Hi, When VMM(s) simultaneously create a large number of irqfds and register their irqfds in the global consumers list, the global mutex contention exponentially increases the average wait latency, which is no longer tolerable on modern systems with a large number of CPU cores. The patch set is intended to reduce this source of latency by converting producers/consumers single linked list to XArray. Please feel free to run more tests and share comments to move forward. V1 -> V2 Changelog: - Send the prerequisite fix as a series; (Alex W.) - Keep producer and consumer connected and tracked; (Alex W.) V1: https://lore.kernel.org/kvm/20230801115646.33990-1-likexu@xxxxxxxxxxx/ https://lore.kernel.org/kvm/20230801085408.69597-1-likexu@xxxxxxxxxxx/ Like Xu (2): KVM: eventfd: Fix NULL deref irqbypass producer KVM: irqbypass: Convert producers/consumers linked list to XArray include/linux/irqbypass.h | 8 +-- virt/lib/irqbypass.c | 127 +++++++++++++++++++------------------- 2 files changed, 63 insertions(+), 72 deletions(-) base-commit: 5a7591176c47cce363c1eed704241e5d1c42c5a6 -- 2.41.0
