Skip to contentSkip to conversation

libbpf: replace '.' with '_' in legacy kprobe event name

PATCH Applied [email protected] 4 participants 7 messages Jan 13, 2023 – Jan 17, 2023

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

PATCH 1/1 libbpf: replace '.' with '_' in legacy kprobe event name

From: Menglong Dong <[email protected]>

'.' is not allowed in the event name of kprobe. Therefore, we will get a EINVAL if the kernel function name has a '.' in legacy kprobe attach case, such as 'icmp_reply.constprop.0'.

In order to adapt this case, we need to replace the '.' with other char in gen_kprobe_legacy_event_name(). And I use '_' for this propose.

---
 tools/lib/bpf/libbpf.c | 7 +++++++
 1 file changed, 7 insertions(+)
tools/lib/bpf/libbpf.c +7 -0
index fdfb1ca34ced..5d6f6675c2f2 100644
@@ -9994,9 +9994,16 @@ static void gen_kprobe_legacy_event_name(char *buf, size_t buf_sz,
 					 const char *kfunc_name, size_t offset)
 {
 	static int index = 0;
+	int i = 0;
 
 	snprintf(buf, buf_sz, "libbpf_%u_%s_0x%zx_%d", getpid(), kfunc_name, offset,
 		 __sync_fetch_and_add(&index, 1));
+
+	while (buf[i] != '\0') {
+		if (buf[i] == '.')
+			buf[i] = '_';
+		i++;
+	}
 }
 
 static int add_kprobe_event_legacy(const char *probe_name, bool retprobe,
Signature · 1 line
2.39.0
Reviewed-by on v1 1/1
#1 lore raw
Quoted 23 lines · On 13/01/2023 09:34, [email protected] wrote:

From: Menglong Dong <[email protected]>

'.' is not allowed in the event name of kprobe. Therefore, we will get a EINVAL if the kernel function name has a '.' in legacy kprobe attach case, such as 'icmp_reply.constprop.0'.

In order to adapt this case, we need to replace the '.' with other char in gen_kprobe_legacy_event_name(). And I use '_' for this propose.

---
 tools/lib/bpf/libbpf.c | 7 +++++++
 1 file changed, 7 insertions(+)
tools/lib/bpf/libbpf.c +1 -0
index fdfb1ca34ced..5d6f6675c2f2 100644
@@ -9994,9 +9994,16 @@ static void gen_kprobe_legacy_event_name(char *buf, size_t buf_sz,
 					 const char *kfunc_name, size_t offset)
 {
 	static int index = 0;
+	int i = 0;

tools/lib/bpf/libbpf.c static void gen_kprobe_legacy_event_name(char *buf, size_t buf_sz, · line 10007

Quoted 10 lines
tools/lib/bpf/libbpf.c +6 -0
 
 	snprintf(buf, buf_sz, "libbpf_%u_%s_0x%zx_%d", getpid(), kfunc_name, offset,
 		 __sync_fetch_and_add(&index, 1));
+
+	while (buf[i] != '\0') {
+		if (buf[i] == '.')
+			buf[i] = '_';
+		i++;
+	}
 }

probably more naturally expressed as a for() loop as is done in gen_uprobe_legacy_event_name(), but not a big deal.

One issue with the legacy kprobe code is that we don't get test coverage with it on new kernels - I wonder if it would be worth adding a force_legacy option to bpf_kprobe_opts? A separate issue to this change of course, but if we had that we could add some legacy kprobe tests that would run for new kernels as well.

Alan

Quoted 3 lines
tools/lib/bpf/libbpf.c +0 -0
 static int add_kprobe_event_legacy(const char *probe_name, bool retprobe,
 
#2 lore raw
Quoted 39 lines · Alan Maguire
Quoted 23 lines · On 13/01/2023 09:34, [email protected] wrote:

From: Menglong Dong <[email protected]>

'.' is not allowed in the event name of kprobe. Therefore, we will get a EINVAL if the kernel function name has a '.' in legacy kprobe attach case, such as 'icmp_reply.constprop.0'.

