The kernel just send out a SIGTRAP signal when handling ptrace breakpoint in
debug exception, so it sounds safe to have interrupt enabled if it is not
disabled by the parent process.
Signed-off-by: Yang Shi <[email protected]>
---
arch/arm64/kernel/debug-monitors.c | 10 ++++++++++
1 file changed, 10 insertions(+)
diff --git a/arch/arm64/kernel/debug-monitors.c b/arch/arm64/kernel/debug-monitors.c
index 8aee3ae..90d70e4 100644
--- a/arch/arm64/kernel/debug-monitors.c
+++ b/arch/arm64/kernel/debug-monitors.c
@@ -239,6 +239,9 @@ static int single_step_handler(unsigned long addr, unsigned int esr,
return 0;
if (user_mode(regs)) {
+ if (interrupts_enabled(regs))
+ local_irq_enable();
+
info.si_signo = SIGTRAP;
info.si_errno = 0;
info.si_code = TRAP_HWBKPT;
@@ -310,6 +313,9 @@ static int brk_handler(unsigned long addr, unsigned int esr,
siginfo_t info;
if (user_mode(regs)) {
+ if (interrupts_enabled(regs))
+ local_irq_enable();
+
info = (siginfo_t) {
.si_signo = SIGTRAP,
.si_errno = 0,
@@ -337,6 +343,10 @@ int aarch32_break_handler(struct pt_regs *regs)
if (!compat_user_mode(regs))
return -EFAULT;
+ /* COMPAT_PSR_I_BIT has the same bit mask with non-compat one */
+ if (interrupts_enabled(regs))
+ local_irq_enable();
+
if (compat_thumb_mode(regs)) {
/* get 16-bit Thumb instruction */
get_user(thumb_instr, (u16 __user *)pc);
--
2.0.2
On Tue, Dec 15, 2015 at 04:18:08PM -0800, Yang Shi wrote:
> The kernel just send out a SIGTRAP signal when handling ptrace breakpoint in
> debug exception, so it sounds safe to have interrupt enabled if it is not
> disabled by the parent process.
Is this actually fixing an issue you're seeing, or did you just spot this?
Given that force_sig_info disable interrupts, I don't think this is really
worth doing.
> Signed-off-by: Yang Shi <[email protected]>
> ---
> arch/arm64/kernel/debug-monitors.c | 10 ++++++++++
> 1 file changed, 10 insertions(+)
>
> diff --git a/arch/arm64/kernel/debug-monitors.c b/arch/arm64/kernel/debug-monitors.c
> index 8aee3ae..90d70e4 100644
> --- a/arch/arm64/kernel/debug-monitors.c
> +++ b/arch/arm64/kernel/debug-monitors.c
> @@ -239,6 +239,9 @@ static int single_step_handler(unsigned long addr, unsigned int esr,
> return 0;
>
> if (user_mode(regs)) {
> + if (interrupts_enabled(regs))
> + local_irq_enable();
> +
My worry here is that we take an interrupt and, on the return path,
decide to reschedule due to CONFIG_PREEMPT. If we somehow end up back
in the debugger, I'm concerned that it could remove the breakpoint and
then later see an unexpected SIGTRAP from the child.
Having said that, I've failed to construct a non-racy scenario in which
that can happen, but I'm just really uncomfortable making this change
unless there's a real problem being solved.
Will
On 12/16/2015 3:13 AM, Will Deacon wrote:
> On Tue, Dec 15, 2015 at 04:18:08PM -0800, Yang Shi wrote:
>> The kernel just send out a SIGTRAP signal when handling ptrace breakpoint in
>> debug exception, so it sounds safe to have interrupt enabled if it is not
>> disabled by the parent process.
>
> Is this actually fixing an issue you're seeing, or did you just spot this?
> Given that force_sig_info disable interrupts, I don't think this is really
> worth doing.
I should made more comments at the first place, sorry for the inconvenience.
