The problem is if you are not holding the spinlock, some parallel thread might have added a structure into the list that makes our no longer needed. So upon spinlock reacquisition we'll need to do another lookup, find the now added structure and throw ours away - even more expensive.
On Jun 17, 2016, at 10:32 AM, Matthew Jacob wrote: > At the risk of sounding annoying, the problem here isn't in the locking but > in the flow. > If the structure already exists, spinlock is fine. Return after dropping > spinlock. > > If the structure doesn't exist, you can drop the spinlock to create it (and > initialize the lock &and& grab the lock, should you want to) *prior* to > adding it to the list because nobody else could know about it until you put > it on the list. > > On Fri, Jun 17, 2016 at 7:24 AM, Oleg Drokin <[email protected]> wrote: > > On Jun 17, 2016, at 10:19 AM, Peter Zijlstra wrote: > > > On Fri, Jun 17, 2016 at 10:14:10AM -0400, Oleg Drokin wrote: > >> > >> On Jun 17, 2016, at 4:25 AM, Peter Zijlstra wrote: > >> > >>> On Wed, Jun 15, 2016 at 02:23:35PM -0400, Oleg Drokin wrote: > >>>> Hello! > >>>> > >>>> To my surprise I found out that it's not possible to initialise a mutex > >>>> into > >>>> a locked state. > >>>> I discussed it with Arjan and apparently there's no fundamental reason > >>>> not to allow this. > >>> > >>> There is. A mutex _must_ have an owner. If you can initialize it in > >>> locked state, you could do so statically, ie. outside of the context of > >>> a task. > >> > >> What's wrong with disallowing only static initializers, but allowing > >> dynamic ones? > >> Then there is a clear owner. > > > > At which point, what wrong with the simple: > > > > mutex_init(&m); > > mutex_lock(&m); > > > > Sequence? Its obvious, has clear semantics and doesn't extend the API. > > The problem is: > > spin_lock(somelock); > structure = some_internal_list_lookup(list); > if (structure) > goto out; > > init_new_structure(new_structure); > mutex_init(&new_structure->s_mutex); > mutex_lock(&new_structure->s_mutex); // XXX CANNOT DO THIS UNDER SPINLOCK! > > list_add(list, new_structure->s_list); > structure = new_structure; > out: > spin_unlock(somelock); > return structure; > >
