[PATCH] arm64 ptrace.c

[PATCH] arm64 ptrace.c

[Will Deacon]

On Tue, Dec 10, 2013 at 06:12:24AM +0000, ??? wrote:
> Hi all,

Hello,

> In these dayes, I use gdb to debug my program on ARM foundation model and
> fast model with Linux 3.12.0-4.12 from ubuntu.  The gdb reports errors
> about setting hardware debug register when stepping pthread_create().
> After investigating gdb 7.6.1 source and Linux 3.12 source, I prepare a
> patch for arch/arm64/kernel/ptrace.c.  Although the patch could solve gdb
> error, I am not sure it has no side effects.  May somebody help to give
> suggestions?  Thank you.
> 
> Ref. to https://bugs.launchpad.net/gdb-linaro/+bug/1205391
> 
> diff --git a/arch/arm64/kernel/ptrace.c b/arch/arm64/kernel/ptrace.c
> index 6777a21..981f961 100644
> --- a/arch/arm64/kernel/ptrace.c
> +++ b/arch/arm64/kernel/ptrace.c
> @@ -214,9 +214,16 @@ static int ptrace_hbp_fill_attr_ctrl(unsigned int note_type,
>  {
>   int err, len, type, disabled = !ctrl.enabled;
> 
> + /*
> + * Does not change type and len in disabled case because it will
> + * break the assertion that only 2 types (TYPE_DATA and TYPE_INST) in
> + * the event linked list in ./kernel/events/hw_breakpoint.c
> + */
>   if (disabled) {
> - len = 0;
> - type = HW_BREAKPOINT_EMPTY;
> + /*
> + * len = 0;
> + * type = HW_BREAKPOINT_EMPTY;
> + */
>   } else {
>   err = arch_bp_generic_fields(ctrl, &len, &type);
>   if (err)
> @@ -234,10 +241,10 @@ static int ptrace_hbp_fill_attr_ctrl(unsigned int note_type,
>   default:
>   return -EINVAL;
>   }
> + attr->bp_len = len;
> + attr->bp_type = type;
>   }
> 
> - attr->bp_len = len;
> - attr->bp_type = type;
>   attr->disabled = disabled;
> 
>   return 0;

I don't understand what this achieves. Do you have a better description of
the problem and/or strace logs showing all the hw_breakpoint ptrace requests
issued by GDB up until the first failure please?

That should help to figure out what the problem is.

Cheers,

Will

[PATCH] arm64 ptrace.c

[Aaron Liu]

Hi Will,

gdb calls ptrace twice to set watch points and break points as
following. After first ptrace call is completed, second ptrace call
fails with no space error. The first ptrace will reserve 16 watch
points in static LIST_HEAD(bp_task_head)@kernel/events/hw_breakpoint.c
and each entry, ie. struct perf_event, has a member called
attr.bp_type. The bp_type in the entry should be used as indicator for
watch points or break points. You could see
__reserve_bp_slot at kernel/events/hw_breakpoint.c and find it does a
basic check bp_type should be not HW_BREAKPOINT_EMPTY and
HW_BREAKPOINT_INVALID. After the __reserve_bp_slot call, the reserved
slot's bp_type is HW_BREAKPOINT_RW or HW_BREAKPOINT_X. However,
ptrace_hbp_fill_attr_ctrl at arch/arm64/kernel/ptrace.c set the bp_type
to HW_BREAKPOINT_EMPTY if it's disabled. This causes the
find_slot_idx at kernel/event/hw_breakpoint.c to judge the existing
entries as TYPE_INST. So, after first 16 watch points are reserved in
list, the following break points determine no space to reserve, ie.
max 16 break points. The patch only set disable bit even user zeroing
control bits and it could still pass
modify_user_hw_breakpoint at kernel/events/hw_breakpoint.c.

I post the test program that causes gdb error.
It just creates a threads and runs forever.

