On Tue, 21 Dec 2021 at 11:39, Marek Szyprowski <m.szyprowski@xxxxxxxxxxx> wrote:
>
> Hi,
>
> On 22.11.2021 10:28, Ard Biesheuvel wrote:
> > Wire up the generic support for managing task stack allocations via vmalloc,
> > and implement the entry code that detects whether we faulted because of a
> > stack overrun (or future stack overrun caused by pushing the pt_regs array)
> >
> > While this adds a fair amount of tricky entry asm code, it should be
> > noted that it only adds a TST + branch to the svc_entry path. The code
> > implementing the non-trivial handling of the overflow stack is emitted
> > out-of-line into the .text section.
> >
> > Since on ARM, we rely on do_translation_fault() to keep PMD level page
> > table entries that cover the vmalloc region up to date, we need to
> > ensure that we don't hit such a stale PMD entry when accessing the
> > stack. So we do a dummy read from the new stack while still running from
> > the old one on the context switch path, and bump the vmalloc_seq counter
> > when PMD level entries in the vmalloc range are modified, so that the MM
> > switch fetches the latest version of the entries.
> >
> > Note that we need to increase the per-mode stack by 1 word, to gain some
> > space to stash a GPR until we know it is safe to touch the stack.
> > However, due to the cacheline alignment of the struct, this does not
> > actually increase the memory footprint of the struct stack array at all.
> >
> > Signed-off-by: Ard Biesheuvel <ardb@xxxxxxxxxx>
> > Tested-by: Keith Packard <keithpac@xxxxxxxxxx>
>
>
> This patch landed recently in linux-next 20211220 as commit a1c510d0adc6
> ("ARM: implement support for vmap'ed stacks"). Sadly it breaks
> suspend/resume operation on all ARM 32bit Exynos SoCs. Probably the
> suspend/resume related code must be updated somehow (it partially works
> on physical addresses and disabled MMU), but I didn't analyze it yet. If
> you have any hints, let me know.
>
Are there any such systems in KernelCI? We caught a suspend/resume
related issue in development, which is why the hunk below was added.
In general, any virt-to-phys translation involving and address on the
stack will become problematic.
Could you please confirm whether the issue persists with the patch
applied but with CONFIG_VMAP_STACK turned off? Just so we know we are
looking in the right place?
> diff --git a/arch/arm/kernel/sleep.S b/arch/arm/kernel/sleep.S
> index 43077e11dafd..803b51e5cba0 100644
> --- a/arch/arm/kernel/sleep.S
> +++ b/arch/arm/kernel/sleep.S
> @@ -67,6 +67,14 @@ ENTRY(__cpu_suspend)
> ldr r4, =cpu_suspend_size
> #endif
> mov r5, sp @ current virtual SP
> +#ifdef CONFIG_VMAP_STACK
> + @ Run the suspend code from the overflow stack so we don't have to rely
> + @ on vmalloc-to-phys conversions anywhere in the arch suspend code.
> + @ The original SP value captured in R5 will be restored on the way out.
> + mov_l r6, overflow_stack_ptr @ Base pointer
> + mrc p15, 0, r7, c13, c0, 4 @ Get per-CPU offset
> + ldr sp, [r6, r7] @ Address of this CPU's overflow stack
> +#endif
> add r4, r4, #12 @ Space for pgd, virt sp, phys resume fn
> sub sp, sp, r4 @ allocate CPU state on stack
> ldr r3, =sleep_save_sp
