Re: ipc,sem: sysv semaphore scalability

[Linus Torvalds <torvalds@linux-foundation.org>]

On Fri, Mar 29, 2013 at 12:06 PM, Dave Jones <davej@redhat.com> wrote:
>
> Here's an oops I just hit..
>
> BUG: unable to handle kernel NULL pointer dereference at 000000000000000f
> IP: [<ffffffff812c24ca>] testmsg.isra.5+0x1a/0x60

Btw, looking at the code leading up to this, what the f*ck is wrong
with the IPC stuff?

It's using the generic list stuff for most of the lists, but then it
open-codes the accesses.

So instead of using

   for_each_entry(walk_msg, &msq->q_messages, m_list) {
      ..
   }

the ipc/msg.c code does all that by hand, with

   tmp = msq->q_messages.next;
   while (tmp != &msq->q_messages) {
      struct msg_msg *walk_msg;

      walk_msg = list_entry(tmp, struct msg_msg, m_list);
      ...
      tmp = tmp->next;
   }

Ugh. The code is near unreadable. And then it has magic memory
barriers etc, implying that it doesn't lock the data structures, but
no comments about them. See expunge_all() and pipelined_send().

The code seems entirely random, and it's badly set up (annoyance of
the day: crazy helper functions in ipc/msgutil.c to make sure that (a)
you have to spend more effort looking for them, and (b) they won't get
inlined).

Clearly nobody has cared for the crazy IPC message code in a long time.

              Linus