Skip to contentSkip to conversation

A correct method to use the x86 breakpoint registers (DR0-7)

DISCUSSION Jeff Carr 2 participants 4 messages Jul 5, 2005 – Jul 7, 2005

lore
Share & tools
lore

Jeff Carr

author
#0 lore raw

I'm trying to set the x86 breakpoint registers to trip on write data.

After they are set, nothing seems to happen when I trigger them.

It's possible I'm not setting them correctly, I tried putting the virt_to_phys() value in them. And, I tried looking at what KPROBE puts in them, but it looks like kprobe doesn't use them at all.

In 2.6.11, arch/i386/ has 5 places where it modifies the db regs:

	do_debug()              in traps.c    (??)
	do_signal()             in signal.c   (re-enable them)
	fix_processor_context() in cpu.c      (reload them)

__switch_to() in process.c (reload them)
cpu_init() in common.c (clears them)

Just FYI: sometime after 2.6.11, the macros get_debugreg() and set_debugreg() were defined in include/asm-i386/processor.h and set to be used in the 5 places above. In any case, the functionality seems the same the above routines. (And the registers names corrected s/db/dr/g )

In any case, setting these registers never seems to do anything. No INT3 or INT1 (is it really supposed to generate an interrupt?) Perhaps I need to have kgdb setup.

The closest I can get to making anything happen is if I set bit 13 of DR7 (triggers on the next access of the breakpoint registers) then when I insmod I get:

root@foxtrot:~/dbregtest# insmod ./dbregtest.ko
Trace/breakpoint trap
root@foxtrot:~/dbregtest#

If I set BR0 with the value from virt_to_phys() I don't get this trace/breakpoint trap.

Enjoy,
Jeff

// Current documentation for these registers is in Vol 3 Section 15.2:
// Also note: EFLAGS BIT 16 (Resume) section 2.3 disables #DB exceptions // http://developer.intel.com/design/pentium4/manuals/index_new.htm

static int __init db_reg_test(void)

{
        u32 *i;
        unsigned int phys_addr;

        i = kmalloc( 0x1000, GFP_DMA );
        printk("i == 0x%08X\n", (int) i);
        phys_addr = virt_to_phys(i);

        printk("virt_to_phys(i)  == 0x%08X\n", (int) phys_addr);
        __asm__ __volatile__( "movl %0, %%dr0\n" : : "r" (phys_addr) );

        // clear out the DR6 status register
        __asm__ __volatile__( "movl %0, %%dr6\n" : : "r" (0xFFFF0FF0) );

        // Enable DR0 as a global breakpoint
        __asm__ __volatile__( "movl %0, %%dr7\n" : : "r" (0x00030002) );

        // Enables all four BR registers as global breakpoints
        __asm__ __volatile__( "movl %0, %%dr7\n" : : "r" (0x333300AA) );

        // shouldn't this trigger a breakpoint exception?
        i[0] = 0xDEADBEEF;

        kfree(i);

	// this will correctly trigger a breakpoint
	// __asm__ ( "movl %0, %%dr7\n" : : "r" (0x333320AA) );
        // __asm__ ( "movl %0, %%dr7\n" : : "r" (0x333300AA) );
        return 0;
}

module_init(db_reg_test);

root@foxtrot:~/dbregtest# tail /var/log/kern.log
Jul  5 09:01:16 localhost kernel: i == 0xC046E000
Jul  5 09:01:16 localhost kernel: virt_to_phys(i)  == 0x0046E000
author
#1 lore raw

kprobe support for memory access watchpoints

I was wondering if there are plans to support a method to register watchpoints for memory data access with kprobe. On x86, it's possible to watch for read/write access to arbitrary memory locations via DR memory registers.

Perhaps register_kprobe() could be modified to support this or perhaps some new function. This would probably be difficult based on how differently kprobe works vs. how the DR registers work. I thought I would send an email because you might be doing or thinking something similar.

Enjoy,
Jeff

#2 lore raw

Re: kprobe support for memory access watchpoints

Jeff,

Quoted 4 lines

I was wondering if there are plans to support a method to register watchpoints for memory data access with kprobe. On x86, it's possible to watch for read/write access to arbitrary memory locations via DR memory registers.

