Skip to contentSkip to conversation

heads up: apparent GAS bug in current (CVS) binutils

DISCUSSION David Mosberger 4 participants 11 messages May 12, 2005 – May 17, 2005

Subsystems (5)
lore
Share & tools
lore
#0 lore raw

Just a quick heads up (it's late here and I'll be on travel tomorrow so I don't have time for a proper investigation/bug-report at the moment):

If I build the kernel with the assembler from the current CVS binutils, I see these errors:

ERROR: proc_set_irq_affinity: 179742 slots, total region length = 36
ERROR: idle_regs: 68847 slots, total region length = 15

The error in proc_set_irq_affinity() is definitely real and a result of the assembler (the assembly code generated by GCC looks fine). I didn't check on idle_regs.

readelf -u vmlinux shows this:

<proc_set_irq_affinity>: [0xa00000010000ef60-0xa0000001000f9000], info at +0x765 c50
v1, flags=0x0 (), len=8 bytes

    R2:prologue_gr(mask=[rp,ar.pfs],grsave=r34,rlen=8)
        P7:pfs_when(t=0)
        P7:rp_when(t=7)
    R1:body(rlen()

<move_irq>: [0xa00000010000f0c0-0xa00000010000f250], info at +0x75da00

Note how the end-address of proc_set_irq_affinity() is _way_ beyond the start-address of the next function, move_irq. Clearly something weird happened here.

If I use the standard Debian/sarge assembler, these errors go away and the unwind-info looks sane.

	--david
#1 lore raw

I tried binutils 2.15.92.0.2 20040927 and got the same result. What is the version of assembler which you think is good?

H.J. Lu
Intel Corporation

Quoted 42 lines

-----Original Message-----
From: David Mosberger [mailto:[email protected]]
Sent: Thursday, May 12, 2005 2:56 PM
To: [email protected]
Cc: Lu, Hongjiu
Subject: heads up: apparent GAS bug in current (CVS) binutils

Just a quick heads up (it's late here and I'll be on travel tomorrow so I don't have time for a proper investigation/bug-report at the moment):

If I build the kernel with the assembler from the current CVS binutils, I see these errors:

ERROR: proc_set_irq_affinity: 179742 slots, total region length = 36
ERROR: idle_regs: 68847 slots, total region length = 15

The error in proc_set_irq_affinity() is definitely real and a result of the assembler (the assembly code generated by GCC looks fine). I didn't check on idle_regs.

readelf -u vmlinux shows this:

<proc_set_irq_affinity>: [0xa00000010000ef60-0xa0000001000f9000], info at +0x765
c50
v1, flags=0x0 (), len=8 bytes
R2:prologue_gr(mask=[rp,ar.pfs],grsave=r34,rlen=8)

       P7:pfs_when(t=0)
       P7:rp_when(t=7)

R1:body(rlen()

<move_irq>: [0xa00000010000f0c0-0xa00000010000f250], info at +0x75da00

Note how the end-address of proc_set_irq_affinity() is _way_ beyond the start-address of the next function, move_irq. Clearly something weird happened here.

If I use the standard Debian/sarge assembler, these errors go away and the unwind-info looks sane.

	--david
#2 lore raw
In reply to David Mosberger

Hi David,

It is a linker issue. There are 2 definitions for those functions, the generic is weak and the ia64 one is strong. The generic one has unwind info and the ia64 one doesn't. When linker combines them, it keeps the strong definition, but uses the weak unwind info. I will see what I can do about it.

H.J. Lu
Intel Corporation

Quoted 42 lines

-----Original Message-----
From: David Mosberger [mailto:[email protected]]
Sent: Thursday, May 12, 2005 2:56 PM
To: [email protected]
Cc: Lu, Hongjiu
Subject: heads up: apparent GAS bug in current (CVS) binutils

Just a quick heads up (it's late here and I'll be on travel tomorrow so I don't have time for a proper investigation/bug-report at the moment):

If I build the kernel with the assembler from the current CVS binutils, I see these errors:

ERROR: proc_set_irq_affinity: 179742 slots, total region length = 36
ERROR: idle_regs: 68847 slots, total region length = 15

The error in proc_set_irq_affinity() is definitely real and a result of the assembler (the assembly code generated by GCC looks fine). I didn't check on idle_regs.

readelf -u vmlinux shows this:

<proc_set_irq_affinity>: [0xa00000010000ef60-0xa0000001000f9000], info at +0x765
c50
v1, flags=0x0 (), len=8 bytes
R2:prologue_gr(mask=[rp,ar.pfs],grsave=r34,rlen=8)

       P7:pfs_when(t=0)
       P7:rp_when(t=7)

R1:body(rlen()

<move_irq>: [0xa00000010000f0c0-0xa00000010000f250], info at +0x75da00

Note how the end-address of proc_set_irq_affinity() is _way_ beyond the start-address of the next function, move_irq. Clearly something weird happened here.

