On Sat, Dec 28, 2019 at 10:13 PM Dan Williams <dan.j.williams@xxxxxxxxx> wrote: > > On Sat, Dec 28, 2019 at 12:54 PM Dan Williams > <dan.j.williams.korg@xxxxxxxxx> wrote: > > > > On Tue, Dec 3, 2019 at 11:53 PM Dave Young <dyoung@xxxxxxxxxx> wrote: > > > > > > Michael Weiser reported he got below error during a kexec rebooting: > > > esrt: Unsupported ESRT version 2904149718861218184. > > > > > > The ESRT memory stays in EFI boot services data, and it was reserved > > > in kernel via efi_mem_reserve(). The initial purpose of the reservation > > > is to reuse the EFI boot services data across kexec reboot. For example > > > the BGRT image data and some ESRT memory like Michael reported. > > > > > > But although the memory is reserved it is not updated in X86 e820 table. > > > And kexec_file_load iterate system ram in io resource list to find places > > > for kernel, initramfs and other stuff. In Michael's case the kexec loaded > > > initramfs overwritten the ESRT memory and then the failure happened. > > > > > > Since kexec_file_load depends on the e820 to be updated, just fix this > > > by updating the reserved EFI boot services memory as reserved type in e820. > > > > > > Originally any memory descriptors with EFI_MEMORY_RUNTIME attribute are > > > bypassed in the reservation code path because they are assumed as reserved. > > > But the reservation is still needed for multiple kexec reboot. > > > And it is the only possible case we come here thus just drop the code > > > chunk then everything works without side effects. > > > > > > On my machine the ESRT memory sits in an EFI runtime data range, it does > > > not trigger the problem, but I successfully tested with BGRT instead. > > > both kexec_load and kexec_file_load work and kdump works as well. > > > > > > Signed-off-by: Dave Young <dyoung@xxxxxxxxxx> > > > --- > > > arch/x86/platform/efi/quirks.c | 6 ++---- > > > 1 file changed, 2 insertions(+), 4 deletions(-) > > > > > > --- linux-x86.orig/arch/x86/platform/efi/quirks.c > > > +++ linux-x86/arch/x86/platform/efi/quirks.c > > > @@ -260,10 +260,6 @@ void __init efi_arch_mem_reserve(phys_ad > > > return; > > > } > > > > > > - /* No need to reserve regions that will never be freed. */ > > > - if (md.attribute & EFI_MEMORY_RUNTIME) > > > - return; > > > - > > > size += addr % EFI_PAGE_SIZE; > > > size = round_up(size, EFI_PAGE_SIZE); > > > addr = round_down(addr, EFI_PAGE_SIZE); > > > @@ -293,6 +289,8 @@ void __init efi_arch_mem_reserve(phys_ad > > > early_memunmap(new, new_size); > > > > > > efi_memmap_install(new_phys, num_entries); > > > + e820__range_update(addr, size, E820_TYPE_RAM, E820_TYPE_RESERVED); > > > + e820__update_table(e820_table); > > > } > > > > > > /* > > > > > > > Bisect says this change (commit af1648984828) is triggering a > > regression, likely not urgent, in my testing of the new efi_fake_mem= > > facility to allow memory to be marked "soft reserved" via the kernel > > command line (commit 199c84717612 x86/efi: Add efi_fake_mem support > > for EFI_MEMORY_SP). The following command line triggers the crash > > signature below: > > > > efi_fake_mem=4G@9G:0x40000,4G@13G:0x40000 > > > > However, this command line works ok: > > > > efi_fake_mem=8G@9G:0x40000 > > > > So, something about multiple efi_fake_mem statements interacts badly > > with this change. Nothing obvious occurs to me at the moment, I'll > > keep debugging, but wanted to highlight this in the meantime in case > > someone else sees a deeper issue or the root cause. > > Still looking, but this failure does not seem to be specific to the > "soft reservation" changes. Any update to the efi memmap that pushes > it over a page boundary triggers this failure. I.e. I can fix the > problem by over-allocating the efi memmap and then page aligning the > result. __early_ioremap "should" be handling this case, but it appears > something else is messing this up. Found it. Neither this patch nor the soft reservation changes are at fault, they are just helping to trigger a long standing bug in efi_fake_memmap(). Its usage of efi_memmap_split_count() can over count the number of splits needed for new entries. Consider the case of 2 contiguous fake entries intersecting the end of a single entry. The first call to efi_memmap_split_count() determines the resulting split will be (old1, new1, old2), the second call determines (old1, new2). The result is 2 splits when only 1 is needed to get a result of (old1, new1, new2) and the new map ends up with an empty entry. efi_memmap_install() interprets an empty entry as start = 0 end = 0xffffffffffffffff and attempts an extra split / copy past the end of the new map. I'll send a patch to fix up efi_fake_memmap().
