Re: [PATCH v8 2/2] lib: checksum: Use aligned accesses for ip_fast_csum and csum_ipv6_magic tests

[Charlie Jenkins <charlie@rivosinc.com>]

On Thu, Feb 15, 2024 at 10:42:29AM -0800, Guenter Roeck wrote:
> On 2/15/24 09:25, John David Anglin wrote:
> > On 2024-02-15 12:06 p.m., Guenter Roeck wrote:
> > > On 2/15/24 08:51, John David Anglin wrote:
> > > > On 2024-02-15 10:44 a.m., Guenter Roeck wrote:
> > > > > On 2/15/24 02:27, David Laight wrote:
> > > > > > ...
> > > > > > > It would be worthwhile tracking this down since there are
> > > > > > > lots of unaligned data accesses (8-byte accesses on 4-byte aligned addresses)
> > > > > > > when running the kernel in 64-bit mode.
> > > > > > 
> > > > > > Hmmm....
> > > > > > For performance reasons you really don't want any of them.
> > > > > > The misaligned 64bit fields need an __attribute((aligned(4)) marker.
> > > > > > 
> > > > > > If the checksum code can do them it really needs to detect
> > > > > > and handle the misalignment.
> > > > > > 
> > > > > > The misaligned trap handler probably ought to contain a
> > > > > > warn_on_once() to dump stack on the first such error.
> > > > > > They can then be fixed one at a time.
> > > > > > 
> > > > > 
> > > > > Unaligned LDD at unwind_once+0x4a8/0x5e0
> > > > > 
> > > > > Decoded:
> > > > > 
> > > > > Unaligned LDD at unwind_once (arch/parisc/kernel/unwind.c:212 arch/parisc/kernel/unwind.c:243 arch/parisc/kernel/unwind.c:371 arch/parisc/kernel/unwind.c:445)
> > > > > 
> > > > > Source:
> > > > > 
> > > > > static bool pc_is_kernel_fn(unsigned long pc, void *fn)
> > > > > {
> > > > >         return (unsigned long)dereference_kernel_function_descriptor(fn) == pc;
> > > > This looks wrong to me.  Function descriptors should always be 8-byte aligned.  I think this
> > > > routine should return false if fn isn't 8-byte aligned.
> > > 
> > > Below you state "Code entry points only need 4-byte alignment."
> > > 
> > > I think that contradicts each other. Also, the calling code is,
> > > for example,
> > >     pc_is_kernel_fn(pc, syscall_exit)
> > > 
> > > I fail to see how this can be consolidated if it is ok
> > > that syscall_exit is 4-byte aligned but, at the same time,
> > > must be 8-byte aligned to be considered to be a kernel function.
> > In the above call, syscall_exit is treated as a function pointer. It points to an 8-byte aligned
> > function descriptor.  The descriptor holds the actual address of the function.  It only needs
> > 4-byte alignment.
> > 
> > Descriptors need 8-byte alignment for efficiency on 64-bit parisc. The pc and gp are accessed
> > using ldd instructions.
> > 
> 
> How about the patch below ?
> 
> Guenter
> 
> ---
> diff --git a/arch/parisc/kernel/unwind.c b/arch/parisc/kernel/unwind.c
> index 27ae40a443b8..c2b9e23cbc0a 100644
> --- a/arch/parisc/kernel/unwind.c
> +++ b/arch/parisc/kernel/unwind.c
> @@ -214,24 +214,14 @@ static bool pc_is_kernel_fn(unsigned long pc, void *fn)
> 
>  static int unwind_special(struct unwind_frame_info *info, unsigned long pc, int frame_size)
>  {
> -       /*
> -        * We have to use void * instead of a function pointer, because
> -        * function pointers aren't a pointer to the function on 64-bit.
> -        * Make them const so the compiler knows they live in .text
> -        * Note: We could use dereference_kernel_function_descriptor()
> -        * instead but we want to keep it simple here.
> -        */
> -       extern void * const ret_from_kernel_thread;
> -       extern void * const syscall_exit;
> -       extern void * const intr_return;
> -       extern void * const _switch_to_ret;
> +       void (*ret_from_kernel_thread)(void);
> +       void (*syscall_exit)(void);
> +       void (*intr_return)(void);
> +       void (*_switch_to_ret)(void);
>  #ifdef CONFIG_IRQSTACKS
> -       extern void * const _call_on_stack;
> +       void (*_call_on_stack)(void);
>  #endif /* CONFIG_IRQSTACKS */
> -       void *ptr;
> -
> -       ptr = dereference_kernel_function_descriptor(&handle_interruption);
> -       if (pc_is_kernel_fn(pc, ptr)) {
> +       if (pc_is_kernel_fn(pc, handle_interruption)) {
>                 struct pt_regs *regs = (struct pt_regs *)(info->sp - frame_size - PT_SZ_ALGN);
>                 dbg("Unwinding through handle_interruption()\n");
>                 info->prev_sp = regs->gr[30];
> 

Seems like a promising direction.

It feels like we have hit a point when we should "close" this thread and
start potentially a couple new ones to correct the behavior of
saving/restoring the PSW and this behavior with unwind.

I don't know what the proper etiquitte is for reverting back to a
previous patch, should I send a v9 that is just the same as the v7?

Thank you Guenter and John for looking into the parisc behavior!

- Charlie