padata: avoid race in reordering
authorJason A. Donenfeld <Jason@zx2c4.com>
Thu, 23 Mar 2017 11:24:43 +0000 (12:24 +0100)
committerWilly Tarreau <w@1wt.eu>
Tue, 20 Jun 2017 12:04:40 +0000 (14:04 +0200)
commit de5540d088fe97ad583cc7d396586437b32149a5 upstream.

Under extremely heavy uses of padata, crashes occur, and with list
debugging turned on, this happens instead:

[87487.298728] WARNING: CPU: 1 PID: 882 at lib/list_debug.c:33
__list_add+0xae/0x130
[87487.301868] list_add corruption. prev->next should be next
(ffffb17abfc043d0), but was ffff8dba70872c80. (prev=ffff8dba70872b00).
[87487.339011]  [<ffffffff9a53d075>] dump_stack+0x68/0xa3
[87487.342198]  [<ffffffff99e119a1>] ? console_unlock+0x281/0x6d0
[87487.345364]  [<ffffffff99d6b91f>] __warn+0xff/0x140
[87487.348513]  [<ffffffff99d6b9aa>] warn_slowpath_fmt+0x4a/0x50
[87487.351659]  [<ffffffff9a58b5de>] __list_add+0xae/0x130
[87487.354772]  [<ffffffff9add5094>] ? _raw_spin_lock+0x64/0x70
[87487.357915]  [<ffffffff99eefd66>] padata_reorder+0x1e6/0x420
[87487.361084]  [<ffffffff99ef0055>] padata_do_serial+0xa5/0x120

padata_reorder calls list_add_tail with the list to which its adding
locked, which seems correct:

spin_lock(&squeue->serial.lock);
list_add_tail(&padata->list, &squeue->serial.list);
spin_unlock(&squeue->serial.lock);

This therefore leaves only place where such inconsistency could occur:
if padata->list is added at the same time on two different threads.
This pdata pointer comes from the function call to
padata_get_next(pd), which has in it the following block:

next_queue = per_cpu_ptr(pd->pqueue, cpu);
padata = NULL;
reorder = &next_queue->reorder;
if (!list_empty(&reorder->list)) {
       padata = list_entry(reorder->list.next,
                           struct padata_priv, list);
       spin_lock(&reorder->lock);
       list_del_init(&padata->list);
       atomic_dec(&pd->reorder_objects);
       spin_unlock(&reorder->lock);

       pd->processed++;

       goto out;
}
out:
return padata;

I strongly suspect that the problem here is that two threads can race
on reorder list. Even though the deletion is locked, call to
list_entry is not locked, which means it's feasible that two threads
pick up the same padata object and subsequently call list_add_tail on
them at the same time. The fix is thus be hoist that lock outside of
that block.

Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Acked-by: Steffen Klassert <steffen.klassert@secunet.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Willy Tarreau <w@1wt.eu>
kernel/padata.c

index 072f4ee4eb893a5341942146775c2ce9fcc1dbc8..0925ccf92c7fefbb2e2523cae9381db3407cd693 100644 (file)
@@ -190,19 +190,20 @@ static struct padata_priv *padata_get_next(struct parallel_data *pd)
 
        reorder = &next_queue->reorder;
 
+       spin_lock(&reorder->lock);
        if (!list_empty(&reorder->list)) {
                padata = list_entry(reorder->list.next,
                                    struct padata_priv, list);
 
-               spin_lock(&reorder->lock);
                list_del_init(&padata->list);
                atomic_dec(&pd->reorder_objects);
-               spin_unlock(&reorder->lock);
 
                pd->processed++;
 
+               spin_unlock(&reorder->lock);
                goto out;
        }
+       spin_unlock(&reorder->lock);
 
        if (__this_cpu_read(pd->pqueue->cpu_index) == next_queue->cpu_index) {
                padata = ERR_PTR(-ENODATA);