On Tue, Dec 12, 2017 at 07:20:31PM -0700, Jason Gunthorpe wrote: > On Tue, Dec 12, 2017 at 04:08:16PM -0800, Bryan Tan wrote: > > On Tue, Dec 12, 2017 at 11:55:30AM -0700, Jason Gunthorpe wrote: > > > Hm, it's still not clear to me what expectations are not followed here, > > and how you're differentiating 'refcount' from 'usecnt'. > > A refcount is something that has exactly one 'refcount_dec_and_test' > that then goes on to 'free' the thing being refcountedt. It never has > a wait queue. > > A usecnt is something that has an async path that waits for all users > to stop using it then 'frees' it. It always includes a wait > queue. I see, thanks for the explanation. > > And not that it makes it right, but I did notice that mlx4 uses the > > same pattern as what I've done here [1]. > > Ah, this is good, I was trying to find something like it. Kees > approves so I have no problem taking your patch. > > Though you might want to reflect on if using a completion is better > than a wait_event.. Not immediately clear to me. Oof, thanks for pointing that out. There's a problem here it seems: 0. refcount for a resource is 2 after entering an event handler 1. refcount_dec_and_test gets called by the resource destroy 2. refcount_dec_and_test gets called by the resource event handler 3. resource destroy call checks the condition in wait_event and proceeds to free the resource 4. resource event handler calls wake_up() on free'd resource I'll use a completion here instead. -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
