On Mon, Aug 14, 2017 at 11:12:05PM -0700, Dan Williams wrote: > Changes since v3 [1]: > * Move from an fallocate(2) interface to a new mmap(2) flag and rename > 'immutable' to 'sealed'. > > * Do not record the sealed state in permanent metadata it is now purely > a temporary state for as long as a MAP_DIRECT vma is referencing the > inode (Christoph) > > * Drop the CAP_IMMUTABLE requirement, but do require a PROT_WRITE > mapping. > > [1]: https://lwn.net/Articles/730570/ > > --- > > This is the next revision of a patch series that aims to enable > applications that otherwise need to resort to DAX mapping a raw device > file to instead move to a filesystem. > > In the course of reviewing a previous posting, Christoph said: > > That being said I think we absolutely should support RDMA memory > registrations for DAX mappings. I'm just not sure how S_IOMAP_IMMUTABLE > helps with that. We'll want a MAP_SYNC | MAP_POPULATE to make sure all > the blocks are populated and all ptes are set up. Second we need to > make sure get_user_page works, which for now means we'll need a struct > page mapping for the region (which will be really annoying for PCIe > mappings, like the upcoming NVMe persistent memory region), and we need > to guarantee that the extent mapping won't change while the > get_user_pages holds the pages inside it. I think that is true due to > side effects even with the current DAX code, but we'll need to make it > explicit. And maybe that's where we need to converge - "sealing" the > extent map makes sense as such a temporary measure that is not persisted > on disk, which automatically gets released when the holding process > exits, because we sort of already do this implicitly. It might also > make sense to have explicitly breakable seals similar to what I do for > the pNFS blocks kernel server, as any userspace RDMA file server would > also need those semantics. > > So, this is an attempt to converge on the idea that we need an explicit > and process-lifetime-temporary mechanism for a process to be able to > make assumptions about the mapping to physical page to dax-file-offset > relationship. The "explicitly breakable seals" aspect is not addressed > in these patches, but I wonder if it might be a voluntary mechanism that > can implemented via userfaultfd. > > These pass a basic smoke test and are meant to just gauge 'right track' > / 'wrong track'. The main question it seems is whether the pinning done > in this patchset is too early (applies before get_user_pages()) and too > coarse (applies to the whole file). Perhaps this is where I discarded > too easily Jan's suggestion to look at Peter Z's mm_mpin() syscall [2]? On > the other hand, the coarseness and simple lifetime rules of MAP_DIRECT > make it an easy mechanism to implement and explain. > > Another reason I kept the scope of S_IOMAP_SEALED coarsely defined was > to support Dave's desired use case of sealing for operating on reflinked > files [3]. Which really needs a fcntl() interface to set/clear iomap seals. Which, now that I look at it, already has a bunch of "file sealing" commands defined which arrived in 3.17. It appears to be a special purpose access control interface for memfd_create() to manage shared access to anonymous tmpfs files and will EINVAL on any fd that points to a real file. Oh, even more problematic: Seals are a property of an inode. [....] Furthermore, seals can never be removed, only added. That seems somewhat difficult to reconcile with how I need F_SEAL_IOMAP to operate. /me calls it a day and goes looking for the hard liquor..... Cheers, Dave. -- Dave Chinner david@xxxxxxxxxxxxx -- To unsubscribe from this list: send the line "unsubscribe linux-xfs" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
