Hi!
On Sat 15-02-20 21:54:08, Yufen Yu wrote:
> On 2020/2/14 22:05, Tejun Heo wrote:
> > Hello,
> >
> > On Fri, Feb 14, 2020 at 10:50:01AM +0800, Yufen Yu wrote:
> > > > So, unregistering can leave ->dev along and re-registering can test
> > > > whether it's NULL and if not put the existing one and put a new one
> > > > there. Wouldn't that work?
> > >
> > > Do you mean set bdi->dev as 'NULL' in call_rcu() callback function
> > > (i.e. bdi_release_device()) and test 'bdi->dev' in bdi_register_va()?
> > >
> > > I think that may do not work.
> > > We cannot make sure the order of rcu callback function and re-registering.
> > > Then bdi_release_device() may put the new allocated device by re-registering.
> >
> > No, I meant not freeing bdi->dev on deregistration and only doing so
> > when it actually needs to - on re-registration or release. So, sth
> > like the following.
> >
> > * Unregister: Unregister bdi->dev but don't free it. Leave the pointer
> > alone.
> >
> > * Re-register: If bdi->dev is not null, initiate RCU-free and update
> > bdi->dev to the new dev.
> >
> > * Release: If bdi->dev is not NULL, initiate RCU-free of it.
>
> Okay, I think we can do that.
>
> When do re-register, we need to update bdi->dev as 'NULL' or the new dev
> before invoking call_rcu() to free '->dev'. So readers started after call_rcu()
> cannot read the old dev, like:
>
> bdi_register_va()
> {
> ...
> if (bdi->dev) {
> //rcu_assgin_pointer(bdi->dev, new_dev);
> rcu_assgin_pointer(bdi->dev, NULL);
> call_rcu();//rcu callback function will free the old dev
> }
> ...
> }
>
> After assigning new value for bdi->dev, rcu callback function cannot get
> the old device. So I think we may need to replace the '->dev' with a new
> struct pointer, in which includes '->dev' and 'rcu_head'. We pass the
> struct.rcu_head pointer to call_rcu() and then it can get old dev address.
>
> IMO, maybe we can maintain the original code logic, fix the problem like:
I've now noticed there's commit 68f23b8906 "memcg: fix a crash in wb_workfn
when a device disappears" from end of January which tries to address the
issue you're looking into. Now AFAIU the code is till somewhat racy after
that commit so I wanted to mention this mostly so that you fixup also the
new bdi_dev_name() while you're fixing blkg_dev_name().
Also I was wondering about one thing: If we really care about bdi->dev only
for the name, won't we be much better off with just copying the name to
bdi->name on registration? Sure it would consume a bit of memory for the
name copy but I don't think we really care and things would be IMO *much*
simpler that way... Yufen, Tejun, what do you think?
Honza
>
> diff --git a/include/linux/backing-dev-defs.h b/include/linux/backing-dev-defs.h
> index 4fc87dee005a..e2de4a4e5392 100644
> --- a/include/linux/backing-dev-defs.h
> +++ b/include/linux/backing-dev-defs.h
> @@ -185,6 +185,11 @@ struct bdi_writeback {
> #endif
> };
>
> +struct bdi_rcu_device {
> + struct device *dev;
> + struct rcu_head rcu_head;
> +};
> +
> struct backing_dev_info {
> u64 id;
> struct rb_node rb_node; /* keyed by ->id */
> @@ -219,7 +224,7 @@ struct backing_dev_info {
> #endif
> wait_queue_head_t wb_waitq;
>
> - struct device *dev;
> + struct bdi_rcu_device *rcu_dev;
> struct device *owner;
>
> struct timer_list laptop_mode_wb_timer;
> diff --git a/mm/backing-dev.c b/mm/backing-dev.c
> index 62f05f605fb5..05f07ce19091 100644
> --- a/mm/backing-dev.c
> +++ b/mm/backing-dev.c
> @@ -850,7 +850,7 @@ static int bdi_init(struct backing_dev_info *bdi)
> {
> int ret;
>
> - bdi->dev = NULL;
> + bdi->rcu_dev = NULL;
>
> kref_init(&bdi->refcnt);
> bdi->min_ratio = 0;
> @@ -932,20 +932,28 @@ struct backing_dev_info *bdi_get_by_id(u64 id)
>
> int bdi_register_va(struct backing_dev_info *bdi, const char *fmt, va_list args)
> {
> - struct device *dev;
> struct rb_node *parent, **p;
> + struct bdi_rcu_device *rcu_dev;
>
> - if (bdi->dev) /* The driver needs to use separate queues per device */
> + /* The driver needs to use separate queues per device */
> + if (bdi->rcu_dev)
> return 0;
>
> - dev = device_create_vargs(bdi_class, NULL, MKDEV(0, 0), bdi, fmt, args);
> - if (IS_ERR(dev))
> - return PTR_ERR(dev);
> + rcu_dev = kzalloc(sizeof(struct bdi_rcu_device), GFP_KERNEL);
> + if (!rcu_dev)
> + return -ENOMEM;
> +
> + rcu_dev->dev = device_create_vargs(bdi_class, NULL, MKDEV(0, 0),
> + bdi, fmt, args);
> + if (IS_ERR(rcu_dev->dev)) {
> + kfree(rcu_dev);
> + return PTR_ERR(rcu_dev->dev);
> + }
>
> cgwb_bdi_register(bdi);
> - bdi->dev = dev;
> + bdi->rcu_dev = rcu_dev;
>
> - bdi_debug_register(bdi, dev_name(dev));
> + bdi_debug_register(bdi, dev_name(rcu_dev->dev));
> set_bit(WB_registered, &bdi->wb.state);
>
> spin_lock_bh(&bdi_lock);
> @@ -1005,17 +1013,28 @@ static void bdi_remove_from_list(struct backing_dev_info *bdi)
> synchronize_rcu_expedited();
> }
>
> +static void bdi_put_device_rcu(struct rcu_head *rcu)
> +{
> + struct bdi_rcu_device *rcu_dev = container_of(rcu,
> + struct bdi_rcu_device, rcu_head);
> + put_device(rcu_dev->dev);
> + kfree(rcu_dev);
> +}
> +
> void bdi_unregister(struct backing_dev_info *bdi)
> {
> + struct bdi_rcu_device *rcu_dev = bdi->rcu_dev;
> /* make sure nobody finds us on the bdi_list anymore */
> bdi_remove_from_list(bdi);
> wb_shutdown(&bdi->wb);
> cgwb_bdi_unregister(bdi);
>
> - if (bdi->dev) {
> + if (rcu_dev) {
> bdi_debug_unregister(bdi);
> - device_unregister(bdi->dev);
> - bdi->dev = NULL;
> + get_device(rcu_dev->dev);
> + device_unregister(rcu_dev->dev);
> + rcu_assign_pointer(bdi->rcu_dev, NULL);
> + call_rcu(&rcu_dev->rcu_head, bdi_put_device_rcu);
> }
>
> if (bdi->owner) {
> @@ -1031,7 +1050,7 @@ static void release_bdi(struct kref *ref)
>
> if (test_bit(WB_registered, &bdi->wb.state))
> bdi_unregister(bdi);
> - WARN_ON_ONCE(bdi->dev);
> + WARN_ON_ONCE(bdi->rcu_dev);
> wb_exit(&bdi->wb);
> cgwb_bdi_exit(bdi);
> kfree(bdi);
>
>
>
>
>
>
>
--
Jan Kara <jack@xxxxxxxx>
SUSE Labs, CR
