RE: [PATCH next v2 2/3] printk: change @clear_seq to atomic64_t

From: David Laight
Date: Mon Dec 07 2020 - 11:48:22 EST


From: John Ogness
> Sent: 07 December 2020 10:04
>
> On 2020-12-07, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> >> Yes, and it is read-only access. Perhaps atomic64_t is the wrong thing
> >> to use here. We could use a seqcount_latch and a shadow variable so that
> >> if a writer has been preempted, we can use the previous value. (Only
> >> kmsg_dump would need to use the lockless variant to read the value.)
> >>
> >> void clear_seq_set(u64 val)
> >> {
> >> spin_lock_irq(&clear_lock);
> >> raw_write_seqcount_latch(&clear_latch);
> >> clear_seq[0] = val;
> >> raw_write_seqcount_latch(&clear_latch);
> >> clear_seq[1] = val;
> >> spin_unlock_irq(&clear_lock);
> >> }
> >>
> >> u64 clear_seq_get_nolock(void)
> >> {
> >> unsigned int seq, idx;
> >> u64 val;
> >>
> >> do {
> >> seq = raw_read_seqcount_latch(&clear_latch);
> >> idx = seq & 0x1;
> >> val = clear_seq[idx];
> >> } while (read_seqcount_latch_retry(&clear_latch, seq));
> >>
> >> return val;
> >> }
> >
> > That's overly complicated.
> >
> > If you're going to double the storage you can simply do:
> >
> >
> > seq = val
> > smp_wmb();
> > seq_copy = val;
> >
> > vs
> >
> > do {
> > tmp = seq_copy;
> > smp_rmb();
> > val = seq;
> > } while (val != tmp);
>
> That will not work. We are talking about a situation where the writer is
> preempted. So seq will never equal seq_copy in that situation. I expect
> that the seqcount_latch is necessary.

Is the value just being incremented??
If so you can do:
seq_hi_0 = val >> 32;
smp_wmb();
seq_lo = val;
smp_wmb();
seq_hi_1 = val >> 32;

Then the reader can assume that seq_lo is zero if seq_h1_0 and
seq_hi_1 differ.

David



-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)