On Thu, Feb 23, 2023 at 03:48:26PM +0800, Xiongfeng Wang wrote: > From: Xiongfeng Wang <wangxiongfeng2@xxxxxxxxxx> > > When NOHZ_FULL is enabled, such as in HPC situation, CPUs are divided > into housekeeping CPUs and non-housekeeping CPUs. Non-housekeeping CPUs > are NOHZ_FULL CPUs and are often monopolized by the userspace process, > such HPC application process. Any sort of interruption is not expected. > > blk_mq_hctx_next_cpu() selects each cpu in 'hctx->cpumask' alternately > to schedule the work thread blk_mq_run_work_fn(). When 'hctx->cpumask' > contains housekeeping CPU and non-housekeeping CPU at the same time, a > housekeeping CPU, which want to request a IO, may schedule a worker on a > non-housekeeping CPU. This may affect the performance of the userspace > application running on non-housekeeping CPUs. > > So let's just schedule the worker thread on the current CPU when the > current CPU is housekeeping CPU. This looks like an odd non-systemic bandaid. Shouldn't we have a more generic way nothing ever gets onto these non-housekeeping CPUs by making sure they never show up in the cpumask, and never get completion IPIs?
