Skip to contentSkip to conversation

arm64: implement KPROBES_ON_FTRACE

PATCH Superseded Jisheng Zhang 2 participants 6 messages Aug 22, 2019

Subsystems (3)

v4 of 7 Superseded Latest is v7 →

lore
Share & tools
lore
#0 lore raw

PATCH v4 1/1 arm64: implement KPROBES_ON_FTRACE

KPROBES_ON_FTRACE avoids much of the overhead with regular kprobes as it eliminates the need for a trap, as well as the need to emulate or single-step instructions.

Tested on berlin arm64 platform.

~ # mount -t debugfs debugfs /sys/kernel/debug/
~ # cd /sys/kernel/debug/
/sys/kernel/debug # echo 'p _do_fork' > tracing/kprobe_events

before the patch:

/sys/kernel/debug # cat kprobes/list
ffffff801009fe28 k _do_fork+0x0 [DISABLED]

after the patch:

/sys/kernel/debug # cat kprobes/list
ffffff801009ff54 k _do_fork+0x4 [DISABLED][FTRACE]

---
KPROBES_ON_FTRACE avoids much of the overhead with regular kprobes as it eliminates the need for a trap, as well as the need to emulate or single-step instructions.

Applied after arm64 FTRACE_WITH_REGS:
http://lists.infradead.org/pipermail/linux-arm-kernel/2019-August/674404.html

Changes since v3:
- move kprobe_lookup_name() and arch_kprobe_on_func_entry to ftrace.c since we only want to choose the ftrace entry for KPROBES_ON_FTRACE.
- only choose ftrace entry if (addr && !offset)

Changes since v2:
- remove patch1, make it a single cleanup patch
- remove "This patch" in the change log
- implement arm64's kprobe_lookup_name() and arch_kprobe_on_func_entry instead of patching the common kprobes code

Changes since v1:
- make the kprobes/x86: use instruction_pointer and instruction_pointer_set as patch1
- add Masami's ACK to patch1
- add some description about KPROBES_ON_FTRACE and why we need it on

    arm64

- correct the log before the patch
- remove the consolidation patch, make it as TODO
- only adjust kprobe's addr when KPROBE_FLAG_FTRACE is set
- if KPROBES_ON_FTRACE, ftrace_call_adjust() the kprobe's addr before calling ftrace_location()
- update the kprobes-on-ftrace/arch-support.txt in doc

 .../debug/kprobes-on-ftrace/arch-support.txt  |  2 +-
 arch/arm64/Kconfig                            |  1 +
 arch/arm64/kernel/probes/Makefile             |  1 +
 arch/arm64/kernel/probes/ftrace.c             | 84 +++++++++++++++++++
 4 files changed, 87 insertions(+), 1 deletion(-)
 create mode 100644 arch/arm64/kernel/probes/ftrace.c
Documentation/features/debug/kprobes-on-ftrace/arch-support.txt +1 -1
index 68f266944d5f..e8358a38981c 100644
@@ -9,7 +9,7 @@
     |       alpha: | TODO |
     |         arc: | TODO |
     |         arm: | TODO |
-    |       arm64: | TODO |
+    |       arm64: |  ok  |
     |         c6x: | TODO |
     |        csky: | TODO |
     |       h8300: | TODO |
arch/arm64/Kconfig +1 -0
index 663392d1eae2..928700f15e23 100644
@@ -167,6 +167,7 @@ config ARM64
 	select HAVE_STACKPROTECTOR
 	select HAVE_SYSCALL_TRACEPOINTS
 	select HAVE_KPROBES
+	select HAVE_KPROBES_ON_FTRACE
 	select HAVE_KRETPROBES
 	select HAVE_GENERIC_VDSO
 	select IOMMU_DMA if IOMMU_SUPPORT
arch/arm64/kernel/probes/Makefile +1 -0
index 8e4be92e25b1..4020cfc66564 100644
@@ -4,3 +4,4 @@ obj-$(CONFIG_KPROBES)		+= kprobes.o decode-insn.o	\
 				   simulate-insn.o
 obj-$(CONFIG_UPROBES)		+= uprobes.o decode-insn.o	\
 				   simulate-insn.o
+obj-$(CONFIG_KPROBES_ON_FTRACE)	+= ftrace.o
arch/arm64/kernel/probes/ftrace.c new file +84 -0

84 lines not on this page · raw message

Signature · 7 lines
2.23.0.rc1


_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Reviewed-by on v4 1/1
#1 lore raw
Quoted 15 lines · Jisheng Zhang

KPROBES_ON_FTRACE avoids much of the overhead with regular kprobes as it eliminates the need for a trap, as well as the need to emulate or single-step instructions.