#include <pthread.h>
#include <stdio.h>
#include <unistd.h>

static void* thread_body(void* usr) {
  int c = 0;
  printf("thread running...\n");
  while(1) {
    sleep(1);
    printf("thread count: %d\n", ++c);
  }
  return 0;
}

int main(int argc, const char** argv) {
  pthread_t t;
  void* tr = 0;
  int r = 0;
  r = pthread_create(&t, 0, thread_body, 0);
  if (r) {
    printf("pthread_create failed\n");
    return -1;
  }
  pthread_join(t, &tr);
  return 0;
}

Here is the strace log that gdb debug the program. I discards the
others because the log is too much. If you need full log, please tell
me.

ptrace(PTRACE_GETEVENTMSG, 559, 0, 0x7fe563aba8) = 0
ptrace(PTRACE_SETREGSET, 562, 0x403 /* NT_??? */, [{0x7fe563a970, 264}]) = 0
ptrace(PTRACE_SETREGSET, 562, 0x402 /* NT_??? */, [{0x7fe563a970,
264}]) = -1 ENOSPC (No space left on device)
rt_sigprocmask(SIG_SETMASK, [], NULL, 8) = 0
rt_sigaction(SIGTTOU, {SIG_IGN, [TTOU], SA_RESTART}, {SIG_DFL, [TTOU],
SA_RESTART}, 8) = 0
ioctl(0, SNDCTL_TMR_TIMEBASE or SNDRV_TIMER_IOCTL_NEXT_DEVICE or
TCGETS, {B38400 opost isig icanon echo ...}) = 0
ioctl(0, TIOCGPGRP, [559])              = 0
ioctl(0, SNDCTL_TMR_TIMEBASE or SNDRV_TIMER_IOCTL_NEXT_DEVICE or
TCGETS, {B38400 opost isig icanon echo ...}) = 0
ioctl(0, SNDCTL_TMR_START or SNDRV_TIMER_IOCTL_TREAD or TCSETS,
{B38400 opost isig icanon echo ...}) = 0
ioctl(0, SNDCTL_TMR_TIMEBASE or SNDRV_TIMER_IOCTL_NEXT_DEVICE or
TCGETS, {B38400 opost isig icanon echo ...}) = 0
ioctl(0, SNDRV_TIMER_IOCTL_SELECT or TIOCSPGRP, [554]) = 0
rt_sigaction(SIGTTOU, {SIG_DFL, [TTOU], SA_RESTART}, {SIG_IGN, [TTOU],
SA_RESTART}, 8) = 0
fcntl(0, F_GETFL)                       = 0x20002 (flags O_RDWR|0x20000)
fcntl(0, F_SETFL, O_RDWR|0x20000)       = 0
fcntl(0, F_SETFL, O_RDWR|0x20000)       = 0
ioctl(1, TCSBRK, 0x1)                   = 0
write(2, "Unexpected error setting hardwar"..., 49Unexpected error
setting hardware debug registers) = 49
write(2, "\n", 1
)                       = 1

