An rmdir pushes css's ref count to zero. However, if the associated
directory is open at the time, the dentry ref count is non-zero. If
the fd for this directory is then passed into perf_event_open, it
does a css_get(). This bounces the ref count back up from zero. This
is a problem by itself. But what makes it turn into a crash is the
fact that we end up doing an extra dput, since we perform a dput
when css_put sees the ref count go down to zero.
css_tryget() does not fall into that trap. So, we use that instead.
Reproduction test-case for the bug:
#include <unistd.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <linux/unistd.h>
#include <linux/perf_event.h>
#include <string.h>
#include <errno.h>
#include <stdio.h>
#define PERF_FLAG_PID_CGROUP (1U << 2)
int perf_event_open(struct perf_event_attr *hw_event_uptr,
pid_t pid, int cpu, int group_fd, unsigned long flags) {
return syscall(__NR_perf_event_open,hw_event_uptr, pid, cpu,
group_fd, flags);
}
/*
* Directly poke at the perf_event bug, since it's proving hard to repro
* depending on where in the kernel tree. what moved?
*/
int main(int argc, char **argv)
{
int fd;
struct perf_event_attr attr;
memset(&attr, 0, sizeof(attr));
attr.exclude_kernel = 1;
attr.size = sizeof(attr);
mkdir("/dev/cgroup/perf_event/blah", 0777);
fd = open("/dev/cgroup/perf_event/blah", O_RDONLY);
perror("open");
rmdir("/dev/cgroup/perf_event/blah");
sleep(2);
perf_event_open(&attr, fd, 0, -1, PERF_FLAG_PID_CGROUP);
perror("perf_event_open");
close(fd);
return 0;
}
Signed-off-by: Salman Qazi <sqazi@google.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Tejun Heo <tj@kernel.org>
Link: http://lkml.kernel.org/r/20120614223108.1025.2503.stgit@dungbeetle.mtv.corp.google.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>