[Bug 1989446] Re: [SRU] unable to boot guest with large memory when SEV is enabled on host

Tue Nov 29 08:31:51 UTC 2022

I have no plans for individual Backports as we plan to copy back 2.06
after the security update is out.

The security update is currently blocked by getting the 2022v1 signing
key installed in a PPA and having the previous security update copied
out of updates to security.

-- 
You received this bug notification because you are a member of Ubuntu
Foundations Bugs, which is subscribed to grub2-unsigned in Ubuntu.
https://bugs.launchpad.net/bugs/1989446

Title:
  [SRU] unable to boot guest with large memory when SEV is enabled on
  host

Status in grub2-unsigned package in Ubuntu:
  Fix Released
Status in grub2-unsigned source package in Focal:
  Triaged
Status in grub2-unsigned source package in Jammy:
  Triaged
Status in grub2-unsigned source package in Kinetic:
  Fix Committed

Bug description:
  [ Impact ]

  When booting a large memory guest (both focal and jammy) with 5.15 kernel on a SEV enabled host
  it fails to boot and shows the following error in dmesg:
  software IO TLB: Cannot allocate buffer

  But booting a Fedora36 guest works fine on a SEV enabled host

  With this kernel commit:
  https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=e998879d4fb7991856916972168cf27c0d86ed12

  SWIOTLB could allocate from 64MB to 1G top contiguous memory according
  to how much memory the system has

  in sev_setup_arch:

  size = total_mem * 6 / 100;
  size = clamp_val(size, IO_TLB_DEFAULT_SIZE, SZ_1G);
  swiotlb_adjust_size(size);

  Look into the memory block layout from Fedora grub, the available memory blocks are:
  [ 0.005879] memory[0x0] [0x0000000000001000-0x000000000009ffff], 0x000000000009f000 bytes flags: 0x0
  [ 0.005881] memory[0x1] [0x0000000000100000-0x000000007e9ecfff], 0x000000007e8ed000 bytes flags: 0x0
  [ 0.005883] memory[0x2] [0x000000007eb1b000-0x000000007fb9afff], 0x0000000001080000 bytes flags: 0x0
  [ 0.005885] memory[0x3] [0x000000007fbff000-0x000000007ffdffff], 0x00000000003e1000 bytes flags: 0x0
  [ 0.005886] memory[0x4] [0x0000000100000000-0x00000004ffffffff], 0x0000000400000000 bytes flags: 0x0

  The biggest one is:
  [ 0.005881] memory[0x1] [0x0000000000100000-0x000000007e9ecfff], 0x000000007e8ed000 bytes flags: 0x0

  The size is close to 2G and sufficient for SWIOTLB to allocate 1G
  contiguous memory

  Then we need to exclude reserved memory blocks overlapped with this region, below is the list
  [ 0.005892] reserved[0x2] [0x00000000574a7000-0x0000000059313fff], 0x0000000001e6d000 bytes flags: 0x0
  [ 0.005894] reserved[0x3] [0x000000007e133018-0x000000007e17e057], 0x000000000004b040 bytes flags: 0x0
  [ 0.005896] reserved[0x4] [0x000000007e845018-0x000000007e845857], 0x0000000000000840 bytes flags: 0x0
  [ 0.005897] reserved[0x5] [0x000000007ee95698-0x000000007ee95af7], 0x0000000000000460 bytes flags: 0x0

  Now the biggest available range is
  [0x0000000000100000-0x00000000574a7000]

  Before SWIOTLB allocates memory block, EFI also reserves some memory
  the one that overlapped with the above range is
  [ 0.005942] memblock_reserve: [0x000000007bfbe000-0x000000007bfddfff] efi_reserve_boot_services+0x8a/0xdb

  It’s fine that SWIOTLB can still allocate 1G contiguous memory from [0x0000000000100000-0x00000000574a7000]:
  [ 1.089832] software IO TLB: mapped [mem 0x00000000174a7000-0x00000000574a7000] (1024MB)

  But if we look into the memory block layout from Ubuntu grub, the available memory blocks are:
  [ 0.005833] memory[0x0] [0x0000000000001000-0x000000000009ffff], 0x000000000009f000 bytes flags: 0x0
  [ 0.005835] memory[0x1] [0x0000000000100000-0x000000007e9ecfff], 0x000000007e8ed000 bytes flags: 0x0
  [ 0.005837] memory[0x2] [0x000000007eb1b000-0x000000007fb9afff], 0x0000000001080000 bytes flags: 0x0
  [ 0.005838] memory[0x3] [0x000000007fbff000-0x000000007ffdffff], 0x00000000003e1000 bytes flags: 0x0
  [ 0.005840] memory[0x4] [0x0000000100000000-0x00000004ffffffff], 0x0000000400000000 bytes flags: 0x0

