Skip to contentSkip to conversation

lkdtm: fix irq handler entry for arm64

PATCH Superseded AKASHI Takahiro 6 participants 13 messages Feb 1, 2018 – Feb 27, 2018

Subsystems (2)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

cover lkdtm: fix irq handler entry for arm64

Lkdtm is a handy way for testing kdump in case of panic in interrupt context:
echo PANIC > /sys/kernel/debug/provoke-crash/INT_HARDWARE_ENTRY

But this just doesn't work on arm64 since "do_IRQ" function doesn't exist on arm64. This patch series addresses this problem.

Patch#1 is a preliminary fix for patch#2 and allows for probing any exception entries, including irq handler.

For further improvements, some of other hooks doesn't work:
* handle_IRQ_event() was removed in v2.6.39
* hrtimer_start() was converted to an inline function in v4.2

but I don't know how we should deal with them.

AKASHI Takahiro (1):
lkdtm: fix irq handler entry for arm64

Masami Hiramatsu (1):
arm64: kprobes: Remove unneeded address sanity check

 arch/arm64/kernel/probes/kprobes.c |  8 --------
 drivers/misc/lkdtm_core.c          | 20 ++++++++++++++++++--
 2 files changed, 18 insertions(+), 10 deletions(-)
Signature · 1 line
2.15.1
#1 lore raw

PATCH 1/2 arm64: kprobes: Remove unneeded address sanity check

From: Masami Hiramatsu <[email protected]>

Remove unneeded address sanity check in arch_prepare_kprobe(). Since do_debug_exception() is already blacklisted for kprobes, no need to reject all __exception functions. Also, since generic kprobe framework already ensures the address is in kernel text, no need to check it is in rodata again.

---
 arch/arm64/kernel/probes/kprobes.c | 8 --------
 1 file changed, 8 deletions(-)
arch/arm64/kernel/probes/kprobes.c +0 -8
index d849d9804011..3c487a389252 100644
@@ -78,8 +78,6 @@ static void __kprobes arch_simulate_insn(struct kprobe *p, struct pt_regs *regs)
 int __kprobes arch_prepare_kprobe(struct kprobe *p)
 {
 	unsigned long probe_addr = (unsigned long)p->addr;
-	extern char __start_rodata[];
-	extern char __end_rodata[];
 
 	if (probe_addr & 0x3)
 		return -EINVAL;
@@ -87,12 +85,6 @@ int __kprobes arch_prepare_kprobe(struct kprobe *p)
 	/* copy instruction */
 	p->opcode = le32_to_cpu(*p->addr);
 
-	if (in_exception_text(probe_addr))
-		return -EINVAL;
-	if (probe_addr >= (unsigned long) __start_rodata &&
-	    probe_addr <= (unsigned long) __end_rodata)
-		return -EINVAL;
-
 	/* decode instruction */
 	switch (arm_kprobe_decode_insn(p->addr, &p->ainsn)) {
 	case INSN_REJECTED:	/* insn not supported */
Signature · 1 line
2.15.1
#2 lore raw

PATCH 1/2 arm64: kprobes: Remove unneeded address sanity check

Quoted 28 lines · AKASHI Takahiro

From: Masami Hiramatsu <[email protected]>

Remove unneeded address sanity check in arch_prepare_kprobe(). Since do_debug_exception() is already blacklisted for kprobes, no need to reject all __exception functions. Also, since generic kprobe framework already ensures the address is in kernel text, no need to check it is in rodata again.

---
 arch/arm64/kernel/probes/kprobes.c | 8 --------
 1 file changed, 8 deletions(-)
arch/arm64/kernel/probes/kprobes.c +0 -2
index d849d9804011..3c487a389252 100644
@@ -78,8 +78,6 @@ static void __kprobes arch_simulate_insn(struct kprobe *p, struct pt_regs *regs)
 int __kprobes arch_prepare_kprobe(struct kprobe *p)
 {
 	unsigned long probe_addr = (unsigned long)p->addr;
-	extern char __start_rodata[];
-	extern char __end_rodata[];
 
 	if (probe_addr & 0x3)
 		return -EINVAL;

arch/arm64/kernel/probes/kprobes.c int __kprobes arch_prepare_kprobe(struct kprobe *p) · line 95

Quoted 9 lines
arch/arm64/kernel/probes/kprobes.c +0 -6
@@ -87,12 +85,6 @@ int __kprobes arch_prepare_kprobe(struct kprobe *p)
 	/* copy instruction */
 	p->opcode = le32_to_cpu(*p->addr);
 
-	if (in_exception_text(probe_addr))
-		return -EINVAL;
-	if (probe_addr >= (unsigned long) __start_rodata &&
-	    probe_addr <= (unsigned long) __end_rodata)
-		return -EINVAL;
-

Does this mean we can remove in_exception_text and the corresponding __exception annotations from arm64?