Here are couple of patches providing debug register allocation mechanism and kernel API to register watchpoints. These patches were posted and reviewed on lkml some time back, Please see the URL below for details.

http://seclists.org/lists/linux-kernel/2004/Oct/4730.html
http://seclists.org/lists/linux-kernel/2004/Oct/4729.html

Thanks
Prasanna

Signature · 5 lines
Prasanna S Panchamukhi
Linux Technology Center
India Software Labs, IBM Bangalore
Ph: 91-80-25044636
<[email protected]>
author
#3 lore raw
In reply to Jeff Carr

I setup KDB with 2.6.13-rc1. It works well. continuing past a BUG() is potentially very useful. Well, it's at least fun to see it happen :)

I wanted to set the BR0/7 registers from within running code. I thought that would trigger KDB, but it doesn't. It seems that the breakpoint trap is passed through the kernel and into userspace and caught by glibc. I tried to track this down, but am lost in how or where a int3 actually is handled.

I looked through the Documentation/kdb/ man pages (they are nice) but didn't see anything on how one might register a breakpoint with KDB from
some exported function.

The kprobe guys have a patch that I can use to try to solve my problem (described below), but I thought I would try kdb some more before I switched to trying kprobe.

Thanks,
Jeff

BTW, I don't know enough about kbuild, but if it is possible, you might want to force CONFIG_KPROBE off when you have kdb enabled or you get duplicate do_int3() functions defined in arch/i386/traps.c

Quoted 88 lines · Jeff Carr

I'm trying to set the x86 breakpoint registers to trip on write data.

After they are set, nothing seems to happen when I trigger them.

It's possible I'm not setting them correctly, I tried putting the virt_to_phys() value in them. And, I tried looking at what KPROBE puts in them, but it looks like kprobe doesn't use them at all.

In 2.6.11, arch/i386/ has 5 places where it modifies the db regs:

	do_debug()              in traps.c    (??)
	do_signal()             in signal.c   (re-enable them)
	fix_processor_context() in cpu.c      (reload them)

__switch_to() in process.c (reload them)
cpu_init() in common.c (clears them)

Just FYI: sometime after 2.6.11, the macros get_debugreg() and set_debugreg() were defined in include/asm-i386/processor.h and set to be used in the 5 places above. In any case, the functionality seems the same the above routines. (And the registers names corrected s/db/dr/g )

In any case, setting these registers never seems to do anything. No INT3 or INT1 (is it really supposed to generate an interrupt?) Perhaps I need to have kgdb setup.

The closest I can get to making anything happen is if I set bit 13 of DR7 (triggers on the next access of the breakpoint registers) then when I insmod I get:

root@foxtrot:~/dbregtest# insmod ./dbregtest.ko
Trace/breakpoint trap
root@foxtrot:~/dbregtest#

If I set BR0 with the value from virt_to_phys() I don't get this trace/breakpoint trap.

Enjoy,
Jeff

// Current documentation for these registers is in Vol 3 Section 15.2:
// Also note: EFLAGS BIT 16 (Resume) section 2.3 disables #DB exceptions // http://developer.intel.com/design/pentium4/manuals/index_new.htm

static int __init db_reg_test(void)

{
        u32 *i;
        unsigned int phys_addr;

        i = kmalloc( 0x1000, GFP_DMA );
        printk("i == 0x%08X\n", (int) i);
        phys_addr = virt_to_phys(i);

        printk("virt_to_phys(i)  == 0x%08X\n", (int) phys_addr);
        __asm__ __volatile__( "movl %0, %%dr0\n" : : "r" (phys_addr) );

        // clear out the DR6 status register
        __asm__ __volatile__( "movl %0, %%dr6\n" : : "r" (0xFFFF0FF0) );

        // Enable DR0 as a global breakpoint
        __asm__ __volatile__( "movl %0, %%dr7\n" : : "r" (0x00030002) );

        // Enables all four BR registers as global breakpoints
        __asm__ __volatile__( "movl %0, %%dr7\n" : : "r" (0x333300AA) );

        // shouldn't this trigger a breakpoint exception?
        i[0] = 0xDEADBEEF;

        kfree(i);

	// this will correctly trigger a breakpoint
	// __asm__ ( "movl %0, %%dr7\n" : : "r" (0x333320AA) );
        // __asm__ ( "movl %0, %%dr7\n" : : "r" (0x333300AA) );
        return 0;
}

module_init(db_reg_test);

root@foxtrot:~/dbregtest# tail /var/log/kern.log
Jul  5 09:01:16 localhost kernel: i == 0xC046E000
Jul  5 09:01:16 localhost kernel: virt_to_phys(i)  == 0x0046E000

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/