Tested on berlin arm64 platform.

~ # mount -t debugfs debugfs /sys/kernel/debug/
~ # cd /sys/kernel/debug/
/sys/kernel/debug # echo 'p _do_fork' > tracing/kprobe_events

before the patch:

/sys/kernel/debug # cat kprobes/list
ffffff801009fe28 k _do_fork+0x0 [DISABLED]

Quoted 6 lines

after the patch:

/sys/kernel/debug # cat kprobes/list
ffffff801009ff54 k _do_fork+0x4 [DISABLED][FTRACE]

This looks good to me. Except for a small confirmation below:

Quoted 100 lines

---
KPROBES_ON_FTRACE avoids much of the overhead with regular kprobes as it eliminates the need for a trap, as well as the need to emulate or single-step instructions.

Applied after arm64 FTRACE_WITH_REGS:
http://lists.infradead.org/pipermail/linux-arm-kernel/2019-August/674404.html

Changes since v3:
- move kprobe_lookup_name() and arch_kprobe_on_func_entry to ftrace.c since we only want to choose the ftrace entry for KPROBES_ON_FTRACE.
- only choose ftrace entry if (addr && !offset)

Changes since v2:
- remove patch1, make it a single cleanup patch
- remove "This patch" in the change log
- implement arm64's kprobe_lookup_name() and arch_kprobe_on_func_entry instead of patching the common kprobes code

Changes since v1:
- make the kprobes/x86: use instruction_pointer and instruction_pointer_set as patch1
- add Masami's ACK to patch1
- add some description about KPROBES_ON_FTRACE and why we need it on

    arm64

- correct the log before the patch
- remove the consolidation patch, make it as TODO
- only adjust kprobe's addr when KPROBE_FLAG_FTRACE is set
- if KPROBES_ON_FTRACE, ftrace_call_adjust() the kprobe's addr before calling ftrace_location()
- update the kprobes-on-ftrace/arch-support.txt in doc

 .../debug/kprobes-on-ftrace/arch-support.txt  |  2 +-
 arch/arm64/Kconfig                            |  1 +
 arch/arm64/kernel/probes/Makefile             |  1 +
 arch/arm64/kernel/probes/ftrace.c             | 84 +++++++++++++++++++
 4 files changed, 87 insertions(+), 1 deletion(-)
 create mode 100644 arch/arm64/kernel/probes/ftrace.c
Documentation/features/debug/kprobes-on-ftrace/arch-support.txt +1 -1
index 68f266944d5f..e8358a38981c 100644
@@ -9,7 +9,7 @@
     |       alpha: | TODO |
     |         arc: | TODO |
     |         arm: | TODO |
-    |       arm64: | TODO |
+    |       arm64: |  ok  |
     |         c6x: | TODO |
     |        csky: | TODO |
     |       h8300: | TODO |
arch/arm64/Kconfig +1 -0
index 663392d1eae2..928700f15e23 100644
@@ -167,6 +167,7 @@ config ARM64
 	select HAVE_STACKPROTECTOR
 	select HAVE_SYSCALL_TRACEPOINTS
 	select HAVE_KPROBES
+	select HAVE_KPROBES_ON_FTRACE
 	select HAVE_KRETPROBES
 	select HAVE_GENERIC_VDSO
 	select IOMMU_DMA if IOMMU_SUPPORT
arch/arm64/kernel/probes/Makefile +1 -0
index 8e4be92e25b1..4020cfc66564 100644
@@ -4,3 +4,4 @@ obj-$(CONFIG_KPROBES)		+= kprobes.o decode-insn.o	\
 				   simulate-insn.o
 obj-$(CONFIG_UPROBES)		+= uprobes.o decode-insn.o	\
 				   simulate-insn.o
+obj-$(CONFIG_KPROBES_ON_FTRACE)	+= ftrace.o
arch/arm64/kernel/probes/ftrace.c new file +20 -0
new file mode 100644
index 000000000000..5989c57660f3
@@ -0,0 +1,84 @@
+// SPDX-License-Identifier: GPL-2.0-or-later
+/*
+ * Dynamic Ftrace based Kprobes Optimization
+ *
+ * Copyright (C) Hitachi Ltd., 2012
+ * Copyright (C) 2019 Jisheng Zhang <[email protected]>
+ *		      Synaptics Incorporated
+ */
+
+#include <linux/kprobes.h>
+
+/* Ftrace callback handler for kprobes -- called under preepmt disabed */
+void kprobe_ftrace_handler(unsigned long ip, unsigned long parent_ip,
+			   struct ftrace_ops *ops, struct pt_regs *regs)
+{
+	struct kprobe *p;
+	struct kprobe_ctlblk *kcb;
+
+	/* Preempt is disabled by ftrace */
+	p = get_kprobe((kprobe_opcode_t *)ip);