If I use the standard Debian/sarge assembler, these errors go away and the unwind-info looks sane.

	--david
#3 lore raw
In reply to David Mosberger

Hi David,

I think it is the result of the usage of weak function. I am enclosing a testcase to show it. The correct approach is to use comdat section group or linkonce section. The problem will go away. I will see if I can come up with a kernel patch.

H.J. Lu
Intel Corporation

Quoted 42 lines

-----Original Message-----
From: David Mosberger [mailto:[email protected]]
Sent: Thursday, May 12, 2005 2:56 PM
To: [email protected]
Cc: Lu, Hongjiu
Subject: heads up: apparent GAS bug in current (CVS) binutils

Just a quick heads up (it's late here and I'll be on travel tomorrow so I don't have time for a proper investigation/bug-report at the moment):

If I build the kernel with the assembler from the current CVS binutils, I see these errors:

ERROR: proc_set_irq_affinity: 179742 slots, total region length = 36
ERROR: idle_regs: 68847 slots, total region length = 15

The error in proc_set_irq_affinity() is definitely real and a result of the assembler (the assembly code generated by GCC looks fine). I didn't check on idle_regs.

readelf -u vmlinux shows this:

<proc_set_irq_affinity>: [0xa00000010000ef60-0xa0000001000f9000], info at +0x765
c50
v1, flags=0x0 (), len=8 bytes
R2:prologue_gr(mask=[rp,ar.pfs],grsave=r34,rlen=8)

       P7:pfs_when(t=0)
       P7:rp_when(t=7)

R1:body(rlen=28)

<move_irq>: [0xa00000010000f0c0-0xa00000010000f250], info at +0x75da00

Note how the end-address of proc_set_irq_affinity() is _way_ beyond the start-address of the next function, move_irq. Clearly something weird happened here.

If I use the standard Debian/sarge assembler, these errors go away and the unwind-info looks sane.

	--david
#4 lore raw
In reply to David Mosberger
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Hongjiu

HJ> I tried binutils 2.15.92.0.2 20040927 and got the same HJ> result. What is the version of assembler which you think is HJ> good?

The standard Debian/sarge as seems fine:

 $ as -v
 GNU assembler version 2.15 (ia64-linux) using BFD version 2.15
	--david
#5 lore raw
In reply to David Mosberger
Quoted 11 lines · David Mosberger
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Hongjiu

HJ> I tried binutils 2.15.92.0.2 20040927 and got the same HJ> result. What is the version of assembler which you think is HJ> good?

The standard Debian/sarge as seems fine:

 $ as -v
 GNU assembler version 2.15 (ia64-linux) using BFD version 2.15

As I have mentioned before, it is an IPF specific kernel issue. Basically, when you use weak functions, unwind info may be wrong. I am enclosing an updated kernel patch.

H.J.

arch/ia64/kernel/irq.c renamed from arch/ia64/kernel/irq.c.weak +3 -1
@@ -103,7 +103,9 @@ static char irq_redir [NR_IRQS]; // = { 
  * Arch specific routine for deferred write to iosapic rte to reprogram
  * intr destination.
  */
-void proc_set_irq_affinity(unsigned int irq, cpumask_t mask_val)
+void
+__comdatfunc (proc_set_irq_affinity)
+proc_set_irq_affinity(unsigned int irq, cpumask_t mask_val)
 {
 	pending_irq_cpumask[irq] = mask_val;
 }
arch/ia64/kernel/smpboot.c renamed from arch/ia64/kernel/smpboot.c.weak +3 -1
@@ -404,7 +404,9 @@ start_secondary (void *unused)
 	return 0;
 }
 
-struct pt_regs * __devinit idle_regs(struct pt_regs *regs)
+struct pt_regs *
+__devinit __comdatfunc (idle_regs)
+idle_regs(struct pt_regs *regs)
 {
 	return NULL;
 }
include/linux/compiler.h renamed from include/linux/compiler.h.weak +5 -0
@@ -155,4 +155,9 @@ extern void __chk_io_ptr(void __iomem *)
 #define __always_inline inline
 #endif
 
+#ifndef __comdatfunc
+#define __comdatfunc(f) \
+  __attribute__ ((__section__ (".gnu.linkonce.t." #f )))
+#endif
+
 #endif /* __LINUX_COMPILER_H */
kernel/fork.c renamed from kernel/fork.c.weak +3 -1
@@ -1148,7 +1148,9 @@ bad_fork_free:
 	goto fork_out;
 }
 
-struct pt_regs * __devinit __attribute__((weak)) idle_regs(struct pt_regs *regs)
+struct pt_regs *
+__devinit __comdatfunc (idle_regs)
+idle_regs(struct pt_regs *regs)
 {
 	memset(regs, 0, sizeof(struct pt_regs));
 	return regs;
kernel/irq/proc.c renamed from kernel/irq/proc.c.weak +2 -1
@@ -19,7 +19,8 @@ static struct proc_dir_entry *root_irq_d
  */
 static struct proc_dir_entry *smp_affinity_entry[NR_IRQS];
 
