Skip to contentSkip to conversation

ppc_set_hwdebug vs ptrace_set_debugreg

DISCUSSION Andreas Schwab 2 participants 11 messages Nov 27, 2010 – Jan 2, 2011

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

Why does ptrace_set_debugreg call register_user_hw_breakpoint, but ppc_set_hwdebug doesn't? Shouldn't ppc_set_hwdebug set the DABR_DATA_(READ|WRITE|TRANSLATION) bits in the dabr?

Andreas.

Signature · 3 lines
Andreas Schwab, [email protected]
GPG Key fingerprint = 58CA 54C7 6D53 942B 1756  01D3 44D5 214B 8276 4ED5
"And now for something completely different."
#1 lore raw
Quoted 5 lines · Andreas Schwab

Why does ptrace_set_debugreg call register_user_hw_breakpoint, but ppc_set_hwdebug doesn't? Shouldn't ppc_set_hwdebug set the DABR_DATA_(READ|WRITE|TRANSLATION) bits in the dabr?

Andreas.

The hw-breakpoint interfaces were initially planned for the old ptrace option PTRACE_SET_DEBUGREG,while, the newer ptrace options are mostly to exploit the advanced debug features of Book3E processors.

Although ppc_set_hwdebug() can set DABR through set_dabr() in arch/powerpc/kernel/process.c, it is good to have it converted to use register_user_hw_breakpoint(). This was planned to be done alongside the conversion of all ptrace options enabled by CONFIG_PPC_ADV_DEBUG_REGS, which is yet to be done.

Are you looking to use debug registers through perf, or somesuch, that you need register_user_hw_breakpoint() to be used for these new ptrace flags?

Thanks,
K.Prasad

#2 lore raw
Quoted 3 lines · K.Prasad

Although ppc_set_hwdebug() can set DABR through set_dabr() in arch/powerpc/kernel/process.c, it is good to have it converted to use register_user_hw_breakpoint().

What do you mean with "good to have"? It doesn't work without it unless I disable PERF_EVENTS (which is the only way to disable HAVE_HW_BREAKPOINT).

Andreas.

Signature · 3 lines
Andreas Schwab, [email protected]
GPG Key fingerprint = 58CA 54C7 6D53 942B 1756  01D3 44D5 214B 8276 4ED5
"And now for something completely different."
#3 lore raw
Quoted 6 lines · Andreas Schwab
Quoted 3 lines · K.Prasad

Although ppc_set_hwdebug() can set DABR through set_dabr() in arch/powerpc/kernel/process.c, it is good to have it converted to use register_user_hw_breakpoint().

Quoted 6 lines

What do you mean with "good to have"? It doesn't work without it unless I disable PERF_EVENTS (which is the only way to disable HAVE_HW_BREAKPOINT).

Andreas.

Let me see if I can cook up a patch for this i.e. make set_dabr() invoke register_user_hw_breakpoint() when CONFIG_PPC_BOOK3S is defined; before I head out on my vacation (starting second week of this month).

Thanks,
K.Prasad

#4 lore raw
Quoted 14 lines · K.Prasad
Quoted 6 lines · Andreas Schwab
Quoted 3 lines · K.Prasad

Although ppc_set_hwdebug() can set DABR through set_dabr() in arch/powerpc/kernel/process.c, it is good to have it converted to use register_user_hw_breakpoint().

Quoted 6 lines

What do you mean with "good to have"? It doesn't work without it unless I disable PERF_EVENTS (which is the only way to disable HAVE_HW_BREAKPOINT).

Andreas.

Quoted 7 lines

Let me see if I can cook up a patch for this i.e. make set_dabr() invoke register_user_hw_breakpoint() when CONFIG_PPC_BOOK3S is defined; before I head out on my vacation (starting second week of this month).

Thanks,
K.Prasad

Hi,
Can you check if the following patch (compile tested only) works
for you?

Thanks,
K.Prasad

---
 arch/powerpc/kernel/ptrace.c |   34 ++++++++++++++++++++++++++++++++++
 1 file changed, 34 insertions(+)
arch/powerpc/kernel/ptrace.c +34 -0
Index: linux-2.6.set_hwdebug/arch/powerpc/kernel/ptrace.c
===================================================================
@@ -1316,6 +1316,11 @@ static int set_dac_range(struct task_str
 static long ppc_set_hwdebug(struct task_struct *child,
 		     struct ppc_hw_breakpoint *bp_info)
 {
+#ifdef CONFIG_HAVE_HW_BREAKPOINT
+	struct perf_event *bp;
+	struct perf_event_attr attr;
+#endif /* CONFIG_HAVE_HW_BREAKPOINT */
+
 	if (bp_info->version != 1)
 		return -ENOTSUPP;
 #ifdef CONFIG_PPC_ADV_DEBUG_REGS
@@ -1362,10 +1367,29 @@ static long ppc_set_hwdebug(struct task_
 
 	if (child->thread.dabr)
 		return -ENOSPC;
+#ifdef CONFIG_HAVE_HW_BREAKPOINT
+	if (child->thread->ptrace_bps[0])
+		return -ENOSPC;
+#endif /* CONFIG_HAVE_HW_BREAKPOINT */
 
 	if ((unsigned long)bp_info->addr >= TASK_SIZE)
 		return -EIO;
 