Thank you
Aaron


2013/12/10 Will Deacon <will.deacon@arm.com>:
> On Tue, Dec 10, 2013 at 06:12:24AM +0000, ??? wrote:
>> Hi all,
>
> Hello,
>
>> In these dayes, I use gdb to debug my program on ARM foundation model and
>> fast model with Linux 3.12.0-4.12 from ubuntu.  The gdb reports errors
>> about setting hardware debug register when stepping pthread_create().
>> After investigating gdb 7.6.1 source and Linux 3.12 source, I prepare a
>> patch for arch/arm64/kernel/ptrace.c.  Although the patch could solve gdb
>> error, I am not sure it has no side effects.  May somebody help to give
>> suggestions?  Thank you.
>>
>> Ref. to https://bugs.launchpad.net/gdb-linaro/+bug/1205391
>>
>> diff --git a/arch/arm64/kernel/ptrace.c b/arch/arm64/kernel/ptrace.c
>> index 6777a21..981f961 100644
>> --- a/arch/arm64/kernel/ptrace.c
>> +++ b/arch/arm64/kernel/ptrace.c
>> @@ -214,9 +214,16 @@ static int ptrace_hbp_fill_attr_ctrl(unsigned int note_type,
>>  {
>>   int err, len, type, disabled = !ctrl.enabled;
>>
>> + /*
>> + * Does not change type and len in disabled case because it will
>> + * break the assertion that only 2 types (TYPE_DATA and TYPE_INST) in
>> + * the event linked list in ./kernel/events/hw_breakpoint.c
>> + */
>>   if (disabled) {
>> - len = 0;
>> - type = HW_BREAKPOINT_EMPTY;
>> + /*
>> + * len = 0;
>> + * type = HW_BREAKPOINT_EMPTY;
>> + */
>>   } else {
>>   err = arch_bp_generic_fields(ctrl, &len, &type);
>>   if (err)
>> @@ -234,10 +241,10 @@ static int ptrace_hbp_fill_attr_ctrl(unsigned int note_type,
>>   default:
>>   return -EINVAL;
>>   }
>> + attr->bp_len = len;
>> + attr->bp_type = type;
>>   }
>>
>> - attr->bp_len = len;
>> - attr->bp_type = type;
>>   attr->disabled = disabled;
>>
>>   return 0;
>
> I don't understand what this achieves. Do you have a better description of
> the problem and/or strace logs showing all the hw_breakpoint ptrace requests
> issued by GDB up until the first failure please?
>
> That should help to figure out what the problem is.
>
> Cheers,
>
> Will

[PATCH] arm64 ptrace.c

[Will Deacon]

On Wed, Dec 11, 2013 at 07:16:11AM +0000, Aaron Liu wrote:
> Hi Will,

Hi Aaron,

> gdb calls ptrace twice to set watch points and break points as
> following. After first ptrace call is completed, second ptrace call
> fails with no space error. The first ptrace will reserve 16 watch
> points in static LIST_HEAD(bp_task_head)@kernel/events/hw_breakpoint.c
> and each entry, ie. struct perf_event, has a member called
> attr.bp_type. The bp_type in the entry should be used as indicator for
> watch points or break points. You could see
> __reserve_bp_slot at kernel/events/hw_breakpoint.c and find it does a
> basic check bp_type should be not HW_BREAKPOINT_EMPTY and
> HW_BREAKPOINT_INVALID. After the __reserve_bp_slot call, the reserved
> slot's bp_type is HW_BREAKPOINT_RW or HW_BREAKPOINT_X. However,
> ptrace_hbp_fill_attr_ctrl at arch/arm64/kernel/ptrace.c set the bp_type
> to HW_BREAKPOINT_EMPTY if it's disabled. This causes the
> find_slot_idx at kernel/event/hw_breakpoint.c to judge the existing
> entries as TYPE_INST. So, after first 16 watch points are reserved in
> list, the following break points determine no space to reserve, ie.
> max 16 break points. The patch only set disable bit even user zeroing
> control bits and it could still pass
> modify_user_hw_breakpoint at kernel/events/hw_breakpoint.c.

Ok, thanks for the explanation. This used to work, but I suspect something
changed in the core code which caused this to start breaking with
HW_BREAKPOINT_EMPTY.

Lucky for us, our ptrace_hbp_create function will infer the type/len for us
using values that are known to pass validation. In that case, the use of
HW_BREAKPOINT_EMPTY is redundant -- we just need to make sure we avoid
populating anything other than the disabled field in the perf_event_attr.

A curious behaviour here (which also exists in the current code) is that the
breakpoint register effectively becomes read-only apart from the enable bit
if you are disabling it. That shouldn't effect gdb, though, and is required
to allow nuking of the breakpoint without changing its type.

Can you test the following patch please?

