On Mon, Nov 20, 2006 at 09:57:12PM +0300, Oleg Nesterov wrote:
> On 11/20, Alan Stern wrote:
> >
> > @@ -158,6 +199,11 @@ void synchronize_srcu(struct srcu_struct
> >
> > [... snip ...]
> >
> > +#ifdef SMP__STORE_MB_LOAD_WORKS /* The fast path */
> > + if (srcu_readers_active_idx(sp, idx) == 0)
> > + goto done;
> > +#endif
>
> I guess this is connected to another message from you,
>
> > But of course it _is_ needed for the fastpath to work. In fact, it might
> > not be good enough, depending on the architecture. Here's what the
> > fastpath ends up looking like (using c[idx] is essentially the same as
> > using hardluckref):
> >
> > WRITER READER
> > ------ ------
> > dataptr = &(new data) atomic_inc(&hardluckref)
> > mb mb
> > while (hardluckref > 0) ; access *dataptr
> >
> > Notice the pattern: Each CPU does store-mb-load. It is known that on
> > some architectures each CPU can end up loading the old value (the value
> > from before the other CPU's store). This would mean the writer would see
> > hardluckref == 0 right away and the reader would see the old dataptr.
>
> So, if we have global A == B == 0,
>
> CPU_0 CPU_1
>
> A = 1; B = 2;
> mb(); mb();
> b = B; a = A;
>
> It could happen that a == b == 0, yes? Isn't this contradicts with definition
> of mb?
It can and does happen. -Which- definition of mb()? ;-)
To see how this can happen, thing of the SMP system as a message-passing
system, and consider the following sequence of events:
o The cache line for A is initially in CPU 1's cache, and the
cache line for B is initially in CPU 0's cache (backwards of
what you would want knowing about the upcoming writes).
o CPU 0 stores to A, but because A is not in cache, places it in
CPU 0's store queue. It also puts out a request for ownership
of the cache line containing A.
o CPU 1 stores to B, with the same situation as for CPU 0's store
to A.
o Both CPUs execute an mb(), which ensures that any subsequent writes
follow the writes to A and B, respectively. Since neither CPU
has yet received the other CPU's request for ownership, there is
no ordering effects on subsequent reads.
o CPU 0 executes "b = B", and since B is in CPU 0's cache, it loads
the current value, which is zero.
o Ditto for CPU 1 and A.
o CPUs 0 and 1 now receive each other's requests for ownership, so
exchange the cache lines containing A and B.
o Once CPUs 0 and 1 receive ownership of the respective cache lines,
they complete their writes to A and B (moving the values from the
store buffers to the cache lines).
> By definition, when CPU_0 issues 'b = B', 'A = 1' should be visible to other
> CPUs, yes? Now, b == 0 means that CPU_1 did not read 'a = A' yet, otherwise
> 'B = 2' should be visible to all CPUs (by definition again).
>
> Could you please clarify this?
See above...
> Btw, this is funny, but I was going to suggest _exactly_ same cleanup for
> srcu_read_lock :)
;-)
Thanx, Paul
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
[Index of Archives]
[Kernel Newbies]
[Netfilter]
[Bugtraq]
[Photo]
[Stuff]
[Gimp]
[Yosemite News]
[MIPS Linux]
[ARM Linux]
[Linux Security]
[Linux RAID]
[Video 4 Linux]
[Linux for the blind]
[Linux Resources]
