Re: [PATCH] string: Disable read_word_at_a_time() optimizations if kernel MTE is enabled

[Mark Rutland]

On Tue, Mar 11, 2025 at 11:45:21AM +0000, Catalin Marinas wrote:
> On Mon, Mar 10, 2025 at 07:37:32PM +0000, Mark Rutland wrote:
> > I was worried that ex_handler_load_unaligned_zeropad() might not do the
> > right thing in response to a tag check fault (e.g. access the wrong 8
> > bytes), but it looks as though that's ok due to the way it generates the
> > offset and the aligned pointer.
> > 
> > If load_unaligned_zeropad() is handed a string that starts with an
> > unexpected tag (and even if that starts off aligned),
> > ex_handler_load_unaligned_zeropad() will access that and cause another
> > tag check fault, which will be reported.
> 
> Yes, it will report an async tag check fault on the
> exit_to_kernel_mode() path _if_ load_unaligned_zeropad() triggered the
> fault for other reasons (end of page).

Sorry, yes. The aligned case I mentioned shouldn't apply here.

> It's slightly inconsistent, we could set TCO for the async case in
> ex_handler_load_unaligned_zeropad() as well.

Yep, I think that'd be necessary for async mode.

> For sync checks, we'd get the first fault ending up in
> ex_handler_load_unaligned_zeropad() and a second tag check fault while
> processing the first. This ends up in do_tag_recovery and we disable
> tag checking after the report. Not ideal but not that bad.

Yep; that's what I was describing in the second paragraph above, though
I forgot to say that was assuming sync or asymm mode.

> We could adjust ex_handler_load_unaligned_zeropad() to return false if
> the pointer is already aligned but we need to check the semantics of
> load_unaligned_zeropad(), is it allowed to fault on the first byte?

IIUC today it's only expected to fault due to misalignment, and the
gneral expectation is that for a sequence of load_unaligned_zeropad()
calls, we should get at least one byte without faulting (for the NUL
terminator).

I reckon it'd be better to figure this out based on the ESR if possible.
Kristina's patches for MOPS would give us that.

Mark.