Re: [Xen-devel] PV audio drivers for Linux

[Stefano Stabellini <sstabellini@xxxxxxxxxx>]

On Wed, 18 Jan 2017, Ughreja, Rakesh A wrote:
> >-----Original Message-----
> >From: Stefano Stabellini [mailto:sstabellini@xxxxxxxxxx]
> >Sent: Wednesday, January 18, 2017 5:41 AM
> >To: Ughreja, Rakesh A <rakesh.a.ughreja@xxxxxxxxx>
> >Cc: xen-devel@xxxxxxxxxxxxx; Oleksandr_Andrushchenko@xxxxxxxx;
> >Oleksandr_Grytsov@xxxxxxxx; oleksandr.dmytryshyn@xxxxxxxxxxxxxxx;
> >iurii.konovalenko@xxxxxxxxxxxxxxx; konrad.wilk@xxxxxxxxxx
> >Subject: Re: [Xen-devel] PV audio drivers for Linux
> >
> >On Tue, 17 Jan 2017, Ughreja, Rakesh A wrote:
> >> Hi,
> >>
> >> I am trying to develop PV audio drivers and facing one issue to
> >> achieve zero copy of the buffers between Front End (DOM1) and
> >> Back End (DOM0) drivers.
> >
> >You might want to take a look at the existing PV sound proposal:
> >
> >http://marc.info/?l=xen-devel&m=148094319010445
> >
> Sure, let me look into this.
> Thank you very much for the quick reply and the reference.
> 
> >
> >> When the buffer is allocated using __get_free_pages() on the DOM0
> >> OS, I am able to grant the access using gnttab_grant_foreign_access()
> >> to DOM1 as well as I am able to map it in the DOM1 virtual space
> >> using xenbus_map_ring_valloc().
> >>
> >> However the existing audio driver allocates buffer using
> >> dma_alloc_coherent(). In that case I am able to grant the access using
> >> gnttab_grant_foreign_access() to DOM1 but when I try to map in the
> >> DOM1 virtual space using xenbus_map_ring_valloc(), it returns an error.
> >>
> >> [1] Code returns from here.
> >>
> >> 507                         xenbus_dev_fatal(dev, map[i].status,
> >> 508                                          "mapping in shared page %d 
> >> from domain %d",
> >> 509                                          gnt_refs[i], 
> >> dev->otherend_id);
> >>
> >> gnttab_batch_map(map, i) is unable to map the page, but I am unable to
> >> understand why. May be its due to the difference in the way buffers
> >> are allocated dma_alloc_coherent() vs __get_free_pages().
> >>
> >> Since I don't want to touch existing audio driver, I need to figure out
> >> how to map buffer to DOM1 space with dma_alloc_coherent().
> >>
> >> Any pointers would be really helpful. Thank you in advance.
> >
> >Pages allocated by dma_alloc_coherent can be a bit special. Are you
> >going through the swiotlb-xen
> >(drivers/xen/swiotlb-xen.c:xen_swiotlb_alloc_coherent) in Dom0?
> >
> 
> No, I am not using this.

Keep in mind that when swiotlb-xen is used, it is transparent from the
device driver point of view. You might be using it without knowing. I
suggest you add a printk in there to be sure.


> Actually I am trying to reuse the existing
> HDA driver and just opening the ALSA streams at kernel level in
> PV Backend driver.
> 
> Buffers are allocated by the existing HDA driver.
> http://lxr.free-electrons.com/source/sound/core/memalloc.c#L83
> 
> >I would probably add a few printks to Xen in
> >xen/common/grant_table.c:do_grant_table_op to understand what is the
> >error exactly.
> 
> In the gnttab_retry_eagain_gop function, when it tries to
> get the status it always receives status as GNTST_eagain. After the retries
> the status is marked as GNTST_bad_page.
> 
> I am unable to figure out what properties of dma_alloc_coherent allocated
> buffers makes it un-mappable at Dom1.

Nothing comes to mind, but maybe the x86 maintainers (CC'ed) know.

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
https://lists.xen.org/xen-devel