arch/arm64/kernel/probes/ftrace.c line 30

Quoted 10 lines
arch/arm64/kernel/probes/ftrace.c new file +10 -0
+	if (unlikely(!p) || kprobe_disabled(p))
+		return;
+
+	kcb = get_kprobe_ctlblk();
+	if (kprobe_running()) {
+		kprobes_inc_nmissed_count(p);
+	} else {
+		unsigned long orig_ip = instruction_pointer(regs);
+		/* Kprobe handler expects regs->pc = pc + 4 as breakpoint hit */
+		instruction_pointer_set(regs, ip + sizeof(kprobe_opcode_t));

Just want to make sure that you've confirmed that this is what happens with a regular trap/brk based kprobe on ARM64. The reason for setting the instruction pointer here is to ensure that it is set to the same value as would be set if there was a trap/brk instruction at the ftrace location. This ensures that the kprobe pre handler sees the same value regardless.

Further changes to the instruction pointer are to achieve the same effect for kprobe post handlers.

- Naveen

#2 lore raw

Hi,

Quoted 29 lines · On Thu, 22 Aug 2019 12:23:58 +0530 "Naveen N. Rao"…
Quoted 15 lines · Jisheng Zhang

KPROBES_ON_FTRACE avoids much of the overhead with regular kprobes as it eliminates the need for a trap, as well as the need to emulate or single-step instructions.

Tested on berlin arm64 platform.

~ # mount -t debugfs debugfs /sys/kernel/debug/
~ # cd /sys/kernel/debug/
/sys/kernel/debug # echo 'p _do_fork' > tracing/kprobe_events

before the patch:

/sys/kernel/debug # cat kprobes/list
ffffff801009fe28 k _do_fork+0x0 [DISABLED]

Quoted 6 lines

after the patch:

/sys/kernel/debug # cat kprobes/list
ffffff801009ff54 k _do_fork+0x4 [DISABLED][FTRACE]

This looks good to me. Except for a small confirmation below:

<...>

Quoted 18 lines
Quoted 18 lines
arch/arm64/kernel/probes/ftrace.c new file +18 -0
+/* Ftrace callback handler for kprobes -- called under preepmt disabed */
+void kprobe_ftrace_handler(unsigned long ip, unsigned long parent_ip,
+                        struct ftrace_ops *ops, struct pt_regs *regs)
+{
+     struct kprobe *p;
+     struct kprobe_ctlblk *kcb;
+
+     /* Preempt is disabled by ftrace */
+     p = get_kprobe((kprobe_opcode_t *)ip);
+     if (unlikely(!p) || kprobe_disabled(p))
+             return;
+
+     kcb = get_kprobe_ctlblk();
+     if (kprobe_running()) {
+             kprobes_inc_nmissed_count(p);
+     } else {
+             unsigned long orig_ip = instruction_pointer(regs);
+             /* Kprobe handler expects regs->pc = pc + 4 as breakpoint hit */
Quoted 8 lines
Quoted 1 line

+ instruction_pointer_set(regs, ip + sizeof(kprobe_opcode_t));

Just want to make sure that you've confirmed that this is what happens with a regular trap/brk based kprobe on ARM64. The reason for setting the instruction pointer here is to ensure that it is set to the same value as would be set if there was a trap/brk instruction at the ftrace location. This ensures that the kprobe pre handler sees the same value regardless.

