On Wed, 20 May 2026 16:26:42 +0100 Steven Price <[email protected]> wrote:
> On 18/05/2026 09:43, Boris Brezillon wrote: > > On Thu, 14 May 2026 10:09:10 -0700 > > Chia-I Wu <[email protected]> wrote: > > > >> On Thu, May 14, 2026 at 6:24 AM Steven Price <[email protected]> wrote: > >> > >>> > >>> On 13/05/2026 17:58, Boris Brezillon wrote: > >>>> Right now panthor is mixed bag of manual locks and guards. Let's > >>>> make that more consitent and thus encourage new submissions to go > >>>> for guards. > >>> > >>> I'm fine with encouraging guards for future code - but I'm a little wary > >>> of a big change like this - it's hard to review it and check that > >>> everything works the same. And it's a little dubious that the mechanical > >>> refactoring produces more readable code in some cases. > >> I agree with Steven in general, although I am in favor of landing now > >> that you've gone through the trouble. > > > > Honestly, I agree with you. The only reason I went for it is > > because the mix we have right now is pretty confusing. This has to do > > with the fact the scopes are often loosely defined unless you used > > scoped_guard(), so it's pretty easy to mess up the lock/unlock > > ordering. For instance, > > > > mutex_lock(locka); > > guard(lockb); > > mutex_unlock(locka); > > > > ... > > > > once expanded, turns into inconsistent locked sections, where the inner > > lock (lockb) is released after the outer one (locka). > > I think that's a good argument for getting all the guard forms available > before tackling the conversion. Yep, makes sense. The reason I didn't go for that in v1 is because I wasn't sure how well the new guard definition would be received. Now that we know there's a general consensus to define those, I'll re-order the patches accordingly. > Mostly I feel like it would be benefit > from being split up into multiple patches (maybe one per file?) so that > there are smaller units to review. Sure, I can do that. > > >> > >> I also have mixed feelings about some of the non-scoped guards. Their > >> scopes are extended slightly than before, supposedly to avoid adding > >> another level of indentation. But other than slightly slower, > > > > I tried to used scoped_guard()s every where the extra non-guarded > > section could be CPU heavy (the only bits left are some very simple > > bit/arithmetic ops, and a couple queue_work() IIRC). > > > >> it also > >> becomes less clear what exactly do the guards protect. > > > > I know, and I have pretty much the same feeling, but we've crossed that > > bridge when we started accepting non-scoped guard()s, unfortunately. > > The problem with scoped guards is the extra level of indentation. Yep. > Personally I find a mixture of all three is appropriate depending on the > case. > > E.g. > > int small_simple_function() { > if (simple_condition) > return early; > > guard(lock); > > if (condition_that_needs_lock) > return early; > /* more work */ > return late; > } > > Here it's easy to reason because the lock is just held for the duration > of the function after the initial early-out condition is checked. > > int short_lock() { > /* bunch of work */ > > scoped_guard(lock) { > tmp = read_value(); > if (tmp == 42) > return -ESOLONGANDTHANKSFORALLTHEFISH; > tmp++; > write_value(tmp); > } > > /* more work */ > } > > Here there's a small section of code which is working on the lock, so it > makes sense to indent it to show the boundaries of it. The other nice > thing is that the error return handles the locks for us. > > int old_fashioned() { > if (lock_required) > mutex_lock(lock); > > /* some work */ > > if (lock_required) > mutex_unlock(lock); > } > > Generally a pattern to be avoided if possible, Yeah, honestly I try my best to never end up with that sort of conditional locks. > but IMHO this is much > better than the equivalent of: > > int dodgy_function() { > /* some work */ > } > > int outer_function() { > if (lock_required) { > scoped_guard(lock) > dodgy_function(); > } else { > dodgy_function(); > } > } If I were to choose, I'd probably go for this version, but luckily we don't seem to have this conditional-locking pattern in panthor.
