lkml.org 
[lkml]   [2019]   [Jun]   [11]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 1/8] iommu: Add I/O ASID allocator
On Tue, 11 Jun 2019 15:37:42 +0100
Jean-Philippe Brucker <jean-philippe.brucker@arm.com> wrote:

> On 11/06/2019 13:26, Jacob Pan wrote:
> >> +/**
> >> + * ioasid_set_data - Set private data for an allocated ioasid
> >> + * @ioasid: the ID to set data
> >> + * @data: the private data
> >> + *
> >> + * For IOASID that is already allocated, private data can be set
> >> + * via this API. Future lookup can be done via ioasid_find.
> >> + */
> >> +int ioasid_set_data(ioasid_t ioasid, void *data)
> >> +{
> >> + struct ioasid_data *ioasid_data;
> >> + int ret = 0;
> >> +
> >> + xa_lock(&ioasid_xa);
> > Just wondering if this is necessary, since xa_load is under
> > rcu_read_lock and we are not changing anything internal to xa. For
> > custom allocator I still need to have the mutex against allocator
> > removal.
>
> I think we do need this because of a possible race with ioasid_free():
>
> CPU1 CPU2
> ioasid_free(ioasid) ioasid_set_data(ioasid, foo)
> data = xa_load(...)
> xa_erase(...)
> kfree_rcu(data) (no RCU lock held)
> ...free(data)
> data->private = foo;
>
make sense, thanks for explaining.

> The issue is theoretical at the moment because no users do this, but
> I'd be more comfortable taking the xa_lock, which prevents a
> concurrent xa_erase()+free(). (I commented on your v3 but you might
> have missed it)
>
Did you reply to my v3? I did not see it. I only saw your comments about
v3 in your commit message.

> >> + ioasid_data = xa_load(&ioasid_xa, ioasid);
> >> + if (ioasid_data)
> >> + rcu_assign_pointer(ioasid_data->private, data);
> > it is good to publish and have barrier here. But I just wonder even
> > for weakly ordered machine, this pointer update is quite far away
> > from its data update.
>
> I don't know, it could be right before calling ioasid_set_data():
>
> mydata = kzalloc(sizeof(*mydata));
> mydata->ops = &my_ops; (1)
> ioasid_set_data(ioasid, mydata);
> ... /* no write barrier here */
> data->private = mydata; (2)
>
> And then another thread calls ioasid_find():
>
> mydata = ioasid_find(ioasid);
> if (mydata)
> mydata->ops->do_something();
>
> On a weakly ordered machine, this thread could observe the pointer
> assignment (2) before the ops assignment (1), and dereference NULL.
> Using rcu_assign_pointer() should fix that
>
I agree it is better to have the barrier. Just thought there is already
a rcu_read_lock() in xa_load() in between. rcu_read_lock() may have
barrier in some case but better not count on it. No issues here. I will
integrate this in the next version.

> Thanks,
> Jean

[Jacob Pan]

\
 
 \ /
  Last update: 2019-06-11 19:08    [W:0.101 / U:1.544 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site