Skip to contentSkip to conversation

IA32 process segfaults on exit

Linux-ia64 Michael Madore 2 participants 9 messages Apr 20, 2001 – May 7, 2001

lore
Share & tools
lore
#0 lore raw

Hi,

The attached 32 bit code will run without error on IA32, but the same binary will generate the following errors when it returns on IA64.

sizeof(elf_gregset_t) (1024) != sizeof(struct pt_regs) (400) Segmentation fault (core dumped)

Is this expected behavior?

Signature · 3 lines
Mike Madore
Software Engineer
TurboLinux, Inc.
#1 lore raw

Mike-

How very serendipitous :-)

This is not supposed to happen but I think I'm working on the solution. Some code to initialize EDX on process startup (EDX points to a potential atexit routine) got deleted. I put the code back but it's not working properly so there's something else wrong. Stay tuned.

Quoted 19 lines · Michael Madore

Hi,

The attached 32 bit code will run without error on IA32, but the same binary will generate the following errors when it returns on IA64.

sizeof(elf_gregset_t) (1024) != sizeof(struct pt_regs) (400) Segmentation fault (core dumped)

Is this expected behavior?

Signature · 8 lines
Mike Madore
Software Engineer
TurboLinux, Inc.

_______________________________________________
Linux-IA64 mailing list
[email protected]
http://lists.linuxia64.org/lists/listinfo/linux-ia64
Signature · 4 lines
Don Dugger
"Censeo Toto nos in Kansa esse decisse." - D. Gale
[email protected]
Ph: 303/938-9838
#2 lore raw
In reply to Michael Madore

Mike-

Curious. I used `gdb' to clear out the registers and you program is still getting a segv. I'll look into it.

Quoted 40 lines · Don Dugger

Mike-

How very serendipitous :-)

This is not supposed to happen but I think I'm working on the solution. Some code to initialize EDX on process startup (EDX points to a potential atexit routine) got deleted. I put the code back but it's not working properly so there's something else wrong. Stay tuned.

Quoted 19 lines · Michael Madore

Hi,

The attached 32 bit code will run without error on IA32, but the same binary will generate the following errors when it returns on IA64.

sizeof(elf_gregset_t) (1024) != sizeof(struct pt_regs) (400) Segmentation fault (core dumped)

Is this expected behavior?

Signature · 8 lines
Mike Madore
Software Engineer
TurboLinux, Inc.

_______________________________________________
Linux-IA64 mailing list
[email protected]
http://lists.linuxia64.org/lists/listinfo/linux-ia64
Signature · 9 lines
Don Dugger
"Censeo Toto nos in Kansa esse decisse." - D. Gale
[email protected]
Ph: 303/938-9838

_______________________________________________
Linux-IA64 mailing list
[email protected]
http://lists.linuxia64.org/lists/listinfo/linux-ia64
Signature · 4 lines
Don Dugger
"Censeo Toto nos in Kansa esse decisse." - D. Gale
[email protected]
Ph: 303/938-9838
#3 lore raw
In reply to Michael Madore

Mike-

Well, I'll very confused. I thought your problem was some registers that we stopped initializing that the LIBC startup code expected to point to an exit handler. I spent over a week debugging my initialization code (the register was always non-zero in user mode) only to disciver today that when I link the process static everything works fine, including your test program. When I try and run a dynamic program the initial registers are bogus. (Did I ever mention that I HATE dynamic linking.)

Bottom line is I still don't know what the problem is but I think I'm making progress.

Quoted 19 lines · Michael Madore

Hi,

The attached 32 bit code will run without error on IA32, but the same binary will generate the following errors when it returns on IA64.

sizeof(elf_gregset_t) (1024) != sizeof(struct pt_regs) (400) Segmentation fault (core dumped)

Is this expected behavior?

Signature · 8 lines
Mike Madore
Software Engineer
TurboLinux, Inc.

