On Mon, Oct 28, 2024 at 04:00:29PM -0400, Liam R. Howlett wrote: > * Liam R. Howlett <Liam.Howlett@xxxxxxxxxx> [241028 15:50]: > > * Lorenzo Stoakes <lorenzo.stoakes@xxxxxxxxxx> [241028 15:14]: > > > On Mon, Oct 28, 2024 at 09:05:44AM -1000, Linus Torvalds wrote: > > > > On Mon, 28 Oct 2024 at 08:57, Lorenzo Stoakes > > > > <lorenzo.stoakes@xxxxxxxxxx> wrote: > > > > > > > > > > So likely hook on your mapping changes flags to set VM_MTE | VM_MTE_ALLOWED and > > > > > expects this to be checked after (ugh). > > > > > > > > Gaah. Yes. mm/shmem.c: shmem_mmap() does > > > > > > > > /* arm64 - allow memory tagging on RAM-based files */ > > > > vm_flags_set(vma, VM_MTE_ALLOWED); > > > > > > > > and while I found the equivalent hack for the VM_SPARC_ADI case, I > > > > hadn't noticed that MTE thing. > > > > > > > > How very annoying. > > > > > > > > So the arch_validate_flags() case does need to be done after the ->mmap() call. > > > > > > > > How about just finalizing everything, and then doing a regular > > > > munmap() afterwards and returning an error (all still holding the mmap > > > > semaphore, of course). > > > > > > > > That still avoids the whole "partially completed mmap" case. > > > > > > > > Linus > > > > > > Yeah I was thinking the same... just bite the bullet, go through the whole damn > > > process and revert if arch_validate_flags() chokes. It also removes the ugly > > > #ifdef CONFIG_SPARC64 hack... > > > > > > This will litearlly only be applicable for these two cases and (hopefully) most > > > of the time you'd not fail it. > > > > > > I mean by then it'll be added into the rmap and such but nothing will be > > > populated yet and we shouldn't be able to fault as vma_start_write() should have > > > incremented the vma lock seqnum. > > > > > > Any issues from the RCU visibility stuff Liam? > > > > It is probably fine? We would see a mapping appear then disappear. > > We'd have a (benign) race with rmap for truncating the PTEs (but it's > > safe). Page faults would be stopped though. > > > > Unfortunately, we'd have to write to the vma tree so that we could... > > write to the vma tree. We'd have to somehow ensure munmap() is done > > with a gfp flag to ensure no failures as well... > > > > Maybe we should just call close on the vma again (and do whatever > > call_mmap() needs to undo)? > > I take it back, that won't work. > > > > > > > > > Any security problems Jann...? > > > > > > It'd suck to have to bring back a partial complete case. Though I do note we > > > handle errors from mmap_file() ok so we could still potentially handle that > > > there, but would sort of semi-undo some of the point of the series. I'm genuinely not opposed to a horrible, awful: #ifdef CONFIG_ARM64 if (file && file->f_ops == shmem_file_operations) vm_flags |= VM_MTE_ALLOWED; #endif Early in the operation prior to the arch_validate_flags() check. Just to get this over the finish line (original idea credit to Vlastimil, insane ugliness credit to me). Could be in a __arch_workarounds() type function heavily commented... I mean this is pretty gross but...