Will

#3 lore raw

PATCH 1/2 arm64: kprobes: Remove unneeded address sanity check

Hi Will,

2018-02-06 23:36 GMT+09:00 Will Deacon <[email protected]>:

Quoted 32 lines · AKASHI Takahiro
Quoted 31 lines

From: Masami Hiramatsu <[email protected]>

Remove unneeded address sanity check in arch_prepare_kprobe(). Since do_debug_exception() is already blacklisted for kprobes, no need to reject all __exception functions. Also, since generic kprobe framework already ensures the address is in kernel text, no need to check it is in rodata again.

---
 arch/arm64/kernel/probes/kprobes.c | 8 --------
 1 file changed, 8 deletions(-)
arch/arm64/kernel/probes/kprobes.c +0 -2
index d849d9804011..3c487a389252 100644
@@ -78,8 +78,6 @@ static void __kprobes arch_simulate_insn(struct kprobe *p, struct pt_regs *regs)
 int __kprobes arch_prepare_kprobe(struct kprobe *p)
 {
      unsigned long probe_addr = (unsigned long)p->addr;
-     extern char __start_rodata[];
-     extern char __end_rodata[];
 
      if (probe_addr & 0x3)
              return -EINVAL;
@@ -87,12 +85,6 @@ int __kprobes arch_prepare_kprobe(struct kprobe *p)
      /* copy instruction */
      p->opcode = le32_to_cpu(*p->addr);
 
Quoted 9 lines

arch/arm64/kernel/probes/kprobes.c int __kprobes arch_prepare_kprobe(struct kprobe *p) · line 95

Quoted 6 lines
arch/arm64/kernel/probes/kprobes.c +0 -6
-     if (in_exception_text(probe_addr))
-             return -EINVAL;
-     if (probe_addr >= (unsigned long) __start_rodata &&
-         probe_addr <= (unsigned long) __end_rodata)
-             return -EINVAL;
-

Does this mean we can remove in_exception_text and the corresponding __exception annotations from arm64?

Good catch. As far as I can see, there is no other in_exception_text() user under arch/arm64,
so we can remove it all.

Thank you,

Signature · 1 line
Masami Hiramatsu
#4 lore raw

PATCH 1/2 arm64: kprobes: Remove unneeded address sanity check

In reply to AKASHI Takahiro
Quoted 31 lines · AKASHI Takahiro

From: Masami Hiramatsu <[email protected]>

Remove unneeded address sanity check in arch_prepare_kprobe(). Since do_debug_exception() is already blacklisted for kprobes, no need to reject all __exception functions. Also, since generic kprobe framework already ensures the address is in kernel text, no need to check it is in rodata again.

---
  arch/arm64/kernel/probes/kprobes.c | 8 --------
  1 file changed, 8 deletions(-)
arch/arm64/kernel/probes/kprobes.c +0 -2
index d849d9804011..3c487a389252 100644
@@ -78,8 +78,6 @@ static void __kprobes arch_simulate_insn(struct kprobe *p, struct pt_regs *regs)
  int __kprobes arch_prepare_kprobe(struct kprobe *p)
  {
  	unsigned long probe_addr = (unsigned long)p->addr;
-	extern char __start_rodata[];
-	extern char __end_rodata[];
 
  	if (probe_addr & 0x3)
  		return -EINVAL;
@@ -87,12 +85,6 @@ int __kprobes arch_prepare_kprobe(struct kprobe *p)
  	/* copy instruction */
  	p->opcode = le32_to_cpu(*p->addr);

arch/arm64/kernel/probes/kprobes.c int __kprobes arch_prepare_kprobe(struct kprobe *p) · line 90

Quoted 10 lines
arch/arm64/kernel/probes/kprobes.c +0 -6
 
-	if (in_exception_text(probe_addr))
-		return -EINVAL;
-	if (probe_addr >= (unsigned long) __start_rodata &&
-	    probe_addr <= (unsigned long) __end_rodata)
-		return -EINVAL;
-
  	/* decode instruction */
  	switch (arm_kprobe_decode_insn(p->addr, &p->ainsn)) {
  	case INSN_REJECTED:	/* insn not supported */

I have tested this change on v4.15 using kprobes events and I find it allows kprobes to be placed in exception text when they were previously rejected. Is there some other recent change I need to test this with for the previous behavior to be preserved?

