On Mon, 15 Dec 2025 at 14:38, Marco Elver <[email protected]> wrote: > > On Fri, Dec 12, 2025 at 12:09PM +0100, Peter Zijlstra wrote: > > On Fri, Dec 12, 2025 at 11:15:29AM +0100, Marco Elver wrote: > > > On Fri, 12 Dec 2025 at 10:43, Peter Zijlstra <[email protected]> wrote: > > > [..] > > > > > Correct. We're trading false negatives over false positives at this > > > > > point, just to get things to compile cleanly. > > > > > > > > Right, and this all 'works' right up to the point someone sticks a > > > > must_not_hold somewhere. > > > > > > > > > > > Better support for Linux's scoped guard design could be added in > > > > > > > future if deemed critical. > > > > > > > > > > > > I would think so, per the above I don't think this is 'right'. > > > > > > > > > > It's not sound, but we'll avoid false positives for the time being. > > > > > Maybe we can wrangle the jigsaw of macros to let it correctly acquire > > > > > and then release (via a 2nd cleanup function), it might be as simple > > > > > as marking the 'constructor' with the right __acquires(..), and then > > > > > have a 2nd __attribute__((cleanup)) variable that just does a no-op > > > > > release via __release(..) so we get the already supported pattern > > > > > above. > > > > > > > > Right, like I mentioned in my previous email; it would be lovely if at > > > > the very least __always_inline would get a *very* early pass such that > > > > the above could be resolved without inter-procedural bits. I really > > > > don't consider an __always_inline as another procedure. > > > > > > > > Because as I already noted yesterday, cleanup is now all > > > > __always_inline, and as such *should* all end up in the one function. > > > > > > > > But yes, if we can get a magical mash-up of __cleanup and __release (let > > > > it be knows as __release_on_cleanup ?) that might also work I suppose. > > > > But I vastly prefer __always_inline actually 'working' ;-) > > > > > > The truth is that __always_inline working in this way is currently > > > infeasible. Clang and LLVM's architecture simply disallow this today: > > > the semantic analysis that -Wthread-safety does happens over the AST, > > > whereas always_inline is processed by early passes in the middle-end > > > already within LLVM's pipeline, well after semantic analysis. There's > > > a complexity budget limit for semantic analysis (type checking, > > > warnings, assorted other errors), and path-sensitive & > > > intra-procedural analysis over the plain AST is outside that budget. > > > Which is why tools like clang-analyzer exist (symbolic execution), > > > where it's possible to afford that complexity since that's not > > > something that runs for a normal compile. > > > > > > I think I've pushed the current version of Clang's -Wthread-safety > > > already far beyond what folks were thinking is possible (a variant of > > > alias analysis), but even my healthy disregard for the impossible > > > tells me that making path-sensitive intra-procedural analysis even if > > > just for __always_inline functions is quite possibly a fool's errand. > > > > Well, I had to propose it. Gotta push the envelope :-) > > > > > So either we get it to work with what we have, or give up. > > > > So I think as is, we can start. But I really do want the cleanup thing > > sorted, even if just with that __release_on_cleanup mashup or so. > > Working on rebasing this to v6.19-rc1 and saw this new scoped seqlock > abstraction. For that one I was able to make it work like I thought we > could (below). Some awkwardness is required to make it work in > for-loops, which only let you define variables with the same type. > > For <linux/cleanup.h> it needs some more thought due to extra levels of > indirection.
For cleanup.h, the problem is that to instantiate we use "guard(class)(args..)". If it had been designed as "guard(class, args...)", i.e. just use __VA_ARGS__ explicitly instead of the implicit 'args...', it might have been possible to add a second cleanup variable to do the same (with some additional magic to extract the first arg if one exists). Unfortunately, the use of the current guard()() idiom has become so pervasive that this is a bigger refactor. I'm going to leave cleanup.h as-is for now, if we think we want to give this a go in the current state. One observation from the rebase: Generally synchronization primitives do not change much and the annotations are relatively stable, but e.g. RCU & sched (latter is optional and depends on the sched-enablement patch) receive disproportionally more changes, and while new annotations required for v6.19-rc1 were trivial, it does require compiling with a Clang version that does produce the warnings to notice. While Clang 22-dev is being tested on CI, I doubt maintainers already use it, so it's possible we'll see some late warnings due to missing annotations when things hit -next. This might be an acceptable churn cost, if we think the outcome is worthwhile. Things should get better when Clang 22 is released properly, but until then things might be a little bumpy if there are large changes across the core synchronization primitives. Thanks, -- Marco
