Re: [PATCH 02/12] arm64/module: Fix BTI exceptions caused by omitted landing pads in Clang 21

["Ard Biesheuvel" <ardb@kernel.org>]

On Sat, 15 Aug 2026, at 21:57, Josh Poimboeuf wrote:
> On Sat, Aug 15, 2026 at 12:56:11PM +0300, Ard Biesheuvel wrote:
>> Hi Josh,
>> 
>> On Sat, 15 Aug 2026, at 07:45, Josh Poimboeuf wrote:
>> > The following BTI exception was seen when loading a livepatch module:
>> >
>> >   Internal error: Oops - BTI: 0000000036000001 [#1]  SMP
>> >   pstate: 634004c9 (nZCv daIF +PAN -UAO +TCO +DIT -SSBS BTYPE=jc)
>> >   pc : kill_orphaned_pgrp+0x0/0x150
>> >   lr : do_exit+0x498/0xaf0 [livepatch_combined]
>> >
>> > The problem is that the patch module's do_exit() is branching to a
>> > static function in vmlinux using a module PLT veneer (indirect branch),
>> > but the target function doesn't have a BTI landing pad.
>> >
>> > Clang 21+ omits the landing pad for static functions which can only be
>> > reached by a direct branch.  But livepatch modules use klp relocations
>> > to reference arbitrary kernel symbols, and with
>> > CONFIG_RANDOMIZE_MODULE_REGION_FULL the module is far enough away that
>> > every call to vmlinux needs a PLT.
>> >
>> > Note this problem is actually not specific to livepatch.  It's possible
>> > for any module's .init section to be allocated > 128MB away from its
>> > .text section.  So calls from .init to .text via a PLT can trigger a BTI
>> > exception when the target function doesn't have a landing pad.
>> >
>> > GCC has always omitted the landing pad when possible, so kernel BTI is
>> > already considered incompatible with GCC since commit c0a454b9044f
>> > ("arm64/bti: Disable in kernel BTI when cross section thunks are
>> > broken").
>> >
>> > When missing landing pads are detected, allocate a page close to the
>> > target which can be used to hold BTI veneers which receive PLT veneer
>> > indirect branches and direct branch to the final target:
>> >
>> 
>> This does not work for cross-section calls from .init.text to .text.
>> 
>> If .init.text is far away from .text, it is likely because .text
>> ended up in the 128M 'near' module region, and .init.text did not.
>> (They tend to end up in direct branching range of each otherwise.)
>> 
>> Given that the module init code is typically small, I don't think
>> it is safe to assume that allocating a single page close enough to
>> .text is going to be possible if allocating the space for .init.*
>> was not.
>> 
>> IOW, the fix I proposed for cross-section calls is still needed
>> with this approach.
>
> But the BTI veneer page is allocated from a *256MB* window, of which the
> near region is only a 128MB subset.
>

OK, so this window would be [_etext - 128M, _text + 128M), right? Or
whichever the equivalent is for a kernel module?

I agree that this considerably reduces the likelihood of running out
of veneer space within direct branching range.

> There is a theoretical case where the 256MB window around the target is
> completely full without any fragmentation, but I would think that there
> would almost always be some fragmentation.  If that window is modules
> stacked together, most modules have at least .init.plt and .init.text,
> and many have .init.data.
>

Are you saying these are freed again once the module is initialized, and
therefore there is fragmentation? Not sure if it works like that if modules
are loaded one after the other.

> Right now it needs two pages (because of the default guard page) but we
> could maybe fall back to VM_NO_GUARD in case of emergency.
>

I like the advantage of only allocating veneers if they are really needed,
as opposed to allocating the space unconditionally, which is what my
solution does. But I need more convincing that this will actually cover
all these cases, given that the cross-section call issue only occurs when
running out of direct branching space to begin with.