FDPIC: Add coredump capability for the ELF-FDPIC binfmt [try #4]
Quoted 11 lines · David Howells
(quoted diff) +7
-0+static int elf_fdpic_core_dump(long signr, struct pt_regs *regs, + struct file *file) +{ + read_lock(&tasklist_lock); + do_each_thread(g,p) + if (current->mm == p->mm && current != p) { + tmp = kzalloc(sizeof(*tmp), GFP_ATOMIC);
Review
Quoted 10 lines
(quoted diff) +10
-0+ if (!tmp) { + read_unlock(&tasklist_lock); + goto cleanup; + } + INIT_LIST_HEAD(&tmp->list); + tmp->thread = p; + list_add(&tmp->list, &thread_list); + } + while_each_thread(g,p); + read_unlock(&tasklist_lock);
Do you see any reason for tasklist_lock here (and in elf_core_dump) ?
do_each_thread() is rcu-safe, and all tasks which use this ->mm must sleep in wait_for_completion(&mm->core_done) at this point.
Oleg.
Quoted 4 lines · Oleg Nesterov
Do you see any reason for
tasklist_lockhere (and inelf_core_dump) ?
do_each_thread()is rcu-safe, and all tasks which use this ->mm must sleep inwait_for_completion(&mm->core_done)at this point.
Hmmm... do_each_thread() does not call rcu_read_lock/unlock(), but you may well be right. What about kernel threads running on another CPU with active_mm set to this mm (assuming I'm remembering correctly how that works)? I'm not sure they'd be a problem, though.
It does sound like you've got a valid point, though.
David
Quoted 10 lines · David Howells
Quoted 4 lines · Oleg Nesterov
Do you see any reason for
tasklist_lockhere (and inelf_core_dump) ?
do_each_thread()is rcu-safe, and all tasks which use this ->mm must sleep inwait_for_completion(&mm->core_done)at this point.Hmmm...
do_each_thread()does not call rcu_read_lock/unlock(), but you may well be right.
Yes, we can't just kill tasklist_lock, we should replace it with rcu_read_lock.
Quoted 3 lines
What about kernel threads running on another CPU with
active_mmset to this mm (assuming I'm remembering correctly how that works)? I'm not sure they'd be a problem, though.
Those threads have ->mm == NULL, this loop will not count them. But I think tasklist_lock can't make any difference for this issue. I am not sure I understand your concern, it is ok to use this ->mm (current uses it), and a kernel thread incremented ->mm_count.
Oleg.