Due to the arm64's DYNAMIC_FTRACE_WITH_REGS implementation, the code itself is correct. But this doesn't look like "there was a trap instruction at the ftrace location".

W/O KPROBE_ON_FTRACE:

foo:
00 insA
04 insB
08 insC

kprobe's pre_handler() will see pc points to 00.

W/ KPROBE_ON_FTRACE:

foo:
00 lr saver
04 nop // will be modified to ftrace call ins when KPROBE is armed
08 insA
0c insB

later, kprobe_ftrace_handler() will see pc points to 04, so pc + 4 will point to 08 the same as the one w/o KPROBE_ON_FTRACE.

It seems I need to fix the comment.

Quoted 5 lines

Further changes to the instruction pointer are to achieve the same effect for kprobe post handlers.

#3 lore raw
Quoted 5 lines · Jisheng Zhang

Hi,

Quoted 1 line · On Thu, 22 Aug 2019 12:23:58 +0530 "Naveen N. Rao"…

Jisheng Zhang wrote:

...

Quoted 32 lines
Quoted 19 lines
Quoted 19 lines

+/* Ftrace callback handler for kprobes -- called under preepmt disabed */

arch/arm64/kernel/probes/ftrace.c new file +17 -0
+void kprobe_ftrace_handler(unsigned long ip, unsigned long parent_ip,
+                        struct ftrace_ops *ops, struct pt_regs *regs)
+{
+     struct kprobe *p;
+     struct kprobe_ctlblk *kcb;
+
+     /* Preempt is disabled by ftrace */
+     p = get_kprobe((kprobe_opcode_t *)ip);
+     if (unlikely(!p) || kprobe_disabled(p))
+             return;
+
+     kcb = get_kprobe_ctlblk();
+     if (kprobe_running()) {
+             kprobes_inc_nmissed_count(p);
+     } else {
+             unsigned long orig_ip = instruction_pointer(regs);
+             /* Kprobe handler expects regs->pc = pc + 4 as breakpoint hit */
Quoted 8 lines
Quoted 1 line

+ instruction_pointer_set(regs, ip + sizeof(kprobe_opcode_t));

Just want to make sure that you've confirmed that this is what happens with a regular trap/brk based kprobe on ARM64. The reason for setting the instruction pointer here is to ensure that it is set to the same value as would be set if there was a trap/brk instruction at the ftrace location. This ensures that the kprobe pre handler sees the same value regardless.

Due to the arm64's DYNAMIC_FTRACE_WITH_REGS implementation, the code itself is correct. But this doesn't look like "there was a trap instruction at the ftrace location".

Quoted 8 lines

W/O KPROBE_ON_FTRACE:

foo:
00 insA
04 insB
08 insC

kprobe's pre_handler() will see pc points to 00.

In this case, the probe will be placed at foo+0x00, so pre_handler() seeing that address in pt_regs is correct behavior - as long as arm64 'brk' instruction causes an exception with the instruction pointer set *to* the 'brk' instruction. This is similar to how powerpc 'trap' works. However, x86 'int3' causes an exception *after* execution of the instruction.

Quoted 8 lines

W/ KPROBE_ON_FTRACE:

foo:
00 lr saver
04 nop // will be modified to ftrace call ins when KPROBE is armed
08 insA
0c insB

In this case, if user asks for a probe to be placed at 'foo', we will choose foo+0x04 and from that point on, the behavior should reflect that a kprobe was placed at foo+0x04. In particular, the pre_handler() should see foo+0x04 in pt_regs. The post_handler() would then see foo+0x08.

Quoted 3 lines

later, kprobe_ftrace_handler() will see pc points to 04, so pc + 4 will point to 08 the same as the one w/o KPROBE_ON_FTRACE.

I didn't mean to compare regular trap/brk based kprobes with KPROBES_ON_FTRACE. The only important aspect is that the handlers see consistent pt_regs in both cases, depending on where the kprobe was placed. Choosing a different address/offset to place a kprobe during its registration is an orthogonal aspect.

Quoted 2 lines

It seems I need to fix the comment.

Given your explanation above, I think you can simply drop the first adjustment to the instruction pointer before the pre handler invocation. The rest of the code looks fine.