In order to adapt this case, we need to replace the '.' with other char in gen_kprobe_legacy_event_name(). And I use '_' for this propose.

---
 tools/lib/bpf/libbpf.c | 7 +++++++
 1 file changed, 7 insertions(+)
tools/lib/bpf/libbpf.c +1 -0
index fdfb1ca34ced..5d6f6675c2f2 100644
@@ -9994,9 +9994,16 @@ static void gen_kprobe_legacy_event_name(char *buf, size_t buf_sz,
                                       const char *kfunc_name, size_t offset)
 {
      static int index = 0;
+     int i = 0;

tools/lib/bpf/libbpf.c static void gen_kprobe_legacy_event_name(char *buf, size_t buf_sz, · line 10007

Quoted 10 lines
tools/lib/bpf/libbpf.c +6 -0
 
      snprintf(buf, buf_sz, "libbpf_%u_%s_0x%zx_%d", getpid(), kfunc_name, offset,
               __sync_fetch_and_add(&index, 1));
+
+     while (buf[i] != '\0') {
+             if (buf[i] == '.')
+                     buf[i] = '_';
+             i++;
+     }
 }

probably more naturally expressed as a for() loop as is done in gen_uprobe_legacy_event_name(), but not a big deal.

Quoted 1 line

Applied, but tuned to be exactly the same loop as in gen_uprobe_legacy_event_name. Thanks.

Quoted 6 lines

One issue with the legacy kprobe code is that we don't get test coverage with it on new kernels - I wonder if it would be worth adding a force_legacy option to bpf_kprobe_opts? A separate issue to this change of course, but if we had that we could add some legacy kprobe tests that would run for new kernels as well.

Yep, good idea. If we ever have some bug in the latest greatest kprobe implementation, users will have an option to work around that with this.

The only thing is that we already have 3 modes: legacy, perf-based through ioctl, and bpf_link-based, so I think it should be something like

enum kprobe_mode {
KPROBE_MODE_DEFAULT = 0, /* latest supported by kernel */ KPROBE_MODE_LEGACY,
KPROBE_MODE_PERF,
KPROBE_MODE_LINK,
};

LEGACY/PERF/LINK naming should be thought through, just a quick example.

And then just have enum kprobe_mode mode; in kprobe_opts, which would default to 0 (KPROBE_MODE_DEFAULT).

Would that work?

Quoted 5 lines

Alan

Quoted 3 lines
tools/lib/bpf/libbpf.c +0 -0
 static int add_kprobe_event_legacy(const char *probe_name, bool retprobe,
 
#3 lore raw

Hello,

Quoted 41 lines · Andrii Nakryiko
Quoted 39 lines · Alan Maguire
Quoted 23 lines · On 13/01/2023 09:34, [email protected] wrote:

From: Menglong Dong <[email protected]>

'.' is not allowed in the event name of kprobe. Therefore, we will get a EINVAL if the kernel function name has a '.' in legacy kprobe attach case, such as 'icmp_reply.constprop.0'.

In order to adapt this case, we need to replace the '.' with other char in gen_kprobe_legacy_event_name(). And I use '_' for this propose.

---
 tools/lib/bpf/libbpf.c | 7 +++++++
 1 file changed, 7 insertions(+)
tools/lib/bpf/libbpf.c +1 -0
index fdfb1ca34ced..5d6f6675c2f2 100644
@@ -9994,9 +9994,16 @@ static void gen_kprobe_legacy_event_name(char *buf, size_t buf_sz,
                                       const char *kfunc_name, size_t offset)
 {
      static int index = 0;
+     int i = 0;

tools/lib/bpf/libbpf.c static void gen_kprobe_legacy_event_name(char *buf, size_t buf_sz, · line 10007

Quoted 10 lines
tools/lib/bpf/libbpf.c +6 -0
 
      snprintf(buf, buf_sz, "libbpf_%u_%s_0x%zx_%d", getpid(), kfunc_name, offset,
               __sync_fetch_and_add(&index, 1));
+
+     while (buf[i] != '\0') {
+             if (buf[i] == '.')
+                     buf[i] = '_';
+             i++;
+     }
 }

probably more naturally expressed as a for() loop as is done in gen_uprobe_legacy_event_name(), but not a big deal.

