On Wed, Oct 14, 2020 at 07:46:12AM -0700, Yonghong Song wrote:
> Commit 4fc427e05158 ("ipv6_route_seq_next should increase position index")
> tried to fix the issue where seq_file pos is not increased
> if a NULL element is returned with seq_ops->next(). See bug
> https://bugzilla.kernel.org/show_bug.cgi?id=206283
> The commit effectively does:
> - increase pos for all seq_ops->start()
> - increase pos for all seq_ops->next()
>
> For ipv6_route, increasing pos for all seq_ops->next() is correct.
> But increasing pos for seq_ops->start() is not correct
> since pos is used to determine how many items to skip during
> seq_ops->start():
> iter->skip = *pos;
> seq_ops->start() just fetches the *current* pos item.
> The item can be skipped only after seq_ops->show() which essentially
> is the beginning of seq_ops->next().
>
> For example, I have 7 ipv6 route entries,
> root@arch-fb-vm1:~/net-next dd if=/proc/net/ipv6_route bs=4096
> 00000000000000000000000000000000 40 00000000000000000000000000000000 00 00000000000000000000000000000000 00000400 00000001 00000000 00000001 eth0
> fe800000000000000000000000000000 40 00000000000000000000000000000000 00 00000000000000000000000000000000 00000100 00000001 00000000 00000001 eth0
> 00000000000000000000000000000000 00 00000000000000000000000000000000 00 00000000000000000000000000000000 ffffffff 00000001 00000000 00200200 lo
> 00000000000000000000000000000001 80 00000000000000000000000000000000 00 00000000000000000000000000000000 00000000 00000003 00000000 80200001 lo
> fe800000000000002050e3fffebd3be8 80 00000000000000000000000000000000 00 00000000000000000000000000000000 00000000 00000002 00000000 80200001 eth0
> ff000000000000000000000000000000 08 00000000000000000000000000000000 00 00000000000000000000000000000000 00000100 00000004 00000000 00000001 eth0
> 00000000000000000000000000000000 00 00000000000000000000000000000000 00 00000000000000000000000000000000 ffffffff 00000001 00000000 00200200 lo
> 0+1 records in
> 0+1 records out
> 1050 bytes (1.0 kB, 1.0 KiB) copied, 0.00707908 s, 148 kB/s
> root@arch-fb-vm1:~/net-next
>
> In the above, I specify buffer size 4096, so all records can be returned
> to user space with a single trip to the kernel.
>
> If I use buffer size 128, since each record size is 149, internally
> kernel seq_read() will read 149 into its internal buffer and return the data
> to user space in two read() syscalls. Then user read() syscall will trigger
> next seq_ops->start(). Since the current implementation increased pos even
> for seq_ops->start(), it will skip record #2, #4 and #6, assuming the first
> record is #1.
>
> root@arch-fb-vm1:~/net-next dd if=/proc/net/ipv6_route bs=128
> 00000000000000000000000000000000 40 00000000000000000000000000000000 00 00000000000000000000000000000000 00000400 00000001 00000000 00000001 eth0
> 00000000000000000000000000000000 00 00000000000000000000000000000000 00 00000000000000000000000000000000 ffffffff 00000001 00000000 00200200 lo
> fe800000000000002050e3fffebd3be8 80 00000000000000000000000000000000 00 00000000000000000000000000000000 00000000 00000002 00000000 80200001 eth0
> 00000000000000000000000000000000 00 00000000000000000000000000000000 00 00000000000000000000000000000000 ffffffff 00000001 00000000 00200200 lo
> 4+1 records in
> 4+1 records out
> 600 bytes copied, 0.00127758 s, 470 kB/s
>
> To fix the problem, create a fake pos pointer so seq_ops->start()
> won't actually increase seq_file pos. With this fix, the
> above `dd` command with `bs=128` will show correct result.
>
> Fixes: 4fc427e05158 ("ipv6_route_seq_next should increase position index")
> Cc: Andrii Nakryiko <andriin@xxxxxx>
> Cc: Alexei Starovoitov <ast@xxxxxxxxxx>
> Suggested-by: Vasily Averin <vvs@xxxxxxxxxxxxx>
> Reviewed-by: Vasily Averin <vvs@xxxxxxxxxxxxx>
> Signed-off-by: Yonghong Song <yhs@xxxxxx>
> ---
> net/ipv6/ip6_fib.c | 4 +++-
> 1 file changed, 3 insertions(+), 1 deletion(-)
>
> Changelog:
> v2 -> v3:
> - initialize local variable "p" to avoid potential syzbot complaint. (Eric)
Acked-by: Martin KaFai Lau <kafai@xxxxxx>
