On Wed, Mar 25, 2026 at 5:08 PM Pasha Tatashin <[email protected]> wrote: > > On Wed, Mar 25, 2026 at 4:34 PM David Matlack <[email protected]> wrote: > > > > On Wed, Mar 25, 2026 at 1:20 PM Pratyush Yadav <[email protected]> wrote: > > > > > For memfd and hugetlb at least, we serialize the _inode_ not the file. > > > The inode has the contents that we care to preserve. > > > > > > So if two FDs point to the same inode, this will break. You can do this > > > by first creating a memfd and then by opening "/proc/self/fd/<fd>". Then > > > you would be able to trigger the preservation twice, causing all sorts > > > of problems. Same on the retrieve side. > > Hm. > > > > > > So unless I am missing something, I don't think this approach will work. > > > As much as I hate to suggest it, I think we need to move this check to > > > each caller so they can find out the object they need to serialize and > > > check if it already is. > > > > I think LUO can still enforce that the file is not preserved twice. > > HugeTLB and memfd's preserve() functions just need to also check that > > the associated inode has not already been preserved? > > For memfd/hugetlbs the true state is in inode > For vfio/kvm the shared anonymous inode is just a dummy wrapper, and > the true state is in file->private_data. > > I wonder if we could use the XArray to track inodes for standard > files, but track the struct file itself for anonymous files (we would > need a new function from FS that allows us to determine if "struct > file" has anonymous inode or not).
Actually, let's not modify the fs layer, instead, add an optional get_id(struct file *file) callback to the luo file handler (struct liveupdate_file_ops). This would return a unique deduplication key. For example, the memfd callback would return (unsigned long)file->f_inode. If get_id() is not implemented for a given handler, the LUO would default to using the struct file pointer directly > > Pasha