Quoted 5 lines
Quoted 1 line

Applied, but tuned to be exactly the same loop as in gen_uprobe_legacy_event_name. Thanks.

Thanks for your modification, it looks much better now!

Quoted 22 lines
Quoted 6 lines

One issue with the legacy kprobe code is that we don't get test coverage with it on new kernels - I wonder if it would be worth adding a force_legacy option to bpf_kprobe_opts? A separate issue to this change of course, but if we had that we could add some legacy kprobe tests that would run for new kernels as well.

Yep, good idea. If we ever have some bug in the latest greatest kprobe implementation, users will have an option to work around that with this.

The only thing is that we already have 3 modes: legacy, perf-based through ioctl, and bpf_link-based, so I think it should be something like

enum kprobe_mode {
KPROBE_MODE_DEFAULT = 0, /* latest supported by kernel */ KPROBE_MODE_LEGACY,
KPROBE_MODE_PERF,
KPROBE_MODE_LINK,
};

Quoted 7 lines

LEGACY/PERF/LINK naming should be thought through, just a quick example.

And then just have enum kprobe_mode mode; in kprobe_opts, which would default to 0 (KPROBE_MODE_DEFAULT).

Would that work?

Sounds great, which means I don't have to switch to an older kernel to test this function for my app.

BTW, should I do this job, (which is my pleasure), or Alan?

Thanks!
Menglong Dong

Quoted 5 lines
Quoted 5 lines

Alan

Quoted 3 lines
tools/lib/bpf/libbpf.c +0 -0
 static int add_kprobe_event_legacy(const char *probe_name, bool retprobe,
 
#4 lore raw
Quoted 45 lines · Menglong Dong

Hello,

Quoted 41 lines · Andrii Nakryiko
Quoted 39 lines · Alan Maguire
Quoted 23 lines · On 13/01/2023 09:34, [email protected] wrote:

From: Menglong Dong <[email protected]>

'.' is not allowed in the event name of kprobe. Therefore, we will get a EINVAL if the kernel function name has a '.' in legacy kprobe attach case, such as 'icmp_reply.constprop.0'.

In order to adapt this case, we need to replace the '.' with other char in gen_kprobe_legacy_event_name(). And I use '_' for this propose.

---
 tools/lib/bpf/libbpf.c | 7 +++++++
 1 file changed, 7 insertions(+)
tools/lib/bpf/libbpf.c +1 -0
index fdfb1ca34ced..5d6f6675c2f2 100644
@@ -9994,9 +9994,16 @@ static void gen_kprobe_legacy_event_name(char *buf, size_t buf_sz,
                                       const char *kfunc_name, size_t offset)
 {
      static int index = 0;
+     int i = 0;

tools/lib/bpf/libbpf.c static void gen_kprobe_legacy_event_name(char *buf, size_t buf_sz, · line 10007

Quoted 10 lines
tools/lib/bpf/libbpf.c +6 -0
 
      snprintf(buf, buf_sz, "libbpf_%u_%s_0x%zx_%d", getpid(), kfunc_name, offset,
               __sync_fetch_and_add(&index, 1));
+
+     while (buf[i] != '\0') {
+             if (buf[i] == '.')
+                     buf[i] = '_';
+             i++;
+     }
 }

probably more naturally expressed as a for() loop as is done in gen_uprobe_legacy_event_name(), but not a big deal.

Quoted 37 lines
Quoted 5 lines
Quoted 1 line

Applied, but tuned to be exactly the same loop as in gen_uprobe_legacy_event_name. Thanks.

Thanks for your modification, it looks much better now!

Quoted 29 lines
Quoted 6 lines

One issue with the legacy kprobe code is that we don't get test coverage with it on new kernels - I wonder if it would be worth adding a force_legacy option to bpf_kprobe_opts? A separate issue to this change of course, but if we had that we could add some legacy kprobe tests that would run for new kernels as well.

