Re: [PATCH v2] sparc64: increase kernel thread stack size to 32K

[Tony Rodriguez <unixpro1970@gmail.com>]

Hi Stian,

Thanks for following up regarding this, I am currently busy with other 
tasks.  Back in May/June 2026, I also mentioned the following to Andreas:

The combined stack usage is already very close to the 32K limit. If you 
need an immediate workaround, increasing the stack to 64 K may be the 
better option. That should provide enough headroom to avoid the stack 
overflow panics we are seeing.  As of kernel 7.1, I haven't noticed any 
panics using a 32K stack, but once again a stack size that is so close 
to the 32K boundary is a concern. But once again, sparc64 definitely 
crashes with a 16K stack.  The Nvidia/Mellonox mlx5 driver also 
allocates a big stack on sparc64, which pushes it to the 32K boundary.  
There may be other drivers that may do so as well.

The 64K stack is more ideal.

When I compile the kernel with -fstack-usage to generate .su files, on 
7.1 kerner, the static analysis shows small stack frames for all USB 
core functions.
For example:

hub_event:      2457 bytes  (static)
hub_activate:   1892 bytes  (static)
usb_control_msg: 1248 bytes (static)

However, my runtime stack tracing shows a dramatically different picture:

STACKTRACE: hub_event():entry: 31856 bytes used
STACKTRACE: hub_activate():entry: 31680 bytes used
STACKTRACE: usb_control_msg():entry: 30768 bytes used


PS - I have not tested a 64K stack yet, only 32K, and this is a heads-up 
recommendation.


Tony

On 8/31/26 10:29 AM, Stian Halseth wrote:
> From: Tony Rodriguez <unixpro1970@gmail.com>
>
> Kernel stacks on sparc64 are 16K and this is no longer enough:
> several machines (SPARC T5-2 among them) panic early in boot during
> USB hub enumeration with "corrupted stack end detected inside
> scheduler". sparc has not been converted to THREAD_INFO_IN_TASK, so
> thread_info sits at the bottom of the kernel stack and a marginal
> overflow corrupts it first; CONFIG_SCHED_STACK_END_CHECK then fires
> from __schedule long after the deep path has unwound, which is why
> the reported backtraces look shallow.
>
> Measurements with CONFIG_STACK_TRACER on an UltraSPARC T4-1 show the
> problem is frame count, not any single large frame. The high-water
> mark of an ordinary successful boot is 12616 of 16384 bytes (77%),
> reached in hub_probe() with a printk console flush and then a timer
> interrupt (which runs on the task stack, and whose scheduler tick
> performs load balancing and IPI delivery) stacked on top. Of the 66
> frames in that path the largest is 408 bytes, and ~85% of them are
> 176-224 bytes - at or just above the SPARC V9 ABI minimum frame
> (128-byte register window save area plus 48-byte argument save
> area). An equivalent call chain on x86-64 costs roughly a third of
> the stack, so a 16K stack on sparc64 provides far less effective
> call depth than on other 64-bit architectures.
>
> Double THREAD_SIZE to 32K (four 8K pages). The PAGE_SHIFT
> conditionals are dropped: sparc64 only supports 8K base pages, so
> the other branches were dead code. Kernel stacks become order-2
> allocations; sparc64 has no VMAP_STACK, but stacks are allocated
> once per thread and the trade against boot-time panics is a good
> one.
>
> Link: https://lore.kernel.org/all/20260519075809.8993-1-unixpro1970@gmail.com/
> Signed-off-by: Tony Rodriguez <unixpro1970@gmail.com>
> [stian: reduced the diff to the THREAD_* defines, measured stack
>   usage with CONFIG_STACK_TRACER and rewrote the changelog]
> Signed-off-by: Stian Halseth <stian@itx.no>
> ---
> v2:
>   - drop the CONFIG_SPARC64 / PAGE_SHIFT conditional chain from v1;
>     thread_info_64.h is only built on sparc64 and only 8K pages are
>     supported, so define the three constants unconditionally
>   - replace the panic backtrace in the changelog with stack tracer
>     measurements answering David Laight's review comments:
>     https://lore.kernel.org/all/20260520144104.618c75ca@pumpkin/
>   - retitled from "unify thread stack sizing and add explicit 32KB
>     stack"; the sizing logic for other configurations is unchanged
>
> Tested on an UltraSPARC T4-1, booted with the stack tracer armed
> ("stacktrace") before and after this patch. The boot high-water mark
> is 12616 bytes on both kernels - the worst path (hub_probe with a
> printk and a timer interrupt on top) is deterministic - i.e. 77% of
> the 16K stack before, 38% of the 32K stack after. DEBUG_STACK_USAGE
> agrees: the peak boot-time task shows 9208 bytes left of 16K before
> vs 25592 bytes left of 32K after (7176 bytes used in both).
>
>   arch/sparc/include/asm/thread_info_64.h | 15 +++------------
>   1 file changed, 3 insertions(+), 12 deletions(-)
>
> diff --git a/arch/sparc/include/asm/thread_info_64.h b/arch/sparc/include/asm/thread_info_64.h
> --- a/arch/sparc/include/asm/thread_info_64.h
> +++ b/arch/sparc/include/asm/thread_info_64.h
> @@ -99,13 +99,8 @@
>   #define FAULT_CODE_BLKCOMMIT	0x10	/* Use blk-commit ASI in copy_page */
>   #define	FAULT_CODE_BAD_RA	0x20	/* Bad RA for sun4v		   */
>   
> -#if PAGE_SHIFT == 13
> -#define THREAD_SIZE (2*PAGE_SIZE)
> -#define THREAD_SHIFT (PAGE_SHIFT + 1)
> -#else /* PAGE_SHIFT == 13 */
> -#define THREAD_SIZE PAGE_SIZE
> -#define THREAD_SHIFT PAGE_SHIFT
> -#endif /* PAGE_SHIFT == 13 */
> +#define THREAD_SIZE (4 * PAGE_SIZE)
> +#define THREAD_SHIFT (PAGE_SHIFT + 2)
>   
>   /*
>    * macros/functions for gaining access to the thread information structure
> @@ -128,11 +123,7 @@
>   #endif
>   
>   /* thread information allocation */
> -#if PAGE_SHIFT == 13
> -#define THREAD_SIZE_ORDER	1
> -#else /* PAGE_SHIFT == 13 */
> -#define THREAD_SIZE_ORDER	0
> -#endif /* PAGE_SHIFT == 13 */
> +#define THREAD_SIZE_ORDER	2
>   
>   #define __thread_flag_byte_ptr(ti)	\
>   	((unsigned char *)(&((ti)->flags)))
> --
> 2.53.0