- 1 System emulation
- 1.1 Incompatible changes
- 1.2 New deprecated options and features
- 1.3 Alpha
- 1.4 Arm
- 1.5 Cris
- 1.6 HPPA
- 1.7 M68k
- 1.8 Microblaze
- 1.9 MIPS
- 1.10 Nios2
- 1.11 OpenRISC
- 1.12 PowerPC
- 1.13 RISC-V
- 1.14 s390
- 1.15 SH
- 1.16 SPARC
- 1.17 TileGX
- 1.18 Tricore
- 1.19 x86
- 1.20 Xtensa
- 1.21 Device emulation and assignment
- 1.22 Audio
- 1.23 Character devices
- 1.24 Crypto subsystem
- 1.25 GUI
- 1.26 Host support
- 1.27 Memory backends
- 1.28 Monitor
- 1.29 Migration
- 1.30 Network
- 1.31 Block devices and tools
- 1.32 Tracing
- 1.33 Miscellaneous
- 2 User-mode emulation
- 3 TCG
- 4 Guest agent
- 5 Build Information
- 6 Known issues
- The "handle" option to -fsdev and -virtfs has been removed. The "local" or "proxy" options should be used instead.
- The "-virtioconsole" option has been removed. Use "-device virtconsole" instead.
- The "-no-frame" option has been removed. It was only usable with SDL1.2, and support for this library has been suspended now.
- The "-enable-hax" option has been removed. Use "-accel hax" instead.
- The "-clock" option has been removed. It was only a "dummy" option without meaning since QEMU 1.7, so there is no replacement.
- The legacy "ivshmem" device has been removed. Use "ivshmem-doorbell" or "ivshmem-plain" instead.
- The x86 machine types "pc-0.10" and "pc-0.11" have been removed. Use a newer machine type instead.
- The "irq" property of the "spapr-vscsi", "spapr-vlan" and "spapr-vty" devices has been removed with no replacement.
- The "memory-backend-memfd" backend object type will be reported as unavailable on hosts systems without memfd sealing support. On previous versions, "memory-backend-memfd" was reported as available, but didn't work properly without sealing support.
- HMP snapshot commands (such as "savevm", "loadvm" and "delvm") use only the snapshot tag, and not the ID any more, to identify snapshots. This removes ambiguity from the interface, but may require changing the HMP command in existing scripts.
New deprecated options and features
query-eventsQMP command has been superseded by the more powerful and accurate
acloption to the "-vnc" argument has been replaced by the
acl_removecommands are deprecated with no replacement. Authorization for VNC should be performed using the pluggable QAuthZ objects.
- QEMU_AUDIO_ environment variables and "-audio-help" are now deprecated. Use "-audiodev" instead.
Consult the "Deprecated Features" appendix for the full list of historically deprecated features/options.
- Updated PALcode firmware supports machine check exceptions.
- Implement the ARMv8.0-SB extension
- Implement the ARMv8.0-PredInv extension
- Implement the ARMv8.1-HPD extension
- Implement the ARMv8.1-LOR extension (as the trivial "no limited ordering regions provided" minimum)
- Implement the ARMv8.2-FHM extension
- Implement the ARMv8.2-AA32HPD extension
- Implement the ARMv8.3-PAuth extension
- Implement the ARMv8.3-JSConv extension
- Implement the ARMv8.4-CondM extension
- Implement the ARMv8.5-CondM extension
- Implement the ARMv8.5-FRINT extension
- Implement the Armv8.5-BTI extension for system emulation mode
- New machines "musca-a" and "musca-b1" -- these model the Arm "Musca" development boards
- New machine "mps2-an521" -- this is a model of the AN521 FPGA image for the MPS2 devboard
- Support TBI (top-byte-ignore) properly for linux-user mode
- The micro:bit board now boots simple MicroPython programs since device emulation for the timer, GPIO, NVMC and RNG has been added
- The cubieboard model now implements the 'A' SRAM
- AArch64 processors can now boot from a kernel placed over 4GB into RAM
- The stellaris boards ("lm3s6965evb", "lm3s811evb") now implement the watchdog timer device
- The BLK_MAX register in the TZ MPC device now reports the correct value
- The u-boot "noload" image type is now supported for the Arm virt board
- The Arm virt board now permits more than 255GB of RAM
- stm32f2xx_usart: Do not update data register when device is disabled
- virt board ACPI tables: COHACC override flag now correctly set in IORT SMMUv3 node
- AArch32 exception return is fixed to permit a switch from Mon->Hyp mode
