Message ID | 20240124204909.105952-1-oliver.upton@linux.dev (mailing list archive) |
---|---|
Headers | show |
Series | KVM: arm64: Improvements to GICv3 LPI injection | expand |
Hi Oliver, On Wed, 24 Jan 2024 20:48:54 +0000, Oliver Upton <oliver.upton@linux.dev> wrote: > > The unfortunate reality is there are increasingly large systems that are > shipping today without support for GICv4 vLPI injection. Serialization > in KVM's LPI routing/injection code has been a significant bottleneck > for VMs on these machines when under a high load of LPIs (e.g. a > multi-queue NIC). > > Even though the long-term solution is quite clearly **direct > injection**, we really ought to do something about the LPI scaling > issues within KVM. > > This series aims to improve the performance of LPI routing/injection in > KVM by moving readers of LPI configuration data away from the > lpi_list_lock in favor or using RCU. > > Patches 1-5 change out the representation of LPIs in KVM from a > linked-list to an xarray. While not strictly necessary for making the > locking improvements, this seems to be an opportune time to switch to a > data structure that can actually be indexed. > > Patches 6-10 transition vgic_get_lpi() and vgic_put_lpi() away from > taking the lpi_list_lock in favor of using RCU for protection. Note that > this requires some rework to the way references are taken on LPIs and > how reclaim works to be RCU safe. > > Lastly, patches 11-15 rework the LRU policy on the LPI translation cache > to not require moving elements in the linked-list and take advantage of > this to make it an rculist readable outside of the lpi_list_lock. I quite like the overall direction. I've left a few comments here and there, and will probably get back to it after I try to run some tests on a big-ish box. > All of this was tested on top of v6.8-rc1. Apologies if any of the > changelogs are a bit too light, I'm happy to rework those further in > subsequent revisions. > > I would've liked to have benchmark data showing the improvement on top > of upstream with this series, but I'm currently having issues with our > internal infrastructure and upstream kernels. However, this series has > been found to have a near 2x performance improvement to redis-memtier [*] > benchmarks on our kernel tree. It'd be really good to have upstream-based numbers, with details of the actual setup (device assignment? virtio?) so that we can compare things and even track regressions in the future. Thanks, M.
On Thu, Jan 25, 2024 at 11:02:01AM +0000, Marc Zyngier wrote: [...] > > I would've liked to have benchmark data showing the improvement on top > > of upstream with this series, but I'm currently having issues with our > > internal infrastructure and upstream kernels. However, this series has > > been found to have a near 2x performance improvement to redis-memtier [*] > > benchmarks on our kernel tree. > > It'd be really good to have upstream-based numbers, with details of > the actual setup (device assignment? virtio?) so that we can compare > things and even track regressions in the future. Yeah, that sort of thing isn't optional IMO, I just figured that getting reviews on this would be a bit more productive while I try and recreate the test correctly on top of upstream. The test setup I based my "2x" statement on is 4 16 vCPU client VMs talking to 1 16 vCPU server VM over NIC VFs assigned to the respective VMs. 16 TX + 16 RX queues for each NIC. As I'm sure you're aware, I know damn near nothing about the Redis setup itself, and I'll need to do a bit of work to translate the thing I was using into a script.