
> | #1 0xffffffff8488de01 in skb_prepare_for_shift (skb=<optimized out>) at ./arch/x86/include/asm/atomic.h:29
> | #2 skb_prepare_for_shift (skb=0xffff888117f55ec0) at net/core/skbuff.c:3276
> | #3 0xffffffff848936b1 in skb_shift (tgt=0xffff888117f549c0, skb=0xffff888117f55ec0, shiftlen=674) at net/core/skbuff.c:3351
> | #4 0xffffffff84d264de in tcp_skb_shift (shiftlen=<optimized out>, pcount=<optimized out>, from=<optimized out>,
> | to=<optimized out>) at net/ipv4/tcp_input.c:1497
> | #5 tcp_shift_skb_data (dup_sack=<optimized out>, end_seq=<optimized out>, start_seq=<optimized out>, state=<optimized out>,
> | skb=<optimized out>, sk=<optimized out>) at net/ipv4/tcp_input.c:1605
> | #6 tcp_sacktag_walk (skb=0xffff888117f55ec0, sk=0xffff8881176cc000, next_dup=0x894,
> | state=0xffffffff88fa1aa0 <watchpoints+192>, start_seq=0, end_seq=401956752, dup_sack_in=false)
> | at net/ipv4/tcp_input.c:1670
> | #7 0xffffffff84d276de in tcp_sacktag_write_queue (sk=0xffff888117f55f90, ack_skb=0x1888117f55f90, prior_snd_una=2196,
> | state=0xffffffff88fa1aa0 <watchpoints+192>) at net/ipv4/tcp_input.c:1931
> | #8 0xffffffff84d2ca1d in tcp_ack (sk=0xffff8881176cc000, skb=0x1888117f55f90, flag=16643) at net/ipv4/tcp_input.c:3758
> | #9 0xffffffff84d32387 in tcp_rcv_established (sk=0xffff8881176cc000, skb=0xffff888117f54020) at net/ipv4/tcp_input.c:5858
> | #10 0xffffffff84d56731 in tcp_v4_do_rcv (sk=0xffff8881176cc000, skb=0xffff888117f54020) at net/ipv4/tcp_ipv4.c:1668
> [...]
>
>
> Any of this make sense?
Very nice debugging !
I guess we could fix this in skb_prepare_for_shift(), eventually
caring for the truesize manipulation
(or reverting the change done in pskb_expand_head(), since only kfence
is having this issue.
(All TCP skbs in output path have the same allocation size for skb->head)
diff --git a/net/core/skbuff.c b/net/core/skbuff.c
index e578544b2cc7110ec2f6bcf4c29d93e4b4b1ad14..798b51eeeaa4fbed65d41d9eab207dbbf438dab3
100644
--- a/net/core/skbuff.c
+++ b/net/core/skbuff.c
@@ -3270,7 +3270,14 @@ EXPORT_SYMBOL(skb_split);
*/
static int skb_prepare_for_shift(struct sk_buff *skb)
{
- return skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
+ unsigned int ret = 0, save;
+
+ if (skb_cloned(skb)) {
+ save = skb->truesize;
+ ret = pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
+ skb->truesize = save;
+ }
+ return ret;
}