I did run into some problems on -rt kernel with CRIU restore, please see
the below kernel bug log:
BUG: sleeping function called from invalid context at
/kernel-source/kernel/locking/rtmutex.c:917
in_atomic(): 0, irqs_disabled(): 128, pid: 551, name: test.sh
CPU: 5 PID: 551 Comm: test.sh Not tainted 4.1.13-rt13 #7
Hardware name: Freescale Layerscape 2085a RDB Board (DT)
Call trace:
[<ffff8000000885f0>] dump_backtrace+0x0/0x128
[<ffff80000008873c>] show_stack+0x24/0x30
[<ffff800000798e84>] dump_stack+0x80/0xa0
[<ffff8000000bd858>] ___might_sleep+0x128/0x1a0
[<ffff8000007a072c>] rt_spin_lock+0x2c/0x40
[<ffff8000000a477c>] force_sig_info+0xcc/0x210
[<ffff800000085174>] brk_handler.part.2+0x6c/0x80
[<ffff800000085260>] brk_handler+0xd8/0xe8
[<ffff800000082368>] do_debug_exception+0x58/0xb8
Exception stack(0xffff80834b6e3e30 to 0xffff80834b6e3f50)
3e20: 00000000 00000000 004e6000
00000000
3e40: ffffffff ffffffff 00400004 00000000 0000d280 00000000 00000007
00000000
3e60: 00000021 00000000 ffffffff ffffffff 0000011a 00000000 000000de
00000000
3e80: 4ab9ef50 ffff8083 00086324 ffff8000 e587f780 0000ffff 000839b0
ffff8000
3ea0: 00000000 00000000 004e6000 00000000 ffffffff ffffffff 00400004
00000000
3ec0: 60000000 00000000 00000015 00000000 aa3e6000 0000ffff 0000d280
00000000
3ee0: 00000007 00000000 00000021 00000000 ffffffff ffffffff 00000000
00000000
3f00: 00000000 00000000 00000000 00000000 000000de 00000000 00000008
00000000
3f20: 004eff00 00000000 004e4ff0 00000000 004f0490 00000000 e587f730
0000ffff
3f40: 0046f508 00000000 00000028 00000000
It is because force_sig_info called spin_lock_irqsave which could sleep
on -rt kernel with irq disabled.
However, it just happens at brk_handler in my test. But, I saw
single_step has the same code path, so I expanded it to single step too.
Since this is rt related, cc to rt mailing list too.
>
>> Signed-off-by: Yang Shi <[email protected]>
>> ---
>> arch/arm64/kernel/debug-monitors.c | 10 ++++++++++
>> 1 file changed, 10 insertions(+)
>>
>> diff --git a/arch/arm64/kernel/debug-monitors.c b/arch/arm64/kernel/debug-monitors.c
>> index 8aee3ae..90d70e4 100644
>> --- a/arch/arm64/kernel/debug-monitors.c
>> +++ b/arch/arm64/kernel/debug-monitors.c
>> @@ -239,6 +239,9 @@ static int single_step_handler(unsigned long addr, unsigned int esr,
>> return 0;
>>
>> if (user_mode(regs)) {
>> + if (interrupts_enabled(regs))
>> + local_irq_enable();
>> +
>
> My worry here is that we take an interrupt and, on the return path,
> decide to reschedule due to CONFIG_PREEMPT. If we somehow end up back
> in the debugger, I'm concerned that it could remove the breakpoint and
> then later see an unexpected SIGTRAP from the child.
>
> Having said that, I've failed to construct a non-racy scenario in which
> that can happen, but I'm just really uncomfortable making this change
> unless there's a real problem being solved.
The patch is inspired by the similar code in other architectures, e.g.
x86 and powerpc which have hardware debug exception to handle breakpoint
and single step like arm64. And, they have interrupt enabled in both
breakpoint and single step. So, I'm supposed arm64 could do the same thing.
For the preempt case, it might be possible, but it sounds like a
exception handling problem to me. The preempt should not be allowed in
debug exception (current arm64 kernel does it), and in interrupt return
path the code should check if debug is on or not. If debug is on,
preempt should be just skipped. Or we could disable preempt in debug
exception.
I also checked the handling in x86 and powerpc, they go different way.
1. x86
Disable preempt in IST exception since it uses per CPU stack.
2. powerpc
Check if debug is on in interrupt return path. Powerpc has DBCR0_IDM
indicate if the processor is in debug mode.
For ARM64, I don't find such bit. So, I may consider to have preempt
disabled.
Thanks,
Yang
>
> Will
>
On Wed, Dec 16, 2015 at 12:45:15PM -0800, Shi, Yang wrote:
> On 12/16/2015 3:13 AM, Will Deacon wrote:
> >On Tue, Dec 15, 2015 at 04:18:08PM -0800, Yang Shi wrote:
> >>The kernel just send out a SIGTRAP signal when handling ptrace breakpoint in
> >>debug exception, so it sounds safe to have interrupt enabled if it is not
> >>disabled by the parent process.