Thanks,

Will

--->8

diff --git a/arch/arm64/kernel/ptrace.c b/arch/arm64/kernel/ptrace.c
index 6777a2192b83..6a8928bba03c 100644
--- a/arch/arm64/kernel/ptrace.c
+++ b/arch/arm64/kernel/ptrace.c
@@ -214,31 +214,29 @@ static int ptrace_hbp_fill_attr_ctrl(unsigned int note_type,
 {
 	int err, len, type, disabled = !ctrl.enabled;
 
-	if (disabled) {
-		len = 0;
-		type = HW_BREAKPOINT_EMPTY;
-	} else {
-		err = arch_bp_generic_fields(ctrl, &len, &type);
-		if (err)
-			return err;
-
-		switch (note_type) {
-		case NT_ARM_HW_BREAK:
-			if ((type & HW_BREAKPOINT_X) != type)
-				return -EINVAL;
-			break;
-		case NT_ARM_HW_WATCH:
-			if ((type & HW_BREAKPOINT_RW) != type)
-				return -EINVAL;
-			break;
-		default:
+	attr->disabled = disabled;
+	if (disabled)
+		return 0;
+
+	err = arch_bp_generic_fields(ctrl, &len, &type);
+	if (err)
+		return err;
+
+	switch (note_type) {
+	case NT_ARM_HW_BREAK:
+		if ((type & HW_BREAKPOINT_X) != type)
 			return -EINVAL;
-		}
+		break;
+	case NT_ARM_HW_WATCH:
+		if ((type & HW_BREAKPOINT_RW) != type)
+			return -EINVAL;
+		break;
+	default:
+		return -EINVAL;
 	}
 
 	attr->bp_len	= len;
 	attr->bp_type	= type;
-	attr->disabled	= disabled;
 
 	return 0;
 }

[PATCH] arm64 ptrace.c

[Aaron Liu]

Hi Will,

Your patch works without errors.
Thanks
Aaron

2013/12/12 Will Deacon <will.deacon@arm.com>:
> On Wed, Dec 11, 2013 at 07:16:11AM +0000, Aaron Liu wrote:
>> Hi Will,
>
> Hi Aaron,
>
>> gdb calls ptrace twice to set watch points and break points as
>> following. After first ptrace call is completed, second ptrace call
>> fails with no space error. The first ptrace will reserve 16 watch
>> points in static LIST_HEAD(bp_task_head)@kernel/events/hw_breakpoint.c
>> and each entry, ie. struct perf_event, has a member called
>> attr.bp_type. The bp_type in the entry should be used as indicator for
>> watch points or break points. You could see
>> __reserve_bp_slot at kernel/events/hw_breakpoint.c and find it does a
>> basic check bp_type should be not HW_BREAKPOINT_EMPTY and
>> HW_BREAKPOINT_INVALID. After the __reserve_bp_slot call, the reserved
>> slot's bp_type is HW_BREAKPOINT_RW or HW_BREAKPOINT_X. However,
>> ptrace_hbp_fill_attr_ctrl at arch/arm64/kernel/ptrace.c set the bp_type
>> to HW_BREAKPOINT_EMPTY if it's disabled. This causes the
>> find_slot_idx at kernel/event/hw_breakpoint.c to judge the existing
>> entries as TYPE_INST. So, after first 16 watch points are reserved in
>> list, the following break points determine no space to reserve, ie.
>> max 16 break points. The patch only set disable bit even user zeroing
>> control bits and it could still pass
>> modify_user_hw_breakpoint at kernel/events/hw_breakpoint.c.
>
> Ok, thanks for the explanation. This used to work, but I suspect something
> changed in the core code which caused this to start breaking with
> HW_BREAKPOINT_EMPTY.
>
> Lucky for us, our ptrace_hbp_create function will infer the type/len for us
> using values that are known to pass validation. In that case, the use of
> HW_BREAKPOINT_EMPTY is redundant -- we just need to make sure we avoid
> populating anything other than the disabled field in the perf_event_attr.
>
> A curious behaviour here (which also exists in the current code) is that the
> breakpoint register effectively becomes read-only apart from the enable bit
> if you are disabling it. That shouldn't effect gdb, though, and is required
> to allow nuking of the breakpoint without changing its type.
>
> Can you test the following patch please?
>
> Thanks,
>
> Will
>
> --->8
>
> diff --git a/arch/arm64/kernel/ptrace.c b/arch/arm64/kernel/ptrace.c
> index 6777a2192b83..6a8928bba03c 100644
> --- a/arch/arm64/kernel/ptrace.c
> +++ b/arch/arm64/kernel/ptrace.c
> @@ -214,31 +214,29 @@ static int ptrace_hbp_fill_attr_ctrl(unsigned int note_type,
>  {
>         int err, len, type, disabled = !ctrl.enabled;
>
> -       if (disabled) {
> -               len = 0;
> -               type = HW_BREAKPOINT_EMPTY;
> -       } else {
> -               err = arch_bp_generic_fields(ctrl, &len, &type);
> -               if (err)
> -                       return err;
> -
> -               switch (note_type) {
> -               case NT_ARM_HW_BREAK:
> -                       if ((type & HW_BREAKPOINT_X) != type)
> -                               return -EINVAL;
> -                       break;
> -               case NT_ARM_HW_WATCH:
> -                       if ((type & HW_BREAKPOINT_RW) != type)
> -                               return -EINVAL;
> -                       break;
> -               default:
> +       attr->disabled = disabled;
> +       if (disabled)
> +               return 0;
> +
> +       err = arch_bp_generic_fields(ctrl, &len, &type);
> +       if (err)
> +               return err;
> +
> +       switch (note_type) {
> +       case NT_ARM_HW_BREAK:
> +               if ((type & HW_BREAKPOINT_X) != type)
>                         return -EINVAL;
> -               }
> +               break;
> +       case NT_ARM_HW_WATCH:
> +               if ((type & HW_BREAKPOINT_RW) != type)
> +                       return -EINVAL;
> +               break;
> +       default:
> +               return -EINVAL;
>         }
>
>         attr->bp_len    = len;
>         attr->bp_type   = type;
> -       attr->disabled  = disabled;
>
>         return 0;
>  }