-dl

#5 lore raw

PATCH 1/2 arm64: kprobes: Remove unneeded address sanity check

Hi David,

Quoted 32 lines · On Wed, 14 Feb 2018 21:08:03 -0500 David Long <[email protected]> wrote:
Quoted 31 lines · AKASHI Takahiro

From: Masami Hiramatsu <[email protected]>

Remove unneeded address sanity check in arch_prepare_kprobe(). Since do_debug_exception() is already blacklisted for kprobes, no need to reject all __exception functions. Also, since generic kprobe framework already ensures the address is in kernel text, no need to check it is in rodata again.

---
  arch/arm64/kernel/probes/kprobes.c | 8 --------
  1 file changed, 8 deletions(-)
arch/arm64/kernel/probes/kprobes.c +0 -2
index d849d9804011..3c487a389252 100644
@@ -78,8 +78,6 @@ static void __kprobes arch_simulate_insn(struct kprobe *p, struct pt_regs *regs)
  int __kprobes arch_prepare_kprobe(struct kprobe *p)
  {
  	unsigned long probe_addr = (unsigned long)p->addr;
-	extern char __start_rodata[];
-	extern char __end_rodata[];
 
  	if (probe_addr & 0x3)
  		return -EINVAL;
@@ -87,12 +85,6 @@ int __kprobes arch_prepare_kprobe(struct kprobe *p)
  	/* copy instruction */
  	p->opcode = le32_to_cpu(*p->addr);
 
Quoted 15 lines

arch/arm64/kernel/probes/kprobes.c int __kprobes arch_prepare_kprobe(struct kprobe *p) · line 90

Quoted 10 lines
arch/arm64/kernel/probes/kprobes.c +0 -6
-	if (in_exception_text(probe_addr))
-		return -EINVAL;
-	if (probe_addr >= (unsigned long) __start_rodata &&
-	    probe_addr <= (unsigned long) __end_rodata)
-		return -EINVAL;
-
  	/* decode instruction */
  	switch (arm_kprobe_decode_insn(p->addr, &p->ainsn)) {
  	case INSN_REJECTED:	/* insn not supported */

I have tested this change on v4.15 using kprobes events and I find it allows kprobes to be placed in exception text when they were previously rejected. Is there some other recent change I need to test this with for the previous behavior to be preserved?

Hmm, the latest change is to avoid retpoline thunk functions on x86. Since the retpoline may not be applied on aarch64, it can be ignored. However, I found there were still many "__kprobes" tags under arch/arm64. That was replaced with NOKPROBE_SYMBOL() (and nokprobe_inline for inline function). It should be done on arm/arm64 too because the functions marked with NOKPROBE_SYMBOL are listed in <debugfs>/kprobes/blacklist.

Thank you,

Quoted 3 lines

-dl

Signature · 1 line
Masami Hiramatsu <[email protected]>
#6 lore raw

PATCH 1/2 arm64: kprobes: Remove unneeded address sanity check

Quoted 53 lines · Masami Hiramatsu

Hi David,

Quoted 32 lines · On Wed, 14 Feb 2018 21:08:03 -0500 David Long <[email protected]> wrote:
Quoted 31 lines · AKASHI Takahiro

From: Masami Hiramatsu <[email protected]>

Remove unneeded address sanity check in arch_prepare_kprobe(). Since do_debug_exception() is already blacklisted for kprobes, no need to reject all __exception functions. Also, since generic kprobe framework already ensures the address is in kernel text, no need to check it is in rodata again.

---
   arch/arm64/kernel/probes/kprobes.c | 8 --------
   1 file changed, 8 deletions(-)
arch/arm64/kernel/probes/kprobes.c +0 -2
index d849d9804011..3c487a389252 100644
@@ -78,8 +78,6 @@ static void __kprobes arch_simulate_insn(struct kprobe *p, struct pt_regs *regs)
   int __kprobes arch_prepare_kprobe(struct kprobe *p)
   {
   	unsigned long probe_addr = (unsigned long)p->addr;
-	extern char __start_rodata[];
-	extern char __end_rodata[];
 
   	if (probe_addr & 0x3)
   		return -EINVAL;
@@ -87,12 +85,6 @@ int __kprobes arch_prepare_kprobe(struct kprobe *p)
   	/* copy instruction */
   	p->opcode = le32_to_cpu(*p->addr);
 
Quoted 15 lines

arch/arm64/kernel/probes/kprobes.c int __kprobes arch_prepare_kprobe(struct kprobe *p) · line 90

Quoted 10 lines
arch/arm64/kernel/probes/kprobes.c +0 -6
-	if (in_exception_text(probe_addr))
-		return -EINVAL;
-	if (probe_addr >= (unsigned long) __start_rodata &&
-	    probe_addr <= (unsigned long) __end_rodata)
-		return -EINVAL;
-
   	/* decode instruction */
   	switch (arm_kprobe_decode_insn(p->addr, &p->ainsn)) {
   	case INSN_REJECTED:	/* insn not supported */

I have tested this change on v4.15 using kprobes events and I find it allows kprobes to be placed in exception text when they were previously rejected. Is there some other recent change I need to test this with for the previous behavior to be preserved?

