Messages in this thread | | | Date | Wed, 21 Aug 2019 09:14:26 -0700 | From | "Paul E. McKenney" <> | Subject | Re: [PATCH 1/1] Fix: trace sched switch start/stop racy updates |
| |
On Wed, Aug 21, 2019 at 11:48:43AM -0400, Mathieu Desnoyers wrote: > ----- On Aug 21, 2019, at 8:33 AM, Peter Zijlstra peterz@infradead.org wrote: > > > On Wed, Aug 21, 2019 at 06:23:10AM -0700, Paul E. McKenney wrote: > >> On Wed, Aug 21, 2019 at 11:32:01AM +0100, Will Deacon wrote: > > > >> > and so it is using a store-pair instruction to reduce the complexity in > >> > the immediate generation. Thus, the 64-bit store will only have 32-bit > >> > atomicity. In fact, this is scary because if I change bar to: > >> > > >> > void bar(u64 *x) > >> > { > >> > *(volatile u64 *)x = 0xabcdef10abcdef10; > >> > } > >> > > >> > then I get: > >> > > >> > bar: > >> > mov w1, 61200 > >> > movk w1, 0xabcd, lsl 16 > >> > str w1, [x0] > >> > str w1, [x0, 4] > >> > ret > >> > > >> > so I'm not sure that WRITE_ONCE would even help :/ > >> > >> Well, I can have the LWN article cite your email, then. So thank you > >> very much! > >> > >> Is generation of this code for a 64-bit volatile store considered a bug? > >> Or does ARMv8 exclude the possibility of 64-bit MMIO registers? And I > >> would guess that Thomas and Linus would ask a similar bugginess question > >> for normal stores. ;-) > > > > I'm calling this a compiler bug; the way I understand volatile this is > > very much against the intentended use case. That is, this is buggy even > > on UP vs signals or MMIO. > > And here is a simpler reproducer on my gcc-8.3.0 (aarch64) compiled with O2: > > volatile unsigned long a; > > void fct(void) > { > a = 0x1234567812345678ULL; > } > > void fct(void) > { > a = 0x1234567812345678ULL; > 0: 90000000 adrp x0, 8 <fct+0x8> > 4: 528acf01 mov w1, #0x5678 // #22136 > 8: 72a24681 movk w1, #0x1234, lsl #16 > c: f9400000 ldr x0, [x0] > 10: b9000001 str w1, [x0] > 14: b9000401 str w1, [x0, #4] > } > 18: d65f03c0 ret > > And the non-volatile case uses stp (is it a single store to memory ?): > > unsigned long a; > > void fct(void) > { > a = 0x1234567812345678ULL; > } > > void fct(void) > { > a = 0x1234567812345678ULL; > 0: 90000000 adrp x0, 8 <fct+0x8> > 4: 528acf01 mov w1, #0x5678 // #22136 > 8: 72a24681 movk w1, #0x1234, lsl #16 > c: f9400000 ldr x0, [x0] > 10: 29000401 stp w1, w1, [x0] > } > 14: d65f03c0 ret > > It would probably be a good idea to audit other architectures, since this > is done by the compiler backend.
That does seem like a very good idea!
Thanx, Paul
| |