On Fri, Aug 19, 2022 at 05:04:31PM +0200, Christoph Lameter wrote: > On Thu, 18 Aug 2022, Vlastimil Babka wrote: > > > On 8/18/22 11:42, Christoph Lameter wrote: > > > On Wed, 17 Aug 2022, Sebastian Andrzej Siewior wrote: > > > > > >> + * On PREEMPT_RT, the local lock neither disables interrupts nor preemption > > >> + * which means the lockless fastpath cannot be used as it might interfere with > > >> + * an in-progress slow path operations. In this case the local lock is always > > >> + * taken but it still utilizes the freelist for the common operations. > > > > > > The slub fastpath does not interfere with slow path operations and the > > > > That's true on !PREEMPT_RT because a slowpath operation under > > local_lock_irqsave() will disable interrupts, so there can't be a > > fastpath operation in an interrupt handler appearing in the middle of a > > slowpath operation. > > > > On PREEMPT_RT local_lock_irqsave() doesn't actually disable interrupts, > > so that can happen. IIRC we learned that the hard way when Mike > > Galbraith was testing early versions of my PREEMPT_RT changes for SLUB. > > Well yes if you enable interrupts during the slowpath then interrupts may > use the fastpath. That is a basic design change to the way concurrency is > handled in the allocators. > > There needs to be some fix here to restore the exclusion of the fastpath > during slow path processing. This could be > > A) Exclude the fastpath during slowpath operations > > This can be accomplished by setting things up like in the debug mode > that also excludes the fastpath. I think we can do that by disabling preemption (for a short period, I think) in slowpath on RT (like disabling irq in non-RT) But I wonder if RT guys will prefer that? > B) Force interrupt allocations to the slowpath. > > Check some flag that indicates an interrupt allocation is occurring and > then bypass the fastpath. There is nothing special about interrupt allocation on RT. All users of SLUB on RT must not be in hardirq context. So I don't think it is possible to distingush between a thread being preempted and another thread that preempts it. -- Thanks, Hyeonggon
