Re: [PATCH v3] skbuff: fix a data race in skb_queue_len()
From: Jason A. Donenfeld
Date: Mon Feb 17 2020 - 02:39:52 EST
On 2/17/20, Herbert Xu <herbert@xxxxxxxxxxxxxxxxxxx> wrote:
> Jason A. Donenfeld <Jason@xxxxxxxxx> wrote:
>> Hi Eric,
>> On Tue, Feb 04, 2020 at 01:40:29PM -0500, Qian Cai wrote:
>>> - list->qlen--;
>>> + WRITE_ONCE(list->qlen, list->qlen - 1);
>> Sorry I'm a bit late to the party here, but this immediately jumped out.
>> This generates worse code with a bigger race in some sense:
>> list->qlen-- is:
>> 0: 83 6f 10 01 subl $0x1,0x10(%rdi)
>> whereas WRITE_ONCE(list->qlen, list->qlen - 1) is:
>> 0: 8b 47 10 mov 0x10(%rdi),%eax
>> 3: 83 e8 01 sub $0x1,%eax
>> 6: 89 47 10 mov %eax,0x10(%rdi)
>> Are you sure that's what we want?
> Fixing these KCSAN warnings is actively making the kernel worse.
> Why are we still doing this?
Not necessarily a big fan of this either, but just for the record here
in case it helps, while you might complain about instruction size
blowing up a bit, cycle-wise these wind up being about the same
anyway. On x86, one instruction != one cycle.