> >
> >Is this actually fixing an issue you're seeing, or did you just spot this?
> >Given that force_sig_info disable interrupts, I don't think this is really
> >worth doing.
>
> I should made more comments at the first place, sorry for the inconvenience.
>
> I did run into some problems on -rt kernel with CRIU restore, please see the
> below kernel bug log:
Thanks.
> >My worry here is that we take an interrupt and, on the return path,
> >decide to reschedule due to CONFIG_PREEMPT. If we somehow end up back
> >in the debugger, I'm concerned that it could remove the breakpoint and
> >then later see an unexpected SIGTRAP from the child.
> >
> >Having said that, I've failed to construct a non-racy scenario in which
> >that can happen, but I'm just really uncomfortable making this change
> >unless there's a real problem being solved.
>
> The patch is inspired by the similar code in other architectures, e.g. x86
> and powerpc which have hardware debug exception to handle breakpoint and
> single step like arm64. And, they have interrupt enabled in both breakpoint
> and single step. So, I'm supposed arm64 could do the same thing.
>
> For the preempt case, it might be possible, but it sounds like a exception
> handling problem to me. The preempt should not be allowed in debug exception
> (current arm64 kernel does it), and in interrupt return path the code should
> check if debug is on or not. If debug is on, preempt should be just skipped.
> Or we could disable preempt in debug exception.
Yeah, disabling preemption during the debug handler and then enabling
interrupts if it came from userspace sounds like the best option. However,
that's a fairly invasive change to entry.S at this point, so maybe we're
better off with something like the patch below in the meantime?
Will
--->8
diff --git a/arch/arm64/kernel/debug-monitors.c b/arch/arm64/kernel/debug-monitors.c
index 8aee3aeec3e6..7f4913f2ea3c 100644
--- a/arch/arm64/kernel/debug-monitors.c
+++ b/arch/arm64/kernel/debug-monitors.c
@@ -226,11 +226,29 @@ static int call_step_hook(struct pt_regs *regs, unsigned int esr)
return retval;
}
+static void send_user_sigtrap(int si_code)
+{
+ struct pt_regs *regs = current_pt_regs();
+ siginfo_t info = {
+ .si_signo = SIGTRAP,
+ .si_errno = 0,
+ .si_code = si_code,
+ .si_addr = (void __user *)instruction_pointer(regs),
+ };
+
+ if (WARN_ON(!user_mode(regs)))
+ return;
+
+ preempt_disable();
+ local_irq_enable();
+ force_sig_info(SIGTRAP, &info, current);
+ local_irq_disable();
+ preempt_enable();
+}
+
static int single_step_handler(unsigned long addr, unsigned int esr,
struct pt_regs *regs)
{
- siginfo_t info;
-
/*
* If we are stepping a pending breakpoint, call the hw_breakpoint
* handler first.
@@ -239,11 +257,7 @@ static int single_step_handler(unsigned long addr, unsigned int esr,
return 0;
if (user_mode(regs)) {
- info.si_signo = SIGTRAP;
- info.si_errno = 0;
- info.si_code = TRAP_HWBKPT;
- info.si_addr = (void __user *)instruction_pointer(regs);
- force_sig_info(SIGTRAP, &info, current);
+ send_user_sigtrap(TRAP_HWBKPT);
/*
* ptrace will disable single step unless explicitly
@@ -307,17 +321,8 @@ static int call_break_hook(struct pt_regs *regs, unsigned int esr)
static int brk_handler(unsigned long addr, unsigned int esr,
struct pt_regs *regs)
{
- siginfo_t info;
-
if (user_mode(regs)) {
- info = (siginfo_t) {
- .si_signo = SIGTRAP,
- .si_errno = 0,
- .si_code = TRAP_BRKPT,
- .si_addr = (void __user *)instruction_pointer(regs),
- };
-
- force_sig_info(SIGTRAP, &info, current);
+ send_user_sigtrap(TRAP_BRKPT);
} else if (call_break_hook(regs, esr) != DBG_HOOK_HANDLED) {
pr_warning("Unexpected kernel BRK exception at EL1\n");
return -EFAULT;
@@ -328,7 +333,6 @@ static int brk_handler(unsigned long addr, unsigned int esr,
int aarch32_break_handler(struct pt_regs *regs)
{
- siginfo_t info;
u32 arm_instr;
u16 thumb_instr;
bool bp = false;
@@ -359,14 +363,7 @@ int aarch32_break_handler(struct pt_regs *regs)
if (!bp)
return -EFAULT;
- info = (siginfo_t) {
- .si_signo = SIGTRAP,
- .si_errno = 0,
- .si_code = TRAP_BRKPT,
- .si_addr = pc,
- };
-
- force_sig_info(SIGTRAP, &info, current);
+ send_user_sigtrap(TRAP_BRKPT);
return 0;
}
On Mon, 21 Dec 2015, Will Deacon wrote:
> +static void send_user_sigtrap(int si_code)
> +{
> + struct pt_regs *regs = current_pt_regs();
> + siginfo_t info = {
> + .si_signo = SIGTRAP,
> + .si_errno = 0,
> + .si_code = si_code,
> + .si_addr = (void __user *)instruction_pointer(regs),
> + };
> +
> + if (WARN_ON(!user_mode(regs)))
> + return;
> +
> + preempt_disable();
That doesn't work on RT either. force_sig_info() takes task->sighand->siglock,
which is a 'sleeping' spinlock on RT.
