Now the writing operation return the count of writes whether events are enabled or disabled. Fix this by just return -ENOENT when events are disabled.
v1 -> v2: - Change the returh vale from -EFAULT to -ENOENT
sunliming (3): tracing/user_events: Fix incorrect return value for writing operation when events are disabled selftests/user_events: Enable the event before write_fault test in ftrace self-test selftests/user_events: Add test cases when event is disabled
kernel/trace/trace_events_user.c | 3 ++- tools/testing/selftests/user_events/ftrace_test.c | 8 ++++++++ 2 files changed, 10 insertions(+), 1 deletion(-)
The writing operation return the count of writes whether events are enabled or disabled. This is incorrect when events are disabled. Fix this by just return -ENOENT when events are disabled.
Signed-off-by: sunliming sunliming@kylinos.cn --- kernel/trace/trace_events_user.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/kernel/trace/trace_events_user.c b/kernel/trace/trace_events_user.c index 1ac5ba5685ed..92204bbe79da 100644 --- a/kernel/trace/trace_events_user.c +++ b/kernel/trace/trace_events_user.c @@ -1957,7 +1957,8 @@ static ssize_t user_events_write_core(struct file *file, struct iov_iter *i)
if (unlikely(faulted)) return -EFAULT; - } + } else + return -ENOENT;
return ret; }
On Fri, Jun 09, 2023 at 11:03:00AM +0800, sunliming wrote:
The writing operation return the count of writes whether events are enabled or disabled. This is incorrect when events are disabled. Fix this by just return -ENOENT when events are disabled.
When testing this patch locally I found that we would occasionally get -ENOENT when events were enabled, but then become disabled, since writes do not have any locking around the tracepoint checks for performance reasons.
I've asked a few peers of mine their thoughts on this, whether an error should result when there are no enabled events. The consensus I've heard back is that they would not consider this case an actual error, just as writing to /dev/null does not actually return an error.
However, if you feel strongly we need this and have a good use case, it seems better to enable this logic behind a flag instead of having it default based on my conversations with others.
Thanks, -Beau
Signed-off-by: sunliming sunliming@kylinos.cn
kernel/trace/trace_events_user.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/kernel/trace/trace_events_user.c b/kernel/trace/trace_events_user.c index 1ac5ba5685ed..92204bbe79da 100644 --- a/kernel/trace/trace_events_user.c +++ b/kernel/trace/trace_events_user.c @@ -1957,7 +1957,8 @@ static ssize_t user_events_write_core(struct file *file, struct iov_iter *i) if (unlikely(faulted)) return -EFAULT;
- }
- } else
return -ENOENT;return ret; } -- 2.25.1
Beau Belgrave beaub@linux.microsoft.com 于2023年6月17日周六 00:08写道：
On Fri, Jun 09, 2023 at 11:03:00AM +0800, sunliming wrote:
The writing operation return the count of writes whether events are enabled or disabled. This is incorrect when events are disabled. Fix this by just return -ENOENT when events are disabled.
When testing this patch locally I found that we would occasionally get -ENOENT when events were enabled, but then become disabled, since writes do not have any locking around the tracepoint checks for performance reasons.
I've asked a few peers of mine their thoughts on this, whether an error should result when there are no enabled events. The consensus I've heard back is that they would not consider this case an actual error, just as writing to /dev/null does not actually return an error.
However, if you feel strongly we need this and have a good use case, it seems better to enable this logic behind a flag instead of having it default based on my conversations with others.
Thanks, -Beau
There is indeed a problem. Once enabled, perform the write operation immediately.
Now，when the event is disabled, the trace record appears to be lost. In some situations where data timing is sensitive, it may cause confusion. In this case, not returning an error (as mentioned in your reply, it is not considered this case an actual error) and returning 0 ( meaning that the number of data to be written is 0) may be a good way to handle it? Thanks, -Sunliming
Signed-off-by: sunliming sunliming@kylinos.cn
kernel/trace/trace_events_user.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/kernel/trace/trace_events_user.c b/kernel/trace/trace_events_user.c index 1ac5ba5685ed..92204bbe79da 100644 --- a/kernel/trace/trace_events_user.c +++ b/kernel/trace/trace_events_user.c @@ -1957,7 +1957,8 @@ static ssize_t user_events_write_core(struct file *file, struct iov_iter *i)
if (unlikely(faulted)) return -EFAULT;
}
} elsereturn -ENOENT; return ret;}
2.25.1
On Mon, Jun 19, 2023 at 04:51:56PM +0800, sunliming wrote:
Beau Belgrave beaub@linux.microsoft.com 于2023年6月17日周六 00:08写道：
On Fri, Jun 09, 2023 at 11:03:00AM +0800, sunliming wrote:
The writing operation return the count of writes whether events are enabled or disabled. This is incorrect when events are disabled. Fix this by just return -ENOENT when events are disabled.