Yep, good idea. If we ever have some bug in the latest greatest kprobe implementation, users will have an option to work around that with this.

The only thing is that we already have 3 modes: legacy, perf-based through ioctl, and bpf_link-based, so I think it should be something like

enum kprobe_mode {
KPROBE_MODE_DEFAULT = 0, /* latest supported by kernel */ KPROBE_MODE_LEGACY,
KPROBE_MODE_PERF,
KPROBE_MODE_LINK,
};

LEGACY/PERF/LINK naming should be thought through, just a quick example.

And then just have enum kprobe_mode mode; in kprobe_opts, which would default to 0 (KPROBE_MODE_DEFAULT).

Would that work?

Looks good - I'd missed the "no BPF link" case, it'd be great to test that too.

So for legacy mode, we'd force using the legacy codepath, and to simulate the PERF mode where BPF link isn't supported I think we'd need to add to bpf_perf_event_opts so that we could choose the "no bpf link" code path rather than purely relying on the kernel support test.

This would be nice as it would allow us to test other "pre-BPF link" attach targets too.

So I think we need add an option to bpf_perf_event_opts for when KPROBE_MODE_PERF is set, such as PE_MODE_PERF or PE_MODE_NO_BPF_LINK.

All of this would generalize to uprobe too I think; having a perf option makes that straightforward I suspect.

Quoted 7 lines

Sounds great, which means I don't have to switch to an older kernel to test this function for my app.

BTW, should I do this job, (which is my pleasure), or Alan?

Feel free to take this on if you've got time; I'm trying to get the dwarves patches covering support for .isra functions out as soon as possible so it would probably be a week or so before I get to this. Something like the above combined with updating the attach_probe selftests to run through the various attach modes would be great for testing legacy codepaths on newer kernels. We could perhaps rework the test_attach_probe() function to take an attach mode argument, and add subtests for each attach mode (skipping if it wasn't supported). Thanks!

Alan

Quoted 8 lines

Thanks!
Menglong Dong

Quoted 5 lines
Quoted 5 lines

Alan

Quoted 3 lines
tools/lib/bpf/libbpf.c +0 -0
 static int add_kprobe_event_legacy(const char *probe_name, bool retprobe,
 
#5 lore raw
Quoted 79 lines · Alan Maguire
Quoted 77 lines · Menglong Dong

Hello,

Quoted 41 lines · Andrii Nakryiko
Quoted 39 lines · Alan Maguire
Quoted 23 lines · On 13/01/2023 09:34, [email protected] wrote:

From: Menglong Dong <[email protected]>

'.' is not allowed in the event name of kprobe. Therefore, we will get a EINVAL if the kernel function name has a '.' in legacy kprobe attach case, such as 'icmp_reply.constprop.0'.

In order to adapt this case, we need to replace the '.' with other char in gen_kprobe_legacy_event_name(). And I use '_' for this propose.

---
 tools/lib/bpf/libbpf.c | 7 +++++++
 1 file changed, 7 insertions(+)
tools/lib/bpf/libbpf.c +1 -0
index fdfb1ca34ced..5d6f6675c2f2 100644
@@ -9994,9 +9994,16 @@ static void gen_kprobe_legacy_event_name(char *buf, size_t buf_sz,
                                       const char *kfunc_name, size_t offset)
 {
      static int index = 0;
+     int i = 0;

tools/lib/bpf/libbpf.c static void gen_kprobe_legacy_event_name(char *buf, size_t buf_sz, · line 10007

Quoted 10 lines
tools/lib/bpf/libbpf.c +6 -0
 
      snprintf(buf, buf_sz, "libbpf_%u_%s_0x%zx_%d", getpid(), kfunc_name, offset,
               __sync_fetch_and_add(&index, 1));
+
+     while (buf[i] != '\0') {
+             if (buf[i] == '.')
+                     buf[i] = '_';
+             i++;
+     }
 }

probably more naturally expressed as a for() loop as is done in gen_uprobe_legacy_event_name(), but not a big deal.

