On Tue, Aug 15, 2023 at 10:36:18AM +0800, Yuan Yao wrote: > On Mon, Aug 14, 2023 at 05:09:18PM +0800, Yan Zhao wrote: > > On Fri, Aug 11, 2023 at 12:35:27PM -0700, John Hubbard wrote: > > > On 8/11/23 11:39, David Hildenbrand wrote: > > > ... > > > > > > Should we want to disable NUMA hinting for such VMAs instead (for example, by QEMU/hypervisor) that knows that any NUMA hinting activity on these ranges would be a complete waste of time? I recall that John H. once mentioned that there are > > > > > similar issues with GPU memory: NUMA hinting is actually counter-productive and they end up disabling it. > > > > > > > > > > > > > > > > Yes, NUMA balancing is incredibly harmful to performance, for GPU and > > > > > accelerators that map memory...and VMs as well, it seems. Basically, > > > > > anything that has its own processors and page tables needs to be left > > > > > strictly alone by NUMA balancing. Because the kernel is (still, even > > > > > today) unaware of what those processors are doing, and so it has no way > > > > > to do productive NUMA balancing. > > > > > > > > Is there any existing way we could handle that better on a per-VMA level, or on the process level? Any magic toggles? > > > > > > > > MMF_HAS_PINNED might be too restrictive. MMF_HAS_PINNED_LONGTERM might be better, but with things like iouring still too restrictive eventually. > > > > > > > > I recall that setting a mempolicy could prevent auto-numa from getting active, but that might be undesired. > > > > > > > > CCing Mel. > > > > > > > > > > Let's discern between page pinning situations, and HMM-style situations. > > > Page pinning of CPU memory is unnecessary when setting up for using that > > > memory by modern GPUs or accelerators, because the latter can handle > > > replayable page faults. So for such cases, the pages are in use by a GPU > > > or accelerator, but unpinned. > > > > > > The performance problem occurs because for those pages, the NUMA > > > balancing causes unmapping, which generates callbacks to the device > > > driver, which dutifully unmaps the pages from the GPU or accelerator, > > > even if the GPU might be busy using those pages. The device promptly > > > causes a device page fault, and the driver then re-establishes the > > > device page table mapping, which is good until the next round of > > > unmapping from the NUMA balancer. > > > > > > hmm_range_fault()-based memory management in particular might benefit > > > from having NUMA balancing disabled entirely for the memremap_pages() > > > region, come to think of it. That seems relatively easy and clean at > > > first glance anyway. > > > > > > For other regions (allocated by the device driver), a per-VMA flag > > > seems about right: VM_NO_NUMA_BALANCING ? > > > > > Thanks a lot for those good suggestions! > > For VMs, when could a per-VMA flag be set? > > Might be hard in mmap() in QEMU because a VMA may not be used for DMA until > > after it's mapped into VFIO. > > Then, should VFIO set this flag on after it maps a range? > > Could this flag be unset after device hot-unplug? > > Emm... syscall madvise() in my mind, it does things like change flags > on VMA, e.g madvise(MADV_DONTFORK) adds VM_DONTCOPY to the VMA. Yes, madvise() might work. And setting this flag might be an easy decision, while unsetting it might be hard unless some counters introduced.