Quoted 6 lines

Hmm, the latest change is to avoid retpoline thunk functions on x86. Since the retpoline may not be applied on aarch64, it can be ignored. However, I found there were still many "__kprobes" tags under arch/arm64. That was replaced with NOKPROBE_SYMBOL() (and nokprobe_inline for inline function). It should be done on arm/arm64 too because the functions marked with NOKPROBE_SYMBOL are listed in <debugfs>/kprobes/blacklist.

My bad for not reading the whole patch set before commenting. I understand the goal now.

I see NOKPROBE_SYMBOL is only used for a few architectures so far, with arm64 widely using both methods. I'm presuming this is work in progress.

I verified do_debug_exception is still rejected by kprobes. The other global functions in there are accepted after the change. Do we think that's safe? I can't immediately come up with a reason it wouldn't be. Has it been tested, beyond the IRQ stuff?

I remember adding the rodata test. Seems to me there was a reason for that at the time, but I've verified that probes in rodata are still rejected after the patch.

-dl

#7 lore raw

PATCH 1/2 arm64: kprobes: Remove unneeded address sanity check

Hi David,

2018-02-22 14:19 GMT+09:00 David Long <[email protected]>:

Quoted 73 lines · Masami Hiramatsu
Quoted 58 lines

Hi David,

Quoted 35 lines · On Wed, 14 Feb 2018 21:08:03 -0500 David Long <[email protected]> wrote:
Quoted 34 lines · AKASHI Takahiro

From: Masami Hiramatsu <[email protected]>

Remove unneeded address sanity check in arch_prepare_kprobe(). Since do_debug_exception() is already blacklisted for kprobes, no need to reject all __exception functions. Also, since generic kprobe framework already ensures the address is in kernel text, no need to check it is in rodata again.

---
   arch/arm64/kernel/probes/kprobes.c | 8 --------
   1 file changed, 8 deletions(-)
arch/arm64/kernel/probes/kprobes.c +0 -2
index d849d9804011..3c487a389252 100644
@@ -78,8 +78,6 @@ static void __kprobes arch_simulate_insn(struct kprobe
 *p, struct pt_regs *regs)
   int __kprobes arch_prepare_kprobe(struct kprobe *p)
   {
        unsigned long probe_addr = (unsigned long)p->addr;
-       extern char __start_rodata[];
-       extern char __end_rodata[];
 
        if (probe_addr & 0x3)
                return -EINVAL;
arch/arm64/kernel/probes/kprobes.c +0 -0
@@ -87,12 +85,6 @@ int __kprobes arch_prepare_kprobe(struct kprobe *p)
        /* copy instruction */
        p->opcode = le32_to_cpu(*p->addr);
 
Quoted 15 lines

arch/arm64/kernel/probes/kprobes.c int __kprobes arch_prepare_kprobe(struct kprobe *p) · line 90

Quoted 10 lines
arch/arm64/kernel/probes/kprobes.c +0 -6
-       if (in_exception_text(probe_addr))
-               return -EINVAL;
-       if (probe_addr >= (unsigned long) __start_rodata &&
-           probe_addr <= (unsigned long) __end_rodata)
-               return -EINVAL;
-
        /* decode instruction */
        switch (arm_kprobe_decode_insn(p->addr, &p->ainsn)) {
        case INSN_REJECTED:     /* insn not supported */

I have tested this change on v4.15 using kprobes events and I find it allows kprobes to be placed in exception text when they were previously rejected. Is there some other recent change I need to test this with for the previous behavior to be preserved?

Quoted 9 lines

Hmm, the latest change is to avoid retpoline thunk functions on x86. Since the
retpoline may not be applied on aarch64, it can be ignored. However, I found there were still many "__kprobes" tags under arch/arm64. That
was replaced with NOKPROBE_SYMBOL() (and nokprobe_inline for inline function).
It should be done on arm/arm64 too because the functions marked with NOKPROBE_SYMBOL are listed in <debugfs>/kprobes/blacklist.

My bad for not reading the whole patch set before commenting. I understand the goal now.

Quoted 7 lines

I see NOKPROBE_SYMBOL is only used for a few architectures so far, with arm64 widely using both methods. I'm presuming this is work in progress.

I verified do_debug_exception is still rejected by kprobes. The other global functions in there are accepted after the change. Do we think that's safe? I can't immediately come up with a reason it wouldn't be. Has it been tested, beyond the IRQ stuff?

