On Tue, Jan 19, 2021 at 12:34:23PM +0000, Alessio Balsini wrote: > On Tue, Jan 19, 2021 at 07:06:54PM +0800, Rokudo Yan wrote: > > on Mon, Jan 18, 2021 at 5:27 PM Alessio Balsini <balsini@xxxxxxxxxxx> wrote: > > > > > > This is the 11th version of the series, rebased on top of v5.11-rc4. > > > Please find the changelog at the bottom of this cover letter. > > > > > > Add support for file system passthrough read/write of files when enabled > > > in userspace through the option FUSE_PASSTHROUGH. > > [...] > > > > > > Hi Allesio, > > > > Could you please add support for passthrough mmap too ? > > If the fuse file opened with passthrough actived, and then map (shared) to (another) process > > address space using mmap interface. As access the file with mmap will pass the vfs cache of fuse, > > but access the file with read/write will bypass the vfs cache of fuse, this may cause inconsistency. > > eg. the reader read the fuse file with mmap() and the writer modify the file with write(), the reader > > may not see the modification immediately since the writer bypass the vfs cache of fuse. > > Actually we have already meet an issue caused by the inconsistency after applying fuse passthrough > > scheme to our product. > > > > Thanks, > > yanwu. > > Hi yanwu, > > Thank you for your interest in this change. > > FUSE passthrough for mmap is an extension that is already in my TODO > list, together with passthrough for directories. > For now I would prefer to keep this series minimal to make the review > process leaner and simpler. > I will start working on extending this series with new features and > addressing more corner cases as soon as these changes get merged, what > do you think? > > Thanks, > Alessio Hi yanwu, Sorry if I overlooked this issue. I added memory-mapping to my tests and could reproduce/verify this wrong behavior you mentioned. I created this WIP (history may change) branch that has the missing mmap implementation: https://github.com/balsini/linux/commits/fuse-passthrough-v12-develop-v5.11-rc4 I did some mmap testing in the last days with this extra mmap implementation and couldn't find any issues, everything seems to be working as expected with the extra mmap patch. Can you please confirm this is fixed on your end too? I'm also going to revert in this branch the stacking policy changes to how they were in V10 as suggested by Amir if there are no concerns with that. I'm waiting for some extra tests to complete and, if no issue is detected, I'll post the V12 series super soon. Thanks, Alessio
