When there are a lot of IO requests queued for the HW, the IRQ handler of the mpxsas driver spins forever until all requests are serviced. Spinning in the IRQ handler forever is a bad design and causes stalls on the other cores that is unrelated to this request. A simple solution has been posted here but rejected by the maintainer as the maintainer has been looking to offload only the slow path to task context. https://patchwork.kernel.org/patch/7579951/ Further requests for details to resolve the issue has been unanswered. Logging the issue here for awareness.
Created attachment 255737 [details] Propsed fix