When testing this patch locally I found that we would occasionally get -ENOENT when events were enabled, but then become disabled, since writes do not have any locking around the tracepoint checks for performance reasons.
I've asked a few peers of mine their thoughts on this, whether an error should result when there are no enabled events. The consensus I've heard back is that they would not consider this case an actual error, just as writing to /dev/null does not actually return an error.
However, if you feel strongly we need this and have a good use case, it seems better to enable this logic behind a flag instead of having it default based on my conversations with others.
Thanks, -Beau
There is indeed a problem. Once enabled, perform the write operation immediately.
The immediate write does work, and gets put into a buffer. The ftrace and perf self tests do the above case. So, no worries at this point.
Now，when the event is disabled, the trace record appears to be lost.
I'm taking this to mean, if in between the time of the bit check and the actual write() /writev() syscall the event becomes disabled, the event won't write to the buffer. Yes, that is expected.
In some situations where data timing is sensitive, it may cause confusion. In this case, not returning an error (as mentioned in your reply, it is not considered this case an actual error) and returning 0 ( meaning that the number of data to be written is 0) may be a good way to handle it?
This is where I get a little lost. What would a user process do with a return of 0 bytes? It shouldn't retry, since it just hit that small timing window. In reality, it just incurred a temporary excessive syscall cost, but no real data loss (the operator/admin turned the event off).
I'm missing why you feel it's important the user process know such a window was hit?
Can you help me understand that?
I do think returning 0 bytes is better than an error here, but I'd really like to know why the user process wants to know at all. Maybe they have user-space only logging and want to be able to mark there if it's in both spots (kernel and user buffers)?
Thanks, -Beau
Thanks, -Sunliming
Signed-off-by: sunliming sunliming@kylinos.cn
kernel/trace/trace_events_user.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/kernel/trace/trace_events_user.c b/kernel/trace/trace_events_user.c index 1ac5ba5685ed..92204bbe79da 100644 --- a/kernel/trace/trace_events_user.c +++ b/kernel/trace/trace_events_user.c @@ -1957,7 +1957,8 @@ static ssize_t user_events_write_core(struct file *file, struct iov_iter *i)
if (unlikely(faulted)) return -EFAULT;
}
} elsereturn -ENOENT; return ret;}
2.25.1
Beau Belgrave beaub@linux.microsoft.com 于2023年6月20日周二 02:40写道：
On Mon, Jun 19, 2023 at 04:51:56PM +0800, sunliming wrote:
Beau Belgrave beaub@linux.microsoft.com 于2023年6月17日周六 00:08写道：
On Fri, Jun 09, 2023 at 11:03:00AM +0800, sunliming wrote:
The writing operation return the count of writes whether events are enabled or disabled. This is incorrect when events are disabled. Fix this by just return -ENOENT when events are disabled.
When testing this patch locally I found that we would occasionally get -ENOENT when events were enabled, but then become disabled, since writes do not have any locking around the tracepoint checks for performance reasons.
I've asked a few peers of mine their thoughts on this, whether an error should result when there are no enabled events. The consensus I've heard back is that they would not consider this case an actual error, just as writing to /dev/null does not actually return an error.
However, if you feel strongly we need this and have a good use case, it seems better to enable this logic behind a flag instead of having it default based on my conversations with others.
Thanks, -Beau
There is indeed a problem. Once enabled, perform the write operation immediately.
The immediate write does work, and gets put into a buffer. The ftrace and perf self tests do the above case. So, no worries at this point.
Now，when the event is disabled, the trace record appears to be lost.
I'm taking this to mean, if in between the time of the bit check and the actual write() /writev() syscall the event becomes disabled, the event won't write to the buffer. Yes, that is expected.
Yes , got it, thank you for your explanation.
In some situations where data timing is sensitive, it may cause confusion. In this case, not returning an error (as mentioned in your reply, it is not considered this case an actual error) and returning 0 ( meaning that the number of data to be written is 0) may be a good way to handle it?
This is where I get a little lost. What would a user process do with a return of 0 bytes? It shouldn't retry, since it just hit that small timing window. In reality, it just incurred a temporary excessive syscall cost, but no real data loss (the operator/admin turned the event off).