- Naveen

#4 lore raw
Quoted 48 lines · On Thu, 22 Aug 2019 15:52:05 +0530 "Naveen N. Rao"…
Quoted 5 lines · Jisheng Zhang

Hi,

Quoted 1 line · On Thu, 22 Aug 2019 12:23:58 +0530 "Naveen N. Rao"…

Jisheng Zhang wrote:

...

Quoted 39 lines
Quoted 19 lines
Quoted 19 lines

+/* Ftrace callback handler for kprobes -- called under preepmt disabed */

arch/arm64/kernel/probes/ftrace.c new file +17 -0
+void kprobe_ftrace_handler(unsigned long ip, unsigned long parent_ip,
+                        struct ftrace_ops *ops, struct pt_regs *regs)
+{
+     struct kprobe *p;
+     struct kprobe_ctlblk *kcb;
+
+     /* Preempt is disabled by ftrace */
+     p = get_kprobe((kprobe_opcode_t *)ip);
+     if (unlikely(!p) || kprobe_disabled(p))
+             return;
+
+     kcb = get_kprobe_ctlblk();
+     if (kprobe_running()) {
+             kprobes_inc_nmissed_count(p);
+     } else {
+             unsigned long orig_ip = instruction_pointer(regs);
+             /* Kprobe handler expects regs->pc = pc + 4 as breakpoint hit */
Quoted 8 lines
Quoted 1 line

+ instruction_pointer_set(regs, ip + sizeof(kprobe_opcode_t));

Just want to make sure that you've confirmed that this is what happens with a regular trap/brk based kprobe on ARM64. The reason for setting the instruction pointer here is to ensure that it is set to the same value as would be set if there was a trap/brk instruction at the ftrace location. This ensures that the kprobe pre handler sees the same value regardless.

Due to the arm64's DYNAMIC_FTRACE_WITH_REGS implementation, the code itself is correct. But this doesn't look like "there was a trap instruction at the ftrace location".

W/O KPROBE_ON_FTRACE:

foo:

00    insA
04    insB
08    insC
Quoted 5 lines
Quoted 1 line

kprobe's pre_handler() will see pc points to 00.

In this case, the probe will be placed at foo+0x00, so pre_handler() seeing that address in pt_regs is correct behavior - as long as arm64 'brk' instruction causes an exception with the instruction pointer set

Yep, confirmed with regular trap/brk based kprobes, I do see PC set to the "brk" instruction.

Quoted 3 lines

*to* the 'brk' instruction. This is similar to how powerpc 'trap' works. However, x86 'int3' causes an exception *after* execution of the instruction.

Got it. I understand where's the comment "expects regs->pc = pc + 1" from.

Quoted 16 lines
Quoted 9 lines

W/ KPROBE_ON_FTRACE:

foo:

00    lr saver
04    nop     // will be modified to ftrace call ins when KPROBE is armed
08    insA
0c    insB  

In this case, if user asks for a probe to be placed at 'foo', we will choose foo+0x04 and from that point on, the behavior should reflect that a kprobe was placed at foo+0x04. In particular, the pre_handler() should see foo+0x04 in pt_regs. The post_handler() would then see foo+0x08.

Quoted 1 line
Quoted 8 lines
Quoted 2 lines

later, kprobe_ftrace_handler() will see pc points to 04, so pc + 4 will point to 08 the same as the one w/o KPROBE_ON_FTRACE.

I didn't mean to compare regular trap/brk based kprobes with KPROBES_ON_FTRACE. The only important aspect is that the handlers see consistent pt_regs in both cases, depending on where the kprobe was placed. Choosing a different address/offset to place a kprobe during its registration is an orthogonal aspect.

Indeed, previously, I want to let the PC point to the same instruction, it seems I misunderstood the "consistent" meaning.

Quoted 9 lines
Quoted 3 lines

It seems I need to fix the comment.

Given your explanation above, I think you can simply drop the first adjustment to the instruction pointer before the pre handler invocation. The rest of the code looks fine.

Yep, thanks a lot. Will send out a new version soon.

