|
From: Arnd B. <ar...@ke...> - 2021-09-28 08:20:28
|
On Tue, Sep 28, 2021 at 1:21 AM Anton Altaparmakov <an...@tu...> wrote:
>
> Hi Arnd,
>
> Thanks for the patch but what is the problem with the stack usage exceeding 2048 bytes?
>
> I am not aware of any architectures that implements kernel stack size (THREAD_SIZE)
> less than page size and by default most architectures with 4kiB page size even use two
> pages to make the stack 8kiB.
The two are decoupled these days unless CONFIG_VMAP_STACK is enabled at build
time, in which case the THREAD_SIZE is always a multiple of STACK_SIZE.
No architecture currently forces the use of VMAP_STACK though, so the allocation
is done in alloc_thread_stack_node() using this kmem_cache:
thread_stack_cache = kmem_cache_create_usercopy("thread_stack",
THREAD_SIZE, THREAD_SIZE, 0, 0,
THREAD_SIZE, NULL);
64K pages are allowed on arm64, powerpc, mips, microblaze, ia64, sh, hexagon
and the upcoming loongarch port. The respective THREAD_SHIFT/THREAD_SIZE
values on these are
arch/arm64/include/asm/memory.h:#define MIN_THREAD_SHIFT (14 +
KASAN_THREAD_SHIFT)
arch/powerpc/Kconfig:config THREAD_SHIFT
arch/powerpc/Kconfig- default "14" if PPC64
arch/mips/include/asm/thread_info.h:#define THREAD_SIZE_ORDER (0)
arch/mips/include/asm/thread_info.h:#define THREAD_SIZE (PAGE_SIZE <<
THREAD_SIZE_ORDER)
arch/microblaze/include/asm/thread_info.h:#define THREAD_SHIFT 13
arch/ia64/include/asm/ptrace.h:# define KERNEL_STACK_SIZE_ORDER 0
arch/ia64/include/asm/ptrace.h:#define IA64_STK_OFFSET
((1 << KERNEL_STACK_SIZE_ORDER)*PAGE_SIZE)
arch/ia64/include/asm/ptrace.h:#define KERNEL_STACK_SIZE
IA64_STK_OFFSET
arch/ia64/include/asm/thread_info.h:#define THREAD_SIZE
KERNEL_STACK_SIZE
arch/sh/include/asm/thread_info.h:#define THREAD_SHIFT 12
arch/hexagon/include/asm/thread_info.h:#define THREAD_SHIFT 12
As far as I can tell, only mips and ia64 require the kernel stack to
be a multiple
of the page size here, and I would consider that a bug: This is extremely
wasteful, especially considering that those machines typically won't have the
vast amounts of RAM that modern arm64 and powerpc64 servers have.
On a hexagon or sh system with 4KB stacks, using over 2KB in one function
is definitely excessive. Those machines wouldn't normally need NTFS support,
but that was kind-of the point of my patch.
Arnd
|
