> From: Ian Campbell [mailto:ian.campb...@citrix.com] > Sent: Thursday, February 04, 2016 6:06 PM > > On Wed, 2016-02-03 at 17:41 +0000, George Dunlap wrote: > > But of course, since they they aren't actually ranges but just gpfns, > > they're scattered randomly throughout the guest physical address > > space. > > (Possibly) stupid question: > > Since, AIUI, the in-guest GPU driver is XenGT aware could it not allocate a > contiguous range of pages at start of day to use as GPU PTs? Or even just N > contiguous regions, e.g. i think the "8K" refers to pages, which is 16 2M > allocations, which is a far more manageable number of ranges to track than > 8096 individual pages. >
We add XenGT awareness in guest driver at minimum requirement (e.g. to handle address space ballooning due to graphics memory partitioning), which only impacts hardware specific initialization code (thinking vgpu as a new sku). However we're hesitating to touch other general driver code, such as allocation policy you mentioned earlier. Changing that part specifically for one sku needs very strong reason to convince driver maintainer. And we cannot assume above allocation policy can be always met. Thanks Kevin _______________________________________________ Xen-devel mailing list Xen-devel@lists.xen.org http://lists.xen.org/xen-devel