On Tue, Jul 15, 2025 at 01:24:36PM -0400, Mark Johnston wrote: > On Tue, Jul 15, 2025 at 09:40:18AM -0700, Gleb Smirnoff wrote: > > On Tue, Jul 15, 2025 at 03:19:04PM +0000, Mark Johnston wrote: > > M> The branch main has been updated by markj: > > M> > > M> URL: > > https://cgit.FreeBSD.org/src/commit/?id=fad79db405052f3faad7184ea2c8bfe9f92a700d > > M> > > M> commit fad79db405052f3faad7184ea2c8bfe9f92a700d > > M> Author: Mark Johnston <ma...@freebsd.org> > > M> AuthorDate: 2025-07-15 15:16:40 +0000 > > M> Commit: Mark Johnston <ma...@freebsd.org> > > M> CommitDate: 2025-07-15 15:16:40 +0000 > > M> > > M> vm_pageout: Remove a volatile qualifier from some vm_domain members > > M> > > M> These are always accessed using atomic(9) intrinsics, so do not need > > the > > M> qualifier. No functional change intended. > > M> > > M> Reviewed by: alc, kib > > M> MFC after: 2 weeks > > M> Sponsored by: Modirum MDPay > > M> Sponsored by: Klara, Inc. > > M> Differential Revision: https://reviews.freebsd.org/D51322 > > > > What's the benefit of removing the qualifiers? They act as documentation > > and they match atomic(9) prototypes. To me this looks like removing a > > const qualifier with a reasoning that we use the variable only as an > > argument to functions that have const qualifier. > > Note that I updated the comments as well to indicate that accesses to > the fields should be done through atomic(9), so the documentation is > preserved. > > The reason atomic(9) prototypes include the volatile qualifier is to > permit their use with volatile-qualified variables without a cast, not > because the interface expects only volatile-qualified parameters. > > More generally, I believe that new code should always avoid using > volatile to provide any kind of synchronization, ignoring the case of > accesses to memory mapped with non-default attributes. atomic(9) > intrinsics should be used where they are needed, and some comments > should explain the synchronization protocol if it's not obvious.
Hmm, when I wrote atomic_load/store(), volatile casts were used to utilize compiler-specific semantic of volatile accesses, that happens to match what loads and stores should do (access that place now). I.e. it is not for the allowance to use volatile-qualified locations, but to provide the C11-compatible semantic. After stating that, it is clear why qualifying the vars with volatile is not what we want: the semantic of volatile is compiler-dependent, and it only happens to match what is really used for code. Atomics loads and stores do provide the primitive ops we need, and hide the compiler-specific implementation under. I.e., volatile should *not* be used.
