On Wed, Jun 07, 2023 at 10:05:11AM -0400, Steven Sistare wrote: > On 2/8/2023 2:22 PM, Peter Xu wrote: > > On Wed, Feb 08, 2023 at 01:34:18PM -0500, Steven Sistare wrote: > >> On 2/7/2023 3:23 PM, Peter Xu wrote: > >>> On Tue, Feb 07, 2023 at 11:03:33AM -0800, Steve Sistare wrote: > >>>> migrate_ignore_shared() is an optimization that avoids copying memory > >>>> that is visible and can be mapped on the target. However, a > >>>> memory-backend-ram or a memory-backend-memfd block with the RAM_SHARED > >>>> flag set is not migrated when migrate_ignore_shared() is true. This is > >>>> wrong, because the block has no named backing store, and its contents > >>>> will > >>>> be lost. To fix, ignore shared memory iff it is a named file. Define a > >>>> new flag RAM_NAMED_FILE to distinguish this case. > >>> > >>> There's also TYPE_MEMORY_BACKEND_EPC. Reading the commit message it seems > >>> it can still be used in similar ways. Pasting commit message from > >>> c6c0232: > >>> > >>> Because of its unique requirements, Linux manages EPC separately from > >>> normal memory. Similar to memfd, the device /dev/sgx_vepc can be > >>> opened to obtain a file descriptor which can in turn be used to mmap() > >>> EPC memory. > >>> > >>> I'm not sure whether it means that should apply for RAM_NAMED_FILE too, > >>> neither do I think it's super important.. Still better to define it > >>> properly. > >> > >> The RAM_NAMED_FILE flag will be false for TYPE_MEMORY_BACKEND_EPC, so > >> ramblock_is_ignored will return false, and the contents will be copied > >> over the socket to the target, and the segment will be recreated there. > >> > >> However, perhaps I do not understand your point. > > > > My point was it looked like it should apply RAM_NAMED_FILE too, because > > it's also a named file. > > > > But.. I don't think another QEMU can share the same data if opening the > > same file. Based on my zero knowledge on it... I quickly looked up > > sgx_vepc_open() in the Linux impl where sgx_vepc.page_array plays a role of > > page cache iiuc, while it's private per vma even if VM_SHARED. > > > > So please ignore my comment.. > > > >> > >>> Another comment is, AFAIK this patch will modify senamtics of the old > >>> capability "x-ignore-shared". But I'd say in a sensible way. Maybe worth > >>> directly modify qapi/migration.json to reflect it (especially it's x- > >>> prefixed) to something like: > >>> > >>> # @x-ignore-shared: If enabled, QEMU will not migrate named shared memory > >>> # (since 4.0) > >> > >> Good idea. I propose: > >> > >> # @x-ignore-shared: If enabled, QEMU will not migrate shared memory that is > >> # accessible on the target. (since 4.0) > > > > I would use s/target/destination host/ because target can be misread as the > > VM itself (at least a general term in migration code base). Other than > > that looks good to me. > > Hi Peter, I will rebase to the tip, modify the wording, and repost. > Can I add your RB?
Yes. thanks, -- Peter Xu