[PATCH] arm64 ptrace.c

[Aaron Liu]

Hi all,

In these dayes, I use gdb to debug my program on ARM foundation model
and fast model with Linux 3.12.0-4.12 from ubuntu.
The gdb reports errors about setting hardware debug register when
stepping pthread_create().
After investigating gdb 7.6.1 source and Linux 3.12 source, I prepare
a patch for arch/arm64/kernel/ptrace.c.
Although the patch could solve gdb error, I am not sure it has no side effects.
May somebody help to give suggestions?
Thank you.

Ref. to https://bugs.launchpad.net/gdb-linaro/+bug/1205391

diff --git a/arch/arm64/kernel/ptrace.c b/arch/arm64/kernel/ptrace.c
index 6777a21..981f961 100644
--- a/arch/arm64/kernel/ptrace.c
+++ b/arch/arm64/kernel/ptrace.c
@@ -214,9 +214,16 @@ static int ptrace_hbp_fill_attr_ctrl(unsigned int
note_type,
 {
  int err, len, type, disabled = !ctrl.enabled;

+ /*
+  * Does not change type and len in disabled case because it will
+  * break the assertion that only 2 types (TYPE_DATA and TYPE_INST) in
+  * the event linked list in ./kernel/events/hw_breakpoint.c
+  */
  if (disabled) {
- len = 0;
- type = HW_BREAKPOINT_EMPTY;
+ /*
+  * len = 0;
+  * type = HW_BREAKPOINT_EMPTY;
+  */
  } else {
  err = arch_bp_generic_fields(ctrl, &len, &type);
  if (err)
@@ -234,10 +241,10 @@ static int ptrace_hbp_fill_attr_ctrl(unsigned
int note_type,
  default:
  return -EINVAL;
  }
+ attr->bp_len = len;
+ attr->bp_type = type;
  }

- attr->bp_len = len;
- attr->bp_type = type;
  attr->disabled = disabled;

  return 0;
--
Aaron Liu