- ftgmac100: implement the new MDIO interface on Aspeed SoC
- Emulation of the ARM PMU has been improved
- Fix various places where we failed to UNDEF invalid A64 instructions
- Don't UNDEF a valid FCMLA on 32-bit inputs
- Make FPSCR/FPCR trapped-exception bits RAZ/WI
- Fixed a bug in handling clearing of FPSCR/FPSR exception status bits
- armv7m_nvic: Allow byte accesses to SHPR1 register
- linux-user: support HWCAP_CPUID which exposes ID registers to user code
- Fixed a bug where the v8M MPU was always using the background region if it was enabled, rather than only if there was no hit on a more specific region
- Fixed bug where SVE ADDVL/ADDPL/RDVL instructions were missing the "is SVE enabled" check
- The raspi2 and raspi3 models now implement the "local timer" in the bcm2836_control block
- Fixed bug preventing EL0 from writing to architected timer registers
- Fixed SMMUv3 interrupt line numbers in ACPI tables for 'virt' board
- Fixed bug in emulation of CASP instruction
- check-tcg now builds and runs the cris tests using the fedora-cris-cross docker image
- Fix condition code generation for "add,<" & "add,<=", "addb".
- Fix the output for "dcor".
- Fix pci config address access.
- Fix the output for "b,gate".
- Map CPU HPA regions into PCI address space.
- Support TLB protection IDs
- Accept but ignore the diag instruction
- Add TLB trace events.
- Update seabios-hppa to upstream latest.
- The "mcf5208evb" machine now supports loading of firmware binaries with the "-bios" option.
- Added support for I7200 CPU (nanoMIPS32 ISA + DSP ASE; system mode only).
- Added support for I6500 CPU (MIPS64R6 ISA + MSA ASE + multicore features).
- Added support for QMP-based querying of the available CPU types.
- Added support for SAARI and SAAR configuration registers.
- Added support for MTTCG (multi-threaded TCG).
- Improved support for ITU (Interthread Communication Unit).
- Improved support for Fulong 2E machine.
- Improved end user documentation.
- Fixed build for MIPS n32 hosts.
- POWER9 large decrementer is now supported (pseries & powernv machines, TCG & KVM)
- sam460ex can now have 2GB memory
- pseries default CPU type is now POWER9
- pseries now supports the XIVE interrupt controller of POWER9 (emulated version only)
- pseries can now select its interrupt controller through a new ic-mode machine option
- pseries can now control the exposure of the host model and system-id through the new host-model and host-serial machine options
- pseries now supports "-vga cirrus"
- pseries now supports hot plug and unplug of PCI Host Bridges (PHBs)
- pseries now enables Spectre/Meltdown mitigations by default (this will cause warnings with TCG which doesn't have the mitigations implemented)
- pseries now supports the proposed count-cache-flush mitigation for Spectre (this requires guest co-operation, but should be lower cost than the current approaches)
- powernv now supports POWER9 cpus
- powernv now allows kernel images up to 256MiB
- gdbstub can now access SPRs
- mac99 machine now defaults to sungem NIC instead of ne2k_pci
- e500 now has a device for its I2C controller
- e500 has had an update to a newer U-Boot version
- ref405, sam460ex and taihu now have flash memory of fixed size matching the physical hardware instead of deriving it from -drive if-pflash
- The virt board now supports PCI and USB.
- The FS field of mstatus now supports three states (dirty, clean, and off).
- The TSR, TW, and TVM fields of mstatus are now implemented.
- The misa CSR is now writable.
- The built-in gdbserver supports register lists via XML files.
- The sifive_u machine supports SMP.
- The SiFive UART supports TX interrupts.
- The sifive_u machine has the correct number of PLIC interrupts.
- The 'zpci' feature bit is now indicated by default in the 'qemu' cpu model.
- The z14 cpu model now includes the multiple epoch and PTFF enhancement features per default.
- A cpu model for the z14 GA 2 has been added.
- vfio-ap now no longer inhibits usage of memory ballooners.
- zPCI devices now provide some instruction counters to the guest (for a Linux guest, check /sys/kernel/debug/pci/<function>/statistics).
