On Sat, 1 Feb 2025 19:42:32 +0100 Mateusz Guzik <mjguzik@xxxxxxxxx> wrote: > On Sat, Feb 1, 2025 at 7:19 PM David Laight > <david.laight.linux@xxxxxxxxx> wrote: > > > > On Sat, 1 Feb 2025 17:31:06 +0100 > > Mateusz Guzik <mjguzik@xxxxxxxxx> wrote: > > > > > It no longer serves any purpose now that the tasklist_lock -> > > > pidmap_lock ordering got eliminated. > > > > Not disabling interrupts may make thing worse. > > It is a trade off between 'interrupt latency' and 'lock hold time'. > > > > If interrupts are disabled then (clearly) they can get delayed because > > the lock is held. > > Provided the lock is only held for a short time it probably doesn't matter. > > Indeed, unless it is the worst one, it probably doesn't matter at all. > > After all spin locks shouldn't really be held for significant periods. > > > > OTOH if the lock doesn't disable interrupts then an interrupt will > > increase the length of time a lock is held for. > > This can be significant - and I mean upwards of 1ms. > > Network interrupts can tale a while - and then the work that is deferred > > to 'softint' context happens as well (I don't think a spinlock stops > > the softint code). > > > > I've a feeling that unless a spin lock is held for 'far longer than one > > should ever be held for' then you really want to disable interrupts. > > > > Note that taking the interrupt trip increases single-threaded overhead. I'm not sure what you mean. Disabling interrupts isn't as cheap as it ought to be, but probably isn't that bad. > Per your own description, if the lock is contested and interrupts are > disabled, handling them also get delayed by CPUs which are busy just > waiting (and which would otherwise take care of them). The slow path for spin locks ought to have interrupts enabled. But, in any case, the interrupt is only delayed for the short time the spin lock is held for. > So while this is indeed a tradeoff, as I understand the sane default > is to *not* disable interrupts unless necessary. I bet to differ. If an interrupt is taken by a cpu that holds a spin lock then any other cpu that attempts to acquire the lock spins for the additional time that the interrupt takes. If interrupts are disabled then an interrupt is delayed for the time that the spin lock is held. The execution time of an interrupt can easily be a lot longer than most spin locks are held for. Either because of the work that an ethernet interrupt does (even before deferring to softint), or because even a single PCIe read (eg to an Altera Cyclone V fpga) can easily take 1000s of clocks. Now the execution cost of the interrupt has to happen some time. But you really don't want multiple cpu spinning waiting for it to finish. David
