[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] [PATCH v5 1/4] VT-d PI: track the number of vcpus on pi blocking list

>>> On 01.09.17 at 09:55, <chao.gao@xxxxxxxxx> wrote:
> On Fri, Sep 01, 2017 at 02:24:08AM -0600, Jan Beulich wrote:
>>>>> On 01.09.17 at 03:39, <chao.gao@xxxxxxxxx> wrote:
>>> After thinking it again, I want to define the counter as
>>> a unsigned int variable for the following reasion:
>>> 1. It is definite that the counter is closely related with
>>> list_add() and list_del(). If the list is protected by the
>>> lock, it is straightforward that the counter is also protected
>>> by the lock.
>>> 2. In patch 3, althought there are some lock-less readers, we
>>> will check the counter still meets our requirement with the lock
>>> held. Thus, I don't think there is a racing issue.
>>I think that's fine, but then you still don't need LOCKed accesses
>>to the counter for updating it; write_atomic() will suffice afaict.
> A stupid question.
> Is it contradictory that you think the counter can be protected by
> the lock while suggesting using write_atomic() instead of LOCKed
> accesses?
> updating the counter is always accompanied by updating list and updating
> list should in locked region. I meaned things like:
> spin_lock()
> list_add()
> counter++
> spin_unlock()
> However, I am afraid that not using LOCKed accesses but using
> write_atomic() means something like (separating updating the counter
> from updating the list I think is not good):
> spin_lock()
> list_add()
> spin_unlock()
> write_atomic()

No, I mean


whereas ...

> And I think this version is:
> spin_lock()
> list_add()
> add_sized()
> spin_unlock()

... this produces a needless LOCKed instruction redundant with being
inside the locked region).


Xen-devel mailing list



Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.