- zPCI devices are now explicitly marked as unmigratable. No change in functionality, migration support for zPCI devices had never been implemented.
- vfio-ap now supports hot(un)plug of the vfio-ap device.
- Support for the floating-point extension facility has been added.
- Code for vector support instructions has been added.
- Machine r2d's flash memory size doubled to 16MiB to match physical hardware.
- Fix Solaris boot on SS-10/SS-20 machines with OpenBIOS (OpenBIOS)
- Fixed mixed up operands in CADDN and CADD
- The HAX accelerator is now supported for POSIX hosts other than Darwin, including Linux and NetBSD.
- Machines pc-* now support configuring firmware with -machine pflash0=ID0,pflash1=ID1 in addition to -drive if=pflash,... This permits use of -blockdev.
- MPX is considered a failed experiment by Intel, and has thus been removed from all named CPU models. It is still accessible via "-cpu host".
- PVH Linux images can be booted with "-kernel".
- xtfpga boards provide SMP support expected by linux (interrupt distributor, IPI and runstall)
- New test_mmuhifi_c3 core configuration capable of running SMP linux
- Flexible length instructions extension (FLIX) is now supported
- check-tcg now builds and runs xtensa system tests
Device emulation and assignment
- IDE/via: Implement PCI IDE mode
- virtio-blk: DISCARD and WRITE_ZEROES support
- smbus_eeprom: Will now transfer its state
- pm_smbus: I2C block transfers will now work properly
- pm_smbus: state transfer will now work properly
- pvrdma: Add support for RDMA MAD
- pvrdma: Removed the dev-caps-max-sge parameter
- Generic PCIe root port link speed and width enhancements: Starting with the Q35 QEMU 4.0 machine type, generic pcie-root-port will default to the maximum PCIe link speed (16GT/s) and width (x32) provided by the PCIe 4.0 specification. Experimental options x-speed= and x-width= are provided for custom tuning, but it is expected that the default over-provisioning of bandwidth is optimal for the vast majority of use cases. Previous machine versions and ioh3420 root ports will continue to default to 2.5GT/x1 links.
- Added the device_id property for SCSI disks that specifies which value to use for the vendor specific designator in the Device Identification VPD page
- Fixed erroneously detected multipath setup with multiple disks created with node-names (these would use an empty string as the vendor specific designator before, now the designator is left out)
- Several bugfixes in the LSI53C8xxA and ESP/PCscsi parallel SCSI adapters.
- ACPI HID for TPM TIS for TPM 2.0 has been corrected to MSFT0101; this should allow all Operating Systems to use the TPM 2.0 through the TIS interface
- QEMU support for TPM PPI (Physical Presence Interface) allows OS to use PPI functionality if the firmware supports it
- EDID interface for supported mdevs (Intel vGPU, host kernel v5.0+). Use options xres= and yres= to specify display resolution.
- New 'xen-disk' device which can create a Xen PV disk backend (instead of having to create it via xenstore).
- Improved performance of the Xen PV disk backend.
- Xen PV disk backend can now handle resize.
- The "wait" option for the socket backend is now forbidden for client sockets. It previously had no functional effect except for server sockets.
- The "reconnect" option for the socket backend is now forbidden for server sockets. It previously had no functional effect except for client sockets.
- The "tls-authz" option for the socket backend allows associating an authorization object with the character device to check against the TLS client's x509 certificate identity.
- The websocket protocol on socket character devices now correctly handles end of file when a client disconnects
- The block storage encryption backends are now capable of using multiple threads for encryption/decryption
-display spice-app: configure & launch a Spice client. With virt-viewer >= 8.0, this will present a UI similar to QEMU GTK (with monitor & console etc), but running in a separate process.
- Support for building against SDL1.2 has been deleted. SDL2 should be used instead.
- The VNC server will no longer accidentally delete its UNIX listener socket when clients disconnect
- The VNC server supports two new options "tls-authz" and "sasl-authz". These provide the ID of a previously create "authz" object to be used for authorization checks on incoming client identities.
- The VNC server "acl" option has been deprecated in favour of "tls-authz" and "sasl-authz"
- QMP can now execute a few commands "out of band". This is useful for postcopy recovery. For details, see docs/interop/qmp-spec.txt.
query-qmp-schemareflects QEMU's build configuration more closely. For instance, stuff related to replication is properly absent when QEMU was built with
--disable-replication, and stuff specific to other targets is properly absent.
