On Thu, Dec 15, 2022 at 09:08:34AM -0800, Paul E. McKenney wrote:
> On Thu, Dec 15, 2022 at 05:54:52PM +0100, Frederic Weisbecker wrote:
> > On Wed, Dec 14, 2022 at 11:13:55AM -0800, Paul E. McKenney wrote:
> > > The comment in srcu_readers_active_idx_check() following the smp_mb()
> > > is out of date, hailing from a simpler time when preemption was disabled
> > > across the bulk of __srcu_read_lock(). The fact that preemption was
> > > disabled meant that the number of tasks that had fetched the old index
> > > but not yet incremented counters was limited by the number of CPUs.
> > >
> > > In our more complex modern times, the number of CPUs is no longer a limit.
> > > This commit therefore updates this comment, additionally giving more
> > > memory-ordering detail.
> > >
> > > Reported-by: Boqun Feng <boqun.feng@xxxxxxxxx>
> > > Reported-by: Frederic Weisbecker <frederic@xxxxxxxxxx>
> >
> > Not really, while you guys were debating on that comment, I was still starring
> > at the previous one (as usual).
> >
> > Or to put it in an SRCU way, while you guys saw the flipped idx, I was still
> > using the old one :)
> >
> > > - * OK, how about nesting? This does impose a limit on nesting
> > > - * of floor(ULONG_MAX/NR_CPUS/2), which should be sufficient,
> > > - * especially on 64-bit systems.
> > > + * It can clearly do so once, given that it has already fetched
> > > + * the old value of ->srcu_idx and is just about to use that value
> > > + * to index its increment of ->srcu_lock_count[idx]. But as soon as
> > > + * it leaves that SRCU read-side critical section, it will increment
> > > + * ->srcu_unlock_count[idx], which must follow the updater's above
> > > + * read from that same value. Thus, as soon the reading task does
> > > + * an smp_mb() and a later fetch from ->srcu_idx, that task will be
> > > + * guaranteed to get the new index. Except that the increment of
> > > + * ->srcu_unlock_count[idx] in __srcu_read_unlock() is after the
> > > + * smp_mb(), and the fetch from ->srcu_idx in __srcu_read_lock()
> > > + * is before the smp_mb(). Thus, that task might not see the new
> > > + * value of ->srcu_idx until the -second- __srcu_read_lock(),
> > > + * which in turn means that this task might well increment
> > > + * ->srcu_lock_count[idx] for the old value of ->srcu_idx twice,
> > > + * not just once.
> >
> > You lost me on that one.
> >
> > UPDATER READER
> > ------- ------
> > //srcu_readers_lock_idx //srcu_read_lock
> > idx = ssp->srcu_idx; idx = ssp->srcu_idx;
> > READ srcu_lock_count[idx ^ 1] srcu_lock_count[idx]++
>
> Shouldn't this be "READ srcu_unlock_count[idx ^ 1]"?
>
> And then the above paragraph assumes that the updater gets stuck here...
Right I ignored the unlock part on purpose. But ok let's add it (later note: just switch
directly to the next paragraph to see how I realize I'm wrong)
UPDATER READER
------- ------
idx = ssp->srcu_idx; idx = ssp->srcu_idx;
READ srcu_unlock_count[idx ^ 1] srcu_lock_count[idx]++
smp_mb(); smp_mb();
READ srcu_lock_count[idx ^ 1] // read side crit
smp_mb(); smp_mb();
idx = ssp->srcu_idx; srcu_unlock_count[old_idx]++
ssp->srcu_idx++; idx = ssp->srcu_idx;
smp_mb();
READ srcu_unlock_count[idx ^ 1]
smp_mb();
READ srcu_lock_count[idx ^ 1]
> Unless I am missing something, the reader must reference the
> srcu_unlock_count[old_idx] and then do smp_mb() before it will be
> absolutely guaranteed of seeing the new value of ->srcu_idx.
>
> So what am I missing?
But there is the smp_mb() between the srcu_lock_count[idx]++ of the 1st
srcu_read_lock() and the idx READ from the second srcu_read_lock():
WRITER READER
----- -------
WRITE idx WRITE srcu_lock_count[old_idx]
smp_mb() smp_mb()
READ srcu_lock_count[new_idx] READ idx
Ah wait! On SCAN2 we are reading the count from the _new_ idx, not the old one, ok
that's why it doesn't work. So then for it to write twice on the old idx we have:
_ idx is initially 0
_ READER fetches idx (idx=0) and is preempted
_ New GP: Updater goes through its whole thing and flips idx
_ Yet another new GP: Updater goes again but is preempted in the middle of
SCAN1: it has read unlock_count but not yet lock_count
_ READER increments lock_count, then unlock_count, for the old idx (0).
_ New READER: indeed we don't have a barrier between unlock_count and idx read.
So we read again the idx unordered against the previous WRITE to unlock_count.
So this may be still the old idx (0): we increment lock_count, there goes the
desired smp_mb(), we increment unlock_count of the old idx (0).
_ Yet another READER: finally we see the new idx (1).
Phew! Did I get it right this time? :))
