On Tue, Aug 23, 2016 at 08:39:18AM +0200, Geert Uytterhoeven wrote:
> Hi Paul,
>
> On Mon, Aug 22, 2016 at 11:16 PM, Paul E. McKenney
> <paul...@linux.vnet.ibm.com> wrote:
> > On Mon, Aug 22, 2016 at 10:48:57PM +0200, Geert Uytterhoeven wrote:
> >> On Mon, Aug 22, 2016 at 9:54 PM, Paul E. McKenney
> >> <paul...@linux.vnet.ibm.com> wrote:
> >> > On Mon, Aug 22, 2016 at 03:18:54PM -0400, Steven Rostedt wrote:
> >> >> On Mon, 22 Aug 2016 20:56:09 +0200
> >> >> Peter Zijlstra <pet...@infradead.org> wrote:
> >> >>
> >> >> > > Don't we have __alignof__(void *) to avoid #ifdef CONFIG_M68K and
> >> >> > > other new macros ?
> >> >
> >> > Hmmm... Does __alignof__(void *) give two-byte alignment on m68k,
> >> > allowing something like this? Heh!!! It is already there. ;-)
> >> >
> >> > struct callback_head {
> >> > struct callback_head *next;
> >> > void (*func)(struct callback_head *head);
> >> > } __attribute__((aligned(sizeof(void *))));
> >>
> >> No, it's aligning to sizeof(void *) (4 on m68k), not __alignof__(void *).
> >
> > Right you are. Commit 720abae3d68ae from Kirill A. Shutemov in November
> > 2015.
> >
> > Given that you haven't complained, I am guessing that this works for you.
> > If so, I can make the __call_rcu() WARN_ON() more strict.
> >
> >> > #define rcu_head callback_head
> >> >
> >> > If so, that does sound quite attractive! Might need the WARN_ON()
> >> > anyway, to flag wild pointers if nothing else.
> >> >
> >> > Adding Geert on CC for his thoughts.
> >>
> >> __alignof__(void *) is indeed 2 on m68k, and h8300.
> >>
> >> Note that it is 1 on crisv32!
> >
> > Gah... ((__alignof__(void *) + 1) & ~0x1), eh?
> >
> >> It's 4 or 8 on anything else I have a cross-compiler for.
> >>
> >> $ cat a.c
> >> unsigned x = __alignof__(void *);
> >> $ for i in /opt/cross/*/*/bin/*gcc; do echo +++ $i +++; $i -c -S a.c;
> >> cat a.s; done | less
> >
> > Thank you for checking!
> >
> > Again, does the current state work for you?
>
> Yes it does. See also your commit 1146edcbef378922 ("rcu: Loosen
> __call_rcu()'s
> rcu_head alignment constraint").
Understood!
But given that all architectures now provide at least four-byte alignment
for the rcu_head structure, isn't it now OK for me to tighten up __call_rcu()'s
check, for example, to this?
WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
Thanx, Paul
