On Fri, 23 Jun 2017, Jarvis Roach wrote: > > -----Original Message----- > > From: Stefano Stabellini [mailto:sstabellini@kernel.org] > > Sent: Friday, June 23, 2017 2:21 PM > > To: Julien Grall > > Cc: Stefano Stabellini ; Zhongze Liu > > ; xen-devel@lists.xenproject.org; Wei Liu > > ; Ian Jackson ; Jarvis Roach > > ; edgari@xilinx.com; Edgar E. Iglesias > > > > Subject: Re: [RFC v2]Proposal to allow setting up shared memory areas > > between VMs from xl config file > > > > On Fri, 23 Jun 2017, Julien Grall wrote: > > > Hi, > > > > > > On 22/06/17 22:05, Stefano Stabellini wrote: > > > > > When we encounter an id IDx during "xl create": > > > > > > > > > > + If it’s not under /local/shared_mem: > > > > > + If the corresponding entry has a "master" tag, create the > > > > > corresponding entries for IDx in xenstore > > > > > + If there isn't a "master" tag, say error. > > > > > > > > > > + If it’s found under /local/shared_mem: > > > > > + If the corresponding entry has a "master" tag, say error > > > > > + If there isn't a "master" tag, map the pages to the newly > > > > > created domain, and add the current domain and necessary > > information > > > > > under /local/shared_mem/IDx/slaves. > > > > > > > > Aside from using "gfn" instead of gmfn everywhere, I think it looks > > > > pretty good. > > > > > > > > I would leave out permissions and cacheability attributes from this > > > > version of the work. I would just add a note saying that memory will > > > > be mapped as RW regular cacheable RAM. Other permissions and > > > > cacheability will be possible, but they are not implemented yet. > > > > > > Well, I think we should design the interface correctly from the > > > beginning to facilitate future extension. > > > > Which interface are you speaking about? > > > > I don't think we should attemp to write how the hypercall interface might > > look like in the future to support setting permissions and cacheability > > attributes. > > > > > > > Also, you need to clarify what you mean by "regular cacheable RAM". > > > Are they write-through, write-back...? But, on ARM, this would only be > > > the caching attribute in stage-2 page table. The final caching, memory > > > type, shareability would be a combination of stage-2 and stage-1 attributes. > > > > The very same that is used today for the ram of virtual machines, do we need > > to say any more than that? (For ARM, p2m_ram_rw and MATTR_MEM, > > LPAE_SH_INNER. For stage1, we should refer to > > xen/include/public/arch-arm.h.) > > I have customers who need some buffers LPAE_SH_OUTER and others who need NORMAL non-cacheable or inner-cacheable buffers, so my suggestion is to provide a way to support the full combination of configurations. > > While the stage 1/stage 2 combination results allow guests (via the stage 1 translation regime) to force the two combinations I specifically mentioned, in the first case the customers want LPAE_SH_OUTER for cache coherency with a DMA-capable I/O device. In that case, Xen needs to set the shareability attribute to OUTER in the stage 2 table since that's what is used for the SMMU. In the second case, NORMAL non-cacheable or inner-cacheable, the customers are in a position where they can't trust the guests to disable their cache or set it for inner-cacheable, so it would be good for a way to Xen or privileged/trusted domain to do so. Let me premise that I would be happy to see the whole set of configurations implemented in the long run, we might just not get there on day1. We could spec out how the VM config option should look like, but leave the cacheability and shareability parameteres unimplemented for now (also to address Julien't comment on defining future proof interfaces). I understand the need for cache-coherent buffers for dma to/from devices, but I think that problem should be solved with the iomem config option. This project was meant to setup shared memory regions for VM-to-VM communications. It doesn't look like that is the kind of requirement that this framework is meant to meet, unless I am missing something? Normal non-cacheable buffers are more interesting: do you actually see guests running on non-cacheable memory? If not, could you make an example of a use-case for two VMs sharing a non-cacheable page?