  The biggest one is also:
  [ 0.005835] memory[0x1] [0x0000000000100000-0x000000007e9ecfff], 0x000000007e8ed000 bytes flags: 0x0

  Then excluding the reserved memory blocks:
  [ 0.005846] reserved[0x2] [0x000000003a9ba000-0x000000003c7cdfff], 0x0000000001e14000 bytes flags: 0x0
  [ 0.005848] reserved[0x3] [0x000000007e133018-0x000000007e17e057], 0x000000000004b040 bytes flags: 0x0
  [ 0.005849] reserved[0x4] [0x000000007e847018-0x000000007e847887], 0x0000000000000870 bytes flags: 0x0
  [ 0.005851] reserved[0x5] [0x000000007ee95698-0x000000007ee95af7], 0x0000000000000460 bytes flags: 0x0

  Now the biggest one is:
  [0x000000003e133018-0x000000007e133018]

  Then excluding EFI reserved memory block that overlapped with the above range:
  [ 0.005896] memblock_reserve: [0x000000007bfbe000-0x000000007bfddfff] efi_reserve_boot_services+0x8a/0xdb

  So now, the biggest contiguous memory becomes
  [0x000000003c7ce000-0x000000007bfbe000]

  Which is less than 1G, this is why SWIOTLB can not allocate 1G
  contiguous memory

  This commit from rhboot/grub2 fixes this issue:
  https://github.com/rhboot/grub2/commit/9e6c1d803ade111b8719502ff25e86d8b4564de8

  it adjusts the memory block layout, so SWIOTLB or any other drivers
  that need more than 1G contiguous memory can be satisfied

  [ Test Plan ]

  Enable SEV on a AMD machine, refer to
  https://docs.ovh.com/us/en/dedicated/enable-and-use-amd-sme-
  sev/#references-and-additional-resources_1

  create a ubuntu VM with SEV enabled (--launchSecurity sev) and 18G memory as below:
  virt-install --name <guest-name> --memory 18874368 --memtune hard_limit=36507216 --boot uefi --disk /var/lib/libvirt/images/<guest-name.img>,device=disk,bus=scsi --disk /var/lib/libvirt/images/<guest-name>-config.iso,device=cdrom --os-type linux --os-variant <variant> --import --controller type=scsi,model=virtio-scsi,driver.iommu=on --controller type=virtio-serial,driver.iommu=on --network network=default,model=virtio,driver.iommu=on --memballoon driver.iommu=on --graphics none --launchSecurity sev --noautoconsole

  Make sure the running kernel in VM is 5.15

  Then check if it can boot successfully with the above patch
  dmesg should show SWIOTLB is correctly mapped to 1G memory
  [    0.005713] software IO TLB: SWIOTLB bounce buffer size adjusted to 1024MB
  [    0.821746] PCI-DMA: Using software bounce buffering for IO (SWIOTLB)
  [    0.822210] software IO TLB: mapped [mem 0x0000000014fb4000-0x0000000054fb4000] (1024MB)
  [    0.933346] software IO TLB: Memory encryption is active and system is using DMA bounce buffers

  [ Where problems could occur ]

  Originally, allocate memory for initrd/params/cmdline/kernel tries to
  get a memory block from less than 0x3fffffff, if it can not find a
  contiguous memory block, it will fail

  This patch only adjusts the memory allocation that
  1. check if it can get memory from less than 0x7fffffff (GRUB_EFI_MAX_ALLOCATION_ADDRESS)
  2. if step 1 fails, then check if it can get memory from less than 0xffffffffffffffff (GRUB_EFI_MAX_USABLE_ADDRESS)

  With this patch, initrd/params/cmdline/kernel will be located in
  higher address, e.g. between 0x3fffffff - 0x7fffffff, so it gives more
  room for drivers like SWIOTLB to allocate a larger memory, so it
  shouldn't affect other functions

  The only issue is that if initrd is too big that needs to be allocated from higher than 4G (GRUB_EFI_MAX_USABLE_ADDRESS), but even without this patch, this issue still exists, because the original policy is to allocate memory from less than 2G (0x3fffffff)
  But this issue is being handled by lp:1842320

  [ Other Info ]

  Related bugs:
  lp:1983625
  lp:1842320

To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/grub2-unsigned/+bug/1989446/+subscriptions