[PATCH v2] sched: move stack_canary to the start of the randomizable region

[Charles Mirabile]

Hi Ruidong, Valentin, (+ copying in RISC-V list/maintainers)—

On Tue, May 12, 2026 at 06:16:43PM +0200, Valentin Schneider wrote:
> On 09/05/26 11:50, Ruidong Tian wrote:
> > task_struct keeps growing over time.  On architectures that compute the
> > per-task stack canary offset from asm-offsets.h and pass it to the
> > compiler via -mstack-protector-guard-offset=, this growth eventually
> > pushes stack_canary beyond what the target ISA can encode.
> >
> > On RISC-V, canary loads are emitted as
> >
> >     ld	t0, TSK_STACK_CANARY(tp)
> >
> > where TSK_STACK_CANARY must fit into a 12-bit signed immediate, i.e.
> > [-2048, 2047].  Once stack_canary sits past byte 2047 of task_struct,
> > the build fails with
> >
> >     cc1: error: '<N>' is not a valid offset in
> >         '-mstack-protector-guard-offset='

Thanks for this patch. We just ran into this exact issue during a CKI build
https://datawarehouse.cki-project.org/kcidb/builds/1639782 and this seems
like exactly the fix we need.

> >
> >   * On RISC-V, CONFIG_STACKPROTECTOR_PER_TASK depends on !RANDSTRUCT,
> >     so randomized_struct_fields_start/end always expand to nothing and
> >     stack_canary lands at a small, stable offset well within the
> >     12-bit signed immediate range.  The build error goes away.
> >
> >   * On architectures that enable RANDSTRUCT for hardening, stack_canary
> >     stays inside the randomized region and is still shuffled together
> >     with the other fields by the layout randomization, so its hardening
> >     coverage is preserved.  asm-offsets-based architectures read the
> >     shuffled offset at build time, so the generated canary accesses
> >     remain correct.
> >
> > pahole on a typical 64-bit config shows that the area around the
> > wakee_* fields already contains a usable hole:
> >
> >     struct __call_single_node  wake_entry;           /*    56    16 */
> >     /* --- cacheline 1 boundary (64 bytes) ---       */
> >     unsigned int               wakee_flips;          /*    72     4 */
> >     /* XXX 4 bytes hole, try to pack */
> >     unsigned long int          wakee_flip_decay_ts;  /*    80     8 */
> >     struct task_struct *       last_wakee;           /*    88     8 */
> >
> > Move wakee_flips to sit after last_wakee.  That opens up a clean
> > 8-byte slot at offset 72 into which stack_canary fits exactly:
> >
> >     struct __call_single_node  wake_entry;           /*    56    16 */
> >     /* --- cacheline 1 boundary (64 bytes) ---       */
> >     unsigned long              stack_canary;         /*    72     8 */
> >     unsigned long int          wakee_flip_decay_ts;  /*    80     8 */
> >     struct task_struct *       last_wakee;           /*    88     8 */
> >     unsigned int               wakee_flips;          /*    96     4 */
> >
> > Suggested-by: K Prateek Nayak <kprateek.nayak at amd.com>
> > Signed-off-by: Ruidong Tian <tianruidong at linux.alibaba.com>
> 
> Reviewed-by: Valentin Schneider <vschneid at redhat.com>
> 

Thanks for the review, is there any hope of this patch going in during
the 7.3 merge window?

I have also copied the RISC-V list as well, maybe if it is ok with the
sched folks, the RISC-V maintainers could take it through their tree as
a RISC-V specific urgent fix?

Best—Charlie