[prev in list] [next in list] [prev in thread] [next in thread] 

List:       glibc-alpha
Subject:    Re: [PATCH] Memory fencing problem in pthread cancellation
From:       Jeff Law <law () redhat ! com>
Date:       2013-01-31 22:07:49
Message-ID: 510AEB35.6090909 () redhat ! com
[Download RAW message or body]

On 01/14/2013 03:58 PM, Carlos O'Donell wrote:
> You are correct, we need another entry for libgcc_s_resume, and we need
> to load that after the barrier, otherwise we won't call libgcc_s_resume.
> 
> Something like this:
> 
> diff --git a/ports/sysdeps/unix/sysv/linux/arm/nptl/unwind-forcedunwind.c \
> b/ports/sysdeps/unix/sysv/linux/arm/nptl/unwind-forcedunwind.c index \
>                 58ca9ac..1e03b13 100644
> --- a/ports/sysdeps/unix/sysv/linux/arm/nptl/unwind-forcedunwind.c
> +++ b/ports/sysdeps/unix/sysv/linux/arm/nptl/unwind-forcedunwind.c
> @@ -104,7 +104,12 @@ asm (
> "      mov     r6, r0\n"
> "      cmp     r3, #0\n"
> "      beq     4f\n"
> -"5:    mov     r0, r6\n"
> +       atomic_asm_full_barrier()
> +"5:    ldr     r4, 6f\n"
> +"      ldr     r5, 7f\n"
> +"8:    add     r4, pc, r4\n"
> +"      ldr     r3, [r4, r5]\n"
> +"      mov     r0, r6\n"
> "      ldmfd   sp!, {r4, r5, r6, lr}\n"
> "      " CFI_ADJUST_CFA_OFFSET (-16) "\n"
> "      " CFI_RESTORE (r4) "\n"
> @@ -123,7 +128,13 @@ asm (
> #else
> "1:    .word   _GLOBAL_OFFSET_TABLE_ - 3b - 8\n"
> #endif
> -"2:    .word   libgcc_s_resume(GOTOFF)\n"
> +"2:    .word   libgcc_s_handle(GOTOFF)\n"
> +#ifdef __thumb2__
> +"6:    .word   _GLOBAL_OFFSET_TABLE_ - 8b - 4\n"
> +#else
> +"6:    .word   _GLOBAL_OFFSET_TABLE_ - 8b - 8\n"
> +#endif
> +"7:    .word   libgcc_s_resume(GOTOFF)\n"
> "      .size   _Unwind_Resume, .-_Unwind_Resume\n"
> );
Added.  Likewise for the other changes.  With one tweak:

[ ... ]

> @@ -76,7 +92,7 @@ asm (
> "      " CFI_RESTORE (lr) "\n"
> "      bx      r3\n"
> "      " CFI_RESTORE_STATE "\n"
> -"4:    bl      init\n"
> +"4:    bl      pthread_cancel_init\n"
> "      ldr     r3, [r4, r5]\n"
> "      b       5b\n"
> "      " CFI_ENDPROC "\n"
I got the impression via a later message from Joseph that this code was 
supposed to call init rather than pthread_cancel_init.  So I didn't 
apply that hunk.

Updated patch with all the updates/fixes to be posted tomorrow AM.

Jeff


[prev in list] [next in list] [prev in thread] [next in thread]