Yeah, it is reasonable concern. We already have ftrace/kprobe interface in debugfs (tracefs)
so we can start testing it now with, something like below shell script;

n=0
cd /sys/kernel/debug/tracing
cut -f 3 -d " " /proc/kallsyms | while read sym; do
echo "probing $sym"
echo "p $sym" >> kprobe_events
n=$((n+1))
[ $n -gt 3000 ] && break
done

I think it could be enough to test first 2-3k syms since most of arch dependent code are placed there.

See my old slide.
https://events.static.linuxfound.org/sites/events/files/slides/Handling%20the%20Massive%20Multiple%20Kprobes%20v2_1.pdf

Unfortunately, kpcache and hash-table expansion (most critical features) were not accepted, so take care of performance degradation.

Quoted 3 lines

I remember adding the rodata test. Seems to me there was a reason for that at the time, but I've verified that probes in rodata are still rejected after the patch.

Yes, because kernel/kprobes.c accepts the probes probing the kernel .text.

Thank you,

Signature · 1 line
Masami Hiramatsu
#8 lore raw

PATCH 2/2 lkdtm: fix irq handler entry for arm64

In reply to AKASHI Takahiro

Arm64 doesn't have "do_IRQ" function, instead *handle_arch_irq, which is initialized by irq chip (gic), is called from exception entry. This patch fixes this problem.

---
 drivers/misc/lkdtm_core.c | 20 ++++++++++++++++++--
 1 file changed, 18 insertions(+), 2 deletions(-)
drivers/misc/lkdtm_core.c +18 -2
index ba92291508dc..e20343543053 100644
@@ -249,13 +249,29 @@ static int lkdtm_register_cpoint(struct crashpoint *crashpoint,
 	if (lkdtm_kprobe != NULL)
 		unregister_kprobe(lkdtm_kprobe);
 
+	if (IS_ENABLED(CONFIG_ARM64) &&
+			!strcmp(crashpoint->name, "INT_HARDWARE_ENTRY")) {
+		extern void (*handle_arch_irq)(struct pt_regs *regs);
+
+		crashpoint->kprobe.addr = (kprobe_opcode_t *)*handle_arch_irq;
+		/*
+		 * Instantiating kprobe.symbol_name here, say
+		 * with lookup_symbol_name(*handle_arch_irq),
+		 * would cause register_kprobe() to fail.
+		 */
+		crashpoint->kprobe.symbol_name = NULL;
+	}
 	lkdtm_crashpoint = crashpoint;
 	lkdtm_crashtype = crashtype;
 	lkdtm_kprobe = &crashpoint->kprobe;
 	ret = register_kprobe(lkdtm_kprobe);
 	if (ret < 0) {
-		pr_info("Couldn't register kprobe %s\n",
-			crashpoint->kprobe.symbol_name);
+		if (IS_ENABLED(CONFIG_ARM64))
+			pr_info("Couldn't register kprobe 0x%lx\n",
+				(unsigned long)crashpoint->kprobe.addr);
+		else
+			pr_info("Couldn't register kprobe %s\n",
+				crashpoint->kprobe.symbol_name);
 		lkdtm_kprobe = NULL;
 		lkdtm_crashpoint = NULL;
 		lkdtm_crashtype = NULL;
Signature · 1 line
2.15.1
#9 lore raw

PATCH 2/2 lkdtm: fix irq handler entry for arm64

Quoted 3 lines · AKASHI Takahiro

Arm64 doesn't have "do_IRQ" function, instead *handle_arch_irq, which is initialized by irq chip (gic), is called from exception entry. This patch fixes this problem.

As in, this symbol is not known a lkdtm setup time? Hm, seems like we'd want a more generalized approach here.

Quoted 11 lines
---
 drivers/misc/lkdtm_core.c | 20 ++++++++++++++++++--
 1 file changed, 18 insertions(+), 2 deletions(-)
drivers/misc/lkdtm_core.c +0 -0
index ba92291508dc..e20343543053 100644

drivers/misc/lkdtm_core.c static int lkdtm_register_cpoint(struct crashpoint *crashpoint, · line 254

Quoted 6 lines
drivers/misc/lkdtm_core.c +3 -0
@@ -249,13 +249,29 @@ static int lkdtm_register_cpoint(struct crashpoint *crashpoint,
        if (lkdtm_kprobe != NULL)
                unregister_kprobe(lkdtm_kprobe);
 
+       if (IS_ENABLED(CONFIG_ARM64) &&
+                       !strcmp(crashpoint->name, "INT_HARDWARE_ENTRY")) {
+               extern void (*handle_arch_irq)(struct pt_regs *regs);

I don't like this extern -- can handle_arch_irq be properly exported somewhere?

