[Qemu-devel] [Bug 1497479] [NEW] memory corruption with migrate/savevm in TCG mode

[Pavel Boldin]

Public bug reported:

[ISSUE]

QEMU releases 2.3.1 and lower are forgetting to flush TLBs before
enabling the global dirty pages log and entering the final stage of
saving the VM.

[DESCRIPTION]

The situation is the following:
1. TLB misses is the only way for page dirtying in the TCG mode.
2. If TLB is hit by a running VM during the execution of the `ram_save_iterate' 
by migration thread (e.g. if VM is mostly idling) then some pages are missing 
in the dirty log.
3. These pages are then not migrated during `ram_save_complete'.
4. This makes memory content in a saved VM state differ from the actual VM 
memory.
5. If the affected area includes some Kernel data structures such as trees or 
lists this can cause Kernel to Oops after loading the saved state.

[SOLUTION]

A proposed solution is to flush TLB when `log_global_start' is called.
Here is the patch: 
https://bugs.launchpad.net/ubuntu/+source/qemu/+bug/1493049/+attachment/4459905/+files/tcg-commit-on-log-global-start.patch

[LINKS]

Ubuntu bug: https://bugs.launchpad.net/ubuntu/+source/qemu/+bug/1493049

** Affects: qemu
     Importance: Undecided
         Status: New

-- 
You received this bug notification because you are a member of qemu-
devel-ml, which is subscribed to QEMU.
https://bugs.launchpad.net/bugs/1497479

Title:
  memory corruption with migrate/savevm in TCG mode

Status in QEMU:
  New

Bug description:
  [ISSUE]

  QEMU releases 2.3.1 and lower are forgetting to flush TLBs before
  enabling the global dirty pages log and entering the final stage of
  saving the VM.

  [DESCRIPTION]

  The situation is the following:
  1. TLB misses is the only way for page dirtying in the TCG mode.
  2. If TLB is hit by a running VM during the execution of the 
`ram_save_iterate' by migration thread (e.g. if VM is mostly idling) then some 
pages are missing in the dirty log.
  3. These pages are then not migrated during `ram_save_complete'.
  4. This makes memory content in a saved VM state differ from the actual VM 
memory.
  5. If the affected area includes some Kernel data structures such as trees or 
lists this can cause Kernel to Oops after loading the saved state.

  [SOLUTION]

  A proposed solution is to flush TLB when `log_global_start' is called.
  Here is the patch: 
https://bugs.launchpad.net/ubuntu/+source/qemu/+bug/1493049/+attachment/4459905/+files/tcg-commit-on-log-global-start.patch

  [LINKS]

  Ubuntu bug:
  https://bugs.launchpad.net/ubuntu/+source/qemu/+bug/1493049

To manage notifications about this bug go to:
https://bugs.launchpad.net/qemu/+bug/1497479/+subscriptions