Messages in this thread | | | Date | Wed, 2 Jul 2008 09:56:04 +0200 | From | Haavard Skinnemoen <> | Subject | Re: [PATCH v4 2/6] dmaengine: Add dma_chan_is_in_use() function |
| |
"Dan Williams" <dan.j.williams@intel.com> wrote: > On Thu, Jun 26, 2008 at 6:23 AM, Haavard Skinnemoen > <haavard.skinnemoen@atmel.com> wrote: > > This moves the code checking if a DMA channel is in use from > > show_in_use() into an inline helper function, dma_is_in_use(). DMA > > controllers can use this in order to give clients exclusive access to > > channels (usually necessary when setting up slave DMA.) > > > > I have to admit that I don't really understand the channel refcounting > > logic at all... dma_chan_get() simply increments a per-cpu value. How > > can we be sure that whatever CPU calls dma_chan_is_in_use() sees the > > same value? > > As Chris noted in the comments at the top of dmaengine.c this is an > implementation Rusty's 'bigref'. It seeks to avoid the > cache-line-bouncing overhead of maintaining a single global refcount > in hot paths like tcp_v{4,6}_rcv(). When the channel is being > removed, a rare event, we transition to the accurate, yet slow, global > method.
Ok, I was sort of wondering what happens if you call dma_chan_get() on one cpu and dma_chan_put() on a different cpu later on. But it looks like when it really matters, the sum across all cpus is used, so the end result will be correct.
> Your observation is correct, dma_chan_is_in_use() may lie in the case > when the current cpu is not using the channel. For this particular > test I think you can look to see if this channel's resources are > already allocated. If they are then some other client got a hold of > this channel before the current attempt. Hmm... that would also > require that we free the channel's resources in the case where the > client replies with DMA_NAK, probably something we should do anyways.
Yes, I think that's good thing to do in general. In fact, I think the dw_dmac driver will waste a channel for each slave because it always assigns the channel to the client even if the client may NAK or DUP it later on. I haven't seen this actually happening because I only have one slave client at the moment.
Another reason to do this is to reclaim the memory used for descriptors. Currently, a channel that was NAK'ed or DUP'ed will still have a lot of preallocated descriptors, possibly with client-specific parameters already set up.
Haavard
| |