Why would we need to disable preemption here at all? What's the problem of
being preempted or even migrated?
Thanks,
tglx
On Mon, Dec 21, 2015 at 05:51:22PM +0100, Thomas Gleixner wrote:
> On Mon, 21 Dec 2015, Will Deacon wrote:
> > +static void send_user_sigtrap(int si_code)
> > +{
> > + struct pt_regs *regs = current_pt_regs();
> > + siginfo_t info = {
> > + .si_signo = SIGTRAP,
> > + .si_errno = 0,
> > + .si_code = si_code,
> > + .si_addr = (void __user *)instruction_pointer(regs),
> > + };
> > +
> > + if (WARN_ON(!user_mode(regs)))
> > + return;
> > +
> > + preempt_disable();
>
> That doesn't work on RT either. force_sig_info() takes task->sighand->siglock,
> which is a 'sleeping' spinlock on RT.
Ah, I missed that :/
> Why would we need to disable preemption here at all? What's the problem of
> being preempted or even migrated?
There *might* not be a problem, I'm just really nervous about changing
the behaviour on the debug path and subtly changing how ptrace behaves.
My worry was that you could somehow get back into the tracer, and it
could remove a software breakpoint in the knowledge that it wouldn't
see any future (spurious) SIGTRAPs for that location.
Without a concrete example, however, I guess I'll bite the bullet and
enable irqs across the call to force_sig_info, since there is clearly a
real issue here on RT.
Will
On 12/21/2015 9:00 AM, Will Deacon wrote:
> On Mon, Dec 21, 2015 at 05:51:22PM +0100, Thomas Gleixner wrote:
>> On Mon, 21 Dec 2015, Will Deacon wrote:
>>> +static void send_user_sigtrap(int si_code)
>>> +{
>>> + struct pt_regs *regs = current_pt_regs();
>>> + siginfo_t info = {
>>> + .si_signo = SIGTRAP,
>>> + .si_errno = 0,
>>> + .si_code = si_code,
>>> + .si_addr = (void __user *)instruction_pointer(regs),
>>> + };
>>> +
>>> + if (WARN_ON(!user_mode(regs)))
>>> + return;
>>> +
>>> + preempt_disable();
>>
>> That doesn't work on RT either. force_sig_info() takes task->sighand->siglock,
>> which is a 'sleeping' spinlock on RT.
>
> Ah, I missed that :/
>
>> Why would we need to disable preemption here at all? What's the problem of
>> being preempted or even migrated?
>
> There *might* not be a problem, I'm just really nervous about changing
> the behaviour on the debug path and subtly changing how ptrace behaves.
>
> My worry was that you could somehow get back into the tracer, and it
> could remove a software breakpoint in the knowledge that it wouldn't
> see any future (spurious) SIGTRAPs for that location.
>
> Without a concrete example, however, I guess I'll bite the bullet and
> enable irqs across the call to force_sig_info, since there is clearly a
> real issue here on RT.
Thanks for the review.
I don't have any concrete usecase for that race condition too since all
ptrace tests from ltp passed. However, even though we need preemption
disabled at some point in debug exception code path in the future, I
think we could just extend the "signal delay send" approach from x86-64
to arm64, which is currently used by x86-64 on -rt kernel only.
Regards,
Yang
>
> Will
>