Hi guys, On 21/10/2019 10:53, Anshuman Khandual wrote: > On 10/18/2019 03:18 PM, Catalin Marinas wrote: >> On Fri, Oct 11, 2019 at 08:26:32AM +0530, Anshuman Khandual wrote: >>> On 10/10/2019 05:04 PM, Catalin Marinas wrote: >>>> Mark Rutland mentioned at some point that, as a preparatory patch to >>>> this series, we'd need to make sure we don't hot-remove memory already >>>> given to the kernel at boot. Any plans here? >>> >>> Hmm, this series just enables platform memory hot remove as required from >>> generic memory hotplug framework. The path here is triggered either from >>> remove_memory() or __remove_memory() which takes physical memory range >>> arguments like (nid, start, size) and do the needful. arch_remove_memory() >>> should never be required to test given memory range for anything including >>> being part of the boot memory. >> >> Assuming arch_remove_memory() doesn't (cannot) check, is there a risk on > > Platform can definitely enumerate boot memory ranges. But checking on it in > arch_remove_memory() which deals with actual procedural details might not be > ideal IMHO. Refusing a requested removal attempt should have been done up in > the call chain. This will require making generic hot plug reject any removal > request which falls within enumerated boot memory. IFAICS currently there is > no generic way to remember which memory came as part of the boot process. > Probably be a new MEMBLOCK flag will do. Memblock flags are fun because they have to be provided to the walkers like for_each_mem_range(). Unless hot remove is a hot path, it should be enough to check against the UEFI memory map or DT memory node. (we already have helpers to query the attributes from the memory map at runtime, so it is still available). >> arm64 that, for example, one removes memory available at boot and then >> kexecs a new kernel? Does the kexec tool present the new kernel with the >> original memory map? > I dont know, probably James can help here. But as I had mentioned earlier, > the callers of remove_memory() should be able to control that. ACPI should > definitely be aware about which ranges were part of boot memory and refrain > from removing any subset, if the platform is known to have problems with > any subsequent kexec operation because the way boot memory map get used. > > Though I am not much aware about kexec internals, it should inherit the > memory state at given point in time It does, but t = first-boot > accommodating all previous memory hot and remove operations. This would imply we rewrite the tables we get from firmware as the facts about the platform change ... that way madness lies! ACPI doesn't describe memory, the UEFI memory map does. You may be using the UEFI memory map on either a DT or ACPI system. If you don't have UEFI, you're using the DT memory-node. Linux passes on exactly what it had at boot through kexec. We don't rewrite the tables. Memory is either described in DT, or in the UEFI memory map that was left in memory by the EFI stub. Linux remembers where the UEFI memory map is through kexec using the additional entries in the DT chosen node that were put there by the EFI stub. The bootloader (including the EFI stub) needs to know what memory is removable. Certain allocations can't move once they have been made: * The kernel's randomised physical address should not be in removable memory. With UEFI, the EFI stub does this. * Firmware structures like the DT or ACPI tables should not be in removable memory. Neither should reservations for runtime use, like the RAS CPER regions, or the UEFI runtime services. * The EFI stub should not allocate the authoritative copy of the memory map in removable memory. (we have runtime helpers to lookup the attributes. we pass the boot-time memory map to the next OS via kexec). * During paging_init() we allocate memory for swapper_pg_dir. This isn't something we can easily move around. Its not just software!: * The GIC ITS property/pending (?) tables should not be in removable memory. The simplest thing to do here is decree that all memory present at boot, is non-removable. Firmware may need to trim the memory available to UEFI to the minimum needed to boot the system, we can hot-add the rest of it once we're up and running. > As an example cloud environment scenario, memory > resources might have increased or decreased during a guest lifetime, so > when the guest needs to have new OS image why should not it have all the > memory ? I dont know if it's feasible for the guest to expect previous hot > add or remove operations to be played again after the kexec. Firmware can't know that we kexec'd, so it can't replay the operations. I think we need a way of determining whether a particular block of removable memory is present or not. If we do this during boot, then kexec works in the same way as a normal boot. > There is another fundamental question here. Is there a notion of a minimum > subset of boot memory which cannot be hot removed no matter what ? If yes, > how that is being conveyed to the kernel currently ? Yes. The UEFI memory map. See drivers/firmware/efi/libstub/fdt.c::exit_boot_func() the EFI stub calls efi_get_virtmap() to get the running memory map, then stores in the DT with update_fdt_memmap(). The memory described at this stage may not be removed as allocations from the EFI stub can't be moved. The biggest of these, is the kernel, which relocates itself to a random physical address during the EFI stub. See drivers/firmware/efi/libstub/arm64-stub.c::handle_kernel_image() The memcpy() is at the end. > The point is that all these need to be established between ACPI, EFI and > kernel. AFAICS this problem is for MM subsystem (including the platform > part of it) to solve instead. >> I can see x86 has CONFIG_FIRMWARE_MEMMAP suggesting that it is used by >> kexec. try_remove_memory() calls firmware_map_remove() so maybe they >> solve this problem differently. >> >> Correspondingly, after an arch_add_memory(), do we want a kexec kernel >> to access it? x86 seems to use the firmware_map_add_hotplug() mechanism. > > Hmm, kexec could use it instead on arm64 as well ? Mmm, a linux specific description of the platform that we have to keep over kexec. How do we describe this if we kexec something that isn't linux? How do we tell a version of linux that doesn't support hotplug not to overwrite it? It would be better if we had something in ACPI to tell us at runtime whether a hot pluggable range of memory was populated. (I haven't looked to see whether ACPI can already do this) Thanks, James