_______________________________________________
Linux-IA64 mailing list
[email protected]
http://lists.linuxia64.org/lists/listinfo/linux-ia64
Signature · 4 lines
Don Dugger
"Censeo Toto nos in Kansa esse decisse." - D. Gale
[email protected]
Ph: 303/938-9838
#4 lore raw
In reply to Michael Madore

Hi Don,

Thanks for looking into it. It's not a critical problem for us at the moment. Our IA32 installer was invoking some BIOS code to get geometry information for the hard disk. We replaced that with basically the same code that fdisk uses.

So you can just work on this for fun now. ;-)

Quoted 39 lines · Don Dugger

Mike-

Well, I'll very confused. I thought your problem was some registers that we stopped initializing that the LIBC startup code expected to point to an exit handler. I spent over a week debugging my initialization code (the register was always non-zero in user mode) only to disciver today that when I link the process static everything works fine, including your test program. When I try and run a dynamic program the initial registers are bogus. (Did I ever mention that I HATE dynamic linking.)

Bottom line is I still don't know what the problem is but I think I'm making progress.

Quoted 19 lines · Michael Madore

Hi,

The attached 32 bit code will run without error on IA32, but the same binary will generate the following errors when it returns on IA64.

sizeof(elf_gregset_t) (1024) != sizeof(struct pt_regs) (400) Segmentation fault (core dumped)

Is this expected behavior?

Signature · 8 lines
Mike Madore
Software Engineer
TurboLinux, Inc.

_______________________________________________
Linux-IA64 mailing list
[email protected]
http://lists.linuxia64.org/lists/listinfo/linux-ia64
Signature · 4 lines
Don Dugger
"Censeo Toto nos in Kansa esse decisse." - D. Gale
[email protected]
Ph: 303/938-9838
Signature · 4 lines
Mike Madore
Senior Software Engineer
TurboLinux, Inc.
(650)228-5203
#5 lore raw
In reply to Michael Madore

Michael-

Um, I think I found your problem and I'm pretty sure it's program error. The test case you sent effectively does:

	fd_zero = open("/dev/zero", O_RDONLY);
	m = mmap(0x10000, 0x10000, ..., fd_zero, 0);

This mapps zeros to loation 0x10000. What makes you think that the address 0x10000 is free? On an IA32 machine the Run Time LoaDer is loaded at 0x2aaab000 so this location, strictly by chance, is free. On Trillian the RTLD is loaded at 0x4000 (actually, it's the first page boundary) so you've mapped zero's on top of RTLD.

I'm really curious, why do you think 0x10000 is free?

Quoted 19 lines · Michael Madore

Hi,

The attached 32 bit code will run without error on IA32, but the same binary will generate the following errors when it returns on IA64.

sizeof(elf_gregset_t) (1024) != sizeof(struct pt_regs) (400) Segmentation fault (core dumped)

Is this expected behavior?

Signature · 8 lines
Mike Madore
Software Engineer
TurboLinux, Inc.

_______________________________________________
Linux-IA64 mailing list
[email protected]
http://lists.linuxia64.org/lists/listinfo/linux-ia64
Signature · 4 lines
Don Dugger
"Censeo Toto nos in Kansa esse decisse." - D. Gale
[email protected]
Ph: 303/938-9838
#6 lore raw
In reply to Michael Madore

Hi Don,

I'm not intimately familiar with what the author (not me, really!) was trying to accomplish. I believe that the code belongs to part of the IA32 installer that is making BIOS calls to get drive geometry. I guess since it didn't outright break on IA32, nobody questioned whether it was actually correct or not.

If you only saw the RPM handling code that we just ported... ugh! Integer -> pointer -> integer conversions all over the place.

Anyway, as I said, we removed this code and replaced it with something that doesn't suck as much. ;-)

Quoted 42 lines · Don Dugger

Michael-

Um, I think I found your problem and I'm pretty sure it's program error. The test case you sent effectively does:

	fd_zero = open("/dev/zero", O_RDONLY);
	m = mmap(0x10000, 0x10000, ..., fd_zero, 0);

This mapps zeros to loation 0x10000. What makes you think that the address 0x10000 is free? On an IA32 machine the Run Time LoaDer is loaded at 0x2aaab000 so this location, strictly by chance, is free. On Trillian the RTLD is loaded at 0x4000 (actually, it's the first page boundary) so you've mapped zero's on top of RTLD.

