On Tue, Sep 26, 2017 at 09:47:10AM -0500, Reza Arbab wrote: > On Tue, Sep 26, 2017 at 01:37:07PM +0000, Michal Hocko wrote: > > On Fri 22-09-17 15:13:56, Reza Arbab wrote: > > > The move_pages() syscall can be used to find the numa node where a page > > > currently resides. This is not working for device public memory pages, > > > which erroneously report -EFAULT (unmapped or zero page). > > > > > > Enable by adding a FOLL_DEVICE flag for follow_page(), which > > > move_pages() will use. This could be done unconditionally, but adding a > > > flag seems like a safer change. > > > > I do not understand purpose of this patch. What is the numa node of a > > device memory? > > Well, using hmm_devmem_pages_create() it is added to this node: > > nid = dev_to_node(device); > if (nid < 0) > nid = numa_mem_id(); > > I understand it's minimally useful information to userspace, but the memory > does have a nid and move_pages() is supposed to be able to return what that > is. I ran into this using a testcase which tries to verify that user > addresses were correctly migrated to coherent device memory. > > That said, I'm okay with dropping this if you don't think it's worthwhile. Just to add a data point, PCIE devices are tie to one CPU (architecturaly PCIE lane are connected to CPU at least on x86/ppc AFAIK) and thus to one numa node. Right now i am traveling but i want to check that this patch does not allow user to inadvertaly pin device memory page. I will look into it once i am back. Cheers, Jérôme -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>
