On Mon, Apr 27, 2020 at 05:56:42PM +0100, Dave P Martin wrote: > On Tue, Apr 21, 2020 at 03:26:00PM +0100, Catalin Marinas wrote: > > The copy_mount_options() function takes a user pointer argument but not > > a size. It tries to read up to a PAGE_SIZE. However, copy_from_user() is > > not guaranteed to return all the accessible bytes if, for example, the > > access crosses a page boundary and gets a fault on the second page. To > > work around this, the current copy_mount_options() implementations > > performs to copy_from_user() passes, first to the end of the current > > page and the second to what's left in the subsequent page. > > > > Some architectures like arm64 can guarantee an exact copy_from_user() > > depending on the size (since the arch function performs some alignment > > on the source register). Introduce an arch_has_exact_copy_from_user() > > function and allow copy_mount_options() to perform the user access in a > > single pass. > > > > While this function is not on a critical path, the single-pass behaviour > > is required for arm64 MTE (memory tagging) support where a uaccess can > > trigger intra-page faults (tag not matching). With the current > > implementation, if this happens during the first page, the function will > > return -EFAULT. > > Do you know how much extra overhead we'd incur if we read at must one > tag granule at a time, instead of PAGE_SIZE? Our copy routines already read 16 bytes at a time, so that's the tag granule. With current copy_mount_options() we have the issue that it assumes a fault in the first page is fatal. Even if we change it to a loop of smaller uaccess, we still have the issue of unaligned accesses which can fail without reading all that's possible (i.e. the access goes across a tag granule boundary). The previous copy_mount_options() implementation (from couple of months ago I think) had a fallback to byte-by-byte, didn't have this issue. > I'm guessing that in practice strcpy_from_user() type operations copy > much less than a page most of the time, so what we lose in uaccess > overheads we _might_ regain in less redundant copying. strncpy_from_user() has a fallback to byte by byte, so we don't have an issue here. The above is only for synchronous accesses. For async, in v3 I disabled such checks for the uaccess routines. -- Catalin