+#ifdef CONFIG_HAVE_HW_BREAKPOINT
+	/* Create a new breakpoint request if one doesn't exist already */
+	hw_breakpoint_init(&attr);
+	attr.bp_addr = bp_info->addr & ~HW_BREAKPOINT_ALIGN;
+	arch_bp_generic_fields(bp_info->addr &
+				(DABR_DATA_WRITE | DABR_DATA_READ),
+							&attr.bp_type);
+
+	bp = register_user_hw_breakpoint(&attr, ptrace_triggered, task);
+	if (IS_ERR(bp))
+		return PTR_ERR(bp);
+
+	child->thread.ptrace_bps[0] = bp;
+#endif /* CONFIG_HAVE_HW_BREAKPOINT */
+
 	child->thread.dabr = (unsigned long)bp_info->addr;
 
 	return 1;
@@ -1395,6 +1419,16 @@ static long ppc_del_hwdebug(struct task_
 		return -EINVAL;
 	if (child->thread.dabr == 0)
 		return -ENOENT;
+#ifdef CONFIG_HAVE_HW_BREAKPOINT
+	/*
+	 * There is no way by which address in ptrace_bps[0] and thread.dabr
+	 * can be different. So we don't explicitly check if they're the same
+	 */
+	if (child->thread.ptrace_bps[0]) {
+		unregister_hw_breakpoint(child->thread.ptrace_bps[0]);
+		child->thread.ptrace_bps[0] = NULL;
+	}
+#endif /* CONFIG_HAVE_HW_BREAKPOINT */
 
 	child->thread.dabr = 0;
 
#5 lore raw
Quoted 6 lines · K.Prasad
arch/powerpc/kernel/ptrace.c +6 -0
+#ifdef CONFIG_HAVE_HW_BREAKPOINT
+	/* Create a new breakpoint request if one doesn't exist already */
+	hw_breakpoint_init(&attr);
+	attr.bp_addr = bp_info->addr & ~HW_BREAKPOINT_ALIGN;
+	arch_bp_generic_fields(bp_info->addr &
+				(DABR_DATA_WRITE | DABR_DATA_READ),

arch/powerpc/kernel/ptrace.c static long ppc_set_hwdebug(struct task_ · line 1393

Quoted 10 lines
arch/powerpc/kernel/ptrace.c +9 -0
+							&attr.bp_type);
+
+	bp = register_user_hw_breakpoint(&attr, ptrace_triggered, task);
+	if (IS_ERR(bp))
+		return PTR_ERR(bp);
+
+	child->thread.ptrace_bps[0] = bp;
+#endif /* CONFIG_HAVE_HW_BREAKPOINT */
+
 	child->thread.dabr = (unsigned long)bp_info->addr;

That cannot work, see
<http://permalink.gmane.org/gmane.linux.ports.ppc64.devel/71418>.

Andreas.

Signature · 3 lines
Andreas Schwab, [email protected]
GPG Key fingerprint = 58CA 54C7 6D53 942B 1756  01D3 44D5 214B 8276 4ED5
"And now for something completely different."
#6 lore raw
Quoted 13 lines · Andreas Schwab
Quoted 11 lines · K.Prasad
(quoted diff) +11 -0
+#ifdef CONFIG_HAVE_HW_BREAKPOINT
+	/* Create a new breakpoint request if one doesn't exist already */
+	hw_breakpoint_init(&attr);
+	attr.bp_addr = bp_info->addr & ~HW_BREAKPOINT_ALIGN;
+	arch_bp_generic_fields(bp_info->addr &
+				(DABR_DATA_WRITE | DABR_DATA_READ),
+							&attr.bp_type);
+
+	bp = register_user_hw_breakpoint(&attr, ptrace_triggered, task);
+	if (IS_ERR(bp))
+		return PTR_ERR(bp);
Quoted 9 lines

Review

Quoted 5 lines
(quoted diff) +4 -0
+
+	child->thread.ptrace_bps[0] = bp;
+#endif /* CONFIG_HAVE_HW_BREAKPOINT */
+
 	child->thread.dabr = (unsigned long)bp_info->addr;

That cannot work, see
<http://permalink.gmane.org/gmane.linux.ports.ppc64.devel/71418>.

Ok. The above patch makes it a bit easy.

How about the revised patch below? It is only compile-tested; have you got a quick test case that I can run?

Enable PPC_PTRACE_SETHWDEBUG and PPC_PTRACE_DELHWDEBUG to use the generic hardware breakpoint interfaces. This helps prevent conflict for the use of DABR register in the absence of CONFIG_PPC_ADV_DEBUG_REGS and when PTRACE_SET_DEBUGREG/PTRACE_GET_DEBUGREG flags are used by ptrace.

