Messages in this thread | | | Date | Tue, 9 Apr 2024 17:34:35 +0100 | From | Mark Rutland <> | Subject | Re: [PATCH 2/6] locking/atomic/x86: Rewrite x86_32 arch_atomic64_{,fetch}_{and,or,xor}() functions |
| |
On Tue, Apr 09, 2024 at 02:50:19PM +0200, Uros Bizjak wrote: > On Tue, Apr 9, 2024 at 2:03 PM Uros Bizjak <ubizjak@gmail.com> wrote: > > > > On Tue, Apr 9, 2024 at 1:13 PM Mark Rutland <mark.rutland@arm.com> wrote: > > > > > > static __always_inline void arch_atomic64_and(s64 i, atomic64_t *v) > > > > { > > > > - s64 old, c = 0; > > > > + s64 val = __READ_ONCE(v->counter); > > > > > > I reckon it's worth placing this in a helper with a big comment, e.g. > > > > > > static __always_inline s64 arch_atomic64_read_tearable(atomic64_t *v) > > > { > > > /* > > > * TODO: explain that this might be torn, but it occurs *once*, and can > > > * safely be consumed by atomic64_try_cmpxchg(). > > > * > > > * TODO: point to the existing commentary regarding why we use > > > * __READ_ONCE() for KASAN reasons. > > > */ > > > return __READ_ONCE(v->counter); > > > } > > > > > > ... and then use that in each of the instances below. > > > > > > That way the subtlety is clearly documented, and it'd more clearly align with > > > the x86_64 verions. > > > > This is an excellent idea. The separate definitions needs to be placed > > in atomic64_32.h and atomic_64_64.h (due to use of atomic64_t > > typedef), but it will allow the same unification of functions between > > x64_32 and x64_64 as the approach with __READ_ONCE(). > > Something like this: > > --cut here-- > /* > * This function is intended to preload the value from atomic64_t > * location in a non-atomic way. The read might be torn, but can > * safely be consumed by the compare-and-swap loop. > */ > static __always_inline s64 arch_atomic64_read_tearable(atomic64_t *v) > { > /* > * See the comment in arch_atomic_read() on why we use > * __READ_ONCE() instead of READ_ONCE_NOCHECK() here. > */ > return __READ_ONCE(v->counter); > } > --cut here-- > > Thanks, > Uros.
Yeah, something of that shape.
Having thought for a bit longer, it's probably better to use '_torn' rather than '_tearable' (i.e. name this arch_atomic64_read_torn()).
It'd be nice if we could specify the usage restrictions a bit more clearly, since this can only be used for compare-and-swap loops that implement unconditional atomics. (e.g. arch_atomic64_and(), but not arch_atomic_add_unless()).
So I'd suggest:
/* * Read an atomic64_t non-atomically. * * This is intended to be used in cases where a subsequent atomic operation * will handle the torn value, and can be used to prime the first iteration of * unconditional try_cmpxchg() loops, e.g. * * s64 val = arch_atomic64_read_torn(v); * do { } while (!arch_atomic_try_cmpxchg(v, &val, val OP i); * * This is NOT safe to use where the value is not always checked by a * subsequent atomic operation, such as in conditional try_cmpxchg() loops that * can break before the atomic, e.g. * * s64 val = arch_atomic64_read_torn(v); * do { * if (condition(val)) * break; * } while (!arch_atomic_try_cmpxchg(v, &val, val OP i); */ static __always_inline s64 arch_atomic64_read_torn(atomic64_t *v) { /* See comment in arch_atomic_read() */ return __READ_ONCE(v->counter); }
Mark.
| |