- QMP events SHUTDOWN and RESET now carry a reason
- New QMP command
- QMP/HMP command
system_wakeupnow fails when the guest isn't suspended, or doesn't even support suspend.
- QMP/HMP command
cpu-addis now deprecated
- New QMP commands
block-dirty-bitmap-merge, plus enhancements to
block-dirty-bitmap-add. See Bitmaps & Incremental Backups for greater detail.
- The HMP commands
acl_removeare deprecated with no direct replacement. The new authorization framework provides an alternative object based approach to the authorization / access control problem, with pluggable implementations.
- New experimental QMP command
x-blockdev-reopenthat allows reopening any block device in use with a new set of options.
- free page hinting through virtio-balloon to avoid migrating unused pages
- ignore-shared feature for skipping shared memory blocks for migration-on-same-host hacks
- The new "tls-authz" migration parameter can be used on the incoming migration server, to specify an authorization policy to an check incoming client connections' TLS certificate. This ensures only the intended source host may initiate the incoming migration.
- new QMP/HMP command
announce_selfto trigger generation of broadcast RARP frames to update network switches.
- Slirp license has been clarified again as BSD-3, and the code has been adjusted to build as a standalone project (https://gitlab.freedesktop.org/slirp)
Block devices and tools
- Improved tracing and error diagnostics for NBD code
- Many fixes related to iothreads
- auto-read-only=on for the file driver switches dynamically between read-only and read-write file descriptors now, depending on whether writers (such as read-write guest devices, or block jobs that write to the node) are attached to the image. This allows starting without write permissions for backing files, but automatically making them writable e.g. for a commit block job.
- HMP snapshot commands use only the snapshot tag, and not the ID any more, to identify snapshots. This removes ambiguity from the interface, but may require changing the HMP command in existing scripts.
- The block-latency-histogram QMP command (used to be x-block-latency-histogram) is now considered stable and not marked as experimental any more
- Added the experimental x-blockdev-reopen QMP command to change options of block nodes after their creation
- NBD client code no longer attempts to access beyond end-of-file for a compliant server that advertised a size that is not a multiple of a sector, allowing for better interoperability with nbdkit
qemu-img map --output=humannow works for NBD disks
- NBD client code now gracefully works around several alignment compliance bugs present in the qemu 3.1 server code, rather than dropping the connection
Bitmaps & Incremental Backups
- Transactions now abort in reverse order, fixing crashes involving bitmap modifying commands
- Failed transactions involving disabled bitmaps should unwind cleanly
- Adding/Removing bitmaps to dataplane disks will no longer crash QEMU
- read-only bitmaps are now prohibited for use in incremental push backups instead of trying and failing at conclusion.
- busy bitmaps are now prohibited from being used as the source of x-block-dirty-bitmap-merge.
x-block-dirty-bitmap-mergenow accepts multiple source bitmaps.
- Stable QMP promotions:
- all matching transactions (-enable, -disable, -merge) are promoted as well.
BlockDirtyInfo, the query-block field
[i]/dirty-bitmaps[j]structure, has changed:
BlockDirtyInfo.statusis deprecated, to be removed in 3 releases.
BlockDirtyInfo.recordingis added. It's true when the bitmap is recording guest writes.
BlockDirtyInfo.persistentis added. It's true when the bitmap is either stored on-disk or scheduled to be written out to disk.
BlockDirtyInfo.busyis added. It's true when the bitmap is in-use by an operation.
BlockDirtyInfo.inconsistentis added. It's true when a persistent bitmap is no longer reliable.
BlockDirtyInfo.statuswill report "inconsistent" when the inconsistent bit is true.
- Pull Mode Incremental Backup:
block-dirty-bitmap-merge, plus enhancements to
nbd-server-addallow a management application to perform incremental backups with an NBD client as a consumer learning which portions of the disk were changed while the bitmap was enabled. The experimental commands
x-nbd-server-add-bitmapwere removed in favor of the stable commands.
- Bitmap checkpoints: The addition of
block-dirty-bitmap-disableallow for a management API to enable/disable bitmaps at critical moments (push, pull backups) to create discrete deltas per-bitmap. These deltas can be combined with
block-dirty-bitmap-mergeto create arbitrary differential-style backups including only the deltas chosen by the user.
