In the work to remove proc_mnt I noticed that we were calling
proc_flush_task now proc_flush_pid possibly multiple times for the same
pid because of how de_thread works.
This is a bare minimal patchset to sort out de_thread, by introducing
exchange_tids and the helper of exchange_tids hlists_swap_heads_rcu.
The actual call of exchange_tids should be slowpath so I have
prioritized readability over getting every last drop of performance.
I have also read through a bunch of the code to see if I could find
anything that would be affected by this change. Users of
has_group_leader_pid were a good canidates. But I also looked at other
cases that might have a pid->task->pid transition. I ignored other
sources of races with de_thread and exec as those are preexisting.
I found a close call with send_signals user of task_active_pid_ns, but
all pids of a thread group are guaranteeds to be in the same pid
namespace so there is not a problem.
I found a few pieces of debugging code that do:
task = pid_task(pid, PIDTYPE_PID);
if (task) {
printk("%u\n", task->pid);
}
But I can't see how we care if it happens at the wrong moment that
task->pid might not match pid_nr(pid);
Similarly because the code in posix-cpu-timers goes pid->task->pid it
feels like there should be a problem. But as the code that works with
PIDTYPE_PID is only available within the thread group, and as de_thread
kills all of the other threads before it makes any changes of this
kind the race can not happen.
In short I don't think this change will introduce any regressions.
Eric W. Biederman (2):
rculist: Add hlists_swap_heads_rcu
proc: Ensure we see the exit of each process tid exactly once
fs/exec.c | 5 +----
include/linux/pid.h | 1 +
include/linux/rculist.h | 21 +++++++++++++++++++++
kernel/pid.c | 19 +++++++++++++++++++
4 files changed, 42 insertions(+), 4 deletions(-)
Eric
