PATCH? exit_mm: update the comments, move atomic_inc(mm_count) outside of mmap_sem
(textually depends on
"[PATCH 2/2] coredump: exit_mm: clear ->mm first, then play with ->core_state")
More the "is my understanding correct" than a patch. Assuming it is correct:
- explain the subtle atomic_inc(&mm->mm_count).
- move it outside of mm->mmap_sem, otherwise it is not clear why do we need this lock to bump ->mm_count.
- "fix" the comment above task_lock(). This lock is really needed for get_task_mm/etc. Yes, we can change the code to use smp_mb()
+ spin_unlock_wait() but I guess we also have another reason.
Signed-off-by Oleg Nesterov <[email protected]>
kernel/exit.c renamed from kernel/exit.c~EXIT_MM +9 -3
@@ -680,6 +680,14 @@ static void exit_mm(struct task_struct *mm_release (tsk, mm);if (!mm)return ; + +/* +* We are going to clear ->mm. The next schedule() after that +* will assume we did context_switch()->atomic_inc(mm_count). +* finish_task_switch()->mmdrop() will "fix" the counter. +*/ +atomic_inc (&mm->mm_count); +BUG_ON (mm != tsk->active_mm);/* * Serialize with any possible pending coredump. * We must hold mmap_sem around checking core_state
@@ -689,9 +697,7 @@ static void exit_mm(struct task_struct **/ down_read (&mm->mmap_sem); core_state = mm->core_state; -atomic_inc (&mm->mm_count); -BUG_ON (mm != tsk->active_mm); -/* more a memory barrier than a real lock */ +/* disables preemption, we must not schedule before enter_lazy_tlb() */ task_lock (tsk); tsk->mm =NULL ;up_read (&mm->mmap_sem);