I'm missing why you feel it's important the user process know such a window was hit?
Can you help me understand that?
I haven't encountered a specific scenario that it's important the user process know such a window was hit. This may be a mistake in my understanding. When someone uses user events checking the output of an event to confirm the execution status of a program, it may cause confusion if someone else prohibits the event. This shouldn't be a serious issue, this patch just makes things look better.
Thanks, -Sunliming
I do think returning 0 bytes is better than an error here, but I'd really like to know why the user process wants to know at all. Maybe they have user-space only logging and want to be able to mark there if it's in both spots (kernel and user buffers)?
Thanks, -Beau
Thanks, -Sunliming
Signed-off-by: sunliming sunliming@kylinos.cn
kernel/trace/trace_events_user.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/kernel/trace/trace_events_user.c b/kernel/trace/trace_events_user.c index 1ac5ba5685ed..92204bbe79da 100644 --- a/kernel/trace/trace_events_user.c +++ b/kernel/trace/trace_events_user.c @@ -1957,7 +1957,8 @@ static ssize_t user_events_write_core(struct file *file, struct iov_iter *i)
if (unlikely(faulted)) return -EFAULT;
}
} elsereturn -ENOENT; return ret;}
2.25.1
On Mon, 19 Jun 2023 11:40:44 -0700 Beau Belgrave beaub@linux.microsoft.com wrote:
Now，when the event is disabled, the trace record appears to be lost.
I'm taking this to mean, if in between the time of the bit check and the actual write() /writev() syscall the event becomes disabled, the event won't write to the buffer. Yes, that is expected.
In some situations where data timing is sensitive, it may cause confusion. In this case, not returning an error (as mentioned in your reply, it is not considered this case an actual error) and returning 0 ( meaning that the number of data to be written is 0) may be a good way to handle it?
This is where I get a little lost. What would a user process do with a return of 0 bytes? It shouldn't retry, since it just hit that small timing window. In reality, it just incurred a temporary excessive syscall cost, but no real data loss (the operator/admin turned the event off).
Note, this is similar to the race in the kernel with several tracing activities. If a disable happens and the buffer is now off, but the trace is still attempted, zero or NULL (depending on the function) is returned. This just means that tracing is off, and the event should just be dropped.
-- Steve
The user_event has not be enabled in write_fault test in ftrace self-test, Just enable it.
Signed-off-by: sunliming sunliming@kylinos.cn --- tools/testing/selftests/user_events/ftrace_test.c | 4 ++++ 1 file changed, 4 insertions(+)
diff --git a/tools/testing/selftests/user_events/ftrace_test.c b/tools/testing/selftests/user_events/ftrace_test.c index abfb49558a26..d33bd31425db 100644 --- a/tools/testing/selftests/user_events/ftrace_test.c +++ b/tools/testing/selftests/user_events/ftrace_test.c @@ -373,6 +373,10 @@ TEST_F(user, write_fault) { ASSERT_EQ(0, ioctl(self->data_fd, DIAG_IOCSREG, ®)); ASSERT_EQ(0, reg.write_index);
+ /* Enable event */ + self->enable_fd = open(enable_file, O_RDWR); + ASSERT_NE(-1, write(self->enable_fd, "1", sizeof("1"))) + /* Write should work normally */ ASSERT_NE(-1, writev(self->data_fd, (const struct iovec *)io, 2));
When user_events are disabled, it's write operation should be fail. Add this test cases.
Signed-off-by: sunliming sunliming@kylinos.cn --- tools/testing/selftests/user_events/ftrace_test.c | 4 ++++ 1 file changed, 4 insertions(+)
diff --git a/tools/testing/selftests/user_events/ftrace_test.c b/tools/testing/selftests/user_events/ftrace_test.c index d33bd31425db..cf52789bafda 100644 --- a/tools/testing/selftests/user_events/ftrace_test.c +++ b/tools/testing/selftests/user_events/ftrace_test.c @@ -297,6 +297,10 @@ TEST_F(user, write_events) { io[0].iov_base = ®.write_index; io[0].iov_len = sizeof(reg.write_index);
+ /* Write should fail when event is not enabled */ + ASSERT_EQ(-1, writev(self->data_fd, (const struct iovec *)io, 3)); + ASSERT_EQ(ENOENT, errno); + /* Enable event */ self->enable_fd = open(enable_file, O_RDWR); ASSERT_NE(-1, write(self->enable_fd, "1", sizeof("1")))
linux-kselftest-mirror@lists.linaro.org
-
Beau Belgrave -
Steven Rostedt -
sunliming -
sunliming