Re: [PATCH v4] [RFC] trace: Add kprobe on tracepoint

[Masami Hiramatsu <mhiramat@kernel.org>]

On Thu, 12 Aug 2021 18:44:29 +0900
Masami Hiramatsu <mhiramat@kernel.org> wrote:

> On Wed, 11 Aug 2021 23:46:48 -0400
> Steven Rostedt <rostedt@goodmis.org> wrote:
> 
> > On Thu, 12 Aug 2021 10:27:35 +0900
> > Masami Hiramatsu <mhiramat@kernel.org> wrote:
> > 
> > > Let me confirm this, so eprobes can be attached to synthetic event?
> > > IMHO, I rather like to prevent attaching eprobe_event on the other
> > > dynamic events. It makes hard to check when removing the base dynamic
> > > events...
> > > 
> > > For the above example, we can rewrite it as below to trace filename
> > > without attaching eprobe_events on the synthetic event.
> > > 
> > >   echo 'my_open pid_t pid; char file[]' > synthetic_events
> > > 
> > >   echo 'e:myopen syscalls.sys_enter_open file=+0($filename):ustring' > dynamic_events
> > >   echo 'e:myopen_ret syscalls.sys_exit_open ret=$ret' > dynamic_events
> > >  
> > >   echo 'hist:keys=common_pid:fname=file' > events/eprobes/myopen/trigger
> > >   echo 'hist:keys=common_pid:fname=$fname:onmatch(eprobes.myopen).trace(my_open,common_pid,$fname)' > events/eprobes/myopen_ret
> > > 
> > 
> > The problem is that the above wont work :-(
> > 
> > For example, I can use this program:
> > 
> > #include <stdio.h>
> > #include <unistd.h>
> > #include <fcntl.h>
> > #include <sys/types.h>
> > 
> > static const char *file = "/etc/passwd";
> > 
> > int main (int argc, char **argv)
> > {
> > 	int fd;
> > 
> > 	fd = open(file, O_RDONLY);
> > 	if (fd < 0)
> > 		perror(file);
> > 	close(fd);
> > 	return 0;
> > }
> > 
> > Which if you do the above, all you'll get from the myopen is "(null)".
> > 
> > That's because the "/etc/passwd" is not paged in at the start of the
> > system call, and because tracepoints can not fault, the "ustring" will
> > not be mapped yet, it can not give you the content of the file pointer.
> > This was the entire reason we are working on eprobes to attach to
> > synthetic events in the first place.
> 
> I think that is another limitation. If you run this program,
> 
> static const char *file = "/etc/passwd";
> 
> int main (int argc, char **argv)
> {
> 	char buf[BUFSIZE];
> 	int fd;
> 
> 	strlcpy(buf, file, BUFSIZE);
> 	fd = open(buf, O_RDONLY);
> 	if (fd < 0)
> 		perror(file);
> 	read(fd, buf, BUFSIZE);
> 	close(fd);
> 	return 0;
> }
> 
> you'll not see any filename from the "myopen_ret" or the synthetic event.
> Thus, the user-space page fault must be handled by the other way. (e.g.
> making a special worker thread and run it before the task returns to
> user space.)
> Using eprobe over synthetic event does not solve the root cause (and
> it can introduce another issue.)

Oops, I missed that is the exit of open(), not close(). OK so filename 
should be accessible at that point.

> > 
> > The trick is to use the synthetic event to pass the filename pointer to
> > the exit of the system call, which the system call itself would map the
> > pointer to "file", and when the eprobe reads it with ":ustring" from
> > the exit of the system call it gets "/etc/passwd" instead of "(null)".
> > 
> > Your above example doesn't fix this.

OK, I got it.

Thanks, 


-- 
Masami Hiramatsu <mhiramat@kernel.org>