-void __attribute__((weak))
+void
+__comdatfunc (proc_set_irq_affinity)
 proc_set_irq_affinity(unsigned int irq, cpumask_t mask_val)
 {
 	irq_affinity[irq] = mask_val;
#6 lore raw
In reply to David Mosberger
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · H. J. Lu
Quoted 2 lines · David Mosberger
Quoted 2 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

On Fri, 13 May 2005 10:32:19 -0700, "Lu, Hongjiu"

<[email protected]> said:

HJ> I tried binutils 2.15.92.0.2 20040927 and got the same HJ> result. What is the version of assembler which you think is HJ> good?

Quoted 4 lines
Quoted 4 lines

The standard Debian/sarge as seems fine:

$ as -v GNU assembler version 2.15 (ia64-linux) using BFD version
2.15

HJ> As I have mentioned before, it is an IPF specific kernel issue. HJ> Basically, when you use weak functions, unwind info may be HJ> wrong. I am enclosing an updated kernel patch.

I agree that this is weird code that needs to be fixed, preferably by not using weak functions at all (as Andreas suggested).

However, there still seems to be a binutils issues here: if this is something binutils cannot properly support, it should issue an error, not silently generate wrong code, no?

	--david
#7 lore raw
In reply to David Mosberger
Quoted 18 lines · David Mosberger
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · H. J. Lu
Quoted 2 lines · David Mosberger
Quoted 2 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

On Fri, 13 May 2005 10:32:19 -0700, "Lu, Hongjiu"

<[email protected]> said:

HJ> I tried binutils 2.15.92.0.2 20040927 and got the same HJ> result. What is the version of assembler which you think is HJ> good?

Quoted 4 lines
Quoted 4 lines

The standard Debian/sarge as seems fine:

$ as -v GNU assembler version 2.15 (ia64-linux) using BFD version
2.15

HJ> As I have mentioned before, it is an IPF specific kernel issue. HJ> Basically, when you use weak functions, unwind info may be HJ> wrong. I am enclosing an updated kernel patch.

Quoted 6 lines

I agree that this is weird code that needs to be fixed, preferably by not using weak functions at all (as Andreas suggested).

However, there still seems to be a binutils issues here: if this is something binutils cannot properly support, it should issue an error, not silently generate wrong code, no?

I will see what I can do.

H.J.

#8 lore raw
In reply to David Mosberger
Quoted 5 lines
Quoted 5 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line · H. J. Lu

However, there still seems to be a binutils issues here: if this is something binutils cannot properly support, it should issue an error, not silently generate wrong code, no?

HJ> I will see what I can do.

BTW: Thanks for tracking this down! I had no idea this weird "overloading" was going on.

	--david
#9 lore raw
In reply to David Mosberger

Review

Quoted 13 lines

+++ linux/arch/ia64/kernel/smpboot.c 2005-05-13

(quoted diff) +4 -2
-struct pt_regs * __devinit idle_regs(struct pt_regs *regs)
+struct pt_regs *
+__devinit __comdatfunc (idle_regs)
+idle_regs(struct pt_regs *regs)
 
+++ linux/kernel/fork.c	2005-05-13 13:19:40.000000000 -0700
-struct pt_regs * __devinit __attribute__((weak)) 

idle_regs(struct pt_regs *regs)

(quoted diff) +3 -0
+struct pt_regs *
+__devinit __comdatfunc (idle_regs)
+idle_regs(struct pt_regs *regs)

So is there some deeper magic that I don't understand, or is this wrong? After this patch is applied, both of these definitions look identical to me. So how does the linker know which one to use?

-Tony

#10 lore raw
In reply to David Mosberger
Quoted 8 lines
Quoted 7 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line · Tony

+++ linux/arch/ia64/kernel/smpboot.c 2005-05-13

(quoted diff) +3 -1
-struct pt_regs * __devinit idle_regs(struct pt_regs *regs)
+struct pt_regs *
+__devinit __comdatfunc (idle_regs)
+idle_regs(struct pt_regs *regs)
Quoted 6 lines

Review

Quoted 6 lines

+++ linux/kernel/fork.c 2005-05-13 13:19:40.000000000 -0700
-struct pt_regs * __devinit __attribute__((weak))
idle_regs(struct pt_regs *regs)

(quoted diff) +3 -0
+struct pt_regs *
+__devinit __comdatfunc (idle_regs)
+idle_regs(struct pt_regs *regs)

Tony> So is there some deeper magic that I don't understand, or is this Tony> wrong? After this patch is applied, both of these definitions look Tony> identical to me. So how does the linker know which one to use?

Please do not apply this patch. The bug is in the toolchain. (The kernel may be ugly in this corner and we may want to fix that by getting rid of the weak definitions but that issue is entirely independent of the toolchain bug.)

	--david