I'm really curious, why do you think 0x10000 is free?

Quoted 19 lines · Michael Madore

Hi,

The attached 32 bit code will run without error on IA32, but the same binary will generate the following errors when it returns on IA64.

sizeof(elf_gregset_t) (1024) != sizeof(struct pt_regs) (400) Segmentation fault (core dumped)

Is this expected behavior?

Signature · 8 lines
Mike Madore
Software Engineer
TurboLinux, Inc.

_______________________________________________
Linux-IA64 mailing list
[email protected]
http://lists.linuxia64.org/lists/listinfo/linux-ia64
Signature · 4 lines
Don Dugger
"Censeo Toto nos in Kansa esse decisse." - D. Gale
[email protected]
Ph: 303/938-9838
Signature · 4 lines
Mike Madore
Senior Software Engineer
TurboLinux, Inc.
(650)228-5203
#7 lore raw
In reply to Michael Madore

Michael-

Well, unless anyone else can prove different, I'm just going to chalk this one up as application error and declare the kernel innocent.

PS: Yes, I feel your RPM pain. It's one of those programs that, when it works, it's great but, when it doesn't work ...

Quoted 63 lines · Michael Madore

Hi Don,

I'm not intimately familiar with what the author (not me, really!) was trying to accomplish. I believe that the code belongs to part of the IA32 installer that is making BIOS calls to get drive geometry. I guess since it didn't outright break on IA32, nobody questioned whether it was actually correct or not.

If you only saw the RPM handling code that we just ported... ugh! Integer -> pointer -> integer conversions all over the place.

Anyway, as I said, we removed this code and replaced it with something that doesn't suck as much. ;-)

Quoted 42 lines · Don Dugger

Michael-

Um, I think I found your problem and I'm pretty sure it's program error. The test case you sent effectively does:

	fd_zero = open("/dev/zero", O_RDONLY);
	m = mmap(0x10000, 0x10000, ..., fd_zero, 0);

This mapps zeros to loation 0x10000. What makes you think that the address 0x10000 is free? On an IA32 machine the Run Time LoaDer is loaded at 0x2aaab000 so this location, strictly by chance, is free. On Trillian the RTLD is loaded at 0x4000 (actually, it's the first page boundary) so you've mapped zero's on top of RTLD.

I'm really curious, why do you think 0x10000 is free?

Quoted 19 lines · Michael Madore

Hi,

The attached 32 bit code will run without error on IA32, but the same binary will generate the following errors when it returns on IA64.

sizeof(elf_gregset_t) (1024) != sizeof(struct pt_regs) (400) Segmentation fault (core dumped)

Is this expected behavior?

Signature · 8 lines
Mike Madore
Software Engineer
TurboLinux, Inc.

_______________________________________________
Linux-IA64 mailing list
[email protected]
http://lists.linuxia64.org/lists/listinfo/linux-ia64
Signature · 4 lines
Don Dugger
"Censeo Toto nos in Kansa esse decisse." - D. Gale
[email protected]
Ph: 303/938-9838
Signature · 4 lines
Mike Madore
Senior Software Engineer
TurboLinux, Inc.
(650)228-5203
Signature · 4 lines
Don Dugger
"Censeo Toto nos in Kansa esse decisse." - D. Gale
[email protected]
Ph: 303/938-9838
#8 lore raw
In reply to Michael Madore
Quoted 4 lines · Don Dugger

Michael-

Well, unless anyone else can prove different, I'm just going to chalk this one up as application error and declare the kernel innocent.

Sounds good. Now about real player... ;-)

Signature · 4 lines
Mike Madore
Senior Software Engineer
TurboLinux, Inc.
(650)228-5203