Hi,
stress testing module loading and unloading on a derived scheduler, I
had crashes on cfq caused by what seemed to be an error in the caching of
cic lookup results in the ioc_data field of io contexts.
As what's happening is a little bit involved (at least for me), I've put
together some of the collected oopses with and without debug patches, the
patches themselves, a script used to reproduce the problem and the .config
used on kvm/qemu here, hoping that they can explain the problem better than
words:
http://feanor.sssup.it/~fabio/linux/cfq-ioc-data/
The patch below should fix the problem.
Subject: cfq-iosched: do not leak ioc_data across iosched switches
When switching scheduler from cfq, cfq_exit_queue() does not clear
ioc->ioc_data, leaving a dangling pointer that can deceive the following
lookups when the iosched is switched back to cfq. The pattern that can
trigger that is the following:
- elevator switch from cfq to something else;
- module unloading, with elv_unregister() that calls cfq_free_io_context()
on ioc freeing the cic (via the .trim op);
- module gets reloaded and the elevator switches back to cfq;
- reallocation of a cic at the same address as before (with a valid key).
To fix it just assign NULL to ioc_data in __cfq_exit_single_io_context(),
that is called from the regular exit path and from the elevator switching
code. The only path that frees a cic and is not covered is the error handling
one, but cic's freed in this way are never cached in ioc_data.
Signed-off-by: Fabio Checconi <fabio@gandalf.sssup.it>
---
diff --git a/block/cfq-iosched.c b/block/cfq-iosched.c
index 0f962ec..67cd023 100644
--- a/block/cfq-iosched.c
+++ b/block/cfq-iosched.c
@@ -1207,6 +1207,8 @@ static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
struct cfq_io_context *cic)
{
+ struct io_context *ioc = cic->ioc;
+
list_del_init(&cic->queue_list);
/*
@@ -1216,6 +1218,9 @@ static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
cic->dead_key = (unsigned long) cic->key;
cic->key = NULL;
+ if (ioc->ioc_data == cic)
+ rcu_assign_pointer(ioc->ioc_data, NULL);
+
if (cic->cfqq[ASYNC]) {
cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
cic->cfqq[ASYNC] = NULL;
@@ -1248,8 +1253,7 @@ static void cfq_exit_single_io_context(struct io_context *ioc,
*/
static void cfq_exit_io_context(struct io_context *ioc)
{
- rcu_assign_pointer(ioc->ioc_data, NULL);
call_for_each_cic(ioc, cfq_exit_single_io_context);
}
static struct cfq_io_context *
@@ -1480,8 +1485,7 @@ cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
spin_lock_irqsave(&ioc->lock, flags);
- if (ioc->ioc_data == cic)
- rcu_assign_pointer(ioc->ioc_data, NULL);
+ BUG_ON(ioc->ioc_data == cic);
radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
hlist_del_rcu(&cic->cic_list);
--
| H. Peter Anvin | Re: [RFC 00/15] x86_64: Optimize percpu accesses |
| Tarkan Erimer | Re: Dual-Licensing Linux Kernel with GPL V2 and GPL V3 |
| Eric W. Biederman | Remaining straight forward kthread API conversions... |
| Greg Kroah-Hartman | [PATCH 001/196] Chinese: Add the known_regression URI to the HOWTO |
| David Miller | [GIT]: Networking |
| Gerrit Renker | [PATCH 27/37] dccp: Integration of dynamic feature activation - part 2 (server side) |
| David Miller | Re: [PATCH] pkt_sched: Destroy gen estimators under rtnl_lock(). |
| Frans Pop | svc: failed to register lockdv1 RPC service (errno 97). |
git: | |