drivers/misc/lkdtm_core.c static int lkdtm_register_cpoint(struct crashpoint *crashpoint, · line 256

Quoted 1 line
drivers/misc/lkdtm_core.c +1 -0
+               crashpoint->kprobe.addr = (kprobe_opcode_t *)*handle_arch_irq;

I don't think the * is needed here: it's already a function pointer.

drivers/misc/lkdtm_core.c static int lkdtm_register_cpoint(struct crashpoint *crashpoint, · line 262

Quoted 6 lines
drivers/misc/lkdtm_core.c +6 -0
+               /*
+                * Instantiating kprobe.symbol_name here, say
+                * with lookup_symbol_name(*handle_arch_irq),
+                * would cause register_kprobe() to fail.
+                */
+               crashpoint->kprobe.symbol_name = NULL;

Is kprobe.addr sufficient for register_kprobe?

Quoted 7 lines
drivers/misc/lkdtm_core.c +1 -1
+       }
        lkdtm_crashpoint = crashpoint;
        lkdtm_crashtype = crashtype;
        lkdtm_kprobe = &crashpoint->kprobe;
        ret = register_kprobe(lkdtm_kprobe);
        if (ret < 0) {
-               pr_info("Couldn't register kprobe %s\n",

drivers/misc/lkdtm_core.c static int lkdtm_register_cpoint(struct crashpoint *crashpoint, · line 277

Quoted 10 lines
drivers/misc/lkdtm_core.c +6 -1
-                       crashpoint->kprobe.symbol_name);
+               if (IS_ENABLED(CONFIG_ARM64))
+                       pr_info("Couldn't register kprobe 0x%lx\n",
+                               (unsigned long)crashpoint->kprobe.addr);
+               else
+                       pr_info("Couldn't register kprobe %s\n",
+                               crashpoint->kprobe.symbol_name);
                lkdtm_kprobe = NULL;
                lkdtm_crashpoint = NULL;
                lkdtm_crashtype = NULL;

So I can replicate, how did you test this?

-Kees

Signature · 2 lines
Kees Cook
Pixel Security
#10 lore raw

PATCH 2/2 lkdtm: fix irq handler entry for arm64

Hi,

2018-02-27 12:57 GMT+09:00 Kees Cook <[email protected]>:

Quoted 29 lines
Quoted 3 lines · AKASHI Takahiro

Arm64 doesn't have "do_IRQ" function, instead *handle_arch_irq, which is initialized by irq chip (gic), is called from exception entry. This patch fixes this problem.

As in, this symbol is not known a lkdtm setup time? Hm, seems like we'd want a more generalized approach here.

Quoted 11 lines
---
 drivers/misc/lkdtm_core.c | 20 ++++++++++++++++++--
 1 file changed, 18 insertions(+), 2 deletions(-)
drivers/misc/lkdtm_core.c +0 -0
index ba92291508dc..e20343543053 100644

drivers/misc/lkdtm_core.c static int lkdtm_register_cpoint(struct crashpoint *crashpoint, · line 254

Quoted 6 lines
drivers/misc/lkdtm_core.c +3 -0
@@ -249,13 +249,29 @@ static int lkdtm_register_cpoint(struct crashpoint *crashpoint,
        if (lkdtm_kprobe != NULL)
                unregister_kprobe(lkdtm_kprobe);
 
+       if (IS_ENABLED(CONFIG_ARM64) &&
+                       !strcmp(crashpoint->name, "INT_HARDWARE_ENTRY")) {
+               extern void (*handle_arch_irq)(struct pt_regs *regs);

I don't like this extern -- can handle_arch_irq be properly exported somewhere?

Quoted 3 lines

drivers/misc/lkdtm_core.c static int lkdtm_register_cpoint(struct crashpoint *crashpoint, · line 256

Quoted 1 line
drivers/misc/lkdtm_core.c +1 -0
+               crashpoint->kprobe.addr = (kprobe_opcode_t *)*handle_arch_irq;

I don't think the * is needed here: it's already a function pointer.

Since the addr is no void *, gcc warns this assignment from incompatible pointer type.
Hmm, maybe better casting it to void *.

Quoted 9 lines

drivers/misc/lkdtm_core.c static int lkdtm_register_cpoint(struct crashpoint *crashpoint, · line 262

Quoted 7 lines
drivers/misc/lkdtm_core.c +6 -0
+               /*
+                * Instantiating kprobe.symbol_name here, say
+                * with lookup_symbol_name(*handle_arch_irq),
+                * would cause register_kprobe() to fail.
+                */
+               crashpoint->kprobe.symbol_name = NULL;

Is kprobe.addr sufficient for register_kprobe?

Yes, if symbol_name is NULL, register_kprobe uses only kprobe.addr to find the probe point.

