Re: [PATCH] core dump: remain dumpable

[Oleg Nesterov <oleg@tv-sign.ru>]

On 10/10, Roland McGrath wrote:
> 
> As far as I know, set_dumpable was only used in do_coredump for
> synchronization and not intended for any security purpose.

I don't understand why do_coredump() clears DUMPABLE too.

> This changes do_coredump to use a separate bit for its
> synchronization, so the "dumpable" bits remain the same.
> ...
>
> @@ -1727,7 +1727,8 @@ int do_coredump(long signr, int exit_code, struct pt_regs * regs)
>  	if (!binfmt || !binfmt->core_dump)
>  		goto fail;
>  	down_write(&mm->mmap_sem);
> -	if (!get_dumpable(mm)) {
> +	if (!get_dumpable(mm) ||
> +	    test_and_set_bit(MMF_DUMP_STARTED, &mm->flags)) {
>  		up_write(&mm->mmap_sem);
>  		goto fail;

Do we really need the new MMF_DUMP_STARTED flag? We are holding ->mmap_sem,
can't we check "mm->core_waiters != 0" instead?

Hmm. Actually, do we need any check? If another thread (or CLONE_VM task)
already started do_coredump(), we must see SIGNAL_GROUP_EXIT when we take
->mmap_sem. In that case coredump_wait() does nothing but returns -EAGAIN.

> -	set_dumpable(mm, 0);

Looks like this change is all we need, no? The only problem is that we
return -EAGAIN if we race with another thread (the current code returns 0),
but nobody checks the returned value.

What do you think?

Oleg.