On Sat, Sep 10, 2022, at 2:52 PM, Guo Ren wrote:
> On Thu, Sep 8, 2022 at 3:37 PM Arnd Bergmann <arnd@xxxxxxxx> wrote:
>> On Thu, Sep 8, 2022, at 4:25 AM, guoren@xxxxxxxxxx wrote:
>> > From: Guo Ren <guoren@xxxxxxxxxxxxxxxxx>
>> - When VMAP_STACK is set, make it possible to select non-power-of-two
>> stack sizes. Most importantly, 12KB should be a really interesting
>> choice as 8KB is probably still not enough for many 64-bit workloads,
>> but 16KB is often more than what you need. You probably don't
>> want to allow 64BIT/8KB without VMAP_STACK anyway since that just
>> makes it really hard to debug, so hiding the option when VMAP_STACK
>> is disabled may also be a good idea.
> I don't want this config to depend on VMAP_STACK. Some D1 chips would
> run with an 8K stack size and !VMAP_STACK.
That sounds like a really bad idea, why would you want to risk
using such a small stack without CONFIG_VMAP_STACK?
Are you worried about increased memory usage or something else?
> /* thread information allocation */
> -#ifdef CONFIG_64BIT
> -#define THREAD_SIZE_ORDER (2 + KASAN_STACK_ORDER)
> -#else
> -#define THREAD_SIZE_ORDER (1 + KASAN_STACK_ORDER)
> -#endif
> +#define THREAD_SIZE_ORDER CONFIG_THREAD_SIZE_ORDER
> #define THREAD_SIZE (PAGE_SIZE << THREAD_SIZE_ORDER)
This doesn't actually allow additional THREAD_SIZE values, as you
still round up to the nearest power of two.
I think all the non-arch code can deal with non-power-of-2
sizes, so you'd just need
#define THREAD_SIZE round_up(CONFIG_THREAD_SIZE, PAGE_SIZE)
and fix up the risc-v specific code to do the right thing
as well. I now see that THREAD_SIZE_ORDER is not actually
used anywhere with CONFIG_VMAP_STACK, so I suppose that
definition can be skipped, but you still need a THREAD_ALIGN
definition that is a power of two and at least a page larger
than THREAD_SIZE.
Arnd
