On Fri, Oct 31, 2014 at 11:44:52PM +0000, Geoff Levand wrote: > Hi, > > On Fri, 2014-10-24 at 13:27 +0100, Mark Rutland wrote: > > On Fri, Oct 24, 2014 at 11:59:38AM +0100, Grant Likely wrote: > > > On Fri, Oct 24, 2014 at 11:51 AM, Mark Rutland <mark.rutland at arm.com> wrote: > > > > On Fri, Oct 24, 2014 at 12:10:58AM +0100, Geoff Levand wrote: > > > >> Device tree regions described by /memreserve/ entries are not available in > > > >> /proc/device-tree, and hence are not available to user space utilities that use > > > >> /proc/device-tree. In order for these regions to be available, convert any > > > >> arm64 DTS files using /memreserve/ entries to use reserved-memory nodes. > > > > > > > > The limitation here is in the kernel (and a partially in userspace), so > > > > modifying the dts files is a workaround rather than a fix. > > > > > > > > It's perfectly valid for people to remain using /memreserve/, so this > > > > isn't sufficient. There are also existing DTBs using /memreserve/ which > > > > we can't rely on being modified to use reserved-memory. > > > > > > > > I think we need to expose memreserves to userspace somehow, potentially > > > > along with other DTB header fields. Grant, ideas? > > > > > > Yes, I suggested the same thing to Geoff in a separate thread. Here's > > > what I wrote: > > > > > > >> Geoff Levand wrote: > > > >>> I did some work on this and I think we just need to convert all the > > > >>> arm64 dts from /memreserve/ to reserved-memory nodes and update the > > > >>> arm64 booting.txt to specify using reserved-memory. I'll prepare a > > > >>> patch for it. > > > >> > > > >> I don't think that is going to be entirely sufficient. There will be > > > >> platforms that don't get converted over, and this is a generic problem > > > >> that covers all architectures using DT, not just aarch64. The solution > > > >> probably needs to include exposing the /memreserve/ sections to > > > >> userspace. I can see two ways to do this: > > > >> > > > >> 1) Create a new property in /sysfs with all the memreserve sctions > > > >> 2) Live-modify the device tree to put the memreserve data into a node > > > >> at boot time. > > > >> > > > >> Option 2 is probably the most generic solution, but it will require > > > >> some care to make sure there aren't any overlaps if a reserved-memory > > > >> node already exists. > > > > I would prefer the former currently. While I currently believe that > > modifying the tree is something we're going to have to do for stateful > > properties, it's not someting I want to have to do unless absolutely > > necessary. > > Current user space kexec utilities use /proc/device-tree and nothing > else. The intension of the device tree is to describe the system > sufficiently for a kernel to boot, so I think we should put the > /memreserve/ info into /proc/device-tree. I don't follow. The intention of device tree generically has precisely nothing to do with the Linux-specific filesystem hierarchy for exposing device tree to userspace. The memreserve entries aren't the same as other elements in /proc/device-tree. They're a flat table pointed to by the header rather than a hierarchy of nodes and properties. There's no way to expose that difference under /proc/device-tree, so exposing that as a parallel hierarchy makes the most sense to me. > We could put the /memreserve/ entries in there directly, or convert > to reserved-memory nodes. At the moment I like the idea to convert to > reserved-memory nodes. I think we need a parallel hierarchy to expose header information (though we likely only need the memreserve entries). I'm not keen on rewriting the device tree to work around a historical mistake. > > > > For reference, here is an old conversation about this exact thing. > > > > Reading through it, the opinions I expressed then don't necessarily > > > > match what I think now. I still don't think it is a good idea to > > > > expose the physical address of the old .dtb blob, but I do agree that > > > > the memreserve sections need to be exposed. > > > > > > > > https://lists.ozlabs.org/pipermail/linuxppc-dev/2010-July/083993.html > > > > Another option would be to expose the original DTB as a (read-only) > > binary file somewhere. That might interact poorly with live tree > > modification in future, however. > > I don't like the idea of having two interfaces to get essentially the same > info. I think it will be a maintenance problem over time. Given Grant's comments, I'm perfectly happy to not expose the raw DTB for situations otehr than debugging. Thanks, Mark.