On 19/10/2023 8:48 am, Jan Beulich wrote: > On 13.10.2023 11:42, Juergen Gross wrote: >> Instead of being able to use normal spinlocks as recursive ones, too, >> make recursive spinlocks a special lock type. >> >> This will make the spinlock structure smaller in production builds and >> add type-safety. >> >> This allows to increase the maximum number of physical cpus from 8191 >> to 65535 without increasing the size of the lock structure in production >> builds (the size of recursive spinlocks in debug builds will grow to >> 12 bytes due to that change). >> >> Changes in V2: >> - addressed comments by Jan Beulich >> - lots of additional cleanups >> - reorganized complete series >> >> Juergen Gross (13): >> xen/spinlock: fix coding style issues >> xen/spinlock: reduce lock profile ifdefs >> xen/spinlock: make spinlock initializers more readable >> xen/spinlock: introduce new type for recursive spinlocks >> xen/spinlock: rename recursive lock functions >> xen/spinlock: add rspin_[un]lock_irq[save|restore]() >> xen/spinlock: make struct lock_profile rspinlock_t aware >> xen/spinlock: add explicit non-recursive locking functions >> xen/spinlock: add another function level >> xen/spinlock: add missing rspin_is_locked() and rspin_barrier() >> xen/spinlock: split recursive spinlocks from normal ones >> xen/spinlock: remove indirection through macros for spin_*() functions >> xen/spinlock: support higher number of cpus > Before looking at patches 4 and onwards, I'd like us to settle on the future > of recursive locking in Xen, considering in particular Andrew's objections > to their use in the code base. If the plan was to eliminate them, I'd see > little point in reworking the infrastructure. I'd like to suggest that one > of us tries to remember to put this up as an agenda item for the next > Community Call. But of course the discussion can also happen right here; I > merely expect there might not be much of a reaction.
Actually, I consider this series an improvement. The CPU limit is the most urgent problem to fix. XenServer has just jumped to NR_CPUS=2k in order to support 2024's range of hardware, and it's only going to be a couple of years more before we're stuck given the current spinlocks. I do genuinely think the code and logic would be better without recursive locks, but making that happen is going to be very invasive and complicated. But in the meantime with spinlocks properly separated from recursive locks, it becomes easier IMO to dissuade the introduction of new cases while we unpick the existing ones. And so what if we do end up deleting recursive locks in a few years time? That's not an argument against doing this untangling now. ~Andrew
