[Qemu-devel] [PATCH v3 0/2] pc: memhp: enforce gaps between DIMMs

[Igor Mammedov <imammedo@redhat.com>]

v3:
 -  use pc_machine_class_init() for default. (Eduardo)
 -  make 'gap' boolean and just add 1 byte gap
    when calculating new_addr for DIMM.
v2:
   make inter_dimm_gap a boolean and inster gap in 1 byte
   instead of 2Mb, due to alignment that gap would be extended
   up to natural backend alignment value.

it's a simplier way suggested by Michael S. Tsirkin
to workaround virtio bug reported earlier:
http://lists.nongnu.org/archive/html/qemu-devel/2015-08/msg00522.html
where virtio can't handle buffer that crosses border
between 2 DIMM's (i.e. 2 MemoryRegions).

idea is to leave gaps between DIMMs, making their GPAs
non contiguous, which effectively forces kmalloc
to not use DIMM if buffer doesn't fit inside of it.

Simpler reproducer:
 qemu-system-x86_64 -enable-kvm  -m 256,slots=250,maxmem=32G  -drive if=virtio,file=/dev/slow/rhel72  \
  `for i in $(seq 0 58); do echo -n "-object memory-backend-ram,id=m$i,size=10M -device pc-dimm,id=dimm$i,memdev=m$i "; done` \
   -nodefaults -snapshot -serial stdio -nographic -M pc-i440fx-2.4

if guest manages to boot just login and do:
  dd if=/dev/vda of=/dev/null bs=128M

it will crash QEMU in 99% cases with following message:
  qemu-system-x86_64: virtio: error trying to map MMIO memory

Tested vs master and mst/PCI trees.

Igor Mammedov (2):
  memhp: extend address auto assignment to support gaps
  pc: memhp: force gaps between DIMM's GPA

 hw/i386/pc.c             |  5 ++++-
 hw/i386/pc_piix.c        |  1 +
 hw/i386/pc_q35.c         |  1 +
 hw/mem/pc-dimm.c         | 15 +++++++++------
 hw/ppc/spapr.c           |  2 +-
 include/hw/i386/pc.h     |  1 +
 include/hw/mem/pc-dimm.h |  7 ++++---
 7 files changed, 21 insertions(+), 11 deletions(-)

-- 
1.8.3.1