- QEMU now allows resizing of qcow2 files with persistent bitmaps.
- QEMU will now open qcow2 files with improperly saved bitmaps, with status='inconsistent' and inconsistent=true.
qemu-img infowill show persistent dirty bitmaps
- QAPI documentation updates to reflect the new interface(s) described above.
- A new
qemu-nbd --bitmapoption allows the exposure of a persistent dirty bitmap for a qcow2 image not in use by a guest, in a simpler manner than the older procedure of using QMP commands to a temporary qemu process attached to the file.
- A new
qemu-nbd --listoption enables the ability to probe a remote NBD server for information about what it is exporting.
qemu-nbd --partition=Noption has been deprecated; it does not support GPT partitions, and has always been broken for MBR logical partition 6 and beyond. Its functionality of exporting a subset of the guest-visible data can still be accomplished with
- A new
qemu-nbd --tls-authzcommand line argument and counterpart tls-authz argument to QMP
nbd-server-startpermit an NBD server to fine-tune which clients may connect to a TLS session.
- Several bugs in qemu-nbd providing non-aligned responses to NBD_CMD_READ and NBD_CMD_BLOCK_STATUS have been fixed, although there are still some corner cases left if a backing file has smaller granularity than the active layer.
- dmg: lzfse compression support
- dmg: Fixed infinite loop
- qcow2: Support for external data files
- qcow2: The size of L2 cache entries is reduced to 4k by default to make better use of the memory and increase the I/O performance
- qcow2: Decompression is now delegated to worker threads, improving the read performance for compressed images
- qcow2: Include LUKS overhead in qemu-img measure output
- vmdk: Support for blockdev-create
- The new qemu-trace-stap script makes it convenient to collect traces without writing SystemTap scripts. See "man qemu-trace-stap" for details.
- QEMU's builtin gdbstub now supports the gdb multiprocess extension. Boards which have more than one cluster of CPUs (like the 'xlnx-zcu102' board when run with '-smp 6') will report them as being two processes (each of which has one thread per CPU in the cluster). You will need at least GDB 7.2. Attach to QEMU with a GDB command sequence like:
target extended :1234
- and then the "info threads" command should show 2 processes.
- QEMU will default to KVM when compiled with --disable-tcg or when launched from an executable whose name ends with "kvm".
- Common Python code now lives under "python", instead of under "scripts". That directory now contains a proper "qemu" Python module. qmp Python libraries, previously under "scripts/qmp" have also moved to the same location.
- The curses front-end now allows to type non-ascii letters. Emulating pressing escape now takes 25ms instead of 1s.
- The curses front-end now supports printing all non-ascii letters. The VGA font encoding can now be set with e.g. "-display curses,charset=CP850" . The default charset is CP437 (default VGA font).
- The sandbox seccomp filter for resource resource syscalls will no longer kill the QEMU process. It will report EPERM instead. This will avoid virgl dieing when using recent Mesa that tries to do CPU pinning.
- The docker.py script now checks the path and persistence status of binfmt_misc entries when installing and updating user-mode docker containers
- SoftFloat acceleration - where it is safe to do so, the softfloat helpers will use host floating-point instructions yielding significant performance increases
- Dynamic sizing of software TLBs, yielding noticable performance increases
- Host support for 64-bit RISC-V.
CI & Testing
- FreeBSD builds are now done via https://cirrus-ci.com/github/qemu/qemu
- Additional CI builds are now run if you host your repo on gitlab
- The check-tcg build framework can now be used to build system-mode/softmmu based tests
- QEMU now requires at least GCC version 4.8 or Clang version 3.4 (equivalent to Clang from Xcode 5.1 on macOS) to be built
- Running the QEMU testsuite now requires the Perl Test::Harness module. Most Linux and BSD distributions however install it by default together with Perl.
- For macOS hosts, our minimum supported version is now macOS 10.10. We have fixed the bug which meant that some guests would crash when running with the Cocoa UI on macOS 10.14 Mojave.
- We now have some documentation in rST format which we process with Sphinx. The host system must now have the 'build-sphinx' program available (version 1.3 or better) in order for any of the documentation to be built.
- see Planning/4.0