On Fri, Sep 23, 2022 at 01:08:31PM -0600, Logan Gunthorpe wrote: > > > On 2022-09-23 12:13, Jason Gunthorpe wrote: > > On Thu, Sep 22, 2022 at 10:39:19AM -0600, Logan Gunthorpe wrote: > >> GUP Callers that expect PCI P2PDMA pages can now set FOLL_PCI_P2PDMA to > >> allow obtaining P2PDMA pages. If GUP is called without the flag and a > >> P2PDMA page is found, it will return an error. > >> > >> FOLL_PCI_P2PDMA cannot be set if FOLL_LONGTERM is set. > > > > What is causing this? It is really troublesome, I would like to fix > > it. eg I would like to have P2PDMA pages in VFIO iommu page tables and > > in RDMA MR's - both require longterm. > > You had said it was required if we were relying on unmap_mapping_range()... Ah.. Ok. Dan and I have been talking about this a lot, and it turns out the DAX approach of unmap_mapping_range() still has problems, really the same problem as FOLL_LONGTERM: https://lore.kernel.org/all/Yy2pC%2FupZNEkVmc5@xxxxxxxxxx/ ie nothing actually waits for the page refs to go to zero during memunmap_pages(). (indeed they are not actually zero because currently they are instantly reset to 1 if they become zero) The current design requires that the pgmap user hold the pgmap_ref in a way that it remains elevated until page_free() is called for every page that was ever used. I'm encouraging Dan to work on better infrastructure in pgmap core because every pgmap implementation has this issue currently. For that reason it is probably not so relavent to this series. Perhaps just clarify in the commit message that the FOLL_LONGTERM restriction is to copy DAX until the pgmap page refcounts are fixed. > > Is it just because ZONE_DEVICE was created for DAX and carried that > > revocable assumption over? Does anything in your series require > > revocable? > > We still rely on unmap_mapping_range() indirectly in the unbind > path. So I expect if something takes a LONGERM mapping that would > block until whatever process holds the pin releases it. That's less > than ideal and I'm not sure what can be done about it. We could improve the blocking with some kind of FOLL_LONGTERM notifier thingy eg after the unmap_mapping_rage() broadcast that a range of PFNs is going away and FOLL_LONGTERM users can do a revoke if they support it. It is a rare enough we don't necessarily need to optimize this alot, and blocking unbind until some FDs close is annoying not critical.. (eg you already can't unmount a filesystem to unbind the device on the nvme while FS FDs are open) Jason
