Jump to navigation Jump to search

System emulation

Incompatible changes

  • The number of allow PCI host bridges for pSeries machine was reduced from 256 to 31 (more can be configured by setting up MMIO windows manually).
  • Removed support for tftp:// in the block layer, since this has been broken forever for files bigger than 256KB.

Future incompatible changes

  • Three options are using different names on the command line and in configuration file. In particular:
    • The "acpi" configuration file section matches command-line option "acpitable";
    • The "boot-opts" configuration file section matches command-line option "boot";
    • The "smp-opts" configuration file section matches command-line option "smp".
-readconfig will standardize on the name for the command line option.
  • Behavior of automatic calculation of SMP topology when some SMP topology options for -smp are omitted (sockets, cores, threads) will change in the future. If guest ABI needs to be preserved on upgrades while using the SMP topology options, users should either set set all options explicitly (sockets, cores, threads), or omit all of them.
  • Devices "allwinner-a10", "pc87312", "ssi-sd" will be configured with explicit properties instead of implicitly. This is unlikely to affect users.
  • QMP command blockdev-add is still a work in progress. It doesn't support all block drivers, it lacks a matching blockdev-del, and more. It might change incompatibly.
  • For x86, specifying a CPUID feature with both "+feature/-feature" and "feature=on/off" will cause a warning. The current behavior for this combination ("+feature/-feature" wins over "feature=on/off") will be changed so that "+feature" and "-feature" will be synonyms for "feature=on" and "feature=off" respectively).


  • Improvements to the Aspeed board.
  • Support for HLT semihosting traps in AArch32 mode (both ARM and Thumb).
  • The ACPI tables for the "virt" machine type support ITS.
  • The Cadence GEM device now supports multiple priority queues through the num-priority-queues property.
  • The STM32F2xx board (Netduino 2) now includes ADC and SPI devices.



  • Support for 24KEc CPUs.


  • Support for POWER9 CPUs.
  • Improvements for the new "powernv" platform.


  • PCI host bridges can be associated to NUMA nodes.
  • Support for more than 1 TiB of guest memory.
  • Support for more than 64 GiB of MMIO window in a PCI host bridge.
  • Support for the "-prom-env" parameter


  • Support for CPU models.
  • Support for virtio-ccw revision 2.






  • Support for several new CPUID features related to AVX-512 instruction set extensions.
  • The emulated IOAPIC (used by TCG and, with KVM, if the "-machine kernel_irqchip" option has the value "off" or "split") now defaults to version 0x20, which supports directed end-of-interrupt messages.
  • Support for Extended Interrupt Mode (EIM) in the intel_iommu device. EIM requires KVM (Linux v4.7 or newer, for x2APIC support) and "-machine kernel-irqchip=split"; it is enabled automatically if interrupt remapping is enabled ("-machine kernel-irqchip=split -device intel_iommu,intremap=on").
  • Support for up to 288 CPUs with the Q35 machine types. 256 or more CPUs are only supported if IOMMU and EIM are enabled.


  • Support for unplugging SCSI disk.
  • Support for SUSE xenlinux-compatible device unplug.

Device emulation and assignment

  • QEMU now includes a generic loader pseudo-device that lets you load multiple images or values into memory at startup. This device is documented in docs/generic-loader.txt.


  • Support for hotplugging of NVDIMM devices (_FIT)

Block devices

Network devices

  • Support for fault tolerance based on coarse-grained lock stepping (COLO).






  • New device vhost-vsock.
  • Initial support for graceful handling of guest errors (i.e. QEMU should not exit on guest errors).
  • Support for new virtio-crypto device.


  • Support for grant copy.

Character devices

Crypto subsystem

  • Support for more hash algorithms for PBKDF.
  • Support for CTR mode.


  • SPICE can use pure OpenGL rendering if "gl=on" is specified.



  • Support for fault tolerance based on coarse-grained lock stepping (COLO).


Block devices and tools

  • More QMP commands support node-name (block-stream, block-commit, blockdev-backup, blockdev-mirror, blockdev-snapshot-delete-internal-sync, blockdev-snapshot-internal-sync, change-backing-file, drive-backup, drive-mirror, nbd-server-add).
  • The BLOCK_IO_ERROR event now includes the node name.
  • More QMP commands accept device model names (block_set_io_throttle, blockdev-change-medium, eject, x-blockdev-remove-medium, x-blockdev-insert-medium, blockdev-open-tray, blockdev-close-tray)
  • The DEVICE_TRAY_MOVED event now includes the device id.
  • Throttling now applies to the guest device only, and not to block jobs or the NBD server.
  • drive-backup and blockdev-backup support writing out backups in compressed format.
  • The LUKS format now can configure the PBKDF iteration count.
  • block-stream supports streaming from a backing file to another backing file.
  • Support for replication, for coarse-grained lock stepping (COLO) fault tolerance.
  • New "dd" subcomamand of qemu-img.
  • The DMG driver can be compiled to a separate driver, so as to make QEMU's dependency on libbz2 optional.
  • Support for iSER in QEMU's iSCSI initiator through a iser:// URI.
  • The NBD client and server support the NBD_CMD_WRITE_ZEROES extension.
  • Raw images support "offset" and "size" options to access only a part of the file or device.


  • New tracing backend "syslog".
  • Support for multiple "-d trace:PATTERN" command-line arguments.

CLI options

User-mode emulation

Removed target support

  • The unicore32-linux-user target implemented a different system call ABI from mainline Linux for this architecture. Support for it has been dropped.

New functionality

  • Added support for more syscalls including preadv, pwritev, syslog.
  • Major scalability improvements for multi-threaded programs (ARM, SPARC, x86).
  • QEMU can now understand and generate fence and cmpxchg operations.


  • New TCG primitives have been added for safely modelling architectural synchronisation instructions (e.g. atomics, LL/SC, LOCK prefixes). arm, aarch64, alpha and x86 targets now use these primitives for multi-threaded linux-user programs. TCG target maintainers are encouraged to port their front-ends to use the new facilities.
  • The TCG backends now emit appropriate barrier instructions for frontend barriers when running multi-threaded programs. However, emulating a strongly-ordered architecture (e.g., x86) on a weakly-ordered one (e.g., ARM or POWER) will not work yet.
  • tb_flush() is finally thread-safe meaning multi-threaded programs are less likely to crash when the translation buffer is reset
  • lock contention in the main cpu run-loop has been reduced improving performance for multi-threaded code
  • a number of races were identified and fixed

A lot of the TCG work merged in this cycle where prerequisites for supporting multi-threaded system emulation (MTTCG). While full MTTCG support is expected to be merged in the next development cycle, multi-threaded linux-user programs will already benefit from this work.

Build Information


  • The qemu-tech manual has been merged with qemu-doc.

Build dependencies

  • QEMU does not depend anymore on libuuid.

Docker Build Targets

  • A new "travis" target has been added to help debugging Travis CI failures
  • The docker targets can now be run against a specified docker tag, so:
   make docker-run TEST="test-quick" IMAGE="debian:arm64" EXECUTABLE=./aarch64-linux-user/qemu-aarch64

Allows you to run the normal docker build test against a binfmt_misc powered image you have built yourself.

Known issues

  • postcopy migration can't be used with KVM PR on POWER (KVM HV and TCG work well). This is a kernel issue (USERFAULTFD).
  • QEMU for Windows won't work when it was compiled with SDL 2.0 support (terminates with "Fatal Error: Out of memory - aborting").
  • x86 vapic with TCG mode may lead to double-free due to a race condition, use -global apic-common.vapic=off as a work around. This is an existing bug, info here.