Quoted 5 lines
Quoted 1 line

Applied, but tuned to be exactly the same loop as in gen_uprobe_legacy_event_name. Thanks.

Thanks for your modification, it looks much better now!

Quoted 24 lines
Quoted 6 lines

One issue with the legacy kprobe code is that we don't get test coverage with it on new kernels - I wonder if it would be worth adding a force_legacy option to bpf_kprobe_opts? A separate issue to this change of course, but if we had that we could add some legacy kprobe tests that would run for new kernels as well.

Yep, good idea. If we ever have some bug in the latest greatest kprobe implementation, users will have an option to work around that with this.

The only thing is that we already have 3 modes: legacy, perf-based through ioctl, and bpf_link-based, so I think it should be something like

enum kprobe_mode {
KPROBE_MODE_DEFAULT = 0, /* latest supported by kernel */ KPROBE_MODE_LEGACY,
KPROBE_MODE_PERF,
KPROBE_MODE_LINK,
};

LEGACY/PERF/LINK naming should be thought through, just a quick example.

Quoted 8 lines
Quoted 5 lines
Quoted 5 lines

And then just have enum kprobe_mode mode; in kprobe_opts, which would default to 0 (KPROBE_MODE_DEFAULT).

Would that work?

Looks good - I'd missed the "no BPF link" case, it'd be great to test that too.

My mistake, I forgot to add the 'bpf-next' tag :)

Quoted 22 lines

So for legacy mode, we'd force using the legacy codepath, and to simulate the PERF mode where BPF link isn't supported I think we'd need to add to bpf_perf_event_opts so that we could choose the "no bpf link" code path rather than purely relying on the kernel support test.

This would be nice as it would allow us to test other "pre-BPF link" attach targets too.

So I think we need add an option to bpf_perf_event_opts for when KPROBE_MODE_PERF is set, such as PE_MODE_PERF or PE_MODE_NO_BPF_LINK.

All of this would generalize to uprobe too I think; having a perf option makes that straightforward I suspect.

Quoted 7 lines

Sounds great, which means I don't have to switch to an older kernel to test this function for my app.

BTW, should I do this job, (which is my pleasure), or Alan?

Quoted 8 lines

Feel free to take this on if you've got time; I'm trying to get the dwarves patches covering support for .isra functions out as soon as possible so it would probably be a week or so before I get to this. Something like the above combined with updating the attach_probe selftests to run through the various attach modes would be great for testing legacy codepaths on newer kernels. We could perhaps rework the test_attach_probe() function to take an attach mode argument, and add subtests for each attach mode (skipping if it wasn't supported). Thanks!

Okay, I'll have a try!

Quoted 10 lines

Alan

Quoted 8 lines

Thanks!
Menglong Dong

Quoted 5 lines
Quoted 5 lines

Alan

Quoted 3 lines
tools/lib/bpf/libbpf.c +0 -0
 static int add_kprobe_event_legacy(const char *probe_name, bool retprobe,
 
#6 lore raw
In reply to [email protected]
This patch was applied to bpf/bpf-next.git (master) by Andrii Nakryiko <[email protected]>:

Hello:

This patch was applied to bpf/bpf-next.git (master) by Andrii Nakryiko <[email protected]>:

Quoted 10 lines · On Fri, 13 Jan 2023 17:34:27 +0800 you wrote:

From: Menglong Dong <[email protected]>

'.' is not allowed in the event name of kprobe. Therefore, we will get a EINVAL if the kernel function name has a '.' in legacy kprobe attach case, such as 'icmp_reply.constprop.0'.

In order to adapt this case, we need to replace the '.' with other char in gen_kprobe_legacy_event_name(). And I use '_' for this propose.

[...]

Here is the summary with links:
- libbpf: replace '.' with '_' in legacy kprobe event name

    https://git.kernel.org/bpf/bpf-next/c/2fa074536590

You are awesome, thank you!

Signature · 2 lines
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html