---
 arch/powerpc/kernel/ptrace.c |   32 ++++++++++++++++++++++++++++++++
 1 file changed, 32 insertions(+)
arch/powerpc/kernel/ptrace.c +32 -0
Index: linux-2.6.set_hwdebug/arch/powerpc/kernel/ptrace.c
===================================================================
@@ -1316,6 +1316,10 @@ static int set_dac_range(struct task_str
 static long ppc_set_hwdebug(struct task_struct *child,
 		     struct ppc_hw_breakpoint *bp_info)
 {
+#ifdef CONFIG_HAVE_HW_BREAKPOINT
+	struct perf_event *bp;
+	struct perf_event_attr attr;
+#endif /* CONFIG_HAVE_HW_BREAKPOINT */
 #ifndef CONFIG_PPC_ADV_DEBUG_REGS
 	unsigned long dabr;
 #endif
@@ -1365,6 +1369,10 @@ static long ppc_set_hwdebug(struct task_
 
 	if (child->thread.dabr)
 		return -ENOSPC;
+#ifdef CONFIG_HAVE_HW_BREAKPOINT
+	if (child->thread.ptrace_bps[0])
+		return -ENOSPC;
+#endif /* CONFIG_HAVE_HW_BREAKPOINT */
 
 	if ((unsigned long)bp_info->addr >= TASK_SIZE)
 		return -EIO;
@@ -1376,6 +1384,20 @@ static long ppc_set_hwdebug(struct task_
 	if (bp_info->trigger_type & PPC_BREAKPOINT_TRIGGER_WRITE)
 		dabr |= DABR_DATA_WRITE;
 
+#ifdef CONFIG_HAVE_HW_BREAKPOINT
+	/* Create a new breakpoint request if one doesn't exist already */
+	hw_breakpoint_init(&attr);
+	attr.bp_addr = dabr & ~HW_BREAKPOINT_ALIGN;
+	arch_bp_generic_fields(dabr & (DABR_DATA_WRITE | DABR_DATA_READ),
+							&attr.bp_type);
+
+	bp = register_user_hw_breakpoint(&attr, ptrace_triggered, child);
+	if (IS_ERR(bp))
+		return PTR_ERR(bp);
+
+	child->thread.ptrace_bps[0] = bp;
+#endif /* CONFIG_HAVE_HW_BREAKPOINT */
+
 	child->thread.dabr = dabr;
 
 	return 1;
@@ -1405,6 +1427,16 @@ static long ppc_del_hwdebug(struct task_
 		return -EINVAL;
 	if (child->thread.dabr == 0)
 		return -ENOENT;
+#ifdef CONFIG_HAVE_HW_BREAKPOINT
+	/*
+	 * There is no way by which address in ptrace_bps[0] and thread.dabr
+	 * can be different. So we don't explicitly check if they're the same
+	 */
+	if (child->thread.ptrace_bps[0]) {
+		unregister_hw_breakpoint(child->thread.ptrace_bps[0]);
+		child->thread.ptrace_bps[0] = NULL;
+	}
+#endif /* CONFIG_HAVE_HW_BREAKPOINT */
 
 	child->thread.dabr = 0;
 
#7 lore raw
Quoted 2 lines · K.Prasad

How about the revised patch below? It is only compile-tested; have you got a quick test case that I can run?

Try the watchpoint tests in gdb.

Andreas.

Signature · 3 lines
Andreas Schwab, [email protected]
GPG Key fingerprint = 58CA 54C7 6D53 942B 1756  01D3 44D5 214B 8276 4ED5
"And now for something completely different."
#8 lore raw
In reply to K.Prasad
Quoted 2 lines · K.Prasad

How about the revised patch below? It is only compile-tested; have you got a quick test case that I can run?

It crashes the kernel when running the watch-vfork test.

Andreas.

Signature · 3 lines
Andreas Schwab, [email protected]
GPG Key fingerprint = 58CA 54C7 6D53 942B 1756  01D3 44D5 214B 8276 4ED5
"And now for something completely different."
#9 lore raw
Quoted 9 lines · Andreas Schwab
Quoted 2 lines · K.Prasad

How about the revised patch below? It is only compile-tested; have you got a quick test case that I can run?

It crashes the kernel when running the watch-vfork test.

Andreas.

Hi Andreas,
I tried running it multiple times but saw no crash (or error
messages in dmesg). Can you send me the crash logs? What's the behaviour when the testcase is run on an unpatched kernel?

The watch-vfork test actually fails on my system (4 unexpected failures) irrespective of the kernel containing the patch or not.

Thanks,
K.Prasad
P.S.: I'd been on vacation and couldn't look at this issue during then.

#10 lore raw
Quoted 1 line · K.Prasad

The watch-vfork test actually fails on my system (4 unexpected failures)

It should pass all four tests. If gdb cannot even set a watchpoint it cannot trigger the crash, of course.

Andreas.

Signature · 3 lines
Andreas Schwab, [email protected]
GPG Key fingerprint = 58CA 54C7 6D53 942B 1756  01D3 44D5 214B 8276 4ED5
"And now for something completely different."