Re: "kernel ade access" oops on LoongArch

[Xi Ruoyao]

On Wed, 2023-02-15 at 20:51 +0800, Xi Ruoyao wrote:
> On Wed, 2023-02-15 at 19:52 +0800, Xi Ruoyao wrote:
> > On Wed, 2023-02-15 at 16:35 +0800, Xi Ruoyao wrote:
> > > On Wed, 2023-02-15 at 16:25 +0800, Youling Tang wrote:
> > > > Can you modify the kernel as follows and test it, so as to avoid
> > > > possible relationship with the exception table data link
> > > > position
> > > > and
> > > > alignment rules (or use EXCEPTION_TABLE(12))?
> > > > 
> > > > --- a/arch/loongarch/kernel/vmlinux.lds.S
> > > > +++ b/arch/loongarch/kernel/vmlinux.lds.S
> > > > @@ -4,7 +4,6 @@
> > > >   #include <asm/thread_info.h>
> > > > 
> > > >   #define PAGE_SIZE _PAGE_SIZE
> > > > -#define RO_EXCEPTION_TABLE_ALIGN       4
> > > > 
> > > >   /*
> > > >    * Put .bss..swapper_pg_dir as the first thing in .bss. This
> > > > will
> > > > @@ -54,6 +53,8 @@ SECTIONS
> > > >          . = ALIGN(PECOFF_SEGMENT_ALIGN);
> > > >          _etext = .;
> > > > 
> > > > +       EXCEPTION_TABLE(16)
> > > > +
> > > 
> > > It seems the kernel refuses to boot after the change, but I'm not
> > > completely sure: I'm 5 km away from the board and operating it via
> > > ssh
> > > so maybe it's a reboot failure or network failure.  I'll report
> > > again in
> > > the evening.
> > 
> > It was a reboot failure.
> > 
> > Now it has booted successfully, but the stack trace still shows
> > (during
> > the 25th run of the make test t=... command).
> 
> Ouch, I know what's happening...
> 
> In the architecture-independent code we have something like
> 
> extern struct exception_table_entry a[], b[];
> bsearch(a, b - a);
> 
> According to the C standard, when you write "b - a" where a and b are
> pointers to type T, "b" and "a" must be pointers to elements in the
> same
> array of T.  So the compiler can assume ((uintptr_t)b - (uintptr_t)a)
> %
> 12 == 0 and optimize "b - a" to something like
> 
> (((uintptr_t)b - (uintptr_t)a) >> 2) * inv3
> 
> Here inv3 is the inversion of 3 in the modulo-2**64 integer ring, so
> the
> compiler can avoid an expensive divide instruction.  But in my vmlinux
> ((uintptr_t)b - (uintptr_t)a) is somehow not a multiple of 12:
> 
> (gdb) p ((uintptr_t)__stop___ex_table - (uintptr_t)__start___ex_table)
> % sizeof(struct exception_table_entry)
> $9 = 8
> 
> So I guess
> 
> #define RO_EXCEPTION_TABLE_ALIGN       12
> 
> will work.  I'll take a try...

No, it's not related...

The reason is the "LoongArch: Add vector extensions support" in my local
repo is not same as the version in loongarch-next!  My local version
contains some ".section __ex_table", and the content seems predates the
relative exception table change.

Sorry for wasted you guys an afternoon :(.

-- 
Xi Ruoyao <xry111@xxxxxxxxxxx>
School of Aerospace Science and Technology, Xidian University