On 5 Dec 2021, at 13:22, Konstantin Belousov <kostik...@gmail.com> wrote: > > On Sun, Dec 05, 2021 at 03:03:26AM +0000, Jessica Clarke wrote: >> On 4 Dec 2021, at 22:21, Konstantin Belousov <k...@freebsd.org> wrote: >>> >>> The branch main has been updated by kib: >>> >>> URL: >>> https://cgit.FreeBSD.org/src/commit/?id=a4e4132fa3bfadb6047fc0fa5f399f4640460300 >>> >>> commit a4e4132fa3bfadb6047fc0fa5f399f4640460300 >>> Author: Konstantin Belousov <k...@freebsd.org> >>> AuthorDate: 2021-11-29 16:26:31 +0000 >>> Commit: Konstantin Belousov <k...@freebsd.org> >>> CommitDate: 2021-12-04 22:20:58 +0000 >>> >>> swapoff(2): replace special device name argument with a structure >>> >>> For compatibility, add a placeholder pointer to the start of the >>> added struct swapoff_new_args, and use it to distinguish old vs. new >>> style of syscall invocation. >>> >>> Reviewed by: markj >>> Discussed with: alc >>> Sponsored by: The FreeBSD Foundation >>> MFC after: 1 week >>> Differential revision: https://reviews.freebsd.org/D33165 >>> --- >>> sys/vm/swap_pager.c | 27 +++++++++++++++++++++++++-- >>> sys/vm/swap_pager.h | 8 ++++++++ >>> 2 files changed, 33 insertions(+), 2 deletions(-) >>> >>> diff --git a/sys/vm/swap_pager.c b/sys/vm/swap_pager.c >>> index 165373d1b527..dc1df79f4fcd 100644 >>> --- a/sys/vm/swap_pager.c >>> +++ b/sys/vm/swap_pager.c >>> @@ -2491,15 +2491,38 @@ sys_swapoff(struct thread *td, struct swapoff_args >>> *uap) >>> struct vnode *vp; >>> struct nameidata nd; >>> struct swdevt *sp; >>> - int error; >>> + struct swapoff_new_args sa; >>> + int error, probe_byte; >>> >>> error = priv_check(td, PRIV_SWAPOFF); >>> if (error) >>> return (error); >>> >>> + /* >>> + * Detect old vs. new-style swapoff(2) syscall. The first >>> + * pointer in the memory pointed to by uap->name is NULL for >>> + * the new variant. >>> + */ >>> + probe_byte = fubyte(uap->name); >>> + switch (probe_byte) { >>> + case -1: >>> + return (EFAULT); >>> + case 0: >>> + error = copyin(uap->name, &sa, sizeof(sa)); >>> + if (error != 0) >>> + return (error); >>> + if (sa.flags != 0) >>> + return (EINVAL); >>> + break; >>> + default: >>> + bzero(&sa, sizeof(sa)); >>> + sa.name = uap->name; >>> + break; >>> + } >> >> Doesn’t this change the semantics of swapoff("")? >> >> Previously it would fail deterministically, presumably with ENOENT or >> something, but now it reinterprets whatever follows that string in >> memory as the new argument structure. It probably doesn’t matter, but >> this approach is ugly. Can we not just define a new syscall rather than >> this kind of bodge? > > Having two swapoff() syscalls is worse, and having them only differ in > semantic by single flag is kind of crime. > > I do not see swapoff("") as problematic, we are changing a minor semantic of > the management syscall. I only wanted to avoid flag day for swapoff binaries. > > BTW, I considered requiring proper alignment for uap->name, and then checking > the whole uap->name_old_syscall for NULL, but then decided that this is > overkill. If you think that swapoff("") that important, I can add that > additional verification.
Why’s it worse? It’s just a syscall number, you deprecate the old one and move on, we do that for things relatively regularly. This is really not a good solution; harder to use as a caller since the prototype is wrong, impossible to ensure you preserve the semantics for the existing interface in all cases, and ugly to implement. You don’t need a flag day for a new syscall, either, you can continue to only use the new method for -f for a release and then switch over to the new syscall entirely. Or switch over to the new syscall entirely now and fall back on the old syscall if -f isn’t passed. Defining a new syscall also lets you not need the name_old_syscall member in the struct, and gives you a clean, fully-extensible syscall to which future features can be added in a backwards-compatible way, rather than forever keeping around this legacy mess. Jess
