Hello, here is second version of my patches to implement synchronous page faults for DAX mappings to make flushing of DAX mappings possible from userspace so that they can be flushed on finer than page granularity and also avoid the overhead of a syscall. We use a new mmap flag MAP_SYNC to indicate that page faults for the mapping should be synchronous. The guarantee provided by this flag is: While a block is writeably mapped into page tables of this mapping, it is guaranteed to be visible in the file at that offset also after a crash. How I implement this is that ->iomap_begin() indicates by a flag that inode block mapping metadata is unstable and may need flushing (use the same test as whether fdatasync() has metadata to write). If yes, DAX fault handler refrains from inserting / write-enabling the page table entry and returns special flag VM_FAULT_NEEDDSYNC together with a PFN to map to the filesystem fault handler. The handler then calls fdatasync() (vfs_fsync_range()) for the affected range and after that calls DAX code to update the page table entry appropriately. >From my (fairly limited) knowledge of XFS it seems XFS should be able to do the same and it should be even possible for filesystem to implement safe remapping of a file offset to a different block (i.e. break reflink, do defrag, or similar stuff) like: 1) Block page faults 2) fdatasync() remapped range (there can be outstanding data modifications not yet flushed) 3) unmap_mapping_range() 4) Now remap blocks 5) Unblock page faults Basically we do the same on events like punch hole so there is not much new there. Note that the implementation of MAP_SYNC flag is pretty crude for now just to enable testing since Dan is working in the same area to implement another mmap flag. Once the decision on how to implement new mmap flag is settled, I can clean up that patch. I did some basic performance testing on the patches over ramdisk - timed latency of page faults when faulting 512 pages. I did several tests: with file preallocated / with file empty, with background file copying going on / without it, with / without MAP_SYNC (so that we get comparison). The results are (numbers are in microseconds): File preallocated, no background load no MAP_SYNC: min=5 avg=6 max=42 4 - 7 us: 398 8 - 15 us: 110 16 - 31 us: 2 32 - 63 us: 2 File preallocated, no background load, MAP_SYNC: min=10 avg=10 max=43 8 - 15 us: 509 16 - 31 us: 2 32 - 63 us: 1 File empty, no background load, no MAP_SYNC: min=21 avg=23 max=76 16 - 31 us: 503 32 - 63 us: 8 64 - 127 us: 1 File empty, no background load, MAP_SYNC: min=91 avg=108 max=234 64 - 127 us: 467 128 - 255 us: 45 File empty, background load, no MAP_SYNC: min=21 avg=23 max=67 16 - 31 us: 507 32 - 63 us: 4 64 - 127 us: 1 File empty, background load, MAP_SYNC: min=94 avg=112 max=181 64 - 127 us: 489 128 - 255 us: 23 So here we can see the difference between MAP_SYNC vs non MAP_SYNC is about 100-200 us when we need to wait for transaction commit in this setup. Anyway, here are the patches, comments are welcome. Changes since v1: * switched to using mmap flag MAP_SYNC * cleaned up fault handlers to avoid passing pfn in vmf->orig_pte * switched to not touching page tables before we are ready to insert final entry as it was unnecessary and not really simplifying anything * renamed fault flag to VM_FAULT_NEEDDSYNC * other smaller fixes found by reviewers Honza
