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

[Xen-devel] gnttab_map_refs and contiguous buffer


  • To: xen-devel <xen-devel@xxxxxxxxxxxxxxxxxxxx>
  • From: Oleksandr Andrushchenko <andr2000@xxxxxxxxx>
  • Date: Thu, 23 Mar 2017 17:40:49 +0200
  • Delivery-date: Thu, 23 Mar 2017 15:40:55 +0000
  • List-id: Xen developer discussion <xen-devel.lists.xen.org>

Hi, all!

I am trying to implement a zero-copy scenario for DRM front/back,
e.g. buffers allocated by DomU in the DRM frontend used as is
by Dom0. The requirement I have is that the buffer is contiguous.

So, what I currently have is:
1. DomU is 1:1 mapped and is able to allocate physically contiguous
buffer, e.g. PFNs and MFNs are sequential.
2. On Dom0 side I allocate ballooned pages (because I need reservation)
and am able to map those (gnttab_map_refs with grefs from DomU):
pfn 52a35 dev_bus_addr 6c100000
pfn 52a34 dev_bus_addr 6c101000
...
pfn 52a2c dev_bus_addr 6c109000

3. Then I have to create an SG table from these, so I can pass the buffer
to real HW DRM driver: as you can see from the above PFNs they are not
sequential as those were allocated from the balloon. Thus, sg_table
is also has number of segments != 1 which is not ok for my use-case.

Still, if I do __pfn_to_mfn(page_to_pfn(balloon_page[i])) I get the

correct MFN, which is expected because p2m translation happens.

4. If I try to do a hack:
dev_bus_pages[i] = pfn_to_page(MFN[i]) and then create an SG table from
these pages then of course the sg table is configuous and
I can share the SG with HW DRM driver.
What is naturally happens next is:
"Unable to handle kernel paging request at virtual address ffff80002c100000"
[   86.263197] PC is at __clean_dcache_area_poc+0x20/0x40
[   86.268377] LR is at __swiotlb_map_page+0x80/0x98
for the very first maddr == 6c100000, because my dev_bus_addr[i] and
dev_bus_pages[i] have no translation set up.

So, the question: is there any way I can make those ballooned pages
to convert into contiguous scatter-gather? So, sg table consists
of a single entry?

I was thinking of:
1. Is it possible to update translation manually so dev_bus_addr[i]
has corresponding continuous pages, e.g. so I can do
dev_bus_pages[i] = pfn_to_page(MFN[i]) and CPU can access those?
2. I can of course allocate contiguous buffer in Dom0 and
manually (unfortunately there is no API to do that) increase/decrease
reservation for the pages as balloon driver does and use those for mapping.
This way MFN will follow PFN (we are in Dom0 which is 1:1).
But this is going to be clumsy, as I'll have to copy-paste part
of the balloon driver into mine (decrease_reservation/increase_reservation)
3. Any other neat solution?

Thank you,
Oleksandr


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

 


Rackspace

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