On Thu, Oct 21, 2021 at 04:42:33PM +0200, Andreas Gruenbacher wrote: > On Thu, Oct 21, 2021 at 12:06 PM Catalin Marinas > <catalin.marinas@xxxxxxx> wrote: > > On Thu, Oct 21, 2021 at 02:46:10AM +0200, Andreas Gruenbacher wrote: > > > When a page fault would occur, we > > > get back an error instead, and then we try to fault in the offending > > > pages. If a page is resident and we still get a fault trying to access > > > it, trying to fault in the same page again isn't going to help and we > > > have a true error. > > > > You can't be sure the second fault is a true error. The unlocked > > fault_in_*() may race with some LRU scheme making the pte not accessible > > or a write-back making it clean/read-only. copy_to_user() with > > pagefault_disabled() fails again but that's a benign fault. The > > filesystem should re-attempt the fault-in (gup would correct the pte), > > disable page faults and copy_to_user(), potentially in an infinite loop. > > If you bail out on the second/third uaccess following a fault_in_*() > > call, you may get some unexpected errors (though very rare). Maybe the > > filesystems avoid this problem somehow but I couldn't figure it out. > > Good point, we can indeed only bail out if both the user copy and the > fault-in fail. > > But probing the entire memory range in fault domain granularity in the > page fault-in functions still doesn't actually make sense. Those > functions really only need to guarantee that we'll be able to make > progress eventually. From that point of view, it should be enough to > probe the first byte of the requested memory range, so when one of > those functions reports that the next N bytes should be accessible, > this really means that the first byte surely isn't permanently > inaccessible and that the rest is likely accessible. Functions > fault_in_readable and fault_in_writeable already work that way, so > this only leaves function fault_in_safe_writeable to worry about. I agree, that's why generic_perform_write() works. It does a get_user() from the first byte in that range and the subsequent copy_from_user() will make progress of at least one byte if it was readable. Eventually it will hit the byte that faults. The gup-based fault_in_*() are a bit more problematic. Your series introduces fault_in_safe_writeable() and I think for MTE doing a _single_ get_user(uaddr) (in addition to the gup checks for write) would be sufficient as long as generic_file_read_iter() advances by at least one byte (eventually). This discussion started with the btrfs search_ioctl() where, even if some bytes were written in copy_to_sk(), it always restarts from an earlier position, reattempting to write the same bytes. Since copy_to_sk() doesn't guarantee forward progress even if some bytes are writable, Linus' suggestion was for fault_in_writable() to probe the whole range. I consider this overkill since btrfs is the only one that needs probing every 16 bytes. The other cases like the new fault_in_safe_writeable() can be fixed by probing the first byte only followed by gup. I think we need to better define the semantics of the fault_in + uaccess sequences. For uaccess, we document "a hard requirement that not storing anything at all (i.e. returning size) should happen only when nothing could be copied" (from linux/uaccess.h). I think we can add a requirement for the new size_t-based fault_in_* variants without mandating that the whole range is probed, something like: "returning leftover < size guarantees that a subsequent user access at uaddr copies at least one byte eventually". I said "eventually" but maybe we can come up with some clearer wording for a liveness property. Such requirement would ensure that infinite loops of fault_in_* + uaccess make progress as long as they don't reset the probed range. Code like btrfs search_ioctl() would need to be adjusted to avoid such range reset and guarantee forward progress. -- Catalin