Thank you,

Signature · 1 line
Masami Hiramatsu
#11 lore raw

PATCH 2/2 lkdtm: fix irq handler entry for arm64

In reply to Kees Cook

Hi Kees,

Quoted 8 lines · Kees Cook
Quoted 3 lines · AKASHI Takahiro

Arm64 doesn't have "do_IRQ" function, instead *handle_arch_irq, which is initialized by irq chip (gic), is called from exception entry. This patch fixes this problem.

As in, this symbol is not known a lkdtm setup time? Hm, seems like we'd want a more generalized approach here.

Hmm. See my comments below.

Quoted 14 lines
Quoted 14 lines
---
 drivers/misc/lkdtm_core.c | 20 ++++++++++++++++++--
 1 file changed, 18 insertions(+), 2 deletions(-)
drivers/misc/lkdtm_core.c +0 -0
index ba92291508dc..e20343543053 100644
@@ -249,13 +249,29 @@ static int lkdtm_register_cpoint(struct crashpoint *crashpoint,
        if (lkdtm_kprobe != NULL)
                unregister_kprobe(lkdtm_kprobe);
 
Quoted 5 lines

drivers/misc/lkdtm_core.c static int lkdtm_register_cpoint(struct crashpoint *crashpoint, · line 254

Quoted 3 lines
drivers/misc/lkdtm_core.c +3 -0
+       if (IS_ENABLED(CONFIG_ARM64) &&
+                       !strcmp(crashpoint->name, "INT_HARDWARE_ENTRY")) {
+               extern void (*handle_arch_irq)(struct pt_regs *regs);

I don't like this extern -- can handle_arch_irq be properly exported somewhere?

Define a weak function, get_handle_irq(), in linux/irq.h and a real one in arch code. Then

        if (!kallsyms_lookup_name(crashpoint->symbol_name)) {
                if (!strcmp(crashpoint->name, "INT_HARDWARE_ENTRY")) {
                        func = get_handle_irq();
                        if (func) {
                                crashpoint->kprobe.addr = func;
                                crashpoint->kprobe.symbol_name = NULL;
                        } else {
                                /* error */
                        }
                } /* anything else? */
        }

Do you like this code better?

Quoted 4 lines

drivers/misc/lkdtm_core.c static int lkdtm_register_cpoint(struct crashpoint *crashpoint, · line 256

Quoted 2 lines
drivers/misc/lkdtm_core.c +1 -0
+               crashpoint->kprobe.addr = (kprobe_opcode_t *)*handle_arch_irq;

I don't think the * is needed here: it's already a function pointer.

Will check.

Quoted 8 lines

drivers/misc/lkdtm_core.c static int lkdtm_register_cpoint(struct crashpoint *crashpoint, · line 262

Quoted 6 lines
drivers/misc/lkdtm_core.c +6 -0
+               /*
+                * Instantiating kprobe.symbol_name here, say
+                * with lookup_symbol_name(*handle_arch_irq),
+                * would cause register_kprobe() to fail.
+                */
+               crashpoint->kprobe.symbol_name = NULL;

Is kprobe.addr sufficient for register_kprobe?

Yes as Masami explained.
Leaving symbol_name ends up failure of register_kprobe().

Quoted 18 lines
Quoted 7 lines
drivers/misc/lkdtm_core.c +1 -1
+       }
        lkdtm_crashpoint = crashpoint;
        lkdtm_crashtype = crashtype;
        lkdtm_kprobe = &crashpoint->kprobe;
        ret = register_kprobe(lkdtm_kprobe);
        if (ret < 0) {
-               pr_info("Couldn't register kprobe %s\n",

drivers/misc/lkdtm_core.c static int lkdtm_register_cpoint(struct crashpoint *crashpoint, · line 277

Quoted 10 lines
drivers/misc/lkdtm_core.c +6 -1
-                       crashpoint->kprobe.symbol_name);
+               if (IS_ENABLED(CONFIG_ARM64))
+                       pr_info("Couldn't register kprobe 0x%lx\n",
+                               (unsigned long)crashpoint->kprobe.addr);
+               else
+                       pr_info("Couldn't register kprobe %s\n",
+                               crashpoint->kprobe.symbol_name);
                lkdtm_kprobe = NULL;
                lkdtm_crashpoint = NULL;
                lkdtm_crashtype = NULL;
Quoted 1 line

So I can replicate, how did you test this?

All what I did in my arm64 test is

# echo PANIC > /sys/kernel/debug/provoke-crash/INT_HARDWARE_ENTRY

The probe point will hit sooner or later and we will see a panic (and kdump kicks in).

