Skip to contentSkip to conversation

Current utrace breaks UML

DISCUSSION Jeff Dike 2 participants 6 messages May 22, 2007 – May 26, 2007

lore
Share & tools
lore

Jeff Dike

author
#0 lore raw

This chunk from linux-2.6-utrace.patch breaks PTRACE_SYSEMU, which UML rather relies on.

@@ -514,9 +514,6 @@ syscall_trace_entry:
movl %esp, %eax
xorl %edx,%edx
call do_syscall_trace
- cmpl $0, %eax
- jne resume_userspace # ret != 0 -> running under PTRACE_SYSEMU,
- # so must skip actual syscall
movl PT_ORIG_EAX(%esp), %eax
cmpl $(nr_syscalls), %eax
jnae syscall_call

If the point of this patch was to completely remove ptrace support, with later patches adding it back, then that chunk should be added back in utrace-ptrace-compat.patch.

Jeff

Signature · 1 line
Work email - jdike at linux dot intel dot com
#1 lore raw
Quoted 2 lines

This chunk from linux-2.6-utrace.patch breaks PTRACE_SYSEMU, which UML rather relies on.

PTRACE_SYSEMU and PTRACE_SYSEMU_SINGLESTEP are implemented in a different way (see kernel/ptrace.c), which does not require that assembly code. (It also works for free on other arch's if you want to #define the constants there.)

Do you have a test case for PTRACE_SYSEMU that does not work right?

Thanks,
Roland

author
#2 lore raw
Quoted 1 line · Roland McGrath

Do you have a test case for PTRACE_SYSEMU that does not work right?

UML, obviously. Below is a smaller test. orig_eax is wrong, so you can't read the system call number from the process.

With kernel-2.6.20-1.2948, it prints out a bunch of -1's. On FC5 kernels, you get more reasonable numbers.

Jeff

Signature · 80 lines
Work email - jdike at linux dot intel dot com


#include <stdlib.h>
#include <stdio.h>
#include <asm/unistd.h>
#include <sys/mman.h>
#include <sys/ptrace.h>
#include <asm/ptrace.h>
#include <sys/wait.h>

#define PTRACE_SYSEMU		  31
#define PAGE_SIZE 4096

static inline long stub_syscall0(long syscall)
{
	long ret;

	__asm__ volatile ("int $0x80" : "=a" (ret) : "0" (syscall));

	return ret;
}

static void child(void)
{
	if(ptrace(PTRACE_TRACEME, 0, 0, 0) < 0){
		perror("traceme");
		exit(1);
	}

	kill(getpid(), SIGUSR1);
	stub_syscall0(__NR_getpid);
}

int main(void)
{
	void *stack;
	unsigned long sp, regs[FRAME_SIZE];
	int pid, n, status;

	stack = mmap(NULL, PAGE_SIZE, PROT_READ | PROT_WRITE | PROT_EXEC,
		     MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
	if(stack == MAP_FAILED){
		perror("mmap");
		exit(1);
	}

	sp = (unsigned long) stack + PAGE_SIZE - sizeof(void *);
	pid = fork();
	if(pid < 0)
		perror("fork");
	else if(pid == 0){
		child();
	}
	else {
		while(1){
			n = waitpid(pid, &status, WUNTRACED);
			if(n < 0){
				perror("waitpid");
				exit(1);
			}

			n = ptrace(PTRACE_GETREGS, pid, 0, regs);
			if(n < 0){
				perror("PTRACE_GETREGS");
				exit(1);
			}

			printf("Status 0x%x orig_eax 0x%x\n", status,
			       regs[ORIG_EAX]);

			if(ptrace(PTRACE_SYSEMU, pid, 0, 0)){
				perror("PTRACE_SYSEMU");
				exit(1);
			}
		}
	}

	return 0;
}
#3 lore raw
Quoted 2 lines

UML, obviously. Below is a smaller test. orig_eax is wrong, so you can't read the system call number from the process.

Oops! I overlooked the need to preserve the orig_eax value, though its necessity is obvious. This makes me wonder about those previous reports that UML was working OK.

I refined the test case a little and that's in
https://bugzilla.redhat.com/bugzilla/show_bug.cgi?id=241408

I've fixed this in the latest utrace patch set. I also wired up sysemu on x86_64. This means 32-bit processes calling ptrace now support it for full compatibility with native i386, which the vanilla kernel does not. It also means it works for 64-bit ptrace calls, whether operating on a 64-bit or a 32-bit target process.

Thanks,
Roland

author
#4 lore raw
Quoted 3 lines · Roland McGrath

Oops! I overlooked the need to preserve the orig_eax value, though its necessity is obvious. This makes me wonder about those previous reports that UML was working OK.

The one from me was on x86_64, where PTRACE_SYSEMU isn't an issue yet.

Quoted 5 lines

I've fixed this in the latest utrace patch set. I also wired up sysemu on x86_64. This means 32-bit processes calling ptrace now support it for full compatibility with native i386, which the vanilla kernel does not. It also means it works for 64-bit ptrace calls, whether operating on a 64-bit or a 32-bit target process.

Beautiful!

Jeff

Signature · 1 line
Work email - jdike at linux dot intel dot com
author
#5 lore raw
In reply to Roland McGrath
Quoted 2 lines · Roland McGrath

(It also works for free on other arch's if you want to #define the constants there.)

(Forgot to mention...) sweet

Jeff

Signature · 1 line
Work email - jdike at linux dot intel dot com