Re: [PATCH v5 3/8] arm64: entry: add unwind info for various kernel entries

On Wed, Apr 29, 2026 at 8:26=E2=80=AFAM Mark Rutland <[email protected]>=
 wrote:
>
> Hi Dylan,
>
> On Tue, Apr 28, 2026 at 06:36:38PM +0000, Dylan Hatch wrote:
> > From: Weinan Liu <[email protected]>
> >
> > DWARF CFI (Call Frame Information) specifies how to recover the return
> > address and callee-saved registers at each PC in a given function.
> > Compilers are able to generate the CFI annotations when they compile
> > the code to assembly language. For handcrafted assembly, we need to
> > annotate them by hand.
> >
> > Annotate minimal CFI to enable stacktracing using SFrame for kernel
> > exception entries through el1*_64_*() paths
>
> I thought we were only consuming SFrame when unwinding an exeption
> boundary?
>
> We shouldn't be taking exceptions _from_ the entry assembly functions
> unless something has gone horribly wrong, and so I don't see why we'd
> need CFI entries for the entry assembly functions.
>
> Am I missing some reason we need CFI entries for the entry assembly
> functions? I strongly suspect it is not necessary to add these, and I'd
> prefer to omit them.

I believe the el1 entry functions are called in an exception, and are
called before call_on_irq_stack. Example stacktrace segment:

[  262.119564]  handle_percpu_devid_irq+0xb4/0x348
[  262.119913]  handle_irq_desc+0x3c/0x68
[  262.120196]  generic_handle_domain_irq+0x20/0x40
[  262.120678]  gic_handle_irq+0x48/0xe0
[  262.121005]  call_on_irq_stack+0x30/0x48
[  262.121412]  do_interrupt_handler+0x88/0xa0
[  262.121779]  el1_interrupt+0x38/0x58
[  262.122089]  el1h_64_irq_handler+0x18/0x30
[  262.122617]  el1h_64_irq+0x6c/0x70
[  262.123159]  _raw_spin_unlock_irq+0x10/0x60 (P)
[  262.123720]  __filemap_add_folio+0x200/0x580 (L)
[  262.124145]  filemap_add_folio+0xec/0x300
[  262.124674]  page_cache_ra_unbounded+0x128/0x368
[  262.125338]  do_page_cache_ra+0x70/0x98
[  262.125875]  page_cache_ra_order+0x460/0x4e0

Here, el1h_64_irq is the last function that appears in the exception
stack before _raw_spin_unlock_irq and __filemap_add_folio are
recovered from the saved PC and LR, respectively. So we therefore need
the CFI annotations in order to unwind through the full exception
boundary.

Is my interpretation here correct?

>
> > and irq entries through call_on_irq_stack()
>
> Needing some sort of unwind annotations for call_on_irq_stack() makes
> sense to me, but don't we need something for other assembly functions
> too?
>
> We can interrupt things like memset(); I assume we'll treat those as
> unreliable until annotated?

While looking into adding these annotations, I noticed a pattern where
a sibling call is made to a local function:

SYM_FUNC_START(__pi_memset)
alternative_if_not ARM64_HAS_MOPS
        b       __pi_memset_generic
alternative_else_nop_endif

        mov     dst, dstin
        setp    [dst]!, count!, val_x
        setm    [dst]!, count!, val_x
        sete    [dst]!, count!, val_x
        ret
SYM_FUNC_END(__pi_memset)

In this case, do we consider the stacktrace unreliable since
__pi_memset may not appear in the trace? Or is this not important
because assembly functions cannot be directly livepatched anyway?

Thanks,
Dylan