On 08/05/2024 14:07, David Hildenbrand wrote: > On 08.05.24 14:54, Ryan Roberts wrote: >> On 08/05/2024 13:45, David Hildenbrand wrote: >>> On 08.05.24 14:43, Ryan Roberts wrote: >>>> On 08/05/2024 13:10, David Hildenbrand wrote: >>>>> On 08.05.24 14:02, David Hildenbrand wrote: >>>>>> On 08.05.24 11:02, Ryan Roberts wrote: >>>>>>> On 08/05/2024 08:12, David Hildenbrand wrote: >>>>>>>> On 08.05.24 09:08, David Hildenbrand wrote: >>>>>>>>> On 08.05.24 06:45, Baolin Wang wrote: >>>>>>>>>> >>>>>>>>>> >>>>>>>>>> On 2024/5/7 18:52, Ryan Roberts wrote: >>>>>>>>>>> On 06/05/2024 09:46, Baolin Wang wrote: >>>>>>>>>>>> To support the use of mTHP with anonymous shmem, add a new sysfs >>>>>>>>>>>> interface >>>>>>>>>>>> 'shmem_enabled' in the >>>>>>>>>>>> '/sys/kernel/mm/transparent_hugepage/hugepages-kB/' >>>>>>>>>>>> directory for each mTHP to control whether shmem is enabled for that >>>>>>>>>>>> mTHP, >>>>>>>>>>>> with a value similar to the top level 'shmem_enabled', which can be >>>>>>>>>>>> set to: >>>>>>>>>>>> "always", "inherit (to inherit the top level setting)", "within_size", >>>>>>>>>>>> "advise", >>>>>>>>>>>> "never", "deny", "force". These values follow the same semantics as >>>>>>>>>>>> the top >>>>>>>>>>>> level, except the 'deny' is equivalent to 'never', and 'force' is >>>>>>>>>>>> equivalent >>>>>>>>>>>> to 'always' to keep compatibility. >>>>>>>>>>> >>>>>>>>>>> We decided at [1] to not allow 'force' for non-PMD-sizes. >>>>>>>>>>> >>>>>>>>>>> [1] >>>>>>>>>>> https://lore.kernel.org/linux-mm/533f37e9-81bf-4fa2-9b72-12cdcb1edb3f@xxxxxxxxxx/ >>>>>>>>>>> >>>>>>>>>>> However, thinking about this a bit more, I wonder if the decision we >>>>>>>>>>> made to >>>>>>>>>>> allow all hugepages-xxkB/enabled controls to take "inherit" was the >>>>>>>>>>> wrong >>>>>>>>>>> one. >>>>>>>>>>> Perhaps we should have only allowed the PMD-sized enable=inherit >>>>>>>>>>> (this is >>>>>>>>>>> just >>>>>>>>>>> for legacy back compat after all, I don't think there is any use case >>>>>>>>>>> where >>>>>>>>>>> changing multiple mTHP size controls atomically is actually useful). >>>>>>>>>>> Applying >>>>>>>>>> >>>>>>>>>> Agree. This is also our usage of 'inherit'. >>>>>>>> >>>>>>>> Missed that one: there might be use cases in the future once we would start >>>>>>>> defaulting to "inherit" for all knobs (a distro might default to that) and >>>>>>>> default-enable THP in the global knob. Then, it would be easy to disable >>>>>>>> any >>>>>>>> THP >>>>>>>> by disabling the global knob. (I think that's the future we're heading to, >>>>>>>> where >>>>>>>> we'd have an "auto" mode that can be set on the global toggle). >>>>>>>> >>>>>>>> But I am just making up use cases ;) I think it will be valuable and just >>>>>>>> doing >>>>>>>> it consistently now might be cleaner. >>>>>>> >>>>>>> I agree that consistency between enabled and shmem_enabled is top priority. >>>>>>> And >>>>>>> yes, I had forgotten about the glorious "auto" future. So probably >>>>>>> continuing >>>>>>> all sizes to select "inherit" is best. >>>>>>> >>>>>>> But for shmem_enabled, that means we need the following error checking: >>>>>>> >>>>>>> - It is an error to set "force" for any size except PMD-size >>>>>>> >>>>>>> - It is an error to set "force" for the global control if any size >>>>>>> except >>>>>>> PMD- >>>>>>> size is set to "inherit" >>>>>>> >>>>>>> - It is an error to set "inherit" for any size except PMD-size if the >>>>>>> global >>>>>>> control is set to "force". >>>>>>> >>>>>>> Certainly not too difficult to code and prove to be correct, but not the >>>>>>> nicest >>>>>>> UX from the user's point of view when they start seeing errors. >>>>>>> >>>>>>> I think we previously said this would likely be temporary, and if/when tmpfs >>>>>>> gets mTHP support, we could simplify and allow all sizes to be set to >>>>>>> "force". >>>>>>> But I wonder if tmpfs would ever need explicit mTHP control? Maybe it >>>>>>> would be >>>>>>> more suited to the approach the page cache takes to transparently ramp up >>>>>>> the >>>>>>> folio size as it faults more in. (Just saying there is a chance that this >>>>>>> error >>>>>>> checking becomes permanent). >>>>>> >>>>>> Note that with shmem you're inherently facing the same memory waste >>>>>> issues etc as you would with anonymous memory. (sometimes even worse, if >>>>>> you're running shmem that's configured to be unswappable!). >>>>> >>>>> Also noting that memory waste is not really a problem when a write to a shmem >>>>> file allocates a large folio that stays within boundaries of that write; >>>>> issues >>>>> only pop up if you end up over-allocating, especially, during page faults >>>>> where >>>>> you have not that much clue about what to do (single address, no real range >>>>> provided). >>>>> >>>>> There is the other issue that wasting large chunks of contiguous memory on >>>>> stuff >>>>> that barely benefits from it. With memory that maybe never gets evicted, there >>>>> is no automatic "handing back" of that memory to the system to be used by >>>>> something else. With ordinary files, that's a bit different. But I did not >>>>> look >>>>> closer into that issue yet, it's one of the reasons MADV_HUGEPAGE was added >>>>> IIRC. >>>> >>>> OK understood. Although, with tmpfs you're not going to mmap it then randomly >>>> extend the file through page faults - mmap doesn't permit that, I don't think? >>>> So presumably the user must explicitly set the size of the file first? Are you >>>> suggesting there are a lot of use cases where a large tmpfs file is created, >>>> mmaped then only accessed sparsely? >>> >>> I don't know about "a lot of use cases", but for VMs that's certainly how it's >>> used. >> > > There are more details around that and the sparsity (memory ballooning, > virtio-mem, free page reporting), but it might distract here :) I'll note that > shmem+THP is known to be problematic with memory ballooning. > >> Gottya, thanks. And out of curiosity, what's the benefit of using tmpfs rather >> than private (or shared) anonymous memory for VMs? > > The primary use case I know of is sharing VM memory with other processes > (usually not child processes): DPDK/SPDK and other vhost-user variants (such as > virtiofs) mmap() all guest memory to access it directly (some sort of > multi-process hypervisors). They either use real-file-based shmem or memfd > (essentially the same without a named file) for that. > > Then, there is live-hypervisor upgrade, whereby you start a second hypervisor > process that will take over. People use shmem for that, so you can minimize > downtime by migrating guest memory simply by mmap'ing the shmem file into the > new hypervisor. > > Shared anonymous memory is basically never used (I only know of one corner case > in QEMU). > > I would assume that there are also DBs making use of rather sparse shmem? But no > expert on that. > That all makes sense - thanks for the lesson!