Thanks,
-Takahiro AKASHI

Quoted 4 lines
Signature · 3 lines
-- 
Kees Cook
Pixel Security
#12 lore raw

PATCH 2/2 lkdtm: fix irq handler entry for arm64

Quoted 49 lines · AKASHI Takahiro

Hi Kees,

Quoted 8 lines · Kees Cook
Quoted 3 lines · AKASHI Takahiro

Arm64 doesn't have "do_IRQ" function, instead *handle_arch_irq, which is initialized by irq chip (gic), is called from exception entry. This patch fixes this problem.

As in, this symbol is not known a lkdtm setup time? Hm, seems like we'd want a more generalized approach here.

Hmm. See my comments below.

Quoted 14 lines
Quoted 14 lines
---
 drivers/misc/lkdtm_core.c | 20 ++++++++++++++++++--
 1 file changed, 18 insertions(+), 2 deletions(-)
drivers/misc/lkdtm_core.c +0 -0
index ba92291508dc..e20343543053 100644
@@ -249,13 +249,29 @@ static int lkdtm_register_cpoint(struct crashpoint *crashpoint,
        if (lkdtm_kprobe != NULL)
                unregister_kprobe(lkdtm_kprobe);
 
Quoted 5 lines

drivers/misc/lkdtm_core.c static int lkdtm_register_cpoint(struct crashpoint *crashpoint, · line 254

Quoted 3 lines
drivers/misc/lkdtm_core.c +3 -0
+       if (IS_ENABLED(CONFIG_ARM64) &&
+                       !strcmp(crashpoint->name, "INT_HARDWARE_ENTRY")) {
+               extern void (*handle_arch_irq)(struct pt_regs *regs);

I don't like this extern -- can handle_arch_irq be properly exported somewhere?

Define a weak function, get_handle_irq(), in linux/irq.h and a real one in arch code. Then

        if (!kallsyms_lookup_name(crashpoint->symbol_name)) {
                if (!strcmp(crashpoint->name, "INT_HARDWARE_ENTRY")) {
                        func = get_handle_irq();
                        if (func) {
                                crashpoint->kprobe.addr = func;
                                crashpoint->kprobe.symbol_name = NULL;
                        } else {
                                /* error */
                        }
                } /* anything else? */
        }
Quoted 1 line

Do you like this code better?

Yeah, this is more generalized; thanks! If we end up with other late-defined functions we can further generalize this, but this is fine for our one case. :)

Quoted 38 lines
Quoted 5 lines

drivers/misc/lkdtm_core.c static int lkdtm_register_cpoint(struct crashpoint *crashpoint, · line 256

Quoted 3 lines
drivers/misc/lkdtm_core.c +1 -0
+               crashpoint->kprobe.addr = (kprobe_opcode_t *)*handle_arch_irq;

I don't think the * is needed here: it's already a function pointer.

Will check.

Quoted 8 lines

drivers/misc/lkdtm_core.c static int lkdtm_register_cpoint(struct crashpoint *crashpoint, · line 262

Quoted 6 lines
drivers/misc/lkdtm_core.c +6 -0
+               /*
+                * Instantiating kprobe.symbol_name here, say
+                * with lookup_symbol_name(*handle_arch_irq),
+                * would cause register_kprobe() to fail.
+                */
+               crashpoint->kprobe.symbol_name = NULL;

Is kprobe.addr sufficient for register_kprobe?

Yes as Masami explained.
Leaving symbol_name ends up failure of register_kprobe().

Quoted 18 lines
Quoted 17 lines
drivers/misc/lkdtm_core.c +7 -2
+       }
        lkdtm_crashpoint = crashpoint;
        lkdtm_crashtype = crashtype;
        lkdtm_kprobe = &crashpoint->kprobe;
        ret = register_kprobe(lkdtm_kprobe);
        if (ret < 0) {
-               pr_info("Couldn't register kprobe %s\n",
-                       crashpoint->kprobe.symbol_name);
+               if (IS_ENABLED(CONFIG_ARM64))
+                       pr_info("Couldn't register kprobe 0x%lx\n",
+                               (unsigned long)crashpoint->kprobe.addr);
+               else
+                       pr_info("Couldn't register kprobe %s\n",
+                               crashpoint->kprobe.symbol_name);
                lkdtm_kprobe = NULL;
                lkdtm_crashpoint = NULL;
                lkdtm_crashtype = NULL;
Quoted 7 lines
Quoted 1 line

So I can replicate, how did you test this?

All what I did in my arm64 test is

# echo PANIC > /sys/kernel/debug/provoke-crash/INT_HARDWARE_ENTRY

The probe point will hit sooner or later and we will see a panic (and kdump kicks in).

Great, thanks!

-Kees

Signature · 2 lines
Kees Cook
Pixel Security