#5 lore raw

Hi,

Quoted 85 lines · On Thu, 22 Aug 2019 18:32:54 +0800 Jisheng Zhang <[email protected]>…
Quoted 48 lines · On Thu, 22 Aug 2019 15:52:05 +0530 "Naveen N. Rao"…
Quoted 5 lines · Jisheng Zhang

Hi,

Quoted 1 line · On Thu, 22 Aug 2019 12:23:58 +0530 "Naveen N. Rao"…

Jisheng Zhang wrote:

...

Quoted 39 lines
Quoted 19 lines
Quoted 19 lines

+/* Ftrace callback handler for kprobes -- called under preepmt disabed */

arch/arm64/kernel/probes/ftrace.c new file +17 -0
+void kprobe_ftrace_handler(unsigned long ip, unsigned long parent_ip,
+                        struct ftrace_ops *ops, struct pt_regs *regs)
+{
+     struct kprobe *p;
+     struct kprobe_ctlblk *kcb;
+
+     /* Preempt is disabled by ftrace */
+     p = get_kprobe((kprobe_opcode_t *)ip);
+     if (unlikely(!p) || kprobe_disabled(p))
+             return;
+
+     kcb = get_kprobe_ctlblk();
+     if (kprobe_running()) {
+             kprobes_inc_nmissed_count(p);
+     } else {
+             unsigned long orig_ip = instruction_pointer(regs);
+             /* Kprobe handler expects regs->pc = pc + 4 as breakpoint hit */
Quoted 8 lines
Quoted 1 line

+ instruction_pointer_set(regs, ip + sizeof(kprobe_opcode_t));

Just want to make sure that you've confirmed that this is what happens with a regular trap/brk based kprobe on ARM64. The reason for setting the instruction pointer here is to ensure that it is set to the same value as would be set if there was a trap/brk instruction at the ftrace location. This ensures that the kprobe pre handler sees the same value regardless.

Due to the arm64's DYNAMIC_FTRACE_WITH_REGS implementation, the code itself is correct. But this doesn't look like "there was a trap instruction at the ftrace location".

W/O KPROBE_ON_FTRACE:

foo:

00    insA
04    insB
08    insC
Quoted 5 lines
Quoted 1 line

kprobe's pre_handler() will see pc points to 00.

In this case, the probe will be placed at foo+0x00, so pre_handler() seeing that address in pt_regs is correct behavior - as long as arm64 'brk' instruction causes an exception with the instruction pointer set

Yep, confirmed with regular trap/brk based kprobes, I do see PC set to the "brk" instruction.

Quoted 3 lines

*to* the 'brk' instruction. This is similar to how powerpc 'trap' works. However, x86 'int3' causes an exception *after* execution of the instruction.

Got it. I understand where's the comment "expects regs->pc = pc + 1" from.

Quoted 19 lines
Quoted 9 lines

W/ KPROBE_ON_FTRACE:

foo:

00    lr saver
04    nop     // will be modified to ftrace call ins when KPROBE is armed
08    insA
0c    insB    

In this case, if user asks for a probe to be placed at 'foo', we will choose foo+0x04 and from that point on, the behavior should reflect that a kprobe was placed at foo+0x04. In particular, the pre_handler() should see foo+0x04 in pt_regs. The post_handler() would then see foo+0x08.

Quoted 3 lines

later, kprobe_ftrace_handler() will see pc points to 04, so pc + 4 will point to 08 the same as the one w/o KPROBE_ON_FTRACE.

Quoted 15 lines
Quoted 5 lines

I didn't mean to compare regular trap/brk based kprobes with KPROBES_ON_FTRACE. The only important aspect is that the handlers see consistent pt_regs in both cases, depending on where the kprobe was placed. Choosing a different address/offset to place a kprobe during its registration is an orthogonal aspect.

Indeed, previously, I want to let the PC point to the same instruction, it seems I misunderstood the "consistent" meaning.

Quoted 6 lines
Quoted 3 lines

It seems I need to fix the comment.

Given your explanation above, I think you can simply drop the first adjustment to the instruction pointer before the pre handler invocation.

Just send out v5. But the first adjustment is modified as instruction_pointer_set(regs, ip);

Because in entry of kprobe_ftrace_handler() pc/ip(the first parameter) points to foo+0x4, while regs->pc points to foo+0x8. Based on your previous explanation, I think we should instruction_pointer_set(regs, ip) to let the pre_handler see foo+0x4

Thanks a lot for your help