__tls_get_addr problem with recent current

[Kostik Belousov]

On Sat, Aug 30, 2008 at 09:46:05AM -0700, Artem Belevich wrote:
> Hi,
> 
> Did someone figure out what was/is causing this problem? I believe
> I've ran into the same issue on a freshly built (Aug 29th) RELENG_7 on
> amd64 (Q9450 w/ 8GB RAM).
> 
> What's strange is that in my case the problem pops up when I run
> 32-bit binaries from RELENG_6. Otherwise the system is stable. I can
> buildworld continuously with -j8 without any problems. However, when I
> try to build my code using hacked make from FreeBSD6/i386 things start
> falling apart. All of a sudden I get cores from FreeBSD7/amd64
> binaries. All of them, as far as I can tell crash in __tls_get_addr.
> So far I've been unable to reproduce the crashes in any other
> scenarios, but I'm still trying.
> 
> Interestingly enough, crash always occurs on "mov    %fs:0x0,%rdi" instruction,
> yet the same code in the same binaries works fine when binaries are used
> without mixing amd64/i386 binaries. Could that be that amd64 might
> somehow inherit invalid registers from i386 binaries?
Quite possible, almost sure. Could you, please, check whether the
change below would fix it for you ?

diff --git a/sys/amd64/amd64/cpu_switch.S b/sys/amd64/amd64/cpu_switch.S
index f34b0cc..9e13ef4 100644
--- a/sys/amd64/amd64/cpu_switch.S
+++ b/sys/amd64/amd64/cpu_switch.S
@@ -265,6 +265,10 @@ load_seg:
 	movl	PCB_DS(%r8),%ds
 	movl	PCB_ES(%r8),%es
 	movl	PCB_FS(%r8),%fs
+	movl	$MSR_FSBASE,%ecx
+	rdmsr
+	shlq	$32,%rdx
+	leaq	(%rax,%rdx),%r9
 	jmp	done_load_seg
 	/* Restore userland %gs while preserving kernel gsbase */
 2:	movq	PCB_GS32P(%r8),%rax
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 195 bytes
Desc: not available
Url : http://lists.freebsd.org/pipermail/freebsd-current/attachments/20080830/21de45ff/attachment.pgp